Computational Toxicology: Methods and Protocols

Methods in
Molecular Biology 1800
Orazio Nicolotti Editor
Computational
Toxicology
Methods and Protocols
Methods in M o l e c u l a r B i o lo g y
Series Editor
John M. Walker
School of Life and Medical Sciences
University of Hertfordshire
Hatfield, Hertfordshire, AL10 9AB, UK
For further volumes:

http://www.springer.com/series/7651
Computational Toxicology
Methods and Protocols
Edited by
Orazio Nicolotti
Dipartimento di Farmacia-Scienze del Farmaco,
Università degli Studi di Bari Aldo Moro, Bari, Italy
Editor
Orazio Nicolotti
Dipartimento di Farmacia-Scienze del Farmaco
Università degli Studi di Bari Aldo Moro
Bari, Italy
ISSN 1064-3745 ISSN 1940-6029 (electronic)

Methods in Molecular Biology
ISBN 978-1-4939-7898-4 ISBN 978-1-4939-7899-1 (eBook)
https://doi.org/10.1007/978-1-4939-7899-1
Library of Congress Control Number: 2018943843
© Springer Science+Business Media, LLC, part of Springer Nature 2018

This work is subject to copyright. All rights are reserved by the Publisher, whether the whole or part of the material is
concerned, specifically the rights of translation, reprinting, reuse of illustrations, recitation, broadcasting, reproduction
on microfilms or in any other physical way, and transmission or information storage and retrieval, electronic adaptation,
computer software, or by similar or dissimilar methodology now known or hereafter developed.
The use of general descriptive names, registered names, trademarks, service marks, etc. in this publication does not imply,
even in the absence of a specific statement, that such names are exempt from the relevant protective laws and regulations
and therefore free for general use.
The publisher, the authors and the editors are safe to assume that the advice and information in this book are believed to
be true and accurate at the date of publication. Neither the publisher nor the authors or the editors give a warranty,
express or implied, with respect to the material contained herein or for any errors or omissions that may have been made.
The publisher remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.
Printed on acid-free paper
This Humana Press imprint is published by the registered company Springer Science+Business Media, LLC part of
Springer Nature.
The registered company address is: 233 Spring Street, New York, NY 10013, U.S.A.
Dedication
To Chiara and Maria Giovanna
v
Preface
I first heard about computational toxicology about ten years ago when I was asked to apply
elsewhere modeling predictive strategies, typical of rational drug design. This was the way
I started this challenging journey. Walking toward this emerging field was quite natural
although goals and objectives were rather different. At present, I would say that computa-
tional toxicology is a blend of knowledge embracing medicinal chemistry, pharmacology,
organic chemistry, biochemistry, clinical and forensic medicine, and other attractive disci-
plines. Anyhow, I would say that computing has evolved toxicology from a mostly empirical
level based on disease-specific observational measures to target-specific models aimed at
properly predicting the risk/benefit ratio of chemicals.
This book comprises excellent contributions from colleagues working all over the
world. Its structure reflects my personal experience, initiated from medicinal chemistry.
Chapters 1–4 provide a comprehensive view of molecular descriptors and QSAR, the fasci-
nating root from which all comes from. Molecular and data modeling methods needed to
comply both scientific and regulatory sides are discussed in Chapters 5–10 while the rele-
vance of computational toxicology in drug discovery is mostly highlighted in Chapters
11–19. The last part, including Chapters 20–27, explains how to predict some relevant
human-health toxicology endpoints.
The book collects methods and protocols currently used in computational toxicology
by sharing the vision of top scientists in the understanding of solid target-specific models.
Last but not least, the ultimate aim of the book is to arouse the curiosity and interest of the
reader.
Bari, Italy Orazio Nicolotti
vii
Contents
Dedication. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . v
Preface. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . vii
Contributors . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . xiii
Part I Where We Are and Where We Are Going To

1 Molecular Descriptors for Structure–Activity Applications:
A Hands-On Approach�� 3
Francesca Grisoni, Davide Ballabio, Roberto Todeschini,
and Viviana Consonni
2 The OECD QSAR Toolbox Starts Its Second Decade�� 55
Terry W. Schultz, Robert Diderich, Chanita D. Kuseva,
and Ovanes G. Mekenyan
3 QSAR: What Else?�� 79
Giuseppina Gini
4 (Q)SARs as Adaptations to REACH Information Requirements�� 107
Toni Alasuvanto, Andrea Gissi, Tomasz Sobanski,
Panagiotis Karamertzanis, and Mike Rasenberg
Part II Molecular and Data Modeling

5 Machine Learning Methods in Computational Toxicology�� 119
Igor I. Baskin
6 Applicability Domain: A Step Toward Confident Predictions
and Decidability for QSAR Modeling�� 141
Supratik Kar, Kunal Roy, and Jerzy Leszczynski
7 Molecular Similarity in Computational Toxicology�� 171
Matteo Floris and Stefania Olla
8 Molecular Docking for Predictive Toxicology�� 181
Daniela Trisciuzzi, Domenico Alberga, Francesco Leonetti,
Ettore Novellino, Orazio Nicolotti, and Giuseppe F. Mangiatordi
9 Criteria and Application on the Use of Nontesting Methods
within a Weight of Evidence Strategy�� 199
Anna Lombardo, Giuseppa Raitano, Domenico Gadaleta,
and Emilio Benfenati
10 Characterization and Management of Uncertainties
in Toxicological Risk Assessment: Examples from the Opinions
of the European Food Safety Authority�� 219
Alberto Mantovani
ix
x Contents
Part III Impact in Drug Discovery and Development

11 Computational Toxicology and Drug Discovery�� 233
Catrin Hasselgren and Glenn J. Myatt
12 Approaching Pharmacological Space: Events and Components�� 245
Giulio Vistoli, Alessandro Pedretti, Angelica Mazzolari,
and Bernard Testa
13 Computational Toxicology Methods in Chemical Library
Design and High-Throughput Screening Hit Validation�� 275
Kirk E. Hevener
14 Enalos Suite: New Cheminformatics Platform for Drug Discovery
and Computational Toxicology�� 287
Dimitra-Danai Varsou, Spyridon Nikolakopoulos, Andreas Tsoumanis,
Georgia Melagraki, and Antreas Afantitis
15 Ion Channels in Drug Discovery and Safety Pharmacology�� 313
Paola Imbrici, Orazio Nicolotti, Francesco Leonetti, Diana Conte,
and Antonella Liantonio
16 Computational Approaches in Multitarget Drug Discovery�� 327
Luciana Scotti, Hamilton Mitsugu Ishiki, Marcelo Cavalcante Duarte,
Tiago Branquinho Oliveira, and Marcus T. Scotti
17 Nanoformulations for Drug Delivery: Safety, Toxicity, and Efficacy�� 347
Antonio Lopalco and Nunzio Denora
18 Toxicity Potential of Nutraceuticals�� 367
Ramesh C. Gupta, Ajay Srivastava, and Rajiv Lall
19 Impact of Pharmaceuticals on the Environment: Risk Assessment
Using QSAR Modeling Approach�� 395
Part IV Predicting Human Health Toxicology Endpoints

20 (Q)SAR Methods for Predicting Genotoxicity and Carcinogenicity:
Scientific Rationale and Regulatory Frameworks�� 447
Cecilia Bossa, Romualdo Benigni, Olga Tcheremenskaia,
and Chiara Laura Battistelli
21 Stem Cell-Based Methods to Predict Developmental Chemical Toxicity�� 475
Hiroki Takahashi, Xian-Yang Qin, Hideko Sone, and Wataru Fujibuchi
22 Predicting Chemically Induced Skin Sensitization by Using
In Chemico/In Vitro Methods�� 485
Laura H. Rossi and Janine Ezendam
23 Hepatotoxicity Prediction by Systems Biology Modeling
of Disturbed Metabolic Pathways Using Gene Expression Data�� 505
Oriol López-Massaguer, Manuel Pastor, Ferran Sanz, and Pablo Carbonell
24 Nontest Methods to Predict Acute Toxicity: State of the Art
for Applications of In Silico Methods�� 519
Ronan Bureau
Contents xi
25 Predictive Systems Toxicology�� 535

Narsis A. Kiani, Ming-Mei Shang, Hector Zenil, and Jesper Tegner
26 Chemoinformatic Approach to Assess Toxicity of Ionic Liquids�� 559
Anita Sosnowska, Anna Rybinska-Fryca, Maciej Barycki,
Karolina Jagiello, and Tomasz Puzyn
27 Prediction of Biochemical Endpoints by the CORAL Software:
Prejudices, Paradoxes, and Results�� 573
Andrey A. Toropov, Alla P. Toropova, Alessandra Roncaglioni,
Index�� 585
Contributors
Antreas Afantitis • NovaMechanics Ltd, Nicosia, Cyprus

Toni Alasuvanto • European Chemicals Agency, Helsinki, Finland
Domenico Alberga • Dipartimento di Farmacia-Scienze del Farmaco,
Università degli Studi di Bari “Aldo Moro”, Bari, Italy
Davide Ballabio • Department of Earth and Environmental Sciences, Milano
Chemometrics and QSAR Research Group, University of Milano-Bicocca, Milan, Italy
Maciej Barycki • Laboratory of Environmental Chemometrics, Faculty of Chemistry,
University of Gdansk, Gdansk, Poland
Igor I. Baskin • Faculty of Physics, M.V. Lomonosov Moscow State University, Moscow,
Russian Federation; Butlerov Institute of Chemistry, Kazan Federal University, Kazan,
Russian Federation
Chiara Laura Battistelli • Environment and Health Department,
Istituto Superiore di Sanità, Roma, Italy
Emilio Benfenati • IRCCS–Istituto di Ricerche Farmacologiche Mario Negri, Milan,
Italy
Romualdo Benigni • Alpha-Pretox, Roma, Italy
Cecilia Bossa • Environment and Health Department, Istituto Superiore di Sanità,
Roma, Italy
Ronan Bureau • Centre d’Etudes et de Recherche sur le Médicament de Normandie
(CERMN), Normandie Univ, UNICAEN, Caen, France
Pablo Carbonell • Manchester Centre for Fine and Speciality Chemicals
(SYNBIOCHEM), Manchester Institute of Biotechnology, University of Manchester,
Manchester, UK
Viviana Consonni • Department of Earth and Environmental Sciences, Milano
Diana Conte • Department of Pharmacy–Drug Sciences, University of Bari “Aldo Moro”,
Bari, Italy
Nunzio Denora • Department of Pharmacy–Drug Sciences, The University of Bari Aldo
Moro, Bari, Italy
Robert Diderich • Organisation for Economic Cooperation and Development (OECD),
PARIS CEDEX 16, France
Marcelo Cavalcante Duarte • Federal University of Sergipe, Sergipe, Brazil
Janine Ezendam • National Institute for Public Health and the Environment (RIVM),
Centre for Health Protection, Bitlhoven, The Netherlands
Matteo Floris • Department of Biomedical Sciences, University of Sassari, Sassari, Italy;
IRGB–CNR, National Research Council, Institute of Genetics and Biomedical Research,
Monserrato, CA, Italy
Wataru Fujibuchi • Center for iPS Cell Research and Application, Kyoto University,
Kyoto, Japan
xiii
xiv Contributors
Domenico Gadaleta • IRCCS–Istituto di Ricerche Farmacologiche Mario Negri,

Milan, Italy
Giuseppina Gini • DEIB, Politecnico di Milano, Milan, Italy
Andrea Gissi • European Chemicals Agency, Helsinki, Finland
Francesca Grisoni • Department of Earth and Environmental Sciences, Milano
Ramesh C. Gupta • Toxicology Department, Breathitt Veterinary Center, Murray State
University, Hopkinsville, KY, USA
Catrin Hasselgren • PureInfo Discovery Inc., Albuquerque, NM, USA; Leadscope Inc.,
Columbus, OH, USA
Kirk E. Hevener • Department of Pharmaceutical Sciences, University of Tennessee
Health Science Center, Memphis, TN, USA
Paola Imbrici • Department of Pharmacy – Drug Sciences, University of Bari “Aldo
Moro”, Bari, Italy
Hamilton Mitsugu Ishiki • University of Western São Paulo (Unoeste), Presidente
Prudente, SP, Brazil
Karolina Jagiello • Laboratory of Environmental Chemometrics, Faculty of Chemistry,
Supratik Kar • Interdisciplinary Center for Nanotoxicity, Department of Chemistry and
Biochemistry, Jackson State University, Jackson, MS, USA
Panagiotis Karamertzanis • European Chemicals Agency, Helsinki, Finland
Narsis A. Kiani • Unit of Computational Medicine, Center for Molecular Medicine,
Department of Medicine, Karolinska Institutet, Solna, Stockholm, Sweden; Algorithmic
Dynamics Lab, Center for Molecular Medicine, Karolinska Institutet, Stockholm, Sweden;
Science for Life Laboratory, Solna, Sweden
Chanita D. Kuseva • Laboratory of Mathematical Chemistry (LMC), As. Zlatarov
University, Bourgas, Bulgaria
Rajiv Lall • Vets Plus Inc., Menomonie, WI, USA
Francesco Leonetti • Dipartimento di Farmacia-Scienze del Farmaco, Università degli
Studi di Bari “Aldo Moro”, Bari, Italy
Jerzy Leszczynski • Interdisciplinary Center for Nanotoxicity, Department of Chemistry
and Biochemistry, Jackson State University, Jackson, MS, USA
Antonella Liantonio • Department of Pharmacy–Drug Sciences, University of Bari
“Aldo Moro”, Bari, Italy
Anna Lombardo • IRCCS–Istituto di Ricerche Farmacologiche Mario Negri, Milan, Italy
Antonio Lopalco • Department of Pharmacy–Drug Sciences, The University of Bari Aldo
Moro, Bari, Italy
Oriol López-Massaguer • Research Programme on Biomedical Informatics (GRIB),
Dept. of Experimental and Health Sciences, Institut Hospital del Mar d’Investigacions
Mèdiques (IMIM), Universitat Pompeu Fabra, Barcelona, Spain
Giuseppe F. Mangiatordi • Dipartimento di Farmacia-Scienze del Farmaco, Università
degli Studi di Bari “Aldo Moro”, Bari, Italy
Alberto Mantovani • Department of Food Safety, Nutrition and Veterinary Public
Health, Istituto Superiore di Sanità (ISS) viale Regina Elena, Rome, Italy
Angelica Mazzolari • Dipartimento di Scienze Farmaceutiche, Università degli Studi di
Milano, Milan, Italy
Contributors xv
Ovanes G. Mekenyan • Laboratory of Mathematical Chemistry (LMC), As. Zlatarov

University, Bourgas, Bulgaria
Georgia Melagraki • NovaMechanics Ltd, Nicosia, Cyprus
Glenn J. Myatt • PureInfo Discovery Inc., Albuquerque, NM, USA
Orazio Nicolotti • Dipartimento di Farmacia-Scienze del Farmaco, Università degli
Studi di Bari Aldo Moro, Bari, Italy
Spyridon Nikolakopoulos • NovaMechanics Ltd, Nicosia, Cyprus
Ettore Novellino • Dipartimento di Farmacia, Università degli Studi di Napoli
“Federico II”, Naples, Italy
Tiago Branquinho Oliveira • Federal University of Sergipe, Sergipe, Brazil
Stefania Olla • Department of Biomedical Sciences, University of Sassari, Sassari, Italy
Manuel Pastor • Research Programme on Biomedical Informatics (GRIB), Dept. of
Experimental and Health Sciences, Institut Hospital del Mar d’Investigacions Mèdiques
(IMIM), Universitat Pompeu Fabra, Barcelona, Spain
Alessandro Pedretti • Dipartimento di Scienze Farmaceutiche, Università degli Studi
di Milano, Milan, Italy
Tomasz Puzyn • Laboratory of Environmental Chemometrics, Faculty of Chemistry,
Xian-Yang Qin • Center for Health and Environmental Risk Research, National Institute
for Environmental Studies, Ibaraki, Japan
Giuseppa Raitano • IRCCS–Istituto di Ricerche Farmacologiche Mario Negri, Milan,
Italy
Mike Rasenberg • European Chemicals Agency, Helsinki, Finland
Alessandra Roncaglioni • Laboratory of Environmental Chemistry and Toxicology,
IRCCS–Istituto di Ricerche Farmacologiche Mario Negri, Milan, Italy
Laura H. Rossi • European Chemicals Agency (ECHA), Helsinki, Finland
Kunal Roy • Drug Theoretics and Cheminformatics Laboratory, Department of
Pharmaceutical Technology, Jadavpur University, Kolkata, India
Anna Rybinska-Fryca • Laboratory of Environmental Chemometrics, Faculty of
Chemistry, University of Gdansk, Gdansk, Poland
Ferran Sanz • Research Programme on Biomedical Informatics (GRIB), Dept. of
Experimental and Health Sciences, Institut Hospital del Mar d’Investigacions Mèdiques
(IMIM), Universitat Pompeu Fabra, Barcelona, Spain
Terry W. Schultz • The University of Tennessee, College of Veterinary Medicine,
Knoxville, TN, USA
Luciana Scotti • Postgraduate Program in Natural Products and Synthetic Bioactive,
Federal University of Paraíba, João Pessoa, PB, Brazil; Teaching and Research
Management–University Hospital, Federal University of Paraíba, João Pessoa, PB, Brazil
Marcus T. Scotti • Postgraduate Program in Natural Products and Synthetic Bioactive,
Federal University of Paraíba, João Pessoa, PB, Brazil
Ming-Mei Shang • Unit of Computational Medicine, Center for Molecular Medicine,
Department of Medicine, Karolinska Institutet, Solna, Stockholm, Sweden; Science for
Life Laboratory, Solna, Sweden
Tomasz Sobanski • European Chemicals Agency, Helsinki, Finland
Hideko Sone • Center for Health and Environmental Risk Research, National Institute
for Environmental Studies, Ibaraki, Japan
xvi Contributors
Anita Sosnowska • Laboratory of Environmental Chemometrics, Faculty of Chemistry,

Ajay Srivastava • Vets Plus Inc., Menomonie, WI, USA
Hiroki Takahashi • Center for iPS Cell Research and Application, Kyoto University,
Kyoto, Japan
Olga Tcheremenskaia • Environment and Health Department, Istituto Superiore di
Sanità, Roma, Italy
Jesper Tegner • Unit of Computational Medicine, Center for Molecular Medicine,
Department of Medicine, Karolinska Institutet, Solna, Stockholm, Sweden; Science for
Life Laboratory, Solna, Sweden; Biological and Environmental Sciences and Engineering
Division, Computer, Electrical and Mathematical Sciences and Engineering Division,
King Abdullah University of Science and Technology (KAUST), Thuwal, Kingdom of
Saudi Arabia
Bernard Testa • University of Lausanne, Lausanne, Switzerland
Roberto Todeschini • Department of Earth and Environmental Sciences, Milano
Andrey A. Toropov • Laboratory of Environmental Chemistry and Toxicology,
Alla P. Toropova • Laboratory of Environmental Chemistry and Toxicology,
Daniela Trisciuzzi • Dipartimento di Farmacia-Scienze del Farmaco,
Università degli Studi di Bari “Aldo Moro”, Bari, Italy
Andreas Tsoumanis • NovaMechanics Ltd, Nicosia, Cyprus
Dimitra-Danai Varsou • NovaMechanics Ltd, Nicosia, Cyprus
Giulio Vistoli • Dipartimento di Scienze Farmaceutiche, Università degli Studi di
Milano, Milan, Italy
Hector Zenil • Unit of Computational Medicine, Center for Molecular Medicine,
Department of Medicine, Karolinska Institutet, Solna, Stockholm, Sweden; Algorithmic
Dynamics Lab, Center for Molecular Medicine, Karolinska Institutet, Stockholm, Sweden;
Science for Life Laboratory, Solna, Sweden
Part I
Where We Are and Where We Are Going To

Chapter 1
Molecular Descriptors for Structure–Activity Applications:

A Hands-On Approach
Francesca Grisoni, Davide Ballabio, Roberto Todeschini,
and Viviana Consonni
Abstract
Molecular descriptors capture diverse parts of the structural information of molecules and they are the
support of many contemporary computer-assisted toxicological and chemical applications. After briefly
introducing some fundamental concepts of structure–activity applications (e.g., molecular descriptor
dimensionality, classical vs. fingerprint description, and activity landscapes), this chapter guides the readers
through a step-by-step explanation of molecular descriptors rationale and application. To this end, the
chapter illustrates a case study of a recently published application of molecular descriptors for modeling the
activity on cytochrome P450.
Key words Molecular descriptors, Molecular similarity, Chemical space, Mathematical chemistry,
QSAR
1 Introduction
Many toxicological applications rely on the principle that the bio-

logical properties of any chemical are the effects of its structural
characteristics [1, 2]. Analogously, compounds with similar molec-
ular structures will be likely to show the same biological and physi-
cochemical profile, according to the long-held guiding principle of
similarity [3].
This concept was first formalized in 1868 by Crum-Brown and
Fraser [4], who noticed the existence of a correlation between the
biological activity of different alkaloids and their molecular consti-
tution. More specifically, according to the authors, the physiologi-
cal action of a substance in a biological system (Φ) is a function (f )
of its chemical constitution (C), namely: Φ = f (C). Analogously, to
Francesca Grisoni and Viviana Consonni contributed equally to this chapter.
Orazio Nicolotti (ed.), Computational Toxicology: Methods and Protocols, Methods in Molecular Biology, vol. 1800,
https://doi.org/10.1007/978-1-4939-7899-1_1, © Springer Science+Business Media, LLC, part of Springer Nature 2018
3
4 Francesca Grisoni et al.
any alteration in the chemical constitution (ΔC), there corresponds

a change in the biological activity (ΔΦ). Crum-Brown’s and
Fraser’s statement paved the way to many studies linking several
molecular characteristics to experimentally measurable properties.
For instance, the biological activity of phenoxyacetic acids was esti-
mated by Hammett substituent constants and the partition coeffi-
cients [5], the anesthetic potency of aliphatic alcohols was related
to their carbon chain length and molecular weight [6], and the
toxicity of narcotics to their water solubility [7]. Later, the first
theoretical, structure-related numbers were also developed, such as
the Wiener index (i.e., the sum of all the shortest through-bond
distances between atoms [8]) and the Platt number (i.e., twice the
number of non-hydrogen bonds in the molecule [9]), which were
related to the boiling point of hydrocarbons. These are the first
examples of the conversion of molecular characteristics into useful
numbers that allow for a mathematical treatment of molecules.
These numbers are nowadays defined as “molecular descriptors”,
which are mathematical representations of a molecule obtained by
a well-specified algorithm applied to a defined molecular represen-
tation or by an experimental protocol [10].
Each molecular descriptor captures a small part of the global
chemical information contained in the real molecular structure.
Therefore, the number of descriptors is continuously increasing
with the increasing request of deeper investigations of chemical
and biological systems. Evidence of the interest of the scientific
community in the molecular descriptors is provided by the huge
number of descriptors proposed up today: thousands of descriptors
derived from different theories and approaches are actually defined
and computable by using dedicated software tools [10].
Because of their numeric nature, molecular descriptors allow
one to capture the theoretical information arising from the molec-
ular structure (e.g., geometric, steric, and electronic properties)
and to link it to some experimental evidence on the molecule [11]
(e.g., acute/chronic toxicity, receptor binding). Thus, molecular
descriptors have become the support to many computational toxi-
cology applications. Under this perspective, the structure–activity
paradigm can be formulated as the following:

(
P = f x 1 ,x 2 , …,x p ) (1)
where P is the biological/physicochemical property of a com-

pound, which is considered as a mathematical function of some
structural characteristics, encoded within p molecular descriptors
(x1, x2,…, xp).
Once the relationship (f) of Eq. 1 has been estimated, the
property (P) of a new or untested chemical can be inferred from its
molecular structure, by calculating the selected molecular descrip-
tors. This approach is generally known as QSAR (quantitative
Molecular Descriptors for Structure–Activity Applications: A Hands-On Approach 5
structure–activity relationship) or QSPR (quantitative structure–

property relationship), according to the type of modeled property,
that is, biological or physicochemical, respectively.
The development of QSAR/QSPR models is a nontrivial pro-
cess (Fig. 1). Important stages of the procedure are:
(a) Data selection and curation. Any structure-based application
relies and heavily depends on the starting set of molecules. On
the one hand, the choice of the starting set of molecules and
properties will influence the quantity and type of information
captured and transferred to the models, for instance in terms
of modes of action included. On the other hand, any error in
the data will propagate into the developed model and will limit
its reliability and applicability. Molecular structures (and their
representation) are the starting point of the descriptor calcula-
tion and, thus, a proper structure curation has a direct impact
on the modeling outcome [12]. Errors in the structures may
reflect in erroneous descriptor values and, accordingly, in
unreliable model outcomes. At the same time, also ensuring
that the annotated experimental properties (e.g., biological
and toxicological endpoints to model) are reliable and do not
contain errors and anomalous values is crucial to ensure the
model usefulness [13]. At this stage, data curation, in terms of
the unambiguity of the annotated structures and chemical
identifiers [14], and of outlier analysis and experimental proto-
col evaluation [15], is key to ensure that no errors are intro-
duced in the modeling workflow.
(b) Molecular descriptor calculation. Once all of the molecular
structures have been checked and curated, molecular descrip-
tors are calculated from the chosen molecular representation,
using one or more available software tools. The descriptors to
use are generally chosen based on a priori knowledge on the
property to model or on their previous performance for the
problem under analysis (if available). The computed molecular
descriptors become then the new independent variables to be
used for developing the model(s) of interest.
(c) Information extraction. In this phase, information regarding
the relationship between molecular descriptors and property
to be predicted is extracted and formalized into a mathemati-
cal model. Several steps are usually required, such as (1) the
data splitting into a training and a test set, the former used for
model calibration, the latter for model evaluation; (2) the
choice of the appropriate modeling technique, according to
the project scopes and the performance; (3) supervised vari-
able selection (e.g., [16–20]) to identify the best descriptors to
model the property of interest, and increase model stability,
performance and interpretability; (4) model evaluation
through dedicated metrics, such as Root Mean Squared Error
Fig. 1 Principal steps of quantitative structure–activity relationship (QSAR) development and use: starting from
a set of molecules with annotated experimental properties (e.g., physicochemical, toxicological, and biologi-
cal), several types of molecular descriptors can be calculated. The obtained dataset (molecular descrip-
tors + experimental properties) is then used in the phase of information extraction, to obtain a reliable and
validated QSAR model. The model can be later applied to predict the properties of untested molecules, to
obtain mechanistic insights through the interpretation of the molecular descriptors, or to design novel
molecules
(RMSE) and the model predictive ability (Q2) [21, 22] for
quantitative responses, and Sensitivity and Non-Error Rate
[23] for qualitative responses. It is important to keep in mind
that structure–activity models are reductionist, that is, they
capture only a portion of information relevant for the process
under investigation. In this framework, one additional factor
to consider when developing structure–activity models is the
so-called applicability domain (AD) [24, 25]. AD can be
defined as the region of the chemical space (or, more precisely,
of the descriptor space) where the predictions can be consid-
ered as reliable and the model assumptions are met. Molecules
falling outside the AD may be too dissimilar from the mole-
cules used to calibrate the model and, thus, they may be sub-
ject to different biological pathways, have diverse mechanisms
of action or be characterized by structural features not repre-
sented in the training data.
(d) Model application. The validated model can be then used for
several applications, such as to predict the properties of
untested molecules (e.g., [15, 26, 27]), to design new mole-
cules with desirable properties (e.g., [28, 29]) and/or to glean
useful insights into the relationships among the structural fea-
tures and the property of interest (e.g., [30–33]).
In addition to the QSAR/QSPR approach, the problem of
property estimation can also be addressed indirectly, that is, with-
out the need of mathematically expressing the relationship between
descriptors and property. This is done through the similarity prin-
ciple, according to which, molecules with similar descriptor values
will be likely to have similar bioactivities [34]. In toxicology and
related fields, this approach is often referred to as read-across [35],
while in other fields, such as drug discovery, it is referred to as simi-
larity search [36].
Since molecular descriptors are usually fast and inexpensive to
compute, QSAR and descriptor-based methods for toxicological
and ecotoxicological applications have been in the spotlight of
industries, researchers and regulatory agencies, especially as alter-
natives to animal testing [37]. In computational toxicology and
related fields, molecular descriptors have, for instance, been applied
for testing prioritization purposes [38], in vivo acute [39–41] and
chronic [42–44] toxicity prediction, as well as organ [45, 46] and
receptor-mediated toxicity modelling [27, 47, 48]. In addition,
descriptor-based approaches have become valuable tools for screen-
ing and designing efficient hits and pharmaceuticals [29, 49, 50],
protein–ligand interaction prediction [51–53], environmental fate
[15, 26, 54, 55] and risk assessment [56, 57], as well as for food
chemistry applications [58, 59].
1.1 Molecular The information captured by molecular descriptors can vary from
Representation simple bulk properties to complex three-dimensional definitions or
and Descriptor substructure frequency. In particular, different levels of complexity
Dimensionality (also known as “dimensionality”) can be used to represent any
given molecule (Fig. 2), as follows:
(a) 0-Dimensional (0D). The simplest molecular representation
is the chemical formula, that is, the specification of the
chemical elements and their occurrence in a molecule. For
instance, the chemical formula of
2,3,7,8-Tetrachlorodibenzodioxin (a contaminant known
for its toxicity to humans and ecosystems [60, 61]) is
C12H4Cl4O2, which indicates the presence of 12 Carbon, 4
Hydrogen, 4 Chlorine, and 2 Oxygen atoms. This represen-
tation is independent of any knowledge about atom connec-
tivity and bond types. Hence, molecular descriptors obtained
from the chemical formula are referred to as 0D descriptors
and capture bulk properties. 0D descriptors are very simple
to compute and interpret, but show a low information con-
tent and a high degeneration degree, that is, they may have
equal values for different molecules. Some examples of 0D
descriptors are atom counts (e.g., number of carbon atoms),
molecular weight, and sum or average of atomic properties
(e.g., atomic van der Waals volumes).
Fig. 2 Graphical example of different molecular representations of the same structure (ibuprofen, here depicted
as a 2D structure). The relationship between chosen dimensionality and information content/ease of calcula-
tion is also depicted
(b) 1-Dimensional (1D). According to this representation, mol-

ecules are perceived as a set of substructures, such as functional
groups or atom-centered fragments. This representation does
not require the complete knowledge of molecular structures.
The 1D representation of the molecule is reflected in the
derived descriptors, which usually are binary (encoding for the
presence/absence of given substructures) or occurrence
frequencies.
(c) 2-Dimensional (2D). This representation adds an additional
information level to the 1D representation, by also considering
how the atoms are connected, in terms of both presence and
nature of chemical bonds. Usually, the molecule is represented
as a graph, whose vertexes are the atoms and edges are the
bonds. From a graph representation, several numerical quanti-
fiers of molecular topology are mathematically derived in a
direct and unambiguous manner. They are commonly known
as topological indices (TIs). TIs encode topological properties
(e.g., adjacency, connectivity) and are usually sensitive to struc-
tural features such as size, shape, symmetry, branching, and
cyclicity. Often, also specific chemical properties of atoms are
considered, e.g., mass and polarizability [62], or the presence
of hydrogen bond donors/acceptors [36, 63, 64]. Thus, topo-
logical indices can be logically divided into two categories [65]:
(1) topostructural indices, which encode only information
about adjacency and through-bond distances between atoms,
and (2) topochemical indices, which quantify information
about topology but also specific chemical properties of atoms,
such as their chemical identity and hybridization state.
(d) 3-Dimensional (3D). An additional level of complexity may
be added by perceiving the molecule not only in terms of atom
type, connectivity, and adjacency but also by viewing it as a
geometrical object in space, characterized by the spatial con-
figuration of the atoms. In other words, the molecule is defined
in terms of atom types and their x-y-z coordinates. Descriptors
deriving from 3D representation have a high information con-
tent [66] and can be particularly useful for modeling pharma-
ceutical and biological properties [67–69]. When dealing with
the 3D representation, users have to keep in mind several issues
connected to the geometric optimization of molecules, such as
(1) the influence of the optimization method on the coordinate
values [70]; (2) the presence of more than one similar mini-
mum energy conformer for highly flexible molecules; and (3)
the difference between the bioactive geometry and the opti-
mized geometry, the degree of deformation depending upon
the number of freely rotatable bonds in the molecule [71]. For
these reasons, the cost/benefit of using 3D descriptors is case-
dependent and has to be carefully evaluated [68].
(e) 4-Dimensional (4D). In addition to the molecular geometry,

also a “fourth dimension” can be introduced, usually aiming
to identify and characterize quantitatively the interactions
between the molecule(s) and the active site(s) of a biological
receptor. For instance, a grid-based representation can be
obtained by placing molecules in a 3D grid of several thou-
sands of evenly spaced grid points and by using a probe (steric,
electrostatic, hydrophilic, etc.) to map the surface of the mol-
ecule. The molecule can be then described through its molec-
ular interactions with the probe (e.g., see Comparative
Molecular Field Analysis [CoMFA] [32] and Comparative
Molecular Similarity Indices Analysis [CoMSIA] [72] descrip-
tors). 4D representations may also be “ensemble-based,” that
is, they can include conformational flexibility and freedom of
alignment, through an ensemble of the spatial features of dif-
ferent members of a training set [73, 74], or by representing
each ligand by an ensemble of conformations, protonation
states, and/or orientations [75, 76].
Dimensions higher than the fourth (i.e., 5D-QSAR and
6D-QSAR) can be added to the modeling problem, for instance,
by considering different induced-fit models and/or different solva-
tion scenarios (e.g., [77, 78]).
1.2 Classical Descriptors are usually calculated from the chosen molecular rep-
Descriptors vs. resentation. They can be chosen based on an a priori knowledge
Fingerprints and/or on their performance for the problem under analysis.
Molecular descriptors can be grouped according to the rationale
underlying their design, which influences their applicability to
computational problems and the required modeling steps. In par-
ticular, molecular descriptors can be divided into classical molecu-
lar descriptors and binary fingerprints, as follows:
1. Classical molecular descriptors (MDs) are designed to encode
a precise structural/chemical feature (or a set of features of dif-
ferent complexity) into one, single number. Thus, each descrip-
tor can be used alone or in combination with other descriptors.
Classical descriptors can have different measurement scales:
they can be integers (e.g., number of double bonds and counts
of atom types), binary (e.g., presence/absence of a given sub-
stituent) or can have continuous values (e.g., molecular weight).
MDs may be subject to scaling, reduction and selection tech-
niques, as explained in the next paragraph. The majority of clas-
sical molecular descriptors are usually interpretable to a certain
extent, and, in some cases, they can be mapped back onto sets
of structural features (i.e., reversible decoding).
2. Binary fingerprints (FPs) give a complete representation of all
the structural fragments of a molecule in a binary form. Unlike
classical descriptors, fingerprints encode information about 2D
molecular structure in a series of binary digits (bits) that represent

the presence or absence of particular substructures in the mole-
cule and are meaningful only when used as a whole. Typically, a
set of patterns (e.g., branched/linear fragments or substructures)
are generated from a given molecule, the presence and absence of
a pattern are encoded within a string of a given length and marked
as “1” or “0”, respectively. Hashing algorithms are often applied,
leading to a “collision” of multiple features in the same bit(s) and
to a loss of one-to-one correspondence with molecular features.
Fingerprints allow performing quick calculations for molecule
similarity/diversity problems [79, 80] but lack the possibility of
an immediate connection with precise structural features.
Nonetheless, the frequency of the molecular fragments encoded
into FPs can be used to interpret the structural features underly-
ing the observed bioactivity patterns (e.g., [30]). A variant of the
classical binary fingerprints is that of the count-based fingerprints
[81–83], which, instead of being a vector of presence/absences of
fragments, represent the molecules as a count-vector of fragment
frequencies. Count fingerprints have been investigated as an alter-
native to binary fingerprints, but no systematic improvement over
their binary counterpart has been demonstrated yet [84].
1.3 Descriptor When dealing with a modeling campaign, one can hypothesize that
Choice and Activity not all of the structural features are relevant in determining the
Landscapes experimental property of interest. Likewise, also not all of the MDs
will be relevant for the modeling purposes. Thus, the molecular
descriptor choice directly affects the outcome of the respective
computer-aided project (e.g., [49, 63]). The choice of molecular
descriptors has shown to have much greater influence on the pre-
diction performance of QSAR models than the nature of modeling
techniques [85, 86]. Thus, it is always crucial to determine the
optimal set(s) of descriptors to best address the problem of inter-
est. In addition to the choice of molecular descriptors, also their
processing and use have a crucial impact on the corresponding
computational models, such as on how the similarity between mol-
ecules is expressed [49].
A relevant factor to consider when dealing with descriptor-
based applications is the so-called “activity landscape” [87–90].
The activity landscape can be thought of as the relationship between
the molecular descriptor space and the experimental a ctivity space.
In other words, if the former is imagined as a set of geographical
coordinates of a map (i.e., latitude and longitude), the latter would
represent the height of the landscape at each point of the map
(Fig. 3). Activity landscapes depend on the nature of the endpoint/
assay of interest, the chemical space covered by the training com-
pounds, the density distribution of the compounds in these regions,
and, most importantly, on the nature of the molecular descriptors
used [90]. Molecular activity landscapes can be similar to gently
rolling hills, rugged canyons [91] or can be an intermediate sce-

nario. In the presence of gently rolling hills (Fig. 3a), small changes
in molecular structure (or, more precisely, in the descriptors used to
capture the structural information) will have a small effect on
the activity. In this case, the neighborhood of a given molecule will
be populated by a spectrum of increasingly diverse structures shar-
ing a similar activity. Conversely, in the presence of “canyon-like”
landscapes (Fig. 3b), small changes in structure may have dramatic
effects on the activity. These drastic changes are commonly known
as “activity cliffs” [90]. Activity cliffs are generally defined as regions
of the chemical space where neighboring molecules show large dif-
ferences in potency, activity or toxicity. The study of activity cliffs is
of high interest in many structure–activity based applications, given
their “small chemical change = large activity change” character
[92]. While, on the one hand, activity cliffs provide computational
chemists with fundamental information to understand the underly-
ing structure–activity relationships, on the other hand they
also affect the generalization ability of the corresponding models
[93]. In this context, molecular descriptors play a fundamental role,
as they determine how the molecular structure is converted into
information and, accordingly, how the similarity between molecules
is perceived. Thus, the appropriate choice of the molecular descrip-
tor space may reduce or emphasize the presence of activity cliffs,
leading to the optimal structure–activity scenario for the problem
under analysis, such as gently rolling hills for toxicological problems
and rugged landscapes for lead optimization purposes.
Finally, because of their unambiguous connection with the
underlying molecular structure(s), molecular descriptors can be
leveraged to obtain mechanistic insights into the underlying bio-
logical processes (e.g., [30, 31, 94, 95]). To this end, the choice of
easily interpretable molecular descriptors is generally advisable for
Fig. 3 Schematic representation of two activity landscapes [91], given two hypothetical descriptors (x1 and x2)
and a given biological property (A): (a) “gently rolling hills” landscape, where, to small changes in the descrip-
tor values correspond small changes of the corresponding biological activity; (b) “rugged canyons” landscape,
where, to small changes in the descriptor values correspond drastic changes in the activity
most of the modeling and descriptor-based protocols. In addition,

the model interpretability can be fostered by choosing “parsimoni-
ous” models, that is, models comprised of as few as (and as simple
as) possible descriptors. Finally, even when a linkage with a struc-
tural interpretation is not immediate, understanding the encoded
structural features may set the basis to the introduction of new
chemical and biological concepts [10].
2 Materials
For the guided example, we considered two QSAR models to pre-

dict the cytochrome P450 activity [30]. The two models are differ-
ent in their nature, in terms of included molecular descriptors and
utilized modeling techniques. This paragraph describes the chosen
endpoint and the training data, along with the QSAR models, their
performance and the software/material requirements for follow-
ing/reproducing the guided exercise on a step-by-step basis.
2.1 Endpoint Cytochromes P450 (CYP) are a family of monooxygenase enzymes

Description known for their crucial role in the metabolism of xenobiotics, as
they are involved in the oxidation of the majority of compounds
[96]. At the same time, these enzymes may also metabolically acti-
vate biologically inert compounds to electrophilic derivatives that
can cause toxicity, cell death, and sometimes cellular transforma-
tion resulting in cancer [97]. Thus, the evaluation of the interac-
tion of CYP with chemicals represents a fundamental step for
toxicity assessment, as well as for drug discovery and design
[98–100].
In the present chapter, the 3A4 isoform of the receptor
(CYP3A4) was considered [101], as it interacts with more than a
half of all clinically used drugs [99].
2.2 Data We chose the CYP3A4 dataset developed and curated by Nembri
et al. [30]. The data were retrieved from the publicly available CYP
bioactivity database of Veith et al. [102], which contains the
potency values of 17,143 drug-like compounds on five Cytochrome
P450 isoforms (3A4, 2D6, 2C9, 2C19, 1A2). The dataset was
retrieved from PubChem [103] (PubChem ID = AID: 1851). The
database provides the class of activity (active/inactive) for each
compound, identified by a SMILES (Simplified Molecular Input
Line Entry System, see Subheading 3.1.2) string.
Data were curated by: (1) removing the records without
SMILES and/or activity class; (2) removing duplicate struc-
tures with mismatching class activity; (3) removing discon-
nected structures. The CYP3A4 dataset is composed of 9122
molecules, 6385 of which (70%) were used as the training set to
build the models, while 2737 (30%) were used as the validation
set for model selection. The final models were recalibrated on

the full set of 9122 molecules. Additional details can be found
in the original publication [30].
The original dataset can be downloaded free of any charge from
Milano Chemometrics & QSAR Research Group website [104].
2.3 Classifi In the guided example, after describing the molecular descrip-
cation QSARs tors under analysis, we will show how they can be used for
structure–activity classification tasks. Classification commonly
refers to the application of statistical and mathematical tech-
niques to the problem of predicting the “label” of a given object
(in our case, a molecule). In the SAR context, the labels are
generally experimental values expressed in a categorical form,
such as “active” vs. “inactive” compounds, or “toxic” vs. “non-
toxic” compounds. Classification problems differ from the so-
called regression problems, where the value to be predicted is
numerical and continuous, such as the half maximal effective
concentration (EC50), or the octanol–water partitioning coef-
ficient (KOW).
In this chapter, two classification models for predicting the
activity towards CYP3A4 will be taken into account, namely a
decision-tree-based QSAR and a similarity-based QSAR. This sec-
tion will describe the rationale of each modeling approach, while
details on each single model along with the definition of the
selected molecular descriptors will be given in the Methods section
on a step-by-step basis.
2.3.1 Decision-Tree This model was calculated by the Classification and Regression
Based QSAR Model (CART) Tree (CART) approach [105]. CART is a machine-learning algo-
rithm based on a recursive partitioning of data using one descriptor
at a time: at each univariate split, data are divided in two groups (as
homogeneous as possible) according to their descriptor values; the
splitting procedure is further applied to each group separately until
a stop criterion is fulfilled. The model is graphically represented as
a decision tree (Fig. 4): each node is a univariate split that parti-
tions the molecules in the following branches, while the leaves are
the predicted classes for the molecules that fall in them. In addition
to its simplicity and interpretability, CART technique can deal with
nonlinear relationships between variables, thus it is particularly
well-suited for complex biological problems (e.g., [31, 106]). The
power of CART also lies in its ability to select the descriptors that
provide the best class separation automatically, by neglecting those
that are not relevant for the problem under analysis. However, vali-
dation protocols are fundamental to prune the classification tree
and avoid overfitting [107].
2.3.2 Similarity-Based The second model relies on the similarity between molecules,
QSAR Model (KNN) according to the k-Nearest Neighbors (KNN) method [108]. For
a new molecule to be predicted, first a fixed number (k) of similar
Fig. 4 Exemplary scheme of a Classification and Regression Tree (CART) model with three nodes (rounded
rectangles) and four leaves (circles). Each node corresponds to a univariate split of data according to a selected
descriptor (MDi) and threshold value (thr). A new molecule is assigned the successive node/leaf according to
the comparison of its descriptor value with the node threshold (in this example, y [yes] indicates that the
descriptor value is smaller than the threshold, while n [no] that is equal to or larger than the node threshold).
Leaves correspond to the assigned class (e.g., class 1 and 2 in this example)
Fig. 5 Graphical example of the k-Nearest Neighbors approach (KNN). The model is based on two variables (x1
and x2) and k = 3. Given a target molecule with unknown class (a), the three closer molecules in the variable
space are selected (b) and their experimental classes are used to predict the unknown class of the target as a
majority vote (c). In this case, 2 out of 3 neighbors belong to the “orange” class, and, thus, the unknown mol-
ecule will be assigned the “orange” class
compounds (i.e., the nearest neighbors) with known experimental

class are identified according to their molecular descriptor values;
then, the class of the neighbors is used to assign a class to the new
compound as a majority vote (Fig. 5). In other words, the class
that is most frequent amongst the neighbors will be chosen as the
most likely for the compound with unknown activity.
2.4 Model The predictive ability of the two classification models was quanti-
Performance fied using Sensitivity (Sn), Specificity (Sp), and Non-Error Rate
Evaluation (NER), which, in the two-class cases, are defined as follows:
TP
Sn% = × 100
TP + FN
TN (2)
Sp% = × 100
TN + FP
Sn + Sp%
NER% = %
2
where TP, TN, FP and FN are the number of true positives (active
molecules correctly predicted), true negatives (inactive molecules
correctly predicted), false positives (inactive molecules predicted as
active) and false negatives (active molecules predicted as inactive).
All of the parameters range from 0 to 100%, the higher, the better.
Sn and Sp are class-based parameters, that is, they quantify the abil-
ity to correctly classify positive and negative compounds, respec-
tively. NER, which is the average between Sn and Sp, is a measure
of the global classification ability of a model. Sn, Sp and NER were
chosen as they are optimal parameters for both binary- and multi-
class classification tasks [109] .
The classification statistics for the selected models are reported
in Table 1. Both the models correctly identify around 74% of
active compounds (Sn), the CART model being slightly better.
The KNN model, on the other hand, has a slightly better ability
to correctly identify inactive compounds (Sp% = 79.8%) than
CART (Sp% = 75.1).
2.5 Example Six molecules were chosen to illustrate the basic steps of descriptor
Molecules calculation, processing and use. The molecules are reported in
Table 2 along with their identification label (ID), chemical name
and chemical representation (see Note 1).
2.6 Software Molecular descriptors were calculated using the software Dragon
and Requirements 7.0 [110], which can be used to reproduce the results. In addition,
this chapter will illustrate how to manually calculate them starting
from a 2D molecular representation. Besides this, several free alter-
Table 1
Classification performance of the selected QSAR models on the training data, in terms of NER, Sn
and Sp. Number of molecular descriptors (p) and of nodes/neighbors (k) are also reported. Statistics
have been calculated on the 6385 molecules used to calibrate the models (“Fitting”), on the test set
of 2737 molecules (“Validation”) and on all of the 9122 molecules (“All data”) of the original dataset
Fitting Validation All data
Model p k NER% Sn% Sp% NER% Sn% Sp% NER% Sn% Sp%
CART 3 4 74.6 74.7 74.6 73.4 73.4 73.4 74.7 74.3 75.1
KNN 2 14 76.5 73.7 79.3 74.9 71.0 79.3 76.4 72.9 79.8
Table 2
List of the molecules used in the guided example, annotated for their ID, name and 2D representation
ID Name 2D Structure
Mol1 N-[4-(2,5-dimethylphenyl)-5-methyl-1,3-
thiazol-2-yl]-3-ethyl-5-methyl-1,2-oxazol-
4-carboxamide
Mol2 1,1′-(1,3-propanediyldisulfonyl)
bis(4-methylbenzene)
Mol3 2-oximino-3-butanone
Mol4 2-{[2-Hydroxy-3-(4-methylphenoxy)propyl]
carbamoyl}benzoic acid
Mol5 3-bromophenyl nicotinate
Mol6 2-(9H-xanthen-9-yl)benzoic acid
native software exist, such as E-Dragon [111] (see Note 2).

MATLAB v. R2016 [112] or Python v. 2.7 [113] are required to
perform the model predictions. Throughout the chapter and in the
section Notes, readers will find the programming code to repro-
duce the results and/or apply the models to any new molecule.
2.7 Supplementary The training and example data (i.e., SMILES, descriptor values,
Material experimental classes) are provided free of any charge as both
spreadsheet (.xlsx) and MATLAB (.mat) files. They can be down-
loaded from Milano Chemometrics and QSAR Research Group
website at the following link: http://michem.disat.unimib.it/
chm/download/mol_desc_example.htm. The Excel-format data-
set can be found under the name “CYP3A4_dataset.xlsx” (Table 3).
Table 3
Example portion of the file “CYP3A4_dataset.xls”, sheet “Training set”
ID SMILES Class Class code nBO nBM C% …

Tr1 CN(C)c1nc(nc(n1)N(C)C)N(C)C Inactive 2 15 6 27.3 …
Tr2 CC(C)C(O)(C(C)C)C(C)C Inactive 2 10 0 30.3 …
… … … … … … … …
Tr16 CCN(CC)C(=S)SSC(=S)N(CC)CC Active 1 15 2 27.8 …
Fig. 6 Items provided in the MATLAB file (“CYP3A4_dataset.mat”), available at

http://michem.disat.unimib.it/chm/download/mol_desc_example.htm
The first sheet (“Training set”) contains the list of the molecules
that were used in the original work to develop the models. Each
row corresponds to a molecule, with an assigned ID, a structure
representation (SMILES, see Subheading 3.1.2), the experimental
class (annotated both as a string and a numeric code), along with
the values of the selected molecular descriptors. The second sheet
(“Example molecules”) contains the molecules used in the step-
by-step guided example. It is organized as the previous sheet (i.e.,
IDs, SMILES and descriptor values), except for the experimental
class, which is not reported it being unknown.
The MATLAB file “CYP3A4_dataset.mat” contains several
items (Fig. 6):
(a) class: the experimental classes (1 = active, 2 = inactive);
(b) ID: the molecule IDs, which are the same as those provided in
the Excel file;
(c) SMILES: the SMILES strings of the molecules;
(d) X_CART: the molecular descriptors for the CART model;
(e) X_KNN: the molecular descriptors for the KNN model;
(f) lab_CART and lab_KNN: contain the descriptor labels of X_
CART and X_KNN, respectively.
All of the item rows have a row-to-row correspondence, that
is, each row corresponds univocally to the same molecule.
3 Methods
In this section, we present a step-by-step tutorial to apply the previ-

ously described QSAR models to predict the activity on CYP3A4.
We will show the necessary steps to (1) represent a new molecule,
(2) calculate the relevant descriptors, (3) use the computed descrip-
tor values to predict the activity on CYP3A4. The chapter workflow
is divided in four logical phases, namely: (1) structure representation
(Subheading 3.1), (2) descriptor calculation (Subheading 3.2), (3)
predictions using the CART model (Subheading 3.3), and (4) pre-
dictions using the KNN model (Subheading 3.4). Finally, some con-
siderations about the use, the choice and the interpretation of
molecular descriptors are made in Subheading 3.5. The illustrated
steps can be applied to any type of descriptor-based computational
workflow, beyond the described application.
3.1 Structure To make molecules machine-readable and calculate molecular

Representation descriptors, it is fundamental to represent them symbolically,
through formal procedures and conventional rules. The methodol-
ogy and the symbolic representation directly influence the quantity
and type of chemical information preserved and, in turn, the type
of computable descriptors. Under this perspective, it is crucial to
identify the optimal level of complexity to address the problem
under analysis on a case-by-case basis.
As explained in the introductory section, any molecular struc-
ture can be represented in many ways and with different dimen-
sionality. When dealing with a 2D perception of the molecule, the
most widely used representations are: (1) the molecular graph,
which allows for a manual calculation of the selected molecular
descriptors or (2) linear string notations, (e.g., SMILES [114]),
which are a common graph-based input to many software for auto-
matic descriptor calculation. These representations will be explained
in the following subheadings.
3.1.1 Molecular Graph A graph is the fundamental mathematical object of graph theory
[115], which can be adapted to the description of molecules. The
so-called “molecular graph” is a topological representation of a
chemical compound; it is usually denoted as follows:
G = (V ,E ) (3)

where V is a set of vertices that correspond to the molecule atoms
and E is a set of elements representing the binary relationships
between pairs of vertices; unordered vertex pairs are called edges,
which correspond to the bonds between atom pairs. If two vertices
occur as an unordered pair more than once, they define a multiple
edge. A molecular graph obtained by excluding all the hydrogen
atoms is called H-depleted molecular graph (Fig. 7a), while a
Fig. 7 Examples of the generation of different types of graphs for Mol3: (a)
H-depleted graph, (b) H-filled graph, and (c) H-depleted multigraph
molecular graph that also includes hydrogens is called H-filled

molecular graph (Fig. 7b). The multigraph (or multiple graph) is a
graph including multiple edges (i.e., multiple bonds) between at
least a pair of vertices (Fig. 7c).
Some of the salient elements and concepts relative to a molec-
ular graph are:
(a) The vertex degree (δ), which characterizes each vertex (atom).
It is the count of its σ electrons in the H-depleted molecular
graph, that is, the number of adjacent (bonded) non-hydrogen
atoms. It is often referred to as simple vertex degree.
(b) The valence vertex degree (δV), which, unlike the simple ver-
tex degree, considers all of the atom valence electrons. In par-
ticular, it is equal to the difference between the number (Zv) of
valence electrons (σ electrons, π electrons and lone pair elec-
trons) and the number of attached hydrogen atoms (h).
(c) The conventional bond order (π*), which quantifies the edge
multiplicity. It can be equal to 1, 2, 3, and 1.5, for single,
double, triple and aromatic bonds, respectively;
(d) The path (or self-avoiding walk), which is defined as the

sequence of pairwise adjacent edges (bonds) connecting any
two vertices (atoms) vi and vj, without including any repeated
vertices. The path length is the number of edges associated
with the path, that is, the number of bonds separating the two
vertices of interest.
Fig. 8 Examples of the generation of SMILES from a given molecular representation for Mol3 (a) and Mol5 (b).
Grey numbers represent the order used for generating the SMILES string
(e) The topological distance, which is the length of the shortest

path connecting two vertices.
3.1.2 Simplified One of the most widely used graph-based (2D) representations is the
Molecular Input Line Entry Simplified Molecular Input Line Entry System (SMILES). SMILES
System (SMILES) are a chemical notation language specifically designed for computer
use by chemists [116]. According to their rationale, after representing
the chemical as a molecular graph, it is converted to a linear notation
by specifying atom types and connectivity, as well as other chemical
information through predefined rules, as follows (Fig. 8):
(a) Atoms are represented by their atomic symbols, with the pos-
sibility to omit H;
(b) Single, double, triple and aromatic bonds can be represented
with the following symbols: “−”, “=”, “#”, and “:”, respec-
tively. Single bonds can be omitted.
(c) Branches are specified by enclosures in parentheses;
(d) Cyclic structures are represented by breaking one single or
aromatic bond in each ring and starting from one of the ring
atoms. Ring “opening”/”closure” bonds are then indicated
by a digit immediately following the atomic symbol at each
ring closure (Fig. 8b). Aromaticity on carbon atoms can be
written with lower-case letters or by alternating single and
double bonds (Kekulé notation). For instance, benzene can be
both written as “c1ccccc1” and “C1=CC=CC=C1”;
(e) Local chirality can be specified using the symbols “/” and “\”.
For instance, E- and Z-1,2-difluoroethene can be written as
F/C=C/F and F/C=C\F, respectively. Additionally, tetrahe-
dral centers are often indicated using “@” (or “@@”), follow-
ing the atomic symbol of the chiral atom. “@” indicates that
the listed neighbors are arranged anticlockwise, while “@@”

that they appear in a clockwise order.
(f) The chosen atom order for generating the SMILES does not
affect the encoded 2D structure. However, several types of
standardized (also known as canonicalized) SMILES genera-
tion procedures exist [117–120].
SMILES allow to easily store and handle molecular struc-
tures. They are useful for many chemoinformatic applications,
such as database search, molecular descriptor calculation and
Table 4
ID, 2D structure and corresponding SMILES strings for the example molecules. The red dot in the
structural representation indicates the starting atom used to generate the SMILES
ID 2D Structure SMILES
Mol1 CCc1noc(C)c1C(=O)Nc2nc(c(C)s2)
c3cc(C)ccc3C
Mol2 Cc1ccc(cc1)S(=O)(=O)CCCS(=O)(=O)
c2ccc(C)cc2
Mol3 CC(=O)C(=NO)C
Mol4 Cc1ccc(OCC(O)CNC(=O)
c2ccccc2C(=O)O)cc1
Mol5 Brc1cccc(OC(=O)c2cccnc2)c1
Mol6 OC(=O)c1ccccc1C2c3ccccc3Oc4ccccc24
Table 5
Molecular descriptors for the six example molecules under analysis
ID nBM nBO C% SIC3 ATSC4p ATSC6i nROH NaasC

Mol1 27 17 41.3 0.867 9.042 1.306 0 9
Mol2 24 16 39.5 0.676 12.844 0.880 0 4
Mol3 6 2 28.6 0.822 0.936 0.008 1 0
Mol4 25 14 41.9 0.902 7.733 0.829 2 4
Mol5 17 13 50.0 1.000 4.169 0.295 0 3
Mol6 26 19 54.1 0.808 8.346 1.045 1 6
Mol7 27 17 41.3 0.867 9.042 1.306 0 9
data management. SMILES can be easily generated starting from

any given chemical file format using software like OpenBabel
[121] or obtained manually through dedicated software (see
Note 1). The SMILES strings of our example molecules are pro-
vided in Table 4.
3.2 Descriptor This section will define the selected descriptors and provide the
Calculation users with a step-by-step guide on how to calculate them manu-
ally starting from a 2D representation of the molecule (e.g.,
molecular graph). Where necessary, some of the example mole-
cules will be used and the descriptor values for all of them will
also be provided.
Amongst the high number of software for molecular descriptor
calculation, we chose Dragon 7 [110], which is regarded as a
benchmark software and can compute more than 5000 0D to 3D
descriptors. However, the descriptors can be calculated with other
software (see Note 2) or manually, and the considerations drawn
here always apply.
The models include in total eight descriptors, namely: nBM,
nBO, C%, SIC3, ATSC4p, ATSC6i, nROH, and NaasC. Their
numerical values for the example molecules are collected in Table 5.
The MDs can be divided into four logical blocks, according to
their chemical meaning and the molecular representation they
derive from, namely (1) constitutional indices (nBM, nBO, C%,
and nROH), (2) autocorrelation descriptors (ATSC4p and
ATSC6i), (3) indices of neighborhood symmetry (SIC3), and (4)
atom-type E-state indices (NaasC). The descriptor theory, the
calculation formulas, and some numerical examples will be pre-
sented according to this logical division.
3.2.1 Constitutional Constitutional descriptors are the simplest descriptors. They reflect
Descriptors the chemical composition of a compound, without encoding any
Table 6
Calculation example of the indices of neighborhood symmetry for 2-oximino-3-butanone (Mol3).
Equivalent vertices, their probabilities, and IC and SIC descriptor values, from order 0 to order 3. The
process of vertex partitioning can be found in Fig. 10
Order Descriptor
(m) Equivalent vertices Probability (p) values
0 [C1, C2, C4, C7]; [O3, O6]; 4 2 1 7 IC0 = 1.689
[N5]; [H8, H9, H10, H11,  14  ;  14  ;  14  ;  14  SIC0 = 0.444
       
H12, H13, H14]
1 [C1, C7]; [C2]; [C4]; [O3]; 2 1 1 1 1 1 1 IC1 = 2.557
[O6]; [N5]; [H11]; [H8,  14  ; ; ; ; ; ; ; SIC1 = 0.672
   14   14   14   14   14   14 
H9, H10, H12, H13, H14]
6
 14 
 
2 [C1]; [C2]; [C4]; [C7]; [O3]; 1 1 1 1  1   1   1  IC2 = 2.700
[O6]; [N5]; [H11]; [H8,  14  ;  14  ;  14  ;  14  ;   ;   ;   ; SIC2 = 0.709
         14   14   14 
H9, H10, H12, H13, H14]
1 6
 14  ;  14 
   
3 [C1]; [C2]; [C4]; [C7]; [O3]; 1 1 1 1  1   1   1  IC3 = 3.128
[O6]; [N5]; [H11]; [H8,  14  ;  14  ;  14   14  ;   ;   ;   ; SIC = 0.822
     ;    14   14   14  3
H9, H10]; [H12, H13,
H14] 1 3 3
 14  ;  14  ;  14 
     
information about its molecular geometry and overall topology.

Most of them are derived from the chemical formula. Typical con-
stitutional descriptors are the molecular weight, the counts of
atoms and bonds, the absolute and relative occurrence frequencies
of specific atom- and bond-types, as well as some ring descriptors
that are calculated only considering the molecular structure com-
position. Among the constitutional descriptors, four will be used in
this chapter: (1) nBO, which is the count of non-H bonds in the
molecule (i.e., the total number of edges in the H-depleted graph),
(2) nBM, which is the number of multiple bonds (i.e., the count of
double, triple and aromatic bonds in a molecule), (3) C%, which is
the percentage of carbon atoms in the molecule, and (4) nROH,
which is the number of hydroxyl groups.
3.2.2 Autocorrelation Autocorrelation indices are based on the concept of spatial auto-
Descriptors correlation, which is a measure of the degree to which a spatial
phenomenon is correlated to itself in space, or, in other words, the
degree to which the observed value of a variable at one region
depends on values of the same variable at neighboring regions. The
spatial pattern of a property distribution is defined by the arrange-
ment of individual entities in space and the spatial relationships

among them. Analogously, molecular autocorrelation descriptors
are based on (1) a conceptual dissection of the molecular structure
into “spatial regions”, and (2) the application of an autocorrelation
function to molecular properties measured in the molecular
regions.
Let f(x) be a function of x, the autocorrelation function of
order k is defined as the integration of the products of the function
f values calculated at x and x + k. In other words, the autocorrela-
tion expresses how numerical values of the function are correlated
at intervals k. Formally, the autocorrelation function (ACk) for
ordered discrete sequences of n values f(xi) can be written as the
summation of the products of the ith value and the (i + k)th value,
as follows:
1 n −k
AC k = ⋅ ∑ ( f ( x i ) − µ ) ⋅ ( f ( x i +k ) − µ )  (4)
(n − k ) ⋅ σ 2 i =1 
where f(x) is any function of the variable x, and k is the lag repre-
senting an interval of x. σ2 is the variance of the function values,
while μ is their mean. The lag assumes values between 1 and K,
where the maximum value K can be n–1; however, in several appli-
cations, K is chosen equal to a smaller number (e.g., K ≤ 8). A lag
value of zero corresponds to the sum of the squared centered val-
ues of the function. Autocovariances are calculated in the same
way, but omitting the standardization by σ2. As visible in Eq. 4,
spatial autocorrelation indicates the extent to which the occur-
rence of one feature is influenced by similar features in the adjacent
area. Such influence exists when there is a systematic spatial varia-
tion in the values of a given variable. This variation can exist in two
forms: positive and negative spatial autocorrelation. In the case of
positive autocorrelation, the value of a variable at a given location
tends to be similar to the values of that variable in nearby locations.
If the value of a variable is low in a given location, the presence of
positive spatial autocorrelation indicates that nearby values are also
low. Conversely, negative spatial autocorrelation is characterized
by dissimilar values in nearby locations (e.g., a small value of a vari-
able may be surrounded by large values in nearby locations). A
salient feature of the autocorrelation function is its invariance to
translation and rotation, as it does not change when the origin of
the x variable is shifted.
Autocorrelation descriptors of chemical compounds are calcu-
lated by using molecular properties of different complexity, such as
atomic- or molecular surface-based levels. These molecular descrip-
tors inherit the same invariance property of the autocorrelation func-
tion, and are thus, independent of any translation and rotation.
When a molecular graph is used as input representation, the
lag k coincides with the topological distance between any pair of
vertices, that is, the number of bonds along the shortest path con-
necting two atoms. Moreau and Broto [122–124] are the first who
applied an autocorrelation function to the molecular graph to mea-
sure the distribution of atomic properties on the molecule t opology.
The final vector of autocorrelation functions at different lags (k)
was defined by the authors as autocorrelation of a topological
structure (ATS). ATS at a given k (ATSk) can be calculated as
follows:
nAT −1 nAT
ATSk = ∑ (
∑ wi ⋅ w j ⋅ δ dij ;k ) (5)
i =1 j =i +1

where w is any atomic property, nAT is the number of atoms in a
molecule, k is the lag, and dij is the topological distance between
the ith and jth atom; δ(dij; k) is a Dirac-delta function equal to 1 if
dij = k, zero otherwise.
The centered Broto–Moreau autocorrelations are calculated
by replacing atomic properties (w) with their centered values (w’),
which are derived by subtracting the average property value w of
the molecule from each w value:
nAT −1 nAT nAT −1 nAT
ATSCkw = ∑ ∑ ( ) (
| (wi − w ) | ⋅ | w j − w | ⋅δ dij ; k = ) ∑ ∑w ⋅w ′
i
′
j (
⋅ δ dij ; k ) (6)
i =1 j =i +1 i =1 j =i +1

Hollas [125] demonstrated that, only if properties are centered, all
autocorrelation descriptors are uncorrelated, thus resulting more
suitable for the subsequent statistical analysis.
The selected autocorrelation descriptors for this example are
ATSC4p and ATSC6i. They are the centered Broto–Moreau auto-
correlation descriptors, respectively of lag 4 weighted by polariz-
ability (p) and of lag 6 weighted by ionization potential (i). A
calculation example is reported in Fig. 9 for 2-oximino-3-butanone
(Mol3). The molecule is represented by the H-filled molecular
graph (Fig. 9a), which shows the chemical identity and the sequen-
tial identification number of the vertices. The topological distance
matrix (Fig. 9b) can be used to find the pairs of atoms that enter
the summation. To calculate the centered Broto–Moreau autocor-
relation of lag 4 (ATSC4p) with polarizability as the atomic weight-
ings (Fig. 9c), one has to select the atom pairs with distance 4
(highlighted in red boldface into the matrix) and add the products
of the corresponding polarizability values (Fig. 9d).
3.2.3 Indices Indices of neighborhood symmetry are topological informa-

of Neighborhood Symmetry tion indices calculated from the H-filled molecular multigraph
and based on the concepts of neighbor degrees and edge mul-
tiplicity [126, 127]. They account for the equivalence rela-
tionships amongst the topological neighborhood of atoms.
The topological neighborhood of order m of an atom (vertex)
can be thought of as an open sphere comprising all the vertices
Fig. 9 Calculation of the centered Broto–Moreau autocorrelation for 2-oximino-3-butanone (Mol3): (a) repre-
sentation of the molecule as an H-filled molecular graph; (b) topological distance matrix, atom-pairs with a
topological distance of 4 and of 6 are highlighted in red and blue boldface, respectively; (c) atomic weightings,
where w refers to carbon-scaled atomic weightings (polarizability [p] and ionization potential [i]) and w’ to
centered atomic weightings on the molecule mean; (d) example of descriptor calculation (ATSC4p and ATSC6i),
using the obtained w’ values
with a through-bond distance from the considered vertex

equal to m. Two atoms are considered as “topologically equiv-
alent” at the mth order, if they have the same topological
neighborhood of mth order.
The indices of neighborhood symmetry, in particular, are cal-
culated by partitioning graph vertices into equivalence classes of
different orders. According to what previously said, two vertices vi
and vj of a multigraph are topologically equivalent at the mth order
if and only if: (1) they have the same chemical identity (i.e., they
are the same element), (2) they have the same vertex degree (i.e.,
same number of bonded non-hydrogen atoms), and (3) at each
mth order path starting from vi there is a mth order path starting
from vj having the same conventional bond order of the edges in
Fig. 10 Neighborhood symmetry for 2-oximino-3-butanone (Mol3). Illustrative process of vertex partitioning
into equivalence classes considering a neighborhood of order 0 to 3. Note that for the sake of simplicity, the
unitary equivalence classes (i.e., classes comprised of only one vertex), which appear in one level, are not
repeated in the subsequent level of the partition process. The calculated class probabilities, along with the
corresponding IC and SIC values, are reported in Table 6
the path and the same chemical element and vertex degree of the
involved vertices. In other words, two atoms vi and vj are equiva-
lent at the mth level if it is possible to obtain two equal sub-
structures (in terms of connected atoms and bonds, through paths
of maximum m bonds), one rooted in vi and one in vj,
respectively (Fig. 10).
Once the equivalence classes of the vertices have been deter-
mined from the H-filled multigraph, for each mth order (usually
m ranges from 0 to 5), the neighborhood Information Content
(ICm) is calculated by Shannon’s entropy formula as follows:
G Ag Ag G
ICm = − ∑ ⋅ log 2 = − ∑ p g ⋅ log 2 p g (7)
g =1 nAT nAT g =1

where the summation goes over the G atom equivalence classes, Ag
is the cardinality of the gth equivalence class (i.e., the number of
atoms grouped in the same class), nAT is the total number of ver-
tices (i.e., atoms) and pg can be thought of as the probability of
randomly selecting a vertex of the gth class. ICm represents a mea-
sure of structural complexity per vertex. The larger the number of
equivalence classes at a given mth order, the higher the ICm value.
Its normalized counterpart is called structural information content
(SICm) and is calculated as follows:
ICm
SICm = (8)
log 2 nAT
SICm is normalized according to a function of the number of

atoms, thus it being independent of the molecular size.
The crucial point to calculate information indices is to deter-
mine the vertex equivalence classes. An example of vertex parti-
tioning into the equivalence classes is shown in Fig. 10 for Mol3.
As it can be observed, for Mol3, which is small and not very com-
plex, the third order is the one leading to the highest possible
number of equivalence classes.
The KNN model under analysis, which will be illustrated in
Subheading 3.4, in particular, uses the Structural Information
Index of 3rd order (i.e., SIC3), along with other molecular
descriptors.
3.2.4 Atom-Type E-State Atom-type electrotopological state (or E-state) indices are molec-
Indices ular descriptors that encode topological and electronic informa-
tion related to particular atom-types in the molecule. They
combine structural information about (1) the electron accessibil-
ity associated with each atom-type, (2) the presence or absence of
a given atom-type and (3) counts of the atoms of a given atom-
type [128, 129].
To compute atom-type E-state indices, the first step is to
identify precise atom-types in the molecule, according to three
factors:
(a) The atom identity, based on the atomic number Z;
(b) The valence state indicator (VSI), which can be calculated as
the sum between the vertex degree and the valence vertex
degree (VSI = δv + δ);
(c) The aromatic indicator (IAR), which is equal to 1 if the atom
belongs to an aromatic system, while 0 otherwise.
Once the atoms in the molecule are assigned their specific
atom-types, two different atom-type E-state indices can be com-
puted: the atom-type E-state sums and the atom-type E-state
counts, as described below.
The atom-type E-state sums are calculated by adding the
electrotopological state (Si) of all the atoms of the same atom-type
in the molecule. The electrotopological state (or E-state) is an
atomic index encoding information related to the electronic and
topological state of the atoms in the molecule [130]. It is calcu-
lated from the H-depleted molecular graph as follows:
V Ii − I j
Si = I i + ∆I i = I i + ∑ (9)
(d )
k
j =1
ij +1

where Ii is the intrinsic state of the ith atom and ΔIi is the field
effect on the ith atom calculated as perturbation of the intrinsic
state of ith atom by all other non-H atoms in the molecule (V); dij
is the topological distance between the ith and the jth atoms. The
exponent k is a parameter to modify the influence of distant or
nearby atoms for particular studies. Usually, k = 2. The intrinsic
state Ii of the ith atom is calculated as follows:
( 2 / Li )
2
⋅ δiv + 1
Ii = (10)
δi

where Li is the principal quantum number (i.e., 2 for C, N, O, F
atoms, 3 for Si, S, Cl, etc.), while δi and δiv are the simple and the
valence vertex degrees, respectively. For any atom, the intrinsic
state can be thought of as the ratio of π and lone pair electrons over
the count of the σ bonds in the molecular graph. Therefore, the
intrinsic state reflects the possible partitioning of non-σ electrons
influence along the paths, starting from the considered atom. The
smaller the partitioning of the electron influence, the more avail-
able are the valence electrons for intermolecular interactions.
The atom-type E-state counts are also based on the assign-
ment of the molecule atoms to the given atom-types. However,
unlike the previous ones, the atoms of the same atom-type in a
molecule are simply counted [131].
The symbol of each atom-type E-state index is composed of
three parts. The first part is “S” or “N”, depending on whether the
E-states of the atoms of the same type are summed up or simply
counted. The second part is a string representing the bond-types
associated with the atom (“s”, “d”, “t”, “a” for single, double,
triple, and aromatic bonds, respectively). The third part is the sym-
bol identifying the chemical element and the eventual bonded
hydrogens, such as CH3, CH2, and F.
In the case of Mol5, as illustrated in Fig. 11, seven different atom-
types are identified according to the mentioned criteria (i.e., atomic
number [Z], valence state indicator [VSI] and aromatic indicator [IAR]
of non-H atoms). The atom-types are –Br, --C(–)--, --CH--, –O–,
=C(–)–, =O, --N--, where the letter indicates the chemical element
and the symbols “–“, “=” and “--” indicate single, double and aro-
matic bonds, respectively. To each atom-type there corresponds an
atom-type E-state count (Fig. 11), namely: NsBr = 1, NaasC = 3,
NaaCH = 8, NssO = 1, NdssC = 1; NdO = 1, and NaaN = 1. To
obtain the atom-type E-state sums, the electrotopological states Si of
the atoms of the same type are instead summed up; thus, for instance,
the descriptor SaasC = 1.8006 derives from the summation of the
Fig. 11 H-depleted molecular graph of 3-Bromophenyl nicotinate (Mol5) and the atomic indices for atom-type
assignment and E-state descriptor calculation. The atom-type was represented as the chemical element plus
the specification of its bond(s), i.e., single (−), double (=) and aromatic (--) bonds
Fig. 12 Graphical representation of the CART model. Rounded rectangles denote univariate splits (i.e., nodes)
according to the selected descriptors and thresholds, while round boxes (i.e., leafs) denote the assigned class.
Yes (y) and no (n) indicate whether the condition specified at each node is satisfied or not by any molecule to
be predicted
S-states of the atoms 2, 6, and 10. In our guided exercise, the descrip-
tor involved in the KNN model is NaasC; its values for the six example
molecules in analysis are collected in Table 5.
3.3 Predictions The CART model used in this chapter is depicted in Fig. 12. It is
with the CART Model comprised of three constitutional descriptors (nBM, nBO, and
nROH) and constituted by four nodes (Table 5).
The model can be written as a pseudo-code as follows:
if nBMi < 13: predicted_classi = “inactive”
else:
if nROHi < 1:
if nBMi > 16:

if nBOi < 25: predicted_classi =
“inactive”

else: predicted_classi = “active”

else: predicted_classi = “active”
else: predicted_classi = “inactive”
where i denotes the ith compound, while nBMi, nBOi, nROHi its
descriptor values, and predicted_classi the predicted activity,
respectively.
The advantage of CART models lies in their being easily imple-
mented and manually applicable, if necessary. In a MATLAB
environment, for instance, this can be done as follows:
n = size(X,1); % number of molecules to be predicted
class = zeros(n,1); % preallocation of predicted
class
for i = 1:n % runs over the molecules
if X.nBM(i) < 13; class(i,1) = 2; % node 1
else
if X.nROH(i) >= 1; predicted_class(i,1) = 2; % node 2
else
if X.nBM(i) > 16; predicted_class(i,1) = 1; %
node 3
else
if X.nBO(i) < 25; predicted_class(i,1) = 2;
% node 4
else predicted_class(i,1) = 1;
end
end
end
end
end
where X is the numerical matrix containing the descriptor values,

and predicted_class contains the predicted class (1 = active; 2 = inac-
tive) of each compound.
Figure 13 shows the application of the CART model to predict
the activity of the example molecules. For each prediction, the
used nodes, along with the associated molecular descriptors/
thresholds, and the branches/leaves are highlighted.
Fig. 13 Application of the CART model to the example molecules (as indicated by boldface titles), along with
their descriptor values, the predicted class and the set of nodes, branches and leaves used for the prediction.
In analogy with Fig. 12, y [yes] indicates that the condition specified in the node is satisfied, while n [no] that
it is not satisfied
3.4 Predictions The KNN model under analysis uses six molecular descriptors,
with the KNN Model namely C%, nROH, SIC3, ATSC4p, ATSC6i, and NaasC, which
have been described in Subheading 3.2, and a number of neigh-
bors (k) equal to 14. For any new molecule to be predicted, the
neighbors, which are the most similar compounds, are identified
on the basis of their molecular descriptors, after the appropriate
data-pretreatment and Euclidean distance calculation [132], as

illustrated on a step-by-step basis in the following paragraphs.
3.4.1 Data Scaling The scaling of molecular descriptors has a crucial influence on the
outcome of many modeling techniques, especially on those based
on molecular similarity/diversity analysis. In fact, when dealing
with molecular descriptors expressed in different measuring units
(e.g., molecular weight vs. number of carbon atoms), the scaling is
necessary to have comparable descriptor ranges and to avoid biased
distance/similarity calculations [132]. The values of the relevant
molecular descriptors in a model change according to the chosen
scaling function, and, as recently shown, the data scaling can have
a major influence on the modeling output of several types of
descriptors [49].
The KNN model was trained on range scaled data, that is, each
descriptor was transformed into a new numerical scale with mini-
mum and maximum values equal to 0 and 1, respectively. This is
achieved as follows:
x ij − min j
x ’ij = (11)
max j − min j

where xij is the original-scale value of the ith molecule for the jth
descriptor, while minj and maxj are the minimum and maximum
value of the jth descriptor in the training set, respectively.
Using MATLAB programming language, the range scaling on
the training data can be performed as follows:
X_scal = (X - min(X))./(max(X)-min(X));
where X is the original data matrix, in which each row is a molecule
and the columns correspond to the selected molecular descriptors,
and X_scal is the range-scaled data matrix. The effect of the data
scaling on the descriptor comparability can be then investigated
using a boxplot (Fig. 14), as follows:
figure;
subplot(1,2,1); boxplot(X); title 'Original data'
subplot(1,2,2); boxplot(X_scal); title 'Scaled data'
In the case of new molecules to be predicted, their descriptors

need to be scaled in the same way as the training molecule descrip-
tors. This is achieved by using the training set minimum and maxi-
mum values for each descriptor, as follows:
x ijnew − min j
x ijnew’ = (12)
max j − min j

where x ijnew is the value of the ith new molecule for the jth descrip-
tor, while minj and maxj are the minimum and maximum value of
Fig. 14 Effect of the range scaling procedure on the measure scales of the selected molecular descriptor:
(a) original data, (b) range-scaled data. Boxplots show median, 1st and 3rd quartiles (solid lines), and
minimum/maximum values (asterisks). Grey dots represent the underlying descriptor values for the train-
ing set molecules
the jth descriptor in the training set of molecules (as in Eq. 11),

respectively.
In MATLAB, the range scaling of the descriptors for new mol-
ecules (i.e., external molecules that are not included in the training
set) can be performed as follows:
X_scal_new = (X_new - min(X))./(max(X)-min(X));
where X_new is the MATLAB array containing the (p) descriptor
values for all the new molcules (nnew x p), while X is the training set
descriptor matrix (ntrain x p) and X_scal_new is the range scaled data
matrix of the new molecules (nnew x p). The Python code for train-
ing/external set scaling and result visualization can be found in
Note 3.
The scaled-descriptor values for the example molecules are
reported in Table 7, along with the training set parameters.
3.4.2 Distance Molecular similarity is often expressed through a distance measure

Calculation (the smaller the distance, the higher the similarity), calculated by
pairwise comparisons of the descriptor values [132]. Like the case
of descriptor scaling, also the choice of the distance metric can
affect how the chemical similarity between molecular entities is
perceived (e.g., [49, 133]).
The selected KNN model, in particular, relies on the Euclidean
distance to determine the closest compounds to a given molecule
(i.e., those with the minimum distance to the target molecule),
which can be mathematically formalized as follows:
36
Table 7
Original and range-scaled KNN descriptor values for the example molecules, along with minimum and maximum training set values
Francesca Grisoni et al.
Original data Range-scaled data
ID C% SIC3 ATSC4p ATSC6i nROH NaasC C% SIC3 ATSC4p ATSC6i nROH NaasC
Mol1 41.3 0.867 9.042 1.306 0 9 0.666 0.789 0.132 0.177 0.000 0.45
Mol2 39.5 0.676 12.844 0.880 0 4 0.637 0.486 0.187 0.119 0.000 0.20
Mol3 28.6 0.822 0.936 0.008 1 0 0.461 0.717 0.014 0.001 0.083 0.00
Mol4 41.9 0.902 7.733 0.829 2 4 0.676 0.845 0.113 0.112 0.167 0.20
Mol5 50.0 1.000 4.169 0.295 0 3 0.806 1.000 0.061 0.041 0.000 0.15
Mol6 54.1 0.808 8.346 1.045 1 6 0.871 0.695 0.122 0.142 0.083 0.30
min* 0.0 0.370 0.000 0.000 0 0 0.000 0.000 0.000 0.000 0.000 0.00
max* 62.0 1.000 68.590 7.386 12 20 1.000 1.000 1.000 1.000 1.000 1.00
*Training set parameters used for range-scaling

Table 8
Scaled descriptor values for the two molecules (Mol1 and Tr1) that are used in the calculation example
of Euclidean distance (see text)
Scaled descriptor values
ID 2D Structure C% SIC3 ATSC4p ATSC6i nROH NaaSC

Mol1 0.666 0.789 0.132 0.177 0.000 0.45
Tr1 0.440 0.000 0.050 0.083 0.000 0.15
p
( )
2
dit = ∑ x ij − x tj (13)
j =1

where dit is the Euclidean distance between molecules i and t, p is
the number of molecular descriptors, while xij and xti are the values
of the jth descriptor for the molecules i and t, respectively. Whenever
one wants to apply the KNN model, it is necessary to calculate the
distance of any new molecule from all of the training set molecules.
Note that, for the reasons explained above, the data used to com-
pute the distance must be scaled.
Using the KNN molecular descriptors, the Euclidean distance
between any ith new molecule to be predicted and any tth training
molecule, can be obtained as follows:

(C %i − C %t ) + ( SIC 3i − SIC 3t ) + ( ATSC 4 pi − ATSC 4 pt )
2 2 2
dit =
+ ( ATSC 6ii − ATTSC 4it ) + (nROHi − nROHt ) + ( NaasCi − NaasCt )
2 2 2
For instance, when considering the molecules reported in Table 8

(Mol1 and Tr1), their Euclidean distance is calculated as follows:

( 0.666 − 0.440 ) + ( 0.789 − 0.000 ) + ( 0.132 − 0.050 ) + ( 0.177 − 0.083 )
2 2 2 2
dit =
+ ( 0.00 − 0.00 ) + ( 0.45 − 0.15)
2 2
= 0.2262 + 0.7892 + 0.0822 + 0.0942 + 0.302

= 0.88
Note that the distance of i from t (dit) is necessarily equal to

the distance of t from i (dti).
In a MATLAB environment, the Euclidean distance between
the scaled training set descriptors (X_scal) and the new molecules
can be computed using the function pdist2, as follows:
D = pdist2(X_scal_new,X_scal)
where X_scal_new is the MATLAB array containing the range-

scaled descriptor values for all the new molcules (nnew x p), while
X_scal is the training range-scaled data matrix (ntrain x p). Note that
the pdist2 function is included in the Statistics and Machine
Learning Toolbox of MATLAB.
The output (D) is a matrix with dimensions (nnew × ntrain), that
is, it has as many rows as the number of new molecules (nnew), and
as many columns as the number of training set molecules (ntrain).
Entries will have a row-to-row correspondence with the new mol-
ecules and a column-to-row correspondence with the training set
molecules. Each cell will contain the distance values for the cor-
responding pair of new and training molecules. The larger the
distance, the lower the molecular similarity according to the

selected descriptors.
3.4.3 Nearest-Neighbor- According to the KNN strategy, once the descriptors have been
Based Prediction used to compute the distances between any new molecule and all
of the training set compounds, the k closest neighbors are identi-
fied, along with their experimental class. The neighbors’ experi-
mental class is then used to predict the class of the new molecule,
using a majority voting criterion, that is, by selecting the class that
is the most frequent amongst the neighbors. In the case of Mol1,
for instance, 9 out of 14 neighbors are active and, thus, the mole-
cule was predicted as active by the KNN model (Table 9).
In MATLAB, this can be carried out as follows:
% selects the neighbors according to D
[D_sort,ind_sort] = sort(D,2); % sorts D (ascending) column-
wise
D_sort = D_sort(:,1:k); % distance of the k neighbors
ind_sort = ind_sort(:,1:k); % numerical identifiers
of the k neighbors
class_neigh = class(ind_sort); % class of the k neighbors
% counts the frequency of each class amongst the k
neighbors
G = max(class); % number of classes
for i = 1:G % runs over all the classes

freq(i) = sum(class_neigh == i,2)/k;
end
% assigns the class and checks the frequency
[freq_neigh,predicted_class] = max(freq,[],2);
In the above code, the class is represented with consecutive

numbers, from 1 to G, where G is the number of classes (i.e., 2 in
our guided exercise). Note that the chosen numeric class label does
not affect the final classification. The array “freq” allows users to
check how frequently any class occurs amongst the neighbors, the
closer to 1, the more in agreement the neighbors’ classes are and,
thus, the more reliable the prediction.
In some cases, two or more classes may equally occur amongst
the neighbors of a given molecule. For binary classification tasks,
this may happen only for even k (as in this case, where k = 14),
while with more than two classes, this could happen indepen-
dently of the chosen k. In the case of ties, one option is to assign
the molecule to the class whose neighbors have the smallest aver-
age distance. The presuppose is that, amongst the classes that are
equally represented within the neighborhood, the most likely one
for the new compound will be the one of the class that is on aver-
age chemically more similar to the compound.
The MATLAB code to predict the final class of any new mol-
ecule using the KNN algorithm in the presence of ties is the
following:
% no. of classes with same frequency
check_frequency = sum(freq == freq_neigh,2);
% index of molecules with ties
mol_ties = find(check_frequency > 1);
% check the distance and correct the prediction
for i = 1:numel(mol_ties)
mol = mol_ties(i); % index of the molecule
for j = 1:G % mean distance of each class
d_mean(j) = mean(D_sort(mol,class_neigh(mol,:) ==
j));
end
% correct the predictions according to the closest
class
[~,closer_class] = min(d_mean);
predicted_class(mol) = closer_class;
end
For any new molecule, the class-relative frequencies of the 14

neighbors can be graphically compared through a bar plot (Fig. 15),
as follows:
figure
label = {'Active','Inactive'};
bar(freq(i,:));
ylabel 'Frequency'
ylim([0 1.1])
title (label(i))
Table 9
The 14 training neighbors of Mol1, sorted according to their distance (d ), along with their 2D depiction
and the experimental class. Since 9 out of 14 neighboring molecules are active (64%), Mol1 is predicted
as active by the KNN model
2D structure d Class 2D structure d Class

0.0521 Active 0.0638 Active
0.0526 Inactive 0.0658 Active
0.0531 Active 0.0682 Inactive
0.0578 Inactive 0.0717 Inactive
0.0590 Inactive 0.0766 Active

Fig. 15 Bar plot of the relative frequencies of the neighbors’ classes for all of the example molecules. Bold
numbers represent the example molecules (from Mol1 to Mol6), while bold titles (“Active” or “Inactive”)
represent the final prediction according to the KNN model
Note that the frequency of the neighborhood class adds a

valuable information regarding the reliability of the prediction.
For instance, Mol3 and Mol4 have 100% inactive neighbors,
thus, the prediction (inactive) could be considered as more reli-
able than that of Mol5, whose neighbors are 57% active and 43%
inactive, respectively. In addition to the information regarding
the frequency, also the distance of the neighbors is useful to
gather insights into the prediction reliability. For instance, a pre-
diction made with very far neighbors, that is, very dissimilar
neighbors, would be less reliable than a prediction made with
very similar neighbors. For similarity-based methods, this can be
evaluated by delimiting the chemical space reliability (AD) with a
threshold on the neighbors’ distance (e.g., [30, 40, 134–136]).
The Python code for distance calculation and KNN-based predic-
tions can be found in Note 4.
Fig. 16 List of the two most similar training molecules to Mol5, according to the type molecular descriptors
used (Euclidean distance). Descriptor values and corresponding distance (d) are reported for each conceptual
group of descriptors. When more than one descriptor was used, the data were range scaled as explained in the
text (Eq. 11). Note that, in the case of one descriptor considered (i.e., NaasC and SIC3) more than two training
molecules had a distance equal to 0 from Mol5, but, for the sake of simplicity, only two randomly selected
molecules were shown
3.5 Bringing it all As shown in the previous sections, molecular descriptors grasp
Together diverse characteristics of the molecular structure. This reflects in the
information captured by structure–activity relationship models.
To better understand the effect of the chosen molecular
descriptors on the perceived similarity between molecules, the
Euclidean distance from one example molecule (Mol5) and all of
the training compounds was computed on each conceptual group
of molecular descriptors separately. Figure 16 depicts the two most
similar training molecules to Mol5 according to the chosen set of
descriptors.
For instance, since constitutional indices reflect only the infor-
mation about the chemical composition and the single atom con-
nectivity, the identified neighbors share with Mol5 the same
number of multiple bonds (nBM = 13) and the same percentage of
carbon atoms (C% = 50.0), as well as similar number of non-
hydrogen bonds (nBO = 17 and 16). As no higher-level informa-
Fig. 17 First two components of the Principal Component Analysis (PCA) performed on constitutional descrip-
tors (a, c) and autocorrelation descriptors (b, d). EV indicates the percentage of variance explained by each
component. (a) score plot obtained using constitutional descriptors; (b) score plot obtained using autocorrela-
tion descriptors; (c) loading plot of constitutional descriptors; (d) loading plot of autocorrelation descriptors.
Compounds are colored according to their activity against CYP3A4 (red = active, grey = inactive)
tion is accounted for, the selected molecules are similar to Mol5

from a 0- and 1D-point of view, but differ from it in terms of
shape, branching, cyclicity, and bulk/steric effects. On the con-
trary, the information about the molecular shape is well captured
by SIC3 (information indices), which allows one to identify two
neighbors with a similar basic scaffold.
The effect of molecular descriptors in determining the chemi-
cal space can also be observed through a Principal Component
Analysis (PCA) [137]. PCA is a statistical technique that linearly
combines the original variables (in our case, the molecular descrip-
tors) to obtain new orthogonal variables, termed principal compo-
nents (PCs). PCs are determined in such a way that the first PC
explains the largest data variance, the second one (orthogonal to
the first) the second largest variance, and so on, up to a number of
components equal to the number of starting variables. The objects
(molecules) can be then projected into the new PC space. Thereby,
one can comprehend the linear relationships among the original
variables, the objects, and the PCs, through: (1) the scores, which
are the object (i.e., molecule) coordinates in the PC space, and (2)
the loadings, which represent the contribution of each variable
(descriptor) to each PC. The MATLAB code for performing the
PCA can be found in Note 5. The PCA was performed on consti-
tutional (i.e., nROH, nBO, nBM, and C%) and autocorrelation
descriptors (i.e., ATSC6i and ATSC4p) separately (Fig. 17).
The space determined by the first two principal compo-
nents explains 83% and 100% of the total variance for constitu-
tional and autocorrelation descriptors, respectively. In the case of
autocorrelation descriptors, 100% of the variance is explained as
expected, due to the use of two initial variables, which are com-
bined in maximum number of two PCs. As it can be noticed, the
molecular information captured by the two molecular descriptor
types are very different, in terms of compound spatial distribution
and the relative positioning of active and inactive compounds. For
instance, in the chemical space defined by constitutional descrip-
tors (Fig. 17a), active compounds are more clustered in the center
of the PC space than inactive compounds, the latter having smaller
scores on PC1 and larger scores on PC2. On the contrary, when
autocorrelation descriptors are used, this separation is less visible,
and only a few of inactive compounds isolate with very high scores
on PC1 and very small scores on PC2.
The information encoded by molecular descriptors can be lev-
eraged to gather some insights into the structure–activity relation-
ships. In the case of PCA, for instance, the compounds distribution
can be interpreted using the loading plot (Fig. 17c and d): the
loadings represent the linear coefficients used to generate the new
PCs starting from any variable value. High loadings (in absolute
values) indicate that the descriptor has a great contribution in
determining a given PC, while loadings close to 0 indicate that a
given variable is not relevant for that PC.
In the case of constitutional descriptors (Fig. 17c), for instance,
we can observe that nBM, C% and nBO have positive loadings on
PC1: this means that compounds with positive scores on PC1 will
be characterized by a high number of multiple bonds (nBM), a
high percentage of carbon atoms (C%) and a high number of non-
hydrogen bonds (nBO). As nROH and nBO have positive load-
ings on PC2, compounds with high PC2 scores will have a higher
number of hydroxyl groups and of non-hydrogen bonds than
compounds with low scores on PC2. This information, combined
with the relative distribution of active and inactive compounds on

the score plot, allows one to infer some information regarding the
activity. Many inactive molecules, which have higher PC2 scores
than active compounds and lower PC1 scores than active com-
pounds, are, for instance, characterized by more hydroxyl groups
(higher nROH), more heteroatoms (smaller C%) and less multiple
bonds (smaller nBM) than the active compounds.
In the case of autocorrelation descriptors (Fig. 17d), both
ATSC6i and ATSC4p have positive loadings on PC1, that is, the
higher their values, the higher the compound score on PC1. On
the contrary, PC2 explains the differential contribution of the two
descriptors: high ATSC6i values correspond to high PC2 scores
and, vice versa, high ATSC4p values correspond to low scores on
PC2. As already observed, when projecting molecules in this PC
space, the separation between active and inactive groups is not
much apparent. However, a few of inactive compounds are easily
distinguished in the bottom-right region of the space, since they
have relatively higher ATSC4p values than the remaining com-
pounds, along with lower values of ATSC6i. The comparison
between the two PCAs obtained on the diverse molecular descrip-
tions allows one to understand how each numerical representation
leads to a different perception of the chemical space and of similar-
ity/diversity relationships between the molecules, such as, for
example, between actives and inactives.
The chapter’s considerations highlight the role of molecular
descriptors in capturing distinct types of structural features and the
importance of selecting the best subset according to the problem
under analysis.
For every modeling/analysis purpose, our recommendation is
to test and evaluate the largest number of molecular descriptors
available, to explore the largest amount of chemical information
possible and then identify the most suitable set. Molecular descrip-
tor interpretability and derived biological/chemical insights are
also key factors to consider when choosing amongst structure–
activity models with the same performance.
4 Notes
1. Molecular representations and chemical file formats. The gener-

ation of molecular representations can be done through several
molecular drawing software/websites, such as Marvin Sketch
[138] or eMolecules (https://www.emolecules.com/). The
Chemical Identifier Resolver [139] allows to convert a given
molecule identifier (e.g., Name, CAS) into a structural repre-
sentation (or a different molecular identifier) in an automated
way. In addition, many chemical file formats exist, but some-
times not all of them are compatible with the descriptor calcu-
lation software. Usually, an easy way to overcome this issue is
using dedicated software, such as OpenBabel [121], to trans-

late of diverse formats into the needed ones. Users should keep
in mind that the conversion to some chemical file formats may
lead to a loss of information, such as the conversion from for-
mats containing 3D information to 2D representations (e.g.,
from MDL Molfile [140] to SMILES notations).
2. Software for descriptor calculation. The model descriptors were
calculated by the proprietary software Dragon 7. Most of the
Dragon descriptors can be calculated using E-Dragon software
[111], which is freely available online. In addition to Dragon-
based tools, several software for molecular descriptor calculation
exist, such as (1) RDKit Descriptor calculator [141], which
allows calculating 117 0D to 2D descriptors and properties
(http://www.rdkit.org/docs/GettingStartedInPython.
html#list-of-available-descriptors); (2) RDKit (Count-based)
Fingerprints [141], for calculating four of the most common
types of binary/count-based fingerprints; (3) Fingerprints
(CDK) [142, 143], for calculating five types of well-established
fingerprints; (4) MOE descriptor calculation [144] (requires
MOE license), to calculate hundreds of the most popular 0-2D
molecular descriptors; (5) EnalosMold2 [145], to calculate
777 2D molecular descriptors.
3. Range scaling (Python). Given x as a Python N-dimensional
array containing the training set data (nmol × nMD), the training
set can be range-scaled as follows:
# range-scale data
x_scal = (x - x.min(axis=0))/(x.max(axis=0) -
x.min(axis=0))
The outcome of the scaling procedure can be then visualized
through two boxplots, with the following code:
# plot non-scaled and scaled data
plt.figure(1)
plt.subplot(121)
plt.boxplot(x)
plt.title('Original data')
plt.subplot(122)
plt.boxplot(x_scal)
plt.title('Scaled data')
plt.show()
The descriptors of new molecules or of a test/evaluation set

can be scaled as follows:
# scaling of new molecules using training set parameters

x_new_scal = (x_new - x.min(axis=0))/(x.max(axis=0) - x.min(axis=0))
where x_new is a N-dimensional array containing the descrip-

tors of the molecules to be predicted (nne × nvariables).
4. Distance calculation and KNN-based class prediction (Python).

Given x_scal and x_new_scal as Python N-dimensional arrays
containing the scaled training set and new molecule data,
respectively, the Euclidean distance can be calculated as
follows:
from scipy.spatial.distance import cdist # pairwise distance

d = cdist(x_new, x, metric='euclidean')
where d is the nnew × ntrain Euclidean distance matrix. This
requires users to install the SciPy [146] module. The neighbor-
based predictions can be then performed as follows:
# select neighbors according to d

sort_index = np.argsort(d) # index for sorting according to D
neigh = sort_index[:,0:k] # neigh ID
d_neig = d[:,sort_index[:,0:k]] # sorted distance
# preallocation
n_new = x_new_scal.shape[0] # number of new molecules
class_neig = np.zeros(shape = (n_new,k))
G = np.amax(class_X)+1 # number of classes - 1
freq = np.zeros(shape = (n_new,G-1))
d_mean = np.zeros(shape = (n_new,2))
# extracts the class of the neighbors and makes the predictions
for i in range(n_new):
class_neig[i,:] = class_X[neigh[i,:]] # class of neighbors
for j in range(G-1):
class_here = class_neig[i,:] == (j + 1)
freq[i,j] = sum(class_here)
# make predictions according to the frequency
freq[0,j] = 1
predicted_class[i] = np.argmax(freq[i,:]) + 1
freq_used = np.amax(freq,axis = 1)
# check for the presence of ties
ties = sum(np.equal(freq[i,:],freq_used[i]))
if ties>1:
d_here = d[i,sort_index[i,0:k]] # distance from each class
d_mean[i,j] = sum(d_here[class_here])
predicted_class[i] = np.argmin(d_mean[i,j]) + 1
freq /= k
5. MATLAB code for Principal Component Analysis. Let X be the

data matrix, the PCA can be performed as follows:
% Gaussian normalization (for comparability)

X_scal = (X - mean(X)./std(X);
% pca
[load,score] = princomp(X_scal)
% loading and score plots (for visualization)
figure, scatter(load(:,1),load(:,2)); % first two PCs – loading plot
figure, scatter(score(:,1),score(:,2)); % first two PCs – score plot
Note that, to use the function “princomp,” the Statistics and

Machine Learning Toolbox of MATLAB is required. As an alterna-
tive, a freely available MATLAB toolbox with an easy-to-use
graphical interface [147] can be downloaded at Milano
Chemometrics & QSAR Research Group website (http://michem.
disat.unimib.it/chm/download/pcainfo.htm).
References
1. Schultz TW, Cronin MTD, Walker JD, Aptula cal structure curation in cheminformatics and
AO (2003) Quantitative structure–activity QSAR modeling research. J Chem Inf Model
relationships (QSARs) in toxicology: a his- 50:1189–1204
torical perspective. J Mol Struct THEOCHEM 13. Furusjö E, Svenson A, Rahmberg M,
622:1–22 Andersson M (2006) The importance of out-
2. McKinney JD, Richard A, Waller C, Newman lier detection and training set selection for
MC, Gerberick F (2000) The practice of reliable environmental QSAR predictions.
structure activity relationships (SAR) in toxi- Chemosphere 63:99–108
cology. Toxicol Sci 56:8–17 14. Mansouri K, Grulke CM, Richard AM,
3. Johnson MA, Maggiora GM (1990) Concepts Judson RS, Williams AJ (2016) An automated
and applications of molecular similarity. Wiley, curation procedure for addressing chemical
New York errors and inconsistencies in public datasets
4. Crum-Brown A, Fraser T (1868) On the con- used in QSAR modelling. SAR QSAR Environ
nection between chemical constitution and Res 27:911–937
physiological action. Part 1. On the physio- 15. Grisoni F, Consonni V, Villa S, Vighi M,
logical action of the ammonium bases, derived Todeschini R (2015) QSAR models for bio-
from Strychia, Brucia, Thebaia, Codeia, concentration: is the increase in the complex-
Morphia and Nicotia. Trans R Soc Edinb ity justified by more accurate predictions?
25:151–203 Chemosphere 127:171–179
5. Hansch C, Maloney PP, Fujita T, Muir RM 16. Goldberg DE, Holland JH (1988) Genetic
(1962) Correlation of biological activity of algorithms and machine learning. Mach Learn
phenoxyacetic acids with Hammett substitu- 3:95–99
ent constants and partition coefficients. 17. Grisoni F, Cassotti M, Todeschini R (2014)
Nature 194:178–180 Reshaped sequential replacement for vari-
6. Richardson B (1869) Physiological able selection in QSPR: comparison with
research on alcohols. Med Times Gazzette other reference methods. J Chemom
703:706 28:249–259
7. Richet M (1893) Note sur le rapport entre la 18. Cassotti M, Grisoni F, Todeschini R (2014)
toxicité et les propriétés physiques des corps. Reshaped sequential replacement algorithm:
Compt Rend Soc Biol Paris 45:775–776 an efficient approach to variable selection.
8. Wiener H (1947) Influence of interatomic Chemom Intell Lab Syst 133:136–148
forces on paraffin properties. J Chem Phys 19. Shen Q, Jiang J-H, Jiao C-X, Shen G, Yu R-Q
15:766–766 (2004) Modified particle swarm optimization
9. Platt JR (1947) Influence of neighbor bonds algorithm for variable selection in MLR and
on additive bond properties in paraffins. PLS modeling: QSAR studies of antagonism
J Chem Phys 15:419–420 of angiotensin II antagonists. Eur J Pharm Sci
10. Todeschini R, Consonni V (2009) Molecular 22:145–152
descriptors for chemoinformatics, vol 2. 20. Derksen S, Keselman HJ (1992) Backward,
Wiley-VCH Verlag GmbH, Weinheim, forward and stepwise automated subset selec-
Germany, Weinheim tion algorithms: frequency of obtaining
11. Todeschini R, Consonni V, Gramatica P authentic and noise variables. Br J Math Stat
(2009) Chemometrics in QSAR. In: Psychol 45:265–282
Comprehensive Chemometrics. Elsevier, 21. Cramer RD, Bunce JD, Patterson DE, Frank
Oxford, pp 129–172 IE (1988) Crossvalidation, bootstrapping,
12. Fourches D, Muratov E, Tropsha A (2010) and partial least squares compared with mul-
Trust, but verify: on the importance of chemi- tiple regression in conventional QSAR stud-
ies. Quant Struct Act Relat 7:18–25
22. Todeschini R, Ballabio D, Grisoni F (2016) 35. Patlewicz G, Ball N, Booth ED, Hulzebos E,
Beware of unreliable Q2! A comparative study Zvinavashe E, Hennes C (2013) Use of cate-
of regression metrics for predictivity assess- gory approaches, read-across and (Q)SAR:
ment of QSAR models. J Chem Inf Model general considerations. Regul Toxicol
56(10):1905–1913 Pharmacol 67:1–12
23. Sokolova M, Lapalme G (2009) A systematic 36. Schneider G, Neidhart W, Giller T, Schmid G
analysis of performance measures for classifi- (1999) “Scaffold-hopping” by topological
cation tasks. Inf Process Manag 45:427–437 pharmacophore search: a contribution to vir-
24. Sahigara F, Mansouri K, Ballabio D, Mauri A, tual screening. Angew Chem Int Ed
Consonni V, Todeschini R (2012) 38:2894–2896
Comparison of different approaches to define 37. Höfer T, Gerner I, Gundert-Remy U, Liebsch
the applicability domain of QSAR models. M, Schulte A, Spielmann H, Vogel R, Wettig
Molecules 17:4791–4810 K (2004) Animal testing and alternative
25. Dragos H, Gilles M, Alexandre V (2009) approaches for the human health risk assess-
Predicting the predictability: a unified ment under the proposed new European
approach to the applicability domain problem chemicals regulation. Arch Toxicol
of QSAR models. J Chem Inf Model 78:549–564
49:1762–1776 38. Mansouri K, Abdelaziz A, Rybacka A et al
26. Sabljic A (2001) QSAR models for estimating (2016) CERAPP: collaborative estrogen
properties of persistent organic pollutants receptor activity prediction project. Environ
required in evaluation of their environmental Health Perspect 124(7):1023–1033. https://
fate and risk. Chemosphere 43:363–375 doi.org/10.1289/ehp.1510267
27. Novič M, Vračko M (2010) QSAR models for 39. Sedykh A, Zhu H, Tang H, Zhang L, Richard
reproductive toxicity and endocrine disrup- A, Rusyn I, Tropsha A (2011) Use of in vitro
tion activity. Molecules 15:1987–1999 HTS-derived concentration–response data as
28. Miyao T, Arakawa M, Funatsu K (2010) biological descriptors improves the accuracy
Exhaustive structure generation for inverse- of QSAR models of in vivo toxicity. Environ
QSPR/QSAR. Mol Inform 29:111–125 Health Perspect 119:364–370
29. Munteanu RC, Fernandez-Blanco E, Seoane 40. Cassotti M, Ballabio D, Todeschini R,
AJ, Izquierdo-Novo P, Angel Rodriguez- Consonni V (2015) A similarity-based QSAR
Fernandez J, Maria Prieto-Gonzalez J, model for predicting acute toxicity towards
Rabunal RJ, Pazos A (2010) Drug discovery the fathead minnow (Pimephales promelas).
and design for complex diseases through SAR QSAR Environ Res 26:217–243
QSAR computational methods. Curr Pharm 41. Belanger SE, Brill JL, Rawlings JM, Price BB
Des 16:2640–2655 (2016) Development of acute toxicity quanti-
30. Nembri S, Grisoni F, Consonni V, Todeschini tative structure activity relationships (QSAR)
R (2016) In silico prediction of cytochrome and their use in linear alkylbenzene sulfonate
P450-drug interaction: QSARs for CYP3A4 species sensitivity distributions. Chemosphere
and CYP2C9. Int J Mol Sci 17:914 155:18–27
31. Grisoni F, Consonni V, Vighi M, Villa S, 42. Wang C, Lu GH, Li YM (2005) QSARs for
Todeschini R (2016) Investigating the the chronic toxicity of halogenated benzenes
mechanisms of bioconcentration through to bacteria in natural waters. Bull Environ
QSAR classification trees. Environ Int Contam Toxicol 75:102–108
88:198–205 43. Fan D, Liu J, Wang L, Yang X, Zhang S,
32. Cramer RD, Patterson DE, Bunce JD (1988) Zhang Y, Shi L (2016) Development of quan-
Comparative molecular field analysis titative structure–activity relationship models
(CoMFA). 1. Effect of shape on binding of for predicting chronic toxicity of substituted
steroids to carrier proteins. J Am Chem Soc benzenes to daphnia magna. Bull Environ
110:5959–5967 Contam Toxicol 96:664–670
33. Marrero Ponce Y (2004) Total and local 44. Austin TJ, Eadsforth CV (2014) Development
(atom and atom type) molecular quadratic of a chronic fish toxicity model for predicting
indices: significance interpretation, compari- sub-lethal NOEC values for non-polar nar-
son to other molecular descriptors, and cotics. SAR QSAR Environ Res 25:147–160
QSPR/QSAR applications. Bioorg Med 45. Schöning V, Hammann F, Peinl M, Drewe
Chem 12:6351–6369 J (2017) Identification of any structure-spe-
34. Bender A, Glen CR (2004) Molecular similar- cific hepatotoxic potential of different pyrroli-
ity: a key technique in molecular informatics. zidine alkaloids using random forest and
Org Biomol Chem 2:3204–3218 artificial neural network. Toxicol Sci
160(2):361–370. https://doi.org/10.1093/ ular structure. Environ Sci Technol

toxsci/kfx187 41:2833–2839
46. Myshkin E, Brennan R, Khasanova T, Sitnik 58. Rojas C, Todeschini R, Ballabio D, Mauri A,
T, Serebriyskaya T, Litvinova E, Guryanov A, Consonni V, Tripaldi P, Grisoni F (2017) A
Nikolsky Y, Nikolskaya T, Bureeva S (2012) QSTR-based expert system to predict sweet-
Prediction of organ toxicity endpoints by ness of molecules. Front Chem 5:53. https://
QSAR modeling based on precise chemical- doi.org/10.3389/fchem.2017.00053
histopathology annotations. Chem Biol Drug 59. Martinez-Mayorga K, Medina-Franco JL
Des 80:406–416 (2009) Chapter 2 chemoinformatics—appli-
47. Gu C, Goodarzi M, Yang X, Bian Y, Sun C, cations in food chemistry. Adv Food Nutr Res
Jiang X (2012) Predictive insight into the 58:33–56
relationship between AhR binding property 60. Sweeney MH, Mocarelli P (2000) Human
and toxicity of polybrominated diphenyl health effects after exposure to 2,3,7,8-
ethers by PLS-derived QSAR. Toxicol Lett TCDD. Food Addit Contam 17:303–316
208:269–274 61. Walker MK, Spitsbergen JM, Olson JR,
48. Tong W, Fang H, Hong H, Xie Q, Perkins R, Peterson RE (1991)
Sheehan DM (2004) Receptor-mediated tox- 2 , 3 , 7 , 8 - Te t r a c h l o r o d i b e n z o - p - d i o x i n
icity: QSARs for estrogen receptor binding (TCDD) toxicity during early life stage devel-
and priority setting of potential estrogenic opment of lake trout (Salvelinus namaycush).
endocrine disruptors. CRC Press, Boca Can J Fish Aquat Sci 48:875–883
Raton, FL, USA 62. Consonni V, Todeschini R (2012) Multivariate
49. Grisoni F, Reker D, Schneider P, Friedrich L, analysis of molecular descriptors. In: Dehmer
Consonni V, Todeschini R, Koeberle A, Werz M, Varmuza K, Bonchev D (eds) Statistical
O, Schneider G (2017) Matrix-based molecu- modelling of molecular descriptors in QSAR/
lar descriptors for prospective virtual com- QSPR. Wiley-VCH Verlag GmbH & Co,
pound screening. Mol Inform 36:1–7 KGaA, pp 111–147
50. Ekins S, Mestres J, Testa B (2007) In silico 63. Reutlinger M, Koch CP, Reker D, Todoroff
pharmacology for drug discovery: methods N, Schneider P, Rodrigues T, Schneider G
for virtual ligand screening and profiling. Br (2013) Chemically advanced template search
J Pharmacol 152:9–20 (CATS) for scaffold-hopping and prospective
51. Jacob L, Vert J-P (2008) Protein-ligand intertarget prediction for “orphan” molecules.
action prediction: an improved chemogenom- Mol Inform 32:133–138
ics approach. Bioinformatics 24:2149–2156 64. Fechner U, Franke L, Renner S, Schneider P,
52. Rognan D (2007) Chemogenomic Schneider G (2003) Comparison of correla-
approaches to rational drug design. Br tion vector methods for ligand-based similar-
J Pharmacol 152:38–52 ity searching. J Comput Aided Mol Des
53. Strömbergsson H, Kleywegt GJ (2009) A 17:687–698
chemogenomics view on protein-ligand 65. Basak SC, Gute BD, Grunwald GD (1997)
spaces. BMC Bioinformatics 10:1–11 Use of topostructural, topochemical, and
54. Cronin MTD, Walker JD, Jaworska JS, geometric parameters in the prediction of
Comber MHI, Watts CD, Worth AP (2003) vapor pressure: a hierarchical QSAR approach.
Use of QSARs in international decision-mak- J Chem Inf Comput Sci 37:651–655
ing frameworks to predict ecologic effects and 66. Kubinyi H (1993) 3D QSAR in drug design.
environmental fate of chemical substances. In: Theory methods and applications, vol 1.
Environ Health Perspect 111:1376–1390 Springer Science & Business Media, Berlin
55. Mansouri K, Ringsted T, Ballabio D, 67. Consonni V, Todeschini R, Pavan M (2002)
Todeschini R, Consonni V (2013) Structure/response correlations and similarity/
Quantitative structure–activity relationship diversity analysis by GETAWAY descriptors. 1.
models for ready biodegradability of chemi- Theory of the novel 3D molecular descriptors.
cals. J Chem Inf Model 53:867–878 J Chem Inf Comput Sci 42:682–692
56. Carlsen L, Walker JD (2003) QSARs for pri- 68. Nettles JH, Jenkins JL, Bender A, Deng Z,
oritizing PBT substances to promote pollu- Davies JW, Glick M (2006) Bridging chemi-
tion prevention. QSAR Comb Sci 22:49–57 cal and biological space: “target fishing” using
57. Gramatica P, Papa E (2007) Screening and 2D and 3D molecular descriptors. J Med
ranking of POPs for global half-life: QSAR Chem 49:6802–6810
approaches for prioritization based on molec- 69. Schuur JH, Selzer P, Gasteiger J (1996) The
coding of the three-dimensional structure of
molecules by molecular transforms and its 82. Watson P (2008) Naïve bayes classification
application to structure-spectra correlations using 2D pharmacophore feature triplet vec-
and studies of biological activity. J Chem Inf tors. J Chem Inf Model 48:166–178
Comput Sci 36:334–344 83. Klon AE, Diller DJ (2007) Library finger-
70. Rybinska A, Sosnowska A, Barycki M, Puzyn prints: a novel approach to the screening of
T (2016) Geometry optimization method virtual libraries. J Chem Inf Model
versus predictive ability in QSPR modeling 47:1354–1365
for ionic liquids. J Comput Aided Mol Des 84. Geppert H, Bajorath J (2010) Advances in
30:165–176 2D fingerprint similarity searching. Expert
71. Nicklaus MC, Wang S, Driscoll JS, Milne Opin Drug Discov 5:529–542
GWA (1995) Conformational changes of 85. Tetko IV, Sushko I, Pandey AK, Zhu H,
small molecules binding to proteins. Bioorg Tropsha A, Papa E, Oberg T, Todeschini R,
Med Chem 3:411–428 Fourches D, Varnek A (2008) Critical assess-
72. Klebe G, Abraham U, Mietzner T (1994) ment of QSAR models of environmental tox-
Molecular similarity indices in a comparative icity against Tetrahymena pyriformis: focusing
analysis (CoMSIA) of drug molecules to cor- on applicability domain and overfitting by
relate and predict their biological activity. variable selection. J Chem Inf Model
J Med Chem 37:4130–4146 48:1733–1746
73. Hopfinger AJ, Wang S, Tokarski JS, Jin B, 86. Zhu H, Tropsha A, Fourches D, Varnek A,
Albuquerque M, Madhav PJ, Duraiswami C Papa E, Gramatica P, Oberg T, Dao P,
(1997) Construction of 3D-QSAR models Cherkasov A, Tetko IV (2008) Combinatorial
using the 4D-QSAR analysis formalism. J Am QSAR modeling of chemical toxicants tested
Chem Soc 119:10509–10524 against Tetrahymena pyriformis. J Chem Inf
74. Andrade CH, Pasqualoto KFM, Ferreira EI, Model 48:766–784
Hopfinger AJ (2010) 4D-QSAR: perspectives 87. Guha R (2011) The ups and downs of struc-
in drug design. Mol Basel Switz ture-activity landscapes. Methods Mol Biol
15:3281–3294 672:101–117
75. Vedani A, McMasters DR, Dobler M (2000) 88. Bajorath J, Peltason L, Wawer M, Guha R,
Multi-conformational ligand representation Lajiness MS, Van Drie JH (2009) Navigating
in 4D-QSAR: reducing the bias associated structure–activity landscapes. Drug Discov
with ligand alignment. Quant Struct Act Today 14:698–705
Relat 19:149–161 89. Wassermann AM, Wawer M, Bajorath
76. Vedani A, Briem H, Dobler M, Dollinger H, J (2010) Activity landscape representations
McMasters DR (2000) Multiple-conformation for structure−activity relationship analysis.
and protonation-state representation in J Med Chem 53:8209–8223
4D-QSAR: the Neurokinin-1 receptor system. 90. Maggiora GM (2006) On outliers and activity
J Med Chem 43:4416–4427 cliffs: why QSAR often disappoints. J Chem
77. Vedani A, Dobler M (2002) 5D-QSAR: the Inf Model 46:1535–1535
key for simulating induced fit? J Med Chem 91. Eckert H, Bajorath J (2007) Molecular simi-
45:2139–2149 larity analysis in virtual screening: founda-
78. Vedani A, Dobler M, Lill MA (2005) tions, limitations and novel approaches. Drug
Combining protein modeling and Discov Today 12:225–233
6D-QSAR. Simulating the binding of struc- 92. Hu Y, Bajorath J (2012) Extending the
turally diverse ligands to the estrogen recep- activity cliff concept: structural categoriza-
tor. J Med Chem 48:3700–3703 tion of activity cliffs and systematic identifi-
79. Willett P (2006) Similarity-based virtual cation of different types of cliffs in the
screening using 2D fingerprints. Drug Discov ChEMBL database. J Chem Inf Model
Today 11:1046–1053 52:1806–1811
80. Cassotti M, Grisoni F, Nembri S, Todeschini 93. Cruz-Monteagudo M, Medina-Franco JL,
R (2016) Application of the weighted power- Pérez-Castillo Y, Nicolotti O, Cordeiro
weakness ratio (wPWR) as a fusion rule in MNDS, Borges F (2014) Activity cliffs in
ligand–based virtual screening. MATCH drug discovery: Dr Jekyll or Mr Hyde? Drug
Comm Math Comp Chem 76:359–376 Discov Today 19:1069–1080
81. Ewing T, Baber JC, Feher M (2006) Novel 94. Guha R, Jurs PC (2004) Development of
2D fingerprints for ligand-based virtual QSAR models to predict and interpret the
screening. J Chem Inf Model 46:2423–2431
biological activity of artemisinin analogues. 109. Ballabio D, Grisoni F, Todeschini R (2017)

J Chem Inf Comput Sci 44:1440–1449 Multivariate comparison of classification per-
95. McCarty LS, Dixon DG, MacKay D, Smith formance measures. Chemom Intell Lab Syst
AD, Ozburn GW (1992) Residue-based 174:33–44
interpretation of toxicity and bioconcentra- 110. Kode SRL (2016) Dragon (software for

tion QSARs from aquatic bioassays: neutral molecular descriptor calculation) version 7.0–
narcotic organics. Environ Toxicol Chem 2016–https://chm.kode-solutions.net
11:917–930 111. E-Dragon Software. http://www.vcclab.
96. Munro AW, Girvan HM, Mason AE, Dunford org/lab/edragon/. Accessed 4 Sep 2017
AJ, McLean KJ (2013) What makes a P450 112. MathWorks Inc. (2016) MATLAB R2016b.
tick? Trends Biochem Sci 38:140–150 https://it.mathworks.com/. Accessed 6 Sep
97. Gonzalez FJ (2005) Role of cytochromes 2017
P450 in chemical toxicity and oxidative stress: 113. Python. In: Python.org. https://www.
studies with CYP2E1. Mutat Res python.org/. Accessed 23 Feb 2017
569:101–110 114. Daylight Theory: SMILES. http://www.day-
98. Gonzalez FJ, Gelboin HV (1994) Role of light.com/dayhtml/doc/theor y/theor y.
human cytochromes P450 in the metabolic smiles.html. Accessed 9 Jun 2016
activation of chemical carcinogens and toxins. 115. West DB (2001) Introduction to graph theory.
Drug Metab Rev 26:165–183 Pearson, Prentice hall Upper Saddle River
99. Zanger UM, Schwab M (2013) Cytochrome 116. Weininger D (1988) SMILES, a chemical lan-
P450 enzymes in drug metabolism: regula- guage and information system. 1. Introduction
tion of gene expression, enzyme activities, to methodology and encoding rules. J Chem
and impact of genetic variation. Pharmacol Inf Comput Sci 28:31–36
Ther 138:103–141
117. Schneider N, Sayle RA, Landrum GA (2015)
100. Guengerich FP (2006) Cytochrome P450s
Get your atoms in order—an open-source
and other enzymes in drug metabolism and implementation of a novel and robust molec-
toxicity. AAPS J 8:E101–E111 ular canonicalization algorithm. J Chem Inf
101. Protein Data Bank (2013) Crystal structure Model 55:2111–2120
of CYP3A4 in complex with an inhibitor. 118. Weininger D, Weininger A, Weininger JL

PDB ID: 4NY4 (1989) SMILES. 2. Algorithm for generation
102. Veith H, Southall N, Huang R et al (2009) of unique SMILES notation. J Chem Inf
Comprehensive characterization of cyto- Comput Sci 29:97–101
chrome P450 isozyme selectivity across chem- 119. O’Boyle NM (2012) Towards a universal

ical libraries. Nat Biotechnol 27:1050–1055 SMILES representation - a standard method
103. The PubChem Project. https://pubchem.
to generate canonical SMILES based on the
ncbi.nlm.nih.gov/. Accessed 11 Sep 2017 InChI. J Cheminform 4:1–14
104. Nembri S, Grisoni F, Consonni V, Todeschini 120. Koichi S, Iwata S, Uno T, Koshino H, Satoh
R (2016) Cytochrome P450–Drug interac- H (2007) Algorithm for advanced canonical
tion dataset, available at http://michem. coding of planar chemical structures that con-
disat.unimib.it/chm/download/cyto- siders stereochemical and symmetric informa-
chrome.htm. http://michem.disat.unimib. tion. J Chem Inf Model 47:1734–1746
it/chm/download/cytochrome.htm. 121. O’Boyle NM, Banck M, James CA, Morley C,
Accessed 29 Sep 2017 Vandermeersch T, Hutchison GR (2011)
105. Breiman L, Friedman J, Stone CJ, Olshen RA Open babel: an open chemical toolbox.
(1984) Classification and regression trees. J Cheminform 3:33
CRC press 122. Broto P, Moreau G, Vandycke C (1984)

106. Daszykowski M, Walczak B, Xu Q-S et al
Molecular structures: perception, autocorre-
(2004) Classification and regression trees– lation descriptor and Sar studies: system of
studies of HIV reverse transcriptase inhibi- atomic contributions for the calculation of the
tors. J Chem Inf Comput Sci 44:716–726 n-octanol/water partition coefficients. Eur
107. Steinberg D, Colla P (2009) CART: classifi- J Med Chem 19:71–78
cation and regression trees. Top Ten 123. Broto P, Moreau G, Vandycke C (1984)

Algorithms Data Min 9:179 Molecular structures: perception, autocorre-
108. Cover T, Hart P (1967) Nearest neighbor
lation descriptor and Sar studies. Use of the
pattern classification. IEEE Trans Inf Theory autocorrelation descriptor in the qsar study of
13:21–27 two non-narcotic analgesic series. Eur J Med
Chem 19:79–84
124. Moreau G, Turpin C (1996) Use of similarity est neighbours approach to assess the applica-
analysis to reduce large molecular libraries to bility domain of a QSAR model for reliable
smaller sets of representative molecules: predictions. J Cheminform 5:27
Informatique et analyse. I. Analysis 136. Dimitrov S, Dimitrova G, Pavlov T, Dimitrova
24:M17–M21 N, Patlewicz G, Niemela J, Mekenyan O
125. Hollas B (2002) Correlation properties of the (2005) A stepwise approach for defining the
autocorrelation descriptor for molecules. applicability domain of SAR and QSAR mod-
MATCH–Commun math. Comput Chem els. J Chem Inf Model 45:839–849
45:27 137. Jolliffe IT (1986) Principal component analy-
126. Magnuson V, Harriss D, Basak S (1983)
sis and factor analysis. In: Principal compo-
Topological indices based on neighborhood nent analysis. Springer, New York, NY,
symmetry: chemical and biological applica- pp 115–128
tions. In: Chemical applications of topology 138. Marvin Sketch 5.1.11 ChemAxon, (2013).

and graph theory. Elsevier, Amsterdam, http://www.chemaxon.com
pp 178–191 139.
NCI/CADD Group, (2013) Chemical
127. Roy A, Basak S, Harriss D, Magnuson V
Identifier Resolver. Available at: http://cac-
(1984) Neighborhood complexities and sym- tus.nci.nih.gov/chemical/structure
metry of chemical graphs and their biological 140.
Dalby A, Nourse JG, Hounshell WD,
applications. Pergamon Press, New York Gushurst AK, Grier DL, Leland BA, Laufer
128. Hall LH, Kier LB, Brown BB (1995)
J (1992) Description of several chemical
Molecular similarity based on novel atom- structure file formats used by computer pro-
type electrotopological state indices. J Chem grams developed at molecular design limited.
Inf Comput Sci 35:1074–1080 J Chem Inf Comput Sci 32:244–255
129. Hall LH, Kier LB (1995) Electrotopological 141.
RDKit: Open-source cheminformatics;
state indices for atom types: a novel combina- http://www.rdkit.org
tion of electronic, topological, and valence 142. Steinbeck C, Han Y, Kuhn S, Horlacher O,
state information. J Chem Inf Comput Sci Luttmann E, Willighagen E (2003) The
35:1039–1045 chemistry development kit (CDK): an open-
130. Kier LB, Hall LH (1990) An electrotopolog- source java library for chemo- and bioinfor-
ical-state index for atoms in molecules. Pharm matics. J Chem Inf Comput Sci
Res 7:801–807 43:493–500
131. Butina D (2004) Performance of kier-hall
143. Steinbeck C, Hoppe C, Kuhn S, Floris M,
E-state descriptors in quantitative structure Guha R, Willighagen EL (2006) Recent
activity relationship (QSAR) studies of multi- developments of the chemistry development
functional molecules. Molecules kit (CDK)-an open-source java library for
9:1004–1009 chemo-and bioinformatics. Curr Pharm Des
132. Todeschini R, Ballabio D, Consonni V (2015) 12:2111–2120
Distances and other dissimilarity measures in 144. Chemical Computing Group Inc., (2013)

chemometrics. In: Encyclopedia of analytical Molecular operating environment (MOE).
chemistry. John Wiley & Sons Ltd, Hoboken 1010 Sherbooke St West Suite 910 Montr.
133. Todeschini R, Ballabio D, Consonni V,
QC Can. H3A 2R7 2014
Grisoni F (2016) A new concept of higher- 145. Hong H, Xie Q, Ge W, Qian F, Fang H, Shi
order similarity and the role of distance/simi- L, Su Z, Perkins R, Tong W (2008) Mold2,
larity measures in local classification methods. molecular descriptors from 2D structures for
Chemom Intell Lab Syst 157:50–57 chemoinformatics and toxicoinformatics.
134. Cassotti M, Ballabio D, Consonni V, Mauri J Chem Inf Model 48:1337–1344
A, Tetko IV, Todeschini R (2014) Prediction 146. SciPy.org—SciPy.org. https://www.scipy.
of acute aquatic toxicity toward Daphnia org/. Accessed 5 Sep 2017
magna by using the GA-kNN method. Altern 147. Ballabio D (2015) A MATLAB toolbox for
Lab Anim 42:31–41 principal component analysis and unsuper-
135.
Sahigara F, Ballabio D, Todeschini R, vised exploration of data structure. Chemom
Consonni V (2013) Defining a novel k-near- Intell Lab Syst 149:1–9
Chapter 2
The OECD QSAR Toolbox Starts Its Second Decade

Terry W. Schultz, Robert Diderich, Chanita D. Kuseva,
and Ovanes G. Mekenyan
Abstract
The OECD QSAR Toolbox is a computer software designed to make pragmatic qualitative and quantita-
tive structure–activity relationship methods-based predictions of toxicity, including read-across, available
to the user in a comprehensible and transparent manner. The Toolbox, provide information on chemicals
in structure-searchable, standardized files that are associated with chemical and toxicity data to ensure that
proper structural analogs can be identified. This chapter describes the advantages of the Toolbox, the aims,
approach, and workflow of it, as well as reviews its history. Additionally, key functional elements of it use
are explained and features new to Version 4.1 are reported. Lastly, the further development of the Toolbox,
likely needed to transform it into a more comprehensive Chemical Management System, is considered.
Key words OECD QSAR Toolbox, Chemical category, Data gap filling, Adverse outcome pathways,
Weight of evidence
1 Introduction
The Organisation for Economic Cooperation and Development

(OECD) (Q)SAR Application Toolbox, now referred to as the
“OECD QSAR Toolbox” or simply the ‘Toolbox’ has reached its
10th anniversary. For a number of reasons, this stand-alone in
silico workflow has become highly beneficial to regulators, indus-
try, and others. Seminal among these reasons are the following:
(a) It is freely available.
(b) It continues to be updated regularly.
(c) It was designed to be applied to issues important in assessing
the safetyof organic substances.
Briefly, the Toolbox is a computer software designed to make prag-
matic qualitative and quantitative structure–activity relationship
((Q)SAR) methods-based predictions, including read-across, avail-
able to the user in a comprehensible and transparent manner [1].
55
56 Terry W. Schultz et al.
Computational toxicology frameworks, such as the Toolbox, pro-

vide information on chemicals in structure-searchable, standard-
ized files that are associated with chemical and toxicity data to
ensure that structural analogs can be identified. In the near future,
such frameworks must be able to integrate divergent data sets,
namely, those being generated by omics methods and the next
generation of in vitro tests.
History has shown that integrating new science and modern
structure–activity into the regulatory process is being done in an
iterative fashion. This iterative integration of relevant information
will continue to expand in the future. We are likely to see coevolu-
tion of new testing methods and assessment paradigms. Advances
in science will be used as the basis to improve toxicity testing and
assessment with better biological understanding as the driver of
this improvement. Specifically, this understanding will come by
building on advances in information sciences, molecular, cellular,
and systems biology, as well as reliable new methods (e.g., high
throughput screening (HTS)). Toxicology is clearly on the brink of
changing into a science able to efficiently determine the biological
“pathways” by which chemicals are capable of exerting adverse
effects [2]. Not only will this transformation allow assessing more
chemicals faster, but it also will allow for the development of a
more robust knowledge base of how chemicals perturb biological
processes.
The aims of this chapter are to review the history of the
Toolbox, point out its advantages, summarize the current status of
the Toolbox, and discuss where further development is likely to
lead.
2 Advantages
2.1 Toolbox One of the little-known advantages of the Toolbox is the structure
Governance of its governance. From the beginning, the establishment of guid-
ing principles and workflow, as well as the development of Toolbox
modules, was under the direction of an administrative hierarchy.
While the day-to-day development was under the direction of the
primary contractor, the Laboratory of Mathematic Chemistry
http://oasis-lmc.org/, and various sub-contractors, the overall
governance was through the OECD. The OECD, via the
Environmental Health and Safety Division of the Environment
Directorate, supervised the Toolbox on a monthly basis. The
decision-making body was the OECD QSAR Toolbox Management
Group which consisted of representatives of the OECD member
countries. While this group was formed largely of regulators famil-
iar with (Q)SAR, it also included representatives of industry and
nongovernmental organizations. Financial support was provided,
The OECD QSAR Toolbox Starts Its Second Decade 57
first by the European Commission and then by the European

Chemicals Agency.
The Toolbox was developed in phases. Each phase, approxi-
mately 4 years in duration, was divided into workplans. Each work-
plan was approximately 6 months in duration. Each workplan
consisted of a number of deliverables (4–8/workplan). The topics
for each phase were developed by the OECD. For each phase, the
topics were reviewed and approved by the Management Group.
Once a phase was initiated, the Management Group met semian-
nually to review progress with each deliverable from the previous
work period. This report included demonstrating the functionality
of the deliverable and how it was mapped to the Toolbox. For each
of these deliverables, the Management Group either accepted the
work or sent it back to the contractors for further development. In
addition, at each meeting, the contractors presented a proposal for
work to be done in the subsequent 6 months. Again, the
Management Group accepted or rejected the proposal. This some-
what elaborate system was designed to ensure the Toolbox contin-
ued to be developed with the aim of providing an instrument that
was useful to regulators in their day-to-day work.
2.2 Toolbox Training Another key feature contributing to the success of the Toolbox is
the continuous development and expansion of training material
(http://oasis-lmc.org/products/software/toolbox/toolbox-sup-
port.aspx). New training exercises in the form of slide shows con-
tinue to be developed. Additionally, LMC provides on-site training
sessions, as well as regular courses (http://oasis-lmc.org/prod-
ucts/software/toolbox/toolbox/training.aspx).
2.3 Docking to Other A third advantage of the Toolbox is its ability to dock to other soft-
Software ware platforms. The docking of other software platforms to the
Toolbox allows the user to make predictions via the docked software
while using the knowledge from the Toolbox. The key element of
docking is the knowledge that the Toolbox represented by the pro-
filers can be used to search analogs in the training sets of the docked
software and their models. This capability is important for providing
reliable analogs and for establishing a weight-of-evidence.
The ECOSAR models for predicting aquatic toxicity and the
models of the Danish EPA QSAR Database are docked to the
Toolbox. In addition, the OASIS software platforms TIMES and
Catalogic are capable of being docked to the Toolbox and predic-
tions from respective models (TIMES model for predicting Skin
sensitization, AMES Mutagenicity, Catalogic 301C model, etc.)
can be done within the Toolbox. Importantly, the predictions from
the docked platforms can be reported via the Toolbox. Specific
information unique to the docked software (such as metabolic
maps, quantitative distribution of metabolites, the effect of meta-
bolic activation, and applicability domains) can be provided in the

Toolbox report.
3 Aims, Approach, and Workflow
3.1 The Aims The long-term goal of the tool is to group organic substances into
of the Toolbox chemical categories for apical outcomes of regulatory interest and
using data from tested category members to fill data gaps for
untested category members. To be useful in a regulatory setting,
this means the Toolbox needed to provide all the information nec-
essary to ensure, as far as possible, that the Toolbox user would
actually use the prediction(s) coming from it as part of its regula-
tory assessment. To enhance the likelihood of acceptance, it was
critical that the Toolbox first gets the chemistry correct, second
gets the biology correct and thirdly, when appropriate, adds statis-
tical assurance. The problem is, while some regulatory endpoints,
such as acute aquatic toxicity are amenable to data gap filling by
classic QSAR modeling (e.g., trend analysis; Y = aX + b), many of
the more critical endpoints (e.g., human health effects) were not
amenable to such QSAR prediction.
While decades of work have led to the development of a variety
QSARs using various descriptors and various modeling approaches,
most of these models fail to achieve regulatory use. This lack of
acceptance is often a reflection of putting statistics ahead of chem-
istry and biology. The net result is typically a “black box” predic-
tion that lacks transparency and mechanistic understanding. One
answer to this dilemma, the one employed in the Toolbox, is the
category approach and read-across.
3.2 The Category The category approach is the basis for the Toolbox predictions [3].
Approach In the OECD chemical category approach, a number of chemicals
are grouped based on their similarity. Available experimental results
from one or more members of the category, the source substances,
are used to fill data gap for other members of the category, the
target substances.
According to the OECD guidance [3], similarity is context
dependent and not only similarity in chemical structure and physi-
cochemical properties, but also similarity of mechanism of interac-
tion with different biomolecular targets (e.g., proteins, DNA), as
well as toxicokinetic and toxicodynamic properties, should be con-
sidered [4].
3.3 Toolbox To implement the category approach, the Toolbox [1]:

Workflow

(a)
Identifies the relevant structural features, potential
mechanism(s) of chemical interaction and likely mode of
action of a target substance,
(b) Identifies other substances that have the same features, and
(c) Uses existing experimental data to fill the data gap.
Six modules (i.e., Chemical Input, Profiling, Endpoints, Category
Definition, Filling Data Gap, and Report) [5] guide the use
through a logical workflow based on the category approach.
Guidance suggests these modules be employed in a sequential
work flow [6]. The “Chemical Input” module provides the user
with several means of entering the target chemical. Since all subse-
quent functions are based on chemical structure, the goal of this
module is to make sure the molecular structure assigned to the
target chemical is the correct one. The “Profiling” module elec-
tronically retrieves relevant information in the form of automated
alerts on the target compound. The “Endpoints” module elec-
tronically retrieves experimental results for regulatory endpoints
(e.g., data on environmental fate, ecotoxicity, in vitro or in vivo
mammalian toxicity). These data are stored in the Toolbox. This
data gathering can be executed in a global fashion (i.e., collecting
all data of all endpoints) or more often, on a more narrowly defined
basis (e.g., collecting data for a single or limited number of end-
points). The “Category Definition” module provides the user with
several methods of grouping chemicals into a toxicologically mean-
ingful category. It is key that the category includes the target mol-
ecule and at least one source substance. As previously pointed out,
this is the critical step in the work flow; several options are available
in the Toolbox to assist the user in refining the category definition
via subcategorization [5]. The “Filling Data Gaps” module pro-
vides the user with three options for making an endpoint-specific
prediction for the target chemical; these options, in increasing
order of complexity, are by read-across, by trend analysis, and
through the use of QSAR models. The ‘Report’ module provides
the user several means of downloading a written audit trail of the
sequence of Toolbox functions the user performed in arriving at
the prediction.
Further elaborations of the workflow of the Toolbox, includ-
ing examples, have been presented elsewhere [1, 5, 6, http://
oasis-lmc.org/products/software/toolbox/toolbox-support.
aspx].
4 Chronology of Early Toolbox Development
In 2003, the OECD established the Toolbox strategy under the

umbrella of its Quantitative Structure–Activity Relationships [(Q)
SAR] Project of the Environmental Health and Safety Programme.
Initially, the (Q)SAR Project endeavored to focus on the
following:

(a) Developing “OECD Principles for the Validation, for
Regulatory Purposes, of (Q)SAR Models.”
(b) Garnering the experiences of member countries in applying
(Q)SAR.
(c) Developing a prototype (Q)SAR Application Toolbox.
The establishment of the “OECD Principles for the Validation, for
Regulatory Purposes, of (Q)SAR Models” [7, 8] largely built on
the discussions and conclusions of the 2002 workshop held in
Setubal, Portugal and organized by the European Chemical
Industry Council and the International Council of Chemical
Associations [9]. A report on the regulatory uses and applications
in OECD Member countries of (Q)SAR models in the assessment
of new and existing Chemicals was released in 2006 [10].
In June 2005, the OECD Member countries endorsed the
plan for Phase 1 of Toolbox development. The concept for Phase
1 of the Toolbox, inspired by Gilman Veith, relied on the concept
of chemical similarity [11]. The Phase 1 work (2006–2008), done
in conjunction with the European Commission, focused largely on
proof-of-concepts. The objective of Phase 1 was to develop a work-
ing prototype designed to integrate, in a single in silico platform,
knowledge and data, as well as ‘profilers’ for grouping chemicals.
The initial knowledge and data was donated by the member coun-
tries or provided by the contractor. The profilers, along with the in
silico platform, were developed by the Laboratory of Mathematical
Chemistry (LMC). Briefly, using the previously derived OASIS
platform, LMC implemented the agreed upon work flow.
Phase 1 was completed with the release of Version 1.0 of the
Toolbox, in October of 2008. Key to the member countries agree-
ing to its release were several case studies which demonstrated that
within mechanistically consistent categories, developed by the
Toolbox, existing experimental data within the category could be
used to fill data gaps for other chemicals in the same category.
These examples were designed to show the Toolbox Management
Group how the Toolbox’s chemical profilers and integrated toxi-
cology databases can be used to make transparent predictions. The
case studies focused on regulatory endpoints where high quality
databases and a basic understanding of the mechanisms of action
leading to the apical outcome existed. Particularly useful were the
case studies demonstrating the following:
(a) The skin sensitization potential of a chemical by read-across
based on protein-binding similarity.
(b) The Ames mutagenicity potential of a chemical by read-across
based on DNA-binding similarity.
(c) The acute aquatic toxicity to fish and Daphnia, based on the
correlation between toxic potency and hydrophobicity for the
narcotic mode of action (i.e., trend analysis).
All three of these examples started with a target chemical and

were based on automated profiling constructing a category of sim-
ilar molecules. Subsequently, with data pruning via subcategoriza-
tion, a final prediction was reported.
Version 1.0 of the Toolbox consisted of 21 profilers and 18
databases. Its minimal functionality included:
(a) Profiling—the process of retrieving information about the
functional groups identified in the molecules by applying the
knowledge implemented in the profilers.
(b) Collecting experimental data for a single chemical or a list of
chemicals.
(c) Defining a category—the process of collecting analogs by
applying specific profilers.
(d) Data gap filling—the process of making prediction by using
read across, trend analysis or by applying an external QSAR
model.
In 2008, the OECD initiated Phase 2 (2008–2012) of the
Toolbox development project. In collaboration with the European
Chemicals Agency (ECHA), this work addressed 48 specific topics
which fell into one of five focus areas:
(a) Information technology,
(b) Chassis development and additional functionalities,
(c) Database compilation,
(d) QSAR library and expert system compilation, and
(e) Training.
During Phase 2, the Toolbox became more streamlined. The chas-
sis was further developed to include advanced chemical identifica-
tion and query tools. Ways of handling metabolic activation,
prediction of the properties of mixtures and accounting for tau-
tomerism were also introduced. New profilers, including a
repeated-dose toxicity profiler that contained boundaries based on
repeated-dose toxicity test data extracted from the database of the
Hazard Evaluation Support System (HESS) (see [12] for details)
were added. New simulators for autooxidation, as well as acid, base
and neutral hydrolysis were included. Other profilers and meta-
bolic simulators also were updated. Additionally, during Phase 2,
existing databases were expanded and new databases added. At the
same time, the QSAR library was only minimally expanded, as a
greater emphasis was placed on read-across and trend analysis
based on internal data. However, expert system compilation took
an exciting turn with the introduction of the Adverse Outcome
Pathway concept for estimating the skin sensitization potential of
chemicals (see below). Training included development of guidance
documents on profiling strategies for aquatic toxicity, genotoxicity,
and carcinogenicity, and skin sensitization. It also included training

material in the form of slide shows. The latter were designed to
take a novice user through a step-by-step exercise pointing out the
feature and key strokes in going through the Toolbox workflow for
an example chemical and a particular endpoint. The continuous
improvement in the Toolbox during Phase 2 led to numerous revi-
sions; Version 2.0 of the Toolbox was released at the end of 2010,
while Version 3.0 was released at the end of 2012. With the goals
of improving the usability of Toolbox, especially in terms of stream-
lining its operation and facilitating users in their assessments of
chemical hazard, a variety of advances were implemented from
2008 to 2013. Seminal advances realized during Phase 2 were:
(a) Better integration of mechanistic and endpoint-specific profil-
ers with the categories approach to fill data gaps,
(b) Expansion of the applicability of the Toolbox to long-term
human health endpoints, and
(c) The introduction of the Adverse Outcome Pathway concept to
predict human health endpoints from a matrix of key events.
Further details on version 3.4 can been seen at http://oasis-lmc.
org/products/software/toolbox/toolbox-support.aspx. The suc-
cess of the Toolbox developed in Phase 2 led to the start of Phase
3 in 2014. Specifically, the main areas of development in Phase 3
are usability improvements, further scientific developments, addi-
tional functionalities and improvement of information technology.
A key element of Phase 3 work is the development and release of
Version 4.0 in the second quarter of 2017 and Version 4.1 in the
third quarter of 2017.
5 Key Elements of Using the Toolbox
5.1 Read-Across Read-across is a process of assessing a toxic endpoint of an untested

substance (i.e., target chemical) founded on the results for the
same endpoint for a tested substance (i.e., source chemical) con-
sidered ‘similar’ [1]. As such, it is a (Q)SAR-based method for
filling data gaps but not one based on a mathematical algorithm. It
is a prevailing concept and several in silico tools have evolved to
expedite the process [13]. Since the conception of the Toolbox,
read-across has become a common method employed in data gap
filling. While read-across is conceptually simple, in practice it is dif-
ficult, especially for complex health endpoints such as repeated-
dose toxicity [14].
The quality of a read-across prediction is driven by:

(a) Quality and quantity of the data available for source
chemicals,
(b) Adequacy and reliability associated with the underlying simi-

larity hypothesis, including through data from appropriate
nonanimal methods providing evidence of the mechanism of
action [15].
As further described [16], the category formation and read-
across process have to be transparent, reproducible and clearly
documented. Further, while there can be an over-arching rationale
for grouping organic substances based on molecular structure and
chemical properties, these similarities alone are often not sufficient
to justify a read-across prediction. Lastly, sources of uncertainty
must be identified and addressed. While the limitations to using
read-across predictions include the lack of suitable in vivo data to
be read across and the lack of toxicologically relevant profilers and
data from in vitro or alternative methods, a major limitation is
often the lack of toxicokinetic understanding [16]. Much of the
work in Phase 2 was aimed at improving quality of data gap filling
by read-across by increasing transparency and mechanistic proba-
bility, especially through the use of Adverse Outcome Pathways
(AOPs) and addressing uncertainties through a Weight-of-
Evidence (WoE) approach.
5.2 Adverse An AOP delineates the documented, plausible and testable pro-
Outcome Pathways cesses by which a chemical induces a molecular perturbation and
the associated biological responses at the sub-cellular, cellular, tis-
sue, organ, whole animal, and/or, when appropriate, population
levels of observation [17, 18]. As such, an AOP depicts existing
knowledge concerning the association between the extremes of a
toxic sequence, a Molecular Initiating Event (MIE) and an Adverse
Outcome (AO). These two extremes are coupled by a succession
of Key Events (KEs) and, where possible, the relationships between
the KEs (KERs). An AOP is typically represented by moving from
one KE to another, as compensatory mechanisms and feedback
loops are overcome [19]. The KEs are by design limited to a few
measurable and toxicologically relevant events that are fundamen-
tal to the progression of biological occurrences leading to the
AO. An AOP is not expected to postulate a comprehensive descrip-
tion of every aspect of the chemistry and biology; rather it focuses
on the crucial steps along the sequence [18].
Since an AOP provides a means of recording and formalizing
toxicity pathway information, it has the capability of assisting in
developing a chemical category. For example, the estrogen recep-
tor (ER) binding AOP links ER binding to reproductive impair-
ment in fish [20]. This AOP is diagrammed in Fig. 1.
In the above example, experimental binding potency in the
Rainbow trout estrogen-receptor binding assay is linked to experi-
mental vitellogenin production in the Rainbow trout liver slice assay.
These data, when combined, lead to a rule-based expert system [21]
Fig. 1 The AOP for estrogen-binding leading to fish reproductive impairment
and the ER-binding profiler, which can be used to develop catego-

ries for assessing the apical endpoint reproductive toxicity through
this pathway. For example, the profiler can identify chemicals that
are structurally capable of binding with a specific site on the ER
receptor. Subsequently, available data on vitellogenin production
can be used to refine the grouping to develop a homogeneous group
of chemicals inside of which available test results on reproductive
impairment can be used to fill data gaps.
As part of Phase 2 developed, a few high-risk deliverables were
included in the workplans. In response to the topic “The Use of
Adverse Outcome Pathways in the Development of Categories,” a
workshop was organized through the OECD Toolbox Management
Group. This workshop focused on how to best use “Mechanistic
Information in Forming Chemical Categories” [22].
As a result of the success of the mechanistic information work-
shop, and at the instruction of the OECD Toolbox Management
Group, the AOP for skin sensitization initiated by protein-binding
was developed and published [23, 24]. The OECD skin sensitiza-
tion AOP is based on an electrophilic substance (parent compound
or metabolite) undergoing the MIE (i.e., covalent interaction with
skin proteins), initiating a cascade of other KEs, leading to the AO
of skin sensitization. The MIE (KE 1) is protein binding which can
be predicted from the molecular structure. The potency of the
reactivity can be measured by various in chemico assays (labeled as
KE 2 in Fig. 2). KE 3 is a gene expression response which can be
measured in vitro in genetically engineered keratinocytes. KE 4: is
a cell maker expression response which can be measured in vitro in
dendritic cells. The KE 5 (i.e., T-cell activation) is captured in the
in vivo organ response of the Local Lymph Node Assay (LLNA).
Fig. 2 Key nodes of the AOP for protein-binding leading to skin sensitization
The integration of in chemico and in vitro data relevant to the

assessment of skin sensitization and mapping of the skin sensitiza-
tion AOP to the Toolbox was incorporated in Version 3.3. With
Version 4, the AOP module for skin sensitization was streamlined
to include automated and standardized work flows and relevant
profilers were redeveloped to reflect new knowledge.
Adding AOPs to the Toolbox provides a means of combining
results from mechanistically relevant in silico, in chemico, and in
vitro assays to derive an assessment of hazard which benefits from
decreased uncertainty. Within the Toolbox, the skin sensitization
AOP scheme is a directed graph including a sequence of roots. The
AOP workflow uses Toolbox functionalities for filtering informa-
tion. An overview of the scheme implemented in the Toolbox is
presented in Fig. 2.
5.3 Reducing There are uncertainties inherent to the current in vivo toxicity test
Uncertainty paradigm. While some of these uncertainties are known, others
Through Weight have yet to be identified or characterized. The level of comfort
of Evidence (WoE) with these uncertainties has been bolstered by nearly 40 years of
experience with the methods and the data which they provide.
The rise in the development of new alternative test systems
provide additional information that, when used appropriately, pro-
vide additional WoE. There will be uncertainties in the alternative
methods as well. It will be important to characterize the nature of

these uncertainties and objectively determine whether they are
more or less acceptable than those of the in vivo tests they are
designed to augment or replace. It is important to realize that
uncertainties will be a part of toxicity testing and assessment for
the foreseeable future, regardless of the strategies implemented.
The challenge will be to identify these uncertainties and develop
science-based strategies to address them.
Where feasible and relevant, uncertainty in predictions should
be evaluated in the light of a number of considerations [24].
While the underlying uncertainty of a prediction based on read-
across is related to the quality and quantity of the data available
for the source chemical(s), uncertainty also includes a variety of
other factors primary of which is the uncertainty associated with
the similarity justification. There are inherent uncertainties asso-
ciated with the presumption that the results of the in vivo study
on the source chemical can be read across to the target analog.
The justification for this presumption is based on two interrelated
rationales. First, the target and source chemicals are sufficiently
similar to be relevant to the apical endpoint under consideration.
Second, any differences in similarity are not toxicologically rele-
vant. For example, differences in similarity in chemical properties
often only have a limited effect on the differences in toxicological
responses between the target and source chemicals [14]. It has
been demonstrated that for chronic health endpoints, toxicoki-
netic, as well as toxicodynamic similarity, are more important
[12, 18]. In vitro assays and in silico tools often provide critical
information needed to strengthen a toxicodynamic similarity
description [25]. It also been established that toxicokinetic simi-
larity (i.e., ADME), especially metabolism, most often drives the
overall uncertainty assessment [26].
To facilitate the consideration of a Toolbox prediction in data
gap filling for regulatory purposes, the OECD established princi-
ples for “validation” [8].
Key among the five principles are a defined endpoint, a defined
domain of applicability, and a mechanistic interpretation. The
intent of “a defined endpoint” is to ensure clarity in the endpoint
being predicted by a given model. However, it is the lack of clarity
in the endpoint that often leads to the use of read-across. As noted,
a no-observed-adverse-effect (NOAEL) value may not be regarded
as a defined endpoint in the scientific sense of referring to a specific
effect within a specific tissue/organ under specified conditions [8].
Read-across is often used to fill data gaps for repeated-dose toxicity
[27, 28] based on a NOAEL defined by the lowest-observed-
adverse-effect(s), assuming that the target chemical triggers the
same systemic effects than the source chemical.
The similarity assessment within the Toolbox work flow is

designed to ensure the applicability domain. Basically, if a good
source analog is identified, by definition the prediction is made
within the domain. Within the Toolbox, the principle of mechanis-
tic interpretation is fulfilled when the grouping is based on mecha-
nistic considerations, especially through the use of an AOP and
appropriate profilers.
The WoE approach is a validation tool in toxicology [29]. It
combines information from several sources with the intent of pro-
viding evidence that a prediction is sufficient to fulfill a require-
ment. The weight given to the available evidence depends on a
variety of factors including the following:
(a) Quality of the data.
(b) Consistency of results.
(c) Nature and severity of effects.
(d) Relevance of the information/data.
WoE is used to describe considerations which are made where
there is toxicological uncertainty. For example, it may be used to
ascertain whether the evidence (data and/or information) sup-
porting an assertion (e.g., both the target substance and the source
substances belong to the same chemical category) is greater than
that supporting the opposite contention. WoE assessments of read-
across predictions typically take the form of qualitatively integrat-
ing different lines of evidence through systematic narrative reviews,
causal criteria, and/or qualitative or semiquantitative measures of
toxicologically relevant studies [30, 31]. As previously noted [16],
of particular concern in such assessments are the:
(a) Plausibility, coherence, and consistency of the chemical and
biological experimental evidence between the target and
source substances,
(b) Strength, consistency, and specificity of the association between
the apical outcome and the initiating event, and
(c) Strength, consistency, and specificity of the association between
apical endpoint relevant experimental data.
These concerns are typically addressed in a WoE assessment.
The WoE for each hypothesis should be the comparison and
contrast of the extent and nature of the supporting data versus
potentially inconsistent data or missing information. Within the
Toolbox, WoE may be improved by the use of appropriate end-
point profilers, as well as in chemico and in vitro data. For
example, in Fig. 2, node 2c, in chemico GSH depletion, is
incorporated into the Toolbox as a reactivity potency profiler
and the GSH RC50 database.
Fig. 3 Progression of relevant information in read-across in Version 1 of the

Toolbox
5.4 Summary of Key In summary, much of the more recent Toolbox development has
Elements been aimed at improving acceptance of data gap filling by read-
across (see Subheading 5.1). It is generally acknowledged that one
increases acceptance of a read-across prediction by decreasing the
uncertainty surrounding its preparation. The latter is attained by
increasing transparency and establishing better mechanistic prob-
ability (see Subheading 5.2), as well as assessing the uncertainty,
including the plausible mode of action, and WoE (see Subheading
5.3). Acceptance of a Toolbox prediction is driven by:
(a) Quality and quantity of the apical endpoint data,
(b) Confidence (e.g., adequacy and reliability) associated with the
underlying similarity hypothesis, and
(c) Good relevant supporting information, including data from
appropriate in vitro and in chemico methods.
Earlier versions of the Toolbox relied on simple profilers to estab-
lish a chemical category for reading across for an apical endpoint.
Within the context of an AOP, such a progression typically jumped
from the MIE to the AO with little regard to what happened in
between (Fig. 3). However, by adding toxicity pathway informa-
tion and appropriate information and data for relevant intermedi-
ate events, a more robust category could be established (Fig. 4).
Specifically, one that reduced uncertainty and adds WoE.
6 Current Toolbox Features and Functionalities
The current version of the Toolbox is Version 4.1. Starting with

Version 4.0, the Toolbox was completely rewritten using
Microsoft’s .NET framework. A number of modifications have
Fig. 4 Progression of relevant information in read-across in Version 4 of the Toolbox
been implemented related to improving the core module fea-

tures, new functionalities and user interface modifications, as
outlined below.
For more adequate identification of the chemical substance
two additional descriptors were added to characterize chemical IDs
(in addition to CAS RN and SMILES)—Substance Type
(Monoconstituent, Multiconstituent, UVCB, etc.) and Composition
(including information on constituents, additives, impurities, and
their quantities).
An important modification is that chemical connectivity is now
presented in Daylight SMILES known as a standard format for
chemical structure representation (http://www.daylight.com/
dayhtml/doc/theory/theory.smiles.html). All sub-searching
modules in Toolbox 4.0, including the profilers, have been rewrit-
ten in the SMARTS language (SMiles ARbitrary Target
Specification; http://www.daylight.com/dayhtml/doc/theory/
theory.smarts.html)—which is able to specify (sub)structural pat-
terns in molecules. The 2D editor was also redesigned to work
with SMARTS.
The following key improvements were implemented in the lat-
est Version of the Toolbox:
(a) Defining a category accounting for metabolism allows for
searches of analogs having the same distribution pattern of
metabolites as the target chemical with respect to selected pro-
filing criteria. For example, it is possible to collect analogs hav-
ing the same distribution pattern of metabolites as the target
with respect to protein binding alerts or analogs having a com-

mon metabolite with the target chemical.
(b) PubChem features [ftp://ftp.ncbi.nlm.nih.gov/pubchem/

specifications/pubchem_fingerprints.pdf] were added for
assessing the structural similarity between a target structure
and its analogs. The PubChem System generates a binary sub-
structure fingerprint for chemical structures. These finger-
prints are used by PubChem for similarity neighboring and
similarity searching.
(c) The species taxonomy bank has been expanded with a full tax-
onomy tree as represented in the Integrated Taxonomic
Information System (https://www.itis.gov/).
(d) The Report generator has been redesigned in Version 4.0 of
the Toolbox based on a more compact template which gener-
ates a data matrix with selected physicochemical parameters,
profiling results and experimental data for the target chemical
and analogs. This matrix allows the user to visually evaluate the
consistency of the category.
(e) The database access module was rewritten on C# and the data-
base engine is changed from Firebird to PostgreSQL. The new
module is less coupled with the rest of the Toolbox server
modules which allows a relatively easy replacement of the
underlying database management system (DBMS) or even
developing a Toolbox-based distribution using different data-
base management systems. The current database in version 4.1
of the Toolbox contains ∼79,000 chemicals and more than
2,000,000 data points related to physicochemical properties,
environmental fate and transport, ecotoxicological and human
health endpoints.
A variety of new functionalities were also added in the Toolbox,
such as:
(a) A functionality which defines the target endpoint along with
the target structure. Once the user has selected an endpoint to
predict, the system advises the user as to what knowledge will
be useful for categorization, as well as highlights the databases
within which specified endpoint values may be found. This
functionality assists the user in the category building process
and is considered a significant contribution to the user-
friendliness of the system.
(b) New statistical information for each database is provided, such
as distribution of data by endpoints, bioassays, year, species
and others. This information will lead to a more effective use
of the data in the Toolbox.
(c) It is now possible to evaluate the performance of alerts or

functional groups with respect to the predictability of a target
endpoint. In other words, the predictive capability of any func-
tional group (e.g., an alert for protein or DNA binding, or an
inert chemical group such as “aliphatic alcohol”) with respect
to a defined endpoint can be evaluated. The alert performance
(in %) indicates how many of the analogs in the category hav-
ing the same alert elicit a positive (or negative) endpoint value
out of the total number of analogs in the category. An alert
performance can also be executed accounting for the meta-
bolic activation of the chemicals. Alert performance allows fil-
tering some of the alerts or functional groups having a low
capacity for predicting specific endpoints for specific chemi-
cals. Lastly, the evaluation of the alert performance justifies the
selection of specific alerts in case of multifunctional
chemicals.
(d) In order to increase the user-friendliness and prediction reli-
ability of the Toolbox, automated and standardized workflows
have been developed for selected endpoints. These workflows
follow a well-defined logic based on the specificity of the
observed effects. The aim of the automated workflows is to
make predictions for selected endpoints without the user’s
interaction. Once a target substance is added, the automated
workflow can be executed, finishing with the prediction for
the selected endpoint. This feature is intended to be applied to
large lists of chemicals, e.g., for priority setting. The standard-
ized workflows apply the same logic used in the automated
workflows; however, asking the user to make a selection from
a provided list of possibilities at each step of the workflow.
(e) Accessibility to the document tree is easier with the current
Version of Toolbox, as the document tree stays visible in all the
Toolbox modules, including the gap filling stage. This pro-
vides a clear understanding what actions were performed in
the data gap filling session and provides an easy navigation to
all the results.
(f) Lastly, other new interface modifications include the possibil-
ity to filter, sort, and color by different attributes the profilers
and databases. Help texts with information related to structure
and prediction reliability also have been implemented.
More information about the new version of Toolbox can be
acquired at http://oasis-lmc.org/products/software/toolbox/
toolbox-support.aspx.
7 Further Development of the Toolbox
Transforming the Toolbox into a Chemical Management System

that has regulatory and stakeholder acceptance will required fur-
ther development. In the near-term, the Toolbox will be a crucial
component in linking the in silico instruments and data sets neces-
sary for the implementation of Integrated Approaches to Testing
and Assessment (IATA). IATA are pragmatic approaches designed
to span the transition from the in vivo testing paradigm to assess-
ments which are based on better mechanistic understanding of
both the chemistry and biology behind toxicity. IATA follow an
iterative approach to answer a defined question in a specific regula-
tory context. They take into account the acceptable level of uncer-
tainty associated within the decision setting. There is a range of
IATA—from more flexible, nonformalized judgment based
approaches (e.g., grouping and read-across) to more structured
rule based approaches (http://www.oecd.org/chemicalsafety/
risk-assessment/iata-integrated-approaches-to-testing-and-assess-
ment.htm#reporting). IATA can include a combination of meth-
ods and can be informed by integrating results from one or many
methodological approaches (e.g., (Q)SAR, read-across, in chem-
ico, in vitro, ex vivo, in vivo) or “omic” technologies (e.g., high-
throughput molecular screening, toxicogenomics). The IATA
approach may be implemented in a tiered in design, where failure
to satisfy the structural or toxicity requirements at a lower tier typi-
cally rule out further testing at a higher tier.
Fundamental to using any IATA is having pathway informa-
tion (e.g., an AOP) which can provide a framework comprised of
key events, preferability at the different levels of biological
organization, causally relating the in vivo endpoint of interest. Key
events are often measured by alternative methods (either testing or
nontesting) typically targeting specific cellular or physiological
responses. Alternative methods preclude validation with in vivo
data by a one-for-one approach. The AOP allows for the use of a
battery of assays and subsequent databases designed to target par-
ticular steps along a specific pathway. Each assay/dataset in a suite
of information would inform the next tier of the IATA. The scien-
tific justification of an alternative method or dataset should focus
on comparing the test outcome to what is known about the under-
lying biology as described in the AOP and thus, aid in the decision-
making process. Not all key events or all tiers in an IATA may have
to be satisfied to make an assessment. Today, the most often used
IATA is read-across.
One long-term vision of the OECD QSAR Toolbox sees it
having a battery of structure–activity applications so there will be
little or no need for testing of any kind. However, before the
Toolbox is fully capable of making such assessments, the near-term
future will most likely include toxicity testing in the form of a bat-
tery of assays which are performed in silico, in chemico, and in
vitro, with a reduced need for in vivo testing. The long-term vision
will only be realized by incremental advances in both scientific
knowledge and regulatory acceptance. Moreover, incremental
integration of SAR and new test methods to augment and subse-
quently replace existing in vivo toxicity testing requirements also
will facilitate public acceptance of these alternative approaches. It
must be stressed that this long-term vision of the future will not be
realized for decades to come. By combining state-of-the-art
approaches in a transparent and scientifically defensible manner,
Toolbox-aided assessments will be compatible with the future
vision of toxicity testing and assessment.
Toxicology, especially in the regulatory sense, must be looked
at as both a science and an art. Predictive toxicology makes use of
data-gathering and observational processes to develop hypotheses
and models that can be used to make informed predictions about
adverse effects of chemicals for which there is little available experi-
mental data. The future of regulatory toxicology, as well as the
Toolbox, lies in enhancing the science and improving predictive
capacity.
Increasing the strength of the Toolbox of the future will lie in
enhancing the breadth of information that is used to develop an
understanding of the toxicological profile of a chemical. The
Toolbox must continue to seek new means of deriving toxicologi-
cal information from existing and future data and to infer toxico-
logical potential and potency based on chemical properties and
similarities to other chemicals. The Toolbox must continue to
advance the use of both in vitro and in vivo data; however, the
Toolbox will, in the near-term at least, continue to target only
those in vivo endpoints that have relevance to the toxicity profile of
the target chemical or chemical category. To this end, new in silico,
in chemico, and in vitro methods will be mapped to the Toolbox
and used as categorizing or prioritizing tools.
As noted above, AOPs are focused on developing an under-
standing of the underlying biological response that results in a
regulatory endpoint. Since one goal of the Toolbox is to base pre-
dictive toxicology on such understanding, AOPs will be important
to further Toolbox development. While simple AOPs have been
mapped to the Toolbox, future AOPs will be complex and highly
integrated. A complete mechanistic understanding of the biologi-
cal responses underlying most toxicity outcomes is arguably many
years in the future. However, work is underway at the OECD
Cooperative Chemicals Assessment Programme to develop this
kind of biological systems-level understanding. In order to develop
an understanding of the key responses that trigger an adverse out-
come, it will continue to be necessary to understand how an organ-
ism functions at all biological levels of organization and how these
levels interact with one another. It will not be enough to simply

map out the physical components and interactions of a system;
rather, it will be equally important to know how information is
transferred through the system in response to a perturbation.
Computational biology in the form of both knowledge discov-
eries (e.g., data-mining and the elucidation of patterns from exper-
imental data and simulation-based analyses) will have a greater role
in future Toolbox development. Collating and analyzing the large
volumes of data that come from molecular biology and related
subject and the adoption of common ontologies will play a larger
role in future of the Toolbox. Any computational toxicological
approach is only as good as the data on which it is based. Therefore,
essential to the success of computational toxicology is the develop-
ment of chemical informatics. In chemical informatics, toxicologi-
cal and other databases are developed from standardized guidelines
which are developed in conjunction with subject matter experts for
specific areas of toxicology and populated with chemical-specific
data sets. Such databases are essential to predictive toxicology, as
they form the basis for the in vivo endpoints of regulatory
interest.
HTS is an experimental-based method that simultaneously
conducts thousands of assays (using robotics systems that have
facilitated the automation of the process, from sample preparation
through data collection). Ideally, HTS must test chemicals over a
range of concentrations, preferably up to aqueous saturation, with
a low tolerance for both false positives and false negatives. To max-
imize HTS requires having cells that are robust, have unlimited
capacity for self-renewal, and closely imitate the behavior of normal
cells in vivo. The objective of HTS is to develop bioactivity signa-
tures with which to predict in vivo pathological outcomes from
alterations of cellular phenotypes.
Bioactivity signatures have the potential to be a novel way of
developing profilers and forming chemical categories within the
Toolbox. Combining HTS results with physical chemical proper-
ties creates the opportunity to develop novel profiling strategies.
Much of the information on which these new profilers will be cre-
ated will take the form of knowledge bases. These knowledge bases
will be organized in terms of ontologies that permit automated
knowledge extraction from the data.
The OECD QSAR Toolbox of the future will incorporate
exposure into its predictions. Physiologically based pharmacoki-
netic modeling is a computational approach that facilitates the
translation of in vitro data into estimates of exposure. This approach
considers the physiology and anatomy of the body, as well as the
biochemistry of the chemical of interest. Briefly, pharmacokinetics
considers how a substance is introduced into a biological system,
what it does after introduction, how long it remains in the system,
and how it is eventually eliminated from the system. This informa-
tion is captured with the basic properties of absorption, distribu-

tion, metabolism, and excretion. Absorption, distribution,
metabolism, and excretion properties and physiologically based
pharmacokinetics will be further integrated into the Toolbox.
In summary, further development of the Toolbox must con-
tinue to focus on transparent data gap filling, documenting, plau-
sible and testable processes by which chemicals induce molecular
perturbations and their associated biological responses, thereby
reducing uncertainties associated with predictions. However, to
meet future needs, the Toolbox will undergo a stepwise progres-
sion into a more complete Chemical Management System.
Acknowledgments
The authors gratefully acknowledge the financial and intellectual

contributions of the European Commission, European Chemical
Agency, and OECD member countries, as well as industry and
other organizations. Without these contributions the Toolbox
would not have been a success.
References
1. Diderich R (2010) Tools for category forma- safety series on testing and assessment no. 102.
tion and read-across overview of the OECD ENV/JM/MONO(2009)
(Q)SAR application toolbox. In: Cronin MTD, 6. Dimitrov SD, Diderich R, Sobanski T,
Madden JC (eds) In Silico toxicology: princi- Pavlov TS, Chankov GV, Chapkanov AS,
ples and applications. RSC Publishing, Karakolev YH, Temelkov SG, Vasilev RA,
Cambridge, pp 385–407 Gerova KD, Kuseva CD, Todorova ND,
2. Schultz TW, Dimitrova G, Dimitrov S, Mehmed AM, Rasenberg M, Mekenyan OG
Mekenyan OG (2016) The adverse outcome (2016) QSAR toolbox—workflow and
pathway for skin sensitisation: moving closer to major functionalities. SAR QSAR Environ
replacing animal testing. Altern Lab Anim Res 27:203–219
44:1–8 7. Organisation for Economic Cooperation and
3. Organisation for Economic Cooperation and Development (OECD) 2004. The Report from
Development (OECD) 2007. Guidance on the expert group on (quantitative) structure–
grouping of chemicals, OECD environmental activity relationships (Q) SARs. on the princi-
health and safety series on testing and assess- ples for the validation of (Q)SARs, ENV/JM/
ment no. 80. ENV/JM/MONO(2007)28 TG(2004)27/REV, Organisation for Economic
4. Organisation for Economic Cooperation and Cooperation and Development, Paris, FR
Development (OECD) 2014. Guidance on 8. Organisation for Economic Cooperation and
grouping of chemicals. 2nd edn, OECD envi- Development (OECD) 2007. Guidance docu-
ronmental health and safety series on testing ment on the validation of (quantitative) struc-
and assessment no. 194. ENV/JM/ ture–activity relationships (Q) SARs. Models,
MONO(2014)4 OECD environmental health and safety series
5. Organisation for Economic Cooperation and on testing and assessment no. 69. ENV/JM/
Development (OECD) 2009. Guidance docu- MONO(2007)2
ment for using the OECD (Q)SAR application 9. Jaworska JS, Comber M, Auer C, Van leeuwen
toolbox to develop chemical categories accord- CJ (2003) Summary of a workshop on regula-
ing to the OECD guidance on grouping of tory acceptance of (Q)SARs for human health
chemicals, OECD environmental health and and environmental endpoints. Environ Health
Perspect 111:1358–1360
10. Organisation for Economic Cooperation and 19. Organisation for Economic Co-operation and
Development (OECD) 2006. Report on the Development (OECD) 2009. Report of the
regulatory uses and applications in oecd mem- expert consultation to evaluate an estrogen
ber countries of (Q)SAR models in the assess- receptor binding affinity model for hazard
ment of new and existing chemicals, OECD identification. OECD environmental health
environmental health and safety series on test- and safety series on testing and assessment, no.
ing and assessment no. 58. ENV/JM/ 111, ENV/JM/MONO(2009)33
MONO(2006)25 20. Schmieder PK, Kolanczyk RC, Hornung MW,
11. Bradbury SP, Russom CL, Schmieder PK, Tapper MA, Denny JS, Sheedy BR, Aladjov H
Henry TR, Schultz TW, Diderich R, Auer (2014) A rule-based expert system for chemical
CM2014 Advancing computational toxicology prioritization using effects-based chemical cat-
in a regulatory setting: a selected review of the egories. SAR QSAR Environ Res 25:253–287
accomplishments of Gilman D. Veith (1944– 21. Organisation for Economic Co-operation and
2013). Appl In Vitro Toxicol 1:11–20 Development (OECD) 2011. Report of the
12. Sakuratani Y, Zhang HQ, Nishikawa S, workshop on using mechanistic information in
Yamazaki K, Yamada T, Yamada J, Gerova K, forming chemical categories, 8–10 december
Chankov G, Mekenyan O, Hayashi M (2013) 2010, Crystal City, VA, USA. OECD environ-
Hazard evaluation support system (HESS) for mental health and safety series on series on test-
predicting repeated dose toxicity using toxico- ing and assessment no. 138, ENV/JM/
logical categories. SAR QSAR Environ Res MONO(2011)8
24:351–363 22. Organisation for Economic Co-operation and
13. Patlewicz G, Helman G, Pradeep P, Shah I Development (OECD) 2012. The adverse out-
(2017) Navigating through the minefield of come pathway for skin sensitisation initiated by
read-across tools: a review of in silico tools for covalent binding to proteins. Part 1: scientific
grouping. Comput. Toxicology 3:1–18 evidence. OECD environmental health and
14. Schultz TW, Amcoff P, Berggren E, Gautier F, safety series on testing and assessment no. 168,
Klaric M, Knight DJ, Mahony C, Schwarz M, ENV/JM/ MONO(2012)10/PART 1
White A, Cronin MTD (2015) A strategy for 23. Organisation for Economic Cooperation and
structuring and reporting a read-across predic- Development (OECD) 2016. Guidance docu-
tion of toxicity. Regul Toxicol Pharmacol ment on the reporting of defined approaches to
72:586–601 be used within integrated approaches to testing
15. Organisation for Economic Co-operation and and assessment. OECD environmental health
Development (OECD) 2014. Guidance on and safety series on testing & assessment no.
grouping of chemicals, 2nd edn, OECD envi- 255. ENV/JM/MONO(2016)28
ronmental health and safety series on testing 24. Schultz TW, Przybylak KR, Richarz AN,
and assessment no. 194 ENV/JM/ Bradbury SP, Cronin MTD (2017) Read-
MONO(2014)4 across for 90-day rat oral repeated-dose toxicity
16. Schultz TW, Cronin MTD (2017) Lessons for selected 2-alkyl-1-alkanols: a case study.
learned from read-across case studies for Comput Toxicol 2:28–38
repeated-dose toxicity. Regul Toxicol 25. Organisation for Economic Co-operation and
Pharmacol 88:185–191 Development (OECD) 2012. The adverse out-
17. Tollefsen KE, Scholz S, Cronin MT, Edwards come pathway for skin sensitisation initiated by
SW, de Knecht J, Crofton K, Garcia-Reyero N, covalent binding to proteins. Part 2: use of the
Hartung T, Worth A, Patlewicz G (2014) AOP to develop chemical categories and inte-
Applying adverse outcome pathways (AOPs) to grated assessment and testing approaches.
support integrated approaches to testing and OECD environmental health and safety series
assessment (IATA). Regul Toxicol Pharmacol on testing and assessment no 168: ENV/JM/
70:629–640 MONO(2012) 10/ PART 2
18. Organisation for Economic Co-operation and 26. Organisation for Economic Co-operation and
Development (OECD) 2016. Guidance docu- Development (OECD) 2016. Case study on
ment for the use of adverse outcome pathways the use of integrated approaches for testing and
in developing integrated approach to testing assessment for repeated dose toxicity of substi-
and assessment (IATA). OECD environmental tuted diphenylamines (SDPA). OECD envi-
health and safety series on testing & assessment ronmental health and safety series on series on
no. 260 ENV/JM/MONO(2016)67 testing and assessment no. 252, ENV/JM/
MONO(2016)50
27. Organisation for Economic Co-operation and 29. Ellison CM, Madden JC, Judson P, Cronin
Development (OECD) 2016. Case study on MTD (2010) Using in silico tools in a weight
the use of an integrated approach to testing of evidence approach to aid toxicological
and assessment for hepatotoxicity of allyl esters. assessment. Mol Inform 29:97–110
OECD environmental health and safety series 30. Borgert CJ, Mihairch EM, Ortego LS, Bentley
on testing and assessment no. 253, ENV/JM/ KS, Holmes CM, Levine SL, Becker RA (2011)
MONO(2016)51 Hypothesis-driven weight of evidence frame-
28. Balls M, Amcoff P, Bremer S, Casati S, Coecke work for evaluating data within the US EPA's
S, Clothier R, Combes R, Corvi R, Curren R, endocrine disruptor screening program. Regul
Eskes C, Fentem J, Gribaldo L, Halder M, Toxicol Pharmacol 61:185–191
Hartung T, Hoffmann S, Schechtman L, Scott 3 1. Schultz TW, Przybylak KR, Richarz A-N,
L, Spielmann H, Stokes W, Tice R, Wagner D, Mellor CL, Escher SE, Bradbury SP,
Zuang V (2006) The principles of weight of Cronin MTD (2017) Read-across for
evidence validation of test methods and testing 90-day rat oral repeated-dose toxicity for
strategies: the report and recommendations of selected n-alkanols: a case study. Comput
ECVAM workshop 58. Altern Lab Anim Toxicol 2:12–19
34:603–620
Chapter 3
QSAR: What Else?

Giuseppina Gini
Abstract
QSAR (quantitative structure–activity relationship) is a method for predicting the physical and biological
properties of small molecules; it is today in large use in companies and public services. However, as any
scientific method, it is nowadays challenged by more and more requests, especially considering its possible
role in assessing the safety of new chemicals. Posing the question whether QSAR is a way not only to
exploit available knowledge but also to build new knowledge, we shortly review QSAR history, thus
searching for a QSAR epistemology. We consider the three pillars on which QSAR stands: biological data,
chemical knowledge, and modeling algorithms. Most of the time we assume that biological data is a true
picture of the world (as they result from good experimental practice), that chemical knowledge is scientifi-
cally true; so if a QSAR is not working, blame modeling. This opens the way to look at the role of modeling
in developing scientific theories, and in producing knowledge. QSAR is a mature technology; however,
debate is still active in many topics, in particular about the acceptability of the models and how they are
explained. After an excursus in inductive reasoning, we relate the QSAR methodology to open debates in
the philosophy of science.
Key words QSAR, Predictive modeling, Induction, Validation, Acceptability
1 Introduction
QSAR (quantitative structure–activity relationship) evolved from

the preliminary hypothesis that a link exists between biological
effects and molecular structure. The QSAR method originated
from the pioneering work of Hansch [1–3] who introduced mod-
els of multilinear equations of physicochemical parameters of mol-
ecules. The parameters represented the hydrophobic, steric, and
electrostatic properties of whole molecules, and were considered
responsible for the effect. Small sets of congeneric compounds
were studied within this initial QSAR.
In contrast with this approach Free and Wilson [4] proposed
an additive system considering that the contributions from relevant
molecular fragments sum up to give the endpoint value. This work
opened the way to the development of molecular descriptors that
account for the various aspects of the molecular structure. For
79
80 Giuseppina Gini
about 20 years the definition of new molecular descriptors was an

active research topic. A large number of descriptors, such as topo-
logical [5], electro-topological [6], geometrical [7], and quantum-
chemical [8] were developed. Thousands of them are commonly
available in specialized software systems, making it impossible to
simply use multilinear regression due to many collinear variables
and to the small size of available datasets [9]. Methods for variable
selection and other statistical methods were also introduced to deal
with this more complex scenario [10].
QSAR exploited the rapid development of machine learning
and data mining technologies to become an active area of building
models from data. This opened the road to the introduction of a
plethora of new computing methods, each with theoretical and
practical pros and cons, making it necessary to state in the model
which learning method used and how.
Historical overviews about QSAR are available in the literature
(see for instance [11, 12]). Many trends appeared in the QSAR
history: from congeneric compounds to large chemical spaces,
from toxicity as dose to mode of action, from linear to nonlinear
models, from designing the best descriptors to selecting the best
ones for modeling, and from generating a single strong model to
ensemble many weak models.
In the pharmaceutical field, QSARs are commonly used to
screen candidate drugs in the initial development. To study phar-
macophore features in interactions between the receptor and the
drug, other specific methods, as COMFA (Comparative Molecular
field Analysis) [13], are used.
QSAR models enlarged their interest from the pharmaceutical
industry to other industries as well as regulatory bodies. Public
bodies started the practice of creating databases of molecular struc-
tures and measured effects for many physical and biological end-
points, making it possible to create models on larger training sets.
Since the introduction of protection measures against possible
risks of chemical products (as regulated in USA by EPA, and in the
European Union in the REACH regulation), QSAR has become a
routine tool, together with other instrumental and expert evalua-
tions, for regulatory bodies. So the way to test the model became
an urgent priority for any regulatory use; efforts proposing new
validation technologies for acceptable QSAR have been a hot topic
[14, 15].
In 2004, the Organization for Economic Co-operation and
Development (OECD) issued a document stating the five indis-
pensable components of an acceptable QSAR: a defined endpoint,
an unambiguous algorithm, a limited applicability domain, appro-
priate assessment for internal performance and external predictiv-
ity, and possibly a mechanistic interpretation [16]. These principles
will be discussed in the next section.
QSAR: What Else? 81
Today, there are two main streams in the development and use
of QSAR: QSAR as a method for molecular design, and QSAR as
a method for evaluation of properties. The first stream uses QSAR
together with computational chemistry tools and docking meth-
ods, and is aimed at designing new products with desired proper-
ties, in particular drugs. The second line is aimed at screening
molecular structures for predicting properties relevant for humans
and nature.
We will not consider the tools necessary for molecular design,
but only will concentrate on the common aspects of both QSARs,
namely the building of predictive models and the meaning of these
models.
2 The Three Pillars of QSAR: Biological Data, Chemical Knowledge,

and Modeling Algorithms
Is QSAR building knowledge or exploiting knowledge? And what

is knowledge? We start answering these questions by looking at the
basic multidisciplinary pillars of QSAR, and the knowledge they
represent.
2.1 Biological Data Biological data used in QSAR originate from laboratory testing on
living systems (animal or vegetal). Many models use principles sim-
ilar to QSARs to predict physicochemical properties (as boiling
point or logP) and are usually called QSPRs (quantitative struc-
ture–property relationships). In the following, we focus only on
biological data, since the effects of chemical substances on the liv-
ing systems are the final interest of regulatory bodies and industrial
developers.
While chemical and physical knowledge are often enough to
understand the physicochemical properties of the molecules, the
biological effects are hard to model and are affected by a larger
variability. The main reason is the individual variability; it is impos-
sible to find two perfectly identical instances of any biological
entity, thus affecting the reproducibility of the experiments. If a
substance has different effects on different individuals, how is it
possible to generalize the results to a whole population? The prac-
tice in toxicology is standardizing the experimental results from
individual answers to average over a population. The purpose of
the biological endpoints LD50 or IC50 is exactly that of eliminating
the individual variability from the experimental results.
A more basic discussion about how to make use of biological
observations to make science has been a hot topic in the philoso-
phy of science in the last century. How to explain the biological
phenomena that generate the data? Reductionist and antireduc-
tionist positions in biology appeared [17]. Pure reductionists con-
tend that the only biological explanation to seek is at the level of
82 Giuseppina Gini
physicochemical processes, while antireductionists argue that such

explanations do not belong in the realm of biology.
The basic question of reductionism is whether theories devel-
oped in one science can be special cases of theories developed in
other branches of science. Reduction of one branch of science to
another was possible in the past; for instance, the reduction of
thermodynamics to statistical mechanics after finding that the tem-
perature of a gas reflects its kinetic energy, or the reduction of part
of chemistry to physics after discovering that valence relates to the
number of electrons in the external orbit of an atom. Many aspects
of biology have been reduced to chemistry after the discovery of
particular molecules, as DNA and RNA. This can lead to the con-
clusion that the ultimate goal of biology is to explain biological
laws as special cases of chemical and physical laws.
Reductionism and antireductionism have existed since the ori-
gin of philosophy, as exemplified by Democritus and Aristotle.
Democritus tried to explain the world in terms of atomic struc-
tures, while Aristotle had a holistic and teleological view of the
world.
Consider that some events cannot be predicted as single occur-
rences but only statistically. Could this principle, well known in
physics, be applied to biology? In practice, this principle states that
randomness is not an appearance due to our incomplete knowl-
edge but an essential property of the world. So the possibility of
prediction in biological sciences may be only statistically assessed.
According to Popper, reductionism is a mode of research
which is never completely successful. Pure reductionism is a mis-
take, since it wants to reduce everything to an ultimate explana-
tion, something that cannot be explained anymore. Popper
contends that the physical world is not causally closed, as in the
excerpt from [18]: “We live in a world of emergent evolution; of
problems whose solution, if they are solved, beget new and deeper
problems. Thus we live in a world of emergent novelty; of a novelty
which, as a rule, is not completely reducible to any of the preced-
ing stages.”
Going further, Jacques Monod stated the “postulate of
Objectivity”: true knowledge cannot be obtained on the basis of
assumptions or theories that pretend to explain things in the uni-
verse by final causes.
2.2 Chemical Chemistry is usually absent in the philosophy of science, as noted

Knowledge by Schummer [19]. One of the reasons may be that the target of
science (and of the philosophy of science) is to explain the reality,
while the main target of chemistry is to change the reality by creat-
ing new substances. So, unlike other sciences, the meaning of
chemical knowledge has not been fully addressed.
Most people assume that scientific knowledge is somehow
stored in theories, not in databases. Modern science makes an
QSAR: What Else? 83
extensive use of data mining on big data to find new associations

and rules from data.
Where is chemistry in this trend? Chemistry has produced and
is still producing an enormous quantity of new substances. Have
they produced new theories or only new databases? New data in
some sense increase our chemical knowledge, but only in quantita-
tive terms. If we consider that the number of possible chemical
substances is somehow limited, we may in theory obtain an almost
full knowledge in some final moment. This raises the following
question: is this number finite, and the possible chemical knowl-
edge obtained finite or infinite? Chemical properties, such as the
reactivity of substances, tell us that the more substances we have
the more new chemicals that could be developed, in an exponential
process. But real world is not a collection of substances; it is a com-
plex dynamic system. When adding a new chemical we have to
understand that we are adding an entire new class of possible trans-
formations, and that we are changing our environment. This expo-
nential growth of chemicals challenges both our theoretical
understanding of what chemistry is and our ethical concern about
what risks may derive from such new chemicals.
In conclusion, should we be optimistic or pessimistic about the
completeness of chemical knowledge? Schummer [19] says: “We
have no reason at all to assume that the realm of possible sub-
stances is limited. If we take that seriously, we must assess the finite
growth of chemical knowledge against the background of an infin-
ity of possible knowledge. An infinite realm of possible substances
corresponds to an infinite amount of possible knowledge that we
not yet have. To be sure, the fast increase of our chemical knowl-
edge decreases our lack of knowledge in a certain sense. But that
does not matter. Mathematics forces us to accept that a finite
decrease of an infinite amount does not affect the infinity at all. As
a consequence, whatever the rates of growth of chemical knowl-
edge will be, that does not change the fact that our knowledge gap
is infinite and will remain infinite in the future.”
In this view QSAR is not only using available knowledge but
also producing new knowledge.
2.3 Modeling Modeling is the last, but not least, pillar of QSAR. Roughly, there
Algorithms are two main streams for making models: data modeling and
algorithmic modeling.
Data modeling is the stream commonly developed by statisti-
cians: from the data analysis they postulate the kind of relation
between data and response, and use a large set of mathematical
tools to derive the model.
Algorithmic modeling has been developed more recently,
starting in the mid-1980s when powerful new algorithms for fit-
ting data became available. The community using these tools
aimed at building predictive models for data sets so difficult that
84 Giuseppina Gini
the traditional statistical analysis was unable to solve: speech rec-

ognition, image recognition, nonlinear time series prediction,
handwriting recognition, prediction in financial markets. Statistical
data models were rarely used in this community. Today algorith-
mic methods include decision trees, production rules, artificial
neural networks, genetic algorithms and optimization, and sup-
port vector machine, just to name the main families. This field is
in rapid evolution, and is boosted by the introduction of massively
parallel hardware that can make computation of millions of data
readily available. This trend also impacts on the concept of chemi-
cal descriptors. For instance, deep learning (neural networks with
many internal layers) has been tested for generating QSAR using
directly as input the 2D drawing of molecules [20]; this is a fur-
ther step after using only the string representation (SMILES) of
molecules to create QSAR [21].
The large extent of possible modeling methods has created,
even in QSAR, a plethora of models; very different models with
similar performance are often available for the same endpoint. This
fact opens a series of new problems: How to develop the best
model? What algorithm and which features to use? How to com-
pare the obtained models? Does it, the best model, exist?
The development of a good model is generally intended as the
development of a simple model. In his “Statistical Modelling: The
Two Cultures” [22], Breiman argues that the preference for simple
statistical linear models on the claim that they are interpretable, is
unsound. The goal of a (statistical) model is to get useful informa-
tion about the relation between the response and predictor vari-
ables. Interpretability is a way of getting information, but a model
does not have to be simple to provide reliable information about
the relation between predictor and response variable. This point
will be discussed later on.
Traditionally, science prefers simple theories, as exemplified
by the Occam’s Razor: the best scientific theory is the smallest
one that explains the data. Can simplicity be stated in an equa-
tion? The minimum description length (MLD) principle pro-
duces such an equation [23]. The MLD principle is a way to
measure the “complexity” of a classifier according to the bits
necessary to codify it.
(a) The theory T requires a number of bits L(T) to encode itself.
(b) The training set E of examples requires a number of bits L(E|
T) to encode it, given the theory T.
The MLD principle establishes a connection to the probabil-
ity theory, i.e., minimizing the description length is equivalent to
maximizing the probability Pr(T|E)—the probability of the the-
ory after the examples have been considered. This result is a use-
ful shortcut since applying the Bayes rule has a high cost due to
QSAR: What Else? 85
finding a priori probabilistic distributions for the Pr(T), while

applying MLD is easier.
But there are also reasons against MLD. Epicurus is consid-
ered the proposer of the “Principle of maximum explanation.” As
reported by Lucretius: “There are also some things for which it is
not enough to state a single cause, but several, of which one, how-
ever, is the case. Just as if you were to see the lifeless corpse of a
man lying far away, it would be fitting to list all the causes of death
in order to make sure that the single cause of this death may be
stated. For you would not be able to establish conclusively that he
died by the sword or of cold or of illness or perhaps by poison, but
we know that there is something of this kind that happened to
him” 1. In practice, if more than one theory is consistent with data,
keep all the consistent hypotheses. Each model can better explain
a part of the data, and together they can achieve a higher precision.
This is the principle of the ensemble methods.
A family of models can be derived from the same data set, and
many of them can be valid models. Is there something considered
“the best possible model”? The question is raised by Wolpert’s “no
free-lunch” (NFL) theorems [24]. Wolpert wrote: “The impor-
tance of the NFL theorems is their implication that, for any two
learning algorithms A and B, there are just as many situations in
which algorithm A is superior to algorithm B as vice versa. So if we
know that learning algorithm A is superior to B averaged over
some set of targets F, then the NFL theorems tell us that B must
be superior to A if one averages over all targets not in F.”
This note is only to remember that the problem of selecting
the best algorithm has huge theoretical limitations. Those limita-
tions are partially managed by specific methods, as the definition of
the applicability domain, used in QSAR.
3 Success Stories, Pitfalls, and Trends in QSAR
The evolution of QSAR has made it possible to apply QSAR mod-

els to real cases in science, industry, and regulations. We briefly
look at this evolution and review the concerns it raised.
3.1 Success Stories Classical QSARs are based on the Hansh hypothesis that three
main molecular properties are required to explain variations in a set
of congeneric compounds: electronic, hydrophobic, and steric
properties. The intent of those initial QSARs was not predictive
but explanatory, as a way to increase the understanding of the bio-
chemical properties under consideration. The initial QSAR equa-
tions were linearly relating the three descriptors with the logarithm
1
Titus Lucretius Carus, Of The Nature of Things,
http://www.gutenberg.org/ebooks/785?msg=welcome_stranger
86 Giuseppina Gini
of the activity. But nonlinear models were soon introduced to

explain some properties, for instance lipophilicity.
The availability of larger data sets and the needs of real applica-
tions have gradually extended the classical QSAR to consider other
molecular properties and noncongeneric compounds, applying all
the modern statistical and algorithmic knowledge to create large
models. Those nonclassical QSARs are today the most used.
QSAR’s main role is prediction; however, its role covers also the
following:
(a) Knowledge mining. QSAR models are tools for virtual screen-
ing of molecules (without the need to make wet experiments);
their prediction can be used for hypothesis confirmation, and
their interpretation for hypothesis generation, and eventually
molecular design.
(b) Knowledge validation. If the developed model explains the
experimental observations, and is in agreement with the
expert knowledge, then the model is considered as a valid
tool. If they disagree, either the model is incorrect or the
available knowledge about the phenomenon should be
revised. The statistical validation is important to check the
correctness of the model, but biased data sets are a possible
source of this disagreement.
(c) Knowledge exploitation. Validated models with a high confi-
dence level can be used for regulatory purposes.
Successful examples of QSAR can be traced back to the eight-
ies of last century, especially with the successful study of new drugs.
In the environmental and ecotoxicological domains the use of
QSAR together with other nontesting methods is a necessity since
the introduction of regulations in Europe and elsewhere that limit
or forbid the use of animal testing. Most of the success in using
QSAR derives from the principles of the 3Rs (Replacement,
Reduction, and Refinement), developed about 50 years ago and
providing a framework for performing more ethical animal research.
An index of success of QSAR methods can be also the number
of models developed and companies selling QSAR technology. To
give some numbers, the ANTARES EU project2 listed the systems
(proprietary or publicly available) in 2014 for properties of rele-
vance for the REACH regulation.
The number of regulations including QSAR in their assess-
ment is also increasing both for the registration of chemicals (by
ECHA in Europe and EPA in USA) and for food safety (by FDA
in USA and EFSA in EU). The same happens in almost all the
industrialized countries (Canada, Japan, Germany, Denmark, and
many more) for their internal regulations. Notably, the ECHA
2
http://www.antares-life.eu/index.php?sec=modellist
QSAR: What Else? 87
regulation gave the opportunity to test QSAR models in regula-

tory tasks, opening a debate that is still underway [25, 26].
3.2 Pitfalls Since modern QSARs have large domain of applicability and use a
large set of molecular properties (often implicitly contained into
molecular descriptors), there is no more a common structural core.
They are no more local models of a specific “mechanism.” The
molecular descriptors are not easily connected to the biological
interactions of ligand and protein, so QSAR models using those
descriptors cannot say directly what biochemical property is rele-
vant to explain the effect. This situation is partially modified by the
use of fingerprint descriptors that conceptually can be interpreted
as fragments relevant to some biological process.
Cronin and Schultz in 2003 made a warning about bad prac-
tices in QSAR. They wrote: “At the end of the day, however,
QSARs are predictive techniques based on the relationship, for a
series of chemicals, between some form of biological activity and
some measure(s) of physico-chemical or structural properties. As
such, there are a number of limitations to the use and application
of QSARs. It is the concern of the authors that these are often not
appreciated, or may be forgotten by the developers of QSAR”
[27]. Considering the three pillars of QSARs, they indicate the
most common pitfalls in:
1. Biology: experimental errors, reproducibility of data;
2. Chemistry: errors in descriptors;
3. Statistical analysis: overfitting, using unnecessarily complex

models.
After a decade, Cherkasov et al. [12] listed twenty-one prob-
lems that can make it hard to develop or to accept QSARs in prac-
tice. We group them in five categories:
(a) Data curating: Failure to take account of data heterogeneity—
Use of inappropriate end point units.
(b) Data preprocessing: Use of confounded descriptors—Use of
noninterpretable descriptors—Errors in descriptor values.
(c) Model construction: Overfitting of data—Use of excessive
numbers of descriptors in a QSAR—Incorrect calculation—
Lack of descriptor auto-scaling—Misuse or misrepresentation
of statistics—No consideration of distribution of residuals—
Replication of chemicals in a data set.
(d) Model validation: Inadequate training and/or test set selec-
tion—Inadequate QSAR model validation—Inadequate or
missing statistic measures.
(e) Model usability and delivery: Poor transferability of QSARs-
Inadequate or undefined applicability domainC—
88 Giuseppina Gini
Unacknowledged omission of data points—Use of inadequate

data—Narrow range of end point values—Lack of mechanistic
interpretation.
Recent literature in QSAR is proposing practical solutions for
most of those problems.
To address good practice in developing QSAR for regulatory
tasks, OECD stated its five “principles”: (1) a well-defined end-
point; (2) an unambiguous algorithm for model derivation; (3) a
clearly defined applicability domain (AD); (4) appropriate mea-
sures of goodness-of-fit, robustness, and predictivity; and (5) a
mechanistic interpretation, if possible [16].
The first requisite is a clear indication that we cannot generalize
from one endpoint to another, and the endpoint should be clearly
defined. As a consequence, to predict a property we need to indi-
viduate a precise procedure used to measure a precise quantity; all
the data used should be obtained with that procedure. Most of the
data sets available in open repositories are built with this principle
in mind; however, this does not mean that they do not contain
errors, so a complete check of any data set is necessary before using
it for any modeling purposes.
The second requisite is much more demanding. Unambiguous
algorithm is quite cryptic; it can be interpreted as the indication
that the model cannot be obtained with any method that uses ran-
domness (as in GA, ANN, etc.). Since many models of real use
incorporate such random techniques, it is plausible that for OECD
it means that all the steps in obtaining the model and the predic-
tion should be fully reproducible; this guarantees that the result for
one molecule will be the same in any laboratory at any time. This
requisite is hard to obtain in practice since any software application
can produce numerical results that depend on the hardware (num-
ber of bits in the memory word, implementation of floating point,
etc.) and on the software environment. To be sure that the repro-
ducibility is full the software should be open source, or the pro-
ducer of the software should provide new releases to cope with
upgrading hardware and software. This is impractical, so this req-
uisite can be only partially fulfilled.
The third requisite, about the AD, has received a lot of atten-
tion in the QSAR community. This principle seems straightfor-
ward, but different AD can be obtained with different methods,
and the sharp answer yes/no is rarely reached.
The fourth requisite is about the validation of the models used;
they should be validated for how they fit the data, how they are
robust to noise, and how they are expected to predict new data.
This principle is the result of a large body of literature that has dis-
cussed the validation methods in QSAR. At the same time, other
application domains that use similar machine learning and data
mining methods evolved many measures and procedures to validate
QSAR: What Else? 89
models. The valuable answers are many, and validation can be ful-
filled with a plethora of methods. We will explore this point in
Subheading 5.
The fifth requisite is about model interpretation. A mechanistic
interpretation strictly means that, similarly to a mechanical system,
the predictors (or most of them) used by the model can be played
in a simulation exercise to show that really their values activate a
process and produce the observed results. This property of the
models seldom can be demonstrated, for various reasons. For
instance, there are tens or hundreds of molecular descriptors that
are correlated with shape, physical properties, steric properties, and
so on. This correlation is a coarse interpretation of the “mecha-
nism,” if any, involved. Second, the hidden variables may be more
important than the predictors. Third, different good quality mod-
els can use completely different descriptors, giving the idea that
any kind of mechanistic interpretation has more a didactic value
than a true interpretation value. The problem of interpretation of
a predictive model is the subject of philosophical speculation, and
will be discussed in Subheading 5.
3.3 Trends Trends in QSAR as we have so far seen include the change of focus:
1. From laboratory use to regulatory use.
2. From a chemical family to the chemical space.
3. From linear to nonlinear models.
4. From single model to ensemble models.
Other trends appear and disappear, and one of them is related
to the right representation level. A recurrent issue is whether to
consider the whole molecule or focus on the presence of specific
fragments. The whole molecule is represented through molecular
descriptors, while fragments are represented as simple strings or
bits. Moreover, does the effect depend on the whole structure
(including energy and chemical properties arising from the 3D
shape) or on the presence of specific functional groups, usually
called structural alerts (SA)?
In the development of predictive models SAs are used as a
transparent way to decide the possible hazards of a chemical. Many
systems of large use rely on rules that check for SAs. SAs were cre-
ated by human experience after observing in many cases that the
presence of a given structure was associated with the effect. Most
of them are accepted since a full set of plausible chemical transfor-
mations and its binding to a receptor is individuated. The role of
functional subgroups in design of drugs is of primary importance.
The drawback of their use for regulatory purposes is that often
the association observed is not a causal relationship but is still a
statistical relationship. In some cases the alert is positively corre-
lated to the effect in about 50% of the observations. In general the
only use of SAs tends to overestimate the positive effect, and has to
90 Giuseppina Gini
be mediated with other tools, for instance going back to a QSAR

model.
In [28] the reliability of SAs is discussed considering the results
from using structural alerts and QSAR models on skin sensitiza-
tion. The authors conclude that QSAR models afford higher accu-
racy than structural alerts. Similar results are obtained in [29] for
mutagenicity. However SAs are an important communication tool
and can be integrated with QSAR and read across to level off their
too conservative results. For instance, ToxRead and ToxDelta [30,
31] make such integration, exploiting both similarity and differ-
ences among the molecule under scrutiny.
Another trend that affects QSAR is related to the search of a
deep causal explanation (in terms of sequence of steps) to explain
the arising of the toxicity. This approach is part of the before men-
tioned reductionist approach, and is made today possible by the
growing availability of large in vitro data sets.
4 QSAR as Induction
QSAR is basically a way to extract knowledge from observation. It

is a case of the inductive methods defined in philosophy, and uses
the mathematical tools of statistics and probability.
4.1 Induction The online Oxford English Dictionary defines induction as: “The
process of inferring a general law or principle from the observation
of particular instances (opposed to deduction).”
The problem of induction is the question of whether inductive
reasoning leads to knowledge. Inductive methods are essential in
science (as well as in everyday life), but they come to a cost.
Induction seems to lack of justification; it either makes a general-
ization of the properties of a class of objects based on a large num-
ber of observations of particular instances, or presupposes that a
sequence of events in the future will occur as it always has in the
past. To give an example, after observing that “all birds we have
seen fly,” the induction” all birds fly,” is true only before the obser-
vation that ostriches cannot fly.
A principle found by induction cannot be proved deductively.
That induction is opposed to deduction (as stated in the Oxford
Dictionary) is not always right. However, deductive logic is demon-
strable: the premises of an argument constructed according to the
rules of logics imply the argument’s conclusion. For induction
there are no complete theories to distinguish good from bad
inductions.
David Hume is considered the father of inductive reasoning.
Hume was interested in how we make causal connections, an
argument central in his project of developing the empirical science
of human nature and belief. Hume divided all reasoning into
QSAR: What Else? 91
demonstrative, i.e., deductive, and probabilistic, i.e., the general-

ization of causal reasoning. According to Hume, causal relations
are found not by reason, but by induction; this is because for any
cause, multiple effects are conceivable, and the actual effect can-
not be determined by reasoning about the cause. One must
observe occurrences of the causal relation to discover that it holds.
Since causal relations are central to the human reasoning and
all causal relations are found by induction, Hume worked on the
justification of induction. His justification introduced probabilistic
concepts: if one has found that this object has been always followed
by this effect, one foresees that other similar objects will be fol-
lowed by similar effects. In this way, the problem of induction is
concerned with the uncertainty of the conclusions. The Stanford
Encyclopaedia of Philosophy3 has a large covering of induction.
To connect the induction problem to QSAR, we can state that
the tools of statistics and probability “empirically” justify
induction.
4.2 The Role The Merriam-Webster dictionary defines statistics as: “a branch of
of Statistics mathematics dealing with the collection, analysis, interpretation,
and presentation of masses of numerical data.” Statistical analysis
infers properties about a population, which is assumed to be larger
than the observed data set.
Statistic is a well-developed tool used to derive information
from multiple observations. As such it is connected with induction.
Statistical inference is the process of deducing properties of an
underlying distribution by analyzing data.
Since all empirical scientific research uses statistics, the philoso-
phy of statistics is of key importance to the philosophy of science,
and is connected with the problem of induction, which concerns
how to justify the inferences that extrapolate from data to predic-
tions and general facts.
In 1953 Rudner wrote: “Since no hypothesis is ever completely
verified, in accepting a hypothesis the scientist must make the deci-
sion that the evidence is sufficiently strong or that the probability is
sufficiently high to warrant the acceptance of the hypothesis” [32].
“Sufficiently” is the key word. Consider when making experiments
with known error probability (the probability of rejecting a true
hypothesis or of accepting a false hypothesis). Here the problem of
induction is in reality a problem of decision, and the acceptability
of the results is in practice an optimization problem.
In the 70s of the last century, there was a flourishing of statistic
predictive models. Most of them were linear models, in particular,
proper models that used weights of the features (obtained in a way
to optimize the model properties), or improper models using ran-
dom weights (or unitary weights). These systems aimed at improving
3
https://plato.stanford.edu/entries/induction-problem/
92 Giuseppina Gini
decision-making in areas such as diagnosis and prognosis of diseases,

academic career, and student admissions. About these models Lovie
and Lovie [33] observed the flat maximum principle (FMP): in
many cases the improper unit-weight models are more reliable than
proper models. This may depend on the overfitting, since the proper
model fits some of the random peculiarities of the training data set.
FMP says that for a certain class of prediction problems, as long as
the signs of the coefficients are right, any linear model will predict
about as well as any other. FMP is restricted to difficult problems, for
which no proper model is reliable, and there are redundant and pre-
dictive evidential cues.
At that time the literature on predictive modeling was flourish-
ing. Trout and Bishop [34] wrote: “The lesson of this literature is
straightforward: For a very wide range of prediction problems, sta-
tistical prediction rules (SPRs), often rules that are very easy to
implement, make predictions than are as reliable as, and typically
more reliable than, human experts.” However, ethicists and episte-
mologists neglected the results obtained by statistical models.
Why? Trout and Bishop [34] in their work indicate that we have to
rethink our views about explanation, understanding, and good
reasoning.
4.3 The Role The problems considered by probability and statistics are inverse to
of Probability each other. In probability theory we consider some underlying
process, which has some uncertainty modeled by random variables,
and we figure out what happens. In statistics we observe some-
thing that has happened, and try to figure out what underlying
process would explain the observations.
The connection between induction and probability is given by
the Bayes rule. In an inductive inference problem there are data
D = d1, d2, … dn, and a set of hypotheses H = h1, h2, … hm. The
problem is to find out which of the hypothesis is the true hypoth-
esis that explains the data, or is the most plausible one. The Bayes
formula P(h|D) = P(D|h)P(h)P(D) computes the probability of
each hypothesis h given the data D (in the formula is P(h\D)),
using the inverse probability of observing such data given the
hypothesis (in the formula P(D\h)). P(D) is an independent con-
stant, leaving the computational burden on defining the P(D\h),
and on assigning the prior probabilities P(h). Conceptually, it is
hard to assign such P(h) probabilities before observing the data D.
Solomonoff [35] proposed a universal prior distribution that
unifies probability and uncertainty, thus answering the question:
given data about an unknown phenomenon, how to rate different
hypotheses and thereby select the hypothesis that best explains that
phenomenon? And how to use this hypothesis to predict new data
and measure the likelihood of that prediction being the right one?
The Solomonoff answer is the algorithmic probability theory
(APT). APT integrates philosophical principles with mathematics.
QSAR: What Else? 93
According to the principle of multiple explanations (see Epicurus)

we should keep all the hypotheses consistent with data: that is, in
Bayes terms, the ones with positive a priori probability. Moreover,
according to the Occam’s Razor, we prefer the simplest hypotheses
consistent with the observations: in Bayes terms, we give higher a
priori probabilities to the simplest hypotheses. If data and pro-
grams are considered as strings (as in the Turing machine), a short
program is preferred to a long one, given the Occam’s Razor prin-
ciple, so a high-probability observation string is the one computed
by a short program.
This idea serves to construct the prior probability distribution
for the given data; this distribution can be used in the Bayes rule
where priors are unknown to enable prediction under uncertainty.
The method is theoretically sound but not easily computable.
4.4 From Models People think that using statistics is a declaration of ignorance: we
to Theory do not have a deterministic knowledge, so we have to use statistics.
to Knowledge This debate has been active for about a century in atomic physics,
with the so-called Copenhagen interpretation, accepted by Bohr
and Heisenberg, and rejected by Planck and Einstein. The
Copenhagen interpretation states that physical systems generally
do not have definite properties prior to being measured, and quan-
tum mechanics can only predict the probabilities that measure-
ments will produce certain results; the act of measurement affects
the system, causing the set of probabilities to reduce to only one of
the possible values immediately after the measurement.
Even though models born to explain the property of atoms
cannot be exported to other fields of science, many of the consid-
erations advanced by the Copenhagen interpretation were indeed
taken from Hume.
Scientific models represent a phenomenon that is interesting in
science, as for instance the Bohr model of the atom. But we could
represent the same subject matter in different ways. Any model,
and QSAR too, has two different representational functions: it can
model a target or a theory. In the first case the model can be a rep-
resentation of a selected part of the world, the target system. In the
second case the model represents a theory, in the sense that it
interprets the laws of that theory. These two notions are not mutu-
ally exclusive in scientific models. Usually the knowledge about the
model can be translated into knowledge about the target system.
Models of data, introduced by [36], are an idealized version of
the data obtained from immediate observation, the raw data. After
the steps of data reduction and curve fitting what we get is a model
of data. Models of data play a crucial role in confirming theories
because it is the model of data and not raw data that we compare
to a theoretical prediction.
94 Giuseppina Gini
Models of theory are defined in formal logic; a model is a

structure that makes all sentences of a theory true. The theory is a
set of sentences in a formal language [37].
How do models relate to theories? The separation between
models and theory is often difficult, if not impossible. In the syn-
tactic view, if a theory is a set of sentences in first order logic, then
a model is just a system of rules. Such models are irrelevant to sci-
ence; they are at best of pedagogical, aesthetical, or psychological
value, since they are too coarse for many real applications [38].
The semantic view of theories [39] declares that a theory is indeed
just a family of models.
An extreme case is the use of a model when there are no theo-
ries at all available. We encounter this situation in all domains, but
in particular in biology and economy, where the models have the
role of substituting a not-available theory.
In many cases we expect that explaining a model mean indicat-
ing the laws of nature that play a role in the model. Models are
considered as tools to find out the causal relations that hold
between observed facts and processes, and those relations are con-
sidered explanatory. Cartwright [40] suggested the so-called simu-
lacrum account of explanation: we explain a phenomenon by
constructing a model that fits the phenomenon into the basic
framework of a theory [40]. On this account, the model itself is the
explanation we seek.
Hempel and Oppenheim [41], searching for a formal logical
definition of laws and theories, indicated that there are different
levels of explanation. The first level is connecting observable char-
acteristics under a general law, while higher-level explanations use
different procedures. One of them is to rely on theories of micro-
structures, as in the case of the kinetic theory of heat. Scientific
explanations in many fields rely on those micro-theories.
Often, some properties are unexpected, and unpredicted by
the available theories. For instance, the transparency of water can-
not be simply derived from properties of the constituent oxygen
and hydrogen atoms, but will require knowledge about how atoms
combine into molecules. If a property is unexpected and not
explainable with available theories, we call it emergent. Some prop-
erties are emergent, while other properties, for instance the molec-
ular weight, can be easily predicted by computing a formula. A
characteristic of a whole system is deemed emergent if it cannot be
predicted from the knowledge of its parts. So the concepts of
explanation and prediction share some properties.
Many general rules would have never been found without the
observation of some phenomena; on the basis of data available at a
given time science makes generalizations, and those generaliza-
tions are good until they are no more able to predict newly
observed phenomena. For instance, the periodicity of certain
QSAR: What Else? 95
chemical elements allowed Mendeleev to predict the existence of

elements discovered later on.
Emergence of a characteristic is not an ontological quality of
the phenomenon; it is instead an indication about the scope of our
knowledge, at a given time. The usual assertion that the phenom-
ena of life are emergent has also a more precise definition: life can-
not be explained with the physicochemical knowledge available
today. In fact, the description of these phenomena requires terms
that are not available in the present scientific vocabulary, and hence
cannot be inferred from the available knowledge.
5 Epistemology and QSAR: Validation, Justification, and Model Interpretability
Epistemology studies the nature and the extent of knowledge. It

proposes different definitions for the different kinds of knowledge,
making a distinction among procedural knowledge (something
one is able to do), acquaintance knowledge (something we know
by observation), and propositional knowledge (a set of declarative
sentences describing the state of the world).
Propositional knowledge can derive from both experience and
reasoning. Rationalists state that all knowledge is grounded on rea-
son, while empiricists view all knowledge as grounded upon expe-
rience. The debate between these extreme views has animated the
research in artificial intelligence and cognitive development.
But cognition is also a reflexive activity: one knows something
if they are aware of that something. Knowledge is a kind of belief,
but not every belief is knowledge, since beliefs can be either true or
false. Only true beliefs constitute knowledge, so beliefs should be
justified. Truth and justification should be connected, there should
be a way to matching one’s mind and the world. So there are two
approaches to constructing justification: in terms of the believer’s
mind, and in terms of the world. This is the origin of the internalist
versus externalist debate in epistemology.4
According to the internalist view, belief is a mental state, so
beliefs are justified by examining the thought-processes of the
believer during their formation, for example using the logic con-
cepts of consistency, implication, and coherence. According to
internalism, external justification is not requited for knowledge.
But focusing solely on factors internal to the believer’s mind could
lead to a mistaken account of justification.
According to externalism, the only way to ensure that knowl-
edge is sound is to consider some factors other than the individu-
al’s other beliefs. So justification requires using other processes
than beliefs.
4
https://plato.stanford.edu/entries/justep-intext/
96 Giuseppina Gini
This debate is important when discussing about causal reason-

ing. Internalists claim that every condition that determines the jus-
tification of a belief is internal, but causal relations are not internal.
Causal reasoning requires an externalist view.
In the externalist view, the concept of justification is analyzed
in terms of degree of evidential support. The acceptability of the
results of models largely depends on the quality of the model itself
(i.e., its validation), and on the way people understand and com-
municate it (i.e., justification and interpretability). This is true also
for QSARs.
5.1 Validation QSAR model validation is an essential activity to verify whether a

model inferred from data has the necessary predictive power.
Machine learning and statistics are the main tools used to build
QSAR, with a continuum of the two methods: statistics tradition-
ally has been concerned with testing hypotheses whereas machine
learning has been more interested in knowledge representation.
Many new methods evolved in the machine learning and data min-
ing communities; see as an example, WEKA, a popular tool for
developing data mining applications [42].
Today QSARs use a mix of statistics (for data visualization,
selection of attributes, and model validation) and of machine learn-
ing (in the algorithms). In [42], learning is distinguished from sta-
tistical approaches in that it is a search through a space of possible
concept descriptions to obtain the one that fits the data. In machine
learning important decisions are about the following:
–– The description language: if-then rules and so on.
–– The order in which the space is searched, as for instance greedy
search, generalization, specification, or other.
–– The way to avoid overfitting: stopping criterion, pruning, and
so on.
Different leaning algorithms correspond to different concept
description spaces, and different search criteria.
Evaluating what has been learned is the final step of modeling.
Moreover, the cost of making errors may depend on the kind of
error: false positive or false negative can have very different conse-
quences in the application domain. In general, sensitivity and spec-
ificity are important as much as accuracy.
A good model should always minimize the error in prediction.
The typical curves about the error are presented in Fig. 1: the error
is related to the complexity of the model, and the error is mini-
mized in a trade-off between underfitting and overfitting the
model.
The performance of the model is usually expressed as the error
rate. The error rate on the training set is not a good indicator of
future performance on new data, so the error rate should be computed
QSAR: What Else? 97
Fig. 1 The error of underfitted, good, and overfitted models
on a test set, a set of data not used in any way in the process of
building the model. The implicit assumption is that both the train-
ing and the test sets are taken from the same population. To avoid
the bias of choosing a bad test set, the process should be repeated
more times with randomly selected test and training sets, and the
error averaged.
Other validation methods include cross-validation; it should
assure that the random sampling of the test guarantees that each
class is properly represented. In cross-validation we fix a number of
folds, or partitions of the data. In the case of tenfold cross-
validation, the dataset is split in 10 equal partitions; 9 are used for
training and 1 for testing, repeating the procedure ten times so
that any instance is used at once for testing. Finally, the ten errors
are averaged to obtain the overall error estimate. The choice of 10
has practical and theoretical evidence, but other numbers, as 5, can
work as well. Cross-validation provides a method of high accuracy
and low variability in estimating the model performance. Leave-
one-out is n-fold cross-validation for n = 1. The procedure is inter-
esting since it uses a large training set and no random split is
involved. However, it can give artificial error rates, since the test set
contains only one element.
In case there is a plenty of data (as in the big data applications
on web data) it is straightforward to use many of them for training
and the remaining for testing. The results on the test set are con-
sidered to be a true picture of the predictive capability of the
model. But there is a trade-off: the larger the training set, the bet-
ter the classifier; the larger the testing set, the better the error esti-
mate on future new items.
A common practice in QSAR is to use an external test set [43].
However, both statistical and experimental analyses show that the
external test set is not always the best choice. Gütlein et al. [44]
report a large evaluation of cross-validation and external test set
validations using a big dataset of 500 and a small dataset of 100
compounds; the authors consistently found that cross-validation
gives a more pessimistic view of the model performance than exter-
98 Giuseppina Gini
nal test set. Moreover, since in QSAR often data are scarce, cross-
validation can be the best evaluation choice for small data sets.
The bootstrap is another estimation method based on sam-
pling with replacement. The idea is to use the same instance twice,
thus sampling the data set of n instances to obtain a new data set
of n instances where some elements are missing and some are
repeated. The bootstrap procedure is repeated many times with
different replacements and the errors averaged. This could be
another good way to measure the error in small datasets.
The coefficient of determination, which indicates the propor-
tion of the variance in the dependent variable that is predictable
from the independent variable(s), called R2, is usually adopted in
regression QSARs. Its value ranges in (0–1) and the values com-
puted on the training set and on the test sets should be as near as
possible to 1 and similar. Several other measures are available, as
the commonly mean squared error (MSE) and the mean absolute
error (MAE).
Another useful and often neglected way to assess the value of a
classifier is the ROC (receiver operating characteristic) curve. The
ROC curve plots the percentage of positive on the total number of
positive in the test set on the vertical axis, and the percentage of
negatives on the horizontal axis (this applies also to the training
set). The result is a line that starts in (0, 0) and ends in (1, 1). The
curve of a perfect classifier immediately reaches (0, 1) and is then
constant to (1, 1), while good classifiers approach that ideal curve.
Different classifiers have different ROC curves that can be com-
pared and interpreted.
ROC graphics has been extended to regression models; REC
(Regression Error Characteristic) curves use a different range of
values on the x-axis, giving an effective representation, seldom
used in QSAR [45]. See Fig. 2 for an example REC curve compar-
ing three models: the best model, which corresponds to the lowest
squared residual, is characterized by the upper curve.
Sometimes we are interested in comparing different training
schemes on the same dataset. In this case it is not completely cor-
rect to compare their errors in any of the before mentioned meth-
ods, but it is necessary to apply statistics tests, as the t-test or the
student’s-t-test. Often in the QSAR literature only the R2 are
reported; they are biased since the coefficient of determination
increases with the number of independent variables.
5.2 Justification According to Stanford Encyclopaedia of Philosophy,5 justification

can be a priori or a posteriori. A priori justification is independent
of experience: it rests on rational intuitions, or insights. One has a
priori justification in believing some proposition if, and only if, that
justification rests solely on one’s understanding the proposition
5
<https://plato.stanford.edu/archives/fall2014/entries/justep-intext/>
QSAR: What Else? 99
0.9
mean(0.47838)
0.8 NNgdx4(0.35436)
NNbr3(0.32655)
0.7
0.6
Accuracy
0.5
0.4
0.3
0.2
0.1
0
0 0.5 1 1.5 2 2.5
Squared residual
Fig. 2 Comparing three different regression models represented as REC curves
that is the object of one’s belief. By contrast, a posteriori justifica-

tion does not rest solely on understanding such a proposition.
Rationalist philosophers think that there can be a priori justifica-
tion and knowledge of the world, while empiricists deny this.
In conclusion, one knows some proposition on the basis of
evidence, or good reasons, or experience. The same happens for
justified beliefs, i.e., beliefs justified on the basis of evidence, or
reasons, or experience, or on the basis of the process that produces
those beliefs. Do justified beliefs represent knowledge?
According to internalism, a person has an access to the basis
for knowledge and justified belief: what justifies any belief is some
mental state. Externalists, by contrast, deny that one always can
have access to the basis for one’s knowledge and justified belief;
they propose that something other than mental states operates as
justifiers.
Another point is whether justification should be based on true
theories, or on evidential reasoning, or on induction. So let us go
to the definition of true theories, as given by Popper, and to his
solution of the induction problem. In inductive reasoning one
makes a series of observations and infers a new claim based on
them. Is this the way to produce new knowledge? As we have seen,
Hume raised this problem in philosophy, arguing that causal rela-
tionships are found by induction, not by reasoning. This means
that a deductive justification for induction cannot be provided.
Two centuries later Popper rejected the inductivist view on the
scientific method in favor of what he called empirical falsification.
100 Giuseppina Gini
According to Popper, theories are not generated by induction but

by conjectures; a theory in the empirical sciences can never be
proven, but only falsified. Every theory should be scrutinized by
experiments to be eventually rejected. Scientific theories, and
human knowledge, are conjectural and generated by the i magination
of people aimed at solving real problems encountered in specific
historical situations.
Popper claimed having solved the philosophical problem of
induction: it is not possible to prove that a theory is true, but a
theory that gives results in accordance with the observations can be
used until new data will falsify it. According to Popper, the most
easily falsifiable theory should be preferred. This view is opposite
to the principle of positivism, which states that one should prefer
the theory that is most likely to be true. According to Popper, it is
impossible to ensure a theory to be true; it is more important that
its falsity can be detected as easily as possible. No number of posi-
tive outcomes from experimental testing can confirm a scientific
theory, but a single counterexample can show that the theory from
which the implication is derived is false.
Popper thinks that once we understand that falsifiability is the
mark of science, we are able to respond to Hume’s problem of
induction. The solution is that science does not use induction; it
does not proceed by reaching conclusions about unobserved things
on the basis of observed ones. Science never believes that its theo-
ries are true; rather, it treats its theories as mere provisional conjec-
tures and accepts them only tentatively.
5.3 Interpretability We are inclined to an increasing trend in using models learned

from data in monitoring and control applications, as in driverless
cars, drones, and autopilots. In toxicology the main reason for
using models is that analytical models, derived from first principles,
are largely unknown. The problem is common to other domains,
where physical rules do exist but their direct application gives low
accuracy, because of nonmodeled randomness or simplifications
introduced.
Deploying data-driven models in applications where incorrect
outputs of the model may have fatal consequences requires ensur-
ing that the model is correct, or quantifying the error it may intro-
duce in making predictions. This is not enough in the case where
people want to “understand” the model before accepting its
predictions.
Linear regression models have a long tradition of use in many
problems. Regulators and medical people are accepting linear
regression models based on few predictors, as they seem easy to
understand and to communicate.
Interpreting supervised statistical models is to understand how
a set of predictor variables X is associated with a response variable
Y, and what is the structure of the association.
QSAR: What Else? 101
Nonparametric regression assumes no specific form of the

regression function, and adapts to the unknown regression func-
tion a smooth curve derived from the empirical data; its
interpretation is more complicated, but possible with various tech-
niques. Other kinds of models, as rules and trees, are considered
being interpretable.
For real applications, are we looking for an accurate and pre-
dictive model or for a model simple to explain? There is a trade-off
with accuracy. For example, a tree model remains interpretable as
long as the number of nodes is small, which means that the model
is quite coarse. In contrast, Random Forests have higher accuracy
but loose interpretability.
In general the models that best emulate nature in terms of
predictive accuracy are also the most complex. But this dilemma
can be resolved by realizing the wrong question is being asked: the
goal of statistical analysis is to get useful information about the
relation between the response and predictor variables.
Interpretability is also a way of getting information, but a model
does not have to be simple to provide reliable information about
the relation between predictor and response variables. According
to Brieman [22], “The approach is that nature produces data in a
black box whose insides are complex, mysterious, and, at least,
partly unknowable. There is a set of x’s that go in and a subsequent
set of y’s that come out from the box. The problem is to find an
algorithm f(x) such that for future x in a test set, f(x) will be a good
predictor of y”.
According to Breiman, three important lessons are derived
from the algorithmic view of modeling [22]:
1. Rashomon, or the “multiplicity of good models.” Rashomon (a
Japanese movie in which four people witness an incident, and
when testifying report the fact with four very different individ-
ual stories) indicates that there is a multitude of different mod-
els giving about the same minimum error rate. Which one is
better is a wrong question.
2. Occam, or the “conflict between simplicity and accuracy.” The
Occam’s Razor is the lesson that simpler is better. Unfortunately,
in prediction, accuracy and simplicity are in conflict. Accuracy
generally requires more complex prediction methods.
3. Bellman, and the reduction of “dimensionality.” The Bellman’s
phrase “the curse of dimensionality” has been interpreted as
searching for the few features that contain most of the informa-
tion, considering high dimensionality as dangerous. But reduc-
ing dimensionality reduces also the amount of information
available for prediction. Instead of reducing dimensionality,
increase it; this is the lesson implicit in well performing meth-
ods, such as Support Vector Machines [46].
102 Giuseppina Gini
Going back to specific problems of QSAR, the recent contribution

of Polishchuk [47] discusses how the concept of interpretation in
QSAR evolved from the paradigm of interpreting descriptors, to
implicitly understand the role of structure, and then to reason
directly from model to structure. Even in case no interpretable
descriptors are used, any model can be made interpretable by esti-
mating the contributions of the molecule substructures. This
approach, implemented in various cases, is similar to what is pro-
posed in [30].
6 Conclusions
Today, the dominant paradigms used to interpret nature are based

on Newtonian and Darwinian approaches. As Ulanowicz [49]
observes, neither is enough to reach a comprehensive description
of complex living systems that do not obey the Newtonian view,
characterized by causal closure, atomism, determinism, and univer-
sality of physical laws. Biological systems obey specific postulates,
where chance and feedback have a role. New models should be
actively pursued.
QSAR is an alive and evolving set of methods that embraces
different sciences and constitutes one of the most advanced body
of scientific models. Using QSAR is either a necessity or a plus in
many human activities, from chemical design to chemical
assessment.
Today, toxicologists propose a combined approach to toxicity,
where the definition of toxicity in the whole organism (as repre-
sented by most of the classical endpoints) is contrasted with an ab
initio analysis that starts from the chemical interaction between the
substance and the cell and tries to draw the sequences of following
events, using in vitro and toxicogenomics analysis [48] (see Fig. 3).
Fig. 3 Top-down and bottom up strategies for toxicity evaluation

This process is meant to be mechanistic, and also to be expressed

as a complete set of transformation rules. The ideal encounter of
both lines of analysis is, at an abstraction level not yet defined, a
core micro-model that explains how the function of the cell is dis-
rupted by the chemical.
Besides some data that are already available at the organ and
the cell level, what QSAR tells us about what happens inside a cell
is only a hypothesis. So to create this ideal micro-model it is neces-
sary to use other experimental data and methods.
It is worth mentioning that the massive experimental data at
cell and gene levels are managed through statistical tools, and that
the results also obtained by induction share the same problems of
QSAR: either they are tentatively accepted or they need to be falsi-
fied by other experiments. This is a crucial point, since the
falsification of results from genetic analyses is quite common, con-
sidering the low repeatability of these tests.
In conclusion, this ambitious project for producing a complete
explanation for all the transformations that allow the toxic effect to
manifest has again to deal with all the epistemological problems so
far encountered in QSAR. This long-time process may use some of
the methods already developed in personalized medicine, but at
the cost of making it difficult to generalize.
Statistical/predictive methods may exit through the window
and reenter by the door.
QSAR, what else?
References
1. Hansch C, Maloney PP, Fujita T, Muir RM state information. J Chem Inf Comput Sci
(1962) Correlation of biological activity of 35:1039–1045
phenoxyacetic acids with hammett substitu- 7. Connolly ML (1985) Computation of molecu-
ent constants and partition coefficients. Nature lar volume. J Am Chem Soc 107:1118–1124
194:178–180 8. Karelson K, Lobanov VS, Katritzky AR (1996)
2. Hansch C, Fujita T (1964) p-σ-π analysis. A Quantum-chemical descriptors in QSAR/
method for the correlation of biological activ- QSPR studies. Chem Rev 96:1027–1044
ity and chemical structure. J Am Chem Soc 9. Wold S, Sjostrom M, Eriksson L (2001) PLS-
86:1616–1626 regression: a basic tool of chemometrics.
3. Hansch C (1969) Quantitative approach to Chemom Intell Lab Syst 58:109–130
biochemical structure-activity relationships. 10. Rogers D, Hopfinger AJ (1994) Application of
Acc Chem Res 2:232–239 genetic function approximation to quantitative
4. Free SM, Wilson JW (1964) A mathemati- structure-activity relationships and quantitative
cal contribution to structure-activity studies. structure-property relationships. J Chem Inf
J Med Chem 7:395–399 Comput Sci 34:854–866
5. Kier LB, Hall LH, Murray WJ, Randić M 11. Li L, Hu J, Ho Y-S (2014) Global performance
(1975) Molecular connectivity I: relationship and trend of QSAR/QSPR research: a biblio-
to non specific local anesthesia. J Pharm Sci metric analysis. Mol Inform 33:655–668
64:1971–1974 12. Cherkasov A, Muratov EN, Fourches D,
6. Hall LH, Kier LB (1995) Electrotopological Varnek A, Baskin II, Cronin MTD et al (2014)
state indices for atom types: a novel combi- QSAR modeling: where have you been? Where
nation of electronic, topological, and valence are you going to? J Med Chem 57:4977−5010
104 Giuseppina Gini
13. Cramer RD, Patterson DE, Bunce JD (1988) 27. Cronin MTD, Schultz W (2003) Pitfalls
Comparative molecular field analysis (CoMFA). in QSAR. J Mol Struct (THEOCHEM)
1. Effect of shape on binding of steroids to car- 622:39–51
rier proteins. J Am Chem Soc 110:5959–5967 28. Alves V, Muratov E, Capuzzi S, Politi R, Low
14. Golbraikh A, Tropsha A (2002) Beware of q2! Y, Braga R et al (2016) Alarms about structural
J Mol Graph 20(4):269–276 alerts. Green Chem 18:4348–4360
15. Gramatica P (2007) Principles of QSAR mod- 29. Ferrari T, Gini G (2010) An open source multi-
els validation: internal and external. QSAR step model to predict mutagenicity from statis-
Comb Sci 26:694–701 tic analysis and relevant structural alerts. Chem
16. OECD principles for the validation, for regu- Cent J 4(Suppl 1):S2. (online http://www.
latory purposes, of (quantitative) structure- journal.chemistrycentral.com/content/4/
activity relationship models. Organization S1/S2)
for economic co-operation and development 30. Gini G, Franchi AM, Manganaro A, Golbamaki
(2004) http://www.oecd.org/env/ehs/risk- A, Benfenati E (2014) ToxRead: a tool to assist
assessment/37849783.pdf in read across and its use to assess mutagen-
17. José Ayala F, Dobzhansky T (eds) (1974) icity of chemicals. SAR QSAR Environ Res
Studies in the philosophy of biology: reduction 25:999–1011
and related problems. University of California 31. Benfenati E, Roncaglioni A, Petoumenaou M,
Press, California Cappelli C, Gini G (2015) Integrating QSAR
18. Popper KR (1974) Scientific reduction and and read across for environmental assessment.
the essential incompleteness of all science. In: SAR QSAR Environ Res 26:605–618
Ayala FJ, Dobzhansky T (eds) Studies in the 32. Rudner R (1953) The scientist qua scientist
philosophy of biology. Palgrave, London makes value judgments. Philos Sci 20:1–6
19. Schummer J (1999) Coping with the growth 33. Lovie AD, Lovie P (1986) The flat maximum
of chemical knowledge: challenges for chem- effect and linear scoring models for prediction.
istry documentation, education, and working J Forecast 5:159–168
chemists. Educación Química 10:92–101 34. Trout JD, Bishop M (2002) 50 years of suc-
20. Gòmez Bombarelli R, Duvenaud DK, cessful predictive modeling should be enough:
Hernàndez Lobato JM, Aguilera-Iparraguirre lessons for philosophy of science. Philos Sci
J, Hirzel TD, Adams RP, Aspuru-Guzik A 69(S3):S197–S208
(2016) Automatic chemical design using a 35. Solomonoff RJ (1964) A formal theory of
data-driven continuous representation of mol- inductive inference: parts 1 and 2. Inf Control
ecules. ACS Central Science, Washington, DC 7:1-22–224-254
21. Gini G, Ferrari T, Cattaneo D, Golbamaki 36. Suppes P (1962) Models of data. In Studies
Bakhtyari N, Manganaro A, Benfenati E (2013) in the methodology and foundations of sci-
Automatic knowledge extraction from chemi- ence. Selected Papers from 1951 to 1969,
cal structures: the case of mutagenicity predic- Dordrecht, Reidel. pp. 24–35
tion. SAR and QSAR Environ Res 24:365–383 37. Hodges W (1997) A shorter model theory.
22. Brieman L (2001) Statistical modeling: the Cambridge University Press, Cambridge
two cultures (with comment and a rejoinder by 38. Bailer-Jones DM (2003) When scientific mod-
the author). Stat Sci 16:199–231 els represent. Int Stud Philos Sci 17:59–74
23. Rissanen J (1978) Modeling by shortest data 39. Giere R (1988) Explaining science: a cogni-
description. Automatica 14:465–658 tive approach. University of Chicago Press,
24. Wolpert D (1996) The lack of a priori distinc- Chicago
tions between learning algorithms. Neural 40. Cartwright N (1983) How the laws of physics
Comput 8:1341–1390 lie. Clarendon Press, Oxford
25. Benfenati E, Gini G, Hoffmann S, Luttik R 41. Hempel CG, Oppenheim P (1948) Studies in
(2010) Comparing in vivo, in vitro, in Silico the logic of explanation. Philos Sci 15:135–175
methods and integrated strategies for chemi-
cal assessment: problems and prospects. ATLA 42. Witten H, Frank E (2000) Data mining: prac-
38:153–166 tical machine learning tools and techniques
with java implementations. Morgan Kaufmann
26. Benfenati E, Gonella Diaza R, Cassano Publishers, London
A, Pardoe S, Gini G, Mays C et al (2011)
The acceptance of in silico models for 43. Benfenati E, Crètien JR, Gini G, Piclin N,
REACH. Requirements, barriers, and perspec- Pintore M, Roncaglioni A (2007) Validation of
tives. Chem Cent J 5:58 the models. In: Benfenati E (ed) Quantitative
structure-activity relationships (QSAR) for
pesticides regulatory purposes. Elsevier, 46. Vapnik VN (1995) The nature of statistical
Amsterdam, pp 185–200 learning theory. Springer-Verlag, Berlin
44. Gütlein M, Helma C, Karwath A, Kramer S 47. Polishchuk PG (2017) Interpretation of QSAR
(2013) A large-scale empirical evaluation of models: past, present and future. J Chem Inf
cross-validation and external test set validation Model 57(11):2618–2639
in (Q)SAR. Mol Inform 32:516–528 48. Hartung T (2017) Food for thought. Opinion
45. Bi J, Bennett K P (2003) Regression error versus evidence for the need to move away
characteristic curves. Procs of the Twentieth from animal testing. ALTEX 34:193–200
international conference on machine learning 49. Ulanowicz RE (2009) A third window: natural
(ICML-2003), Washington DC life beyond Newton and Darwin. Templeton
Foundation Press, West Conshohocken
Chapter 4
(Q)SARs as Adaptations to REACH Information

Requirements
Toni Alasuvanto, Andrea Gissi, Tomasz Sobanski,
Panagiotis Karamertzanis, and Mike Rasenberg
Abstract
REACH is a regulation of the European Union adopted to improve the safe use of chemicals with regard
to human health and the environment. The safe use of chemicals can be achieved only if the hazard and
the exposure of the substances are well characterized. Testing on animals has been traditionally the main
tool for hazard assessment. For ethical and economic reasons, alternative ways of testing that do not use
laboratory animals have been developed by different parties (regulatory agencies, researchers, industry)
over the recent decades, and their proper use in hazard assessment is encouraged under REACH. In this
chapter, we describe how (Q)SAR models and predictions are included into REACH and their adequate
use promoted by the European Chemicals Agency (ECHA).
Key words (Q)SAR, REACH, EU regulation, Alternative methods, ECHA
1 Introduction
Regulation (EC) No 1907/2006 of the European Parliament and

of the Council of 18 December 2006 concerning the Registration,
Evaluation, Authorisation and Restriction of Chemicals [1, 2]
(commonly known as REACH) is an EU legal act for chemicals
and entered into force on 1 June 2007.
The main objectives of REACH are stated in Article 1: to
ensure high level of protection of human health and the environ-
ment, to promote free movement of substances (on their own, in
mixtures and in articles) on the internal market while enhancing
competitiveness and innovation. The regulation promotes the
development of alternative methods for assessment of hazards of
substances.
*The views and opinions expressed in this chapter represent exclusively the personal ideas of the authors and do not
represent the official position of the Agency.
107
108 Toni Alasuvanto et al.
REACH ensures safe use of “industrial chemicals” through all

life cycle stages of products that include chemicals up until the
waste stage. Risk management measures are foreseen, both at user
(company) level as well as at EU community level, for instance by
the restriction of certain uses of chemicals of very high concern.
One of the crucial elements of REACH is the requirement for
companies manufacturing or importing more than one metric ton
of a substance per year, to submit a registration dossier to the
European Chemicals Agency (ECHA). This dossier should contain
all relevant and available information on the intrinsic properties,
including information to fulfill the minimum requirements. The
intrinsic properties include (eco)toxicological information, envi-
ronmental fate, and physical-chemical properties. The minimum
requirements are defined to ensure a basic understanding of the
(hazardous) properties of the chemicals to be able to do a risk
assessment, using also the use and exposure information that is to
be submitted as well as part of the dossier.
The minimum requirements increase per increased volume,
e.g., a 1–10 ton per year dossier has less minimum requirements
than a more than 1000 ton per year dossier. This approach is com-
mon practice in regulatory systems worldwide to balance the
resource needs versus the expected exposure, where annual volume
is taken as a proxy for potential exposure.
To ensure consistency in the outcome and to avoid duplication
of testing, companies registering under REACH need to gather
and generate information where needed, together with other com-
panies who are registering the same substance.
REACH indeed requires companies to gather and share all
available and relevant information. When minimum information
requirements are not met with the available information, addi-
tional testing might be needed. This testing should be done as a
last resort, as REACH gives opportunities to either avoid testing
(e.g., testing is not possible) or to generate alternative informa-
tion, not using animal studies. When alternative information is
used, this information should still be sufficient to characterize the
hazards to such a level that a risk assessment can be performed. If
tests are still required which would involve vertebrate animal test-
ing, registrants are in most cases obliged to first ask permission
from the agency and the EU Member States to conduct the tests,
via the so-called testing proposal process.
In the following subheadings, we will discuss how companies
in general can adapt testing requirements and how they have used
these adaptations in their registrations until now. Finally, we will
also see how (Q)SAR methods fit into this context and can be used
adequately under REACH, including advice on good practices.
(Q)SARs as Adaptations to REACH Information Requirements 109
2 Adaptations to Standard Information Requirements
REACH Annexes VII–X list the so-called “standard information

requirements,” i.e., the minimum information that has to be
included in the registration depending on the tonnage of the sub-
stance. These requirements include information to be obtained
by testing the substance, for some endpoints on animals. However,
column two of the same Annexes give specific rules for adapta-
tion from the standard requirements. As an example of this, the
bioaccumulation study in aquatic species need not to be con-
ducted if the substance is of low potential for bioaccumulation
(e.g., log Kow ≤3) and/or of low potential to cross biological
membranes. Another example of specific rules for adaptation is
that the in vivo testing for skin sensitization does not need to be
conducted if the substance is a strong acid or base. Further on,
Annex XI lists general rules for adaptations and the conditions to
use such adaptations. The possibilities for adaptations listed in
Annex XI are as follows:
(a) Testing does not appear scientifically necessary.
(b) Testing is technically not possible.
(c) Substance-tailored exposure-driven testing.
Point (a) contains the following subcategories: use of existing data,
weight of evidence, qualitative or quantitative structure–activity
relationship (Q)SAR, in vitro methods, grouping of substances,
and read-across approach. In the next chapter, the requirements
for the use of (Q)SARs are further described.
Adaptations can be used either individually or combined in a
weight-of-evidence approach (e.g., use of (Q)SAR and informa-
tion from read-across in combination with literature evidence
and/or some properties indicating the possible fate of a sub-
stance). In all cases, the data used must be adequate, reliable, and
relevant for the particular endpoints, and must follow the criteria
set out in Annex XI.
According to REACH Art. 117.3, every 3 years ECHA has to
submit a report on the status of implementation and use of nonani-
mal test methods and testing strategies used to generate informa-
tion on intrinsic properties and for risk assessment to meet the
REACH requirements. Figure 1, extracted from the latest report
[3], published in 2017, summarizes the use of adaptations in
REACH registrations so far.
Registrants make an extensive use of existing information and
the various adaptation possibilities instead of conducting new
studies or proposing new high tier vertebrate animal tests. In gen-
eral, for all analyzed substances for the endpoints concerning ver-
tebrate animals:
Fig. 1 Relative proportions of the options used by registrants to cover REACH information requirements. In this
figure, omitted studies include both REACH Annexes VII–X Column 2 and Annex XI adaptations
(a) 89% contain at least one endpoint in the dossiers where an

adaptation or other argument was provided instead of a study
result;
(b) 63% contain at least one read-across adaptation;
(c) 43% contain at least one weight-of-evidence argument; and
(d) 34% contain at least one (Q)SAR prediction.
Based on the relative amount of experimental data available
and adaptations used by registrants, three groups of endpoints can
be identified: low tier endpoints, high tier human health endpoints
and high tier environmental endpoints. Experimental data are
available for 66%, 40%, and 9% of substances, on average across
endpoints within the three groups, respectively.
For low tier endpoints (acute rodent toxicity, skin corrosion/
irritation, serious eye damage/eye irritation, skin sensitization, and
short-term toxicity to fish), the main source of information is
experimental studies, with a high percentage of them carried out
before REACH.
Less experimental data are available for high tier human health
endpoints (repeated dose toxicity (all routes of administration, all
durations), genetic toxicity in vivo, developmental toxicity, toxicity
to reproduction, and carcinogenicity) compared to low tier end-

points. Read-across is the most used alternative approach, followed
by weight of evidence.
For high tier environmental endpoints (bioaccumulation, long-
term fish toxicity, and long-term toxicity to birds), adaptations are
much more common than experimental data and much less exper-
imental data is available compared to the low tier endpoint short-
term fish toxicity. Data waiving is used most frequently, followed
by (Q)SARs and read-across. The endpoints where (Q)SARs have
been (in proportion) more used so far are bioaccumulation and
aquatic toxicity.
The finding that (Q)SARs were not largely used for human
health high tier endpoints is in line with ECHA’s report on “regu-
latory applicability of non-animal approaches under the REACH,
CLP and Biocidal Products Regulations” [4]. The report states
there is progress regarding alternatives to animal testing on the
lower tier properties of substances, but a full replacement of chem-
ical toxicity testing by nonanimal approaches is not yet foreseeable.
Further dialogue between research and regulatory communities is
needed to put scientific developments faster to regulatory use.
To increase the robustness and regulatory acceptance of those
adaptations for high tier human health endpoints, additional data is
needed, particularly related to toxicological mechanisms and absorp-
tion, distribution, metabolism, and excretion (ADME) properties.
New approach methodologies (e.g., high throughput in vitro
screening) have also a potential to further substantiate the justifica-
tions for adaptations. As these approaches often use starting points
which are directly relevant for humans (e.g., human liver cells),
more relevant data can be obtained.
More results and details on the analysis can be found in
ECHA’s report “The use of alternatives to testing on animals for
the REACH Regulation” [3].
3 How to Use and Report (Q)SAR Predictions for REACH
As stated in Subheading 2, (Q)SAR predictions are one way to

adapt information requirements.
REACH Annex XI par. 1.3 defines the general rules for using
(Q)SAR as an adaptation of the standard testing regime:
1. Results are derived from (Q)SAR models whose scientific
validity has been established.
2. The substance falls within the applicability domain of the (Q)
SAR model.
3. Results are adequate for the purpose of classification and label-
ing and/or risk assessment.
4. Adequate and reliable documentation of the applied method is

provided.
This subheading analyzes good practices to satisfy the above
conditions, by summarizing the content of ECHA’s “Practical
guide on how to use and report (Q)SARs” [5]. More detailed
guidance is otherwise given in chapter R6 of ECHA’s Guidance on
information requirements and chemical safety assessment [6].
Before elaborating on the four rules defined by REACH, some
preliminary considerations are needed.
REACH allows registrants to use information from (Q)SAR
models as stand-alone for full replacement of animal testing and as
part of weight of evidence. Because additional information is in any
case needed to assess the reliability of the prediction, it is preferable
to use the prediction as a part in a more general weight of evidence
context. Especially for high tier endpoints, the (Q)SAR models
available today do not capture the full complexity of biological
systems and cannot be used as stand-alone tools. Only for lower
tier endpoints, (Q)SARs might be accepted to fulfill information
requirements, when they are scientifically valid, adequate, and well
documented.
3.1 Substance The vast majority of (Q)SAR models require the input of a SMILES
Characterization (i.e., a structure) to predict a result. Substances registered under
REACH can be of different types: monoconstituents, multicon-
stituents, and substances with unknown or variable composition,
complex reaction products, or biological materials (UVCB).
Monoconstituents are substances where the main constituent has a
concentration ≥80%, and the remaining part of the composition is
impurities and additives. Multiconstituents are substances consist-
ing of several main constituents, each of them at a concentration
between 10% and 80%. UVCBs are substances where the composi-
tion can be variable or difficult to predict, and some constituents
are unknown. The first aspect worth taking into account is the
concept of substance type and compositions in REACH.
While for monoconstituent substances, the selection of the
input structure might be simple (and yet the potential (lack of)
toxicity of impurities and additives needs to be discussed), the
choice of the input for multiconstituents and UVCBs is not trivial.
One way to address multiconstituents is to predict the different
structures individually, and then select the highest toxicity among
the predicted values for further hazard and risk assessment
calculations. Such an approach does not take into account possible
synergies in the toxicological mode of action of the component,
and therefore needs to be justified. For UVCBs, the situation is
even more complex, because some of the structures may be even
unknown. In this case, the input may include one or more repre-
sentative structures supposed to cover the range of expected toxic-
ity of the UVCB substance [7].
3.2 Results Are The validity of a model, for regulatory purposes, of (Q)SARs can
Derived from (Q)SAR be assessed according to the principles set by OECD [8]. The
Models Whose OECD document states that “To facilitate the consideration of a
Scientific Validity Has (Q)SAR model for regulatory purposes, it should be associated with
Been Established the following information:
1. a defined endpoint,
2. an unambiguous algorithm,
3. a defined domain of applicability,
4. appropriate measures of goodness-of-fit, robustness and
predictivity,
5. a mechanistic interpretation, if possible.”
Principle 1 ensures that there is clarity on the endpoint being
predicted and the experimental system being modeled. The rele-
vance of the endpoint will be analyzed when discussing the third
REACH requirement—adequacy of the results.
Principle 2 is set to ensure transparency and reproducibility of
the results, which are essential for regulatory acceptance.
Principle 3 is needed to be able to assess whether the chemical
under investigation falls within the applicability domain of the
model. A definition of applicability domain for (Q)SARs accepted
worldwide is lacking, therefore model developers define the appli-
cability domain differently. From the concept of applicability
derives the assessment of the reliability of a prediction, which will
be further discussed when dealing with the second REACH
requirement—the substance falls within the applicability domain.
Principle 4 gives the possibility to estimate uncertainties and
error probabilities associated with the predictions. This is very
important in the regulatory field and for risk assessment. The level
of uncertainty that can be accepted depends on the purpose of the
prediction (higher for screening, lower for definitive
considerations).
Principle 5 is optional in the OECD principles; however, expe-
rience indicates that there is less confidence in predictions without
a mechanistic rationale, which are therefore rarely accepted.
It is important to note that the validity of the model used to
obtain the prediction is only the first REACH requirement for hav-
ing an acceptable prediction.
3.3 The Substance The concept of applicability domain is linked to the assessment of
Falls Within the reliability of the prediction. There are some points that can be
the Applicability addressed to evaluate the reliability of the specific prediction and
Domain of the (Q) these points apply regardless of the definition of applicability
SAR Model domain. Results derived from interpolation have lower uncertainty
than results of extrapolation. Therefore, it is important that the
substance under investigation falls within the descriptor ranges,
whose mechanistic and metabolic domain are defined by the mol-
ecules in the training set of the model. On the same line, it can also
be checked that all the structural fragments of the target chemical
are represented in the training set of the models. Further on, pre-
dictions are more reliable when the models are trained with struc-
tures similar enough to the target, and these analogues are well
predicted. Finally, some types of substances (e.g., salts) may need
specific considerations because some models have difficulties han-
dling them.
3.4 Results Are Once the model is considered valid and the prediction within its
Adequate applicability domain, the adequacy of the prediction (i.e., its rele-
for the Purpose vance for the purpose it is being used) needs to be verified. The
of Classification adequacy of the prediction is assessed by comparing it to the regu-
and Labeling and/or latory requirements, both in term of endpoint and results. First,
Risk Assessment the type and quality of information obtained with the prediction
has to be of the same level than that of the test. A subchronic toxic-
ity study on rats would give information on the target organs, types
of effects and the doses at which these effects are observed. (Q)
SAR models available today cannot reliably provide such a wealth
of information and therefore should not be used as standalone to
replace standard information requirements for “high-tier” end-
points when it comes to definitive hazard assessment. On the same
line, a classification model predicting positive or negative outcomes
cannot replace a test providing quantitative information on the
toxic dose. Moreover, the model should be developed using train-
ing data of high quality from validated tests, comparable to the
standard requirement. Ames test should be performed on five bac-
terial strains in the absence and presence of metabolic activation,
and a model trained on data from four bacterial strains cannot
replace the standard test. If the standard requirement foresees a
test with an exposure to the substance of 28 days, the training set
data cannot refer to 14 day studies. A model predicting biodegra-
dation half-life (typically measured in simulation tests) should also
provide the structure of the degradation products to replace the
standard test.
3.5 Adequate REACH dossiers are submitted to ECHA in IUCLID (International

and Reliable Uniform ChemicaL Information Database) format. Endpoint spe-
Documentation cific results have to be reported in sections called “Endpoint Study
of the Applied Method Records” (ESRs). It is possible to attach files to the ESRs, and the
Is Provided QSAR model reporting format (QMRF) and the QSAR prediction
reporting format (QPRF) can be included there after filling the
basic fields of the ESR (e.g., “Type of information” = “(Q)SAR”).
The information has to be transparent and complete to allow an
independent evaluation by ECHA’s assessors.
4 Conclusions
REACH requires testing as a last resort and offers many alternative

ways of generating the required information. Registrants are using
these opportunities extensively. The challenge is to generate alter-
native information that is sufficient for a robust hazard assessment
and risk assessment.
From this perspective (Q)SARs have a place as a stand-alone
method for the endpoints that are less complex from a toxicologi-
cal/biological point of view. (Q)SAR methods should however not
be used as stand-alone for high tier (e.g., carcinogenicity, toxicity
to reproduction) properties. The models available today cannot
reliably provide comparable information to in vivo testing. Overall,
for any level of complexity for the predicted properties, ideally (Q)
SAR methods should be used as part of weight of evidence to sup-
port decisions. When applying a QSAR prediction (stand-alone or
in a weight of evidence approach), or when developing a QSAR
model, it is important that the considerations as reflected in
Chapter 3 are observed: (1) Results are derived from (Q)SAR
models whose scientific validity has been established; (2) The sub-
stance falls within the applicability domain of the (Q)SAR model;
(3) Results are adequate for the purpose of classification and label-
ing and/or risk assessment; and (4) Adequate and reliable docu-
mentation of the applied method is provided.
References
1. Chemicals Legislation–European Commission. 5. European Chemicals Agency (2016) Practical
https://ec.europa.eu/growth/sectors/chemi- Guide How to use and report (Q)SARs.
cals/legislation_en. Accessed 29 Nov 2017 https://echa.europa.eu/documents/10162/
2. Regulation (EC) No 1907/2006 of the 13655/pg_report_qsars_en.pdf/407dff11-
European Parliament and of the Council of 18 aa4a-4eef-a1ce-9300f8460099. Accessed 29
December 2006 concerning the Registration, Nov 2017
Evaluation, Authorisation and Restriction of 6. European Chemicals Agency (2008) Guidance
Chemicals–REACH Legislation on ECHA’s on information requirements and chemical
website. https://echa.europa.eu/regulations/ safety assessment. Chapter R.6: QSARs and
reach/legislation. Accessed 29 Nov 2017 grouping of chemicals. https://echa.europa.
3. European Chemicals Agency The use of alter- eu/documents/10162/13632/information_
natives to testing on animals for the REACH requirements_r6_en.pdf. Accessed 29 Nov
Regulation. https://echa.europa.eu/docu- 2017
ments/10162/13639/alternatives_test_ani- 7. Dimitrov SD, Georgieva DG, Pavlov TS et al
mals_2017_en.pdf. Accessed 29 Nov 2017 (2015) UVCB substances: methodology for
4. European Chemicals Agency (2017) Non- structural description and application to fate
animal approaches - regulatory applicability of and hazard assessment. Environ Toxicol Chem
non-animal approaches under the REACH, 34:2450–2462. https://doi.org/10.1002/
CLP and Biocidal Products Regulations. etc.3100
https://echa.europa.eu/documents/ 8. OECD Principles for the Validation, for
10162/22931011/non_animal_approcches_ Regulatory Purposes, of (Quantitative)
en.pdf/87ebb68f-2038-f597-fc33- Structure-Activity Relationship Models. https://
f4003e9e7d7d. Accessed 2 Dec 2017 www.oecd.org/chemicalsafety/risk-assessment/
37849783.pdf. Accessed 29 Nov 2017
Part II
Molecular and Data Modeling

Chapter 5
Machine Learning Methods in Computational Toxicology

Igor I. Baskin
Abstract
Various methods of machine learning, supervised and unsupervised, linear and nonlinear, classification and
regression, in combination with various types of molecular descriptors, both “handcrafted” and “data-
driven,” are considered in the context of their use in computational toxicology. The use of multiple linear
regression, variants of naïve Bayes classifier, k-nearest neighbors, support vector machine, decision trees,
ensemble learning, random forest, several types of neural networks, and deep learning is the focus of atten-
tion of this review. The role of fragment descriptors, graph mining, and graph kernels is highlighted. The
application of unsupervised methods, such as Kohonen’s self-organizing maps and related approaches,
which allow for combining predictions with data analysis and visualization, is also considered. The neces-
sity of applying a wide range of machine learning methods in computational toxicology is underlined.
Key words Computational toxicology, Machine learning, Support vector machines, Random forest,
Neural networks, Deep learning
1 Introduction
In view of the wide variety of different types of toxicity, endpoints,

and mechanisms of action, computational toxicology intensively
uses approaches of different types, including rule-based expert sys-
tems, molecular docking, pharmacophore (toxicophore) model-
ing, quantum chemistry studies, and building structure–activity
models using machine learning methods. In recent years, due to
the accumulation of a large amount of data on different toxicity
endpoints in databases, the rapid grows of computer power, the
development of sophisticated data analysis algorithms, the role of
machine learning methods in computational toxicology has
become crucial. This is reflected in several reviews [1–11] concern-
ing computational toxicology. The use of machine learning in
building structure–activity models, including quantitative struc-
ture–activity relationships (QSAR) models, is comprehensively
analyzed in several reviews [12, 13].
In this chapter, the application of various methods of machine
learning in computational toxicology is surveyed. Unlike other
119
120 Igor I. Baskin
reviews on this topic, the material presented is organized in accor-

dance with the logic of machine learning. First, different types of
molecular descriptors widely used for predicting different end-
points in toxicology with the help of structure–activity models are
considered. Then, the application of different types of machine
learning algorithms proved to be useful in computational toxicol-
ogy is discussed. Classical linear models are followed by sophisti-
cated nonlinear approaches and, finally, by neural networks with
deep architecture. Trends in the use of machine learning for pre-
dicting toxicity are summarized at the end of the chapter.
2 Methods
2.1 Molecular The first stage in the construction of any structure–property model
Descriptors is the selection and calculation of molecular descriptors, i.e., numer-
ical characteristics used to represent chemical structures [14]. The
most commonly used descriptors in computational toxicology are
various modifications of fragment descriptors [15–17], which indi-
cate the presence of certain fragments (substructures) in molecular
structures. The use of fragment descriptors allows for a natural way
for chemists to describe structure–property/activity relationships in
terms of the presence or absence of certain structural fragments.
This, in turn, makes it possible to interpret the structure–property/
activity models visually by tinting atoms with colors indicating their
contribution to activity [18]. Of the other types of descriptors often
used in computational toxicology, one should mention physical-
chemical and quantum-chemical. So the physicochemical descriptor
logP evaluates the lipophilicity of the molecule, which is very impor-
tant for assessing the fate of xenobiotics in different tissues, whereas
the quantum-chemical descriptor ELUMO, the energy of the lowest
unoccupied molecular orbital, assesses the ease of metabolic activa-
tion of xenobiotics through reduction reactions with antioxidants.
Electronic descriptors representing local atomic properties (partial
charges, residual electronegativity, effective polarizability, etc.),
connectivity descriptors (topological indices), and shape descriptors
(molecular volume, surface areas, etc.) are also actively used for
building QSAR models for toxicity endpoints.
2.2 Machine Most of machine learning methods belong to a very wide category
Learning Methods of supervised methods, which can build models by analyzing the
training sets containing the structures of chemical compounds rep-
resented by descriptors and information about their physicochemi-
cal properties or biological activities [12]. Such models allow
predicting the properties/activities for any new molecule given a
set of descriptors computed for it. Classes of activity (toxicity) are
predicted with classification models, whereas regression models are
used for predicting numeric property/activity values.
Machine Learning Methods in Computational Toxicology 121
2.2.1 Supervised Linear Multiple linear regression. The simplest and still very popular super-
Methods vised linear method is multiple linear regression (MLR) [19],
which produces linear models of the following form:
Activity = w0 + w1 ´ x 1 + w2 ´ x 2 + ¼, (1)
where Activity is the activity value, xi is the value of the i-th descrip-
tor, wi is the corresponding regression coefficient, w0 is the con-
stant (offset) term. Regression coefficients in such models can be
interpreted as contributions of the corresponding descriptors to
the value of activity. A lot of linear models of this type were con-
structed for various endpoints in toxicology. A typical example is
the following model reported in paper [20] for predicting loga-
rithm of his+ revertant number, ln(Nhis+), in the Ames Salmonella
histidine revertant assay for substituted polycyclic compounds con-
taining biphenyl substructure:
ln (N his + ) = -0.52 - 1.77 ´ E LUMO - 0.22 ´ log P + 0.90N para　 nitro , (2)

where ELUMO is the energy of the lowest unoccupied molecular
orbital (quantum-chemical descriptor), logP is the lipophilicity
(physicochemical descriptor), Npara-nitro is the number of para-nitro
groups in biphenyls.
Multiple linear regression equations of type (1) are used for
predicting continuous toxicity metrics, such as LD50, LC50, and
EC50. On the other hand, for dichotomous toxicity metrics (active
or inactive, 1 or 0), e.g., for mutagenicity, carcinogenicity, embryo-
toxicity, teratogenicity, and biodegradability, either eq. (3) or (4) is
used. They can be obtained using machine learning methods for
classification, such as linear discriminant analysis.
PActive = s (w0 + w1 ´ x 1 + w2 ´ x 2 + ¼) ,
(3)
s ( y ) = 1 / (1 + exp ( - y ) )

Activity = 1 if (w1 ´ x 1 + w2 ´ x 2 + ¼) >= threshold

(4)
Activity = 0 if (w1 ´ x 1 + w2 ´ x 2 + ¼) < threshold

where PActive is the probability that the molecule is active.
Numerous linear regression and classification models are col-
lected in the TopKat software [21]. The models are based on the
use of electronic, connectivity, shape, and fragmental (substruc-
ture) descriptors. The latter type of descriptors are based on a
predefined library of “3000 molecular fragments representing

chemically and/or biological important functional groups, hetero-
cyclic, aliphatic, and aromatic fused and unfused ring systems,
electron-donating and -withdrawing groups and their environ-
ments, etc” [21]. The presence of such a large number of descrip-
tors dictates the need to select the most important descriptors.
122 Igor I. Baskin
This is accomplished in the framework of the TopKat software by

discarding rare and correlated fragmental descriptors, followed by
application of a stepwise discriminant analysis procedure in order
to select descriptors with the greatest ability for distinguishing
between active and inactive compounds.
The obvious drawback of using a predefined library of frag-
ment descriptors, as in the case of the TopKat software, is the sub-
jectivity of the process of its formation. An alternative approach is
to systematically generate fragment descriptors of a certain type in
accordance with some constraints. According to this, Klopman
developed a program that extracts from a training set of com-
pounds all linear fragments (chains of atoms) subject to constraints
on minimum and maximum chain length [22]. The use of such
fragments underlies the functioning of several programs for toxic-
ity (mostly genotoxicity and carcinogenicity) prediction: CASE
[23], MULTICASE [24, 25], META-CASETOX [26], MCASE
[27–29], MC4PC [30]. Two kinds of fragments are identified:
those that are responsible for biological activity (the so-called “bio-
phores” or “toxophores”) and those that do not cause activity, but
increase or decrease its level. As soon as a biophore (toxophore)
has been identified, QSAR models can be built by analyzing struc-
tures containing it. A large number of models constructed in this
way cover many types of toxicity endpoints: acute mammalian tox-
icity, hepatotoxicity, renal toxicity, cardiac toxicity, carcinogenicity,
developmental toxicity, skin and eye irritation, etc. They are avail-
able from Multicase Inc. Besides the linear fragments, other types
of substructures are also used in conjunction with machine learn-
ing methods in computational toxicology. Among them, the most
popular are “atom pairs” [31] and “atom-centered fragments”
[32], “circular fingerprints” [33], and “multilevel neighborhoods
of atoms” [34].
Naïve Bayes. Another supervised linear machine learning method,
which is currently one of the most important in computational
toxicology, is the Naïve Bayes classifier—a classical machine learn-
ing method, which is discussed in numerous textbooks on machine
learning and data mining. It is based on the well-known Bayes’
theorem and assumes the conditional independence among molec-
ular descriptors. For each test compound with molecular descrip-
tors x1, x2, …, it computes a discriminant function d(x1, x2, …) by
taking a linear combination of the molecular descriptors:

d ( x 1 ,,x 2 ,, ¼) = log P (1 | x 1 ,,x 2 ,, ¼) - log P ( 0 | x 1 ,,x 2 ,, ¼) = w0 + w1 ´ x 1 + w2 ´ x 2 + ¼, (5)

where P(1|x1, x2, …) is the posterior probability that the test com-
pound is active, P(0|x1, x2, …) is the posterior probability that the
test compound is inactive, wi is a Bayesian score for i-th descriptor.
So the test compound is predicted to be active if the value of the
discriminant function for it is positive, and it is predicted to be

inactive if the value of the discriminant function for it is negative.
The absolute value of the Bayesian score indicates the importance
of the corresponding descriptor, so descriptors with near-zero val-
ues can be neglected. The Naïve Bayes classifier is typically used in
conjunction with fragment descriptors indicating the presence of a
certain fragment in chemical structure. In this case, the positive
value of wi indicates that the presence of the corresponding sub-
structure is favorable for activity, whereas its negative value indi-
cates the undesirability of its presence in active compounds. In the
toxicology domain, fragments with high positive value of the
Bayesian score can be considered as “alerts.” This allows interpret-
ing structure–activity models in structural terms understandable
for chemists. Such information is very important for drug design.
The value wi can be computed by counting active and inactive
compounds with the corresponding fragment in the training set.
Consider several typical examples of the use of the Naïve Bayes
classifier in computational toxicology. Metz et al. applied this
method in combination with extended connectivity fingerprints
(ECFP) [35], which are currently a widely used type of fragment
descriptors, for predicting reactivity towards protein thiol groups
[36]. This property is known to be closely related to several toxic-
ity endpoints. Langdon et al. used this method in combination
with Scitegic FCFP_6 fingerprints (another type of fragment
descriptors) implemented in the Pipeline Pilot software (Accelrys,
Inc., San Diego, CA) for cytotoxicity prediction [37]. Xia et al.
applied the same methodology for performing classification of
kinase inhibitors [38]. Liew et al. applied the Naïve Bayes classifier
to predict hepatotoxicity [39]. Several cases of using this method
in computational toxicology are reviewed in paper [11].
PASS. A special modification of the Naïve Bayes classifier was
implemented in the PASS computer program [40], which utilizes
a special kind of atom-centered fragment descriptors—MNA
(Multilevel Neighborhood of Atoms) [34]. For each activity, it
predicts the “probability” of a test compound to be active, as well
as its “probability” to be inactive. PASS can predict the biological
activity spectra of organic compounds for a large set of different
types of biological activity, including various toxicity endpoints. A
special study deals with the use of PASS for predicting rodent car-
cinogenicity [41].
The abovementioned PASS program formed the basis for the
PASS-BioTransfo program for predicting the feasibility of different
classes of biotransformation reactions for a given organic com-
pound [42]. The classification models used for this purpose were
built by utilizing the information on metabolic transformations
contained in the Metabolite (MDL) and Metabolism (Accelrys)
databases. This methodology was further extended to the predic-
124 Igor I. Baskin
tion of feasible sites of metabolic transformation, for which special

modifications of the MNA descriptors, RMNA (Reacting Multilevel
Neighborhood of Atom) [43] and LMNA (Labeled Multilevel
Neighborhood of Atom) [44–46], were developed. On this basis,
a MetaTox Web application was created to predict the structures
and toxicity of xenobiotic metabolites [47].
Graph mining. All the above-discussed approaches involve genera-
tion of some particular families of fragment descriptors, e.g., based
on sequences of atoms and bonds along paths in molecular graphs
or atoms with their closest neighbors, followed by application of a
descriptor selection procedure. The main drawback of this approach
stems from the fact that it is impossible to generate all possible
types of fragment descriptors because of the enormous number of
possible substructures, and therefore some important fragment
descriptors not belonging to the chosen families will inevitably be
lost. A promising approach to solving this problem is to use the
so-called graph mining methods, which can extract from molecular
graphs important for a given task fragment descriptors based on
substructures of any complexity [48]. Such “mined” fragment
descriptors are usually combined with some supervised linear
machine learning method. As an example, Saigo et al. embedded a
graph mining algorithm into the linear programming boosting
method to build a QSAR model for the activity of endocrine dis-
ruptors [49]. The resulting “mined” fragment descriptors are
based not only on simple chains but also on branched and cyclic
substructures, which could not have been generated using popular
software for computing fragment descriptors [15].
2.2.2 Supervised k-nearest neighbors. Perhaps the simplest supervised nonlinear

Nonlinear Machine machine learning method is k-nearest neighbors (kNN) classifier,
Learning Methods which classifies a test compound by looking for the training com-
pounds with the shortest distance to it. The majority class of its
k-nearest neighbors determines the predicted class for the test
compound. The Euclidean distance is a common choice for real-
valued descriptors for detecting the nearest neighbors, whereas the
Tanimoto similarity index is usually used for this purpose with
binary descriptors. Since all functions used for distance computa-
tion are nonlinear with respect to descriptors, this method can be
considered nonlinear. One of the drawbacks of this method is that
all descriptors are taken with the same weight for calculating
distances, and so they equally affect the prognosis. In many cases,
however, such a simplification is not acceptable. This is especially
true for toxicology, where only a small number of fragments in
molecules are often responsible for the observed effect. That is
why in the field of computational toxicology the kNN classifier is
usually used in combination with a descriptor selection procedure.
A common choice is to apply stochastic optimization algorithms to
select important descriptors. For example, the simulated annealing

algorithm was used for this purpose for building QSAR model for
estrogen receptor ligands [50]. For modeling liver-related adverse
effects of drugs, a stochastic sampling procedure was used to select
important descriptors [51]. The number of nearest neighbors, k,
the distance measures (metric), and the data normalization method
can also be optimized to find best kNN models [39].
Support vector machine. Support vector machine (SVM) is a
machine learning method based on Vapnik’s statistical learning
theory [52–54]. Using a kernel function, SVM maps descriptor
vectors into a higher dimensional feature space, in which it tries to
build a maximal margin hyperplane to separate active from inactive
compounds. Although an SVM model is linear in the feature space,
it can be highly nonlinear in the original descriptor space, provided
the kernel function is nonlinear. The most popular type of nonlin-
ear kernel function is the Gaussian radial basis function, which is
capable of modeling nonlinear dependencies of any complexity. It
consistently provides models with better performance than other
types of kernels [55]. The cosine between the perpendicular to the
separating hyperplane in the feature space and an axis in this space
indicates the relative importance of the corresponding feature for
classification. This implicitly weights descriptors and their combi-
nations by their relative importance. Because of this, there is no
need to preselect descriptors for building SVM models. Besides,
SVM tolerates strong correlations among descriptors. All this
determines the important role played by the SVM modeling in
structure–property modeling.
Consider several typical examples of using SVM models in
computational toxicology. Khandelwal et al. applied SVM model-
ing to predicting human pregnane X receptor (PXR) activation and
compared the results with several other machine learning methods
and results of docking [56]. Fourches et al. built SVM models in
combination with ISIDA fragment descriptors [17] for predicting
drug-induced liver injury (DILI) in different species [57]. The
same endpoint was also analyzed using the SVM method by Liew
et al. [39]. Ekins compared the SVM and the Naïve Bayes methods
for hepatotoxicity (e.g., for DILI), cardiotoxicity, renal toxicity
and genotoxicity [11]. Both methods were shown to perform simi-
larly on cross-validation data. In a benchmarking study aimed at
predicting mutagenicity in Ames test, the SVM method in
combination with NASAWIN fragment descriptors [58–60]

showed one of the best results [61].
Graph kernels. Graph kernels are a special kind of kernel functions
that are able to map molecular graphs representing chemical struc-
tures directly into a high-dimensional feature space without the
need to precompute fixed sets of molecular descriptors [62, 63].
Their main advantage over traditional kernels built on descriptor
126 Igor I. Baskin
vectors, such as the aforementioned Gaussian kernel, is the ability

to work implicitly with a huge number of fragments without the
need to generate them explicitly. Thus, they provide an effective
mechanism to enumerate all potentially useful fragments and, due
to the internal mechanism of building SVM models, to weight
them implicitly according to their importance. Thanks to this, they
can reveal very complex substructures or hidden patterns of local
properties, which could not have been found by computing tradi-
tional families of molecular descriptors. Therefore, graph kernels
could be of great interest for computational toxicology.
Several studies deal with the use of graph kernels for predicting
toxicity of organic compounds. Kashima et al. pioneered in this
direction by using marginalized kernels to predict several genotox-
icity endpoints [64]. In the follow-up publications, Menchetti
et al. used weighted decomposition kernel [65], Swamidass et al.
used several classes of kernels for comparing small molecules [66],
Mahé et al. used graph kernels based on random walk on chemical
structures [67], whereas Ralaivola et al. applied several kinds of
graph kernels [62] for this purpose.
Decision trees. Decision trees is a very popular group of nonlinear
machine learning methods, in which each model is a set of rules
organized in the form of a tree [68]. Each tree contains a single
root node, any number of internal nodes and several leaf nodes,
while directed branches connect the root and the internal nodes
with other internal nodes and leaves. The root and each internal
node represent a test on some descriptor (e.g., whether its value
exceeds some threshold), the outgoing branches represent alterna-
tive outcomes of the test, whereas each leaf represents a class label
(e.g., a type of biological activity for a compound being tested).
The path from the root to a leaf represents a sequence of classifica-
tion rules predicting class label (activity/toxicity) for any com-
pound “falling” into the leaf. This provides unique opportunity to
easily interpret the models and the results of their application in
terms of simple rules easily understandable by human researchers,
which makes this approach very appealing and popular in compu-
tational toxicology. Due to the simplicity of interpretation of such
models, toxicologists can compare them with the background
knowledge concerning the mechanisms of action as well as with
rules in expert systems, which is extremely important for decision-
making. In combination with fragment descriptors, decision trees
allow to build the models that are easily interpretable in terms of
structural alerts, which is important particularly for dealing with
mutagenicity and carcinogenicity. In this case, decision-trees mod-
els can be used as automatic extractors of rules and structural alerts
from raw toxicity data.
Decision trees (DT) can be used for building both classification
models for making qualitative predictions (e.g., whether a given
compound is toxic) and regression models for making quantitative

predictions (e.g., the value of LD50 for a given compound). DT have
been used in numerous applications in the field of computational
toxicology. Consider several examples. Dixon with coauthors used
DT with structural descriptors to develop models for predicting
dose-dependent drug-induced hepatotoxicity [69] and human cyto-
chrome P450 inhibition [70]. Feng et al. made a benchmarking
study to compare the performance of several machine leaning meth-
ods, including DT, in combination with several types of molecular
descriptors, for predicting several toxicity endpoints [71].
Toxtree (https://eurl-ecvam.jrc.ec.europa.eu/laboratories-
research/predictive_toxicology/qsar_tools/toxtree), which has
been developed by Ideaconsult Ltd (Sofia, Bulgaria) under the
terms of a contract with the Joint Research Center at European
Union Reference Laboratory for Alternatives to Animal Testing
(EURL ECVAM), is an open source application, which is able to
estimate toxic hazard by applying the DT approach. Currently, it
includes Cramer rules [72], which was the first DT application in
toxicology, Verhaar scheme for predicting toxicity mode of actions
[73], several DT models for estimating skin irritation and corro-
sion potential [74, 75], and a DT model built by Benigni and
Bossa for estimating carcinogenicity and mutagenicity [76].
DT are usually built using a special procedure called recursive
partitioning, which selects partitioning descriptors and thresholds
for their values in a greedy manner to optimize a given measure of
class purity of leaves. Although this procedure is computationally
very efficient, it does not always lead to the construction of optimal
trees. An alternative approach to building DT is to use genetic
algorithm, which mimics evolution of species according to Darwin’s
theory [77]. Although by sacrificing computational efficiency, this
approach can lead to the construction of more optimal DT charac-
terized by a higher predictive ability. The advantage of this approach
in the application to building DT for predicting hepatotoxicity was
successfully demonstrated by DeLisle and Dixon [78].
Ensemble learning and random forest. An even more efficient
approach to improving the predictive ability of classification and
regression models based on DT is to use them as base models in
the framework of ensemble learning, although at the expense of a
partial loss of interpretability [79–84]. Ensemble learning consists
in combining several base models into a more predictive one. The
most popular types of ensemble modeling are bagging [85], ran-
dom spaces [86], boosting [80, 87], and stacking [88]. In bagging
all base models are built on randomly resampled training sets [85],
while in random spaces the base models are built with random
subsets of descriptors [86]. Combining the bagging with the ran-
dom spaces approaches in application to decision trees base models
leads to the Random Forest (RF) method [89, 90], which is cur-
128 Igor I. Baskin
rently considered as one of the best approaches to building classi-

fication and regression models. The main advantages of RF is very
high predictive performance of models, very high computational
efficiency of model building models, and the ease of use, because
there is no need to configure or optimize model settings.
Currently, the RF method is actively used to build classification
and regression structure–property/activity models almost in all
fields of chemoinformatics [90], and especially in the domain of
computational toxicology [90–94]. Consider several examples.
Svetnik et al. were among the first to apply RF to build classifica-
tion and regression SAR/QSAR models for several types of bio-
logical activity, including a classification model for estrogen
receptor binding [90]. Li et al. applied RF to predict Local Lymph
Node Assay based skin sensitization activity [91]. Zhang and Aires-
de-Sousa used RF for predicting mutagenicity from empirical
physicochemical descriptors [92]. Polishchuk et al. applied RF
with special simplex descriptors to build a QSAR model for pre-
dicting aquatic toxicity [93]. Vasanthanathan et al. compared the
performance of several machine learning methods for performing
classification of cytochrome P450 1A2 inhibitors and noninhibi-
tors and found RF to be the best one [94]. In another benchmark-
ing study, several machine learning methods in combination with
various types of descriptors were used to build classification model
for mutagenicity measured in Ames test, and again RF appeared to
provide the most predictive models [61].
2.2.3 Artificial Neural Artificial neural networks (ANNs) are a broad category of machine
Networks and Deep learning methods, based on a simplified simulation of the opera-
Learning tion of human brain cells called neurons [95, 96]. There are three
types of neurons: (1) input neurons receiving input signals from
outside, (2) output neurons that form output signals, and (3) hid-
den neurons serving for intermediate computations. In structure–
property/activity modeling, input neurons correspond to
molecular descriptors, while output neurons correspond to the
predicted properties/activities. After training by adjusting the
weights of interneural connections, the ANN are able to predict
the values of the properties/activities of chemical compounds rep-
resented by molecular descriptors. Since the beginning of the 90s
of the last century, ANNs are actively used in all areas of structure–
property/activity modeling, see comprehensive reviews [97–99].
Backpropagation neural networks. The most widely used type
(architecture) of ANNs is multilayer feed-forward neural networks,
also known as backpropagation neural networks (BPNNs). In
BPNNs, all neurons are organized into three layers, and informa-
tion flow proceeds from the (first) layer of input neurons to the
(second) layer of hidden neurons, and from there—to the (third)
layer of output neurons. First applications of BPNNs in computa-
tional toxicology date back to 1994. So Vellemin et al. predicted

carcinogenicity of polycyclic aromatic compounds using BPNNs
with molecular descriptors derived from graph theory [100]. Xu
et al. applied BPNN for building QSAR models for predicting
polar narcosis toxicity of phenols [101]. In subsequent publica-
tions, BPNNs were used for predicting numerous other toxicity
endpoints, including toxicity of organic chemicals to luminescent
bacteria (Microtox test) [102], cytotoxicity [103], toxicity of ben-
zothiazolium salts [104], toxicity of amide herbicides [105], acute
mammalian toxicity of pesticides [106–109], acute and sublethal
toxicity endpoints for fish, invertebrate, protozoan, and bacterial
species [110], toxicity of DDT-type analogs [111], fathead min-
now acute toxicity [71, 112–114], mutagenicity in different classes
of compounds [20, 71, 115], and ecotoxicity of environmental
pollutants [116]. Dearden and Rowe have recently surveyed the
use of ANNs (mostly BPNNs) in the prediction of toxicities for
REACH legislation [117].
Bayesian-regularized neural networks. It is known that standard
BPNNs can suffer from overtraining, overfitting, inherent instabil-
ity, and other factors [99, 118]. One of the ways to solve these
problems is to apply regularization, where the model complexity is
balanced against the accuracy of reproducing training data by
searching for an optimal value of a special parameter called regular-
ization coefficient [119]. Such search can be performed with the
help of Bayesian methods, and in this case the resulting neural net-
work is called Bayesian-regularized neural network (BRNN) [120–
122]. Application of Bayesian methods with sparse priors allows
for performing automatic descriptor selection. As a result, the
aforementioned problems of standard BPNNs appear to be mostly
solved, and QSAR models become robust, interpretable, and char-
acterized by enhanced predictive performance. This makes BRNNs
a valuable tool for building structure–property/activity models.
BRNNs have already been used for predicting the acute toxicity of
substituted benzenes to Tetrahymena pyriformis [123, 124],
human intestinal absorption [125], and for modeling toxicity of
nanoparticles [126].
Associative neural networks. An alternative approach to solve the
abovementioned problems of BPNNs is to apply ensemble learn-
ing (see above). Tetko developed a special type of ANNs, called
associative neural network (ASNN) [127], which is based on the
principles of ensemble learning. For building an ASNN model, a
set of BPNNs are trained independently on subsamples of the orig-
inal training set, and the predictions are averaged. Besides, the pre-
diction can be corrected using information on the properties of
compounds stored in memory. The predictive ability of ASNN
models considerably exceeds that of individual BPNN models.
ASNN models have provided top-ranked models in the challenges
130 Igor I. Baskin
organized by US Environment Protection Agency ToxCast [128]

and National Institutes of Health Tox21 programs [129]. ASNN
models can be built and applied for making predictions via the
online chemical modeling environment OCHEM [130].
Deep learning. Deep learning (DL) is a recently emerged field of
machine learning based on the use of ANNs with multiple hidden
layers, which form multiple levels of representations corresponding
to different levels of abstraction [131, 132]. DL has recently revo-
lutionized image processing, computer vision, speech recognition,
natural language processing and other domains usually associated
with the notion of artificial intelligence. The tremendous success
of DL is due not only to the use of novel algorithms and architec-
tures of ANNs, but also to the availability of very fast computers
and large datasets. Over the past few years, DL has also become a
powerful tool in structure–property/activity studies and drug dis-
covery [99, 133–135].
In each layer, DL combines more simple features taken from
the previous layer into more complex and abstract features, so very
high-level features can automatically be formed in the top layers.
For example, in image processing, low-level raw pixel data are
combined to edges, which are combined in the next levels to more
complex parts of objects, such as eyes or noses, which are finally
combined in the last layers to faces. In chemistry, low-level atom-
based descriptors can be combined to simple functional groups,
which, in turn, can be combined to complex structural features,
pharmacophores, or toxicophores. So a hierarchy of chemical fea-
tures is automatically constructed by DL. This makes DL well
suited to computational toxicology.
In order to apply DL to toxicity prediction, Mayr et al. devel-
oped the DeepTox pipeline, which in the Tox21 Data Challenge
demonstrated clear advantage of DL over other machine learning
methods, like Naïve Bayes, SVM, and RF [136]. By carefully ana-
lyzing the activation of hidden neurons in different layers, it has
been clearly shown in that paper that in lower (i.e., closer to the
input layer) hidden layers the features formed by neurons code
small substructures of toxicophores, while in higher (i.e., closer to
the input layer) layers they correspond to larger substructures or
even whole toxicophores. Hence, deep ANNs can learn from data
complex toxicophores features, and this leads to high predictive
power for toxicity.
2.2.4 Unsupervised The goal of unsupervised machine learning methods is to reveal

Methods regularities in chemical data and form new features to simplify data
description. Unsupervised methods comprise dimensionality
reduction, finding data clusters, approximating probability density
in datasets, etc. The models built with their help allow for a thor-
ough analysis of chemical data, as well as visualizing chemical data
in 2D or 3D. Furthermore, unsupervised modeling can be consid-

ered as representation learning [137], so new features revealed by
analyzing chemical data can be used as “data-driven” descriptors
that can lead to stronger correlations with physicochemical proper-
ties and biological activity of chemical compounds.
Unsupervised machine learning methods, like the supervised
ones, can be either linear (in which the revealed features are linear
combinations of original descriptors) or nonlinear. A prototype of
unsupervised linear methods is the principal component analysis
(PCA), which is a classical multivariate statistical method widely
used in numerous studies for data analysis. Currently, in view of
severe limitations of PCA and other linear approaches, nonlinear
unsupervised machine learning methods are becoming more and
more popular for chemical data analysis and visualization.
In chemoinformatics, the most popular nonlinear dimension-
ality reduction method is Kohonen’s self-organizing maps (SOM),
which can map molecules from original descriptor space onto a 2D
grid of neurons [138]. The training algorithm of SOM guarantees
that similar molecules are mapped to the same or closely located
neurons in the grid, so projection of molecules produces maps
with preserved neighborhood relations. Provided similar molecules
have similar properties/activities, the molecules belonging to the
same activity class are mapped either to the same neuron or to
neighboring neurons. Since neurons in SOM can be colored
according to the property values or activity class of molecules
mapped to them, the grid of colored neurons can be used for pre-
dicting properties/activities of new molecules projected onto it, as
well as for producing colored maps for visualizing datasets of mol-
ecules in relation with their properties/activities. Such maps enable
the use of SOM for drug design [139]. As for toxicity prediction,
SOM were used in several studies as providers of “data-driven”
descriptors for supervised machine learning methods. For example,
the MOLMAP descriptors produced by SOM from bond proper-
ties were successfully used in combination with RF by Zhang and
Aires-de-Sousa to predict mutagenicity in Ames test [92]. In this
case, the use of intermediate unsupervised SOM model makes pre-
dicted values independent on the size of molecules and the num-
bering of atoms and bonds in them.
Another way to make predictions with SOM is to augment the
2D grid of neurons with an additional layer of neurons that can
learn in the supervised manner. Such combined architecture is
used in counterpropagation neural networks (CPNN) [140]. Their
main advantage over the use of pure supervised methods is the
ability to provide data visualization, which may be important for
the interpretation of models. For example, using CPNN, Vracko
built a model for predicting carcinogenic potency of benzene
derivatives [141], Mazzatorta et al. modeled acute toxicity for the
132 Igor I. Baskin
fathead minnow [142], while Spycher et al. built a model to dis-

criminate between modes of toxic action of phenols [143].
Generative topographic mapping (GTM) is a probabilistic ana-
log of SOM which also projects molecules onto a 2D grid [144].
GTM has recently been shown to be a universal tool for various
applications in chemoinformatics, because in addition to visualiz-
ing chemical data, it can also be used as an efficient classification
and regression modeling tool [145–150]. GTM is able to provide
a much richer visualization capabilities and lead to models with
significantly higher predictive power in comparison with SOM-
based methods. In the field of computational toxicity, Kireeva et al.
has successfully applied this approach to identify HERG channel
blockers [151].
3 Conclusions and Outlook
The task of predicting the toxic effects of chemical compounds on

the human and animal organism is of paramount importance. Its
successful solution will allow to avoid very slow, expensive, and
often morally unacceptable animal testing, not waste money on the
development of the drugs that may prove toxic, and protect people
from the adverse effects of toxic pollutants. Computational toxi-
cology aims to solve this problem, and the methods of machine
learning play a very important role in this. Starting from the classi-
cal linear regression, which allows for working only with small
series of similar compounds, these methods have been largely
improved, and currently they can be used to build complex struc-
ture–property/activity models, linear and nonlinear, regression
and classification, supervised and unsupervised, on large sets of
structurally diverse compounds and for a large number of toxic
endpoints. In addition to the prediction task, modern methods of
machine learning allow for profound analysis, visualization, and
interpretation of chemical data and structure–property/activity
relationships.
With the development of new and improvement of existing
methods of machine learning, the notion of the state-of-the-art
approach is constantly changing. Multiple linear regression (MLR),
k-nearest neighbors (kNN), naïve Bayes (NB), backpropagation
neural networks (BPNN), support vector machines (SVM), deci-
sion trees (DT), and random forest (RF)—all of them were once
considered the best methods for structure–property/activity mod-
eling. The key event in the past few years is the rapid growth in
popularity of the deep learning (DL), which has already revolu-
tionized the whole domain of artificial intelligence and is rapidly
being introduced into various application areas, including drug
discovery and computational toxicology. Impressive results shown
by DL in the Kaggle competition on drug discovery and the Tox21
Data Challenge competition on toxicity prediction indicate the

high prospects for further development and implementation of the
DL technology. Availability of large datasets due to high-
throughput screening (“big data”), fast hardware accelerated by
graphical processors (GPU), very high-quality and easy-to-use
libraries of program code, and large financial investments are the
factors that determine the success of DL in these areas. Nevertheless,
this does not mean that all other methods of machine learning can
be forgotten. The advantages of DL compared to other machine
learning methods in building structure–property/activity models
are not large, and in many cases, especially with small amount of
data, the use of alternative approaches may be more preferable.
The success of further development of computational toxicology
can only be ensured by profound understanding and correct appli-
cation of a wide range of machine learning methods.
References
1. Barratt MD, Rodford RA (2001) The com- 10. Knudsen T, Martin M, Chandler K,
putational prediction of toxicity. Curr Opin Kleinstreuer N, Judson R, Sipes N (2013)
Chem Biol 5:383–388 Predictive models and computational toxicol-
2. Kavlock RJ, Ankley G, Blancato J, Breen M, ogy. In: Barrow PC (ed) Teratogenicity test-
Conolly R, Dix D, Houck K, Hubal E, Judson ing: methods and protocols. Humana Press,
R, Rabinowitz J, Richard A, Setzer RW, Shah Totowa, NJ, pp 343–374. https://doi.
I, Villeneuve D, Weber E (2008) org/10.1007/978-1-62703-131-8_26
Computational toxicology—a state of the sci- 11. Ekins S (2014) Progress in computational
ence mini review. Toxicol Sci 103:14–27 toxicology. J Pharmacol Toxicol Methods
3. Muster W, Breidenbach A, Fischer H, 69:115–140
Kirchner S, Müller L, Pähler A (2008) 12. Varnek A, Baskin I (2012) Machine learning
Computational toxicology in drug develop- methods for property prediction in chemoin-
ment. Drug Discov Today 13:303–310 formatics: quo vadis? J Chem Inf Mod
4. Valerio LG (2009) In silico toxicology for the 52:1413–1437
pharmaceutical sciences. Toxicol Appl 13. Cherkasov A, Muratov EN, Fourches D,
Pharmacol 241:356–370 Varnek A, Baskin II, Cronin M, Dearden J,
5. Nigsch F, Macaluso NJM, Mitchell JBO, Gramatica P, Martin YC, Todeschini R,
Zmuidinavicius D (2009) Computational Consonni V, Kuz'min VE, Cramer R, Benigni
toxicology: an overview of the sources of data R, Yang C, Rathman J, Terfloth L, Gasteiger
and of modelling methods. Expert Opin J, Richard A, Tropsha A (2015) QSAR mod-
Drug Metab Toxicol 5:1–14 eling: where have you been? Where are you
6. Merlot C (2010) Computational toxicol- going to? J Med Chem 57:4977–5010
ogy—a tool for early safety evaluation. Drug 14. Todeschini R, Consonni V (2009) Molecular
Discov Today 15:16–22 descriptors for chemoinformatics. In:
7. Raunio H (2011) In silico toxicology – non- Methods and principles in medicinal chemis-
testing methods. Front Pharmacol 2:33 try, vol 41. Wiley-VCH, Weinheim
8. Sun HM, Xia MH, Austin CP, Huang RL 15. Baskin I, Varnek A (2008) Fragment descrip-
(2012) Paradigm shift in toxicity testing and tors in SAR/QSAR/QSPR studies, molecular
modeling. AAPS J 14:473–480 similarity analysis and in virtual screening. In:
Varnek A, Tropsha A (eds) Chemoinformatics
9. Reisfeld B, Mayeno AN (2012) What is com- approaches to virtual screening. RSC
putational toxicology? In: Reisfeld B, Mayeno Publisher, Cambridge, pp 1–43
AN (eds) Computational toxicology, vol
Volume I. Humana Press, Totowa, NJ, 16. Baskin I, Varnek A (2008) Building a chemi-
pp 3–7 cal space based on fragment descriptors.
134 Igor I. Baskin
Comb Chem High Throughput Screen genicity using the MCASE QSAR expert sys-
11:661–668 tem. SAR QSAR Environ Res 14:165–180
17. Varnek A, Fourches D, Hoonakker F, Solov’ev 29. Klopman G, Chakravarti SK, Zhu H, Ivanov
V (2005) Substructural fragments: an univer- JM, Saiakhov RD (2004) ESP: a method to
sal language to encode reactions, molecular predict toxicity and pharmacological proper-
and supramolecular structures. J Comput ties of chemicals using multiple MCASE data-
Aided Mol Des 19:693–703 bases. J Chem Inf Comput Sci 44:704–715
18. Marcou G, Horvath D, Solov'ev V, Arrault A, 30. Klopman G, Ivanov J, Saiakhov R, Chakravarti
Vayer P, Varnek A (2012) Interpretability of S (2005) MC4PC–an artificial intelligence
SAR/QSAR models of any complexity by approach to the discovery of structure toxic
atomic contributions. Mol Inform activity relationships (STAR). In: Helma C
31:639–642 (ed) Predictive toxicology. CRC Press, Boca
19. Draper NR, Smith H (1998) Applied regres- Raton, pp 423–457
sion analysis, 3rd edn. John Wiley, 31. Carhart RE, Smith DH, Venkataraghavan R
New York (1985) Atom pairs as molecular features in
20. Lyubimova IK, Abilev SK, Gal'berstam NM, structure-activity studies: definition and appli-
Baskin II, Palyulin VA, Zefirov NS (2001) cations. J Chem Inf Comput Sci 2:64–73
Computer-aided prediction of the mutagenic 32. Xiao Y, Qiao Y, Zhang J, Lin S, Zhang W
activity of substituted polycyclic compounds. (1997) A method for substructure search by
Biol Bull 28:139–145 atom-centered multilayer code. J Chem Inf
21. Enslein K, Gombar VK, Blake BW (1994) Comput Sci 37:701–704
Use of SAR in computer-assisted prediction 33. Glen RC, Bender A, Arnby CH, Carlsson L,
of carcinogenicity and mutagenicity of chemi- Boyer S, Smith J (2006) Circular fingerprints:
cals by the TOPKAT program. Mutat Res flexible molecular descriptors with applica-
305:47–61 tions from physical chemistry to
22. Klopman G (1984) Artificial intelligence ADME. IDrugs 9:199–204
approach to structure-activity studies. 34. Filimonov D, Poroikov V, Borodina Y,
Computer automated structure evaluation of Gloriozova T (1999) Chemical similarity
biological activity of organic molecules. J Am assessment through multilevel neighborhoods
Chem Soc 106:7315–7321 of atoms: definition and comparison with the
23. Rosenkranz HS, Klopman G (1988) CASE, other descriptors. J Chem Inf Comput Sci
the computer-automated structure evaluation 39:666–670
system, as an alternative to extensive animal 35. Hassan M, Brown RD, Varma-O'Brien S,
testing. Toxicol Ind Health 4:533–540 Rogers D (2006) Cheminformatics analysis
24. Klopman G (1992) MULTICASE. 1. A hier- and learning in a data pipelining environ-
archical computer automated structure evalu- ment. Mol Divers 10(3):283–299
ation program. Quant Struct-Act Relat 36. Metz JT, Huth JR, Hajduk PJ (2007)
11(2):176–184. https://doi.org/10.1002/ Enhancement of chemical rules for predicting
qsar.19920110208 compound reactivity towards protein thiol
25. Klopman G (1998) The MultiCASE program groups. J Comput Aided Mol Des
II. Baseline activity identification algorithm 21:139–144
(BAIA). J Chem Inf Comput Sci 38:78–81 37. Langdon SR, Mulgrew J, Paolini GV, van
26. Klopman G (1996) The META-CASETOX Hoorn WP (2010) Predicting cytotoxicity
system. In: Puijnenburg WJGM, Damborsky from heterogeneous data sources with
J (eds) Biodegradability prediction. Springer, Bayesian learning. J Cheminform 2:11
Berlin, pp 27–40 38. Xia X, Maliski EG, Gallant P, Rogers D
27. Matthews EJ, Contrera JF (1998) A new (2004) Classification of kinase inhibitors
highly specific method for predicting the car- using a Bayesian model. J Med Chem
cinogenic potential of pharmaceuticals in 47:4463–4470
rodents using enhanced MCASE QSAR-ES 39. Liew CY, Lim YC, Yap CW (2011) Mixed
software. Regul Toxicol Pharmacol learning algorithms and features ensemble in
28:242–264 hepatotoxicity prediction. J Comput Aided
28. Klopman G, Chakravarti SK, Harris N, Ivanov Mol Des 25:855
J, Saiakhov RD (2003) In-silico screening of 40. Poroikov VV, Filimonov DA, Borodina YV,
high production volume chemicals for muta- Lagunin AA, Kos A (2000) Robustness of
biological activity spectra predicting by com- 51. Rodgers AD, Zhu H, Fourches D, Rusyn I,
puter program PASS for noncongeneric sets Tropsha A (2010) Modeling liver-related
of chemical compounds. J Chem Inf Comput adverse effects of drugs using k nearest neigh-
Sci 4:1349–1355 bor quantitative structure−activity relation-
41. Lagunin AA, Dearden JC, Filimonov DA, ship method. Chem Res Toxicol
Poroikov VV (2005) Computer-aided rodent 23:724–732
carcinogenicity prediction. Mutat Res 52. Vapnik V (1998) Statistical learning theory.
586:138–146 Wiley-Interscience, New York
42. Borodina Y, Sadym A, Filimonov D, Blinova 53. Vapnik VN (1995) The nature of statistical
V, Dmitriev A, Poroikov V (2003) Predicting learning theory. Springer, Berlin
biotransformation potential from molecular 54. Cortes C, Vapnik V (1995) Support-vector
structure. J Chem Inf Comput Sci networks. Mach Learn 20:273–297
43:1636–1646 55. Czermiński R, Yasri A, Hartsough D (2001)
43. Borodina Y, Rudik A, Filimonov D, Use of support vector machine in pattern clas-
Kharchevnikova N, Dmitriev A, Blinova V, sification: application to QSAR studies. Mol
Poroikov V (2004) A new statistical approach Inform 20:227–240
to predicting aromatic hydroxylation sites. 56. Khandelwal A, Krasowski MD, Reschly EJ,
Comparison with model-based approaches. Sinz MW, Swaan PW, Ekins S (2008) Machine
J Chem Inf Comput Sci 44:1998–2009 learning methods and docking for predicting
44. Rudik AV, Dmitriev AV, Lagunin AA, human pregnane X receptor activation. Chem
Filimonov DA, Poroikov VV (2014) Res Toxicol 21:1457–1467
Metabolism site prediction based on xenobi- 57. Fourches D, Barnes JC, Day NC, Bradley P,
otic structural formulas and PASS prediction Reed JZ, Tropsha A (2010) Cheminformatics
algorithm. J Chem Inf Mod 54:498–507 analysis of assertions mined from literature
45. Rudik A, Dmitriev A, Lagunin A, Filimonov that describe drug-induced liver injury in dif-
D, Poroikov V (2015) SOMP: web server for ferent species. Chem Res Toxicol
in silico prediction of sites of metabolism for 23:171–183
drug-like compounds. Bioinformatics 58. Artemenko NV, Baskin II, Palyulin VA,
31:2046–2048 Zefirov NS (2001) Prediction of physical
46. Rudik AV, Dmitriev AV, Lagunin AA, properties of organic compounds using
Filimonov DA, Poroikov VV (2016) artificial neural networks within the sub-
Prediction of reacting atoms for the major structure approach. Dokl Chem
biotransformation reactions of organic xeno- 381:317–320
biotics. J Cheminf 8:68 59. Artemenko NV, Baskin II, Palyulin VA,
47. Rudik AV, Bezhentsev VM, Dmitriev AV, Zefirov NS (2003) Artificial neural network
Druzhilovskiy DS, Lagunin AA, Filimonov and fragmental approach in prediction of
DA, Poroikov VV (2017) MetaTox: web physicochemical properties of organic com-
application for predicting structure and toxic- pounds. Russ Chem Bull 52:20–29
ity of xenobiotics’ metabolites. J Chem Inf 60. Zhokhova NI, Baskin II, Palyulin VA, Zefirov
Mod 57:638–642 AN, Zefirov NS (2007) Fragmental descrip-
48. Saigo H, Tsuda K (2010) Graph mining in tors with labeled atoms and their application
chemoinformatics. In: Lodhi H, Yamanishi Y in QSAR/QSPR studies. Dokl Chem
(eds) Chemoinformatics and advanced 417:282–284
machine learning perspectives: complex com- 61. Sushko I, Novotarskyi S, Korner R, Pandey
putational methods and collaborative tech- AK, Cherkasov A, Li J, Gramatica P, Hansen
niques. IGI Global, Hershey, PA, pp 95–128 K, Schroeter T, Muller KR, Xi L, Liu H, Yao
49. Saigo H, Kadowaki T, Tsuda K (2006) A lin- X, Oberg T, Hormozdiari F, Dao P, Sahinalp
ear programming approach for molecular C, Todeschini R, Polishchuk P, Artemenko A,
QSAR analysis. Paper presented at the Kuz'min V, Martin TM, Young DM, Fourches
International Workshop on Mining and D, Muratov E, Tropsha A, Baskin I, Horvath
Learning with Graphs 2006, Berlin D, Marcou G, Muller C, Varnek A,
50. Zheng W, Tropsha A (2000) Novel variable Prokopenko VV, Tetko IV (2010)
selection quantitative structure-property rela- Applicability domains for classification prob-
tionship approach based on the k-nearest- lems: benchmarking of distance to models for
neighbor principle. J Chem Inf Comput Sci Ames mutagenicity set. J Chem Inf Model
40:185–194 50:2094–2111
136 Igor I. Baskin
62. Ralaivola L, Swamidass SJ, Saigo H, Baldi P Draize rabbit eye test: the use of QSARs and
(2005) Graph kernels for chemical informat- in vitro tests for the classification of eye irrita-
ics. Neural Netw 18:1093–1110 tion. Altern Lab Anim 33:215–237
63. Rupp M, Schneider G (2010) Graph kernels 76. Benigni R, Bossa C (2008) Predictivity and
for molecular similarity. Mol Inform reliability of QSAR models: the case of muta-
29:266–273 gens and carcinogens. Toxicol Mech Methods
64. Kashima H, Tsuda K, Inokuchi A (2003) 18:137–147
Marginalized kernels between labeled graphs. 77. Goldberg DE (1989) Genetic algorithms in
In: Proceedings, twentieth international con- search, optimization, and machine learning.
ference on machine learning, vol 1. AAAI Addison-Wesley Professional, New York
Press, Washington D.C., pp 321–328 78. DeLisle RK, Dixon SL (2004) Induction of
65. Menchetti S, Costa F, Frasconi P 2005 decision trees via evolutionary programming.
Weighted decomposition kernels. In: J Chem Inf Comput Sci 44:862–870
Proceedings of the 22nd international confer- 79. Dietterichl TG (2002) Ensemble learning. In:
ence on Machine learning. ACM, Arbib M (ed) The handbook of brain theory
pp. 585–592 and neural networks. MIT Press, Cambridge,
66. Swamidass SJ, Chen J, Phung P, Ralaivola L, pp 405–408
Baldi P (2005) Kernels for small molecules 80. Svetnik V, Wang T, Tong C, Liaw A, Sheridan
and the prediction of mutagenicity, toxicity RP, Song Q (2005) Boosting: an ensemble
and anti-cancer activity. Bioinformatics learning tool for compound classification and
21:I359–I368 QSAR modeling. J Chem Inf Mod
67. Mahé P, Ueda N, Akutsu T, Perret J-L, Vert 45:786–799
J-P (2005) Graph kernels for molecular struc- 81. Baskin II, Marcou G, Horvath D, Varnek A
ture-activity relationship analysis with support (2017) Bagging and boosting of classification
vector machines. J Chem Inf Mod models. In: Tutorials in chemoinformatics.
45:939–951 John Wiley & Sons, Ltd, Hoboken,
68. Breiman L, Friedman J, Stone CJ, Olshen RA pp 241–247
(1984) Classification and regression trees. 82. Baskin II, Marcou G, Horvath D, Varnek A
Chapman & Hall/CRC, Wadsworth, (2017) Bagging and boosting of regression
California models. In: Tutorials in chemoinformatics.
69. Cheng A, Dixon SL (2003) In silico models John Wiley & Sons, Ltd, Hoboken,
for the prediction of dose-dependent human pp 249–255
hepatotoxicity. J Comput Aided Mol Des 83. Baskin II, Marcou G, Horvath D, Varnek A
17:811–823 (2017) Random subspaces and random for-
70. Susnow RG, Dixon SL (2003) Use of robust est. In: Tutorials in chemoinformatics.
classification techniques for the prediction of John Wiley & Sons, Ltd, Hoboken,
human cytochrome P450 2D6 inhibition. pp 263–269
J Chem Inf Comput Sci 43:1308–1315 84. Baskin II, Marcou G, Horvath D, Varnek A
71. Feng J, Lurati L, Ouyang H, Robinson T, (2017) Stacking. In: Tutorials in chemoinfor-
Wang Y, Yuan S, Young SS (2003) Predictive matics. John Wiley & Sons, Ltd, Hoboken,
toxicology: benchmarking molecular descrip- pp 271–278
tors and statistical methods. J Chem Inf 85. Breiman L (1996) Bagging predictors. Mach
Comput Sci 43:1463–1470 Learn 24:123–140
72. Cramer GM, Ford RA, Hall RL (1976) 86. Ho TK (1998) The random subspace method
Estimation of toxic hazard—a decision tree for constructing decision forests. IEEE Trans
approach. Food Cosmet Toxicol 16:255–276 Pattern Anal 20:832–844
73. Verhaar HJM, van Leeuwen CJ, Hermens 87. Friedman JH (2002) Stochastic gradient
JLM (1992) Classifying environmental pol- boosting. Comput Stat Data Anal
lutants. Chemosphere 25:471–491 38:367–378
74. Walker JD, Gerner I, Hulzebos E, Schlegel K 88. Breiman L (1996) Stacked regressions. Mach
(2005) The skin irritation corrosion rules Learn 24:49–64
estimation tool (SICRET). QSAR Comb Sci 89. Breiman L (2001) Random forests. Mach
24:378–384 Learn 45:5–32
75. Gerner I, Liebsch M, Spielmann H (2005) 90. Svetnik V, Liaw A, Tong C, Culberson JC,
Assessment of the eye irritating properties of Sheridan RP, Feuston BP (2003) Random
chemicals by applying alternatives to the forest: a classification and regression tool for
compound classification and QSAR model- 103. Molnar L, Keseru GM, Papp A, Lorincz Z,
ing. J Chem Inf Comput Sci Ambrus G, Darvas F (2006) A neural net-
43:1947–1958 work based classification scheme for cytotox-
91. Li S, Fedorowicz A, Singh H, Soderholm SC icity predictions: validation on 30,000
(2005) Application of the random forest compounds. Bioorg Med Chem Lett
method in studies of local lymph node assay 16(4):1037–1039
based skin sensitization data. J Chem Inf Mod 104. Hatrik S, Zahradnik P (1996) Neural network
45:952–964 approach to the prediction of the toxicity of
92. Zhang Q-Y, Aires-de-Sousa J (2007) Random benzothiazolium salts from molecular struc-
forest prediction of mutagenicity from ture. J Chem Inf Comput Sci 36:992–995
empirical
physicochemical descriptors. 105. Zakarya D, Larfaoui EM, Boulaamail A,

J Chem Inf Mod 47:1–8 Lakhlifi T (1996) Analysis of structure-
93. Polishchuk PG, Muratov EN, Artemenko toxicity relationships for a series of amide her-
AG, Kolumbin OG, Muratov NN, bicides using statistical methods and neural
Kuz'min VE (2009) Application of ran- network. SAR QSAR Environ Res
dom forest approach to QSAR prediction 5:269–279
of aquatic toxicity. J Chem Inf Model 106. Eldred DV, Jurs PC (1999) Prediction of

49:2481–2488 acute mammalian toxicity of organophospho-
94. Vasanthanathan P, Taboureau O, Oostenbrink rus pesticide compounds from molecular
C, Vermeulen NPE, Olsen L, Jorgensen FS structure. SAR QSAR Environ Res 10:75–99
(2009) Classification of cytochrome P450 107. Devillers J, Flatin J (2000) A general QSAR
1A2 inhibitors and noninhibitors by machine model for predicting the acute toxicity of pes-
learning techniques. Drug Metab Dispos ticides to Oncorhynchus mykiss. SAR QSAR
37:658–664 Environ Res 1:25–43
95. Rumelhart DE, McClelland JL (1986) 108. Devillers J (2001) A general QSAR model for
Parallel distributed processing, vol 1,2. MIT predicting the acute toxicity of pesticides to
Press, Cambridge, MA Lepomis macrochirus. SAR QSAR Environ
96. Gasteiger J, Zupan J (1993) Neural networks Res 11:397–417
in chemistry. Angew Chem Int Ed Engl 109. Devillers J, Pham-Delegue MH, Decourtye
105:503–527 A, Budzinski H, Cluzeau S, Maurin G (2002)
97. Halberstam NM, Baskin II, Palyulin VA, Structure-toxicity modeling of pesticides to
Zefirov NS (2003) Neural networks as a honey bees. SAR QSAR Environ Res
method for elucidating structure-property 13:641–648
relationships for organic compounds. Russ 110. Kaiser KLE (2003) The use of neural net-
Chem Rev 72:629–649 works in QSARs for acute aquatic toxicologi-
98. Baskin II, Palyulin VA, Zefirov NS (2008) cal endpoints. J Mol Struct (THEOCHEM)
Neural networks in building QSAR models. 622:85–95
Methods Mol Biol 458:137–158 111. Zakarya D, Boulaamail A, Larfaoui EM,

99. Baskin II, Winkler D, Tetko IV (2016) A Lakhlifi T (1997) QSARs for toxicity of
renaissance of neural networks in drug discov- DDT-type analogs using neural network. SAR
ery. Expert Opin Drug Discovery QSAR Environ Res 6:183–203
11:785–795 112. Eldred DV, Weikel CL, Jurs PC, Kaiser KLE
100. Villemin D, Cherqaoui D, Mesbah A (1994) (1999) Prediction of fathead minnow acute
Predicting carcinogenicity of polycyclic aro- toxicity of organic compounds from molecu-
matic hydrocarbons from back-propagation lar structure. Chem Res Toxicol 12:670–678
neural network. J Chem Inf Comput Sci 113. Martin TM, Young DM (2001) Prediction of
34:1288–1293 the acute toxicity (96-h LC50) of organic
101. Xu L, Ball JW, Dixon SL, Jurs PC (1994) compounds to the fathead minnow
Quantitative structure-activity relationships (Pimephales promelas) using a group contri-
for toxicity of phenols using regression analy- bution method. Chem Res Toxicol
sis and computational neural networks. 14:1378–1385
Environ Toxicol Chem 13:841–851 114. Moore DRJ, Breton RL, MacDonald DB

102. Devillers J, Bintein S, Domine D, Karcher W (2003) A comparison of model performance
(1995) A general QSAR model for predicting for six quantitative structure-activity rela-
the toxicity of organic chemicals to lumines- tionship packages that predict acute toxicity
cent bacteria (Microtox test). SAR QSAR to fish. Environ Toxicol Chem
Environ Res 4:29–38 22:1799–1809
138 Igor I. Baskin
115.
Garg A, Bhat KL, Bock CW (2002) 128. Novotarskyi S, Abdelaziz A, Sushko Y, Körner
Mutagenicity of aminoazobenzene dyes and R, Vogt J, Tetko IV (2016) ToxCast EPA
related structures: a QSAR/QPAR investiga- in vitro to in vivo challenge: insight into the
tion. Dyes Pigments 55:35–52 rank-I model. Chem Res Toxicol
116. Shoji R (2005) The potential performance of 29:768–775
artificial neural networks in QSTRs for pre- 129. Abdelaziz A, Spahn-Langguth H, Schramm
dicting ecotoxicity of environmental pollut- K-W, Tetko IV (2016) Consensus modeling
ants. Curr Comput Aided Drug Des for HTS assays using in silico descriptors cal-
1:65–72 culates the best balanced accuracy in Tox21
117. Dearden JC, Rowe PH (2015) Use of artifi- challenge. Front Environ Sci 4. https://doi.
cial neural networks in the QSAR prediction org/10.3389/fenvs.2016.00002
of physicochemical properties and toxicities 130. Sushko I, Novotarskyi S, Körner R, Pandey
for REACH legislation. Methods Mol Biol AK, Rupp M, Teetz W, Brandmaier S,
1260:65–88 Abdelaziz A, Prokopenko VV, Tanchuk VY,
118. Tetko IV, Livingstone DJ, Luik AI (1995) Todeschini R, Varnek A, Marcou G, Ertl P,
Neural network studies. 1. Comparison of Potemkin V, Grishina M, Gasteiger J,
overfitting and overtraining. J Chem Inf Schwab C, Baskin II, Palyulin VA,
Comput Sci 35:826–833 Radchenko EV, Welsh WJ, Kholodovych V,
119. Tikhonov AN, Arsenin VA (1977) Solution of Chekmarev D, Cherkasov A, Aires-De-
ill-posed problems. Winston & Sons, Sousa J, Zhang QY, Bender A, Nigsch F,
Washington Patiny L, Williams A, Tkachenko V, Tetko
IV (2011) Online chemical modeling envi-
120. Winkler DA, Burden FR (2004) Bayesian
ronment (OCHEM): web platform for data
neural nets for modeling in drug discovery. storage, model development and publishing
Drug Discov Today: BIOSILICO of chemical information. J Comput Aided
2:104–111 Mol Des 25:533–554
121. Burden F, Winkler D (2008) Bayesian regu- 131. LeCun Y, Bengio Y, Hinton G (2015) Deep
larization of neural networks. Methods Mol learning. Nature 521:436–444
Biol 458:25–44
132. Bengio Y (2009) Learning deep architectures
122. Burden FR, Ford MG, Whitley DC, Winkler for AI. Found Trends Mach Learn 2:1–127
DA (2000) Use of automatic relevance deter-
mination in QSAR studies using Bayesian 1
33. Gawehn E, Hiss JA, Schneider G (2016)
neural networks. J Chem Inf Comput Sci Deep learning in drug discovery. Mol Inform
40:1423–1430 35:3–14
123. Burden FR, Winkler DA (2000) A quantita- 134. Goh GB, Hodas NO, Vishnu A (2017) Deep
tive structure-activity relationships model for learning for computational chemistry. J Comp
the acute toxicity of substituted benzenes to Chem 38:1291–1307
Tetrahymena pyriformis using Bayesian- 1
35. Ekins S (2016) The next era: deep learning in
regularized neural networks. Chem Res pharmaceutical research. Pharm Res
Toxicol 13:436–440 33:2594–2603
124.
Cronin MTD, Schultz TW (2001) 136. Mayr A, Klambauer G, Unterthiner T,
Development of quantitative structure- Hochreiter S (2016) DeepTox: toxicity pre-
activity relationships for the toxicity of aro- diction using deep learning. Front Environ
matic compounds to tetrahymena pyriformis: Sci 3:80
comparative assessment of the methodolo- 137. Bengio Y, Courville A, Vincent P (2013)
gies. Chem Res Toxicol 14:1284–1295 Representation learning: a review and new
125. Polley MJ, Burden FR, Winkler DA (2005) perspectives. Pattern Anal Mach Intell IEEE
Predictive human intestinal absorption QSAR Trans 35:1798–1828
models using Bayesian regularized neural net- 138. Kohonen T (2001) Self-organizing maps.
works. Aust J Chem 58:859–863 Springer, Berlin Heidelberg
126. Epa VC, Burden FR, Tassa C, Weissleder R, 139. Anzali S, Barnickel G, Krug M, Sadowski J,
Shaw S, Winkler DA (2012) Modeling bio- Wagener M, Gasteiger J, Polanski J (1996)
logical activities of nanoparticles. Nano Lett The comparison of geometric and electronic
12:5808–5812 properties of molecular surfaces by neural
127. Tetko IV (2002) Neural network studies. 4. networks: application to the analysis of
Introduction to associative neural networks. corticosteroid-binding globulin activity of
J Chem Inf Comput Sci 42:717–728 steroids. J Comput Aided Mol Des
10:521–534
140. Hecht-Nielsen R (1987) Counterpropagation 147. Gaspar HA, Baskin II, Marcou G, Horvath
networks. Appl Opt 26:4979–4984 D, Varnek A (2015) GTM-based QSAR mod-
141. Vracko M (1997) A study of structure-
els and their applicability domains. Mol
carcinogenic potency relationship with artifi- Inform 34:348–356
cial neural networks. The using of descriptors 148. Gaspar HA, Baskin II, Marcou G, Horvath
related to geometrical and electronic struc- D, Varnek A (2015) Stargate GTM: bridging
tures. J Chem Inf Comput Sci descriptor and activity spaces. J Chem Inf
37:1037–1043 Model 55:2403–2410
142.
Mazzatorta P, Vracko M, Jezierska A, 149. Gaspar HA, Baskin II, Varnek A (2016)

Benfenati E (2003) Modeling toxicity by Visualization of a multidimensional descrip-
using supervised Kohonen neural networks. tor space. In: Frontiers in molecular design
J Chem Inf Comput Sci 43:485–492 and chemical information science–Herman
143. Spycher S, Pellegrini E, Gasteiger J (2005) Skolnik Award Symposium 2015: Jürgen
Use of structure descriptors to discriminate Bajorath, vol 1222. ACS Symposium Series,
between modes of toxic action of phenols. vol 1222. American Chemical Society,
J Chem Inf Model 45:200–208 pp. 243–267
144. Bishop CM, Svensén M, Williams CKI (1998) 150. Gaspar HA, Sidorov P, Horvath D, Baskin II,
GTM: the generative topographic mapping. Marcou G, Varnek A (2016) Generative top-
Neural Comput 10:215–234 ographic mapping approach to chemical space
145. Kireeva N, Baskin II, Gaspar HA, Horvath D, analysis. In: Frontiers in molecular design and
Marcou G, Varnek A (2012) Generative top- chemical information science–Herman
ographic mapping (GTM): universal tool for Skolnik Award Symposium 2015: Jürgen
data visualization, structure-activity modeling Bajorath, vol 1222. ACS Symposium Series,
and dataset comparison. Mol Inform vol 1222. American Chemical Society,
31:301–312 pp. 211–241
146. Gaspar HA, Baskin II, Marcou G, Horvath 151. Kireeva N, Kuznetsov SL, Bykov AA, Tsivadze
D, Varnek A (2015) Chemical data visualiza- AY (2012) Towards in silico identification of
tion and analysis with incremental generative the human ether-a-go-go-related gene chan-
topographic mapping: big data challenge. nel blockers: discriminative vs. generative
J Chem Inf Mod 55:84–94 classification models. SAR QSAR Environ Res
24:103–117
Chapter 6
Applicability Domain: A Step Toward Confident Predictions

and Decidability for QSAR Modeling
Abstract
In the context of human safety assessment through quantitative structure–activity relationship (QSAR)
modeling, the concept of applicability domain (AD) has an enormous role to play. The Organization of
Economic Co-operation and Development (OECD) for QSAR model validation recommended as prin-
ciple 3 “A defined domain of applicability” to be present for a predictive QSAR model. The study of AD
allows estimating the uncertainty in the prediction for a particular molecule based on how similar it is to
the training compounds which are used in the model development. In the current scenario, AD represents
an active research topic, and many methods have been designed to estimate the competence of a model
and the confidence in its outcome for a given prediction task. Thus, characterization of interpolation space
is significant in defining the AD. The diverse set of reported AD methods was constructed through differ-
ent hypotheses and algorithms. These multiplicities of methodologies mystify the end users and make the
comparison of the AD for different models a complex issue to address. We have attempted to summarize
in this chapter the important concepts of AD including particulars of the available methods to compute the
AD along with their thresholds and criteria for estimating AD through training set interpolation in the
descriptor space. The idea about transparent domain and decision domain are also discussed. To help read-
ers determine the AD in their projects, practical examples together with available open source software
tools are provided.
Key words Applicability domain, Confidence, In silico, QSAR, Reliability
1 Introduction
A considerable amount of current chemical research has been

oriented toward the drug designing and risk assessment utilizing
the quantitative structure–activity relationships (QSARs) approach.
Validation plays an imperative role in the construction of predictive
QSAR models which may be reliable for the future prediction of
new chemical entity (NCE) [1]. Thus, a new area is directed to
introduction of efficient validation approaches for precise and pre-
dictive QSAR model. In this viewpoint, the objective of QSAR
modeling is to predict the response of the NCE falling within the
applicability domain (AD) of the developed model development.
141
142 Supratik Kar et al.
The reliability of the QSAR model relies on the confident predic-

tions of untested and new molecules based on the AD of the model,
and therein lies the importance of the AD study [2].
In the context of human safety assessment and management,
the global performance of a model is of little help when estimating
the confidence in a specific individual or local prediction. From a
local prediction perspective, the question is “How much one can
trust the model’s prediction for a query compound?” regardless of
the overall statistical performance of the model. To answer the
question, the idea of AD is introduced. The task of the AD is to
define the boundaries within which a model can be utilized and
offer reliable predictions. A well-defined AD space is a “must have”
characteristic for in silico prediction systems, and the Organization
of Economic Co-operation and Development (OECD) has
included AD as part of the requirements for a predictive QSAR
model [3, 4]. Thus, majority of in silico prediction systems feature
the capability to evaluate if a query molecule is a part of developed
model’s AD or not, and therefore give an improved sense of confi-
dence at the individual prediction level.
The AD articulates the fact that QSARs are inevitably associ-
ated with restrictions in terms of the grouping of chemical struc-
tures, physicochemical properties, and mechanisms of action
(MOA) for which the models can produce reliable predictions [5].
It is tremendously helpful for the QSAR modelers to have informa-
tion about the AD space of the developed model to identify inter-
polation (true prediction) or extrapolation (less reliable prediction)
[6]. In this chapter we explore the concept of AD and decision
domain (DD), available methods for AD as well as accessible open
source software tools to calculate AD with solved examples.
2 Concept of Applicability Domain
A QSAR model is fundamentally assessed in terms of its predict-

ability, suggesting how well it is capable of predicting the response
values of the query or test compounds. A model efficiently vali-
dated internally and externally can be considered reliable for both
scientific and regulatory purposes [7]. In this background, QSAR
model predictions are most consistent if they come from the mod-
el’s AD which is broadly defined under OECD principle 3 [8]. The
OECD includes AD assessment as one of the QSAR acceptance
criteria for regulatory purposes [8]. The Setúbal, Portugal work-
shop held in March 2002 set up the guidelines for the validation of
QSAR models, in particular for regulatory purposes [9]. These
principles were approved by the OECD member countries, QSAR
and regulatory communities at the 37th Joint Meeting of the
Chemicals Committee and Working Party on Chemicals, Pesticides
and Biotechnology in November 2004. The report [9] presented
Applicability Domain: A Step Toward Confident Predictions and Decidability for QSAR… 143
the following regulation for AD assessment under OECD Principle

3-a defined domain of applicability:
“The applicability domain of a (Q)SAR is the physico-chemical,
structural, or biological space, knowledge or information on which the
training set of the model has been developed, and for which it is appli-
cable to make predictions for new compounds. The applicability
domain of a (Q)SAR should be described in terms of the most relevant
parameters, i.e. usually those that are descriptors of the model. Ideally
the (Q)SAR should only be used to make predictions within that
domain by interpolation not extrapolation.”
This depiction is useful for explaining the instinctive meaning
of the “applicability domain” approach.
The AD [10] is a theoretical region in the chemical space sur-
rounding both the model descriptors and modeled response. While
building a QSAR model, the AD plays a deciding role for checking
the uncertainty in the prediction of a specific molecule based on
how similar it is to the compounds used to build the model. Thus,
the prediction of a modeled response employing QSAR is applica-
ble only if the molecule being predicted falls within the AD of the
model as it is unfeasible to predict a complete universe of chemicals
using a single QSAR model [11–13]. It is interesting to point out
that the choice of the training and test set compounds has a critical
effect on the AD of the QSAR model. Thus, while dividing a data-
set for external validation, the training set molecules should be
selected in such a way that they cover the entire chemical space for
all the dataset. To achieve reliable predictions, a QSAR model
should always be used for compounds within its AD.
3 Concept of Decision Domain
Hanser et al. [14] proposed the significance of the decision domain

with a three-step approach: applicability domain, reliability domain,
and decidability domain. Putting together all three steps the deci-
sion domain can be defined as follows: “The Decision Domain
(DD) is the scope within which it is possible to make a decision based
on a valid, reliable and non-equivocal prediction”.
Applicability: It is the first step to check if the model is valid to
apply for the compound being predicted or to confirm that the
model’s specifications are compliant with the intended use case. If
the model is not appropriate for the intended prediction assign-
ment, one can state that the query compounds are outside the
model’s AD. If the specifications of the model are compatible with
the query molecule then they are inside the AD.
Reliability: Once it is certain that the model is appropriate for use;
one can proceed to the next step by making a prediction and con-
sider its reliability. The intention of this reliability metric is to notify
the user about the quantity, quality and relevance of the information
available to the model to perform the prediction task. If the out-
come is good enough, the resulting prediction is expected to be
more reliable than the case when the input information is bad. The
reliability generally captures the relevance and quality of the infor-
mation available to the model for a given prediction.
Decidability: To conclude when the prediction has been supposed
to be valid and reliable, one can consider its actual outcome. The
degree of assertiveness of the conclusion that can be resultant from
the prediction will rely on the weight of confirmation that supports
this conclusion. Decidability is directly related to the error estimate
of an individual prediction and it measures how much one can
believe the conclusion derived from a prediction. It is imperative
not to confuse the term reliability with decidability. The former
defines how much one can trust the prediction itself, whereas the
latter captures how much one can trust the conclusion derived
from this prediction. The idea and steps behind decision domain is
reported in Fig. 1.
Fig. 1 The concept of confidence level and decision domain

4 Methods of Applicability Domain
An ideal AD approach estimates interpolation regions in a multi-

variate space. Major in silico prediction systems feature the capac-
ity to evaluate if a molecule is a part of their AD or not, and thus
provide a better judgment of confidence at the individual predic-
tion level. The wide spectrum of methods developed for this pur-
pose is based on various definitions of the AD concept, and often
takes into account different types of information [4, 6, 15]. Most
of the AD approaches follow a range in descriptor values or dis-
tance based hypothesis. The threshold for distance based meth-
ods is the largest distance between the training set data points
and the center of the training data set. Along with the common
approaches, Stanforth et al. [16] proposed a cluster-based
approach to evaluate the AD of any QSAR model. The method
applies an intelligent version of the k-means clustering algorithm
for modeling the training set as a compilation of clusters in the
descriptors space. A classification- based approach calculates
regression residuals for identification of good and bad classes as
proposed by Guha & Jurs [17]. To determine AD, all available
approaches (Graphically explained in Fig. 2) are discussed here
thoroughly based on their hypotheses, threshold criteria along
with their strength and weakness to enlighten the interpolation
space.
4.1 Range Based Hypothesis: Modeled descriptors are considered with a uniform dis-
Approaches in the tribution, defining an n-dimensional hyperrectangle developed on
Descriptor Space the basis of the maximum and minimum values of each descriptor
with sides parallel to the coordinate axes.
4.1.1 Bounding Box
Threshold: As suggested, this is based on the highest and lowest
values of X variables (descriptors of the QSAR model) and Y vari-
able (response) of the training set. Any test molecules, which are
outside of these particular ranges, are considered out of the AD
and their predictions are less reliable [15]. For the “bounding
box,” the zone of AD is the smallest axis-aligned rectangular box
encloses all the data points presented in Fig. 3.
Flaws: (a) The method encloses substantial empty space in case of
nonuniformly distributed data points, (b)As only descriptor ranges
are employed for AD space determination, empty regions in the
interpolation space cannot be identified, (c) Correlation among
descriptors cannot be considered [15].
Fig. 2 Available AD techniques classified under different hypothesis
Fig. 3 Concept of bounding box plot

Fig. 4 Graphical demonstration of the AD in principal component (PC) space
4.1.2 PCA Bounding Box Hypothesis: Principal components (PCs) transform the initial data
into a new orthogonal coordinate system by the rotation of axes
and facilitate to correct for correlations among descriptors. Newly
formed axes are defined as PCs presenting the maximum variance
of the total dataset. The points between the lowest and highest
value of each PC defined by M-dimensional (M is the number of
significant components) hyper-rectangle with sides parallel to the
PCs [18, 19].
Threshold: The AD in PC space is reported in Fig. 4 where the train-
ing set is represented by the biggest circle. The predictions of
query compounds within the circle are considered reliable. Query
molecules located outside the model space would be expected to
be less reliably predicted.
4.1.3 TOPKAT Optimal Hypothesis: Variation of principal component analysis (PCA) is

Prediction Space implemented by the optimum prediction space (OPS) from
TOPKAT OPS [20]. In the PCA approach, instead of the stan-
dardized mean value, the data are centered on the mean of indi-
vidual parameter range ([xmax–xmin]/2). Therefore, it generates a
new orthogonal coordinate system which is known as OPS coordi-
nate system. The OPS boundary is defined by the minimum and
maximum values of the generated data points on each axis of the
OPS coordinate system.
Threshold: the property sensitive object similarity (PSS) is executed
in the TOPKAT as a heuristic solution to imitate the data set’s
dense and spare regions, and includes the response variable (y).
Precision of prediction between the training and query points is
checked by the PSS. Similarity search technique is employed to
estimate the performance of TOPKAT in predicting the effects of

a chemical that is structurally analogous to the training data points.
4.2 Geometrical Hypothesis: This method estimates the direct coverage of an

Methods n-dimensional set utilizing the convex hull calculation [21] per-
formed based on complex but efficient algorithms. The approach
4.2.1 Convex Hull
recognizes the boundary of the dataset considering the degree of
data distribution.
Threshold: Interpolation space is suggested by the smallest axis-
aligned convex regions containing the entire training set illustrated
in Fig. 5.
Flaws: (a) Increase in data dimensions contributes to the order of
data complexity. Generally, convex hull calculation is good for
two and three dimensions. The complexity swiftly amplifies in
higher dimensions. For n points and d dimensions, the complex-
ity is of order O which can be defined as O = [n[d/2] + 1] [21], (b)
The method simply analyses the set boundaries without bearing
in mind the actual data distribution, and (c) It cannot categorize
the probable internal empty regions within the interpolation
space [15].
Fig. 5 The idea of convex hull plot

Fig. 6 The extrapolation regions of distance-to-centroid plot
4.3 Distance-Based Distance-based approaches are focused on the “distance-to-

Methods centroid” principle. The ellipsoidal region is centered on the data-
set grand mean and has its principal axes based on the eigenvectors
of the dataset variance–covariance matrix. Most frequently
employed distance-based AD approaches are discussed in the fol-
lowing sections. In Fig. 6, the AD space of “distance-to-centroid”
is presented.
4.3.1 Leverage Approach Hypothesis: The leverage (h) of a molecule in the variable space is
computed based on the HAT matrix as:
H = (X T ( X T X )−1 X ) (1)

In Eq. (1), H is a [n × n] matrix that orthogonally projects vectors
into the space spanned by the columns of X [3, 10]. The AD space
of the model is defined as a squared area within the ±3 band for
standardized residuals (σ) and the leverage threshold is defined as
h* = 3(p + 1)/n, where p is the number of descriptors and n is the
number of molecules. The leverage values (h) are calculated for
and plotted vs. cross-validated standardized residuals (σ) (Y-axis)
labeled as the Williams plot.
Threshold criteria: Williams plot (Fig. 7) confirms the occurrence
of response outliers and training compounds that are structurally
very influential in determining model parameters. The data
predicted for high leverage chemicals in the prediction set are
extrapolated and could be less reliable.
Fig. 7 Concept of Williams plot
4.3.2 Euclidean Distance Hypothesis: Euclidean method calculates the distance from every
other point to a particular point in the data set. A distance score,
dij, for two different compounds Xi and Xj can be measured by the
Euclidean distance norm. The Euclidean distance can be expressed
by the following equation:
k =1
( )
2
dij = ∑ x ik − x jk (2)
m
The mean distances of one sample to the residual ones are calcu-
lated as follows:
j =1
∑ dij
di = n (3)
n −1
where, i = 1,2, …, n.
The mean distances are then normalized within the interval of
zero to one. It is appropriate only for statistically independent vari-
ables [22].
Threshold criteria: The mean normalized distances are measured
for both training and test molecules. The boundary area created by
normalized mean distance scores of the training set are defined as
the zone of AD for test molecules. If a test compound resides
inside the domain covered by the training set, it suggests that this
molecule is inside the AD, otherwise not. An example of Euclidean
distance plot is reported in Fig. 8.
Fig. 8 Euclidean distance plot for AD study
4.3.3 Mahalanobis Hypothesis: The Mahalanobis distance method calculates the dis-
Distance tance of an observation from the mean values of the independent
variables but not considering the effect on the predicted value. It
provides one of the exclusive and straightforward approaches for
identification of outliers. Mahalanobis distance is unique because it
automatically takes into account the correlation between descrip-
tor axes [23].
Threshold criteria: Observations values much higher than those of
the remaining ones may be considered to be outside the domain.
4.3.4 City Block Distance City-block distance is the summed difference across dimensions
and is computed from the following equation:
i =1
d ( x ,y ) = ∑ x i − y i (4)
n
It examines the absolute differences between coordinates of a pair
of objects (xi and yi) and assumes a triangular distribution. This
method is valuable for the discrete type of descriptors. It is used
only for training sets which are uniformly distributed with respect
to count-based descriptors or fragment mapping counts [22].
4.3.5 Hotelling T2 Test Hypothesis: The Hotelling T2 is a multivariate student’s t-test and
proportional to leverage and Mahalanobis distance methods. It
presumes a normal data distribution like the leverage approach and
utilized to estimate the statistical impact of the difference on the
means of two or more variables between two groups. Hotelling T2
corrects for collinear descriptors through the use of the covariance
matrix and measures the distance of an observation from the center

of a set of X observations. A tolerance volume is derived for
Hotelling T2 [23].
Threshold criteria: Query compound is inside or not, evaluated
based on the t value.
4.3.6 k-Nearest Hypothesis: The theory is based on similarity search for a new
Neighbors Approach chemical entity with respect to the space generated by the training
set compounds. The similarity is evaluated by taking the distance
of a query molecule from nearest training compound or its dis-
tances from k-nearest neighbors in the training set. Therefore,
similarity to the training set molecules is noteworthy for this
method to facilitate a query compound with trustworthy predic-
tion [24].
Threshold criteria: If the calculated distance values of test or query
compounds are within the user mentioned threshold set by the
training molecules, then the prediction of these molecules are reli-
able. A k-nearest neighbors plot is presented in Fig. 9.
4.3.7 DModX (Distance Hypothesis: The method was developed by Wold et al. [5] and usu-
to the Model in X-Space) ally applied for partial least squares (PLS) regression models. The
fundamental hypothesis lies in the residuals of Y and X which are of
diagnostic value for the quality of the model. As there are number
Fig. 9 k-nearest neighbors plot

of X-residuals, one needs a summary for individual observation

which is accomplished by the residual standard deviation (SD) of
the X-residuals of the corresponding row of the residual matrix
E. As this SD is proportional to the distance between the data
point and the model plane in X-space, it is known as DModX (dis-
tance to the model in X-space). Here, X is the matrix of predictor
variables, of size (N × K), Y is the matrix of response variables, of
size (N × M) and E is the (N × K) matrix of X-residuals, N is num-
ber of objects (cases, observations), k is the index of X-variables
(k = 1, 2, ..., K) and m is the index of Y-variables (m = 1, 2, ..., M).
Threshold criteria: A DModXS value higher than around 2.5 times
the overall SD of the X residuals (related to an F-value of 6.25)
indicates that the query molecule is outside the AD. In the DModX
plot, the threshold line is attributed as D-critical line and the plot
can be simply drawn in SIMCA-P software for PLS models [25]. A
DModX plot is reported in Fig. 10 where the plot is generated at
95% confidence level by SIMCA-P. The DModXS values of 16 test
molecules out of total 20 compounds are within the stipulated
critical value of 1.97. Compounds 5, 6, 12, and 18 are outside the
D-critical value indicating that they are outside the AD of the
developed model.
4.3.8 Tanimoto Similarity The Tanimoto index measures the similarity between two com-
pounds based on the number of common molecular fragments
[26]. To calculate the Tanimoto similarity, all unique fragments of
Fig. 10 DModXS plot for PLS based QSAR model

a particular length in two compounds are calculated. The Tanimoto

similarity between the compounds J and I is defined as:
i =1
∑ ( x J ,i ⋅ x K ,i )
TANIMOTO ( J ,K ) = i =1
N
i =1 i =1 (5)
∑ ( x J ,i ⋅ x J ,i ) + ∑ ( x K ,i .x K ,i ) − ∑ ( x J ,i .x K ,i )
N N N
where, N is the number of unique fragments in both the com-

pounds, xJ, i and xK, i are the counts of the i-th fragment in the
compounds J and K. Based on eq. 5, the distance between two
compounds J and K is “1– TANIMOTO(J, K)”, and the distance
of a molecule to the model is the minimum distance between the
query compound and compounds from the training set.
4.3.9 Standard Deviation Theory: The principle behind this idea is easy and it depicts that if
of the Ensemble different models give significantly different predictions for a par-
Predictions (STD) ticular molecule, then the prediction for this compound is more
likely to be unreliable. The sample standard deviation can be ide-
ally used as an estimator of model uncertainty [26].
For example, consider that Y(J) = {yi(J), i = 1 ... N} is a set of
predictions for a molecule J given by a set of N trained models, the
consequent distance to model STD can be defined as:
∑ ( yi ( J ) − y )
2
dSTD ( J ) = stdev (Y ( J ) ) = (6)

N −1
Based on the category of an ensemble used to assess the stan-

dard deviation, the approach can be classified into several subtypes.
(a) Consensus STD or CONS-STD for models developed based on
machine learning techniques, (b) BAGGING-STD for an ensem-
ble of models utilizing the bagging technique, and (c) Associative
neural networks or ASNN-STD for an ensemble of neural network
models [26, 27]. A graphical demonstration of the STD approach
has been provided in Fig. 11 where reliable prediction has a low
prediction spread and unreliable prediction has higher prediction
spread.
4.3.10 Correlation Hypothesis: The approach is based on the correlation of vectors of

of Prediction Vectors ensemble’s predictions for the query molecule and compounds
(CORREL) from the training set [28]. Similar to the STD method, this calcu-
lation is applicable only for ensembles of models. Especially,
CORREL measure for the target compound J is calculated accord-
ing to the following expression:

i =1…n 
(
dCORREL ( J ) = 1 − max corr y (Ti ), y ( J ) ) (7)
Fig. 11 STD graph for AD estimation

In Eq. (7), y (Ti ) and y ( J ) define the vectors of ensemble’s pre-
dictions for the training set compound Ti and the target compound
J, corr is Spearman rank correlation coefficient between the two
vectors and N is the number of compounds in the training set.
Threshold criteria: A low value of CORREL indicates that for tar-
get compound J, there is a compound T from the training set for
which predictions of the ensemble of models are highly correlated.
If a compound T has the same descriptors as the target compound
J, then predictions of models will be identical for both molecules
and thus resulted CORREL(J) will be 0. Compounds having high
correlation coefficient values are considered to be “closer to the
model” [29, 30].
4.4 Probability Hypothesis: Probability density distribution is established based on

Density Distribution estimating the probability density function for the given data. The
approach is classified into two sub-classes. First one is a parametric
method which assumes a standard distribution such as Gaussian
and Poisson distributions, and the second one is a nonparametric
method which does not rely on such assumptions considering the
data distribution [6, 15]. These methods are performed by esti-
mating probability density of the molecules followed by identifying
highest density region which consists of a known fraction from the
total probability mass [6]. A potential is formed for individual mol-
ecule in the training set such that it is maximum for that molecule
and decreases with increasing distance. Once the potential is com-
puted for all the molecules, global potential is obtained by sum-
ming up the individual potentials thus representing the probability
density [31]. The significant feature of these approaches is their
ability to recognize the internal empty space. The real data distri-
bution can be exposed by generating concave regions around the
interpolation space borders.
Threshold criteria: Given two molecules xi and xj, it can be deter-

mined as below:
 
1 −1
ϕ= .exp   (8)
( 2π )s 
(
 ( 2s ) x i − x j ) 
2
2
where, Ф (xi and xj) is the potential induced on xj by xi and width

of the curve is defined by smoothing parameter s. The cutoff value
associated with Gaussian potential functions, namely fp can be cal-
culated by methods based on sample percentile [31]:

(
f p = f i + (q − j ) f j +1 − f j ) (9)
with q = p × n , where p is the percentile value of probability-

100
density, n is the number of compounds in the training set and j is
the nearest integer value of q. Query molecules with potential
function values lower than the defined threshold are identified as
outside the AD.
4.5 Range A compound can be identified outside the AD by checking distri-

of the Response bution plot or evaluating the range of the response variable. In a
Variable dataset, if the response value of a specific test molecule is largely
different from the mean response value of the training set mole-
cules, then the compound can be considered as out of the AD.
4.6 Miscellaneous There are numerous miscellaneous approaches introduced in the

Approaches recent times to determine the AD of constructed QSAR models.
4.6.1 Standardization Theory: The approach is proposed by Roy et al. [32]. According to
Technique ideal data distribution, 99.7% of the population would stay within
the range mean ± 3 standard deviation (SD). Thus, mean ± 3SD
represents the zone where majority of the training compounds
belong to. Any molecule outside this zone is different from the rest
of the compounds. Thus, after a descriptor column is standardized
based on the corresponding mean and standard deviation for the
training set compounds only, if the corresponding standardized
value for descriptor i of compound k (Ski) is more than 3, then the
compound should be a X-outlier (if in the training set) or outside
AD (if in the test set) based on descriptor i.
Algorithm and methodology:
1. Standardize all the Descriptors (Training and Test Set) Using

the Following Formula:
X ki − X i
Ski = (10)
σ Xi

Where, k = 1, 2, 3… nComp (here, nComp = total number of

compounds).
i = 1, 2, 3… nDes (here, nDes = total number of descriptors).
Ski = Standardized descriptor i for compound k (from the train-
ing or test set),
Xki = Original descriptor i for compound k (from the training
or test set).
X i = Mean value of the descriptor Xi for the training set com-
pound only.
σ Xi = standard deviation of the descriptor Xi for the training
set compounds only.
The above calculation should run for all descriptor values
(nComp × nDes) present in the model. So, one will have now nDes
number of the standardized descriptor Si(k) (i = 1 to nDes) values for
any compound k.
2. Thereafter, one needs to calculate the maximum Si(k) value
([Si]max(k)) for the molecule k. If [Si]max(k) is lower than or equal
to 3, then that compound is not an X-outlier (if in the training
set) or is within applicability domain (if in the test set).
3. If [Si]max(k) is above 3, then one should compute [Si]min(k). If
[Si]min(k) > 3, then the compound is a X-outlier (if training
compound) or is not within AD (if test compound).
4. If [Si]max(k) > 3 and [Si]min(k) < 3, then one should compute
Snew(k) from the following equation:
S new (k ) = Sk + 1.28 × σ Sk (11)

Where, Snew(k)= Snew value for the compound k.
Sk =Mean of Si(k) values of the compound k.
σ Sk = standard deviation of Si(k) values of the compound k.
If the calculatedSnew(k)value is lower than or equal to 3, then
that compound is not an X-outlier (if in the training set) or is
within AD (if in the test set).
Drawbacks:
(a) It does not consider intercorrelation among descriptors.
(b) It does not consider relative contribution of descriptors (a
lower contributing descriptor to model should also be less
influential in determining the AD).
4.6.2 Stepwise Based Hypothesis: It is a stepwise based method proposed by Dimitrov

Approach et al. [33] helpful for identifying mechanistic rationality and trans-
parency of the QSAR model. This approach is divided in four
stages:
1. General domain: It checks the variation of the physicochemical

properties of the query chemicals in the training set
compounds.
2. Structural domain: It defines the structural similarity found
within the chemicals that are correctly predicted by the model.
3. Mechanistic domain: The model domain merges the reliability
of specific functional groups assumed to cause the effect and
the domain of explanatory variables.
4. Domain of metabolic simulation: If simulated metabolism of
chemicals is a part of the developed QSAR model then the
dependability of simulated metabolism is considered in assess-
ing the reliability of the predictions.
Threshold criteria: A test molecule needs to meet all the conditions
within the mentioned four stages to be considered within the
AD. The approach is a rigorous one as a compound is evaluated for
similarity, metabolic, and mechanistic check to address the reliabil-
ity of predictions and better assessment of model’s AD [33].
4.6.3 Decision Trees The AD space is identified based on the consensus prediction of
and Decision Forests Decision Trees (DT) and Decision Forests (DF). The hypothesis is
to minimize the overfitting which can be attained by merging the
DTs and maintaining the differences within different DTs to maxi-
mum possible. Predictions from all the combined DTs are aver-
aged in order to find the prediction confidence for a particular
compound, while domain extrapolation offers the prediction preci-
sion for that compound outside the training space [34, 35].
4.6.4 Kernel-Based Most machine learning approaches for QSAR rely on a vectorial
Applicability Domain presentation of the compounds. Thus, the AD is expressed as a
subspace of the vector space with one dimension for each descrip-
tor used. However, this vectorial perception cannot be straight for-
wardly applied to kernel-based techniques like support vector
machines (SVM). Thus, these methods have to rely on an implicit
feature space which is only defined by the applied kernel similarity
and with unknown dimensions. Therefore the domain of applica-
bility of a kernel-based model has to be defined by means of the
kernel. This also offers us to utilize the structured similarity mea-
sures like the Optimal Assignment Kernel and its extension, instead
of a numerical encoding. The concept of kernel density estimation
has incorporated additional information in a trained model. The
added information can be achieved by employing a weighted aver-
age kernel similarity of a predicted molecule to the training data
set. The weights can be attained either by using the knowledge
contained in the learned model or by methods that describe the
feature space structure using the kernel [36].
4.6.5 Intelligent k-Means An intelligent version of the k-means clustering algorithm-based

Clustering measure of distance-to-domain for QSAR models was designed by
Stanforth et al. [16]. This approach combines the modeling of the
training molecules as a compilation of the intelligent k-means clus-
ters algorithm in the descriptor space with a new interpretation of
a usual optimization criterion in fuzzy clustering which leads to a
modified harmonic mean measure. A test compound is assigned
fuzzy membership of individual cluster, from which an overall dis-
tance may be computed. Stanforth et al. [16] reported that the
approach is more stable than existing methods and more indicative
of the prediction error.
k-means clustering is obtained based on the approximation of
each point in a dataset by the centroid of that point’s cluster. This
assessment can be quantified by decomposing the data scatter into
contributions that are explained and unexplained by the cluster
model as presented in the following equation [37]:
k =1 k =1
∑ x i ⋅ x i = ∑ N kc k ⋅ c k + ∑ ∑ ( x i − c k ) ⋅ ( x i − c k ) (12)
i ∈T k k i ∈Ck

In eq. (12), {xi: i ϵ T} is the training set and the k clusters Ck have
sizes Nk and centroids Ck, respectively. This suggests an optimiza-
tion of the so-called square-error k-means criterion
k =1
∑ ∑ ( xi − ck ) ⋅ ( xi − ck ) (13)
k i ∈Ck

Equation 13 suggests for any inner product in the descriptor space.
k-means clustering needs an initialization step to indicate the
number of clusters k and the preliminary positions of their cen-
troids. The initialization of the intelligent k-means algorithm has
been achieved by the principle of anomalous pattern clustering
(APC) [37, 38].
5 Software Employed for Applicability Domain Determination
This is the era of computer simulation and molecular modeling

without any doubt. A huge number of chemicals, pharmaceuticals,
food products, and agrochemicals are designed by different
branches of modeling. Therefore, to improve the prediction
acceptability and reliability, the AD techniques have emerged as
important contributors. With the improvement of different AD
approaches, a good number of open source software has also
arrived, which are freely available to everyone. Here, we discuss
open source software tools for evaluation of AD, which are also
illustrated in Fig. 12.
Fig. 12 Commonly employed open source software for defining AD of a QSAR model
5.1 Cheminformatics Professor Kunal Roy and his group have developed open source AD
Tools by DTC Lab software available under following links: http://teqip.jdvu.ac.in/
QSAR_Tools/ and http://dtclab.webs.com/software-tools. Users
can employ three different approaches for The AD study.
(a) AD (using standardization approach) v1.0: It is a tool to find
out test/query molecules that are outside the AD of the devel-
oped QSAR model, and it also detects outliers present in the
training set compounds by standardization technique.
(b) Euclidean AD 1.0: It is used to ensure that the compounds of
the test set are representative of the training set. It is based on
distance scores calculated by the Euclidean distance norm.
(c) AD-MDI GUI 1.2: Applicability domain-Model Disturbance
Index (AD-MDI) program is a tool to define AD of unknown
samples.
5.2 QSARINS Professor Paola Gramatica and her group have developed QSARINS
software under which user can develop and validate QSAR model
along with AD study employing leverage based technique. The
software is available from http://www.qsar.it/. Along with the
leverage based AD evaluation, user can obtain Williams plot
explaining critical HAT value and zone of applicability for the
developed model.
5.3 Domain Manager The Domain Manager (Presently Version 1.06) is a software tool
developed by Laboratory of Mathematical Chemistry, University
“Prof. Assen Zlatarov” Bourgas,Bulgaria, obtainable from http://
oasis-lmc.org/products/software/domain-manager.aspx. The
tool can be implemented for automatic extraction of features from
training set compounds utilized for QSAR modeling purpose.
These features are encoded and ordered as layers in the final model
domain ordered in the following way:
Parameter ranges: It corresponds to ranges of the molecular
parameters for compounds in the training set, which is the descrip-
tor space of the training set;
Structural domain: It defines the structural similarity between
molecules that are correctly predicted by the model. The structural
neighborhood of atom-centered fragments is employed to deter-
mine this similarity;
Mechanistic domain: It combines the consistency of definite reac-
tive groups hypothesized to cause the effect and the domain of
explanatory variables determining the parametric requirements in
order for functional groups to elicit their reactivity;
Metabolism domain: It accounts the reliability of simulated metab-
olism (metabolites, pathways, and maps) if metabolic activation of
chemicals is a part of the QSAR model.
6 How to Determine Applicability Domain: Solved Examples
To determine the AD for a regression based QSAR model, we have

reported here a set of hypothetical data in Table 1 and a corre-
sponding QSAR model developed with multiple linear regression
(MLR) tool with three descriptors. For QSAR learners, most fre-
quently used and simple to calculate AD approaches like bounding
box, range of the response variable, leverage method, Euclidean
distance based method, and standardization technique have been
used here. All these methods can be used by QSAR learners with
open access software or by hand calculation.
(a) Bounding box and Range of the response variable: Both
approaches are firmly reliant on the boundary formed by inde-
162
Supratik Kar et al.
Table 1
An arbitrary example representing calculation of AD for a regression-based QSAR model employing methods like bounding box, leverage method,
Euclidean distance based method, and standardization approach
ID Y* (Observed) X1* X2* X3* Y (Calculated) ha Standardized residual Distance scoreb Mean distancec Normalized Mean distance
1 3.50 1.64 2 1.64 3.45 0.41 1.63 16.174 1.348 0.841
3 3.21 1.62 3 1.10 2.74 0.14 0.48 8.708 0.726 0.014
5 2.87 2.12 3 1.37 2.68 0.39 1.56 10.072 0.839 0.165
7 2.68 2.01 3 1.35 2.68 0.22 0.83 9.422 0.785 0.093
9 2.27 1.85 3 1.23 2.71 0.17 0.61 8.709 0.726 0.014
10 2.22 1.91 4 0.97 1.98 0.32 1.26 12.334 1.028 0.416
11 2.18 1.36 4 0.68 2.05 0.36 1.41 13.068 1.089 0.497
13 1.79 1.52 4 0.78 2.02 0.29 1.14 12.136 1.011 0.394
15 3.47 1.93 2 1.94 3.41 0.66 2.70 17.605 1.467 1
16 2.53 1.68 3 1.12 2.70 0.16 0.56 8.582 0.715 0
17 2.81 1.12 3 0.74 2.85 0.65 2.67 12.139 1.012 0.394
19 1.96 1.61 4 0.79 1.91 0.22 0.83 12.019 1.002 0.381
ID Y* (Observed) X1* X2* X3* Y (Calculated) ha Standardized residual Distance scoreb Mean distancec Normalized Mean distance
Test set
2 3.22 1.96 2 1.92 3.40 0.38 1.49 17.698 1.475 1.01
4 2.82 1.24 3 0.83 2.82 0.32 1.23 11.151 0.929 0.285
6 2.86 1.80 3 1.19 2.73 0.15 0.51 8.65 0.721 0.007
8 2.52 1.63 3 1.07 2.75 0.17 0.60 8.761 0.73 0.02
12 2.12 1.98 4 0.98 1.98 0.27 1.03 12.672 1.056 0.453
14 3.56 1.91 2 1.88 3.40 0.36 1.41 17.373 1.448 0.974
18 2.21 1.93 3 1.93 2.25 0.81 3.33 13.526 1.127 0.548
* a b
Y is the dependent variable; X1, X2, and X3 are the modeled descriptors, h is leverage value, h* is1, calculated from eq. 1, calculated from eq. 2, and ccalculated from eq. 3
Applicability Domain: A Step Toward Confident Predictions and Decidability for QSAR…
163
pendent and dependent variables of the training set molecules.

Analyzing the descriptor values for the training and test sets,
not a single test compound falls outside the AD boundary cre-
ated by the independent variables. On the other hand, one
compound (14) can be considered outside the AD boundary
considering the response variable range. In concluding remark,
considering both methods, only one compound (14) falls
outside the AD although it is very close to the boundary
created by response variable.
(b) Euclidean distance approach: Based on eqs. 2 and 3,
Euclidean distance scores and the mean distance scores are cal-
culated. Followed by this, mean distances are normalized
within the interval of zero to one. From the normalized mean
distances, one test compound (2) resides outside the AD,
although it is just above the boundary. The Euclidean graph
(Fig. 13a) is generated plotting the mean normalized distance
in Y-axis and compound numbers in X-axis. Also, open source
software Euclidean Applicability domain 1.0 under DTC lab
software can be used to get the output at a single click.
(c) Leverage approach: The standardized residual (SR) value is
calculated based on the following equation:
Residual value − Average Residualvalue ( Training )

SR = (14)
Standard deviation of Residualvalue( Training )

Based on the calculated leverage (h) and standardized
residual values, we have constructed Williams plot (Fig. 13b)
taking standardized values in Y-axis and leverage values in
X-axis. As the number of descriptors is three and number of
compounds in the training set is twelve, the critical leverage
(h*/HAT) value is = [3 × (3 + 1)]/12 = 1.00. Taking into con-
sideration the threshold criteria, all test compounds are within
the AD. On the contrary, one of the test compounds (18)
belongs to AD as it lies within the critical HAT value but the
prediction for this compound is not precise or reliable (greater
than 3σ unit). As mentioned earlier, one can also simply plot
Williams plot by using QSARINS software.
(d) Standardization technique: Employing training and test set
in Applicability domain 1.0 (using standardization approach)
under DTC lab software, one can easily get output on the out-
lier information of training set and AD information of test set
compounds (See Fig. 13c). All calculations are performed
based on the mentioned eq. 10.
It is important to remember that no single method of AD
determination can be considered as the universally the best one.
The results of one particular method may not match with another
Fig. 13 (a) Euclidean distance plot, (b) Williams plot, (c) Training set outliers and test set AD information from
standardization technique employing dependent and independent variable values provided in Table 1
approach as each method is based on different statistical algorithm

for the determination of AD. As a consequence, manifold
approaches should be considered before coming to concluding
remarks about AD space of a QSAR model.
7 Future Direction
Apart from acceptable statistical prediction and a well-defined

decision domain, a QSAR model needs to provide interpretable
predictions with meaningful hypotheses and evidence to support
the prediction. Hanser et al. [14] proposed TARDIS principle
for Transparency, Applicability, Reliability, Decidability,
Interpretability, and Support to capture these mentioned require-
Fig. 14 The TARDIS principle
ments (Fig. 14). The prediction tools which maintain and follow

the TARDIS principle, offer experts means to understand the pre-
diction and decision making steps.
The QSAR model developers need to realize that single AD
method cannot be relied upon as the ideal one to recognize the
interpolation region for any QSAR model. As the background
hypothesis of every AD study is different, one needs to consider
combination of approaches for reliable predictions. Widespread
research is being carried out on some new approaches or amend-
ment of already accessible approaches to make them more accept-
able and decision worthy ones. Future research regarding AD
studies should be directed considering following points.
(a) Global similarity test is required to identify whether the struc-
tural features of an untested molecules are enclosed in the
training set of compounds.
(b) The regulatory agencies and QSAR experts should maintain a
conceptual framework for the assessment of AD maintaining
the OECD guidelines.
(c) Acceptable confidence limits for various AD approaches need
to be defined.
(d) To make the concept of AD estimation more user friendly,
QSAR developers have to develop open access software tools.
(e) QSAR users must possess the AD concept and its implementa-
tion in the QSAR models, and no model should be accepted
without proper evidence of AD study.
(f) The hypothesis and assessment criteria of any newly developed
AD approaches should be transparent.
8 Conclusion
The study of AD for QSAR models is critical to estimate how much

one can rely on individual predictions, especially in case of drug
development and human safety assessment. All implemented AD

approaches have their own strengths and flaws. But there is no
doubt about strong statistical hypothesis and algorithms behind
the identification of interpolation regions by the discussed AD
methods. Therefore, it entirely depends on the developer how one
chooses the specific approach according to the requirements to
define the AD for the model more precisely and in robust way.
Consequently, it is recommended to evaluate the applicability
space through different promising strategies before assessing an
untested molecule to get the result under confidence level. The
concept of decision domain is established on three steps that suc-
cessively assess the applicability, reliability, and decidability of a pre-
diction and help developers’ transparent decision making and
support the conclusion with proper interpretation. In concluding
remarks, we may mention:
(a) As the most important criteria of any QSAR model is predic-
tion capability for new molecules, the AD study assures the
reliability of prediction.
(b) Before prediction of any test compound employing a QSAR
model, one should check whether the prediction from that
model is reliable or not for that compound.
(c) The choice of the training set molecules is very critical as it
should consist of the characteristics of the total dataset to
reflect the effect of each structural feature in the developed
model.
(d) The reliability of experimentally determined response values
should be carefully checked before developing the QSAR
model, as error in the response value will deceive the QSAR
model which will lead to flawed interpolation space for AD.
(e) Uncertainty aspect should be considered. If the developed
QSAR model is not reliable, one cannot get confidence in AD
assessment.
(f) Estimation of AD is dependent on a number of factors like the
model dimensionality, employed descriptors, endpoint value,
underlying data distribution, and algorithm of AD
determination. Therefore, all factors must be kept in mind and
employed in the study.
Acknowledgments
S.K. and J.L. thank the National Science Foundation (NSF/

CREST HRD-1547754, and NSF/RISE HRD-1547836) for
financial support. K.R. is thankful to the UGC, New Delhi for
financial assistance under the UPE II scheme.
References
1. Roy K, Kar S, Das RN (2015) Understanding formal definition. SAR QSAR Environ Res
the basics of QSAR for applications in pharma- 27:865–881
ceutical sciences and risk assessment. Academic 15. Jaworska J, Nikolova-Jeliazkova N, Aldenberg
Press, San Diego, CA, USA T (2005) QSAR applicability domain estima-
2. Roy K, Kar S (2015) Importance of applicability tion by projection of the training set descriptor
domain of QSAR models. In: Roy K (ed) space: a review. Altern Lab Anim 33:445–459
Quantitative structure-activity relationships in drug 16. Stanforth RW, Kolossov E, Mirkin B (2007) A
design, predictive toxicology, and risk assessment. measure of domain of applicability for QSAR
IGI Global, Hershey PA, USA, pp 180–211 modeling based on intelligent K-means cluster-
3. Gadaleta D, Mangiatordi GF, Catto M, Carotti ing. QSAR Comb Sci 26:837–844
A, Nicolotti O (2016) Applicability domain for 17. Guha R, Jurs PC (2005) Determining the
QSAR models: where theory meets reality. Int validity of a QSAR model-a classification
J Quant Struct Prop Relat J 1:45–63 approach. J Chem Inf Model 45:65–73
4. Mathea M, Klingspohn W, Baumann K (2016) 18. Nikolova-Jeliazkova N, Jaworska J (2005) An
Chemoinformatic classification methods and approach to determining applicability domain
their applicability domain. Mol Inform for QSAR group contribution models: an analy-
35:160–180 sis of SRC KOWWIN. Altern Lab Anim
5. Wold S, Sjostrom M, Eriksson L (2001) PLS- 33:461–470
regression: a basic tool of chemometrics. 19. Worth AP, Bassan A, Gallegos A, Netzeva TI,
Chemom Intell Lab Syst 58:109–130 Patlewicz G, Pavan M et al (2005) The charac-
6. Netzeva TI, Worth AP, Aldenberg T, Benigni terisation of (quantitative) structure- activity
R, Cronin MTD, Gramatica P et al (2005) relationships: preliminary guidance. ECB
Current status of methods for defining the Report EUR 21866 EN, European
applicability domain of (quantitative) structure- Commission, Joint Research Centre; Ispra,
activity relationships. Altern Lab Anim Italy, pp. 95
33:155–173 20. Topkat OPS (2000). U.S. Patent 6, 036, 349
7. Golbraikh A, Tropsha A (2002) Beware of q2! 21. Preparata FP, Shamos MI (1991) In: Preparata
J Mol Graph Model 20:269–276 FP, Shamos MI (eds) Computational geome-
8. OECD, Principles for the validation of (Q) try: an introduction. Springer-Verlag,
SARs (2004). http://www.oecd.org/datao- New York
ecd/33/37/37849783.pdf (Accessed 20 May, 22. Jaworska JS, Nikolova-Jeliazkova N, Aldenberg
2017) T (2004) Review of methods for applicability
9. Jaworska JS, Comber M, Auer C, Van Leeuwen domain estimation. Report, The European
CJ (2003) Summary of a workshop on regula- Commission-Joint Research Centre, Ispra,
tory acceptance of (Q)SARs for human health Italy
and environmental endpoints. Environ Health 23. Hair JF Jr, Anderson RE, Tatham RL, Black
Perspect 111:1358–1360 WC (2005) Multivariate data analysis. Pearson
10. Gramatica P (2007) Principles of QSAR mod- Education, Singapore
els validation: internal and external. QSAR 24. Sheridan R, Feuston RP, Maiorov VN, Kearsley
Comb Sci 26:694–701 S (2004) Similarity to molecules in the training
11. Weaver S, Paul Gleeson M (2008) The impor- set is a good discriminator for prediction accu-
tance of the domain of applicability in QSAR racy in QSAR. J Chem Inform Comput Sci
modeling. J Mol Graph Model 44:1912–1928
26:1315–1326 25. SIMCA-P 10.0. (2002) info@umetrics.com,
12. Roy K, Kar S, Das RN (2015) A primer on UMETRICS, Umea, Sweden, www.umetrics.
QSAR/QSPR modeling: fundamental con- com
cepts (SpringerBriefs in Molecular Science). 26. Tetko IV, Sushko I, Pandey AK, Zhu H,
Springer, Berlin Tropsha A, Papa E et al (2008) Critical assess-
13. Roy K, Kar S (2015) How to judge predictive ment of QSAR models of environmental toxic-
quality of classification and regression based ity against Tetrahymena pyriformis: focusing
QSAR models? In: Haq ZU, Madura J (eds) on applicability domain and overfitting by vari-
Frontiers of computational chemistry. able selection. J Chem Inform Comput Sci
Bentham, Sharjah, pp 71–120 48:1733–1746
14. Hanser T, Barber C, Marchaland JF, Werner S 27. Manallack DT, Tehan BG, Gancia E, Hudson
(2016) Applicability domain: towards a more BD, Ford MG, Livingstone DJ et al (2003) A
consensus neural network-based technique for approach for defining the applicability domain
discriminating soluble and poorly soluble com- of SAR and QSAR models. J Chem Inform
pounds. J Chem Inform Comput Sci Model 45:839–849
43:674–679 34. Tong W, Hong H, Fang H, Xie Q, Perkins R
28. Tetko IV (2008) Associative neural network. (2003) Decision forest: combining the predic-
Methods Mol Biol 458:185–202 tions of multiple independent decision tree
29. Tetko IV, Tanchuk VY (2002) Application of models. J Chem Inform Comput Sci
associative neural networks for prediction of 43:525–531
lipophilicity in ALOGPS 2.1 program. J Chem 35. Tong W, Hong H, Xie Q, Xie L, Fang H,
Inform Comput Sci 42:1136–1145 Perkins R (2004) Assessing QSAR limitations–
30. Chen JJ, Tsai CA, Young JF, Kodell RL (2005) a regulatory perspective. Curr Comput Aided
Classification ensembles for unbalanced class Drug Des 1:195–205
sizes in predictive toxicology. SAR QSAR 36. Fechner N, Jahn A, Hinselmann G, Zell A
Environ Res 16:517–529 (2009) Atomic local neighborhood flexibility
31. Jouan-Rimbaud D, Bouveresse E, Massart DL, incorporation into a structured similarity mea-
de Noord OE (1999) Detection of prediction sure for QSAR. J Chem Inform Model
outliers and inliers in multivariate calibration. 49:549–560
AnalyticaChimicaActa 388:283–301 37. Mirkin B (2005) Clustering for data mining: a
32. Roy K, Kar S, Ambure P (2015) On a simple data recovery approach. Chapman & Hall/
approach for determining applicability domain CRC, London
of QSAR models. Chemom Intell Lab Syst 38. Smellie A (2004) Accelerated K-means cluster-
145:22–29 ing in metric spaces. J Chem Inform Comput
33. Dimitrov S, Dimitrova G, Pavlov T, Dimitrova Sci 44:1929–1935
N, Patlewicz G, Niemela J et al (2005) Stepwise
Chapter 7
Molecular Similarity in Computational Toxicology

Matteo Floris and Stefania Olla
Abstract
The concept of chemical similarity has many applications in several fields of cheminformatics. One com-
mon use of chemical similarity measurements, based on the principle that similar molecules have similar
properties, is in the context of the read-across approach, where estimates of a specific endpoint for a chemi-
cal are obtained starting from experimental data available from highly similar compounds.
This chapter reports an implementation of chemical similarity and the analysis of multiple combina-
tions of binary fingerprints and similarity metrics in the context of the read-across technique.
This analysis demonstrates that the classical similarity measurements can be improved with a general-
izable model of similarity. The approach presented here has been implemented in two open-source soft-
ware tools for computational toxicology (CAESAR and VEGA).
Key words Chemical similarity, QSAR, Toxicity prediction, Similarity searching, Read-across
1 Introduction
Animal models have been used for a long time as classical tools for
toxicity testing. However, in vivo animal tests are mainly limited by
ethical considerations and financial burden. Therefore, when pos-
sible, alternative testing methods for estimating the toxicity of
chemicals-such as computational methods-are preferred. In silico
toxicology refers to the class of computational methods used to
predict the toxicity of chemicals; such methods are intended not to
replace but to complement classical toxicity tests in different appli-
cations, such as toxicity prediction and prioritization of chemicals,
and in drug design for minimizing late-stage failures. Computational
methods have the unique advantage of returning an early toxicity
estimate even before the synthesis of chemicals [1].
Such non-testing data can be generated by three main
approaches: (1) grouping approaches, which include read-across
and chemical category formation; (2) structure–activity relation-
ship (SAR) and quantitative SAR (QSAR, a term used in the fol-
lowing text to imply both); and (3) expert systems. The development
171
172 Matteo Floris and Stefania Olla
and application of all kinds of non-testing methods is based on the

similarity principle, i.e., the hypothesis that similar compounds
should have similar biological activities or properties.
Methods that provide a measure of chemical similarity are
strongly popular in many fields of cheminformatics applications, as
they allow to predict the molecular behavior and the fate of struc-
turally related compounds. One common application of chemical
similarity measurements is the read-across approach, where an esti-
mation of a specific endpoint for a chemical is provided using
experimental data available from highly similar compounds. Read-
across involves the use of relevant information from analogous
substances to predict properties for another substance under con-
sideration [2].
The underlying assumption is that substances sharing physico-
chemical, toxicological, and eco-toxicological properties are likely
to be structurally and chemically related or retain common fea-
tures. Structural similarity is thus a necessary prerequisite for any
grouping and read-across approach. Ideally, similarity among dif-
ferent substances may be due to a number of factors: (1) the pres-
ence of common functional groups; (2) the existence of common
precursors or likelihood of common breakdown products through
physical and/or mutual degradation products (following a process
of bio-transformation); (3) or, eventually, a well-defined pattern in
the changing of the endpoint across the group of substances.
However, a still open and debated issue behind such different
approaches turns precisely around to the concept of similarity. Indeed,
there is no way to define in an unambiguous fashion (and, conse-
quently, with an unequivocal or universal algorithm) the degree of
similarity of two chemical entities. In fact, two compounds can be
perceived as more or less similar with respect to the chemical features
taken into attention or chosen as a priority. For instance, if a similarity
measurement is needed for QSAR purposes, the same category of
molecular descriptors used for developing the QSAR model should
have higher priority. In other situations, the similarity could rely on
approaches based on a broader description of the chemical structure
and on the experience and sensibility of the chemist.
Another point leading to different approaches is related to
practical applications of the similarity measurement. An excess of
complexity should be avoided to obtain algorithms that can be
calculated in a reasonable time and whose results can be easily
interpreted.
Broadly speaking, we can describe three main components of a
similarity measure: the method used for the description of the
structure (a numerical description of the chemical structure), the
weighting scheme (used to prioritize the contributions of each part
of the descriptors schema), and the similarity measure (the
quantitative measure of similarity between two numerical

representations).
Molecular Similarity in Computational Toxicology 173
1.1 Description Molecular descriptors can be calculated with both commercial

of the Chemical and open access tools. Drug discovery software packages, such
Structure by Molecular as MOE [3], Schroedinger [4], and Discovery Studio [5],
Descriptors implement their own molecular description calculation libraries.
A specialized software for descriptor calculation implements
more than 5000 different algorithms (Dragon 6) [6].
ADRIANA.Code [7] and its successor CORINA Symphony
provide a workflow for calculation of 3D descriptors. Many
freeware tools are available: many of them are based on the
Chemistry Development Kit (CDK) [8], an open source Java
library for cheminformatics. Examples are PaDEL-Descriptor, a
software that currently calculates 1875 descriptors [9];
ChemDes which has the capability of computing 3679 molecu-
lar descriptors and 59 types of molecular fingerprints [10];
Molgen [11]; Molconn-Z [12].
1.2 Description Several binary fingerprints are available. A paper by Cereto-

of the Chemical Massagué et al. concluded in 2014 that most of the commonly
Structure by Molecular used and popular fingerprints have highly similar performances,
Fingerprints intertarget differences for the same fingerprint being usually
and Structural Keys greater than the differences for different fingerprints for the same
target molecule [13]. Among others, fragment-based Daylight
[14] and Tripos UNITY 2D fingerprints [15] are some of the
best known commercial examples. The fingerprints available in
the CDK [16, 17] are free and open-source implementations of
different algorithms.
1.3 Similarity Similarity coefficients are used to calculate the similarities between
Coefficients a reference and a target fingerprint. Several binary similarity coef-
ficients are available; a comprehensive and up-to-date list has been
summarized by Todeschini et al. [18]. Remarkably, Todeschini
listed 51 similarity coefficients for binary variables (or 44 different
coefficients) extracted from the literature and compared using
both simulated and real data. Todeschini identified also seven simi-
larity coefficients, five pairs and a triplet of coefficients, which are
completely correlated.
Similarity-based methods have been successfully applied to
solve various cheminformatics related issues, such as predictions
of biological targets [19], likelihood of therapeutic indications
[20] or side-effects estimates [21]. In particular, state-of-the-art
machine learning approaches-such as k-nearest neighbors (kNN),
naïve Bayes models, support vector machines (SVM), random
forests (RF), or ensembles of different methods-can implement
the toxicity endpoints predictions on the basis of different chemi-
cal similarity algorithms. This concept has been applied to predic-
tions of various toxicological endpoints in different research areas,
such as in oral toxicity [22] and persistence in the sediment com-
partment [23].
Examples of fast and successful methods for the prediction of

different outcomes have been developed in the context of Tox21
Data Challenges in 2014 [24]. One of these was based on the
combination of a simple molecular similarity calculation with a
naïve Bayes machine learning algorithm [25]. Three different two-
dimensional (2D) molecular representation methods as well as
their combination were compared and the prediction methods
were optimized individually for every target.
1.4 Applications One common application of chemical similarity measurements is

in Read-Across the read-across approach, where an estimation of a specific end-
point for a chemical is provided using experimental data available
from highly similar compounds. A way to automatize this process
starts from blending fingerprints with nonbinary structural keys
based on constitutional molecular descriptors. The basic idea is
that such a combination can help to overcome the drawbacks of a
plain fingerprint approach and thus to increase the accuracy of
similarity measurements, yet avoiding an excessive calculation
complexity. In this respect, an integrated similarity index resulting
from the weighted combination of a fingerprint array and three
structural keys based on molecular descriptors has been developed
and is here presented [26]. The performances of different finger-
prints, different similarity coefficients, and different weighting
schemes for the elements contained in the final index have been
evaluated through the implementation of a batch implementation
of the read-across approach on two distinct datasets, in order to
find an acceptable criterion of choice of elements and weighting
scheme for the similarity index in a generic application. Such com-
parative analysis of the performances on the two datasets of all the
possible combinations of 9 fingerprint implementations and 44
similarity coefficients, followed by an exploration of a reasonable
subset of all the possible weighting schemes for the fingerprint
and the structural keys based on molecular descriptors, provided
good performances on both datasets and has finally been chosen
to build the similarity index, actually implemented in the VEGA
platform (an open-source online platform providing several QSAR
models) [27].
2 Material and Methods
2.1 Fingerprints The performance of nine different fingerprint algorithms has been
evaluated. All of them are implemented in the CDK libraries. While
they fall under the generic definition of fingerprints, some of them
can be classified as structural keys and not as hashing-based finger-
prints. More specifically, the fingerprints here considered are the
following:
(a) Default Fingerprints (as defined by Daylight) [14].

(b) Extended Fingerprints (same as Default, but with additional
bits that take into account ring features).
(c) Graph-only fingerprints (same as Default, but do not take
bond orders into account).
(d) Hybridization fingerprints (same as Default, but do not per-
form aromaticity perception).
(e) E-State fragments (79-bit fingerprints described by Kier and
Hall [28].
(f) Klekota–Roth fingerprints (set of 4860 chemical substructures
enriched for biological activity [29].
(g) MACCS keys (structural key made of a set of 166 bits [30]).
(h) PubChem fingerprints (structural key made of 881 keys [31]).
(i) Substructure fingerprints (structural key made of 307 bits
[32]).
2.2 Molecular Three structural keys fingerprints made of molecular descriptors

Descriptors-Based related to constitutional issues have been implemented. The
Structural Keys hypothesis that led to these keys was to test if such information
could fill the information gap of hashing-based fingerprints. These
keys are made of molecular descriptors, calculated by an in-house
JAVA software module, based on the CDK libraries; for the defini-
tion of the descriptors the commercial software Dragon [6] has
been taken as reference.
The three keys are: (1) Constitutional descriptors (CD): this
key is made of 35 constitutional descriptors; (2) Heteroatoms
descriptors (HD): this key is made of 11 counters for different
types of heteroatoms. These descriptors are a subset of the consti-
tutional descriptors. This schema has been implemented on the
basis of the observation that often the generic idea of chemical
similarity is strongly influenced by small differences in the number
and type of heteroatoms, i.e., molecules with several similar fea-
tures (molecular weight, number and type of rings, bonds, etc.)
can often behave with remarkably differences due to the presence/
absence of some heteroatoms. (3) Functional Groups (FG): this
key is made of 154 functional groups, as defined in Dragon.
2.3 Similarity Two sets of similarity coefficients were implemented and tested
Coefficients respectively with the chosen fingerprints (binary coefficients) and
descriptors based keys (nonbinary coefficients). The chosen binary
coefficients are 44, coming from the work of Todeschini et al. [18].
The chosen nonbinary coefficients are six, from the work of
Holliday [33]. All the coefficients have been implemented in an
in-house JAVA software module.
2.4 Similarity Index In order to combine the fingerprint with the descriptors-based
keys, a generic scheme for the similarity index SI is defined as
follows:

L L L
SI ( A,B ) = éë Sb ( FPa,FPb ) ùû Wfp ´ éë Snb (CDa,CDb ) ùû Wcd ´ éë Snb ( HDa,HDb ) ùû Whd
´ éë Snb ( FGa,FGb ) ùû ´ Wfg

where A and B are two molecules to be compared, while FPa, CDa,
HDa, FGa, FPb, CDb, HDb, and FGb are, respectively, the
Fingerprint, the Constitutional Descriptors, the Heteroatom
Descriptors, and the Functional Groups keys as defined before,
calculated for the two molecules A and B; Sb (Xa,Xb) is the result
of the application of a binary similarity coefficient to two finger-
prints Xa and Xb, where the resulting values are in the interval
[0,1]; Snb (Xa,Xb) is the result of the application of a nonbinary
similarity coefficient to two descriptors based keys Xa and Xb,
where the resulting values are in the interval [0,1]; Wfp, Wcd,
Whd, and Wfg are the relative weights of the four contributions,
under the condition that their sum is equal to 1.
2.5 Datasets Two publicly available datasets where extracted from the VEGA
and Read-Across project [27].
Model The bioconcentration factor in fish (BCF) dataset composed by
473 compounds with corresponding experimental BCF values, and
the water–octanol partition coefficient (logP) dataset composed by
10,005 compounds, each with an experimental logP value. The
choice of these two different datasets is aimed at finding an optimal
setting for the SI on different kinds of data, thus implementing a
“generic” idea of chemical similarity. Indeed, this experiment is
based on two endpoints with relevance for toxicity (BCF) and on a
physical-chemical property (logP) with several applications, with
two datasets with marked different size (BCF: 860 molecules; logP
10,005 molecules).
For the purpose of testing the performances of the proposed
Similarity Index with different settings, an in-house JAVA imple-
mentation of a simple read-across based prediction model has been
used, where BCF or logP is predicted for a query compound by
finding the three most similar compounds of the dataset according
to the SI, then calculating the mean of their three experimental
values, weighted by their SI values.
In this procedure, the leave-one-out strategy was adopted for
cross-validation. Iteratively, one molecule at a time was left out of
the dataset to be predicted using the read-across approach on the
remaining molecules.
Finally, as the above described model approach is analogous to
a regression model, the values of the coefficient of determination
(R) and of the root mean square error (RMSE) on all the predic-
tions of the dataset were calculated to quantify the quality of the
model, that is a measure of the goodness of the SI.
2.6 Evaluation A combinatorial strategy was applied to test all the possible permu-
Process tation (N = 400) of different settings (similarity coefficient, binary
fingerprints, nonbinary descriptors, weighting scheme). The best
combinations of these settings were the ones identified on the basis
of R and RMSE.
A second analysis was then performed using the selected cou-
ple of fingerprint/coefficient and a set of combinations of the
weights for the SI contributions and of nonbinary similarity coef-
ficients for the descriptors keys. The batch process generated a
total of ca. 7200 combinations of weights and coefficients.
3 Notes
The fingerprints found in the ten best solutions are the Extended
Fingerprints, PubChem structural keys, and Default Fingerprint.
For the fingerprints, it is not surprising that the Extended yield
better results than the Default, as Extended are the same as default
but with some extra bits encoding information about rings. The
best coefficients found in combination with the fingerprints are 37
(Maxwell–Pilliner), 34 (Cohen), 18 (Rogot–Goldberg), 42 (CT4),
13 (Sokal–Sneath), and 1 (simple matching).
In the second step, having identified as best solutions the
Extended fingerprints and the coefficient no. 37 (Maxwell–
Pilliner), about 7200 combinations of weights and nonbinary simi-
larity coefficients have been analysed. All the 10 best solutions use
the coefficient no. 3 (Bray–Curtis) for the measurement of the
nonbinary keys of descriptors. Subsequently, it can be easily
observed that all the 10 solutions have a similar distribution of the
weight values. In the best solution the fingerprints block represent
the most important contribution (weight of 0.4), followed by the
Constitutional Descriptors block (0.35), the Functional Groups
Descriptors block (0.15), and the Heteroatoms Descriptors block
(0.1). This result can be interpreted as follows: the SI is mainly
constituted by the classical fingerprint-based comparison, strongly
corrected with some constitutional information like number (and
type) of atoms and number (and type) of bonds; this part of the SI
could be considered as the core contribution to generalizability of
the SI; furthermore, a smaller contribution of functional and het-
eroatoms descriptors is sufficient to extend the information embed-
ded in the fingerprint and constitutional descriptor blocks; this
block essentially explains the “fine chemical differences” within the
dataset.
The computation of similarities between chemical compounds

is usually limited to a simple combination of a common binary
representations of chemical structures and a single similarity coef-
ficient. This experiment demonstrated how a higher accuracy in
measures of chemical similarity can be achieved by combining fin-
gerprints with nonbinary structural keys based on constitutional
molecular descriptors. This combination can overcome the draw-
backs of a simple traditional fingerprint approach by introducing a
similarity index combining one fingerprint and three structural
keys based on molecular descriptors of different weights. A combi-
natorial process was designed to evaluate the performances of dif-
ferent settings (fingerprints, similarity coefficients, and different
weighting schemes) for the identification of the best combination
of elements in the context of two heterogeneous datasets.
References
1. Madan AK, Bajaj S, Dureja H (2013) 17. Steinbeck C, Hoppe C, Kuhn S, Floris M,
Classification models for safe drug molecules. Guha R, Willighagen EL (2006) Recent devel-
In: Reisfeld B, Mayeno AN (eds) Computational opments of the chemistry development kit
toxicology, vol 930. Humana Press, New York, (CDK)–an open-source java library for chemo-
pp 99–124 and bioinformatics. Curr Pharm Des
2. Read-Across Assessment Framework (RAAF), 12(17):2111–2120
accessed Sept 2017 18. Todeschini R, Consonni V, Xiang H, Holliday
3. http://www.chemcomp.com/ J, Buscema M, Willett P (2012) Similarity coef-
4. https://www.schrodinger.com/ ficients for binary chemoinformatics data: over-
view and extended comparison using simulated
5. h t t p : / / a c c e l r y s . c o m / p r o d u c t s / and real data sets. J Chem Inf Model
discovery-studio/ 52:2884–2901
6. http://www.talete.mi.it/, accessed Sept 2017 19. Campillos M, Kuhn M, Gavin A-C, Jensen LJ,
7. h t t p s : / / w w w. m n - a m . c o m / p r o d u c t s / Bork P (2008) Drug target identification using
adrianacode side-effect similarity. Science 321:263–266
8. h t t p s : / / w w w . n c b i . n l m . n i h . g o v / 20. Nickel J, Gohlke B-O, Erehman J, Banerjee P,
pubmed/29086040,cdk.sf.net Rong WW, Goede A, Dunkel M, Preissner R
9. h t t p : / / w w w . y a p c w s o f t . c o m / d d / (2014) SuperPred: update on drug classifica-
padeldescriptor/ tion and target prediction. Nucleic Acids Res
10. h t t p s : / / w w w . n c b i . n l m . n i h . g o v / 42:W26–W31
pubmed/26664458 21. Lounkine E, Keiser MJ, Whitebread S,
11. http://www.molgen.de/ Mikhailov S, Hamon J, Jenkins JL, Lavan P,
12. http://www.edusoft-lc.com/molconn/ Weber E, Doak AK, Côté S, Shoichet BK,
Urban L (2012) Large-scale prediction and
13. Cereto-Massagué A, Ojeda MJ, Valls C, Mulero testing of drug activity on side-effect targets.
M, Garcia-Vallvé S, Pujadas G (2015) Nature 486:361–367
Molecular fingerprint similarity search in vir-
tual screening. Methods 71:58–63 22. Drwal MN, Banerjee P, Dunkel M, Wettig MR,
Preissner R (2014) ProTox: a web server for
14. Daylight Chemical Information Systems Inc., the in silico prediction of rodent oral toxicity.
http://www.daylight.com/ Nucleic Acids Res 42:W53–W58
15. Tripos Inc., http://www.tripos.com
23. Manganaro A, Pizzo F, Lombardo A,
16. Steinbeck C, Han Y, Kuhn S, Horlacher O, Pogliaghi A, Benfenati E (2016) Predicting
Luttmann E, Willighagen E (2003) The chem- persistence in the sediment compartment with
istry development kit (CDK): an open-source a new automatic software based on the k-near-
java library for chemo- and bioinformatics. est neighbor (k-NN) algorithm. Chemosphere
J Chem Inf Comput Sci 43:493–500 144:1624–1630
24. https://tripod.nih.gov/tox21/challenge/ 30. MACCS Structural Keys. CA (USA): Symyx

25. Drwal MN, Siramshetty VB, Banerjee P, Goede Software S. R
A, Preissner R, Dunkel M (2015) Molecular 31. National Center for Biotechnology
similarity-based predictions of the Tox21 Information: “PubChem Substructure
screening outcome. Front Environ Sci 3:1–9 Fingerprint v1.3.” PubChem Data Specification
26. Floris M, Manganaro A, Nicolotti O, Medda 2009 (ftp://ftp.ncbi.nlm.nih.gov/pubchem/
R, Mangiatordi GF, Benfenati E (2014) A gen- specifications/pubchem_fingerprints.txt)
eralizable definition of chemical similarity for 32. O r g . o p e n s c i e n c e . c d k . f i n g e r p r i n t :
read-across. J Chem 6:39 SubstructureFingerprinter. http://pele.farm-
27. VEGA project website: http://www.vega-qsar. bio.uu.se/nightly/api/org/openscience/
eu/ cdk/fingerprint/SubstructureFingerprinter.
28. Hall LH, Kier LB (1995) Electrotopological html
state indices for atom types: a novel combina- 33. Al Khalifa A, Haranczyk M, Holliday J (2009)
tion of electronic, topological, and valence Comparison of nonbinary similarity coefficients
state information. J Chem Inf Comput Sci for similarity searching, clustering and com-
35:1039–1045 pound selection. J Chem Inf Model
29. Klekota J, Roth FP (2008) Chemical substruc- 49:1193–1201
tures that enrich for biological activity.
Bioinformatics 24:2518–2525
Chapter 8
Molecular Docking for Predictive Toxicology

Daniela Trisciuzzi, Domenico Alberga, Francesco Leonetti,
Ettore Novellino, Orazio Nicolotti, and Giuseppe F. Mangiatordi
Abstract
Molecular docking is an in silico method widely applied in drug discovery programs to predict the binding
mode of a given molecule interacting with a specific biological target. This computational technique is
today emerging also in the field of predictive toxicology for regulatory purposes, being for instance suc-
cessfully applied to develop classification models for the prediction of the endocrine disruptor potential of
chemicals. Herein, we describe the protocol for adapting molecular docking to the purposes of predictive
toxicology.
Key words Molecular docking, Classification model, Predictive toxicology, Endocrine potential,
Applicability domain
1 Introduction
Molecular docking is a computational strategy whose primary aims

are (1) predicting a preferred orientation (pose) of a given molecule
with respect to a target macromolecule and (2) scoring the strength
of the established binding interactions [1]. Widely applied by both
academia and industry to drug discovery programs [2–5], this tech-
nique is today emerging in the field of predictive toxicology for
regulatory purposes, an outstanding research area whose main aim
is that of weighing the risk-benefit ratio of chemicals [6–10]. This
issue can be properly addressed by employing molecular docking,
provided that the following conditions are met:(1) the endpoint of
interest (e.g., androgenic potential) is causatively related to the
occurrence of interactions between a given chemical and one (or
more) protein targets (e.g., androgen receptor, AR); (2) 3D infor-
mation of the protein target is available (e.g., AR X-ray solved
structures free released from the Protein Data Bank (PDB)); (3) a
training set (TS) comprising high quality experimental data and
curated structures can be employed (e.g., dataset with trustworthy
androgenic experimental data). Herein we describe the necessary
181
182 Daniela Trisciuzzi et al.
steps to appropriately exploit molecular docking for deriving highly

predictive classification models [11], paying attention to (1) pre-
cautions to be taken in order to fairly adapt this methodology, well
known to medicinal chemists, to the needs of predictive toxicology
and (2) the importance of implementing a suitable applicability
domain (AD) [12] to border the interpolative chemical space from
which obtaining reliable predictions. As a first phase, the available
TS must be properly curated. Next, tautomers and ionization states
at a physiological pH are generated for each TS compound. In par-
allel, the X-ray solved protein structures taken from the PDB should
be prepared for docking simulations. In the case of ligand-protein
complexes, docking parameters can be set by optimizing the over-
lap between the obtained docking pose and the experimentally
solved conformation of the cognate ligand, that is, by minimizing
the root mean square deviation (RMSD) values. Nevertheless, these
procedures are the same adopted for drug discovery programs.
Major differences are instead in the analysis of the docking results
that is merely qualitative in the case of drug discovery mostly
addressed to the lead optimization through a trial and error
approach while is rather quantitative in the case of predictive toxi-
cology whose main concern is to keep as low as possible the number
of false negatives by employing classification models to discern safe
from unsafe chemicals. Having said that, it becomes clear that dock-
ing score represents the basis to designate harmful or alternatively
harmless chemicals. Basically, a docking classification model is based
on the rank of the docking scores computed for each considered
target structure. To quantify right and wrong predictions, statistical
confusion matrixes are used from which descend several descriptive
statistical parameters such as sensitivity (SE), negative predictive
values (NPV), negative likelihood ratio (−LR), and the balanced
classification rate (BCR). Once these parameters have been calcu-
lated, the most promising models must be further challenged on ad
hoc external validation sets (VS). The herein described protocol has
been successfully applied to assess the estrogenic and androgenic
potential of chemicals within the two initiatives supervised by the
US Environmental Protection Agency (US EPA), the Collaborative
Estrogen Receptor Activity Prediction Project (CERAPP) [13] and
the Collaborative Modeling Project for Androgen Receptor Activity
(CoMPARA) [14], respectively. More specifically, all the examples
in the following refer to a recent study [15] aimed at developing
highly predictive docking based-classification models of the andro-
genic potential of chemicals, within the CoMPARA consortium.
2 Materials
2.1 Training Set 1. Retrieve a suitable training data set. We used a curated col-
lection of 1689 chemicals (hereafter referred to as
EPA-DB) having high quality experimental data (androgenic
Molecular Docking for Predictive Toxicology 183
potential) provided by the US EPA. All TS compounds are

available as SMILES in the Supporting Information
Material of reference [15].
2. Make sure 3D information of all the chemicals is available in
the TS (see Note 1).
3. Assess the toxicological data quality. Three key aspects are nor-
mally taken into account in order to assess data quality: ade-
quacy, reliability, and relevance [16]. We employed a TS
provided with chemicals whose AR bioactivity was quantified
based on the results of 11 AR-related in vitro assays, conducted
in high-throughput screening (HTS) measuring the androgen-
related activity at multiple points along the AR signaling path-
way [17]. In addition, avoid, if possible, TS provided with
unbalanced data, comprising an uneven distribution of com-
pounds between active (hazard) and inactive (safe) chemicals.
If necessary, consider ad hoc statistical metrics independent
from the TS data distribution in order to correctly assess model
quality (see Subheading 3.8). Specifically, we employed a TS
containing 205 (i.e., approximately 12%) hazard (compounds
with androgenic potential) and 1484 (i.e., approximately 88%)
safe chemicals (compounds without androgenic potential).
2.2 Validation Set(s) 1. Build appropriate validation sets (VS) to challenge model(s)
performance. Data can be extracted from commercial and pub-
lic sources. An example is given by the DUD-E (Directory of
Useful Decoys Enhanced) database [18], a web accessible and
free of charge repository comprising decoys and active chemi-
cals against 102 targets, included AR. In order to properly vali-
date a classification model, it is advisable to build at least three
equally sized VS (namely, VS1, VS2, and VS3) comprising the
same number of active (hazard) compounds. Accordingly, we
employed three VS obtained by extracting data from DUD-E,
each comprising 2590 decoys randomly selected and the same
pool of 259 compounds (10% of the set size) experimentally
proved to bind AR (the interested reader if refereed to http://
dude.docking.org/targets/ and for additional details).
2. Make sure 3D information of all the chemicals is available in
the VS (see Note 1).
3. Identify and remove duplicate compounds (same entry in both
the VS and TS).
2.3 Protein Target 1. Retrieve the structural information of the protein target (e.g.,
Coordinates AR X-ray crystal solved structures). Nowadays, several public
repositories containing all the information of experimentally
determined structures of proteins are available (see Note 2).
It is advisable to select, if possible, protein structure(s) pro-
vided with high quality data having a resolution better than
2.0 Å and, more importantly, a structurally unambiguous
Fig. 1 Example of slider chart relative to an X-ray crystal structure provided with high-quality indicators.
Reproduced from RCSB PDB (www.rcsb.org) of PDB ID 1CBS with the permission from (Kleywegt GJ, Bergfors
T, Senn H, Le Motte P, Gsell B, Shudo K, Jones TA (1994) Crystal structures of cellular retinoic acid binding
proteins I and II in complex with all-trans-retinoic acid and a synthetic retinoid. Struct Lond Engl 1993
2:1241–1258)
binding site (e.g., no missing side chains). The chart reported

in the Fig. 1, for instance, provides at first glimpse a global
assessment of the quality of a structure model and highlights
specific concerns (see Note 3). It is available for each PDB
entry and is associated to the wwPDB Validation Report
(https://files.rcsb.org/pub/pdb/validation_reports/).
Specifically, we retrieved 9 AR high-quality crystal structures
from PDB having resolution ranging from 1.55 Å to 2.10 Å.
Additionally, other metrics (e.g., the B factor and free R-factor) can
be used to further assess crystallographic data quality (see Note 4).
2.4 Docking 1. Select the docking software to be employed. For example, free
Software automated Autodock suite [19] as well as the licensed Grid-
based Ligand Docking with Energetics (Glide) [20] or the
Genetic Optimization for Ligand Docking (Gold) [21] are
popular software for drug discovery recently exploited also in
the toxicological context (see ref. [11, 22] and Note 5). In
particular, we used Gold v.5.2.
2. For the sake of time, docking software employing an efficient
parallel implementation should be preferred.
2.5 Scripts for Data 1. Retrieve/write appropriate computer scripts capable of autom-

Analysis atizing data analysis. A large variety of open source program-
ming languages is today available (e.g., Python, R, Perl), each
one providing wide sets of packages that can be adapted to the
specific analysis. In order to properly evaluate model perfor-
mance, we employed in-house Python scripts able to compute
relevant statistical parameters based on the obtained docking
score ranking (see Note 6).
2.6 Computing 1. In order to efficiently process large chemical libraries, it is

Facilities advisable to employ parallel calculation techniques and high-
performance computing (HPC) facilities.
3 Methods
3.1 TS Data Handling Before carrying out docking simulations, the TS must be properly
handled following two different phases.
3.1.1 Data Curation 1. Desalt structures and remove inorganic counterions.

2. Check for the presence of carbon atoms and remove inorganic
compounds.
3. Identify/remove structural duplicates using InChI (IUPAC
International Chemical Identifier) codes, unequivocal
identifiers.
4. Remove chemicals containing metals, which often cause prob-
lems in docking simulations.
3.1.2 Data Filtration 1. Remove compounds with undesirable properties. In particular,

oversized and too flexible molecules with a high number of
rotatable bonds should be removed, being the exploration of
their conformational space too demanding [23].
3.2 Experimental 1. Set a given score as experimental threshold to discriminate

Threshold Setting hazard from safe chemicals. For instance, EPA-DB was pro-
vided by US EPA with a binary (0/1) value assigned to each
compound as function of biological response (nonandrogenic/
androgenic) [17].
3.3 Ligand 1. Generate all the tautomers and ionization states at a pH value
Preparation of 7.0 ± 2.0. We processed all the chemicals belonging to TS
and VS using LigPrep (Schrodinger Suite 2016-3) [24].
3.4 Target Structures 1. Pretreat the selected protein target structures adding missing
Preparation hydrogen atoms, filling missing side chains and loops (see Note
7) and assigning bond and metals orders.
2. Delete water molecules distant from HET groups (see Note 8)
and retain water molecules in binding site having a functional
role.
3. Optimize the hydrogen bond network by reorienting hydroxyl
and thiol groups, amide groups of asparagine and glutamine,
and the imidazole ring in histidine.
4. Predict optimal protonation states of histidine, aspartic acid
and glutamic acid residues at a pH value of 7.0 ± 2.0.
5. Determine tautomeric states of histidine at a pH value of

7.0 ± 2.0.
We used the Protein Preparation Wizard tool available from
Schrodinger suite [25] to refine all the nine AR crystal structures
retrieved from PDB.
3.5 Grid Generation 1. Generate a grid defining the binding site of the target proteins.
According to the selected docking software (see Subheading
2.4), the grid can be set as a sphere (Gold [21]) or as a cube
(Glide [20]) centered on the centroid of the cognate ligand or
of some selected residues. Usually, it is possible to specify the
resolution of the grid as the number of points in each dimen-
sion of the 3D space, or as the spacing between points
(Autodock [19]). Grid size should be carefully set in order to
minimize, in a reasonable computational time, the number of
undocked compounds. In our study, we used a spherical grid
having a radius of 10 Å centered on the center of mass of the
cognate ligands for all the docking simulations.
3.6 Docking Protocol 1. Challenge the reliability of the simulation protocol by per-
Calibration forming preliminary docking simulations of the cognate
ligands, if present. In particular, for each selected target struc-
ture (see Subheading 2.3), compare the Cartesian coordinates
of the pose of the cognate ligands resulting from docking sim-
ulation with the experimental ones by computing RMSD val-
ues (see Note 9). For the 9 AR crystal structures considered,
we computed RMSD values ranging from 0.436 Å to 2.069 Å.
3.7 Docking of TS 1. Subject the TS to molecular docking simulations. In particular,

and Models Derivation use the prepared ligands (Subheading 3.3) and pretreated pro-
tein structures (Subheading 3.4) and apply the protocol result-
ing from the calibration step (Subheading 3.6). Dock the
entire TS into a grid defining the binding site (Subheading
3.5) of the selected target structures, thus obtaining, for each
of which, a docking-score based ranking.
2. Evaluate docking performance by considering, for each
docking-score based ranking, the receiver operating character-
istic (ROC) curve and calculating the corresponding AUC and
Enrichment Factor (EF) at a given percentage of ranked TS.
Notice that the coordinates to graph ROC curves can be
derived based on a typical confusion matrix (see Note 10). AUC
values, ranging from 0 to 1, provide an idea of the performances
in classification. AUC = 1 indicates a perfect classifier while
AUC = 0.5 a random one (see Note 11). We docked EPA-DB
on 9 AR crystals and computed AUC values ranging from 0.70
to 0.76.
3. Define the binding classes by setting appropriate thresholds on

the docking-score based ranking. For instance, based on the
confusion matrix, we set three sensitivity (SE) values equal to
0.25, 0.50, and 0.75, in order to define, for each target struc-
ture, four probability classes as follows: SE ≤0.25 (hazard mol-
ecules), 0.25 < SE ≤ 0.50 (warning molecules),
0.25 < SE ≤ 0.75 (suspicious molecules), and SE >0.75 (safe
molecules) [15]. Hence, docking scores whose ranks are near-
est to such SE thresholds were identified as values for chemical
classification. To do that, we employed an in-house developed
Python script.
3.8 Model Selection 1. Measure the goodness of the derived classification models by
computing several statistical parameters. First, we calculated
the Negative predictive values (NPV) and Positive predictive
values (PPV) (see Note 12) for each SE threshold. Notably, in
the case of unbalanced TS, ad hoc statistical metrics are recom-
mended (see Note 13 for methodological details). Accordingly,
we computed the positive–negative likelihood ratio (+/−LR)
and balance classification rate (BCR) to select the best per-
forming docking-based classification model. It should be noted
that the performance of a classification model conceived for
toxicological purposes chiefly depends on its ability to mini-
mize the rate of FNs (e.g., low −LR at the SE threshold equals
to 0.75). We computed, depending on the employed target
structure, −LR values ranging from 0.38 to 0.48, thus indicat-
ing, for all the developed models, an optimal capability to min-
imize FNs. Based on the computed statistical parameters (see
Fig. 2), we selected one AR structure (PDB code 2PNU [26])
as that providing the best performing classification model.
3.9 AD 1. Implement an AD definition for the selected classification

Implementation model. In this respect, different approaches can be employed
based on physicochemical, structural, and response domains
(the interested reader is referred to the references [12, 27] for
a survey of the most used AD implementation strategies).
In our study, we define the AD based on a hierarchical two-
steps approach (see Note 14). In particular, we applied a range-
based method, named bounding-box, and a convex-hull strategy.
The former is based on the calculation of molecular descriptors: a
chemical is considered within the AD if all the computed molecular
descriptors fall in the range limited by the maximum and minimum
values computed for the entire TS. The latter is a geometric-based
approach: the AD is defined by the smallest convex area whose
borders design a polygon containing the TS compounds. Moreover,
in order to avoid the presence of empty regions, the polygon can
be restricted to an inner region containing the top 95% of the EPA-
Fig. 2 Summary of the 9 classification models ranking obtained by the three statistical metrics (NPV, −LR and
BCR) computed at SE = 0.75. (NPV Negative Predictive Value, −LR Negative likelihood ratio, BCR Balanced
Classification Rate). Each model is specifically indicated as the PDB entry employed for docking simulation.
Reprinted (adapted) with permission from (Trisciuzzi D, Alberga D, Mansouri K, Judson RS, Novellino E,
Mangiatordi GF, Nicolotti O (2017) Predictive structure-based toxicology approaches to assess the androgenic
potential of chemicals. J Chem Inf Model. doi: https://doi.org/10.1021/acs.jcim.7b00420). Copyright 2017
American Chemical Society
ARDB chemicals on the basis of their closeness to the EPA-ARDB

centroid. In particular, the interpolation space is defined by the
Cartesian coordinates of the top two principal components (PCs)
based on 162 descriptors (see Fig. 3).
3.10 Model 1. Assess whether or not each chemical belonging to the VS is

Validation within the model AD.
2. Dock VS compounds within the AD into the binding site of
the selected target structure (i.e., AR structure 2PNU [26])
following the same protocol previously applied to the TS for
model derivation.
3. Compute the percentage of hazard compounds within the safe
class predicted by the selected classification model and com-
pare this value to the percentage of hazard compounds in the
VS. We detected 4.42% (VS1), 4.59% (VS2) and 5.09% (VS3)
of hazard compounds in the class set at SE >0.75, thus con-
firming a high performance of the selected classification model
after the application of the two steps AD approach. Each VS,
in fact, taken as a whole, contained a number of hazard com-
pounds corresponding to a set size of 10%.
3.11 Real-Life 1. Subject the selected classification model to a further challenge

Prediction Power by predicting reference hazard compounds (i.e., well-known
Evaluation androgenic substances). In our study, we chose 12 substances:
6 drugs and 6 chemicals widely employed for industrial and
commercial uses. Importantly, all of them resulted within the
model AD and were accurately classified based on a rationale at
molecular level.
Fig. 3 Projection of both TS and VS into the top two PCs obtained from the 162 descriptors computed for each
compound of the EPA-DB. The outer polygon (dashed line) takes into accounts all the chemicals in the EPA-DB
(black circles), while the inner polygon (solid line) retains the 95% of them. Chemicals of the VS (red circles)
outside the inner 95% polygon are flagged as outside AD. Reprinted (adapted) with permission from (Trisciuzzi
D, Alberga D, Mansouri K, Judson RS, Novellino E, Mangiatordi GF, Nicolotti O (2017) Predictive structure-based
toxicology approaches to assess the androgenic potential of chemicals. J Chem Inf Model. doi: https://doi.
org/10.1021/acs.jcim.7b00420). Copyright 2017 American Chemical Society
4 Notes
1. TS 3D information is mandatory to ensure that compounds

with chiral centers are properly treated.
2. All the information of experimentally determined crystallo-
graphic protein structures are available at Protein Data Bank
(PDB, https://www.rcsb.org/pdb/home/home.do) [28],
Worldwide Protein Data Bank (wwPDB, https://www.
wwpdb.org/) [29] and Protein Data Bank Japan (PDBj,
https://pdbj.org/) [30].
3. The slider chart compares important global quality indicators
(e.g., clash score, Ramachandran outliers, sidechain outliers)
for a given structure within the PDB repository. The interested
reader can also examine the wwPDB Validation Report associ-
ated to each PDB entry. Global percentile ranks (black vertical
lines) are computed with respect to all X-ray structures avail-
able prior to 2011, whereas the ranks of the resolution-specific
percentile (white vertical lines) are calculated with respect to a
subset of X-ray entries of the PDB archive with comparable
resolution to the target query.
4. (a) The B factor (also named temperature factor, B value, or
Debye-Waller factor) indicates the mobility of an atom and
it can be used as quality indicator of X-ray crystallographic
model. The B-factor is defined by
Bi = 8π 2U i2
where Ui2 stands for the mean square displacement of atom i from
its rest position. As U rises, the B-factor increases and the contri-
bution of the atom to the scattering decreases [31, 32]. In the
PDB file, the value in the last column represents the B-factor for
each atom in the structure. NMR models provide no information
in the temperature value fields of PDB files. B-values normally
range from 15 to 30 Å2 and are often higher than 30 Å2 for atoms
in flexible regions (e.g., atoms at side-chains are expected to exhibit
higher degree of freedom). In other words, very high B-factors
designate atoms incorrectly positioned in the model, thus indicat-
ing a local disorder due to purely thermal motions.
(b) The R-factor is a metric used to assess the quality of the
crystallographic model fitting the original X-ray diffraction
data. It is defined as follows:
∑h ,k ,l Fobs (h ,k,l ) − ∑h ,k ,l Fcalc (h ,k,l )
R=
∑h ,k ,l Fobs (h ,k,l )

where h, k, l are the reciprocal lattice points of the crystal struc-
ture. In this equation, |Fobs (h, k, l)| and |Fcalc (h, k, l)| are derived
from the intensity of a reflection measured in the diffraction pat-
tern and the intensity of the same reflection calculated from the
model, respectively [33]. Furthermore, a different version of the
R factor is the free R-factor (Rfree). If the R factor measures how
well the model predicts the whole data set that produced the
model, the Rfree evaluates how the atomic model predicts a small
“test set” (about 5–10% of observed intensities) of the diffraction
data that were not included in the building and refinement pro-
cess. Low values and small differences between the R and Rfree
factors indicate a high prediction power of the crystallographic
model [32, 33].
5. Each docking algorithm adds different penalties to the final
docking score. For instance, if a compound has been processed
(see Subheading 3.3) using Epik, a tool available from
Schrodinger Suite, ionization or tautomeric states penalties are
added to the docking score for adopting higher-energy states
[34]. Compounds without this information are not penalized
and will therefore show better scores.
6. Here an example of in-house python script used in our study to
compute, based on the obtained docking-based ranking, the
statistical parameters SE, SP, PPV, and NPV (see Fig. 4).
7. If the protein has residues in the binding site with atom types
that are mis-assigned, overlapped or having alternate positions,
the user should carefully inspect this region.
Fig. 4 Example of an in-house Python script employed to assess the performance of the obtained docking-
based classification models
8. HET groups consist in nonstandard residues (e.g., prosthetic

groups, solvent molecules, cognate ligands) and ions for which
coordinates are supplied.
9. The RMSD of atomic positions is the measure of the average
distance between the atoms of superimposed molecules. It is
calculated as follows:
1 i =1 2
RMSD = ∑ δi
N N
where δi is the distance between the atom i and either a refer-
ence structure or the mean position of the N equivalent atoms.
The RMSD is computed for the backbone heavy atoms (i.e., C, O,
N, and C-alpha) or considering only C-alpha atoms.
In the case of a high RMSD values (>2.5 Å) returned by the
performed docking protocol calibration (see Subheading 3.6), it is
advisable to modify the docking protocol (i.e., increase the accu-
racy at the expanse of the computational time) or, alternatively,
discard the corresponding target structure [35].
10. A confusion matrix includes information about experimental
and predicted matches and mismatches returned by each clas-
sification system (see Table 1) [36].
In particular, the confusion matrix takes into account the
number of:
(a) true positives (TPs), i.e., the number of experimental posi-
tives cases that are correctly identified;
(b) true negatives (TNs), i.e., the number of experimental neg-
atives cases that are classified correctly;
(c) false positives (FPs), i.e., the number of experimental nega-
tives cases that are incorrectly classified as positive;
(d) false negatives (FNs), i.e., the number of experimental posi-
tives cases that are incorrectly classified as negative.
The correctly classified proportion of binders (hazard
compounds) and nonbinders (safe compounds) represent
the Sensitivity (SE) and Specificity (SP), respectively. They
are defined as follows:
TP
SE =
TP + FN
and
TP
SP =
TN + FP
SE and SP range from 0 to 1 and are used to draw ROC curves

(which report the variation of SE with respect to 1-SP values), to
Table 1
Synoptic view of a confusion matrix
Experimental class
P N
Predicted class P True positive False positive
N False negative True negative
evaluate the performance of a classification model and to set user-

dependent thresholds.
11. The performance of a docking based-classification model can
be assessed by considering the Area Under (AUC) of the ROC
curve and the Enrichment Factor at early 1% (EF1%) [37].
The AUC value stands for the probability that a classification
model will rank a randomly chosen positive instance (e.g., a
hazard chemical) higher than a randomly chosen negative one
(e.g., a safe chemical). It provides, at first glance, an overall
evaluation of the classifier quality independent from a particu-
lar threshold. Additionally, AUC curve can also be considered
as a linear transformation of the Youden Index (J) [38]. This
index, ranging from 0 to 1, is defined as follows:
J = SE − (1 − SP )

It is computed for all points of a ROC curve and it is the maxi-
mum vertical distance. The maximum value of the J could be used
for selecting the docking protocol [39].
ROC-based EF refers to the percentage of known hazard com-
pounds found at a given percentage (e.g., 1%, 2%, 5%, or 20%) of
the ranked TS and can be calculated with the following equation:
H SCR TSTOT
EF = ×
H TOT TSSCR

where HSCR is the number of hazard compounds recovered at a
specific percentage level of the binders–nonbinders ratio of the
ranked TS, HTOT is the total number of hazard compounds for a
given protein target, TSSCR is the number of compounds screened
at a specific percentage level of the TS and TSTOT is the total num-
ber of compounds in the TS. Remarkably, EF1% depends on the
hazard–safe ratio (i.e., androgenic/nonandrogenic), and its value
should be compared to the ideal EF (EFmax) obtained by dividing
the total number of TS compounds by the total number of hazard
compounds. A small gap between EF1% and EFmax indicates a
high-performing classification model [40].
12. The goodness of the classification can be evaluated computing

PPV and NPV values, defined as follows:
TP
PPV =
TP + FP
and
TN
NPV =
TN + FN
At a given threshold of SE, PPV is related to the probability that a

chemical predicted as a hazard compound (over threshold) is actu-
ally hazard, whereas NPV is related to the probability that a chemi-
cal predicted as a safe compound (under-threshold) is actually safe.
13. To overcome the intrinsic limitation of unbalance TS, differ-
ent statistical metrics, i.e., F-measure, Matthews correlation
coefficient (MCC), the balance classification rate (BCR) or
alternatively the positive–negative likelihood ratio (+/−LR)
can be employed to evaluate model performance.
(a) The F-measure (also known as F1-score or balanced
F-score) [41], bounding by 0 and 1, is the harmonic aver-
age of PPV and SE, defined as follows:
TP TP
×
F − measure = 2 × TP + FP TP + FN
TP TP
+
TP + FP TP + FN
(b) MCC introduced by Brian W. Matthews in 1975 [42] is

widely used in machine learning as a measure of the quality
of binary classifications. MCC, ranging from −1 to 1, can
be calculated directly from the confusion matrix using the
formula:
TP × TN − FP × FN
MCC =

( TP + FP ) ( TP + FN ) ( TN + FP ) ( TN + FN )
It is generally refereed as a balanced measure and thus can be
used also when data are strongly unbalanced.
(c) BCR is a modification of the well-established correct clas-
sification rate [43]. Models showing an optimal balance
between SE and SP have higher BCR scores [44, 45]. This
statistical parameter is computed as follows:
SE + SP
BCR = × (1 − SE − SP )
2
(d) +/−LR can be calculated at each SE threshold as follows:

SE
+LR =
1 − SP
and
1 − SE
−LR =
SP
A +LR value equal to three indicates a threefold increase (with

respect to the initial condition before the classification) of the
probability that a given chemical is hazard; similarly, a −LR = 0.3,
indicates that, for an under-threshold chemical, the probability to
be hazard is equal to 3/10 with respect to that at the initial condi-
tion. Widely applied to assess the reliability of diagnostic tests, the
likelihood ratios could be adapted to assess the performance of
classification models in the toxicological field. In doing this, the
lower is the −LR at SE = 0.75, the better the performance of the
classification model.
14. It is worth to note that the approach to be used for AD imple-
mentation depends on the starting training dataset [27]. In
other words, a case-by-case evaluation is required [46].
Acknowledgments
This work was supported by FIRB [Futuro in Ricerca 2012,

RBFR12SJA8_003] and the Programma IDEA 2011. We acknowl-
edge the US Environmental Protection Agency (US-EPA) for pro-
viding us high-quality androgenic experimental data.
References
1. Kitchen DB, Decornez H, Furr JR, Bajorath inhibitors: structure-based studies and efficacy
J (2004) Docking and scoring in virtual screen- in hypertension and associated CLC-K poly-
ing for drug discovery: methods and applica- morphisms. J Hypertens 34:981–992
tions. Nat Rev Drug Discov 3:935–949 6. Nicolotti O, Benfenati E, Carotti A, Gadaleta
2. Shoichet BK (2004) Virtual screening of chem- D, Gissi A, Mangiatordi GF, Novellino E
ical libraries. Nature 432:862–865 (2014) REACH and in silico methods: an
3. Meng XY, Zhang HX, Mezei M, Cui M (2011) attractive opportunity for medicinal chemists.
Molecular docking: a powerful approach for Drug Discov Today 19:1757–1768
structure-based drug discovery. Curr Comput 7. Merlot C (2010) Computational toxicology--a
Aided Drug Des 7:146–157 tool for early safety evaluation. Drug Discov
4. Wang G, Zhu W (2016) Molecular docking for Today 15:16–22
drug discovery and development: a widely used 8. Kavlock R, Dix D (2010) Computational toxi-
approach but far from perfect. Future Med cology as implemented by the U.S. EPA: pro-
Chem 8:1707–1710 viding high throughput decision support tools
5. Liantonio A, Imbrici P, Camerino GM et al for screening and assessing chemical exposure,
(2016) Kidney CLC-K chloride channels hazard and risk. J Toxicol Environ Health B
Crit Rev 13:197–217
9. Gissi A, Mangiatordi GF, Sobański T, Netzeva 21. Jones G, Willett P, Glen RC, Leach AR, Taylor
T, Nicolotti O (2017) Non-test methods for R (1997) Development and validation of a
REACH legislation. Comprehensive Medicinal genetic algorithm for flexible docking. J Mol
Chemistry 3rd ed, Volume 1 Biol 267:727–748
10. Gissi A, Gadaleta D, Floris M, Olla S, Carotti 22. Kolšek K, Mavri J, Sollner Dolenc M, Gobec S,
A, Novellino E, Benfenati E, Nicolotti O Turk S (2014) Endocrine disruptome–an open
(2014) An alternative QSAR-based approach source prediction tool for assessing endocrine
for predicting the bioconcentration factor for disruption potential through nuclear receptor
regulatory purposes. ALTEX 31:23–36 binding. J Chem Inf Model 54:1254–1267
11. Trisciuzzi D, Alberga D, Mansouri K et al 23. Lyne PD (2002) Structure-based virtual
(2015) Docking-based classification models for screening: an overview. Drug Discov Today
exploratory toxicology studies on high-quality 7:1047–1055
estrogenic experimental data. Future Med 24. Schrödinger Release 2016–3: LigPrep,
Chem 7:1921–1936 Schrödinger, LLC, New York, NY, 2016
12. Gadaleta D, Mangiatordi GF, Catto M, Carotti 25. Schrödinger Suite 2016–3 Protein Preparation
A, Nicolotti O (2016) Applicability domain for Wizard; Epik, Schrödinger, LLC, New York,
QSAR models: where theory meets reality. Int NY, 2016; Impact, Schrödinger, LLC,
J Quant Struct-Prop Relatsh IJQSPR 1:45–63 New York, NY, 2016; Prime, Schrödinger,
13. Mansouri K, Abdelaziz A, Rybacka A et al LLC, New York, NY, 2016
(2016) CERAPP: collaborative estrogen recep- 26. Cantin L, Faucher F, Couture JF et al (2007)
tor activity prediction project. Environ Health Structural characterization of the human
Perspect 124:1023–1033 androgen receptor ligand-binding domain
14. Kamel M, Kleinstreuer N, Watt E, Harris J, complexed with EM5744, a rationally
Judson R (2017) CoMPARA: collaborative designed steroidal ligand bearing a bulky
modeling project for androgen receptor activ- chain directed toward helix 12. J Biol Chem
ity conference: SOT meeting 56th annual 282:30910–30919
meeting and ToxExpo. doi: https://doi. 27. Sahigara F, Mansouri K, Ballabio D, Mauri A,
org/10.13140/rg.2.2.16791.78241 Consonni V, Todeschini R (2012) Comparison
15. Trisciuzzi D, Alberga D, Mansouri K, Judson of different approaches to define the applicabil-
RS, Novellino E, Mangiatordi GF, Nicolotti O ity domain of QSAR models. Mol Basel Switz
(2017) Predictive structure-based toxicology 17:4791–4810
approaches to assess the androgenic potential 28. Berman HM, Westbrook J, Feng Z, Gilliland
of chemicals. J Chem Inf Model G, Bhat TN, Weissig H, Shindyalov IN, Bourne
57:2874–2884 PE (2000) The protein data bank. Nucleic
16. Klimisch HJ, Andreae M, Tillmann U (1997) Acids Res 28:235–242
A systematic approach for evaluating the qual- 29. Berman H, Henrick K, Nakamura H (2003)
ity of experimental toxicological and ecotoxi- Announcing the worldwide protein data bank.
cological data. Regul Toxicol Pharmacol RTP Nat Struct Biol 10:980
25:1–5 30. Kinjo AR, Suzuki H, Yamashita R et al (2012)
17. Kleinstreuer NC, Ceger P, Watt ED et al Protein Data Bank Japan (PDBj): maintaining
(2017) Development and validation of a com- a structural data archive and resource descrip-
putational model for androgen receptor activ- tion framework format. Nucleic Acids Res
ity. Chem Res Toxicol 30:946–964 40:D453–D460
18. Mysinger MM, Carchia M, Irwin JJ, Shoichet 31. Trueblood KN, Bürgi H-B, Burzlaff H, Dunitz
BK (2012) Directory of useful decoys, JD, Gramaccioli CM, Schulz HH, Shmueli U,
enhanced (DUD-E): better ligands and decoys Abrahams SC (1996) Atomic displacement
for better benchmarking. J Med Chem parameter nomenclature. Report of a subcom-
55:6582–6594 mittee on atomic displacement parameter
19. Morris GM, Huey R, Lindstrom W, Sanner nomenclature. Acta Crystallogr A 52:
MF, Belew RK, Goodsell DS, Olson AJ (2009) 770–781
AutoDock4 and AutoDockTools4: automated 32. Rupp B (2007) Biomolecular crystallography:
docking with selective receptor flexibility. principles, practice, and application to struc-
J Comput Chem 30:2785–2791 tural biology. Garland Science, Taylor and
20. Friesner RA, Banks JL, Murphy RB et al (2004) Francis Group, New York
Glide: a new approach for rapid, accurate dock- 33. Brünger AT (1992) Free R value: a novel statis-
ing and scoring. 1. Method and assessment of tical quantity for assessing the accuracy of crys-
docking accuracy. J Med Chem 47:1739–1749 tal structures. Nature 355:472–475
34. Greenwood JR, Calkins D, Sullivan AP, Shelley multi-objective optimization algorithm. BMC
JC (2010) Towards the comprehensive, rapid, Bioinformatics 10:58
and accurate prediction of the favorable tauto- 41. Powers DM (2011) Evaluation: from preci-
meric states of drug-like molecules in aqueous sion, recall and F-measure to ROC, informed-
solution. J Comput Aided Mol Des 24: ness, markedness and correlation. J Mach
591–604 Learn Technol 2:37–63
35. Wilantho A, Tongsima S, Jenwitheesuk E 42. Matthews BW (1975) Comparison of the pre-
(2008) Pre-docking filter for protein and ligand dicted and observed secondary structure of T4
3D structures. Bioinformation 3:189–193 phage lysozyme. Biochim Biophys Acta
36. Provost F, Kohavi R (1998) Guest editors’ 405:442–451
introduction: on applied research in machine 43. Truchon JF, Bayly CI (2007) Evaluating vir-
learning. Mach Learn 30:127–132 tual screening methods: good and bad metrics
37. Triballeau N, Acher F, Brabet I, Pin JP, for the “early recognition” problem. J Chem
Bertrand HO (2005) Virtual screening work- Inf Model 47:488–508
flow development guided by the “receiver 44. Sokolova M, Lapalme G (2009) A systematic
operating characteristic” curve approach. analysis of performance measures for classifica-
Application to high-throughput docking on tion tasks. Inf Process Manage 45:427–437
metabotropic glutamate receptor subtype 4. 45. Sánchez-Rodríguez A, Pérez-Castillo Y,
J Med Chem 48:2534–2547 Schürer SC, Nicolotti O, Mangiatordi GF,
38. Youden WJ (1950) Index for rating diagnostic Borges F, Cordeiro MNDS, Tejera E, Medina-
tests. Cancer 3:32–35 Franco JL, Cruz-Monteagudo M (2017) From
39. Schisterman EF, Perkins NJ, Liu A, Bondell H flamingo dance to (desirable) drug discovery: a
(2005) Optimal cut-point and its corresponding nature-inspired approach. Drug Discov Today
Youden index to discriminate individuals using 22:1489–1502
pooled blood samples. Epidemiology 16:73–81 46. Nembri S, Grisoni F, Consonni V, Todeschini
40. Li H, Zhang H, Zheng M, Luo J, Kang L, Liu R (2016) In silico prediction of cytochrome
X, Wang X, Jiang H (2009) An effective dock- P450-drug interaction: QSARs for CYP3A4
ing strategy for virtual screening based on and CYP2C9. Int J Mol Sci 17:914–933
Chapter 9
Criteria and Application on the Use of Nontesting

Methods within a Weight of Evidence Strategy
Anna Lombardo, Giuseppa Raitano, Domenico Gadaleta,
Abstract
Nontesting methods (NTM) proved to be a valuable resource for risk assessment of chemical substances.
Indeed, they can be particularly useful when the information provided by different sources was integrated
to increase the confidence in the final result. This integration can be sometimes difficult because different
methods can lead to conflicting results, and because a clear guideline for integrating information from dif-
ferent sources was not available in the recent past. In this chapter, we present and discuss the recently
published guideline from EFSA for integrating and weighting evidence for scientific assessment. Moreover,
a practical example on the application of these integration principles on evidence from different in silico
models was shown for the assessment of bioconcentration factor (BCF). This example represents a dem-
onstration of the suitability and effectiveness of in silico methods for risk assessment, as well as a practical
guide to end-users to perform similar analyses on likely hazardous chemicals.
Key words Nontesting methods, Weight of evidence, BCF, QSAR models, Read-across
1 Introduction
We will dissert the use of weight of evidence in the case of nontest-

ing methods (NTM). NTM include in silico models and read-
across. In this case, the use of weight of evidence is highly
recommended, to increase the confidence on the results and mini-
mize the chance of wrong results, based on a single approach
(ECHA 2016). The drawback of using multiple tools is that there
may be conflicting results. The issue of conflicting results in the
case of QSAR models [1] and of read-across [2] has been discussed
[3]. The two main errors are the following: (1) renouncing to take
a decision, because only the predicted value was taken into consid-
eration and not the supporting information provided; (2) limiting
the exploration to one single tool. What we will discuss below
should prevent these errors.
199
200 Anna Lombardo et al.
Using multiple models requires identifying criteria and bound-

aries for the choice of the tools. The number of tools to be used
depends on the cost (several of them are commercial) and on the
time available. Generally, freely available models have similar per-
formance of commercial ones [4–8]. In addition, they may have
the advantage of a greater transparency, because commercial ones
have part of the information which is confidential (e.g., the algo-
rithm, or the set of compounds used to build the model). Some
tools are faster. Most importantly, several models provide not only
the predicted value but also information very useful to evaluate the
reliability of the result, such as the applicability domain assessment.
The user should start using these fast, freely available models,
which provide supporting information and the available commer-
cial ones. Further models should be used, in particular if there is
too much uncertainty from the obtained results.
Recently, EFSA published a guidance describing the process of
integrating and weighting evidence for scientific assessment [9].
This document describes the weight of evidence in three basic
steps: (1) assembling the evidence; (2) weighing the evidence; and
(3) integrating the evidence.
The first step involves searching for and selecting single pieces of
evidence (e.g., outcome of single studies) relevant for answering the
question, and deciding how to group it into lines of evidence (i.e., set
of evidence of similar type). The second step involves a detailed evalu-
ation to assign a weight to evidence. When weighting single line of
evidence, one should consider its reliability (i.e., the extent of scien-
tific correctness of evidence) and its relevance (i.e., the contribution
that a line of evidence would give in answering the question). In the
third step, the lines of evidence are integrated to give a final answer.
In this case consistency must be considered, i.e., compatibility of the
information of different lines of evidence contributing to the final
answer. Uncertainty and variability of data must also be taken into
account in the overall integration approach of single evidence [9].
We will go through a practical example which applies the crite-
ria identified by EFSA and will show a substance predicted non-
toxic in the presence of other results indicating toxicity.
This example evaluates the bioaccumulation potential through
the bioconcentration factor (BCF). It shows a continuous endpoint in
which all the predictions give a range of values. This is a further step,
compared to the example presented within the EFSA guidance [9]
which simply consider a categorical output (toxic or not).
2 The Used Nontesting Methods
Here four platforms—VEGA, T.E.S.T., EPISuite™, and ToxRead—

are used to estimate the toxicity/property of the chemical [10].
VEGA and T.E.S.T. platforms include more number of models for
Criteria and Application on the Use of Nontesting Methods within a Weight of Evidence… 201
the same endpoint, and already integrate the results. EPISuite™

has more number of models, but they are not integrated.
Additionally, the use of the read-across approach using ToxRead
has been made to further support the findings. The in silico plat-
forms used in the assessments are briefly described below.
T.E.S.T. program v 4.2.1: The Toxicity Estimation Software Tool
(T.E.S.T.) has been developed by scientists at the US Environmental
Protection Agency (EPA) to allow for easy estimation of toxicity
from molecular structure using a variety of QSAR methods (http://
www.epa.gov/nrmrl/std/qsar/qsar.html#TEST). In the consensus
method, the predicted toxicity is simply the average of the predicted
toxicities from QSAR models, taking into account the applicability
domain of each method. This method typically provides the highest
prediction accuracy since any inaccurate prediction by one model is
corrected/compensated by the other methods. In addition, this
method provides the highest prediction coverage because several
methods with slightly different applicability domains are used to
make a prediction. It also includes tools to evaluate if the test chemi-
cal is inside the applicability domain of the models. If it is outside,
the prediction is not provided. The software, in the case of BCF,
includes four methods in addition to a consensus model: hierarchical
clustering, single model, group contribution, and nearest neighbor.
VEGA platform v 1.1.4 beta2: VEGA is a Java based platform
(http://www.vega-qsar.eu). For BCF it includes three models:
CAESAR, Meylan (It is an implementation of the model included
in EPISuite™.), and KNN/read-across. The VEGA platform also
offers a tool to measure the reliability of the prediction, through
the applicability domain index (ADI). This index identifies the rea-
sons of possible concern, and guides the user to a careful further
evaluation of the results. It is based on a similarity check to com-
pare the queried substances with those used to develop the model
and to verify how accurate the predicted values are.
EPISuite™ v 4.1: the Estimations Programs Interface (EPI) Suite™
(https://www.epa.gov/tsca-screening-tools/epi-suitetm-estima-
tion-program-interface) is a Windows-based suite of physical/
chemical property and environmental fate estimation programs
developed by the US EPA and Syracuse Research Corporation
(SRC). The BCFBAF program predicts the bioaccumulation
potential using different approaches: a logP-based (the logP is the
logarithm of the octanol–water partition coefficient) equation to
estimate the BCF and a model that gives both BCF and BAF (bio-
accumulation factors) estimations for three trophic levels consider-
ing the metabolism and an estimation for one trophic level without
the metabolism. The KOWWIN program predicts the log Kow.
For all the programs included in the EPISuite™ the AD should be
manually checked.
ToxRead beta 0.11: ToxRead (http://www.toxread.eu) allows the

user making reproducible read-across evaluations between tested
and untested compounds. The program uses a query chemical
structure to show other structurally most similar compounds, and
any structural alerts and relevant features that may be common
between them. The platform contains libraries of chemicals with
associated experimental values, structural alerts, and algorithms of
relevant features.
3 Case Study
Figure 1 presents the chemical we used. Table 1 shows the results

of the different models. All models predict the substance as non-
bioaccumulative (considering the 3.3 l.u. threshold).
The BCFBAF program (see Fig. 2) predicts the test chemical
with a log BCF value that spread from 2.137 to 2.766 l.u. The
highest value corresponds to the BCF/BAF, Arnot–Gobas, upper
trophic level, kM = 0 model that can be considered as a worst-case
model since it do not consider the metabolism. The AD has to be
checked manually in the case of EPISuite. For BCF, Meylan model,
the AD check consists in the examination of the log Kow and the
molecular weight (MW) used. The MW should be between 68.08
and 959.17 (in this case, 226.45). The log Kow should be in the
training set range (from −1.37 to 11.26). In this case, it is of 3.74.
This value is a predicted one. Therefore, the quality of this predic-
tion should be checked too. The program used is KOWWIN v.
1.68. In this case, the AD check considers the MW (maximum
1000) and the fragments used. The fragments should cover all the
molecules and they should be present with a number of instances
lower or equal to those in the molecules of the training set. It
results that this molecule is inside the AD also for this model. The
BCF/BAF, Arnot–Gobas gives low reliable estimations for pig-
ments, dyes, and perfluorinated compounds with log Kow above 9
or that appreciably ionize. Also these requirements are satisfied,
and thus the target compound is in the AD.
Fig. 1 Chemical structure of the target substance

Table 1
Summary of the results obtained by nontesting methods
Applicability
Programs used Models used Results domain
EPISuite™ v.4.1 BCF, Meylan model 2.137 (137.2 L/kg ww) Manually
checked
BCF/BAF, Arnot–Gobas, 2.516/2.516
upper trophic level model (328/328.3 L/kg ww)
mid trophic level model (276.6/278.4 L/kg
ww)
lower trophic level model (256.2/261.7 L/kg
ww)
upper trophic level, (583.1/1001 L/kg ww)
kM = 0 model
T.E.S.T. v. 4.2.1 CONSENSUS method 2.36 (230.04 L/kg) Internally
2012 U.S. EPA checked
Hierarchical clustering method 2.31 (202.85 L/kg);
PI* = 0.39
Single model method 2.30 (199.95 L/kg);
PI* = 1.24
Group contribution method 2.46 (288.98 L/kg);
PI* = 2.62
FDA 2.29 (196.35 L/kg);
PI* = 1.19
Nearest neighbor method 2.45 (279.89 L/kg)
VEGA platform BCF model (CAESAR) 2.1.14 1.61 (40 L/kg) 0.85 ADI
v.1.1.3
BCF model (Meylan) 1.0.3 2.14 (137 L/kg) 1 ADI
BCF (KNN/read-across) 1.1.0 2.47 0.7 ADI
*PI = Prediction interval
T.E.S.T. applies statistical models that also predict the substance

as nonbioaccumulative (see Fig. 3). Five models are present in this
platform, and then their results are integrated (see Table 1). The
models predict log BCF from 2.29 to 2.46 with a consensus of 2.36.
Each prediction is supported by several analyses: (1) the predic-
tion interval (excluded the Nearest neighbor method); (2) the list of
the most similar compounds of the external set with a similarity
score ≥0.5 (see Fig. 4); (3) the mean absolute error (MAE, in log10)
for both the entire external set and for the most similar compounds
(see Fig. 5); (4) the list of the most similar compounds of the training
set with a similarity score ≥0.5 (see Fig. 6); (5) the MAE (in log10)
Fig. 2 Output of EPISuite™
for both the entire training set and for the most similar compounds.
Analyzing each prediction is it possible to observe that the Single
model method, the Group contribution method and the FDA
method give low reliable prediction due to the high prediction inter-
val (of 1.18/3.42, 1.15/3.77, and 1.70/2.89 respectively). The
Hierarchical clustering gives a reliable prediction (with a prediction
interval of 2.11/2.50). The MAE values confirm this tendency (data
not shown). In any case, all the MAE are below or equal to the
experimental variability (that spreads from 0.42 to 0.75 [11, 12]).
Observing the list of the most similar compounds of the training set,
we can see that the most similar compound (S1, CAS 117-18-0) is very
similar. Indeed, it has only one added chlorine atom. It has an experi-
mental log BCF value of 3.25. The second one (S2, CAS 609-89-2)
has an added hydroxyl group; therefore, it is probably more soluble
Fig. 3 Summary of the results of T.E.S.T.
in water. For this reason, it should be considered not so similar. The

third one (S3, CAS 18708-70-8) has three chlorine atoms and a
nitro group attached to the benzene ring as the target, but in differ-
ent positions, whereas the fourth one (S4, CAS 99-54-7) has only
two chlorine atoms. In this case, the third compound is the most
similar one. All models underestimate S1, 3 and 4 with a maximum
error of 1 l.u. Furthermore, the experimental log BCF increases
passing from two to four chlorine atoms (from 1.92 to 3.25).
Therefore, the target that has three chlorine atoms should have a log
BCF included in this range. In this case, the maximum log BCF is
below the threshold. Therefore, the target compound should be
considered nonbioaccumulative.
The VEGA platform has three models that predict the BCF. All
of them predict the target compound as nonbioaccumulative.
VEGA gives for each model an applicability domain evalua-
tion, the ADI. The assessment of the first model, the CAESAR
model for BCF version 2.1.14 (see Fig. 7), indicates that the pre-
diction compound could be out of the AD. The reason is that the
first two similar compounds of the training set (S1, CAS 99-54-7;
S2, CAS 609-89-2) have errors in prediction of about 0.6 with a
maximum error of 0.737. These errors are comparable to the
experimental variability. Observing the list of the six most similar
compounds (Fig. 8) we can see that S2 has an alert not found in
the target: the hydroxyl group (PG 06). This is the same problem
of the S2 of the T.E.S.T. model. However, VEGA helps the user to
identify the differences showing the list of alerts identified in the
similar chemical that are not present in the target one.
Similarity Experimental value Predicted value

CAS Structure
Coefficient Log10 Log10
1
N/A 2,36
(test chemical)
88-73-3 0,86 1,71 1,31
95-95-4 0,74 2,46 2,03
88-72-2 0,68 1,34 1,04
6130-75-2 0,66 2,78 2,73
54135-80-7 0,66 2,96 2,71
119-33-5 0,65 0,96 0,85
121-14-2 0,64 0,83 0,81
98-15-7 0,59 2,32 2,23
555-03-3 0,57 0,72 0,96
108-43-0 0,56 1,08 1,28
Fig. 4 List of similar compounds of the external set of T.E.S.T.

Fig. 5 Prediction for the similar compounds of the external set of T.E.S.T.
The second model (see Fig. 9), the BCF model (Meylan) 1.0.3,
assesses the target compound into the Applicability Domain of the
model. Indeed, the two most similar compounds (see Fig. 10) are
very similar to the target (with three chlorine atoms and one nitro
group attached to a benzene ring). Moreover, they are well pre-
dicted and concordant with the prediction for the target compound
(of about 2 l.u.). We notice that this model is the same model as in
EPISuite, but in this case the AD is checked by the VEGA platform,
and more information is provided, such as the similar compounds.
The third model, the BCF (KNN/read-across) 1.1.0, assesses the
target compound as outside the Applicability Domain of the model
(see Fig. 11). The reason is the wrong prediction of the second most
similar compound (CAS 117-18-0), with an error in prediction above
1 l.u. This model uses the first four similar compounds (see Fig. 12).
The first one (CAS 18708-70-8) is similar to the target substance
(three chlorine atoms and a nitro group attached to the benzene ring),
the second one (CAS 117-18-0) has an additional chlorine atom, the
third one (CAS 99-54-7) a chlorine atom less and the fourth one
(CAS 99-30-9) has a chlorine atom less and an additional secondary
amine. As for the T.E.S.T. model similar compounds, the molecules
with four chlorine atoms have a higher log BCF value, whereas the
ones with two chlorine atoms a lower value.
The BCF CAESAR and Meylan models in VEGA offer also
another tool of analysis for the users: the scatter plots log BCF vs
MlogP (see Fig. 13). MlogP is a log Kow calculated by the model. For
each model, a scatter plot with all the compounds of the training set
Similarity Experimental value Predicted value

CAS Structure
Coefficient Log10 Log10
1
N/A 2,36
(test chemical)
117-18-0 0,96 3,25 2,75
609-89-2 0,88 1,37 1,64
18708-70-8 0,83 2,72 2,30
99-54-7 0,79 1,92 1,89
50-31-7 0,76 0,07 0,76
58-90-2 0,76 2,01 2,35
935-95-5 0,71 2,15 2,31
87-86-5 0,71 2,67 2,89
88-75-5 0,71 1,00 0,76
636-30-6 0,70 2,33 2,07
Fig. 6 List of similar compounds of the training set of T.E.S.T.

Fig. 7 Summary prediction of the BCF CAESAR model
(Fig. 13a and c) and one with the three most similar compounds
(Fig. 13b and d) are shown. If the target compound is in the cloud,
as in the first plot, it means that it is similar to the compounds of the
training set (They are represented by their experimental value.). The
second plot allows for verifying if there is a trend between the three
most similar compounds and the target. It shows both the experi-
mental (circles) and predicted (black dot) value of the similar sub-
stances (the dimension of the circle represents the similarity).
Figure 13a indicates that the target compound is inside the cloud for
the CAESAR model but borderline, whereas Fig. 13c that it is inside
the cloud of the Meylan model. This is confirmed by Fig. 13b and d:
in the first one the target has a higher MlogP than the similar com-
pounds. Considering that the log BCF increases if the MlogP
increases, this plot may indicate an underestimation of the log BCF
for the target compound: all the similar compounds have the experi-
Fig. 8 List of the most similar compounds found in the BCF CAESAR model
Fig. 9 Summary prediction of the BCF Meylan model
mental values (the open, white circles in the figure) higher than the
predicted values (the black dots). Extrapolating these experimental
values, or adding the error done in prediction, we can assume that the
correct value is about 2.2. In Fig. 13d all similar compounds and the
target have the same MlogP, and log BCF spreads from 1.84 to 2.47.
Figure 14 shows as additional analysis: the uncertainty assess-
ment. In this assessment, a safety margin is added to the predicted
value. The safety margin is calculated on the bases of the ADI and
the threshold (3.3 or 3.7 l.u.) considered. In this case, with both
the thresholds the values remain widely below 3.3 l.u.
ToxRead gives details about the most similar compounds (their
number is selected by the user), the rules identified (with a descrip-
tion and the list of the compounds in which the rule appear), and
the interpolation chart (logP vs log BCF). Figure 15 shows the out-
put with three most similar molecules. ToxRead found two rules,
the nitro aromatic and the acceptor atoms for H-bonds (N, O, F).
Fig. 10 List of the most similar compounds found in the BCF Meylan model
The three most similar compounds have a nitro group and two,
three, or four chlorine atoms attached to the benzene ring. The
same rules are present in all of them. In the interpolation chart we
can see that the most similar compound (CAS 18708-70-8), which
Fig. 11 Summary prediction of the BCF KNN/read-across model
differs only for the chlorine atoms position, has a very similar logP
value and an experimental log BCF value of 2.72. The other two
chemicals have lower logP and log BCF or higher logP and log
BCF. The log BCF value of the target molecule should be between
1.92 and 3.26 and should be very similar to the one of the CAS
18708-70-8.
Figure 16 shows the behavior of a larger number of similar
compounds (10 in this case). The interpolation chart shows also in
this case a linear trend: the higher the logP, the higher the log
BCF. This confirms the conclusion obtained with the analysis done
with three similar compounds.
Fig. 12 List of the most similar compounds found in the BCF KNN/read-across model
a c
LogBCF LogBCF
5 6
4 5
3 4
2 3
1 2
1
0
0
-1
-3 -2 -1 0 1 2 3 4 5 6 7 8 9 -6 -5 -4 -3 -2 -1 0 1 2 3 4 5 6 7 8 9 10 11 12 13
MLogP MLogP
b d
LogBCF
2 2.5
2.4
1.5 2.3
2.2
1 2.1
2
0.5 1.9
1.8
2 2.5 3 3.5 4 3.7 3.8

MLogP MLogP
Fig. 13 Scatter plots for the CAESAR model (a, b) and for the Meylan model (c, d)
4 Conclusion
In conclusion, all the QSAR models used agree and predict the
target compound as nonbioaccumulative, with a log BCF of
about 2.5, and values from 1.61 to 2.766. The lowest value is
clearly underestimated, as we discussed (Fig. 13). Since the
majority of the QSAR models have, in this case, the tendency to
underestimate the similar compounds, the most probable value
is about 2.7–2.8 l.u. The same conclusion can be reached also
using the ToxRead software that shows a similar compound
with a very similar logP with a log BCF of 2.72.
Here we show how in silico methods and, in general, NTMs
are a valuable resource for risk assessment of substances when
multiple pieces of evidence from multiple tools are available.
Evidence can be integrated and compared in order to obtain a
highly confident assessment, when it results fromdifferent tools
yet shows concordance.
Fig. 14 Uncertainty assessment of the CAESAR model
Acknowledgments
We acknowledge the financial contribution of the to DIVINE

project FKZ 3717 65 417 0, funded by the German Environment
Agency (UBA).
Fig. 15 ToxRead output with three similar compounds (lower part of the figure), the rules (on the right), the
interpolation chart with logP (upper part of the figure, in the middle) and the graph with the overview of the
similar compounds and rules (upper, left)
Fig. 16 ToxRead output with ten similar compounds

References
1. Benfenati E, Pardoe S, Martin T, Gonella 8. Milan C, Schifanella O, Roncaglioni A,
Diaza R, Lombardo A, Manganaro A, Gissi Benfenati E (2011) Comparison and possible
A (2013) Using toxicological evidence from use of in silico tools for carcinogenicity within
QSAR models in practice. ALTEX 30:19–40 REACH legislation. J Environ Sci Health
2. Benfenati E, Belli M, Borges T, Casimiro E, Part C Environ Carcinog Ecotoxicol Rev
Cester J, Fernandez A, Gini G, Honma M, 29:300–323
Kinzl M, Knauf R, Manganaro A, Mombelli E, 9. Scientific Committee EFSA, Hardy A, Benford
Petoumenou MI, Paparella M, Paris P, Raitano D, Halldorsson T, Jeger MJ, Knutsen HK,
G (2016) Results of a round-robin exer- More S, Naegeli H, Noteborn H, Ockleford
cise on read-across. SAR QSAR Environ Res C, Ricci A, Rychen G, Schlatter JR, Silano V,
27:371–384 Solecki R, Turck D, Benfenati E, Chaudhry
3. Benfenati E, Roncaglioni A, Petoumenou MI, QM, Craig P, Frampton G, Greiner M, Hart A,
Cappelli CI, Gini G (2015) Integrating QSAR Hogstrand C, Lambre C, Luttik R, Makowski
and read-across for environmental assessment. D, Siani A, Wahlstroem H, Aguilera J, Dorne
SAR QSAR Environ Res 26:605–618 J-L, Fernandez Dumont A, Hempen M,
4. Cappelli CI, Benfenati E, Cester J (2015) Valtuena Martınez S, Martino L, Smeraldi C,
Evaluation of QSAR models for predicting Terron A, Georgiadis N, Younes M (2017)
the partition coefficient (logP) of chemicals Guidance on the use of the weight of evidence
under the REACH regulation. Environ Res approach in scientific assessments. EFSA J 15.
143:26–32 https://doi.org/10.2903/j.efsa.2017.4971
5. Cappelli CI, Cassano A, Golbamaki A, Moggio 10. Marzo M, Roncaglioni A, Kulkarni S, Barton-
Y, Lombardo A, Colafranceschi M, Benfenati E Maclaren TS, Benfenati E (2016) In Silico
(2015) Assessment of in silico models for acute model for developmental toxicity: how to
aquatic toxicity towards fish under REACH reg- use QSAR models and interpret their results.
ulation. SAR QSAR Environ Res 26:977–999 Methods Mol Biol 1425:139–161
6. Diaza RG, Manganelli S, Esposito A, 11. Lombardo A, Roncaglioni A, Boriani E, Milan
Roncaglioni A, Manganaro A, Benfenati C, Benfenati E (2010) Assessment and vali-
E (2015) Comparison of in silico tools for dation of the CAESAR predictive model for
evaluating rat oral acute toxicity. SAR QSAR bioconcentration factor (BCF) in fish. Chem
Environ Res 26:1–27 Cent J 4:S1
7. Bakhtyari NG, Raitano G, Benfenati E, 12. Dimitrov S, Dimitrova N, Parkerton T,
Martin T, Young D (2013) Comparison of in Comber M, Bonnell M, Mekenyan O (2005)
silico models for prediction of mutagenicity. Base-line model for identifying the bioaccu-
J Environ Sci Health Part C Environ Carcinog mulation potential of chemicals. SAR QSAR
Ecotoxicol Rev 31:45–66 Environ Res 16:531–554
Chapter 10
Characterization and Management of Uncertainties

in Toxicological Risk Assessment: Examples
from the Opinions of the European Food Safety Authority
Alberto Mantovani
Abstract
Uncertainties can be defined as the gaps of knowledge and/or of data sets and/or of methodologies that
can exert an unwanted influence on the outcome of a risk assessment. In principle, uncertainties are
unavoidable, and thus, a transparent description and weighing of relevant uncertainties should be a neces-
sary component of risk assessment. Examples are provided of uncertainty analysis in recent opinions of the
European Food Safety Authority concerning additives, pesticides, and contaminants. Whereas it is difficult
to quantify the impact of each specific uncertainty on the outcome, it should be possible to quantify the
combined effect of identified uncertainties; also, a stepwise approach may be envisaged, focusing on those
issues where a detailed appraisal of uncertainties is needed. On a more general ground, consideration of
uncertainty and its sources meets the general requirement for transparency in scientific assessment.
Key words Food safety, Contaminant, Pesticide, Additive, Exposure, Point of departure, Adverse
outcome pathway, Benchmark dose
1 Introduction
Toxicological risk assessment aims at providing scientific advice to

risk managers and policy makers concerning chemicals to which
humans and/or the environment may be exposed. Such advice is a
basis for weighing options and taking decisions (e.g., in order to
set limits or to restrict or ban certain chemicals). Risk assessment is
a formalized, multistep process [1] encompassing hazard identifi-
cation, hazard characterization, exposure assessment, and risk
characterization. The whole process rests on the pillars of up-to-
date scientific evidence and of transparency. Yet one main compo-
nent of process transparency is the description of the knowledge
gaps or data inconsistencies and the assessment of how these may
impact on the assessment outcome, i.e., on the output that risk
managers expect from risk assessors [2, 3]. The knowledge gaps
and data inconsistencies that can be relevant to the outcome of risk
219
220 Alberto Mantovani
assessment are defined as “uncertainties.” It is widely recognized,

in principle, that uncertainties need to be dealt with by a formal-
ized and consistent approach across the different fields of risk
assessment; however, the elaboration of such process and its practi-
cal use are still a work in progress. Indeed, the guidance developed
by the European Food Safety Authority on how to cope with
uncertainty in risk assessment, after a public consultation in 2016,
is still (January, 2018) under internal testing [4].
As stated above, the uncertainties that are of greater interest
for scientists are those stemming from knowledge gaps, defined as
“epistemic” uncertainties. There are other relevant sources of
uncertainties as well [5]. Due to natural variability as well as ran-
domness, it is difficult that two successive measurements in the
same population yield exactly the same results: this is defined as
“aleatory” uncertainty. Another source of uncertainty is the expert
judgment: risk assessment usually requires quantitative estimates
(the intake level of a contaminant starting to cause toxicity in
exposed individuals, the amount of a pesticide residue in a vegeta-
ble that is made up by the very toxic metabolite Y, etc.): these
estimates of a “true” value are unavoidably influenced by the sub-
jective judgments of experts, and, in practice, groups of experts
with different compositions may produce different interpretations
of the same data sets. An approach to cope in a transparent way
with uncertainties due to expert judgment is the expert elicitation
of probabilistic distributions; for example, experts are elicited to
provide evidence-based estimates of the dose level that have a 5%,
50%, or 95% probability to exceed the “real” no-effect-level for a
critical toxicological effect [4]. Needless to say, expert knowledge
elicitation is a time-consuming process that should be used when
uncertainties may substantially weaken the risk assessment output.
Indeed, the uncertainty concept has been incorporated in
time-honored, established outputs of toxicological risk assessment.
The acceptable daily intake (ADI) is the standard measure to iden-
tify an intake of a food contaminant, additive or residue that, based
on available knowledge, would not cause any appreciable health
risk [1]. The ADI is calculated using the most conservative and
reliable no observed adverse effect level (NOAEL) for the toxico-
logical effects of the substance X (where “observed” means “within
the boundaries of the available data set): the NOAEL is divided by
“safety” or, more accurately, “uncertainty” factors to take into
account to the potential higher human susceptibility compared to
laboratory animals (when the NOAEL is derived from animal stud-
ies, as in most cases) and of the intraspecies variability in humans
(such as due to health status, age, genetic background, etc.). In
some cases the uncertainty factors are data based, but in most cases
they are conventional (usually 10 for interspecies and 10 for intra-
species variability: 10 × 10 = 100). Thus, when dissecting the
established ADI parameter, it is apparent that effort is needed to
Characterization and Management of Uncertainties in Toxicological Risk Assessment… 221
cope with different systemic uncertainty sources. The use of the

benchmark dose (BMD) approach in risk assessment is considered
to reduce the uncertainties related to the conventional, NOAEL-
based ADI calculation [6, 7]. The BMD is a dose level, derived
from the estimated dose–response curve, associated with a speci-
fied change in response (e.g., a 5% or 10% increase of the incidence
of a given quantal effect); therefore the BMD approach, contrary
to the use of NOAEL, makes an extended, consistent, and system-
atic use of all available data relevant to the dose–response curve.
Most important, it is recommended to always report the BMD
confidence interval (usually the 95% confidence limits): the lower
bound (BMDL) is used as a reference point to set the ADI or other
health guidance values, and the upper bound (BMDU) is needed
for establishing the BMDU–BMDL ratio, which reflects the uncer-
tainty in the BMD estimate. Noticeably, the use of BMD cannot,
per se, avoid the need for the uncertainty factors accounting for
interspecies and intraspecies variability.
The above paragraphs concern the elaboration of uncertainty in
regard to the first component of risk assessment, the identification
and characterization of hazards using toxicological studies. The sec-
ond component, exposure assessment, necessarily relies on robust
chemical measurements. Chemists and metrologists have considered
the sources of measurement. Uncertainties of specific analytical
methods. The consequences on the interpretation of data and the
potential remedies, also by dissecting the constituents of accuracy of
results from different analyte–matrix–technique combinations [8].
The current scenario sees a still ongoing development of con-
ceptual framework on how to cope with uncertainties in toxico-
logical testing and in the use of data for risk assessment; in the
meanwhile, there is growing set of assessments by international
agencies where the identification of epistemic uncertainties is seri-
ously taken into account.
In Europe, and also worldwide, the EFSA (http://www.efsa.
eueopa.eu) is becoming both a hub for methodologies and tools used
in risk assessment and a major repository of risk assessment opinions
since 2003. The following chapters will deal with two recent exam-
ples of EFSA opinions where uncertainties were identified and char-
acterized: risk assessment of specific chemicals, namely nitrite salts
used as food additives [9]: uncertainties in the development of a
methodological approach, namely the use of Adverse Outcome
Pathways for the hazard characterization of pesticides [10].
2 Case Study: Nitrites as Food Additives
Potassium and sodium nitrite are food additives authorized in the

European Union: they have been recently reevaluated by EFSA
[9]. An ADI of 0.07 mg nitrite ion/kg bw per day is established
using the BMD approach and the increase in methemoglobin

blood level, observed in human and animals, as relevant effect
[11]. The estimated exposure from the use as food additive will
not lead to exceed the ADI, except for children age group, where
a slight exceedance is estimated to occur in high consumers. In the
meanwhile, the contribution of the use as food additives represents
only 17% (range 1.5–36.0%) of the overall dietary exposure to
nitrites, which includes the natural occurrence in foods and envi-
ronmental contamination (e.g., waste, fertilizers); therefore, the
overall dietary exposure in estimated to exceed the ADI in infants
(up to 1 year), toddlers (1–3 years), and children (up to 9 years)
[9].
The main uncertainties concerned the ADI derivation and the
exposure assessment.
The induction of methemoglobin formation following expo-
sure to nitrites is well established, as well as dose responses in rats
and mice exposed to nitrite and also in some human studies. The
EFSA expert panel made the following specific considerations:
●● Although the efficiency of formation of methemoglobin from
nitrite in rats and humans could be different, the mechanism is
identical; therefore, rodent data are relevant for hazard
characterization.
●● Whereas the ADI should normally be derived from chronic
studies covering lifetime exposures, a 14-week (subchronic) rat
study provided the best dose-response data in order to derive a
BMDL. However, the same effects were observed at compa-
rable dose levels in longer term studies on rodents, indicating
no need for an additional uncertainty factor for time extrapola-
tion from subchronic to chronic exposure.
●● Human data were available, but were too limited to derive an
ADI: therefore a default uncertainty factor of 100 applied to
rodent data was used to derive the ADI. Better kinetic data in
humans would yield more information on intraspecies and
interspecies differences, respectively. Resulting in a specific
uncertainty factor, likely smaller than 100.
A more significant uncertainty might be represented by the
identification of the critical threshold (i.e., “point of departure”)
for hazard characterization. This was the dose level estimated to
induce a twofold increase of the mean background concentration
of methemoglobin. It might be uncertain whether such increase
in methemoglobin represents an adverse effect and/or is simply a
biomarker of effective dose (i.e., of nitrite internal dose inducing
any biological activity). Also the background methemoglobin
levels are quite variable, as observed in control groups of the dif-
ferent rat studies., However, some population groups are highly
susceptible to methemoglobin adverse effects, in particular infants
up to 4 months of age and subjects with certain genetic

conditions.
Overall, the identified uncertainties were considered to have a
minimal impact on the risk assessment and the ADI was sufficiently
conservative to cover the most susceptible population subgroups.
As for exposure, the food consumption data were collected by
different methodologies in different countries; in particular, differ-
ent criteria and levels of detail for food categorization are a source
of uncertainties. Moreover, and rather obviously, the exposure esti-
mates depended on the availability and quality of data from the
different European Countries. In general, the use of data from
food consumption survey lasting a few days to estimate long-term
(chronic) exposure is a significant source of uncertainty, especially
for assessing the intake of high consumers [12]. Other uncertain-
ties were the use of scenarios where relevant foods contained the
additive at the maximum permitted level, without considerations
for the several restrictions/exceptions concerning specific prod-
ucts, due to insufficient consumption data. Overall, such uncer-
tainties likely led to an overestimation of the exposure to nitrites as
food additives.
A main issue is the endogenous formation of nitrosamines
from nitrites: these potentially carcinogenic metabolites are the
plausible reason for the observed epidemiological links between
(overall) dietary nitrite and gastric and colorectal cancers in humans
[13]. As for other carcinogenic compounds the EFSA has used an
approach based on the margin of exposure (MoE) between the
BMD for nitrosamine carcinogenicity in laboratory animals and
the nitrosamine intake resulting from the ingestion of nitrites at
the ADI: a MoE <10,000 is considered to indicate a possible health
concern. The MoE for exogenous nitrosamines in meat products
was estimated to be <10,000 for high consumers in all age groups.
The assessment of nitrosamine content in foods had a further sig-
nificant uncertainty, relevant to risk managers: it was impossible to
discriminate between the nitrosamines resulting from the nitrite
added at the authorized levels and those resulting from nitrite in
the food matrix. Since the use as additive represents a smaller frac-
tion of dietary nitrite compared to other sources, it is likely that
this uncertainty does not reduce the conservativeness of the assess-
ment. Therefore, whereas the uncertainties per se were high, their
impact on the overall risk assessment of nitrites as a food additive
was, rather reasonably, considered as low in view of the definitely
greater weight of other dietary sources [9]. Finally, nonvolatile
nitrosamines may also increase after nitrite addition into cured
meat. Even though these are considered of lower toxicological
concern based on their chemical structure, there are gaps of knowl-
edge concerning their actual toxicological hazards and levels of
formation, leading to an uncertainty in risk assessment: such uncer-
tainty can only be reduced by the generation of experimental data.
3 Use of Adverse Outcome Pathways in the Hazard Characterization

of Pesticides
An adverse outcome pathway (AOP) describes the chain of events

leading from the first interaction of any chemical with a molecular
target (molecular initiating event) to an adverse outcome, which
can be an effect accepted as adverse in regulatory toxicity testing,
as well as a health disorder in humans or in wildlife: these upstream
and downstream terminals are sequentially linked by a series of
biologically plausible and essential key events at the levels of cell
organelles, cells, and tissue/organ. Indeed, the AOP framework
may be considered akin to the traditional concept of “pathogene-
sis,” in addition it is presented in a formal and consistent way in
accordance with a set of quality criteria. AOP are not toxicity
mechanisms of specific chemicals; indeed AOP are “chemically
agnostic”, even though studies on chemicals can provide empirical
evidence to build a specific AOP. As a result, any chemical trigger-
ing a given upstream event with sufficient intensity is flagged as
able to perturb adversely a physiological pathway, therefore to
induce a certain kind of toxicity [14, 15].
The use of AOP in risk assessment is a new development that
still needs a lot of refinement and practical implementation. EFSA
has used AOP in order to deal with a hot and emerging issue: the
plausible involvement of pesticide exposure as a risk factor for
Parkinson’s disease and childhood leukemia, suggested by the
review of available epidemiological studies [10]. Four AOP at a
different degree of definition are described in this EFSA opinion:
two for Parkinson’s disease in relation to two molecular initiating
events that were confluent downstream, and one each for infant
leukemia and childhood leukemia. The biological plausibility of
epidemiological associations for these three diseases is supported,
albeit with different weight of evidence, by combining multiple
information (in vitro, in vivo, and clinical studies) on biological
pathways and species differences and similarities. The conclusions
identify research needs and suggest possible improvements of reg-
ulatory strategies in order to incorporate novel markers relevant to
the investigated adverse outcomes [16, 17]. In the meanwhile, a
number of uncertainties were identified a regarding each AOP as
well as general aspects.
Specific uncertainties in the development of AOP for Parkinson
disease concern a few inconsistencies and gaps about the relation-
ships among the upstream events, in particular mitochondrial dys-
function leading to disturbed proteostasis and the role played by
oxidative stress.
For infant leukemia the main specific uncertainties are due to
the lack of animal models, the lack of robust studies to compare
the sensitivity to chemical stressors of fetal hematopoietic stem
cells (the primary site of the AO) to their mature counterparts; the
rarity of the disease itself is a source of uncertainty, as it limits the
power of human studies investigating the biology of the disease.
For infant leukemia the in utero disruption of DNA topoisom-
erase II is the molecular initiating event; conversely, the main spe-
cific uncertainty for childhood leukemia is the lack of an identified
molecular initiating event, although this is considered to occur in
utero. Pesticides authorized in Europe do not show genotoxic
effects in regulatory studies; on the other hand, some studies in the
open literature suggest genotoxicity by using different biomarkers
and some epidemiological studies on agricultural workers exposed
to pesticides have reported DNA damage. These uncertainties and
inconsistencies warrant further research to delineate whether and
how specific pesticides may interact with DNA in stem cells in
utero and produce genetic lesions.
General aspects of uncertainty concerned the three major areas
explored during the development of the opinion: epidemiological
studies, experimental studies, and AOP development.
●● Epidemiological studies: uncertainties include the definition of
outcome. In particular, epidemiological studies grouped infant
and childhood leukemia which in fact have different pathogen-
esis, leading to significant data heterogeneity. The main uncer-
tainty in epidemiological studies are exposure estimates, from
both standpoints of the generic definition of the substances of
concern (e.g., “insecticides” which include substances with
largely different toxicological profiles) and of the lack of quan-
titative information concerning internal exposure. Furthermore,
in a realistic field scenario, humans are exposed to several sub-
stances and coformulants contained in pesticide products, rep-
resenting an additional source of uncertainty. Finally, the
limitation of knowledge about the etiology of multifactorial
human diseases, involving the presence of different pheno-
types, genetic and environmental factors, is an uncertainty
directly relevant to AOP development.
●● Experimental studies. For most pesticides, the bulk of experi-
mental studies is provided by regulatory studies which usually
do not provide information on upstream events. For some
human diseases, including infant and childhood leukemias,
predictive in vivo models are not available.
●● The most relevant uncertainties concerned the process of AOP
development. Most of the current empirical support for molec-
ular initiating events and upstream key events is derived from in
vitro assays, which use ad hoc nonstandardized models; the
availability of such models and their characteristics may affect
the full description and characterization of the AOP. The differ-
ences in the exposure (time, concentrations, and route of
administration) between in vitro models and in vivo scenarios
may be considered an additional uncertainty; however, in any

case the AOP model can provide proof of concept that environ-
mental chemicals eliciting certain molecular/cellular events can
cause the AO.
Another important uncertainty is the limited empirical sup-
port, when an AOP is illustrated by a single or few chemical
stressors.
Finally, a conceptual uncertainty is how to include the role of
modulating factors (i.e., not essential, but influencing the relation-
ships among key events): these can be relevant when dealing with
AOs derived from complex human diseases involving gene–envi-
ronment interactions.
AOP may not need to be extremely detailed in the description
of process, whereas they must be fit-to-purpose and compliant to a
set of quality criteria that qualify the key events and their relation-
ships. Indeed, the identification of several uncertainties can improve
the transparency of the AOP approach [18].
4 General Considerations and Conclusions
In practical terms, uncertainties can be defined as gaps of knowl-

edge and/or of data sets and/or of methodologies that can exert
an unwanted influence on the outcome of a risk assessment. In
principle, uncertainties are unavoidable, and thus, a transparent
description and weighing of relevant uncertainties should be a nec-
essary component of risk assessment. It is apparent that uncertain-
ties are expected to be much lower when risk assessment addresses
substances that are supported by a dossier presented by an appli-
cant, since the data set must be compliant to regulatory require-
ments. Also in this case, however, significant uncertainties can be
encountered. For example, in the case of a feed additive based on
lanthanum and cerium in weaned piglets, the EFSA panel of experts
on substances used in animal feeds (FEEDAP) considered that
major and unclarified data gaps in toxicology and residues of these
two trace elements raised significant uncertainties and prevented
the evaluation of potential risks for consumers resulting from the
intake of edible tissues from treated pigs [19]. In other cases, as in
[10], new findings may raise issues that are not covered by current
regulatory tests; in its turn, this calls for expert elicitation and
development of novel approaches, with the associated assessment
of uncertainties.
Overall, the weight of uncertainties can be much higher for
contaminants where the assessment cannot make avail of any regu-
latory dossier: the significant uncertainty burden is evident in a
number of EFSA opinions such as on undesirable process by-
products in edible fats [20] and on the mycotoxin deoxynivalenol

[21]. Broadly speaking, a major group of uncertainties concerns
exposure data, especially for emerging contaminants. Factors
include uneven geographical distribution of data collection, use of
nonstandardized methods for sampling and analysis, and lack of
normal distribution of values: the assessors have to clearly describe
the assumptions they make, as well as whether these assumptions
are likely to result in underestimation or overestimation of risk. Also
when biomonitoring data are available, uncertainty is associated
with gaps in knowledge on the factors influencing variability within
and among populations [21]. Another major cluster of uncertain-
ties concerns the derivation of the point of departure; contrary to
most substances used in food chains, for contaminants a structured
data set of toxicological studies is normally not available: in most
cases, the toxicological database depends on how and to what
extent and on what aspects research has devoted attention to that
substance. Uncertainties can result, for instance, from the lack of a
NOAEL, the insufficient information on toxicokinetics, including
the formation of toxic metabolites, and the extrapolation of toxico-
kinetic pathways from animals to humans. The assessors may have
to report that uncertainties are high and this might lead to an
underestimation of risk [20].
Assessment of uncertainties in toxicological risk assessment is
still a work in progress [4]; nevertheless, a number of points are
now established.
Consideration of uncertainty is required to meet general
requirement for transparency in scientific assessment: sources of
uncertainty have to be identified in a clear and understandable way,
and their impact on the final assessment outcome has to be charac-
terized considering both the range and the likelihood of possible
outcomes.
Uncertainty analysis should be flexible and apt to different sce-
narios, yet it should follow a robust and consistent framework.
Such a framework can gain strength from the growing set of risk
assessments that include a formalized evaluation of uncertainties.
Whereas it is difficult to quantify the impact of each specific uncer-
tainty on the outcome, it should be possible to quantify the
combined effect of identified uncertainties: indeed, in the same
assessment specific uncertainties may lead to an overestimation
while others to an underestimation of risk [9, 20, 21]. Finally, in
the current European and international scenario the requests to
risk assessors are ever increasing: for the sake of time effectiveness
and efficient use of resources, a stepwise approach may screen the
cases that really need a detailed appraisal of uncertainties.
References
1. FAO (Food and Agriculture Organization of 11. Mensinga TT, Speijers GJA, Meulenbelt
the United Nations) and WHO (World Health J (2003) Health implications of exposure
Organization) (2009). Principles and methods to environmental nitrogenous compounds.
for the risk assessment of chemicals in food. Toxicol Rev 22:41–51
Environmental Health Criteria 240, http:// 12. European Food Safety Authority (2007)
www.who.int/foodsafety/publications/chem- Scientific opinion of the scientific committee
ical-food/en/ (last visit September 2017) related to uncertainties in dietary exposure
2. Dong Z, Liu Y, Duan L, Bekele D, Naidu R assessment. EFSA J 5:438–492
(2015) Uncertainties in human health risk 13. IARC - International Agency for Research on
assessment of environmental contaminants: a Cancer (2010) Ingested nitrate and nitrite, and
review and perspective. Environ Int 85:120–132 cyanobacterial peptide toxins. IARC Monogr
3. Aiassa E, Higgins JP, Frampton GK, Greiner Eval Carcinog Risks Hum 94:v–vii, 1–412
M, Afonso A, Amzal B, Deeks J, Dorne JL, 14. Organisation for Economic Co-operation
Glanville J, Lövei GL, Nienstedt K, O'connor and Development - OECD (2013) Guidance
AM, Pullin AS, Rajić A, Verloo D (2015) document on developing and assessing adverse
Applicability and feasibility of systematic review outcome pathways. OECD, Paris. ENV/JM/
for performing evidence-based risk assessment MONO 6, p 45
in food and feed safety. Crit Rev Food Sci Nutr 15. Mantovani A (2017) Some considerations on
55:1026–1034 adverse outcome pathways: how to use them
4. European Food Safety Authority (2017) in toxicological risk assessment. Adjacent
Guidance on Uncertainty in EFSA Scientific Government August 2017, pp .130–131.
Assessment. Revised Draft for Internal Testing. http://edition.pagesuite-professional.co.uk/
https://www.efsa.europa.eu/sites/default/fil html5/reader/production/default.aspx?
es/160321DraftGDUncertaintyInScientificAs pubname=&edid=f8abf7ad-bc1b-4a51-b42b-
sessment.pdf (last visit, January 2018) d7b202a02971 (last visit, September 2017)
5. European Food Safety Authority (2014) 16. Fujita KA, Ostaszewski M, Matsuoka Y, Ghosh
Guidance on expert knowledge elicitation in S, Glaab E, Trefois C, Crespo I, Perumal TM,
food and feed safety risk assessment. EFSA Jurkowski W, Antony PM, Diederich N, Buttini
J 12:3734. [278 pp] M, Kodama A, Satagopam VP, Eifes S, Del Sol
6. European Food Safety Authority, Scientific A, Schneider R, Kitano H, Balling R (2014)
Committee (2009) Guidance of the scientific Integrating pathways of Parkinson’s disease in
committee on a request from EFSA on the use a molecular interaction map. Mol Neurobiol
of the benchmark dose approach in risk assess- 49:88–102
ment. EFSA J 1150:1–72 17. Pelkonen O, Terron A, Hernandez AF,
7. European Food Safety Authority, Scientific Menendez P, Bennekou SH, EFSA WG EPI1
Committee (2017) Update: guidance on the and its other members (Angeli K, Fritsche E,
use of the benchmark dose approach in risk Leist M, Mantovani A, Price A, Viviani B)
assessment. EFSA J 15:4658–5079 (2017) Chemical exposure and infant leukae-
8. Frazzoli C, Robouch P, Caroli S (2010) mia: development of an adverse outcome path-
Analytical accuracy for trace elements in food: way (AOP) for aetiology and risk assessment
a graphical approach to support uncertainty research. Arch Toxicol 91:2763–2780
analysis in assessing dietary exposure. Toxicol 18. Leist M, Ghallab A, Graepel R, Marchan R,
Environ Chem 92:641–654 Hassan R, Bennekou SH, Limonciel A, Vinken
9. European Food Safety Authority, Panel on M, Schildknecht S, Waldmann T, Danen E, van
Food Additives and Nutrient Sources added Ravenzwaay B, Kamp H, Gardner I, Godoy
to Food (2017) Scientific opinion on the re- P, Bois FY, Braeuning A, Reif R, Oesch F,
evaluation of potassium nitrite (E 249) and Drasdo D, Höhme S, Schwarz M, Hartung T,
sodium nitrite (E 250) as food additives. EFSA Braunbeck T, Beltman J, Vrieling H, Sanz F,
J 15:4786–4943 Forsby A, Gadaleta D, Fisher C, Kelm J, Fluri

10. European Food Safety Authority, Pamel on D, Ecker G, Zdrazil B, Terron A, Jennings
Plant Protection Producys and their Residues P, van der Burg B, Dooley S, Meijer AH,
(2017) Scientific opinion on the investigation Willighagen E, Martens M, Evelo C, Mombelli
into experimental toxicological properties of E, Taboureau O, Mantovani A, Hardy B, Koch
plant protection products having a potential B, Escher S, van Thriel C, Cadenas C, Kroese
link to Parkinson’s disease and childhood leu- D, van de Water B, Hengstler JG (2017)
kaemia. EFSA J 15:4691–5016 Adverse outcome pathways: opportunities,
limitations and open questions. Arch Toxicol health related to the presence of 3- and
91:3477–3505 2-monochloropropanediol (MCPD), and their

19. European Food Safety Authority, Panel on fatty acid esters, and glycidyl fatty acid esters in
Additives and Products or Substances used in food. EFSA J 14(4426):159
Animal Feed (2016) Scientific opinion on the 21. European Food Safety Authority, Panel on
safety of lancer (lanthanide citrate) as a zoo- Contaminants in the Food Chain (2017)
technical additive for weaned piglets. EFSA Scientific Opinion on the risks to human
J 214(4477):10 and animal health related to the presence of

20. European Food Safety Authority, Panel on deoxynivalenol and its acetylated and modi-
Contaminants in the Food Chain (2017) fied forms in food and feed. EFSA J 15:4718.
Scientific opinion on the risks for human [345 pp]
Part III
Impact in Drug Discovery and Development

Chapter 11
Computational Toxicology and Drug Discovery

Catrin Hasselgren and Glenn J. Myatt
Abstract
The use of computational toxicology methods within drug discovery began in the early 2000s with appli-
cations such as predicting bacterial mutagenicity and hERG inhibition. The field has been continuously
expanding ever since and the tasks at hand have become more complex. These approaches are now strate-
gically integrated into the risk assessment process, as a complement to in vitro and in vivo methods. Today,
computational toxicology can be used in every phase of drug discovery and development, from profiling
large libraries early on, to predicting off-target effects in the mid-discovery phase, to assessing potential
mutagenic impurities in development and degradants as part of life-cycle management. This chapter pro-
vides an overview of the field and describes the application of computational toxicology throughout the
entire discovery and development process.
Key words Computational toxicology, Drug discovery, Hit identification, Lead identification, Lead
optimization
1 Introduction
The pharmaceutical industry generates a broad range of products

to treat and prevent diseases. The discovery of each new product
is currently estimated to cost over $2 billion and take approxi-
mately 14 years [1]. The cost is, to a large extent, a result of the
high attrition rate which is due to a variety of reasons, and the
industry’s inability to reliably predict these failures. Attrition
occurs throughout the drug discovery and development process;
however, the later a potential drug fails, the costlier it is. Failure
of drugs in late phases of clinical trials or withdrawal of a drug
from the market is particularly expensive, and the “fail early–fail
cheap” paradigm is thus widely accepted. Clinical efficacy and
safety are two of the most common reasons that drugs fail in the
later stages of development [2, 3].
To ensure patient safety, substantial in vitro and in vivo toxicity
testing is essential, and the preclinical safety package for a new
drug to gain approval for clinical testing is strongly driven by regu-
latory requirements. It is designed to cover all the relevant
233
234 Catrin Hasselgren and Glenn J. Myatt
e ndpoints such as reproductive and developmental toxicity, cardio-

toxicity, genetic toxicity, and carcinogenicity. It involves general
repeated dose toxicity studies in animals for identification of organ
toxicities, as well as specific studies designed to mechanistically
understand the relationship between the molecular structure and
biological effects or pathways. The amount of time and expense
(often tens to hundreds of thousands of dollars) to generate and
interpret in vivo results, as well as animal welfare considerations,
prohibits animal studies from being used unnecessarily. In early
stages of drug discovery, a limiting factor in experimental testing is
also the small amounts of compound that are available. These fac-
tors are driving the development of alternatives such as in vitro or
in silico methods. Both the cost of generating in vivo data and the
ethical considerations prompt a conscientious approach of maxi-
mizing value in the generated data.
In silico approaches, or “computational toxicology,” have
gained importance over the last 10–15 years and are now well
embedded into the safety assessment process in many pharmaceuti-
cal companies. Computational toxicology uses computer gener-
ated models to predict toxicological hazards and results are
generated quickly and cheaply with no need for test material or
animals, since the prediction is often based on the chemical struc-
ture alone. In silico approaches can be used either in a predictive
sense or in a problem-solving setting to aid in understanding an
underlying mechanism, for example. They are in many ways agnos-
tic to the origin of the data and can be applied throughout the
early discovery phase all the way to clinical studies and through
life-cycle management for certain applications. In vitro, in vivo, or
clinical data can be used and examples will be given in the follow-
ing sections of this chapter.
2 Computational Toxicology: Fit for Purpose
The starting point in the generation of most in silico models is a

high quality curated database of historical study results. This
involves collecting the results from the literature or from studies
reported in an online database or generated within an organiza-
tion. Typically, in vitro results are more easily obtained as they are
traditionally stored in databases and the results are relatively stan-
dardized, especially if they are collected within an organization and
are part of a large standard screen. Harmonization is however
required such as converting different types of measurements (e.g.,
Ki values vs IC50s) and this may sometimes pose problems. Care
also needs to be taken to assure that data generated in different
assays can be merged which is often not the case if the experimental
protocols are not sufficiently similar. In vivo results from toxicology
studies tend to be compiled in text reports, which makes extraction
Computational Toxicology and Drug Discovery 235
of the results more complicated. The raw data is often subject to

interpretation (histopathology results for example) which some-
times introduces a subjective component. Another issue which is
often missing, especially in publically available data, is information
pertaining to what exposure a particular finding was observed at.
Sources of clinical adverse effects information may also be useful in
computational toxicology applications. These data can be extracted,
curated and sometimes normalized from adverse events reporting
systems or from drug labels [4, 5]. Once databases have been con-
structed or data sets compiled, computational tools and models
can be generated that encode the often complex, structure–activity
(toxicity) relationships. They can then be used to predict a toxico-
logical effect or mechanism or to provide flags for potential issues
that require follow-up.
There are multiple methodologies to support in silico assess-
ment that are commonly used including quantitative structure–
activity relationship (QSAR) models and structural alerts, clustering
methods, and structural or biological similarity tools [6–8].
Sometimes, QSAR models are used as an ensemble to predict a
more complex endpoint. In addition, there are problems that
require more detailed 3D methods such as docking or molecular
dynamics. As with other applications of computational property
predictions, the most crucial components for a successful outcome
are the data and the right choice of methodology. Understanding
the quality and applicability of the data as well as the availability or
abundance of data is of utmost importance. This information, and
the nature of the question to be answered, will dictate what meth-
odologies are most well suited.
Drug discovery focuses on optimizing desirable biological
effects including potency, metabolic stability, and so on, while
avoiding adverse effects. To this end, computational toxicology
can support several important use cases (shown in Fig. 1) includ-
ing prioritizing lead series and lead optimization, analyzing
actual or theoretical metabolites, understanding in vitro or
in vivo results, and supporting the strategic selection of follow-
up toxicological tests. In addition, any first-in-man or new drug
marketing application can be supported with in silico results,
and in some cases (such as for drug impurities) can be used as a
regulatory test. Finally, in silico assessment can support the
manufacturing process including the assessment of impurities
and degradants, selection of alternative materials to use in the
manufacturing process (e.g., green chemistry), any necessary
extractables and leachables analyses, analysis of chemicals dis-
charged into the water systems (including the active pharma-
ceutical ingredient [API]), classification and labeling, and
worker safety assessments.
The following chapter is structured such that applications of
computational toxicology are discussed with respect to the differ-
Selection of product development candidates

Metabolite analysis As part of the weight of evidence
Rationalization of in vivo or in vitro study results in regulatory studies
Prioritizing testing of chemicals As a main regulatory submission
Target identification Hit identification Lead optimization Preclinical Clinical trial Market
Manufacture
Impurities and degradation products

Green chemistry and safer alternatives
Emergency response situations
Extractables and leachables
Ecotoxicology
Classification and labeling
Workers safety and occupational
health
Fig. 1 Drug discovery pipeline
ent phases of drug discovery and development where they are

commonly applied. This does not mean that they are exclusively
used at certain time points; organizational differences and the
nature of the drug project (therapy area, drug target, availability of
experimental data, for example) may also influence when particular
activities are performed, as is the case with experimental methods.
Often, simpler, or “filter-like,” methods are applied early on, and
more complex methods increase in usage in the later phases as the
tasks become less routine and more bespoke. Technical details of
modeling methodologies will not be covered here as these are dis-
cussed in detail in other chapters relating to specific areas of appli-
cation. The chapter concludes with a discussion on key issues and
future directions.
3 Hit Identification
3.1 Compound Early hit identification is focused on identifying molecules that

Collection bind to a specific biological target. This phase of drug discovery
Enhancement involves screening large numbers (thousands to hundreds of thou-
sands) of compounds in high throughput in vitro screens (HTS).
This step may be done on entire compound libraries or it may be
preceded by computational methods, such as virtual screening,
where the potential ligands are matched against structural require-
ments of the target by applying automated docking methods or
pharmacophore methods. Many companies will profile their com-
pound collections to have chemically diverse libraries with pre-
ferred physicochemical properties to ensure that future lead
compounds have a good chance of successfully advancing through
the discovery and development pipeline. Several publications have
discussed possible associations between physicochemical properties
and toxicity and have aimed at defining what property space

increases the chances of success. There are different views on both
the usefulness and precisely where that space is [9–12]. It is, how-
ever, generally accepted that properties such as high lipophilicity
and reactivity should be avoided to reduce the potential for toxic-
ity. Structural moieties related to genotoxicity, such as particular
aromatic amines or functionalities linked to the formation of reac-
tive metabolites, are sometimes avoided, and this is most often
done using structural alerts (described below).
4 Lead Identification: Lead Optimization
4.1 Structural Alerts Once in vitro active compounds are identified, it is desirable to
and Warning Systems develop series around the hit molecules and to computationally
profile the series for hazard identification purposes. This is pref-
erentially done before synthesis to have an idea of what the series
liabilities are, or in the ideal case of having multiple series to pur-
sue, which one might be more favorable. This type of profiling is
typically done using QSAR methods or by matching against
structural alerts and most commonly involves predicting poten-
tial in vitro effects. Structural alerts for toxicology are used in
many different phases of drug discovery and development. They
were first introduced for genotoxicity related carcinogenicity by
Ashby and Tennant [13] and their initial alerts have since been
added to, modified and refined for genotoxicity by several groups
[14–18]. Structural alerts are suitable for any application where
there is knowledge of the relationship between a particular sub-
structure and a certain biological effect (or in this case, toxicity).
It is preferable that the alerts are based on mechanistic under-
standing relating the structure to the effects but sometimes they
are purely based on a hypothesis and are not confirmed. It is
important that alerts are defined as specifically as possible to avoid
generating too many false positive predictions by flagging very
generic substructures, when they are applied. In addition to
genotoxicity, reactive metabolite formation [19–21] and skin
sensitization [22, 23] (mainly for occupational toxicity) are two
endpoints where structural alerts are often used.
Structural alerts should be used as hazard identification flags
only and require assessment as to the relevance of the alert within
the compound they are contained in. They can be used on their
own but are better used in a weight-of-evidence scenario together
with, for example, a QSAR prediction for the same endpoint and
the assessment of structural analogs with available experimental
data (read-across) to understand better how predictive the alert is in
a specific chemical series. This type of tool can be used very early in
lead identification to profile compounds in a series, or across series,
but is also useful in lead optimization to check new experimental
hypotheses before synthesis. Some industrial applications of such

comprehensive tools have been described in Hasselgren et al. [8].
4.2 Secondary In vitro pharmacological profiling is performed to understand the

Pharmacology selectivity of the compound and the potential off-target effects that
Profiling may occur. Publications discussing important targets related to
toxicity have outlined the ones most commonly used [24] but it is
often the case that the further a compound moves in the discovery
process, the larger the in vitro panel used for screening. Panels usu-
ally include kinases, proteases, cytochrome P450s, and G-protein
coupled receptors (GPCRs), and the importance of using such sec-
ondary pharmacology profiling has been shown to be beneficial in
terms of avoiding clinical adverse events [24, 25].
One example of off-target profiling concerns cardiac related
off-target interactions. The most well-known interaction is the
binding to the voltage-gated ion-channel, Kv11.1 (hERG), protein
which is part of regulating the cardiac action potential. Blocking the
hERG channel has been linked to long QT syndrome [26, 27]
which has further been linked to torsades de pointes, a potentially
fatal condition [28]. Experimentally, a battery of in vitro assays is
screened for this application, and depending on the availability of
data, it is also common to build models to predict the inhibition
potential of other ion channels related to cardiotoxicity including
the sodium and calcium channels, in addition to hERG. The effects
of introducing a computational filter for hERG have been discussed
in the context of significantly reducing the number of highly potent
hERG binders that were synthesized in projects [8]. Kinases are
another target class where selectivity is highly desired and where
promiscuity in binding is common due to the highly conserved
ATP binding site across the kinome. For oncology applications, this
can be a useful characteristic in terms of overcoming drug resistance
and blocking multiple targets or pathways simultaneously. However,
off-target or nonselective kinase interactions have been linked to
both genotoxicity [29, 30], inducing chromosome damage, and to
cardiotoxicity [31, 32], highlighting the importance of thoroughly
understanding the pharmacological profile of drug candidates.
Computational profiling of a ligand can be done either by
building ligand based QSAR (or other) models for the targets of
interest, either as single models to be used for optimizing particu-
lar safety interactions, or as an ensemble of models to generate a
predicted profile that can be used to understand potential for
promiscuity in a more general sense. It is also common to perform
structural similarity searching, or data mining, by comparing the
query compound to compound with measured in vitro bioactivi-
ties, building on the well-established principle that similar com-
pounds will have similar pharmacological profiles [33]. Ideally, the
link between the in vitro profile and potential adverse events should
be made. An example of such a methodology, linking compound-
target and target-adverse event information, has been published by

Remez et al. [34]. In some cases, it may also be possible to use 3D
structures of proteins, if these are available, to further understand
potential interactions or to look for targets with similar binding
sites and thereby identifying when a small molecule may bind to
several targets with similar sites. Methods for off-target predictions
can be applied in both a prospective sense to predict potential risk,
or in a problem-solving capacity to aid in understanding the mech-
anistic basis for an in vivo finding. Schmidt et al. have published a
thorough review of available methods and described applications
for off-target predictions [35].
Most companies have individual preferences on the endpoints
that are considered relevant and these often follow the experimen-
tal strategies that are applied in later stages or are based on the
therapeutic areas they work in as well as prior experience of toxicol-
ogy issues within the organization. QSAR models applied in this
early stage are usually models built using cross-project data or
sometimes a mixture of proprietary and publicly available data and
are referred to as “global” models. The datasets are usually chemi-
cally diverse and the data may come from more than one particular
assay for an endpoint. The models are designed to be used with
novel chemical classes in contrast to “local” models that are built
from a congeneric series of compounds and are more typically used
in later stages when data has been generated within a drug project.
As mentioned earlier, when constructing global datasets, care
needs to be taken to ensure that the assays are comparable and that
the data can be merged. This is particularly important when using
data from public database sources as these may have been gener-
ated using several different protocols.
4.3 Predicting In an extension to looking at the pharmacological profile of a com-

In Vivo Effects pound and relating it to potential adverse events via the compound
structures, Muthas et al. published a methodology comparing only
the pharmacological profiles of compounds, irrespective of chemi-
cal structures [36]. The methodology compares the biological fin-
gerprints by comparing the measured in vitro outputs using
common similarity metrics. The basis of this tool is the concept that
similar in vitro profiles may result in similar manifestations in vivo,
even though the compounds themselves are not chemically similar.
This is a complement to the concept of similar chemical structures
binding to similar targets [33] as it is compound agnostic and can
generate otherwise unintuitive predictions. Mining in vivo data to
extract the toxicological findings for each compound enables the
link to be made between pharmacological profile and in vivo effects.
Comparing a query molecule’s in vitro profile to a database of
stored profiles enables the prediction of findings for the novel com-
pound. Processing this data and understanding at what effect level
findings are relevant as well as identifying true negatives and har-
monizing terminology are complexities that need to be addressed

through data curation.
In vivo effects may also be modeled through traditional QSAR
methodologies. An example is work done by Mulliner et al. on
predicting in vivo hepatotoxicity [37]. In this work, 3700 com-
pounds with liver findings in humans and/or animals were used to
build 21 different QSAR models based on specific mechanistic
classes of liver findings. The models can be used to predict poten-
tial in vivo effects in novel compounds. This work was extended by
Cross et al. [38] to support an understanding of the biological
significance of hepatotoxicity structural alerts. One hundred and
seventeen alerts were identified and associations between these
alerts and the biological activity (animal, clinical, and postmarket
adverse events data) and to elucidate where preclinical signals were
not predictive of human effects.
5 Development
During the development process, it is important to understand

whether there are any safety concerns as part of the manufacturing
of the pharmaceutical and prior to clinical trials. The syntheses of
any drug will invariably introduce impurities into the final product.
It is important to assess whether any of these impurities that are
present above a specified threshold pose any safety concerns. These
impurities are usually reviewed for their potential to be DNA-
reactive mutagens. The International Conference on Harmonization
(ICH) M7 guideline (“Assessment and control of DNA reactive
(mutagenic) impurities in pharmaceuticals to limit potential carci-
nogenic risk”) [39, 40] specifies that two (Q)SAR methodology
(when no adequate data is available) may be used to assess impuri-
ties that are present above the level specified in the guideline. One
methodology is an expert rule-based and the other a statistical-
based methodology, and both methodologies should predict the
results of the bacterial reverse mutation test.
If a drug is being administered as part of a medical device, such
as an inhaled aerosol, then further assessment of any chemicals that
may have leached from any of the devices materials (e.g., the
plastics) should be performed. In many cases a similar (Q)SAR
exercise as performed under the ICH M7 guideline is used. Also,
actual or predicted metabolites of the active ingredient are often
assessed in a similar manner.
It is critical to further assess materials that workers may be
exposed to as part of the production, transfer, storage, and delivery
processes. Computational approaches are regularly used for such
assessment using models for DNA-reactive mutagenicity and der-
mal/respiratory sensitization and irritation. Computational
approaches may support the selection of alternative chemicals used
in the manufacturing that have less impact on the environment, i.e.,

green chemistry. In addition, whenever chemicals are discharged
into the environment, they may be assessed for their potential eco-
toxicology using computational approaches such as QSAR models.
Finally, in emergency situations, for example when a chemical is
unexpectedly released into a production process, a fast assessment is
necessary and in silico tools can provide such a rapid response.
6 Future of Computational Toxicology
There appears to be two branching fields for future applications of

computational toxicology. The first applies to regulatory usage and
the second to applications generated toward systems toxicology and
more complex applications. In vitro and in vivo toxicity methods
are used with confidence throughout the entire drug discovery and
development process. The acceptance of such methods is, in part, a
result of the availability of generally accepted test guidelines from
organizations such as the Organization for Economic Co-operation
and Development (OECD) [41] that ensure such methods are per-
formed and documented in a reliable and consistent manner.
Regulatory usage of computational toxicology for drug impurities
was initiated by the acceptance of the ICH M7 guideline for poten-
tial mutagenic impurities. Further incorporation of in silico meth-
ods for simulation of the cardio action potential using in vitro
results is under discussion for the revision of the S7B guideline
relating to cardiotoxicity. It is likely that we will see off-target pre-
dictions based on various computational approaches being used in
a regulatory context in the future. This prompts further exploita-
tion of in silico tools for regulatory usage to develop in a stricter
and more transparent manner, much in the same way as the ICH
M7 application. The lack of standard protocols for in silico methods
is a factor in their current limited regulatory usage and without
such documented protocols, there is uncertainty around generally
accepted approaches. Questions like what type of model or models
should be used to predict what type of effects or mechanisms need
to be addressed. In silico methods, just like in vivo or in vitro meth-
ods, should be critically assessed and their applicability clearly
defined. Furthermore, a thorough understanding of how to assess
confidence in the results is required to support their extended use.
For example, in early phases where many candidates are being
screened a lower confidence result may be appropriate for prioriti-
zation and filtering, yet in the development phase a higher confi-
dence in the results is necessary to satisfy regulatory requirements.
Consequently, new initiatives to develop in silico toxicology
protocols for hazard assessment have been initiated [42]. These
protocols are being generated through a large international con-
sortium composed of developers and users of in silico technology
including regulatory agencies, government agencies, and multi-

ple industries (e.g., pharmaceuticals, pesticides, food, and cos-
metics). The objective for the consortium is to outline an
acceptable series of steps for generation of in silico assessments
for each major toxicological endpoint (e.g., genetic toxicity, car-
cinogenicity, and reproductive toxicity). These protocols will
ensure that any in silico analysis is performed and documented in
a consistent and transparent manner. The protocols will cover
how to plan, conduct the in silico analysis, perform an expert
review, and document the results.
The benefits of such protocols will only be realized once the
protocols are deployed. Data sharing initiatives such as the eTox
consortium [43] will be important in the future, but new solutions
to share information may also need to be developed to enable more
precompetitive collaboration. This will need to involve new inte-
grated toxicity databases, models, and tools for documentation
and expert review. To further support the development of such
new and improved in silico methods, it will be necessary to gener-
ate partnership and new technologies for sharing public and pro-
prietary information. One such example is the SAR fingerprinting
method where information concerning classes of compounds is
shared to support refining the structure–activity (toxicity) relation-
ships, leading to improved prediction of toxicity [18]. This method
avoids sharing data or structures that are classified as confidential.
In addition to the regulatory context applications, a holistic
view of data and how the complex biological interactions contrib-
ute to both efficacy and toxicity is required to solve some of the
more challenging problems. The field has moved from modeling
simple, single compound–target interactions to modeling entire
in vitro profiles and relating them to in vivo or clinical effects. As
we move forward, thorough understanding of underlying mecha-
nisms of toxicity as well as limitations of data will become even
more crucial as the questions become more complex.
6.1 Summary This chapter provides an overview of the use of different types of
and Conclusion computational toxicology methodologies associated with the
drug discovery and development process, including their use in
early stage of discovery to make strategic decisions concerning
research direction, and their use in development to support
safety assessment, regulatory submissions, and manufacture of
the drug product. The chapter also reviews some of the current
challenges and identifies a number of recent initiatives to over-
come such challenges, including how to increase the quality and
reproducibility of computational toxicology by incorporation of
good “in silico toxicology practice” documented as part of gen-
erally agreed protocols.
References
1. DiMasi JA, Grabowski HG, Hansen RW (2016) 12. Waring MJ, Arrowsmith J, Leach AR, Leeson
Innovation in the pharmaceutical industry: PD, Mandrell S, Owen RM, Pairaudeau G,
new estimates of R&D costs. J Health Econ Pennie WD, Pickett SD, Wang J, Wallace O,
47:20–33 Weir A (2015) An analysis of the attrition of
2. Kola I, Landis J (2004) Can the pharmaceu- drug candidates from four major pharma-
tical industry reduce attrition rates? Nat Rev ceutical companies. Nat Rev Drug Discov
Drug Discov 3:711–716 14:475–486
3. Cook D, Brown D, Alexander R, March R, 13. Ashby J, Tennant RW (1988) Chemical struc-
Morgan P, Satterthwaite G, Pangalos MN ture, Salmonella mutagenicity and extent of
(2014) Lessons learned from the fate of carcinogenicity as indicators of genotoxic car-
AstraZeneca’s drug pipeline: a five-dimensional cinogenesis among 222 chemicals tested in
framework. Nat Rev Drug Discov 13:419–431 rodents by the U.S. NCI/NTP. Mutat Res
4. Harpaz R, Callahan A, Tamang S, Low Y, 204:17–115
Odgers D, Finlayson S, Jung K, LePendu P, 14. Tennant RW, Ashby J (1991) Classification
Shah NH (2014) Text mining for adverse drug according to chemical structure, mutagenicity
events: the promise, challenges, and state of the to Salmonella and level of carcinogenicity of a
art. Drug Saf 37:777–790 further 39 chemicals tested for carcinogenic-
5. Administration FaD FDA Adverse ity by the U.S. National Toxicology Program.
Event Reporting System (FAERS) Mutat Res 257:209–227
(2017). https://www.fda.gov/Drugs/ 15. Kazius J, McGuire R, Bursi R (2005) Derivation
GuidanceComplianceRegulatoryInformation/ and validation of toxicophores for mutagenic-
Surveillance/AdverseDrugEffects/default.htm ity prediction. J Med Chem 48:312–320
6. Myatt GJ, Beilke LD, Cross KP (2017) 16. Benigni R, Bossa C (2008) Structure alerts
4.09 – In silico tools and their application. for carcinogenicity, and the Salmonella assay
In: Chackalamannil S, Rotella D, Ward SE system: a novel insight through the chemical
(eds) Comprehensive medicinal chemistry relational databases technology. Mutat Res
III. Elsevier, Oxford, pp 156–176 659:248–261
7. Mitchell JBO (2014) Machine learning meth- 17. Romualdo B, Cecilia B (2006) Structural alerts
ods in chemoinformatics. Wiley Interdiscip Rev of mutagens and carcinogens. Curr Comput
Comput Mol Sci 4:468–481 Aided Drug Des 2:169–176
8. Hasselgren C, Muthas D, Ahlberg E, Andersson 18. Ahlberg E, Amberg A, Beilke LD, Bower D,
S, Carlsson L, Noeske T, Stålring J, Boyer S Cross KP, Custer L, Ford KA, Van Gompel
(2013) Chemoinformatics and beyond. In: J, Harvey J, Honma M, Jolly R, Joossens E,
Chemoinformatics for drug discovery. Wiley, Kemper RA, Kenyon M, Kruhlak N, Kuhnke
Hoboken, pp 267–290 L, Leavitt P, Naven R, Neilan C, Quigley DP,
9. Luker T, Alcaraz L, Chohan KK, Blomberg N, Shuey D, Spirkl HP, Stavitskaya L, Teasdale
Brown D, Butlin R, Elebring T, Griffin MA, A, White A, Wichard J, Zwickl C, Myatt GJ
Guile S, St-Gallay S, Swahn B-M, Swallow (2016) Extending (Q)SARs to incorporate
S, Waring M, Wenlock M, Leeson P (2011) proprietary knowledge for regulatory pur-
Strategies to improve in vivo toxicology out- poses: a case study using aromatic amine muta-
comes for basic candidate drug. Molecules genicity. Regul Toxicol Pharmacol 77:1–12
21:5673–5682 19. Kalgutkar AS, Dalvie D, Obach RS, Smith DA

10. Hughes JD, Blagg J, Price DA, Bailey S, (2012) Pathways of reactive metabolite forma-
Decrescenzo GA, Devraj RV, Ellsworth E, tion with toxicophores/-structural alerts. In:
Fobian YM, Gibbs ME, Gilles RW, Greene N, Reactive drug metabolites. Wiley-VCH Verlag
Huang E, Krieger-Burke T, Loesel J, Wager T, GmbH & Co, KGaA, pp 93–129
Whiteley L, Zhang Y (2008) Physiochemical 20. Stepan AF, Walker DP, Bauman J, Price DA,
drug properties associated with in vivo toxi- Baillie TA, Kalgutkar AS, Aleo MD (2011)
cological outcomes. Bioorg Med Chem Lett Structural alert/reactive metabolite concept as
18:4872–4875 applied in medicinal chemistry to mitigate the

11. Muthas D, Boyer S, Hasselgren C (2013) risk of idiosyncratic drug toxicity: a perspec-
A critical assessment of modeling safety- tive based on the critical examination of trends
related drug attrition. Med Chem Commun in the top 200 drugs marketed in the United
4:1058–1065 States. Chem Res Toxicol 24:1345–1410
21. Blagg J, Abraham DJ (2003) Structural alerts kinase inhibitors. Acta Oncol (Stockholm,
for toxicity. In: Burger's medicinal chemistry Sweden) 48(7):964–970
and drug discovery. Wiley, Hoboken 33. Johnson M, Lajiness M, Maggiora G (1989)
22. Patlewicz G, Aptula AO, Roberts DW, Uriarte Molecular similarity: a basis for designing
E (2008) A minireview of available skin sensiti- drug screening programs. Prog Clin Biol Res
zation (Q)SARs/expert systems. QSAR Comb 291:167–171
Sci 27:60–76 34. Remez N, Garcia-Serna R, Vidal D, Mestres
23. Verheyen GR, Braeken E, Van Deun K, Van J (2016) The in vitro pharmacological pro-
Miert S (2017) Evaluation of in silico tools file of drugs as a proxy indicator of potential
to predict the skin sensitization potential of in vivo organ toxicities. Chem Res Toxicol
chemicals. SAR QSAR Environ Res 28:59–73 29:637–648
24. Bowes J, Brown AJ, Hamon J, Jarolimek W, 35. Schmidt F, Matter H, Hessler G, Czich A
Sridhar A, Waldron G, Whitebread S (2012) (2014) Predictive in silico off-target profil-
Reducing safety-related drug attrition: the use ing in drug discovery. Future Med Chem
of in vitro pharmacological profiling. Nat Rev 6:295–317
Drug Discov 11:909–922 36. Muthas D, Boyer S (2013) Exploiting pharma-
25. Whitebread S, Dumotier B, Armstrong D, cological similarity to identify safety concerns –
Fekete A, Chen S, Hartmann A, Muller PY, listen to what the data tells you. Mol Inform
Urban L (2016) Secondary pharmacology: 32:37–45
screening and interpretation of off-target activ- 37. Mulliner D, Schmidt F, Stolte M, Spirkl HP,
ities - focus on translation. Drug Discov Today Czich A, Amberg A (2016) Computational
21:1232–1242 models for human and animal hepatotoxic-
26. Zhou Z, Gong Q, Epstein ML, January ity with a global application scope. Chem Res
CT (1998) HERG channel dysfunction in Toxicol 29:757–767
human long QT syndrome. Intracellular 38. Cross KP, Hasselgren C, Myatt GJ (2017)
transport and functional defects. J Biol Chem Integrated in silico methods for predicting
273:21061–21066 human hepatotoxicity. Poster presented at the
27. Recanatini M, Poluzzi E, Masetti M, Cavalli society of toxicology, Baltimore, USA
A, De Ponti F (2005) QT prolongation 39. ICH (2014) M7 assessment and control of
through hERG K(+) channel blockade: current DNA reactive (mutagenic) impurities in phar-
knowledge and strategies for the early predic- maceuticals to limit potential carcinogenic
tion during drug development. Med Res Rev risk. http://www.ich.org/fileadmin/Public_
25:133–166 Web_Site/ICH_Pr oducts/Guidelines/
28. Viskin S (1999) Long QT syndromes and tor- Multidisciplinary/M7/M7_R1_Addendum_
sade de pointes. Lancet (London, England) Step_4_31Mar2017.pdf
354:1625–1633 40. ICH (2015) Addendum to ICH m7: assess-
29. Olaharski AJ, Gonzaludo N, Bitter H, ment and control of DNA reactive (mutagenic)
Goldstein D, Kirchner S, Uppal H, Kolaja K impurities in pharmaceuticals to limit poten-
(2009) Identification of a kinase profile that tial carcinogenic risk. http://www.ich.org/
predicts chromosome damage induced by small fileadmin/Public_Web_Site/ICH_Products/
molecule kinase inhibitors. PLoS Comput Biol Guidelines/Multidisciplinary/M7/M7_R1_
5:e1000446 Addendum_Step_4_31Mar2017.pdf
30. Kirchner S (2012) Kinases as antitargets in geno- 41. OECD (2017) OECD guidelines for the test-
toxicity. In: Polypharmacology in drug discovery. ing of chemicals. http://www.oecd.org/
John Wiley & Sons, Inc., Hoboken, pp 63–81 chemicalsafety/testing/oecdguidelinesforth-
31. Force T, Kolaja KL (2011) Cardiotoxicity of etestingofchemicals.htm
kinase inhibitors: the prediction and translation 42. Bower DC, Cross KP, Hasselgren C, Miller S,
of preclinical models to clinical outcomes. Nat Myatt GJ, Quigley PD (2017) In silico toxicol-
Rev Drug Discov 10:111–126 ogy protocols and software platforms. Poster
32. Orphanos GS, Ioannidis GN, Ardavanis AG presented at the EuroTox, Bratislava, Slovakia
(2009) Cardiotoxicity induced by tyrosine 43. http://www.etoxproject.eu/ (2017)
Chapter 12
Approaching Pharmacological Space: Events

and Components
Giulio Vistoli, Alessandro Pedretti, Angelica Mazzolari, and Bernard Testa
Abstract
With a view to introducing the concept of pharmacological space and its potential applications in investi-
gating and predicting the toxic mechanisms of xenobiotics, this opening chapter describes the logical rela-
tions between conformational behavior, physicochemical properties and binding spaces, which are seen as
the three key elements composing the pharmacological space. While the concept of conformational space
is routinely used to encode molecular flexibility, the concepts of property spaces and, particularly, of bind-
ing spaces are more innovative. Indeed, their descriptors can find fruitful applications (a) in describing the
dynamic adaptability a given ligand experiences when inserted into a specific environment, and (b) in
parameterizing the flexibility a ligand retains when bound to a biological target. Overall, these descriptors
can conveniently account for the often disregarded entropic factors and as such they prove successful when
inserted in ligand- or structure-based predictive models. Notably, and although binding space parameters
can clearly be derived from MD simulations, the chapter will illustrate how docking calculations, despite
their static nature, are able to evaluate ligand’s flexibility by analyzing several poses for each ligand. Such
an approach, which represents the founding core of the binding space concept, can find various applica-
tions in which the related descriptors show an impressive enhancing effect on the statistical performances
of the resulting predictive models.
Key words Conformational space, Property space, Binding space, Ensemble docking simulations,
Ligand mobility, Entropic factors
1 Pharmacological Space: Events and Components
The aim of this Introductory Chapter is to present and discuss the

somewhat fuzzy concept of Pharmacological Space, with a view to
offering readers of the volume a hopefully informative and struc-
tured view of the field. First, we define a Space as an ensemble of
observable phenomena (Events), of physical entities taken individ-
ually (Components) or interacting between them to exchange infor-
mation (Interactions). Given the current state of knowledge, our
presentation begins with a survey of pharmacological events, con-
tinues with a focus on components (agents, products, etc.), and
ends with some examples of interactions between molecular agents
245
246 Giulio Vistoli et al.
and functional biomacromolecules. A schematic summary of main

categories in pharmacological space is proposed in Table 1.
As shown, pharmacological events are conveniently divided
into (a) pharmacodynamic (PD) events, and (b) pharmacokinetic
(PK) events, namely the well-known ADME acronym [1–3].
These events are also depicted in Fig. 1, where the further fate of
the generated metabolites, namely their distribution, further
metabolism and excretion, is left implicit. In a schematic way,
pharmacodynamic events can be encapsulated as “What the xeno-
biotic does to the biosystem(s) involved.” As for pharmacokinetic
events (ADME), they do fit the broad definition: “What the biosys-
tem does to the xenobiotic.” How these PK events can be catego-
rized is suggested in Fig. 2 [4].
Figures 1 and 2 focus on events that say little about the nature
of the components in such events. While the left side of Fig. 1
specifies the general nature of the molecules involved as actors
(drugs and other xenobiotics) or products (metabolites), it is
silent about the biosystems involved. These, as is well known,
may range from the (a) molecular or macromolecular levels
(receptors, enzymes, transporters, and so on), to (b) functional
assemblies such as membranes, to (c) organelles, (d) cells and
monocellular organisms, (e) tissues, (f) organs, (g) multicellular
organisms, and (h) populations.
This book focuses mainly on the toxicity of drugs and other
xenobiotics, obviously including their metabolites when relevant.
Specifically, it aims at making a fruitful contribution to the explora-
tion, analysis and prediction of the interactions of xenobiotics with
biotargets, as well as the outcomes of such interactions. The com-
putational approaches developed toward these goals aim at unveil-
ing correlations based on structural features and physicochemical
properties of toxic agents. Such properties and features are, for
example, absolute configuration, conformational behavior, lipo-
philicity and electrostatic potential, and biochemical reactivity, as
illustrated here for nicotine taken as a convenient example (Fig. 3a,
b). These aspects are given attention in subsequent Sections.
Table 1
Main subdivisions of the pharmacological space
(1) Pharmacological Space = Pharmacodynamic (PD) space + pharmacokinetic (PK) space

(1.1) PD Space = ACT/TOXa Space = Events + components + interactions
(1.2) PK Space = ADMEb Space = Events + components + interactions
Activity and toxicity

a
Absorption, distribution, metabolism (i.e., biotransformation) and excretion

b
Approaching Pharmacological Space: Events and Components 247
Fig. 1 The subdivision of pharmacological events (a component of pharmacologi-

cal space) into pharmacodynamic (PD) events and pharmacokinetic (PK) events
Fig. 2 A schematic representation of the ADME events, extending from administration and absorption to distri-
bution and storage, to wanted and unwanted effects, and ending with elimination, namely metabolism (chemi-
cal elimination) and/or excretion (physical elimination). (Modified from ref. 4)
One of the major challenges faced by predictive methods in

toxicology is the relevance of toxicity models and the data they
yield. Indeed, there are various biochemical and biological mecha-
nisms by which a xenobiotic may elicit toxicity (see Table 2), and
Fig. 3 (a) Molecular properties of nicotine taken as an example, showing the structural/geometric features
of this xenobiotic. Note the conformational behavior of the compound around its single relevant rotor. The
conformation of lowest energy (M2) is a staggered form; the conformations of highest energy are the
eclipsed ones (M1) and (M3). (b) The distribution in space of the electrostatic field and the lipophilicity
potential are shown here for the M2-conformer (see a). These potentials necessarily vary in space with the
conformation considered
Table 2
The TOX space of drugs and other xenobiotics
Types Mechanisms of adverse drug reactions (ADRs)

(1) On-target ADRs Predictable in principle and generally dose-dependent.
Based on the pharmacology of the drug/xenobiotic and
its metabolite(s), often an exaggerated response or a
response in a nontarget tissue.
(2) Off-target ADRs Predictable in principle and generally dose-dependent.
Resulting from the interaction of the drug or a
metabolite with a nonintended target.
(3) ADRs involving reactive metabolites Predictable in principle and generally dose-dependent.
A major mechanism is covalent binding to
macromolecules (adduct formation) resulting in
cytotoxic responses, DNA damage, or hypersensitivity
and immunological reactions. A distinct (and
synergetic) mechanism is the formation of reactive
oxygen species (ROS) and oxidative stress.
(4) Idiosyncratic drug reactions (IDRs) Unpredictable, apparently dose-independent, and rare
(<1 case in 5000). They might result from a
combination of genetic and external factors, but their
nature is poorly understood. IDRs include anaphylaxis,
blood dyscrasias, hepatotoxicity, and skin reactions.
such mechanisms may lead to innumerable outcomes able to con-

fuse the computational scientist [5–8]. In fact, the North Face of
computational toxicology is the relevance and value of the biologi-
cal models used. Whatever the quality of computational methods
and protocols, the correlations, rationalizations and predictions
uncovered in such projects can only be less reliable—even if only
slightly less reliable—than the biodata used as input.
2 The Biological Role of the Conformational, Property and Binding Spaces
From a mathematical standpoint, search algorithms are computa-

tional procedures able to find the best possible solutions within a
well-defined space and selected according to a targeted quality
(or cost) function. Stated differently, these algorithms are able to
generate the solutions, which optimize a given function, the sim-
ple minimization (or maximization) of which is not feasible or
would not allow a comprehensive exploration of local minima (or
maxima) [9].
In computational chemistry, search algorithms have found sev-
eral relevant applications, among which conformational analyses
[10, 11] and docking simulations [12] are the most important ones.
In the former application, the solutions to be found are represented
by the preferred conformers a flexible molecule can assume, the driv-

ing quality function is the corresponding potential energy as com-
puted by quantum-mechanical or empirical approaches, and the
space into which the search is performed corresponds to the confor-
mational space the constraints of which are defined by the geometri-
cal features of the simulated molecules [13]. Docking simulations
similarly involve the search of the arrangements by which a given
ligand optimizes its interactions within a binding cavity (the so-
called ligand poses); the quality function is the free energy of interac-
tion as estimated by approximated scoring functions and the explored
space corresponds to a hypothetical binding space which combines
the geometrical features of the protein binding cavity plus the ligand
conformational space [14].
When considering the clear similarity between conformational
searches and docking simulations, one finds a noteworthy differ-
ence in the way by which the obtained results are commonly ana-
lyzed [15]. Indeed, conformational searches are used to generate a
set of different possible conformers for each simulated molecule
[16]. Clearly, greater attention is given to the lowest energy struc-
ture since it is the most representative conformer, but all local min-
ima are considered since they represent other conformations a
flexible molecule can assume [17]. The analysis of all possible min-
ima is particularly relevant in medicinal chemistry studies since the
conformer(s) involved in the interaction with the protein target do
not necessarily correspond to the global minimum [18]. The
capacity of a flexible molecule to assume different conformers is
reflected in its physicochemical and stereoelectronic profiles since a
flexible molecule may exhibit different property values for each
conformer-dependent property, the ensemble of which will repre-
sent the corresponding property space [19].
In contrast and even though docking simulations also generate
different ligand poses, attention is routinely focused on the best
pose only as selected from a scoring function, visual scrutiny, or
comparison with experimental data [see for example ref. 20]. The
resulting complex may be refined by diverse post-docking simula-
tions whose primary objective should be to explore the dynamic
and mutual adaptability between protein and ligand, while the
other computed poses are completely ignored. Hence, the stan-
dard paradigm by which docking simulations are performed can be
summarized as: one selected ligand plus one optimized target to
yield one best pose which will be refined into one final complex.
Such a final result could be used to predict a ligand’s bioactivity or
to discriminate between active and inactive compounds (as done in
typical virtual screening campaigns). This implies that the (some-
times subjective) criteria by which the best pose is selected among
the generated solutions play a crucial role in determining the reli-
ability of the obtained docking results, but this also implies that the
procedures by which the structures of both ligand and protein tar-

get are selected and prepared play a similarly critical role [21].
On these grounds and as schematized in Fig. 4, the following
sections will analyze how conformational, property and binding
spaces can influence the measured ligand bioactivity. In detail, a
first part will be focused on the effects on bioactivity of the confor-
mational and property spaces, which will be considered together
since they mutually depend on ligand flexibility. A second part will
analyze how docking simulations can be used to analyze the effects
of ligand and protein flexibility (along with their molecular diver-
sity) on the explored binding space and consequently how binding
space descriptors can be used to enhance the predictive power of
the resulting docking scores.
Fig. 4 A schematic representation of the logical relations between conformational, property and binding
spaces. In detail, the relations between conformational and property spaces encode the molecular sensitivity
which in fact can be similarly described also between conformational and property spaces. The relations
between the free and bound ligand spaces define the flexibility and property constraints a ligand experiences
upon binding and the relations between these constraints parameterize the ligand plasticity which can be seen
as an index of the residual mobility a ligand retains when bound to its target
3 Conformational and Property Spaces
As mentioned above, the typical output of a conformational search

takes the form of a set of representative conformers a given mole-
cule can assume. From a structural point of view, these energeti-
cally relevant conformers can be analyzed by monitoring a set of
geometrical descriptors among which the calculation of the corre-
sponding root mean square deviations (that is RMSD) of the
atomic coordinates is the simplest way to compare pairs of con-
formers [22]. The profile of the RMSD values also represents an
easy method to evaluate the molecular flexibility, the higher the
RMSD values the greater the molecular flexibility. As an aside, it
should be noted that, while physicochemical and stereoelectronic
descriptors have progressed amazingly in the last decades, very few
descriptors have been proposed for molecular flexibility. As a mat-
ter of fact, the number of rotatable bonds and the mentioned
RMSD values remain the most commonly used flexibility (and
similarity) descriptor even for large biomacromolecules [23, 24].
When considering conformer-dependent (3D) properties, a
specific value for each collected conformer can be computed and
the resulting set of property values constitutes the corresponding
property space. The capacity of a given flexible molecule to
exhibit different property values raises two major issues. The first
concerns the selection of the best property value to account for
such a variability, the property average (or weighted average)
being the simplest solution. The second issue is related to the
observation that property averages, while being more informative
compared to conformer-specific values, does not account for the
dynamic behavior a conformer-dependent property can exhibit.
To do this, conceptually different descriptors for the property
space have been developed. A first kind of property space descrip-
tors was intended to encode for the spread of the computed
property values, the property range being the most intuitive one.
To account for the ease with which a flexible molecule can vary in
its properties, a second kind of descriptors was proposed by defin-
ing the so-called molecular sensitivity concept. This concept
results from the observation that some molecules are able to
finely modulate their properties by modest conformational
changes, while other molecules seem to be unable to significantly
vary their properties even with marked conformational shifts. As
shown in Fig. 4, molecular sensitivity descriptors encode the rela-
tion between conformational and property spaces and can be cal-
culated by using slightly different equations even though their
general formula corresponds to a ratio between a property vari-
ability (for example, a property range) and a flexibility descriptor
(e.g., the RMSD average or the number of rotors, see below).
Conceivably, the property space descriptors, which encode the
dynamic behavior of a given flexible molecule and thus its capac-

ity to adapt spontaneously to the environment, might play fruit-
ful roles in rationalizing biological processes [25].
Even though property space descriptors also proved successful
in predicting ADME parameters [26], in rationalizing chemical
reactivity [27, 28] and in QSPR studies [29], attention is focused
here on their applications in describing molecular recognition
events, where they found fertile applications in predicting both
ligand selectivity as exemplified by a study involving a set of ligands
for the human adrenergic receptors, [30] as well as enzyme activity
as shown by a recent study focused on a set of butyrylcholinester-
ase substrates [31]. While pursuing clearly different objectives,
these studies revealed common features of the property space
descriptors. While being more informative, property averages do
not perform markedly better than conformer-specific property val-
ues when inserted into the predictive models. This negative result
can be explained by considering that property averages, while indi-
rectly accounting for property variability, do not convey informa-
tion on the dynamic behavior of a conformer-dependent property.
In contrast, property ranges and, to a minor extent, property sen-
sitivities proved successful in enhancing the predictive power of the
resulting correlative equations, thus confirming that the above
described dynamic profiles of these molecular properties play a key
role in all analyzed cases. Reasonably, these innovative property
space descriptors can describe the molecular flexibility in a more
informative way–compared to the available flexibility descriptors,
see above—thus accounting for the often disregarded entropic fac-
tors. Finally and in all studies carried out, sensitivity values based
on the number of rotatable bonds performed better than those
computed by using RMSD values, thus suggesting that the former
probably represents an optimal balance between intuitiveness,
meaningfulness and ease of calculation [27].
Besides being used to analyze in detail the results of conforma-
tional analyses and to develop novel ligand-based descriptors for
correlative models, the property space concept can also be used to
better characterize the conformational and physicochemical con-
straints a molecule experiences when inserted into environments of
different complexity. Such an analysis, which usually requires
molecular dynamics (MD) simulations, can be exploited to investi-
gate how a flexible molecule adapts itself in media of different
polarity as seen for acetylcholine inserted in both isotropic and
anisotropic media [32, 33]. Nevertheless and as depicted in Fig. 4,
this kind of analysis finds its most interesting application when
exploring the constraints a ligand experiences when interacting
with a biological target. Indeed and as demonstrated by many
experimental and in silico studies, the ligands retain part of their
flexibility during the interaction with the target and this mobility
can play key roles in determining their biological activity [34].
Such a flexibility allows the ligands to move within the binding

pocket by assuming multiple binding modes (see for example ref.
35). Even though such a mobility can be enthalpically disfavored
since it requires the continuous breaking and restabilization of the
ligand–protein interactions, it plays a key entropic role especially
for proteins endowed with flexible binding pockets [36]. Clearly,
such a mobility is reflected in the capacity to preserve part of the
ligand property space. On these grounds, two major questions
arise: firstly, is there a relation between the conformational and the
physicochemical constraints experienced during binding and, sec-
ondly, can these constraints (and the relative relations) have a role
in influencing the observed bioactivity.
A first study focused on the interaction between acetylcholine
and muscarinic receptors revealed that the constraints experienced
by acetylcholine in its property spaces are more marked than those
observed in the corresponding conformational space, thus empha-
sizing that the muscarinic receptors do not completely freeze
ligand’s mobility but allow more than one interacting conforma-
tion provided that they exhibit optimal property profiles [37]. In
other words, the well-known concept of bioactive conformation(s)
which inspired most pharmacophore mapping approaches should
be replaced by the concept of bioactive property value(s) which
optimizes the molecular recognition while involving different
ligand conformations.
In a subsequent study, [38] conformational and property
spaces of an extended set of ligand–protein complexes were com-
pared, revealing that the relation between the constraints observed
in the conformational and property spaces might depend on the
ligand bioactivity (see Fig. 4). Hence, the ratio between the
property constraints and the conformational constraints was

defined as the molecular plasticity concept since it encodes for the
capacity of a given ligand to adapt its properties to those of the
binding cavity while partly retaining its mobility. In detail, this
study highlighted the following findings:
(a) Receptor agonists and enzyme substrates show property con-
straints greater than conformational constraints;
(b) Inactive ligands or enzyme products show property constraints
equal to conformational constraints;
(c) Antagonists and inhibitors show property constraints lower
than conformational constraints.
These trends suggest that protein activation (as seen in case a)
is mostly driven by a precise property complementarity between
ligand and target while protein inactivation (i.e., case c) is often
due to a physical packing of the binding pocket. Conceivably, non-
binders are marginally and unsystematically constrained by a pro-
tein (i.e., case b).
These studies, and more generally all in silico investigations

dealing with ligand mobility within a binding site, are commonly
based on MD simulations. Nevertheless, docking simulations can
also be exploited to this end by extending the analysis to all gener-
ated poses without focusing attention only to the best one (as usu-
ally done). The fruitful opportunity of considering more than one
docking pose will be extensively discussed in the next section; here,
attention is focused on the possibility of using this approach to
explore the property and the conformational constraints exerted
by molecular recognition. It should be noted that docking simula-
tions commonly consider the ligand as flexible while the protein
target is often kept fixed, especially since accounting for protein
flexibility is hugely time-demanding even when flexibility is limited
to few residues within the binding site [39]. This means that prop-
erty space analysis as derived from MD simulations accounts for
mutual adaptability between both interacting partners, while the
analysis based on docking results can explore only the effects of
ligand flexibility, thus representing a clear approximation com-
pared to the MD-based studies. Notwithstanding this, and as the
next section will clearly show, considering all generated ligand
poses can represent a straightforward approach to evaluate the
ligand’s mobility within a binding cavity, so avoiding the signifi-
cant computational costs required by more complex simulations.
To this end, a recently published docking study of a set of
serotoninergic agonists was used retrospectively to investigate the
constraints exerted by their interactions with the human 5HT1a
receptor [40]. This study was selected for several reasons which
can be summarized as follows. Firstly, it includes ligands endowed
with a marked flexibility due to the eight rotatable bonds, which
characterize their common scaffold. Secondly, docking simula-
tions involved the receptors homology models in both active and
inactive states thus allowing the effects of different protein con-
formations to be explored. Finally, the simulated ligands include
chiral atoms thus allowing the influence of configuration factors
on the monitored constraints to be analyzed. Not to mention that
docking results afforded encouraging predictive models (see
below), thus emphasizing the reasonable reliability of these simu-
lations. For simplicity, the analysis was focused on the property
space parameters of virtual log P as computed by the MLP
approach [41] and of the molecular surface chosen as example of
physicochemical and geometrical descriptors, respectively. As
mentioned above, RMSD values were used as flexibility indices.
However and as recently proposed, [22] the comparison of dock-
ing poses and the relative RMSD analysis can be performed with
or without ligands’ superimposition. In the first case, the obtained
RMSD values encode for the sole conformational differences
between the compared poses, while the values as computed
without superimposition account for conformational plus

r ototranslational differences; hence, the differences between them

parameterize the rototranslational differences and can be used to
describe the ligand mobility within a binding pocket. For the
explored property spaces, the following descriptors were calcu-
lated both in the two receptor models and in vacuo (as derived
from clustered MonteCarlo procedures):
(a) Property averages.
(b) Property ranges.
(c) RMSD-based flexibility averages as described above.
(d) Property constraints calculated as the ratios between the prop-
erty ranges for the bound ligand and in vacuo.
(e) Flexibility constraints similarly calculated as the ratio between
the RMSD averages.
Table 3 collects the so-computed average values which allow
for some relevant considerations. In detail, the receptor interaction
does not vastly influence the property averages, while significant
effects are seen in the property ranges even though with different
trends. Indeed and as found in the previous study, the agonists
simulated here show physicochemical constraints greater than geo-
metrical constraints as confirmed by RMSD values which are
slightly reduced upon binding, while lipophilicity ranges are more
than halved compared to the vacuum. Surprisingly, surface ranges
increase when bound to the receptor compared to the vacuum.
Collectively, these results confirm that a flexible ligand retains a
significant part of its conformational space within a binding cavity,
and even that receptor interactions stabilize conformations which
are disfavored in vacuo as suggested by an increase in the surface
ranges. Nevertheless, all these allowed conformations have to pres-
ent a well-defined physicochemical profile as indicated by the
restrained lipophilicity range. When comparing the obtained
descriptors for active and inactive states, one may observe that,
although the rototranslational mobility markedly decreases within
the active binding cavity, a finding clearly ascribable to its reduced
size, the simulated ligands when bound to active state retain a con-
formational flexibility comparable (or even larger) than that exhib-
ited into the inactive state. These apparently contrasting results
suggest that a ligand can modulate conformational and rototrans-
lational freedom in an unrelated manner and can maintain a signifi-
cant flexibility even when strongly reducing its mobility. With
regard to property spaces, the descriptors computed in active
receptor confirm the trends already observed in the inactive state.
An interesting result is afforded by the lipophilicity averages which
emphasize that the active state recognizes slightly more lipophilic
ligand conformations, a result in line with the greater relevance of
hydrophobic contacts in the ligand recognition by the active state
as also revealed by docking simulations (see below).
Table 3
Average values (± standard deviations) for selected conformational and property spaces, as computed
by Monte Carlo searches in vacuo as well as by docking simulations in the binding cavity of the human
5HT1a receptors in its inactive and active states (rmsd and mobility are expressed in Å, surface values
(means and ranges) are expressed in Å2, while all constraint values are dimensionless)
Descriptors vacuum 5HT1a_inactive 5HT1a_active

Rmsd mean 2.57 ± 0.16 2.12 ± 0.23 2.31 ± 0.17
(S) Flexibility constraint – 0.86 ± 0.08 0.92 ± 0.08
(R) Flexibility constraint – 0.79 ± 0.06 0.89 ± 0.07
Flexibility constraint – 0.82 ± 0.07 0.90 ± 0.07
(S) Mobility mean – 3.26 ± 0.83 1.78 ± 1.28
(R) Mobility mean – 2.40 ± 0.68 0.62 ± 0.66
Mobility mean – 2.83 ± 0.78 1.26 ± 1.08
(S) Mobility constraint – – 0.49 ± 0.34
(R) Mobility constraint – – 0.27 ± 0.15
Mobility constraint – – 0.40 ± 0.31
(S) Surface mean – 462.4 ± 22.2 462.2 ± 23.5
(R) Surface mean – 465.2 ± 23.9 462.0 ± 21.8
Surface mean 479.5 ± 25.0 463.8 ± 22.6 462.1 ± 22.6
(S) Surface range – 25.7 ± 7.0 23.8 ± 5.7
(R) Surface range – 23.6 ± 5.1 20.7 ± 5.4
Surface range 19.9 ± 1.4 24.6 ± 5.7 22.3 ± 5.3
(S) Surface constraint – 1.29 ± 0.32 1.19 ± 0.25
(R) Surface constraint – 1.18 ± 0.28 1.04 ± 0.25
Surface constraint – 1.23 ± 0.26 1.15 ± 0.23
(S) logPMLP mean – 2.28 ± 0.51 2.32 ± 0.46
(R) logPMLP mean – 2.28 ± 0.47 2.37 ± 0.44
logPMLP_mean 2.21 ± 0.43 2.28 ± 0.47 2.34 ± 0.43
(S) logPMLP range – 0.61 ± 0.18 0.60 ± 0.17
(R) logPMLP range – 0.55 ± 0.15 0.49 ± 0.15
logPMLP_range 1.27 ± 0.16 0.58 ± 0.15 0.54 ± 0.14
(S) logPMLP constraint – 0.49 ± 0.17 0.48 ± 0.13
(R) logPMLP constraint – 0.44 ± 0.13 0.39 ± 0.11
logPMLP_constraint – 0.47 ± 0.14 0.43 ± 0.12
With regard to the configuration effects, the explored property

spaces reveal marked differences in the averages of the flexibility
descriptors, the (S)-enantiomers being markedly more mobile than
the (R)-enantiomers. Such geometrical and physicochemical dif-
ferences between enantiomers are possible only because the ligands
are inserted in an asymmetric environment such as a binding site.
Since the (S)-enantiomers has greater affinity, one may assume that
such a mobility has a positive impact on bioactivity reasonably due
to entropic factors. While showing more limited differences, such
a greater mobility is reflected in the greater ranges of both moni-
tored properties, thus suggesting that both a greater mobility and
a larger property variability can explain the greater affinity values of
the (S)-enantiomers.
In summary, the results so generated allow tentative answers to
be given to the above major questions raised. Firstly, target binding
seems to be able to constrain conformational and physicochemical
spaces in a rather unrelated manner by affecting the property spaces
clearly more than the conformational space. In other words, the
receptor is able to select well defined ligand properties rather than
specific (bioactive) ligand conformations. Clearly, this rule cannot
be true for all ligands and all protein targets but one may suppose
that this is conceivably true for entropy-driven interactions.
Although the limited number of ligands considered does not allow
the development of robust predictive models, the observed differ-
ences between stereoisomers seem to indicate that these
docking-based property space descriptors could find fruitful appli-
cations in accounting for often neglected entropic factors. What is
more, the binding effects on conformational and property spaces
as derived by docking simulations are in encouraging agreement
with those resulting from MD runs [see ref. 29]. While remember-
ing the mentioned approximations, this result suggests that dock-
ing simulations can be conveniently used to explore ligand mobility
by extending the analysis to all generated poses instead of focusing
on the best complex only. Such a straightforward approach can be
also exploited to enhance the reliability of the docking scores as
emphasized by the applications of the binding space concept on
which the next section is focused.
4 Binding Space as Assessed by Docking Simulations
4.1 Overview As mentioned in the Introduction, docking simulations are driven

by scoring functions, which correspond to approximated evalua-
tions of the free energy of interaction. Docking scores are used by
the search algorithms to generate and to prioritize the most rele-
vant poses, thus allowing a proper selection of the best docking
results [42, 43]. Hopefully, the docking scores of the so-selected
best poses should be utilized to develop correlative models and to
predict the bioactivity of novel molecules [44]. Moreover, these

predictive models may also play a methodological role, since they
allow a suitable optimization of both the input structures and the
docking parameters by iteratively modifying the computational
protocol as driven by the resulting predictive equations. Stated dif-
ferently, one may suppose that the optimal docking protocol is the
one producing the best predictive model. Regardless of the adopted
computational details, these correlative analyses are as yet focused
on the best poses by strictly adopting the above mentioned dock-
ing paradigm (see for example ref. 45).
On the contrary and as shown in Fig. 4, the basic concept of
the binding space is that docking simulations can be used to
describe a ligand’s mobility and to account for its structural diver-
sity by simultaneously considering more than one computed pose
[22]. In other words and while considering their intrinsic limita-
tions and particularly their static nature (as discussed above), dock-
ing simulations are supposed to be able to account for the dynamic
behavior of a given flexible ligand within a binding cavity, exactly
as the static search algorithms (e.g., MonteCarlo simulations) are
able to describe molecular flexibility in vacuo.
The ensemble of these multiple poses and computable docking
scores constitutes the corresponding binding space which can be
explored by descriptors conceptually similar to those already
developed for the property space. Hence, a binding space can be
defined by the following:
(a) Score averages (or weighted averages), which represent the
simplest way to parameterize the role of multiple binding
modes.
(b) Score ranges, which can be seen as an energy-based index of
ligand mobility.
(c) Score sensitivity values, which are obtained by subdividing the
score ranges for the three above-described RMSD values as
well as per the number of rotors.
(d) Score constraints, which can be derived only when comparing
different protein targets.
Besides analyzing all the poses generated by a single docking
search, the binding space concept can be also useful in treating the
docking scores from simulations involving different protein (as
typically done in ensemble docking) or ligand structures (e.g.,
when considering the role of different ligand tautomers). As sum-
marized in Fig. 5, the binding space can find relevant applications
in the analysis of multiple docking scores as obtained from:
(a) the best poses as derived using different ligand conformations;
(b) the best poses as derived using different ligand states;
(c) the best poses as derived using different protein conformations;
(d) the best poses as derived using different protein states;

(e) all generated poses as derived by a single docking simulation;
( f) the various combinations of the previous points.
Even though the case (e) is clearly the most innovative way in
which the binding space can be applied, the other possibilities
also play a meaningful role especially when synergistically com-
bined as indicated by point (f). When considering that docking
simulations are usually performed by inserting a flexible ligand
into a rigid protein structure, one may argue that the cases (c)
and (d) are the most interesting ones among the first four cases.
Since most ligands exhibit complex tautomeric and/or ionization
equilibria, also case (b) can play a relevant role to explore how the
resulting ligand states can influence the docking results. Even
though ligand flexibility is directly considered by most docking
algorithms, case (a) might be yet useful when considering that
the generated ligand geometries can be somewhat influenced by
the starting ligand conformations. Hence, the first case could be
relevant when treating quite complex ligands (e.g., peptides), as
well as to minimize the randomness, which unavoidably charac-
terizes most search approaches. Nevertheless and for simplicity,
the following sections will be focused on the three most produc-
tive cases, namely to simulate:
Fig. 5 A schematic representation of the different cases in which the binding space concept can be fruitfully
exploited. The cases differ for the number of ligand states (cases a and b), protein states (cases c and d) or
generated poses (case e) considered for each ligand. The last case corresponds to the various combinations
of the previous cases. Adapted from ref. 31 with permission of ACS, 2017
(a) Ligand structural diversity (Subheading 4.2, i.e., case b); (b)
protein flexibility (Subheading 4.3, i.e., case c); and (c) multiple
binding modes (Subheading 4.4, i.e., cases e & f).
4.2 Molecular A selection of the optimal ligand’s input structure plays a critical
Diversity and Binding role in determining the reliability of a docking simulation [13, 46].
Space of Ligands On the one hand, the choice of the starting ligand conformation
has a marginal role since ligand flexibility is directly considered by
most docking programs. Hence, the use of the lowest energy con-
formation as derived by canonical conformational searches is a
commonly adopted choice [47]. On the other hand, the selection
of the correct ligand state becomes vastly more critical especially
for ligands endowed with complex tautomeric and/or ionization
equilibria. In these cases, selection should fall on the ligand form,
which optimizes the interaction with the protein target while being
compatible with the physiological conditions [48].
Even though the effect of ionization equilibria on ligand affin-
ity can very often be accounted for by simply focusing attention on
the predominant forms at physiological conditions, [49] there are
relevant cases in which the simulated protein target seems to
preferentially interact with specific ionization states regardless of
the pKa values of the involved ligands [50]. Here, the comparison
of the docking results obtained by simulating different ionization
states as well as the calculation of the corresponding score averages
can allow the specific role of each state to be revealed. For example,
comparative docking studies on human carboxylesterases (CES1
and CES2) showed that the best predictive equations were devel-
oped by considering all simulated substrates in their neutral form,
while including the scores of the ionized species or the correspond-
ing score averages worsened the resulting correlative models [51,
52]. This finding, which is in line with the well-known preference
of metabolizing enzymes for lipophilic substrates, emphasizes that
the protein binding site can create specific microenvironments able
to markedly influence the ionization equilibria of the bound ligand
thus selecting specific ionization states [53]. Such a phenomenon
surely becomes more relevant when the simulated ligands have pKa
values very close to the physiological pH. In such cases, the various
ionization states are similarly populated and one might suppose
they are synergistically involved in the observed affinity [54].
Thus, a benchmarking docking study performed by Park and
coworkers [55] compared the performances of different protocols
amenable to estimating the role of distinct ionization states in
docking simulations. In detail, the study analyzed the predictive
power of the correlations as developed by using (a) only the pre-
dominant ionization state; (b) the ionization state affording the
best score and/or the best pose compared to the experimental one
(when available); (c) a weighted ensemble of all possible states.
Even though the results are influenced by the simulated proteins,
they revealed that the calculation of protonation ensembles afford

on average the greatest improvements in the predictive models,
while focusing attention on the predominant states does not neces-
sarily provide the best performing models. Interestingly, similar
results are also obtained when simulating the different protonation
states of the ionizable residues in the target proteins.
While the role of ionization equilibria is routinely investigated,
the effect of tautomeric states is less easily accounted for by docking
calculations despite their relevant role in molecular recognition as
repeatedly demonstrated by many studies [56]. Indeed, tautomers
differ in both physicochemical and stereoelectronic properties and
these differences clearly impact on ligand affinity, thus posing the
problem of how to simulate the tautomeric effects in any computa-
tional analysis [57]. Such a problem is particularly relevant when
considering that molecules showing tautomeric equilibrium repre-
sent about the 25% of all marketed drugs and this percentage is also
maintained in commonly utilized molecular libraries [58].
Docking simulations are usually focused on the tautomers
which best fit the features of the binding site, with a penalty value
which is usually added to the corresponding docking scores when
the selected tautomers do not correspond to the most populated
ones in the physiological conditions. The best interacting tautomers
can be selected according to experimental data or by performing
preliminary docking simulations [59]. Regardless of the criteria by
which these tautomers are chosen, most docking procedures are
designed by assuming that only one tautomer is responsible for the
observed bioactivity. Such an approximation might be justified if
the rate at which tautomers interconvert is slower than the biologi-
cal processes, while, if it is faster, one may suppose that multiple
tautomers can be involved in molecular recognition [47]. In this
context, the binding space concept can be conveniently exploited to
account for the simultaneous contribution of more than one tauto-
mer to bioactivity.
An illustrative example of how the binding space concept can
support the modeling of tautomeric species is offered by a recent
study focused on the carrier-mediated gastrointestinal absorp-
tion of a set of anthocyanins extracted from bilberry juice [60].
These glycosides are indeed recognized and transported by the
same carriers involved in the secondary active transport of glu-
cose. In detail, the Na+ glucose symporter (sGLT1) mediates the
apical transport between the intestinal lumen and the epithelial
cells, [61] while the glucose uniporter GLUT2 is involved in the
basal transport by which anthocyanins reach the blood [62]. As
shown in Fig. 6, anthocyanins are characterized by complex ion-
ization and tautomeric equilibria. Under acidic conditions, they
can exist in a flavylium cationic form which however appears to
be stable only within the gastric environment. At the intestinal
pH, the neutral quinoidal bases should be vastly predominant,
while at higher pH values they can also exist in the correspond-

ing anionic forms, which however should be substantially absent
in the intestinal lumen. Again, the neutral bases can exist in three
major tautomers and thus docking simulations were repeated by
considering these three forms within the 3D structures of the
two mentioned transporters [63]. Along with the prediction of
the gastrointestinal absorption of novel derivatives, docking sim-
ulations and the so generated predictive models were here also
utilized to evaluate the specific role of the three neutral tauto-
mers. For brevity, the here described results will focus on dock-
ing results obtained for sGLT1 only.
As reported above, one may assume that the most relevant
tautomer is that affording the best correlative model, but Fig. 6
clearly shows that all three tautomers allowed the generation of
reliable equations and in particular the first two tautomers (A1
and A2) produced quite similar equations in terms of statistical
parameters. This may indicate that more than one tautomer can
be involved in the recognition by sGLT1 and the calculation of
the corresponding docking score averages can meaningfully
account for this. Figure 6 reveals that the score averages permit
Fig. 6 Application of the binding space concept for accounting for the synergistic role of multiple tautomers as
exemplified by the interactions between anthocyanins and sGLT1
a remarkable enhancement of the statistics, while average values

obtained by considering only two tautomers do not provide
such encouraging results. This result can be seen as an indirect
confirmation of the simultaneous role of the three tautomers.
Even though less probable within the intestinal lumen, docking
simulations also involved the corresponding cationic and anionic
forms which however afforded clearly worse results even when
including their scores in the computed averages, thus suggest-
ing that ionized forms are not properly recognized by the
transporter.
In conclusion, the described examples underline that the
calculation of score averages even when they do not improve
the obtained predictive models, can still be useful in clarifying
the specific role of each possible state. As a matter of fact, the
assumption that only one preferred ligand state is always
involved in receptor recognition can lead to significant inaccu-
racies thus markedly worsening the predictive power of the
obtained results. In this context, the calculation of the binding
space parameters and in particular of the score averages can be
truly useful in understanding the specific role of each ioniza-
tion/tautomeric state.
4.3 Protein Flexibility As mentioned above, protein flexibility is rarely accounted for by
and Binding Space docking analyses and thus the selection of the optimal protein con-
formation represents a crucial step in all docking simulations [39].
As a preamble, it should be noted that considering protein
flexibility in docking simulations can result in two different sce-
narios. On one hand, docking calculations can involve ensembles
of protein structures, which slightly differ in the fine architecture
of their binding cavity [64]. These structural ensembles can include
diverse resolved protein structures or can be generated by various
MD/MM simulations [65]. These ensembles account for the lim-
ited conformational changes the binding site experiences when
interacting with ligands of different shape and size. In other words,
they allow the ligand-induced adaptability of the receptor to be
simulated a priori [66]. Ensemble docking simulations are substan-
tially performed by selecting the single protein structure which
affords the best complex for each docked ligand. The target selec-
tion can involve exhaustive and combinatorial docking simulations
or can be based on the similarity between simulated and cocrystal-
lized ligands, but simulations rarely combine the docking results
from more than one protein structure for each ligand [see for
example ref. 67] An exception is represented by a recent bench-
marking analysis which proposed to consider multiple protein
structures and to generate multiple ligand poses which are clus-
tered and scored based on the Ruvinksy Colony Entropy approach
[68]. However, such a method did not reveal increased
erformances when compared with canonical approaches based on

p
a single protein structure [69].
On the other hand, docking simulations can involve markedly
different protein states, the relevance of which derives from the
observation that upon binding most proteins exhibit large confor-
mational changes by which they shift from inactive to active states
[70]. Very often, such a conformational activation involves a clo-
sure of the overall folding in which also the binding site reveals a
narrower architecture around the bound ligand. For example, the
comparison of the resolved GPCR structures revealed that they can
assume distinct active and inactive conformations which differ in
the arrangement of their binding site and in the induced (or
blocked) signaling pathways [71]. Active states are characterized
by a set of common rearrangements in the TM helices contacting
the G proteins, which mostly result in an outer movement of the
TM6 helix, and by a narrower binding cavity depending on a set of
rearrangements of a “transmission switch” located in the TM3–
TM5–TM6 helices [72]. The recent increase in the number of
resolved GPCR structures in their active states allows both an
increasingly detailed knowledge of the processes modulating
GPCR activation and a reliable homology modeling of the as yet
unresolved GPCRs in their corresponding active states [73, 74].
While ensemble docking simulations account for the mutual adapt-
ability between the interaction partners, simulating distinct protein
states would reveal different interaction patterns which might
allow a better rationalization of the intrinsic activity of the simu-
lated ligands. In other words, one may figure out that agonists or
substrates can interact with both active and inactive states, while
antagonists or inhibitors should interact only with inactive pro-
teins. However and even though the comparison of docking results
from different protein states should allow a more precise character-
ization of the simulated ligands, most studies are focused on a
single protein state suitably selected based on the ligand’s intrinsic
activity (see for example [75]).
The following examples will illustrate how the binding space
concept can be conveniently employed to simulate different pro-
tein structures both for ensemble docking analyses and when con-
sidering distinct protein states. With regard to ensemble docking,
a recent study focused on butyrylcholinesterase (BChE) substrates
and based on simulations involving five resolved protein structures
showed that the inclusion in the correlative models of the score
averages as computed by considering the best poses obtained for
the five simulated enzyme structures allows the generation of cor-
relative models which are comparable with those obtained by using
the best performing single structure (see blue bars in Fig. 8) [31].
Such a result suggests that the predictive power of docking scores,
as calculated by a reliable target structure, is not worsened when
combined with docking scores arising from other less effective
Fig. 7 Application of the binding space concept for accounting for the different protein states as exemplified
by the interactions between a set of dioxane-based serotoninergic agonists and the human 5HT1a receptor in
its active and inactive conformations
rotein structures and also suggests that the calculation of the

p
score averages can be seen as a straightforward procedure to opti-
mize the corresponding predictive power while avoiding system-
atic and combinatorial analyses. What is more, this study revealed
that the inclusion of the corresponding score ranges and sensitivity
values is able to markedly improve the statistics of the obtained
models. One may suppose that these binding space descriptors
account for the entropic factors thus explaining their beneficial role
in the predictive studies based on ensemble docking simulations.
The docking study based on the serotoninergic agonists,
already utilized to analyze the constraints experienced by the
ligands when bound to their protein target (see above), offers an
insightful example of the potential of the binding space concept
when simulating distinct protein states [31]. In detail, docking
simulations involved the 5HT1a receptor as modeled in its inactive
and active states by using homology techniques and by combining
purposely chosen active and inactive templates. Figure 7 shows
that the docking simulations in both receptor states allowed the
generation of satisfactory predictive models including only the pri-
mary ChemPLP score. More importantly, a markedly better pre-

dictive model can be developed when inserting into the same
equation the scores from both receptor states. In fact, a similarly
satisfactory relationship can be derived by considering the corre-
sponding score averages as has previously been done. Nevertheless,
one may suppose that using score averages is appropriate when
combining the docking results from multiple (similar) protein
structures since one may assume that their contribution to bioac-
tivity is comparable. In contrast, one may hypothesize that distinct
protein states differently contribute to bioactivity and thus the cor-
responding docking scores should be separately inserted into the
relationships. Not to mention that calculation of score averages
becomes mandatory when considering several protein structures
since the inclusion of the specific docking scores as single variables
in an equation would be statistically unfeasible.
Figure 7 also shows that the comparison of the docking scores
as computed for the inactive and active states revealed that the latter
are, on average, characterized by more stable complexes (as evi-
denced by the primary ChemPLP scores). When decomposing the
overall score into its major components, Fig. 7 evidences that the
“active” complexes are stabilized by stronger van der Waals contacts
while the ionic interactions appear to be somewhat worsened. These
results can be explained by considering that a narrower active bind-
ing cavity promotes the nonpolar contacts which require very short
interatomic distances at the expense of the salt bridges which, in
turn, play a key role during the initial activation process. Based on
these observations, the score differences between the two simulated
states were included in the correlative equations affording a further
enhancement of the corresponding statistics thus suggesting that
the capacity of a given ligand to adapt itself to the different protein
states beneficially contributes to ligand affinity.
In conclusion, the binding space concept appears to be par-
ticularly meaningful when considering distinct protein conforma-
tions in docking analyses. For ensemble docking simulations, score
averaging represents a very simple procedure to combine the con-
tribution of several protein structures avoiding time-consuming
combinatorial analyses. More importantly, score ranges and
sensitivity values allow a remarkable statistical improvement of the
obtained predictive models presumably because they are able to
take the entropic factors into consideration. Moreover and since
the range values always exert a negative contribution in the consid-
ered relationships, one may argue that they encode for destabiliz-
ing effects which a ligand experiences to parallel the protein
flexibility. When considering that an extremity of the range corre-
sponds to the best score, the score ranges also parameterize how
much a given ligand has to worsen its interactions to adapt itself to
protein mobility. When simulating distinct protein states, the
binding space concept has proven to be successful in revealing the
specific contribution of each state and how a ligand adapts its con-
tacts during receptor activation.
4.4 Ligand Mobility The capacity of a ligand to assume multiple poses within a binding
as Described site is finding countless experimental confirmations. Such a phe-
by Binding Space nomenon is usually studied by various MD simulations which allow
the ligand mobility within a binding cavity to be dynamically mon-
itored [see for example ref. 76]. In contrast and although the
ligands are routinely considered as flexible, docking simulations are
considered to be unsuitable to investigate ligand mobility due to
their static nature and the fact that protein structures are kept
fixed. Even though the contribution of protein flexibility cannot be
clearly accounted for by docking simulations [77], one may ask
whether the solely ligand flexibility can have a role in describing
ligand mobility. As mentioned above, the basic idea of the binding
space concept is that ligand mobility can be simulated by taking
into consideration all generated poses instead of focusing attention
on the best pose only and the scores of these poses can be analyzed
by calculating the discussed binding space descriptors. Such an
approach was preliminarily tested on a comparative docking study
focused on BChE substrates and involving the same five resolved
protein structures already seen in the previously reported study
[31]. Figure 8 reports the statistics of the thus developed equa-
Fig. 8 Effect of the binding space parameters (i.e., mean scores, score ranges, and sensitivities) on the reli-
ability of the predictive models as parameterized by the r2 values by considering minimized or nonminimized
complexes for the five utilized BChE structures plus the overall parameters. Adapted from ref. 31 with
permission of ACS, 2017
tions revealing that averaging the scores of all generated poses

allows for a clear enhancement compared to the corresponding
equations based on the best scores. This means that combining the
scores of the best pose with those of the other poses has a beneficial
role in the correlative performances thus indicating that the other
included poses do not represent incorrect results but instead they
are alternative binding modes which beneficially contribute to the
observed bioactivity. Also, Fig. 8 evidences that the inclusion of
ranges and sensitivity values as computed by considering all gener-
ated poses affords a further enhancement of the corresponding sta-
tistics. Notably, the ranges and sensitivity values positively
contribute to bioactivity probably as they here encode for the
capacity of a ligand to retain part of its mobility even within the
binding site. Overall, these results emphasize a very remarkable
difference in the role of ranges and sensitivity values when com-
puted by considering the best poses from different protein struc-
tures or different poses from a single protein structure. Indeed and
while beneficially contributing to the statistics of the resulting
models in both cases, in the former they have a negative impact on
bioactivity presumably as they encode for the pressure exerted by
protein flexibility on the bound ligands, while they have a positive
effect in the latter since they parameterize the mobility/flexibility
degree a ligand conserves upon binding.
Finally, the same study based on BChE substrates was utilized
to investigate the possibility of combining multiple poses from mul-
tiple protein structures (i.e., case f, see above). The results are con-
ceptually similar to those obtained for the ensemble docking analysis
based on the best poses only, since the overall score averages afford
equations with statistical performances similar to those developed
by using the score averages from the best performing single protein.
In contrast, overall ranges and sensitivity values do not induce sig-
nificant improvements in the resulting models presumably because
of the contrasting role that such descriptors play in the relationships
based on single (best) poses from multiple structures or in those
based on multiple poses from a single structure. This finding, which
can also explain the limited performances of the recently proposed
protocol based on the Ruvinksy Colony Entropy approach, empha-
sizes that the pose variability from ensembles of protein structures
encodes for the entropic pressure a ligand experiences to follow the
protein flexibility, while the pose variability within a single protein
structure describes the mobility degree a ligand can retain upon
binding. In other words, the former encodes for the entropic pen-
alty a ligand has to pay to interact with a target, while the latter
parameterizes the residual conformational entropy a ligand can
conserve upon binding.
When comparing the results obtained in the BChE docking
studies, one may observe that there are marked differences in the
Table 4
Key structural features of the simulated BChE structures plus their corresponding binding space
performances (as encoded by the r2 of the best model). Hydrophobic score and volume of the pockets
were calculated by FPocket [80], while the protein parameters were taken from PDB
BChE Hydrophobic score R-value free

PDB Id best r2 pocket Volume pocket (Å3) Resolution protein (Å) protein
4xii 0.74 33.74 3276.6 2.7 0.257
4bds 0.68 30.24 5168.8 2.1 0.209
4tpk 0.7 30.91 3178.9 2.7 0.239
2wij 0.73 32.15 4000.3 2.1 0.253
3o9m 0.7 33.21 4102.4 2.9 0.262
predictive performances offered by the use of binding space

parameters and notably the best performing protein structures
when considering all generated poses do not correspond with
those when using the best pose only (compare for example blue
and green bars in Fig. 8). An open question involves the structural
features able to explain these markedly different performances.
Table 4 compares some key structural features of the simulated
BChE proteins with the statistical performance of the binding
space parameters in the simulated protein structures (as encoded
by the r2 value of the best model). These relationships allow for
some meaningful considerations. Firstly, there is a direct relation
between performances and hydrophobicity of the binding cavity
(r = 0.74) probably as apolar contacts permit a certain degree of
mobility to the bound ligands, while polar interactions freeze the
binding modes thus reducing the role of binding space parame-
ters. Secondly, there is an inverse relation between performances
and volume of the cavity (r = −0.63) probably as too large cavities
allow the ligand to assume unrealistic binding modes which
worsen the reliability of binding space descriptors. Thirdly, the
performances increase with R-value (r = 0.71) which in well
resolved structures (such as those used here) can be seen as a mea-
sure of the protein flexibility which allows different binding modes
to be assumed [78]. Finally, there is no relation between perfor-
mances and protein resolution (r = 0.08) thus suggesting that
binding space parameters might be a successful strategy to enhance
docking simulations when using low resolution proteins.
Finally and considering together its various applications, the
binding space concept represent a powerful approach to take into
considerations the dynamic behavior of the molecular recognition

events by using docking simulations. Such a concept could find
meaningful applications for better analyzing (a) the often neglected
entropic factors; (b) the role of the ligand mobility in terms of
both conformational and property constraints; (c) the multiple
involvement of more than one ligand or protein states. Not to
mention that the binding space concept can also allow for a richer
integration of ligand- and structure-based techniques with poten-
tial benefits, in particular, to 3D-QSAR methods [79].
5 Conclusion
Within the scientific field of Drug Research and Development, a

methodology known as quantitative structure–activity relation-
ships has progressed immensely in recent decades. As discussed
repeatedly in the previous sections, QSAR investigations have
three objects of study, namely and firstly, “agents” such as sub-
strates, inhibitors, agonists and antagonists, secondly “targets”
such as enzymes, receptors, transporters and other functional
macromolecular complexes, and thirdly all types of interactions
between agents and targets (e.g., recognition, binding, activa-
tion, inhibition).
For many years, the time dimension was almost completely
absent from QSAR investigations, which have been restricted to
biochemical and pharmacological parameters and their single val-
ues for each test agent. In recent years, however, major progress
was made in expressing the molecular properties of agents. Thus,
it became clear that agents are not frozen statues but should be
compared to dancers and their fluid bodily movements and atti-
tudes. And indeed, the concept of property space allowed for
example the flexibility and conformational range of agents to yield
a rich set of data used as valuable descriptors in QSAR investiga-
tions. A comparable phenomenon is also occurring with the tar-
gets, whose conformational behavior and concomitant binding
and functional properties are attracting attention as they have
become attainable by experimental and computational studies.
In concrete terms, time is now ripe to integrate the temporal
dimension into the modeling of agent–target complexes. Stated
differently, the reductionistic approach which investigates sepa-
rately the properties of agents and those of targets should now be
replaced by a more integrative (the word “holistic” comes to mind)
approach. The challenge is now for QSAR to take into account for
capturing and describing the dynamics of the mutual adaptation of
agents and targets. This is what we try to bring to the attention of
readers in this chapter.
References
1. Testa B (1987) Pharmacokinetic and pharma- 17. Hatfield MP, Lovas S (2014) Conformational
codynamic events: can they always be distin- sampling techniques. Curr Pharm Des
guished? Trends Pharmacol Sci 8:381–383 20:3303–3313
2. Testa B, Krämer SD (2009) The biochemistry 18. Zheng Y, Tice CM, Singh SB (2017)
of drug metabolism – an introduction. Part 5: Conformational control in structure- based
metabolism and bioactivity. Chem Biodivers drug design. Bioorg Med Chem Lett
6:591–684 27:2825–2837
3. Testa B (2009) Drug metabolism for the per- 19. Vistoli G, Pedretti A, Testa B (2008) Assessing
plexed medicinal chemist. Chem Biodivers drug-likeness – what are we missing? Drug
6:2055–2070 Discov Today 13:285–294
4. van de Waterbeemd H, Testa B (2009) 20. Ballante F, Marshall GR (2016) An automated
Introduction: the how and why of bioavailabil- strategy for binding-pose selection and docking
ity research. In: van de Waterbeemd H, Testa B assessment in structure-based drug design.
(eds) Drug bioavailability – estimation of solu- J Chem Inf Model 56:54–72
bility, permeability, absorption and bioavailabil- 21. Salmaso V, Sturlese M, Cuzzolin A, Moro S
ity, 2nd edn. Wiley-VCH, Weinheim, pp 1–6 (2017) Combining self- and cross-docking as
5. Guengerich FP (2006) Cytochrome P450s and benchmark tools: the performance of
other enzymes in drug metabolism and toxic- DockBench in the D3R grand challenge 2.
ity. AAPS J 8:E101–E111 J Comput Aided Mol Des Aug 32(1):251–264.
6. Williams DP, Naisbitt DJ (2002) Toxicophores: https://doi.org/10.1007/s10822-017-
groups and metabolic routes associated with 0051-4
increased safety risks. Curr Opin Drug Discov 22. Kabsch W (1978) A discussion of the solution
Devel 5:104–115 for the best rotation to relate two sets of vec-
7. Pirmohamed M, Park BK (2001) Genetic sus- tors. Acta Crystallogr A34:827
ceptibility to adverse drug reactions. Trends 23. Veber DF, Johnson SR, Cheng HY, Smith BR,
Pharmacol Sci 22:298–230 Ward KW, Kopple KD (2002) Molecular prop-
8. Park BK, Pirmohamed M, Kitteringham NR erties that influence the oral bioavailability of
(1998) Role of drug disposition in drug drug candidates. J Med Chem 45:2615–2623
hypersensitivity: a chemical, molecular, and 24. Jamroz M, Kolinski A, Kihara D (2016)
clinical perspective. Chem Res Toxicol Ensemble-based evaluation for protein struc-
11:969–988 ture models. Bioinformatics 32:i314–i321
9. Hofmann KL (2000) Combinatorial optimiza- 25. Testa B, Vistoli G, Pedretti A, Bojarski AJ
tion: current successes and directions for the (2009) Atomic diversity, molecular diversity,
future. J Comput Appl Math 124:341–360 and chemical diversity: the concept of chemo-
10. Eliel E, Allinger N, Angyal S, Morrison G diversity. Chem Biodivers 6:1145–1151
(2007) Conformational analysis. Wiley, 26. Vistoli G, Pedretti A, Testa B (2009) Partition
New York, p 1965 coefficient and molecular flexibility: the con-
11. Agrafiotis DK, Gibbs AC, Zhu F, Izrailev S, cept of lipophilicity space. Chem Biodivers
Martin E (2007) Conformational sampling of 6:1152–1169
bioactive molecules: a comparative study. 27. Vistoli G, De Maddis D, Straniero V, Pedretti
J Chem Inf Model 47:1067–1086 A, Pallavicini M, Valoti E, Carini M, Testa B,
12. Pagadala NS, Syed K, Tuszynski J (2017) Aldini G (2013) Exploring the space of histi-
Software for molecular docking: a review. dine containing dipeptides in search of novel
Biophys Rev 9:91–102 efficient RCS sequestering agents. Eur J Med
13. Wales DJ, Scheraga HA (1999) Global optimi- Chem 66:153–160
zation of clusters. Cryst Biomol Sci 28. Vistoli G, Colzani M, Mazzolari A, Maddis
285:1368–1372 DD, Grazioso G, Pedretti A, Carini M, Aldini
14. Guedes IA, de Magalhães CS, Dardenne LE G (2016) Computational approaches in the
(2014) Receptor-ligand molecular docking. rational design of improved carbonyl quench-
Biophys Rev 6:75–87 ers: focus on histidine containing dipeptides.
Future Med Chem 8:1721–1737
15. Vajda S, Hall DR, Kozakov D (2013) Sampling
and scoring: a marriage made in heaven. 29. Vistoli G, Straniero V, Pedretti A, Fumagalli L,
Proteins 81:1874–1884 Bolchi C, Pallavicini M, Valoti E, Testa B
(2012) Predicting the physicochemical profile
16. Mitsutake A, Mori Y, Okamoto Y (2013) of diastereoisomeric histidine-containing
Enhanced sampling algorithms. Methods Mol dipeptides by property space analysis. Chirality
Biol 924:153–195 24:566–576
30. Vistoli G, Pedretti A, Villa L, Testa B (2005) 41. Gaillard P, Carrupt PA, Testa B, Boudon A
Range and sensitivity as descriptors of molecu- (1994) Molecular lipophilicity potential, a tool
lar property spaces in dynamic QSAR analyses. in 3D QSAR: method and applications.
J Med Chem 48:4947–4952 J Comput Aided Mol Des 8:83–96
31. Vistoli G, Mazzolari A, Testa B, Pedretti A 42. Weill N, Therrien E, Campagna-Slater V,
(2017) Binding space concept: a new approach Moitessier N (2014) Methods for docking
to enhance the reliability of docking scores and small molecules to macromolecules: a user’s
its application to predicting butyrylcholinester- perspective. 1. The theory. Curr Pharm Des
ase hydrolytic activity. J Chem Inf Model 20:3338–3359
57:1691–1702 43. Campagna-Slater V, Therrien E, Weill N,
32. Vistoli G, Pedretti A, Villa L, Testa B (2005) Moitessier N (2014) Methods for docking
Solvent constraints on the property space of small molecules to macromolecules: a user’s
acetylcholine. I. Isotropic solvents. J Med perspective. 2. Applications. Curr Pharm Des
Chem 48:1759–1767 20:3360–3372
33. Vistoli G, Pedretti A, Villa L, Testa B (2005) 44. Moitessier N, Englebienne P, Lee D, Lawandi
Solvent constraints on the property space of J, Corbeil CR (2008) Towards the develop-
acetylcholine. 2. Ordered media. J Med Chem ment of universal, fast and highly accurate
48:6926–6935 docking/scoring methods: a long way to go.
34. McAuley M, Timson DJ (2017) Modulating Br J Pharmacol 153(Suppl 1):S7–S26
mobility: a paradigm for protein engineering? 45. Gao YD, Hu Y, Crespo A, Wang D, Armacost
Appl Biochem Biotechnol 181:83–90 KA, Fells JI, Fradera X, Wang H, Wang H,
35. Alterio V, Di Fiore A, D’Ambrosio K, Supuran Sherborne B, Verras A, Peng Z (2018)
CT, De Simone G (2012) Multiple binding Workflows and performances in the ranking
modes of inhibitors to carbonic anhydrases: prediction of 2016 D3R grand challenge 2: les-
how to design specific drugs targeting 15 dif- sons learned from a collaborative effort.
ferent isoforms? Chem Rev 112:4421–4468 J Comput Aided Mol Des 32:129–142
36. Chakraborti S, Chakravarty D, Gupta S, 46. Sastry GM, Adzhigirey M, Day T, Annabhimoju
Chatterji BP, Dhar G, Poddar A, Panda D, R, Sherman W (2013) Protein and ligand prep-
Chakrabarti P, Ghosh Dastidar S, Bhattacharyya aration: parameters, protocols, and influence
B (2012) Discrimination of ligands with differ- on virtual screening enrichments. J Comput
ent flexibilities resulting from the plasticity of Aided Mol Des 27:221–234
the binding site in tubulin. Biochemistry 47. Oda A, Yamaotsu N, Hirono S, Watanabe Y,
51:7138–7148 Fukuyoshi S, Takahashi O (2015) Effects of
37. Vistoli G, Pedretti A, Testa B, Matucci R initial settings on computational protein–ligand
(2007) The conformational and property space docking accuracies for several docking pro-
of acetylcholine bound to muscarinic receptors: grams. Mol Simul 41:10–12
an entropy component accounts for the sub- 48. ten Brink T, Exner TE (2009) Influence of
type selectivity of acetylcholine. Arch Biochem protonation, tautomeric, and stereoisomeric
Biophys 464:112–121 states on protein-ligand docking results.
38. Vistoli G, Pedretti A, Testa B (2011) J Chem Inf Model 49:1535–1546
Chemodiversity and molecular plasticity: rec- 49. Krieger E, Dunbrack RL Jr, Hooft RW, Krieger
ognition processes as explored by property B (2012) Assignment of protonation states in
spaces. Future Med Chem 3:995–1010 proteins and ligands: combining pKa p rediction
39. Wong CF (2015) Flexible receptor docking for with hydrogen bonding network optimization.
drug discovery. Expert Opin Drug Discov Methods Mol Biol 819:405–421
10:1189–1200 50. Onufriev AV, Alexov E (2013) Protonation and
40. Del Bello F, Bonifazi A, Giannella M, Giorgioni pK changes in protein-ligand binding. Q Rev
G, Piergentili A, Petrelli R, Cifani C, Micioni Di Biophys 46:181–209
Bonaventura MV, Keck TM, Mazzolari A, 51. Vistoli G, Pedretti A, Mazzolari A, Testa B
Vistoli G, Cilia A, Poggesi E, Matucci R, Quaglia (2010) In silico prediction of human carboxy-
W (2017) The replacement of the 2-methoxy lesterase-1 (hCES1) metabolism combining
substituent of N-((6,6-diphenyl-1,4-dioxan- docking analyses and MD simulations. Bioorg
2-yl)methyl)-2-(2-methoxyphenoxy)Ethan-1- Med Chem 18:320–329
amine improves the selectivity for 5-HT(1A) 52. Vistoli G, Pedretti A, Mazzolari A, Testa B
receptor over α(1)-adrenoceptor and D(2)- (2010) Homology modeling and metabolism
like receptor subtypes. Eur J Med Chem prediction of human carboxylesterase-2 using
125:233–244 docking analyses by GriDock: a parallelized
tool based on AutoDock 4.0. J Comput Aided approach. J Comput Aided Mol Des
Mol Des 24:771–787 32:187–198
53. Aguilar B, Anandakrishnan R, Ruscio JZ, 68. Ruvinsky AM, Kozintsev AV (2006) Novel
Onufriev AV (2010) Statistics and physical ori- statistical-thermodynamic methods to predict
gins of pK and ionization state changes upon protein-ligand binding positions using proba-
protein-ligand binding. Biophys J 98:872–880 bility distribution functions. Proteins
54. Petukh M, Stefl S, Alexov E (2013) The role of 62:202–208
protonation states in ligand-receptor recogni- 69. Fradera X, Verras A, Hu Y, Wang D, Wang H,
tion and binding. Curr Pharm Des Fells JI, Armacost KA, Crespo A, Sherborne B,
19:4182–4190 Wang H, Peng Z, Gao YD (2018) Performance
55. Park MS, Gao C, Stern HA (2011) Estimating of multiple docking and refinement methods in
binding affinities by docking/scoring methods the pose prediction D3R prospective grand
using variable protonation states. Proteins challenge 2016. J Comput Aided Mol Des
79:304–314 32:113–127
56. Sayle RA (2010) So you think you understand 70. Hoeppner A, Schmitt L, SHJ S (2013) Proteins
tautomerism? J Comput Aided Mol Des and their ligands: their importance and how to
24:485–496 crystallize them. In: Ferreira SO (ed) Advanced
57. Katritzky AR, Hall CD, El-Gendy B-D, topics on crystal growth. Rijeka, InTech
Draghici B (2010) Tautomerism in drug dis- 71. Manglik A, Kruse AC (2017) Structural basis
covery. J Comput Aided Mol Des 24:475–484 for G protein-coupled receptor activation.
58. Martin YC (2009) Let's not forget tautomers. Biochemistry 56:5628–5634
J Comput Aided Mol Des 23:693–704 72. Tehan BG, Bortolato A, Blaney FE, Weir MP,
59. Milletti F, Vulpetti A (2010) Tautomer prefer- Mason JS (2014) Unifying family a GPCR
ence in PDB complexes and its impact on theories of activation. Pharmacol Ther
structure-based drug discovery. J Chem Inf 143:51–60
Model 50:1062–1074 73. Lu M, Wu B (2016) Structural studies of G
60. Baron G, Altomare A, Regazzoni L, Redaelli V, protein-coupled receptors. IUBMB Life
Grandi S, Riva A, Morazzoni P, Mazzolari A, 68:894–903
Carini M, Vistoli G, Aldini G (2017) 74. Sengupta D, Joshi M, Athale CA,
Pharmacokinetic profile of bilberry anthocya- Chattopadhyay A (2016) What can simulations
nins in rats and the role of glucose transporters: tell us about GPCRs: integrating the scales.
LC-MS/MS and computational studies. Methods Cell Biol 132:429–452
J Pharm Biomed Anal 144:112–121 75. Rodríguez D, Gao ZG, Moss SM, Jacobson
61. Wright EM, Ghezzi C, Loo DDF (2017) KA, Carlsson J (2015) Molecular docking
Novel and unexpected functions of SGLTs. screening using agonist-bound GPCR struc-
Physiology (Bethesda) 32:435–443 tures: probing the A2A adenosine receptor.
62. Yan N (2017) A glimpse of membrane trans- J Chem Inf Model 55:550–563
port through structures-advances in the struc- 76. Anselmi M, Pisabarro MT (2015) Exploring
tural biology of the GLUT glucose transporters. multiple binding modes using confined replica
J Mol Biol 429:2710–2725 exchange molecular dynamics. J Chem Theory
63. Smeriglio A, Barreca D, Bellocco E, Trombetta Comput 11:3906–3918
D (2016) Chemistry, pharmacology and health 77. Buonfiglio R, Recanatini M, Masetti M (2015)
benefits of Anthocyanins. Phytother Res Protein flexibility in drug discovery: from the-
30:1265–1286 ory to computation. ChemMedChem
64. Okamoto Y, Kokubo H, Tanaka T (2013) 10:1141–1148
Ligand docking simulations by generalized- 78. Brünger AT (1992) Free R value: a novel statis-
ensemble algorithms. Adv Protein Chem Struct tical quantity for assessing the accuracy of crys-
Biol 92:63–91 tal structures. Nature 355:472–475
65. Hospital A, Goñi JR, Orozco M, Gelpí JL (2015) 79. Nicolotti O, Giangreco I, Miscioscia TF,
Molecular dynamics simulations: advances and Carotti A (2009) Improving quantitative struc-
applications. Adv Appl Bioinform Chem 8:37–47 ture-activity relationships through multiobjec-
66. Lin JH (2011) Accommodating protein flexi- tive optimization. J Chem Inf Model
bility for structure-based drug design. Curr 49:2290–2302
Top Med Chem 11:171–178 80. Le Guilloux V, Schmidtke P, Tuffery P (2009)
67. Lam PC, Abagyan R, Totrov M (2018) Ligand- Fpocket: an open source platform for ligand
biased ensemble receptor docking (LigBEnD): pocket detection. BMC Bioinformatics
a hybrid ligand/receptor structure- based 10:168
Chapter 13
Computational Toxicology Methods in Chemical Library

Design and High-Throughput Screening Hit Validation
Kirk E. Hevener
Abstract
The discovery of molecular toxicity in a clinical drug candidate can have a significant impact on both the
cost and timeline of the drug discovery process. Early identification of potentially toxic compounds during
screening library preparation or, alternatively, during the hit validation process, is critical to ensure that
valuable time and resources are not spent pursuing compounds that may possess a high propensity for
human toxicity. This chapter focuses on the application of computational molecular filters, applied either
prescreening or postscreening, to identify and remove known reactive and/or potentially toxic compounds
from consideration in drug discovery campaigns.
Key words Molecular toxicity, Computational filter, High-throughput screening, Virtual screening,
Library design, Drug discovery
1 Introduction
Toxicity is now one of the leading causes of compound failure in

clinical drug development. A recent analysis of drug candidate
attrition from several large pharmaceutical companies showed that
safety and toxicity are now the greatest sources of failure [1]. It is
well known that the physicochemical properties of drug candidates
are associated with their toxicological outcomes [2–4], and decades
of medicinal chemistry experience have resulted in the identifica-
tion of specific functional groups and chemical motifs that are
strongly associated with toxicological issues in vivo (toxicophores),
often due to a higher propensity for chemical reactivity [5, 6]. The
application of this knowledge in the screening and removal of
potentially toxic compounds from consideration early in the drug
discovery process will be a critical factor in efforts to lower drug
candidate attrition rates and mitigate the high costs associated with
attrition late in the drug discovery process.
To date, a variety of techniques and methods have been
developed to predict potential toxicity in clinical drug candidates
275
276 Kirk E. Hevener
early in the discovery process. Two general methods are often

employed. The first involves the application of computationally
developed algorithms or models to the identification and elimina-
tion of potentially toxic compounds from screening sets. These
methods include quantitative structure-toxicity relationship
models (QSTR), machine learning and pattern recognition tech-
niques, toxicophore-mapping, and knowledge-based approaches
[7–15]. The second method commonly employed is the applica-
tion of chemical filters to screening sets designed to eliminate
compounds from screening that have undesirable physicochemi-
cal properties (non-leadlikeness) and/or possess known or sus-
pected toxicophores. A variety of filters have been developed,
published, and successfully employed including early non-leadlike
rules and properties [16], the Bristol-Myers Squibb published
filters [17], the Eli Lilly rules for identifying potentially reactive
compounds [6], the REOS filters (rapid elimination of swill)
developed by Vertex [18], and the Astra Zeneca filters [19].
Additionally, there are several online structural alert web servers
(e.g., ToxAlerts) that can be employed pre- or post-screening to
flag potentially problematic compounds [20].
As a common underlying mechanism of in vivo toxicity is
chemical reactivity, this chapter will primarily focus on the appli-
cation of specific filters for prediction of chemically reactive, and
thereby potentially toxic, functional groups or chemical motifs in
the development of screening libraries, prefiltering of libraries
prior to virtual (HTVS) or high-throughput screening (HTS),
and in post-HTS/HTVS hit confirmation studies. Chemical fil-
ters for leadlike properties (rule of five, etc.), drug-like character-
istics, and those that influence desirable pharmacokinetic
properties, will not be explicitly discussed herein [21–23].
Further, rules for predicting activity assay promiscuity (frequent
hitters), e.g., PAINS (pan- assay interference compounds), are
not explicitly discussed [24, 25]. While there may be some over-
lap, a frequent hitting compound does not necessarily result in
human toxicity; rather the hit frequency may be related to assay
interference. Finally, it should be noted that chemical reactivity
does not necessarily predict human toxicity, and many of the
functional groups detailed below have not explicitly been linked
to human toxicities, as such. However, their higher propensity
for reactivity with biological macromolecules, i.e., protein acyla-
tion, lead toward a higher predictivity toward adverse human
effects in general, as well as common assay interference, and it is
recommended that they be removed from screening libraries if
there is not sufficient reason to retain them, such as biochemical
screens for covalently acting agents.
Computational Toxicology Methods in Chemical Library Design and High-Throughput… 277
2 Applications
2.1 Screening Consideration of potentially reactive or toxic compounds should be

Library done as early as possible, including the screening library design
Development Tools stage. Many physicochemical properties are considered in the
design of a chemical library. Properties influencing “drug-likeness,”
favorable pharmacokinetic properties, molecular diversity (if a tar-
geted library is not the goal), potential assay interference, synthetic
accessibility, and chemical reactivity are commonly evaluated and
used in the decision process to retain or remove compounds when
building a small-molecule screening library [26]. Several tools are
available to aid researchers in the identification of potentially toxic
compounds while developing screening libraries, including predic-
tive tools and chemical filters based upon known or suspected reac-
tive or toxic chemical moieties.
Among the research tools that can be directly applied to chemical
library design, FAF-Drugs4 (Free ADME-Tox Filtering Tool) is a
useful online program that can be used to prescreen chemical libraries
during development or prior to high-throughput virtual or experi-
mental screening [27]. Key features include the ability of the software
to computationally predict ADME (Absorption, Distribution,
Metabolism, and Excretion) properties and to remove salts, counter-
ions, and duplicate compounds during the library design stage. The
server includes a large set of predefined toxicophore filters and allows
users to custom design their own filters as well.
The ZINC15 database presents a valuable resource to the drug
discovery community for screening library development [28].
Provided by the Irwin and Shoichet laboratories at University of
California, San Francisco, the ZINC15 database curates more than
120 million commercially available compounds from nearly 400
vendors or other catalogs for virtual screening. Compounds can be
downloaded in a docking-ready, 3D format prefiltered for drug-
likeness and to remove compounds with potentially problematic
chemical structures, including known toxicophores. Molecular
properties are annotated for each compound, allowing researchers
to download custom screening sets, such as fragment-like or lead-
like sets. Additionally, compounds with known activity have been
annotated with biological data allowing for researchers to design
focused or targeted screening libraries. A robust web interface and
search tool allows for rapid compound scaffold and similarity search-
ing. A potential disadvantage is that the database is somewhat dated
(2015), thus screening compounds downloaded may not be readily
available for purchase if vendor stocks have been recently depleted.
There have also been a large variety of software reported that
allow for in silico combinatorial library design, many of which
incorporate tools for the prediction of physicochemical properties
and toxicity risk assessment. Representative examples include
278 Kirk E. Hevener
proprietary software such as Schrödinger’s CombiGlide, QikProp,

and LigPrep tools; Biovia’s (Accelrys) Discovery Studio; the
Design Module of MedChem Studio™; and the open-source
ChemT package from BioChemCore [29]. Table 1 provides a list
of software and web-based tools for prediction of chemical reactiv-
ity and toxicity, many of which are applicable to the design of
chemical screening libraries.
2.2 Tools A wide variety of tools have been developed for the prediction of
for Prediction molecular toxicity, including predictive software, servers, and data-
of Reactivity/Toxicity bases. A representative selection is discussed here, with a more
complete listing, including system classification and web address
given in Table 1. One example is the eTOX project. Sponsored by
the European Innovative Medicines Initiative (IMI), the eTOX
project collected toxicological data from pharmaceutical industry
and academia into an online database that has been used to create
a series of toxicity prediction models, including the online toxicity
prediction system, eTOXsys [30]. Key features of this system are
the ability to query the database for toxicities related to the target
rather than the drug and chemical substructures showing a high
correlation with specific toxicities.
Several companies offer model-based toxicity prediction ser-
vices, including Leadscope’s® Toxicity Database comprising over
180,000 compounds and over 400,000 toxicity study results.
Leadscope® models include hepatobiliary, cardiological, and uri-
nary effects as well as developmental toxicity, genetic toxicity, and
neurotoxicity. Derek Nexus (Lhasa Limited) is another member-
based (proprietary) service that offers a rapid toxicity assessment
for compounds submitted.
ToxAlerts is an open-source, web-based server integrated with
the Online Chemical Modeling Environment (OCHEM) that col-
lects and stores toxicological data collected from existing literature
or submitted by users [20]. Structural alerts in the form of
SMARTS patterns are generated and can be used to screen indi-
vidual compounds for toxicity prediction or to prefilter libraries
during screening library design. Structure alerts currently exist for
endpoints including mutagenicity, carcinogenicity, skin sensitiza-
tion, and compounds that can form reactive metabolites.
Lastly, the US National Library of Medicine’s Toxicology Data
Network (TOXNET) is a publicly available resource that allows
users to screen specific compounds against a large variety of toxi-
cology databases (HSDB, CCRIS, GENETOX, etc.) and literature
references (TOXLINE, DART) [31, 32]. More of a data mining
than predictive tool, the TOXNET databases are useful for assess-
ing hit compounds from virtual or experimental screening during
the hit validation process to assess a compound for further advance-
ment. A note of caution is merited here: the absence of data for a
submitted compound does not necessarily mean there is no risk for
toxicity, rather the absence of a literature or a database report.
Table 1
Selected software and web-based tools for prediction of chemical reactivity/toxicity
Software/tool Classification Website or Web reference

ACD/Percepta Software http://www.acdlabs.com/products/percepta/predictors.php
ADMET Predictor™ Software http://www.simulations-plus.com/software/admet-property-prediction-qsar/
BIOVIA Discovery Studio Software http://accelrys.com/products/collaborative-science/biovia-discovery-studio/
CompuDrug/HazardExpert Pro Software http://www.compudrug.com/?q=node/35
Derek Nexus Software https://www.lhasalimited.org/products/derek-nexus.htm
Leadscope® Software http://www.leadscope.com/model_appliers/
MultiCASE/CASE Ultra Software http://www.multicase.com/case-ultra
Schrödinger/QikProp Software https://www.schrodinger.com/qikprop
Way2Drug GUSAR Software http://www.way2drug.com/mg/about.php
FAF-Drugs4 Server http://mobyle.rpbs.univ-paris-diderot.fr/cgi-bin/portal.py?form=admetox#forms::FAF-Drugs4
eTOXsys Server https://etoxsys.eu/etoxsys.v3-demo-bk/dashboard/
Mcule/Toxicity Checker Server https://mcule.com/apps/toxicity-checker/
PASS Online Server http://pharmaexpert.ru/Passonline/index.php
OpenTox/ToxPredict Server http://www.opentox.net/library/toxicity-prediction
ToxAlert Server https://ochem.eu/home/show.do
ToxiPred Server http://crdd.osdd.net/oscadd/toxipred/
VirtualToxLab™ Server http://www.biograf.ch/index.php?id=projects&subid=virtualtoxlab
Lazar Server https://lazar.in-silico.de/predict
Computational Toxicology Methods in Chemical Library Design and High-Throughput…
Leadscope® Database http://www.leadscope.com/product_info.php?products_id=78

NIH/NLM TOXNET Database https://toxnet.nlm.nih.gov/
279
SuperToxic Database http://bioinformatics.charite.de/supertoxic/index.php?site=home

280 Kirk E. Hevener
2.3 Structure Alerts The use of structure alerts for the identification of potentially
for Reactive and/or reactive or toxic compounds is regularly employed in both prefilter-
Toxic Functional ing of chemical libraries during the library design stage and postfil-
Groups tering of screening compounds during the hit validation stage
[33–36]. There are advantages and disadvantages to both prefilter-
ing and postfiltering strategies. For the former, the unilateral
removal of compounds containing known reactive or toxic func-
tional groups may result in the loss of a valid hit compound which
might be modified during the optimization process to remove the
offending chemical moiety. For the latter, there is an increased per-
sonnel and infrastructure expense related to the cost of screening
larger libraries that have not been prefiltered, as well as a potential
high cost of late-stage failure of a clinical candidate. One possible
solution is the use of customizable threshold cutoffs, an available
option for most structure alert algorithms, based on the number of
occurrences of a reactive substructure (e.g. no more than two nitro
groups). Alternatively, toxicophore filtering may be cautiously
employed postscreening to “flag” compounds identified with
potentially problematic groups that may warrant close attention.
Lastly, it is possible to employ a prescreen filter for the elimination
of particularly high-risk compounds, coupled with a postscreen filter
to flag lower-risk compounds, such as PAINS compounds or com-
pounds with less reactive functional groups occasionally still seen in
approved drugs (e.g., aniline and nitro groups).
In most cases, structure alert algorithms employ the use of
SMARTS (SMILES Arbitrary Target Specification) patterns to
identify predetermined chemical patterns in compounds and
chemical libraries. SMARTS, developed by Daylight Chemical
Information Systems, Inc., is a SMILES-based 2D line notation
that allows for the incorporation of variability, wildcards, atomic
properties, and connectivity in the search [37]. Using SMARTS,
atoms can be represented by atomic number, capital or lowercase
letters. As an example, carbon can be represented as C (aliphatic
carbon atom), c (aromatic carbon atom), or [#6] (any carbon
atom). Wildcard values can be included in SMARTS patterns to
represent * (any atom), a (aromatic atom), A (aliphatic atom), R
(ring membership), r (ring size), X (connectivity), charge, chiral-
ity, valence, mass, and several others. Values for atoms can be
coupled together for greater specificity using brackets and semico-
lons, as with [C;X2] (aliphatic carbon with two total bonds,
including implicit hydrogens). Variability at atomic positions can
be specified using brackets and commas, as with [O,N,S;X2;r3]
(oxygen, nitrogen, or sulfur with two total bonds in a three-mem-
bered ring system). Various symbols are used to represent atomic
bonds making connections between atoms, including - (single
bond), = (double bond), # (triple bond), : (aromatic bond), ~ (any
bond), @ (any ring bond), and others. A missing bond symbol is
interpreted as “single or aromatic,” which can be used to prevent
the possibility of missing an aromatic system described using double

bonds. Positional branching is represented, as with SMILES, by
parentheses and logical operators, ! (not) and & (and) can be
included for additional specificity. Recursive expressions and
component-level grouping can also be incorporated into SMARTS
patterns. Interested readers are referred to the DAYLIGHT website
for additional details and manuals [37].
3 Summary of Reactive Structure Filters
As mentioned above, several groups have published useful rules

and filters that have been successfully employed in the library
design and screening processes, and there have been several helpful
reviews published in the area. The seventh edition of Burger’s
Medicinal Chemistry, Drug Discovery, and Development includes
a very useful chapter summarizing structural alerts for toxicity
[34]. Structure alerts are categorized by compounds with inherent
chemical reactivity (e.g., acylating and alkylating groups), com-
pounds requiring metabolism to generate a reactive compound
(e.g., anilines, nitro-aromatics, hydrazines), compounds exhibiting
CYP450 interference (e.g., imidazoles, triazoles, 2,6-unsubstituted
pyridines), and compounds exhibiting a high DNA binding pro-
pensity (e.g., polycyclic aromatic compounds).
Several works have been published by Rishton discussing non-
leadlikeness and compound reactivity in drug discovery and include
several suggested functional groups that may lead to false positives
or potential toxicity [5, 16]. Reactive groups explicitly detailed
include sulfonyl, acyl and alkyl halides, anhydrides, aldehydes,
imines, epoxides, sulfonate and phosphonate esters, Michael accep-
tors, and several others. Pearce and coworkers at Bristol-Myers
Squibb published a useful set of filters for use in the design of high-
throughput screening libraries [17]. The filters designed by this
group were categorized as exclusion filters, flagging compounds
for removal from the library, and information filters, annotating
potentially problematic compounds but not removing them. Bruns
and coworkers at Eli Lilly recently published a similar set of filters
and described a demerit-based system employed at Lilly to identify
potentially problematic compounds in their screening sets. They
describe 275 rules for the identification of reactive compounds,
compounds that may interfere with assays (e.g., fluorescence,
absorbance, quenching), compounds with intrinsic protein damag-
ing capability (e.g., oxidizing agents and detergents), unstable
compounds, and compounds lacking drug-like features. The REOS
(Rapid Elimination of Swill) rules described by Walters and cowork-
ers at Vertex Pharmaceuticals includes a set of more than 200 func-
tional group filters, which include reactive and other undesirable
functional groups [18, 38]. Lastly, researchers at Astra Zeneca
282 Kirk E. Hevener
have recently published a set of filters (AZ-Filters) for use in library

design and screening hit validation [19]. The AZ filters include
patterns for “bland structures” (primarily non-druglike features),
reactive structures (potential toxicity), frequent hitters, dyes,
natural products, and others.
Table 2 includes a listing of commonly filtered functional
groups associated with reactivity or toxicity and their associated
Table 2
Commonly filtered reactive groups and their SMARTS patterns
Group Name SMARTS patternsa

1,2-dicarbonyls [C;X3](=O)([C;X3](=O))
Acyl halides [F,Cl,Br,I][C;X3]=[O,S]
Aldehydes [#6][C;H1]=[O;X1]
Alkyl halides, P/S halides, mustards [Cl,Br,I][P,S,C&H2&X4,C&H3&X4]
Alkyl sulfonates, sulfate esters [#6]O[S;X4](=O)=O
Alpha-halocarbonyls [#6][C;X3](=[O;X1])-[C;H1,H2]-[F,Cl,Br,I]
Alpha-beta unsaturated nitriles [#6]=[#6]C#N
Anhydride [O;X2]([CX3,S,P]=O)([CX3,S,P]=O)
Azides [#6][N;X2]=[N;X2]=[N;X1]
Beta-carbonyl quaternary nitrogen [C;X3](=O)[C][N,n;X4]
Beta-heterosubstituted carbonyls [O;X1]=C[C;H2]C[F,Cl,Br,I]
Carbodiimides [#6][N;X2]=[C;X2]=[N;X2][#6]
Diazos, diazoniums [#6]~[NX2;+0,+1]~[NX1;-1,+1,+0]
Disulfides [S;X2]~[S;X2]
Epoxides, thioepoxides, aziridines [O,N,S;X2;r3](C)C
Formates [O;X2][C;H1]=O
Halopyrimidines [F,Cl,Br,I]c(nc)nc
Heteroatom-heteroatom single bonds [O,N,S;X2]~[O,N,S;X2]
Imines (Schiff’s base) [N;X2]([!#1])=[C;X3][C;H2,H3]
Isocyanates, isothiocyanates [#6][N;X2]=C=[O,S&X1]
Isonitriles [#6][N;X2]#[C;X1]
Michael acceptors [#6]=[#6][#6,#16]=[O]
Nitroaromatic [c;X3][$([NX3](=O)=O),$([NX3+](=O)[O-])]
Nitrosos, nitrosamines [#6,#7][N;X2](=O)
(continued)
Table 2
(continued)
Group Name SMARTS patternsa

Perhalomethylketones [#6][C;X3](=O)[C;X4]([F,Cl,Br,I])([F,Cl,Br,I])
[F,Cl,Br,I]
Phosphines, phosphoranes [#6][#15&X3,#15&X5]([#6])~[#6]
Phosphinyl halides [P;X3][Cl,Br,I]
Reactive cyanides N#C[C&X4,C&X3]~[O&X1,O&H1&X2]
Sulfenyl, sulfinyl, sulfonyl halides [F,Cl,Br,I][$([SX2]),$([S;X3]=O),$([S;X4]
(=O)=O)]
Thiocyanate [#6][S]C#N
Thioesters *SC(=O)*
Thiourea [SX1]~C([N&!R&X3,N&!R&X2])
[N&!R&X3,N&!R&X2]
Vinyl halides [C;X2]=[C;X2]-[F,Cl,Br,I]
a
SMARTS patterns were generated using the Schrödinger Ligand Filtering utility and verified using SMARTSviewer
(http://smartsview.zbh.uni-hamburg.de/)
SMARTS patterns. The table represents functional groups common

to all the publications discussed above and is not intended to pres-
ent an all-encompassing list. Readers are referred to the resources
discussed above for additional structural alerts and filters that may
be applicable to their specific research. Representative chemically
reactive groups commonly included in structure alerts or filters
include acylating and alkylating agents, aldehydes and ketones,
Michael acceptors, reactive esters and thioesters, anhydrides,
imines, cyanates, and cyanides. Other known or suspected toxico-
phores that are often included in filters are quinones (or quinone-
forming groups), nitroaromatic and aromatic amines, nitrosamines,
acylhydrazides, thioureas, sulfur and nitrogen mustards, polycyclic
aromatic systems, triazenes, epoxides and aziridines, and aminopy-
rines [39, 40]. Additionally, many structure alert filters are
employed to remove groups associated both with non- drug-
likeness as well as potential toxicity, including the lanthanides,
actinides, noble gases, alkali metals, alkali earth metals, and transi-
tion metals. Lastly, common protecting groups, chemical reagents
and intermediates (triflate, esters of hydroxybenzotriazole,
Lawesson’s reagent, pentafluorophenyl esters, chloramidine,
triacyloxime, reactive cyanides, reactive azo groups, etc.) should
be considered for removal by filters prior to screening or in the
library design stage. It should be noted that there is considerable
overlap between the reactive and toxicophoric moieties discussed
here and groups associated with promiscuity, assay interference,
and general non-leadlikeness.
284 Kirk E. Hevener
4 Conclusions
The early identification of potential toxicity in drug discovery

campaigns is critical to prevent significant financial loss resulting
from pursuing compounds that may fail in late stages of develop-
ment. Methods introduced here include the application of resources
such as servers, databases, and predictive algorithms to identify or
flag potentially toxic compounds based upon their physicochemi-
cal properties and the use of structure alerts, or functional group
filters, to identify compounds based upon the presence of known
reactive functional groups or toxicophores. Both strategies can be
employed in the library design stage as well as the hit validation
stage and can be used prescreening or postscreening, depending
on the goals of the program and the resources available. Finally, the
strategies discussed here are applicable both to virtual and experi-
mental library screening.
Acknowledgments
This work was supported by NIH grant AI126755 and faculty

development program funding from UTHSC College of Pharmacy
to K.E.H.
References
1. Waring MJ, Arrowsmith J, Leach AR, Leeson 6. Bruns RF, Watson IA (2012) Rules for identi-
PD, Mandrell S, Owen RM, Pairaudeau G, fying potentially reactive or promiscuous com-
Pennie WD, Pickett SD, Wang J, Wallace O, pounds. J Med Chem 55:9763–9772
Weir A (2015) An analysis of the attrition of 7. Goh GB, Hodas NO, Vishnu A (2017) Deep
drug candidates from four major pharmaceuti- learning for computational chemistry.
cal companies. Nat Rev Drug Discov J Comput Chem 38:1291–1307
14:475–486 8. Gawehn E, Hiss JA, Schneider G (2016) Deep
2. Hughes JD, Blagg J, Price DA, Bailey S, learning in drug discovery. Mol Inform
Decrescenzo GA, Devraj RV, Ellsworth E, 35:3–14
Fobian YM, Gibbs ME, Gilles RW, Greene N, 9. van de Waterbeemd H, Gifford E (2003)
Huang E, Krieger-Burke T, Loesel J, Wager T, ADMET in silico modelling: towards predic-
Whiteley L, Zhang Y (2008) Physiochemical tion paradise? Nat Rev Drug Discov
drug properties associated with in vivo toxico- 2:192–204
logical outcomes. Bioorg Med Chem Lett
18:4872–4875 10. Bugrim A, Nikolskaya T, Nikolsky Y (2004)
Early prediction of drug metabolism and toxic-
3. Price DA, Blagg J, Jones L, Greene N, Wager ity: systems biology approach and modeling.
T (2009) Physicochemical drug properties Drug Discov Today 9:127–135
associated with in vivo toxicological outcomes:
a review. Expert Opin Drug Metab Toxicol 11. Segall MD, Barber C (2014) Addressing toxic-
5:921–931 ity risk when designing and selecting com-
pounds in early drug discovery. Drug Discov
4. Barratt MD (2000) Prediction of toxicity from Today 19:688–693
chemical structure. Cell Biol Toxicol 16:1–13
12. Gertrudes JC, Maltarollo VG, Silva RA,
5. Rishton GM (1997) Reactive compounds and Oliveira PR, Honorio KM, da Silva AB (2012)
in vitro false positives in HTS. Drug Discov Machine learning techniques and drug design.
Today 2:382–384 Curr Med Chem 19:4289–4297
13. Moroy G, Martiny VY, Vayer P, Villoutreix 27. Lagorce D, Sperandio O, Baell JB, Miteva MA,
BO, Miteva MA (2012) Toward in silico Villoutreix BO (2015) FAF-Drugs3: a web
structure- based ADMET prediction in drug server for compound property calculation and
discovery. Drug Discov Today 17:44–55 chemical library design. Nucleic Acids Res
14. Gini G (2016) QSAR methods. Methods Mol 43(W1):W200–W207
Biol 1425:1–20 28. Sterling T, Irwin JJ (2015) ZINC 15 – ligand
15. Singh PK, Negi A, Gupta PK, Chauhan M, discovery for everyone. J Chem Inf Model
Kumar R (2016) Toxicophore exploration as a 55(11):2324–2337
screening technology for drug design and dis- 29. Abreu RM, Froufe HJ, Daniel PO, Queiroz
covery: techniques, scope and limitations. Arch MJ, Ferreira IC (2011) ChemT, an open-
Toxicol 90:1785–1802 source software for building template-based
16. Rishton GM (2003) Nonleadlikeness and lead- chemical libraries. SAR QSAR Environ Res
likeness in biochemical screening. Drug Discov 22:603–610
Today 8:86–96 30. Sanz F, Carrio P, Lopez O, Capoferri L, Kooi
17. Pearce BC, Sofia MJ, Good AC, Drexler DM, DP, Vermeulen NP, Geerke DP, Montanari F,
Stock DA (2006) An empirical process for the Ecker GF, Schwab CH, Kleinoder T, Magdziarz
design of high-throughput screening deck fil- T, Pastor M (2015) Integrative modeling strat-
ters. J Chem Inf Model 46:1060–1068 egies for predicting drug toxicities at the eTOX
18. Walters WP, Ajay, Murcko MA (1999) project. Mol Inform 34:477–484
Recognizing molecules with drug-like proper- 31. Fowler S, Schnall JG (2014) TOXNET: infor-
ties. Curr Opin Chem Biol 3:384–387 mation on toxicology and environmental
19. Cumming JG, Davis AM, Muresan S, health. Am J Nurs 114:61–63
Haeberlein M, Chen H (2013) Chemical pre- 32. Wexler P (2001) TOXNET: an evolving web
dictive modelling to improve compound qual- resource for toxicology and environmental
ity. Nat Rev Drug Discov 12:948–962 health information. Toxicology 157:3–10
20. Sushko I, Salmina E, Potemkin VA, Poda G, 33. Zhu T, Cao S, Su PC, Patel R, Shah D, Chokshi
Tetko IV (2012) ToxAlerts: a Web server of HB, Szukala R, Johnson ME, Hevener KE
structural alerts for toxic chemicals and com- (2013) Hit identification and optimization in
pounds with potential adverse reactions. virtual screening: practical recommendations
J Chem Inf Model 52:2310–2316 based on a critical literature analysis. J Med
21. Lipinski CA (2000) Drug-like properties and Chem 56:6560–6572
the causes of poor solubility and poor permea- 34. Blagg J (2010) Structural alerts for toxicity. In:
bility. J Pharmacol Toxicol Methods Abraham DJ, Rotella DP (eds) Burger’s medic-
44:235–249 inal chemistry and drug discovery, 7th edn.
22. Veber DF, Johnson SR, Cheng HY, Smith BR, Wiley, Hoboken, pp 301–334
Ward KW, Kopple KD (2002) Molecular prop- 35. Smith GF (2011) Designing drugs to avoid
erties that influence the oral bioavailability of toxicity. Prog Med Chem 50:1–47
drug candidates. J Med Chem 45:2615–2623 36. Kazius J, McGuire R, Bursi R (2005)
23. Teague SJ, Davis AM, Leeson PD, Oprea T Derivation and validation of toxicophores for
(1999) The design of leadlike combinatorial mutagenicity prediction. J Med Chem
libraries. Angew Chem Int Ed Engl 48:312–320
38:3743–3748 37. SMARTS – a language for describing molecu-
24. Baell JB, Holloway GA (2010) New substruc- lar patterns. Daylight Chemical Information
ture filters for removal of pan assay interference Systems, Inc. http://www.daylight.com/
compounds (PAINS) from screening libraries dayhtml/doc/theor y/theor y.smarts.html.
and for their exclusion in bioassays. J Med Accessed 20 Dec 2017
Chem 53:2719–2740 38. Walters WP, Stahl MT, Murcko MA (1998)
25. Dahlin JL, Nissink JW, Strasser JM, Francis S, Virtual screening - an overview. Drug Discov
Higgins L, Zhou H, Zhang Z, Walters MA Today 3:160–178
(2015) PAINS in the assay: chemical mecha- 39. Williams DP, Naisbitt DJ (2002) Toxicophores:
nisms of assay interference and promiscuous groups and metabolic routes associated with
enzymatic inhibition observed during a sulfhy- increased safety risk. Curr Opin Drug Discov
dryl-scavenging HTS. J Med Chem Dev 5:104–115
58:2091–2113 40. Hakimelahi GH, Khodarahmi GA (2005) The
26. Dolle RE (2011) Historical overview of chemi- identification of toxicophores for the predic-
cal library design. Methods Mol Biol tion of mutagenicity, hepatotoxicity and car-
685:3–25 diotoxicity. J Iran Chem Soc 2:244–267
Chapter 14
Enalos Suite: New Cheminformatics Platform for Drug

Discovery and Computational Toxicology
Dimitra-Danai Varsou, Spyridon Nikolakopoulos, Andreas Tsoumanis,
Georgia Melagraki, and Antreas Afantitis
Abstract
In this chapter we present and discuss, with the aid of several representative case studies from drug discov-
ery and computational toxicology, a new cheminformatics platform, Enalos Suite, that was developed with
open source and freely available software. Enalos Suite (http://enalossuite.novamechanics.com/) was
designed and developed as a useful tool to address a variety of cheminformatics problems, given that it
expedites tasks performed in predictive modeling and allows access, data mining and manipulation for
multiple chemical databases (PubChem, UniChem, etc.). Enalos Suite was carefully designed to permit its
extension and adjustment to the special field of interest of each user, including, for instance, nanoinformat-
ics, biomedical, and other applications. To demonstrate the functionalities of Enalos Suite that are useful
in different cheminformatics applications, we present indicative case studies that include the exploitation
of chemical databases within a drug discovery project, the calculation of molecular descriptors, and finally
the development of a predictive QSAR model validated according to OECD principles. We aspire that at
the end of this chapter, the reader will capture the effectiveness of different functionalities included in the
Enalos Suite that could be of significant value in a multitude of cheminformatics applications.
Key words Cheminformatics, Computational toxicology, Drug discovery, QSAR
1 Introduction
The identification of the structural characteristics of compounds

that could be related to biological effects is a very critical proce-
dure within drug discovery or risk assessment that could help high-
light, from a pool of available compounds, the most suitable
candidates for a specific application [1]. In this direction, combina-
torial chemistry and high-throughput screening (HTS) techniques
significantly contributed in the production of a great number of
data including structures and their corresponding activity and/or
properties. This abundance of data, together with the need of sav-
ing time and experimental resources, led to the development of
well-structured chemical databases such as PubChem, a
287
288 Dimitra-Danai Varsou et al.
ell-
w organized database including among others, compounds
along with their activity information [2]. A systematic, reproduc-
ible, and sustainable exploitation of this chemical information can
be significantly promoted with the aid of cheminformatics in silico
tools [3].
Cheminformatics employ several strategies and methods to
deal with a range of problems, including, among others, data
mining and predictive modeling [4–6]. For a subsequent chemin-
formatics analysis, the chemical information extracted from chemi-
cal databases must often be transformed into mathematical
representations (known as molecular descriptors) that can be later
analyzed and used in data mining, modeling, virtual screening,
similarity analysis, and more [7]. Predictive modeling techniques,
including quantitative structure–activity relationship (QSAR)
modeling, can be then employed to correlate the structural prop-
erties of the compounds to their bioactivities or other properties in
an effort to discover patterns in the structure of molecules that can
help in explaining their activity profile [7–9]. These models can be
later used to obtain reliable predictions for the activities of novel
compounds [1, 7].
QSAR model building includes different steps such as data
curation, modeling, internal and external validation, and defini-
tion of the domain of their applicability, in order to afford robust
models with high predictive power according to OECD princi-
ples [8]. These steps can be performed using different software
tools or programming languages, which often require specific
programming skills.
An important challenge in cheminformatics is the employ-
ment of various computational techniques and developed models
in real drug discovery applications to eliminate the time spent
and the cost required for experimental procedures. This often
involves animal testing that can be significantly eliminated by
cheminformatics procedures [10]. Experimental researchers usu-
ally do not have a strong programming or computational back-
ground, and they are not expected to easily handle workflows and
scripts in different programming languages. Therefore, the devel-
opment of handy tools that can be used directly from the experi-
mentalists to predict properties of compounds, explore and
interpret easily the results, within just a few steps, is highly desired
[11, 12]. Enalos Suite software [13] aims to fill the gap of the
lack of user-friendly and ready-to-use tools for cheminformatics
and aspires to become a useful tool in computer-aided drug dis-
covery process, as it provides many functionalities including the
calculation of molecular descriptors, data mining and manipula-
tion from popular chemical databases like PubChem, UniChem,
SureChEMBL, IBM patents, and the employment and develop-
ment of custom-made predictive models.
Enalos Suite: New Cheminformatics Platform for Drug Discovery and Computational… 289
2 Enalos Suite
Enalos Suite consists of a simple and user-friendly interface familiar

for the most medicinal chemists that gives access to multiple chem-
ical databases (PubChem, UniChem SureChEMBL, IBM patents,
etc.), facilitating data retrieval for thousands of chemical com-
pounds (e.g., patent and vendor information), and allows descrip-
tor calculation and activity/property predictions with minimum
steps required.
The Enalos Suite workspace is composed of three tabs allow-
ing for structures to be imported in different formats:
–– The main tab (Design Molecule), where the user can design and
submit several chemical compounds using a friendly drawing
tool (Fig. 1).
–– The Import Smile tab, which enables the user to enter and sub-
mit chemical species in SMILES notation (Fig. 2).
–– The Import sdf tab, where the user can browse for the .sdf file
of the chemical compound(s) of interest (Fig. 3).
Fig. 1 Enalos Suite “Design Molecule” tab

Fig. 2 Enalos Suite “Import SMILES” tab
Fig. 3 Enalos Suite “Import sdf” tab
In the main tab, the sketcher [14] enables the user to easily draw
any chemical compound that is then submitted for further analysis.
Several possibilities are available in different panels like drawing dou-
ble and triple bonds, creating chains and making any stereo bonds.
Propane, butane, pentane, hexane, octane, and benzene rings are
also available for use, along with some more complex structure tem-
plates like alkaloids, amino acids, beta lactams, carbohydrates, and
steroids. The user can also select among different heteroatoms (N,
P, S, F, Cl, Br, I) that usually exist in organic chemical molecules,
enter an element or group symbol via keyboard and select new draw-
ing symbols from the periodic table. More functionalities are avail-
able including selecting, deleting, rotating, moving, and removing
whole or parts of the designed molecules. Finally, the sketcher
enables the user to open, save, and convert files with a variety of
chemical formats such as SMILES or IUPAC Chemical Identifier.
If the SMILES notation is known, the user can directly submit it
in the Import SMILES tab. However, if the SMILES notation is not
initially known, the chemical sketcher included gives the users the
opportunity to draw the chemical structure and then convert the
structure in SMILES format. This facilitates the generation of several
structures, by allowing for multiple modifications to be performed
using the sketcher and then all structures can be transferred as
SMILES to the appropriate tab allowing for the analysis for the whole
set of produced structures.
The .sdf files contain chemical structure records, used as a
standard exchange format for chemicals’ information. The
compound structure in .sdf format can be extracted from PubChem
database or other repositories and uploaded to Enalos Suite
through the corresponding field.
2.1 Enalos Suite A main area of interest in the cheminformatics field is the applica-
Database Functions tion of computational methods and the development of predictive
models, in an effort to reduce the time and resources spent on
experiments and to make faster decisions in the drug discovery
framework. This in silico approach demands the availability of
accessible data that could be mined, refined, and proceeded within
a model development or virtual screening procedure [5, 6].
Enalos Suite Database functions allow for access and data
retrieval from CIR, PubChem, and UniChem. More specifically,
these functions give access to NCI, PubChem, and UniChem
chemical databases for data mining and manipulation and also con-
tain more functions using PubChem Database related to Assay,
Patent Coverage Information, Similar Compounds, and Vendor.
2.1.1 CIR Function The CIR function enables the user to get direct access to CIR
(Chemical Identifier Resolver). CIR works as a resolver for differ-
ent chemical structure identifiers and allows the conversion of a
given structure identifier into another representation or structure
identifier. Several output formats can be selected through a GUI
menu. The available identifiers are: Standard InChI, Standard
InChiKey, InChiKey Simplified, SMILES, NCI/CADD FICTS
Identifier, NCI/CADD FICuS Identifier, NCI/CADD Identifier,
CACTVS HASHISY Hashcode, IUPAC Name, CAS Registry
Number, ChemSpider ID, Molecular Weight, Chemical Formula,
Number of Hydrogen Bond Donors, Number of Hydrogen Bond
Acceptors, Lipinski Rule of five Violations, Number of Effectively
Rotatable Bonds, list of chemical names for the structure, and SD
file of the structure [2, 15].
2.1.2 PubChem PubChem (https://pubchem.ncbi.nlm.nih.gov), hosted by the

Functions US National Institutes of Health (NIH) is a public repository for
chemical information including molecular structures and their bio-
activities. It is administered as part of the Molecular Libraries
Roadmap Initiative (MLI) [16] and includes, at the time of writ-
ing, information for more than 90 million compounds, thus is cur-
rently one of the largest publicly available molecular databases. The
available data come from various sources such as university labora-
tories, chemical vendors, assay providers, pharmaceutical compa-
nies, and journals, and include information about small molecules,
iRNAs, lipids, peptides, carbohydrates, and more [17, 18].
PubChem also offers online services supportive to drug inno-
vation, such as refinement and analysis of the stored substances and
assays that have many applications in chemical biology, medicinal
chemistry, and informatics research. Finally, it is a resource of sec-
ondary databases and web services, contributing in this way to
drug discovery process [19].
The Enalos Suite PubChem category contains five functions
that give direct access to PubChem database to extract useful
information:
The PubChem function enables the user to search the PubChem

database and obtain information for thousands of compounds with
one request. The function exports for each query compound the
PubChem CID, the IUPAC name, the InChI, the InChI-Key, its’
molecular formula and molecular weight, the canonical SMILES,
and the direct PubChem URL.
The Assay function gives the user access to PubChem database
via substance or compound IDs (SID and CID respectively), to
retrieve the Assays in which a particular compound was tested.
Using this function, the user can download within Enalos Suite,
information about the Assay and the Assay outcome, and later use
this information for modeling activities within the same software.
The Patent Coverage Information function enables the user to
extract information about the patent coverage for thousands of
compounds.
Via Similar Compounds function, the user can search the whole
PubChem database for similar compounds (using Tanimoto
Similarity) and obtain the PubChem CID, the molecular formula
and weight, and the number of rotatable bonds of the similar com-
pounds to the query compound.
The Vendor function enables the user to obtain information
about the commercial availability for thousands of chemical
compounds.
As it can be easily concluded the above functions greatly facili-
tate the researchers to retrieve useful information for thousands of
compounds in a few steps, eliminating significantly the time spent
in the monotonous and tedious search of information from differ-
ent resources, within the drug discovery process.
2.1.3 UniChem Function UniChem is a simple, large-scale nonredundant database of point-

ers between chemical structures and the European Molecular
Biology Laboratory (EMBL)-European Bioinformatics Institute
(EBI) chemistry resources. The advantage of UniChem is that it
implements an easily scalable and maintainable system of chemical
structure integration within the EBI-Search infrastructure, allow-
ing future new chemistry resources within the EBI to be included
with minimal effort.
The UniChem function gives the user direct access to the 27
available databases under UniChem that are grouped in five friendly
and easily recognizable categories. Exporting data from UniChem
databases, offers a great flexibility as permits their direct analysis
and handling and offers fast and automated modeling.
2.1.4 Searching This case study demonstrates the process of searching Chemical
Chemical Databases: Databases such as PubChem and UniChem using the Enalos Suite
A Case Study Databases Functions as described above. The specific paradigm
deals with the chemical compound abacavir, which is a medication
used to prevent and treat HIV/AIDS.
Fig. 4 Designing abacavir molecule
Fig. 5 Importing SMILES notation of chemical compounds
Through the main Enalos Suite sketcher the abacavir can be

designed as shown in Fig. 4.
Alternatively, the molecule of interest can be imported as
SMILES (Fig. 5) or as an .sdf file (Fig. 6).
Subsequently the user can select one or more of the available
Databases Functions, namely PubChem, Similar Compounds,
Vendor, Patent Coverage Information, Assay, UniChem, and CIR.
The user can browse and choose the results directory, where the
results will be saved. The results of the Enalos Suite functions are
displayed in different Excel files which enables the easy manipula-
tion of the results in further steps.
Fig. 6 Importing an .sdf file
Fig. 7 PubChem function results
From the PubChem function, information is extracted from

PubChem Database including the PubChem CID, the IUPAC
Name, the InChI, the InChiKey, the Molecular Formula, the
Molecular Weight, the Canonical SMILES, and the PubChem
URL (Fig. 7).
From the Similar Compounds function information is extracted
about the PubChem CID, the Molecular Formula, the Molecular
Weight, and the Number of Rotatable Bonds of the similar com-
pounds found. The search for similar compounds is based on the
Tanimoto Similarity, which is the percentage of desirable similarity
among the input compound and the compounds included in the
PubChem database (Fig. 8).
Fig. 8 Similar Compounds results
Fig. 9 Vendor function results
The Vendor function extracts information from PubChem

about the PubChem SID, the Compound URL, the Supplier, and
the Category of the Vendor (Fig. 9).
The Patent Coverage Information function exports informa-
tion from PubChem about the Patent’s IDs (Fig. 10).
Fig. 10 Patent Coverage Information results
Fig. 11 Assay function results
From the Assay function, information is extracted from

PubChem about the Assay name, the AID (BioAssay identification
number), the Bioassay Type, the Bioactivity Outcome, the Target
IG, the Target GeneID, the Activity Value in μM (whenever avail-
able), and the Activity Name (Fig. 11).
The UniChem function exports the InChiKeys found in all
UniChem Databases and the InChiKeys not found in these
Databases (Fig. 12).
The CIR function outputs a data-table with all available CIR
identifiers (Standard InChI, Standard InChI-Key, Smiles of the
structure, etc.) (Fig. 13).
All functions presented above were executed for one submit-
ted chemical compound. The user can also design several chemical
molecules and import their SMILES notation. The functions can
be then executed for all requested compounds. For example,
PubChem function can be used for the chemical compounds abaca-
vir, sofosbuvir, and lasofoxifene (Figs. 14 and 15) with the option
to extend the list to thousands of compounds.
Fig. 12 UniChem function results (InChiKeys found)
Fig. 13 CIR function results
Fig. 14 Transferring SMILES of several compounds
Fig. 15 PubChem function results for several compounds

2.2 Enalos Suite The molecular descriptors play an essential part in model develop-
Molecular Descriptors ment as they convey molecular structure key data [1]. More spe-
cifically, molecular descriptors encode with numerical values the
properties and the features of the molecules that can be later linked
through modeling techniques, to the biological activities or other
properties of the molecules [7, 20]. The calculation of molecular
descriptors and their use in modeling can have direct applications
in different research fields such as health and pharmaceutical sci-
ences, toxicology, and environmental chemistry [7].
There are different methodologies and software tools that
can be applied in order to determine molecular descriptors that
employ organic and quantum chemistry, graph theory, etc. [1,
21, 22]. The National Center for Toxicological Research of FDA
designed and released the Mold2, a freely available software, that
can be employed for the calculation of 777 important molecular
descriptors that encode two-dimensional chemical structure
information, including topological, geometric, and structural
characteristics of compounds [7, 12]. Comparison of Mold2
descriptors with descriptors calculated from commercial software
on several p ublished datasets showed that Mold2 descriptors pro-
duce models with higher quality than other packages. The calcu-
lation of molecular descriptors using the Mold2 software through
Enalos Suite was made available via the Calculate Through Mold2
function [2, 7].
2.2.1 Calculation This example is designed to help the interested reader in calculat-
of Molecular Descriptors: ing the molecular descriptors of one or more chemical compounds,
A Case Study using the Calculate Through Mold2 Enalos Suite function. Two
case studies will be considered. The first one deals with the calcula-
tion of the molecular descriptors for a single chemical compound,
abacavir (CID 441300 in PubChem), and the second one deals
with all the small molecules associated with a specific assay in
PubChem: Luminescence Cell-Based Counter screen to Identify
Inhibitors of A1 Apoptosis (AID 449761 in PubChem).
In the first case, the user has three options to insert the com-
pound in the Suite as presented in the previous case study. The
Calculate Through Mold2 function calculates a large and diverse set
of molecular descriptors (777), encoding two-dimensional chemi-
cal structure information, including topological, geometric, and
structural characteristics of compounds. The output is presented as
an .xls file (Fig. 16).
In the second case, the user first searches the PubChem data-
base for the assay with AID 449761 and downloads the .sdf file
including both active and inactive compounds that have been
tested in the specific assay (Fig. 17).
Consequently, the .sdf file can be inserted in the Import sdf
window. The calculated molecular descriptors for all the active and
inactive compounds are presented in Figs. 18 and 19 respectively.
Fig. 16 Calculate Through Mold2 results for a single compound
Fig. 17 PubChem AID 449761
Fig. 18 Molecular descriptors for the active compounds in PubChem Assay, AID 449761
Fig. 19 Molecular descriptors for inactive compounds in the PubChem Assay, AID 449761
2.3 Enalos Suite Recent efforts within the drug discovery and the risk assessment
Predictive Models framework, are focused in the substitution of the in vivo and in
vitro experiments with computational in silico approaches. The
computer-aided research in a drug discovery process addresses the
need of eliminating the expensive and labor-intensive experimental
procedures, and can also be the answer to the question of ethics
that arises by the use of laboratory animals. One of the most well-
known in silico approaches is the development of predictive mod-
els that connect the structural characteristics of the molecules to
their activity or toxicity profile (quantitative structure–activity/
toxicity relationships models, QSAR/QSTR) [9]. Using this kind
of models, the researchers can seek strong patterns among the
experimental data and use these patterns in order to make accurate
predictions in future data [23, 24]. Modeling procedures require a
series of important steps, such as variable selection, and internal
and external evaluation, in order to build a robust model that pro-
duces trustworthy predictions. When a model is developed it is
crucial that its results are disseminated to the scientific community,
including researchers with no computational background, in order
to use them in real-life applications.
KNIME (Konstanz Information Miner) platform is used as the
basic infrastructure in the development of Enalos Suite. KNIME is
a powerful tool for data analysis, integration and modeling, which
is open-source and freely available. This software offers a user-
friendly interface for creating visual data flows which consist of
nodes and connections between them. The ease of visualization of
data “pipelines” gives the user the flexibility to interact and selec-
tively execute some or all analysis steps and investigate the results.
Within the same workflow it is possible to combine tools from dif-
ferent suites in short time, including different cheminformatics
tools (CDK, RDKit, ChEMBL, etc.) and other modeling tools
(e.g., WEKA) [25].
The development of appropriate KNIME workflows can sig-
nificantly facilitate data mining, analysis, and modeling in the area
of cheminformatics and nanoinformatics. Within KNIME,
NovaMechanics has developed its’ proprietary nodes, Enalos+
nodes, that are designed to fill some cheminformatics related tasks
lacking from KNIME platform, that are associated with data pre-
processing, modeling and data mining of chemical databases
(PubChem, UniChem, NCI, etc.). Enalos Suite, as a part of the
Enalos software family, can easily integrate every KNIME w orkflow,
offering in this way a friendly user interface for any functionality or
model that is already or will be developed in KNIME.
One of the main advantages of Enalos Suite is that it can host
any predictive model developed within KNIME. Enalos Suite con-
tains several custom-made predictive and validated models includ-
ing the MouseTox and K562 models described below.
2.3.1 MouseTox Model Within Enalos Suite a fully validated QSTR model, MouseTox
for Cytotoxicity Prediction model, that can predict the cytotoxic effects of a wide range of
compounds, is incorporated [10]. The developed model was based
on calculated molecular descriptors from the Mold2 software and
the Random Forest machine learning methodology. MouseTox
model was first released as a web service via Enalos Cloud Platform
(http://enalos.insilicotox.com/MouseTox/) that allows for
online predictions for novel structures that are designed or
uploaded to the server.
The dataset selected for MouseTox model development,
included 5416 compounds that were tested for cytotoxic effects to
NIH/3T3 cells, as part of a project for the identification of new
drugs for the treatment of Chagas disease [10, 26]. The cytotoxic-
ity profile of compounds was considered as the output variable and
compounds included in the original dataset were classified as
“actives” (compounds cytotoxic to NIH/3T3s) or “inactives”
(compounds noncytotoxic to NIH/3T3s).
The MouseTox model was built based on a workflow devel-
oped in KNIME analytics platform that included the Enalos+ pro-
prietary KNIME nodes developed by NovaMechanics [13]. The
workflow included all steps required to afford a fully validated pre-
dictive model. Initially, for each compound in the original dataset,
777 molecular descriptors were calculated using the Mold2 soft-
ware. During the correlation analysis, some descriptors were fil-
tered out, leaving 424 to be used as inputs for the QSTR model
development. For validation purposes the initial dataset was divided
into training and test sets, using the Kennard and Stones algo-
rithm. The training set was used for model development, whereas
the test set was used during the external validation process. The
InfoGain variable selection along with Ranker evaluator method
was applied on the training set, in order to identify 15 descriptors
as the most critical for the model development. A detailed analysis
of the physical meaning of these descriptors can be found in the
original publication [10].
Special attention was given to the validation of the proposed
model, by employing different strategies for external and internal
validation to meet the criteria recommended by OECD. For inter-
nal validation, the model’s stability to the inclusion-exclusion of
data was tested by performing leave-k-out cross-validation tests. In
addition, external validation, was performed using the blank test
set. Tables 1 and 2 present the confusion matrices for the training
and the test set respectively, while Table 3 presents the relevant
statistics that prove the accuracy and the robustness of the devel-
oped model. Finally, the Y-randomization test, when applied on
the data, demonstrated the robustness and the statistical signifi-
cance of the proposed model [10].
Table 1
Confusion matrix (training set)
Inactive
Active (predicted) (predicted)
Active 2114 165
Inactive 181 1602
Table 2
Confusion matrix (test set)
Active (predicted) Inactive (predicted)

Active 719 111
Inactive 116 408
One of our main purposes of the MouseTox model is to be

used in a virtual screening framework for compounds not tested in
the original assay. Toward this goal it was important, to denote
whether a prediction could be considered as reliable or not; some-
thing that was achieved by providing, along with the proposing
validated model, a well-defined domain of applicability [27–29].
In the present work, domain of applicability was calculated using
similarity measurements based on the Euclidean distance among all
training and test compounds [13, 25].
The incorporation of MouseTox model within Enalos Suite
greatly facilitated virtual screening to assess the the cytotoxicity of
novel compounds. Based on the developed model and in order to
obtain cytotoxicity predictions for a given compound of interest,
the user should submit one or several structures either by using the
chemical sketcher or by submitting a SMILES notation or an .sdf
file. Within a virtual screening context, the user can upload differ-
ent datasets and examine the relationships among the different
inputs with the generated predictions.
As an indicative case study we selected among millions of
compounds in PubChem database three similar (Tanimoto simi-
larity ≥90%) and two dissimilar compounds (Tanimoto similarity
≤80%) to the compound C24H23FN2O2 (CID 54649948) included
in the initial data set. This compound is an “active” one (cytotoxic
to NIH/3T3s), and our purpose was to test whether its similar
compounds are also cytotoxic and therefore define exactly in a
future study the structural characteristics that affect the toxicity of
these compounds.
Table 3
Accuracy statistics of the predictive model
Accuracy Sensitivity Specificity

Training set Test set Training set Test set Training set Test set
0.915 0.832 0.928 0.866 0.898 0.779
Fig. 20 Structure of compound #1 (≥90% similar to C24H23FN2O2)
For these compounds, the user can sketch the structure of the
molecules in the Enalos Suite sketcher (Figs. 20, 21, 22, 23, and
24), or import their SMILES (Fig. 25) and execute the MouseTox
model. The results are presented in a .csv file and provide predictions
on the cytotoxic class of given compounds (“active”/“inactive”)
together with an indication of whether the predictions fall within
the applicability domain.
The predictions for the compounds that fall out the model’s
applicability domain limits cannot be considered reliable (Fig. 26).
2.3.2 K562 Apart from the MouseTox model, more predictive models are
Inhibition Model incorporated within Enalos Suite covering a wide range of biologi-
cal activities. Among these, a K562 inhibition predictive model is
included. Recent efforts in beta thalassemia treatment, suggest the
discovery of fetal hemoglobin inducers (HbF) that could compen-
sate the effects caused by this disorder. Toward this goal, an in
silico and fully validated model for the prediction of K562 func-
tional inhibition, possibly associated with HbF induction was pro-
posed [30] This K562 model was made available online through
Enalos Cloud Platform (http://enalos.insilicotox.com/K562)
and was also incorporated within Enalos Suite.
For the model development, an initial dataset of 129 diverse

small molecules tested in human erythroleukemia cells as potential
K562 inhibitors, was selected from the PubChem database [31].
The compounds were classified in the original dataset as “actives”
and “inactives” based on their biological effect. A KNIME work-
flow facilitated by Enalos nodes, was built to include all necessary
modeling steps. For each compound in the original dataset, 777
molecular descriptors using the Mold2 software, were calculated
Fig. 23 Structure of compound #4 (≤80% similar to C24H23FN2O2)
Fig. 24 Structure of compound #5 (≤80% similar to C24H23FN2O2)
and after a filtering step, some of them were excluded from further
analysis, due to their low discrimination power. A consensus model-
ing scheme was then developed: three different models—including
three different modeling methodologies (the kNN, random tree,
and random forest) and two different variable selection techniques
(the Gain Attribute evaluator and the InfoGain Attribute Ratio
Feature)—were used, and later combined based on a consensus
majority vote approach.
Fig. 25 Transferred SMILES for all five compounds
Fig. 26 MouseTox toxicity assessment
All models were validated according to OECD principles, and

the accuracy statistics are presented in Table 4. As seen from the
results, the consensus model outperformed all others, and was
finally proposed for the prediction of the K562 cell growth inhibi-
tion together with a well-defined applicability domain.
This validated model can be accessed through the Enalos Suite
and could be easily employed during drug discovery in a decision-
making framework. The user can upload one or several structures
(in one of the three available formats) and examine whether these
compounds are predicted as “actives” or “inactives” in K562 cell
growth inhibition within the domain of applicability of the model
as shown in the example below.
2.3.3 Virtual Screening All models included within Enalos Suite can be used in combina-
tion to virtually examine the biological effects of a given com-
pound. As an example, we can consider a specific paradigm for the
in silico screening of a structure included in PubChem database.
Assuming that the chemical compound of interest is C18H19ClN4O2
that can be found in PubChem (https://pubchem.ncbi.nlm.nih.
gov/compound/16663089) the following procedure can be
undertaken. In a first step this compound of interest is submitted
through the main Enalos Suite window, as shown in Fig. 27, and
the K562 workflow is selected from the available list, for the execu-
tion of the corresponding model.
Table 4
Model validation results (test set)
Specificity Sensitivity Precision Accuracy

Model I—random tree 0.6 0.733 0.733 0.68
Model II—kNN 0.5 0.933 0.737 0.76
Model III—random forest 0.7 0.733 0.786 0.72
Consensus model 0.7 0.933 0.824 0.84
Fig. 27 Designing C18H19ClN4O2 molecule
The results are exported in a .csv file providing information

about the inhibition activity of the compound and an indication of
whether this prediction can be considered reliable. As can be seen
in Fig. 28 the query compound is predicted as active and therefore
it can be considered a promising compound for K562 inhibition.
In a next step, the cytotoxicity of the submitted structure can
be examined based on the MouseTox model, to in silico assess
whether this compound can provoke undesirable side effects. For
that, the MouseTox model from the Enalos Suite main window is
Fig. 28 K562 prediction results for C18H19ClN4O2
Fig. 29 MouseTox toxicity assessment for C18H19ClN4O2
Fig. 30 Vendor results for C18H19ClN4O2
Fig. 31 SureChEMBLPatents results for C18H19ClN4O2
selected and submitted and results can be inspected as shown in

Fig. 29. As the proposed structure (C18H19ClN4O2) was predicted
as inactive, a noncytotoxic profile is important.
More information on the submitted structure can be retrieved
by selecting one or more of the available Databases Functions,
namely Vendor, PubChemPatents, IBMPatents, Novelty Checking,
SureChEMBLPatents, and Assay, that extract useful information
about its commercial availability (Fig. 30), its patent coverage
(Fig. 31), and Assays in which it is tested.
It is underlined that all the above steps can be performed
simultaneously and for more than one compound.
Fig. 32 Similar Compounds to C18H19ClN4O2 results for Tanimoto similarity 95%
Moreover, other functionalities are available to further explore

a specific structure of interest and its chemical space. The Similar
Compounds function can be used to extract information about the
PubChem CID, the Molecular Formula, the Molecular Weight,
and the Number of Rotatable Bonds of similar compounds to the
query compound. The search for similar compounds is based on
the Tanimoto Similarity, which is the percentage of desirable simi-
larity among the input compound and the rest compounds included
in the PubChem database. For this example, this was set equal to
95% (Fig. 32).
This example has highlighted the multiple functionalities of
Enalos Suite and demonstrated step by step the important finding
that could guide further steps in the drug discovery process.
3 Conclusions
The current needs in drug development process, toxicity assess-

ment, or cosmetics production demand the introduction of
computational tools in the field of research which help to eliminate
labor-intensive and expensive experiments and to reduce the use of
laboratory animals. While various predictive models over the past
years have been developed toward this goal, it is common that such
models remain within the developers’ group and do not reach the
wider community to facilitate further advancements. It is thus very
important to consider, in parallel with model development, the dis-
semination of its results in a way that it could immediately serve as
an important source of information for any interested party.
Moreover, it is also important to provide a user-friendly working
environment, especially for the researchers that do not have a
strong computational background but want to benefit from the
model’s results.
Enalos Suite addresses these needs, given that it provides
access to data, cheminformatics tools, and robust models within a
simple interface. The Calculate Through Mold2 function, incorpo-
rated in Enalos Suite, offers new features in transforming the
structural information of the molecules into mathematical repre-
sentations that can be used as inputs to the cheminformatics tools.
In addition, Enalos Suite includes functionalities that give easy
and immediate access to multiple chemical databases for data min-

ing and manipulation allowing for their direct and automated
analysis. On top of that, Enalos Suite includes custom made mod-
els specified for making accurate predictions for chemical com-
pounds for a variety of applications.
In conclusion, the new Enalos Suite, developed with open
source and freely available software, could be a very useful aid in
the field of cheminformatics as it provides access to fully validated
predictive models as well as new innovative tools that allow for data
handling and analysis upon a very simple and user-friendly
interface.
References
1. Todeschini R, Consonni V (2010) Molecular Ligand (RANKL). PLoS Comput Biol
descriptors for chemoinformatics. Wiley- VCH, 13:e1005372
Weinheim 12. Melagraki G, Afantitis A (2014) Enalos
2. Leonis G et al (2016) Handbook of computa- InSilicoNano platform: an online decision sup-
tional chemistry. Springer, New York, NY port tool for the design and virtual screening of
3. Willett P (2002) Chemistry plans a structural nanoparticles. RSC Adv 4:50713–50725
overhaul The rising tide of data being gener- 13. NovaMechanics Ltd (2017) NovaMechanics
ated by high-throughput. Nature 419:4–7 Ltd. http://www.novamechanics.com/index.
4. Leach Andrew R, Gillet VJ (2007) An intro- php/what-we-do/software/. Accessed 7 Apr
duction to chemoinformatics, revised Ed. 2017
Springer, New York, NY 14. Krause S, Willighagen E, Steinbeck C (2000)
5. Melagraki G, Afantitis A (2016) Editorial: JChemPaint - Using the collaborative forces of
towards open access for cheminformatics. the internet to develop a free editor for 2D
Comb Chem High Throughput Screen chemical structures. Molecules 5:93–98
19:260–261 15. Chemical Identifier Resolver. https://cactus.
6. Vrontaki E, Melagraki G, Mavromoustakos T, nci.nih.gov/chemical/structure. Accessed 22
Afantitis A (2014) Exploiting ChEMBL data- Dec 2017
base to identify indole analogs as HCV replica- 16. Kim S, Thiessen PA, Bolton EE et al (2016)
tion inhibitors. Methods 71:4–13 PubChem substance and compound databases.
7. Hong H, Xie Q, Ge W et al (2008) Mold2, Nucleic Acids Res 44:D1202–D1213
molecular descriptors from 2D structures for 17. Melagraki G, Afantitis A (2015) A risk assess-
chemoinformatics and toxicoinformatics. ment tool for the virtual screening of metal
J Chem Inf Model 48:1337–1344 oxide nanoparticles through enalos insili-
8. Tropsha A (2010) Best practices for QSAR conano platform. Curr Top Med Chem
model development, validation, and exploita- 15:1827–1836
tion. Mol Informatics 29:476–488 18. Chen B, Wild DJ (2010) PubChem BioAssays
9. Melagraki G, Afantitis A, Sarimveis H et al as a data source for predictive models. J Mol
(2006) A novel RBF neural network training Graph Model 28:420–426
methodology to predict toxicity to Vibrio 19. Cheng T, Pan Y, Hao M et al (2014) PubChem
fischeri. Mol Divers 10:213–221 applications in drug discovery: a bibliometric

10. Varsou D-D, Melagraki G, Sarimveis H, analysis. Drug Discov Today 19:1751–1756
Afantitis A (2017) MouseTox: an online toxic- 20. Ojha PK, Roy K (2018) Development of a
ity assessment tool for small molecules through robust and validated 2D-QSPR model for
Enalos Cloud platform. Food Chem Toxicol sweetness potency of diverse functional organic
110:83–93 molecules. Food Chem Toxicol 112:551

11. Melagraki G, Ntougkos E, Rinotas V et al 21. Melagraki G, Afantitis A, Sarimveis H et al
(2017) Cheminformatics-aided discovery of (2007) A novel QSPR model for predicting θ
small-molecule Protein-protein interaction (lower critical solution temperature) in poly-
(PPI) dual inhibitors of Tumor Necrosis Factor mer solutions using molecular descriptors.
(TNF) and Receptor Activator of NF-κB J Mol Model 13:55–64
22. Afantitis A, Melagraki G, Sarimveis H et al 27. Melagraki G, Afantitis A, Sarimveis H et al
(2008) Development and evaluation of a (2010) In silico exploration for identifying
QSPR model for the prediction of diamag- structure-activity relationship of MEK inhibi-
netic susceptibility. QSAR Comb Sci tion and oral bioavailability for isothiazole
27:432–436 derivatives. Chem Biol Drug Des 76:397–406
23. Melagraki G, Afantitis A (2011) Ligand and 28. Papa E, Sangion A, Arnot JA, Gramatica P
structure based virtual screening strategies for (2018) Development of human biotransforma-
hit-finding and optimization of hepatitis C tion QSARs and application for PBT assess-
virus (HCV) inhibitors. Curr Med Chem ment refinement. Food Chem Toxicol 112:535
18:2612–2619 29. Alves VM, Muratov EN, Zakharov A et al (2018)
24. Afantitis A, Melagraki G, Sarimveis H et al Chemical toxicity prediction for major classes of
(2006) A novel QSAR model for evaluating industrial chemicals: is it possible to develop uni-
and predicting the inhibition activity of dipep- versal models covering cosmetics, drugs, and
tidyl aspartyl fluoromethylketones. QSAR pesticides? Food Chem Toxicol 112:526
Comb Sci 25:928–935 30. Afantitis A, Leonis G, Gambari R, Melagraki G
25. Leonis G, Melagraki G, Afantitis A (2016) (2017) Consensus Predictive Model for the
Open source chemoinformatics software prediction of Human K562 Cell Growth
including KNIME analytics platform among a Inhibition through Enalos Cloud Platform.
multitude. Springer, New York, NY, ChemMedChem. https://doi.org/10.1002/
pp 2201–2230 cmdc.201700675
26. National Center for Biotechnology Information 31. National Center for Biotechnology Information
(2012) PubChem BioAssay Database, (2016) SANGER: inhibition of human K-562
AID=651744. https://pubchem.ncbi.nlm.nih. cell growth in a cell viability assay. https://
gov/bioassay/651744. Accessed 3 Jan 2017 pubchem.ncbi.nlm.nih.gov/bioassay/742260.
Accessed 19 Dec 2017
Chapter 15
Ion Channels in Drug Discovery and Safety Pharmacology

Paola Imbrici, Orazio Nicolotti, Francesco Leonetti, Diana Conte,
and Antonella Liantonio
Abstract
Ion channels are membrane proteins involved in almost all physiological processes, including neurotrans-
mission, muscle contraction, pace-making activity, secretion, electrolyte and water balance, immune
response, and cell proliferation. Due to their broad distribution in human body and physiological roles, ion
channels are attractive targets for drug discovery and safety pharmacology. Over the years ion channels
have been associated to many genetic diseases (“channelopathies”). For most of these diseases the therapy
is mainly empirical and symptomatic, often limited by lack of efficacy and tolerability for a number of
patients. The search for the development of new and more specific therapeutic approaches is therefore
strongly pursued. At the same time acquired channelopathies or dangerous side effects (such as proar-
rhythmic risk) can develop as a consequence of drugs unexpectedly targeting ion channels. Several noncar-
diovascular drugs are known to block cardiac ion channels, leading to potentially fatal delayed ventricular
repolarization. Thus, the search of reliable preclinical cardiac safety testing in early stage of drug discovery
is mandatory. To fulfill these needs, both ion channels drug discovery and toxicology strategies are evolv-
ing toward comprehensive research approaches integrating ad hoc designed in silico predictions and exper-
imental studies for a more reliable and quick translation of results to the clinic side.
Here we discuss two examples of how the combination of in silico methods and patch clamp experi-
ments can help addressing drug discovery and safety issues regarding ion channels.
Key words Ion channels, Pharmacovigilance, Patch clamp, Molecular docking, Bartter syndrome,
Cardiotoxicity
1 Introduction
Ion channels are membrane proteins that selectively regulate ion

transport across the membranes of cells and cellular organelles by
shifting from the open to closed state in dependence of membrane
potential, ligand binding, or physicochemical stimuli. The presence
of distinct ion channel isoforms and their age-dependent and tissue-
specific expression throughout the body allow the fine regulation of
many cellular functions, such as cell excitability, contraction, neu-
rotransmitter and hormone release, gene expression, ion and water
homeostasis [1]. About 5% of the available marketed drugs are
313
314 Paola Imbrici et al.
small molecules targeting cationic ion channels to fight a number of

diseases. Sodium channels blockers are on the market as anticonvul-
sants, antiarrhythmics, and local anesthetics [2]. Similarly, diabetes
benefits from sulfonylureas known KATP channels blockers, and
calcium channels blockers represent antiarrhythmics and antihyper-
tensive medications. Conversely, anionic channels such as chloride
channels represent a relatively little explored target class for drug
discovery [3–6]. Dysfunctions in ion channels genes, or their inter-
acting proteins, as a consequence of genetic mutations cause spe-
cific ion channelopathies, ranging from relatively common ones,
such as some cardiac arrhythmias, to very rare diseases, including
Bartter syndrome. For most channelopathies the therapy is based
on relief of symptoms, with limited efficacy and low tolerability pro-
file for a significant part of patients.
Channelopathies can also occur as a secondary defect caused
by xenobiotics exposure, including drugs. Indeed, given the piv-
otal roles played by ion channels and their extensive distribution, it
is not surprising that ion channels can also promote off-target drug
activities. In particular, it is well known that many drugs can inhibit
cardiac ionic currents, including INa, ICaL, and IK, thereby affecting
action potential (AP) properties in a way that can trigger or sustain
cardiac arrhythmias [7]. More recently similar toxicological con-
cerns have been raised also for drugs acting on CNS, kidney, and
other organs mediating predominately off-target activity on neuro-
nal ion-channels, receptors, transporters, and enzymes [8, 9].
Thus, on one hand, new drug discovery paradigms are needed to
develop selective ion channel ligands and, on the other hand, more
comprehensive and unfailing safety testing, prior to first-in-human
trials, are required to refine the clinical risk profile of new drugs.
Ion channels drug development and drug-induced toxicity
assessment are being greatly assisted by the recently identified
X-ray solved crystal structures of several ion channels, by computer
molecular modeling, by high-throughput screening methods, by
novel genome-editing technologies and preclinical models of dis-
eases [5, 10–15]. In some cases, human stem cells technology and
in silico reconstruction of human physiology are supporting and
sometimes overcoming the employment of transgenic and pharma-
cologically induced animal models which are often unsuccessful in
providing reliable and interpretable results due to species-related
differences [16, 17]. In parallel, novel pharmacological research
strategies include time and cost effective approaches favoring the
repurposing of drugs already on the market, with recognized safety
and bioavailability in humans, toward novel disease areas [18]. The
selective optimization of side activities (SOSA approach) of
approved drugs can also be exploited to reprofile drugs by convert-
ing side activities into novel main effects [19]. Thus, we are
assisting to a paradigm shift from common, simplified and not spe-
cific studies to strategies integrating different experimental meth-
Ion Channels in Drug Discovery and Safety Pharmacology 315
ods, for example combining in silico predictions with in vitro

human ion channel assays, enabling an improved validation and
fair interpretation of the results.
Here we first provide an example of how pharmacovigilance
database search, electrophysiological and computational methods
can be integrated to identify novel ligands for kidney ClC-K chlo-
ride channels by exploiting the kidney toxicity of commercial
drugs. We found that valsartan and olmesartan are novel ligands of
ClC-Ka channels and elucidated the molecular requirements for an
effective drug–channel interaction [20].
Then, we briefly summarize the recently proposed
Comprehensive in vitro proarrhythmia assay (CiPA) schema, a
nonclinical mechanism-based paradigm for the preclinical evalu-
ation of proarrhythmic liabilities of novel drug candidates in
early stages of drug discovery [21].
2 Discovery of High Affinity ClC-K Ligands Starting from Pharmacovigilance

Database Search
The first reported example regards the discovery of novel ClC-Ks

ligands among approved drugs by exploiting the pharmacological
side activity of approved drugs. ClC-Ka and ClC-Kb channels are
pivotal in the kidney where they control chloride reabsorption
and water diuresis [22]. Loss-of-function mutations of CLCKNB
and BSND genes, encoding the ClC-Kb channel isoform and
barttin β subunit respectively, cause Bartter syndrome (BS), a rare
condition characterized by salt and fluid loss, leading to kidney
failure and sometimes sensorineural deafness [23, 24]. ClC-K
channels remain a quite neglected target for drug discovery [25]:
so far, no drug on the market is efficacious on these channels and
BS patients receive symptomatic therapies with limited clinical
improvement [6, 26, 27]. Thus, the identification of ClC-K
channels-targeted drugs would be highly desirable. In this study
by Imbrici et al. [20], the discovery of novel ClC-Ks ligands
started from the analysis of the FDA pharmacovigilance database
monitoring drug safety in search of commercial drugs inducing a
Bartter-like syndrome as side effect with the assumption that BS
could result from the block of ClC-K channels. The ability of the
selected BS-causing drugs to bind and block ClC-K channels was
then validated through an integrated experimental and computa-
tional approach based on patch clamp electrophysiology in
HEK293 cells and molecular docking simulations.
2.1 FDA-AERS Database searching was performed using the Food and Drug
Database Searching Administration-Adverse Effects Reporting System (FDA-AERS)
database [28, 29]. AERS collects spontaneous reports of adverse
events following any US-approved drugs. Despite AERS cannot
generally demonstrate causal links between drugs and adverse

events, it can detect signals that can be better investigated with
more rigorous methods such as the experimental one [30]. From
the analysis of this pharmacovigilance registry several commercial
drugs inducing Bartter-like syndrome as side effect were identified
in treated patients. The identified compounds were further filtered
in order to select the best candidates for in vitro screening accord-
ing to: (1) high number of reports of Bartter-like adverse reactions
and (2) physical and/or chemical properties (such as lipophilicity
or bulk volume) compatible with in vitro screening. To limit the
risk of false positives, all compounds determining BS with a mecha-
nism explicitly unrelated with the block of ClC-K channels were
discarded [31–35]. Five molecules were finally selected: mycophe-
nolate mofetil, quinapril, valsartan, candesartan cilexetil, and
acenocoumarol (Table 1).
2.2 Evaluation In order to test whether the selected approved drugs cause Bartter-
of the Blocking like symptoms in patients by targeting ClC-K proteins, ClC-Ks
Efficacy of Marketed channels were expressed in HEK293 cells and chloride currents
Drugs on ClC-Ka recorded through whole-cell patch-clamp before and after the
Channels application of 50 μM of each drug using specific extracellular and
Through Manual Patch intracellular solutions and customized voltage-clamp protocols.
Clamp These studies demonstrated that mycophenolate mofetil, aceno-
coumarol, and quinapril showed a very low affinity for ClC-Ka
channels, whereas the same concentration of valsartan caused ~60%
reduction of ClC-Ka chloride currents [20] (Fig. 1a).
The inhibition of ClC-Ka by this drug was concentration-
dependent with an IC50 = 21 μM. On the other hand, ClC-Kb
channels were much less sensitive to valsartan block, being inhib-
ited by ~40% only at a concentration of 100 μM as well as ClC-1
and ClC-5 channels.
Valsartan belongs to a class of antihypertensive drugs acting as
AT1R antagonists. To detect the molecular determinants of valsar-
tan structure responsible for chloride current inhibition, other
molecules belonging to the same pharmacological class, namely
losartan, telmisartan, candesartan cilexetil, and olmesartan were
then tested in vitro to verify their ability to block ClC-Ka channels.
The considered sartans share a common chemical scaffold made by
a biphenyl bearing a tetrazole ring at ortho position with the
exception of telmisartan where the tetrazole ring is replaced by a
carboxylic group. The performed patch clamp experiments showed
that losartan, candesartan cilexetil, and Telmisartan at 50 μM were
very weak blockers of ClC-Ka/barttin channels reducing chloride
currents by ~10% to ~25%. Unlike them, valsartan and olmesartan,
containing both the tetrazole ring and the carboxylic group, are by
far stronger blockers of ClC-Ka/barttin channels (Fig. 1a, b).
These experiments clarified that differences in the chemical struc-
tures of sartans accounted for a different affinity toward the ClC-Ka
Table 1
Drugs returned from the analysis of FDA pharmacovigilance database causing Bartter-like syndrome
as side effect in the indicated number of case reports
Number
of BS
reports
BS-inducing Primary (2011– Proposed mechanism
drugs Therapeutic class pharmacodynamics 2012) of BS syndrome
Furosemide Loop diuretic NKCC2 blocker 8 PS Inhibits Na/K/2Cl
cotransporter
mimicking type 1
BS [31]
Mycophenolate Immunosuppressant Inosine 7 SS Unknown
Mofetil monophosphate—
Dehydrogenase
inhibitor
Prednisone Anti-inflammatory Glucocorticoid 4 PS, 2 SS Possibly inhibits
glucocorticoid receptor agonist Na/K-ATPase
similarly to the
digitoxigenin
derivative
rostafuroxin [32]
Quinapril Antihypertensive ACE inhibitor 4 SS Unknown
Candesartan Antihypertensive Blocks the AT1R 2 PS, 2 SS Unknown
Cilexetil
Valsartan Antihypertensive Blocks the AT1R 1 PS, 3 SS Unknown
Rituximab Immunosuppressant Monoclonal mAb 4 SS Unknown
anti-CD20 B cell
Tacrolimus Immunosuppressant Phosphatase calcineurin 3 SS, 1 C Stimulates the renal
inhibitor Na–Cl
cotransporter with
hypertension [33]
Methyl- Anti-inflammatory, Glucocorticoid 3 SS Unknown
prednisolone immunosuppressant receptor agonist
Calcitriol Hormone Vitamin D receptor 2C Increases Ca transport
agonist across the epithelial
cells and renal
phosphate excretion
Acenocoumarol Anticoagulant Vitamin K antagonist 2C Unknown
BS Bartter syndrome, PS primary suspected drug, SS secondary suspected drug, C concomitant drug, CaSR calcium
sensing receptor, HMG-CoA 3-hydroxy-3-methylglutaryl-coenzyme A, AT1R angiotensin type 1 receptor, ACE angio-
tensin converting enzyme
A C
ClC-Ka/barttin
F347
S*
L253
1.8
5nA Å
2.6 Å
100ms 2.1
N68
Valsartan 50mM Å
K165
Docking score: -10.77 kcal/mol
B D
ClC-Ka/barttin
F347
L253
5nA
100ms
N68
Olmesartan 50mM K165
Docking score : -9.05 kcal/mol
Fig. 1 Effect of valsartan and olmesartan on ClC-Ka channels expressed in HEK293 cells and docking studies
into a homology model of ClC-Ka. (a, b) Representative current traces of ClC-Ka/barttin channels before and
after the application of valsartan 50 μM (a) and Olmesartan 50 μM (b) in HEK293 cells. Data are mean ± SEM
of n = 8 cells. (c, d) Top-scored docking poses of valsartan (c) and olmesartan (d) in ClC-Ka. Important residues
and docked ligands are rendered as sticks while the protein is shown as a surface. The H-bond interactions
are depicted by a dotted line
channel and that both the carboxylic group and the tetrazole ring
are required for an efficacious in vitro block of ClC-Ka channels, as
in valsartan and olmesartan (Table 2).
2.3 Molecular The following step was to gain insights into the putative binding
Docking Simulations site of ClC-K channels, and obtain valuable information for the
future design of new drugs targeting ClC-Ks. To these aims, dock-
ing simulations of valsartan, losartan, and olmesartan were carried
out on ClC-Ka channels, assuming the same binding site recently
defined for the phenyl-benzofuran carboxylic acid derivatives [36–
38] and using a homology model of ClC-Ka built on the 3D struc-
ture of ClC-ec [39, 40]. In particular, the carboxylic unit of these
compounds is known to interact with three residues, N68, K165,
and H346 located at the extracellular side of the ClC-Ka protein.
Table 2
Structure–activity studies of sartans
Valsartan Losartan Telmisartan Candesartan cilexetil Olmesartan

(58.9 ± 5.5) (12.5 ± 3.9) (24.5 ± 4.0) (11.0 ± 5.9) (50.5 ± 3.2)
Chemical structures of tested sartans and percentage of inhibition of ClC-Ka/barttin currents induced by 50 μM of
drug. Blue and red squares indicate the tetrazole ring and the carboxylic group, respectively. Data are mean ± SEM of
n = 8 cells
Valsartan, losartan, and olmesartan were subjected to flexible recep-

tor docking studies using a multi stage induced fit docking protocol
(IFD) available from the Schrödinger Suite v2015-3 [20, 38, 41].
At physiological pH, Valsartan and Olmesartan present two nega-
tive charges, resulting from the deprotonation of both the tetrazole
ring and the carboxylic acid group. These two negatively ionized
moieties are able to establish salt bridge interactions with K165, the
same positively charged residue crucial for the binding of benzofu-
ran derivatives [38]. In addition, docking simulations allowed to
detect a well-oriented H-bond interaction between the carboxylic
group of the ligand and N68, while π–π stacking interactions were
observed between the proximal aromatic ring of the biphenyl frag-
ment and F347. Finally, in the case of valsartan, hydrophobic con-
tacts were established between the alkyl chain of the butanoic
moiety and L253. Taken together, these interactions explain the
remarkable docking score values obtained for both Valsartan and
Olmesartan (−10.77 kcal/mol and −9.05 kcal/mol respectively)
and, in agreement with the experimental data, indicate the high
affinity binding of these sartans toward ClC-Ka channel.
Interestingly, a very similar posing is observed in the top-scored
docking solution of Losartan. However, due to the lack of the car-
boxylic acid group, only one ionic interaction occurs, which involves
the tetrazole ring and K165. Thus, docking studies suggested that
the lack of the interactions established with N68, which were
instead experienced by Valsartan and Olmesartan, were likely
responsible for the drop of the docking score (−7.18 kcal/mol) and
for the lower inhibitory capacity of losartan toward ClC-Ka, consis-
tently with experimental data (Fig. 1c, d).
In conclusion, the combination of patch clamp experiments
with docking simulations convincingly showed that valsartan and
olmesartan block chloride currents through a specific interaction
between their tetrazole and carboxylic acidic moieties and a blocking
binding site in ClC-Ka channels. These information are crucial in

the perspective of using the identified drugs as lead compounds for
the development of novel and more potent ClC-Ks ligands, openers
and blockers, potentially useful in ClC-K associated diseases [42].
3 Preclinical Cardiac Safety Testing: The CiPA Initiative
One important step during the drug discovery process is the early
and efficient assessment of drug-induced electrophysiological and
structural cardiotoxicity in order to advance novel and safe drug
candidates into clinical trials whereas discarding unsafe com-
pounds. The most dangerous cardiac arrhythmia that can be gen-
erated as a consequence of side effect of noncardiovascular drugs is
torsades de pointes (TdPs), a life-threatening ventricular disorder
predominantly associated with the block of hERG channels
(KCNH2 or Kv11.1 or IKr) and characterized by a marked prolon-
gation of QT interval, considered during the last decade its princi-
pal surrogate biomarker [21]. For this reason, preclinical and
clinical cardiac safety approaches regulated by the International
Committee of Harmonization (ICH) S7B and E14 relied for a
long time on in vitro assays in cell lines aiming at discarding any
new chemical entity proven to block this repolarizing potassium
current and expected to induce a prolongation of QT interval in
animals ECG and in clinical thorough QT (TQT) studies. These
screening tests are now considered too simplistic (single parameter
tests) and not specific. They employ nonhuman species that do not
allow a correct prediction of human structural and contractile car-
diotoxicity. Moreover, IKr blockade and QT prolongation are now
considered incomplete biomarkers of proarrhythmic risk. Cardiac
repolarization is in fact affected by the interplay of multiple ionic
currents, and, following a revised understanding of the TdPs
mechanisms, repolarization instability and early after depolariza-
tions (EADs) are additional critical events for TdPs onset [16].
The drawbacks of these common tests have led to the wrong exclu-
sion of favorable lead compounds from later-stage development
and to the great expense of TQT assays; thus, more reliable pre-
clinical cardiac safety paradigms are being developed. One example
is the Comprehensive in vitro Proarrhythmia Assay (CiPA) schema,
a mechanism-based, nonclinical regulatory paradigm born from a
public–private collaboration, which represents the natural evolu-
tion of previous cardiac safety guidelines ICH S7B and E14. The
CiPA initiative integrates drug effects on multiple cardiac ionic
currents with in silico modeling of human ventricular electrical
activity, and in vitro data obtained from human stem cell-derived
ventricular cardiomyocytes to provide a more accurate assessment
of drug proarrhythmic propensity in a cost-effective and high-
throughput manner [7] (Fig. 2).
Study of drug candidate effects on

multiple ion channels expressed in
vitro in established cell lines
In silico reconstruction of human

CiPA initiative for improved heart ventricular action potential and
assessment of candidate integration of determined drug
drug proarrhythmic liability effects on ionic currents
Study of drug candidate effects on

hSC-CMs to confirm previous
evaluations or envisage missed
effects
Fig. 2 Schematic diagram of CiPA core assays. hSC-CM human stem cell-derived cardiomyocytes
The aim of CiPA is twofold: (1) to discard unsafe compounds

thus reducing late-stage risks to patients and limiting developmen-
tal time and costs and (2) to avoid false-positive and inappropriate
discontinuation of viable candidates.
3.1 In Vitro Studies According to CiPA, the effect of a new compound should be first
evaluated in vitro on seven recombinant ion channels/currents
(Cav1.2 or ICaL, Nav1.5 or INaF and L, Kv11.1 or IKr, Kv7.1 or IKs,
Kv4.2 or IKto, Kir2.1-2.4 or IK1) expressed in heterologous systems.
Rigorous and standardized voltage-clamp protocols and experi-
mental conditions have to be set to ensure robust, reliable and
reproducible dataset from automated planar patch-clamp. To this
aim, the Ion Channel Working Group (ICWG) pinpoints that both
the potency (IC50 determination) and the voltage-, kinetics-, use-
dependence of drug block are required to better predict the proar-
rhythmic potential of new chemical entities [21, 43].
3.2 In Silico Studies The drug effects recorded on multiple human cardiac currents
should be next integrated in silico in a reconstructed undiseased
mathematical model of human ventricular myocyte electrophysiol-
ogy that should help translating drug effects on individual currents
into their propensity for delayed repolarization and early after-
depolarizations. The In Silico Working Group (ISWG) is in charge
of the development and validation of the best computational
reconstruction of the human ventricular myocyte action potential.
The O’Hara-Rudy (OHR) model represents a reliable albeit per-
fectible approach [44]. Its strengths rely on the facts that it is open
Table 3
The 12 CiPA training drugs
Drug Pharmacological class CiPA TdPs category

Quinidine Antiarrhythmic High
Bepridil Angina High
Dofetilide Antiarrhythmic High
Sotalol Antiarrhythmic High
Chlorpromazine Antipsychotic Intermediate
Cisapride Gastrokinetic Intermediate
Terfenadine Antihistamine Intermediate
Ondansetron Antiemetic Intermediate
Diltiazem Hypertension/angina Low/no risk
Mexiletine Antiarrhythmic Low/no risk
Ranolazine Angina Low/no risk
Verapamil Hypertension/angina Low/no risk
source (it is available on the Rudy Laboratory research website

http://rudylab.wustl.edu), it performed better than other pub-
lished models in determining TdPs risk in simulation tests and,
principally, it is based on extensive human experimental data [21,
44]. Proarrhythmic biomarkers relevant for simulation studies
include changes in membrane resistance, change in refractoriness,
dispersion of depolarization across the tissue, beat-to-beat variabil-
ity [16, 45]. Possible additional parameters to more accurately pre-
dict the proarrhythmic risk of any drug candidate include individual
patient characteristics such as sex, inherited and acquired chan-
nelopathies, and pharmacological treatments [45]. To date, 28 ref-
erence drugs have been categorized by experts in high, moderate,
and low proarrhythmic risk groups, 12 of which have been selected
to calibrate and enhance the ventricular myocyte model, including
dofetilide, cisapride, and verapamil (Table 3).
Comparing the results of these simulations with known clinical
responses will provide information on the impact of a candidate
channels in the assay and how the models can be further improved.
Pilot simulation studies using the OHR model and clinical data for
dofetilide (high risk) suggested that, at least for hERG, IC50 alone
is not sufficient to assess the potential for a drug to induce TdPs.
Indeed, simulation tests confirmed that adding information about
state- and time-dependence of drug-induced channel block affects
ventricular repolarization delay, instability, and EAD generation
prediction and enhances the model performance in assessing drug
proarrhythmic liability [7, 43]. Additional studies using this model

demonstrated that the block of three ionic currents (IKr, ICaL, and
INa) by a drug can provide a better prediction of torsadogenic risk
than the block of IKr alone [16, 46].
3.3 Studies on Stem The third step of the CiPA paradigm consists in confirming the
Cell-Derived Myocytes effects of drugs on ionic currents in vitro on human stem
cell-
derived cardiomyocytes (hSC-CM) recording extracellular
field potentials through multielectrode array platform and changes
in transmembrane potential through voltage-sensitive dyes.
However, several aspects when using hSC-CMs still require clarifi-
cation and optimization, including the lack of standardized guide-
lines for the evaluation of hSC-CM phenotypes and functionality
that may influence drug effects [13, 17]. Finally, clinical assess-
ment of ECG in Phase I trials is forecasted to evaluate unantici-
pated electrophysiological effects [21].
4 Conclusions and Perspectives
In conclusion, we reported two successful examples of how ion

channel pharmacology and toxicology are evolving toward more
comprehensive and mechanistic-based approaches that are time
and cost-effective, more reliable and could be used as new para-
digms in ion channel research.
In the first presented case, a multidisciplinary approach com-
bining the analysis of side effects with electrophysiology and dock-
ing experiments was successful for repositioning valsartan and
olmesartan as new ClC-K ligands. In terms of medical advice,
these results may suggest that the block of ClC-Ka channels by
valsartan and olmesartan occurs as an ancillary pharmacological
mechanism of these drugs, possibly concurring to lowering blood
pressure, besides of the antagonism toward AT1 receptor [47]. On
the other hand, the evidence that other drugs belonging to the
same pharmacological class showed very poor inhibitory capacity
toward ClC-K channels may give insights into the toxicological
profile of these sartans, providing helpful suggestions to the clini-
cal management of BS. BS affected patients are currently treated
with nonsteroidal anti-inflammatory drugs, potassium supple-
ments, potassium-sparing diuretics, and angiotensin inhibitors
[26, 27]. These results may suggest the use of losartan or telmis-
artan in the therapeutic scheme of these patients with respect to
valsartan that could instead worsen the clinical phenotype.
The CiPA initiative demonstrates the need to combine a
broader in vitro channel screening with in silico reconstruction
of ventricular activity for an unfailing preclinical proarrhythmia
assessment at early stage of preclinical discovery. This integrated
approach to predict proarrhythmic potential of new compounds

is under consideration as part of future guidelines. Despite the
CiPA paradigm requires implementation, some steps of CiPA
are already being considered earlier in the developmental pro-
cess by pharmaceutical industries, such as increasing the num-
ber of ionic currents being drug-screened and using
human-derived cardiomyocytes [17]. If successful, CiPA will
allow the removal of compounds with cardiac undesirable
effects, limiting the risk and the costs of unmasking a critical
QT signal in clinical trials. Moreover, CiPA will help establish-
ing best practice for ion channels and stem cells-derived cardio-
myocytes research through standardized protocols to limit
intralaboratory and interlaboratory variability, and will make
available a single in silico model to predict the risk of arrhyth-
mia of new chemical entities [21].
References
1. Imbrici P, Liantonio A, Camerino GM et al of in vitro electrophysiological methods in
(2016) Therapeutic Approaches to Genetic Ion CNS safety pharmacology. J Pharmacol Toxicol
Channelopathies and Perspectives in Drug Methods 81:47–59
Discovery. Front Pharmacol 7:121 10. Tikhonov DB, Zhorov BS (2012) Architecture
2. Catterall WA, Swanson TM (2015) Structural and pore block of eukaryotic voltage-gated
basis for pharmacology of voltage-gated sodium channels in view of NavAb bacterial
sodium and calcium channels. Mol Pharmacol sodium channel structure. Mol Pharmacol
88:141–150 82:97–104
3. Imbrici P, Altamura C, Camerino GM et al 11. Feng L, Campbell EB, Hsiung Y, MacKinnon
(2016) Multidisciplinary study of a new ClC-1 R (2010) Structure of a eukaryotic CLC trans-
mutation causing myotonia congenita: a para- porter defines an intermediate state in the
digm to understand and treat ion channelopa- transport cycle. Science 330:635–641
thies. FASEB J 30:3285–3295 12. Zou B (2015) Ion channel profiling to advance
4. Imbrici P, Conte D, Liantonio A (2017) Paving drug discovery and development. Drug Discov
the way for Bartter syndrome type 3 drug dis- Today Technol 18:18–23
covery: a hope from basic research. J Physiol 13. Skalova S, Svadlakova T, Shaikh Qureshi WM,
595:5403–5404 Dev K, Mokry J (2015) Induced pluripotent
5. Jentsch TJ (2015) Discovery of CLC transport stem cells and their use in cardiac and neural
proteins: cloning, structure, function and regenerative medicine. Int J Mol Sci
pathophysiology. J Physiol 593:4091–4109 16:4043–4067
6. Verkman AS, Galietta LJV (2009) Chloride 14. Jan LY, Jan YN (2012) Voltage-gated potas-
channels as drug targets. Nat Rev Drug Discov sium channels and the diversity of electrical sig-
8:153–171 nalling. J Physiol 590:2591–2599
7. Huang H, Pugsley MK, Fermini B, Curtis MJ, 15. Zamponi GW, Striessnig J, Koschak A, Dolphin
Koerner J, Accardi M, Authier S (2017) AC (2015) The physiology, pathology, and
Cardiac voltage-gated ion channels in safety pharmacology of voltage-gated calcium chan-
pharmacology: review of the landscape leading nels and their future therapeutic potential.
to the CiPA initiative. J Pharmacol Toxicol Pharmacol Rev 67:821–870
Methods 87:11–23 16. Gintant G, Sager PT, Stockbridge N (2016)
8. Lynch JJ, Van Vleet TR, Mittelstadt SW, Evolution of strategies to improve preclinical
Blomme EAG (2017) Potential functional and cardiac safety testing. Nat Rev Drug Discov
pathological side effects related to off-target 15:457–471
pharmacological activity. J Pharmacol Toxicol 17. Gintant G, Fermini B, Stockbridge N, Strauss
Methods 87:108–126 D (2017) The evolving roles of human iPSC-
9. Accardi MV, Pugsley MK, Forster R, Troncy E, derived cardiomyocytes in drug safety and dis-
Huang H, Authier S (2016) The emerging role covery. Cell Stem Cell 21:14–17
18. Lavecchia A, Cerchia C (2016) In silico meth- 31. Reinalter SC, Jeck N, Peters M, Seyberth HW
ods to address polypharmacology: current sta- (2004) Pharmacotyping of hypokalaemic salt-
tus, applications and future perspectives. Drug losing tubular disorders. Acta Physiol Scand
Discov Today 21:288–298 181:513–521
19. Langer T, Wermuth C-G (2012) Selective opti- 32. Lupoli S, Salvi E, Barcella M, Barlassina C
mization of side activities (SOSA): a promising (2015) Pharmacogenomics considerations in
way for drug discovery. In: Peters J-U (ed) the control of hypertension. Pharmacogenomics
Polypharmacology in drug discovery, 1st edn. 16:1951–1964
John Wiley & Sons, Inc., New York, NY 33.
Lazelle RA, McCully BH, Terker AS,
20. Imbrici P, Tricarico D, Mangiatordi GF, Himmerkus N, Blankenstein KI, Mutig K,
Nicolotti O, Lograno MD, Conte D, Liantonio Bleich M, Bachmann S, Yang C-L, Ellison
A (2017) Pharmacovigilance database search DH (2016) Renal deletion of 12 kDa FK506-
discloses ClC-K channels as a novel target of binding protein attenuates tacrolimus-
the AT1 receptor blockers valsartan and olmes- induced hypertension. J Am Soc Nephrol
artan. Br J Pharmacol 174:1972–1983 27:1456–1464
21. Fermini B, Hancox JC, Abi-Gerges N et al 34. Chen Y-S, Fang H-C, Chou K-J, Lee P-T, Hsu
(2016) A new perspective in the field of cardiac C-Y, Huang W-C, Chung H-M, Chen C-L
safety testing through the comprehensive (2009) Gentamicin-induced Bartter-like syn-
in vitro proarrhythmia assay paradigm. drome. Am J Kidney Dis 54:1158–1161
J Biomol Screen 21:1–11 35. Zietse R, Zoutendijk R, Hoorn EJ (2009)
22. Fahlke C, Fischer M (2010) Physiology and Fluid, electrolyte and acid-base disorders asso-
pathophysiology of ClC-K/barttin channels. ciated with antibiotic therapy. Nat Rev Nephrol
Front Physiol 1:155 5:193–202
23. Birkenhäger R, Otto E, Schürmann MJ et al 36. Liantonio A, Picollo A, Babini E, Carbonara G,
(2001) Mutation of BSND causes Bartter Fracchiolla G, Loiodice F, Tortorella V, Pusch
syndrome with sensorineural deafness and kid- M, Camerino DC (2006) Activation and
ney failure. Nat Genet 29:310–314 inhibition of kidney CLC-K chloride channels
24. Simon DB, Bindra RS, Mansfield TA et al by fenamates. Mol Pharmacol 69:165–173
(1997) Mutations in the chloride channel gene, 37. Liantonio A, Picollo A, Carbonara G et al
CLCNKB, cause Bartter’s syndrome type (2008) Molecular switch for CLC-K Cl- chan-
III. Nat Genet 17:171–178 nel block/activation: optimal pharmacophoric
25. Imbrici P, Liantonio A, Gradogna A, Pusch M, requirements towards high-affinity ligands.
Camerino DC (2014) Targeting kidney CLC-K Proc Natl Acad Sci U S A 105:1369–1373
channels: pharmacological profile in a human 38. Liantonio A, Imbrici P, Camerino GM et al
cell line versus Xenopus oocytes. Biochim (2016) Kidney CLC-K chloride channels
Biophys Acta 1838:2484–2491 inhibitors: structure-based studies and efficacy
26. Loudon KW, Fry AC (2014) The renal chan- in hypertension and associated CLC-K poly-
nelopathies. Ann Clin Biochem 51:441–458 morphisms. J Hypertens 34:981–992
27. Zieg J, Gonsorcikova L, Landau D (2016) 39. Dutzler R, Campbell EB, Cadene M, Chait BT,
Current views on the diagnosis and manage- MacKinnon R (2002) X-ray structure of a ClC
ment of hypokalaemia in children. Acta Paediatr chloride channel at 3.0 A reveals the molecular
105:762–772 basis of anion selectivity. Nature 415:287–294
28. Mele A, Calzolaro S, Cannone G, Cetrone M, 40. Gradogna A, Imbrici P, Zifarelli G, Liantonio
Conte D, Tricarico D (2014) Database search A, Camerino DC, Pusch M (2014) I–J loop
of spontaneous reports and pharmacological involvement in the pharmacological profile of
investigations on the sulfonylureas and glinides- CLC-K channels expressed in Xenopus oocytes.
induced atrophy in skeletal muscle. Pharmacol Biochim Biophys Acta 1838:2745–2756
Res Perspect 2(1):e00028 41. Small-molecule drug discovery suite 2015-3:
29. Pitts PJ, Louet HL, Moride Y, Conti RM Schrödinger suite 2015-3 induced fit docking
(2016) 21st century pharmacovigilance: protocol; glide version 6.8, Schrödinger, LLC,
efforts, roles, and responsibilities. Lancet New York, NY, 2015; Prime version 4.1,
Oncol 17:e486–e492 Schrödinger, LLC, New York, NY, 2015
30. Evans SJ, Waller PC, Davis S (2001) Use of 42. Cheng C-J, Rodan AR, Huang C-L (2017)
proportional reporting ratios (PRRs) for signal Emerging targets of diuretic therapy. Clin
generation from spontaneous adverse drug Pharmacol Ther 102:420–435
reaction reports. Pharmacoepidemiol Drug Saf 43. Windley MJ, Abi-Gerges N, Fermini B, Hancox
10:483–486 JC, Vandenberg JI, Hill AP (2017) Measuring
kinetics and potency of hERG block for CiPA. J

46. Li Z, Dutta S, Sheng J, Tran PN, Wu W,
Pharmacol Toxicol Methods 87:99–107 Chang K, Mdluli T, Strauss DG, Colatsky T
44. O’Hara T, Virág L, Varró A, Rudy Y (2011) (2017) Improving the in silico assessment of
Simulation of the undiseased human cardiac proarrhythmia risk by combining hERG
ventricular action potential: model formulation (Human Ether-à-go-go-related gene)
and experimental validation. PLoS Comput Channel-drug binding kinetics and multi-
Biol 7:e1002061 channel pharmacology. Circ Arrhythm
45. Cavero I, Holzgrefe H (2015) CiPA: ongoing Electrophysiol 10:e004628
testing, future qualification procedures, and 47. McCallum L, Lip S, Padmanabhan S (2015)
pending issues. J Pharmacol Toxicol Methods The hidden hand of chloride in hypertension.
76:27–37 Pflugers Arch 467:595–603
Chapter 16
Computational Approaches in Multitarget Drug Discovery

Luciana Scotti, Hamilton Mitsugu Ishiki, Marcelo Cavalcante Duarte,
Tiago Branquinho Oliveira, and Marcus T. Scotti
Abstract
Current therapeutic strategies entail identifying and characterizing a single protein receptor whose inhibi-
tion is likely to result in the successful treatment of a disease of interest, and testing experimentally large
libraries of small molecule compounds “in vitro” and “in vivo” to identify promising inhibitors in model
systems and determine if the findings are extensible to humans. This highly complex process is largely
based on tests, errors, risk, time, and intensive costs. The virtual computational study of compounds simu-
lates situations predicting possible drug linkages with multiple protein target atomic structures, taking into
account the dynamic protein inhibitor, and can help identify inhibitors efficiently, particularly for complex
drug-resistant diseases. Some discussions will relate to the potential benefits of this approach, using HIV-1
and Plasmodium falciparum infections as examples. Some authors have proposed a virtual drug discovery
that not only identifies efficient inhibitors but also helps to minimize side effects and toxicity, thus increas-
ing the likelihood of successful therapies. This chapter discusses concepts and research of bioactive multi-
targets related to toxicology.
Key words In silico, Drug discovery, Multitarget, Toxicology, Drug
1 Introduction
1.1 Promise Current therapeutic strategies have been applied to several dis-
of a New Paradigm eases, including human immunodeficiency virus type 1 (HIV-1)
in Drug Discovery infection from an initial single target treatment to several targets
[1]. Antiretroviral drugs are no longer the only regimen recom-
mended for clinical use against HIV-1 due to rapid emergence of
drug resistance after initiation of therapy [2]. A combination of
antiretroviral drugs targeting different viral proteins is more effec-
tive in suppressing viral growth [11]. In many cases, however,
these regimens are expensive and result in increased toxicity and
poor patient compliance [3]. New paradigms in the discovery of
multitarget drugs have emerged, particularly for the treatment of
HIV-1 infection. For example, the multitarget antiretroviral drug
Cosalane was developed to inhibit several HIV-1 Proteins (gp120,
327
328 Luciana Scotti et al.
integrase, protease, and reverse transcriptase) simultaneously [4].

Computational screening of small molecule compounds against
protein targets implicated in a disease of interest has been widely
used to discover inhibitors; it potentially involves identifying the
successes in the interaction of the systematic chemical group with
a compound already known to inhibit a target, as in quantitative
structure–activity relationships (QSARs), or by “fitting” a large
molecule into the database of compounds in the active site of the
three-dimensional (3D) structure of a target protein based on the
calculated binding affinity of the molecule to the target. As the
number of high-resolution protein structures and the processing
power of computers have increased exponentially over the recent
years, so the methods used in the computational database to com-
plement experiments with high-throughput screening (HTS)
methods to improve efficiency and effectiveness of the discovery of
lead inhibitors. In addition, studies have shown that HTS success
rates are increased severalfold when compounds are prefiltered by
computational screening [5].
1.2 Summary We have developed a new computational paradigm for the discov-
and Advantages of Our ery of potential lead inhibitors based on the combination of three
Computing Paradigm principles (Fig. 1) [6].
(1) Incorporation of protein side chains and main chain dynam-
ics during the interaction process to further accurately assess bind-
ing affinities. (2) Selection of single inhibitors that bind to multiple
protein targets simultaneously, (3) Use of a screening library con-
sisting of drugs and drug compounds. Each principle increases the
likelihood of a prediction. The compound will successfully inhibit
the disease; in addition, screening with drug-like compounds spe-
cifically enhances pharmacological viability. Overall, this new para-
digm produces hits that will conveniently and predictably improve
the potential of the compounds analyzed to be viable drugs for
various diseases.
1.3 Justifying Biologically active proteins are in continuous motion, yet most
the Use of Dynamics information on the protein structure is limited to the most stable
form of a protein when crystallized under artificial conditions. The
different conformations of crystallographic structures and unre-
lated crystallographic structures suggest that binding events and
protein motions induce variation and more dimensions and changes
in the electrostatics of the catalytic site. This is likely to, under
physiological conditions, inhibit one of these variant conforma-
tions with affinity greater than that observed for the artificial stabi-
lization structure. Thus, dynamic simulations in addition to using
static crystal structures increase the possibility of a physiologically
relevant conformation.
We perform the dynamic nesting through (1) fitting the ligand
of interest, (2) solvation in a shell of water and salt, (3) applying
100 steps of energy minimization, (4) simulating protein
Computational Approaches in Multitarget Drug Discovery 329
movement through structure cycles random perturbations, and (5)

selecting the most relaxed of these models with a knowledge-based
punctuation function. We have shown that this dynamic plumbing
method is successful, in comparison with static plumbing methods,
for the targets of two important pathogens, HIV-1 [7] and P. fal-
ciparum [6].
1.4 Rationale
for Multitasking In the present study, it was found that the functional activity of a
compound, together with the structural conservation of active sites
and/or binding pockets, allows the activity of this compound in
multiple proteins, as indicated by large regulatory systems, such as
ATP facilitating energy transmission [8]. This same principle would
include a common susceptibility of proteins involved in essential
functions to life, creating a niche for formulating drug advertise-
ments. From a computational perspective, a compound that is pre-
dicted to inhibit multiple targets in a disease has an additive
likelihood of having pharmacological activity against that disease.
More importantly, inhibitor resistance is largely overcome by the
diminished exponential probability of resistant mutations that arise
simultaneously in genes encoding proteins corresponding to all
targets [9].
Living organisms have evolved into comparable chemicals in
environments containing similar sets of organic molecules. This
shared evolutionary chemical setting prepares the scenario for
Fig. 1 Comparison of our protocol for the discovery of multitarget inhibitors with traditional approaches cur-
rently used. The advantages of using our broad-spectrum novel. The protocol for the discovery of multitrigid
(right) inhibitors against key pathogens and diseases is contrasted with traditional approaches (left). The main
differences in our protocol, corresponding to the reasons why it is most effective, are as follows: (1) The use of
a molecular dynamics coupling algorithm that takes the flexibility of proteins and inhibitors completely into
account (http://compbio.washington.edu/papers/therapeutics.html). This algorithm is effective because all
molecules in biology are submitted to thermal motility. Traditional rigid anchoring approaches are not respon-
sible for this movement, resulting in poor prediction of binding energies or inhibitory constants as compared
to our approach. (2) The use of compounds which bind to multiple targets simultaneously. The most effective
drugs in humans (e.g., aspirin or Gleevec) inevitably interact and bind to multiple proteins, a feature that tra-
ditional models based on single target drugs do not take into account. The multitarget approach is a necessary
one because each drug has to be effective at its site of action (For example, HIV-1 protease inhibitors must
bind and inhibit the protease molecule.) and must be promptly metabolized by the body (e,g,, cytochrome P450
(CYP450) enzymes, which are responsible for the metabolism of most drugs) [6]. Computational Screening for
Multitarget binding and inhibition are effective because they exploit the evolving fact that the protein structure
is conserved much more in nature than function or sequence. (3) Use of compounds with drug and drug
approved by the FDA and experimental in the process of computational sorting. Screening of drugs developed
for other conditions against infectious diseases is likely to lead to fewer side effects because the pharmacoki-
netic toxicity, absorption, distribution, metabolism, and excretion is typically well established in human models
and animal models. To our knowledge, this is the first time these three elements have been combined to create
an effective inhibitor and drug discovery protocol with predictions that have been experimentally verified to
produce highly promising lead inhibitors for further drug development. The computational aspects of our pro-
tocol are fully automated, can be executed completely in parallel and require only a fixed initial investment in
the number of processors purchased (i.e., the higher number of CPUs, more targets and compounds that can
be traced). Our new protocol is extremely effective and increases success rates downstream preclinical and
clinical use with a considerable reduction of time, effort, and cost spent
various organisms to use the same compounds to control different

processes, making a molecule relevant for diversified physiological
activity. This principle is supported by evolutionary observation,
and these structures are more conserved than sequence or function.
Similar structures with comparable active and binding sites but
different chemical substances are used to perform a variety of
functions [10]. This observation that structural folds are widely
retained, even when the sequence and function are not, provides
logical evidence that a compound can be an excellent initial candi-
date for many different target proteins.
A drug candidate that can modulate multiple targets simulta-
neously could alter the disease network from an imbalanced disease
state to a normal healthy state [12, 13]. Additionally, multitarget
drugs could circumvent drug resistance considering that simulta-
neous mutations of several targets among distinct pathways or in
other positions of one cascade pathway are unusual events [14,
15]. The computational methods that rationally design drugs
capable of binding multiple targets are arising, and the results have
been promising.
Computational drug designing and development encom-
passes two different approaches, the direct and the indirect drug
designing [16].
The indirect methods in drug design are used whenever the
structure of the pharmaceutical target is unknown. The develop-
ment of new compounds with the desired characteristics for the
target should be determined by studying series of known ligands
for the same target. The most commonly used methods are quan-
titative structure–activity relationship (QSAR), molecular similar-
ity/diversity technique, and combinatorial chemistry. On the other
hand, the direct method is applied when the 3D structural infor-
mation for the biological target is known. This methodology com-
bines information from several fields, such as X-ray crystallography
and/or NMR, molecular modeling, synthetic organic chemistry,
QSAR, and biological assays [17].
In this chapter, we provide an overview of some examples of
multitarget drug design through computational methods includ-
ing both indirect and direct methods.
2 Computational Studies
2.1 Antimalarials Shibi et al. [18] employed a virtual screening of 292 phytochemi-
cals present in 72 traditional herbs, principally found in Africa,
China, and Asia, as effective antimalarials, in order to finding out
inhibitors of plasmepsin-2 and falcipain-2 of P. falciparum.
Bioassay datasets AID 504850, with potential to inhibit the
malarial parasite Plasmodium, and AID 2302, measured on levels
of P. falciparum lactate dehydrogenase as surrogate of parasite
growth, were downloaded from PubChem Bioassay [19]. The

descriptors were generated using the software PowerMV [20]. For
classification and regression analysis, artificial neutral network,
k-nearest neighbor, radial basis function neutral network, and ran-
dom forest approaches were employed. The employment of
Toxtree v1.60 software to screen for carcinogenic and/or muta-
genic compounds resulted in 99 compounds without these charac-
teristics. These 99 compounds were subjected to Lipinski’s rule of
five in order to assess the drug likeness of the compounds and only
46 compounds were approved.
In order to study the interactions between the selected com-
pounds and the falcipain-2, the X-ray crystallography structures of
1YVB, 2GHU, 2OUL, 3BPF, and 3PNR were extracted from
PDB. To study the interactions with the plasmepsin-2, the X-ray
crystallography structure of the proteins 1LF2, 1LYB, 1ME6,
1PFZ, and 2BJU were extracted from PDB. Before the analyses,
the energy of the protein molecules was minimized using Molecular
Operating Environment (MOE 2007.09) tool and the docking of
the selected 46 phytochemicals was performed utilizing the default
parameters.
The docking analyses performed with ten proteins were able to
select eight compounds with potential antimalarial activity and were
subjected to ADME prediction using preADMET tool. The human
intestinal absorption analysis has demonstrated the best absorption
of all compounds into the human intestine. They showed moderate
cellular permeability against Caco-2 cells, and the blood–brain bar-
rier penetration values indicated whether the compounds are able
to pass across the blood–brain barrier or not. All the eight com-
pounds were found to be inhibitors of the protein CYP3A4, and
four of them showed highest values of the plasma protein binding
of the drug that predicts its permanence in the system. Two com-
pounds, namely (2E)-3-[3,4-dihydroxy-5-(3-methylbut-2-en-1-yl)
phenyl]-1-(2,4-dihydroxyphenyl) prop-2- en-1-one and (2S)-2-
[3,4-dihydroxy-5-(3-methylbut-2-en-1-yl)phenyl]-5,7-dihydroxy-
3,4-dihydro-2H-1-benzopyran-4-one, were extracted from the
plant Erythrina abyssinica, and the other two compounds, namely
4-[(2S,3R)-3-[(4-hydroxyphenyl)methyl]-2-methylbutyl]-2-me-
thoxyphenol and 4-[(2R,3S)-3-[(4-hydroxy-3-methoxyphenyl)
methyl]-2-methylbutyl]-2- methoxyphenol, were extracted from
the plant Pycnanthus angolensis. The screened lead compounds can
be used for further studies.
The medicinal spectrum of imidazo-azines, Fig. 2, against
malaria, tuberculosis, and Chagas was studied by Kumar et al. [21]
from the perspective of computational multitarget screening con-
sidering (1) the wide range of therapeutic activities exhibited by
this class of heterocycles, (2) the easy accessibility of N-fused bicy-
clic imidazo-azines, and (3) that these heterocycles have not been
explored for the chosen disease targets.
Fig. 2 Imidazo-azines chemical structure
To perform the docking analysis 30 representative imidazo-

azines were selected and the pdb structures of wild type (WT)
forms of Pf-DHFR-TS (1J3I.pdb) [22], Pf-Enoyl ACP Reductase
(1NHG.pdb) [23], Pf-PK7 (2PMN.pdb) [24], Mt-SK (2DFN.
pdb) [25], Mt-PS (1N2H.pdb) [26], Mt-TMPK (1G3U.pdb)
[27], Mt-MurE ligase (2WTZ.pdb) [28], Tc-TR (1BZL.pdb)
[29], Tc-G3PD (1QXS.pdb) [30] and Tc-TS (1S0I.pdb) [31]
were extracted from PDB.
First, in order to improve the reliability of the model, the clus-
tering pattern and flexibility of the redocked molecules were ana-
lyzed. Redocked ligands with too many flexible bonds generally
generate too many clusters and/or abnormal binding energies. In
this context, four proteins were filtered out after first round of
screening. Considering the different targets, the ranking just based
on binding energy can be biased, and to avoid this, an additional
baseline based on ligand efficiency was also taken into account [32].
The docking results of six proteins showed that Mt-SK has
very poor ligand efficiency and was also eliminated from the analy-
ses. The remaining receptors produced hits with significant ligand
efficiency, and the imidazo-azine scaffold can be accepted as a good
binder for these proteins. The compounds 2-(4-chlorophenyl)-
N-cyclohexyl-6-methylH-imidazo[1,2-a]pyridine-3-amine
(MCL011) and N-cyclohexyl-2-(4-methoxyphenyl)-6-methylH-
imidazo[1,2-a]pyridine-3-amine (MCL017) (Fig. 3) exhibited
maximum binding energy with accepted ligand efficiency against
all the targets.
2.2 Alzheimer Azam et al. [33] considering the therapeutic properties of green
tea to reduce the risk of various neurodegenerative diseases, such
as Parkinson’s disease, studied the binding interactions between
catechins and 18 potential protein receptors through molecular
docking simulations. Green tea contains several polyphenolic
cathechins, such as catechin, epicatechin, epicatechin gallate, epi-
gallocatechin, and epigallocatechin gallate which are believed to be
the active components [34]. It was observed that green tea poly-
phenols were protective in SH-SY5Y cells against apoptosis induced
by the pro-Parkinsonian neurotoxin 6-hydroxydopamine [35].
Eighteen crystal structures of potential protein receptors,
extracted from PDB, and five catechin derivatives and gallic acid
Fig. 3 MCL011 and MCL017 chemical structures
were employed for the docking calculations. For every target pro-
tein, ten poses were visualized for each docked compound in order
to identify the model with minimum binding energy and best
ligand–receptor interaction. From docking analysis, it was observed
that the studied catechins were capable of interacting with all
docked targets, indicating that these ligands have a broad spectrum
of structural features. Drugs that interact with multiple targets
might have a better chance of affecting the complex equilibrium of
whole cellular networks than drugs that act on a single target.
Monoamino oxidase-B (MAO-B), adenosine A2A receptor, and nitric
oxide synthase (NOS) are the most promising anti-Parkinsonian
drug targets, and NMDA (N-methyl-d-aspartate) receptor is the
least favorable anti-Parkinsonian drug target.
A structure–activity relationship analysis identifies the impor-
tance of various functionalities for ligand–receptor interactions.
(a) Except for NMDA and α-amino-3-hydroxy-5-methyl-4-

isoxazole propionic acid receptors, the benzopyran moiety and
aromatic ring CB are essential for activity.
(b) Dihydroxyl group at positions 5 and 7 is required for activity
at almost all targets.
(c) OH substitution at R2 increases the activity at C-Jun-N-
terminal kinase, metabotropic glutamate receptor 1, cyclooxy-
genase-
1, cyclooxygenase-2, glutamate dehydrogenase, and
ionotropic glutamate receptor.
(d) Heterocyclic benzopyran ring contains chiral center at posi-
tions 2 and 3, which is important for the activity at C-Jun-N-
terminal kinase, P38 MAP kinase, catechol-O-methyl
transferase, metabotropic glutamate receptor 3, glycogen syn-
thase kinase 3, and α-amino-3-hydroxy-5-methyl-4-isoxazole
propionic acid.
(e) R1 substitution with OH in epicatechin, epigallocatechin, and
catechin is important for the activity at P38 MAP kinase, gly-
cogen synthase kinase 3, and glutamate dehydrogenase.
(f) R1 substitution with gallate moiety in epigallocatechin gallate

and epicatechin gallate makes these compounds very potent at
all targets except P38 MAP kinase, catechol-O-methyl trans-
ferase, NR1, glycogen synthase kinase 3, and α-amino-3-
hydroxy-5-methyl-4-isoxazole propionic acid receptors.
Additionally, a structure analysis shows the importance of electro-
static interaction and hydrogen bonding in the binding process.
A series of eight cyanopyridine–triazine hybrids was designed,
synthesized, and screened by Maqbool et al. [36] as multitargeted
anti-Alzheimer’s agents. Triazine was selected considering that its
roughly planar structure was expected to intercalate between beta-
amyloid sheets and was expected to enhance the beta-amyloid dis-
aggregation. Addittionally, a weak noncovalent interaction like
H-bonding, p–p stacking interaction, and alkyl-pi-interaction was
expected with the active site residues of acetylcholinesterase
(AChE). The inhibitory activity of the synthesized compounds was
evaluated toward cholinesterases and their A® anti-aggregating
activity. A docking study was performed on the crystal structure of
AChE (PDB ID 1EVE) and on the crystal structure of human
butyryl-cholinesterase (BuChE) (PDB ID 4TPK).
The docking analysis showed that compounds 4d
(2-(4-(4-(3-chloro-4-fluorophenylamino)-6-(3-(trifluoromethyl)
phenylamino)-1,3,5-triazin-2-yl) piperazin-1-yl)nicotinonitrile)
and 4h (2-(4-(4-(4-methoxyphenylamino)-6-(3-(tri-fluoromethyl)
phenylamino)-1,3,5-triazin-2-yl)piperazin-1-yl)nicotinonitrile)
were expected to show better interaction with AChE. An interac-
tion between the 3-chloro-4-fluoroaniline group and triazine moi-
ety with the aromatic ring of Trp279 and Tyr334 via π–π stacking
and the 3-(tri-fluoromethyl)aniline group with the amino acids
Phe330 and Phe331 was expected. Phe330 and Phe331 was
expected to engage with p-anisidine ring of compound 4h through
aryl–aryl interaction, and Trp279 was expected to interact with
3-(trifluoromethyl)aniline, while Tyr334 was expected to interact
with triazine rings. The compound 4d showed the highest inhibi-
tory potency against AChE as expected from docking results. The
compound 4d showed the highest inhibitory potency against
AChE as disclosed by docking results and compound 4h showed a
good inhibitory potency. All synthesized compounds exhibited
greater selectivity toward AChE over BuChE.
Palanimuthu et al. [37] have developed a series of compounds
that contain the benzylpiperidine pharmacophore from Donepezil
and the metal-chelating thiosemicarbazone moiety (Fig. 4). The
compounds were designed to be employed against Alzheimer’s
disease (AD) considering the following:
(a) They may chelate redox-active metals.
The accumulation of redox-active Fe and Cu causes oxida-
tive damage to neurons via the generation of deleterious reac-
Fig. 4 4-(1-benzylpiperidin-4-yl)thiosemicarbazone (BPT) derivatives chemical

structure
tive oxygen species (ROS), and the oxidative damage is

consequently implicated in Aβ peptide aggregation [38].
Chelation therapy has been proposed to mediate an improve-
ment in cognition via inhibition of metal-induced Aβ aggrega-
tion [39, 40].
(b) They inhibit the acetylcholinesterase (AChE) activity.
The AChE inhibitors enhance the cholinergic transmission
by inhibiting the degradation of ACh, a neurotransmitter
released by cholinergic neurons that are responsible for cogni-
tion and that degenerate in AD patients [41].
(c) They can induce autophagy.
Autophagy is a degradation mechanism in the cell that plays
a central role in neuroprotection by facilitating the removal of
aggregated proteins [42].
The drug design strategy considered the potent metal chela-
tion activity of thiosemicarbazones and the antioxidant potential of
hydroxyquinoline. Palanimuthu et al. [37] chose the Donepezil
structure, considering that it is an AChE inhibitor clinically used to
treat AD that consists of the benzylpiperidine and indanone moi-
eties. Molecular docking studies of the most active AChE inhibi-
tor, 2,3-OH-BBPT (IC50 = 1.02 μM), and the lead compound,
PBPT (IC50 = 4.93 μM), in comparison to Donepezil in the active
site of the human AChE enzyme (PDB ID: 4EY7), were performed
in order to understand the mode of interaction.
The highest docking score (−19.63 kcal/mol) with good
binding affinity (−88.30 kcal/mol) was obtained for Donepezil,
suggesting its strong interaction with AChE. The BPT analog,
2,3-OH-BBPT, also showed a strong interaction with AChE with
docking score = −17.12 and binding free energy (−88.29 kcal/
mol). The lead compound, PBPT, also interacts with AChE, but
showed the smaller docking score (−15.12) and the biggest bind-
ing energy (−83.59 kcal/mol). The docking diagrams suggest that
the BPT analogs bind to AChE utilizing a binding mode similar to
that of Donepezil.
2.3 Cancer Multitarget inhibition of cancer-associated kinases is an established

strategy to improve the efficacy and clinical outcome of targeted
therapies [43]. For example, several compounds showing dual
Fig. 5 Dinaciclib chemical structure
phosphoinositide 3-kinase/mammalian target of rapamycin inhibi-

tory activity are under evaluation in clinical trials [15]. Martin
et al. [44] demonstrated that dinaciclib (Fig. 5), a well-known
cyclin-dependent kinase (CDK2) inhibitor, interacts with bromo-
domain epigenetic reader proteins (BRDs). The dinaciclib acts on
a broad spectrum of human cancers both in vitro and in vivo;
moreover, the compound has a better in vivo therapeutic index
compared to other CDK inhibitors [45]. The studies with dinaci-
clib showed that a new generation of BRD inhibitors could be
rationally designed by using the chemical space of kinase inhibi-
tors. As a result, the rational identification of dual kinase/BRD
inhibitors has emerged as a promising interfamily polypharmacol-
ogy approach against cancer and inflammatory diseases.
The large amount of structural information available on kinases
and BRDs may allow ligand-based and structural-based computa-
tional approaches to be useful to facilitate the rational discovery of
more potent dual kinase/BRD inhibitors [46]. Using the Phase
software [47, 48] Martin et al. [44] built three new pharmaco-
phore models. The first model, called Type N Binders, comprises
compounds in which the kinase hinge binding moiety interacts
with the conserved Asn residue located in the BRDs ZA loop. The
second model, called Type PZA/ZA Binders, comprises com-
pounds that have the kinase hinge binding moiety interacting with
Pro82 of the WPF shelf and/or with conserved water molecules
located at the entrance of the ZA channel. And finally, the third
model, called Type I Binders, comprises compounds that do not
interact with the epigenetic target with their kinase hinge binding
group. The pharmacophore model generated from the most active
compound of this class shows three key features, i.e., a hydrogen

bond acceptor, a hydrogen bond acceptor/donor, and an aromatic
feature [46]. The kinase-based and the BRD pharmacophores can
be used in combination to allow for the identification of potential
dual kinase/BRD inhibitors from a wider chemical space. Docking
experiments performed with the crystal structure of BRD4 indicate
that the tested compounds overlap relatively well with the cocrys-
tallized reference ligands.
A successful virtual screening investigation to identify novel
dual kinase/bromodomain inhibitors from millions of commer-
cially available small molecules was reported by Allen et al. [49].
The ligand-based Laplacian-modified Naïve Baysian classifiers [50]
were employed for the compounds of Kinase Knowledgebase. The
models were cross-validated generating enrichment factors and
representative receiver operating characteristic (ROC) curves using
50% of the data as test set, and the derived models were applied to
evaluate over six million commercially available compounds from
the eMolecules database. From the library of more than six million
compounds with kinase activity predictions, only 122,136 com-
pounds from the active class for both EGFR models were then
further filtered to optimize the probability of kinase activity and
the favorable physicochemical properties. Observing the most
likely EGFR actives, it was possible to select 908 compounds.
Considering the limited information about the binding data for
BRD4 nine distinct docking models using cocrystal structures of
BRD4, extracted from Protein Data Bank, were built. The gener-
ated BRD4 data fusion model was validated using 246 known actives
extracted from ChEMBL [51] and 15,240 corresponding decoy
compounds obtained from the Directory of Useful Decoys [52].
After the computational trials 24 compounds comprising five
chemical scaffolds including substituted 2-pyrrolidinones,
quinazolin- 4-amines, thienopyrimidin-4-amines, and imidaz-
olidinones were selected and tested. Several novel BRD4 binders
were identified; however, one novel dual EGFR-BRD4 inhibitor,
namely 2870 (Fig. 6), a first-in-class compound was highlighted
and the studies showed that it can be considered a novel multi-
target inhibitor.
A new study involving several N-[3-(2-oxo-pyrrolidinyl)
phenyl]-benzenesulfonamide derivatives that scored highly in
ensemble docking protocol was performed by Allen et al. [49]. In
this study an AlphaScreen counterassay and differential scanning
fluorimetry were employed to confirm the activity of six BRD4
inhibitors identified employing virtual screening analysis. The
compounds had IC50 values in the range of 11–44 μM and was
determined via a BRD4(BD1)-acetylated histone interaction assay.
Initially, cocrystals of these compounds were made and their
structures resolved. These analyses displayed some interactions
that have been predicted by docking model previously reported
Fig. 6 Compound 2870 chemical structure
[53]. On the other hand, some best docking poses showed a root-
mean- square deviation (RMSD) ranging from 1.57 to 2.35 Å
when compared with the experimentally determined structures. In
order to analyze this discrepancy a 50 ns all-atom explicit water
molecular dynamics simulation was performed for all active deriva-
tives starting from the docking results. The results showed that
MD simulations can reproduce the experimentally observed
cocrystal protein–ligand complexes much more accurately than
docking alone.
The optimized models were used to redock 908 compounds
from which the lead compounds had been originally identified. It
was observed that all N-[3-(2-oxo-pyrrolidinyl)phenyl]-benzene-
sulfonamide derivatives were scored better when compared to the
original models. These compounds contain a pyrrolidinone moiety
which interacts with a hydrogen bond to the conserved N140 in
BRD4-1. On the other hand, unlike many of the most potent BET
inhibitors such as (+)-JQ1 and I-BET151, these compounds lack
many of the π–π stacking and hydrophobic interactions necessary
for further stabilization inside of the acetyl-lysine binding site.
These data are important to design more effective compounds.
2.4 Bacterial Considering the importance of antimicrobial peptides (AMPs),

with some studies suggesting that they can be used to inhibit and/
or kill many pathogenic microorganisms, including those which
exhibit high degrees of resistance against multiple antimicrobial
drugs [54], Speck-Planche et al. [55] employed a multitarget che-
mobioinformatic model to predict the antibacterial activity of pep-
tides against multiple gram-positive bacterial strains. AMPs can
target bacterial cell membranes and cause disintegration of the
lipid bilayer structure [56]. Additionally, they kill bacteria by

inhibiting some important pathways inside the cell such as DNA
replication and protein synthesis [57].
In order to perform this analysis, the Database of Antimicrobial
Activity and Structure of Peptides was selected considering the
degree of curation of the experimental information [58]. An addi-
tional curation was performed employing the StrainInfo website
[59]. Thousand five-hundred and eighty-one different peptide
sequences, ranging from 1 to 119 amino acids, tested against at
least one out of 50 different gram-positive bacterial strains, were
selected to perform the analysis. Considering that several AMPs
were assayed against more than one strain the data set was com-
posed from 2488 cases. The structural information present in the
AMPs was characterized through connectivity-like indices, derived
from the Kier–Hall formalism using the software PROTDCAL. The
mt-chemo-bioinformatic model, presented in Eq. 1, is composed
by six descriptors.

AB ( b ) = -27.065 DMKH1 ( Pb ) - 23.647 DMKH1 ( Pt ) + 0.737 DMKH3 ( Z1)
+ 1.432 DMKH3 ( Z2 ) + 0.141 DMKH 4 ( ISA ) + 0.144 DMKH5 ( M w ) - 0.317 (1)
N = 1991 l = 0.457 p　 value < 10-16 c 2 = 1555.30
Considering that all the descriptors are based on the Kier–Hall

formalism it means that the physicochemical properties were asso-
ciated with the molecular accessibility, In the obtained model, the
most relevant parameter is DMKH1(Pb) that represents the ten-
dency of a peptide to be in a β-sheet conformation is detrimental
to the increment of the antibacterial activity. The second relevant
parameter is DMKH1(Pt) and represents the diminution of the
tendencies of the AMPs to be in β-turn conformations will increase
their inhibitory activities against bacteria. These parameters show
that the conformations in AMPs such as α-helices increasing the
biological activity. The presence of hydrophobic regions in the
AMPs is relevant to the biological activity considering the presence
of parameter DMKH3(Z1) that expresses the increment in the
hydrophobicity of fragments/regions formed by a maximum of
three amino acids. The steric factor either contributes to the bio-
logical activity, considering the presence of DMKH3(Z2),
DMKH4(ISA) and DMKH5(Mw) in equation Eq. 1. DMKH3(Z2)
expresses the increment of the bulkiness of the AMPs in frag-
ments/regions formed with a maximum of three amino acids.
DMKH4(ISA) represents the increase in the isotropic surface area
of the AMPs in fragments/regions containing a maximum of
four amino acids and finally, DMKH5(Mw) that represents the
increment of the molecular weight in fragments/regions containing
a maximum of five amino acids. The training set of the mt-chemo-
bioinformatic model exhibited a sensitivity of 94.36% for active
cases and a sensitivity of 94.39% for inactive cases. The prediction

set showed a sensitivity of 92.89% for active cases and a sensitivity
of 96.51% for inactive cases.
Additionally, in order to determine the contribution of each
amino acid to the antibacterial activity of C-terminal amidated
peptide GLLSVLGSVAKHVLPHVVPVIAEHL the concept of
genetics known as deletion was employed. This study was per-
formed considering that some amino acids have an intrinsic influ-
ence in the activity of the AMPs and this influence depends on the
position of the amino acid in the sequence. This AMP was selected
considering that it was correctly classified by the generated model
as active against all studied bacteria. As results it was observed that
some amino acids such as leucine (L), lysine (K), histidine (H),
isoleucine (I), and valine (V) really contribute to the antibacterial
activity. The glycine (G), serine (S), alanine (A), proline (P), and
glutamic acid (E) decrease the antibacterial activity, independent of
their positions.
The automated multitarget screening with library searching is
already a powerful tool for toxicity analysis of samples of intoxica-
tion cases for clinical and forensic purposes. Thus, the use of pre-
dictive and in silico studies toxicology is called for [60].
One-target strategies prove useful to approach the smallest
toxicity and the utmost safety, where the “one drug, one target”
paradigm, highly potent and specific (single-target) treatments
would be better tolerated due to absence of off-target side effects.
Nevertheless, a disease is often a multifactorial condition involv-
ing a combination of constitutive and/or environmental factors,
a fact that emphasizes the importance of the multitarget drug
studies [61].
Identifying multiple protein targets is critical for side-effect
prediction. One third of potential therapeutic compounds fail in
clinical trials or are later removed from the market due to unac-
ceptable side effects often caused by off-target binding. The
Durrant et al. [62] work introduce a multidimensional strategy for
the identification of secondary targets of known small-molecule
inhibitors in the absence of global structural and sequence homol-
ogy with the primary target protein. To demonstrate the utility of
the strategy, they identify several targets of 4,5-dihydroxy-3-(1-
naphthyldiazenyl)-2,7-naphthalenedisulfonic acid (1), an inhibitor
of Trypanosoma brucei RNA editing ligase 1 (TbREL1). Twelve
human proteins was predicted as secondary targets of 1. Then the
potential side effects of 1 can be predicted by considering the phys-
iological role of these targets. For instance, AutoDock predicted
that 1 partially occupies a cofactor (NADPH) binding site, sug-
gesting that the compound may function as a competitive inhibitor
for NADPH. A limitation of the in silico methods is the prediction
of false positives. Two of the predicted secondary targets, SpPce

and HsPDE9A2, were not experimentally confirmed. Even then,
in silico studies may play a useful role in future efforts to reduce
drug side effects.
The strategy of using compounds with multitarget action also
can be used in order to reduce the pronounced toxic effect of a
drug or recreational drug; for example, it is an established fact that
the repeated use of opiate drugs, both for the relief of chronic or
cancer-related pain and for recreational drug-taking behavior, can
lead to the development of dependence. Addiction to opioids is a
complex syndrome involving tolerance, drug seeking, and physical
dependence with withdrawal avoidance behaviors. It is often char-
acterized as a chronic relapsing condition and is a major health and
social issue in most societies [63, 64]. Detoxification, a necessary
step for many forms of long-term abstinence-based treatments,
makes use of two approaches: tapering using methadone or
buprenorphine, or abrupt termination of opioid use, potentially
precipitated by an opioid antagonist (i.e., naltrexone) with admin-
istration of α2-adrenoreceptor (α2-AR) agonists to reduce with-
drawal symptoms. α2ARs impact neuronal function by classically
coupling to heterotrimeric G proteins of the Gi/o subfamily upon
activation by their endogenous agonists epinephrine and norepi-
nephrine, and have been demonstrated to be extremely sensitive to
opioid exposure. Exploring multitarget interactions, the Del Bello
et al. [64] suggest that 2 or its (S)-enantiomer and 3, endowed
with effective α2C-AR agonism/α2A-AR antagonism/5-HT1A-R
agonism, might represent novel multifunctional tools potentially
useful in the reduction of withdrawal syndrome and associated
depression at very low dose. The compounds 4 and 5–7 producing
efficacious α2C-AR agonism/α2A-AR antagonism/I2−IBS inter-
action might similarly be beneficial to both disorders and also
relieve further withdrawal comorbid neurobiologic conditions.
Such agents, lacking in sedative side effects due to their α2A-AR
antagonism, might afford an improvement over current therapies
with clonidine-like drugs [64].
Another approach of multitarget compounds would be the
search for active compounds in all or at least more than one enzyme
related to a particular pathology. A variety of targets are involved in
the sepsis disease network. Among them, 16 targets were involved
in nearly every aspect of sepsis including inflammation (A5LOX,
PDE4B, PLA24, LTA4H, and IL1RA), coagulation (TXA2R,
PAI-1, Prothrombin, and IXa), and immune dysfunction (A2AR,
eNOS, iNOS, MD-2, MMP9, MMP8, and, MIF) and were evalu-
ated by virtual docking screening against 343 compounds which
were identified in five chinese herbs namely Flos Carthami, Radix
Salviae, Paeonia anomala L., Rhizoma Ligustici chuanxiong, and
Radix Angelicae sinensis. This traditional Chinese medicine (TCM)

is widely used in China for the treatment of sepsis and has shown
satisfactory antiendotoxin and anti-inflammatory effects in clinical
trials (50–100 mL, intravenous drip twice a day for 7 days) using
the chemistry database of the Chinese Academy of Sciences, Dr.
Duke’s Phytochemical and Ethnobotanical Database, and the lit-
erature. To validate the virtual computation results, we investi-
gated the in vitro thrombin inhibition of several compounds. The
researchers chose the compounds that were predicted to have
activity against the target and evaluated their antithrombin effect
with bovine thrombin. The results showed that rosmarinic acid
(8), gallic acid (9), and vanillic acid (10) weakly inhibited throm-
bin activity [65].
The multitarget approach is really effective when the main
objective is to provide an increase in toxicity to the etiological
agent, since a multitarget inhibitor is supposed to be more efficient
and potent for treating and reducing risk of mutations that could
potentially confer resistance.
2.5 Treatment Eighteen anti-influenza herbs which have been widely used his-
of Influenza Virus torically to treat patients with influenza were selected by Gu
Infection et al. [66], from the Traditional Chinese Medicine Database 312
by Computational structures available by compounds of these herbs were collected.
Approaches For simplicity, this study focused on influenza viral proteins
structurally available for docking, omitting related targets of
human proteins. The structure of H7 in complex with its sub-
strate (PDB ID 4DJ7) and a complex structure of N9 with a
sulfate ion at the binding site (PDB ID 2B8H) were chosen. To
select TCM compounds that could potentially inhibit H7 and
N9 it was chosen molecules with a docking score at least 1 kcal/
mol lower than the corresponding positive controls (LSTc and
sialic acid, for H7 and N9 were −6.4 kcal/mol and −7.2 kcal/
mol, respectively). Among the screened compounds seven (11–
17) were predicted as dual inhibitors for H7 and N9. Compounds
11 and 12 from Scutellaria baicalensis, 13 from Radix isatidis,
14 from Glycyrrhiza uralensis, 15 from Artemisia annua, 16
from Cinnamomum aromaticum and, 17 from Folium isatidis.
This research demonstrates TCM’s multitarget/multicompo-
nent strategy for disease treatment. In this way, TCM offers an
efficient combinatorial intervention through mechanistic redun-
dancy and less potential drug resistance. For example, the herbal
drug Radix isatidis is frequently used by the Chinese to prevent
H7N9 and other types of influenza, but its effectiveness is
unclear and often argued. This study data here suggests that
Radix isatidis may prevent influenza, but this should be vali-
dated by additional experimental studies [66].
References
1. Hammer SM, Abergh AJ, Erro JJ et al (2006) 13. Pei J, Yin N, Ma X, Lai L (2014) Systems biol-
Treatment for Adult HIV Infection. JAMA ogy brings new dimensions for structure-based
296(7):410–425 drug design. J Am Chem Soc
2. Hopkins AL, Mason JS, Overington JP (2006) 136(33):11556–11565
Can we rationally design promiscuous drugs? 14. Brötz-Oesterhelt H, Brunner NA (2008) How
Curr Opin Struct Biol 16:127. Available on: many modes of action should an antibiotic
http://www.sciencedirect.com/science/arti- have? Curr Opin Pharmacol 8(5):564–573
cle/pii/S0959440X06000157. Access 20 15. Knight ZA, Lin H, Shokat KM (2010)
October, 2017 Targeting the cancer kinome through polyphar-
3. Jenwitheesuk E, Horst AJ, Rivas KL, Van macology. Nat Rev Cancer 10(2):130–137
Voorhis WC, Samudrala R (2008) Novel para- 16. Kumar R, Sharma A, Tiwari RK (2016)
digms for drug discovery: computational mul- Computational drug designing and develop-
titarget screening. Trends Pharmacol Sci ment: an insight. IRJET 03(05):10–14
29(2):62–71 17. Jadhav MN, Kokil GR, Harak SS, Wagh SB
4. Jenwitheesuk E, Samudrala R (2003) Improved (2013) Direct and indirect drug design
prediction of HIV-1 protease-inhibitor bind- approaches for the development of novel tricy-
ing energies by molecular dynamics simula- clic antipsychotics: potential 5-HT2A antago-
tions. BMC Struct Biol 3(1):2 nist. J Chem 2013:1–10
5. Jenwitheesuk E, Samudrala R (2005) 18. Shibi IG, Aswathy L, Jisha RS, Masand VH,
Identification of potential multitarget antima- Gajbhiye JM (2016) Virtual Screening
larial drugs. JAMA 294(12):1490–1491 Techniques to Probe the Antimalarial Activity
6. Jones R, Gazzard B (2006) The cost of antiret- of some Traditionally Used Phytochemicals.
roviral drugs and influence on prescribing poli- Comb Chem High Throughput Screen
cies. Int J STD AIDS 17(8):499–506 19(7):572–591
7. Metzner KJ, Allers K, Rauch P, Harrer T 19. Wang Y, Xiao J, Suzek TO, Zhang J, Wang J,
(2007) Rapid selection of drug-resistant HIV-1 Bryant SH (2009) PubChem: a public infor-
during the first months of suppressive ART in mation system for analyzing bioactivities of
treatment- naive patients. AIDS small molecules. Nucleic Acids Res 37(Web
21(6):703–711 Server issue):W623–W633
8. Santhosh KC, Paul GC, De Clercq E, 20. Liu K, Feng J, Young SS (2005) PowerMV:a
Pannecouque C, Witvrouw M, Loftus TL, software environment for molecular viewing.;
Turpin JA, Buckheit RW Jr, Cushman M descriptor generation.; data analysis and hit
(2001) Correlation of anti-HIV activity with evaluation. J Chem Inf Model 45(2):515–522
anion spacing in a series of cosalane analogues 21. Kumar M, Makhal B, Gupta VK, Sharma A
with extended polycarboxylate pharmacoph- (2014) In silico investigation of medicinal
ores. J Med Chem 44(5):703–714 spectrum of imidazo-azines from the perspec-
9. Scheeff ED, Bourne PE (2005) Structural evo- tive of multitarget screening against malaria,
lution of the protein kinase–like superfamily. tuberculosis and Chagas disease. J Mol Graph
PLoS Comput Biol 1(5):e49 Model 50:1–9
10. Shoichet BK (2005) Virtual screening of 22. Yuvaniyama J, Chitnumsub P,
chemical libraries. Nature Kamchonwongpaisan S, Vanichtanankul J,
432(7019):862–865 Sirawaraporn W, Taylor P, Walkinshaw MD,
11. Volberding PA, Lagakos SW, Grimes JM, Stein Yuthavong Y (2003) Insights into antifo- late
DS, Rooney J, Meng TC, Fischl MA, Collier resistance from malarial DHFR-TS structures.
AC, Phair JP, Hirsch MS (1995) A comparison Nat Struct Biol 10(5):357–365
of immediate with deferred zidovudine therapy 23. Perozzo R, Kuo M, Sidhu AB, Valiyaveettil JT,
for asymptomatic HIV-infected adults with Bittman R, Jacobs WR Jr, Fidock DA,
CD4 cell counts of 500 or more per cubic mil- Sacchettini JC (2002) Structural elucidation of
limeter. N Engl J Med 333(7):401–407 the specificity of the antibacterial agent triclo-
12. Wierenga RK (2001) The TIM-barrel fold: a san for malarial triclosan for malarial enoyl acyl
versatile framework for efficient enzymes. carrier protein reductase. J Biol Chem
FEBS Lett 492:193. Available on http://doi. 277(15):13106–13114
wiley.com/10.1016/S0014-5 793%2801% 24. Merckx A, Echalier A, Langford K, Sicard A,
2902236-0. Access 9 October, 2017 Langsley G, Joore J, Doerig C, Noble M,
Endicott J (2008) Structures of P. falciparum 35. Guo S, Bezard E, Zhao B (2005) Protective
protein kinase 7 identify an activation motif effect of green tea polyphenols on the SH-SY5Y
and leads for inhibitor design. Structure cells against 6-OHDA induced apoptosis
16(2):228–238 through ROS-NO pathway. Free Radic Biol
25. Dias MV, Faim LM, Vasconcelos IB, De Med 39(5):682–695
Oliveira JS, Basso LA, Santos DS, De Azevedo 36. Maqbool M, Manral A, Jameel E, Kumar J,
W (2007) Structure of shikimate kinase from Saini V, Shandilya A, Tiwari M, Hoda N,
Mycobacterium tuberculosis complexed with Jayaram B (2016) Development of
ADP and shikimate at 1.9 Å of resolution. Acta cyanopyridine-triazine hybrids as lead multitar-
Crystallogr Sect F 63:1–6 get anti-Alzheimer agents. Bioorg Med Chem
26. Wang S, Eisenberg D (2003) Crystal structures 24(12):2777–2788
of a pantothenate synthetase from M. tubercu- 37. Palanimuthu D, Poon R, Sahni S, Anjum R,
losis and its complexes with substrates and a Hibbs D, Lin HY, Bernhardt PV, Kalinowski
reaction intermediate. Protein Sci DS, Richardson DR (2017) A novel class of
12(5):1097–1108 thiosemicarbazones show multi-functional
27. Li de la Sierra I, Munier-Lehmann H, Gilles activity for the treatment of Alzheimer’s dis-
AM, Bârzu O, Delarue M (2001) X-ray struc- ease. Eur J Med Chem 9(139):612–632
ture of TMP kinase from Mycobacterium 38. Barnham KJ, Masters CL, Bush AI (2004)
tuberculosis complexed with TMP at 1.95 Å Neurodegenerative diseases and oxidative
resolution. J Mol Biol 311(1):87–100 stress. Nat Rev Drug Discov 3(3):205–214
28. Basavannacharya C, Robertson G, Munshi T, 39. Ayton S, Lei P, Bush AI (2013) Metallostasis in
Keep NH, Bhakta S (2010) ATP- dependent Alzheimer’s disease. Free Radic Biol Med
MurE ligase in Mycobacterium tuberculosis: 62:76–89
biochemical and structural characterisation. 40. Hegde ML, Bharathi P, Suram A, Venugopal
Tuberculosis 90(1):16–24 C, Jagannathan R, Poddar P, Srinivas P,
29. Bond CS, Zhang Y, Berriman M, Cunningham Sambamurti K, Rao KJ, Scancar J, Messori L,
ML, Fairlamb AH, Hunter WN (1999) Crystal Zecca L, Zatta P (2009) Challenges associated
structure of Trypanosoma cruzi trypanothione with metal chelation therapy in Alzheimer’s
reductase in complex with trypanothione, and disease. J Alzheimers Dis 17(3):457–468
the structure-based discovery of new natural 41. Vickers JC, Dickson TC, Adlard PA, Saunders
product inhibitors. Structure 7(1):81–89 HL, King CE, McCormack G (2000) The
30. Ladame S, Castilho MS, Silva CH, Denier C, cause of neuronal degeneration in Alzheimer’s
Hannaert V, Périé J, Oliva G, Willson M disease. Prog Neurobiol 60(2):139–165
(2003) Crystal structure of Trypanosoma cruzi 42. Wolfe DM, Lee JH, Kumar A, Lee S, Orenstein
glyceraldehyde-3-phosphate dehydrogenase SJ, Nixon RA (2013) Autophagy failure in
complexed with an analogue of 1,3-bisphos- Alzheimer’s disease and the role of defective
pho-d-glyceric acid. Eur J Biochem lysosomal acidification. Eur J Neurosci
270(22):4574–4586 37(12):1949–1961
31. Amaya MF, Watts AG, Damager I, Wehenkel 43. Morphy R (2010) Selectively nonselective
A, Nguyen T, Buschiazzo A, Paris G, Frasch kinase inhibition: striking the right balance.
AC, Withers SG, Alzari PM (2004) Structural J Med Chem 53(4):1413–1437
insights into the catalytic mechanism of 44. Martin MP, Olesen SH, Georg GI, Schönbrunn
Trypanosoma cruzi trans-sialidase. Structure E (2013) Cyclin-dependent kinase inhibitor
12(5):775–784 dinaciclib interacts with the acetyl-lysine recog-
32. Abad-zapatero C, Metz JT (2005) Ligand effi- nition site of bromodomains. ACS Chem Biol
ciency indices as guideposts for drug Discovery. 8(11):2360–2365
Drug Discov Today 10(7):464–469 45. Parry D, Guzi T, Shanahan F, Davis N,
33. Azam F, Mohamed N, Alhussen F (2015) Prabhavalkar D, Wiswell D, Seghezzi W,
Molecular interaction studies of green tea cat- Paruch K, Dwyer MP, Doll R, Nomeir A,
echins as multitarget drug candidates for the Windsor W, Fischmann T, Wang Y, Oft M,
treatment of Parkinson’s disease: computa- Chen T, Kirschmeier P, Lees EM (2010)
tional and structural insights. Network Dinaciclib (SCH 727965), a novel and potent
26(3-4):97–115 cyclin-dependent kinase inhibitor. Mol Cancer
34. Braicu C, Ladomery MR, Chedea VS, Irimie Ther 9(8):2344–2353
A, Berindan-Neagoe I (2013) The relationship 46. Carlino L, Rastelli G (2016) Dual kinase-
between the structure and biological actions of bromodomain inhibitors in anticancer drug
green tea catechins. Food Chem discovery: a structural and pharmacological
141(3):3282–3289 perspective. J Med Chem 59(20):9305–9320
47. Dixon SL, Smondyre AM, Knoll EH, Rao SN, 58. Gogoladze G, Grigolava M, Vishnepolsky B,
Shaw DE, Friesner RA (2006) PHASE: a new Chubinidze M, Duroux P, Lefranc MP,
engine for pharmacophore perception, 3D QSAR Pirtskhalava M (2014) DBAASP: database of
model development, and 3D database screening: antimicrobial activity and structure of peptides.
1. Methodology and preliminary results. FEMS Microbiol Lett 357:63–68
J Comput Aided Mol Des 20(10-11):647–671 59. Verslyppe B, De Smet W, De Baets B, De Vos
48. Dixon SL, Smondyrev AM, Rao SN (2006) P, Dawyndt P (2014) Straininfo introduces
PHASE: a novel approach to pharmacophore electronic passports for microorganisms. Syst
modeling and 3D database searching. Chem Appl Microbiol 37:42–50
Biol Drug Des 67(5):370–372 60. Dresen S, Ferreirós N, Gnann H, Zimmermann
49. Allen BK, Mehta S, Ember SWJ, Schonbrunn R, Weinmann W (2010) Detection and identi-
E, Ayad N, Schürer SC (2015) Large-scale fication of 700 drugs by multi-target screening
computational screening identifies first in class with a 3200 Q TRAP® LC-MS/MS system
multitarget inhibitor of EGFR kinase and and library searching. Anal Bioanal Chem
BRD4. Sci Rep 05:16924 396:2425–2434
50. Schürer SC, Muskal SM (2013) Kinome-wide 61. Kell DB (2013) Finding novel pharmaceuticals
activity modeling from diverse public high-qual- in the systems biology era using multiple effec-
ity data sets. J Chem Inf Model 53(1):27–38 tive drug targets, phenotypic screening and
51. Gaulton A, Bellis LJ, Bento AP, Chambers J, knowledge of transporters: where drug discov-
Davies M, Hersey A, Light Y, McGlinchey S, ery went wrong and how to fix it. FEBS
Michalovich D, Al-Lazikani B, Overington JP J 280:5957–5980
(2012) Chembl: a large-scale bioactivity data- 62. Durrant JD, Amaro RE, Xie L, Urbaniak MD,
base for drug discovery. Nucleic Acids Res Ferguson MAJ, Haapalainen A et al (2010) A
40(Database issue):D1100–D1107 multidimensional strategy to detect polyphar-
52. Huang N, Shoichet BK, Irwin JJ (2006) macological targets in the absence of structural
Benchmarking sets for molecular docking. and sequence homology (2010). PLoS Comput
J Med Chem 49(23):6789–6801 Biol 6:e1000648

53. Allen BK, Mehta S, Ember SWJ, Zhu JY, 63. Ouyang H, Liu S, Zeng W, Levitt RC,
Schönbrunn E, Ayad NG, Schürer SC (2017) Candiotti KA, Hao S (2012) An emerging new
Identification of a novel class of BRD4 inhibi- paradigm in opioid withdrawal: a critical role
tors by computational screening and binding for glia-neuron signaling in the periaqueductal
simulations. ACS Omega 2(8):4760–4771 gray. Sci World J 2012:1–9
54. Engler AC, Wiradharma N, Ong ZY, Coady
64. Del Bello F, Diamanti E, Giannella M,
DJ, Hedrick JL, Yang YY (2016) Emerging Mammoli V, Mattioli L, Titomanlio F (2013)
trends in macromolecular antimicrobials to Exploring multitarget interactions to reduce
fight multi-drug-resistant infections. Nano opiate withdrawal syndrome and psychiatric
Today 7(3):201–222 comorbidity. ACS Med Chem Lett
55. Speck-Planche A, Kleandrova VV, Ruso JM, 4:875–879
Cordeiro MN (2016) First multitarget chemo- 65. Ma S, Feng C, Zhang X, Dai G, Li C, Cheng X
bioinformatic model to enable the discovery of et al (2013) The multi-target capabilities of the
antibacterial peptides against multiple gram-pos- compounds in a TCM used to treat sepsis and
itive pathogens. J Chem Inf Model 56:588–598 their in silico pharmacology. Complement
56. Jenssen H, Hamill P, Hancock RE (2006) Ther Med 21:35–41
Peptide antimicrobial agents. Clin Microbiol 66. Gu S, Yin N, Pei J, Lai L (2013) Understanding
Rev 19:491–511 molecular mechanisms of traditional Chinese
57. Brogden KA (2005) Antimicrobial peptides: medicine for the treatment of influenza viruses
pore formers or metabolic inhibitors in bacte- infection by computational approaches. Mol
ria? Nat Rev Microbiol 3:238–250 Biosyst 9:1–5
Chapter 17
Nanoformulations for Drug Delivery: Safety, Toxicity,

and Efficacy
Antonio Lopalco and Nunzio Denora
Abstract
This chapter presents an outline of the recent available information regarding safety, toxicity, and efficacy
of nano drug delivery systems. Of particular importance is the evaluation of several key factors to design
nontoxic and effective nanoformulations. Among them, we focus on nanostructure materials and synthesis
methods, mechanisms of interactions with biological systems, treatment of nanoparticles, manufacture
impurities, and nanostability. Emphasis is given to in silico, in vitro, and in vivo models used to assess and
predict the toxicity of these new formulations. Additionally, some examples of in vitro and in vivo studies
of specific nanoderivatives are also presented in this chapter.
Key words Nanoformulation, Toxicity, Safety, Drug delivery, Efficacy, Nanotechnology,

Computational modeling
1 Introduction
The aim of this chapter is to present an outline of the recent avail-

able information regarding the safety, toxicity, and efficacy of nano
drug delivery systems. Here, several key factors are considered and
discussed to design a nontoxic and effective nanoformulation:
nanostructure materials and synthesis methods, we are referring to
the material sciences and formulation technology; mechanisms of
interactions with biological systems in the body and cells; treat-
ment of nanoparticles (NPs), their manufacture impurities and
nanostability; in silico, in vitro, and in vivo models adapted to spe-
cific questions about safety and efficacy with “nano.” Additionally,
some examples of in vitro and in vivo studies of specific nanofor-
mulations are also presented in this chapter.
Nanomedicine represents a new area of pharmaceutical tech-
nology, based on the considerably improved properties of nanofor-
mulations that make possible the treatment of several disorders
[1–5]. Nanoformulations of conventional medicines have capti-
vated the attention of several scientists for drug delivery purposes,
347
348 Antonio Lopalco and Nunzio Denora
including the treatment, diagnosis, or prevention of a disease [3,

6]. The goal of a nanoformulation is to increase the properties of
an active substance and be exclusive to the targeted delivery site [3,
7]. Dendrimers, nanocrystals, polymer NPs, and liposomes [8–11]
are examples of nanoformulations that are earning importance in
pharmaceutical research and, during the last two decades, in phar-
maceutical industry. Because of their size (in the range of nanome-
ters) nanocarriers can enter the body by inhalation, ingestion, skin
penetration, or injections and, potentially, interact with biological
systems.
If on one hand new nanoformulations can offer improved clin-
ical benefit, patient compliance, reduced side effects, as compared
to the conventional products, on the other hand they can also
reveal differences in safety, toxicity, and efficacy outcomes.
2 Nanostructured Materials and Methods of Production of NPs
NPs as drug/gene delivery systems have the potential use to carrier

drugs or genes to a targeted site in the body and/or cells. NPs can
be composed of two or more ingredients, one of which is the active
pharmaceutical ingredient (API) [3, 12]. Toxicity of a specific NP
includes not only the interaction of that species but also the inter-
action of its nanostructured materials with biological systems
(organs, tissues, cells) [3, 13]. The choice of materials and meth-
ods of preparation of NPs is based on the size and shape of NPs,
drug incorporation and release, physicochemical properties of an
API, target and biodistribution, formulation stability and its shelf-
life, safety and toxicity while preserving beneficial effects [12, 14].
Moreover, when used only as vehicle the potential unfavorable
consequences of remaining substance after drug delivery should
also be pondered. Materials and methods for the preparation of
NPs are important and can permit the preparation of NPs with
appropriate properties to guarantee suitable drug delivery and tar-
geting. An extensive range of materials and synthesis processes
[12] are accessible for the formulation of nanocarriers to deliver
molecules in living biological systems. Here we summarize the
main different approaches of preparation and the materials that
have been offered and used to make nanoformulations for the
in vivo delivery of drug/gene, in vitro diagnosis, photodynamic,
and imaging (Table 1).
There is a restricted quantity of polymeric material that can be
employed as ingredient of NPs intended to deliver an API in vivo
[12]. To justify this point, one must ponder that an appropriate
NP needs to satisfy numerous requests to be employed in such a
purpose. It is essential that nanocarriers are biodegradable and/or
eliminated from the body in a short period consenting the
repetition of administration without any probability of unrestrained
Nanoformulations for Drug Delivery: Safety, Toxicity, and Efficacy 349
Table 1
Types of chemical structures, methods of preparation and applications of some NPs used as
pharmaceutical carrier system in drug/gene delivery, photodynamic, in vitro diagnosis, and imaging
NP class (nanostructured
material) Method of preparation Application
Polymer carrier (PLA, Colloidal mill, emulsification-solvent Drug/gene
polyalkylcyanoacrylates, PEI, evaporation or diffusion, emulsification- delivery
PCL, PLGA, PEG-PLA, reverse salting-out, gelation of emulsion
PEG-PLGA) droplets, ionic gelation, polymerization,
Polymer carrier of natural and/or interfacial polycondensation,
derivatives materials (chitosan, nanoprecipitation, formation of
dextrane, gelatin, albumin, polyelectrolyte complexes, formation of
alginates, liposomes, starch) NPs from neutral nanogels [7, 10, 12]
[15–17]
Dendrimers (PAMAM, blocked Divergent, convergent, click chemistry [11, Drug/gene
polymers) 18, 19] delivery
Fullerenes (carbon-based carriers) Arc and combustion method [20, 21] Drug delivery,
Photodynamic
Ferrofluid (SPIONS, USPIONS) Physical, chemical, and biological methods Imaging (MRI)
[22, 23]
Quantum dots (Cd/Zn-selenides) Top-down and bottom-up [24] In vitro diagnosis,
Imaging
accumulation. Nanostructured materials and their degradation

products cannot be toxic and immunogenic. Table 1 provides an
inventory of the most commonly employed polymeric materials in
the composition of NPs. Currently, a restricted number of those
materials are recognized by ministry health and other national
authorities for parental administration. Some of them can be added
in oral, topical preparations or in nutrition products.
During the last two decades, many polymers including
poly(ethylene-glycol)-polysaccharides were generated. The ratio-
nale behind the elaboration of these polymers derived from the
necessity of NPs with tunable features to control the interactions
with cells in the circulatory system, proteins and with mucous
membranes, thus determining their in vivo destiny [12]. Some
copolymers were also demonstrated useful to stabilize NPs without
the necessity to add further surfactants [12].
Most of the methods for the formulation of polymeric NPs
involve two principal phases: formulation of an emulsion that cor-
responds to the first phase and formation of NPs during the second
phase of the process. This second phase is realized by precipitation
or gelation of a polymer or by a process in which monomers react
to form polymers (polymerization). Some approaches do not need
the preparation of an emulsified system to obtain the NPs. They
are founded on dispersion/precipitation processes of polymers in
liquid phases and formation of nanogels or polyelectrolyte com-

plexes of macromolecules dispersed in solutions [12].
Dendrimers are branched macromolecules with nanometer-
scale dimensions. They are defined by three components: a core,
internal dendritic branches, and a surface functionalized with
groups [8]. The assorted arrangement of these constituents yields
products of unique shapes and sizes with protected internal cores
that are ideal candidates for applications in biology and pharma-
ceutical technology. Their unique features allow them to collocate
between molecular and polymeric chemical entities. Dendrimers
can be produced by two main methods [18]. In the divergent
method, the synthesis begins from the center of the nanocarrier
and the branches are linked by the addition of building blocks in
an exhaustive/phase-wise mode. In the convergent method, the
synthesis begins from the surface, starting with the part of the mol-
ecule that finally becomes the outmost branch of the ultimate
nanocarrier. In the strategy, the number of repeated branching
cycles or final generation number is preestablished, since it deter-
mines the requisite sizes of the dendrimer [18].
Fullerenes are macromolecules of carbon with a globular, ellip-
soid, or cylindrical shape. They can be fabricated by the arc-
discharge vaporization of graphite, by chemical vapor deposition
methods, and by combustion process [20, 21]. These approaches
are not very proficient, and continuous modifications and develop-
ments in this area are required. The development of new synthetic
approaches is crucial in the progress of fullerene knowledge.
Ferrofluids or magnetic nanofluids (MNPs) are smart materials
containing ultrafine magnetic NPs suspended in a liquid vehicle,
which show both fluidity and magnetic controllability [22, 23].
They are formulated by a one-phase or a two-phase method. The
one-phase process involves the concurrent synthesis/dispersion of
MNPs in a liquid vehicle, thereby eluding processing phases such
as drying, storage, transportation, and dispersion. This technique
reduces the formation of agglomerates and can allow to form
homogeneously dispersed particles in the liquid vehicle. Impurities
such as remaining reagents exist in the liquid vehicle of single-
phase methods. Two-phase process is financially more convenient
for these formulations. In the first phase, MNPs can be made by
several synthetic paths, such as grinding/ball-milling, coprecipita-
tion, hydrothermal method, microwave-assisted synthesis, micro-
emulsion, polyol method, and by thermal decomposition. In the
second phase, MNPs are covered with a stabilizing film, yielding a
very stable magnetic sol [22, 23].
Quantum dots (QDs) are a heterogeneous group of NPs.
Their use for sensitive cellular imaging has been established in
several research areas such as cancer metastasis, lymphocyte immu-
nology, and embryogenesis. Several methods have been employed
to make colloidal QDs but, commonly, procedures for QD synthe-
sis are classified as top-down and bottom-up approaches [24].

Top-down techniques involve molecular beam epitaxy, ion implan-
tation, e-beam lithography, and X-ray lithography. Using the sub-
stitute bottom-up methodology, QDs are made by self-assembling
the solution and subsequent chemical reduction [24].
Nanocrystals are constituted of pure drug particles of nanome-
ter dimensions [25, 26]. Due to the favorable and desirable fea-
tures of high drug loading, stability, and ease of preparing them in
amount, nanocrystals were extensively employed to deliver poorly
water-soluble API. The methods of preparation of nanocrystals can
be summarized in top-down and bottom-up procedures. The big-
gest advantage in top-down process is that it is general technique
to prepare crystalline NPs and is adaptable in manufacture scale.
High-energy mechanical forces are involved in the top-down
methodologies, which can be provided either by media milling
(NanoCrystals®) or high-pressure homogenization (IDD-P®,
DissoCubes® and Nanopure®). The disadvantage of this technol-
ogy includes high energy and time consumption as well as impurity
from pulverizing media, leading to unpredicted toxic undesirable
effects. The bottom-up process grows nanocrystals from solution,
which includes two fundamental phases: nucleation and conse-
quent crystal growing [25]. The first step is critical to realize small
and uniform nanocrystals. In the bottom-up process it is essential
to promote rapid and homogeneous nucleation. Sonication can be
introduced to provide cavitation effects. Spray-drying and freeze-
drying are common ways to remove solvent. Recently, supercritical
fluid (SCF), like supercritical carbon dioxide, has been used to pre-
pare nanocrystals by taking advantage of the unique physical prop-
erties of SCF, with combined diffusivity like gas, solubilizing like
liquid, and low environmental impact [25, 27].
3 Treatment of NPs After Preparation, Manufacture Impurities, and Nanostability
Once NPs are obtained, several modes of treatments must be

applied to them [12]. They can include purification to eliminate
impurities and surplus of unreacted materials required during man-
ufacture, sterilization to destroy microorganisms and eliminate
bacteria spores, drying and/or concentration to stabilize and con-
serve the final formulations. Impurities in nanoformulations can be
present after preparation or be formed during the preparation,
treatment, and storage. Depending on the process of formulation,
impurities comprise organic solvents, oil, surfactants, unreacted
monomers, polymerization initiators, salts, surplus of surfactants
or stabilizing molecules, and sizeable polymeric aggregates.
Purification is needed to achieve an extremely purified suspension
of NPs that can be used as pharmaceutics and administered in vivo
by certain routes [12]. NPs manufactured by a reverse salting-out
method in suspensions having elevated amounts of salt [28] must

be purified by desalination before their administration by intrave-
nous injection. Nanocapsules containing an aqueous core dispersed
in oil can be taken by oral via after being cleaned and transferred in
an aqueous medium [29]. Evaporation under vacuum is a useful
methodology to eliminate bulky amounts of organic chemicals
with high vapor pressure at 25 °C and a portion of water [30].
Filtrations through mesh or filters and centrifugation at low cen-
trifugal forces for the sedimentation of particles can be used to
eliminate outsized polymeric aggregates which produced during
formulations of the polymer NP suspensions [31]. The centrifuga-
tion method does not guarantee the exclusion of all aggregates
with a size above a delimited diameter as filtration on determined
membrane does. This technique is not appropriate to treat metal
colloids or NPs having an elevated density, because they deposit as
aggregates [12]. Ultracentrifugation procedures consist in high
speed centrifugations [32]. Nanospheres can sediment and be con-
centrated in a pellet which is then removed from a dispersing liquid
phase. The dispersing medium having the undesirable impurities
can be eliminated and substituted by fresh medium without impu-
rities. The principal disadvantage of this method is that nano-
spheres cannot be easily redispersed. Aggregates can be found even
after the use of vortex or ultrasound [33].
Dialysis is an efficient method of filtration and purification. It
can be executed using cellulose membranes of determined molecu-
lar weight cutoff consenting components with a specific molecular
weight to permeate the membranes and diffuse toward the counter
dialyzing medium from a region of higher density to a region of
lower density of solutes. The inconveniencies of this procedure of
purification are high probability of microbial impurities and pre-
mature release of NP load that can happen because the prolonged
purification period required and the enormous volumes of counter
dialyzing medium used [12].
For parenteral uses these nanoformulations must be sterile and
meet pharmacopoeia requirements. The more used procedures are
heat sterilization or autoclaving, gamma irradiation and sterile fil-
tration [12]. Heat sterilization implicates temperature of 120 °C
that can determine decomposition or degradation of an API and/
or the NPs materials. Strength, viscoelasticity, and other mechani-
cal properties of polymeric materials with a glass transition or a
melting point below the temperature of 120 °C are intensely
reduced after heat sterilization. Nanocapsule size increased after
autoclaving. This consequence was due to the swelling of the
membrane composed of poly(isobutylcyanoacrylate) component
of the nanocapsules and the dilatation of oily phase [34]. Heat can
also catalyze certain reactions with substances such as surfactants
and polymers constituting NPs [35]. Autoclaving of polymeric
materials based gene transfer complexes can produce entire depri-

vation of transfection ability [36].
Gamma irradiation can disturb the performance of NPs because
energy transfer can provoke fragmentation of covalent bonds and
induce the production of free radical species which can destroy the
polymers forming the drug delivery system or form products with
toxicological risk. Gamma irradiation of PCL NPs caused a growth
of molecular weight due to cross-linking reactions between poly-
meric chains or between the polymer and surfactant [12, 35]. It
has been reported that polymeric chain scissions occurred after
application of gamma irradiation treatment of PLA NPs. As result
to the variation of the polymer constituting the NPs due to the
application of gamma irradiation, the degradation rate of the
nanodelivery system and the release properties of savoxepine
loaded in NPs were modified and speeded [37].
Sterile filtration can be used as a substitute technique for mate-
rials that are chemically or thermally susceptible. No unfavorable
consequence on the polymers or the drugs was described using this
technique of sterilization [38]. Although this process demon-
strated its effectiveness to annihilate all vegetative microorganisms,
enhancements are desired to eliminate bacterial spores.
Storage of many pharmaceutical nanoformulations as suspen-
sions presents risk of microbiological impurity, polymeric degrada-
tion by hydrolysis or oxidation, changing in physicochemical
stability due to aggregation and sedimentation of particles, decrease
of the biological activity of the API [12, 39]. Therefore, pharma-
ceutical formulations need to be stored as a dried form, using
appropriate procedures such as freeze-drying or spray-drying.
During freeze-drying method, numerous complications can
arise and result in a loss of properties and characteristics of the
NPs. Crystallization of ice can generate mechanical stress on NPs
resulting in destabilization of the nanostructures. Higher concen-
trations of NPs in the ultimate dried product can lead to aggrega-
tion and in some situations permanent coalescence of NPs. Adding
cryoprotective agents may enhance resistance of NPs toward freez-
ing and drying damages and improve stability during long term
storage [40]. Ultimately, freeze-dried nanocapsules should be kept
in reserve at temperatures below the glass transition temperature
of the polymeric materials constituting the formulation to avoid
aggregation of the nanoformulations [12, 41].
An alternative to freeze-drying is the spray-drying method,
which converts aqueous nanosuspensions into dried particles under
a nonstop procedure. The main problem is the presence of
aggregates that are difficult to dissociate during the dispersion of
the powder [12, 42].
4 Mechanisms of Interactions with Cellular/Intracellular Systems of NPs
The use of NPs as drug delivery systems to organs or cells is very

important as much as the destiny of them inside the cell, in subcel-
lular environments. Most of the time, particles, once entered the
cells, are degraded in endosomes and lysosomes. For therapeutic
action of the entrapped drug in a nanoformulation release into
cytosolic cellular environment is required. Moreover, NPs of
20 nm can be taken in cells without contribution by endocytosis
mechanisms [43]. Physicochemical properties and characteristics
of NPs such as zeta potential can control their destiny in a cell.
External modification on the surface of gold NPs with polyethyl-
ene glycol caused an important internalization in endosomes and
cytosol, and localized in the nuclear region [12, 44]. Poly(dl-
lactide-co-glycolide) (PLGA) NPs were internalized into cells by
endocytosis [45]. It has been suggested that a modification in zeta
potential from negative to positive of the PLGA nanoformulations
affected their intracellular distribution and caused the escape from
the endosomes into the cytoplasm of cells. This hypothesis was
proved by data showing that negatively charged polystyrene NPs
could not arrive at the cytosol but persisted in the endosomes [45].
QDs (<10 nm) have been employed for definite targeting of
dots covered with peptides to the vascular systems of lungs and
tumors [46]. Specific targeting to the cells of pigmented layer of
retina in the eyes has been showed [47]. Moreover, it has been
observed the uptake of QDs by lymph nodes until 4 months after
administration [48]. PEG coating nullified their uptake by the
endothelial and reticular system of organs like liver and spleen.
PEGylated magnetic NPs with a size of 40–50 nm were quite well
kept by endocytosis [12].
Variations of the surface of NPs give chances for medical and
therapeutic uses like drug targeting in terms of cellular binding,
uptake and transport within the cell. The presence of carbohydrate
binding ligands on the shell of PLGA nanospheres could increase
cellular binding [49].
Passage of NPs (~200 nm) across the blood brain barrier
(BBB) was proposed to be achievable by their toxic effects on
microvascular endothelial cells of the brain [50]. The magnitude of
cerebral uptake of negatively charged NPs at a lower concentration
was higher than neutrally or positively charged preparations at the
same concentration. The charge on the surface of NPs must be
pondered for cerebral toxic effects and distribution profile. NPs
coated with polysorbate (Tween) nonionic surfactant and e mulsifier
determined transport of drugs across the BBB. The adsorption of
plasmatic lipoproteins to NPs could affect their endocytosis and
transport of drugs across the BBB. The mechanism involved in
NPs transportation was proposed to be endocytosis through low
density lipoprotein (LDL) receptor present on the microvascular

endothelial cells. Supplementary studies discovered the role of
apolipoprotein-E and showed that its variants did not recognized
LDL receptors failing in the transport of drugs across the barrier.
It was proposed that the identification/interaction with LDL
receptors on cerebral microvascular endothelial cells was related to
the cerebral uptake of drugs [12, 51–53].
Poly(butyl)cyanoacrylate NPs covered with surfactants have
been formulated to deliver drugs to the central nervous system
(CNS) [54]. The cytotoxic effects of paclitaxel (PX) in cetyl alco-
hol/polysorbate nanoformulations were assessed in an in-situ cere-
bral perfusion model with rat [55]. The data showed that
entrapment of PX in nanoformulations appreciably improved cere-
bral drug uptake and consequently its cytotoxicity toward
P-glycoprotein (P-gp) overexpressing cancer cells. The hypothesis
was that PX NPs reduced PX binding to P-gp and successive efflux
from cancer cells [12].
Other paths for reaching the CNS, by-passing the blood cere-
bral barrier, can be through movement along the olfaction and
trigeminal system after deposition on the mucous of the olfactory
epithelium in the upper region of the nasal cavity [56]. Manganese
oxide nanoformulations were shown to migrate to the CNS by
olfactory path [57].
Copolymer of poly(ethylene glycol) and poly(lactic acid)
(PEG-PLA) was used to prepare NPs functionalized on the surface
with bioactive lectin ligands. These formulations showed higher
uptake by inhalation route. Wheat germ agglutinin was also
employed which bound to N-acetyl-d-glucosamine and sialic acid
richly present in the nasal cavity [12].
NPs can interact with the immune system. The main role of
the immune system is to defend the host from bacteria and xeno-
biotics. When NPs are in the body, they are in contact with cells of
the immune system and can stimulate the immune response. NPs
can also affect the immune system’s recognition of other immuno-
genic xenobiotics and activate or suppress the immune response.
Usually, immune response progressively leads to the elimination of
xenobiotic from the body, but NP contact with immune response
can also develop toxic effects [12, 14, 58].
Most of the compounds used for the development of new
nanoformulations are biodegradable polymers or copolymers, like
PLA, PLGA, and the release of the drug occurs after their degrada-
tion. It has been observed that NPs including biodegradable
nanomaterials can be sequestered by the mononuclear phagocytic
system (MPS) cells in organs like liver and spleen [59–61].
Nanocrystals in the blood circulatory system can be accumu-
late as exogenous components by the cells of MPS in organs like
liver, spleen and lung [12, 16, 62].
5 In Vitro, In Vivo, and In Silico Nanotoxicological Assays
In vitro approaches are convenient in nanotoxicology investigation

since they can generate the same research results quickly and eco-
nomically without using expensive animal models [13]. In vitro
approaches that create unambiguous and quantitative results of
toxicity are valued for primarily assessing the estimated biocompat-
ibility of nanoformulations. Commonly mentioned models include
the test of cell membrane integrity (LDH), mitochondrial function
(MTT), and immunochemistry indicators for cell death triggered
by normal process (apoptosis) and by external perturbations
(necrosis). These techniques offer limited data on the mechanism
and origin of cell toxicity or death. The function of metabolic
activity in tetrazolium-based toxicity assay can measure cell viabil-
ity, but this assay cannot clarify the main mechanism of mitochon-
drial inactivity or cell death. In fact, any effects from NPs contact
(breaking down of the membrane of a cell, cell cycle checkpoint
and programmed cell death) can stop mitochondrial activity.
Colorimetric analyses such as trypan blue and neutral red also offer
limited info about the mechanism of cellular apoptosis/necrosis, as
they can only distinguish live from dead cells. Accuracy and preci-
sion of this approach of investigation for in vitro toxicity of NPs
can also generate problems due to the contacts of the NPs with the
color-generating dyes [13]. Moreover, intrinsic concerns involving
dose and time-course effects, cell–cell and cell–matrix interactions,
and physical and chemical features of nanoformulations in cell cul-
ture environment can give incorrect data. The toxicity is typically
higher in vitro experiments, because higher amounts of NPs can be
commonly employed in in vitro compared with in vivo experi-
ments. Latest findings have showed correlations between the
in vitro and in vivo toxicity of several NPs. Sayes et al. evaluated the
reliability of in vitro screening analyses to give an in vivo toxicity
prediction of nanoformulations in the lungs of rats [63]. In vitro
techniques are generally valuable to detect properties of NPs that
can highlight their toxicity and/or generate a classification of NP
toxicity for mechanistic studies.
NPs can be administrated by several routes. The principal
routes among them are intravenous, transdermal, subcutaneous,
inhalation, intraperitoneal, and oral [64]. In vivo results were
reported and presented from inhalation and intratracheal instillation
in rat, mouse and other different types of rodents, including the
toxicity data for gold NPs, carbon nanotubes and QDs [65].
Animal models could be valuable to identify and study certain
properties that are not evident with in vitro techniques, such as the
toxicity-kinetic in the body, we are referring to the four processes
of absorption, distribution, metabolism and excretion involved
when a pharmaceutical product is administered.
NPs could also show harmful effects on reproductive system of

women and the stages of fetal growth. The adverse outcomes were
connected to NP composition, changes on surfaces, dose, contact
via and types of animals. In pregnancy, NP-based drug formulation
could have the potential to selectively target the placenta and/or
fetus, allowing drugs to be administered with marginal probability
of off-target outcomes. A significant quantity of investigation has
been carried out on maternal-placental-fetal NP uptake, transfer
and toxicity using in vivo, ex vivo, and in vitro models [7, 66].
Because long term consequences of nanoformulations are
challenging to mimic in in vitro and/or in vivo experiments,
structure-based modeling and computational simulation of the
interactions of nanomaterials with biomolecular systems are pre-
cious and useful methodologies to predict potential toxicological
outcomes. Moreover, they offer valuable platforms for designing
important in vitro and in vivo investigations. Computational and
structural modeling approaches can be employed to find critical
patterns in the NPs-biological molecules (i.e., DNA, RNA, pro-
teins) contacts that can result in the elaboration of new methods
and instruments to explore the wide world of nanotech. A note-
worthy example is a new analysis by Calvaresi et al. [67] showing
that hydrophobic pockets in proteins and enzymes can be employed
as kits for classifying and dividing NPs of different dimensions,
forms and chirality. In silico systems to give a prediction of NPs
toxicity can complement and/or substitute some costly and time-
consuming studies, especially early in the design elaboration of
novel nanoformulation. Quantitative structure–activity relation-
ships (QSARs) are theoretical models that can be employed to pre-
dict the physicochemical and biological characteristics of a molecule
[68]. According to the QSAR model, calculating some molecular
parameters for a group of compounds and knowing only part of
the experimental data on the activity of a group of these molecules,
it is possible to interpolate the unknown activity of further com-
pounds from the molecular descriptors using an appropriate math-
ematical model [69]. In silico predictive toxicology methods are
fast and cost efficient alternative to analytical methods to estimate
and identify toxic outcomes of materials with a size of nanometers
[70]. Sayes et al. employed QSAR methodology to develop math-
ematical models for the prediction of cellular membrane perturba-
tions resulting from several physicochemical factors of NPs. They
found that dimension, concentration and charge of particles in
ultrapure aqueous medium were distinctive features on cytoplasm
leaking [71]. Puzyn et al. applied nano-QSAR to predict the toxic-
ity of 17 different metal oxides NPs and could describe the rela-
tionship between NP structure and cytotoxicity to E. coli cells [69].
In silico methodology can be applied to both in vivo and in vitro
results, therefore the quality of the in vivo or in vitro results is of
essential significance. In vitro and in silico approaches for acute
chemical toxicity can offer sufficient results for many bulk materi-
als; but in vivo interaction of nanoformulations and biological sys-
tems is relatively complex and dynamic [13, 72].
6 In Vitro and In Vivo Case Studies for Specific NPs
In the past two decades, several papers have described the safety
and toxicity of NPs formed from biodegradable materials, like
PLGA, and nonbiodegradable materials, including NPs like den-
drimers, carbon nanotubes, and quantum dots [2, 3, 13, 14, 73].
Usually, nanoformulations constituted from biodegradable
polymers are estimated to display scarcer toxic effects than nonbio-
degradable polymers. Semete et al. researched the in vivo toxicity
and biodistribution of poly(lactic-co-glycolic acid) NPs. One week
after oral administration in mouse, 40% percent of those NPs were
concentrated in the liver, and the rest were distributed in brain and
kidney without apparent toxicity [13, 74]. Chemical composition
of biodegradable NPs and their degradation products can affect
the biological effects. Polyesters such as PLGA or polycaprolactone
(PCL) undergo hydrolysis and enzymatic degradation after implan-
tation into the body, forming lactic acid, glycolic or capronic acid,
which are biocompatible entities. Aside from size-dependent toxic-
ity due to ROS-generating ability, dimension of a PLGA or PCL
particle can change the rate of degradation of the polymeric matrix
[75].
Polyacrylate NPs were among the first NPs investigated for
controlled delivery of biological agents. Recently, Song et al. [76]
examined the human toxicity of polyacrylate NPs prepared from
polymerization of unsaturated monomers, such as methyl meth-
acrylic amide or cyanoacrylates. The toxicity of the larger cyanoac-
rylate NPs was correlated with the chemical properties and
molecule chain length; pathological investigation indicated non-
specific pulmonary inflammation, pulmonary fibrosis, and foreign-
body granulomas of pleura after exposure [13, 76–78].
Because of their specific nature dendrimers can find applica-
tions in drug delivery, gene transfection, and technology [8, 14].
Even though their small size limits extensive drug incorporation
into the dendrimers, their nature and polymeric structure allows
for drug loading onto the internal dendritic branches. It has been
shown that functionalization of the surface with specific antibodies
may further enhance potential targeting and reduce side effects
[79]. The limited clinical studies and published data available on
the toxicity of this class of particles make it impossible to define
their safety and toxicity. More recently, they have been explored as
potential antimicrobial agents [80].
In view of their potency for induction of reactive oxygen spe-
cies (ROS) after photoexcitation, fullerenes have also been studied
as antimicrobial molecules [81, 82]. Various reports with fuller-

enes have been published regarding the ecotoxicity of these build-
ing blocks in nanomaterials. Experiments with uncoated,
water-soluble, colloidal fullerenes showed that the 48-h LC50 in
Daphnia magna increased from 460 to 800 ppb [83, 84]. In large-
mouth bass lipid peroxidation was seen in the brain and glutathi-
one depletion in the gill after exposure to 0.5 ppm fullerene nC60
for 2 days [85].
Carbon nanotubes are long carbon-based tubes that can be
single-walled, double-walled, or multiwalled and have the poten-
tial to act as biopersistent fibers. They are also known as cylindrical
fullerenes. In vitro incubation of keratinocytes and bronchial epi-
thelial cells with high doses of single-walled carbon nanotubes
results in ROS generation, lipid peroxidation, oxidative stress,
mitochondrial dysfunction, and changes in cell morphology [86].
Recent findings showed that platelet aggregation was also induced
by nanotubes [87]. Numerous findings using intratracheal instilla-
tion of elevated amounts of nanotubes in rat and mouse confirmed
chronic lung inflammation, including foreign-body granuloma
formation, and interstitial fibrosis [88]. Two in vivo studies showed
that nanotubes induced lung granulomas after intratracheal admin-
istration indicating that they were more toxic than quartz particles
well-known for their lung toxicity [88, 89].
Adsorption, distribution, metabolism, and elimination of QDs,
and consequently their toxic effects, depend on various aspects
connected to their intrinsic physicochemical properties. Studies
performed to investigate QDs toxicity are few [90]. In vitro studies
have indicated that QDs may be toxic and their toxicity could be
due to the surface coating [91, 92]. Choi et al. confirmed that QD
toxicity decreased after changes on the shell with N-acetylcysteine,
whereas the nonmodified cadmium telluride QDs provoked oxida-
tive degradation of lipids in cell membrane, resulting in cell dam-
ages [93]. In other studies, conducted by Lovric et al. QDs were
cytotoxic by stimulation of ROS resulting in cell damages, involv-
ing cytoplasmic membranes, mitochondria, and nucleus [94].
Consideration is necessary in using compounds on the shell of NPs
in terms of induction of toxic effects. Nevertheless, the materials of
the center of QDs influenced the probable toxicity of QDs. The
toxicity of cadmium containing QDs was proposed to be to the
liberation of highly toxic concentrations of cadmium ions [95].
Golden NPs/nanoshells are widely used and commercilally
available in various size ranges. Cells can take up gold NPs without
cytotoxic effects [96]. For gold nanorods the cytotoxicity could be
attributed to the residual presence after washing of the stabilizer
cetrimonium bromide (CTAB). PEG-modified gold nanorods
with removing the excess CTAB did not show cytotoxicity [97].
Gold solutions are also employed to formulate nanoshells consti-
tuted of metals like gold, copper or silver useful in Magnetic
Resonance Imaging (RMI), and gold–silica for photothermal abla-

tion of cancer cells [98, 99]. In vitro experiments of the nontar-
geted nanoshells showed no toxicity for the cancer cells, however
after binding to the cancer cells, death could be achieved after laser
stimulation [99, 100]. Also in vivo positive results were found with
photothermal ablation treatment in a mouse model for colon car-
cinoma after intravenous injection of PEGylated gold nanoshells
with a size of 130 nm [101].
Silica NPs are made with inorganic materials andalso can be
employed as drug delivery systems. For these NPs both in vitro
toxic effect and nontoxic response could be detected. Silica NPs
with a size of 15 and 46 nm showed a dose-dependent toxicity
in vitro [102]. SiO2 exposure increased ROS and reduced gluta-
thione levels, increasing oxidative stress. Chang et al. discovered
that silica NPs were toxic at elevated amounts as shown by a
decrease in cell viability/cell proliferation and membrane damage
indicated by lactate dehydrogenase release from the cells [103]. In
other findings, an alveolar macrophage cell line (MHS) was found
to be more predisposed for NP induced toxicity than the A549
cell line a type II human pulmonary epithelial cell lung. The cyto-
toxicity was suggested to be due to the phagocytic properties of
the macrophage cell line [104]. Low or no cell toxicity was
observed for cationic silica NPs using amino-hexyl-amino-
propyltrimethoxysilane as a surface modification [105].
Conventional amphotericin B (AMB) formulation, Fungizone®
AMB deoxycholate, has been considered the gold standard of ther-
apy for the treatment of systemic fungal infections since 1960s
[106, 107]. Its clinical use is limited by accompanying side effects,
such as high rates of nephrotoxicity. During the 1980s new AMB
delivery systems based on liposomes and lipid-complexes were for-
mulated to allow better doses over a protracted period with reduced
toxicity in the kidneys. In 1990s, FDA approved a lipid-based
AMB formulation (Albelcet®), consisting of one mole of the anti-
biotic complexed with one mole of two different phospholipids.
Clinical investigations of conventional AMB formulation versus
this new nanoformulation did not show any modification in clinical
therapy rates and increased infusion correlated toxicities [2].
7 Final Considerations
Many studies have demonstrated that nanotechnology in medi-

cine and more particularly drug delivery can offer many advan-
tages, such as targeting to specific cells, and modified drug release
kinetics and biodistribution profiles. Nanotoxicology is still an
unexplored wide world, and most of the findings have focused on
acute toxicity. Long-term toxicity and investigations of chronic
contact are important to understanding the toxicology of
nanoformulations in vivo. Evaluation of toxicity and safety of a

nano drug delivery system is a challenge as many factors can cause
nanotoxicity. Many unpredictable interactions with biological sys-
tems may occur. Validated analytic methodologies and accurately
designed experiments can elucidate the mechanisms of toxicity so
that nanoformulations can safely be used as therapeutics or diag-
nostic instruments. The development in nanobiotechnology,
material synthesis, and computer simulation studies can change
how nanoformulations can be engineered that could guarantee
safety and efficacy, and avoid toxicity.
New screening tests and approaches are required to reveal the
biodistribution, and the interactions of nanoformulations with
biological systems and thus help understand the biochemical path-
ways that regulate cell functions.
In the final analysis, even though nanoformulations are pro-
foundly modifying the healthcare area in the prevention, diagnosis,
and treatment of a disease, it is also estimated to change current
medical criteria associated with safety and ethics control. Definite
regulatory institutes and legislatures have been employed in some
countries concerning nanoformulations in pharmaceutical mar-
kets. Currently, EMA and FDA, the two agencies for the evalua-
tion of medicinal products in European Union and the USA,
respectively, have recognized some regulatory agencies and legisla-
tures with rules to manage the use of nanoformulations in
medicine.
References
1. Bawa R (2009) Nanopharmaceuticals for drug N (2017) Bridging pharmaceutical chemistry

delivery - a review. Drug Deliv 3:122–127 with drug and nanoparticle targeting to inves-
2. Weissig V, Pettinger TK, Murdock N (2014) tigate the role of the 18-kDa translocator pro-
Nanopharmaceuticals (part 1): products on tein TSPO. ChemMedChem 12:1261–1274
the market. Int J Nanomedicine 9:4357–4373 7. Lopalco A, Ali H, Denora N, Rytting E
3. De Jong WH, Borm PJ (2008) Drug delivery (2015) Oxcarbazepine-loaded polymeric
and nanoparticles: applications and hazards. nanoparticles: development and permeability
Int J Nanomedicine 3:133–149 studies across in vitro models of the blood-
4. Lopedota A, Cutrignelli A, Laquintana V, brain barrier and human placental tropho-
Denora N, Iacobazzi RM, Perrone M, Fanizza blast. Int J Nanomedicine 10:1985–1996
E, Mastrodonato M, Mentino D, Lopalco A, 8. Denora N, Laquintana V, Lopalco A, Iacobazzi
Depalo N, Franco M (2016) Spray dried chi- RM, Lopedota A, Cutrignelli A, Iacobellis G,
tosan microparticles for intravesical delivery Annese C, Cascione M, Leporatti S, Franco
of celecoxib: preparation and characteriza- M (2013) In vitro targeting and imaging the
tion. Pharm Res 33:2195–2208 translocator protein TSPO 18-kDa through
5. Perrone M, Lopalco A, Lopedota A, G (4)-PAMAM-FITC labeled dendrimer.
Cutrignelli A, Laquintana V, Douglas J, J Control Release 172:1111–1125
Franco M, Liberati E, Russo V, Tongiani 9. Denora N, Lopedota A, Perrone M,
S, Denora N (2017) Preactivated thio- Laquintana V, Iacobazzi RM, Milella A,
lated glycogen as mucoadhesive polymer Fanizza E, Depalo N, Cutrignelli A, Lopalco
for drug delivery. Eur J Pharm Biopharm A, Franco M (2016) Spray-dried mucoad-
119:161–169 hesives for intravesical drug delivery using
6. Iacobazzi RM, Lopalco A, Cutrignelli A, N-acetylcysteine- and glutathione-glycol chi-
Laquintana V, Lopedota A, Franco M, Denora tosan conjugates. Acta Biomater 43:170–184
10. Laquintana V, Denora N, Lopalco A, chemical features, characterization, and appli-

Lopedota A, Cutrignelli A, Lasorsa FM, cations. ChemPlusChem 79:1382–1420
Agostino G, Franco M (2014) Translocator 23. Ali A, Zafar H, Zia M, ul Haq I, Phull AR,
protein ligand-PLGA conjugated nanopar- Ali JS, Hussain A (2016) Synthesis, charac-
ticles for 5-fluorouracil delivery to glioma terization, applications, and challenges of iron
cancer cells. Mol Pharm 11:859–871 oxide nanoparticles. Nanotechnol Sci Appl
11. Iacobazzi RM, Porcelli L, Lopedota AA, 9:49–67
Laquintana V, Lopalco A, Cutrignelli A, 24. Valizadeh A, Mikaeili H, Samiei M, Farkhani
Altamura E, Di Fonte R, Azzariti A, Franco SM, Zarghami N, Kouhi M, Akbarzadeh A,
M, Denora N (2017) Targeting human Davaran S (2012) Quantum dots: synthesis,
liver cancer cells with lactobionic acid-G bioapplications, and toxicity. Nanoscale Res
(4)-PAMAM-FITC sorafenib loaded den- Lett 7:480
drimers. Int J Pharm 528:485–497 25. Lu Y, Li Y, Wu W (2016) Injected nanocrys-
12. Vauthier C, Bouchemal K (2009) Methods for tals for targeted drug delivery. Acta Pharm Sin
the preparation and manufacture of polymeric B 6:106–113
nanoparticles. Pharm Res 26:1025–1058 26. Ibanez M, Cabot A (2013) All change for
13. Sharifi S, Behzadi S, Laurent S, Forrest ML, nanocrystals. Science 340:935–936
Stroeve P, Mahmoudi M (2012) Toxicity of 27. Zhao X, Zu Y, Li Q, Wang M, Zu B, Zhang
nanomaterials. Chem Soc Rev 41:2323–2343 X, Jiang R, Zu C (2010) Preparation and
14. Cauerhff A, Martinez YN, Islan GA, Castro characterization of camptothecin powder
GR (2014) Nanostability. Nanotoxicology. micronized by a supercritical antisolvent
Springer, New York, NY, pp 57–95 (SAS) process. J Supercrit Fluids 51:412–419
15. Tricarico D, Maqoud F, Curci A, Camerino G, 28. Leroux JC, Allemann E, Doelker E, Gurny R
Zizzo N, Denora N, Cutrignelli A, Laquintana (1995) New approach for the preparation of
V, Lopalco A, la Forgia F, Fontana S, Franco nanoparticles by an emulsification-diffusion
M, Lopedota A (2018) Characterization of method. Eur J Pharm Biopharm 41:14–18
minoxidil/hydroxypropyl-β-cyclodextrin 29. Watnasirichaikul S, Rades T, Tucker IG,
inclusion complex in aqueous alginate man- Davies NM (2002) In-vitro release and
agement: efficacy evaluation in male rat. Eur oral bioactivity of insulin in diabetic rats
J Pharm Biopharm 122:146–157 using nanocapsules dispersed in biocom-
16. Lopedota A, Denora N, Laquintana V, patible microemulsion. J Pharm Pharmacol
Cutrignelli A, Lopalco A, Tricarico D, 54:473–480
Maqoud F, Curci A, Mastrodonato M, 30. Legrand P, Lesieur S, Bochot A, Gref R,
la Forgia F, Fontana S, Franco M (2018) Raatjes W, Barratt G, Vauthier C (2007)
Alginate-based hydrogel containing min- Influence of polymer behaviour in organic
o x i d i l / h y d r o x y p r o p y l - β- c y c l o d e x t r i n solution on the production of polylactide
inclusion complex for topical alopecia treat- nanoparticles by nanoprecipitation. Int
ment. J Pharm Sci 107:1046. https://doi. J Pharm 344:33–43
org/10.1016/j.xphs.2017.11.016 31. Murakami H, Kobayashi M, Takeuchi H,
17. Lopedota A, Cutrignelli A, Denora N, Kawashima Y (1999) Preparation of poly
Laquintana V, Lopalco A, Selva S, Ragni L, (DL-lactide-co-glycolide) nanoparticles by
Tongiani S, Franco M (2015) New ethanol modified spontaneous emulsification solvent
and propylene glycol free gel formulations diffusion method. Int J Pharm 187:143–152
containing a minoxidil-methyl-β-cyclodextrin 32. Allemann E, Doelker E, Gurny R (1993)
complex as promising tools for alopecia treat- Drug loaded poly (lactic acid) nanoparticles
ment. Drug Dev Ind Pharm 41:728–736 produced by a reversible salting-out process:
18. Holister P, Vas CR, Harper T (2003) purification of an injectable dosage form. Eur
Dendrimers. Technol White Papers 6:1–15 J Pharm Biopharm 39:13–18
19. Franc G, Kakkar AK (2009) Diels-Alder click 33. Dreis S, Rothweiler F, Michaelis M, Cinatl
chemistry in designing dendritic macromol- J Jr, Kreuter J, Langer K (2007) Preparation,
ecules. Chem A Eur J 15:5630–5639 characterisation and maintenance of drug
20. Osawa E (2012) Perspectives of fullerene nan- efficacy of doxorubicin-loaded human serum
otechnology. Kluwer Academic Publishers, albumin (HSA) nanoparticles. Int J Pharm
New York, NY 341:207–214
21. Mojica M, Alonso JA, Mendez F (2013) 34. Rollot JM, Couvreur P, Roblot-Treupel L,
Synthesis of fullerenes. J Phys Org Chem Puisieux F (1986) Physicochemical and mor-
26:526–539 phological characterization of polyisobutyl
22. Joseph A, Mathew S (2014) Ferrofluids: cyanoacrylate nanocapsules. J Pharm Sci
synthetic strategies, stabilization, physico- 75:361–364
35. Masson V, Maurin F, Fessi H, Devissaguet epithelium using polylactide nanoparticles.

JP (1997) Influence of sterilization pro- Invest Ophthalmol Vis Sci 44:3562–3569
cesses on poly (ε-caprolactone) nanospheres. 48. Ballou B, Lagerholm BC, Ernst LA, Bruchez
Biomaterials 18:327–335 MP, Waggoner AS (2004) Noninvasive imag-
36. Bos GW, Trullas-Jimeno A, Jiskoot W, ing of quantum dots in mice. Bioconjug
Crommelin DJA, Hennink WE (2000) Chem 15:79–86
Sterilization of poly (dimethylamino) ethyl 49. Weissenbock A, Wirth M, Gabor F (2004)
methacrylate-based gene transfer complexes. WGA-grafted PLGA-nanospheres: prepara-
Int J Pharm 211:79–88 tion and association with Caco-2 single cells.
37. Athanasiou KA, Niederauer GG, Agrawal CM J Control Release 99:383–392
(1996) Sterilization, toxicity, biocompatibil- 50. Olivier JC, Fenart L, Chauvet R, Pariat C,
ity and clinical applications of polylactic acid/ Cecchelli R, Couet W (1999) Indirect evi-
polyglycolic acid copolymers. Biomaterials dence that drug brain targeting using poly-
17:93–102 sorbate 80-coated polybutylcyanoacrylate
38. Memisoglu-Bilensoy E, Hincal AA (2006) nanoparticles is related to toxicity. Pharm Res
Sterile, injectable cyclodextrin nanoparticles: 16:1836–1842
effects of gamma irradiation and autoclaving. 51. Lockman PR, Koziara JM, Mumper RJ, Allen
Int J Pharm 311:203–208 DD (2004) Nanoparticle surface charges alter
39. Magenheim B, Benita S (1991) Nanoparticle blood-brain barrier integrity and permeability.
characterization: a comprehensive physi- J Drug Target 12:635–641
cochemical approach. STP Pharma Sci 52. Kreuter J (2001) Nanoparticulate systems for
1:221–241 brain delivery of drugs. Adv Drug Deliv Rev
40. Abdelwahed W, Degobert G, Stainmesse S, 47:65–81
Fessi H (2006) Freeze-drying of nanoparti- 53. Michaelis K, Hoffmann MM, Dreis S,
cles: formulation, process and storage consid- Herbert E, Alyautdin RN, Michaelis M,
erations. Adv Drug Deliv Rev 58:1688–1713 Kreuter J, Langer K (2006) Covalent linkage
41. Abdelwahed W, Degobert G, Fessi H (2006) of apolipoprotein e to albumin nanoparticles
Investigation of nanocapsules stabilization by strongly enhances drug transport into the
amorphous excipients during freeze-drying brain. J Pharmacol Exp Ther 317:1246–1253
and storage. Eur J Pharm Biopharm 63:87–94 54. Alyautdin RN, Petrov VE, Langer K, Berthold
42. Tewa-Tagne P, Briancon S, Fessi H (2006) A, Kharkevich DA, Kreuter J (1997) Delivery
Spray-dried microparticles containing poly- of loperamide across the blood-brain barrier
meric nanocapsules: formulation aspects, with polysorbate 80-coated polybutylcyanoac-
liquid phase interactions and particles charac- rylate nanoparticles. Pharm Res 14:325–328
teristics. Int J Pharm 325:63–74 55. Koziara JM, Lockman PR, Allen DD,
43. Edetsberger M, Gaubitzer E, Valic E, Mumper RJ (2004) Paclitaxel nanoparticles
Waigmann E, Kohler G (2005) Detection of for the potential treatment of brain tumors.
nanometer-sized particles in living cells using J Control Release 99:259–269
modern fluorescence fluctuation methods. 56. Oberdörster G, Sharp Z, Atudorei V, Elder
Biochem Biophys Res Commun 332:109–116 A, Gelein R, Kreyling W, Cox C (2004)
44. Shenoy D, Fu W, Li J, Castro C, Jones G, Translocation of inhaled ultrafine particles to
DiMarzio C, Sridhar S, Amiji M (2006) the brain. Inhal Toxicol 16:437–445
Surface functionalization of gold nanopar- 57. Elder A, Gelein R, Silva V, Feikert T,
ticles using hetero-bifunctional poly (ethylene Opanashuk L, Carter L, Potter R, Maynard
glycol) spacer for intracellular tracking and A, Ito Y, Finkelstein J, Oberdörster G (2006)
delivery. Int J Nanomedicine 1:51–57 Translocation of inhaled ultrafine manganese
45. Panyam J, Zhou WZ, Prabha S, Sahoo SK, oxide particles to the central nervous system.
Labhasetwar V (2002) Rapid endo-lysosomal Environ Health Perspect 114:1172–1178
escape of poly (DL-lactide-co-glycolide) 58. Zolnik BS, Gonzalez-Fernandez A, Sadrieh
nanoparticles: implications for drug and gene N, Dobrovolskaia MA (2010) Minireview:
delivery. FASEB J 16:1217–1226 nanoparticles and the immune system.
46. Akerman ME, Chan WC, Laakkonen P, Bhatia Endocrinology 151:458–465
SN (2002) Nanocrystal targeting in vivo. 59. Lenaerts V, Nagelkerke JF, Van Berkel TJC,
Proc Natl Acad Sci U S A 99:12617–12621 Couvrer P, Grislain L, Roland M, Speiser P
47. Bourges JL, Gautier SE, Delie F, Bejjani (1984) In vivo uptake of polyisobutyl cya-
RA, Jeanny JC, Gurny R, BenEzra D, noacrylate nanoparticles by rat liver Kupffer,
Behar-Cohen FF (2003) Ocular drug deliv- endothelial, and parenchymal cells. J Pharm
ery targeting the retina and retinal pigment Sci 73:980–982
60. Gibaud S, Demoy M, Andreux JP, 72. Yanamala N, Kagan VE, Shvedova AA (2013)
Weingarten C, Gouritin B, Couvrer P (1996) Molecular modeling in structural nano-
Cells involved in the capture of nanopar- toxicology: interactions of nano-particles with
ticles in hematopoietic organs. J Pharm Sci nano-machinery of cells. Adv Drug Deliv Rev
85:944–950 65:2070–2077
61. Demoy M, Gibaud S, Andreux JP, Weingarten 73. Aragao-Santiago L, Hillaireau H, Grabowski
C, Gouritin B, Couvrer P (1997) Splenic N, Mura S, Nascimento TL, Dufort S, Coll
trapping of nanoparticles: complementary JL, Tsapis N, Fattal E (2016) Compared
approaches for in situ studies. Pharm Res in vivo toxicity in mice of lung delivered bio-
14:463–468 degradable and non-biodegradable nanopar-
62. Silva AL, Peres C, Conniot J, Matos AI, ticles. Nanotoxicology 10:292–302
Loura L, Carriera B, Sainz V, Scomparin 74. Semete B, Booysen L, Lemmer Y, Kalombo
A, Satchi-Fainaro R, Preat V, Florindo HF L, Katana L, Verschoor J, Swai HS (2010)
(2017) Nanoparticle impact on innate In vivo evaluation of the biodistribution and
immune cell pattern-recognition recep- safety of PLGA nanoparticles as drug delivery
tors and inflammasomes activation. Semin systems. Nanomedicine 6:662–671
Immunol 34:3–24 75. Panyam J, Dali MM, Sahoo SK, Ma W,
63. Sayes CM, Reed KL, Warheit DB (2007) Chakravarthi SS, Amidon GL, Levy RJ,
Assessing toxicity of fine and nanoparticles: Labhasetwar V (2003) Polymer degradation
comparing in vitro measurements to in vivo and in vitro release of a model protein from
pulmonary toxicity profiles. Toxicol Sci poly (D, L-lactide-co-glycolide) nano-and
97:163–180 microparticles. J Control Release 92:173–187
64. Ryman-Rasmussen JP, Riviere JE, Monteiro- 76. Song Y, Li X, Du X (2009) Exposure to
Riviere NA (2007) Variables influencing nanoparticles is related to pleural effusion,
interactions of untargeted quantum dot pulmonary fibrosis and granuloma. Eur
nanoparticles with skin cells and identifica- Respir J 34:559–567
tion of biochemical modulators. Nano Lett 77. Kreuter J (2007) Nanoparticles- a historical
7:1344–1348 perspective. Int J Pharm 331:1–10
65. Oberdörster G, Maynard A, Donaldson 78. Lherm C, Muller RH, Puisieux F, Couvreur
K, Castranova V, Fitzpatrick J, Ausman K, P (1992) Alkylcyanoacrylate drug carriers:
Carter J, Karn B, Kreyling W, Lai D, Olin II. Cytotoxicity of cyanoacrylate nanopar-
S, Monteiro-Riviere N, Warheit D, Yang ticles with different alkyl chain length. Int
H (2005) Principles for characterizing the J Pharm 84:13–22
potential human health effects from expo- 79. Lee CC, MacKay JA, Frechet JM, Szoka FC
sure to nanomaterials: elements of a screening (2005) Designing dendrimers for biological
strategy. Part Fibre Toxicol 2:8–35 applications. Nat Biotechnol 23:1517–1526
66. Keelan JA, Leong JW, Ho D, Iyer KS (2015) 80. Balogh L, Swanson DR, Tomalia DA,
Therapeutic and safety considerations of Hagnauer GL, McManus AT (2001)
nanoparticle-mediated drug delivery in preg- Dendrimer- silver complexes and
nancy. Nanomedicine 10:2229–2247 nanocomposites as antimicrobial agents.

67. Calvaresi M, Zerbetto F (2011) Fullerene Nano Lett 1:18–21
sorting proteins. Nanoscale 3:2873–2881 81. Kolosnjaj J, Szwarc H, Moussa F (2007)
68. Hansch C (1969) Quantitative approach to Toxicity studies of fullerenes and derivatives.
biochemical structure-activity relationships. Bio-applications of nanoparticles. Springer,
Acc Chem Res 2:232–239 New York, NY, pp 168–180
69. Puzyn T, Rasulev B, Gajewicz A, Hu X, Dasari 82. Yamakoshi Y, Umezawa N, Ryu A, Arakane
TD, Michalkova A, Hwang HM, Toropov A, K, Miyata N, Goda Y, Masumizu T, Nagano
Leszczynska D, Leszczynski J (2011) Using T (2003) Active oxygen species generated
nano-QSAR to predict the cytotoxicity of from photoexcited fullerene (C60) as poten-
metal oxide nanoparticles. Nat Nanotechnol tial medicines: O2-• versus 1O2. J Am Chem
6:175–178 Soc 125:12803–12809
70. Puzyn T, Leszczynska D, Leszczynski 83. Lovern SB, Klaper R (2006) Daphnia magna
J (2009) Toward the development of “nano- mortality when exposed to titanium dioxide
QSARs”: advances and challenges. Small and fullerene (C60) nanoparticles. Environ
5:2494–2509 Toxicol Chem 25:1132–1137
71. Sayes C, Ivanov I (2010) Comparative study 84. Zhu S, Oberdörster E, Haasch ML (2006)
of predictive computational models for Toxicity of an engineered nanoparticle (fuller-
nanoparticle-induced cytotoxicity. Risk Anal ene, C 60) in two aquatic species, Daphnia
30:1723–1734 and fathead minnow. Mar Environ Res
62:S5–S9
85. Oberdörster E (2004) Manufactured nano- 96. Connor EE, Mwamuka J, Gole A, Murphy
materials (fullerene, C60) induce oxidative CJ (2005) Gold nanoparticles are taken up by
stress in the brain of juvenile largemouth bass. human cells but do not cause acute cytotoxic-
Environ Health Perspect 112:1058–1062 ity. Small 1:325–327
86. Shvedova A, Castranova V, Kisin E, 97. Niidome T, Yamagata M, Okamoto Y,
Schwegler-Berry D, Murray A, Gandelsman Akiyama Y, Takahashi H, Kawano T, Katayama
V, Maynard A, Baron P (2003) Exposure Y (2006) PEG-modified gold nanorods with
to carbon nanotube material: assessment of a stealth character for in vivo applications.
nanotube cytotoxicity using human kerati- J Control Release 114:343–347
nocyte cells. J Toxicol Environ Health Part A 98. Su CH, Sheu HS, Lin CY, Huang CC, Lo
66:1909–1926 YW, Pu YC, Weng JC, Shien JH, Chen JH,
87. Radomski A, Jurasz P, Alonso-Escolano D, Yeh CS (2007) Nanoshell magnetic resonance
Drews M, Morandi M, Malinski T, Radomski imaging contrast agents. J Am Chem Soc
MW (2005) Nanoparticle-induced plate- 129:2139–2146
let aggregation and vascular thrombosis. Br 99. Bernardi RJ, Lowery AR, Thompson
J Pharmacol 146:882–893 PA, Blaney SM, West JL (2008)
88. Warheit DB, Laurence BR, Reed KL, Roach Immunonanoshells for targeted photother-
DH, Reynolds GAM, Webb TR (2004) mal ablation in medulloblastoma and glioma:
Comparative pulmonary toxicity assess- an in vitro evaluation using human cell lines.
ment of single-wall carbon nanotubes in rats. J Neurooncol 86:165–172
Toxicol Sci 77:117–125 100. Lowery AR, Gobin AM, Day ES, Halas NJ,
89. Lam CW, James JT, McCluskey R, Hunter West JL (2006) Immunonanoshells for tar-
RL (2004) Pulmonary toxicity of single-wall geted photothermal ablation of tumor cells.
carbon nanotubes in mice 7 and 90 days Int J Nanomedicine 1:149–154
after intra-tracheal instillation. Toxicol Sci 101. O’Neal DP, Hirsch LR, Halas NJ, Payne JD,
77:126–134 West JL (2004) Photo-thermal tumor abla-
90. Hardman R (2006) A toxicologic review of tion in mice using near infrared-absorbing
quantum dots: toxicity depends on physico- nanoparticles. Cancer Lett 209:171–176
chemical and environmental factors. Environ 102. Lin W, Huang YW, Zhou XD, Ma Y (2006) In
Health Perspect 114:165–172 vitro toxicity of silica nanoparticles in human
91. Hoshino A, Fujioka K, Oku T, Suga M, Sasaki lung cancer cells. Toxicol Appl Pharmacol
YF, Ohta T, Yasuhara M, Suzuki K, Yamamoto 217:252–259
K (2004) Physicochemical properties and cel- 103. Chang JS, Chang KLB, Hwang DF, Kong ZL
lular toxicity of nanocrystal quantum dots (2007) In vitro cytotoxicity of silica nanopar-
depend on their surface modification. Nano ticles at high concentrations strongly depends
Lett 4:2163–2169 on the metabolic activity type of the cell line.
92. Hoshino A, Manabe N, Fujioka K, Suzuki K, Environ Sci Technol 41:2064–2068
Yasuhara M, Yamamoto K (2007) Use of flu- 104. Jin Y, Kannan S, Wu M, Zhao JX (2007)

orescent quantum dot bioconjugates for cel- Toxicity of luminescent silica nanoparticles to
lular imaging of immune cells, cell organelle living cells. Chem Res Toxicol 20:1126–1133
labeling, and nanomedicine: surface modifica- 105. Kumar MNV, Sameti M, Mohapatra SS, Kong
tion regulates biological function, including X, Lockey RF, Bakowsky U, Lindenblatt G,
cytotoxicity. J Artif Organs 10:149–157 Schmidt CH, Lehr CM (2004) Cationic
93. Choi AO, Cho SJ, Desbarats J, Lovric J, silica nanoparticles as gene carriers: synthesis,
Maysinger D (2007) Quantum dot-induced characterization and transfection efficiency
cell death involves Fas upregulation and lipid in vitro and in vivo. J Nanosci Nanotechnol
peroxidation in human neuroblastoma cells. 4:876–881
J Nanobiotechnol 5:1 106. Howard DH (1960) Effect of mycostatin

94. Lovric J, Cho SJ, Winnik FM, Maysinger D and fungizone on the growth of Histoplasma
(2005) Unmodified cadmium telluride quan- capsulatum in tissue culture. J Bacteriol
tum dots induce reactive oxygen species for- 79:442–449
mation leading to multiple organelle damage 107. Perlman D, Giuffre NA, Brindle SA (1961)
and cell death. Chem Biol 12:1227–1234 Use of Fungizone® in control of fungi and
95. Derfus AM, Chan WC, Bhatia SN (2004) yeasts in tissue culture. Proc Soc Exp Biol
Probing the cytotoxicity of semiconductor Med 106:880–883
quantum dots. Nano Lett 4:11–18
Chapter 18
Toxicity Potential of Nutraceuticals

Ramesh C. Gupta, Ajay Srivastava, and Rajiv Lall
Abstract
By the turn of the twenty-first century, the use of nutraceuticals became increasingly popular in both
humans and animals due to their easy access, cost-effectiveness, and tolerability with a wide margin of
safety. While some nutraceuticals are safe, others have a toxic potential. For a large number of nutraceuti-
cals, no toxicity/safety data are available due to a lack of pharmacological/toxicological studies. The safety
of some nutraceuticals can be compromised via contamination with toxic plants, metals, mycotoxins, pes-
ticides, fertilizers, drugs of abuse, etc. Knowledge of pharmacokinetic/toxicokinetic studies appears to
play a pivotal role in safety and toxicity assessment of nutraceuticals. Interaction studies are essential to
determine efficacy, safety, and toxicity when nutraceuticals and therapeutic drugs are used concomitantly.
This chapter describes various aspects of nutraceuticals, particularly their toxic potential, and the factors
influencing their safety.
Key words Nutraceuticals, Toxicity testing models, Toxic potential, Safe nutraceuticals, Pesticides,
Metals, Mycotoxins, Plant alkaloids, Drugs of abuse
1 Introduction
In 1989, Dr. Stephen DeFelice coined the term “nutraceutical”

from the words nutrition and pharmaceutical and defined it as “a
food or part of a food that provides medical or health benefits,
including the prevention and/or treatment of a disease.” Later, the
term “nutraceutical” has been defined somewhat differently but
with the same meaning in many countries and by different profes-
sional societies [1, 2]. The term “nutraceutical” may be used inter-
changeably with the term “herbal” (Ayurvedic, Chinese, Tibetan,
African, Amazonian, Herbalism), naturopathy, vitamin and mineral
therapy, and homeopathy [3]. The extracts of medicinal plants have
been used for thousands of years around the world to prevent and/
or treat human and animal diseases. By the turn of the twenty-first
century, the use of nutraceuticals became increasingly popular.
Currently, the nutraceutical industry earns more than a $200 billion
per year. In the USA, most herbal supplements are classified as
dietary supplements and are not subject to the regulations and safety
367
368 Ramesh C. Gupta et al.
standards applied to conventional medicines. The only major US

regulation related to nutraceuticals is the 1994 passage by the US
Congress of the Dietary Supplement Health and Education Act.
Based on this act, dietary supplements are classified as foods, not
drugs, allowing them to be sold without proof of safety and effec-
tiveness [4]. In the EU, current regulations require evidence that
herbal medicinal products meet acceptable standards of quality,
safety, and efficacy before a product license can be issued. However,
unlike food, nutraceuticals are not generally recognized as safe, nor
can one assume that they are all safe.
Pharmacological and toxicological evaluation of nutraceuti-
cals, compared to pure synthetic pharmaceuticals, is complex
due to: (1) multiple phytochemicals present in a single plant;
(2) variability in phytochemical constituents due to geography,
soil characteristics, and climate; (3) use of fertilizers and pesti-
cides, (4) diurnal variation during harvesting, (5) stress, and (6)
quality control standards [2]. These factors and a lack of stan-
dard extraction procedures can influence identity, purity, qual-
ity, quantity, composition, potency/strength, and safety of
active component(s), thereby causing a wide variability in prod-
uct effectiveness from batch to batch and from one manufac-
turer to another.
Nutraceuticals, by having antioxidant, anti-inflammatory,
immunomodulatory, sedative, adaptogenic, anticancer, and
other such properties, are used against many ailments, including
arthritis, diabetes, dermatitis, gastrointestinal, cardiovascular,
respiratory, neurodegenerative, reproductive, and metabolic dis-
eases. Even today, 80% of the world’s populations living in devel-
oping countries rely on traditional medicine. Approximately
150 million Americans consume at least one dietary supplement
daily. As the global use of nutraceuticals has increased for humans
and animals, so have health risks emerging from active compo-
nents as well as from toxic contaminants of supplements. Since
these nutraceuticals have not been tested as rigorously as phar-
maceuticals and no large-scale clinical trials have been done,
safety remains a serious issue. Furthermore, since most nutraceu-
ticals are plant extracts, their contamination with other plant
alkaloids (e.g., pyrrolizidine alkaloids), metals (arsenic, cad-
mium, lead, and mercury), pesticides, and mycotoxins is very
likely. Adulteration of nutraceuticals with drugs of abuse remains
another serious concern. Currently, it is a common practice that
those receiving prescription drugs also consume nutraceuticals,
and some nutraceuticals are likely to alter efficacy and safety of
pharmaceutical drugs by modulating their pharmacodynamics
and pharmacokinetics. This chapter describes the toxic potential
of some common nutraceuticals.
Toxicity Potential of Nutraceuticals 369
2 Common Nutraceuticals
Currently, there are hundreds, if not thousands, of nutraceuticals

on the market for use, and a partial list of common nutraceuticals
is provided in Table 1.
Table 1
List of common nutraceuticals or their ingredients
Nutraceutical/ingredient Nutraceutical/ingredient
Acacia catechu Hawthorn (Crataegus oxycantha)
Acacia senegal Isoflavones
Aloe vera Kava (Piper methysticum)
Andrographis paniculata Krill oil (Euphausia superba)
Anthocyanins and anthocyanidin Lychee (Litchi chinensis Sonn)
Arginine and citrulline Maqui Berry (Aristotelia chilensis)
Ashwagandha, Indian ginseng (Withania somnifera) Melatonin
Astaxanthin (Haematococcus pluvialis) Methylsulfonylmethane (MSM)
Bee pollen Milk thistle (Silybum marianum)
Berberine Momordica charantia
Black kohosh (Actaea racemosa) Monsonia angustifolia
Boswellia serrata Mulberry branch bark powder (Ramulus
mori)
Caffeine Neem (Azadirachta indica) extract/oil
Cannabis sativa Olive leaf (Olea europaea L.)
Caralluma fimbriata Omega-3 fatty acids
Capsicum spp. Panax ginseng (Chinese ginseng)
Cassia fistula Pequi oil (Caryocar brasiliense, Camb)
Cassia occidentalis Perilla frutescens
Chondroitin sulfate Perna canaliculus (Green-lipped mussel)
Chromium (trivalent) polynicotinate Phyllanthus emblica (Amla extract)
Elderberry (Sambucus nigra) Quercetin
Ephedra sinica/Ephedra equisetina/Ma Huang Red marine alga (Alsidium corallinum)
Eremophila maculata Resveratrol
Eyebright Royal jelly proteins
Fenugreek (Trigonella foenum-graecum) Sarcandra glabra
Garcinia cambogia Shilajit
(continued)
Table 1
(continued)
Nutraceutical/ingredient Nutraceutical/ingredient
Garlic and onion (Allium spp.) Spirulina (Arthrospira platensis)
Ginger (Zingiber officinale) St. John’s wort (Hypericum perforatum)
Ginkgo (Ginkgo biloba) Terminalia chebula extract
Glucosamine Terminalia muelleri Benth
Glucosinolates (Brassica napus) Thymoquinone (Nigella sativa)
Goldenseal (Hydrastis canadensis) Tibouchina granulosa
Grape seed extract Turmeric/curcumin (Curcuma longa)
Green coffee bean (Coffea arabica; Coffeea Type II collagen
canephora)
Green tea (Camellia sinensis) extract Yucca glauca (Yucca)
Guava (Psidium guajava) Valerian (Valeriana officinalis)
Guarana (Paullinia cupana) Wild garlic (Allium ursinum)
3 Nutraceuticals with a Toxic Potential
A significant number of nutraceuticals have a toxic potential,

including Aloe vera, St. John’s wort, green tea extract, bitter
melon, goldenseal, Ma Huang, Ginkgo biloba, Garcinia cambogia,
and others. Table 2 provides a list of some nutraceuticals, their
toxic constituents, and their major toxic effects. A brief description
of some nutraceuticals that have a toxic potential is given below.
3.1 Goldenseal Goldenseal (Hydrastis canadensis), also called orangeroot or yel-

low puccoon, is a perennial herb native to Canada and the east-
ern USA. The plant contains isoquinoline alkaloids (hydrastine,
berberine, canadine, canalidine, and many others). Goldenseal is
used both internally and externally for treating skin, mouth, and
eye infections and for the treatment of gastrointestinal and uri-
nary diseases [5]. Side effects may include mouth ulcers, diges-
tive upset (constipation or diarrhea), nausea, vomiting,
nervousness, CNS depression, tachycardia, and seizures. High
doses may cause breathing problems, paralysis, and even death.
Long-term use may lead to vitamin B deficiency, hallucinations,
and delirium. Toxicity of goldenseal appears to be due to berber-
ine and hydrastine.
A standard NTP 2-year toxicology and carcinogenesis study in
rodents showed that oral administration of goldenseal root powder
resulted in increased incidence of hepatocellular adenoma and/or
carcinoma in F344/N rats and male B6C3F1 mice [5]. California

is proposing to list goldenseal root powder as a carcinogen.
Goldenseal has been reported to inhibit cytochrome P450
(CYP2D6, CYP3A4, and CYP3A5) activity by ~40%, which is a
statistically and clinically significant reduction [6]. CYP2D6
specifically is a known metabolizer of many commonly used
pharmaceuticals, such as antidepressants (including selective
serotonin reuptake inhibitors, except for fluvoxamine), neuro-
leptics, and codeine. Combining goldenseal with such medica-
tions should be done with caution and under the supervision of
a physician as it can lead to serious—perhaps fatal—toxicity.
Individuals with a genetic deficiency in these enzymes are at
particular risk [7].
3.2 Ginkgo biloba The Ginkgo biloba, commonly referred to as “living fossil,” is a
hardy tree native to Asia, Europe, North America, Argentina, and
New Zealand. The main constituents in G. biloba are terpenoids
(ginkgolides A, B, and C, and bilobalide), the flavonoids (querce-
tin and its metabolites kaempferol and isorhamnetin), and alkyl
phenols (ginkgolic acid). These phytochemicals possess antioxida-
tive, anti-inflammatory and anticancer activities. G. biloba has been
used for medicinal purposes, including brain disorders, circulatory
disorders, respiratory disorders such as asthma, urinary tract dis-
eases, ear, nose, and throat problems, and as an antiparasitic.
Currently, G. biloba is one of the most commonly used herbal
medicines in both the USA and Europe [8]. G. biloba extract is
regulated as a dietary supplement in the USA, but in Germany and
France it is regulated as a prescription drug.
The long-term use of G. biloba extract has been associated with
spontaneous bleeding (including intracranial, hyphema, and retinal
hemorrhage) in humans [9]. Serious side effects, such as bleeding,
hematoma, hyphema, apraxia, neurological deficits, and even
death, have been reported in humans from the concurrent use of
G. biloba extract and anesthetics, analgesics, anticoagulant or anti-
platelet agents [10–12]. The ginkgotoxins from the seeds of G.
biloba are thought to be responsible for epileptiform seizures,
unconsciousness, and paralysis of the legs [13]. In a number of
toxicological studies, G. biloba components have shown potential
for mutagenicity and carcinogenicity [7, 14].
3.3 Green Tea Green tea (Camellia sinensis) infusions and extracts are used as a
Extract beverage, nutraceutical, and phytopharmaceutical around the world.
Green tea comprises more than 20% of the tea products sold and has
the highest annual increase in consumption [15]. There are more
than 2000 different chemical substances (including many flavan-
3-ols, flavones, and flavonols) associated with green tea. Green tea
chemicals have antioxidant, anti-inflammatory, immunomodulatory,
and anticancer properties. Green tea extract (GTE) is claimed to
Table 2
List of nutraceuticals/dietary supplements/ingredients, and their toxic effects
Plant/nutraceutical Toxic principle(s) Major toxic effect(s) References

Ackee (Blighia Hypoglycin A, hypoglycin B GI and CNS toxicity [135]
sapida)
Aconite alkaloids Aconite/aconitine Ventricular dysrhythmia/ [138]
tachycardia
Allium spp. Organosulfoxides, n-propyl Hemolytic anemia, Heinz [43, 139]
disulfide, allicin body formation,
hemosiderosis, icterus,
tachycardia, tachypnea
Aloe vera Galacturonic acid, Hypoglycemia, genotoxicity, [7, 43, 52]
anthraquinones, aloin, mutagenicity
aloe-emodin, danthron
Aristolochia spp. Aristolochic acid I, and Nephrotoxicity, urothelial [67, 140]
Aristolochic acid II toxicity, urinary tract
carcinomas
Artemisia spp. and Monoterpene ketones Neurotoxicity [141]
Salvia spp. (α-thujone and β-thujone)
Ayahuasca Β-carbolines (harmine, Hallucinations, tachycardia, [142]
(Banisteriopsis tetrahydroharmine, and hypertension, mydriasis,
caapi and harmaline), and DMT and vomiting
Psychotria viridis)
Bitter melon/Karela Momordicins, vicine, charantin Hypoglycemia, miscarriage [43, 45–47]
(Momordica in pregnant women
charantia)
Black kohosh Cimigenol, formononetin Metabolic disturbance, liver [6]
(Cimicifuga toxicity, vertigo, seizures
racemosa)
Blue-green algae Microcystins Hepatotoxicity [143, 144]
Chinese cinnamon Cinnamaldehyde Hepatotoxicity [145]
(Cassia cinnamon)
Coffee Caffeine Anxiety, tremor, tachycardia, [23, 25–27,
diuresis, respiratory failure, 146]
reproductive and
developmental disorders,
death
Comfrey (Symphytum Pyrrolizidine alkaloids Hepatotoxicity [74]
officinale)
Dioscorea bulbifera L. Furanoids: diosbulbin B, and Hepatotoxicity [147]
8-epidiosbulbin E acetate
(continued)
Table 2
(continued)

DMAA 1,3-Dimethylamyl-amine Cardiotoxicity [148]
English hawthorn Vitexin, isovitexin, Heart palpitations, [149]
(Crataegus vitexin-4-O-glucoside mutagenicity, genotoxicity
oxyacantha)
Ephedra Ephedrine alkaloids Cardiotoxicity, nephrotoxicity [53, 57, 58,
137]
Garcinia cambogia (−)-hydroxycitric acid Cardiotoxicity, hepatotoxicity, [30–35,
nephrotoxicity, 150]
rhabdomyolysis,
reproductive toxicity, death
Ginkgo biloba Ginkgotoxins, ginkgolides, Spontaneous bleeding, [7, 8, 12,
ginkgolic acid hyphema, genotoxicity, 14, 137]
seizures, paralysis, death
Goldenseal Berberine, hydrastine, canadine, Tachycardia, seizures, [5–7]
(Hydrastis canalidine carcinogenesis, paralysis,
canadensis) metabolic disturbance, death
Grapefruit/juice Furocoumarins (bergamottin, Interaction with CytP450, [117, 151]
(Citrus paradisi epoxybargamottin, bergaptol, Metabolic disturbance
Macf) and 6′,7′-DHB)
Green tea extract (−)-epigallocatechin-3-gallate Hepatotoxicity, nephrotoxicity [15, 17, 19,
(Camellia sinensis) reproductive toxicity 152]
Kava (Piper Kavalactones, flavokawain B, Hepatotoxicity [6, 37, 153]
methysticum) pipermethysticin, schaftoside
Kratom (Mitragyna Mitragynine, Neurotoxicity, addiction [154]
speciosa) 7-hydroxymitragynine
Lichen (Usnea spp.) Usnic acid Hepatotoxicity, cardiotoxicity [155–157]
Lychee/Litchi Hypoglycin A, Encephalopathy, impaired fat [133, 135]
(Litchi chinensis) methylenecyclopropyl-glycine metabolism, hypoglycemia
(MCPG)
Pennyroyal oil Pulegone, menthofuran Hepatotoxicity, GI bleeding, [61, 62,
(Mentha pulegium) renal failure, coma, death 158]
Pequi pulp oil Not identified Hematological abnormalities [64]
(Caryocar
brasiliense)
Radix sophora Cytisine, matrine, oxymatrine Hepatotoxicity [77]
tonkinensis
Raspberry (Rubus Ellagitannins Genotoxicity [159]
idaeus)
(continued)
Table 2
(continued)

Saw palmetto Β-sitosterol Bleeding, impotence in men, [160]
(Serenoa repens) heart attack
St. John’s wort Hypericin, hyperforin, Seizures, hepatocellular [38, 39,
(Hypericum naphthodianthrones, carcinoma, bone marrow 41–43,
perforatum) pseudohypericin, necrosis 117]
sesquiterpenes
Valerian (Valeriana Valerinic acid Metabolic disturbance, [6, 161]
officinalis) hepatotoxicity, CNS
depression
have beneficial effects in genital warts, obesity, diabetes, inflamma-

tory bowel syndrome, cardiovascular diseases, and cancers.
Kapetanovic et al. [16] carried out a chronic toxicity study in
fasted and nonfasted 5- to 6-month-old Beagle dogs with a modi-
fied GTE (64% (−)-epigallocatechin gallate, 3.4% (−)-epigallocat-
echin, 7% (−)-epicatechin gallate, and 0–1% (−)-epicatechin).
Dogs were given GTE orally at 0, 200, 500, and 1000 mg/kg/
day. The study was terminated at 6.5 months and there were 16
deaths in 24 dogs. The majority of deaths occurred in the first
13 weeks of the study. Male dogs were found to be more sensitive
than female dogs; and fasted dogs were more sensitive than non-
fasted dogs. The liver lesions were described as centrolobular
necrosis and chronic inflammation. Renal lesions were tubular epi-
thelial necrosis, tubular dilation with granular or hyaline protein
casts, epithelial regeneration, acute inflammatory exudate, and
transitional hyperplasia. Epithelial necrosis was observed in the tes-
ticles along with aspermia. Additional histopathology observed
was atrophy of the prostate gland in males and ovarian atrophy in
females. Studies in mice have shown that a single oral dose
(1500 mg/kg) of (−)-epigallocatechin gallate administered to
fasted CF-1 mice caused moderate to severe liver necrosis [17].
Oxidative stress was measured by a fivefold increase in lipid peroxi-
dation, 9.5-fold increase in 8-isoprostane, and increased hepatic
metallothioneine and c-histone 2AX protein expression.
A subchronic study involving rats using green tea infusion
showed adverse changes in the reproductive organs [18]. Rats
received 1 mL extract (1.5%, 2.5%, and 5% solids)/kg body wt/day
for 26 days. The epididymal sperm counts in rats were significantly
decreased in the 2.5% and 5% groups. Serum luteinizing hormone
and testosterone were significantly decreased in a dose-dependent
manner. Spermatogenesis was reduced in the 2.5% and 5% groups.
Side effects of GTE in humans include bloating, nausea, heartburn,
abdominal pain, dizziness, headache, muscle pain, and hepatotoxicity

[19]. Evidence also suggests that increased consumption of green tea
infusions may increase the risk for breast cancer [20]. It needs to be
pointed out that consumption of green tea and GTEs with meals may
reduce the absorption of iron and folate. Green tea contaminants,
such as polycyclic aromatic hydrocarbons and other persistent organic
pollutants, may further exacerbate the toxicity of GTE [21, 22].
3.4 Green Coffee Green coffee bean (GCB) refers to unroasted mature or immature
Bean/Caffeine coffee beans from Coffea arabica and Coffea canephora. Although
coffee bean has a mixture of more than 1000 phytochemicals, caf-
feine is the most significant from a nutraceutical perspective [23].
Caffeine is present in a wide range of products, i.e., coffee (40–
180 mg/150 mL), tea (24–50 mg/150 mL), soft drinks (15–
29 mg/180 mL), cocoa beverages (2–7 mg/150 mL), chocolate
bars (1–36 mg/28 g), and maté (~160 mg/L) [24]. By having
antioxidative, anti-inflammatory, and immunomodulatory proper-
ties, GCB phytochemicals have health claims in cardiovascular,
hepatic, pulmonary, and neurodegenerative diseases, type 2 diabe-
tes, obesity, and cancer.
Although caffeine has GRAS status from the US FDA and falls
into the category of functional foods, its risk assessment appears
complex. Toxic concentrations of caffeine are several times higher
than the concentration required to block adenosine receptors. To
inhibit cyclic nucleotide breakdown through phosphodiesterase
inhibition, 20 times higher concentrations of caffeine are required;
to block GABAA, 40 times higher concentrations; and to mobilize
intracellular calcium depots, 100 times higher concentrations [24].
In rats and other laboratory animal species, the LD50 of caffeine is
determined to be 150–265 mg/kg. In humans, ingestion of caf-
feine at 20 mg/kg body wt is considered toxic, and the lethal dose
has been estimated to be in the range of 150–200 mg/kg [25].
Blood concentrations of caffeine in excess of 30 μg/mL are associ-
ated with clinical signs of intoxication, such as anxiety, restlessness,
and tachycardia. The symptoms of acute and chronic caffeine intox-
ication after consumption of high doses (300–800 mg/person/
day) have been described as caffeinism. These symptoms include
dizziness, restlessness, agitation, anxiety, irritability, muscle tremor,
hyperventilation, arrhythmia, tachycardia, hypertension followed
by hypotension, diuresis, and gastrointestinal disorders, renal fail-
ure, respiratory failure, and seizures. Caffeine overdose may cause
hypokalemia, hyponatremia, impaired iron and zinc absorption,
rhabdomyolysis, and finally, circulatory collapse [23, 26, 27].
During pregnancy, caffeine is considered the most consumed
psychostimulant xenobiotic. The UK Government’s Food
Standards Agency, Committee on Toxicity and the US FDA con-
sider that caffeine intake during pregnancy is safe up to 300 mg/
person/day. Recently, Menezes and Da Silva [27] described in
detail the effects of caffeine on human reproduction (such as

delayed conception, dystocia, and spontaneous abortion or pre-
term birth) and fetal development (such as fetal growth restriction
and neural tube defects). Women who consume eight cups of cof-
fee per day can have a 25% reduction of blood flow in the placenta
and are at double the risk of spontaneous abortions compared to
those who do not consume coffee. In some developmental toxicity
studies, caffeine exposure at higher concentrations that may pro-
duce maternal toxicity can cause reduced fetal body weight, delayed
ossification, resorption, and an increased incidence of malforma-
tion of the limbs and palate [27, 28]. It is suggested that caffeine
at 400 mg/day for adults and 200 mg/day for pregnant and lactat-
ing women can be considered safe [29].
3.5 Garcinia Garcinia cambogia is a tropical tree native to Asia and Africa. The
cambogia extract of fruit and fruit rind has been used in the treatment of vari-
ous ailments, such as expulsion of intestinal parasites, dysentery,
rheumatism, as a cardiotonic in angina, and tumors. The fruit rind
is also used in rickets, enlargement of spleen, and for healing of
bone fractures. The fruit extracts from G. cambogia, G. indica, and
G. atroviridis are very rich in hydroxycitric acid (HCA). Other phy-
tochemicals include garcinol, guttiferones (A, E, F, and M), and
xanthones. HCA is a popular component of dietary supplements for
promotion of weight loss. Garcinia/HCA has been found to be
efficacious in ameliorating obesity-related complications, such as
inflammation, oxidative stress, and insulin resistance. G. cambogia
extract has also been shown to have antiulcerative properties.
In a number of in vivo studies (including two generation repro-
ductive and developmental), HCA/HCA-SX has been found to be
very safe [30]. HCA up to a dose of 2800 mg/day in humans has
been regarded as no observed adverse effect level. Rats receiving
ethanolic seed extract of G. cambogia (100 or 200 mg/kg body
wt/day for 6 days a week for 6 weeks) revealed, in a dose-dependent
manner, an increase in the interstitial spaces, degeneration of the
Leydig cells, and distortion in the arrangement of the cells of sper-
matogenic series in histological examination of the testes [31].
Dietary feeding of high doses of G. cambogia (102 or 154 mmol of
HCA/kg diet) to male Zucker obese rats for 90 days caused toxicity
in the testes. The weight of testes was reduced by half, and histo-
pathological changes included marked testicular atrophy and
impairment of spermatogenesis. Kim et al. [32] fed mice G. cambo-
gia (1%) for 16 weeks, and noted hepatic fibrosis, inflammation,
lipid peroxidation, and increased mRNA levels of genes related to
oxidative stress. A number of cases of hepatotoxicity were reported
with the use of Hydroxycut®, an over-the-counter weight loss prod-
uct in the USA [33–35]. Of course, hepatotoxicity might also have
resulted from co-consumption of (−)-epigallocatechin gallate from
green tea extract, acetaminophen, alcohol, anabolic steroids, drugs,
or a combination of these factors [30, 33, 36]. Interestingly, Lopez

et al. [36] suspected serotonin toxicity in the presence of therapeu-
tic dosing of serotonin reuptake inhibitors, when combined with a
nutritional supplement containing G. cambogia and HCA.
3.6 Kava Kava (Piper methysticum), also known as kava kava, is an herbal
shrub that has been used for centuries in the South Pacific as a
social beverage and in traditional ceremonial rituals. For the last
two decades, kava roots and rhizome extracts have been used in
Western industrialized countries for treating mild to moderate anx-
iety, stress, insomnia, restlessness, and muscle fatigue. Several
reports of severe hepatotoxicity in kava consumers led the US FDA
and authorities in Europe to restrict sales of kava-containing prod-
ucts. Kava is still widely used as a dietary supplement to relieve
anxiety and stress in the USA.
Human studies using kava at therapeutic dosages have failed to
demonstrate any toxic effects. Prolonged use of a dose equivalent
to 400 mg or more of kava lactones/day is likely to cause the char-
acteristic skin lesions of kava toxicity (pigmented, dry, covered
with scales) which heals upon discontinuance of the kava extract.
At doses >9 g/day, liver enzymes can elevate and should be moni-
tored. Kavalactones extracted in organic solvents have been pro-
posed to account for kava-induced liver toxicity. In experimental
studies (in vitro and in vivo), Zhou et al. [37] identified flavoka-
wain B (FKB) as a hepatotoxin, and suggested its mechanism of
hepatotoxicity. FKB inhibits IKK activity (directly or indirectly),
leading to downregulation of NF-κB transcriptional activity, which
is crucial for hepatocellular survival. FKB also alters intracellular
redox levels and induces oxidative stress that is likely to be further
enhanced by blockade of NF-κB-mediated transcription and down-
regulation of SOD2. In essence, FKB induces GSH-sensitive oxi-
dative stress and hepatotoxicity through modulation of IKK/
NF-κB and MAPK signaling pathways.
3.7 St. John’s Wort St. John’s wort (Hypericum perforatum) is native to Europe, Asia,
and North Africa. Extracts of this plant are known to have at least
150 compounds. The side effects observed in large clinical trials
conducted over several weeks include minor gastrointestinal irrita-
tion, allergic reactions, tiredness, and restlessness [38]. Seizures and
confusion were reported in a 16-year-old female who consumed
hypericum extract tablets in an apparent suicide attempt [39]. In
some other cases, hypericum extract might have caused autonomic
arousal [40], hepatocellular carcinoma [41], and bone marrow
necrosis [42]. In addition, the plant extracts can upregulate and
downregulate gut and hepatic CYP enzymes and xenobiotic trans-
porters, and by these mechanisms it can alter the pharmacokinetics
and efficacy of concurrent medications, such as theophylline, warfa-
rin, verapamil, digoxin, ibuprofen, methadone, oxycodone, antidia-

betic drugs, and many others [38, 43].
3.8 Bitter Melon The plant bitter melon (Momordica charantia) originated from
India and was introduced in China in the fourteenth century.
Currently, it is being used worldwide as a food and a medicine. The
fruit has several biologically active chemical compounds such as gly-
cosides, saponins, alkaloids, fixed oils, triterpenes, proteins, and ste-
roids. Consumption of bitter melon is capable of lowering blood
glucose levels by increasing hepatic utilization of glucose and
decreasing hepatic glucose output [44], and is thereby used as an
herbal remedy to control diabetes. The fruit has shown not only
potent antihyperglycemic activity in various animal models and clin-
ical trials, but also a dose-related hypoglycemic effect at higher
doses in normoglycemic rats [45]. Therefore, the use of bitter
melon with insulin, sulfonylureas, or meglitinides should be cau-
tioned for theoretical risk of increased hypoglycemic events in the
pharmacotherapeutic management of type 2 diabetes [46]. Other
side effects of bitter melon are diarrhea, abdominal pain, headache,
fever, hypoglycemia, heart atrial fibrillation, urinary incontinence,
and chest pain. Bitter melon is contraindicated in pregnant women
because it can cause bleeding, contractions, stimulation of men-
struation, and miscarriage [47]. Also, those deficient in the glucose-
6-phosphate dehydrogenase gene should avoid consumption of
bitter melon preparations due to the presence of vicine in the seeds.
3.9 Aloe vera Aloe vera, also referred to as “plant of immortality,” is cultivated
around the world. It is a stemless plant and the whole leaf extract
(WLE: gel and latex) contains more than 200 chemicals, including
amino acids, vitamins, minerals, lignin, and phytosterols. The latex
contains ~80 chemical constituents and most of them are anthrones
and anthraquinones (mainly barbaloin, also known as aloin A; and
isobarbaloin, aloin B; and aloeresin A and aloeresin B). Aloe vera
has many properties, including wound-healing, antiaging, anti-
inflammatory, fungicidal, antiviral, antibacterial, laxative, immuno-
logical, antiseptic, and antitumor [48]. It is claimed to have
therapeutic effects in arthritis, asthma, chronic fatigue syndrome,
digestive and bowel disorders, external and internal ulcers, psoria-
sis, genital herpes, hypertension and diabetes [49].
Use of topical Aloe vera is not associated with significant side
effects, but its oral ingestion may cause abdominal cramps and
diarrhea and thereby decrease the absorption of drugs (http://
nccih.nih.gov/health/aloevera). IARC [50] have found ingested
nondecolorized liquid Aloe vera to be carcinogenic in animals, and
state that it is a possible carcinogen in humans as well. Under the
guidelines of California Proposition 65, orally ingested nondecol-
orized Aloe vera WLE has been listed by the OEHHA among
“chemicals known to the state to cause cancer or reproductive tox-

icity” [51].
Evidence of carcinogenic activity was demonstrated in F344/N
rats following 2-year oral administration of Aloe vera WLE in
drinking water [14]. Studies of aloe latex genotoxicity mainly focus
on a number of anthraquinones, which have been found muta-
genic in most in vitro and in vivo assay systems. Aloe-emodin
induces DNA damage in human lung carcinoma cells through pro-
duction of reactive oxygen species (ROS) and increased chromo-
somal aberration frequencies in Chinese hamster ovary cells.
Danthron induces DNA damage and caspase cascade-mediated
apoptosis through mitochondrial permeability transition pores and
BAX-triggered pathways in SUN-1 human gastric cancer cells. An
in vivo mouse Comet assay was performed on isolated kidney and
colon cells to demonstrate the possible organ-specific genotoxicity
of aloe-emodin [52]. The primary DNA damage in the colon was
observed between 3 and 6 h after two oral administrations at 500,
1000, and 2000 mg/kg body wt, suggesting an in vivo genotoxic
mode of action. These findings suggest that aloe-emodin should
be considered as an in vivo genotoxin and the risk for human expo-
sure should be taken into account [7].
3.10 Ephedrine The ingredient sources of the ephedrine alkaloids in dietary supple-
Alkaloids ments include raw botanicals and extracts from those plant sources
[53]. Ephedra sinica Stapf, Ephedra equisetina Bunge, Ephedra
intermedia var. tibetica Stapf, Ephedra distachya L. (the Ephedras),
Sida cordifolia, and Pinellia ternata (Thunb) are sources of ephed-
rine alkaloids. Other common names that have been used for the
various plants that contain ephedrine alkaloids include sea grape,
yellow horse, joint fir, popotillo, and country mallow. Ephedra spp.
is used to produce Ma huang, an herbal drug used in asthma, allergy,
and cold formulations, diet pills, and various supplements [54–56].
Cardiotoxicity associated with ephedra has been well documented
[57, 58]. Drugs containing ephedrine, in the form of Ma huang, are
often combined with caffeine, in the form of guarana. This combina-
tion of drugs acts synergistically, enhancing the toxicity of this prod-
uct. Doses of 1.3–88.9 mg/kg Ma huang given concurrently with
4.4–296.2 mg/kg guarana have been associated with clinical toxico-
sis in dogs. One dog given a dose of 5.8 mg/kg Ma huang and
19.1 mg/kg guarana died. In a number of studies in rodents, the
combined use of ephedra and caffeine has been associated with car-
diovascular toxicity and death [59, 60]. In the Federal Register of
February 11, 2004 (69 FR 6788), it was concluded by the FDA that
dietary supplements containing ephedrine alkaloids are adulterated
under section 402(f)(1)(A) [21 U.S.C. 342 (f)(1)(A)] of the Act
because they present an unreasonable risk of illness or injury under
the conditions of use recommended or suggested in labeling, or if
no conditions of use are suggested or recommended in labeling,

under ordinary conditions of use [53].
3.11 Pennyroyal Oil Even though pennyroyal (Mentha pulegium) oil is extremely toxic,
people have relied on the fresh and dried herb for centuries.
Pennyroyal oil contains 85% of the ketone pulegone. The oil has
been used in folklore medicine for many years as an abortifacient
and as a means to induce menstruation. The abortive effect of the
oil is thought to be due to irritation of the uterus with subsequent
uterine contractions. Sullivan et al. [61] reported two cases of
pennyroyal oil ingestion for the purpose of abortion, and noted
shock, disseminated intravascular coagulation, massive hepatic
necrosis, and death. Anderson et al. [62] identified pulegone and
menthofuran in patients’ serum causing hepatotoxicity and death.
In addition to hepatotoxicity, the oil is known to cause gastroin-
testinal upset and bleeding, hematuria, vaginal bleeding, renal fail-
ure, and coma.
3.12 Pequi Pequi (Caryocar brasiliense, Camb) almond oil has been reported
Almond Oil to possess unsaturated fatty acid and antioxidant compounds (phe-
nolics, gallic acid, quercetin, and carotenoids) related to beneficial
effects on oxidative stress and inflammatory conditions, and thereby
provides hepatoprotective effects [63]. The oil extracted from
Pequi pulp has been evaluated in oral acute and subchronic toxicity
studies in rats [64]. Acute toxicity conducted in female Wistar rats
revealed LD50 > 2000 mg/kg body wt. In a subchronic study, male
and female Wistar rats received repeated doses of 125, 250, 500, or
1000 mg/kg body wt for 28 days. In both acute and subchronic
studies, Pequi oil was found to be of low toxicity. In subchronic
studies, Pequi oil caused some hematopoietic abnormalities.
4 Nutraceuticals’ Safety Concern During Perioperative Care
Echinacea, ephedra, garlic, ginkgo, ginseng, kava, St. John’s wort,

and valerian are commonly used herbal medications that may pose
a concern during the perioperative period. Complications can arise
from these herbs’ direct and pharmacodynamic or pharmacoki-
netic effects. Direct effects include bleeding from garlic, ginkgo,
and ginseng; cardiovascular instability from ephedra; and hypogly-
cemia from ginseng [65]. Pharmacodynamic herb–drug interac-
tions include potentiation of the sedative effect of anesthetics by
kava and valerian. Pharmacokinetic herb–drug interactions include
increased metabolism of many drugs used in the perioperative
period by St. John’s wort.
5 Toxic Contaminants in Nutraceuticals
A number of toxic contaminants/adulterants, such as phytotoxi-

cants, metals, mycotoxins, pesticides, radiation, therapeutic drugs,
and drugs of abuse can be present in foods and nutraceuticals.
Exposure to these contaminants at higher concentrations can result
in serious adverse health effects or even fatality [3, 66–69]. It is
worth emphasizing that vulnerable subgroups of the population,
such as pregnant women, children, and elderly adults are at greater
risks if the concentrations of these contaminants exceed the tolera-
ble levels. Some of these contaminants are described below in brief.
5.1 Pyrrolizidine Among phytotoxicants, pyrrolizidine alkaloids (PAs) are the most
Alkaloids potent toxins present in plant species, including Crotalaria, Senecio,
Heliotropium, Symphytum, Cynoglossum, Amsinckia, and Echitum
[70–73]. More than 150 PAs have been identified, and some of these
PAs commonly contaminate the food and herbal nutraceuticals. The
PAs contain the pyrrolizidine nucleus and can be represented by the
basic structures of senecionine and heliotrine. The toxic effects of PAs
are somewhat similar, although their potency varies due to their bio-
activation in the liver to toxic metabolites called pyrroles [70]. These
pyrroles are powerful alkylating agents that react with cellular pro-
teins and cross-link DNA, resulting in cellular dysfunction, abnormal
mitosis, and tissue necrosis. Many PAs can cause adverse health
effects, including hepatotoxicity in humans and animals [72, 74–76].
Retrorsine and monocrotaline share the same core structure (ret-
ronecine) and similar metabolic activation pathway, but retrorsine is
more hepatotoxic than monocrotaline [76]. PA toxicity may also dis-
rupt other hepatic functions such as copper metabolism, clotting fac-
tors, NH3 metabolism, and protein metabolism. Recently, Zhu et al.
[77] reported that the significant persistence of PA-derived DNA
adducts in vivo supports their role in serving as a mechanism-based
biomarker of PA exposure and toxicity.
The FDA’s position is clear that PAs are harmful following oral,
nonoral (suppositories), or through broken skin routes of exposure.
Although liver damage is the major documented form of injury to
humans from herbs containing PAs, animal studies suggest that
their toxicity is much broader, affecting the lungs, kidneys, GI tract,
and CNS [53, 70]. In addition, many of the PAs have potential for
cytotoxicity and carcinogenicity. Risk assessment of these alkaloids
is currently based on the carcinogenicity of certain PAs after chronic
application to rats using the sum of detected PAs as the dose metric
[71]. These authors have derived interim Relative Potency (REP)
factors for a number of abundant PAs for use in toxicological risk
management. The FDA ruling banned the internal use of Symphytum
officinale (common comfrey), S. asperum (prickley comfrey), and
Symphytum × uplandicum (Russian comfrey), as well as any other
plant/substance containing PAs.
5.2 Heavy Metals Nutraceuticals, in general, are considered relatively safe, but their
contamination by certain metals can make them unsafe. In fact,
heavy metal contamination in traditional medicines has been well
documented [78–81]. Because of their cumulative properties and
toxicity, heavy metal concentrations could potentially reach levels
that lead to hazardous effects on human and animal health.
Cadmium (Cd), mercury (Hg), lead (Pb), and arsenic (As) are
nonessential toxic elements of special concern because of their
toxicity even at low concentrations [3, 80]. Lead exposure has
been associated with renal tumors, reduced cognitive develop-
ment, and increased blood pressure and cardiovascular toxicity.
In a recent retrospective-observational study, traditional and folk
remedies were found to contain toxic amounts of Pb causing
abdominal pain, constipation, hypertension, anemia, neurologi-
cal symptoms, and kidney and liver dysfunction [82]. Cadmium
may induce toxicity in the kidneys, lungs, bones, and reproduc-
tive and developmental system [83]. Arsenic can cause toxic
effects on dermal, respiratory, GIT, hepatic, cardiovascular, ner-
vous, and hematological systems [83]. Both Cd and As are carci-
nogenic. Toxic effects of Hg poisoning can occur in nervous,
renal, cardiovascular, reproductive and developmental, and
immunological systems [84, 85].
The World Health Organization (WHO) has regulated maxi-
mum permissible limits of toxic metals like As, Hg, Cd, and Pb in
herbal medicines, which amount to 10 ppm, 1.0 ppm, 0.3 ppm,
and 10 ppm, respectively [3]. The USA has standardized recom-
mended daily dietary allowances for essential dietary trace elements,
but not for toxic metals [81]. In Europe, the following limits have
been set: 3 ppm for Pb, 1 ppm for Cd (except for seaweed products,
where the limit is set at 3 ppm); and 0.1 ppm for Hg [86]. Cooper
et al. [87] found that certain traditional Chinese medicines (TCM)
were severely contaminated with As, Pb, and Hg. This unacceptable
finding can place consumers at risk for severe or even fatal heavy
metal/metalloid poisoning from the consumption of these phyto-
medicines. Thus, measuring metal concentrations in nutraceuticals
and food ingredients is important for the assessment of safety and
toxicity [88].
5.3 Pesticides For the large-scale cultivation of herbal and food plants, the use of
pesticides is inevitable. Analysis of crude herbal material has often
shown the presence of pesticide residues [3, 89–92]. Among many
types of pesticides, the use of insecticides and herbicides is maximal,
and of course, insecticides are significantly more toxic than any
other type of pesticides. Insecticides are of several types, including
organophosphates (chlorpyrifos, diazinon, malathion, methyl para-
thion, etc.), carbamates (aldicarb, bendiocarb, carbaryl, carbofuran,
methomyl, etc.), organochlorines (aldrin, DDT, heptachlor, lin-
dane, methoxychlor, etc.), pyrethroids (allethrin, cyphenothrin,
deltamethrin, permethrin, etc.), and neonicotinoids (acetamiprid,
clothianidin, dinotefuran, imidacloprid, thiacloprid, etc.).

Chemically, each class of insecticides is different, so they produce
toxicity via different mechanism of action. Exposure to pesticide
residues exceeding permissible levels can result in many adverse
health consequences from acute problems such as skin rashes and
asthma attacks to chronic problems including neurological, repro-
ductive and developmental disorders, and cancer. It is expected that
the residue of organochlorines can persist in dietary ingredients and
supplements for a longer duration than other types of insecticides.
Following acute exposure, toxic effects are more severe with
organophosphates and carbamates, while following chronic expo-
sure, toxic effects are more devastating from organochlorines.
5.4 Mycotoxins Mycotoxins are secondary metabolites produced by fungi that can
contaminate herbs and dietary supplements. Mycotoxins can be
produced during cultivation of plants, transportation of raw plant
material, and during storage of supplements. Commonly found
mycotoxins in nutraceuticals and dietary ingredients include afla-
toxins, ochratoxin A, citrinin, and others. Raman et al. [93] evalu-
ated several botanical supplements and found the presence of molds
in a great number of samples. The most frequently isolated molds
in medicinal plants were Penicillium, Aspergillus, and Fusarium
[94]. Packed samples of herbal plants have a higher probability of
being infected with molds than nonpacked samples due to increased
humidity inside the pack and unsuitable storage methods [3].
A few studies that found the presence of mycotoxins in herbal
products are cited here in brief. Bungo et al. [95] analyzed raw
medicinal plants collected from Brazilian markets and found con-
tamination with Fusarium and Penicillium. Santos et al. [96]
reported that nearly 87% of medicinal herb samples analyzed in
Spain contained four or more mycotoxins. Romagnoli et al. [97]
detected aflatoxin B1 contamination in different kinds of spices, aro-
matic herbs, and medicinal plants from Italy. Aflatoxin B1 is mainly
produced by the fungus Aspergillus flavus, and its toxicity includes
hepatotoxicity, hepatocarcinogenicity, and mutagenic, teratogenic,
and immunosuppressive effects [98]. The European Pharmacopeia
has set a limit of 2 ppb for aflatoxin B1 and 4 ppb for the sum of
aflatoxins (B1, B2, G1, and G2) for some medicinal herbs [3].
Ochratoxin A (OTA), produced by Penicillium verrucosum,
Aspergillus ochraceus, A. carbonarius, and A. niger, is another
mycotoxin of health concern as it is known to produce nephro-
toxic, hepatotoxic, embryotoxic, teratogenic, neurotoxic, immu-
notoxic, genotoxic, and carcinogenic effects via multiple
mechanisms [3, 99–101]. In a recent study, Gan et al. [99] showed
that OTA at 400 and 800 μg/kg diets significantly increased OTA
concentrations in serum, kidney and spleen, and induced the histo-
pathological lesions in kidney and spleen of pigs. OTA decreased
T-cell receptor (TCR) stimulated T lymphocyte viabilities and IL-2
concentration, increased TNF-α concentration, and decreased
T-AOC levels. Furthermore, OTA increased glucose regulated

protein 78, p38, and ERK½ phosphorylation, and LC3 II and
Atg5 protein expression in kidney and spleen of pigs. Citrinin, pro-
duced by P. citrinum, P. verrucosum, P. radiciola, and P. expansum,
usually co-occurs with OTA, and is known to be found in food and
herbal supplements [102]. These two mycotoxins have been impli-
cated as the cause of Balkan Endemic Nephropathy in humans.
Citrinin is a powerful nephrotoxicant, devoid of mutagenic and
carcinogenic potential. For OTA, a limit of 20 ppb has been
adopted in licorice root. It is noteworthy that a nutraceutical can
be contaminated with multiple mycotoxins, and an additive or
synergistic interaction should be taken into consideration from a
regulatory perspective.
6 Models for Nutraceuticals’ Efficacy, Safety, and Toxicity
Currently, there are many models for efficacy, safety, and toxicity
evaluation of nutraceuticals and food ingredients. The majority of
nutraceuticals are effective against diseases because their ingredi-
ents exert antioxidative, anti-inflammatory, sedative, and adapto-
genic and immunomodulatory properties [103–107]. To evaluate
such properties, in addition to in vivo studies, high-throughput in
vitro, in silico, and omics technology-based assays are available. To
evaluate the safety and toxicity of nutraceuticals, some of the mod-
els utilize vertebrate species using invasive and noninvasive
approaches [108], while others utilize nonvertebrate species, such
as zebrafish and Caenorhabditis elegans [109, 110]. Presently,
alternative in vitro models for safety and toxicity evaluation of
nutraceuticals are considered very informative [111]. Additionally,
there are many novel mechanistic and predictive models to envis-
age adverse outcome pathways and evaluate the safety and toxicity
of nutraceuticals [67, 107, 112–114]. Due to space constraints,
these models are not discussed here, so readers are referred else-
where to a comprehensive resource [2].
7 Nutraceutical–Drug Interaction and Toxicity Outcome
An interaction of a drug with food, herbs, and dietary supplements

can lead to a serious outcome due to multiple factors [43, 112,
115–117]. These authors described that in clinical practices, aging,
concomitant medications, transplant recipients, patients with can-
cer or diabetes, malnutrition, HIV infection, and those receiving
enteral or parenteral feeding may be at increased risk of drug–food
or drug–herb interactions. Medications with narrow therapeutic
index or potential life-threatening toxicity, e.g., the nonsteroidal
anti-inflammatory drugs (NSAIDS), opioid analgesics, cardiovas-
cular medications, warfarin, anticancer drugs, and immunosup-
pressants may be at risk of significant drug–food interactions [117].

Food ingredients can influence bioavailability of nutraceuticals and
therapeutic drugs, and thereby their efficacy.
Some nutraceuticals/dietary supplements/food ingredients
are known to influence the drug metabolizing enzyme cytochrome
P450 (CYP450) and/or transporters, and therefore can reduce or
enhance the effects of therapeutic drugs [43, 68, 118–121].
Multiple constituents of the Ginkgo biloba extract are inhibitors or
inducers of CYP450 [43, 120, 122]. Ginseng is a CYP450 inducer,
while grapefruit juice is a CYP450 inhibitor. Peppermint oil is also
reported to inhibit the CYP450 isoform (CYP1A2, CYP2C19,
CYP2C9, and CYP3A4) [123]. St. John’s wort (Hypericum perfo-
ratum) extracts are known to inhibit and induce CYP enzymes
[124]. In vitro studies showed that hyperforin, hypericin, and
I3-II8-biapigenin inhibited CYP3A4-mediated metabolism of
midazolam and testosterone. Durr et al. [125] observed a 2.5-fold
increase in hepatic activity of CYP3A2 after rats were administered
hypericum extract for 14 days.
The best characterized transporter is P-glycoprotein (P-gp), the
product of the ABCB1 gene, which functions as a drug efflux pump.
P-gp functions at the cellular level in the transport of numerous
structurally and pharmacologically unrelated lipophilic and amphipa-
thic drugs, carcinogens, toxins, and other xenobiotics in intestine,
liver, kidney, brain, and other organs. Currently, the interdepen-
dence of drug metabolism and transport on the disposition of drugs
has been described as “transport–metabolism interplay.” Ginseng is
a P-gp inhibitor. A 3.8-fold increase in intestinal P-gp expression was
reported by hypericum extract in rats treated for 14 days [125].
Hypericum extract can significantly modulate pharmacodynamics
and pharmacokinetics of many drugs and xenobiotics.
The interaction between nutraceuticals and polypharmacy is
noteworthy [126, 127]. Polypharmacy is generally defined as the
use of five or more prescription medications on a regular basis. The
average number of prescribed and over-the-counter medications
used by community-dwelling older adults per day in the USA is six
medications, and the number used by institutionalized older per-
sons is nine medications [128]. Almost all medications affect nutri-
ture, either directly or indirectly, and nutriture affects drug
disposition and effect. Issues surrounding polypharmacy, food–
drug interactions, and the consequences of these interactions for
the older adult are of serious concern.
7.1 Pharmaco- Pharmacokinetic interactions are simply interactions due to similar

kinetics and or different pathways with which the body actively handles ingested
Pharmacodynamics in xenobiotics. Pharmacokinetics relates to absorption, distribution,
Nutraceutical–Drug metabolism and excretion (ADME) of xenobiotics, whereas phar-
Interactions and Toxicity macodynamics relates to the mechanism of action by which a drug
molecule works to produce its therapeutic/toxic effect [46].
An example of pharmacokinetic interaction is the interaction

between ginkgo and phenytoin. Ginkgo by increasing metabolism
reduces the blood level of phenytoin. Pharmacokinetic interactions
are often more difficult to predict without knowing the structures
of the compounds, because molecular configurations in 3D space
often play a critical role in inhibiting transport channels, proteins,
or receptors that are vital in the mechanism of pharmacokinetic
interactions. An example of a pharmacodynamic interaction is the
one between Ma Huang and theophylline due to similar adverse
effect profiles. Another example of a pharmacodynamic interaction
is that between Dang gui and warfarin. Dang gui is a Chinese herb,
which is widely used for regulating menstrual irregularities, pains,
and also for menopausal symptoms. Warfarin, on the other hand, is
used as an anticoagulant to prevent blood clots in those prone to
cardiovascular events. Even though the two agents are indicated for
different disease states, they both are recognized as anticoagulants.
It needs to be pointed out as a confounder that sex differences
in human and animal toxicology may significantly influence toxico-
kinetics, toxicodynamics, and adverse reaction outcome [129]. Aside
from sex-specific organs, sex differences and age interactions occur
for a wide range of disease states as well as hormone-influenced con-
ditions and drug distribution, and women have more adverse drug
reactions than men. The classic sex hormone paradigm (gonadec-
tomy and replacement) reveals a significant interaction between sex
and toxicokinetics (ADME).
Most nutraceuticals are multi-targeted small molecules [126,
130]. They affect multiple signaling pathways to modulate their
effect. There is intensive research to elucidate the exact mechanism
of action of the nutraceuticals, including the signaling pathways as
well as the direct and indirect protein targets involved in those
pathways [126].
8 Biomarkers of Nutraceuticals’ Toxic Potential
Nutraceuticals constitute a wide variety of substances intended for

the well-being of humans and animals. They can be used as preven-
tative and/or therapeutic measures. Often, nutraceuticals are
regarded as safe, but some of them present significant or unreason-
able risk to health, either due to an inherent toxic potential or due
to a toxic interaction. Presence of illegal adulterants in dietary and
herbal supplements is common and of serious concern [69]. An
ingredient or its metabolite with a toxic potential can be detected
in body fluids and/or tissues to ascertain that the nutraceutical has
been consumed. For example, detection of aristolochic acid or pyr-
rolizidine alkaloid in a biological system is indicative of use of
Aristolochia and comfrey, respectively. Presence of ephedrine alka-
loid is indicative of ephedra consumption. The target organs for
toxicity of Aristolochia, comfrey, and ephedra are the renal, hepatic,

and cardiac systems, respectively [53]. In a recent study, Posma
et al. [131] developed an integrated analytical and statistical two-
dimensional spectroscopy strategy for metabolite identification. In
most cases, biomarkers of nutraceuticals’ toxicity are not specific,
so in such a scenario, biomarkers of a target organ’s structure and
function are used. For example, alterations in serum GOT, GPT,
bilirubin, GGT for liver; BUN, creatinine, and NGAL for kidney;
and creatine kinase and troponin for cardiac system are used as
biomarkers of toxic effects. Currently, organ-specific microRNAs
are preferred as biomarkers of organ toxicity, as they show early
changes as compared to conventional biomarkers [132]. By now,
the importance of biomarkers of susceptibility is also well recog-
nized. People deficient in the glucose-6-phosphate dehydrogenase
gene can be vulnerable to the toxicity of bitter melon and other
nutraceuticals, such as Aloe, Ackee, Lychee, and diosmin, having
hypoglycemic effects [46, 133–135]. For details on biomarkers of
nutraceuticals and nutritional dietary ingredients, refer to recent
publications elsewhere [3, 53, 131, 136].
9 Management of Nutraceuticals’ Toxicity
Nutraceutical-induced toxicity in humans and animals is rare. In

isolated cases, when poisoning occurs, institution of antidotal
treatment is difficult due to the complexity of plant-based nutra-
ceuticals having multiple ingredients. So withdrawal of the sus-
pected nutraceutical often resolves in full recovery. Delayed effects
from nutraceuticals use are not reported.
10 Concluding Remarks and Future Directions
Currently, hundreds of herbal and dietary ingredients are used in

thousands of nutraceutical formulations. In general, nutraceuti-
cals are considered safe, as a large number of them have obtained
GRAS status from the FDA. However, a good number of nutra-
ceuticals have a toxic potential, as they can exert minor effects,
such as GI upset, vomiting, and dizziness to severe effects, such as
hepatic, renal, or cardiac failure, seizures or death. Most recently,
Rao et al. [137] noted that there has been an increase in dietary
supplement exposures reported to the US poison control centers.
This chapter describes some nutraceuticals that are inherently
toxic or those that become toxic due to interactions based on
altered toxicokinetics and toxicodynamics. Depending on the
nutraceutical and its intended use, multiple models can be utilized
to assess its efficacy, safety, and toxicity. Biomarkers of use and
effects of nutraceuticals play great roles in detecting early expo-
sure and any events of toxicity to avoid further use of a particular

nutraceutical. Alterations in circulatory microRNAs can give great
insight into early detection of any damage to vital organs. With
rising costs of prescription drugs and dwindling health care bud-
gets around the world, the future of nutraceuticals seems bright.
Novel nutraceuticals will be available for promoting good health
and for prevention and treatment of diseases. However, prior to
marketing, nutraceuticals need to be vigorously evaluated for their
safety and toxicity. Finally, pharmacists and physicians need to be
adequately educated and trained in the field of polypharmacy to
avoid adverse outcomes from the concomitant use of nutraceuti-
cals and therapeutic drugs.
Acknowledgment
The authors would like to thank Ms. Robin B. Doss for her techni-
cal assistance in preparation of this chapter.
References
1. Pandey M, Verma RK, Saraf SA (2010) Nutraceuticals: efficacy, safety and toxicity.
Nutraceuticals: new era of medicine and Academic Press/Elsevier, Amsterdam,
health. Asian J Pharmaceut. Clin Res pp 883–892
3:11–15 8. Heinonen T, Wilhelm G (2015) Cross match-
2. Gupta RC (ed) (2016) Nutraceuticals: effi- ing observations on toxicological and clinical
cacy, safety and toxicity. Academic Press/ data for the assessment of tolerability and
Elsevier, Amsterdam. 1022 pages safety of Ginkgo biloba leaf extract. Toxicology
3. Gil F, Hernández AF, Martín-Domingo MC 327:95–115
(2016) Toxic contamination of nutraceuticals 9. Bent S, Goldberg H, Padula A et al (2005)
and food ingredients. In: Gupta RC (ed) Spontaneous bleeding associated with Ginkgo
Nutraceuticals: efficacy, safety and toxicity. biloba; a case report and systematic review of
Academic Press/Elsevier, Amsterdam, the literature. J Gen Intern Med 20:657–661
pp 825–837 10. Posadzki P, Watson L, Ernst E (2012) Herb-
4. FDA (1994) Dietary supplement health and drug interactions: an overview of systematic
education act of 1994. Congress, Pub. L. reviews. Br J Clin Pharmacol 75:603–618
www.fda.gov/DietarySupplement/default. 11. Diamond B, Baily M (2013) Ginkgo biloba;
htm indications, mechanisms and safety. Psychiatr
5. NTP (2010) Toxicology and carcinogenesis Clin North Am 36:73–78
studies of goldenseal root powder (Hydrastis 12. Dziwenka M, Coppock RW (2016) Ginkgo
canadensis) in F344/N rats and B6C3F1 biloba. In: Gupta RC (ed) Nutraceuticals: effi-
mice (feed studies). Natl Toxicol Program cacy, safety and toxicity. Academic Press/
Tech Rep Ser 562:1–188 Elsevier, Amsterdam, pp 681–691
6. Gurley BJ, Gardner SF, Hubbard MA et al 13. Ude C, Schubert-Zsilavecz M, Wurglics M
(2005) In vivo effects of goldenseal, kava kava, (2013) Ginkgo biloba extracts: a review of the
black kohosh, and valerian on human cyto- pharmacokinetics of the active ingredients.
chrome P450 1A2, 2D6, 2E1, and 3A4/5 Clin Pharmacokinet 52:727–749
phenotypes. Clin Pharmacol Ther 14. NTP (2013) Toxicology and carcinogenesis
77:415–426 studies of Ginkgo biloba extract (CAS No.
7. Zhang Z, Mei N, Chen S et al (2016) 90045-36-6) in F344/N rats and B6C3F1/N
Assessment of genotoxic effects of selected mice (Gavage studies). Nat Toxicol Program
herbal dietary supplements. In: Gupta RC (ed) Tech Report Ser 578:1–183
15. Coppock RW, Dziwenka M (2016) Green tea 29. EFSA (2015) Scientific opinion on the safety
extract. In: Gupta RC (ed) Nutraceuticals: of caffeine. EFSA J 13(5):1–120
efficacy, safety and toxicity. Academic Press/ 30. Raina R, Mondhe DM, Malik JK, Gupta RC
Elsevier, Amsterdam, pp 633–652 (2016) Garcinia cambogia. In: Gupta RC
16. Kapetanovic IM, Crowell JA, Krishnaraj R (ed) Nutraceuticals: efficacy, safety and toxic-
et al (2009) Exposure and toxicity of green ity. Academic Press/Elsevier, Amsterdam,
tea polyphenols in fasted and nonfasted dogs. pp 669–680
Toxicology 260:28–38 31. Kayode OA, Jimoh Olusegun R, Adesanya
17. Lambert JD, Kennett MJ, Sang S et al (2010) Olamide A et al (2007) Effects of crude etha-
Hepatotoxicity of high oral dose (-)- nolic extract of Garcinia cambogia on the
epigallocatechin- 3-gallate in mice. Food reproductive system of male Wistar rats (Rattus
Chem Toxicol 48:409–416 norwegicus). Afr J Biotechnol 6:1236–1238
18. Chandra AK, Choudhury SR, De N et al 32. Kim YJ, Choi MS, Park YB et al (2013)
(2011) Effect of green tea (Camellia sinensis) Garcinia cambogia attenuates diet induced
extract on morphological and functional adiposity but exacerbates hepatic collagen
changes in adult male gonads of albino rats. accumulation and inflammation. World
Indian J Exp Biol 49:689–697 J Gastroenterol 19:4689–4701
19. Shimizu M, Shirakami Y, Sakai H et al (2015) 33. Dara L, Hewett J, Lim JK (2008) Hydroxycut
Chemopreventive potential of green tea cate- hepatotoxicity: a case series and review of liver
chins in hepatocellular carcinoma. Int J Mol toxicity from herbal weight loss supplements.
Sci 16:6124–6139 World J Gastroenterol 14:6999–7004
20. Ogunleye AA, Xue F, Michels KB (2010) 34. Sharma T, Wong L, Tsai N et al (2010)
Green tea consumption and breast cancer risk Hydroxycut® (herbal weight loss supplement)
or recurrence: a meta-analysis. Breast Cancer induced hepatotoxicity: a case report and review
Res Treat 119:477–484 of literature. Hawaii Med J 69:188–190
21. Abd El-Aty AM, Choi JH, Rahman MM et al 35. Kaswala DH, Shah S, Patel N et al (2014)
(2014) Residues and contaminants in tea and Hydroxycut-induced liver toxicity. Ann Med
tea infusions: a review. Food Addit Contam Health Sci Res 4:143–145
Part A 31:1794–1804 36. Lopez AM, Kornegay J, Hendrickson RG
22. Wang J, Cheung W, Leung D (2014) (2014) Serotonin toxicity associated with
Determination of pesticide residue transfer Garcinia cambogia over-the-counter supple-
rates (percent) from dried tea leaves to brewed ment. J Med Toxicol 10:399–401
tea. J Agric Food Chem 62:966–983 37. Zhou P, Gross S, Liu J-H et al (2010)
23. Garg SK (2016) Green coffee bean. In: Gupta Flavokawain B, the hepatotoxic constituent
RC (ed) Nutraceuticals: efficacy, safety and from kava root, induces GSH-sensitive oxida-
toxicity. Academic Press/Elsevier, Amsterdam, tive stress through modulation of IKK/
pp 653–667 NF-κB and MAPK signaling pathways.
24. Fredholm BB, Bättig K, Holmén J et al (1999) FASEB J 24:4722–4732
Actions of caffeine in the brain with special ref- 38. Coppock RW, Dziwenka M (2016) St. John’s
erence to factors that contribute to its wide- wort. In: Gupta RC (ed) Nutraceuticals: effi-
spread use. Pharmacol Rev 51:83–133 cacy, safety and toxicity. Academic Press/
25. Rudolph T, Knudsen K (2010) A case of fatal Elsevier, Amsterdam, pp 619–631
caffeine poisoning. Acta Anesth Scand 39. Becker LC, Bergfeld WF, Belsito DV et al
54:521–523 (2014) Amended safety assessment of
26. Campana C, Griffin PL, Simon EL (2014) Hypericum perforatum-derived ingredients as
Caffeine overdose resulting in severe rhabdo- used in cosmetics. Int J Toxicol 33(3
myolysis and acute renal failure. Am J Emerg suppl):5S–23S
Med 32:111.e3–111.e4 40. Brown TM (2000) Acute St. John’s wort tox-
27. Menezes FP, Da Silva RS (2017) Caffeine. In: icity. Am J Emerg Med 18:231–232
Gupta RC (ed) Reproductive and develop- 41. Lampri ES, Ioachim E, Harissis H et al (2014)
mental toxicology, 2nd edn. Academic Press/ Pleomorphic hepatocellular carcinoma fol-
Elsevier, Amsterdam, pp 399–411 lowing consumption of Hypericum perfora-
28. Roberts A (2016) Caffeine: an evaluation of tum in alcoholic cirrhosis. Wolrd
the safety database. In: Gupta RC (ed) J Gastroenterol 20:2113–2116
Nutraceuticals: efficacy, safety and toxicity. 42. Demiroglu YZ, Yeter TT, Boga C et al (2005)
Academic Press/Elsevier, Amsterdam, Bone marrow necrosis: a rare complication of
pp 417–434 herbal treatment with Hypericum perforatum
(St. John’s Wort). Acta Med Austriaca 56. Ooms TG, Khan S (2001) Suspected caffeine
48:91–94 and ephedrine toxicosis resulting from
43. Gupta RC, Chang D, Nammi S et al (2017) ingestion of an herbal supplement contain-
Interactions between antidiabetic drugs and ing guarana and ma huang in dogs: 47 cases
herbs: an overview of mechanisms of action (1997-1999). J Am Vet Med Assoc
and clinical implications. Diabetol Metab 218:225–229
Syndr 9:59 57. Andraws R, Chawla P, Brown DL (2005)
44. Joseph B, Jini D (2013) Antidiabetic effects Cardiovascular effects of ephedra alkaloids: a
of Momordica charantia (bitter melon) and comprehensive review. Prog Cardiovasc Dis
its medicinal potency. Asian Pac J Trop Dis 47:217–225
3(2):93–102 58. Flanagan CM, Kaesberg JL, Mitchell ES et al
45. Ojewole JA, Adewole SO, Olayiwola G (2010) Coronary artery and thrombosis fol-
(2006) Hypoglycemic and hypotensive effects lowing chronic ephedra use. Int J Cardiol
of Momordica charantia Linn (Cucurbitaceae) 139(1):e11–e13
whole-plant aqueous extract in rats. 59. Nyska A, Murphy E, Foley JE et al (2005)
Cardiovasc J South Afr 17:227–232 Acute hemorrhagic myocardial necrosis and
46. Chan N, Li S, Perez E (2016) Interactions sudden death of rats exposed to a combina-
between Chinese nutraceuticals and western tion of ephedrine and caffeine. Toxicol Sci
medicines. In: Gupta RC (ed) Nutraceuticals: 83:388–396
efficacy, safety and toxicity. Academic Press/ 60. Dunnick JK, Kissling G, Gerken DK et al
Elsevier, Amsterdam, pp 875–882 (2007) Cardiotoxicity of Ma Huang/caffeine
47. Bitter Mellon (2013) Bitter Mellon. http:// in a rodent model system. Toxicol Pathol
www.zmescience.com. Accessed 20 March 35:657–666
2013 61. Sullivan JB, Rumack BH, Thomas H et al
48. Gupta VK, Malhotra S (2012) Pharmaco (1979) Pennyroyal oil poisoning and hepato-
logical attribute of Aloe vera: revalidation toxicity. JAMA 242:2873–2874
through experimental and clinical studies. 62. Anderson IB, Mullen WH, Mecker JE et al
AYU 33:193–196. http://www.ayujournal. (1996) Pennyroyal toxicity: measurement of
org/text.asp?2012/33/2/193/105237 toxic metabolite levels in two cases and review of
49. Choudhary M, Kochhar A, Sangha J (2014) the literature. Ann Intern Med 124:726–734
Hypoglycemic and hypolipidemic effect of 63. Torres LRDO, de Santana FC, Torres-Leal
Aloe vera L. in non-insulin dependent diabet- FL et al (2016) Pequi (Caryocar brasiliense
ics. J Food Sci Technol 51:90–96 Camb) almond oil attenuates carbon
50. IARC (2006) IARC Monograph on the eval- tetrachloride-induced acute hepatic injury in
uation of carcinogenic risks to humans. rats: antioxidant and anti-inflammatory
Preamble. http://monographs.iarc.fr/ effects. Food Chem Toxicol 97:206–216
ENG/Preamble/CurrentPreamble.pdf 64. Traesel GK, Menegati SELT, dos Santos AC
51. Proposition 65 (2015) Chemicals listed effec- et al (2016) Oral acute and subchronic toxic-
tive December 4, 2015 as Known to the State ity studies of the oil extracted from Pequi
of California to cause cancer: Aloe vera, non- (Caryocar brasiliense, Camb) pulp in rats.
decolorized whole leaf extract and goldenseal Food Chem Toxicol 97:224–231
root powder. US Office of Environmental 65. Ang-Lee MK, Moss J, Chen-Su Y (2001)
Health Hazard Assessment. 4 December 2015 Herbal medicines and perioperative care.
52. Nesslany F, Simar-Meintieres S, Ficheux H Review. JAMA 286(2):208–216
et al (2009) Aloe-emodin-induced DNA frag- 66. Chan K (2003) Some aspects of toxic con-
mentation in the mouse in vivo comet assay. taminants in herbal medicines. Chemosphere
Mutat Res 678:13–19 52(9):1361–1371
53. Hilmas CJ, Fabricant DS (2014) Biomarkers 67. Jordan SA, Cunningham DG, Marles RJ
of toxicity for dietary ingredients contained in (2010) Assessment of herbal products: chal-
dietary supplements. In: Gupta RC (ed) lenges, and opportunities to increase the
Biomarkers in toxicology. Academic Press/ knowledge base for safety assessment. Toxicol
Elsevier, Amsterdam, pp 609–627 Appl Pharmacol 243:198–216
54. Means C (1999) Ma huang: all natural but 68. Shi S, Klotz U (2012) Drug interactions with
not always innocuous. Vet Med 94:511–512 herbal medicines. Clin Pharm 51:77–104
55. Means C (2005) Decongestants. In: Plumlee 69. Guo B, Wang M, Liu Y et al (2015) Wide-
K (ed) Clinical veterinary toxicology. Mosby, scope screening of illegal adulterants in dietary
St. Louis, MO, pp 309–311 and herbal supplements via rapid polarity-
switching and multistage accurate mass confir- 82. Mehta V, Midha V, Mahajan R et al (2017)
mation using an LC-IT/TOF hybrid Lead intoxication due to Ayurvedic medica-
instrument. J Agric Food Chem tions as a cause of abdominal pain in adults.
63:6954–6967 Clin Toxicol 55:97–101
70. Panter KE, Welch KD, Gardner DR (2014) 83. Flora SJS, Agrawal S (2017) Arsenic, cad-
Poisonous plants: biomarkers for diagnosis. mium, and lead. In: Gupta RC (ed)
In: Gupta RC (ed) Biomarkers in toxicology. Reproductive and developmental toxicology.
Academic Press/Elsevier, Amsterdam, Academic Press/Elsevier, Amsterdam,
pp 563–589 pp 537–566
71. Merz K-H, Schrenk D (2016) Interim rela- 84. Flora SJS (2014) Metals. In: Gupta RC (ed)
tive potency factors for the toxicological risk Biomarkers in toxicology. Academic Press/
assessment of pyrrolizidine alkaloids in food Elsevier, Amsterdam, pp 485–519
and herbal medicines. Toxicol Lett 85. Bridges CC, Zalpus RK (2017) The aging
263:44–57 kidney and the nephrotoxic effects of mer-
72. Preliasco M, Gardner D, Moraes J et al (2017) cury. J Toxicol Environ Health Part B
Senecio grisebachii Baker: Pyrrolizidine alka- 20:55–80
loids and experimental poisoning in calves. 86. Gasser U, Klier B, Kuhn AV et al (2009)
Toxicon 133:66–73 Current findings on the heavy metal content
73. Wang Y, Xiang L, Yi X et al (2017) Potential in herbal drugs. Pharmeur Sci Notes
anti-inflammatory steroidal saponins from the 1:37–49
barriers of Solanum nigrum (European black 87. Cooper K, Noller B, Connell D et al (2007)
nightshade). J Agric Food Chem Public health risks from heavy metals and met-
65:4262–4272 alloids present in traditional Chinese medi-
74. Winship KA (1991) Toxicity of comfrey. cines. J Toxicol Environ Health A
Adverse Drug React Toxicol Rev 10:47–59 70:1694–1699
75. Fu PP, Xia QS, He XB et al (2017) Detection 88. Bhat R, Kiran K, Arun AB (2010)
of pyrrolizidine alkaloid DNA adducts in liv- Determination of mineral composition and
ers of cattle poisoned with Heliotropium euro- heavy metal content of some nutraceutically
paeum. Chem Res Toxicol 30:851–858 valued plant products. Food Anal Methods
76. Yang XJ, Li WW, Sun Y et al (2017) 3:181–187
Comparative study of hepatotoxicity of pyr- 89. Ahmed MT, Loutfy N, Yousef Y (2001)
rolizidine alkaloids retrorsine and monocrota- Contamination of medicinal herbs with
line. Chem Res Toxicol 30:532–539 organophosphorus insecticides. Bull Environ
77. Zhu L, Xue J, Xia Q et al (2017) The long Contam Toxicol 66:421–426
persistence of pyrrolizidine alkaloid-derived 90. Wong TC, Lee FS, Hu GL et al (2007) A sur-
DNA adducts in vivo: kinetic study following vey of heavy metal and organochlorine pesti-
single and multiple exposures in male ICR cide contaminations on commercial Lingzhi
mice. Genotox Carcinogen 91:949–965 products. J Food Drug Anal 15:472–479
78. Meena AK, Bansal P, Kumar S et al (2010) 91. Sarkhail P, Yunesian M, Ahmadkhaniha R
Estimation of heavy metals in commonly used et al (2012) Levels of organophosphorus pes-
medicinal plants: a market basket survey. ticides in medicinal plants commonly con-
Environ Monit Assess 170:657–660 sumed in Iran. Daru J Pharm Sci 20:9
79. Harris ES, Cao S, Littlefield BA et al (2011) 92. Tong M, Gao W, Jiao W et al (2017) Uptake,
Heavy metal and pesticide content in com- translocation, metabolism, and distribution of
monly prescribed individual raw Chinese glyphosate in nontarget tea plant (Camellia
herbal medicines. Sci Total Environ sinensis L.). J Agric Food Chem
409:4297–4305 65:7638–7646
80. Rao MM, Meena AK et al (2011) Detection 93. Raman P, Patino LC, Nair MG (2004)
of toxic heavy metals and pesticide residue in Evaluation of metal and microbial contamina-
herbal plants which are commonly used in the tion in botanical supplements. J Agric Food
herbal formulations. Environ Monit Assess Chem 52:7822–7827
181:267–271 94. Abou-Arab AAK, Soliman Kawther M,
81. Sarma H, Deka S, Deka H et al (2011) Tantawy MEEI et al (1999) Quantity estima-
Accumulation of heavy metals in selected tion of some contaminants in commonly used
medicinal plants. Rev Environ Contam medicinal plants in the Egyptian market.
Toxicol 214:63–86 Food Chem 67:357–363
95. Bungo A, Almodovar AAB, Pereira TC analysis on big data. In: Gupta RC (ed)
(2006) Occurrence of toxigenic fungi in Nutraceuticals: efficacy, safety and toxicity.
herbal drugs. Braz J Microbiol 37:47–51 Academic Press/Elsevier, Amsterdam,
96. Santos L, Marin S, Sanchis V et al (2009) pp 239–248
Screening of mycotoxin multicontamination 108. Peterson JD (2016) Noninvasive in vivo opti-
in medicinal and aromatic herbs sampled in cal imaging models for safety and toxicity
Spain. J Sci Food Agric 89:1802–1807 testing. In: Gupta RC (ed) Nutraceuticals:
97. Romagnoli B, Menna V, Gruppioni N et al efficacy, safety and toxicity. Academic Press/
(2007) Aflatoxins in spices, aromatic herbs, Elsevier, Amsterdam, pp 305–317
herb-teas, and medicinal plants marketed in 109. Barnet RE, Bailey DC, Hatfield HE, Fitsanakis
Italy. Food Control 18:697–701 VA (2016) Caenorhabditis elegans: a model
98. Prado G, Altoé AF, Gomes TC et al (2012) organism for nutraceutical safety and toxicity
Occurrence of aflatoxin B1 in natural prod- evaluation. In: Gupta RC (ed) Nutraceuticals:
ucts. Braz J Microbiol 43:1428–1436 efficacy, safety and toxicity. Academic Press/
99. Gan F, Hou L, Zhou Y et al (2017) Effects of Elsevier, Amsterdam, pp 341–354
ochratoxin A on ER stress, MAPK signaling 110. Bian W-P, Pei D-S (2016) Zebrafish model
pathway and autophagy of kidney and spleen for safety and toxicity testing of nutraceuti-
in pigs. Environ Toxicol 32:2277–2286 cals. In Gupta RC (ed): Nutraceuticals: effi-
100. Gupta RC, Lasher MA, Miller Mukherjee IR, cacy, safety and toxicity. Academic Press/
Srivastava A, Lall R (2017) Aflatoxins, ochra- Elsevier, Amsterdam, pp 333–339
toxins, and citrinin. In: Gupta RC (ed)
111. Krishna G, Gopalakrishnan G (2016)
Reproductive and developmental toxicology, Alternative in vitro models for safety and toxic-
2nd edn. Academic Press/Elsevier, ity evaluation of nutraceuticals. In: Gupta RC
Amsterdam, pp 945–962 (ed) Nutraceuticals: efficacy, safety and toxicity.
101. Gupta RC, Srivastava A, Lall R (2018)
Academic Press/Elsevier, Amsterdam,
Ochratoxins and citrinin. In: Gupta RC (ed) pp 355–385
Veterinary toxicology: basic and clinical prin- 112. Kadakkuzha BM, Liu X-A, Swarnkar S, Chen
ciples, 3rd edn. Academic Press/Elsevier, Y (2016) Genomic and proteomic mecha-
Amsterdam, pp 1019–1027 nisms and models in toxicity and safety evalu-
102. Ostry V, Malir F, Ruprich J (2013) Producers ation of nutraceuticals. In: Gupta RC (ed)
and important dietary sources of ochratoxin Nutraceuticals: efficacy, safety and toxicity.
A and citrinin. Toxins 5(9):1574–1586 Academic Press/Elsevier, Amsterdam,
pp 227–237
103. Gulati K, Anand R, Ray A (2016) Nutraceuticals
as adaptogens: their role in health and disease. 113. Gonzalez-Suarez I, Martin F, Hoenig J,

In: Gupta RC (ed) Nutraceuticals: efficacy, Peitsch MC (2016) Mechanistic network
safety and toxicity. Academic Press/Elsevier, models in safety and toxicity evaluation of
Amsterdam, pp 193–205 nutraceuticals. In: Gupta RC (ed)
Nutraceuticals: efficacy, safety and toxicity.
104. Ajayi AM, Umukoro S, Ben-Aju B et al
(2017) Toxicity and protective effect of phe- pp 287–304
nolic- enriched ethylacetate fraction of
Ocimum gratissimum (Linn.) leaf against 114. Mindukshew I, Kudryavtsev I, Serebriakova
acute inflammation and oxidative stress in M et al (2016) Flow cytometry and light scat-
rats. Drug Dev Res 78:135–145 tering technique in evaluation of nutraceuti-
cals. In: Gupta RC (ed) Nutraceuticals:
105. Burnaz NA, Kücük M, Akar Z (2017) An on- efficacy, safety and toxicity. Academic Press/
line HPLC system for detection of antioxi- Elsevier, Amsterdam, pp 319–332
dant compounds in some plant extracts by
comparing three different methods. 115. Anadón A, Martínez-Laraaga MR, Ires I et al
J Chromatogr B 1052:66–72 (2016) Interactions between nutraceuticals/
nutrients and therapeutic drugs. In: Gupta
106. Sobrinho AP, Minho AS, Ferreira LLC et al RC (ed) Nutraceuticals: efficacy, safety and
(2017) Characterization of anti-inflammatory toxicity. Academic Press/Elsevier,
effect and possible mechanism of action of Amsterdam, pp 855–874
Tibouchina granulosa. J Pharm Pharmacol
69:706–713 116. Sechi S, Di Cerbo A, Canello S et al (2016)
Effects in dogs with behavioral disorders of a
107. Wang K (2016) Adverse reaction prediction commercial nutraceutical diet on stress and
and pharmacovigilance of nutraceuticals: neuroendocrine parameters. Vet Rec.
examples of computational and statistical https://doi.org/10.1136/vr.103865
117. Mouly S, Lloret-Linares C, Sellier PO et al 128. Heuberger R (2012) Polypharmacy and food-
(2017) Is the clinical relevance of drug-food drug interactions among older persons: a
and drug-herb interactions limited to grape- review. J Nutri Gerontol Geriatr 31:325–403
fruit juice and Saint-John’s Wort? Pharmacol 129. Gochfeld M (2017) Sex differences in human
Res 18:82–92 and animal toxicology: toxicokinetics. Toxicol
118. Mooiman KD, Maas-Bakker RF, Hendrikx JJ Pathol 45:172–118
et al (2014) The effect of complementary and 130. Lee KW, Bode AM, Dong Z (2011) Molecular
alternative medicines on CYP3A4-mediated targets of phytochemicals for cancer preven-
metabolism of three different substrates: tion. Nat Rev Cancer 11:211–218
7-benzyloxy-4-trifluoromethyl-coumarin, 131. Posma JM, Garcia-Perez I, Heaton JC et al
midazolam and docetaxel. J Pharm Pharmacol (2017) Integrated analytical and statistical
166(6):865–874 two-dimensional spectroscopy strategy for
119. Oh HA, Lee H, Kim D et al (2017)
metabolite identification: application to dietary
Development of GC-MS based cytochrome biomarkers. Anal Chem 89:3300–3309
P450 assay for the investigation of multi-herb 132. Penman AD, Kaufman GE, Daniels KK

interaction. Anal Biochem 519:71–83 (2014) MicroRNA expression as an indicator
120. Shao F, Zhang H, Xie L et al (2017)
of tissue toxicity. In: Gupta RC (ed)
Pharmacokinetics of ginkgolides A, B and K Biomarkers in toxicology. Academic Press/
after single and multiple intravenous infusions Elsevier, Amsterdam, pp 1003–1018
and their interactions with midazolam in 133. Srivastava A, Kumar A, Thomas JD et al

healthy Chinese male subjects. Eur J Clin (2017) Association of acute toxic encepha-
Pharmacol 73:537–546 lopathy with litchi consumption in an out-
121. Zhang L, Sparreboom A (2017) Predicting break in Muzaffarpur, India, 2014: a
transporter-mediated drug interactions. Clin case-control study. Lancet Glob Health
Pharmacol Ther 101(4):447–449 5:e458–e466
122. Gaudineau C, Beckerman R, Welbourn S et al 134. Hsu CC, Lin MH, Cheng JT et al (2017)
(2004) Inhibition of human P450 enzymes Antihyperglycemic action of diosmin, a citrus
by multiple constituents of the Ginkgo biloba flavonoid, is induced through endogenous-
extract. Biochem Biophys Res Commun endorphin in type I-like diabetes rats. Clin
318(4):1072–1078 Exp Pharmacol Physiol 44:549–555
123. Unger M, Frank A (2004) Simultaneous
135. Sander J, Terhardt M, Sander S et al (2017)
determination of the inhibitory potency of Quantification of methylenecyclopropyl com-
herbal extracts on the activity of six major pounds and acyl conjugates by UPLC-MS/
cytochrome P450 enzymes using liquid chro- MS in the study of the biochemical effects of
matography/mass spectrometry and auto- the ingestion of canned Ackee (Blighia sap-
mated online extraction. Rapid Commun ida) and Lychee (Litchi chinensis). J Agric
Mass Sptectrom 18:2273–2281 Food Chem 65:2603–2608
124. Goey AK, Mooiman KD, Beijnen JH et al 136. Gupta RC (2014) Biomarkers in toxicology.
(2013) Relevance of in vitro and clinical data Academic Press/Elsevier, Amsterdam. 1128
for predicting CYP3A4-mediated herb-drug pages
interactions in cancer patients. Cancer Treat 137. Rao N, Spiller HA, Hodges NL et al (2017)
Rev 39:773–778 An increase in dietary supplement exposures
125. Durr D, Stieger B, Kullak-Ublick GA et al reported to US poison control centers. J Med
(2000) St. John’s Wort induces intestinal Toxicol 13:227–237
P-glycoprotein/MDR1 and intestinal and 138. Coulson JM, Caparrota TM, Thompson JP
hepatic CYP3A4. Clin Pharmacol Ther (2017) The management of ventricular dys-
68:598–604 rhythmia in aconite poisoning. Clin Toxicol
126. Dormán G, Flachner B, Hajdú I, András CD 55:313–321
(2016) Target identification and polypharma- 139. Cope RB (2005) Toxicology brief: Allium

cology of nutraceuticals. In: Gupta RC (ed) species poisoning in dogs and cats. Vet
Nutraceuticals: efficacy, safety and toxicity. Med:462–566
pp 263–286 140. IARC (2002) Some traditional herbal medi-
cines, some mycotoxins, naphthalene and sty-
127. Herr M, Grondin H, Sanchez S et al (2017) rene. Aristolochia species and aristolochic
Polypharmacy and potentially inappropriate acids. IARC monographs on the evaluation of
medications: a cross-sectional analysis among carcinogenic risks to humans, vol 82. WHO,
451 nursing homes in France. Eur J Clin Lyon, pp 69–128
Pharmacol 73:601–608
141. Pelkonen O, Abass K, Wiesner J (2013)

151. Melough MM, Vance TM, Lee SG et al
Thujone and Thujone-containing herbal (2017) Furocoumarin kinetics in plasma and
medicinal and botanical products: toxicological urine of healthy adults following consump-
assessment. Reg Toxcicol Pharmacol tion of grapefruit (Citrus paradisi Macf.) and
65:100–107 grapefruit juice. J Agric Food Chem
142. Heise CW, Brooks DE (2017) Ayahuasca
65(14):3006–3012
exposure: descriptive analysis of calls to US 152. Mazzanti G, Menniti-Ippolito F, Moro PA
poison control centers from 2005 to 2015. et al (2009) Hepatotoxicity from green tea: a
J Med Toxicol 13:245–248 review of the literature and two unpublished
143. Ma J, Li X (2017) MicroRNAs are involved in cases. Eur J Clin Pharmacol 65:331–341
the toxicity of microcystins. Toxin Rev 153. Teschke R, Schulze J (2010) Risk of kava hep-
36(2):167–175 atotoxicity and the FDA consumer advisory.
144. McClellan NL, Manderville RA (2017) Toxic JAMA 304:2174–2175
mechanisms of microcystins in mammals. 154. Dale O, Ma G, Gemelli C et al (2012) Effects
Toxicol Res 6:391–405 of mitragynine and its derivatives on human
145. NTP-National Toxicology Program (2004)
opioid receptors (delta, kappa, and mu).
NTP Technical report on the toxicology and Planta Med Congress Abstract 78:91
carcinogenesis studies of trans-155. Pessayre D, Mansouri A, Haouzi D et al

cinnamaldehyde (microcapsulated) in (1999) Hepatotoxicity due to mitochondrial
F344/N rats and B6C3F1 mice (feed stud- dysfunction. Cell Biol Toxicol 15:367–373
ies). February 2004, http://ntp.niehs.nih. 156. Liu Q, Zhao X, Lu X et al (2012) Proteomic
gov/ntp/htdocs/LT_rpts/tr514.pdf study on usnic acid-induced hepatotoxicity in
146. Jones AW (2017) Review of caffeine-related rats. J Agric Food Chem 60:7312–7317
fatalities along with postmortem blood con- 157. Yokouchi Y, Imaoka M, Niino N et al (2017)
centrations in 51 poisoning deaths. J Anal Comprehensive evaluation of (+)-usnic acid-
Toxicol 41:167–172 induced cardiotoxicity in rats by sequential
147. Zhang ZY, Lin DJ, Li WM et al (2017)
cross-omics analysis. Toxicol Pathol
Sensitive bromide-based screening of poten- 45:481–492
tial toxic Furanoids in Dioscorea bulbifera L. J 158. Sudekum M, Poppenga RH, Raju N et al
Chromat B. Analyt Techn Biomed. Life Sci (1992) Pennyroyal oil toxicosis in a dog.
1057:1–14 J Am Vet Med Assoc 200:817–818
148. Gee P, Jackson S, Easton J (2010) Another 159. Nowak A, Sojka M, Klewicka E et al (2017)
bitter pill: a case of toxicity from DMAA party Ellagitannins from Rubus idaeus L exert
pill. J N Z Med Assoc 123:124–127 geno- and cytotoxic effects against human
149. De Quadros APO, Mazzeo DEC, Marin-
colon adenocarcinoma cell line Caco-2.
Morales MA et al (2017) Fruit extract of the J Agric Food Chem 65:2947–2955
medicinal plant Crataegus oxyacantha exerts 160. Assmann G, Cullen P, Erbey J et al (2006)
genotoxic and mutagenic effects in cultured Plasma sitosterol elevations are associated
cells. J Toxicol Environ Health Part A with increased incidence of coronary events in
80:161–170 men: results of a nested case-control analysis
150. Saito M, Ueno M, Ogino S et al (2005) of the prospective cardiovascular Münster
High dose of Garcinia cambogia is effective (PROCAM) study. Nutr Metab Cardiovasc
in suppressing fat accumulation in develop- Dis 16:13–21
ing male Zucker obese rats, but highly toxic 161. Macgregor FB, Abernethy VE, Dahabra S
to the testis. Food Chem Toxicol et al (1989) Hepatotoxicity of herbal reme-
43:411–419 dies. Br Med J 299:1156–1157
Chapter 19
Impact of Pharmaceuticals on the Environment: Risk

Assessment Using QSAR Modeling Approach
Abstract
An extensive use of pharmaceuticals and the widespread practices of their erroneous disposal measures have
made these products contaminants of emerging concern (CEC). Especially, active pharmaceutical ingredi-
ents (APIs) are ubiquitously detected in surface water and soil, mainly in the aquatic compartment, where
they do affect the living systems. Unfortunately, there is a huge gap in the availability of ecotoxicological
data on pharmaceuticals’ environmental behavior and ecotoxicity which force EMEA (European Medicines
Agency) to release guidelines for their risk assessment. In silico modeling approaches are vital tools to
exploit the existing information to rapidly emphasize the potentially most hazardous and toxic pharmaceu-
ticals and prioritize the most environmentally hazardous ones for focusing further on their experimental
studies. The quantitative structure–activity relationship (QSAR) models are capable of predicting missing
properties for toxic end-points required to prioritize existing, or newly synthesized chemicals for their
potential hazard. This chapter reviews the information regarding occurrence and impact of pharmaceuti-
cals and their metabolites in the environment along with their persistence, environmental fate, risk assess-
ment, and risk management. A bird’s eye view about the necessity of in silico methods for fate prediction
of pharmaceuticals in the environment as well as existing successful models regarding ecotoxicity of phar-
maceuticals are discussed. Available toxicity endpoints, ecotoxicity databases, and expert systems frequently
used for ecotoxicity predictions of pharmaceuticals are also reported. The overall discussion justifies the
requirement to build up additional in silico models for quick prediction of ecotoxicity of pharmaceuticals
economically, without or involving only limited animal testing.
Key words APIs, Ecotoxicity, CEC, In silico, Pharmaceuticals, QSAR, Risk assessment, Risk manage-
ment, Waste management
1 Introduction
Pharmaceuticals are one of the indispensable products with unques-

tionable benefits to human health and lifestyle. Regrettably, due to
overuse of pharmaceuticals together with their improper disposal,
unwanted residues of active pharmaceutical ingredients (APIs)
have been found in different compartments of environment since
1970 [1, 2]. Pharmaceuticals are deliberately designed to have an
explicit mechanism of action (MOA) and exercise an effect on
395
s pecific organs, tissues, or cells in living systems and many of them

are persistent in the body [3]. Thus, pharmaceuticals and their
unaltered metabolites can affect humans as well as animals when
entered into the environment by diverse sources and routes. Also
the MOA designed for human could be different for another type
of species and even potentially “safe” drugs could have serious
effects on them from biological ladder. This feature makes pharma-
ceuticals unique from other chemicals and this is the one and only
reason to assess the potential acute and chronic effects of pharma-
ceuticals in diverse environmental compartments. It is quite obvi-
ous that the toxicity of pharmaceuticals on organisms in aquatic
and nonaquatic environment is due to their long persistent and
bioaccumulative nature [4]. One of the notable effects is the
increased resistance of infectious microorganisms to numerous
antibiotics due to the overuse of pharmaceuticals in humans and
pet animals [5].
The ridiculously excess use of pharmaceutical products was
well reported [5, 6] showing significant negative consequences for
environment and individual health system in the last decades [6, 7].
Thus, increasing levels of detected and measured medicine residues
in the environment make pharmaceuticals as Contaminants of
Emerging Concern (CEC) [8, 9]. Around 600 APIs or their
metabolites and transformation products have been found in the
environment, specifically in surface water and sewage effluent as
well as in ground water and in the soil samples for more than 71
countries all over the world [7, 10]. More than 200 APIs from
therapeutic classes of painkillers, vascular drugs, antibiotics, and
antidepressants are identified in aquatic and terrestrial compart-
ments in concentration ranging from few ng/L to 1000 μg/L
[11]. Majority of APIs are partially degraded or treated in waste
water treatment plants (WWTPs) and finally discharged in the
aquatic environment [12], leading to a ubiquitous and uninter-
rupted contamination [13]. A good amount of contaminants are
released through improper disposals and excretion through feces
and urines into the sewage system [14]. Other significant sources
of pharmaceuticals are hospitals and industries, whose effluents
are loaded with very high concentration of APIs and their metabo-
lites [15].
One of the frequently used pharmaceutical classes is antibiotics
which are poorly metabolized after ingestion, providing a fraction
from 25% to 75% which may leave the bodies in an unchanged
form after consumption [16]. Under nationwide study of “emerg-
ing pollutants” in 139 rivers in 30 states of the USA, the US
Geological Survey (USGS) detected biologically active compounds
of diverse therapeutic classes [17]. The cardiovascular drug pro-
pranolol and antiepileptic drugs like carbamazepine and clofibrate
have been reported in sewage treatment plants [18, 19]. Commonly
used beta blockers (e.g., Metoprolol around 1.54 μg/L) and
Impact of Pharmaceuticals on the Environment: Risk Assessment Using QSAR Modeling… 397
beta-sympathomimetics, estrogens (e.g., 17β-estradiol equal to

0.013 μg/L) [20], analgesic and anti-inflammatory drugs (e.g.,
Diclofenac up to 1.2 μg/L) [21], lipid lowering agents (e.g., clo-
fibrinic acid up to 0.2 μg/L) [22], and antiepileptic drugs (e.g.,
Carbamazepine up to 2.1 μg/L) [21] were detected in major river
water. Interestingly, there is strong evidence of nonprescription
drugs like cotinine, caffeine, and acetaminophenone in samples of
potable water in Atlanta, Georgia [23]. Gemfibrozil and carbam-
azepine were detected in drinking waters in ten cities in Canada by
Tauber [24]. Under sex hormones, estrogens have been detected
in plasticizers and preservatives, while 17α-ethinylestradiol (EE2),
an active component of contraceptive pills, has been identified in
groundwater and tap water samples [25].
The European Medicines Evaluation Agency (EMEA) and the
Food and Drug Administration (FDA) of USA (US FDA) intro-
duced risk assessment guidelines due to continuous detection of
human and veterinary pharmaceuticals and their residues into the
environment. The EMEA guidelines for the assessment of poten-
tial environmental risks were released in 2006 [26]. According to
the US FDA guidelines, applicants have to present an environmen-
tal assessment report when the probable concentration of the API
in the aquatic environment is ≥1 μg/L [27]. Previously, the FDA
Center for Drug Evaluation and Research (CDER) issued a guid-
ance document ‘Guidance for Industry for the Submission of an
Environmental Assessment in Human drug Application and
Supplements’ in 1995 [28]. Due to the emerging concern of their
hazards, directive 2004/27/EC [29, 30] for human medicine and
directive 2004/28/EC [31] for veterinary medicine required an
Environmental Risk Assessment for marketing authorization of
new pharmaceuticals products. The European Union Directive
2013/39/EU [32] included two pharmaceuticals (diclofenac and
17-alpha-ethynilestradiol (EE2)) and a natural hormone (17-beta-
estradiol (E2)) in a first watch list of ten substances for the European
Union monitoring of water. Furthermore, the watch list was
amended in 2015 with directive 2015/495/EU [33], and three
macrolide antibiotics (azithromycin, clarithromycin, and erythro-
mycin), an additional natural hormone (estrone E1), a UV filter
(octinoxate), and an antioxidant used as a food additive (butylated
hydroxytoluene) were included [34].
The usage of pharmaceuticals is so extensive, and it is expected
to intensify due to ageing of population in Europe and the USA,
that the risk assessment requires a huge number of experimental
data ensuing high costs, high time demand, and animal testing for
in vivo testing. Regrettably, the number of available experimental
data is very limited. The available data are also limited to specific
species and assessed for particular environment compartment. In
absence of sufficient experimental data, quantitative structure–
activity/toxicity relationship (QSAR/QSTR) approach represents
a convincing substitute to predict the possible hazards, from their

chemical structure information [35]. Thus, to fill the data gaps,
government and nongovernment regulatory authorities suggest
the use of in silico methods for prediction of the physicochemical
properties, toxicological activity, distribution, fate, etc. of pharma-
ceuticals along with their effects on environment and living sys-
tems much before they enter into market for usage. Therefore,
usage of QSAR as one of the nonexperimental methods is note-
worthy in order to lessen time, animal usage and cost involvement
in toxicity prediction of pharmaceuticals [36, 37]. Persistent and
Bioaccumulative (PB) behavior of pharmaceuticals was studied by
Howard and Muir [38] employing QSAR models. More than
1200 pharmaceuticals were screened and prioritized for overall
Persistence, Bioaccumulation and Toxicity (PBT) potential using
the Insubria PBT Index and the US-EPA PBT Profiler by Sangion
and Gramatica [39]. The European Union Commission’s Scientific
Committee on Toxicity, Ecotoxicity and Environment (CSTEE)
had recommended the use of QSAR models for screening purposes
of pharmaceutical ingredients [40]. In recent times, a good num-
ber of software or expert systems are available for ecotoxicity pre-
diction. A widely used online modeling tool to predict ecotoxicity
of chemicals by QSAR is ECOSAR [41, 42].
Although few QSAR studies have been performed to fill the
data gap in ecotoxicity of pharmaceuticals [43–46], there is a sig-
nificant lack of knowledge about the environmental fate of a huge
number of pharmaceuticals and their metabolites. Thus, genera-
tion of in silico models for pharmaceuticals’ ecotoxicity is the need
of hour. This chapter intends to provide information regarding
occurrence of pharmaceuticals and their residues in the environ-
ment, their persistence, environmental fate, and toxicity along with
application of QSAR models and expert systems to predict risk and
fate properties of pharmaceuticals. Concise ideas about commonly
used endpoints or test batteries, available databases and expert sys-
tems employed for swift ecotoxicity predictions of pharmaceuticals
are discussed.
2 Ecotoxicity of Pharmaceuticals: A General Overview
2.1 Source and Entry To understand the ecotoxicity of pharmaceuticals, the first step is
Routes to identify their sources and entry routes into the environment.
Major sources and familiar pathways for environmental pollution
of pharmaceuticals are illustrated below.
(a) Household disposal: Due to lack or improper instructions
about medication disposal, in many cases expired and unused
medicines are dumped through the toilet or via waste bins,
before being transferred to landfill sites as terrestrial ecosystem
hazards. According to a study, unused medication were found

due to modification of drugs by the doctor (48.9%), or self-
discontinuation (25.8%), and the study identified that the
common approach of disposal was throwing unused drugs in
the trash (76.5%) or flushing them down the drain (11.2%)
[47].
(b) Industrial waste: One of the major sources of API, in-process
and quality control failed final materials are generated from
pharmaceutical industries. Though industries are following
the Good Manufacturing Practices (GMP), still a large num-
ber of evidence is there for significant pharmaceutical emis-
sions from industries. Concentrations up to several mg/L of
pharmaceutical wastes have been identified in effluents for
single compounds, specifically in Asian countries [48].
(c) Hospital influent and effluent: Hospital influents and efflu-
ents are another noteworthy source according to several
researches and it is proved that the eradication of the pharma-
ceuticals is partial. According to a study, 16 pharmaceuticals
including antiepileptics and anti-inflammatories were detected
in the hospital waste water [49].
(d) Continuous introduction of diagnostic compounds: Iodinated
X-ray contrast media like iomeprol, iohexol, and iopromide
are generally employed as diagnostic tools for capturing
detailed X-ray images of soft tissues and thereafter eliminated
without appropriate treatment which helps in persistence of
these wastes for a long period of time in the ecosystem [50].
(e) Human excreta: APIs as well as their metabolites are excreted
through the human excreta which is another imperative source
of pharmaceutical waste. The complexity of risk assessment is
increased manifold as the risk of metabolites is entirely dissimi-
lar from the API.
(f) Aquaculture: Sewage treatment plant (STP) sludge is rou-
tinely used as fertilizer in agriculture. Along with that, antibi-
otics have been also employed in aquaculture primarily for
therapeutic purposes and as prophylactic agents (erythromy-
cin, oxytetracycline, premix, sarafloxacin, sulfonamides)
according to Food and Agriculture Organization (FAO) [51].
(g) Manure, animal husbandry, and veterinary medicine: Urine
and feces of animals in addition to direct application of veteri-
nary medication in aqua farming leads to soil contamination
besides contaminating both the surrounding surface and
groundwater during rain [52].
(h) Plant agriculture: Antibiotics like streptomycin with oxytet-
racycline are very commonly employed in plant agriculture in
controlling bacterial diseases of flower and fruits. They are
principally used for apple, pear, and related fruit trees for
c ontrolling fire blight caused by Erwinia amylovora. Studies

have suggested that antibiotics applied to plants account for
<0.5% of total antibiotic use in the USA [53].
We have demonstrated most common and significant sources,
routes and fate of pharmaceuticals in Fig. 1.
2.2 Occurrence Pharmaceuticals are frequently employed in healthcare, farming,

and Effects and aquaculture nowadays. The defined daily doses (DDDs) of
consumed drugs can be found in the European Surveillance of
Antimicrobial Consumption (ESAC) homepage [54]. A hefty
number of research covering evidence of pharmaceuticals in water
bodies, sewage and effluent treatment plant, manure, soil, and air
dust have been performed. Pharmaceuticals are multicomponent
mixtures rather than isolated pure substances, which will either be
transformed by physical and chemical means and/or subsequently
biotransformed by microorganisms. Therefore, multi-component
mixtures are the primary concerning issue for the ecotoxicity
assessment. Another point of concern is that majority of molecules
can be neutral, cationic, anionic, or zwitterionic which makes the
risk assessment study trickier. Thus, pharmaceuticals are evaluated
for their acute toxicity by standard tests following the guidelines of
Organisation for Economic Co-operation and Development
(OECD), US EPA, International Organization for Standardization
Fig. 1 Sources, routes and fate of pharmaceuticals in different compartments of environment

(ISO) employing standard laboratory endpoints like zooplankton,

algae, and other invertebrates and fish. The most toxic and concern-
ing classes were antibiotics, antibacterials, analgesics, cardiovascular
drugs, antidepressants, and antipsychotics.
● Antibiotics: Quinolones (mostly ciprofloxacin) have been
identified in the hospital effluent up to a low μg/L range while
β-lactams like carbapenems, monobactams, penicillins, cepha-
losporins, and β-lactamase inhibitors were detected in the
lower μg/L range in hospital effluent as well as in the influent
of a municipal STP [55]. Antibiotics like tetracycline, oxytetra-
cycline, chlortetracycline, sarafloxacin, and cyclosporine A are
largely found in the ecosystem and the most concerning issue
is they have quite slow biodegradability in soil. Chlortetracycline
and tetracycline were detected in ten out of 12 soil samples by
Hamscher et al. [56]. As antibiotics have the ability to affect
the microorganisms in sewage systems and WWTP, the inhibi-
tion of wastewater bacteria may critically influence degradation
and nitrification process of organic matter [57]. Carucci [58]
et al. reported noteworthy inhibition of the nitrification
showed by lincomycin. Ciprofloxacin was identified to be
active against Vibrio fisheri at a concentration of 5 mg/L [59].
Processes like transcription and translation are largely affected
by macrolides, tetracyclines, lincosamides, P-aminoglycosides,
and pleuromutilins for plants. Not only this, metabolic path-
ways like folate biosynthesis, fatty acid biosynthesis, and chlo-
roplast replication are influenced by sulfonamides, triclosan,
and fluoroquinolones, respectively [60]. Along with water
bodies, microorganism and plants, antibiotics have the ability
to affect the degradation of organic matter of soils and sedi-
ments to a large extent. A transitory effect on sulfate reduction
was also identified when antibiotics were present in sediment
[61]. Chloramphenicol was banned for food-producing ani-
mals within the USA and the EU in 1994 as it generated severe
hazardous effects including myelosuppression to farmers [62].
In present times, the most significant effect is the surfacing of
resistance due to improper and uncontrolled usage of antibiot-
ics as medicine for human and animals as well as animal hus-
bandry. The most well-known examples are methicillin-resistant
Staphylococcus aureus (MRSA), vancomycin-resistant entero-
cocci (VRE), and multiresistant pseudomonads [63].
● Blood lipid lowering agents: Proliferation of peroxisomes in
rodent liver is caused by fibrates and statins which have the
ability to suppress synthesis of the juvenile hormone in insects.
Additionally, they produce damaging effect to protozoan
parasites, inhibiting growth and development. Zebrafish

(Danio rerio) and amphibians can be highly affected by fibrates
when present even at micromolar concentrations [3]. Quinn
et al. [64] categorized bezafibrate as damaging for nontarget

organisms with EC50 concentration between 10 and 100 mg/L
and gemfibrozil as toxic with EC50 concentration between 1
and 10 mg/L. Fibrates have been evaluated by usual toxicity
assays and the following no-observed-effect-concentration
(NOEC) was detected for clofibric acid in C. dubia is NOEC
(7 days) = 640 μg/L, for the rotifer B. Calyciflorus is NOEC
(2 days) = 246 μg/L, and in early life stages of zebrafish is
NOEC (10 days) = 70 mg/L [65]. Clofibrate is harmful to
aquatic organisms with LC50 values of 7.7–39.7 mg/L. The
most sensitive organism toward clofibrate is the fish Gambusia
holbrooki [LC50 (96 h) = 7.7 mg/L]. Misra et al. [66] reported
that clofibrate has no effect on in vitro growth of T. bruceii but
reduces the incidence of P. berghei and the invasiveness as well
as development of Acanthomoeba culbertsoni in exposed mam-
malian hosts.
● Analgesics and nonsteroidal anti-inflammatory drugs
(NSAIDs): NSAIDs have been detected in higher concentra-
tions especially in surface water, ranging between 0.4 ng/L
and 15 μg/L. Among NSAIDs, paracetamol, diclofenac, and
ibuprofen are the most quantitatively found [67]. Cleuvers
[68] checked that acute toxicities of NSAIDs were moderately
low with concentration (EC50) attained in Daphnia in the
range from 68 to 166 mg/L and from 72 to 626 mg/L in case
of algae. Observed EC50 values for chronic toxicity are
23.8 mg/L, 23.6 mg/L, and 38.2 mg/L for diclofenac, ibu-
profen, and naproxen, respectively in surface water. The
NASIDs can reach in the environment with concentrations up
to >1 μg/L due to their extensive usage and required pharma-
cokinetic and pharmacodynamic properties. Among NSAIDs,
diclofenac has the highest acute toxicity with the effective con-
centrations below 100 mg/L and frequently detected in waste-
water at a median concentration of 0.81 μg/L, whereas the
maximal concentration in wastewater and surface water is up to
2 μg/L [69]. Acetylsalicylic acid affected reproduction in D.
longispina and D. magna at concentrations of 1.8 mg/L [69].
Another frequently prescribed NSAID is paracetamol which is
present in surface waters with concentration of 78.17 μg/L
and within the range of 20 ng/L to 4.3 μg/L in STP effluents.
The detected concentrations are higher than the stipulated no-
effect concentration (PNEC) of 9.2 μg/L [3].
● Beta-blockers: Propranolol showed the highest acute toxicity
and log Kow supporting the fact that it is one of the strongest
membrane stabilizers among the examined beta-blockers by
Huggett et al. [70]. The NOEC and lowest-observed-effect-
concentration (LOEC) of propranolol affecting reproduction
in C. dubia were 125 and 250 μg/L. In case of H. azteca,
reproduction affected after 27 days of exposure in at 100 μg/L

[70]. In aquatic compartment, propranolol negatively affects
survival, swimming behavior, and phototaxis of free-living
aquatic stages of trematodes. Fathead minnows exposed to
atenolol throughout embryo–larval growth showed LOEC
and NOEC values for growth rate of 10 mg/L and 3.2 mg/L,
respectively [3]. With 48-h exposure to propranolol, LC50 val-
ues of 1.6 mg/L, 29.8 mg/L, and 0.8 mg/L were obtained
for D. magna, H. azteca, and C. dubia respectively, while acute
exposure to nadolol did not affect the survival of the inverte-
brates [3]. Encystment of the protozoan Entamoeba invadens
was inhibited in the presence of metoprolol [71].
● Anticancer drugs: These are one of the most toxic therapeutic
classes, designed to kill the cancer cells. They possess geno-
toxic, mutagenic, carcinogenic, teratogenic, and fetotoxic
properties. One of the interesting points is that 14–53% of the
drugs can be excreted in unaltered form through urine and
feces, making them lethal for aquatic, soil organisms as well as
for living systems and ecosystems [72]. Methotrexate has been
reported to show acute effects in the ciliate Tetrahymena pyri-
formis with an EC50 of 45 mg/L and teratogenicity for fish
embryos with an EC50 of 85 mg/L after 48 h of exposure in
both cases [73]. Cyclophosphamide and methotrexate demon-
strated immunosuppressant property to cause a proliferation in
disease incidence and intensity in host–parasite systems [74].
Highly proliferative species like the ciliate Tetrahymena pyrifor-
mis showed acute toxicity to methotrexate with concentration
of EC50 (48 h) = 45 mg/L [75]. Surprisingly, Methotrexate
has no or little effect on definite protozoans like Babesia bovis,
Toxoplasma gondii, and Leishmania tropica, as these organisms
have dissimilar mechanisms of drug metabolism [76]. Again,
cyclophosphamide emerges to have a minute effect on them.
Development and growth of helminths in mammalian and bird
hosts were destructively effected by cyclophosphamide and
methotrexate. Abnormal teratogenicity was observed in fish
embryos at higher concentrations of methotrexate [EC50
(48 h) = 85 mg/L] [76].
● CNS affecting drugs: Fluoxetine, a serotonin reuptake inhib-
itor (SSRI), is one of the acute toxic pharmaceuticals with tox-
icity ranging from EC50 (48 h, alga) = 0.024 mg/L to LC50
(48 h) = 2 mg/L [18]. Sertraline demonstrated highly toxic
response to rainbow trout (LC50 of 0.38 mg/L) at a 96-h
exposure [77]. SSRIs showed growth inhibitions for algae and
chronic toxicity tests confirmed they were sensitive with NOEC
values below 1 mg/L [78]. On the contrary, organism like C.
vulgaris was identified to be the least sensitive species for all
SSRIs tested. Fluvoxamine showed a rise to the highest EC50
values for all algae species tested with concentration range

between 3563 and 10,208 μg/L. Nitrazepam, benzodiaze-
pines, and diazepam were recognized to amplify the number of
microfilariae of Setavia cervi liberated from the lungs into the
peripheral blood circulation in rats [79]. Caffeine was identi-
fied to stimulate the growth of P. falciparum and P. gallina-
ceum, while the mood stabilizer valproic acid and the
antipsychotic haloperidol efficiently inhibited the in vitro
growth of T. gondii [80]. Antiepileptics like carbamazepine
and diazepam were categorized as potentially harmful to
aquatic organisms as majority of the acute toxicity data was
below 100 mg/L [65].
● Sex hormones: These are one of the significant therapeutic
classes emerged as the most serious aquatic environmental haz-
ards due to their widespread use as human contraceptives. A
synthetic estrogen named Ethinylestradiol (EE2) is generally
found in oral contraceptive pills with evident estrogenic effects
in fish. The EE2 concentrations below 1 ng/L creates striking
effects in fertilization process, egg production and decreased
expression of secondary male sex characteristics to fathead
minnows. Lifelong contact of zebrafish to EE2 (with concen-
tration of 5 ng/L) has led to reproductive failure due to the
nonexistence of secondary male sex characteristics [81].
Exposure to 17β-estradiol caused an increased susceptibility to
the protozoan T. gondii in mice, while increased pathology
occurred in mammals infected with Leishmania mexicana
amazonensis and exposed to either estradiol or testosterone
[82]. Estradiol increased the vulnerability of cyprinids to
hemoflagellates by the repression of lymphocyte proliferation
[83]. Hydrocortisone can increase the intensity of ectoparasitic
infections in fish at reasonably high concentrations. The
detected concentrations of estrogenic products are typically
below 50 ng/L in the effluent of STP and WWTP. On the
contrary, high concentrations of 17α-estradiol and estriol
(about 180 ng/L and 590 ng/L, respectively) were found in
the USA [84].
● Antiparasitic compounds: Antiparasitic compounds dora-
mectin and ivermectin, with concentration of 0.112 mg/kg
and 1.85 mg/kg, respectively were detected in dung of a farm-
house in the UK [85]. Interestingly, the concentrations of
these drugs in soil were noticeably lower up to 0.046 mg/kg
for the same farm [85]. Grønvold et al. [86] found that fen-
bendazole and ivermectin affect the endurance of the nema-
tode Pristionchus maupasi at concentrations of 10–20 mg/kg
and higher than 3 mg/kg, respectively. Svendsen et al. [87]
reported that fenbendazole and ivermectin did not affect
earthworms; however, the disappearance in dung was affected
by the avermectin but not by the fenbendazole. Sun et al. [88]
detected avermectin B1A in soil with the compost worm Eisenia

fetida at a concentration of 17.1 mg/kg (LC50).
● Antivirals: The emergence of Relenzas (zanamivir) and
Tamiflu, neuraminidase inhibitors, in the USA began after the
influenza pandemic (H1N1) in 2009 [89]. Tamiflu overpow-
ered Relenza due to its relative ease of administration. Tamiflu,
a prodrug form, is converted to the active molecule oseltamivir
carboxylate (OC) in the liver. Generally, 80% of an oral dose of
Tamiflu is excreted as OC through urine and the remaining
fractions are excreted as oseltamivir ethylester-phosphate (OP)
in the feces. Thus, both the API and its active metabolite ulti-
mately are projected to attain a mean of 2–12 mg/L in WWTPs
during moderate and severe pandemics, respectively [89]. The
OC concentrations ranging from 293 to 480 ng/L have been
reported in river waters charged with WWTP effluents during
the 2009 pandemic [90].
2.3 Pharmaceutical The definition of medical waste according to EPA is “all waste
Hazardous Wastes materials generated at health care facilities, such as hospitals, clin-
and Their Treatment ics, physicians’ offices, dental practices, blood banks, and veteri-
nary hospitals/clinics, as well as medical research facilities and
laboratories.” Among the medical waste, pharmaceutical waste is
the most prominent and perilous ones. The USA has spent around
$2.5 billion for the disposal of medical waste in 2012. Interestingly,
with annual growth of 4.8%, by 2017, the increased cost is expected
to $3.2 billion [91]. For instance, just hospitals in the USA pro-
duce more than 5.9 million tons of waste annually. Therefore, one
can imagine the level of hazardous intensity generated by medical
waste all over the world as all healthcare activities considered to
humans generated medical wastes. The danger increased to mani-
fold by mishandling or improper disposal of these medical wastes.
Therefore, persons engaged to proper risk assessment and manage-
ment must be aware with types of medical wastes especially phar-
maceutical ones along with different approaches to treat them
efficiently to minimize the hazards to environment and living sys-
tems. A typical list of pharmaceutical hazardous waste with few
examples is illustrated in Fig. 2 and most commonly employed
treatment for pharmaceutical as well as medical waste to avoid high
risk of ecotoxicity is reported in Fig. 3.
3 Environmental Risk Assessment of Pharmaceuticals
Risk assessment is the process of assessing the concentration,

occurrence, and level of environment and human exposure of a
pharmaceutical product [92]. The key aim of environmental risk
assessment (ERA) should be risk mitigation and risk management.
The conclusion of the ERA report should be based on scientific
Fig. 2 Types of pharmaceutical hazardous wastes with few examples [Color of the boxes for the medical waste
represents the color of the waste container]
Fig. 3 Different ways of treatment for medical wastes to avoid high risk of ecotoxicity
reasoning supported by adequate ecotoxicity studies. The outline

of the registration process and the ERA consist of European
Commission and Council directives and regulations on registration,
European policy, case law, and global (trade) agreements.
3.1 Risk Assessment The most commonly employed risk assessment approaches of
Approaches pharmaceuticals and their metabolites in various environment
compartments are discussed below [93].
3.1.1 Hazard The first and foremost step for risk assessment is the identification
Identification of source and occurrence of hazards which supports the intensity
of risk of a pharmaceutical. Majority of scientists highly relied on
in vivo data, but due to huge deficiency of reasonable data for
majority of pharmaceuticals related to specific species and definite
environment compartment, greater effort should be offered on the
efficient usage of in vitro assays and in silico analysis, as well as the
use of computational techniques in systems biology [94].
3.1.2 Dose- Detection of threshold dose of the toxic effect is imperative for
Response scientific risk assessment of any hazards. Dose-response informa-
Assessment tion over a wide range of test concentrations should be performed
through quantitative high throughput screening (q-HTS).
Additionally, sensitive assays should be able to detect toxicity at
very low doses or below environmental levels experienced by living
organisms. There should be sufficient scope available to extrapo-
late adversarial responses and to assess critical concentrations data
employing statistical approaches [95].
3.1.3 Dose and Species Major drawbacks of risk assessment are low-dose toxicity and lack
Extrapolation of interspecies extrapolation data. In some cases, regulatory
authorities and government organizations have implemented in
silico models and expert systems as alternatives to deal with these
problems. In vitro to in vivo extrapolation and physiologically
based pharmacokinetic (PBPK) models are agreeable to sensitivity,
variability, and uncertainty analysis using conventional tools [96].
3.1.4 Risk The final phase of the ecological risk assessment is the risk charac-
Characterization terization which integrates the analyses from the exposure and eco-
logical effects characterization along with the doubts, hypothesis,
strengths and limitations of the analyses. The risk characterization
has two major components: risk estimation and risk description.
Again, risk estimation compares integrated exposure and effects
data in context of Levels of Concern (LOCs) and states the poten-
tial for risk [97].
3.1.5 Deterministic The EPA uses a deterministic approach or the risk quotient (RQ)
Approach and Calculation to evaluate toxicity to environmental exposure which is calculated
of Risk Quotients by dividing a point estimate of exposure by a point estimate of
effects. Calculation of RQ are based upon ecological effects data,
hazards use data, fate and transport data, and estimates of exposure
to the hazards. Thus, the estimated environmental concentration
(EEC) is compared to an effect level, such as an LC50 (the concen-
tration where 50% of the organisms die.)
RQ = Exposure/Toxicity
3.2 Environmental The risk assessment model considers the safety issues and RQ of
Risk Assessment individual pharmaceutical products. The most common approaches
Modeling are offered in the guidance for environmental assessments for reg-
of Pharmaceuticals ulatory drug approvals by the US FDA [28] or by the European
Medicines Agency (EMA) [26]. It is important to evaluate expo-
sure of any pharmaceutical by the following ways previous to model
a toxicological study [98]:
(a) For modeling purpose, the exposure is assessed in the form of
occurrence or the environmental concentration to which the
biological system is exposed along with the duration and fre-
quency being not on the concentrations to which individual
living system is exposed. Exposure is also dependent on many
miscellaneous factors such as sorption effects, metabolism,
transformation processes, and fate.
(b) The life cycle of any organism must be taken into account for
understanding the effect of pharmaceuticals on them.
(c) The MOA of pharmaceuticals needs to be determined to
depict each step of molecular and functional effects.
(d) Proper understanding of the pathways and target sites of phar-
maceuticals in the biological system.
(e) The bioavailability and toxicokinetic properties of the pharma-
ceutical need to be studied.
(f) Complete pharmacokinetic and pharmacodynamic informa-
tion are required to understand the absorption, distribution,
metabolism, excretion, and toxicity pattern.
(g) The hazard generated from inherent toxicity of the pharma-
ceuticals according to their chemical properties is needed to be
studied.
The most important steps for risk assessment and management
process are reported in Fig. 4.
Fig. 4 Reasons for ecotoxicity study along with steps for risk assessment and risk management due to phar-
maceuticals hazards
4 Environmental Risk Management of Pharmaceuticals
The risk management can be defined as follows: “the process of

identifying, evaluating, selecting, and implementing actions to
reduce risk to human health and to ecosystems. The goal of risk
management is scientifically sound, cost-effective, integrated
actions that reduce or prevent risks while taking into account
social, cultural, ethical, political, and legal considerations” [99].
The process of risk management caused by pharmaceuticals
hazards must be balanced with cost benefit and practical to
implement.
4.1 Accomplishment Pharmaceuticals are one of the must have emergency products which
of Preventive cannot be stopped for use but the possible risks of them related to
Measures environmental can be managed by executing apposite preventative
measure and safeguard. A set of guidelines has been set by the EMEA
in 2006 as safety measures for risk management:
1. Initial assessment of risk for individual products,
2. Each pharmaceutical package should have appropriate product
labeling and summary product characteristics (SPC),
3. Educated patients about the possible toxicity toward humans
as well as environment through Package leaflet (PL),
4. Safe and appropriate storage as well as disposal of pharmaceuti-
cal products,
4.2 Minimizing To reduce the input of pharmaceutical products and their metabo-
the Input of lites, the following steps can be employed effectively.
Pharmaceutical
Hazards into the
Environment
4.2.1 Awareness Awareness and training about occurrence and effect of individual
and Training pharmaceutical products along with their corresponding effects
toward environment is the most crucial step. In addition, knowl-
edge about disposal process of diverse types of pharmaceutical
hazards is the first step to reduce the input of those hazards
into the ecosystem. The awareness need to be spread among
the shareholders, stakeholders and community using the pharma-
ceuticals, including patients, doctors and nurses, and pharmacists.
The most important role need to be played by industries as they
are the major source of pharmaceutical hazards and many of them
are APIs when they are released into the environment without
adequate waste treatment. In addition, each raw material, in pro-
cess molecules and API should consist of material safety data sheets
(MSDSs) intended to provide workers and emergency personnel
with process for handling that product safely with information like:
physical data, toxicity and health hazards, first aid, reactivity, stor-
age, disposal, protective equipment, and spill-handling procedures.
People related to risk management should possess information
about the drug flows from the diverse sources of households,
industries, hospitals and pharmacy [100].
4.2.2 High-End Most of the risk management procedures can be controlled with
and Advanced Sewage improvement of sewage treatment. Implication of sophisticated
Treatment and enhanced waste water as well as sewage treatment can diminish
the hormonal effects to living systems, ecotoxicity and pathogenic
effects of the effluent to manifold. Recently, advanced effluent sew-
age treatment has been practiced comprehensively and performed
employing photochemical oxidation, filtration, and adsorption
processes [101].
4.2.3 Green and Viable The final approach is the knowledge of green and sustainable phar-
Pharmacy macy which supports environmentally benign compounds which
after coming into the contact with environment will be degraded
with minimum hazardous effects in no time [100]. In the present
scenario, it is the least practiced methods, but in long term of sus-
tainability, it is the need of hour.
Furthermore, possible measures, roles and action to reduce
the ecotoxicity imposed through pharmaceutical hazards by
diverse stakeholders are addressed in Table 1 for improved
understanding.
Table 1
Roles of different stakeholders to reduce the pharmaceutical induces ecotoxicity
Stakeholders Possible measures, roles and action

Doctors • Prescribed only required medicine.
• Awareness about their hazardous properties.
Pharmacist • Awareness and information about usage and disposal to patients.
• Take back system of unused medicine if possible.
Patients • Improvement of compliance and proper disposal.
• Consumption of medicine only if prescribed by a doctor.
Hospitals • Amalgamation of the delivering pharmacy/wholesaler about the
handling of expired medicaments.
• Proper implementation of rules and regulation by hospital authorities
for the disposal of pharmaceuticals and other medicaments.
Industries • Periodical report of environmental assessment related data of individual
pharmaceuticals along with raw materials, in process ingredients.
• Complete information about analytical methods and results.
• Proper packaging and labeling with proper uses, storage and disposal.
• Improved drug delivery systems so that smaller doses are needed.
• Environment friendly packaging with extended shelf life of packing
material.
Drinking water • Extended monitoring of water for hospital, industry and pathological
centers.
• Advanced treatment of water
• Suitable approach to complete or up to acceptable limit removal of
waste.
Waste water treatment • Tertiary treatment methods such as ozonation, activated carbon
adsorption, or nanofiltration.
• Separate and careful piping between waste water and rain water.
Authorities • Instigation and back up of communication between all stakeholders.
• Implementation of threshold limits for each pharmaceuticals and other
medicaments for different environmental compartments.
Policies and regulations • Annexation of every APIs and formulation product in environmental
legislation.
• Updated regulation for the management of out dated and new as well as
existing medicaments.
Green pharmacy • Fast and trouble free degradability of pharmaceutical products and
supplements.
• Improvement of synthesis and renewable feedstock for preparation of
environment friendly pharmaceuticals.
Table 2
A list of endpoints for modeling purpose under OECD and areas where QSAR models can be employed
Endpoints for modeling under OECD Areas where QSAR models can be used
• Physical-chemical properties: Boiling point, • Prioritization of existing pharmaceuticals
melting point, vapor pressure, octanol–water for toxicity to environment
partition coefficient, organic carbon–water partition • Classification and labeling of new
coefficient and water solubility pharmaceuticals
• Ecological effects on endpoints: long-term toxicity, • Risk assessment of new and existing
acute Daphnia toxicity, Acute fish toxicity, terrestrial pharmaceuticals
toxicity, algal-toxicity, marine organism toxicity, • Guiding experimental design of regulatory
microorganism toxicity in sewage treatment plant tests or testing strategies
• Environmental fate: Biodegradation, hydrolysis in • Providing mechanistic information
water, atmospheric oxidation, and bioaccumulation • Filling up the large data gaps
▪ Human health effects: Acute oral, acute inhalation, • Building a proper database of each
eye irritation, acute dermal, skin sensitization, skin pharmaceutical to different species
irritation, repeated dose toxicity, reproductive regarding ecotoxicity
toxicity, genotoxicity, systemic toxicity, • Development of expert systems for each
developmental toxicity, mutagenicity, carcinogenicity, therapeutic class for diverse compartments
etc. of the environment
• Construction of efficient interspecies
models to extrapolate data from one
species to another species when data of a
specific species is missing
5 Global Regulatory Bodies Concerning Pharmaceuticals Hazard and Risk

Quantification
Increasing exposure of pharmaceutical wastes in the environment

is an affair of anxiety and a hot topic worldwide. The risk of
pharmaceutical hazards are directly related to the environment and
indirectly related to human health to a great extent. There is a
strong need to predict physicochemical properties, environmental
fate, effects of pharmaceuticals and their metabolites, as concerned
experimental data for different compartments of the environment
are absent for huge number of pharmaceuticals till today. A good
number of government regulatory authorities related to environ-
ment safety consider in silico approaches like structure–activity
relationship (SAR) and QSAR to predict the hazardous effect and
fate of untested and newly introduced pharmaceuticals as fast as
possible with economical way using less animal testing [26, 28,
102–106]. To predict the human health or environmental hazards
due to exposure of pharmaceuticals, in silico models are utilized by

the OECD and the developed models are organized as searchable
databases which are intended for providing risk assessment and
management resources. The models are good sources as screening
tools for pharmaceutical databases when there is missing of
chemical-specific data for establishing priorities for risk assessment
and for evaluating issues of potential concern [102]. Commonly
Table 3
Global regulatory agencies which deal with the environmental risk assessment and risk management
of pharmaceuticals
Regulatory
agencies Objective Responsibility and method of risk assessment
AEA (Australian Advises clients on the AEA has undertaken reports for the Department
Environment environmental hazards and of Sustainability, Environment, Water,
Agency) potential risks associated Population and Communities (DSEWPaC),
with the production, use particularly with respect to their environmental
and disposal of chemicals. assessments performed on new and existing
AEA is a member of the agricultural and veterinary chemicals for the
Society of Ecotoxicology Australian Pesticides and Veterinary Medicine
and Chemistry (SETAC) Authority (APVMA), and industrial chemicals
for the National Industrial Chemicals
Notification and Assessment Scheme
(NICNAS).
CDER (Center for CDER reviews New Drug An assessment of risk to the environment is
drug evaluation Applications to ensure that required for manufacture, use and distribution of
and research) the drugs are safe and human drugs under the National Environment
effective. Its primary Policy Act of 1969 and an environmental
objective is to ensure that all assessment procedure was developed by the US
prescription and over-the- Food and Drug Administration (US FDA) as a
counter (OTC) medications part of the registration procedure for new human
are safe and effective when pharmaceutical drugs. Additionally, in 1995, the
used as directed. FDA-CDER issued a new guidance for the
Submission of an Environmental Assessment in
Human drugs. In 1997 the FDA implemented a
Note for Guidance paper in which all drugs
entering the aquatic compartment at levels below
1 μg/L Predicted Environmental Concentration
(PECEFFLUENT) were exempted from a detailed
risk assessment.
(continued)
Table 3
(continued)
Regulatory
EMEA (European EMEA exhibits the scope and Environmental risk assessment is divided into
agency for the legal basis for risk assessment three phases:
evaluation of of pharmaceuticals and (a) Phase I: Pre-screening and estimation of
medicinal outlines the general exposure based on the drug only, irrespective
products) considerations and the of its route of administration, pharmaceutical
recommended stepwise form, metabolism and excretion
procedure for their risk (b) Phase II Tier A: Screening and initial
assessment. The guideline prediction of risk where all relevant data should
considers the specific features be taken into account, e.g., data on physical-
of pharmaceuticals, e.g., the chemical properties, primary and secondary
use of available pharmacodynamics, toxicology, metabolism,
pharmacological information. excretion, degradability, and persistence of the
Previously environmental risk drug substance and/or relevant metabolites
assessments performed mainly (c) Phase II Tier B: Extended and substance and
on acute ecotoxicity data, but compartment-specific risk assessment.
in recent time EMEA draft Information from the refined data set is
has proposed to include available comprising information on route(s)
pharmacokinetic and of excretion; and qualitative and quantitative
pharmacodynamic data for information on excreted compounds, and
environmental risk assessment. possibly additional long-term toxicity data
EU-CSTEE The CSTEE has identified the QSAR is the first step in gaining more general
(European need for a proactive knowledge on the risk assessment issue as an
Union approach in obtaining data alternative to nonanimal method. In contrast
Commission’s on the environmental effects to the amount of analytical data, information
scientific of pharmaceuticals. Thus, it about the ecotoxicological effects of drug
committee on is recognized that a residues is scrubby. To create a broader basis
toxicity, prioritization procedure for the evaluation of the ecotoxicological
ecotoxicity and needs to be developed for relevance of pharmaceutical compounds,
environment) environmental risk proper documentation of their effects and the
assessment of reason are identified and documented.
pharmaceuticals, and that
this should follow the
general scheme for chemicals
described in the White Paper
for future EU chemicals
policy i.e., REACH
guideline.
(continued)
Table 3
(continued)
Regulatory
MHLW (Ministry The MHLW constructed a The risk assessment is judged by the PEC/
of Health, research group to build up a PNEC (Predicted Environmental
Labour and concept on the regulation of Concentration/Predicted No Effect
Welfare of Japan) pharmaceuticals for Concentration) ratio or ΣPECi/PNECi. In
environmental safety in addition, the Organization for Pharmaceutical
2007. The regulation system Safety and Research (OPSR) conducted
is similar to that of general compliance reviews on application data. This
chemicals in Japan and the was followed by the integration of the
Guideline by EMEA. The aforementioned Evaluation Center, OPSR,
main function of this group and part of the Medical Devices Center to
is to establish a risk-benefit form a new independent administrative
analysis committee for the organization, the Pharmaceutical and Medical
pharmaceuticals which have Devices Agency (PMDA). The MHLW and
a high risk for environmental PMDA handle a wide range of activities from
organisms and to human clinical studies to approval reviews, reviews
health. throughout post-marketing stage, and
pharmaceutical safety measures.
NICNAS NICNAS was established in The major responsibility of NICNAS are:
(National July 1990 under the • Assessing new industrial chemicals for human
Industrial Industrial Chemicals health and/or environmental effects
Chemicals (Notification and • Maintaining the Australian Inventory of
Notification and Assessment) Act 1989 by Chemical Substances (AICS)
Assessment Australian Government • Circulation of information on the human
Scheme) Department of Health. A health and environmental impacts of chemicals
range of state, territory and and recommending on their safe use
Commonwealth government • Registering new industrial chemicals
agencies share regulatory
responsibility for chemical
safety in Australia, with each
chemical being regulated
according to its use, whether
as a pharmaceuticals,
veterinary medicine,
pesticide, food additive or
industrial chemical.
(continued)
Table 3
(continued)
Regulatory
REACH Aims to improve the “No data no market”: the REACH Regulation
(Registration, protection of human health places responsibility on industry to manage
Evaluation, and the environment the risks from chemicals and to provide safety
Authorisation through the better and information on the substances. Manufacturers
and Restriction earlier identification of the and importers of substances have a general
of Chemicals) intrinsic properties of obligation to submit a registration to the
chemical substances. European Chemicals Agency for each
substance manufactured or imported in
quantities of 1 tonne or more per year per
company
SECIS (Swedish An authorized regulatory body To improve risk management decision making,
Environmental which was initiated in 2005 sufficient knowledge about environmental
Classification by the Swedish Association exposures and effects in nontarget species for
and Information for the Pharmaceutical all relevant pharmaceutical substances is
System for Industry. The rationale of needed. Within SECIS, the pharmaceutical
pharmaceuticals) the classification system is to companies provide environmental data and
offer the public and health classify their products according to predefined
care sectors with criteria and a guidance document. The
environmental information guidance document is developed for the
about all active purposes of SECIS, but it is based on the
pharmaceutical ingredients European Medicines Agency (EMA) guideline
(API) on the Swedish for environmental risk assessment of
market up to now. pharmaceuticals and the European Commission
Technical Guidance Document (TGD).
UBA (Federal The German Medicines Act The UBA already assessed around 180 veterinary
Environment provides that the UBA is and around 240 human pharmaceutical
Agency) responsible for the formulations. Filtering concepts established
environmental risk between UBA and the authorization agency
assessment. The UBA responsible for veterinary medicines focused
started assessing the the ERA on antibiotics, parasiticidal
environmental impact of substances and analgesics. Cytostatic
veterinary and human medicines, hormones and contrast agents
pharmaceuticals in an dominated the human medicine dossiers
authorization routine in assessed by UBA.
1998 and 2003, respectively.
VICH VICH is a trilateral (EU– Veterinary medicinal products (VMPs) are
(International Japan–USA) program aimed regulated for environment safety as described
Cooperation on at harmonizing technical in Environmental Impact Assessment for
Harmonization requirements for veterinary VMPs; Phase I in 2000 and Phase II in 2004.
of Technical product registration was
Requirements officially launched in April
for Registration 1996. The initiative to begin
of Veterinary the harmonization process
Medicinal came about in 1983 when
Products) the first International
Technical Consultation on
Veterinary Drug Registration
(ITCVDR) was held.
Fig. 5 Areas of risk assessment and modeling for ecotoxicity prediction as stated to the OECD
Fig. 6 Category of information included in predicting health and environmental effects according to the OECD
guidelines
used endpoints for ecotoxicity modeling and areas where QSAR

models can be employed for risk prediction are reported in Table 2.
Regulatory bodies responsible for the risk identification, assess-
ment and management of pharmaceuticals ecotoxicity across the
globe are listed in Table 3. Areas for risk assessment and grouping
of information for risk prediction of human health and environment
according to OECD’s guidelines are illustrated in Figs. 5 and 6,
respectively.
6 In Silico Modeling in Ecotoxicity of Pharmaceuticals
Computational methods intend to harmonize in vitro and in vivo

toxicity tests to potentially curtail the necessity for animal testing,
reduce the cost and time of experiments, and improve toxicity pre-
diction and safety assessment. Additionally, computational
approaches have an exclusive benefit of being competent to approxi-
mate chemicals for its toxicity even before they are synthesized
[102]. With increasing concern about the ecotoxicity and human
health, the storage, distribution and release of pharmaceuticals after
their usage to the environment are controlled and regulated at vari-
ous levels by governments and regulatory bodies. Applications of
assortment of in silico tools are very much constructive option to
provide sufficient information in a regulatory decision making
Fig. 7 Most common in silico tools for the prediction of pharmaceuticals ecotoxicity
context in the absence of experimental data [102]. Most commonly

employed predictive in silico tools are depicted in Fig. 7.
Among the available in silico predictive models for ecotoxicity,
majority of the models are generated employing QSAR techniques,
for toxicity prediction some time termed as quantitative structure–
toxicity relationship (QSTR). Along with the typical QSAR/QSTR
models, toxicophore or structural alerts, read-across, chemical ana-
logue, trend analysis and docking approaches are employed in
many successful prediction researches. The QSTR approach
attempts to correlate structural/molecular properties (x1, x2, …,
xn) with toxicity response (Y), for a set of molecules by means
of statistical methods [107], generating simple mathematical
relationship as follows:
Y = f ( x 1 ,,,x 2 ,,, ¼,,,x n ) (1)

The major objective of QSAR/QSTR modeling is to examine and
recognize the influential factors for the measured activity/toxicity
for a particular system, to have an insight of the mechanism and
behavior of the studied system. This strategy generates a mathe-
matical model which joins experimental measures with a set of
chemical descriptors determined from the molecular structure for
the studied compounds. The constructed model should possess
good predictive abilities in order to predict the studied response
for untested or future compounds. The factors leading the events
in a biological system are depicted by a multitude of physicochemi-
cal descriptors [107]. From its initial days, the QSAR approach has
come a long way. Along with the time, new and modified methods,
algorithms and dimensions (1D to 7D) have been applied in QSAR
studies are discussed elsewhere [36, 37]. Goodness-of-fit and
prediction quality is two most important features for an acceptable
and reliable QSAR model. The predictive quality of the QSAR
model is checked through different validation statistics and met-
rics. Thus, validation of QSAR models is the major step along with
defining the applicability domain for the prediction of untested
and future compounds [108, 109].
6.1 Why In Silico ● The 3Rs concept signifies “Reduction,” “Replacement,” and
Models Required? “Refinement” regarding animal experimentation in scientific
experiments. “Reduction” defines to the lessening the number
of animals used to get precise results, “Replacement” corre-
sponds to the implication of nonliving resources to substitute
conscious living higher animals, and “Refinement” suggests
turn down the severity or cruelty of inhuman methodologies
applied to the experimental animals [110]. Thus to set up the
3Rs concept, in silico techniques are one of the best options
available. The European Centre for the Validation of Alternative
Methods (ECVAM) was established in the year 1991 that

agrees with the principle of 3Rs.
● The ban of animal experimentation by regulatory agencies and
government organizations initiated the need of molecular
modeling approach [111]. Council Directive 86/609/EEC
on the approximation of Laws, Regulations, and Administrative
provisions of the member states regarding the protection of
animals used for experimental and other scientific purposes.
The testing ban on the finished cosmetic products applies since
11 September 2004; the testing ban on ingredients or combi-
nation of ingredients applies since 11 March 2009. The mar-
keting ban applies since 11 March 2009 for all human health
effects with the exception of repeated-dose toxicity, reproduc-
tive toxicity, and toxicokinetics. For these specific health
effects, the marketing ban applies since 11 March 2013, irre-
spective of the availability of alternative nonanimal tests [112].
● Regulatory decision making through SARs and QSARs models
for predicting aquatic toxicity, physicochemical parameters and
environmental fate properties.
● Filling data gaps for ecotoxicity due to pharmaceuticals haz-
ards as acceptable toxicity data of pharmaceuticals to environ-
ment and human health is <5% [113]. In silico prediction has
the proficiency to help out in the prioritization of pharmaceu-
ticals for testing, and for predicting ecotoxicity to allow for
classification. Computer models are a reliable source for toxic-
ity predictions as they can be used as one of the significant
sources for filling the missing data of ecotoxicity.
● Understanding the real mechanism of action for each pharma-
ceutical for specific endpoints and environment compartment
system. For many modeling approaches, it may be assumed
that molecules fitting the similar QSAR models are acting by
the same MOA [114].
● Cost and time saving are two major reasons for the use of in
silico approaches. Even a simple ecotoxicological assay may
cost several thousand dollars. On the contrary, early toxicity
prediction can save a good amount of time and money [115].
A schematic representation is provided in Fig. 8 showing the
requirement of in silico modeling evaluating pharmaceuticals
ecotoxicity.
7 Successful In Silico Models for Ecotoxicity Prediction of Pharmaceuticals
Sanderson et al. [116] provided a baseline to fill the screening data

regarding API environmental toxicity utilizing the US EPA generic
aquatic QSAR model ECOSAR for screening of more than 2800
Fig. 8 The need of in silico modeling assessing the impact of pharmaceuticals

ecotoxicity
pharmaceuticals and. The model can be successfully used to pre-

dict both acute and chronic aquatic toxicity. In another work, toxic
potential of mixtures of the β-blockers and related metabolites are
modeled for the phytotoxicity endpoint by Escher et al. [117]. For
performing the modeling, they assumed two conditions; first, the
metabolites lose their definite response and act as baseline toxi-
cants and second, the metabolites expose the identical specific
MOA like their parent drug. The authors accounted experimen-
tally determined liposome–water partition ratios at pH 7 as inde-
pendent variable for correlating with the response variable to make
the QSAR analysis more reliable.
Sanderson and Thomsen [118] overestimated the toxicity for
70% of the 59 pharmaceuticals and more than 94% underestimated
toxicity predictions by less than a factor of ten for the remaining
30% pharmaceuticals by ECOSAR v3.20. The authors have
reported correlation coefficients ranging from 0.73 to 0.76
between all the modeled Log EC50 and Log Kow. The slopes of the
Log EC50-Log Kow regression based on measured data from the
US National Oceanic and Atmospheric Administration (NOAA)
database for both fish and daphnia equal to −0.86 which suggested
a narcotic MOA. In another study, acute toxicity was predicted

(>92%) employing a QSAR model developed by Sanderson and
Thomsen [45] suggesting a narcotic MOA of 275 pharmaceuti-
cals. Their analysis suggested 68% of the pharmaceuticals have a
nonspecific MOA based on model prediction error. The authors
have also compared the measured effect data to the predicted effect
concentrations utilizing ECOSAR regarding the predictability of
ecotoxicity of pharmaceuticals and accurate hazard categorization
relative to Global Harmonized System (GHS). Pharmaceuticals
were predicted employing the model resulting in 71% algae, 74%
daphnia, and 83% fish datasets that could be compared.
One of the first interspecies QSAR models for 77 pharmaceu-
ticals’ ecotoxicity was reported by Kar and Roy [119] for Daphnia
magna and fish endpoints. Analyzing the interspecies models, the
authors have reported that the keto group and the (aasC)
fragment are predominantly accountable for higher toxicity of
pharmaceuticals to D. magna. On the contrary, along with the
keto group, structural fragments like X=C=X, R–C(=X)–X, and R–
C≡X are principally responsible for fish toxicity. The interspecies
models were also employed to predict fish toxicities of 59 pharma-
ceuticals and Daphnia toxicities of 30 pharmaceuticals when
Daphnia and fish toxicity data were present, respectively.
Christen et al. [120] developed VirtualTox Lab [121] for the
prediction of pharmaceuticals’ effect in the aquatic system. The
study guides to the conclusion that the MOA perception is most
suitable for the classification of highly active compounds (HC).
The authors also reported that modification could be performed
by balancing this concept utilizing the QSAR model (VirtualTox
Lab), whereas the fish plasma model appeared to be less appropri-
ate due to the requisite of environmental concentration above
10 ng/L for the identification of a risk. The Virtual-ToxLab can
support the MOA concept and can be advantageous to distinguish
surplus targets of the pharmaceutical to assess the ecotoxicity.
Das et al. [122] reported interesting interspecies correlation
models using rodent toxicity as dependent variable and fish, daph-
nia and algae toxicity data as independent variables separately for
194 pharmaceuticals. All interspecies extrapolation QSAR models
were generated using multiple statistical tools. Explaining the
models, the authors concluded that heteroatom count and charge
distribution were noteworthy parameters of the rodent toxicity
along with the atom level logP contributions of various structural
fragments and a mixture of extended topochemical atom (ETA)
indices reflecting electronic information and branching pattern of
molecules. The authors also concluded that atom level logP contri-
butions of dissimilar fragments, charge distribution, shape and
ETA parameters were imperative in describing the daphnia and fish
toxicities in the interspecies correlation models with algae toxicity.
Interestingly, the toxicity of pharmaceuticals to rodents bears
minimum interspecies correlation with other mentioned nonverte-

brate and vertebrate toxicity endpoints.
De García et al. [123] performed the environmental risk assess-
ment of 26 pharmaceuticals and personal care products (PPCPs)
based on the ecotoxicity data tested through bioluminescence and
respirometry assays. This was followed by classification and label-
ing of pharmaceuticals by the GHS using the US EPA ecological
structure–activity relationship (ECOSAR™). The risk impact of
these pharmaceuticals in WWTPs and in the aquatic environment
was predicted following the criteria of the EMA. According to
their two ecotoxicity tests, 65.4% of the PPCPs showed prominent
toxicity to aquatic organisms. Pharmaceuticals like
1,4-benzoquinone, ciprofloxacin, acetaminophen, clofibrate, clar-
ithromycin, omeprazole, ibuprofen, triclosan, and parabens
showed risk threat for aquatic environments and/or for activated
sludge of WWTPs according to the performed analysis.
Sangion and Gramatica proposed [39] a screening approach to
evaluate the potential PBT of around 1200 pharmaceuticals
employing two different QSAR models. The authors applied the
Insubria-PBT-Index, a Multiple Linear Regression (MLR) QSAR
model based on simple molecular descriptors, implemented in
QSARINS software. An agreement of 86% was reported between
the two models and a priority list of 35 pharmaceuticals, high-
lighted as potential PBTs by consensus, was suggested for addi-
tional experimental validation. The models can be useful in the
hazard assessment, performing preliminary screening and prioriti-
zation of pharmaceuticals, mainly associated with the potential
PBT behavior of the prioritized pharmaceuticals.
Externally validated QSAR models, specific to predict acute
toxicity of APIs in three aquatic trophic levels endpoints, i.e., algae,
Daphnia and two species of fish were developed using the
QSARINS software by Sangion and Gramatica [124]. The devel-
oped MLR-ordinary least squares (MLR-OLS) models were devel-
oped with theoretical molecular descriptors computed through
PaDEL-Descriptor software. The selections of descriptors was per-
formed by Genetic Algorithm (GA). The generated models were
employed further to predict acute toxicity for a large set of APIs
without experimental data by Principal Component Analysis
(PCA). Further, a trend was set by the combination of toxicities for
all the studied organisms and the trend is termed as Aquatic
Toxicity Index (ATI) which allowed the raking of pharmaceuticals
according to their potential toxicity upon the complete aquatic
environment. Not only that, the accuracy of the models was com-
pared with the accuracy of the frequently used software ECOSAR,
and the authors concluded that their models showed better
performances.
Sangion and Gramatica [125] proposed quantitative activity-
activity relationship (QAAR) models, implemented in QSARINS
by theoretical molecular descriptors to enhance the quality and

predictivity of the interspecies relationships between toxicity
toward Daphnia magna and two fish species, Pimephales promelas
and Oncorhynchus mykiss. The authors claimed that the developed
invertebrate-fish interspecies models could reduce the composite
experimental tests on upper trophic organisms and reduce animal
experiments. They also illustrated that the Daphnia could serve as
a surrogate for fish toxicity and the fish-fish intercorrelations could
be used for evaluating toxicological data when sufficient informa-
tion is unavailable.
Successful in silico models, especially QSAR models on eco-
toxicity of pharmaceuticals were discussed in this section. There is
no doubt that reported number of models is quite low compared
to that of chemical toxicity models. One of the main reasons for
this is the limited experimental data on ecotoxicity of pharmaceu-
ticals. One has to understand that it is impossible to study toxicity
of each pharmaceutical in different species in diverse compart-
ments. Thus, QSAR models will provide the predicted values when
experimental data are absent for specific pharmaceutical. On the
other point, making of more interspecies models is the need of
time to extrapolate toxicity data from one species to another.
8 Endpoints
To generate ecotoxicity data of pharmaceuticals, they should be

assayed employing specific endpoints which are sometime called as
test batteries. The evaluated data are rich source of information for
making ecotoxicity database as well as for making in silico models
and expert systems. Thus a clear understanding is essential about
the ecotoxicity endpoints as they are important to make in silico
models as well as for deeper understanding about the mode of
toxicity of a specific drug for a definite endpoint. With extensive
literature survey, we have enlisted most frequently employed end-
points by the scientist community in Table 4.
9 Databases
The first condition to develop a reliable, accurate, and reproduc-

ible in silico model is to use good quality ecotoxicological data of
pharmaceuticals and their metabolites in diverse environmental
compartments with different concentration. A significant number
of toxicity databases toward environment are accessible to public
and the numbers are increasing with time. Although according to
Table 4
Toxicity endpoints for the modeling of pharmaceuticals ecotoxicity
Endpoints Species Depiction Implication

Algae Chlorella vulgaris Unicellular fresh-water Study the toxic action of organic
Chlorella pyrenoidosa green microalgae compounds
comprises a major part
of phytoplankton
Pseudokirchneriella One of the prime Ecotoxicity is measured by growth
subcapitata producers of the rate inhibition of green algae P.
(Selenastrum aquatic ecosystem and subcapitata
capricornutum) ideal test organisms
for toxicological
studies.
Scenedesmus obliquus Common cosmopolitan Ability to grow in industrial
green algae, often wastewaters of different origins
occurring as almost a showing good adaptation ability
pure culture in fresh and versatile microalgae as toxicity
water plankton. test endpoint
Scenedesmus Green algae of the Employed in the prediction of
vacuolatus Chlorophyceae. photoinduced toxicity of polycyclic
Colonial and aromatic hydrocarbons and ionic
nonmotile in nature. liquids (ILs)
Bacterium Escherichia coli A Gram-negative, E. coli is used as a model organism in
facultative anaerobic, ecotoxicity modeling studies. For
rod-shaped bacterium example, it is used to test
of the genus cytotoxicity of metal oxide
Escherichia. nanoparticles
Vibrio fischeri, Vibrio Gram-negative rod- Employed in the research of
natriegens shaped bacterium microbial bioluminescence,
having bioluminescent quorum sensing along with
properties and found ecotoxicity testing for diverse
predominantly in range of organic chemicals and
symbiosis with various pharmaceuticals
marine animals
Bacillus A genus of gram- Chlorophenol toxicity is tested on
positive, rod-shaped bacillus species.
bacteria and member
of the phylum
Firmicutes
Pseudomonas P. fluorescens has versatile Can be employed for modeling of
fluorescens metabolism. Generally antibiotics toxicity and resistance
found in the soil and studies
in water.
(continued)
Table 4
(continued)

Crustaceans Daphnia magna, Small planktonic Invertebrate species in aquatic food
Daphnia pulex, crustacean and one of webs has been used as a
Daphnia ambigua, the small aquatic representative test species for
Daphnia melanica crustaceans commonly ecotoxicological evaluation of
called water fleas. diverse organic chemicals using
Most commonly immobilization test
employed species is D.
magna.
Thamnocephalus A family of crustaceans 24 h toxicity test employed for
platyurus with wide distribution screening of pure compounds,
including Western effluents, sediments, surface and
Australia and ground waters, wastewaters, and
Southern Africa. biotoxins
Duckweed Lemna minor One form of aquatic Used in modeling of phytotoxicity of
vascular plant ILs and growth inhibition test of
(duckweed) known as duckweeds
thallus, which floats
Lemna gibba Used in testing the phytotoxicity of
on the surface of the
pesticides and other environmental
water.
chemicals to higher plants.
Enzyme Acetylcholinesterase Catalyzes the hydrolysis Enzyme inhibition data of the
of acetylcholinesters acetylcholinesterase from
with a relative Electrophorus electricus, the AMP
specificity for deaminase and the antioxidant
acetylcholine in enzyme system of mouse liver are
autonomic nervous important for toxicity prediction
system function.
Fish Channel Catfish CCO is the cell line of Standard for diagnosing Channel
Ovary (CCO) choice for the Catfish Virus Disease (CCVD) in
propagation and farm reared Channel Catfish.
diagnosis of Channel Prediction of ILs has also been
Catfish Virus (CCV). performed
Zebrafish (Danio A tropical freshwater fish Standardized under the OECD and
rerio) belonging to the employed to test chemicals and
family Cyprinidae pharmaceuticals
Fathead minnow Pimephales promelas is Studied to examine the effects of
(Pimephales the EPA waste materials on the aquatic life.
promelas) recommended Effects induced by progestins are
vertebrate species for largely studied employing fathead
freshwater chronic minnow.
toxicity tests.
(continued)
Table 4
(continued)

Mammalian Human keratinocyte HaCaT cells are a Used for studies of skin biology and
cells cell line (HaCaT) spontaneously cytotoxicity assessment of metal
immortalized, human oxide
keratinocyte line.
CaCo-2 Heterogeneous human Permeability coefficients across the
epithelial colorectal cellular membranes of Caco-2 cells
adenocarcinoma cells. are generally employed for
modeling
HeLa A prototypical cells of Derived from cervical cancer cells
the human epithelium and largely employed for
used in scientific anticancer activity
research.
Prostate cancer cell A human prostate cancer Used in modeling of prostate cancer
line (PC3) cell lines inhibitors
Human malignant Derived from a lymph Modeling of anticancer drugs.
melanoma (Fem-X) node metastasis of a
melanoma patient
HT-29 A human colorectal Sensitive to the chemotherapeutic
adenocarcinoma cell drugs used in standard treatment
line with epithelial for colorectal cancer
morphology
Rat cell line—IPC-81 Promyelotic leukemia rat Employed in cytotoxicity assays of
cell line IPC-81 ILs
Protozoa Tetrahymena Free-living unicellular Commonly employed endpoint for
thermophila ciliated protozoa the assessment of the
Tetrahymena environmental toxicity
pyriformis
Tadpoles Bufo vulgaris Larvae of the frogs, Regularly implicated for toxicity
formosus, Rana typical amphibious testing purposes and risk
japonica animals bridging the assessments, have been
gap between aquatic recommended by the EU-TGD.
and terrestrial animals.
Yeast Saccharomyces Eukaryotic model Important species for modeling of
cerevisiae organism, small in metal oxide nanoparticles
size, reproduction
time quick and
economic.
428
Table 5
Representative list of publicly available databases related to the environmental toxicity due to pharmaceuticals
Database Country/organizations in charge Website

Supratik Kar et al.
ACToR US EPA National Center for Computational Toxicology http://actor.epa.gov/actor/faces/ACToRHome.jsp

BDSM University of Louisville http://systemsanalysis.louisville.edu/
Cal/EPA State of California EPA http://www.oehha.ca.gov/risk/ChemicalDB/index.asp
CCRIS (Chemical US National Library of Medicine (NLM) http://toxnet.nlm.nih.gov/cgi-bin/sis/
carcinogenesis htmlgen?CCRIS
research
information system)
CPDB (Carcinogenic University of California, Berkeley http://potency.berkeley.edu/
potency database)
Danish (Q)SAR Danish EPA http://ecbQSTR.jrc.ec.europa.eu/
DEMETRA EC-funded project http://www.demetra-tox.net/
DevTox German industry and government http://www.devtox.org
DSSTox (Distributed National Center for Computational Toxicology, US EPA http://www.epa.gov/ncct/dsstox/index.html
Structure-
Searchable Toxicity
database)
ECOTOX US EPA http://cfpub.epa.gov/ecotox/
ESIS (European Joint Research Centre of the European Commission http://ecb.jrc.ec.europa.eu/esis/
Chemical
Substances
Information system)
EXTOXNET (The Cooperative effort of University of California-Davis, Oregon State http://extoxnet.orst.edu/ghindex.html
EXtension University, Michigan State University, Cornell University, and the
TOXicology University of Idaho
NETwork)
FDA Poisonous Plant US FDA and Center for Food Safety and Applied Nutrition (CFSAN) http://www.accessdata.fda.gov/scripts/plantox/index.
Database cfm
GAC (Genetic US National Institutes of Health (NIH) and National Institute of http://www.niehs.nih.gov/research/resources/
Alterations in Environmental Health Sciences (NIEHS) databases/gac/index.cfm
Cancer)
Gene-Tox US NLM http://toxnet.nlm.nih.gov/cgi-bin/sis/
htmlgen?GENETOX
HERA (Human and A.I.S.E., the international Association for Soaps, Detergents and http://www.heraproject.com/RiskAssessment.cfm
Environmental Risk Maintenance Products official representative body of industries in
Assessment) Europe.
Household Products US Department of Health and Human Services (DHHS) http://householdproducts.nlm.nih.gov/
Database
IARC Monograph World Health Organization http://monographs.iarc.fr/
(International
Agency for Research
on Cancer
Monograph)
IRIS (Integrated Risk National Center for Environmental Assessment (NCEA) and US EPA http://cfpub.epa.gov/ncea/iris/index.cfm
Information
System)
ISSCAN Istituto Superiore di Sanità, Italy, http://www.iss.it/ampp/dati/cont.
php?id=233&lang=1&tipo=7
(continued)
Impact of Pharmaceuticals on the Environment: Risk Assessment Using QSAR Modeling…
429
430
Table 5
(continued)
Database Country/organizations in charge Website
ITER (International TERA (Toxicity Excellence for Risk Assessment) http://www.tera.org/iter/

Toxicity Estimates
Supratik Kar et al.
for Risk)
IUCLID OECD, EU Biocides and EU REACH http://iuclid.echa.europa.eu/
(International
Uniform ChemicaL
Information
Database)
JECDB Japanese Ministry of Health, Labour and Welfare http://dra4.nihs.go.jp/mhlw_data/jsp/
SearchPageENG.jsp
JRC QSTR Database European Commission, Joint Research Centre’s http://ecb.jrc.ec.europa.eu/QSTR/background/
KATE (KAshinhou Japanese National Institute for Environmental Studies (NIES), http://kate.nies.go.jp
Tool for Ministry of the Environment (MoE), Government of Japan
Ecotoxicity)
MRL (Minimal Risk US DHHS and Agency for Toxic Substances and Disease Registry http://www.atsdr.cdc.gov/mrls/index.html
Levels)
NPIC National Pesticide Information Center through Oregon State 166.1 http://npic.orst.edu/
University and US EPA
NTP (National US NIH/NIEHS http://ntp.niehs.nih.gov/
Toxicology
Program)
PAN Pesticide Pesticide Action Network, North America http://www.pesticideinfo.org/
Pesticide Database Toyohashi University of Technology, Japan http://chrom.tutms.tut.ac.jp/JINNO/
PESDATA/00alphabet.html
RITA (Registry of Fraunhofer Institute of Toxicology and Experimental Medicine http://www.item.fraunhofer.de/reni/public/rita/
Industrial (ITEM) Hannover index.php
Toxicology
Animal-data)
TEXTRATOX The University of Tennessee Institute of Agriculture http://www.vet.utk.edu/TETRATOX/index.php
TOXNET US National Library of Medicine (NLM) http://toxnet.nlm.nih.gov/
ToxRefDB US EPA http://www.epa.gov/ncct/toxrefdb/
Toxtree European Commission, Joint Research Centre http://ecb.jrc.ec.europa.eu/QSTR/QSTR-tools/index.
php?c=TOXTREE
TSCATS (Toxic US EPA https://toxplanet.com/tscats/
Substances Control
Act Test
Submissions)
USGS US Geological Survey http://137.227.231.90/data/acute/acute.html
197.1 WikiPharma Swedish research programme MistraPharma www.wikipharma.org
431
the seriousness of threat of hazard assessment, the available data-

bases are countable with hand in respect to pharmaceuticals library.
As the in silico models are need of the hour, so there must be
expansion and transparency of ecotoxicity database regarding
pharmaceuticals hazard and these must be accessible to the public
at no cost. Publicly accessible toxicity databases describing envi-
ronmental and human health hazards due to pharmaceuticals
implicated in risk assessment, management and hazard character-
ization is illustrated in Table 5.
10 Software
Expert systems are expedient option for toxicity prediction over

the traditional QSAR models as they permit ecotoxicity prediction
with the input of structure of pharmaceuticals only by selecting
endpoints and environment compartment. For speedy and eco-
nomical prediction as well as to make risk management guidelines,
regulatory authorities and industries are largely employing expert
systems. To prioritize the ecotoxicity assessment of pharmaceuti-
cals, the major aim is to distinguish between toxicologically active
and inactive molecules. Again, as multiple mechanisms can lead to
similar effects, therefore this complexity needs high quality predic-
tive tools which are capable of differentiating diverse regions in the
activity space. Expert systems can handle wide structural and
mechanistic complexity regions compare to the local models. We
have tried to summarize open access and commercially available
expert systems to predict pharmaceuticals ecotoxicity in Table 6.
11 Future Avenues
As the situation is very alarming already, there must be apposite

future plans to tackle the ecotoxicity threat due to pharmaceuti-
cals. Therefore, a set of plans must be addressed and followed for
efficient risk assessment along with quick risk management in dif-
ferent compartments of environment which are again directly
related to human health.
(a) Along with pharmaceuticals used by humans, veterinary phar-
maceuticals need to be monitored carefully.
(b) The packaging system should be biodegradable and eco-

friendly to minimize the packaging related hazards.
(c) Dose of drugs should be small and drugs must be ineffective
after expiry date to some extent considering its biological
response.
(d) Green chemistry principles should be followed for risk man-
agement of pharmaceuticals for quick and nuisance free
degradability after its usage. In this perspective, advancement
Table 6
An illustrative list of available expert systems to predict diverse endpoints of environmental toxicity
Expert system Country/Organization Endpoints and source of data Website

ASTER US EPA, NHEERL Assessment Tools for the Evaluation of Risk (ASTER) http://www.epa.gov/med/Prods_Pubs/aster.
designed to provide high quality data for discrete htm
chemicals (databases like ECOTOX and EcoChem),
and QSTR-based estimates
CAESAR EC funded project Computer Assisted Evaluation of industrial chemical http://www.caesar-project.eu/
(Project no. Substances According to Regulations (CAESAR)
022674SSPI) generates reproducible toxicity models. Five
endpoints considered are bioconcentration factor, skin
sensitization, carcinogenicity, mutagenicity,
developmental toxicity
CATALOGIC LMC University “As Platform for models and databases related to the http://oasis-lmc.org/products/software/
Zlatarov”, Bourgas, environment fate of chemicals such as abiotic and catalogic.aspx
Bulgaria biotic degradation, bioaccumulation, and acute
aquatic toxicity.
DEREK Harvard University Deductive estimation of risk from existing Knowledge http://www.lhasalimited.org/index.
Office of Technology (DEREK) consist of 21 structural alerts for php?cat=2&sub_cat=64
Development teratogenicity including four alerts and associated
reasoning rules and examples for estrogenicity.
DfW (Derek for Lhasa Limited Covering 361 toxicological endpoints alerts with https://www.lhasalimited.org/
Windows) and toxicophore. The skin sensitization knowledge base
Meteor Nexus was developed in collaboration with Unilever in 1993
using its database of GPMT data for 294 chemicals.
Version 9.0.0 contains 64 alerts for skin sensitization.
ECOSAR US EPA Hazard assessment of environmentally occurring http://www.epa.gov/oppt/exposure/docs/
pharmaceuticals to fish, daphnia, and green algae episuitedl.htm

using ECOlogical Structure–Activity Relationships
(ECOSAR).
433
(continued)
Table 6
434
(continued)
HazardExpert Pro CompuDrug Inc. Teratogenicty and reproductive toxicity http://www.compudrug.com/

MCASE/MC4PC MultiCASE Inc. Predictive models for blue gill, FHM, rainbow trout, http://www.multicase.com/products/
and red killifish are available. 180 modules covering products.htm
various areas of toxicology and pharmacology
Supratik Kar et al.
endpoints including skin sensitization, retinoids,

developmental toxicity, FDA/TERIS developmental
toxicity, developmental toxicants in FDA
teratogenicity are available
OASIS and TIMES Laboratory of Uses the response-surface approach for modeling acute http://oasis-lmc.org/products/software/
(TIssue Mathematical toxicity for two types of toxicochemical domains: times.aspx
MEtabolism Chemistry, LMC noncovalent (reversible) acting chemicals and
Simulator) University “As irreversible covalent bioreactive chemicals.
Zlatarov”, Bourgas, Interspecies correlations for acute toxicity to 17
Bulgaria aquatic species, such as fish, snail, tadpole, hydrozoan,
crustacean, insect larvae, and bacteria have been
developed
OECD (Q)SAR OECD, Paris A platform that allows for chemical information http://www.oecd.org/document/23/0,3343
Application management, similarity searches, and toxicological ,en_2649_34379_33957015_1_1_1_1,00.
Toolbox profiling. html
ONCOLOGIC US FDA/CDER Predicts cancer-causing potential by applying the rules http://www.epa.gov/oppt/sf/pubs/
of SAR analysis, mimicking the decision logic of oncologic.htm
human experts, and incorporating knowledge of how
chemicals cause cancer
OSIRIS property Actelion Fragments developed from the analysis of 3570 http://www.organic-chemistry.org/prog/
explorer Pharmaceuticals compounds with reproductive effects listed in peo/tox.html
Ltd., Allschwil, RTECS.
Switzerland
PASS Institute of Biomedical Abortion inducer, alkylator, carcinogenic, DNA http://ibmc.p450.ru/PASS//
Chemistry of the intercalator, DNA repair enzyme inhibitor, DNA
Russian Academy of synthesis inhibitor, DNA topoisomerase ATP
Medical Sciences, hydrolyzing Inhibitor, DNA topoisomerase inhibitor,
Moscow DOPA decarboxylase inhibitor, embryotoxic, estradiol
17β-dehydrogenase stimulant, ER modulator, estrone
sulfatase inhibitor, estrone sulfotransferase stimulant,
fertility enhancer, menopausal disorders treatment,
mutagenic, uterine stimulant, and a collection of
diverse set of receptor agonists and antagonists.
SARET MRC Structure–Activity Relationships for Environmental http://www.ibmh.msk.ru
“MEDTOXECO”, Toxicology (SARET) model is designed for statistical
Department of analysis of data and calculation of unknown
General Hygiene, parameters of substances on the basis of QSARs.
Russia and IBMC Provides the information necessary to evaluate the
RAMS, Russia hazard of chemicals and to estimate their unknown
characteristics.
TERA Russia Tools for environmental risk assessment (TERA) http://www.tera.org/
contains information about assessment of
multidomain risk, assessment of carcinogenic potency
risk, prediction of lead concentrations in blood of
fetus, children, adults, health risk connected with lead
exposure and prediction of emission of chemical
substances
TerraQSTR-FHM TerraBase Inc., Stand-alone neural network-based program to compute http://www.terrabase-inc.com
Hamilton, Ontario, the acute toxicity of organic chemicals to the FHM
Canada using a proprietary neural network algorithm
(continued)
435
Table 6
436
(continued)
TOPKAT BIOVIA TOxicity Prediction by C(K)omputer Assisted http://www.accelrys.com/products/topkat/

Technology (TOPKAT) uses a range QSAR models
for assessing acute toxicity to FHM and Daphnia. It
comprises two sets of skin sensitization models. The
Supratik Kar et al.
TOPKAT modeling has been employed by the Danish

EPA in their project to develop QSTR models for
around 47,000 organic substances on the European
Inventory of Existing Commercial chemical
Substances (EINECS)
Toxmatch European Union Open-source computer program of Joint Research http://ecb.jrc.ec.europa.eu/QSTR/QSTR-
Reference Centre (EC) that encodes several chemical similarity tools/index.php?c=TOXMATCH
Laboratory for indices in order to facilitate the grouping of
alternatives to animal chemicals, thereby supporting the development of
testing chemicals and the application of read-across between
(EURL-ECVAM) analogs
of synthesis and renewable feedstock are crucial issues for

preparation of environment friendly pharmaceuticals
[100]. Thus, “benign by design” criteria can be practiced
which asking for easy degradability after application. This
can lead to economic rewards in the long run and will fit
into green pharmacy [126]. Only thing needed to remem-
ber is that pharmaceuticals should not lose its therapeutic
action due to the introduction of green chemistry.
(e) The toxicity of metabolites and mixtures of pharmaceuti-
cals needs to be addressed more carefully as majority of
drugs are combinations of two or more API and some of
them are prodrugs in nature.
(f) There is significant deficiency of data on the effects of
long-term exposure in nontarget organisms, as well as how
an uninterrupted exposure may affect a population is not
explored till today.
(g) Pharmaceuticals’ ecotoxicity databases need to be organized
and classified in terms of endpoints, assay concentrations,
environment compartments with different experimental
conditions.
(h) The expert system should be more equipped with applica-
bility domain, conformal prediction and uncertainty issue
for reliable prediction.
12 Conclusion
The vibrant characteristic of pharmaceuticals is that they have

explicit mode of action and they are designed to exert specific
response to biological system which differentiate them from
other organic chemicals. Thus, once they are released into the
environment in an altered form or in their real form, they affect
the living organism of diverse compartments of environment
which is the sufficient reason to monitor and assess the poten-
tial effects of pharmaceuticals to environment. Assessment of
occurrence level of pharmaceuticals in diverse compartments is
necessary as the observed amount highly varies from one com-
partment to another which makes the situation complex for
environmentalists. On the other hand, most of the interactions
between pharmaceuticals and natural stressors of aquatic and
terrestrial communities are unexplained till date. Thus to eval-
uate the risk of waste hazards pose to wildlife, it has been rec-
ommended to utilize the toxicity data derived from mammals
during the production stages of pharmaceuticals which may be
helpful for future prediction. To quantify the ecotoxic concen-
tration threshold of any pharmaceutical, calculation of the risk
quotient (RQ) is very important. The RQ expresses the ratio

between the predicted concentration in the environment (PEC)
and the concentration at which no effect is expected (PNEC) [21].
The present chapter reviews the most common routes, sources
and occurrence of pharmaceutical hazards along with their fate and
effects in different environments. Apart from studies on individual
pharmaceuticals, the need of risk assessment and management of
their metabolites and mixture of pharmaceuticals are discussed.
The role of government authorities and different policies regulated
regarding identification of risk assessment and management must
be implemented in a proper way with right direction. Conversely,
insufficient ecotoxicity data related to a definite class of pharma-
ceuticals has slowed down the computational modeling to some
point. So, properly documented database is the need of the hour
for environmentalists. On the contrary, though a good number of
in silico models, especially QSAR models are reported for toxicity
prediction of chemicals hazards, but a few successful models exist
for pharmaceutical hazards. Thus, ample number of models should
be generated to get the ecotoxicity data for pharmaceuticals for
diverse endpoints and environmental compartments. Expert sys-
tems have the ability to make fast and reliable predictions with a
single click of mouse, which urges to develop more expert systems
for a set of endpoints. There is no doubt that in silico methods can-
not completely substitute experimental approaches, but they can
be integrated for better understanding and quantification of phar-
maceuticals’ ecotoxicity.
Acknowledgments
S.K. and J.L. wish to thank the National Science Foundation

(NSF/CREST HRD-1547754, and NSF/RISE HRD-1547836)
for financial support. K.R. is thankful to the UGC, New Delhi for
financial assistance under the UPE II scheme.
References
1. Aherne G, English J, Marks V (1985) The 4. Li WC (2014) Occurrence, sources, and fate
role of immunoassay in the analysis of micro- of pharmaceuticals in aquatic environment
contaminants in water samples. Ecotoxicol and soil. Environ Pollut 187:193–201
Environ Saf 9:79–83 5. WHO (2011) The World Medicines Situation
2. Richardson M, Bowron J (1985) The fate of 6. Busfield J (2015) Assessing the over use of
pharmaceutical chemicals in the aquatic envi- medicines. Soc Sci Med 131:199–206
ronment. J Pharm Pharmacol 37:1–12 7. IWW (2014) Pharmaceuticals in the environ-
3. Santosa LHMLM, Araújoa AN, Fachinia A ment: occurence, effects and options for action.
et al (2010) Ecotoxicological aspects related Research project funded by German Federal
to the presence of pharmaceuticals in the Environment Agency (UBA) within the
aquatic environment. J Hazard Mater Environmental Research Plan No.371265408.
175:45–95
8. Roy K, Kar S (2016) In silico models for eco- 20. Adler P, Steger-Hartmann T, Kalbfus W
toxicity of pharmaceuticals. In: Benfenati E (2001) Distribution of natural and synthetic
(ed) In silico methods for predicting drug estrogenic steroid hormones in water samples
toxicity, Methods in molecular biology, vol from southern and middle Germany. Acta
1425. Springer, New York, NY Hydrochim Hydrobiol 29:227–241
9. Taylor D, Senac T (2014) Human pharma- 21. Ternes T (1998) Occurence of drugs in
ceutical products in the environment – the German sewage treatment plants and rivers.
“problem” in perspective. Chemosphere Water Res 32:3245–3260
115:95–99 22. Ahrer W, Scherwenk E, Buchberger W (2001)
10. Kümmerer K (2013) Pharmaceuticals in the Occurence and fate of fluoroquinolone, mac-
environment: sources, fate, effects and risks. rolide, and sulphonamide antibiotics during
Springer Science & Business Media, Berlin wastewater treatment and in ambient waters
11. Hughes SR, Kay P, Brown LE (2013) Global in Switzerland. In: Daughton CG, Jones-
synthesis and critical evaluation of pharma- Lepp T (eds) Pharmaceuticals and personal
ceutical datasets collected from river systems. care products in the environment: scientific
Environ Sci Technol 47:661–677 and regulatory issues. Symposium Series 791.
12. Han GH, Hur HG, Kim SD (2006) American Chemical Society, Washington,
Ecotoxicological risk of pharmaceuticals from DC, pp 56–69
wastewater treatment plants in Korea: occur- 23. Frick EA, Henderson AK, Moll DM, et al
rence and toxicity to Daphnia magna. Environ (2001) Presented at the Georgia Water
Toxicol Chem 25:265–271 Resources Conference, Athens, GA.
13. Fernandez C, Gonzalez-Doncel M, Pro J et al 24. Tauber R (2003) Quantitative analysis of
(2010) Occurrence of pharmaceutically active pharmaceuticals in drinking water from ten
compounds in surface waters of the henares- Canadian cities. Enviro-Test Laboratories,
jarama-tajo river system (Madrid, Spain) and a Winnipeg, MB
potential risk characterization. Sci Total 25. Seiler JP (2002) Pharmacodynamic activity of
Environ 408:543–551 drugs and ecotoxicology can the two be con-
14. Papageorgiou M, Kosma C, Lambropoulou nected? Toxicol Lett 131:105–115
D (2016) Seasonal occurrence, removal, mass 26. EMEA (2006) Guideline on the environmen-
loading and environmental risk assessment of tal impact assessment of medicinal products
55 pharmaceuticals and personal care prod- for human use (Report no. CPMP/
ucts in a municipal wastewater treatment SWP/4447/00). European Agency for the
plant in Central Greece. Sci Total Environ Evaluation of Medicinal Products, London
543:547–569 27. FDA-CDER (1998) Guidance for industry-
15. Oliveira TS, Murphy M, Mendola N et al environmental assessment of human drugs
(2015) Characterization of pharmaceuticals and biologics applications, Revision 1. FDA
and personal care products in hospital effluent Center for Drug Evaluation and Research,
and waste water influent/effluent by direct- Rockville, VA
injection LC-MS-MS. Sci Total Environ 28. FDA: U.S. Department of Health and
518:459–478 Human Services, Food and Drug
16. Rivas J, Encinas A, Beltran F, Grahan N Administration, Center for Drug Evaluation
(2011) Application of advanced oxidation and Research (CDER), Center for Biologics
processes to doxycycline and norfloxacin Evaluation and Research (CBER) (1998)
removal from water. J Environ Sci Health A CMC 6 - Revision 1
Tox Hazard Subst Environ Eng A 29. European Commission, Directive 2006/121/
46:944–951 EC of the European Parliament and of the
17. Kolpin DW, Furlong ET, Meyer MT et al Council of 18 December 2006 amending
(2002) Pharmaceuticals, hormones, and Council Directive 67/548/EEC on the
other organic wastewater contaminants in approximation of laws, regulations and
U.S. streams, 1999-2000: a national recon- administrative provisions relating to the clas-
naissance. Environ Sci Technol sification, packaging and labelling of danger-
36:1202–1211 ous substances in order to adapt it to
18. Halling-Sørensen B, Nors Nielsen S, Lanzky Regulation (EC) No. 1907/2006 concerning
PF et al (1998) Occurrence, fate and effects the Registration, Evaluation, Authorisation
of pharmaceutical substances in the environ- and Restriction of Chemicals (REACH) and
ment-a review. Chemosphere 36(2):357–393 establishing a European Chemicals Agency.
19. Daughton CG, Ternes TA (1999) Off. J. Eur. Union, L 396/850 of 30.12.2006,
Pharmaceuticals and personal care products in Office for Official Publications of the
the environment: agents of subtle change? European Communities (OPOCE),
Environ Health Perspect 107:907–937 Luxembourg
30. Directive 2004/27/EC of the European 42. US EPA (2015) ECOTOX user guide:
Parliament and of the Council of 31 March ECOTOXicology database system. Version
2004 amending Directive 2001/83/EC on 4.0. http:/www.epa.gov/ecotox/. Accessed
the Community code relating to medicinal 4 July, 2017
products for human use. Official Journal L 43. Mendoza A, Acena J, Perez S et al (2015)
136, 30/04/2004 pp. 34-57. 2004 Pharmaceuticals and iodinated contrast media
31. Directive 2004/28/EC of the European in a hospital wastewater: a case study to anal-
Parliament and of the Council of 31 March yse their presence and characterise their envi-
2004 amending Directive 2001/82/EC on ronmental risk and hazard. Environ Res
the Community code relating to veterinary 140:225–241
medicinal products. Official Journal L 136, 44. Ortiz de García S, Pinto GP, García-Encina
30/04/2004 pp. 58-84. 2004 PA et al (2013) Ranking of concern, based on
32. DIRECTIVE 2013/39/EU OF THE environmental indexes, for pharmaceutical
EUROPEAN PARLIAMENT AND OF and personal care products: an application to
THE COUNCIL of 12 August 2013 amend- the Spanish case. J Environ Manage
ing Directives 2000/60/EC and 2008/105/ 129:384–397
EC as regards priority substances in the field 45. Sanderson H, Thomsen M (2009)
of water policy. 2013 Comparative analysis of pharmaceuticals ver-
33. COMMISSION IMPLEMENTING sus industrial chemicals acute aquatic toxicity
DECISION (EU) 2015/495 of 20 March classification according to the United Nations
2015 establishing a watch list of substances classification system for chemicals. Assessment
for Union-wide monitoring in the field of of the (Q)SAR predictability of pharmaceuti-
water policy pursuant to Directive 2008/105/ cals acute aquatic toxicity and their
EC of the European Parliament and of the predominant acute toxicmode-of-action.

Council. 2015 Toxicol Lett 187:84–93
34. Barbosa MO, Moreira NFF, Ribeiro AR et al 46. Singh KP, Gupta S, Basant N (2015) QSTR
(2016) Occurrence and removal of organic modeling for predicting aquatic toxicity of
micropollutants: an overview of the watch list pharmacological active compounds in multi-
of EU Decision 2015/495. Water Res ple test species for regulatory purpose.
94:257–279 Chemosphere 120:680–689
35. Cassani S, Gramatica P (2015) Identification 47. Persson M, Sabelström E, Gunnarsson B
of potential PBT behavior of personal care (2009) Handling of unused prescription
products by structural approaches. Sustain drugs-knowledge, behaviour and attitude
Chem Pharm 1:19–27 among Swedish people. Environ Int
36. Roy K, Kar S, Das RN (2015) Understanding 35:771–774
the Basics of QSAR for Applications in 48. Li D, Yang M, Hu J et al (2008) Determination
Pharmaceutical Sciences and Risk Assessment. and fate of oxytetracycline and related com-
Academic Press, San Diego, CA pounds in oxytetracycline production waste-
37. Roy K, Kar S, Das RN (2015) A primer on water and the receiving river. Environ Toxicol
QSAR/QSPR modeling: fundamental con- Chem 27:80–86
cepts (SpringerBriefs in Molecular Science). 49. José Gómez M, Petrovic M, Fernández-Alba
Springer, New York, NY AR et al (2006) Determination of pharma-
38. Howard PH, Muir DCG (2011) Identifying ceuticals of various therapeutic classes by
new persistent and bioaccumulative organics solid-phase extraction and liquid
among chemicals in commerce II: pharma- chromatographye-tandem mass spectrometry
ceuticals. Environ Sci Technol analysis in hospital effluent wastewaters.
45:6938–6946 J Chromatogr 1114:224–233
39. Sangion A, Gramatica P (2016) PBT assess- 50. Ternes TA, Hirsch R (2000) Occurrence and
ment and prioritization of contaminants of behavior of X-ray contrast media in sewage
emerging concern: pharmaceuticals. Environ facilities and the aquatic environment.
Res 147:297–306 Environ Sci Technol 34:2741–2748
40. European Commission (2001) 51. Serrano PH (2005) Responsible use of antibi-
CSTEE. Discussion paper on environmental otics in aquaculture. fisheries technical paper
risk assessment of medical products for human 469. Food and Agriculture Organization of
use (non-geneticallymodified organisms the United Nations (FAO), Rome
(non-GMO) containing). CPMPpaperRAsses 52. Kreuzig R, Höltge S, Brunotte J et al (2005)
sHumPharm12062001/D(01) Test plat studies on runoff of sulfonamides
41. US EPA (2012) The ECOSAR (ECOlogical from manured soil after sprinkler irrigation.
Structure Activity Relationship) Class Environ Toxicol Chem 24:777–781
Program. 53. http://www.apsnet.org/online/feature/
Antibiotics/. Accessed 4 July, 2017
54. http://www.esac.ua.ac.be/main.aspx?c= of derivatization, gas chromatography and

*ESAC2&n=1066l. Accessed 4 July, 2017 negative-chemical ionization mass spectrom-
55. Ye Z, Weinberg HS, Meyer MT (2007) Trace etry. Clean 35:444–451
analysis of trimethoprim and sulfonamide, 68. Cleuvers M (2004) Mixture toxicity of the
macrolide, quinolone, and tetracycline antibi- anti-inflammatory drugs diclofenac, ibupro-
otics in chlorinated drinking water using liquid fen, naproxen and acetylsalicylic acid.
chromatography electrospray tandem mass Ecotoxicol Environ Saf 59:309–315
spectrometry. Anal Chem 79:1135–1144 69. Schwaiger J, Ferling H, Mallow U et al
56. Hamscher G, Sczesny S, Höper H et al (2002) (2004) Toxic effects of the non-steroidal anti-
Determination of persistent tetracycline resi- inflammatory drug diclofenac: Part I: histo-
dues in soil fertilized with liquid manure by pathological alterations and bioaccumulation
high-performance liquid chromatography in rainbow trout. Aquat Toxicol 68:141
with electrospray ionization tandem mass 70. Huggett DB, Brooks BW, Peterson B et al
spectrometry. Anal Chem 74:1509–1518 (2002) Toxicity of select beta adrenergic
57. Kümmerer K, Alexy R, Hüttig J (2004) receptor-blocking pharmaceuticals
Standardized tests fail to assess the effects of (β-blockers) on aquatic organisms. Arch
antibiotics on environmental bacteria. Water Environ Contam Toxicol 43:229
Res 38:2111–2116 71. Coppi A, Merali S, Eichinger D (2002) The
58. Carucci A, Cappai G, Piredda M (2006) enteric parasite Entamoeba uses an autocrine
Biodegradability and toxicity of pharmaceuti- catecholamine system during differentiation
cals in biological wastewater treatment plants. into the infectious cysts stage. J Biol Chem
Environ Sci Health A Tox Hazard Subst 277:8083–8090
Environ Eng 41:1831–1842 72. Sanderson H, Brain RA, Johnson DJ et al
59. Hernando MD, DeVettori S, Martínez- (2004) Toxicity classification and evaluation
Bueno MJ et al (2007) Toxicity evaluation of four pharmaceuticals classes: antibiotics,
with Vibrio fisheri test of organic chemicals antineoplastics, cardiovascular, and sex hor-
used in aquaculture. Chemosphere mones. Toxicology 203:27–40
68:724–730 73. Henschel K-P, Wenzel A, Diedrich M et al
60. Brain RA, Johnson DJ, Richards SM (2004) (1997) Environmental hazard assessment of
Microcosm evaluation of the effects of an pharmaceuticals. Regul Toxicol Pharmacol
eight pharmaceutical mixture to the aquatic 25:220–225
macrophytes Lemna gibba and Myriophyllum 74. Boroskova Z, Dvoroznakova E, Tomasovicova
sibiricum. Aquat Toxicol 70:23–40 O et al (2001) The effect of cyclophospha-
61. Hansen AJ, Jensen J, Krogh PH (2000) mide on the cellular and humoral immune
Effects of the antibiotics oxytetracycline and responses on experimental Toxocara canis
tylosin on soil fauna. Chemosphere infestation. Helminthologia 38:193–199
40:751–757 75. Grujić S, Vasiljević T, Lauŝević M (2009)
62. Holt D, Harvey D, Hurley R (1993) Determination of multiple pharmaceutical
Chloramphenicol toxicity. Adverse Drug classes in surface and ground waters by liq-
React Toxicol Rev 12:83–95 uid chromatography-ion trap-tandem mass
63. Harrison PF, Lederberg J (1998) spectrometry. J Chromatogr A
Antimicrobial Resistance, Issues and Options. 1216:4989–5000
National Academic Press, Washington DC 76. Ellenberger TE, Wright JE, Rosowsky A et al
64. Isidori M, Nardelli A, Pascarella L et al (2007) (1989) Wild-type and drugresistant
Toxic and genotoxic impact of fibrates and Leishmania major hydrolyze methotrexate to
their photoproducts on non-target organism. N-10-methyl-4-deoxy-4-aminopteroate with-
Environ Int 33:635–641 out accumulation of methotrexate polygluta-
65. Triebskorn R, Casper H, Scheil V et al (2007) mates. J Biol Chem 264:15960–15966
Ultrastructural effects of pharmaceuticals 77. Minagh E, Hernan R, O’Rourke K et al
(carbamazepine, clofibric acid, metoprolol, (2009) Aquatic ecotoxicity of the selective
diclofenac) in rainbow trout (Oncorhynchus serotonin reuptake inhibitor sertraline hydro-
mykiss) and common carp (Cyprinus carpio). chloride in a battery of freshwater test species.
Anal Bioanal Chem 387:1405–1416 Ecotoxicol Environ Saf 72:434–440
66. Misra SK, Sharma AK, Mehdi H et al (1986) 78. Crane M, Watts C, Boucard T (2006) Chronic
Effect of cholesterol and alphap- chloro- aquatic environmental risks from exposure to
phenoxyisobutyyrate on virulence in human pharmaceuticals. Sci Total Environ
Acanthamoeba culbertsoni strain A-1 and C-7. 367:23–41
Int J Parasitol 16:191–196 79. Baqui A, Ansari JA (1987) Investigations on
67. Moder M, Braun P, Lange F et al (2007) the response to various stimuli of the microfi-
Determination of endocrine disrupting com- lariae of Setaria cervi in white rats.
pounds and acidic drugs in water by coupling Helminthologia 24:33–42
80. Jones-Brando L, Torrey EF, Yolkem R (2003) managing the process. National Academies
Drugs used in the treatment of schizophrenia Press, Washington, DC. Accessed 4 July 2017
and bipolar disorder inhibit the replication of 94. NRC (National Research Council) (2007)
Toxoplasma gondii. Schizophr Res Toxicity testing in the 21st century: a vision
62:237–244 and a strategy. National Academies Press,
81. Pawlowski S, van Aerle R, Tyler CR et al Washington, DC. Available: https://down-
(2004) Effects of 17α-ethinylestradiol in a load.nap.edu/login.php?record_
fathead minnow (Pimephales promelas) id=11970&page=%2Fdownload.
gonadal recrudescence assay. Ecotoxicol php%3Frecord_id%3D11970. Accessed 4 July
Environ Saf 57:330–345 2015
82. Arcay L (1985) Influence of sex hormones on 95. Wignall JA, Shapiro AJ, Wright FA et al
the experimental infection produced by a (2014) Standardizing benchmark dose calcu-
strain of Leishmania mexicana amazonensis lations to improve science-based decisions in
from Venezuela. Rev Latinoam Microbiol human health assessments. Environ Health
27:195–207 Perspect 122:499–505
83. Wang R, Belosevic M (1994) Estradiol 96. Wetmore BA, Wambaugh JF, Ferguson SS
increases susceptibility of goldfish to et al (2012) Integration of dosimetry, expo-
Trypanosoma danilewskyi. Dev Comp Immunol sure, and high-throughput screening data in
18:377–387 chemical toxicity assessment. Toxicol Sci
84. Martín J, Camacho-Muñoz D, Santos JL et al 125:157–174
(2012) Occurrence of pharmaceutical com- 97. Jones DP, Park Y, Ziegler TR (2012)
pounds in wastewater and sludge from waste- Nutritional metabolomics: progress in
water treatment plants: removal and addressing complexity in diet and health.
ecotoxicological impact of wastewater dis- Annu Rev Nutr 32:183–202
charges and sludge disposal. J Hazard Mater 98. van Gestel CA, Jonker M, Kammenga JE
239-240:40–47 (2010) Mixture toxicity: linking approaches
85. Boxall AB, Fogg LA, Baird DJ et al (2005) from ecological and human toxicology.
Targeted monitoring study for veterinary Taylor& Francis, New York, NY
medicines in the environment. Environment 99. Presidential/Congressional Commission on
Agency, Bristol Risk Assessment Risk Management (1997).
86. Grønvold J, Svendsen TS, Kraglund HO et al Risk assessment and risk management in regu-
(2004) Effect of the antiparasitic drugs fen- latory decision-making. Final Report. Vol. 2.
bendazole and ivermectin on the soil nema- Washington, DC:PCRARM. Available:
tode Pristionchus maupasi. Vet Parasitol http://www.riskworld.com. Accessed 4 July,
24:91–99 2017
87. Svendsen TS, Hansen PE, Sommer C et al 100. Kümmerer K (2007) Sustainable from the

(2005) Life history characteristics of very beginning: rational design of molecules
Lumbricus terrestris and effects of the veteri- by life cycle engineering as an important
nary antiparasitics compounds Ivermectin and approach for green pharmacy and green
fenbendazole. Soil Biol Biochem 37:927–936 chemistry. Green Chem 9:899–907
88. Sun Y, Diao X, Zhang Q et al (2005) 101. Nowotny N, Epp B, Von Sonntag C (2007)
Bioaccumulation and elimination of avermectin Quantification and modeling of the elimina-
B1a in the earthworm (Eisenia fetida). tion behavior of ecologically problematic
Chemosphere 60:699–704 wastewater micropollutants by adsorption on
89. Singer AC, Johnson AC, Anderson PD powdered and granulated activated carbon.
(2008) Reassessing the risks of Tamiflu use Environ Sci Technol 41:2050–2055
during a pandemic to the Lower Colorado 102. Fjodorova N, Novich M, Vrachko N et al

River. Environ Health Perspect 116:A285– (2008) Directions in QSAR modeling for
A286 regulatory uses in OECD member countries,
90. Soderstrom H, Jarhult JD, Fick J, et al (2010) EU and in Russia. J Environ Sci Health Part
Levels of antivirals and antibiotics in the river C 26:201–236
Thames, UK, during the pandemic 2009. 103. Knacker T, Duis K, Ternes T et al (2005) The
SETAC Europe Annual Meeting, Seville EU-project ERAPharmsIncentives for the
91. http://www.biomedicalwastesolutions.com/ further development of guidance documents?
medical-waste-disposal/. Accessed 4 July, 2017 Environ Sci Pollut Res 12:62–65
92. Rappaport SM (2011) Implications of the 104. Cleuvers M (2002) Aquatic ecotoxicology of
exposome for exposure science. J Expo Sci selected pharmaceutical agents algal and acute
Environ Epidemiol 21:5–9 Daphnia tests. Umweltwissenschaften und
93. NRC (National Research Council) (1983) Schadstoff-Forschung 14:85–89
Risk assessment in the federal government: 105. VICH (2004) Environmental impact assess-
ment (EIAs) for veterinary medicinal prod-
ucts (VMPs)–Phase II, Draft Guidance, 117. Escher BI, Bramaz N, Richter M et al (2006)
August, 2003, International Cooperation on Comparative ecotoxicological hazard assess-
Harmonization of the Technical Requirements ment of beta-blockers and their human
for Registration of Veterinary Medicinal metabolites using a mode-of-action-based
Products. Available at: http://vich.eudra. test battery and a QSTR approach. Environ
org/pdf/10_2003/g138_st4.pdf Sci Technol 40:7402–7408
106. http://www.nicnas.gov.au/. Accessed 4 July, 118.
Sanderson H, Thomsen M (2007)
2017 Ecotoxicological quantitative structure–
107. Dearden JC (2016) The history and develop- activity relationships for pharmaceuticals. Bull
ment of quantitative structure-activity rela- Environ Contam Toxicol 79:331–335
tionships (QSARs). Int J Quant Struct-Prop 119. Kar S, Roy K (2010) First report on

Relat 1:1–44 interspecies quantitative correlation of eco-
108. Roy K, Kar S (2015) How to judge predictive toxicity of pharmaceuticals. Chemosphere
quality of classification and regression based 81:738–747
QSAR models? In: Haq ZU, Madura J (eds) 120. Christen V, Hickmann S, Rechenberg B et al
Frontiers of computational chemistry. Bentham, (2010) Highly active human pharmaceuticals
pp 71–120 in aquatic systems: a concept for their identi-
109. Roy K, Kar S (2015) Importance of applica- fication based on their mode of action. Aquat
bility domain of QSAR models. In: Roy K Toxicol 96:167–181
(ed) Quantitative structure-activity relation- 121. http://www.biograf.ch. Accessed 4 July, 2017
ships in drug design, predictive toxicology, 122. Das RN, Sanderson H, Mwambo AE et al
and risk assessment. IGI Global, Hershey PA, (2013) Preliminary studies on model devel-
pp 180–211 opment for rodent toxicity and its interspecies
110. http://ec.europa.eu/research/biosociety/ correlation with aquatic toxicities of pharma-
pdf/anim_al_see_final_report.pdf. Accessed ceuticals. Bull Environ Contam Toxicol
4 July, 2017 90:375–381
111. Kar S, Roy K (2012) Risk assessment for eco- 123. de García MSAO, Pinto GP, García-Encina
toxicity of pharmaceuticals - an emerging PA et al (2014) Ecotoxicity and environmen-
issue. Expert Opin Drug Saf 11:235–274 tal risk assessment of pharmaceuticals and
112. Kanter J (2013) E.U. Bans cosmetics with
personal care products in aquatic environ-

animal-tested ingredients. http://www. ments and wastewater treatment plants.
nytimes.com. Accessed 4 July, 2017 Ecotoxicology 23:1517–1533
113.
Cronin MTD, Livingstone DJ (2004) 124. Sangion A, Gramatica P (2016) Hazard of
Predicting chemical toxicity and fate. CRC pharmaceuticals for aquatic environment:
press, Washington DC prioritization by structural approaches and

114. Schultz TW, Cronin MTD, Walker JD (2003) prediction of ecotoxicity. Environ Int
Quantitative structure-activity relationships 95:131–143
(QSARs) in toxicology: a historical perspec- 125. Sangion A, Gramatica P (2016) Ecotoxicity
tive. J Mol Struct (THEOCHEM) 622:1–22 interspecies QAAR models from Daphnia
115. Kar S, Roy K (2010) Predictive toxicology toxicity of pharmaceuticals and personal care
using QSAR: a perspective. J Indian Chem products. SAR QSAR Environ Res
Soc 87:1455–1515 27:781–798
116. Sanderson H, Johnson D, Reitsma T et al
126. Kümmerer K (2009) The presence of phar-
(2004) Ranking and prioritization of environ- maceuticals in the environment due to human
mental risks of pharmaceuticals in surface use-present knowledge and future challenges.
waters. Reg Pharm Toxicol 39:158–183 J Environ Manage 90:2354–2366
Part IV
Predicting Human Health Toxicology Endpoints

Chapter 20
(Q)SAR Methods for Predicting Genotoxicity

and Carcinogenicity: Scientific Rationale and Regulatory
Frameworks
Cecilia Bossa, Romualdo Benigni, Olga Tcheremenskaia,
and Chiara Laura Battistelli
Abstract
Knowledge of the genotoxicity and carcinogenicity potential of chemical substances is one of the key sci-
entific elements able to better protect human health. Genotoxicity assessment is also considered as pre-
screening of carcinogenicity. The assessment of both endpoints is a fundamental component of national
and international legislations, for all types of substances, and has stimulated the development of alternative,
nontesting methods. Over the recent decades, much attention has been given to the use and further devel-
opment of structure–activity relationships-based approaches, to be used in isolation or in combination
with in vitro assays for predictive purposes. In this chapter, we briefly introduce the rationale for the main
(Q)SAR approaches, and detail the most important regulatory initiatives and frameworks. It appears that
the existence and needs of regulatory frameworks stimulate the development of better predictive tools; in
turn, this allows the regulators to fine-tune their requirements for an improved defense of human health.
Key words QSAR, Structure–activity relationship, Predictive toxicology, Genotoxicity,

Carcinogenicity, Expert systems, Human health, Alternative testing
1 Introduction
Knowledge of carcinogenicity and genotoxicity potential of chemi-

cal substances is one of the key elements for protect human health.
Their assessment is a fundamental component of national and
international legislations, for all types of substances. Traditionally,
experimental long-term chemical carcinogenesis bioassays in
rodents are carried out to identify potential carcinogenic chemi-
cals. These assays efficiently detect human carcinogens [1, 2], both
genotoxic and nongenotoxic. In the interests of time-effectiveness,
cost-effectiveness, and animal welfare, the research on alternative
methods, with respect to animal assays, has evolved largely both in
the fields of testing and nontesting (in silico) methods.
447
448 Cecilia Bossa et al.
Currently, regulatory policies prescribe genotoxicity testing as

early detection and prescreening for carcinogenic chemicals, as
well as for the evaluation of the genotoxic potential per se. The
assessment has to be carried out with a series of different tests (i.e.,
a testing battery), to ensure a complete coverage of all the end-
points relevant for carcinogenicity and genotoxicity assessment. In
particular, assays measuring both mutagenicity and genotoxicity
have to be considered.
Mutagenicity refers to the induction of permanent transmissi-
ble alterations in the amount or structure of the genetic material in
cells or organisms. These include different types of events such as
base substitutions and deletions, structural chromosomal aberra-
tions (CAs) (break and rearrangements, i.e., clastogenicity), and
numerical CAs (loss or gain of chromosomes, i.e., aneuploidy). It
should be noticed that some of these alterations, such as gene
mutations, are provoked by interactions with DNA, whereas the
other may derive from interactions both with DNA and with other
cellular targets (e.g., proteins). The term genotoxicity has a broader
significance. It includes mutagenic effects, together with DNA
damage, which may not necessarily result in permanent alterations
of the genetic material. In regulatory environments, generally, the
assessment of genotoxic hazard to humans follows a step-wise
approach, beginning with a basic battery of in vitro tests followed
in some cases by confirmatory in vivo follow-up [3]. In Table 1,
different assays are reported, associated with the information of the
endpoint detectable.
Thus, genotoxicity testing is performed (1) to assess the poten-
tial of substances to induce genotoxic effects which may cause heri-
table damage, (2) to predict potential genotoxic carcinogens and
(3) to contribute to the elucidation of carcinogens mechanism of
action.
In the case of carcinogenicity assessment, genotoxicity tests are
indicative of the potential reactivity of DNA-reactive chemicals
[4]. For a complete assessment of carcinogenic potential, also
chemicals acting with non genotoxic mechanism should be
detected. Presently, the only short-term alternative of the rodent
bioassay could be represented by the Cell transformation assays [5,
6]. As per the recently released OECD guidance documents [7, 8],
however, there is no clear consensus for their utilization in regula-
tory environments.
2 (Q)SAR Methodologies
Among nontesting methods, in regulatory environments (Q)SAR

methodologies have gained prominent importance over the last
decades [9, 10]. The term (Q)SAR, (quantitative) structure–
activity relationships, refers to the study of the correlations between
Table 1
Available in vitro and in vivo methods for genotoxicity/mutagenicity assessment
Mutagenicity Genotoxicity
Test method Adopted OECD test In vitro/ Gene Structural Numerical DNA DNA Germ cell
guideline in vivo mutations aberrations aberrations repair damage mutagenicity
activity
Bacterial reverse mutation July TG 471 In vitro ✓
test (Ames test) 1997
In vitro Mammalian July TG 473 In vitro ✓
chromosome aberration 2016
test
In vitro Mammalian cell July TG 476 In vitro ✓ ✓
gene mutation test 2016 TG 490
In vitro Mammalian cell July TG 487 In vitro ✓ ✓
micronucleus test 2016
Mammalian bone marrow October TG 479 In vitro ✓
Sister Chromatid 1986
Exchanges (SCE)
Mammalian erythrocyte July TG 474 In vivo ✓ ✓
micronucleus test 2016
Mammalian bone marrow July TG 475 In vivo ✓
chromosome aberration 2016
test
Transgenic rodent somatic July TG 488 In vivo ✓ ✓
and germ cell gene 2013
mutation assays
(Q)SAR Methods for Predicting Genotoxicity and Carcinogenicity: Scientific Rationale…
(continued)
449
Table 1
450
(continued)
Mutagenicity Genotoxicity
Unscheduled DNA July TG 486 In vivo ✓ ✓
Synthesis (UDS) test 1997
with mammalian liver
cells
Cecilia Bossa et al.
In vivo Mammalian July TG 489 In vivo ✓

Alkaline Comet Assay 2016
Rodent Dominant Lethal July TG 478 In vivo ✓ ✓ ✓
test 2016
Mammalian July TG 483 In vivo ✓ ✓
Spermatogonial 2016
Chromosome
Aberration test
Mouse Heritable October TG 485 In vivo ✓ ✓ ✓
Translocation assay 1986
(Q)SAR Methods for Predicting Genotoxicity and Carcinogenicity: Scientific Rationale… 451
the chemical structure and the biological activity of a substance.

This knowledge can be transferred into models for the quantitative
or qualitative prediction of biological properties of untested chem-
icals, based on structurally related compounds with known activity.
Although the term (Q)SAR has been historically associated to
computer-based models (in silico models), it currently includes a
wide variety of computerized and noncomputerized tools and
approaches, such as QSAR and SAR, Grouping and Read-Across.
Since the concept at the basis of these approaches is the same (i.e.,
toxicity is a function of chemical structure), the methodologies are
strongly interconnected and the terms sometimes used
interchangeably.
For the sake of clarity, we briefly recall the definitions as can be
found in the OECD guidance [11]. These may be slightly different
in other regulatory/research environments (e.g., REACH), but
the basic concepts remain the same.
2.1 Grouping As by the OECD definition, the term (Chemical) Grouping

describes the general approach for considering more than one chemi-
cal at the same time, with the aim of filling gaps in (eco)toxicologi-
cal information necessary to assess hazards of chemicals. Analogue
and Category approaches are included in this definition. In par-
ticular, the Analogue approach concerns the assessment of one spe-
cific chemical, using for the prediction experimental data from
others (one or more) similar (see Note 1) substances. In the
Category approach, chemicals sharing similar characteristics and
(eco)toxicological properties are grouped together. The assess-
ment concerns the category as a whole as data gaps may exist for
different category members and different endpoints. Applying a
Grouping approach, either analogue or category, involves the use
of methodologies to fill the data gaps inside the group. Most popu-
lar are read-across, trend analysis, and (Q)SAR.
As a matter of fact, the grouping approach is a weight of evi-
dence approach (WoE), since it normally integrates different source
of data, experimental and estimated data, with an expert supervi-
sion [12].
2.2 Read-Across Read-across is a technique used for filling data gaps by predicting
endpoint information for one (or more) chemical(s) (referred to as
target chemical), using data for the same endpoint from one or
more similar (see Note 1) substances (the source chemicals). |Read-
across can be qualitative or quantitative, depending on the nature
of the endpoint to be predicted and on what type of data are
required in the specific process.
2.3 Trend Analysis When the substances in a chemical group are related by a trend in
one or more properties (e.g., molecular mass and carbon chain
length), also experimental data for a given endpoint may change
accordingly across the group members. In such a case, the trend

can be used as the basis for the estimation of the missing data value
by interpolation (and possibly extrapolation). Result of the trend
analysis can be the definition of a local QSAR model.
2.4 (Q)SAR QSARs are mathematical equations linking the biological activity
to a limited number of physical chemical or other molecular prop-
erties (descriptors). In order to perform a data gap filling for a
defined endpoint in a group of chemicals, local QSAR models may
be calculated. These kinds of “internal models” are developed as
part of the category formation process, based on the experimental
data available for the category members. Alternatively, external
local or global models could be used for data gap filling. Local
QSARs are calculated on congeneric set of chemicals, i.e., chemi-
cals with similar structure that act through the same mechanism of
action, and are generally more effective models [13, 14]. However,
as the domain of applicability is strictly related to the set of chemi-
cals used for the model definition, local QSARs are narrower in
scope.
Global models, often implemented in software tools, are con-
structed in such a way as to ensure a much broader applicability,
extended to more than one chemical class. Global models can be
classified on the basis of three main modeling approaches referred
to rule-based, statistically based, or hybrid methods. At the basis of
rule-based models, there is the recognition and codification of
functional groups or structural features associated with a potential
reactivity for a defined endpoint, e.g., the structural alerts (SAs).
The structural alerts represent the codification of a mechanistic
understanding, thus having the capability to inform the decision-
making process. The drawback of these methodologies is that, in
general, they are rarely accompanied by a defined applicability
domain. In fact, if a chemical does not include SAs, this does not
necessarily indicate an absence of toxicity; instead it may points to
a lack of knowledge. Statistically based systems use a variety of
machine learning techniques to make associations between struc-
tural features and chemicals activity. These models are driven by
the data available to the computer algorithm, without any expert
supervision, in principle. In this case, whilst mechanistic under-
standing is often not straightforward, negative predictions are gen-
erally more accurate. Hybrid methods, by definition, integrate
both expert knowledge and statistically derived rules, trying to
overcome disadvantages of both approaches. In practice, the dis-
tinction among the methods is seldom absolute.
In recent years, there is a great effort to put these nontesting
methods into a more structured perspective, through the investi-
gation and formalization of integrated approaches to testing and
assessment (IATA). IATA are structured approaches that integrate
and weight information from different methodological approaches,
such as (Q)SAR methodologies, in chemico/in vitro/ex vivo/in

vivo testing, and omic technologies [15, 16]. IATA also have the
potential to be mechanistically informed, by the use of Adverse
Outcome Pathway (AOP) concept in their development [17].
3 Predicting Genotoxicity and Carcinogenicity
Concerning genotoxicity and carcinogenicity prediction, different

tools exist, both in the public domain and as commercial softwares
(Table 2). A detailed description of the models is out of the scope
of this chapter. However, the reader could refer to many excellent
papers (e.g., [12, 18–21]).
In deciding which tool/model is suitable for the prediction of
genotoxicity and/or carcinogenicity for a particular purpose, there
are some key elements to bear in mind. Below some elements,
inspired by the OECD Validation Principles for (Q)SARs [22], are
detailed.
The definition of the endpoint. The number of tools and models
predicting Ames test mutagenicity is remarkably higher with
respect to the other endpoints. This can be attributed to (1) the
larger availability of experimental data; (2) the deeper comprehen-
sion of the underlying mechanisms and consequently (3) the best
performances of the predictive systems; (4) the usefulness of the
test in the prescreening of carcinogenesis.
Concerning the prediction of carcinogenic chemicals, a limited
number of models is available. In particular, the limitation is due to
the difficulty in the elucidation and codification of nongenotoxic
mechanisms of carcinogenic action. In practice, for the prediction
of this endpoint, integrated approaches (such as grouping
approach) appear to be more suitable, especially in regulatory
environments.
The possibility to go back to the rule definition. A transparent
model it is always desirable as to check the details of the prediction,
especially in particular conflicting cases. However, this eventuality
generally requires an expert judgment in order to make a
decision.
The boundaries of acceptability of the prediction. The chemical
to be predicted should fall in the applicability domain of the sys-
tem. Disregarding of such an issue could bring to meaningless
results. In the same time, in cases in which is not possible to define
an applicability domain, special care must be placed in the interpre-
tation of the results, particularly for negative predictions. Also in
this case an expert advice should be considered.
Availability of measures of robustness. Especially in critical cases,
e.g., in the presence of discordant predictions from different mod-
els, measures of goodness-of-fit associated to the particular rule/
alert could help in decision-making. An example is the positive
Table 2
454
Expert systems for carcinogenicity and genotoxicity prediction, a nonexhaustive list
Software Availability Link Relevant features

ACD/Percepta Impurities Suite Commercial http://www.acdlabs.com/products/ Predictive models for the genotoxic and carcinogenic
percepta/impurities.php potential of impurities and degradants.
AMBIT (IdeaConsult Ltd.) Freely available https://ambitlri.ideaconsult.net/tool2 Knowledge-based expert systems for toxicity and
metabolism. Several in silico prediction models (e.g.,
Toxtree) are integrated in AMBIT.

ChemTunes ToxGPS (Altamira Commercial https://www.mn-am.com/ Knowledge based software including in vitro and in vivo
LLC and Molecular Networks toxicity information and in silico models for Ames
GmbH) mutagenicity, chromosome aberration, and in vivo
micronucleus
Danish QSAR Predictions Freely available http://qsar.food.dtu.dk Estimates from more than 200 (Q)SARs from free and
database (DK EPA) commercial platforms, including predictions for in vivo
and in vitro genotoxicity and carcinogenicity in rat/
mouse in male/female
DEREK Nexus Lhasa Ltd. Commercial https://www.lhasalimited.org/ Knowledge based system for prediction of carcinogenicity,
products/derek-nexus.htm mutagenicity, genotoxicity
Sarah Nexus Lhasa Ltd. Commercial https://www.lhasalimited.org/ Statistical software for mutagenicity predictions
products/sarah-nexus.htm
HazardExpert Pro and ToxAlert Commercial http://www.compudrug.com Rule-based systems for different endpoints, including
(CompuDrug Ltd) mutagenicity, carcinogenicity.
LAZAR (in silico toxicology, Freely available https://lazar.in-silico.de/predict Statistically based software, includes model for
GmbH) mutagenicity and carcinogenicity.
Leadscope Model Applier Commercial https:// Statistical models and expert alerts for different endpoints,
(Leadscope, Inc.) www.leadscope.com/model_appliers/ including genotoxicity and carcinogenicity
MolCode Toolbox and REACH Commercial http://molcode.com/ Statistical algorithms for the prediction of different
QSAR (Molcode Ltd) http://reachqsar.com/ endpoints, including Ames mutagenicity and
carcinogenicity.
Multicase CASE Ultra Models Commercial http://www.multicase.com/ Statistical based models for toxicity prediction, including
(MultiCASE Inc.) mutagenicity, genotoxicity, and carcinogenicity.
OECD QSAR Toolbox Freely available http://toolbox.oasis-lmc.org/ Includes “profilers” associated with genotoxicity and
carcinogenicity, and databases of experimental data
Oncologic (US EPA) Freely available https://www.epa.gov/ Knowledge based predictions of carcinogenicity.
reviewing-new-chemicals-under-toxic-
substances-control-act-tsca
PASS (Institute of Biomedical Commercial http://www.pharmaexpert.ru/ Statistically based classification models for mutagenic and
Chemistry of the Russian passonline/ carcinogenic effects, among others.
Academy of Medical Sciences,
Moscow)
T.E.S.T. (US EPA) Freely available https://www.epa.gov/chemical- The software includes models from different external
research/ sources, predicting different endpoints, including Ames
toxicity-estimation-software-tool-test mutagenicity
TIMES (OASIS, LMC) Commercial http://oasis-lmc.org/products/ Hybrid expert system for predicting toxicity of chemicals
software/times.aspx resulting from their metabolic activation. Contains a
suite of models for different endpoints, including
mutagenicity and genotoxicity endpoints
TOPKAT (Accelrys) Commercial http://accelrys.com/products/ Global model for Ames mutagenicity
collaborative-science/biovia-
discovery-studio/qsar-admet-and-
predictive-toxicology.html
Toxtree (EU JRC—Ideaconsult Freely available https://eurl-ecvam.jrc.ec.europa.eu/ Includes SARs for mutagenicity, carcinogenicity, and the
Ltd.) https://apps.ideaconsult.net/data/ui/ in vivo micronucleus assay
toxtree
VEGA (Istituto di Ricerche Freely available https://www.vegahub.eu In silico models and tools assisting the evaluation of
Farmacologiche Mario Negri) different endpoints, including mutagenicity and
carcinogenicity
455
predictivity value associated to each rule/alert. This number mea-

sures the probability for a chemical to be toxic once a SA has been
detected, thus accounting for the intrinsic prediction ability of a
rule/alert.
Association with the underlying mechanism. Mechanistic under-
standing is a key factor to enhance the confidence and the accept-
ability of the prediction. It is of particular importance in improving
a QSAR result on complex endpoints (e.g., carcinogenicity), pro-
viding a ground for interaction and dialogue between model devel-
opers, toxicologists and regulators.
Among the tools freely available in the public domain, it is
worth mentioning a few representative systems suitable for the
genotoxicity/carcinogenicity endpoints, helping in clarify the key
elements described above. These tools are particularly relevant as
they stem directly from regulatory contexts.
3.1 Toxtree The European Commission Joint Research Centre (JRC), formerly
through the European Chemicals Bureau (ECB) and now through
the EU Reference Laboratory for alternatives to animal testing
(EURL ECVAM), has a long-standing commitment in activities
for the development, assessment and application of computational
methods to predict the potential toxicological effects of
chemicals.
JRC is curating an inventory of QSAR models, available at
http://qsardb.jrc.it/qmrf/. The database provides key informa-
tion on the models, reported along a standard template (the QSAR
Model Reporting Format, QMRF). The JRC also commissioned
the development of freely available (Q)SAR tools, including
Toxtree, a computer program capable of estimating different types
of toxic hazard by applying decision tree approaches.
Toxtree (v2.6.13) is an open source software application devel-
oped by Ideaconsult Ltd under a Joint Research Centre contract
[23]. The tool is available for download and documentation at:
https://eurl-ecvam.jrc.ec.europa.eu/laboratories-research/pre-
dictive_toxicology/qsar_tools/toxtree and http://toxtree.source-
forge.net. Modules encoding rulebases for carcinogenicity
(genotoxic and nongenotoxic), and genotoxicity endpoints (Ames
test mutagenicity and in vivo micronucleus) have been curated by
Istituto Superiore di Sanità (Rome, Italy). Each module consists of
a refined compilation of SAs derived from mechanistic knowledge
[6, 24–26]. The open source software allows the user to go in the
deeper details of the rules implemented with full transparency.
Although it is not possible to define an applicability domain, the
alerts, when feasible, are accompanied by detoxifying chemical
functionalities, helping in discriminating positive/negative out-
comes in case of substances belonging to the same chemical class.
However, since neither all the possible mechanisms of toxicity are
coded in the alerts, nor all the possible exceptions to the known
mechanisms, the application of the SAs have to be considered as a

probabilistic approach. Measures of positive predictivity performed
on different data sets are available as supplementary information in
the QSAR Toolbox implementation of these lists, as well as in a
number of publications [24–27]. The in vitro mutagenicity mod-
ule of Toxtree is participating to the AMES/QSAR International
Collaborative Study, promoted by Dr. M. Honma of the Japanese
National Institute of Health Sciences (http://www.nihs.go.jp/
dgm/amesqsar.html). This project will allow many prediction sys-
tems to test and refine the rules for Ames test mutagenicity, giving
access to more than 12,000 new experimental data.
3.2 Oncologic™ In the evaluation and regulation of new and existing chemicals, the
US EPA Office of Pollution Prevention and Toxics (OPPT) makes
routinely use of QSAR methods to assist in evaluating genotoxicity
and carcinogenicity endpoints, among others. OPPT has also
developed and made available to the public a number of (Q)SAR
tools that are used in regulating substances under Toxic Substances
Control Act (TSCA). Among them, it is worth mentioning the
Oncologic software for the carcinogenicity prediction.
OncoLogic is a publicly available, rules-based expert approach
for carcinogenicity potential prediction, developed by the US
Environmental Protection Agency (EPA) (https://www.epa.gov/
tsca-screening-tools/oncologictm-computer-system-evaluate-car-
cinogenic-potential-chemicals). It is a computerized system that
mimics the thinking and reasoning of human experts based on
their knowledge regarding toxicological effects of certain classes of
compounds [28, 29]. Output information includes a prediction of
the carcinogenic potential of the chemical, expressed semiquantita-
tively, and the underlying scientific rationale in the form of a report.
This document is specific for each evaluation, keeping tracks of the
chemical structure submitted together with the details of the rules
that are used to assign a level of carcinogenic concern (namely
Low, Marginal, Low–Moderate, Moderate, High–Moderate,
High). Only a finite number of chemical structures may be entered
into the system and evaluated, thus defining intrinsically a domain
of applicability. Chemicals are classified into predefined potentially
reactive organic classes, to which is assigned a basic level of con-
cern. The evaluation considers the effect of different types of sub-
stituent in modulating the initial concern, potentially giving rise to
a negligible concern, which in practice equals a negative prediction
[18]. These features permit the integration of the information pro-
vided, together with different types of evidence and data, in a wider
risk assessment framework.
3.3 QSAR Toolbox Starting in the 1990s, the Organization for Economic
Cooperation and Development (OECD) is actively committed
in promoting and improving regulatory acceptance of QSAR
methods (http://www.oecd.org/chemicalsafety/risk-assessment/
oecdquantitativestructure-activityrelationshipsprojectqsars.htm).
As a result of the international cooperation among OECD mem-
ber countries, a considerable amount of scientific work has been
produced. Milestones for users and developers of QSAR method-
ologies are (1) the final drafting of the Validation Principles for
[30], (2) the Guidance on Grouping of Chemicals [11], (3) the
OECD QSAR Toolbox (in cooperation with the European
Chemical Agency—ECHA), and (4) the Fundamental And
Guiding Principles For (Q)SAR Analysis Of Chemical Carcinogens
with Mechanistic Considerations [31].
The OECD (Q)SAR Toolbox is a standalone free software
application developed by Laboratory of Mathematical Chemistry
(Burgas, Bulgaria) under the coordination of OECD and ECHA
(www.qsartoolbox.org). It was thought with the purpose of facili-
tating practical application of QSAR approaches within regulatory
frameworks. Crucial to the workflow that culminates in the data
gap filling, is the grouping of chemicals into chemical categories
[32]. The Toolbox incorporates information and tools from many
different sources, such as toxicological databases, expert rulebases,
QSAR models, chemoinformatic and statistical tools, giving rise to
a very powerful and comprehensive instrument, one of a kind in
the public domain. The flexibility of the tool allows for exploiting
different levels of application, depending on the final purpose and
on the experience of the user. These are well summarized as
“Typical Actions performed by the Toolbox” in the brochure
(https://www.qsartoolbox.org/it/support):
(a) Describes the structure of a chemical.
(b) Indicates if a chemical is included in national/regional regula-
tory inventories or existing chemical categories.
(c) Searches for available experimental results for the chemical of
interest.
(d) Explores a chemical list for possible similar chemicals.
(e) Groups chemicals based on mechanism of action and/or struc-
tural similarity.
(f) Groups chemicals based on a common metabolite.
(g) Enables exclusion of different chemicals from the group.
(h) Extracts experimental data for similar chemicals.
(i) Fills data gaps for chemicals using read-across, trend analysis
or QSAR models, where applicable.
(j) Designs a data matrix of a chemical category for printing/
exporting results.
(k) Connects to IUCLID software for direct data exchange.
(l) Generates reports.
The QSAR Toolbox, firstly released in March 2008, is now at

version 4.2 in February 2018. In particular, version 4 of the soft-
ware represents a turning point toward enhanced functionalities,
such as improved user friendliness in advising of suitable tools for
each particular endpoint, automated and standardized workflows
that guide in the predictions, and new features for the manipula-
tion of the data matrix and the reporting.
Concerning the prediction of genotoxicity and carcinogenicity
endpoints, many information and tools are available:
Prediction profilers: Alerts by ISS (same lists available in
Toxtree) for carcinogenicity (genotox and nongenotox) and in
vitro (Ames test)/in vivo (Micronucleus) mutagenicity; Oncologic
Primary classification; DNA alerts for Ames and DNA Alerts for
Chromosomal Aberration and Micronucleus Test by OASIS;
Protein binding Alerts for Chromosomal Aberration by OASIS.
Databases: ECHA CHEM, CPDB, ISSTOX Cluster, ECVAM
Genotoxicity and Carcinogenicity, Genotoxicity and Micronucleus
OASIS, Toxicity Japan MHLW, ToxRefDB.
4 Chemical Toxicity Databases
A critical issue at the basis of QSARs development and application

is the availability of high quality experimental data. In Grouping
approaches the categorization and prediction is based directly on
the experimental data available for similar substances, thus their
adequacy and accuracy is of outmost importance. On the other
hand, in QSAR models, experimental data are fundamental for the
validation and refinement of expert rules and for the derivation of
statistically based rules. In Table 3 databases for carcinogenicity
and genotoxicity endpoints are collected. The list is representative
of what is publicly available, but not inclusive of all possible.
Since all predictive approaches in toxicology require a highly
structured information as a starting point, the definition of ontol-
ogy is a crucial task in order to standardize and organize the chem-
ical and toxicological databases on which the predictive toxicology
methods build on and to improve the interoperability between
toxicology resources. There is an important need for public stan-
dardized toxicology vocabularies to support the QSARs develop-
ment, category formation, and read-across [33–36].
At present, several existing ontologies and standards initiatives
are available and can contribute—to different extents—to the toxi-
cological data standardization. The Open Biomedical Ontology
(OBO) Foundry (http://www.obofoundry.org/) and the
BioPortal (https://bioportal.bioontology.org/) are two impor-
tant sources of ontologies for reuse. Controlled toxicological
vocabularies and interrelations (ontologies) based on the OECD
Harmonised Templates (http://www.oecd.org/ehs/templates/)
460
Table 3
Databases for carcinogenicity and genotoxicity endpoints in the public domain
Database/link Description Main features and use

ACToR Aggregated Computational Toxicology Online Resource (AcTOR); Chemical toxicity data
https://www.epa.gov/chemical-research/ ACToR aggregates data from over 1000 public sources on over 500,000
actor chemicals and is searchable by chemical name, other identifiers and by

chemical structure.
Included in US EPA chemical toxicity database
ATSDR Agency for Toxic Substances and Disease Registry (ATSDR); open access Toxicity experimental data.
https://www.atsdr.cdc.gov/about/index. database, includes toxicological profiles for the hazardous substances
html# including genotoxicity. Searchable by chemical name and other
identifiers
CCRIS Chemical Carcinogenesis Research Information System (CCRIS) by the Toxicity experimental data.
https://toxnet.nlm.nih.gov/cgi-bin/sis/ National Cancer Institute (NCI); contains over 9000 chemical records
htmlgen?CCRIS (from the years 1985–2011) with animal carcinogenicity, mutagenicity,
tumor promotion, and tumor inhibition test results. Test results have
been reviewed by experts in carcinogenesis and mutagenesis. Searchable
by chemical name, CAS number and other advanced searches. Included
in TOXNET database.
ChemIDplus Dictionary of over 400,000 chemicals (names, synonyms, and structures). Chemical database
http://chem.sis.nlm.nih.gov/chemidplus/ Searchable by name, chemical name, other identifiers and by chemical
chemidlite.jsp structure; advanced search available. Included in TOXNET database
CPDB Carcinogenic Potency Database (CPDB) by University of California Carcinogenicity data
https://toxnet.nlm.nih.gov/cpdb/ (Berkeley); contains results of 6540 chronic, long-term animal cancer
tests on 1547 chemicals (1980–2011). Data downloadable in easy-to-
use format (pdf, xls, and txt), and searchable by chemical name, CAS
number, or author. “Chemically-indexed” in the DSSTox database.
Searchable by chemical name, CAS number, or advanced searches.
Included in TOXNET database and in OECD QSAR Toolbox
Dashnoard Chemistry data for over 700,000 chemicals and includes chemical High quality information on
https://comptox.epa.gov/dashboard structures, experimental and predicted physicochemical and toxicity chemicals, including
data. Simple search or advanced search is available. Included in US EPA toxicity data
chemical toxicity database
DSSTox Distributed Structure Searchable Toxicity; downloadable, structure- Chemical toxicity data
https://www.epa.gov/chemical-research/ searchable, standardized chemical structure files associated with chemical
distributed-structure-searchable-toxicity- inventories or toxicity data sets of environmental relevance. DSSTox
dsstox-database provides high quality chemical structures and annotations in association
with toxicity data.
Included in US EPA chemical toxicity database
ECHA database Source of information on the chemicals manufactured and imported in Toxicity experimental data
https://echa.europa.eu/ Europe. It covers their hazardous properties, classification and labeling,
information-on-chemicals and information on how to use them safely.
Searchable by Name, EC or CAS NO; advanced search criteria are available
(such as numerical identifiers, structural information, properties of
concern, classification details, and uses and exposure). Included in
OECD QSAR Toolbox
eChemPortal The Global Portal to Information on Chemical Substances (eChemPortal) Portal of information on
http://www.echemportal.org/echemportal/ by OECD; free public access to information on chemicals properties and chemicals.
index.action links to collections of information prepared for government chemical Direct links to reports and
review programmes at national, regional, and international levels. Data datasets.
on Physical Chemical Properties, (Eco)toxicity, Environmental Fate and
Behaviour, Classification (national/regional classification and GHS),
exposure and use information on chemicals. eChemPortal uses the
OECD Harmonised Templates (www.oecd.org/ehs/templates) for data
formatting. Searchable by chemical and numerical identifiers.
EFSA pesticide genotoxity Database specific for pesticide active substances and their metabolites; Toxicity experimental data
https://data.europa.eu/euodp/data/ includes genotoxicity endpoints. Downloadable in xls format
storage/f/2017-07-19T142131/
GENOTOX%20data%20and%20dictionary.
xls
(continued)
461
Table 3
(continued)
462
EURL ECVAM Genotoxicity and Database by European Reference Laboratory for Alternatives to Animal Toxicity experimental data
Carcinogenicity Consolidated Database Testing. Curated genotoxicity and carcinogenicity data for 726 Ames
https://eurl-ecvam.jrc.ec.europa.eu/news/ positive chemicals compiled from variety of sources. Toxicity results
launch-of-consolidated-database-of-ames- critically reviewed. Downloadable in xls format. Included in OECD
positive-chemicals QSAR Toolbox
GENE-TOX Genetic Toxicology Data Bank. Peer-reviewed genetic toxicology test data Toxicity experimental data
https://www.nlm.nih.gov/databases/ for over 3000 chemicals (1991–1998). Included in TOXNET.
download/genetox.html
HSDB Hazardous Substances Data Bank. Peer-reviewed toxicology data for over Toxicology database
https://toxnet.nlm.nih.gov/newtoxnet/ 5000 hazardous chemicals. Included in TOXNET database
hsdb.htm
IARC Open access International Agency for Research on Cancer (IARC) Experimental data for Risk
http://monographs.iarc.fr/ monographs including carcinogenicity classification Assessment
IPCS INCHEM International Programme on Chemical Safety (IPCS) and Canadian Centre Tool for chemical risk
http://www.inchem.org/ for Occupational Health and Safety (CCOHS). assessment and
Open access to thousands of searchable full-text documents on chemical management
risks and the sound management of chemicals; searchable by CAS
number, Chemical name or Synonym; downloadable the reports in pdf
form
IRIS Integrated Risk Information System. Hazard identification and dose- Risk Assessment
https://toxnet.nlm.nih.gov/newtoxnet/iris. response assessment for over 500 chemicals. Included in TOXNET
htm database
ISSTOX Cluster of databases, by Istituto Superiore di Sanità; each database contains Cluster of chemical relational
http://www.iss.it/ampp/index. experimental results relative to various types of chemical toxicity, such as databases for toxicity
php?id=233&tipo=7&lang=1 long-term carcinogenicity on rodents (ISSCAN), in vitro (ISSSTY) and endpoints
in vivo mutagenicity (ISSMIC), cell transformation (ISSCTA) and
long-term carcinogenicity on rodents and in vitro mutagenicity of
biocides and plant protection products (ISSBIOC). Toxicity results
critically reviewed. Downloadable in xls, sdf, and pdf formats. Included
in OECD QSAR Toolbox
ITER International Toxicity Estimates for Risk. Includes risk information for Risk Assessment
https://toxnet.nlm.nih.gov/newtoxnet/iter. over 600 chemicals from authoritative groups worldwide. Included in
htm TOXNET database
JECDB (Toxicity Japan MHLW) Open access Japanese Existing Chemical Data Base (JECDB) containing Toxicity experimental data
http://dra4.nihs.go.jp/mhlw_data/jsp/ toxicity data (mostly in Japanese) on high production volume chemicals
SearchPageENG.jsp The database contains experimental results from single dose toxicity test
and mutagenicity test results performed under Japan’s Existing
Chemicals Programme. Included in OECD QSAR Toolbox (as Toxicity
Japan MHLW)
NTP National Toxicology Program (NTP) U.S. Department of Health and Historical control database;
https://ntp.niehs.nih.gov/results/ Human Services; NTP evaluates substances that pose a hazard to U.S. Primary toxicology
summaries/chronicstudies/index.html people, for a variety of health-related effects, using rodent models for experimental data
study and protocols specifically designed to fully characterize the toxic
potential; studies are reported as technical reports. Searchable by
chemical name or CAS number, downloadable the reports in pdf form.
“Chemically-indexed” in the DSSTox database
OECD QSAR Toolbox OECD software application to identify and fill (eco)toxicological data gaps Platform that incorporates
https://www.qsartoolbox.org/ for chemicals hazard assessment. The Toolbox contains databases with modules and databases
results from experimental studies. from other sources
Open Food Tox EFSA chemical hazards database; a compilation of chemical and Chemical and toxicological
https://dwh.efsa.europa.eu/bi/asp/Main. toxicological information on chemicals assessed by EFSA since its information on chemicals
aspx?rwtrep=400 creation and included in already published scientific opinions. Summary assessed by EFSA
data sheets for each individual substance downloadable in pdf or xls
format.
RTECS Registry of Toxic Effects of Chemical Substances (RTECS) by The Toxicity data
https://www.cdc.gov/niosh/rtecs/default. National Institute for Occupational Safety and Health (NIOSH);
html collection of basic toxicity information (including mutation and
tumorigenic studies) on substances used in industrial and house
situation, extracted from the open scientific literature.
No attempt to evaluate the studies cited

(continued)
463
Table 3
464
(continued)
PAN Pesticide Database (Eco)Toxicity and regulatory information for pesticide active ingredients Toxicity experimental data
http://pesticideinfo.org/Index.html and their transformation products, as well as adjuvants and solvents used
in pesticide products
PubChem National Center for Biotechnology Information (NCBI); depositor system Toxicological/biological
https://pubchem.ncbi.nlm.nih.gov/ with standardized data from toxicological/biological database sources database

and from literature. Contains chemical structure-annotated data
submissions, with summary bioassay data
TOXNET TOXicology Data NETwork by US National Library Medicine (NLM); Toxicity cluster of databases,
https://toxnet.nlm.nih.gov/ databases on toxicology, hazardous chemicals, environmental health, and including CCRIS, CPDB,
toxic releases. Searchable within and across the databases by chemical GENE-TOX, IRIS, ITER,
name, CAS number, molecular formula, classification code, locator code, HSDB, ChemIDplus
and structure or substructure.
ToxRefDB The Toxicity Reference Database (ToxRefDB) contains thousands of Animal toxicity studies
https://www.epa.gov/chemical-research/ animal toxicity studies on hundreds of chemicals from 30 years of animal
toxicity-forecaster-toxcasttm-data toxicity studies. Included in US EPA chemical toxicity database and in
OECD QSAR Toolbox
have been developed for several toxicological endpoints including

carcinogenicity within the ECHA/OECD ontology program
(https://www.qsartoolbox.org/it/ontologies). The objective was
the improving he the integration of experimental data from differ-
ent sources and facilitating toxicity prediction in the OECD QSAR
Toolbox software.
5 Regulatory Frameworks
In what follows, we describe some relevant regulatory framework,

which explicitly consider the use of QSARs methodologies for car-
cinogenicity and/or genotoxicity endpoint. These initiatives trig-
ger a virtuous circle that stimulates the improvement of QSARs to
be used in chemicals risk assessment, eventually leading to an
increasing regulatory acceptance.
5.1 The REACH One of the fundamental principles of EU REACH Regulation is

Experience that “testing on vertebrate animals shall be undertaken only as a
last resort.” To reduce testing, when inadequate information is
present for a substance, REACH foresees and stimulates (1) shar-
ing of information; (2) use of alternatives to testing on animals
and, in particular, nontesting approaches. Annex XI states the gen-
eral provisions for the use of Weight of Evidence, (Q)SARs,
Grouping and Read-across approaches, further elaborated in chap-
ter R.6 of the “Guidance on Information Requirements and
Chemical Safety Assessment” [37].
As a remarkable result, registrants used extensively nontesting
methods to fill data gaps for hazard or risk assessment. In the third
report on the use of alternatives to animal testing [38], ECHA
documented the use of adaptations to the standard information
requirements, by analyzing substances submitted in the period
2009–2016. In the registration dossiers, the weight of evidence
approach was used for around 15% of substances across all end-
points, whilst usage of QSAR models it is confined to a maximum
of 10%, with the only exception of Bioaccumulation endpoint (for
31% of substances). The fraction of substances with Read-across
adaptations varies strongly among the endpoints, between 8% and
53%. Notably, Read-across usage is considerable for carcinogenic-
ity (around 35%) and genotoxicity (around 45%) endpoints.
A detailed overview of the submitted “endpoint study records”
(ESRs) in dossiers from 2009 to 2016 is presented in Fig. 1. To
gain insight of the overall picture, the histogram shows ESRs pro-
portion for each endpoint and for the three different adaptations
of interest, normalized by the total number of ESRs presented for
each endpoint (computed from [38]). In absolute values, Read-
across was used in 14,000 ESRs for genotoxicity (in vitro/in vivo)
Fig. 1 Overview of the submitted “endpoint study records” (ESRs) in REACH dossiers from 2009 to 2016 [38]
and nearly 3000 ESRs for carcinogenicity. By contrast, testing pro-

posals were submitted only for genotoxicity in vivo, in 60 ESRs.
What is reported above refers to the statistics of adaptation
options as submitted in IUCLID registration dossiers. Detailed
information concerning acceptability and acceptance of the pro-
posed methodologies is not available. However, based on ECHA
evaluation reports [39] and in our own experience, criticalities
observed frequently in Read-across adaptations are the following:

(a)
Shortcomings in the Read-across hypothesis (poorly
documented).
Read-across hypothesis should be clearly stated and dis-
cussed. Explanation of the reasoning and evidences supporting
the read-across are always responsibility of the registrant. All
the arguments have to be justified and adequately
documented.
(b) Deficiencies (incompleteness/inconsistencies) in the data

matrix.
Experimental results collected in the data matrix, should
allow outlining of the information consistency and should not
contradict the read-across hypothesis. Furthermore, high

quality experimental data should be provided/referenced.
(c) Insufficient analysis of similarities/differences of substances
under consideration.
Structural similarity is at the basis of the read-across hypoth-
esis, but is not sufficient per se to justify the possibility to pre-
dict the toxicological effect(s) of the target substance by
read-across. The implications for the predicted property
resulting from similarities but also differences in the chemical
structure have to be discussed thoroughly.
(d) Lack of data on mechanisms of toxicity and ADME

properties
The inclusion of qualitative and quantitative (toxico)kinetic
information represent scientific evidence that enhances robustness
of read-across and improves its acceptability [31].
Above all, it has to be kept in mind that, for any adaptation used,
transparency of the methods is a key feature for regulatory
acceptance.
The results shown demonstrates effectiveness of REACH reg-
ulation in encouraging usage of QSAR methodologies.
Furthermore, ECHA is continuously committed in promoting the
adequate use of nontesting methods. In addition to sponsoring the
development and use of the QSAR Toolbox, many valuable docu-
ments of very high scientific and technical value are available on
the web (e.g., [37–42]). This material is very useful for those who
want to train in the application of QSAR methods in regulatory
frameworks.
5.2 EFSA Panel In the risk assessment of pesticides, according to the requirements
on Plant Protection of Commission Regulation (EU) No 283/2013, a comprehensive
Products and Their toxicological dossier is developed for active substances, tradition-
Residues ally relying on extensive in vivo and in vitro testing. On the other
hand, for substances resulting from pesticides metabolic and deg-
radation processes, toxicological characterization is often very lim-
ited. In this context, nontesting methods have been proposed for
the identification of all possible metabolites and degradates of toxi-
cological relevance that have to be included in the residue defini-
tion for risk assessment.
The usage of nontesting approaches for pesticides assessment
is under evaluation and adoption also in other regulatory frame-
works. Notably, the US EPA’s Office of Pesticide Programs (OPP)
and the Pest Management Regulatory Agency (PMRA) of Health
Canada are developing common strategies to make use of IATA
and (Q)SAR techniques to pesticide assessment [43].
In 2016 the European Food Safety Authority (EFSA) Panel on
Plant Protection Products and their Residues (PPR), published a
practical Guidance on the establishment of the residue definition
Fig. 2 Schematic representation of the EFSA approach for genotoxicity assessment of metabolites, in the
procedure of derivation of the residue definition for dietary risk assessment. Adapted from [44]
for dietary risk assessment, for the harmonization of the procedure

for the derivation of residue definition [44]. The Guidance docu-
ment describes a three modules procedure (Fig. 2) for the inclu-
sion/exclusion of a compound, based on their toxicity and potential
for dietary exposure.
The primary step is represented by the genotoxicity assessment
of the active substance and all the identified metabolites. In this
step, if experimental studies are not available or not sufficient, the
assessment should be assisted by the application of QSAR,
Grouping and read-across approaches. The procedure for the in
silico evaluation illustrated in the guidance is sketched below:
(a) Application of two independent QSAR models (e.g., based on
different training sets and/or algorithms as knowledge based
and statistically based models) for each genotoxicity endpoint,
where possible;
(b) Grouping/read-across analysis of all metabolites, irrespective
to the QSAR models results;
(c) Weight of evidence assessment of all collected evidences;
(d) Definition of the genotoxicity concern.
If no genotoxic concern is raised in this in silico evaluation, the
substances go to the next step, the general toxicity assessment.
Differently, in presence of genotoxicity potential or if no clear
conclusion can be made, after a TTC evaluation of exposure

(optional), a testing battery on representative metabolites should
be performed, according to the Opinion on genotoxicity testing
strategies applicable to food and feed safety assessment [3].
This EFSA guidance represents a solid ground for encouraging

the application of QSAR methods in pesticides genotoxicity assess-
ment. Furthermore, PPR panel is promoting various activities in
order to explore, facilitate and improve the applicability of nontest-
ing methods to pesticides. The initiatives are particularly focused in
expand QSAR models domains of applicability that are often not
sufficiently representative of the structures and mechanisms of
action associated with pesticide active ingredients and metabolites.
As a result, a number of very instructive scientific publications have
been released [45, 46].
In addition, a call on the completion of a database of experimen-
tal data on genotoxicity endpoints for pesticide active substances and
their metabolites has been recently concluded. As a result of the
project, a database has been released (https://data.europa.eu/
euodp/en/data/dataset/database-pesticide-genotoxicity-end-
points). These data will help in improving knowledge on mechanism
of pesticides genotoxic action giving the opportunity to refine QSAR
tools for the applicability to pesticides.
5.3 The ICH M7 Recently, the ICH M7 guideline for the “Assessment and control
Guideline of DNA reactive (mutagenic) impurities in pharmaceuticals” was
finalized [47]. This guideline, elaborated within the mission of the
ICH to support the development and registration of safe and effec-
tive medicines, provides a framework for the minimization of the
risk of human exposure to DNA-reactive chemicals.
As shown in Fig. 3, the strategy to detect potentially carcino-
genic impurities is driven by two key elements: the analysis of
Fig. 3 Schematic representation of the ICH-M7 strategy to detect potentially carcinogenic impurities. Adapted
from [47]
existent experimental data and, when lacking, the use of QSAR

tools. Considering these elements, the hazard assessment is per-
formed, leading to five distinct classes with different actions to be
undertaken for the subsequent control of the impurities. For classes
1, 2, and 3, risk characterization measures should be implemented,
with the definition of acceptable intakes (compound specific for
Class 1 or TTC-based for 2 and 3). Impurities fallen in classes 4
and 5 should be treated as nonmutagenic. Notably, the strategy
recommends a computational toxicology assessment, based on the
use of two complementary (Q)SAR methodologies (expert rule-
based and statistically based), which is considered sufficient, in the
absence of adequate experimental data, to rule out the mutagenic-
ity concern.
Clearly, some difficulties have been experienced during the
practical implementation of the guideline, regarding:
(a) The judgment on the adequacy of the experimental data;
(b) Which QSAR tools to be considered complementary;
(c) How to decide an overall computational outcome in presence
of contradictory in silico results;
(d) The need of expert knowledge, for the final decision.
These issues stimulates the debate in the scientific community
and a number of very interesting papers on the topic have been
published, that could inform the correct implementation of the
procedure (e.g., [48–57]). In general, this guidance acts as a cata-
lyst for the scientific activity and represents another important step
forward in defining and promoting the regulatory use and accep-
tance of in silico predictions.
6 Notes
1. Definition of similarity between two or more substances is not

straightforward and deserves special attention [58]. In particu-
lar, as stated in [12]: “Chemical similarity is not limited to struc-
tural similarity but should consider the factors that drive a given
toxicity and how these can be linked back to chemical properties or
features, e.g. reactivity.”
References
1. Huff J, Haseman J (1991) Long-term chemical carcinogenesis studies. Annu Rev Pharmacol
carcinogenesis experiments for identifying Toxicol 31:621–652
potential human cancer hazards: collective 3. EFSA (2011) Scientific opinion on genotoxic-
database of the National Cancer Institute and ity testing strategies applicable to food and feed
National Toxicology Program (1976-1991). safety assessment. EFSA J 9:2379
Environ Health Perspect 96:23–31 4. Benigni R, Bossa C (2011) Mechanisms of
2. Huff J, Haseman J, Rall D (1991) Scientific chemical carcinogenicity and mutagenicity: a
concepts, value, and significance of chemical review with implications for predictive toxicol-
ogy. Chem Rev 111:2507–2536
5. OECD (2007) Detailed review paper on cell assessment (IATA). Regul Toxicol Pharmacol
transformation assays for detection of chemical 70:629–640
carcinogens. OECD Publishing, Paris. ENV/ 18. Benigni R, Battistelli CL, Bossa C,
JM/MONO(2007)18 Colafranceschi M, Tcheremenskaia O (2013)
6. Benigni R, Bossa C, Tcheremenskaia O, Mutagenicity, carcinogenicity, and other end
Battistelli CL, Giuliani A (2015) The Syrian points. Methods Mol Biol 930:67–98
hamster embryo cells transformation assay 19. Serafimova R, Gatnik MF, Worth A (2010)
identifies efficiently nongenotoxic carcinogens, Review of QSAR models and software tools for
and can contribute to alternative, integrated predicting genotoxicity and carcinogenicity.
testing strategies. Mutat Res Genet Toxicol EUR - Scientific and Technical Research
Environ Mutagen 779:35–38 Reports. EUR 24427 EN
7. OECD (2016) Guidance document on the 20. Worth A, Barroso J, Bremer S, Burton J, Casati
in vitro Bhas 42 cell transformation assay S, Coecke S, Corvi R, Desprez B, Dumont C,
(BHAS 42 CTA). OECD Publishing, Paris. Gouliarmou V, Goumenou M, Gräpel R,
ENV/JM/MONO(2016)1 Griesinger C, Halder M, Roi AJ, Kienzler A,
8. OECD (2015) Guidance document on the Madia F, Munn S, Nepelska M, Paini A, Price
in vitro syrian hamster embryo (SHE) cell A, Prieto P, Rolaki A, Schäffer M, Triebe J,
transformation assay. OECD Publishing, Paris. Whelan M, Wittwehr C, Zuang V (2014)
ENV/JM/MONO(2015)18 Alternative methods for regulatory toxicol-
9. Cherkasov A, Muratov EN, Fourches D et al ogy – a state-of-the-art review. EUR - Scientific
(2014) QSAR modeling: where have you been? and Technical Research Reports. EUR 26797
Where are you going to? J Med Chem 21. Cassano A, Raitano G, Mombelli E, Fernández
57:4977–5010 A, Cester J, Roncaglioni A, Benfenati E (2014)
10. Nicolotti O, Benfenati E, Carotti A, Gadaleta Evaluation of QSAR models for the prediction
D, Gissi A, Mangiatordi GF, Novellino E of ames genotoxicity: a retrospective exercise
(2014) REACH and in silico methods: an on the chemical substances registered under
attractive opportunity for medicinal chemists. the EU REACH regulation. J Environ Sci
Drug Discov Today 19:1757–1768 Health C 32:273–298
11. OECD (2014) Guidance on grouping of 22. OECD (2007) Guidance document on the
chemicals, 2nd edn. OECD Publishing, Paris. validation of (quantitative) structure-activity
ENV/JM/MONO(2014)4 relationship [(Q)SAR] models, vol ENV/JM/
12. ECETOC (2012) Category approaches, Read- MONO(2007)2. OECD Publishing, Paris
across, (Q)SAR. Technical Report no. 116. Brussels 23. Patlewicz G, Jeliazkova N, Safford RJ, Worth
13. Benigni R, Bossa C (2008) Predictivity and AP, Aleksiev B (2008) An evaluation of the
reliability of QSAR models: the case of muta- implementation of the Cramer classification
gens and carcinogens. Toxicol Mech Methods scheme in the Toxtree software. SAR QSAR
18:137–147 Environ Res 19:495–524
14. Benigni R, Bossa C, Netzeva T, Worth A 24. Benigni R, Bossa C (2008) Structure alerts for
(2007) Collection and evaluation of (Q)SAR carcinogenicity, and the Salmonella assay sys-
models for mutagenicity and carcinogenicity. tem: a novel insight through the chemical rela-
EUR - Scientific and Technical Research tional databases technology. Mutat Res
Reports. EUR 22772 EN 659:248–261
15. OECD (2008) Report of a Workshop on 25. Benigni R, Bossa C, Jeliazkova N, Netzeva T,
Integrated Approaches to Testing and Worth A (2008) The Benigni/Bossa rulebase
Assessment (IATA). OECD Publishing, Paris. for mutagenicity and carcinogenicity - a mod-
ENV/JM/MONO(2008)10 ule of toxtree. EUR - Scientific and Technical
Research Reports. EUR 23241 EN
16. USEPA (2011) Integrated approaches to test-
ing and assessment strategy: use of new com- 26. Benigni R, Bossa C, Tcheremenskaia O (2013)
putational and molecular tools. FIFRA Nongenotoxic carcinogenicity of chemicals:
Scientific Advisory Panel Consultation US mechanisms of action and early recognition
Environmental Protection Agency, Office of through a new set of structural alerts. Chem
Pesticide Programs Rev 113(5):2940–2957. https://doi.
org/10.1021/cr300206t
17. Tollefsen KE, Scholz S, Cronin MT, Edwards
SW, de Knecht J, Crofton K, Garcia-Reyero N, 27. Benigni R, Bossa C, Tcheremenskaia O,
Hartung T, Worth A, Patlewicz G (2014) Battistelli CL, Crettaz P (2012) The new
Applying adverse outcome pathways (AOPs) to ISSMIC database on in vivo micronucleus and
support integrated approaches to testing and its role in assessing genotoxicity testing strate-
gies. Mutagenesis 27:87–92
28. Lai D, Woo Y-T (2005) OncoLogic. In: recommendations to registrants. European
Predictive toxicology. CRC Press, Boca Raton, Chemicals Agency
FL, pp 385–413 40. ECHA (2017) Read-across assessment frame-
29. Woo YTLD, Argus MF, Arcos JC (1998) An work (RAAF). European Chemicals Agency
integrative approach of combining mechanisti- 41. ECHA (2016) Practical guide – how to use
cally complementary short-term predictive and report (Q)SARs. European Chemicals
tests as a basis for assessing the carcinogenic Agenc (ECHA)
potential of chemicals. J Environ Sci Health C 42. ECHA (2016) Practical guide: how to use
Environ Carcinog Ecotoxicol Rev C alternatives to animal testing to fulfil the infor-
16(2):101–122 mation requirements for REACH registration.
30. OECD (2004) OECD Principles for the vali- European Chemicals Agency
dation, for regulatory purposes, of (quantita- 43. NAFTA (2012) (Quantitative) Structure
tive) structure-activity relationship models. Activity Relationship [(Q)SAR] Guidance
OECD Publishing, Paris Document. US Environmental Protection
31. OECD (2015) Fundamental and guiding prin- Agency, Technical Working Group on
ciples for (Q)SAR analysis of chemical carcino- Pesticides
gens with mechanistic considerations. OECD 44. EFSA-PPR (2016) Guidance on the establish-
Publishing, Paris. ENV/JM/MONO(2015)46 ment of the residue definition for dietary risk
32. Dimitrov SD, Diderich R, Sobanski T et al assessment. EFSA J 14:4549
(2016) QSAR Toolbox - workflow and major 45. EU-JRC (2010) Applicability of QSAR analysis
functionalities. SAR QSAR Environ Res:1–17 to the evaluation of the toxicological relevance
33. Benigni R, Battistelli CL, Bossa C, of metabolites and degradates of pesticide
Tcheremenskaia O, Crettaz P (2013) New per- active substances for dietary risk assessment.
spectives in toxicological information manage- EFSA Support Publ 7(5):50E
ment, and the role of ISSTOX databases in 46. EFSA-PPR (2012) Scientific opinion on evalu-
assessing chemical mutagenicity and carcinoge- ation of the toxicological relevance of pesticide
nicity. Mutagenesis 28:401–409 metabolites for dietary risk assessment. EFSA
34. Hardy B, Apic G, Carthew P, Clark D, Cook J 10(07):2799
D, Dix I, Escher S, Hastings J, Heard DJ, 47. ICH-M7 (2017) Assessment and control of
Jeliazkova N, Judson P, Matis-Mitchell S, Mitic DNA reactive (mutagenic) impurities in phar-
D, Myatt G, Shah I, Spjuth O, Tcheremenskaia maceuticals to limit potential carcinogenic risk.
O, Toldo L, Watson D, White A, Yang C http://www.ich.org/fileadmin/Public_Web_
(2012) Food for thought ... A toxicology Site/ICH_Products/Guidelines/
ontology roadmap. ALTEX 29(2):129–137 Multidisciplinary/M7/M7_R1_Addendum_
35. Hardy B, Apic G, Carthew P, Clark D, Cook Step_4_2017_0331.pdf
D, Dix I, Escher S, Hastings J, Heard DJ, 48. Greene N, Dobo KL, Kenyon MO, Cheung J,
Jeliazkova N, Judson P, Matis-Mitchell S, Mitic Munzner J, Sobol Z, Sluggett G, Zelesky T,
D, Myatt G, Shah I, Spjuth O, Tcheremenskaia Sutter A, Wichard J (2015) A practical applica-
O, Toldo L, Watson D, White A, Yang C tion of two in silico systems for identification of
(2012) Toxicology ontology perspectives. potentially mutagenic impurities. Regul
ALTEX 29:139–156 Toxicol Pharmacol 72:335–349
36. Tcheremenskaia O, Benigni R, Nikolova I, 49. Amberg A, Beilke L, Bercu J, Bower D, Brigo
Jeliazkova N, Escher SE, Batke M, Baier T, A, Cross KP, Custer L, Dobo K, Dowdy E,
Poroikov V, Lagunin A, Rautenberg M, Hardy Ford KA, Glowienke S, Van Gompel J, Harvey
B (2012) OpenTox predictive toxicology J, Hasselgren C, Honma M, Jolly R, Kemper
framework: toxicological ontology and R, Kenyon M, Kruhlak N, Leavitt P, Miller S,
semantic media wiki-based OpenToxipedia.
Muster W, Nicolette J, Plaper A, Powley M,
J Biomed Semantics 3(Suppl 1):S7 Quigley DP, Reddy MV, Spirkl HP, Stavitskaya
37. ECHA (2008) QSARs and grouping of chemi- L, Teasdale A, Weiner S, Welch DS, White A,
cals, vol R.6. Guidance on information require- Wichard J, Myatt GJ (2016) Principles and
ments and chemical safety assessment. procedures for implementation of ICH M7
Guidance for the implementation of REACH recommended (Q)SAR analyses. Regul Toxicol
38. ECHA (2017) The use of alternatives to test- Pharmacol 77:13–24
ing on animals for the REACH Regulation. 50. Barber C, Amberg A, Custer L, Dobo KL,
European Chemicals Agency Glowienke S, Van Gompel J, Gutsell S, Harvey
39. ECHA (2016) Evaluation under REACH J, Honma M, Kenyon MO, Kruhlak N, Muster
progress report 2016 – executive summary and W, Stavitskaya L, Teasdale A, Vessey J, Wichard
J (2015) Establishing best practise in the appli-
cation of expert review of mutagenicity under 55. Teasdale A (2017) Regulatory highlights. Org
ICH M7. Regul Toxicol Pharmacol Process Res Dev 21:1209–1212
73:367–377 56. Williams RV, Amberg A, Brigo A, Coquin L,
51. Barber C, Cayley A, Hanser T, Harding A, Giddings A, Glowienke S, Greene N, Jolly R,
Heghes C, Vessey JD, Werner S, Weiner SK, Kemper R, O’Leary-Steele C, Parenty A, Spirkl
Wichard J, Giddings A, Glowienke S, Parenty H-P, Stalford SA, Weiner SK, Wichard J (2016)
A, Brigo A, Spirkl H-P, Amberg A, Kemper R, It’s difficult, but important, to make negative
Greene N (2016) Evaluation of a statistics- predictions. Regul Toxicol Pharmacol 76(Suppl
based Ames mutagenicity QSAR model and C):79–86
interpretation of the results obtained. Regul 57. Sutter A, Amberg A, Boyer S, Brigo A,
Toxicol Pharmacol 76(Suppl C):7–20 Contrera JF, Custer LL, Dobo KL, Gervais V,
52. Barber C, Hanser T, Judson P, Williams R Glowienke S, Gompel JV, Greene N, Muster
(2017) Distinguishing between expert and sta- W, Nicolette J, Reddy MV, Thybaud V, Vock
tistical systems for application under ICH M7. E, White AT, Müller L (2013) Use of in silico
Regul Toxicol Pharmacol 84(Suppl systems and expert knowledge for structure-
C):124–130 based assessment of potentially mutagenic
53. Cartus A, Schrenk D (2017) Current methods impurities. Regul Toxicol Pharmacol
in risk assessment of genotoxic chemicals. Food 67:39–52
Chem Toxicol 106(Part B):574–582 58. Floris M, Manganaro A, Nicolotti O, Medda
54. Powley MW (2015) (Q)SAR assessments of R, Mangiatordi GF, Benfenati E (2014) A gen-
potentially mutagenic impurities: a regulatory eralizable definition of chemical similarity for
perspective on the utility of expert knowledge read-across. J Cheminformatics 6:39
and data submission. Regul Toxicol Pharmacol
71:295–300
Chapter 21
Stem Cell-Based Methods to Predict Developmental

Chemical Toxicity
Hiroki Takahashi, Xian-Yang Qin, Hideko Sone, and Wataru Fujibuchi
Abstract
Human pluripotent stem cells such as embryonic stem (ES) and induced pluripotent stem (iPS) cells,
combined with sophisticated bioinformatics methods, are powerful tools to predict developmental chemi-
cal toxicity. Because cell differentiation is not necessary, these cells can facilitate cost-effective assays, thus
providing a practical system for the toxicity assessment of various types of chemicals. Here we describe how
to apply machine learning techniques to different types of data, such as qRT-PCRs, gene networks, and
molecular descriptors, for toxic chemicals, as well as how to integrate these data to predict toxicity catego-
ries. Interestingly, our results using 20 chemical data for neurotoxins (NTs), genotoxic carcinogens (GCs),
and nongenotoxic carcinogens (NGCs) demonstrated that the highest and most robust prediction perfor-
mance was obtained by using gene networks as the input. We also observed that qRT-PCR and molecular
descriptors tend to contribute to specific toxicity categories.
Key words Embryonic stem cell, Chemical toxicity prediction, Developmental effect, Gene network,
Molecular descriptor, Multi-kernel support vector machine
1 Introduction
Chemical toxicity assessment by toxicogenomics study [1] is

important in various fields, including drug screening, food produc-
tion, and the chemical industry. Innovative machine learning
methods that use support vector machines (SVM), random forests,
or other techniques have been devised to predict neurotoxicity [2],
cardiotoxicity [3], nephrotoxicity [4], and other toxicities.
Biological devices that complement these excellent methods are
artificially generated human cells or tissues derived from ES or iPS
cells. More complicated systems with higher physical concepts such
as organ-on-a-chip [5] or body-on-a chip [6] are also under devel-
opment. However, in reality, it is not easy to control the differen-
tiation except toward cardiomyocyte or neuronal cell lineages, and
artificially generated human cells often show similar but not exactly
475
476 Hiroki Takahashi et al.
the same gene expression patterns as natural cells [7]. Therefore, it

may still require several years before making effective use of these
chips. In the past, the ES cell test (EST), which assesses the extent
of embryotoxicity in vitro by using both differentiated and undif-
ferentiated cells from mouse ES cells, was validated by the European
Center for the Validation of Alternative Methods (ECVAM) [8].
In our recent paper, we have further enhanced EST and shown
that it could be used to predict developmental toxicity categories
without having to differentiate human ES cells [9]. In this proto-
col, we focus on computational methods and describe the neces-
sary resources, programs, and procedures to predict chemical
toxicities based on a human ES cell system. We tested all combina-
tions of qRT-PCR, gene networks, and molecular descriptors as
the inputs of SVMs, and the results indicated that gene networks
alone provide accurate and robust prediction. Furthermore, we
tested the preference of multi-kernel SVMs (MK-SVMs) for the
three inputs, which separately represent the three inputs by
weighted kernels. After the learning process, the MK-SVMs calcu-
lated kernel weights, which indicated that the MK-SVMs predomi-
nantly use gene networks in the predictions of NTs and GCs, and
qRT-PCRs in the predictions of NGCs. Molecular descriptors also
contributed to the predictions of GCs and NGCs, concordant with
the prediction performances generated by single-kernel SVMs that
use the three inputs separately.
2 Materials
In this protocol, we focus on only computational methods and do

not describe how to generate the gene expression data themselves.
For experimental materials and methods, please refer to [9, 10].
2.1 Computational 1. Cluster machines with minimum 3.0 GHz, 256 core, 4 GB
Environments memory/core, and 1 TB disk performance
2. Linux CentOS or other operating systems
3. RX-TAOgen (http://stemcellinformatics.org/toxicology/)
[9, 11, 12] or other Bayesian network inference programs
4. Kemba-svm (http://www.net-machine.net/~kato/kemba-
svm1-ts/) [13] or other kernel SVMs
5. SHOGUN package (http://www.shogun-toolbox.org) [14]
or other MK-SVMs
2.2 Microarray 1. Multiple microarray data of human ES cells exposed to various

and qRT-PCR Data toxic chemicals
2. qRT-PCR data of marker and internal control genes (GAPDH,
ACTB, or others) in human ES cells exposed to target toxic
chemicals. The number of points must be at least four times
the number of genes to produce stable Bayesian networks.
Stem Cell-Based Methods to Predict Developmental Chemical Toxicity 477
3 Methods
3.1 Selection We select genes that satisfy the following criteria:

of Marker Genes (a) signal detection p-value < 0.1 for three or more samples (see
from Microarray Data Note 1); (b) the encoded protein is a transcription factor; (c)
within the top 20 highest (or lowest for negative) weights in each
principal component by correspondence analysis; and (d) abun-
dant expression in human ES cells.
1. Prepare microarray data for toxic chemicals that characterize
target toxicities at 48 h after exposure to half the maximum
dose that does not impair cell viability (see Note 2). Normalize
the raw data and log-transform. Select probes for genes coding
for transcription factors by using BioMart (http://www.
biomart.org) or other sites.
2. Scale the log-transformed gene expression values across sam-
ples to obey standard normal distribution N(0, 1). Perform
correspondence analysis [15] by the corresp function of the
MASS library in the statistical language R [16]. Pick out the
top 20 highest (or lowest for negative) weights in each princi-
pal component.
3. Check the expression abundance of candidate genes in human
ES cells and remove genes that have undetectable abundance.
Select ten marker genes for the final list.
3.2 Preprocessing 1. Prepare qRT-PCR data of ten marker genes identified in

of qRT-PCR Data Subheading 3.1 for target chemicals that will be used in predic-
tion (see Note 3). The number of data points must be at least
four times the number of genes to reproduce stable Bayesian
networks, i.e., 40 points are needed to reconstruct a Bayesian
network of ten marker genes for each chemical from data that
include different doses, times, cells, and so on.
2. Adjust the gene expression values by dividing them by the val-
ues of the internal standard gene, such as GAPDH or ACTB,
at the corresponding time and dose points.
3. Log-transform the adjusted gene expression values to obey
normal distribution.
4. Use R package (sva [17] or others) to remove/reduce batch
effects if necessary.
3.3 Inference 1. Scale a series of gene expression data across samples for the
of Bayesian Networks same chemical to standard normal distribution, N(0,1).
by Replica-Exchange 2. (Suppose that intel MPI is installed in the machine.)
Method
% mpdboot –r ssh –n 8 (start MPI daemon with 8 processors)
% I_MPI_DEVICE=ssm mpiexec –configfile mpi.config
< mpi.input (see Note 4).
Example of mpi.config:
-s all
-n 1 -host h01 mpiTaogen
Example of mpi.input:
8
1,1,1,1,1,1,1,1
0.0001,0.0002,0.0004,0.0008,0.002,0.005,0.01,0.02
0.0001,0.0002,0.0004,0.0008,0.002,0.005,0.01,0.02
&input_data
sample=100,
gene=8,
sampling=10,
array_data=”/home/w.fujibuchi/Research/rxtao-
gen.20141211/data.txt”,
replica_exchange=1,
exchange=10000,
replica_exchange_order=1,
r=1,
measure=0,
burn_in=0,
count_in=5000,
verbose=1,
verbose_measure=200,
verbose_parameter=200
/
3. Obtain mpiTaogen.out from the output and select one gene

network with the highest maximum likelihood or posterior
probability. We use beta values, i.e., edge weights, for SVM
inputs.
4. Check the inferred gene networks by yED (http://www.
yworks.com/products/yed) [18] or other visualization pack-
ages (Fig. 1).
3.4 Generation 1. Obtain the Chemical Abstracts Service registry numbers

of QSAR Descriptors (CASRNs) for the chemicals to be studied.
2. Convert CASRNs into SMILES format by Chemical Identifier
Resolver (http://cactus.nci.nih.gov/). Remove ion bonds if
they exist in SMILES.
3. Go to E-Dragon (http://www.vcclab.org/lab/edragon/)
and upload the data by selecting “smiles.” Select ‘Use
CORINA’ from the pull-down menu and push “submit your
task.” Then select “Result as text” from the pull-down menu
Fig. 1 Network diagrams for ten chemicals. Upper: neurotoxins (NTs); Lower: genotoxic carcinogens (GCs)
which will become selectable after a successful calculation on

the top left of window. Finally, push “open ‘result.txt’ in a
browser” button and save the page into your local file.
3.5 Support Vector 1. Once you obtain one or a combination of qRT-PCRs, gene
Machine networks, and molecular descriptors for toxic chemicals, use
t-statistics to rank the inputs (features) in each input type in
learning data so that the order of the inputs reflects the dis-
criminating power of the positive and negative chemicals.
2. Select proper kernels such as linear, polynomial, radial basis
function (RBF), saigoone, saigotwo, kernemb, or others in
SVM analysis [9].
3. Execute grid search with various parameter values and calcu-
late SVM performances (see Note 5).
4. Perform the leave one chemical out prediction (LOCOP), i.e.,
a set of repeat data for each of the chemicals is eliminated dur-
ing the learning process to test the prediction (Table 1; see
Note 6).
5. Instead of combining and ranking the three inputs, i.e.,
qRT-PCRs, gene networks, and molecular descriptors, for
each toxic chemical, you may use distinct kernels for each
input. Use SHOGUN package to perform the LOCOP
analysis by MK-SVM. In Fig. 2, linear, RBF, and polyno-
mial kernels are tested (see Note 7). The highest accuracies
for NTs, GCs, and NGCs are 85.0%, 95.0%, and 100.0%,
respectively (see Note 8).
Table 1
Accuracies for combinations of three inputs: qRT-PCR (PCR), Bayesian network (BN), and quantitative
structure–activity relationships (QSAR)
Toxicity category PCR BN QSAR PCR + BN PCR + QSAR BN + QSAR PCR + BN + QSAR
Kernels: (upper) linear, polynomial, rbf/(lower) saigoone, saigotwo, kernemb

Non-adjusted
NT 70.0 80.0 55.0 70.0 57.5 65.0 65.0
77.5 95.0 75.0 82.5 82.5 75.0 82.5
GC 82.5 90.0 90.0 82.5 90.0 90.0 90.0
82.5 100.0 95.0 97.5 95.0 95.0 95.0
NGC 100.0 90.0 70.0 100.0 100.0 75.0 100.0
100.0 95.0 80.0 100.0 100.0 82.5 100.0
Batch-adjusted
NT 70.0 90.0 – 87.5 65.0 80.0 80.0
80.0 95.0 – 97.5 77.5 95.0 95.0
GC 85.0 85.0 – 82.5 90.0 90.0 90.0
85.0 100.0 – 92.5 95.0 95.0 95.0
NGC 100.0 90.0 – 97.5 97.5 75.0 97.5
100.0 95.0 – 97.5 97.5 80.0 97.5
The highest accuracies in each toxicity category are indicated in bold
4 Notes
1. This threshold is set for Illumina BeadArray chip. Users can set
proper thresholds for different microarray types such as
Affymetrix or Agilent.
2. It is important to select informative marker genes whose
expressions change significantly with exposure to toxic
chemical.
3. In this paper, we use the following ten genes: NANOG, SOX2,
DMTF1, ZNF208, ADRM1, TRIB1, CRY1, SMAD6, SMAD7,
and VHL1.
4. The lines of mpi.input are:
Fig. 2 Weights of kernels from qRT-PCR, BN, and QSAR data by multi-kernel SVM predictions. Vertical axis
indicates the ids of 20 chemicals. Weights that attained the highest accuracy in each toxic category are aver-
aged and shown by heat maps (see Note 8)
The number of replicas: 8

Degree1, or the prior for sampling beta when the edge exists:
1,1,1,1,1,1,1,1
Degree2, or the prior for sampling beta when the edge does not exist: 0.00
01,0.0002,0.0004,0.0008,0.002,0.005,0.01,0.02
Degree3, or dummy (not to be used): 0.0001,0.0002,0.0004,0.0008,0.002,0.00
5,0.01,0.02
The indicator of input data &input_data
- sample --- the number of samples in the gene expression profile
- gene --- the number of genes in the gene expression profile
- sampling --- the number of samplings
- array_data --- path to a file of gene expression profiles (relative paths
are allowed)
- replica_exchang --- switch of replica exchange (0:OFF, 1:ON)
- exchange --- the number of exchanges in replica exchange
- replica_exchange_order --- the order of replica exchange (0:random,
1:broader to narrower of search range)
- r --- upper limit of exchange probability (real values [0,1]) (see [9]
for more details.)
- measure --- method of network score measurement (0:log likelihood,
1:log posterior probability)
- burn_in --- the number of samplings to ignore in the calculation of re-
sults (specify the number to ignore from the first sampling)
- count_in --- the number of samplings to use in the calculation of re-

sults (specify the number to use from the last sampling. To be ignored
when burn_in is specified)
- verbose --- log output format (0:.out+.log, 1:.out+.log+.param+.meas-
ure, 2:.out+.log+.param+.measure+.hist)
- verbose_measure --- the interval (step number) of writing measurements
- verbose_parameter --- the interval (step number) of writing parameters
- verbose_histogram --- the interval (step number) of writing histograms
(beta)
5. For testing, typical settings of parameter ranges in the kemba-

svm package are: C = 10−3, 10−2, 10−1, 1, 10, 102, 103 (7 val-
ues). For kernel parameters, the polynomial kernel values are
D = 1, 2, 3, 4, 5, 6, 7, 8, 9, 10 (10 values); for the RBF, EKM,
and Saigo kernels, they are σ = 10−10, 10−9, 10−8, 10−7, 10−6,
10−5, 10−4, 10−3, 10−2, 10−1, 1 (11 values); and for the ME
kernel, they are G = 2−5, 2−4, 2−3, 2−2, 2−1, 1, 2, 22, 23, 24, 25
(11 values).
6. We test 20 chemicals. The NTs are valproic acid, cyclopamine,
phenytoin, methylmercury, acrylamide, 4-OH-2′,3,3′,4′,5′-
PCB, 2,5-hexanedione, warfarin, and thalidomide; the GCs
are benzo[a]anthracene, 3-methylcholanthrene, benzo[a]
pyrene, diethylnitrosamine, and diethylstilbestrol; and the
NGCs are 2,3,7,8-tetrachlorodibenzodioxin (TCDD), litho-
cholic acid, thioacetamide, butylated hydroxyanisole, metha-
pyrilene hydrochloride, and phenobarbital.
7. The tested parameter ranges are for SVM regularization:
C = 10−3, 10−2, 10−1, 1, 10, 102, 103; for polynomial kernel:
d = 1, 2, 3, 4, 5, 6, 7; and for RBF kernel: τ = 1, 5, 10, 50, 100,
500, 1000. The order of chemical ids is the same as in Note 6.
8. In this example, we show that qRT-PCR extensively contrib-
utes to NGCs and gives 100.0% accuracy in Table 1. QSAR
contributes more to GCs and NGCs than to NTs, which is also
concordant with the predictions in Table 1.
Acknowledgment
This work was partly supported by a Grant-in-Aid for Scientific

Research from the Ministry of Education, Culture, Sports, Science
and Technology of Japan [15H01749] and a grant from Core
Center for iPS Cell Research, Research Center Network for
Realization of Regenerative Medicine, Japan Agency for Medical
Research and Development. We also thank Dr. Peter Karagiannis
for useful comments on the manuscript.
References
1. Waters MD, Fostel JM (2004) Toxicogenomics lines and long-term maintenance with stable
and systems toxicology: aims and prospects. karyotype by enzymatic bulk passage. Biochem
Nat Rev Genet 5:936–948 Biophys Res Commun 345:926–932
2. Schwartz MP, Hou Z, Propson NE et al (2015) 11. Yamanaka T, Toyoshiba H et al (2004) The
Human pluripotent stem cell-derived neural TAO-Gen algorithm for identifying gene inter-
constructs for predicting neural toxicity. Proc action networks with application to SOS repair
Natl Acad Sci 112:12516–12521 in E. coli. Environ Health Perspect
3. Lee EK, Kurokawa YK et al (2015) Machine 112:1614–1621
learning plus optical flow: a simple and sensi- 12. Nagano R, Akanuma H et al (2011) Multi-
tive method to detect cardioactive drugs. Sci parametric profiling network based on gene
Rep 5:11817 expression and phenotype data: a novel
4. Kandasamy K, Chuah JK et al (2015) Prediction approach to developmental neurotoxicity test-
of drug-induced nephrotoxicity and injury ing. Int J Mol Sci 13:187–207
mechanisms with human induced pluripotent 13. Kato T, Fujibuchi W (2010) Kernel classifica-
stem cell-derived cells and machine learning tion methods for cancer microarray data. In:
methods. Sci Rep 5:12337 Emmert-Streib F, Dehmer M (eds) Medical
5. Huh D, Matthews BD et al (2010) biostatistics for complex diseases. John Wiley-
Reconstituting organ-level lung functions on a VCH, Weinheim
chip. Science 328:1662–1668 14. Sonnenburg SĆ, Henschel S et al (2010) The
6. Esch MB, King TL et al (2011) The role of SHOGUN machine learning toolbox. J Mach
body-on-a-chip devices in drug and toxicity Learn Re 11:1799–1802
studies. Annu Rev Biomed Eng 13:55–72 15. Benzécri JP (1973) L’Analyse des Données.
7. Takayama K, Inamura M et al (2012) Efficient Volume II. L’Analyse des Correspondances.
generation of functional hepatocytes from Dunod, Paris
human embryonic stem cells and induced plu- 16. Ripley B (2016). MASS: support functions and
ripotent stem cells by HNF4α transduction. datasets for venables and Ripley’s MASS. R
Mol Ther 20:127–137 package version 7.3-45. https://cran.
8. Seiler AE, Spielmann H (2011) The validated r-project. org/web/packages/MASS
embryonic stem cell test to predict embryotox- 17. Leek JT, Johnson WE et al (2012) The sva
icity in vitro. Nat Protoc 6:961 package for removing batch effects and other
9. Yamane J, Aburatani S et al (2016) Prediction unwanted variation in high-throughput experi-
of developmental chemical toxicity based on ments. Bioinformatics 28:882–883
gene networks of human embryonic stem cells. 18. Wiese R, Eiglsperger M et al (2004) yfiles—
Nucleic Acids Res 44:5515–5528 visualization and automatic layout of graphs.
10. Suemori H, Yasuchika K et al (2006) Efficient In: Jünger M, Mutzel P (eds) Graph drawing
establishment of human embryonic stem cell software. Springer, Berlin
Chapter 22
Predicting Chemically Induced Skin Sensitization by Using

In Chemico / In Vitro Methods
Laura H. Rossi and Janine Ezendam
Abstract
Over the recent years development toward assessing skin sensitization hazard has moved toward non-
animal testing methods. These methods are based on the key events as described in the OECD Adverse
Outcome Pathway (AOP) for skin sensitization initiated by covalent binding to proteins. As these indi-
vidual methods address mainly one mechanistic event (key event) in the initiation of skin sensitization,
combination of different methods are needed to conclude on the skin sensitization hazard. Validated and
regulatory adopted (EU and OECD) in chemico/in vitro methods are available for KEs 1–3 and are pre-
sented here. This chapter also illustrates how individual test methods can be combined by providing two
examples of defined approaches to testing and assessment for skin sensitization hazard identification and
assessment.
Key words In vitro, In chemico, Adverse outcome pathway (AOP), Key event (KE), Skin
sensitization
1 Introduction
Skin sensitization hazard assessment is traditionally based on data

from animal tests, e.g., the murine Local Lymph Node Assay
(OECD test guideline 429) or guinea pig tests, e.g., Guinea Pig
Maximisation Test and Buehler test (OECD test guideline 406).
During recent year’s development toward non-animal test meth-
ods assessing skin sensitization hazard has occurred and many
promising assays were developed, validated, and regulatory
adopted. The currently available non-animal test methods address
different steps of the Adverse Outcome Pathway (AOP) developed
by the OECD (Fig. 1). The AOP consists of four key events (KEs),
i.e., KE1 (molecular initiating event, i.e., MIE)—covalent protein
binding, KE2—keratinocyte activation, KE3—dendritic cell
Laura H. Rossi is a staff member of the European Chemicals Agency; the views and opinions expressed in this article
represent exclusively the personal ideas of the author and do not represent the official position of the Agency.
485
486 Laura H. Rossi and Janine Ezendam
Chemical
Structure/ MIE Cellular Level Organ Level
Properties
Covalent
Electrophilic Keratinocyte Dendritic Cell T-cell Activation Skin
Binding to
Chemicals Activation Activation and Proliferation Sensitization
Skin Proteins
Fig. 1 Representation of different key events from skin sensitization AOP (from AOPWiki)
a ctivation. and KE4—T-cell activation and proliferation [1]. Other

mechanisms leading to skin sensitization such as non-covalent
reactions with metals are not included in this AOP.
In this section, an update of currently available validated and/
or regulatory adopted in chemico/in vitro methods for skin sensi-
tization are described. Due to the applicability domain issues
related to a specific in chemico/in vitro method, not all chemicals
can be assessed by using these methods, e.g., substances that are
poorly soluble or metals.
This section aims to describe the currently available methods,
including their applicability in relation to specific chemicals.
Furthermore, two examples of defined approaches to testing and
assessment (DA) in which different individual test methods are
combined predict either skin sensitizing hazard and/or potency
are provided as well.
2 Methods
2.1 In Vitro/In Currently there is one in chemico assay that has been validated and
Chemico Assay(s) adopted under OECD test guidelines programme, i.e., In chemico
for the AOP Molecular skin sensitization: Direct Peptide Reactivity Assay (DPRA), OECD
Initiation Event/Key TG 442C.
Event 1: Peptide The method is based on the fact that majority of organic chem-
Binding/Protein icals inducing skin sensitization are either inherently electrophilic
Binding or will be converted into electrophilic chemicals via metabolic acti-
vation or autooxidation/abiotic degradation. Electrophilic skin
sensitizing chemicals will bind to nucleophilic amino acids such as
cysteine and/or lysine. This reaction can be measured in this assay.
The DPRA assay uses synthetic heptapeptides containing either
cysteine (-Ac-RFAACAA-COOH) or lysine (-Ac-RFAAKAA-
COOH) and binding to the test chemical is measured by using
high-pressure liquid chromatography (HPLC) using ultraviolet
(UV) detector by determining the concentration of free peptide
Predicting Chemically Induced Skin Sensitization by Using In Chemico/In Vitro Methods 487
Table 1
Reactivity classes of the DPRA
Cysteine 1:10/lysine 1:50 prediction model

Mean cysteine and lysine % Reactivity class DPRA
depletion prediction
0% ≤ mean % No or minimal reactivity Negative
depletion ≤ 6.38%
6.38% < mean % Low reactivity Positive
depletion ≤ 22.62%
22.62% < mean % Moderate reactivity
42.47% < mean % High reactivity
depletion ≤ 100%
Cysteine 1:10 prediction model

Cysteine % depletion Reactivity class DPRA
prediction
0% ≤ Cys % No or minimal reactivity Negative
13.89% < Cys % Low reactivity Positive
23.09% < Cys % Moderate reactivity
98.24% < Cys % High reactivity
depletion ≤ 100%
that is available after incubating with the test chemical for 24 h at
25 ± 2.5 °C. The synthetic peptides contain phenylalanine to aid in
the detection. The standard operating procedures are described in
more detail in the EURL ECVAM DB-ALM protocol no. 154
(https://ecvam-dbalm.jrc.ec.europa.eu/public/datasheet/?id=4
C1CE4B5521172A599E751542A0D8CED59DECC02B7D3
B4D6CB9CD2CDCA67E19FA3291B895581F634) which pro-
vides the step by step instructions how to perform the assay,
reagents needed and how to calculate the results using the predic-
tion model (cysteine 1:10/lysine 1:50, or cysteine 1:10).
Based on the mean peptide depletion values, calculated for test
chemical and positive control substance (Cinnamic aldehyde, CAS
104-55-2; ≥95% food grade purity), a reactivity class is calculated
ranging from “No or minimal reactivity” (negative prediction) to
“High reactivity” (positive prediction). As a cut-off value the mean
cysteine and lysine peptide percent depletion value of 6.38 is used
to discriminate between peptide nonreactive and peptide reactive
chemicals (Table 1).
Even though the method provides information from four dif-

ferent reaction classes, this information cannot be directly used to
predict skin sensitizing potency (UN GHS Cat 1A (extreme/
strong), or Cat 1B (moderate)). However, the information
obtained from DPRA method may provide useful information
when used together with other information such as in vitro data
obtained from other skin sensitization key events.
The assay is not applicable for metals, Chemical Substances of
Unknown or Variable Composition, Complex Reaction Products
and Biological Materials (UVCBs) and chemicals requiring meta-
bolic activation. Sensitizing metals cannot be detected by using
this assay, as they do not react with proteins via covalent binding
[2]. UVCB’s cannot be tested in the assay, as per nature of these
chemicals one cannot determine the exact ratio of the constituents
of those chemicals, which is the prerequisite for this particular
assay. Chemicals requiring metabolic/enzymatic activation (pro-
haptens) for being able to become sensitizing agent cannot be
detected in this assay, as this test system does not contain any meta-
bolic capacity.
The method has gone through validation coordinated by
European Union Reference Laboratory for alternatives to animal
testing (EURL ECVAM). Based on the validation outcome, an
EURL ECVAM recommendation on the test method was pub-
lished in 2013 [3]. Based on the validation study and published
studies [3, 4], the method is considered to be reproducible in pre-
dictions (85% within and 80% between laboratories). The accuracy
of the predictions to discriminate between skin sensitizers (UN
GHS Cat 1) vs. non-sensitizers is 80% with a sensitivity of 80%
(true positive) and specificity of 77% (true negative) when c ompared
to the in vivo Local Lymph Node Assay (OECD TG 429) [4].
2.2 In Vitro Assay(s) Under the OECD test guidelines programme, there is currently
for the AOP Key Event one test guideline that covers this key event (In Vitro Skin
2: Keratinocyte Sensitization—ARE-Nrf2 Luciferase Test Method; OECD
Activation TG442D). Currently, the only ARE-Nrf2 luciferase test method
covered by this test guideline is the Keratinosens™ test method.
Two other keratinocyte-based assays are included in the OECD
work program, the LuSens and SENS-IS.
2.3 KeratinoSens™ KeratinoSens™ identifies skin sensitizers by measuring the Nrf2-

Test Method mediated activation of antioxidant response element (ARE)-
dependent genes in human keratinocytes. Small electrophilic
substances such as skin sensitizers can act on the sensor protein
Keap1 (Kelch-like ECH-associated protein 1), by for example
covalent modification of its cysteine residue, resulting in its disso-
ciation from the transcription factor Nrf2 (nuclear factor-erythroid
2-related factor 2). The dissociated Nrf2 can then activate ARE-
dependent genes such as those coding for phase II detoxifying
enzymes [5]. KeratinoSens™ is an immortalized adherent cell line

derived from HaCaT human keratinocytes transfected with a plas-
mid that contains a luciferase gene under the transcriptional con-
trol of a constitutive promoter fused with an ARE element from
the human AKR1C2 gene that is upregulated by skin sensitizers,
resulting in activation of the luciferase gene. Subsequently, the
luciferase signal can be quantitatively measured using luminescence
detection. This signal is measured after a 48 h exposure of the cells
to different concentrations of the test chemical, vehicle control,
and positive control (Cinnamic Aldehyde, CAS 104-55-2; >95%
food grade purity). Test concentrations are based on preliminary
dose-range findings using cytotoxicity as a readout. After incuba-
tion, cells are harvested and luciferase gene induction is measured
in the cell lysates by luminescence detection. Detailed description
of the standard operating procedures of the KeratinoSensTM can be
found in the EURL ECVAM DB-ALM protocol no. 155 (https://
ecvam-dbalm.jrc.ec.europa.eu/methods-and-protocols/proto-
col/keratinosens-protocol-no.-155/key/p_1541).
The assay is performed at least twice for each test chemical.
Results are used as input to the prediction model which uses four
conditions to determine if a test chemical is positive or negative. If
all four conditions are met in the two repetitions, a test chemical is
considered positive, and if these are not met the test chemical is
considered negative. In case the two repetitions are not concor-
dant a third repetition needs to be performed. According to the
prediction model, the test chemical is considered positive if in two
out of three repetitions the four conditions are met, otherwise the
KeratinoSens™ prediction is considered negative.
The four conditions are as follows:
–– The maximal induction factor of luciferase activity compared
to the solvent (negative) control measured at any test chemical
concentration (Imax) is equal or higher than 1.5 fold and sta-
tistically significantly different as compared to the solvent/
vehicle control.
–– The cellular viability is higher than 70% at the lowest concen-
tration with induction of luciferase activity ≥1.5 fold.
–– The EC1.5 value is <1000 μM (or <200 μg/mL for test chem-
icals with no defined MW).
–– There is an apparent overall dose-response for luciferase
induction
Although the method provides quantitative dose-response
data, the information cannot be directly used to predict skin sensi-
tizing potency (UN GHS Cat 1A (extreme/strong), or Cat 1B
(moderate)). In combination with other test methods, the
KeratinoSens™ may provide an indication of potency [6].
The test method is applicable to test chemicals that are soluble

or form a stable dispersion either in water or DMSO. Test chemi-
cals with a cLogP up to 5 have been successfully tested. Very
hydrophobic molecules with a cLogP above 7 fall outside the
applicability domain of the test. Highly cytotoxic chemicals cannot
be reliable tested. Due to limited metabolic capacity of the cell line
used, pro-haptens (chemicals that require enzymatic activation)
and pre-haptens (chemicals requiring abiotic activation, e.g., via air
oxidation) may not be systematically detected. Negative results
may occur with chemicals with exclusive reactivity toward lysine.
Test chemicals that are known stressors but do not act as skin sen-
sitizers can activate the Nrf2-Keap1 pathway and may lead to false-
positive results. Finally, test chemicals that interfere with luciferase
activity enzyme can confound the activity of this luciferase-based
assay.
The industry-led validation study of the KeratinoSens™ was
evaluated by EURL ECVAM. The assay showed to be transferable
to other labs and had a good within and between laboratory repro-
ducibility of around 85% [7]. The accuracy of this assay for dis-
criminating skin sensitizers (i.e. UN GHS Cat. 1) from
non-sensitizers when compared to LLNA results were in the range
of 77–96% dependent on the dataset used. Overall, the accuracy of
the KeratinoSens™ is higher when compared to human data than
compared to data from the LLNA data [8, 9].
2.4 LuSens Method The LuSens ARE-Nrf2 luciferase test method is based on the same
principle as the KeratinoSens™ assay; it measures activation of the
Nrf2-KEAP1 pathway in a human keratinocyte reporter cell line.
In the LuSens test method the luciferase gene is under the control
of an ARE-element from the rat NADPH Quinone oxidoreductase
(NQO1), instead of the human AKR1C2 in the KeratinoSens™
assay [10].
The protocol and data interpretation procedure of the LuSens
are similar to the KeratinoSens™, with a few minor differences.
The LuSens uses ethylene glycol dimethacrylate (EGDMA, CAS
97-90-5) as a positive control. More details on the standard oper-
ating procedures of the LuSens can be found in the EURL ECVAM
DB-ALM protocol no. 184 (https://ecvam-dbalm.jrc.ec.europa.
eu/methods-and-protocols/protocol/lusens-assay-protocol-
no.-184/key/p_1392).
LuSens underwent a Performance Standards based validation
study based on the KeratinoSens™ as a validated reference method
[11]. An independent review was conducted by ESAC [12], who
concluded that the assay was easily transferred to other labs and
had a very good within- and between-laboratory reproducibility
(100% in both cases). The accuracy of the LuSens was comparable
to the KeratinoSens™ and ranges between 71% and 85% [9, 11].
The accuracy of the LuSens for discriminating skin sensitizers (i.e.,
UN GHS Cat. 1) from non-sensitizers when compared to LLNA

results was 74% [10]. For that reason, the LuSens is considered as
a me-too test method to the KeratinoSens™. OECD test guideline
442D is currently updated to include this me-too keratinocyte-
based assay. At the time of drafting this chapter, the revised test
guideline: “Key event based test guideline 442D: in vitro skin sen-
sitization assays addressing the AOP key event on keratinocyte
activation” is undergoing public consultation.
2.5 SENS-IS SENS-IS is an in vitro test based on a commercial human skin

model Episkin® RhE model in which gene expression is measured
of a panel of biological relevant genes that are known to be induced
by skin sensitizers [13, 14]. Regulation of these genes is measured
using quantitative reverse transcription-polymerase chain reaction
(RT-PCR). SENS-IS determines both the sensitizing potential and
potency of chemicals. The development of a Test Guideline for the
SENS-IS is included in the OECD work program.
The Episkin large model (Skinethic, Lyon, France) is a 3D skin
model that is differentiated from human keratinocytes that are iso-
lated from healthy donors. The epidermis is fully stratified and dif-
ferentiated and covered by a stratum corneum. The protocol of the
SENS-IS is described in detail in Cottrez et al. [14]. Standard
Operation Procedures are not available yet. In short, tissues are
treated by applying the test chemicals on the stratum corneum for
15 min. After washing, the tissues are further incubated for 6 h and
subsequently total RNA is extracted using Qiazol reagent and
RNeasy Mini Kit. Quantitative PCR is performed to measure the
predictive genes. Two panels of genes are measured: 17 genes that
contain Keap1-Nrf2- genes (REDOX panel) and 21 genes involved
in inflammation, danger signals and cell migration (SENS-IS panel)
[14]. These genes were selected from an initial set of 200 genes
involved in the antioxidant and inflammatory responses known to
be induced by skin sensitizers [13].
In the SENS-IS, tissue damage and irritation signals are used
to exclude chemicals that are tested at a too high concentration.
Tissue destruction is assessed by measuring the cycle threshold
(CT) of the HSPAA1 gene. Irritation signals are measured by the
expression of 23 genes. If 20 of these genes are overexpressed the
sample is rejected as well and the chemical is tested at a lower con-
centration. Test chemicals are considered as a sensitizers when at
least seven genes from both the REDOX and SENS-IS gene panels
are overexpressed 1.25 fold compared to the vehicle control. For
positive chemicals, the lowest concentration at which this 1.25 fold
overexpression is induced is used for potency classification. A sen-
sitizer is classified as an extreme, a strong, a moderate or a weak
sensitizer if found positive at 0.1%, 1%, or 10% or 50%, respectively.
If negative at all tested concentrations, a test chemical is consid-
ered a non-sensitizer.
The predictive capacity using a dataset of 150 chemicals was

96.6% compared to LLNA data, compared to human data, the
accuracy was 96% for a set of 130 chemicals [14]. The transferabil-
ity and reproducibility of the SENS-IS was determined in an
industry-led validation study in three laboratories using 19 sub-
stances. The SENS-IS had a within-laboratory reproducibility of
95% and a between-laboratory reproducibility of 100%. The con-
cordance of the potency predictions of the SENS-IS were com-
pared to the LLNA or the human classes. When compared to
LLNA data, the overall accuracy was 92.6% for 150 test chemicals
and compared to human data 90.6% for 64 substances [14].
2.6 In Vitro Assays Under the OECD test guidelines programme, there is currently
for the AOP Key Event one test guideline that covers this key event (In Vitro Skin
3: Activation Sensitization assays addressing AOP Key Event 3: activation of
of Dendritic Cells dendritic cells, OECD TG 442E). Currently, the following test
methods are included in this test guideline: Human Cell Line
Activation test (h-CLAT), U937 cell line activation Test
(U-SENS™), and Interleukin-8 Reporter Gene Assay (IL-8 Luc
assay).
The currently available test methods either quantify the change
in the expression of cell surface markers linked to the activation of
dendritic cells (DCs) when exposed to skin sensitizers (h-CLAT
and U-SENS™) or the changes in a cytokine linked to activation of
DCs, i.e., IL-8 (IL-8 Luc assay).
2.7 Assays The Human Cell Line Activation test (h-CLAT) uses monocyte-
Measuring Expression derived human monocytic leukemia cell line (THP-1) as surrogate
of Cell Surface to dendritic cells to investigate DC activation. In this test method
Markers the expression of two surface markers linked to DC maturation is
measured, i.e., CD86 (costimulatory molecule) and CD54 (adhe-
2.7.1 h-CLAT
sion molecule) by using flow cytometry [15]. Those two surface
markers are upregulated in activated DCs and are important when
DCs presents antigens to T-cells [16, 17]. As such, DC activation
is key in the process of skin sensitization.
In the test method THP-1 cells are exposed for 24 h to eight
different test chemical concentrations in a serial manner. The con-
centrations are selected based on cell viability of 75% (CV75) in
order to verify that the upregulation of the surface markers occurs
at sub-cytotoxic concentrations. Propidium iodide or other cyto-
toxicity markers are to be used to calculate the cell viability. The
test chemical is tested in at least in two independent runs in order
to obtain a single prediction (positive or negative). In each run
concurrent positive (2,4-dinitrochlorobenzene, CAS 97-00-7,
purity ≥99%) and a solvent/vehicle control is to be used.
The prediction is considered to be positive, i.e., indicating that

the test chemical is a skin sensitizer, if at least one of the conditions
are met in two out of two or two out of three independent runs:
1. The Relative Fluorescein Intensity (RFI) of CD86 is equal or

greater that 150% (EC150) at any tested concentration (with
cell viability of >50%);
2. The RFI of CD54 is ≥200% (EC200) at any tested concentra-
tion (with cell viability of >50%).
The standard operating procedures are described more in
detail in the EURL ECVAM DB-ALM protocol no. 158 (https://
ecvam-dbalm.jrc.ec.europa.eu/methods-and-protocols/proto-
col/1558/human-cell-line-activation-test-(h-clat)/datasheet)
which provides step by step instructions how to perform the assay,
reagents needed and how to calculate the results using the predic-
tion model. The results from this assay cannot be directly used to
discriminate between skin sensitization potency classes (UN GHS
Cat 1A (extreme/strong), or Cat 1B (moderate)). However, the
information obtained from h-CLAT assay may provide useful
information when used together with other information such as
in vitro data obtained from other skin sensitization key events.
The assay is not applicable to test chemicals that are either not
soluble or do not form a stable dispersion in an appropriate sol-
vent/vehicle, e.g., saline; medium; or dimethyl sulfoxide (DMSO).
Test chemicals with a Log Kow higher than 3.5 tend to produce
false negative predictions in the assay, hence negative results
obtained with such chemicals should not be considered. Due to
the cell line used and experimental conditions of the assay, pro-
haptens (substances requiring enzymatic activation) and pre-
haptens (substances activated by oxidation) may not be correctly
identified as skin sensitizers. Strong fluorescent chemicals emitting
the same wavelength as the fluorochrome used, may lead to false
negative predictions.
The method has gone through validation coordinated by
European Union Reference Laboratory for alternatives to animal
testing (EURL ECVAM). Based on the validation outcome, an
EURL ECVAM recommendation on the test method was pub-
lished in 2015 [18]. Based on the validation study [18–21] and
published studies [22–24], the method is considered to be repro-
ducible in predictions (on the order of 80% within and between
laboratories). The accuracy of the predictions to discriminate
between skin sensitizers (UN GHS Cat 1) vs. non-sensitizers is
85% with a sensitivity of 93% (true positive) and specificity of 66%
(true negative) when compared to the LLNA. When compared to
human data the accuracy of the test method was 83%, with sensitiv-
ity of 83% and specificity of 67% [23, 24].
2.7.2 U-SENSTM Method The U-SENS™ method uses the human myeloid U937 cell line to
quantify changes in the CD86 cell surface marker following
45 ± 3 h exposure to different concentrations of the test chemical
using flow cytometry. The CD86 surface marker has been selected,
as this is one typical marker of U937 activation and costimulatory
molecule needed for DC activation [15, 25].
In the assay, no dose range finding assay is performed, but
multiple concentrations used directly up to 200 μg/mL either in
complete medium or in 0.4% DMSO. There needs to be at least
four different concentrations and at least two independent runs
that are performed on different days to derive single prediction,
i.e., positive or negative. In the second run, the concentrations
used can be modified to the test chemical depending on the EC150
and cytotoxicity noted in the first run. As in the h-CLAT assay the
EC150 and CV70 is used to derive the prediction, i.e., for each
run if the CD86 induction has been increased by at least 150%
with <30% cytotoxicity, the run is considered positive (pre-
dicted skin sensitizer). In each run concurrent positive
(2,4-dinitrochlorobenzene, CAS 97-00-7, purity ≥99%) control
and a solvent/vehicle control are to be used.
The standard operating procedures are described more in
detail in the EURL ECVAM DB-ALM protocol no. 183 (https://
ecvam-dbalm.jrc.ec.europa.eu/methods-and-protocols/
protocol/1567/u937-cell-line-activation-test-for-skin-sensitiza-
tion-(u-sens-)/datasheet) which provides step by step instructions
how to perform the assay, reagents needed and how to calculate
the results using the prediction model. The results from this assay
cannot be directly used to discriminate between skin sensitization
potency classes (UN GHS Cat 1A (extreme/strong), or Cat 1B
(moderate)). However, the information obtained from the
U-SENS™ assay may provide useful information when used
together with other information such as in vitro data obtained
from other skin sensitization key events.
The assay is not applicable to chemicals that are either not
soluble or do not form a stable dispersion in an appropriate

solvent/vehicle (e.g., complete medium or 0.4% DMSO).

Membrane-
disrupting chemicals, e.g., surfactants, may lead to
false positive predictions due to a nonspecific increase in the induc-
tion CD86 [26]. Strong fluorescent chemicals emitting the same
wavelength as the fluorochrome used, may lead to false negative
predictions.
The test method has gone through industry-led validation.
The ESAC Opinion on the method is available and has been pub-
lished in June 2016 (EURL ECVAM Scientific Advisory
Committee., EURL ECVAM validation available at: https://
ec.eur opa.eu/jr c/en/publication/esac-opinion-lor-al-
coordinated-study-transferability-and-reliability-u-sens-test-
method-skin). The EURL ECVAM recommendation is currently

being prepared (as of January 2018).
The assay showed to be transferable to other laboratories and
had a good within and between laboratory reproducibility in
ranges between 84% and 90% [27], however during the v alidation
lower reproducibility was noted in one of the naïve laboratories,
indicating the need for proper training. The accuracy of this assay
based on the validation and other published data for discriminat-
ing skin sensitizers (i.e., UN GHS Cat. 1) from non-sensitizers
when compared to LLNA or human data is 86% or 77%, respec-
tively. The sensitivity of the assay when compared to LLNA and
human data is 91% and 100% and specificity of 65% and 47%,
respectively [26, 27].
2.8 IL-8 Luc Assay The IL-8 Luc assay addresses KE3 by measuring the induction of
IL-8 mRNA in the THP-1 cell line [28]. IL-8 is a cytokine associ-
ated with activation of dendritic cells [29].
IL-8 Luc is a based on a stable reporter cell line, THP-G8,
which was established by transfection of plasmid vectors into the
THP-1 cells. In these vectors, expression of stable luciferase orange
(SLO) and the stable luciferase red (SLR) genes is under the con-
trol of IL-8 and GAPDH promoters, respectively. In that way IL-8
Luc allows quantitative assessment of IL-8, by measurement of
luciferase expression [28, 30].
THP-G8 cells are treated for 16 h with the test chemical dis-
solved in X-VIVO™15, a commercially available serum-free
medium. Both test chemicals that are fully soluble and those that
are not or poor soluble in this medium are included in this assay.
The poor or insoluble test chemicals are shaken for 30 min and
after that supernatants are used in the assay. Test concentrations
are based on preliminary dose-range findings using GAPDH
expression as a read-out. In each experiment, the positive control
4-nitrobenzyl bromide (4-NBB) (CAS 100-11-8, ≥99% purity)
and the negative control, lactic acid (LA) (CAS 50-21-5, ≥85%
purity) need to be included. After incubation SLO luciferase
activity (IL8LA) reflecting IL-8 promoter activity and SLR lucifer-
ase activity (GAPLA) reflecting GAPDH promoter activity are
measured. The measured values or the quadriplate measurement
are used to calculate the following parameters:
(a) normalised IL8LA (nIL8LA), which is the ratio of IL8LA to
GAPLA;
(b) the induction of nIL8LA (Ind-IL8LA), which is the ratio of the
means of nIL8LA of THP-G8 cells treated with a test chemical
and the values of the nIL8LA of untreated THP-G8 cells;
(c) the inhibition of GAPLA (Inh-GAPLA), which is the ratio of
the values of the GAPLA of THP-G8 cells treated with a test
chemical and the values of the GAPLA of untreated THP-G8

cells, and used as an indicator for cytotoxicity.
The positive control, 4-NBB should produce a positive
response to induce a ≥1.4 fold induction of IL-8 mRNA, while the
negative control LA should produce a negative response (<1.4). A
detailed description of the procedure can be found on the website
of the Japanese Centre for Validation of Alternative Methods [31].
IL8 Luc uses the following criteria to decide if a run is positive
or negative:
(a) an IL-8 Luc assay prediction is judged positive if a test chemi-
cal has a Ind-IL8LA ≥1.4 and the lower limit of the 95% con-
fidence interval of Ind-IL8LA ≥1.0
(b) an IL-8 Luc assay prediction is judged negative if a test chemi-
cal has a Ind-IL8LA <1.4 and/or the lower limit of the 95%
confidence interval of Ind-IL8LA <1.0
According to the prediction model a test chemical is judged to
be positive or negative based on the number of positive or negative
runs, respectively. Test chemicals are considered positive if there
are two positive results among the four runs done, whereas test
chemicals are considered supposed negative, if there are three neg-
ative results among the four runs. For the negative results, an addi-
tional consideration is taken into account. Chemicals that are
soluble at 20 mg/mL in X-VIVO™15 are judged as negative,
whereas test chemicals that are not dissolved at this concentration,
should not be considered, because they might be false-negative
due to solubility issues.
Although the method provides quantitative dose-response
data, the information cannot be directly used to predict skin sensi-
tizing potency (UN GHS Cat 1A (extreme/strong), or Cat 1B
(moderate)). IL-8 Luc assay has not been used in combination
with other individual test methods for skin sensitization. So, it is
unknown whether this assay may contribute to potency assess-
ment, if it is used with other assays.
On the basis of the data currently available, the IL-8 Luc assay
was shown to be applicable to test chemicals covering a variety of
organic functional groups, reaction mechanisms, skin sensitization
potency (as determined in in vivo studies), and physicochemical
properties. IL-8 Luc assay correctly evaluated chemicals with a Log
Kow >3.5 and those with a water solubility of around 100 μg/mL
as calculated by EPI Suite™. Due to the limited metabolic capabil-
ity of the cell line and the experimental conditions, prohaptens
(substances requiring metabolic activation) and prehaptens (sub-
stances activated by air oxidation) might give false-negative results
in the assay. The majority of the prehaptens and prohaptens tested
so far were, however, were correctly identified by the IL-8 Luc
assay. Surfactants tested so far gave (false) positive results irrespec-
tive of their type (e.g., cationic, anionic, or nonionic). Anhydrides

assessed in the validation study showed a high rate of false negative
predictions. Finally, chemicals that interfere with luciferase can
confound its activity/measurement, causing apparent inhibition or
increased luminescence. So surfactants, anhydrides, and chemicals
that interfere with luciferase are therefore outside the applicability
domain of this assay.
The validation study of this assay was evaluated by JaCVAM
[31] and they concluded that IL-8 Luc was transferable to labora-
tories experienced in cell culturing and luciferase measurement.
The assay had a within laboratory reproducibility of 87% and a
between lab reproducibility of 87.5%. The accuracy of this assay to
identify skin sensitizers was 86% for 118 chemicals, compared to
LLNA data. Sensitivity and specificity were 96% and 41%, respec-
tively. Using human data (n = 76), the accuracy was 80%, sensitiv-
ity 93% and specificity 39%. Fourteen chemicals were inconclusive
[28, 31].
2.9 GARD Method The Genomic Allergen Rapid Detection test method (GARD) uses
gene expression profiling of transcripts and machine learning
approaches for predicting skin sensitization hazard and potency.
The human myeloid leukemia cell line (MUTZ-3) is used in this
assay as a surrogate to human dendritic cells.
In the assay, prior the chemical stimulation a qualitative phe-
notypic analysis of the MUTZ-3 cells used is performed to verify
that the proliferating cells are at an immature stage [32]. The test
chemicals are dissolved either in water or in DMSO. An end con-
centration of 500 μM is targeted for noncytotoxic and soluble
chemicals, and a highest possible concentration for chemicals hav-
ing limited solubility or cytotoxic chemicals resulting in relative
cell viability of 90%.
The cells are exposed to the test chemical for 24 h, following
which the cells are lysed with TRIzol reagent (Life Technologies)
for the RNA isolation performed according to the instructions by
the TRIzol supplier. From the RNA, cDNA is prepared for the
gene chip hybridization according to the Affymetrix GeneChip®
protocol using the recommended kits and controls (GeneChip®
Expression Analysis Technical Material. Available at:
https://www.affymetrix.com/support/downloads/manuals/
expression_analysis_plates_manual.pdf). More information about
the protocol can be obtained from the publications of Johansson
et al. and Zeller et al. [32, 33]. The standard operating procedures
are not yet available for this method (as of September, 2017).
In the GARD assay, the prediction signature consists of 200
transcripts (GARD prediction signature) that is proposed to be
refined to 52 transcripts [33]. The changes measured in the tran-
scripts are linked to the maturation and activation of dendritic
cells. The results obtained from the test chemical are compared to
a training set (trained and frozen) in order to obtain either a posi-

tive prediction (sensitizer) or a negative prediction (non-sensitizer).
In addition to the hazard identification (sensitizer vs. non-
sensitizer) a random forest model is used to predict UN GHS skin
sensitization categories, i.e., Cat 1A (extreme/strong sensitizer)
and Cat 1B (moderate sensitizer).
The GARD assay is currently undergoing validation by EURL
ECVAM and is expected to be completed in 2018. Based on the
recently published literature an accuracy of 86% has been reported
for the hazard identification was 86% [34]. Concerning hazard
characterization, i.e., skin sensitization potency an overall accuracy
of 78% was obtained (sensitivity 73%, specificity 80%, and accuracy
76%) [33]. Based on the validation exercise more information in
respect to the test method accuracy, limitations and regulatory rel-
evance will be obtained.
2.10 Defined The currently available and regulatory adopted in vitro/in c hemico
Approaches test methods cover different key events of the skin sensitization
AOP. To replace the LLNA, it is not sufficient to use only one of
such methods addressing a specific key event, but a combination of
test methods covering different steps of the AOP is needed. For
skin sensitization, several approaches combining different methods
including in silico, in chemico and in vitro methods to assess skin
sensitization hazard and/or potency have been proposed [35, 36].
However, there is currently no generally approved and/or vali-
dated way how to combine those methods in order to obtain reli-
able results that provide similar results to the currently available
in vivo methods (OECD TGs 406 and 429). Due to this reason,
an OECD project was approved in 2017 to develop assessment
criteria, how to combine in vitro and other data, to obtained
equivalent information in relation to hazard identification and
characterization when compared to the currently available in vivo
methods. The aim is to generate standardized data set and proce-
dures applied to obtain a prediction of a hazard (or the lack thereof)
without the need to apply expert judgment. For illustrative pur-
poses, two case studies are presented here; one for hazard identifi-
cation and one for potency assessment. More case studies are
included in the OECD GD 256, Annex 1 [37] and are described
in a recent review [35].
2.11 AOP-Based ‘2 The ‘2 out of 3’ ITS as proposed by BASF combines information

out of 3’ Integrated from assays that address the first three key events of the AOP: the
Test Strategy (ITS) DPRA (KE1), the LuSens assay or the KeratinoSens™ (KE2), and
for Skin Sensitization the U-SENS or the h-CLAT (KE3) [4, 9, 38].
Hazard Identification The data interpretation procedure (DIP) uses the predictions
Based on In Chemico of the individual test methods as inputs. If at least two out of three
and In Vitro Data tests give a positive prediction, a substance is considered a skin
sensitizer. In case at least two out of three tests give a negative
prediction, a substance is considered a non-sensitizer. The “2 out

of 3” ITS can only be used for hazard identification, it does not
provide a prediction of the skin sensitizing potency.
The accuracy of this defined approach compared to the LLNA
is 81%, while sensitivity and specificity are 82% and 77%, respec-
tively [4]. More recently, this “2 out of 3” ITS was further evalu-
ated by testing its performance using a dataset of 213 substances
with LLNA data; human data were available for 114 substances
[9]. The predictive capacity of the ITS was evaluated using data
form different test methods, including the DPRA (KE1),
KeratinoSens™ and the LuSens assay (KE2), the h-CLAT, modi-
fied MUSST and U-SENS assays (KE3). It is important to note
that not all assays tested the complete dataset of 213 substances.
Nevertheless, the available information allowed a thorough perfor-
mance analysis and comparison of the individual test methods. In
addition, it was possible to analyze the predictive capacities of ITS
approaches consisting of different assays. For example, an ITS
combining data from the DPRA, KeratinoSens™ and h-CLAT
resulted in the highest accuracy of 90%; sensitivity and specificity
both were 90% as well, compared to human data. This predictivity
was more accurate compared to predictions based on the LLNA
for the same dataset, for which the accuracy, sensitivity and speci-
ficity for the same dataset were 82%, 91% and 64%, respectively [9].
2.12 Integrated The ITS approach published by Kao Corporation combines infor-
Testing Strategy (ITS) mation obtained from DEREK Nexus (version 2.0 from Lhasa
for Skin Sensitization Limited), DPRA (Key event 1, OECD TG 442C) and h-CLAT
Potency Classification (key event 3, OECD TG 442E).
Based on In Silico, The DIP uses the quantitative parameters obtained from
In Chemico DPRA and h-CLAT assay and the outcomes obtained from
and In Vitro Data
Table 2
Conversion of the outcome of the individual assays of the ITS
DPRA depletion (%) DEREK

Score h-CLAT MIT (μg/mL) (Cysteine-only) alert
0 Not calculated <6.376 No
(negative) (<13.89) alert
1 >150 ≤5000 ≥6.367 <22.62 Alert
(≥13.89 <23.09)
2 >10 ≤150 ≥22.62 <42.47 –
(≥23.09 <98.24)
3 ≤10 ≥42.47 –
(≥98.24)
MIT minimum induction threshold

Table 3
Classification of the ITS based on total battery scores
Total battery score Outcome

0–1 Non-sensitizer
2–6 Weak sensitizer
7 Strong sensitizer
Table 4
Predictive performance of the defined approach for hazard identification
ITS
Hazard identification Positive Negative

102 sensitizers 91 11
LLNA 37 non-sensitizers 11 26
Sensitivity (%) 89 (91/102)
Specificity (%) 70 (26/37)
Accuracy (%) 84 (117/139)
DEREK to assign scores that are used to predict the skin sensitiza-
tion potential and potency (Table 2) [39, 40].
For DEREK absence of an alert is assigned a score of 0 and
presence of an alert a score of 1. For DPRA scores are assigned
based on the mean peptide depletion values obtained and for
h-CLAT based on the concentrations for minimum inductions
thresholds.
The individual scores obtained are then summed together into
total battery scores from 0 to 7 to predict skin sensitization hazard
(skin sensitizer vs. non-sensitizer) and the skin sensitization potency
(two rank classes: EC3 <1% in Local Lymph Node Assay (LLNA)
(strong sensitizer), EC3 >1% in LLNA (weak sensitizer)) Table 3.
It is good to note, that the rank classes used here, i.e., EC3
thresholds, are not the same thresholds used to classify skin sensi-
tizers in the UN GHS potency classes, where EC3 ≤2% is for Cat
1A (strong/extreme) and is EC3 >2% is for moderate skin sensitiz-
ers [41], therefore the results obtained cannot be used directly to
classify into UN GHS potency categories.
The predictive capacity for skin sensitization hazard identifica-
tion (sensitizer vs. non-sensitizer) of this defined approach has
been calculated to be 89% for sensitivity, 70% for specificity and
84% for accuracy based on data generated from 139 chemicals
[40], when compared to LLNA (Table 4).
Table 5
Predictive performance of the defined approach for potency estimation
ITS
Potency identification Strong Weak Non-sensitizer

LLNA Extreme/strong 15 14 0
Moderate/weak 5 57 11
Non-sensitizer 0 11 26
Overprediction rate (%) 12 (16/139)
Underprediction rate (%) 18 (25/139)
Accuracy (%) 71 (98/139)
The predictive capacity of this defined approach for predicting

skin sensitization potency has been established to be 12% for over
prediction rate, 18% for under-prediction rate and 71% for accu-
racy based on the data available for 139 chemicals as shown in
Table 5 [40].
3 Conclusion
This chapter provides a short overview of the available nonanimal

test methods for skin sensitization hazard identification and
potency assessment that haven been validated and/or are adopted
by the OECD. Two examples are presented on how these indi-
vidual test methods can be combined in a defined approach to
testing and assessment. It is evident that this field has evolved
greatly, but that more work is needed on independent review of
the defined approaches to better understand their applicability
domain, technical limitations, and uncertainties. Current activities
employed by EURL ECVAM, US EPA, Health Canada, and also
Cosmetics Europe to validate these approaches will increase our
understanding of these factors and facilitate regulatory acceptance
in the future.
References
1. OECD (2012) The Adverse Outcome Pathway 2. Schmidt M, Goebeler M (2015) Immunology
for Skin Sensitisation initiated by Covalent of metal allergies. J Dtsch Dermatol Ges
Binding for Proteins. Part 1. Scientific 13:653–660
Evidence. OECD Environment, Health and 3. EURL-ECVAM (2013) EURL ECVAM rec-
Safety Publications Series on Testing and ommendation on the Direct Peptide
Assessment. No. 168. Available at: http:// Reactivity Assay (DRPA) for skin sensitisa-
www.oecd.org/officialdocuments/publicdispla tion testing. European Commission Joint
ydocumentpdf/?cote=env/jm/ Research Centre. Institute for Health and
mono(2012)10/part1&doclanguage=en Consumer Protection. European Union
Reference Laboratory for Alternatives to ibility and accuracy of the LuSens assay: a
Animal Testing (EURL ECVAM). Available reporter gene-cell line to detect keratinocyte
at:http://ihcp.jrc.ec.europa.eu/our_labs/ activation by skin sensitizers. Toxicol In
eurl-ecvam/eurl-e cvam-r ecommendations/ Vitro 32:278–286
E U R L - E C VA M -R E C - D P R A - D R A F T- 12. ESAC (2016) EURL ECVAM Scientific
ver02August2013.pdf Advisory Commitee (ESAC) opinion on the
4. Natsch A, Ryan CA, Foertsch L, Emter R, BASF-coordinated Performance Standards-
Jaworska J, Gerberick F, Kern P (2013) A data- based validation of the LuSens test method for
set on 145 chemicals tested in alternative assays skin sensitisation testing. Available at: http://
for skin sensitization undergoing prevalidation. publications.jrc.ec.europa.eu/repository/bit-
J Appl Toxicol 33:1337–1352 stream/JRC103706/esac_opinion_2016-04_
5. Natsch A (2010) The Nrf2-Keap1-ARE toxic- lusens_final.pdf
ity pathway as a cellular sensor for skin sensitiz- 13. Cottrez F, Boitel E, Auriault C, Aeby P, Groux
ers--functional relevance and a hypothesis on H (2015) Genes specifically modulated in sen-
innate reactions to skin sensitizers. Toxicol Sci sitized skins allow the detection of sensitizers in
113:284–292 a reconstructed human skin model.
6. Natsch A, Emter R, Gfeller H, Haupt T, Ellis G Development of the SENS-IS assay. Toxicol In
(2015) Predicting skin sensitizer potency based Vitro 29:787–802
on in vitro data from keratinosens and kinetic 14. Cottrez F, Boitel E, Ourlin J-C, Peiffer J-L,
peptide binding: global versus domain- based Fabre I, Henaoui I-S, Mari B, Vallauri A,
assessment. Toxicol Sci 143:319–332 Paquet A, Barbry P, Auriault C, Aeby P, Groux
7. EURL-ECVAM (2013) EURL ECVAM rec- H (2016) SENS-IS, a 3D reconstituted epider-
ommendation on the Keratinosens™ assay for mis based model for quantifying chemical sen-
skin sensitisation testing European Commission sitization potency: reproducibility and
Joint Research Centre. Institute for Health and predictivity results from an inter-laboratory
Consumer Protection. European Union study. Toxicol In Vitro 32:248–260
Reference Laboratory for Alternatives to 15. Ashikaga T, Yoshida Y, Hirota M, Yoneyama K,
Animal Testing (EURL ECVAM). Available Itagaki H, Sakaguchi H, Miyazawa M, Ito Y,
at:https://eurl-ecvam.jrc.ec.europa.eu/eurl- Suzuki H, Toyoda H (2006) Development of
e c v a m -r e c o m m e n d a t i o n s / an in vitro skin sensitization test using human
r e c o m m e n d a t i o n -k e r a t i n o s e n s -s k i n - cell lines: the human Cell Line Activation Test
sensitisation (h-CLAT). I. Optimization of the h-CLAT
8. Bauch C, Kolle SN, Fabian E, Pachel C, protocol. Toxicol In Vitro 20:767–773
Ramirez T, Wiench B, Wruck CJ, Ravenzwaay 16. Harris NL, Ronchese F (1999) The role of B7
BV, Landsiedel R (2011) Intralaboratory vali- costimulation in T-cell immunity. Immunol
dation of four in vitro assays for the prediction Cell Biol 77:304–311
of the skin sensitizing potential of chemicals. 17. Hwang I, Shen X, Sprent J (2003) Direct stim-
Toxicol In Vitro 63:489–504 ulation of naive T cells by membrane vesicles
9. Urbisch D, Mehling A, Guth K, Ramirez T, from antigen-presenting cells: distinct roles for
Honarvar N, Kolle S, Landsiedel R, Jaworska J, CD54 and B7 molecules. Proc Natl Acad Sci U
Kern PS, Gerberick F, Natsch A, Emter R, S A 100:6670–6675
Ashikaga T, Miyazawa M, Sakaguchi H (2015) 18. EURL-ECVAM (2015) EURL ECVAM
Assessing skin sensitization hazard in mice and Recommendation on the human Cell Line
men using non-animal test methods. Regul Activation Test (h-CLAT) for skin sensitisation
Toxicol Pharmacol 71:337–351 testing. European Commission, Joint Research
10. Ramirez T, Mehling A, Kolle SN, Wruck CJ, Centre, Institute for Health and Consumer
Teubner W, Eltze T, Aumann A, Urbisch D, Protection, accesible at https://eurl-ecvam.jrc.
van Ravenzwaay B, Landsiedel R (2014) ec.europa.eu/news/news_docs/
LuSens: a keratinocyte based ARE reporter eurl-ecvam-recommendation-on-the-human-
gene assay for use in integrated testing strate- cell-line-activation-test-h-clat-for-skin-sensiti-
gies for skin sensitization hazard identification. sation-testing
Toxicol In Vitro 28:1482–1497 19. EURL-ECVAM (2015) Re-analysis of the within
11. Ramirez T, Stein N, Aumann A, Remus T, and between laboratory reproducibility of the
Edwards A, Norman KG, Ryan C, Bader JE, human Cell Line Activation Test (h-CLAT).
Fehr M, Burleson F, Foertsch L, Wang X, European Commission, Joint Research Centre,
Gerberick F, Beilstein P, Hoffmann S, Institute for Health and Consumer Protection,
Mehling A, van Ravenzwaay B, Landsiedel R accessible at: https://eurl-ecvam.jrc.ec.europa.
(2016) Intra- and inter-laboratory reproduc- eu/eurl-ecvam-recommendations/
eurl-ecvam-recommendation-on-the-human- stable THP-1-derived IL-8 reporter cell line,

cell-line-activation-test-h-clat-for-skin-sensitisa- THP-G8. Toxicol Sci 124:359–369
tion-testing 29. Toebak MJ, Pohlmann PR, Sampat-
20. Sakaguchi H, Ashikaga T, Miyazawa M, Sardjoepersad SC, von Blomberg BM,
Yoshida Y, Ito Y, Yoneyama K, Hirota M, Bruynzeel DP, Scheper RJ, Rustemeyer T,
Itagaki H, Toyoda H, Suzuki H (2006) Gibbs S (2006) CXCL8 secretion by dendritic
Development of an in vitro skin sensitization cells predicts contact allergens from irritants.
test using human cell lines; human Cell Line Toxicol In Vitro 20:117–124
Activation Test (h-CLAT). II. An inter-labora- 30. Kimura Y, Fujimura C, Ito Y, Takahashi T,
tory study of the h-CLAT. Toxicol In Vitro Nakajima Y, Ohmiya Y, Aiba S (2015)
20:774–784 Optimization of the IL-8 Luc assay as an
21. Sakaguchi H, Ryan C, Ovigne JM, Schroeder in vitro test for skin sensitization. Toxicol In
KR, Ashikaga T (2010) Predicting skin sensiti- Vitro 29:1816–1830
zation potential and inter-laboratory reproduc- 31. JaCVAM (2016) JaCVAM Validation report
ibility of a human Cell Line Activation Test for the international validation study on the
(h-CLAT) in the European Cosmetics IL-8 Luc assay as a test evaluating the skin sen-
Association (COLIPA) ring trials. Toxicol In sitizing potential of chemicals conducted by
Vitro 24:1810–1820 the IL-8 Luc Assay. December 2016. Available
22. Ashikaga T, Sakaguchi H, Sono S, Kosaka N, at: http://www.jacvam.jp/files/
Ishikawa M, Nukada Y, Miyazawa M, Ito Y, doc/03_08/03_08_D1.pdf
Nishiyama N, Itagaki H (2010) A comparative 32. Johansson H, Albrekt AS, Borrebaeck CA,
evaluation of in vitro skin sensitisation tests: the Lindstedt M (2013) The GARD assay for
human cell-line activation test (h-CLAT) ver- assessment of chemical skin sensitizers. Toxicol
sus the local lymph node assay (LLNA). Altern In Vitro 27:1163–1169
Lab Anim 38:275–284 33. Zeller KS, Forreryd A, Lindberg T, Gradin R,
23. Nukada Y, Ashikaga T, Sakaguchi H, Sono S, Chawade A, Lindstedt M (2017) The GARD
Mugita N, Hirota M, Miyazawa M, Ito Y, Sasa platform for potency assessment of skin sensi-
H, Nishiyama N (2011) Predictive perfor- tizing chemicals. ALTEX 34:539. https://doi.
mance for human skin sensitizing potential of org/10.14573/altex.1701121
the human cell line activation test (h-CLAT). 34.
Johansson H, Gradin R, Forreryd A,
Contact Dermatitis 65:343–353 Agemark M, Zeller K, Johansson A, Larne
24. Takenouchi O, Miyazawa M, Saito K, Ashikaga O, van Vliet E, Borrebaeck C, Lindstedt M
T, Sakaguchi H (2013) Predictive performance (2017) Evaluation of the GARD assay in a
of the human Cell Line Activation Test blind Cosmetics Europe study. ALTEX
(h-CLAT) for lipophilic chemicals with high 34:515. https://doi.org/10.14573/
octanol-water partition coefficients. J Toxicol altex.1701121
Sci 38:599–609 35. Ezendam J, Braakhuis HM, Vandebriel RJ
25. Aiba S, Terunuma A, Manome H, Tagami H (2016) State of the art in non-animal
(1997) Dendritic cells differently respond to approaches for skin sensitization testing: from
haptens and irritants by their production of individual test methods towards testing strate-
cytokines and expression of co-stimulatory gies. Arch Toxicol 90:2861–2883
molecules. Eur J Immunol 27:3031–3038 36. OECD (2016) Guidance document on the
26. Piroird C, Ovigne J-M, Rousset F, Martinozzi- reporting of defined approaches and individual
Teissier S, Gomes C, Cotovio J, Alépée N information sources to be used within
(2015) The myeloid U937 skin sensitization Integrated Approaches to Testing and
test (U-SENS) addresses the activation of den- Assessment (IATA) for skin sensitisation Series
dritic cell event in the adverse outcome path- on Testing and Assessment vol ENV/JM/
way for skin sensitization. Toxicol In Vitro HA(2016)11. OECD, Paris
29:901–916 37. OECD (2016) ANNEX I: case studies to the
27. Alépée N, Piroird C, Aujoulat M, Dreyfuss S, guidance document on the reporting of defined
Hoffmann S, Hohenstein A, Meloni M, approaches and individual information sources
Nardelli L, Gerbeix C, Cotovio J (2015) to be used within Integrated Approaches to
Prospective multicentre study of the U-SENS Testing and Assessment (IATA) for skin sensiti-
test method for skin sensitization testing. sation. Series on Testing & Assessment No.
Toxicol In Vitro 30(1, Part B):373–382 256. Available at: http://www.oecd.org/offi-
28. Takahashi T, Kimura Y, Saito R, Nakajima Y, cialdocuments/publicdisplaydocumentpdf/?c
Ohmiya Y, Yamasaki K, Aiba S (2011) An ote=env/jm/mono(2016)29/
in vitro test to screen skin sensitizers using a ann1&doclanguage=en
38. Bauch C, Kolle SN, Ramirez T, Eltze T, Fabian E, Kasahara T, Fujita M, Toyoda A, Sekiya D,
Mehling A, Teubner W, Ravenzwaay BV, Landsiedel Watanabe S, Seto H, Hirota M, Ashikaga T,
R (2012) Putting the parts together: combining Miyazawa M (2015) Test battery with the
in vitro methods to test for skin sensitizing poten- human cell line activation test, direct peptide
tials. Regul Toxicol Pharmacol 63:489–504 reactivity assay and DEREK based on a 139
39. Nukada Y, Miyazawa M, Kazutoshi S, chemical data set for predicting skin sensitizing
Sakaguchi H, Nishiyama N (2013) Data inte- potential and potency of chemicals. J Appl
gration of non-animal tests for the develop- Toxicol 3:1318–1332
ment of a test battery to predict the skin
41. UN (2017) United Nations UN) Globally
sensitizing potential and potency of chemicals. Harmonised System (GHS) for Classification
Toxicol In Vitro 27:609–618 and Labeling. Revison 7, available at https://
40. Takenouchi O, Fukui S, Okamoto K, Kurotani www.unece.org/fileadmin/DAM/trans/dan-
S, Imai N, Fujishiro M, Kyotani D, Kato Y, ger/publi/ghs/ghs_rev07/English/03e_
part3.pdf
Chapter 23
Hepatotoxicity Prediction by Systems Biology

Modeling of Disturbed Metabolic Pathways
Using Gene Expression Data
Oriol López-Massaguer, Manuel Pastor, Ferran Sanz, and Pablo Carbonell
Abstract
The present method describes a systems biology approach for the in silico predictive modeling of drug
toxicity. The data from LINCS were used to determine the type and number of pathways disturbed by each
compound and to estimate the extent of disturbance (network perturbation elasticity). Moreover, the most
frequently disturbed metabolic pathways and reactions were determined across the studied toxicants. The
process was exemplified by successful predictions on various statins. In conclusion, an entirely new approach
linking gene expression alterations to the prediction of complex organ toxicity was developed.
Key words Systems biology, Predictive modeling, Drug toxicity, Hepatotoxicity, Gene regulation
1 Introduction
We develop a method to predict by systems biology approaches the

potential hepatoxicity of a chemical perturbation. The method
consists of the following main steps (Fig. 1):
1. We obtain gene overexpression/underexpression in human

hepatocytes under certain chemical perturbation by querying
the LINCS1000 database [1].
2. We perform a flux variability analysis (FVA) of the Recon 2
model [2] of liver to estimate the upper and lower bounds of
the metabolic reactions.
3. We map the gene expression changes under chemical perturba-
tion to compute the network perturbation according to reac-
tion bounds (perturbation elasticity).
4. We also obtain the pathways altered by using Recon 2 model
and the gene expression alteration by the chemical
perturbation.
505
506 Oriol López-Massaguer et al.
Fig. 1 Global diagram of the method. Panel (1) consists of extracting the chemical-induced gene upregulation/
downregulation data for the substances of interest from the LINCS database. Panel (2) shows how gene–
protein–reactions (GPR) associations of the hepatocyte model are used in order to map the regulated genes
into the set of altered reactions. Panel (3) illustrates how flux variability analysis is applied to the hepatocyte
model in order to determine lower and upper bound limits. These limits are used in panel (4) to perform an
elasticity analysis in order to determine the perturbation in the network. Finally, panel (5) shows how altered
reactions can be mapped to identify perturbed pathways
The described method is based on a previous publication [3],

and each step will be detailed.
We will apply our method to three statins as a running example:
pravastatin, simvastatin, and rosuvastatin, which are cholesterol-
lowering drugs.
2 Materials
The protocol described is computational, so we will describe the

software tools and data sources used as materials:
1. LINCS database to obtain gene expression data.
2. Recon 2 metabolic model for the liver.
3. FVA package COBRApy [4].
Hepatotoxicity Prediction by Systems Biology Modeling of Disturbed Metabolic… 507
Table 1
Supplementary material file listing
File name Description

File S1 GPR rules for liver hepatocytes from
Recon 2
File S2 Flux elasticities for human hepatocytes
File S3 Altered gene expression for statins
from LINCS database
File S4 Perturbed reactions for statins example
File S5 Altered pathways for statins example
4. Custom-made python scripts

(a) We provide a python script that computes FVA analysis for
human hepatocytes. Available in: https://github.com/
phi-grib/Hepatoxicity-SysBio/tree/master/recon2_fva
(b) We provide a jupyter notebook that performs the compu-
tation of the hepatotoxicity by the method described for
the case of statins.
Available in: https://github.com/phi-grib/Hepatoxicity-SysBio/
tree/master/notebook
The whole source code is distributed as an open source under
GPLv3 license [5]:
https://github.com/phi-grib/Hepatoxicity-SysBio
The supplementary material used is listed in Table 1 and available
in: https://github.com/phi-grib/Hepatoxicity-SysBio/tree/
master/material
3 Methods
3.1 Extraction The first step is downloading the data of differential gene expression
of Gene Expression for the chemical perturbation (compound of interest) from the
Data from the LINCS LINCS database, the so-called gene expression signature.
Database In the LINCS database, 22,119 perturbations were available
for compounds with gene expression information based on the
L1000 technology (see Note 1). They focused on the 50 most
highly upregulated/downregulated genes for the chemical pertur-
bation (see Note 2).
Each differential response in the LINCS database is identified

as a gene expression signature. For instance, signature “CPC020_
PC3_24H:BRD-K94173926-001-01-6:10” contains a first part
that identifies the experiment (CPC020), the cell line (PC3), and
the perturbation time (24 h); the second part identifies the pertur-
bation (BRD-K94173926), which in this case corresponds to the
identifier of a small compound, followed by experimental informa-
tion of location on the plate. The final part of the identifier is the
perturbation dose (10 μM).
Refer to the LINCS website for detailed instructions http://
www.lincsproject.org
Output: 50 most highly upregulated/downregulated genes
for the chemical perturbation for the compound of interest (see
Note 3).
3.2 Cell Type- The next step is to map the gene expression data obtained in order
Specific Metabolic to identify the perturbed reactions in the metabolic network of the
Model liver.
1. For this purpose we use Recon 2, a consensus metabolic model
for human cell metabolism [2]. The model can be downloaded
from the Biomodels database [6], MODEL1109130000.1 The
model is represented using SBML and contains 65 cell-type
specific models, including liver hepatocytes. In the models,
each of the 7440 metabolic reactions is annotated with their
associated reactants and corresponding stoichiometric coeffi-
cients (see Note 4).
2. Reactions are annotated with pathway information, enzyme
class, and gene–protein–reaction (GPR) associations. GPRs are
Boolean functions that reflect the requirements in terms of
genes for their associated enzymes and reactions fi(gj), includ-
ing those requirements for enzyme complexes that are formed
by different subunits, the presence of isozymes, enzyme pro-
miscuity, etc. [7]. This step is independent of the compounds
we are interested in because it maps genes to reactions using
GPR rules (see Note 5). The data is available in File S1.
3. We obtained in the previous step from LINCS the upregu-
lated/downregulated gene sets g+(si) and g−(si) for several
chemical perturbations si. In our case the cell type is human
hepatocyte. To compute the perturbation on each reaction we
map the upregulated/downregulated genes into the GPR
Boolean rules defined in the model between genes and reac-
tions (File S2). The way we mapped this information is by
translating AND/OR into min/max relationships as follows,
for a pair of genes gi and gj:
1
https://www.ebi.ac.uk/biomodels-main/MODEL1109130000
(
g i AND g j ® min g i , g j )
g i OR g j ® max ( g , g )
i j (1)
ì -1 if g i Î g -
ï
g i = í 0 otherwise
ï+1 if g Î g +
î i

These previous relationships can be interpreted as follows:
(a)
When two genes are associated with an AND operation
according to the GPR rule, then downregulation can be
caused by any of the two genes (min function). In the
AND case, it suffices for one of the genes to be downregu-
lated to potentially alter the reaction; if one of the genes is
overexpressed, we cannot assume overexpression as the
reaction depends on both genes.
(b) when the genes are associated with an OR rule, then only
one of the genes is needed in order to have upregulation
(max function). In this case, the downregulation of one of
the genes would not be a sufficient condition for expect-
ing negative regulation of the reaction (see Note 6).
4. By applying these criteria systematically to each gene involved
in a reaction flux, as described by the previous equation, we
obtained the overall effect of the gene regulation δijk on that
reaction under the given chemical perturbation si.
For example, the following GPR rule:
(6520.1) and (23,428.1) or (6520.1) and (23,428.2) or

(56,301.1) and (6520.1) → R_CYSGLYexR
is transformed into:
min(max(6520.1, 23,428.1), max(6520.1, 23,428.2), max

(56,301.1, 6520.1)) → R_CYSGLYexR
The pseudo code of the steps:
1. Load liver hepatocytes metabolic model.
2. Obtain from the metabolic model the GPR rules that map
genes to reactions.
3. Map chemical-induced gene perturbation (altered expres-
sion) to reactions using GPR rules (see https://github.
com/phi-grib/Hepatoxicity-SysBio/blob/master/note-
book/Statins.ipynb).
(a)
For each chemical perturbation that produce gene
overexpression and underexpression (obtained from
LINCS database).
(b) Gene expression alteration is mapped using GPR rules
to the associated altered reactions.
4. Apply the mapping to each gene involved in a reaction flux

as described in the previous equation we obtain the overall
effect on the reactions.
We provide a jupyter notebook that performs this mapping
for the example of statins but it can be easily adapted to other
compounds.
Output: Perturbation of the reactions for the chemical
perturbations.
3.3 Flux Variability The procedure described in the previous section provides an estimate
Analysis of the main perturbed reactions in the cell. In the next step, we
perform a simulation of the steady state of the cell model in order to
determine the distribution of fluxes for each reaction. A systems
analysis of the model can determine how critical such perturbations
may be in terms of disrupting the steady state equilibrium.
1. We use a sensitivity analysis method known as flux variability
analysis (FVA), which provides an estimate of the upper and
lower bounds of each steady state reaction flux vi in the cell [8]
(see Note 7). We used the COBRApy package [4].The method
is based on determining the solution, i.e., the distribution of
fluxes in the cell, that corresponds to the optimal state of the
cell in the sense of some fitness objective function (see Note 8).
2. First, we determined the space of solution fluxes vi that maxi-
mizes the objective function defined for each cell in Recon 2
by solving the following constraint-based optimization
problem
Sv = 0
ai £ vi £ bi (2)
max b = c T v
where S is the stoichiometry matrix, where each metabo-
lite has a row and each reaction a column, and stoichiometric
coefficients are mapped into each column; αi and βi are the
nominal lower and upper bounds for each flux vi, and cT is a
vector of coefficients corresponding to the biomass function
in the cell model.
3. Once the optimal biomass b0 has been calculated, in order to
determine the allowed upper and lower bounds for each flux,
vilb, viub respectively, for each reaction flux vi, we calculated an
optimization to maximize/minimize the flux under the addi-
tional constraint of setting the biomass to its optimal b0:
Sv = 0
ai £ vi £ bi
(3)
c T v = b0
min/max vi i ¹ j

4. To compute such FVA we use the COBRApy package [4] (see

Note 9).
The source code that uses the COBRApy package to perform FVA
analysis for the human hepatocytes is provided at https://github.
com/phi-grib/Hepatoxicity-SysBio/tree/master/recon2_fva
5. By solving the given equations, we obtained for each flux the
allowed levels of variability in the model that will not alter cell
fitness objectives (File S2). We will refer to these values as flux
elasticities (see Note 10).
1. Load liver hepatocytes metabolic model.
2. Perform FVA for liver model (see https://github.com/
phi-grib/Hepatoxicity-SysBio/blob/master/recon2_
fva/compute_fva.py).
(a) The constraints are based on:
●● The stoichiometric matrix (S) of the reactions
(metabolites and reactions).
●● The upper and lower bounds of the fluxes of the
reactions (αi and βi).
●● The optimal biomass of the reactions (b0).
(b) Compute maximum and minimum fluxes of the reac-
tions (vi) satisfying these constraints and maximize the
objective function.
Output of the step: List of reactions with upper and lower bounds
that will not alter the cell’s objectives (File S2).
3.4 Estimation In the fourth step, we will compute the global perturbation of the
of the Perturbation metabolic network from the chemical perturbation.
Elasticity
1. From the previous computations, we can define a coefficient in
order to estimate the global network perturbation using the
flux variability information as a measure of robustness of the
system in combination with differential gene expression infor-
mation to identify the points of perturbation in the network,
to ultimately characterize cell responses under chemical per-
turbations. To that end, we can define a coefficient to estimate
the overall level of network elasticity perturbation of a given
reaction flux k due to the chemical perturbation i in cell type j
as follows:

( ) ( )
Dijk = dijk vklb H -dijk + dijk vkub H dijk (4)
where H is the Heavyside function2:

1 dijk ³ 0
( )
H dijk = {
0 dijk < 0
(5)

2. The level of perturbation shown in previous Eq. 4 estimates

the maximum change that can take place in a given reaction k
under perturbation without affecting the cell steady state equi-
librium. We should expect that more toxic perturbations
should be associated with reactions having lower elasticities
and therefore with associated perturbation having a higher
impact on the cell’s state (see Note 11).
1. Use the previous computed upper and lower bound fluxes
( vilb and viub ) .
2. Compute the global perturbation of the metabolic network
(see https://github.com/phi-grib/Hepatoxicity-SysBio/
blob/master/notebook/Statins.ipynb).
By applying Eq. 4 to the δijk (flux k perturbed due to chemical i in
cell type j).
We provide a jupyter notebook that performs network pertur-
bation elasticity for the statins example but can be easily adapted to
other compounds.
3.5 Determination We can determine also the metabolic pathways that are altered
of Altered Pathways by the chemical perturbation. Such information may be used to
obtain additional insights about the toxicity mechanisms.
1. Collect a consensus list of pathways associated with each reac-
tion in the hepatocyte cell model (see Note 12). Such a list can
be obtained from Recon 2, which annotates reactions in the
cell models with associated pathways, also called “subsystems.”
Alternatively, other databases like BioCyc, KEGG or Reactome
associate genes and reactions with pathways and are therefore
valid sources for pathway information (see Note 13).
2. For each compound, according to the list of perturbed reac-
tions obtained from Eq. 1, determine number of perturbed
reactions per pathway.
3. Pathways containing a high percentage of perturbed reactions
are more likely to be associated with a mechanism of response
to toxicity (see Note 14).
2
http://mathworld.wolfram.com/HeavisideStepFunction.html
We provide a jupyter notebook that obtains the pathways altered

for the example of statins but can be easily adapted to other
compounds.
4 Example
We will show an example of the method by applying it to three

statins: pravastatin, simvastatin and rosuvastatin, which are
cholesterol-lowering drugs.
All three drugs have been linked to drug-induced liver injury
[9]. Rosuvastatin appears to be more likely to cause clinically
apparent liver injury [10].
4.1 Extraction We start by searching in the LINCS database the identifiers of

of Gene Expression the chemical perturbations corresponding to the three statins
Data from the LINCS (Table 2).
Database To extract the information for a gene expression profiling from
LINCS database we extract the corresponding information from
the following signatures of the three statins as seen in Table 3.
All perturbations have been observed in hepatocytes (PHH)
and with a dosage of 10 μM for a period of 24 h.
Table 2
LINCS database identifiers corresponding to the three statins
Compound name LINCS chemical perturbation ID

Pravastatin BRD-A71816415
Simvastatin BRD-K22134346
Rosuvastatin BRD-K82941592
Table 3
LINCS gene expression profiling signatures for the three statins
Compound name CHEBI identifier LINCS gene expression profiling signatures

Pravastatin CHEBI:63618 CPC015_PHH_24H:BRD-A71816415-236-03-5:10
Simvastatin CHEBI:9150 CPC015_PHH_24H:BRD-K22134346-001-11-6:10
Rosuvastatin CHEBI:3854 CPC015_PHH_24H:BRD-K82941592-238-02-9:10
We obtain for every statin the 50 most overexpressed and

underexpressed genes.
The full list is shown in Table S3 of supplementary material.
4.2 Mapping Gene We used the gene expression obtained in the previous step (File S3)
Expression to Altered to map it to the reactions altered based on the GPR rules from
Reactions Recon 2 (File S1) using the method described before.
See the jupyter notebook provided for further details of the
implementation.
4.3 FVA for Statins The result of the Flux elasticities obtained from the Recon 2 meta-
bolic model (File S2) was applied to the three statins and we obtain
a list of the upper and lower bounds for each perturbed reaction in
our example for statins. We obtained the pathways altered as well
(fragment). See Table 4.
implementation. The full list is in the S4 file.
4.4 Estimation For the statins example of the network perturbation results are
of the Perturbation shown in Table 5 and Fig. 2.
Elasticity See the jupyter notebook provided for further details of the
implementation.
Table 4
List of pathways altered for the statins example (fragment)
Lower Upper
Reaction Pravastatin Simvastatin Rosuvastatin bound bound Pathway
2OXOADOXm −1 −1 0 0 3 Lysine metabolism
ACACT1r 0 1 1 −1000 333.67 Tryptophan metabolism
ACACT1x 0 1 1 0 774.83 Cholesterol metabolism
ACACT4p 0 0 1 0 501.48 Fatty acid oxidation
Table 5
Network perturbation results for the statins example
Pravastatin Simvastatin Rosuvastatin

Perturbation elasticity 10,160.48 13,254.2 27,857.22
Fig. 2 Network perturbation results bar chart for the statins example (pravastatin, simvastatin, and
rosuvastatin)
4.5 Altered For the example of the statins, File S5 shows the list of pathways
Pathways that were altered in each case and the number of reactions in each
pathway. See Table 6.
implementation.
5 Notes
1. The L1000 platform is a high-throughput gene expression

platform that allows for testing a large number of drug and
compound gene perturbations for different cell types (http://
lincscloud.org). The platform has been successfully applied in
a growing number of studies; the interface is permanently
updated with new cloud big data analysis services.
2. The LINCS dataset can provide more detailed gene expression
data if desired. The advantage of focusing on the 50 top highly
upregulated/downregulated genes is that this information
constitutes a compact and meaningful way of associating a
perturbation signature with each compound, allowing for
instance: (a) comparison between perturbations; (b) pathway
analysis by mapping the signature into the gene network of the
cell (metabolic, signaling, regulatory, etc.).
Table 6
List of pathways altered with the number of reactions in each pathway
Pathway Pravastatin Simvastatin Rosuvastatin

Arginine and proline metabolism 1 0 0
Bile acid synthesis 1 0 0
Cholesterol metabolism 2 8 9
CoA catabolism 0 0 1
CoA synthesis 0 0 1
Fatty acid oxidation 3 4 45
Fructose and mannose metabolism 0 2 1
Glutamate metabolism 1 1 1
Glycine, serine, alanine, and threonine metabolism 1 1 1
Glycolysis/gluconeogenesis 1 2 0
Glyoxylate and dicarboxylate metabolism 0 1 0
Heme synthesis 1 1 0
Inositol phosphate metabolism 3 0 0
Lysine metabolism 1 1 0
Methionine and cysteine metabolism 0 0 1
N-glycan degradation 1 1 0
Nucleotide interconversion 0 0 5
Pentose phosphate pathway 0 0 2
Purine catabolism 1 1 1
Pyruvate metabolism 0 0 5
Sphingolipid metabolism 2 1 1
Transport, endoplasmic reticular 0 1 0
Transport, Golgi apparatus 1 0 0
Tryptophan metabolism 0 1 1
Unassigned 1 1 2
Valine, leucine, and isoleucine metabolism 1 0 0
3. The LINCS database can be used as a training set to develop a

quantitative structure–activity relationship (QSAR) model to
predict chemical induced gene expression signature, as recently
shown in [11]. Such predictive model in combination with the
systems biology hepatotoxicity prediction method presented
here, allows performing virtual screening campaigns for com-
pound libraries.
4. Since its publication in 2013 [2], Recon 2 has been widely

adopted by the community as the reference systems biology
model of human metabolism, from cancer studies to the
development of novel therapies in diabetes and has been inte-
grated with omics data from transcriptomics, proteomics,
metabolomics, and fluxomics.
5. The development of precise GPR rules is essential as it pro-
vides a link between metabolic pathways (reactions) and gene
expression of enzymes. GPR rules are cell type-specific and can
be extended in order to include regulation.
6. The fact that on average only half of the upregulated/down-
regulated genes can be mapped into the GPR introduces some
limitations on the coverage of the method. However, we can
only consider downregulation for AND relationship and
upregulation for OR relationships because these are the only
critical ones that we can guarantee having a perturbation effect
on the model.
7. As with the GPR relationships, the goal of performing FVA is
to determine the boundaries of the equilibrium of the system,
in this case of the reaction fluxes. Our sensitivity analysis is
focused on identifying model perturbations by looking at gene
expression states that are not compatible with the equilibrium
boundaries of the system.
8. The objective function in Recon 2 was obtained for each cell type
through a fitting of growth experimental data with a weighted
linear combination of main essential fluxes in the cell [2].
9. Generally FVA flux boundaries are computed for some thresh-
old of the optimal of the objective function b0, i.e., b < λb0.
The assumption is that there is some tolerance or robustness
to transient perturbations in the system and that only
perturbations above such threshold may elicit inhibitory or
toxic effect in the cell.
10. The elasticity values obtained are independent of the chemical
perturbations studied.
11. We are assuming here that each individual perturbation has an
independent effect on the system. However, more complex
analyses are also possible based on the concerted or synergistic
effect that multiple perturbations have in the system. The
topic is beyond the scope of this chapter, so we advise the
reader to refer to the literature [12].
12. The concept of pathway may vary depending on the context.
In metabolic networks, a pathway is understood as a set of
metabolic reactions, generally converting a set of precursor
metabolites into some end products. Cyclic pathways like the
TCA cycle also exist. In systems biology, another possible defi-
nition of pathway is related to some cell function, and there-
fore not only limited to metabolism but also to signaling,
regulation, transport, etc. In that case, pathway is defined by a

set of genes or their products rather than metabolic reactions.
13. The information that is obtained from flux variability analysis
is given in the domain of reactions. If pathway information is
given with respect to genes, pathways can be mapped into the
reactions using gene–protein–reaction (GPR) associations.
14. In addition to pathway enrichment with perturbed reactions,
another useful information is the connectivity of the perturbed
reactions. If perturbed reactions are connected in the pathway,
there is stronger evidence that the pathway has been per-
turbed. This investigation can be facilitated by using a pathway
visualizer like Escher [13].
Acknowledgment
This work was supported by the Innovative Medicines Initiative

(IMI) Joint Undertaking under grant agreement no. 115002
(eTOX), resources of which are composed of financial contribu-
tion from the European Union’s Seventh Framework Programme
(FP7/2007-2013) and EFPIA companies’ in-kind contributions.
References
1. Duan Q, Flynn C, Niepel M et al (2014) 8. Orth JD, Thiele I, Palsson BØ (2010) What is
LINCS Canvas Browser: interactive web app to flux balance analysis? Nat Biotechnol
query, browse and interrogate LINCS L1000 28:245–248
gene expression signatures. Nucleic Acids Res 9. Zhu X, Kruhlak NL (2014) Construction and
42:W449–W460 analysis of a human hepatotoxicity database
2. Thiele I, Swainston N, Fleming RMT et al suitable for QSAR modeling using post-market
(2013) A community-driven global recon- safety data. Toxicology 321:62–72
struction of human metabolism. Nat 10. Ahmed MH, Al-Atta A, Hamad MA (2012)
Biotechnol 31:419–425 The safety and effectiveness of statins as treat-
3. Carbonell P, Lopez O, Amberg A, Pastor M, ment for HIV-dyslipidemia: the evidence so far
Sanz F (2017) Hepatotoxicity prediction by and the future challenges. Expert Opin
systems biology modeling of disturbed meta- Pharmacother 13:1901–1909
bolic pathways using gene expression data. 11. Liu R, AbdulHameed MDM, Wallqvist A
ALTEX 34:219–234 (2017) Molecular structure-based large-scale
4. Ebrahim A, Lerman JA, Palsson BO, Hyduke prediction of chemical-induced gene expression
DR (2013) COBRApy: COnstraints-Based changes. J Chem Inf Model 57:2194–2202
Reconstruction and Analysis for python. BMC 12. Larhlimi A, Blachon S, Selbig J, Nikoloski Z
Syst Biol 7:74 (2011) Robustness of metabolic networks: a
5. (2007) GNU General Public License, version 3 review of existing definitions. Biosystems
6. Juty N, Ali R, Glont M et al (2015) BioModels: 106:1–8
content, features, functionality, and use. CPT 13. King ZA, Dräger A, Ebrahim A, Sonnenschein
Pharmacometrics Syst Pharmacol 4:55–68 N, Lewis NE, Palsson BO (2015) Escher: a
7. Thiele I, Palsson BØ (2010) A protocol for web application for building, sharing, and
generating a high-quality genome-scale meta- embedding data-rich visualizations of biologi-
bolic reconstruction. Nat Protoc 5:93–121 cal pathways. PLoS Comput Biol 11:e1004321
Chapter 24
Nontest Methods to Predict Acute Toxicity: State of the Art

for Applications of In Silico Methods
Ronan Bureau
Abstract
The assessment of acute toxicity of chemicals by in silico methods is actually done by two methodologies,
read-across and QSAR. The two approaches are strongly based on the similarity between the chemical for
which a risk assessment is required and the reference chemical(s) for which the experimental data are
known. Here, we describe the two methodologies with some main publications as illustrations and the in
silico data associated with acute toxicity endpoints (ECHA, REACH) accessible via eChemPortal.
Key words Acute toxicity, Endpoints, In silico methods, Read-across, QSAR, REACH
1 Introduction
Acute toxicity [1] is usually defined as the adverse change(s)

occurring immediately or a short time following a single or short
period of exposure to a substance or substances, or as adverse
effects occurring within a short time of administration of a single
dose of a substance or multiple doses given within 24 h (for an
inhalation exposure within 4 h). The adverse effects can be seen as
mortality, clinical signs of toxicity, abnormal body weight changes,
and/or pathological changes in organs and tissues. At the cellular
level, acute toxicity can be related to three main types of toxic
effect, i.e., (1) general basal cytotoxicity, (2) selective cytotoxicity,
and (3) cell-specific function toxicity. Toxicity to the whole organ-
ism also depends on the degree of dependence of the organism on
the specific function affected [2]. However, acute toxicity tests in
animals (i.e., rat) use mortality as the main observational endpoint
in order to derive a LD50 or LC50 (see below for their definitions).
Specific target organ toxicity (STOT-SE—Single Exposure) could
be considered for chemicals where there is clear evidence of t oxicity
for a specific organ, when it is observed in the absence of a
classification for mortality. Mortalities during the first 72 h after
519
520 Ronan Bureau
Table 1
GHS acute toxicity scheme
Acute toxicity Cat. 1 Cat. 2 Cat. 3 Cat. 4 Cat. 5

Oral (LD50) ≤5 >5 and ≤50 >50 and ≤300 >300 and ≤2000 Anticipated oral
mg/kg LD50 between
2000 and
Dermal (LD50) ≤50 >50 and ≤200 >200 and ≤1000 >1000 and ≤2000
5000 mg/kg
mg/kg
and other
Vapors (LC50) ≤0.5 >0.5 and ≤2 >2 and ≤10 >10 and ≤20 criteria not
mg/L described here
first treatment (in a repeated dose study) may also be considered

for assessment of acute toxicity [3].
The dose-response descriptor for acute toxicity (oral, dermal)
is classically the LD50 (lethal dose 50%) with a concentration in
mg/kg bw/d as unit (mg of compound per kg of body weight
administered per day). For inhalation toxicity, the LC50 (lethal
concentration 50%) is used with a concentration in mg/L as unit
(estimated air concentration of the compound). These values cor-
respond to the acute toxicity estimate named ATE. For the glob-
ally harmonized scheme (GHS) of classification and labeling of
chemicals [4], five GHS categories (categories 1–5) have been
included in the GHS acute toxicity scheme (see Table 1) in func-
tion of the LD50/LC50 values.
The objectives of an acute toxicity study are to establish (1)
whether an exposure to the substance of interest (when adminis-
tered up to the limit dose of 2000 mg/kg bw (oral or dermal
route), or equivalent concentration (inhalation route)) could be
associated with adverse effects on human health; and/or (2) what
types of toxic effects are induced, their time of onset, duration, and
severity (all to be related to dose); and/or (3) the dose–response
relationship to determine the ATE, the discriminating dose, or the
acute toxicity category; and/or (4) when possible, the slope of the
dose–response curve; and/or (5) when possible, whether there are
marked sex differences in response to the substance [2].
According to REACH, the potential to avoid acute toxicity
testing should be carefully exploited by application of read-across
or other nontesting means. To this end, a procedure (Appendix
R.7.4–1 [2]) on a weight-of-evidence (WoE) adaptation of the
standard information requirement for an acute oral toxicity study
should be considered. The WoE adaptation primarily applies to
low toxicity substances. Besides this phrase, we have this
information: “the complexity of the acute toxicity endpoint (pos-
sibility of multiple mechanisms) is one of the reasons for limited
availability and predictivity of QSAR models for this endpoint. In
the absence of complete validation information, available models
Nontest Methods to Predict Acute Toxicity: State of the Art for Applications of In Silico… 521
could be used as a part of the WoE approach for hazard identifica-

tion and risk assessment purposes after precise evaluation of the
information derived from the model”.
For a particular application in REACH, Fig. 1 [2] shows the
decision tree associated with the WoE for which QSAR methodol-
ogy could be a major point beside a subacute oral toxicity (28-day
RDT study) always required for a tonnage above 10 tpa. The alter-
natives to QSAR are dose range finding (DRF) studies (protocol of
acute toxicity but on a large smaller number of animals) or an in
vitro cytotoxic assay (neutral red uptake). Ahead of the last REACH
registration deadline in 2018, ECHA estimates that registrants of
about 550 substances can omit the in vivo acute oral study by using
this adaptation [5].
In 2016, Blomme et al. [6] observed: “Most computational
toxicology models are not sufficiently performant to justify their
use in drug discovery. One of the reasons is the complexity and
large variety of mechanisms of toxicity for any given tissue, as well
Fig. 1 Decision tree to assess whether an in vivo acute toxicity test is required, when the registrant has to
generate a novel repeated dose subacute oral toxicity study [2]
522 Ronan Bureau
as the extreme diversity of chemical structures associated with simi-

lar toxic endpoints. Because most toxic effects are due to interac-
tions with a variety of proteins or protein binding sites and are
caused by structurally dissimilar chemical structures, developing
robust computational tools, which rely on the similarity principle
to predict general end points (such as organ-specific toxic effects or
broad mechanisms of toxicity), is a daunting and likely not achiev-
able task. A more realistic approach is to focus on very specific
endpoints, such as interactions with a specified target protein (e.g.,
a drug transporter or DNA) or a well characterized, narrowly
defined mechanism of toxicity (e.g., phospholipidosis)”.
Starting from these considerations, it seems to be very difficult to
generate (Q)SAR models associated with acute toxicities. However,
we will see that it is possible to obtain reliable assessment by consider-
ing the molecular similarities between chemical structures.
2 Materials
The databases and the software for QSAR and read-across were
described in recent reviews [7–9] and summarized in Table 2.
In silico methods for the prediction of chemical toxicity were
also described in recent reviews [8, 25, 26]. A guidance on
information requirements and chemical safety assessment related to
QSARs and grouping of chemicals was published by ECHA [27].
Table 2
Software and databases related to acute toxicity
Software Databases
TEST (US EPA) [10] ECHA [11]
ACD/TOX [12] OECD QSAR Toolbox [13]
ADMET Predictor [14] RTECS [15]
Multicase MCASE/MC4PC [16] ACToR/EPA [17]
TerraQSAR [18] TOXNET/NLM [19]
TOPKAT [15] ChemIDplus/NLM [20]
Leadscope [21]
eChemPortal [22]
The Merck Index/RSC [23]
PAN [24]
3 Methods
3.1 Read-Across/ The assessment of the acute toxicity starting from read-across rep-
Category Approaches/ resents one main approach for in silico studies. For this approach,
Structural Analog it is necessary to get robust experimental data of similar chemicals
compared to the reference chemical. From these similar chemicals
(classically called nearest neighbors (NNs) in the next parts), the
prediction for the reference chemical will be the same category of
toxicity (like LD50 > 2000 mg/kg bw in case of nontoxicity) or an
average value of the ATE associated with NNs (potentially weighted
by the molecular similarities toward the reference chemical). The
molecular similarity [28] between two compounds is classically
dependent of two parameters: the molecular fingerprint used to
describe the chemicals and the metric to compare these molecular
fingerprints (similarity coefficient or distance between chemicals).
An large number of fingerprints [28, 29] are available, from sub-
structure keys-based fingerprints (like MACCS keys) to topologi-
cal fingerprints. Circular fingerprints (a kind of topological
fingerprints) are also present with a definition of the environment
of each atom around a determined radius (like ECFP4 (extended-
connectivity fingerprints) with a radius corresponding to four
chemical bonds). The metric [29] is often a similarity coefficient,
like Tanimoto or DICE coefficients, with a range of values between
0 and 1 (1: highly similar). It is also possible to explain how a
chemical is similar to another one, starting from the definition of a
particular substructure associated with a chemical family. This will
be described below.
To analyze the in silico methodologies used, we have consid-
ered the endpoint data recorded in eChemPortal (OECD). Among
the databases recorded in eChemPortal, ECHA CHEM represents
the main database (48,250 substances with 598,842 endpoints
recorded in September, 2017). For the property search, we have
considered the three acute toxicities (oral, inhalation, and dermal)
and the associated in silico study result type ((Q)SAR/read-across
based on grouping of substances (category approach)/read-across
from supporting substance (structural analog or surrogate)).
3.1.1 A Typical Case Prior to analyzing the data from ECHA, a recent publication [30]
described well the principle and discussed about the benefits of using
read-across to fulfill hazard data requirements. This recent publica-
tion is based on a category approach used by the American Cleaning
Institute (ACI) to fill hazard data gaps across 261 substances (high
production volume for these substances). Nine chemical categories
were defined with separations based on common functional group(s)
and an incremental and constant change across the category (e.g., a
chain-length category). Subcategories were also established within a
category based on some specific characteristics. For instance, the
524 Ronan Bureau
Fig. 2 Process for developing categories described in the publication of Stanton et al. [30]
largest grouping, aliphatic acids (78 substances), was subcategorized

into 14 subgroups based on degree of saturation, single or multicon-
stituency, and presence and type of salt.
The authors develop their approaches for a particular subset
named the hydrotrope category associated with eight substances.
The basis is the analysis (Fig. 2) of the actual experimental data for
the similar compounds and an evaluation of their reliabilities accord-
ing to the method established by Klimisch et al. [31]. Only studies
with a Klimisch score of “1” (reliable without restriction) or “2”
(reliable with restriction) were considered as robust. For these eight
substances and 11 endpoints, 125 experimental studies were found.
Among them, three are associated with acute oral toxicity, one for
dermal toxicity, and none for inhalation toxicity. Into this category,
by considering a light modification of the alkyl chain (toluene,
xylene, and cumene sulfonic acid) and the fact that the variation of
the salt should not modify the toxicity, they could extrapolate the
results starting from the known robust experimental data. For
instance, the acute oral toxicity of cumene sulfonic acid, a mmonium
salt could be defined starting from the cumene sulfonic acid, sodium
salt. This is confirmed by the fact that the acute oral toxicity of
xylene sulfonic acid, ammonium salt (>2000 mg/kg) is in the same
range than those recorded for xylene sulfonic acid, sodium salt
(>5000 mg/kg). Moreover, the acute oral toxicological test is not

necessary for xylene sulfonic acid, potassium salt because the tolu-
ene sulfonic acid, sodium and potassium salts have nearly the same
values (>2000 mg/kg). For acute dermal toxicity, only one data is
reliable but with a LD50 > 2000 mg/kg. Another data exists with
the same value but not reliable (Klimisch score). They consider than
this value could be extrapolated for the overall subset. For inhala-
tion toxicity, no reliable data is indicated in this publication. We
have no real information about the process associated with the
other categories but the same logic with the definition of subcate-
gories (small subset) should have been achieved. They claim that on
the 262 substances, 174 tests on acute oral toxicities could be
avoided leading to a total cost saved on around 400,000 USD and
2088 animals not used (12 animals by test).
3.1.2 General View The property search “oral acute toxicity” and a study type associ-
of the Methodology ated with read-across based on grouping of substances (category
Through the Registration approach) lead to 2654 results. By filtering on the substance name
Files (ECHA, eChemPortal) (several records for a same substance), it remains 566 substances.
The term “dermal acute toxicity” leads to 499 substances. The
term “inhalation acute toxicity” leads to 511 substances. By con-
sidering the three property search together, this is associated with
806 unique substances. The property search “oral acute toxicity”
and a study result type associated with read-across from supporting
substances (structural analog or surrogate) lead to 3194 results
with 1119 unique substances (based on the substance name). The
term “dermal acute toxicity” leads to 936 substances. The term
“inhalation acute toxicity” leads to 766 substances. By considering
the three property search together, this is associated with 1568
unique substances. For the registration files, the two in silico meth-
ods (grouping or supporting substances) are very close and the
consideration of supporting substances with robust experimental
data is largely the basis for the assessment of the prediction.
For grouping of substances and read-across, a typical example
is associated with the use of QSAR Toolbox [13]. The main
approach is to consider for the risk assessment the average value
(LD50) of the nearest neighbors (NN, 5 NN classically). Associated
with this molecular similarity, the notion of validation domain is
defined. The predictions were made on a chemical subset which
respect a validation domain specify by a logical expression like: “a”
and “b” (not “c”). An illustration of the parameters of the v alidation
domain is for instance:
–– “a”: the target chemical should be classified as high (Class III)
by Toxic hazard classification by Cramer.
–– “b”: the target chemical should be classified as No alert found
by DNA binding by OASIS.
526 Ronan Bureau
–– “c”: the target chemical should be classified as aromatic amines

OR azo compounds OR benzyl and alkyl halides OR isocya-
nates OR ……
So, the validation domain consists to define some constraints
associated with several properties (physical or biological) described
in the QSAR toolbox or some constraints related to the relation-
ship with specific chemical families (see case “c”). At the end, from
the final chemical subset associated with the validation domain, a
definition of the NNs is done with a molecular fingerprint (Atom
pairs for instance), a metric (Dice for instance), and a threshold for
the cutoff (50% for instance). The characteristics and the values
associated with the molecular fingerprint and the threshold are
potentially variable but the three previous parameters are often
observed (Atom pairs, Dice, 50%).
Another example with a short definition of the validity domain
is to consider two criteria for the structures (“a” and “b”) and two
criteria for the physicochemical properties (“c” and “d”). In this
example, the criterion “a” defines the chemical category associated
with the compound (anilines for instance), the criterion “b”
defined the molecular similarity (fingerprint/metric/cutoff)
toward the members of this chemical family. The criteria “c” and
“d” are related to physicochemical properties of the compounds in
relation with the chemical subset selected. The descriptor logKOW
(vide infra) is logically taken as basis with a max and a min value.
The prediction is made from the average value of the NNs with for
the prediction a Klimisch score of “2” (reliable with restriction).
This score is a function of the quality of the experimental data asso-
ciated with the NNs.
NH2
LD50 between 300 and 2000 mg/kg bw
O
1
NH2 NH2 NH2 NH2
O O
O O
NH 2
Fig. 3 Assessment of two enantiomers of 1 and three similar chemicals by considering 1 as basis
For structure analogs and read-across, a particular example con-

cerned the assessment of several compounds (Fig. 3) from the race-
mic form of 1-(3-methoxyphenyl)ethylamine 1 with in all this case a
reliability of 1 for the prediction (oral acute toxicity). The toxicity of
the enantiomers of 1, the enantiomers of the para-analog of 1 (very
high similarity), and even a compound with a methyl group instead
of a methoxy group was estimated with only the compound 1 as
reference by considering the robustness of the data (OECD guide-
line 423 with GLP compliance for experimental determination) and
the range of values associated with 1 (LD50 between 300 and
2000 mg/kg bw corresponding to the category 4 (GHS)).
For the ester derivatives of aromatic carboxylic acids with long
alkyl chains, analogues with “equivalent” alkyl chains were consid-
ered successfully as basis for the prediction of oral acute toxicity
(reliability of 1 or 2). Other examples concern the assessment of
the dermal acute toxicity of (4-chlorophenyl)hydrazine by consid-
ering phenylhydrazine (reliability of 2), the assessment of the inha-
lation acute toxicity of l-menthol from dl-menthol (reliability of
1) or the assessment of the inhalation acute toxicity of 1,2-hexanediol
from the 1,2-pentanediol (reliability of 2).
3.2 QSAR The partition coefficient between octanol and water (logKOW), as
hydrophobic descriptor, is strongly present for QSAR associated
3.2.1 LogKOW as Main with acute mammalian toxicity. For instance, Eq. 1 is associated
Chemical Descriptor with alcohols (n = 57, s = 0.24, r2 = 0.956). LogKOW is also the
Coupled with Hansch’s main descriptor for the acute toxicity in ecotoxicology for which
Approach the narcosis process represents the main mode of action (MOA) of
the chemicals [32]. Based on the ratio [33] between a predicted
value with logKOW (see Eq. 1, acute oral toxicity) and the experi-
mental value, the notion of Toxic Ratio (TR) was exploited to esti-
mate a particular MOA for some derivatives (this was done also in
ecotoxicology). From their analysis (108 chemicals, acute oral tox-
icity), baseline toxicity may be expected for alcohols, acids, ketones,
and one-ring aromatics. Aldehydes and amines show excess t oxicity,
whereas simple aliphatic hydrocarbons have a miscellaneous mech-
anism of action. Esters show less than baseline toxicity.
- log LD50 ( mol / kg ) = 0.6 log K ow - 0.8 log ( 0.08K ow + 1) + 1.13 (1)

This TR study was also done for inhalation toxicity with
another equation (Eq. 2, n = 18, r2 = 0.964, s = 0.140) related to
the air–water partition coefficient Kaw. The values of Kaw were lack-
ing for 65 chemicals (out of the 108). They also indicated than the
reliability of the TR values of some of the remaining 43 chemicals
may be questionable, due to either problems associated with the
calculation of the baseline toxicity value or errors associated with
the experimental values.
528 Ronan Bureau
- log LC50 ( ppm ) = 0.663 log K ow - 0.589 log K aw - 6.120 (2)

Logically by considering the Hansch’s approach [34] and the
characterization of outliers for equations with logKOW only (values of
TR), steric and electronic descriptors were added for specific chemi-
cal families [35]. For instance with anilines (n = 29), beside logKOW,
the molecular volume (steric descriptor) and the Lowest Unoccupied
Molecular Orbital (LUMO, electronic descriptor) were considered
in the QSAR equation (r2 = 0.9, s = 0.1). LUMO with logKOW were
also the basis for several QSAR equations associated with some sub-
family of chemicals [36]. Always with electronic descriptors, for dif-
ferent chemical families, hardness and ionization potential were used
in combination with logKow (parabolic relationship).
For a complete view of the descriptors, the chemical families,
QSAR equations, we can see the reviews of Tsakovska et al. in 2008
[35], Devillers et al. in 2009 [37], and Lapenna et al. [7]. In the
review of Devillers et al. [37], the potential of interspecies correla-
tions was also particularly analyzed.
3.2.2 General View
of the QSAR Approaches In this case, it is not the prediction of the ATE value which is
in Relation with the GHS straight considered for the reliability of the prediction but the
Category positioning into a particular toxic class related normally to the
GHS acute toxicity scheme (see Table 1).
Fig. 4 Classification performance for in-model-training/out-model-training set compounds from

Diaza et al. [38]
Fig. 5 Modeling workflow. (a) Preparation of the target data set. (b) Modeling procedure for qHTS LD50 data
set from [42]
A publication [38] analyses the potential of governmental

(TEST from US-EPA [10]) and commercial software (TOPKAT
from BIOVIA [15], ToxSuite (ACD-labs [12]), TerraQSAR from
Terrabase [18], ADMET predictors from simulation Inc [14]) on
a set of 7417 chemicals, for the prediction of acute oral toxicity, by
considering the five GHS categories. The quality of the classifica-
tion for each program is described on Fig. 4. Confusion matrix and
statistics on the whole dataset for each software is provided.
Another study [39] considers a specific database (Munro data-
base [40]) associated with 438 chemicals. Acute toxicity values
were discretized in 3 class on the basis of the globally harmonized
scheme: highly toxic, intermediate toxic, and low to nontoxic.
Starting from DRAGON descriptors [41], a k-nearest neighbors
(kNN) classification method was calibrated on 25 molecular
descriptors (genetic algorithm for the selection of the descriptors)
and gave a non-error rate (NER) equal to 0.66 and 0.57 for inter-
nal and external prediction sets, respectively. It is interesting to
notice than the “k” selection with fivefold cross validation gave an
optimal k value of 1 (only one compound for the prediction!). The
authors consider than the analysis of the selected descriptors
(principal component analysis) and their relationship with toxicity
levels constitute a step towards the development of a global QSAR
model for acute toxicity.
Another very interesting study [42] considers beside chemical
descriptors, biological descriptors as key element for improving the
accuracy of the classification (two subclasses toxic and nontoxic
chemicals were considered with a cutoff dependent of the LD50
values). The overall process is schematized on Fig. 5. Dose–
response data points of the quantitative high-throughput screen-
530 Ronan Bureau
Fig. 6 External prediction results of kNN models using the distribution of the predicted values (chemical only
descriptors (left) vs the best hybrid descriptors (right)) [42]
ing (qHTS) assays were served as biological descriptors. A particular

treatment of the dose-response curves was carried out to generate
a fingerprint for each compound (182 descriptors for 13 cell lines
and 14 concentrations for each cell line). An essential novel noise-
filtering algorithm was defined at this level (“baseline threshold”
THR values). 382 descriptors were used for chemical descriptors
(Dragon software [41]). Random forest (RF) and k-nearest neigh-
bors (kNN) algorithms were applied to these data. Figure 6 shows
clearly the improvement of the separation between the two sets by
considering the hybrid methods. The quality of the predictions
(external validation sets) was really higher than classical QSAR
with chemical descriptors and also with a software like TOPKAT
[15]. A sensitivity value of 0.93 (external data set) could be reached
with kNN or RF with the hybrid descriptors (0.45 for TOPKAT
and 0.73 with chemical descriptors only).
3.2.3 QSAR and ECHA
(eChemPortal) The property search “oral acute toxicity” and a study type associated
with (Q)SAR lead to 203 results with 152 unique substances.
The term “dermal acute toxicity” leads to 33 unique sub-
stances. The term “inhalation acute toxicity” leads to 49 unique
substances. The three properties together leads to 183 unique sub-
stances. For oral acute toxicity, 69 out of the 152 unique sub-
stances have really a (Q)SAR analysis described in the registration
files. QSAR Toolbox was used in 27 cases with the same technic
described previously for read-across (average values of nearest
neighbors for the prediction), the software from ACD [12] was
used in 22 cases, TEST [10] from US EPA in 17 cases, an artificial
neural networks (ANN) QSAR method in two cases and no spe-
cific information appears in one case. These studies have, for the
reliability of the prediction, a value of 2 except with TEST V 4.1
for which a value of 3 (not reliable) is observed for 10 cases out of
the 11 cases with this version.
With ACD, the main parameter for the prediction is related to
the value of an internal reliability index (RI) based on the training
set associated with the prediction (RI value superior to 0.5 are reli-
able for the software). RI value takes into account the molecular
similarity between the training set and the compound. The soft-
ware displays up to the five most similar structures with their
experimental data for validating the prediction. When this RI value
is close to the limit (RI = 0.3, no reliability), the prediction is
clearly defined in the conclusions has borderline reliable but, for
the administrative data (registration file), the reliability is still
defined as 2. A QMRF ((Q)SAR Model Reporting Format) is
recorded (ACD/Percepta, Q32-48-43-425 and 426), at the JRC
[43], for rat and mouse acute oral toxicity. It allows to understand
the statistical method and the descriptors used for the prediction.
The statistical method is named GALAS corresponding to a com-
bination of two systems: a first PLS model for the prediction of
LD50 and a local correction based on the model performance for
similar compounds.
With TEST version 4.1, the ten compounds with a low reli-
ability have always the same conclusion: “A prediction is made but
the result may not be reliable as only moderately similar com-
pounds with known experimental values in the training set have
been found”. Beside these ten compounds, four compounds have
a prediction based on TEST version 3.2 and the reliability is 2 (but
it is impossible to have an overall view of the analysis). We have
compared the results with the version 4.2.1 of TEST by consider-
ing two out of the ten compound and one (CAS number: 24650-
42-8) out of the four compounds predicted with the version 3.2.
We have used the consensus method for the prediction (consensus
of hierarchical clustering, FDA and nearest neighbors). An applica-
bility domain is defined before to carry out a prediction [10]. With
the CAS number 24650-42-8 associated with a reliability of 2, the
prediction is explained by considering similar compounds in the
external testing set (Mean Absolute Error (MAE) of 0.78 for com-
pounds with a similarity coefficient ≥0.5). If we start with one of
the derivatives with a reliability of 3 like the 2-phenylisobutyric
acid, the final results seems better by considering the MAE (MAE
of 0.29). The problem could be related to the absence of an acidic
function for the NNs. However, for the dicyclohexyl ketone (CAS
number: 119-60-8), other chemical with a reliability of 3, a very
high similar compound is found (CAS number 90-42-6, similar-
ity = 0.88) and the quality of the prediction for the NNs is correct.
So, it is difficult to understand why we have a reliability of 2 in one
case and a reliability of 3 in the two other cases. The quality of the
data (robust data) associated with the NNs must be the explana-
tion (ChemIDplus database from US National library of Medicine).
Indeed, the source for the LD50 associated with the NN of dicy-
clohexyl ketone (CAS number 90-42-6) is related to “raw material
data handbook, vol. 1, 1974”. This information (experimental
532 Ronan Bureau
data) is not an agreement with the notion of robust data defined by

ECHA (OECD guideline).
Two registration files are associated with an artificial neural
networks (ANN) QSAR model for the prediction of oral acute tox-
icity. This is related to the QMRF number Q17-10-1-297
(Molcode model, [43]). The nonlinear QSAR consist to do a back-
propagation neural network (multilayer perceptron) regression.
Beside the prediction, the applicability domain is based on the
descriptor values (training set values ±30%) and on the presence of
the same of equivalent functional groups in the training set. The
prediction for NNs with known experimental data is also given.
The experimental values associated with the initial dataset are
described to be robust data (OECD protocol). A prediction reli-
ability is also given in one of the two cases.
In a general way, only ACD (ACD/percepta), Molcode
(ANN), and BIOVIA (TOPKAT) have deposited a QMRF for the
oral acute toxicity endpoint. No QMRF is recorded for dermal and
inhalation toxicities.
For dermal acute toxicity (33 substances theoretically), some
assessments with QSAR Toolbox are found (average values of NNs
with the same methodology previously described) but most of the
time, the registrations files are associated with read-across or
experimental studies. For inhalation acute toxicity (49 substances
theoretically), a very large majority of assessments is done with
QSAR Toolbox (average values of NNs with the same methodology
previously described). No other technic is described except the use
ANN QSAR model in some few cases.
4 Conclusion
Structural analogs represent the main approach for the prediction of

the acute toxicity. The reliability is function of the level of molecular
similarity (similarity values, same functional groups, same
substructures) between the chemical compounds and the quality of
the experimental data associated with the reference chemicals. It is
necessary to start from robust data and the best situation is to consider
experimental studies following OECD guideline associated with GLP
compliance. It is possible to reach a reliability of 1 for a set of chemi-
cals with only one robust data for one chemical as a basis. It is the
same logic for QSAR methodology. Indeed, whatever the approaches,
the objective will be to demonstrate that your model is able to explain
the toxicity of the nearest neighbors and for which the experimental
data are robust. The notion of applicability domain is still a problem
and several alternatives are possible. Classically, this domain is the
function of the descriptor values and/or the molecular similarity
between the training set and your compounds. From REACH pro-
gram, robust data are available and others should be available in the
future. These new data will be the basis for a larger application of in
silico methodology. It is sure that actually most of the predictions are
related to derivatives with low toxicity (the WoE proposed primarily
applies to low toxicity substances) and the challenge will be to inte-
grate chemicals with highest toxicity in the future. The consideration
of data related to systemic biology should be an opportunity for
reaching this objective as described briefly with the consideration of
chemical and biological descriptors in a QSAR approach.
References
1. Walum E (1998) Acute oral toxicity. Environ 12. ACD ACD/Percepta. http://www.acdlabs.
Health Perspect 106(Suppl 2):497–503 com/products/percepta/physchem_adme_tox/
2. ECHA (2016) Acute toxicity. Guidance on 13. QSAR Toolbox. https://www.qsartoolbox.
IR&CSA, Section R.7.4. https://echa.europa. org/
eu/guidance-documents/guidance-on-infor- 14. SimulationsPLUS ADMET predictors.
mation-requirements-and-chemical-safety-as- http://www.simulations-plus.com/software/
sessment admet-property-prediction-qsar/
3. ECHA (2016) Guidance on the application 15. BIOVIA QSAR, ADMET and predictive toxi-
of CLP criteria. https://echa.europa.eu/ cology. http://accelrys.com/products/col-
regulations/clp/classification laborative-science/biovia-discovery-studio/
4. UNECE globally harmonized system of clas- qsar-admet-and-predictive-toxicology.html
sification and labelling of chemicals (GHS). 16. Multicase. http://www.multicase.com/
http://www.unece.org/trans/danger/publi/ 17. actor. https://actor.epa.gov/actor/home.
ghs/ghs_rev00/00files_e.html xhtml
5. Gissi A, Louekari K, Hoffstadt L, Bornatowicz 18. Terrabase TerraTox. http://www.terrabase-
N, Aparicio AM (2016) Alternative acute oral inc.com/
toxicity assessment under REACH based on
sub-acute toxicity values. ALTEX 34:353–361 19. TOXNET. https://toxnet.nlm.nih.gov/
6. Blomme EA, Will Y (2016) Toxicology strat- 20. ChemIDPLus. http://sis.nlm.nih.gov/chem/
egies for drug discovery: present and future. alllocators.html
Chem Res Toxicol 29:473–504 21. Leadscope. http://www.leadscope.com/
7. Lapenna S, Fuart-Gatnik M, Worth A (2010) 22. eChemPortal. https://www.echemportal.
Review of QSAR models and software tools org/echemportal/index.action
for predicting acute and chronic systemic 23. The Merck Index. https://www.rsc.org/
toxicity. JRC Scientifica and Technical Merck-Index/
Reports. http://publications.jrc.ec.europa. 24. PAN Pesticide Action Network. http://www.
eu/repositor y/bitstream/JRC61930/ pesticideinfo.org/
eur_24639_en.pdf 25. Raies AB, Bajic VB (2016) In silico toxicol-
8. Nicolotti O, Benfenati E, Carotti A, Gadaleta ogy: computational methods for the predic-
D, Gissi A, Mangiatordi GF, Novellino E tion of chemical toxicity. Wiley Interdiscip Rev
(2014) REACH and in silico methods: an Comput Mol Sci 6:147–172
attractive opportunity for medicinal chemists. 26. Patlewicz G, Fitzpatrick JM (2016) Current
Drug Discov Today 19:1757–1768 and future perspectives on the develop-
9. Toropov AA, Toropova AP, Raska I, ment, evaluation, and application of in silico
Leszczynska D, Leszczynski J (2014) approaches for predicting toxicity. Chem Res
Comprehension of drug toxicity: software and Toxicol 29:438–451
databases. Comput Biol Med 45:20–25 27. ECHA (2008) Guidance on information
10. TEST user’s guide for T.E.S.T. (version 4.2) requirements and chemical safety assessment,
(toxicity estimation software tool) a program QSAR and grouping of chemicals (Chapter
to estimate toxicity from molecular structure. R.6). https://echa.europa.eu/guidance-doc-
https://www.epa.gov/chemical-research/ uments/guidance-on-information-require-
users-guide-test-version-42-toxicity-estima- ments-and-chemical-safety-assessment
tion-software-tool-program-estimate 28. Muegge I, Mukherjee P (2016) An overview
11. ECHA (2017) https://echa.europa.eu/ of molecular fingerprint similarity search in
534 Ronan Bureau
virtual screening. Expert Opin Drug Discov The importance of hydrophobicity and elec-
11:137–148 trophilicity descriptors in mechanistically-
29. Cereto-Massague A, Ojeda MJ, Valls C, based QSARs for toxicological endpoints. SAR
Mulero M, Garcia-Vallve S, Pujadas G (2015) QSAR Environ Res 13:167–176
Molecular fingerprint similarity search in vir- 37. Devillers J, Devillers H (2009) Prediction of
tual screening. Methods 71:58–63 acute mammalian toxicity from QSARs and
30. Stanton K, Kruszewski FH (2016) Quantifying interspecies correlations. SAR QSAR Environ
the benefits of using read-across and in silico Res 20:467–500
techniques to fulfill hazard data require- 38. Gonella Diaza R, Manganelli S, Esposito
ments for chemical categories. Regul Toxicol A, Roncaglioni A, Manganaro A, Benfenati
Pharmacol 81:250–259 E (2015) Comparison of in silico tools for
31. Klimisch HJ, Andreae M, Tillmann U (1997) evaluating rat oral acute toxicity. SAR QSAR
A systematic approach for evaluating the qual- Environ Res 26:1–27
ity of experimental toxicological and ecotoxico- 39. Chavan S, Nicholls IA, Karlsson BC, Rosengren
logical data. Regul Toxicol Pharmacol 25:1–5 AM, Ballabio D, Consonni V, Todeschini R
32. Lipnick RL (1999) Correlative and mechanis- (2014) Towards global QSAR model building
tic QSAR models in toxicology. SAR QSAR for acute toxicity: Munro database case study.
Environ Res 10:239–248 Int J Mol Sci 15:18162–18174
33. de Wolf W, Lieder PH, Walker JD (2004) 40. Munro IC, Ford RA, Kennepohl E, Sprenger
Application of QSARs: correlation of acute JG (1996) Correlation of structural class with
toxicity in the rat following oral or inhalation no-observed-effect levels: a proposal for estab-
exposure. QSAR Comb Sci 23:521–525 lishing a threshold of concern. Food Chem
34. Hansch C, Leo A, Hoekman D (eds) (1995) Toxicol 34:829–867
Exploring QSAR: hydrophobic, electronic, 41. TALETE Dragon 7. http://www.talete.mi.it/
and steric constants, vol 2. ACS professional products/dragon_description.htm
reference book. American Chemical Society, 42. Sedykh A, Zhu H, Tang H, Zhang L, Richard
Washington, DC A, Rusyn I, Tropsha A (2011) Use of in vitro
35. Tsakovska I, Lessigiarska I, Netzeva T, Worth HTS-derived concentration-response data as
AP (2008) A mini review of mammalian toxic- biological descriptors improves the accuracy
ity (Q)SAR models. QSAR Comb Sci 27:41–48 of QSAR models of in vivo toxicity. Environ
36. Cronin MTD, Dearden JC, Duffy JC, Edwards Health Perspect 119:364–370
R, Manga N, Worth AP, Worgan ADP (2002) 43. QMRF, JRC Europen Commission. http://
qsardb.jrc.it/qmrf/search_catalogs.jsp
Chapter 25
Predictive Systems Toxicology

Narsis A. Kiani, Ming-Mei Shang, Hector Zenil, and Jesper Tegner
Abstract
In this review we address to what extent computational techniques can augment our ability to predict
toxicity. The first section provides a brief history of empirical observations on toxicity dating back to the
dawn of Sumerian civilization. Interestingly, the concept of dose emerged very early on, leading up to the
modern emphasis on kinetic properties, which in turn encodes the insight that toxicity is not solely a prop-
erty of a compound but instead depends on the interaction with the host organism. The next logical step
is the current conception of evaluating drugs from a personalized medicine point of view. We review recent
work on integrating what could be referred to as classical pharmacokinetic analysis with emerging systems
biology approaches incorporating multiple omics data. These systems approaches employ advanced statisti-
cal analytical data processing complemented with machine learning techniques and use both pharmacoki-
netic and omics data. We find that such integrated approaches not only provide improved predictions of
toxicity but also enable mechanistic interpretations of the molecular mechanisms underpinning toxicity
and drug resistance. We conclude the chapter by discussing some of the main challenges, such as how to
balance the inherent tension between the predicitive capacity of models, which in practice amounts to
constraining the number of features in the models versus allowing for rich mechanistic interpretability, i.e.,
equipping models with numerous molecular features. This challenge also requires patient-specific predic-
tions on toxicity, which in turn requires proper stratification of patients as regards how they respond, with
or without adverse toxic effects. In summary, the transformation of the ancient concept of dose is currently
successfully operationalized using rich integrative data encoded in patient-specific models.
Key words Toxicology, Systems biology, Network pharmacology, Algorithmic complexity, Omics
1 A Brief History of Toxicology—From Sumerian Drugs to Pharmacokinetic

Analysis of Toxicity
There are numerous examples of “drug” usage in ancient times.

The first documented evidence of drug receipts is believed to be
approximately 5000 years old, on a Sumerian clay slab [1]. In
contrast to the long history of using substances from plants for
therapeutic purposes, it was only a couple of hundred years ago
that people realized the hazards of these substances. This insight
can be expressed as “All substances are poisons; there is none
which is not a poison. The right dose differentiates a poison and
a remedy” [2]. While Paracelsus (1493–1541) had this key
535
536 Narsis A. Kiani et al.
insight, the boundary between poison and remedy is hazy. The

toxicity of individual chemicals is indeed a complex feature
which itself depends on several factors, such as dose, chemistry,
individual genetic make-up and exposure to environmental con-
ditions, which all play key roles, to different degrees, in deter-
mining susceptibility to disease and adverse drug responses. In
modern times it has become increasingly evident that it is not
the case that each medicine works equally well, as regards both
efficacy and safety, in individuals in a population—hence the
rationale behind the idea of personalized medicine [3].
Following the work of Paracelsus, Mathieu Orfila (1787–1853)
first described specific organ damage caused by toxins. Toxicity
studies of individual substances using animals began in 1920.
J.W. Trevan proposed the concept of a 50% lethal dose (LD50),
defining the lethal dose of individual chemicals. As a new sub-
ject, the field of toxicology slowly developed until the occur-
rence of the thalidomide disaster in the early 1960s, one of the
gloomiest episodes in pharmaceutical history. The drug was
approved as a mild sleeping pill with a good safety profile and
beneficial effects on morning sickness in pregnant women.
However, this caused thousands of babies worldwide to be born
with malformed limbs in less than 4 years. Since then, all regula-
tory agencies have made it obligatory to report the toxicity pro-
files of Investigational New Drugs (IND). In the late 1980s, the
Organization for Economic Co-operation and Development
(OECD) and the International Conference on Harmonization
(ICH) brought out the guidelines for the toxicity testing of
pharmaceutical substances, which are still in use, supplemented
with occasional amendments. In the context of regulatory
guidelines, the lowest dose able to induce adverse effects
(LOAEL) and the highest dose without observable adverse
effects (NOAEL) must be tested to extrapolate the derived no-
effect level (DNEL), which is more useful in defining the appro-
priate dose in clinical trials. Other conventional toxicity testing
includes repeated dose toxicity testing, carcinogenicity testing,
one-generation reproduction toxicity testing, and two-genera-
tion reproduction toxicity testing, et al. These depend on the
formulation and indication of the drug. The toxicity testing of
pharmaceuticals depends strongly on different animal models.
Not surprisingly, such an evaluation is expensive (reported to
cost more than $3B per year), time-consuming (two-generation
reproduction toxicity testing takes around 2 years), suffers from
low throughput, and in some cases raises ethical concerns
related to animal welfare [3]. The low throughput of toxicity
testing methods has serious consequences for public health, as
86% of chemicals (not limited to drugs) currently on the market
lack the necessary toxicity data [4, 5]. The most controversial
issue is the translational efficiency of those compounds being
Predictive Systems Toxicology 537
tested in humans [6]. No doubt, the current toxicity model is

not optimal, motivating both regulatory authorities and phar-
maceutical companies to promote innovative alternatives to
limit the use of animals and to better assess the risk of drug
candidates as early as possible. In 2003, an EPA report pro-
posed a computational toxicology research agenda promising
several advantages, including prioritizing candidates and devel-
oping predictive models for quantitative risk assessment. Yet the
use of computational methods to predict toxicity has a history
in toxicology. In 1962, Hansch et al. developed a Quantitative
Structure–Activity Relationship (QSAR) model to estimate the
concentration of chemicals using the octanol–water partition
and the Hammett constant, which laid the foundation for in
silico toxicity prediction [7]. Numerous tools were developed
to predict carcinogenicity, mutagenicity, and developmental
toxicity using prebuilt QSAR models such as TopKat and
METEOR, most of which have been modified and are currently
deployed in academia and the pharmaceutical industry [8].
QSAR models provide a wide range of complexity for toxic end-
points, given flexible feature selection, i.e., qualitative and
quantitative toxicity plus molecular descriptors can be used. Yet
QSARs require a large dataset to produce robust statistics,
which makes the framework less useful in applications where
data is limited. Benezra [9] used structural alerts (SAs) (also
called toxicophores/toxic fragments) for skin sensitization in
1982, which was more practicable and economical with the low
throughput experimental technologies available at the time.
SA-based models flourished in toxicity prediction in almost all
types of toxic endpoints [10, 11]. Several expert systems are
available for toxicity prediction based on prebuilt rules and SAs,
e.g., HazardExpert, Oncologic Cancer Expert System (OCES),
Toxtree, et al. [12–14]. These models are limited to producing
qualitative binary output, i.e., toxic or nontoxic. Chemical simi-
larity cluster methods take into account the structural similarity
of chemicals, physiochemical features, ADME and mechanisms
of action (MoA), which in turn can provide qualitative or quan-
titative predictions depending on the toxicity endpoint [15].
Multiple tools implement this approach, such as AMBIT,
DSSTox, and Toxmatch, with applications including prediction
of environmental risk, reproductive toxicity, skin sensitization
and so on [16–18]. The statistically derived rule-based
approaches mentioned above share a common limitation,
namely, lack of biological insights into the mechanistic basis of
toxicity. Analogous to pharmacokinetics/pharmacodynamics
features indicating the mutual interaction of recipient and
chemicals, toxicokinetics/toxicodynamics analysis selects the
toxic response related to the chemical concentration in vivo.
Importantly, measurement of the internal doses rather than
administered doses and key metabolites provide a more accurate

relationship to the response. In addition, it is a well-developed
practice to extrapolate between various administration routes,
as different species use nonidentical PK/PD and
ADME. However, the toxicity pathway and the MoA can only
be defined with expert knowledge [19–21]. Drug toxicity is a
complex response occurring at system, tissue, cellular, and
molecular levels. Classic toxicity testing and prediction meth-
ods, using either animals or in silico chemicals, similarity based
or PK/PD based models, simplified complexity and left the
mechanistic understanding of the chemical- induced toxicity
pathways out of consideration. In 2007, the NRC released the
report Toxicity testing in the 21st century: A Vision and a Strategy,
in which it addressed future directions that would take com-
plexity and toxicity pathways into account [22].
2 From Systems Biology to Systems Toxicology
The revolution in biomedical science in the post genome era has

made it feasible to study the effects of chemicals using cells, cel-
lular components, and tissues, preferably of human origin.
High-throughput assay technologies, bioinformatics and sys-
tems biology have significantly empowered scientists to deci-
pher how molecular components, different cells or tissues
cooperate to carry out normal physiological functions that are
key to maintaining health [23, 24]. Three high-throughput
assays developed over the recent decades have provided major
impetus to the field of toxicology: omics technologies, image
techniques, and automated robotic platform techniques. The
platforms enable testing of huge numbers of chemicals in a
high-throughput number of samples under standardized condi-
tions. Omics technologies collect the molecular responses to a
substance while image methods decode the phenotypical and
functional change of cells, organs, or organisms in response to
exposure to a compound. Together, these three technologies
allow researchers to characterize toxicity rapidly at affordable
cost [25–27]. As an interdisciplinary field of science, bioinfor-
matics combines computer science, statistics, mathematics, and
engineering to analyze and interpret biological data, and serves
as a key tool with which to decode the enormous quantum of
data generated with high-throughput assays [28]. Since 2000,
Systems Biology had been used widely to “understand biology
at the system level” using computational and mathematical
modeling of complex biological systems [29]. The emergence
of systems toxicology can be characterized as the integration of
classical toxicology with the quantitative analysis of large net-
works of molecular and functional changes occurring across
multiple levels of biological organization. This is in essence a

holistic approach to deciphering the impact of environmental
agents (chemicals, complex mixtures, occupational exposures,
physical agents, biological agents, and lifestyle factors) on com-
plex biological systems using an engineering approach applied
to toxicological research [30]. Systems toxicology is rooted in
the ongoing revolution in biology and biotechnology, and is
founded on the premise that morphological and functional
changes in cellular, tissue, organ, and organism levels are caused
by and cause changes at the omics level. One example is the
Human Toxome project launched by NIH/DDD that is
intended to test the strategies that combine omics data and
computational models, aiming to develop a common, commu-
nity accessible framework [31]. Another is Tox-21c, which
focuses on toxicity pathways, mechanisms/modes of action, and
adverse outcome pathways (AOP) in humans. Tox-21c largely
overlaps with 3Rs (replace, reduce, and refine) proposed half a
century ago [32, 33]. The Systems Toxicology computational
challenge, sbv IMPROVER computational challenge, used
crowd resourcing to demonstrate that gene expression data
from blood cells are sufficiently informative to predict response
to smoking in humans and across species translation [34].
Yet, a comprehensive understanding of the mechanisms of
drug toxicity in specific cases requires the integration of different
data modalities, from changes at the genomic, proteomic, and
metabolomics level across several scales of cellular organization.
In contrast to classical approaches, systems toxicology resides at
the intersection of systems biology and toxicology where chemis-
try incorporates mechanisms into the predictive framework [35].
To understand how this complex interaction system in cells and
tissues leads to toxicity requires the integration of two disciplines
that have been increasingly useful in biomedical research: “Systems
Biology” and “Quantitative Pharmacology.” In systems biology, a
system is generally described as a set of nodes (vertices) connected
by edges describing functional interactions. These edges can rep-
resent physical interactions, functional interactions, and connec-
tions between data across several scales. Similarly, in systems
toxicology biological networks are the basis for the prediction of
drug action in complex biological systems [36].
Systems toxicology models contain expressions that charac-
terize functional interactions within a biological network, which
are very useful when drugs act at multiple targets in the network
or when homeostatic feedback mechanisms are operative [37].
Therefore, these models are particularly useful in describing
complex patterns of drug action such as synergies between dif-
ferent drugs. Although systems toxicology is still in its infancy, it
has tremendous potential to change the way we approach bio-
medical research. It represents a movement beyond a traditional
study-centric approach toward a continuous quantitative inte-

gration of data across studies and the different phases of drug
development. Network-based approaches offer a wide range of
possibilities for deciphering and possibly for understanding the
complexity of human disease, thereby providing new tools with
which to develop novel drugs. Here we review some current
efforts and recent methods through the lens of quantitative sys-
tems pharmacology (QSP).
3 Examples of Predictive Systems Toxicology
The general notion of a network-based approach rests upon the

ambition to connect several entities across the molecular, cellular
pathways, organs, and systems to facilitate the prediction of the
effect of a drug candidate or any kind of perturbation on biological
outcomes of interest [38, 39]. The way in which one defines or
infers a network from data is the main determining factor of the
degree of reliability and applicability of network analysis in drug
design. It is crucial to have a clear definition of network nodes early
on, edges and edge weights in the specific application case, and in
that context to consider data quality and refinements of the data
based on genetic variability, aging, and environmental effects.
Different types of networks such as networks of chemical com-
pounds, signaling networks, gene–gene interaction networks, pro-
tein–protein interaction (PPI) networks or metabolic networks
and disease networks can be (and have been) used in QSP models
and methods [40]. Following the work on inferring a network
comes the analysis of the network and its properties. In the last
step, the result of analysis needs to be converted to a series of
actionable hypotheses, which then need to be tested and validated
or refuted (see Fig. 1).
Drug–target interaction is the first and most common type of
network analysis that has been used in QSP models. Interactions
Public data bases Network structure Drug target

Expanding the analysis identification
network by
combining different
assembly
Network data
Network Drug synergy

Analyzing network
Translation
&Validation
data bases comparison

Predicting new
edges using silico Drug repositioning
methods Network dynamics
Drug efficacy
ADMET side-effects
Fig. 1 Overview of predictive system toxicology approaches

between drugs and targets can facilitate the process of drug discov-
ery by deciphering a drug’s mechanism of action, thereby assisting
researchers seeking new targets for an old (FDA approved) drug as
well as new drug candidates for a known target [41–45]. The main
source of information in reconstruction of the Drug–Target inter-
action network (DTN) is the Drug Bank, which is one of the major
publicly available integrated sources of drugs and targets. It is a
highly comprehensive database combining chemical properties and
detailed clinical information about drugs and their targets. It also
provides drug-related data feeds for well-known databases such as
Uniprot, PubChem, PDB, and KEGG [46, 47].
In spite of the fact that mining drug–target interaction data is
increasing at an amazing rate [42], drug–target interaction data
currently available from public sources are largely incomplete and
biased toward targets of common therapeutic interest [48–50].
Biochemical experiments or in vitro methods for finding drug–tar-
get interaction are costly and time-consuming. An attempt to
address the issue of data completeness of drug–target interaction
involves using in silico methods [51]. For example, docking simu-
lations are extensively used in pharmacology. AutoDock [52] is
one of the most complete suites of free open-source software for
the computational docking and virtual screening of small mole-
cules to macromolecular receptors. Xie et al. identified drug off-
targets by docking the drug into protein binding pockets similar to
those of its primary target, followed by mapping the proteins with
the best docking scores to known biological pathways, thus pre-
dicting potential side effects [53].Classically, the process starts
with a target of known three-dimensional structure, and docking is
used to predict the bound conformation and binding energy. In
most cases, the three-dimensional structure of a target is needed to
compute the binding of each drug candidate to the target, which
for many targets are still unavailable [54–56]. Wallach et al. have
developed a method to mitigate the impact of this important limi-
tation. They utilize a dataset where there is a pairing of drugs with
their observed adverse drug reactions (ADRs), the protein struc-
ture database and in silico virtual docking to identify putative pro-
tein targets for each drug and search for correlated pairs of side
effects and biological pathways [57]. Another challenge when per-
forming docking simulation is that it is computationally expensive
and most of the methods must simplify the problem to make the
computation feasible. The reduction of conformational space by
imposing limitations on the system, such as fixed bond angles and
lengths in the ligand or a simplified scoring function such as those
based on empirical free energies of binding to score poses quickly
at each step of the conformation search, are the most common
short-cuts that are currently used in the field [52, 58].
In a more recent effort, machine-learning approaches have
been used for larger-scale predictions of drug–target interactions.
The new interactions between drugs and targets can lead to poten-
tial insights on previously unidentified side effects for a particular
drug. This idea is the basis of another category of systems toxicol-
ogy methods. Machine learning-based methods mostly use struc-
tural and chemical descriptors of drugs and sequences of targets,
similarity matrix or (and) any other pharmacological information
about drugs as input. Then they use any machine learning method,
such as support vector machines (SVMs) or kernel regression, for
predicting the drug–target interactions [59–63]. Cobanoglu et al.
used the known interactions in the Drug Bank in the form of a
bipartite network to train a model that represents each drug and
target as a vector of latent variables and assigns weights to drug–
target interactions using probabilistic matrix factorization [64].
Approaches that use similarity scores as input are more promising
than other approaches [41].
In general, the use of machine-learning algorithms is one of
most promising approaches to extracting knowledge from big data
using a data-driven framework. However, the performance of
machine-learning algorithms relies heavily on data representations
called features, and identifying which features are more appropri-
ate for the given task is very difficult. Deep Learning has recently
emerged as a promising technique where the features do not need
to be hand-crafted a priori. Recent success has been accomplished
thanks to the availability of fast computations, massive (labeled)
datasets and sophisticated algorithms [65]. Machine learning using
deep learning is defined by neural networks with multiple hidden
layers. Each layer basically constructs a feature from the preceding
layers [66]. The training process allows layers deeper in the net-
work to contribute to the refinement of earlier layers. For this rea-
son, these algorithms can automatically engineer or discover
features that are suitable for representing the data at hand. When
sufficient data are available, these methods construct features
attuned to a specific problem and combine those features into a
predictor [67]. Deep learning algorithms have shown promise in
fields as diverse as high-energy physics [68], dermatology [69],
and translation [70]. DEEPtox is one of the first methods using
Deep Learning for computational toxicity prediction [65].
DeepTox normalizes the chemical representations of the com-
pounds and computes a large number of chemical descriptors that
are used as input in machine learning methods. As a next step,
DeepTox trains several models, evaluates them, and combines the
best of them into ensembles. Finally, DeepTox predicts the toxicity
of new compounds. In DEEPTox SVMs, random forests, and elas-
tic nets are used for cross-checking, supplementing the Deep
Learning models, and for ensemble learning to complement Deep
Neural Networks (DNNs). The networks consist of multiple layers
of rectified linear units (ReLUs) to enforce sparse representations
and counteract the appearance of a vanishing gradient. ReLUs are
followed by a final layer of sigmoid output units, one for each task.
One output unit is used for single-task learning. Stochastic gradi-
ent descent learning has been used to train the DNNs, and both
dropout and L2 weight decay were implemented for the DNNs in
the DeepTox pipeline for regularizing learning and avoiding over-
fitting. Of note is the fact that DEEPtox outperformed many other
computational approaches like naive Bayes, support vector
machines, and random forests in toxicity prediction of 12,000
environmental chemicals.
The output of all the abovementioned methods is a DTN, an
undirected bipartite network composed of two sets of nodes,
drugs and targets. DTN have a complex topology that reflects the
inherently rich polypharmacology of drugs (also known as drug
repurposing) [51]. The analysis of DTN has recently emerged as
an effective means to study targets and to identify new targets for
known drugs. In one of the very first attempts, Ma’ayan et al. [71]
reconstructed such a bipartite network, and the nodes have been
connected if there is an association between a drug and a target on
the basis of data from the Drug Bank. They report several classes
of proteins as better targets for drugs based on network statistics
and gene ontology. A decade later, Lin et al. [72] have followed
the same approach to studying the drug–target interaction and
could characterize the drug–target relations of different kinds of
drugs. They showed that the number of multitarget new molecu-
lar entities (NME) has increased over the years, but less than
single-target NMEs. In both these cases and several other cases in
the literature, it has proven useful to analyze the general structure
of a network in order to extract new knowledge facilitating the
classification of drugs and/or their targets. Structural (graphical)
analysis of a network provides insights into the organization and
topology of the DTN and targets for hypothesis generation and
experimental testing. As a rule this is performed through compu-
tation and analysis of network parameters–parameters that quan-
tify different aspects of the network’s internal structure, such as
parameters measuring centrality, a node or more global parame-
ters such as modularity index, network density, network entropy
or network diameter [36]. Several methods have been developed
and applied based on network topology, graph theory, and cluster
analysis (see [8] for a recent review). Methods based on the simi-
larity of networks are another set of techniques that have been
used to uncover novel target or disease-specific changes [73, 74].
A wide range of similarity measures have been used in the litera-
ture, ranging from intuitive measures such as the number of edge
changes required to get one network from another or the com-
parison of the top-k nodes to the more complicated ones, such as
using an ensemble of different model networks, and the distribu-
tion of the best-fitting ensemble. However, it should be kept in
mind that the fundamental question of checking whether two
given networks have the same structure, network comparison, is

computationally expensive, and despite extensive progress in the
field, it remains one of the greatest challenges in the field. For
example, it is still not known whether graph isomorphism is poly-
nomial solvable or whether it is NP-complete. Therefore most of
the current methods in the network comparison field are heuristic,
which in turn may affect the outcome strongly, depending on
which kind of prior biases exist in the particular method.
All interactions, from protein–protein interactions (PPI) to
gene expression and pathways, are useful in the quest to under-
stand the mechanism(s) of interaction between drugs and complex
diseases. Remez et al. used predicted drug−protein interactions
obtained with a CT-link in combination gene expression data to
obtain a projected anatomical profile of a drug and use it for con-
necting in vitro assays with in vivo outcomes and predict potential
in in vivo organ toxicities [75, 76].
Kuhn et al. used a network based on drug–target interaction
data and drug–ADR interaction data to systematically predict and
characterize proteins that cause drug side effects. They integrated
phenotypic data obtained during clinical trials with known drug–
target relations to identify overrepresented protein–side effect
combinations [77]. Their networks have three types of nodes:
drugs, targets, and side effects, and links are identified side effect
causality predictors. The authors considered overrepresented pro-
tein–side effect pairs, and hypothesized that such overrepresenta-
tion could be indicative of causality. Their approach can make
predictions for proteins that are the targets of a certain number of
drugs. In this context, Yildirim et al. used a bipartite graph com-
posed of FDA-approved drugs and target proteins in the context of
cellular and disease networks and quantitatively demonstrated an
overabundance of “follow-on” drugs [78]. The authors overlaid
the drug–protein network with a network of physical PPI. They
demonstrated a significant increase in the number of interacting
proteins as compared to the average in the PPI network. They used
the distance between drugs and a drug target and the correspond-
ing disease to show that most drug targets are not closer to the
disease genes in the protein interaction network than a randomly
selected group of proteins.
Similarly, several other approaches have been developed based
on the notion of expanded drug–target interactions, combined
with protein–protein interactions data, in order to develop a
network-based pharmacology that could better explain the drug–
phenotype relationship, and this approach has been used to predict
novel targets and drug repositioning [79–84]. For example, Guney
et al. in [85] integrated protein–protein interaction, drug–disease
association, and drug–target association data. They analyzed the
topological characteristics of drug targets with respect to disease
proteins and showed that for a drug to be effective against a dis-
ease, it had to target proteins within or in the immediate vicinity of

the corresponding disease module. Such approaches were also
considered for issues related to drug safety and side effects. Cami
et al. constructed a network representation of drug–ADR associa-
tions for approximately 800 drugs and ADRs and pharmacological
information for toxicity prediction. They exploited network struc-
ture to predict likely unknown adverse events using a trained logis-
tic regression model [86]. Berger et al. used PPI networks to
predict and identify drugs that likely cause Long QT Syndrome
based on both a direct drug–target interaction and separate neigh-
borhood [87].
Complementary to protein–protein interactions, transcriptomic
data, and gene expression differentiation have been used in drug
discovery and safety [87–92]. For example, Gottlieb et al. intro-
duced a method for inferring drug-specific pathways [88]. They
connect known drug-associated genes over protein, metabolic, and
transcriptional interaction networks while preferring high confi-
dence interactions participating in curated cellular processes. They
use their computed pathways to suggest novel drug repositioning
opportunities, gene–side effect associations, and gene–drug inter-
actions. Huang et al. developed a new metric to measure the
strength of network connection between drug targets to predict the
pharmacodynamics of drug–drug interactions [91, 92].
For the purpose of predicting drug toxicity, in most cases we
require a collection of experimental data reflecting molecular
changes in the context of quantifiable cellular changes across dif-
ferent biological scales that are linked to toxicity at the body
level [35]. So in addition to all the abovementioned data, sys-
tems toxicology depends strongly on the quality and scope of
databases annotating side effects (SIDER) and drug-induced
differential gene expression, or a combination thereof [93–96].
As an example, Lounkine et al. developed an association metric
asking how to prioritize those new off-targets that explained side
effects better than any known target of a given drug, thereby creat-
ing a drug–target–adverse drug reaction network [43]. Network-
based approaches allow the generation of hypotheses about
drug–target–phenotype–side effect associations but currently avail-
able interaction data are incomplete and the available parts are
often nonhomogeneous and biased. This situation results in the
fact that the conclusions of such studies strongly depend not only
on the quality, but importantly, on the degree of completeness of
the data [97]. The other relevant point is that most of the sug-
gested approaches in QSP are largely based upon the analysis of the
structure of a network or on comparison of networks, while it has
been shown [98] that network dynamics, the study of temporal
changes in network structures or describing changes of phenotypes
of a complex system in the state-space, is crucial to understanding
the complexity of diseases and the action of drugs [39]. In this
context Mucha et al. [99] developed the technique of multilayer

networks, incorporating different types of nodes and edges, in
order to follow the changes in module structure in a system having
multiple and different types of edges. Interestingly these methods
have also been used to predict drug synergies. However, most of
them are limited to estimating target links on the PPI network.
The advantage of using a network-based approach lies in that it
helps to explain the hidden molecular mechanism of drug synergy
from the interactions. Due to their effectiveness, some approaches
aiming at identifying synergistic drug combinations are based on
the dynamic simulation of specific subnetworks. However, these
models relied on a very detailed dynamical model, where the lack
of information and the uncertainties involved in their kinetics
parameters and lots of artificial constraints often limit the useful-
ness of the simulations, resulting in the model working only for a
few specific pathways. Kiani et al. developed a novel integrative
pipeline for systematic exploration of drug combinations as a com-
prehensive and flexible network-based model in the context of the
DREAM challenge, a pipeline called HotPPI. Here they con-
structed a human protein interaction network from major PPI
resources, and included both experimentally validated and compu-
tationally predicted interactions. The overall procedure resulted in
a vast protein interaction network comprising 15,383 proteins and
337,413 interactions. Next, PPI was filtered based on targets of
the DREAM challenge and the top 50 pathways involving these
targets (Table 1).
The filtered PPI network comprises 6000 proteins and 16,000
interactions. Molecular data are used to weigh interaction in our
PPI. The main goal of HOTPPI is to find the best combination to
eliminate cancer cell lines. Therefore any combination that elimi-
nates most interactions in a network can cause network collapse
followed by death of cancer cells. Thus the heat diffusion algo-
rithm is used to predict potential synergistic drug combinations by
calculating how efficient drugs are in hitting the top 200 selected
nodes in a network based on their betweenness score. The Hot PPI
is generally applicable to high-throughput experimental data where
the challenge is to select a small number of the most promising
combinations for further mechanistic studies. Using this score, we
could rank all possible combinations in a reasonable amount of
time. Interestingly, we learned that we should not include too
many details (i.e., features or molecular components) in our net-
work descriptions, since we may shift our description from optimal
toward the “knowledge of everything,” with the precision of the
method dropping drastically as a result. This underscores the
importance and challenge of pruning a large, but for the given
application reasonable number of features to include in the net-
work model (see Fig. 2).
Table 1
Top 50 pathways
Pathway maps p-Value

Ligand-independent activation of Androgen receptor in Prostate Cancer 1203E−15
Cell adhesion_PLAU signaling 4362E−14
Development_EGFR signaling pathway 8380E−14
Apoptosis and survival_Anti-apoptotic action of Gastrin 1451E−13
K-RAS signaling in pancreatic cancer 1920E−13
Development_G-CSF signaling 7018E−13
Development_Growth factors in regulation of oligodendrocyte precursor cell 1065E−12
proliferation
Main growth factor signaling cascades in multiple myeloma cells 3352E−12
Development_VEGF signaling and activation 5644E−12
Main pathways of Schwann cells transformation in neurofibromatosis type 1 1107E−11
Immune response_IL-5 signaling 1172E−11
Signal transduction_PTEN pathway 1172E−11
Immune response_IL-15 signaling 1331E−11
Ovarian cancer (main signaling cascades) 1595E−11
Tissue Factor signaling in cancer via PAR1 and PAR2 2309E−11
Development_EPO-induced Jak-STAT pathway 2549E−11
Apoptosis and survival_HTR1A signaling 2864E−11
Development_GM-CSF signaling 2864E−11
Cytoskeleton remodeling_TGF, WNT and cytoskeletal remodeling 3189E−11
HBV signaling via protein kinases leading to HCC 3373E−11
Development_Delta-type opioid receptor mediated cardioprotection 4422E−11
Apoptosis and survival_Anti-apoptotic action of membrane-bound ESR1 5750E−11
Translation_Non-genomic (rapid) action of Androgen Receptor 9496E−11
Apoptosis and survival_BAD phosphorylation 1525E−10
Development_Growth hormone signaling via PI3K/AKT and MAPK cascades 1525E−10
Development_Ligand-independent activation of ESR1 and ESR2 2960E−10
Development_Membrane-bound ESR1: interaction with growth factors signaling 2960E−10
Development_VEGF signaling via VEGFR2 - generic cascades 3413E−10
Role of Tissue factor-induced Thrombin signaling in carcinogenesis 5194E−10
Development_CNTF receptor signaling 7526E−10
(continued)
Table 1
(continued)
Pathway maps p-Value

IL-6 signaling in multiple myeloma 9699E−10
Some pathways of EMT in cancer cells 9699E−10
Development_IGF-1 receptor signaling 1164E−09
Development_FGF-family signaling 1164E−09
Signal transduction_Additional pathways of NF-kB activation (in the cytoplasm) 1391E−09
Development_Growth factors in regulation of oligodendrocyte precursor cell survival 1563E−09
Development_Dopamine D2 receptor transactivation of EGFR 1746E−09
Aberrant B-Raf signaling in melanoma progression 1965E−09
Immune response_TSLP signaling 2453E−09
Development_Prolactin receptor signaling 3215E−09
Development_VEGF-family signaling 3751E−09
Immune response_IL-7 signaling in B lymphocytes 5605E−09
Signal transduction_AKT signaling 5605E−09
Regulation of Tissue factor signaling in cancer 5605E−09
Immune response_IL-4 signaling pathway 6797E−09
Influence of smoking on activation of EGFR signaling in lung cancer cells 6797E−09
Immune response_TNF-R2 signaling pathways 8203E−09
Development_Role of IL-8 in angiogenesis 9139E−09
Immune response_IL-4 - antiapoptotic action 9803E−09
4 Information Theoretic Approach to Toxicity
Both network analysis and pharmacokinetic analysis share a focus

and grounding in the physical and functional interactions between
molecules within the cell or tissue and the corresponding drugs.
Here the overarching aim is not only to predict but also to be able
to interpret the mechanism in terms of the underlying biology and
chemistry. Since the design of new drugs for new targets is diffi-
cult, and the prediction problem is easier from an inference point
of view, compared to elucidating the mechanisms driving toxicity,
Dream
Molecular
STRING
Filtering and
HPRD PPI-customized Heat diffusion
wieghting PPI
HI-II
Diffusion score
Dream Drug MetaCore TM Protein in relevent

Target Pathways
Normalizing and
Calculating Pharmacoligcal
Dream weighted
Pharmacological AUC (IC25-75) Score multiplication
Final Synergy Score
Fig. 2 Overview of HotPPI approach
complementary approaches are warranted. For example, instead of

engineering a drug to target the unique pathways or mutations of
a tiny subset of diseases, drug repositioning, such as the one exem-
plified in the DREAM challenge, involves starting with approved
drugs to find combinations that can be used to treat diseases differ-
ent from the ones they have been designed for, with the advantage
that approved drugs can bypass much regulation if correctly con-
trolling for the effects they can have. Thus prediction and simula-
tion are key. This means that the whole field has to move toward
causal modeling and functional inference rather than traditional
statistical classification (e.g., Tanimoto coefficients) or computa-
tional simulation based on classical geometric approaches (e.g.,
distance between molecules, grid-based docking). To this end,
information indexes can facilitate the characterization of drugs by
the combinatorial and structural properties shared with or at a
remove from the structural properties of the targets, because just

as for any molecule, structure means function. Then all these
approaches can contribute to determining drug function based on
the fact that structurally similar molecules usually have similar
properties (known as “neighborhood behavior”). For example,
statins are associated with the heart and cholesterol, while mor-
phine, codeine and heroin share structural properties and effects.
However, algorithmic information-theoretic approaches based on
both classical information and computability theory introduce pre-
dictive causal models that go beyond statistical similarities and can
find, in principle, similar mechanisms shared by sets of drugs with
respect to targets and functions.
It is not difficult to see that complementary regions between
drug and target will have a similar classical and algorithmic infor-
mation content, because the structure of one is the complement of
the other. Another advantage is that these measures are parameter-
free and thus require no training, even though they can comple-
ment and guide machine learning approaches [100, 101]. Because
drug docking is not invariant to, e.g., scaling factors, but informa-
tion theoretic measures are, they may fail to characterize the posi-
tive or negative docking properties of a drug. While coarse-graining
techniques may be introduced, algorithmic complexity has the
advantage of being able to account for scaling effects. The basic
idea is the likelihood of a drug being causally generated by a
mechanistic model (an algorithm). This is, in general, hard if not
impossible to find (the problem is uncomputable), but approxima-
tions are possible and new numerical methods have been advanced
complementary to statistical and lossless compression approaches
that cannot or are very limited at accounting for causation. Drugs,
and molecules in general, can be represented in many ways (see
Fig. 3a–c), some of which are natural networks or networks repre-
senting properties of the molecules. Most of these representations
are lossless representations, meaning that they can reconstruct the
primary representation of the molecule that they encode, e.g., the
simplified molecular-input line-entry system or SMILES. The
SMILES of a molecule is a string obtained by printing the symbol
nodes encountered in a depth-first tree traversal of a chemical
graph. SMILES can be converted back (almost) uniquely to the
two-dimensional representation of a drug.
Figure 3 shows some of these network (a, b) and two-
dimensional representations (c), together with two figures (d, e)
plotting three information-theoretic indexes, two classical and one
algorithmic based on the drugs’ contact networks. While the two
classical indexes are the ones most correlated, as one is extracted
from the other with the additional information of the sequence
valence, the length of the algorithmic complexity (Z-axis) repre-
sents the complexity of a hypothesized model producing the con-
tact network.
Fig. 3 Drug profiling by (algorithmic) information indexes: (a) The molecular (chemical) graph of Atorvastatin
(C33H35FN2O5), a member of the drug class known as statins used primarily as a lipid-lowering agent for pre-
vention associated with treatment of cardiovascular diseases. (b) In a molecular network geographical coordi-
nates and shapes are no longer important, but rather their topology (which element is connected to which
other), which can be built upon (c) the molecular contact map where grey scale (left matrix) indicates proximity
between each element that can be binarized (right) using a cutoff value based on the grey scale median. (d)
Algorithmic information landscape of more than 4000 drugs from the DrugBank (extracted from the Wolfram
Language) constructed by taking the entropy of their SMILES codes, the valence sequences of each of the
elements in their formula (from SMILES), given the importance they have for bonding (e) Algorithmic informa-
tion landscape of the drugs involved in the DREAM challenge. Color is determined by the “contact map com-
plexity”; the less complex (the shorter the length of the algorithm generating it) the closer to blue, the longer
(more algorithmic-random) the closer to red
5 Concluding Remarks
Here we have reviewed different attempts to predict toxicity from

observations (i.e., the Sumerian) to classical pharmacokinetic,
advancing to recent integrative systems oriented approaches taking
more data into account. These systems approaches resort to per-
forming advanced statistical analytical data processing comple-
mented with machine learning techniques to generate paradigms
attempting not only to predict toxicity but also to identify (molec-
ular) mechanisms of toxicity. Information theoretic approaches can
be situated in between, as they are as a rule less dependent upon
biochemical representations in their problem formulation, while
the ones presented here also aim for causal understanding of toxic-
ity in addition to targeting prediction.
In a broader perspective, there are several immediate chal-
lenges where we need more work. These include which features to
include when predicting toxicity? Minimal models may suffer from

being less understandable from a mechanistic standpoint, whereas
including too many features, as in the dream example above, could
hamper the prediction capability of the model. Overall, a systems
biology approach extends the feature space compared to classical
pharmacokinetics, while an (algorithmic) information approach
facilitates predictions in combination, being both scale invariant
and parameter free. Hence there is a tension between predicitive
capacity and mechanistic interpretability.
Furthermore, overtraining and overfitting in solving high-
dimensional and complex nonlinear problems such as toxicity pre-
diction is one of the most common problems of existing machine
learning methods. This originates from the need for estimating
and optimizing numerous hyper parameters. However, a method
such as the relevance vector machine method solves this problem
by incorporating Bayesian criteria into the learning process to
reduce the irrelevant support vectors of the decision boundary in
feature space, thus resulting in a sparser model [102]. Methods
such as Random Forest classifiers are another category of success-
ful methods in systems toxicology. They are one of the most
robust algorithms and are able to identify the patterns important
for the preferred class, even when there is a large imbalance in the
class distribution within the training dataset [103]. Inspecting the
results of the TOX21 data challenge demonstrates that a hybrid
strategy which combines similarity scores for structural finger-
prints and molecular descriptors (features) and machine-learning
based prediction models can readily improve the accuracies of tox-
icity prediction [104]. In general, an ensemble model can be
effective, since taking into account the prediction of other models
can compensate for an incorrect prediction on the part of one of
the individual methods. Certainly, each of the systems toxicology
methods has intrinsic advantages, limitations, and practical con-
straints. Moreover, the performance of these methods depends on
the structural diversity and representativeness of the molecules in
the data set. Therefore, it is quite important to choose the most
suitable machine learning method to develop the prediction
model for a specific toxicity data set. Finally, the computational
cost associated with each method is another practical and impor-
tant factor determining the usability of a given method.
In conclusion, beyond the above challenges and consider-
ations, the grand remaining challenge is to advance the state-of-
the-art toward personalized medicine. This requires patient
specific predictions on toxicity, which in turn requires proper
stratification of patients with regard to how they respond or not,
with or without adverse toxic effects. This most likely requires
integration of multiple layers of information as a background
upon which an individual has to be characterized/described, while
a machinery for toxicity prediction has to be specific enough for a
given patient, given the amount of (sparse) patient-specific infor-

mation. This challenge and perspective will keep the field of data-
driven computational toxicology busy.
Acknowledgments
We thank the contributors to HotPPI, Dr. Gordon Ball, Dr. David

Gomez, Alireza Mazaheri, and Dr. Francesco Marabita. This study
was supported by the Swedish Research Council (J.T. and H. Z.)
and Karolinska Institutet funds (N.K.). The funders had no role in
study design, data collection and analysis, decision to publish, or
preparation of the manuscript.
References
1. Petrovska BB (2012) Historical review of scheme for generating chemical categories
medicinal plants’ usage. Pharmacogn Rev for read-across, structural alerts and insights
6:1–5 into mechanism(s) of action. Crit Rev Toxicol
2. Hunter P (2008) A toxic brew we cannot live 43:537–558
without. Micronutrients give insights into the 11. Gerner I, Barratt MD, Zinke S, Schlegel K,
interplay between geochemistry and evolu- Schlede E (2004) Development and prevali-
tionary biology. EMBO Rep 9:15–18 dation of a list of structure-activity relation-
3. Bottini AA, Amcoff P, Hartung T (2007) ship rules to be used in expert systems for
Food for thought … on globalisation of alter- prediction of the skin-sensitising properties
native methods. ALTEX 24:255–269 of chemicals. Altern Lab Anim ATLA
4. Adeleye Y, Andersen M, Clewell R et al 32:487–509
(2015) Implementing Toxicity Testing in the 12. Milan C, Schifanella O, Roncaglioni A,
21st Century (TT21C): making safety deci- Benfenati E (2011) Comparison and possible
sions using toxicity pathways, and progress use of in silico tools for carcinogenicity within
in a prototype risk assessment. Toxicology REACH legislation. J Environ Sci Health
332:102–111 Part C 29:300–323
5. Pease W (1997) Toxic ignorance: the con- 13. Ellison CM, Enoch SJ, Cronin MTD (2011)
tinuing absence of basic health testing for A review of the use of in silico methods to
top-selling chemicals in the United States. predict the chemistry of molecular initiating
Diane Pub. Co., Darby events related to drug toxicity. Expert Opin
6. Mak IW, Evaniew N, Ghert M (2014) Lost in Drug Metab Toxicol 7:1481–1495
translation: animal models and clinical trials in 14. Bhatia S, Schultz T, Roberts D, Shen J,
cancer treatment. Am J Transl Res 6:114–118 Kromidas L, Marie Api A (2015) Comparison
7. Hansch C, Maloney PP, Fujita T, Muir RM of Cramer classification between Toxtree, the
(1962) Correlation of biological activity OECD QSAR Toolbox and expert judgment.
of phenoxyacetic acids with Hammett sub- Regul Toxicol Pharmacol 71:52–62
stituent constants and partition coefficients. 15. Hardy B, Douglas N, Helma C et al (2010)
Nature 194:178–180 Collaborative development of predictive toxi-
8. Raies AB, Bajic VB (2016) In silico toxicol- cology applications. J Cheminform 2:7
ogy: computational methods for the predic- 16. Gallegos-Saliner A, Poater A, Jeliazkova N,
tion of chemical toxicity. Wiley Interdiscip Patlewicz G, Worth AP (2008) Toxmatch—a
Rev Comput Mol Sci 6:147–172 chemical classification and activity predic-
9. Lepoittevin J-P, Benezra C (1991) Allergic tion tool based on similarity measures. Regul
contact dermatitis caused by naturally occur- Toxicol Pharmacol 52:77–84
ring quinones. Pharm Weekbl Sci 13:119–122 17. Williams-DeVane CR, Wolf MA, Richard
10. Hewitt M, Enoch SJ, Madden JC, Przybylak AM (2009) DSSTox chemical-index files for
KR, Cronin MTD (2013) Hepatotoxicity: a exposure-related experiments in ArrayExpress
and Gene Expression Omnibus: enabling
toxico-chemogenomics data linkages. 32. Flecknell P (2002) Replacement, reduction

Bioinformatics 25:692–694 and refinement. ALTEX 19:73–78
18. Jeliazkova N, Jeliazkov V (2011) AMBIT 33. Hartung T (2010) Lessons learned from
RESTful web services: an implementation of alternative methods and their validation for a
the OpenTox application programming inter- new toxicology in the 21st century. J Toxicol
face. J Cheminform 3:18 Environ Health Part B 13:277–290
19. Ait-Oudhia S, Zhang W, Mager DE (2017) 34. Guryanova S, Guryanova A (2017) sbv
A mechanism-based PK/PD model for hema- IMPROVER: modern approach to systems
tological toxicities induced by antibody-drug biology. Methods Mol Biol 1613:21–29
conjugates. AAPS J 19:1436–1448 35. Kiani NA, Shang M-M, Tegner J (2016)
20. Caldwell GW, Yan Z, Tang W, Dasgupta Systems toxicology: systematic approach
M, Hasting B (2009) ADME optimization to predict toxicity. Curr Pharm Des
and toxicity assessment in early- and late- 22:6911–6917
phase drug discovery. Curr Top Med Chem 36. Kiani NA, Zenil H, Olczak J, Tegnér J (2016)
9:965–980 Evaluating network inference methods in
21. Mager DE, Wyska E, Jusko WJ (2003) terms of their ability to preserve the topology
Diversity of mechanism-based pharmacody- and complexity of genetic networks. Semin
namic models. Drug Metab Dispos Biol Fate Cell Dev Biol 51:44–52
Chem 31:510–518 37. Danhof M (2016) Systems pharmacology –
22. Krewski D, Acosta D, Andersen M et al towards the modeling of network interac-
(2010) Toxicity testing in the 21st century: tions. Eur J Pharm Sci 94:4–14
a vision and a strategy. J Toxicol Environ 38. Arrell DK, Terzic A (2010) Network systems
Health Part B 13:51–138 biology for drug discovery. Clin Pharmacol
23. Tegnér JN, Compte A, Auffray C et al (2009) Ther 88:120–125
Computational disease modeling – fact or fic- 39. Csermely P, Korcsmáros T, Kiss HJM,
tion? BMC Syst Biol 3:56 London G, Nussinov R (2013) Structure
24. Tegnér J, Zenil H, Kiani NA, Ball G, and dynamics of molecular networks: a novel
Gomez-Cabrero D (2016) A perspective on paradigm of drug discovery: A comprehensive
bridging scales and design of models using review. Pharmacol Ther 138:333–408
low-dimensional manifolds and data-driven 40. Knight-Schrijver VR, Chelliah V, Cucurull-
model inference. Philos Transact A Math Phys Sanchez L, Le Novère N (2016) The promises
Eng Sci 374:20160144 of quantitative systems pharmacology model-
25. Blankenburg M, Haberland L, Elvers H-D, ling for drug development. Comput Struct
Tannert C, Jandrig B (2009) High-throughput Biotechnol J 14:363–370
omics technologies: potential tools for the 41. Ding H, Takigawa I, Mamitsuka H, Zhu S
investigation of influences of EMF on biologi- (2014) Similarity-based machine learning
cal systems. Curr Genomics 10:86–92 methods for predicting drug–target inter-
26. Lam F, Ma C, Clifford B, Johnson CL, Liang Z-P actions: a brief review. Brief Bioinform
(2016) High-resolution 1 H-MRSI of the brain 15:734–747
using SPICE: data acquisition and image recon- 42. Hopkins AL (2009) Drug discovery: predict-
struction. Magn Reson Med 76:1059–1070 ing promiscuity. Nature 462:167–168
27. Chapman T (2003) Lab automation and 43. Lounkine E, Keiser MJ, Whitebread S et al
robotics: automation on the move. Nature (2012) Large-scale prediction and testing of
421:661–666 drug activity on side-effect targets. Nature
28. Butte AJ, Ohno-Machado L (2013) Making 486:361–367
it personal: translational bioinformatics. J Am 44. Pauwels E, Stoven V, Yamanishi Y (2011)
Med Inform Assoc 20:595–596 Predicting drug side-effect profiles: a
29. Kitano H (2002) Systems biology: a brief chemical fragment-based approach. BMC
overview. Science 295:1662–1664 Bioinformatics 12:169
30. Plant NJ, Vinken M, Kolodkin A, Boogerd 45. Wang F, Zhang P, Cao N, Hu J, Sorrentino R
FC, Al. E, Borisy AA (2015) An introduction (2014) Exploring the associations between
to systems toxicology. Toxicol Res 4:9–22 drug side-effects and therapeutic indica-
31. Bouhifd M, Hogberg HT, Kleensang A, tions. J Biomed Inform 51:15–23
Maertens A, Zhao L, Hartung T (2014) 46. Barneh F, Jafari M, Mirzaie M (2015)
Mapping the human toxome by systems Updates on drug–target network; facilitat-
toxicology. Basic Clin Pharmacol Toxicol ing polypharmacology and data integra-
115:24–31
tion by growth of DrugBank database. Brief 61. Yabuuchi H, Niijima S, Takematsu H, Ida T,
Bioinform 17:bbv094 Hirokawa T, Hara T, Ogawa T, Minowa Y,
47. Campbell SJ, Gaulton A, Marshall J, Bichko Tsujimoto G, Okuno Y (2014) Analysis of multi-
D, Martin S, Brouwer C, Harland L (2012) ple compound-protein interactions reveals novel
Visualizing the drug target landscape. Drug bioactive molecules. Mol Syst Biol 7:472–472
Discov Today 17:S3–S15 62. Perlman L, Gottlieb A, Atias N, Ruppin E,
48. Swamidass SJ (2011) Mining small-molecule Sharan R (2011) Combining drug and gene
screens to repurpose drugs. Brief Bioinform similarity measures for drug-target elucida-
12:327–335 tion. J Comput Biol 18:133–145
49. Moriaud F, Richard SB, Adcock SA, Chanas- 63. Yamanishi Y, Pauwels E, Kotera M (2012)
Martin L, Surgand J-S, Ben Jelloul M, Drug side-effect prediction based on the
Delfaud F (2011) Identify drug repurposing integration of chemical and biological spaces.
candidates by mining the Protein Data Bank. J Chem Inf Model 52:3284–3292
Brief Bioinform 12:336–340 64. Cobanoglu MC, Liu C, Hu F, Oltvai ZN,
50. Dobson CM (2004) Chemical space and biol- Bahar I (2013) Predicting drug-target inter-
ogy. Nature 432:824–828 actions using probabilistic matrix factoriza-
51. Vogt I, Mestres J (2010) Drug-target net- tion. J Chem Inf Model 53:3399–3409
works. Mol Inform 29:10–14 65. Mayr A, Klambauer G, Unterthiner T,
52. Forli S, Huey R, Pique ME, Sanner MF, Hochreiter S (2016) DeepTox: toxicity pre-
Goodsell DS, Olson AJ (2016) Computational diction using deep learning. Front Environ
protein–ligand docking and virtual drug Sci 3:80
screening with the AutoDock suite. Nat 66. Min S, Lee B, Yoon S (2016) Deep learn-
Protoc 11:905–919 ing in bioinformatics. Brief Bioinform
53. Xie L, Wang J, Bourne PE (2007) In silico 18(5):851–869
elucidation of the molecular mechanism 67. LeCun Y, Bengio Y, Hinton G (2015) Deep
defining the adverse effect of selective estro- learning. Nature 521:436–444
gen receptor modulators. PLoS Comput Biol 68. Baldi P, Sadowski P, Whiteson D (2014)
3:2324–2332 Searching for exotic particles in high-energy
54. Halperin I, Ma B, Wolfson H, Nussinov R physics with deep learning. Nat Commun
(2002) Principles of docking: an overview of 5:ncomms5308
search algorithms and a guide to scoring func- 69. Esteva A, Kuprel B, Novoa RA, Ko J,
tions. Proteins 47:409–443 Swetter SM, Blau HM, Thrun S (2017)
55. Shoichet BK, McGovern SL, Wei B, Irwin JJ Dermatologist-level classification of skin
(2002) Lead discovery using molecular dock- cancer with deep neural networks. Nature
ing. Curr Opin Chem Biol 6:439–446 542:115–118
56. Cheng AC, Coleman RG, Smyth KT, Cao Q, 70. Wu Y, Schuster M, Chen Z, Le QV, Norouzi M
Soulard P, Caffrey DR, Salzberg AC, Huang et al (2016) Google’s neural machine transla-
ES (2007) Structure-based maximal affinity tion system: bridging the gap between human
model predicts small-molecule druggability. and machine translation, arXiv:1609.08144v2
Nat Biotechnol 25:71–75 71. Ma’ayan A, Jenkins SL, Goldfarb J, Iyengar R
57. Wallach I, Jaitly N, Lilien R (2010) A (2007) Network analysis of FDA approved
structure-based approach for mapping adverse drugs and their targets. Mt Sinai J Med
drug reactions to the perturbation of underly- 74:27–32
ing biological pathways. PLoS One 5:e12063 72. Lin H-H, Zhang L-L, Yan R, Lu J-J, Hu Y
58. Huey R, Morris GM, Olson AJ, Goodsell DS (2017) Network analysis of drug-target inter-
(2007) A semiempirical free energy force field actions: a study on FDA-approved new molec-
with charge-based desolvation. J Comput ular entities between 2000 to 2015. Sci Rep
Chem 28:1145–1152 7:12230
59. Xia Z, Wu L-Y, Zhou X, Wong ST (2010) 73. McGary KL, Park TJ, Woods JO, Cha HJ,
Semi-supervised drug-protein interaction Wallingford JB, Marcotte EM (2010)
prediction from heterogeneous biological Systematic discovery of nonobvious human
spaces. BMC Syst Biol 4:S6 disease models through orthologous
60. van Laarhoven T, Nabuurs SB, Marchiori phenotypes. Proc Natl Acad Sci U S A
E (2011) Gaussian interaction profile ker- 107:6544–6549
nels for predicting drug–target interaction. 74. Korcsmáros T, Szalay MS, Rovó P, Palotai R,
Bioinformatics 27:3036–3043 Fazekas D, Lenti K, Farkas IJ, Csermely P,
Vellai T (2011) Signalogs: Orthology-based 88. Gottlieb A, Altman RB (2014) Integrating

identification of novel signaling pathway systems biology sources illuminates drug
components in three metazoans. PLoS One action. Clin Pharmacol Ther 95:1–7
6:e19240 89. Fan S, Geng Q, Pan Z, Li X, Tie L, Pan Y,
75. Garcia-Serna R, Vidal D, Remez N, Mestres Li X (2012) Clarifying off-target effects for
J (2015) Large-scale predictive drug safety: torcetrapib using network pharmacology and
from structural alerts to biological mecha- reverse docking approach. BMC Syst Biol
nisms. Chem Res Toxicol 28:1875–1887 6:152
76. Remez N, Garcia-Serna R, Vidal D, Mestres 90. Azuaje FJ, Zhang L, Devaux Y, Wagner DR
J (2016) The in vitro pharmacological profile (2011) Drug-target network in myocardial
of drugs as a proxy indicator of potential in vivo infarction reveals multiple side effects of unre-
organ toxicities. Chem Res Toxicol 29:637–648 lated drugs. Sci Rep 1:1–10
77. Kuhn M, Al Banchaabouchi M, Campillos M, 91. Huang LC, Wu X, Chen JY (2013) Predicting
Jensen LJ, Gross C, Gavin A-C, Bork P (2014) adverse drug reaction profiles by integrating
Systematic identification of proteins that elicit protein interaction networks with drug struc-
drug side effects. Mol Syst Biol 9:663–663 tures. Proteomics 13:313–324
78. Yildirim MA, Goh K-I, Cusick ME, Barabási 92. Huang J, Niu C, Green CD, Yang L, Mei
A-L, Vidal M (2007) Drug-target network. H, Han JDJ (2013) Systematic prediction
Nat Biotechnol 25:1119–1126 of pharmacodynamic drug-drug interactions
79. Luo H, Chen J, Shi L, Mikailov M, Zhu H, through protein-protein-interaction network.
Wang K, He L, Yang L (2011) DRAR-CPI: PLoS Comput Biol 9:e1002998
a server for identifying drug repositioning 93. Pouliot Y, Chiang AP, Butte AJ (2011)
potential and adverse drug reactions via the Predicting adverse drug reactions using pub-
chemical-protein interactome. Nucleic Acids licly available PubChem BioAssay data. Clin
Res 39:W492–W498 Pharmacol Ther 90:90–99
80. Daminelli S, Haupt VJ, Reimann M, 94. Iskar M, Zeller G, Zhao XM, van Noort V,
Schroeder M (2012) Drug repositioning Bork P (2012) Drug discovery in the age of
through incomplete bi-cliques in an inte- systems biology: the rise of computational
grated drug-target-disease network. Integr approaches for data integration. Curr Opin
Biol (Camb) 4:778–788 Biotechnol 23:609–616
81. Gottlieb A, Stein GY, Ruppin E, Sharan R 95. Bai JPF, Abernethy DR (2013) Systems phar-
(2014) PREDICT: a method for inferring macology to predict drug toxicity: integration
novel drug indications with application to per- across levels of biological organization. Annu
sonalized medicine. Mol Syst Biol 7:496–496 Rev Pharmacol Toxicol 53:451–473
82. Lee H, Bae T, Lee J-H et al (2012) Rational 96. Rahmani H, Weiss G, Méndez-Lucio O,
drug repositioning guided by an integrated Bender A (2016) ARWAR: a network
pharmacological network of protein, disease approach for predicting adverse drug reac-
and drug. BMC Syst Biol 6:80 tions. Comput Biol Med 68:101–108
83. Zhao S, Li S (2012) A co-module approach 97. Dorel M, Barillot E, Zinovyev A, Kuperstein I
for elucidating drug–disease associations and (2015) Network-based approaches for drug
revealing their molecular basis. Bioinformatics response prediction and targeted therapy
28:955–961 development in cancer. Biochem Biophys
84. Guney E, Garcia-Garcia J, Oliva B (2014) Res Commun 464:386–391
GUILDify: a web server for phenotypic char- 98. Pujol A, Mosca R, Farrés J, Aloy P (2010)
acterization of genes through biological data Unveiling the role of network and systems
integration and network-based prioritization biology in drug discovery. Trends Pharmacol
algorithms. Bioinformatics 30:1789–1790 Sci 31:115–123
85. Guney E, Menche J, Vidal M, Barábasi A-L 99. Mucha PJ, Richardson T, Macon K, Porter MA,
(2016) Network-based in silico drug efficacy Onnela J-P (2010) Community structure in
screening. Nat Commun 7:10331 time-dependent, multiscale, and multiplex net-
86. Cami A, Arnold A, Manzi S, Reis B (2011) works. Science 328:876–878
Predicting adverse drug events using phar- 100. Zenil H, Kiani NA, Tegnér J (2016) Methods
macological network models. Sci Transl Med of information theory and algorithmic com-
3:114ra127–114ra127 plexity for network biology. Semin Cell Dev
87. Brouwers L, Iskar M, Zeller G, van Noort V, Biol 51:32–43
Bork P (2011) Network neighbors of drug 101.
Zenil H, Hernández-Orozco S, Kiani
targets contribute to drug side-effect similar- NA, Soler-Toscano F, Rueda-Toicen A
ity. PLoS One 6:e22187 (2016) A decomposition method for global
evaluation of shannon entropy and local Random forest: a classification and regres-
estimations of algorithmic complexity, sion tool for compound classification and
arXiv:1609.00110v5 QSAR modeling. https://doi.org/10.1021/
102. Burden FR, Winkler DA (2015) Relevance CI034160G
vector machines: sparse classification 104. Lei T, Li Y, Song Y, Li D, Sun H, Hou T
methods for QSAR. J Chem Inf Model (2016) ADMET evaluation in drug discov-
55:1529–1534 ery: 15. Accurate prediction of rat oral acute
103. Svetnik V, Liaw A, Tong C, Christopher
toxicity using relevance vector machine and
Culberson J, Sheridan RP, Feuston BP (2003) consensus modeling. J Cheminform 8:6
Chapter 26
Chemoinformatic Approach to Assess Toxicity

of Ionic Liquids
Anita Sosnowska, Anna Rybinska-Fryca, Maciej Barycki,
Karolina Jagiello, and Tomasz Puzyn
Abstract
Chemoinformatic methods, such as multivariable explorative techniques and quantitative structure–activ-
ity relationship (QSAR) modeling, allow for discovering relationships between the activity and the structure
of chemical compounds. These techniques can be applied, as preliminary screening methods for designing
and/or selecting new compounds with defined activity.
Here we describe step by step how to preliminarily screen ionic liquids (a set of 13 ILs) and assess
their cytotoxic activity against leukemia cell line IPC-81 as well as ILs’ potential to inhibit acetylcholines-
terase enzyme using the TRIC method (toxicity ranking index of cations) combined with the QSAR
approach.
Key words Ionic liquids, Multivariable explorative technique, TRIC, Quantitative structure–activity
relationship, Molecular descriptors
1 Introduction
Ionic liquids (ILs) comprise an interesting group of chemical spe-

cies, which can be successfully used for improving various processes
in organic chemistry and chemical technology [1]. Their unique-
ness is expressed in several properties like melting point below
100 °C, high thermal stability [2], or low vapor pressure [3], which
distinguish them from other chemicals commercially available on
the market. The presence of these properties allows for application
of ionic liquids in many branches of today’s chemistry such as
synthesis [1], catalysis [4] or biocatalysis [5]. Moreover, they can
be used in biomass processing [6], in drug delivery systems [7] and
as biocidal agents against microbial biofilms [8].
In the recent past, ionic liquids were considered “green chemi-
cals”—safe for human and environmentally friendly [9, 10].
Nowadays, however, a number of research studies prove that they can
559
560 Anita Sosnowska et al.
be toxic [11–16]. Thus, it is necessary to perform an evaluation of

ILs’ potential impact on living organisms (including humans) before
they are introduced on a large scale into the industrial processes.
Ionic liquids consist of relatively large, typically unsymmetrical
organic cations and usually smaller, inorganic or organic anions.
Possible modifications of cations’ or anions’ structure in a given
ionic liquid lead to an enormous number of new ionic liquids pos-
sible to obtain (each of them with different toxicological activity
and physicochemical properties). Detailed experimental studies on
toxicological effects of such a large chemical group are very
expensive and time-consuming. In this case, the application of
computational screening methods (i.e., multivariate explorative
chemometric techniques combined with the quantitative struc-
ture–activity relationship (QSAR) modeling) can effectively reduce
time and cost of the experimental measurements.
Multivariate explorative techniques are a group of methods
that allow for deeper analysis of the information contained in the
complex data. The aim of this approach is to extract the useful
information from the multivariate data set and transfer that into
more easily interpretable representations (e.g., graphs). Such
methods may be employed to discover relationships between par-
ticular variables or to perform screening of tested objects and
obtain a feature based ranking (for example ranking of chemicals
based on the studied physicochemical properties [17], and/or
activity [18]).
The biological activity of the compounds is determined by
their chemical structure and this relationship can be described by
QSAR models [19]. QSAR methods are based on the assumption,
that for a given set of compounds, the variance in the modeled
endpoint (e.g., activity) is determined by the variance in their
molecular structures (expressed by the so-called “molecular
descriptors”). Consequently, it is possible to find a quantitative
relationship between the descriptors and the endpoint. The idea is
to develop a mathematical model for chemicals, for which the end-
point has been experimentally measured. Thus, when the value of
the endpoint for a (structurally) similar chemical is not available, it
is possible to calculate it from theoretically derived descriptors,
using a mathematical model previously developed.
In this chapter, we present chemoinformatic tools for assessing
the cytotoxicity of the sample set of 13 ILs. We employed the tox-
icity ranking index of cations (TRIC) [18] (a simple metric of the
cation-related toxicity of ILs), as an example of multivariate explor-
ative technique. This was applied to obtain a qualitative c ytotoxicity
screening and ranking of the studied ILs. Then, we used two previ-
ously developed QSAR models predicting (1) cytotoxicity of ILs
toward the leukemia cell line IPC-81 (log EC50 values) [20] and
(2) ILs’ ability of acetylcholinesterase inhibition (log EC50 values).
The presented approach allows for ranking ionic liquids in case of
studied toxicity for further detailed analyses.
Chemoinformatic Approach to Assess Toxicity of Ionic Liquids 561
2 Materials
In order to perform the chemometric analysis several computer

programs are needed. Firstly, the molecular models of studied
compounds should be built. There are various structure generators
like ChemSketch [21] or Avogadro [22]. Those two programs
have user-friendly interface and possibility of generating structures
directly from the SMILES notation. Subsequently, the molecular
descriptors can be calculated in Dragon [23] or Padel [24] soft-
ware. However, in order to calculate the 3D descriptors, geometry
optimization of the molecular structure has to be performed. For
that purpose MOPAC software [25] can be used. It should be
mentioned that each program generates output files in different
formats, and that is why the files conversion should be conducted.
Possible programs allowing for converting the files are for example
Open Babel [26] or Avogadro [22], MOLDEN [27].
3 Methods
3.1 Dataset 1. First, determine the set of ILs that has to be examined. Here
we present the study for 13 ionic liquids with known activ-
ity in order to verify the accuracy of the analysis (IL1—1-
ethyl-3-methylimidazolium bis(pentafluoroethyl)phosphinate;
IL2—butylethyldimethylammonium bis(trifluoromethylsulfonyl)
amide; IL3—1-(3-methoxypropyl)-3-methylimidazolium chloride;
IL4—1-butyl-3-methylimidazolium trifluoromethanesulfonate;
IL5—1-butyl-3 methylimidazolium 1-octylsulfate; IL6—1-hexyl-
3-methylimidazolium 1,1-dioxo-1,2-dihydrobenzo[d]isothiazol-
3-onate; IL7—4 (dimethylamino)-1-butylpyridinium bis(trifluoro
methylsulfonyl)amide; IL8—1-heptyl-3-methylimidazolium tetra
fluoroborate; IL9—3 methyl-1-octylimidazolium chloride;
IL10—3-methyl-1-nonylimidazolium tetrafluoroborate; IL11—
1-decyl-3-methylimidazolium chloride; IL12—1-hexadecyl-3-
methylimidazoliumchloride;IL13—3-methyl-1-octadecylimidazolium
chloride).
2. Try to arrange the data in a clear and manageable manner. We
recommend creating a calculation worksheet (like Excel) and
store all the data in Table 1 (see Note 1).
3.2 Molecular 1. Draw a chemical structure of each cation and anion in an

Descriptors appropriate software (we recommend ChemSketch [21] or
Avogadro [22]) (see Note 2). Save structures in a *.mol or
*.mol2 format, so they can be easily opened in other programs
utilized in the following steps.
Table 1
Scheme presenting a convenient way of data preparation
Descriptors
ID Name SMILES Descriptors for cation Descriptors for anion

IL1
IL2
IL3
…
ILn
2. Convert .mol or .mol2 structures of studied cations and anions

to a MOPAC [25] input .mop files. The simplest way to do
that is to use free AVOGADRO program [22] (see Note 3).
3. Add appropriate keywords in the command line (first line) in
the each of the input *.mop file. For geometry optimization
use the semiempirical PM7 method, simply by adding “PM7”
keyword. It is a good practice to increase the precision of cal-
culations by 100 times with “PRECISE” keyword. For each
ion add “CHARGE=1” or “CHARGE=−1” keyword depend-
ing on ion’s charge (see Note 4). An example first line of *.
mop file for cation: “PM7 PRECISE CHARGE=1”.
4. Run MOPAC to optimize structures. Check .out files whether
the optimization ended without any error (see Note 5). After that
prepare the descriptors software input file (*.mol2 files in case of
calculating the descriptors in DRAGON software) (see Note 6).
5. Calculate molecular descriptors using DRAGON software
[23] (see Notes 7 and 8). In the presented case all (82) WHIM
descriptors for cations, as well as DELS, SpMax8_Bh(e),
RDF035p, Mor05m, MAXDP, nROR, D/Dtr10 and MSD
for cations and QXXi and SNar for anions are needed. Fill the
earlier prepared table (Table 1) with the appropriate obtained
values of the descriptors.
3.3 TRIC Analysis 1. For screening and choosing the least toxic ILs, screening tool
should be used. For this purpose we recommend TRIC metric
[18], which was created for various endpoints and therefore is
capable of predicting toxic activity of ionic liquid against
various organisms (see Note 9).
2. In order to use TRIC, simply multiply the vector of WHIM
descriptors by the vector of loadings. Here, we present the
Fig. 1 Ionic liquids arranged by increasing value of TRIC metric. Blue color represents the ILs with the smallest
TRIC value selected for further analyses; red color represents all the other ILs
values of loadings vector, so that user does not have to look for
it in the literature: 0.172; 0.113; 0.107; 0.110; −0.098;
−0.057; −0.034; −0.049; −0.026; −0.126; 0.007; 0.168;
0.119; 0.113; 0.094; −0.082; −0.057; −0.033; −0.044;
0.069; 0.044; 0.087; 0.172; 0.114; 0.108; 0.110; −0.097;
−0.057; −0.036; −0.052; −0.011; −0.117; 0.017; 0.171;
0.113; 0.103; 0.106; −0.094; −0.063; −0.032; −0.049;
0.053; −0.087; 0.024; 0.172; 0.113; 0.108; 0.111; −0.098;
−0.057; −0.035; −0.048; −0.035; −0.124; 0.007; 0.176;
0.173; 0.176; 0.175; 0.176; 0.167; 0.164; 0.167; 0.165;
0.167; −0.069; −0.073; 0.111; 0.095; 0.110; 0.107; 0.111;
−0.091; 0.095; −0.067; −0.003; −0.091; 0.163; 0.164;
0.163; 0.163; 0.163 (see Note 10).
3. Arrange the ILs according to the values of the multiplication
result. Those values represent relative toxicity of each
IL. Higher values mean higher toxicity. The result for tested
set of 13 ILs is presented on the Fig. 1.
4. Choose ILs with the lowest TRIC values (most promising for
eventual application) for further analyses. In the studied case,
these are IL1–IL5.
3.4 QSAR Predictions 1. This step covers quantitative prediction of toxicity values for
each IL. In order to do so, choose appropriate QSAR models
(see Note 11). In studied case, we present use of two models
capable of predicting ILs’ toxicity against IPC-81 cells in via-
bility test and inhibition of acetylocholinesterase test
respectively.
2. In order to determine ILs’ toxicity against IPC-81 leukemia
cells, substitute descriptors’ values to the QSAR equation
(here we provide model’s equation to make the use of this
methodology fast—superscripts C and A correspond to the

ion (cation or anion) for which the descriptors were c alculated)
(see Note 12).
logEC50 [ IPC-81] = 5.386 − 0.619 * MSD _ C

+0.255 * MAXDP _ C + 0.374 * nROR _ C
−0.023 * D/Dtr10 _ C − 0.090 * SNar _ A
Model statistics : R 2 = 0.772; RMSEC

= 0= .513; Q 2CV 0.758; RMSECV
= = 0.528; Q 2 F 2 0=
.839; RMSEExt 0.434.
3. Reliability of the prediction should be verified by assessing

ILs’ relation with model’s applicability domain (AD). Here we
use the Insubria plot to visualize the AD. The borders of the
AD are determined by the values of minimal and maximal
predictions for the training set (here value predicted for each
tested IL should be between −0.8611 and 4.6487) and the
threshold leverage value h*, which determines whether tested
IL is structurally similar to ILs from the training set. To calcu-
late the leverage value please use the following equation:
h = x kT ( X T X ) x k
−1
In this equation xk stays for the vector of descriptors for tested

IL, in the order given in the QSAR equation. (XTX)−1 part for
IPC-81 cells in the viability test model is as follows:
0.00039 −0.00018 −0.00026 −0.00018 0.00001

−0.00018 0.00062 −0.00070 −0.00028 −0.00004
(XTX)−1= −0.00026 −0.00070 0.00404 −0.00399 0.00003
−0.00018 −0.00028 −0.00399 0.02087 0.00004
0.00001 −0.00004 0.00003 0.00004 0.00005
The leverage value for tested IL should be lower than

h* = 0.0789 (see Note 13).
Determine whether all tested ILs are within the AD of the
analyzed model. The reliability of the QSAR predictions is much
lower for ILs being outside the domain (see Note 14). In the pre-
sented case all predictions can be considered reliable (Fig. 2).
4. Execute the similar procedure for the second QSAR model. In
this case substitute descriptors’ values for cation and anion to
the QSAR equation for estimating toxicity in the
acetylocholinesterase test:
Fig. 2 Insubria plot showing the relation of five tested ILs with the AD QSAR
model for toxicity against IPC-81 leukemia cells
logEC50 [ AChE ] = 2.594 + 0.183∗ DELS _ C

−0.329 SpMax8 _ Bh ( e ) _ C + 0.127∗ RDF035p _ C
+0.644∗ Mor 05m _ C + 0.004∗ QXXi _ A;
=
Model statistics : R 2 0=
.76; RMSEC 0.31; Q 2CV
= 0= .76; RMSECV 0.33; Q 2 F 2
= 0= .71; RMSEExt 0.38;

5. Determine whether tested ILs are within the AD. In this

model minimal and maximal prediction for the training set are
respectively: 0.9723 and 3.6226.
6. Calculate leverage values. The critical leverage value here is
h* = 0.100, and the (XTX)−1 part for that model is as follows:
0.00125 0.00046 −0.00031 0.00236 −0.00001

0.00046 0.03234 −0.00468 0.02888 −0.00012
(XTX)−1= −0.00031 −0.00468 0.00163 −0.00266 0.00001
0.00236 0.02888 −0.00266 0.03620 −0.00006
−0.00001 −0.00012 0.00001 −0.00006 0.00001
In the presented example all five tested ILs belong to the AD

of the model (Fig. 3) (see Note 13).
Fig. 3 Insubria plot showing the relation of five tested ILs with the AD QSAR
model for inhibition of acetylocholinesterase
Table 2
Obtained results for five ILs with assigned ranks (color green indicates IL which shows the least
toxicity in considered tests)
Observed Predicted Ranks
TRIC IPC-81 IPC-81 IPC-81
AChE Test test AChE Test test TRIC AChE Test test
IL3 49.36 2.61 4.49 2.34 3.77 3 1 1
IL1 31.41 2.09 2.83 2.32 3.25 1 2 2
IL2 45.37 2.03 3.14 2.30 3.01 2 3 4
IL4 52.57 1.93 3.02 2.22 3.11 4 4 3
IL5 53.91 1.98 3.23 2.08 2.74 5 5 5
7. Gather all predicted data (from TRIC analysis and QSAR pre-
dictions) in one table.
Create rules according to which you can choose the most
appropriate IL for the further use. In considered case, we assigned
ranks to the qualitative predictions for studied set of ionic liquids
(Table 2). After that, we choose the one with mean lowest value:
IL3—1-(3-methoxypropyl)-3-methylimidazolium chloride
(see Notes 15 and 16).
Fig. 4 Observed versus predicted values of toxicity for used QSAR models: Left panel—QSAR model for toxicity
against IPC-81 leukemia cells; right panel—inhibition of acetylocholinesterase. Blue color represents the ILs
with the smallest TRIC value selected for further analyses; red color represents all other analyzed ILs
To show that the chosen ILs were in fact those one with the
lowest toxicity we present the plots of experimental values vs. pre-
dicted by QSAR models for all studied 13 ionic liquids in both
viability and inhibition tests (Fig. 4).
4 Notes
1. In order to examine a bigger set of ILs find the SMILES nota-

tion for every cation and anion in your dataset. This will save
a lot of time with drawing structures.
2. If you have the SMILES notation for each of ions we strongly
recommend using ChemSketch software. It has a simple-to-
use structure generator, working with SMILES.

3.
Generate a MOPAC input file by choosing
EXTENSIONS ⇒ MOPAC in AVOGADRO software. Our
experience says that input based on z-matrix coordinates is the
most appropriate format for MOPAC calculations.
4. MOPAC generates two output files by default: *.out and
*.arc. Optimized structure can be easily extracted from *.arc
file but only a few computer programs can read this format
(MOLDEN [27] is one of them). We recommend adding
“PDBOUT” keyword in order to receive file with optimized
structure that has a *.pdb extension and can be opened in a
variety of programs.
5. Different things can indicate that something went wrong.

First of all, presence of additional files after geometry optimi-
zation (besides *.out and *.arc) is not good. Usually, MOPAC
generates additional *.den and *.res files if an optimal
geometry of a molecule was not achieved. In such case *.out
file contains the structure with the geometry “so far.” The
problem can be solved by extending number of cycles in the
original *.mop input by adding “CYCLES=<value>” keyword
(for example: CYCLES=2000). Even if the number of output
files is correct, always check the structure from the output file
(.out, or .pbd). “JOB ENDED NORMALLY” statement at
the end of the .out file does not necessarily indicate that every-
thing is correct. Use MOPAC manual for troubleshooting.
6. Dependently on the software chose to descriptors calculation
the inputs should be prepared in a proper format. In case of
DRAGON software which we use, the best way is to convert
*.arc or *.pdb files to .mol2 files in MOLDEN, Open Babel or
AVOGADRO software.
7. After importing the inputs to the DRAGON always check if
the displayed structures are correct. If something is wrong go
back to the previous step and check if the structure geometry
optimization went correctly.
8. Be careful with export options in DRAGON. It excludes con-
stant and near constant columns from the calculated descrip-
tors set by default. Uncheck this option in “export options.”
9. Since the TRIC metric was developed on the basis of trends in
the toxicity of restricted group of tested targets, the predic-
tions obtained by the TRIC should be treated as informative
values of first prescreening. For full assessment of the toxicity
of ILs, additional toxicological tests (for defined organisms)
should be performed.
10. Use an appropriate software like Octave, R, or MATLAB,
especially if you have a big set of ILs.
11. Choosing the best QSAR model is in fact a difficult process.
The best models always follow five OECD rules ((1) a defined
endpoint, (2) an unambiguous algorithm, (3) a defined
domain of applicability, (4) appropriate measures of goodness-
of-fit, robustness, and predictivity, (5) a mechanistic interpre-
tation, if possible). Moreover, a model with best statistics is
not always the best one. Models created on the basis of bigger
training set might be better in many cases, but it is always case
dependent. Also, be aware that some software used for model-
ing (especially those for descriptors calculation) might be
commercial. Make sure that you have access to all needed
software.
12. Common practice is that QSAR models are developed on the
autoscaled (normalized and centered) descriptors. In order to
Fig. 5 Schematic Insubria plot for evaluating whether the compound is in the AD
of the QSAR model
use the model equation, one needs to do the same transforma-

tion on the descriptors. Make sure you substitute correct val-
ues to the QSAR equations.
13. Here we present the example QSAR models, please remember
that the values of minimal and maximal prediction for the
training set, as well as threshold leverage values are dependent
on the dataset that was used during the modeling. Also the
method of determining the AD could be different—to avoid
the mistakes please always check the details in the QSAR mod-
el’s description.
14. The AD is a space, defined by the values of the descriptors (Xi)
and the response of the model (y), in which the predictions are
reliable. Predicted values that are outside the square of AD are
usually the result of the extrapolation of the model and has to
be considered as less reliable, but still can be taken into account
as the estimation (Fig. 5).
15. In the presented analysis we assigned ranks from 1 to 5 to the
predicted by QSAR models values of toxicity. 1 corresponds to
the less toxic IL (highest logEC50 values), whereas 5 to the
most toxic IL (lowest logEC50 values). One could establish
here other ranking methods.
16. The obtained results show that the two presented tools can
give different results. Here the TRIC value for IL3 is third in
order for less toxicity, while in QSAR predictions IL3 gives the
best (less toxic) results. One should remember that TRIC
metric was developed based on trends related to the toxicity of
ten different organisms, so this metric is more informative
than quantitative.
Acknowledgments
This material is based on research funded by the National Science

Center (Poland) (Grant No. UMO-2012/05/E/NZ7/01148).
References
1. Wasserscheid P, Welton T (2002) Ionic liquids benign ionic liquids (1-alkoxymethyl-3-hydro

in synthesis. Wiley-VCH, Weinheim, Germany xypyridinium chloride, saccharinate and ace-
2. Endres F, El Abedin SZ (2006) Air and water sulfamates) on cellular and molecular level.
stable ionic liquids in physical chemistry. Phys Ecotoxicol Environ Saf 71:157–165
Chem Chem Phys 8(18):2101–2116 13. Stock F, Hoffmann J, Ranke J, Stormann R,
3. Aschenbrenner O, Supasitmongkol S, Taylor Ondruschka B, Jastorff B (2004) Effects of
M, Styring P (2009) Measurement of vapour ionic liquids on the acetylcholinesterase – a
pressures of ionic liquids and other low vapour structure-activity relationship consideration.
pressure solvents. Green Chem Green Chem 6:286–290
11(8):1217–1221 14. Ranke J, Molter K, Stock F, Bottin-Weber U,
4. Welton T (1999) Room-temperature ionic liq- Poczobutt J, Hoffmann J, Ondruschka B,
uids. Solvents for synthesis and catalysis. Chem Filser J, Jastorff B (2004) Biological effects of
Rev 99(8):2071–2084 imidazolium ionic liquids with varying chain
5. Sheldon RA, Lau RM, Sorgedrager MJ, van lengths in acute Vibrio fischeri and WST-1 cell
Rantwijk F, Seddon KR (2002) Biocatalysis in viability assays. Ecotoxicol Environ Saf
ionic liquids. Green Chem 4(2):147–151 58:396–404
6. Tan SS, Macfarlane DR (2010) Ionic liquids
15. Frade RFM, Matias A, Branco LC, Afonso
in biomass processing. Top Curr Chem CAM, Duarte CMM (2007) Effect of ionic
290:311–339 liquids on human colon carcinoma HT-29
and CaCo-2 cell lines. Green Chem
7. Adawiyah N, Moniruzzaman M, Hawatulailaa 9(8):873–877
S, Goto M (2016) Ionic liquids as a potential
tool for drug delivery systems. MedChemComm 16. Latala A, Nedzi M, Stepnowski P (2010)
7:1881–1897 Toxicity of imidazolium ionic liquids towards
algae. Influence of salinity variations. Green
8. Gilmore BF, Earle MJ (2011) Development of Chem 12:60–64
ionic liquid biocides against microbial biofilms
Designer microbicides for infection control. 17. Gramatica P, Pilutti P, Papa E (2004) A tool for
Chim Oggi 29:50–53 the assessment of VOC degradability by tropo-
spheric oxidants starting from chemical struc-
9. Das RN, Roy K (2013) Advances in QSPR/ ture. Atmos Environ 38:6167–6175
QSTR models of ionic liquids for the design of
greener solvents of the future. Mol Divers 18. Sosnowska A, Barycki M, Zaborowska M,
17:151–196. https://doi.org/10.1007/ Rybinska A, Puzyn T (2014) Towards design-
s11030-012-9413-y ing environmentally safe ionic liquids: the
influence of the cation structure. Green Chem
10. Dukhande VA, Choksi TS, Sabnis SU, 16:4749–4757
Patwardhan AW, Patwardhan AV (2013)
Separation of toluene from n-heptane using 19. Hansch C, Fujita T (1964) Rho-sigma-pi anal-
monocationic and dicationic ionic liquids. ysis. Method for correlation of biological activ-
Fluid Phase Equilib 342:75–81 ity + chemical structure. J Am Chem Soc
86:1616–1626
11. Viboud S, Papaiconomou N, Cortesi A, Chatel
G, Draye M, Fontvieille D (2012) Correlating 20. Sosnowska A, Grzonkowska M, Puzyn T
the structure and composition of ionic liquids (2017) Global versus local QSAR models for
with their toxicity on Vibrio fischeri: a system- predicting ionic liquids toxicity against IPC-81
atic study. J Hazard Mater 215:40–48 leukemia rat cell line: the predictive ability.
J Mol Liq 231:333–340
12.
Stasiewicz M, Mulkiewicz E, Tomczak-
Wandzel R, Kumirska J, Siedlecka EM, 21. ACD/ChemSketch v, Advanced Chemistry
Golebiowski M, Gajdus J, Czerwicka M, Development, Inc., Toronto, ON, Canada.
Stepnowski P (2008) Assessing toxicity and http://www.acdlabs.com, 2008
biodegradation of novel, environmentally 22. http://avogadro.cc/ Aao-smbavtVX, 2012
23. Dragon Software for Molecular Descriptor 26. O’Boyle NM, Banck M, James CA, Morley C,
Calculation hwtmi, Milano, 2014 Vandermeersch T, Hutchison GR (2011) Open
24. Yap CW (2011) PaDEL-descriptor: an open babel: an open chemical toolbox. J Cheminform
source software to calculate molecular descrip- 3:33
tors and fingerprints. J Comput Chem 27. Schaftenaar G, Noordik JH (2000) Molden: a
32:1466–1474 pre- and post-processing program for molecu-
25. Stewart JJP (2012) Stewart computational lar and electronic structures. J Comput Aided
chemistry, Colorado Springs, CO, USA Mol Des 14:123–134
Chapter 27
Prediction of Biochemical Endpoints by the CORAL

Software: Prejudices, Paradoxes, and Results
Andrey A. Toropov, Alla P. Toropova, Alessandra Roncaglioni,
Abstract
Quantitative structure–activity relationships (QSARs) for prediction of toxicological endpoints built up
with the CORAL software are discussed. Prejudices related to these QSAR models are listed. Possible ways
to improve the software are discussed.
Key words Toxicity, QSAR, SMILES, Monte Carlo technique, CORAL software, Mutagenicity,
Psychotropic drug toxicity, Drug inhibitor activity
1 Introduction
The prediction of toxicological endpoints via quantitative struc-

ture–property/activity relationships (QSPRs/QSARs) is a tool to
understand why a substance (pollutant, cosmetic, drug, food, etc.)
leads to damage in various organisms, i.e., bacteria, algae, fishes,
animals, and humans. The predictions must obey the following
OECD principles (http://www.oecd.org/datao-
ecd/55/35/38130292.pdf):
(a) A defined endpoint.
(b) An unambiguous algorithm.
(c) A defined domain of applicability.
(d) Appropriate measures of goodness-of-fit, robustness, and

predictivity.
(e) A mechanistic interpretation, if possible.
Usually, the relevant factors are some fragments of molecular archi-
tecture, which are indicators of the dangerous behavior of the
molecular structure [1–3].
573
574 Andrey A. Toropov et al.
It is impossible to get experimental data on toxic potential for

all substances involved in agricultural [4] and industrial [5] pro-
cesses, in biochemical research [6], medical problems (cosmetics
[7], drug discovery [8, 9]). Consequently, it is necessary to develop
alternative strategies which are feasible, and one valuable possibil-
ity is the use of computational methods for estimation of the level
of toxic effects of different molecular agents (that may be used for
risk assessment) [10–12].
The development of the abovementioned approaches is based
on the so-called molecular descriptors, i.e., calculable characteris-
tics of molecular architecture and/or phenomena related to toxi-
cological processes [13]. The variations of concepts to define new
descriptors are unlimited [14–16], and most probably, the number
of different descriptors will increase in the future [17–19].
However, it is not clear whether the increase in the number of
descriptors is useful from the point of view of practical applications
[20–22]. In this chapter, we discuss the possible future of different
concepts targeted to predict toxicity of substances, which are not
examined in direct experiments [23, 24].
2 Treatment of the Dependent Variable
The dependent variable to be predicted might refer to different

fields of toxicology [25–27]. In our context, toxicology means
some groups of different measurements of toxic effects in different
directions (endpoints), such as toxicity toward rats [28], Daphnia
magna [29], algae [30], honey bees [31], etc. The environmental
compartment generating the exposure associated to the toxic effects
may be very different (atmosphere, water, soil); therefore, the units
to express these effects can be expressed in different manners [32–
34]. In any case, endpoints related to toxic effect are characterized
by a large dispersion in measurement of toxicity for the same sub-
stance: (1) in experiments which are carried out in different labora-
tories and (2) even in experiments carried out in one laboratory.
Under such circumstances, the expression of corresponding experi-
mental data on various toxicities in the form of the negative decimal
logarithm is the widely used method [35, 36].
3 Types of Parameters Used for QSAR Modeling
Any QSAR model is aimed to predict some endpoint as a mathe-

matical function of molecular structure and/or physicochemical
parameters such as solubility and octanol–water partition coeffi-
cient [37]. However, the experimental data on physicochemical
parameters are available for a limited number of compounds. Thus,
the molecular descriptors, calculated on the basis of the molecular
Prediction of Biochemical Endpoints by the CORAL Software: Prejudices, Paradoxes… 575
structure, become an attractive alternative for physicochemical

descriptors [38, 39]. Software (such as CODESSA [40], DRAGON
[41], and TEST [42]) to calculate large groups of various descrip-
tors are available on the Internet. Besides traditional 2D and 3D
molecular descriptors, which are calculated on molecular graph
[40–43], descriptors calculation based on SMILES [44], SMART
[45], and InChI [46] are also available for users.
4 CORAL Software
The CORAL software is available on the Internet (http://www.

insilico.eu/coral). The basic principle of the software is building
up of a model for external validation set as function of the so-called
correlation weights of molecular features extracted from simplified
molecular input line entry systems (SMILES) [44]. In addition,
the software provides to the user the possibility to involve molecu-
lar features extracted from molecular graph. The correlation weight
for each molecular feature is calculated with Monte Carlo tech-
nique [1, 3, 6, 9–12, 19, 29, 35]. The simplest version of the
Monte Carlo calculation is the numerical data on the correlation
weights which provide maximal correlation coefficient between
experimental (observed) and calculated (predicted) values for an
examined endpoint. However, in addition, other versions of the
Monte Carlo optimization are available [1, 3, 6, 9–12, 19, 29, 35].
In particular, the so-called semicorrelations [19] give possibility to
build up binary classification models (i.e., prediction in form
“Active/Inactive”).
5 Prejudices
There are prejudices related to QSAR in general and to QSAR for

toxicity in particular.
5.1 Prejudices (a) The prejudice that QSAR is a tool to describe only more or less
Related to QSAR nonspecific effects. In other words, the opinion that results of
in General experiments related to toxic effects cannot be predicted with-
out direct experiments exists and remains very popular [47].
(b) Research works in the field of QSAR analysis are not

standardized.
(c) The statistical quality of a QSAR model can be an artifact.
5.2 Prejudices (a) Any kind of toxicity is a confused value depending upon many
Related to QSAR factors, some of which are unruled (weather conditions, health
for Toxicity status of organisms, etc.).
(b) The accuracy of a QSAR prediction for any toxicity is not com-
parable with the accuracy for corresponding experiment. The
modern paradigm believes that it is necessary to develop more
rapid tests. In addition, it is necessary to use primitive organ-
isms (bacteria) in order to avoid the ethical problems.
Unfortunately, this can lead to reducing of heuristic potential
of humanity in the field of toxicology [47].
6 Selection of Endpoints
6.1 Endpoint-1: Data on mutagenic potentials of the set of 95 aromatic and hetero-
Mutagenicity aromatic amines were taken from the literature [48]. The muta-
of Polyaromatic genic activity in Salmonella typhimurium TA98 + S9 microsomal
Amines reparation is expressed as the natural logarithm of R, where R is the
number of revertants per nanomole [49].
6.2 Endpoint-2: Numerical data on toxicity (oral LD50, mg/kg, mice) of psycho-
Toxicity of tropic drugs (phenothiazines, antidepressants, and anxiolitics)
Psychotropic Drugs were taken from the published work [35]. These values were con-
verted into pLD50, i.e., negative decimal logarithm of the LD50
expressed in mmol/kg.
6.3 Endpoint-3: Quantitative data (i.e., yes = 1 and no = −1) for inhibitory activity
Inhibitor Activity against monoamine oxidase B (human) for 1523 compounds were
Against Monoamine taken in the literature [50].
Oxidase B
7 QSAR
QSAR models for the three abovementioned endpoints have been

built up using the CORAL software [51]. For each endpoint a
CORAL method has been defined which is able to provide QSAR
with satisfactory predictive potential.
7.1 QSAR Model QSAR model for mutagenicity of polyaromatic amines character-
for Mutagenicity ized by strange situation: the statistical characteristics of the model
of Polyaromatic for training set are poorer than statistical characteristics for calibra-
Amines tion and validation set. However, this is the result of “natural”
process: building up model is stopped when the maximal quality
for the calibration set is reached (Fig. 1).
7.2 QSAR Model QSAR model for toxicity of psychotropic drugs is characterized by
for Toxicity usual situation: the statistical characteristics of the model for train-
of Psychotropic Drugs ing set are better than statistical characteristics for calibration and
validation set (Fig. 2).
Fig. 1 The CORAL method used for endpoint-1 and its statistical characteristics
Fig. 2 The CORAL method used for endpoint-2 and its statistical characteristics
Table 1
Statistical quality of binary classification of endpoint-3
Set N TP TN FP FN Sensitivity Specificity Accuracy MMC

Training 956 333 460 80 83 0.8005 0.8519 0.8295 0.6529
Calibration 283 104 129 20 30 0.7761 0.8658 0.8233 0.6459
Validation 284 105 126 26 27 0.7955 0.8289 0.8134 0.6247
Fig. 3 The CORAL method used for inhibitor activity against monoamine oxidase B and statistical characteris-
tics for this model in terms of semicorrelations
7.3 Binary Table 1 lists the statistical characteristics for binary classification of
Classification endpoint-3 (inhibitor activity against monoamine oxidase B).
for Inhibitor Activity Figure 3 shows the statistical quality of semicorrelation used to
Against Monoamine build up the model.
Oxidase B
8 Paradoxes
Paradoxically, the statistical characteristics of the suggested models

for mutagenicity for external validation set are better than the sta-
tistical characteristics for the training set. This effect has been
described and explained in published works [52, 53]. The majority
of the substances has an “average” behavior and is the basis for

building up a model. However, there are substances with “atypi-
cal” behavior. When the information contained in the substances
with the “average” behavior runs out then further improvement is
targeted for solely “atypical” substances, i.e., the predictive poten-
tial of the model for the molecules with “average” behavior is
decreasing [52, 53].
Paradoxically, the split into the training and validation sets has
an influence upon the statistical quality of a model, which is com-
parable with the effect of change of list of used descriptors.
9 Results
The suggested models have significant predictive potential. It is

important that the described models in spite of the differences in
the CORAL methods hold the possibility to provide to the user a
standard reporting format [54, 55]. The better statistical quality of
the abovementioned models for the external validation set (as well
as for the calibration set) confirms the hypothesis suggested in the
literature that overtraining is tuning up a model for molecules with
untypical behavior, whereas molecules with ordinary behavior
obtain unsuitable description [53].
10 Possible Ways to Improve the Monte Carlo Technique Used in the CORAL
1. Development of target function of the optimization procedure

[3].
2. Definition of new SMILES attributes and/or new graph invari-
ants [6, 10, 11].
3. Use of new criteria of predictive potential of a model [30].
11 Conclusions
The CORAL software gives QSAR models according to OECD

principles. The software is available on the Internet (http://www.
insilico.eu/coral). The software can be successfully used to solve
various tasks of the QSPR/QSAR practice [10, 56–71]. The soft-
ware can be improved.
Acknowledgments
The authors thank the project EU LIFE-COMBASE (contract

LIFE15 ENV/ES/000416) for financial support.
References
1. Toropov AA, Toropova AP, Lombardo A, prediction by means of the Monte Carlo
Roncaglioni A, Benfenati E, Gini G (2011) method. Ecotoxicol Environ Saf 133:390–394
CORAL: building up the model for biocon- 12. Toropova AP, Toropov AA (2017) CORAL:
centration factor and defining it’s applicability binary classifications (active/inactive) for drug-
domain. Eur J Med Chem 46:1400–1403 induced liver injury. Toxicol Lett 268:51–57
2. Roncaglioni A, Toropov AA, Toropova AP, 13. Park H-G, Yeo M-K (2013) Ecotoxicity esti-
Benfenati E (2013) In silico methods to pre- mation of hazardous air pollutants emitted
dict drug toxicity. Curr Opin Pharmacol from semiconductor manufacturing processes
13:802–806 utilizing QSAR. Bull Kor Chem Soc
3. Toropova AP, Toropov AA, Lombardo A, 34(12):3755–3761
Roncaglioni A, Benfenati E, Gini G (2012) 14. Tong J, Li L, Bai M, Li K (2017) A new
CORAL: QSAR models for acute toxicity in descriptor of amino acids-SVGER and its appli-
fathead minnow (Pimephales promelas). cations in peptide QSAR. Mol Inform
J Comput Chem 33:1218–1223 36(5):1501023
4. Schleifer K-J (2013) Computational approaches 15. Algamal ZY, Lee MH (2017) A new adaptive
in agricultural research. Chapter 2, In book: L1-norm for optimal descriptor selection of
Jeschke P, Krämer W, Schirmer U, Witschel M high-dimensional QSAR classification model
(eds) Modern methods in crop protection for anti-hepatitis C virus activity of thiourea
research, pp 21–41. Wiley-VCH: Verlag&C0. derivatives. SAR QSAR Environ Res 28:75–90
KGaA, Boschstr.12, 69469 Weinheim, 16. Bigdeli A, Hormozi-Nezhad MR, Jalali-Heravi
Germany. Print ISBN: 978-3-527- 33175-8 M, Abedini MR, Sharif-Bakhtiar F (2014)
ePub ISBN: 978-3-527-65592-2 Towards defining new nano-descriptors:
5. Ghorbanzadeh M, Zhang J, Andersson PL extracting morphological features from trans-
(2016) Binary classification model to predict mission electron microscopy images. RSC Adv
developmental toxicity of industrial chemicals 4:60135–60143
in zebrafish. J Chemom 30:298–307 17. Masand VH, Rastija V (2017) PyDescriptor: a
6. Toropov AA, Toropova AP, Marzo M, Dorne new PyMOL plugin for calculating thousands
JL, Georgiadis N, Benfenati E (2017) QSAR of easily understandable molecular descriptors.
models for predicting acute toxicity of pesti- Chemom Intell Lab Syst 169:12–18
cides in rainbow trout using the CORAL soft- 18. Basak SC (2017) The expanding landscape of
ware and EFSA’s OpenFoodTox database. graph theoretic molecular descriptors: develop-
Environ Toxicol Pharmacol 53:158–163 ment, gradual diversification of descriptor
7. Hisaki T, Kaneko MAN, Yamaguchi M, Sasa space, and applications in QSAR/QMSA and
H, Kouzuki H (2015) Development of qsar new drug discovery. Curr Comput Aided Drug
models using artificial neural network analysis Des 13:172–176
for risk assessment of repeated-dose, reproduc- 19. Cassano A, Robinson RLM, Palczewska A,
tive, and developmental toxicities of cosmetic Puzyn T, Gajewicz A, Tran L, Manganelli S,
ingredients. J Toxicol Sci 40:163–180 Cronin MTD (2016) Comparing the CORAL
8. Burello E (2017) Review of (Q)SAR models and random forest approaches for modelling
for regulatory assessment of nanomaterials the in vitro cytotoxicity of silica nanomaterials.
risks. NanoImpact 8:48–58 ATLA Altern Lab Anim 44:533–556
9. Gobbi M, Beeg M, Toropova MA, Toropov 20. Cronin MTD, Schultz TW (2003) Pitfalls in
AA, Salmona M (2016) Monte Carlo method QSAR. J Mol Struct THEOCHEM
for predicting of cardiac toxicity: hERG blocker 622:39–51
compounds. Toxicol Lett 250–251:42–46 21. Schultz TW, Cronin MTD, Netzeva TI (2003)
10. Sokolović D, Aleksić D, Milenković V, Karaleić The present status of QSAR in toxicology.
S, Mitić D, Kocić J, Mekić B, Veselinović JB, J Mol Struct THEOCHEM 622:23–38
Veselinović AM (2016) QSAR modeling of 22. Doweyko AM (2008) QSAR: dead or alive?
bis-quinolinium and bis-isoquinolinium com- J Comput Aided Mol Des 2:81–89
pounds as acetylcholine esterase inhibitors
based on the Monte Carlo method—the impli- 23. Gajewicz A (2017) What if the number of
cation for Myasthenia gravis treatment. Med nanotoxicity data is too small for developing
Chem Res 25:2989–2998 predictive Nano-QSAR models? An alternative
read-across based approach for filling data
11. Toropov AA, Toropova AP, Cappellini L, gaps. Nanoscale 9:8435–8448
Benfenati E, Davoli E (2016) Odor threshold
24. van Leeuwen K, Schultz TW, Henry T, tionships. Curr Comput Aided Drug Des
Diderich B, Veith GD (2009) Using chemical 12:181–250
categories to fill data gaps in hazard assess- 37. Raevsky OA, Razdolskii AN, Liplavskii YV,
ment. SAR QSAR Environ Res 20:207–220 Raevskaya OE, Yarkov AV (2012) Molecular-
25. Auerbach M, Macdougall I (2017) The avail- biological problems of drug design and mecha-
able intravenous iron formulations: history, nism of drug action: acute toxicity evaluation
efficacy, and toxicology. Hemodial Int upon intravenous injection into mice: interspe-
21:S83–S92 cies correlations, lipophilicity parameters, and
26. Campbell ND (2016) Behavior within fortu- physicochemical descriptors. Pharm Chem
itous environments: the entwined history of J 46:69–74
division 28 and the fields of behavioral pharma- 38. Furtula B, Gutman I (2011) Relation between
cology and toxicology. Exp Clin second and third geometric-arithmetic indices
Psychopharmacol 24:209–213 of trees. J Chemom 25:87–91
27. Satoh T (2016) History of Japanese society of 39. Mercader A, Castro EA, Toropov AA (2001)
toxicology. J Toxicol Sci 41:SP1–SP9 Calculation of total molecular electronic ener-
28. Toropov AA, Rasulev BF, Leszczynski J (2007) gies from correlation weighting of local graph
QSAR modeling of acute toxicity for nitroben- invariants. J Mol Model 7:1–5
zene derivatives towards rats: comparative anal- 40. Morrill JA, Topczewski JJ, Lodge AM, Yasapala
ysis by MLRA and optimal descriptors. QSAR N, Quinn DM (2015) Development of quanti-
Comb Sci 26:686–693 tative structure activity relationships for the
29. Toropova AP, Toropov AA, Martyanov SE, binding affinity of methoxypyridinium cations
Benfenati E, Gini G, Leszczynska D, for human acetylcholinesterase. J Mol Graph
Leszczynski J (2012) CORAL: QSAR model- Model 62:181–189
ing of toxicity of organic chemicals towards 41. Mansouri K, Consonni V, Durjava MK, Kolar
Daphnia magna. Chemom Intell Lab Syst B, Öberg T, Todeschini R (2012) Assessing
110:177–181 bioaccumulation of polybrominated diphenyl
30. Gramatica P, Cassani S, Roy PP, Kovarich S, ethers for aquatic species by QSAR modeling.
Yap CW, Papa E (2012) QSAR modeling is not Chemosphere 89:433–444
“push a button and find a correlation”: a case 42. Diaza RG, Manganelli S, Esposito A,
study of toxicity of (Benzo-)triazoles on algae. Roncaglioni A, Manganaro A, Benfenati E
Mol Inform 31:817–835 (2015) Comparison of in silico tools for evalu-
31. Toropov AA, Benfenati E (2007) SMILES as ating rat oral acute toxicity. SAR QSAR
an alternative to the graph in QSAR modelling Environ Res 26:1–27
of bee toxicity. Comput Biol Chem 31:57–60 43. Basant N, Gupta S (2017) QSAR modeling for
32. Wang X, Greene N (2012) Comparing mea- predicting mutagenic toxicity of diverse chemi-
sures of promiscuity and exploring their rela- cals for regulatory purposes. Environ Sci Pollut
tionship to toxicity. Mol Inform 31:145–159 Res 24:14430–14444
33. Venkatapathy R, Wang CY, Bruce RM, 44. Toropov AA, Benfenati E (2008) Additive
Moudgal C (2009) Development of quantita- SMILES-based optimal descriptors in QSAR
tive structure-activity relationship (QSAR) modelling bee toxicity: using rare SMILES
models to predict the carcinogenic potency of attributes to define the applicability domain.
chemicals. I. Alternative toxicity measures as an Bioorg Med Chem 16:4801–4809
estimator of carcinogenic potency. Toxicol 45. Toropov AA, Toropova AP, Benfenati E,
Appl Pharmacol 234:209–221 Manganaro A (2009) QSPR modelling of
34. Gironés X, Carbó-Dorca R (2006) Modelling enthalpies of formation for organometallic
toxicity using molecular quantum similarity compounds by SMART-based optimal descrip-
measures. QSAR Comb Sci 25:579–589 tors. J Comput Chem 30:2576–2582
35. Gissi A, Toropov AA, Toropova AP, Nicolotti 46. Toropov AA, Toropova AP, Benfenati E (2010)
O, Carotti A, Benfenati E (2014) Building up QSAR-modeling of toxicity of organometallic
QSAR model for toxicity of psychotropic drugs compounds by means of the balance of correla-
by the Monte Carlo method. Struct Chem tions for InChI-based optimal descriptors. Mol
25:1067–1073 Divers 14:183–192
36. Sahoo S, Adhikari C, Kuanar M, Mishra BK 47. Rotini A, Manfra L, Spanu F, Pisapia M, Cicero
(2016) A short review of the generation of AM, Migliore L (2017) Ecotoxicological
molecular descriptors and their applications in method with marine bacteria Vibrio anguilla-
quantitative structure property/activity rela- rum to evaluate the acute toxicity of environ-
mental contaminants. J Vis Exp 123:e55211
48. Toropov AA, Toropova AP, Martyanov SE, ture-activity relationships in drug design, pre-
Benfenati E, Gini G, Leszczynska D, dictive toxicology, and risk assessment. IGI
Leszczynski J (2011) Comparison of SMILES Global, Hershey, PA, pp 560–585
and molecular graphs as the representation of 59. Rescifina A, Floresta G, Marrazzo A, Parenti C,
the molecular structure for QSAR analysis for Prezzavento O, Nastasi G, Dichiara M, Amata
mutagenic potential of polyaromatic amines. E (2017) Development of a Sigma-2 receptor
Chemom Intell Lab Syst 109:94–100 affinity filter through a Monte Carlo based
49. Cash GG (2001) Prediction of the genotoxic- QSAR analysis. Eur J Pharm Sci 106:94–101
ity of aromatic and heteroaromatic amines 60. Rescifina A, Floresta G, Marrazzo A, Parenti C,
using electrotopological state indices. Mutat Prezzavento O, Nastasi G, Dichiara M, Amata
Res Genet Toxicol Environ Mutagen E (2017) Sigma-2 receptor ligands QSAR
491:31–37 model dataset. Data Brief 13:514–535
50. Lorenzo VP, Filho JMB, Scotti L, Scotti MT 61. Toropova MA, Raska I, Toropova AP, Raskova
(2015) Combined structure- and ligand-based M (2017) CORAL software: analysis of impacts
virtual screening to evaluate caulerpin analogs of pharmaceutical agents upon metabolism via
with potential inhibitory activity against mono- the optimal descriptors. Curr Drug Metab
amine oxidase B. Braz J Pharmacogn 18:500–510
25:690–697 62. Sokolović D, Ranković J, Stanković V,
51. IRFMN (2017) http://www.insilico.eu/coral. Stefanović R, Karaleić S, Mekić B, Milenković
Accessed 14 Sept 2017 V, Kocić J, Veselinović AM (2017) QSAR
52. Toropova AP, Toropov AA, Diaza RG, study of dipeptidyl peptidase-4 inhibitors based
Benfenati E, Gini G (2011) Analysis of the co- on the Monte Carlo method. Med Chem Res
evolutions of correlations as a tool for QSAR- 26:796–804
modeling of carcinogenicity: an unexpected 63. Kumar A, Chauhan S (2017) QSAR differen-
good prediction based on a model that seems tial model for prediction of SIRT1 modulation
untrustworthy. Cent Eur J Chem 9:165–174 using Monte Carlo method. Drug Res
53. Toropova AP, Toropov AA, Benfenati E, Gini 67:156–162
G (2011) Co-evolutions of correlations for 64. Kumar A, Chauhan S (2017) Use of the Monte
QSAR of toxicity of organometallic and inor- Carlo method for OECD principles-guided
ganic substances: an unexpected good predic- QSAR modeling of SIRT1 inhibitors. Arch
tion based on a model that seems untrustworthy. Pharm (Weinheim) 350(1):e1600268
Chemom Intell Lab Syst 105:215–219 65. Sokolović D, Stanković V, Toskić D, Lilić L,
54. Pajares F, Hartley J, Valiante G (2001) Ranković G, Ranković J, Nedin-Ranković G,
Response format in writing self-efficacy assess- Veselinović AM (2016) Monte Carlo-based
ment: greater discrimination increases predic- QSAR modeling of dimeric pyridinium com-
tion. Meas Eval Couns Dev 33:214–221 pounds and drug design of new potent acetyl-
55. Goodarzi M, Freitas MP, Ferreira EB (2009) choline esterase inhibitors for potential therapy
Influence of changes in 2-D chemical structure of myasthenia gravis. Struct Chem
drawings and image formats on the prediction 27:1511–1519
of biological properties using MIA- 66. Aranda JF, Garro Martinez JC, Castro EA,
QSAR. QSAR Comb Sci 28:458–464 Duchowicz PR (2016) Conformation-
56. Achary PGR (2014) QSPR modelling of independent QSPR approach for the soil sorp-
dielectric constants of π-conjugated organic tion coefficient of heterogeneous compounds.
compounds by means of the CORAL software. Int J Mol Sci 17(8):1247
SAR QSAR Environ Res 25:507–526 67. Islam MA, Pillay TS (2016) Simplified molecu-
57. Veselinović AM, Milosavljević JB, Toropov AA, lar input line entry system-based descriptors in
Nikolić GM (2013) SMILES-based QSAR QSAR modeling for HIV-protease inhibitors.
model for arylpiperazines as high-affinity Chemom Intell Lab Syst 153:67–74
5-HT1A receptor ligands using CORAL. Eur 68. Ghaedi A (2015) Predicting the cytotoxicity of
J Pharm Sci 48:532–541 ionic liquids using QSAR model based on
58. Toropov AA, Toropova AP, Benfenati E, SMILES optimal descriptors. J Mol Liq
Nicolotti O, Carotti A, Nesmerak K, Veselinovic 208:269–279
AM, Veselinovic JB, Duchowicz PR, Bacelo 69. Fioressi SE, Bacelo DE, Cui WP, Saavedra LM,
DE, Castro EA, Rasulev BF, Leszczynska D, Duchowicz PR (2015) QSPR study on refrac-
Leszczynski J (2015) QSPR/QSAR analyses tive indices of solvents commonly used in poly-
by means of the CORAL software: results, mer chemistry using flexible molecular
challenges, perspectives. In: Quantitative struc-
descriptors. SAR QSAR Environ Res

71.
Worachartcheewan A, Nantasenamat C,
26:499–506 Isarankura-Na-Ayudhya C, Prachayasittikul V
70. Li Q, Ding X, Si H, Gao H (2014) QSAR (2014) QSAR study of H1N1 neuraminidase
model based on SMILES of inhibitory rate of inhibitors from influenza a virus. Lett Drug
2,3-diarylpropenoic acids on AKR1C3. Des Discov 11:420–427
Chemom Intell Lab Syst 139:132–138
Index
A Computational filter�� 238, 328

Computational modelling��10, 11, 128, 235, 240, 539
Acceptability��91, 96, 159, 453, 456, 466, 467 Computational toxicology��4, 7, 56, 74,
Active pharmaceutical ingredients (APIs)�� 235, 348, 119–133, 171–178, 233–242, 249, 275–284, 287–310,
351–353, 395, 397, 399, 405, 410, 416, 420 428, 460, 470, 521, 537, 553
Acute toxicity�� 129, 131, 360, 380, 400, Confidence��69, 86, 113, 141–167, 199,
402–404, 422, 423, 434, 435, 519–533 221, 241, 456, 496, 545
Additive��69, 79, 112, 220–223, 226, Conformational space��185, 248–258, 541
329, 384, 397, 415 Contaminant��8, 220, 226, 227, 368,
Adverse Outcome Pathway (AOP)�� 61–65, 221, 375, 381–384, 396
224–226, 384, 485, 539 Contaminants of emerging concern (CEC)��396
Algorithmic complexity��550 CORAL software��573–579
Alternative methods�� 63, 65–66, 72, 107,
419–420, 447, 476, 496 D
Alternative testing�� 7, 65, 73, 111, 126,
Data gap filling�� 58, 61–63, 66, 68, 71, 75
127, 171, 436, 462, 465, 488
Deep learning (DL)��84, 128–130, 132, 542
Applicability domain (AD)��7, 41, 58, 67, 80,
Developmental effect��278
85, 88, 111, 113–115, 141–167, 182, 187–189, 195,
Drug��13, 80, 123, 141, 171, 182, 233, 246,
200–203, 205, 207, 303, 306, 334, 335, 419, 437, 452,
275, 287, 313, 327, 347, 396, 454, 475, 506, 535, 573
453, 456, 486, 490, 497, 501, 531, 532, 564–566, 569
Drug delivery��347–361, 411, 559
B Drug discovery�� 7, 13, 130, 132, 173, 181, 182,
184, 233–242, 275, 277, 281, 284, 287–310, 313–324,
Bartter syndrome (BS)�� 314–317, 323 327–342, 521, 541, 545, 574
Benchmark dose (BMD)��221–223 Drug inhibitor activity��576
Binding space�� 249–252, 258–271 Drugs of abuse�� 368, 381
Bioconcentration factor (BCF)��176, 200–205, Drug toxicity�� 538, 539, 545
207, 209, 211, 213, 215, 433
E
C
Ecotoxicity�� 59, 129, 398–407, 409–412,
Carcinogenicity�� 62, 111, 115, 121–123, 414, 418–420, 422–427, 430, 432, 437, 438
126, 127, 129, 223, 234, 237, 242, 278, 371, 381, 412, Efficacy�� 233, 242, 314, 316, 335, 347–361,
433, 447–470, 536, 537 368, 377, 384, 387, 536
Cardiotoxicity�� 125, 234, 238, 241, 320, Embryonic stem cell��475
373, 379, 475 Endocrine potential��181
Chemical category�� 58, 63, 67, 69, 73, 171, 458, 526 Endpoints�� 5, 58, 79, 109, 119, 167, 172–174,
Chemical similarity��35, 60, 152, 158, 172–176, 176, 181, 200, 234, 278, 398, 448, 519, 537, 560, 573
178, 436, 470, 526, 537, 538 Ensemble docking simulations�� 259, 264–267
Chemical space�� 7, 11, 12, 41, 43, 45, 80, Entropic factors�� 253, 258, 266, 267, 271
89, 143, 182, 309, 336 EU regulation�� 465, 467
Chemical toxicity prediction�� 475–482, 522 European Chemicals Agency (ECHA)��57, 61, 86,
Chemoinformatics�� 22, 128, 131, 132, 108, 109, 114, 199, 416, 458, 459, 461, 465–467,
287–310, 559–569 521–523, 525–527, 530–532
Classification model�� 14, 15, 114, 120, 121, Expert systems��61, 63, 119, 126, 171, 398,
123, 128, 182, 183, 187, 188, 191, 193, 195, 455, 575 407, 412, 424, 432–438, 454–455, 537
https://doi.org/10.1007/978-1-4939-7899-1, © Springer Science+Business Media, LLC, part of Springer Nature 2018
585
Computational Toxicology: Methods and Protocols
586 Index

Exposures�� 74, 108, 109, 114, 219, 221–225, M
227, 235, 314, 341, 358–360, 376, 379, 381–383, 387,
403–405, 407, 408, 412–414, 416, 435, 437, 461, 468, Machine learning��14, 38, 48, 80, 88, 96,
469, 477, 480, 489, 494, 519, 520, 536, 538, 539, 574 119–133, 154, 158, 173, 174, 194, 276, 301, 452, 475,
497, 541, 542, 550–552
F Mathematical chemistry�� 60, 161, 434, 458, 560
Metals�� 185, 283, 334, 335, 352, 357, 359,
Food safety��86, 219–227, 429, 467
368, 381, 425, 427, 486, 488
G Molecular descriptor�� 3, 6, 16, 23, 35, 42,
79, 87, 89, 120, 122, 125–129, 172–175, 178, 187, 288,
Gene network��476, 478, 479, 515 298, 299, 301, 304, 357, 423, 424, 476, 479, 529, 537,
Gene regulation��509 552, 560–562, 574
Genotoxicity��61, 122, 125, 126, 225, 237, Molecular docking��119, 181–195, 315, 318, 332, 335
238, 372, 373, 379, 412, 447–470 Molecular similarity�� 10, 34, 35, 38, 171–178,
330, 522, 523, 525, 526, 531, 532
H Molecular toxicity��278
Hepatotoxicity�� 122, 123, 125, 127, 240, Monte Carlo technique�� 257, 575, 579
249, 372–377, 380, 381, 383, 505–518 Multi-kernel support vector machine�� 476, 481
High-throughput screening (HTS)��56, 74, 133, Multitarget��327–342, 386, 543
183, 236, 275–284, 287, 314, 328, 407, 530 Multivariable explorative technique��559
Hit identification��236–237 Mutagenicity��57, 60, 121, 125–129, 131,
Human health��58, 62, 70, 107, 110, 111, 240, 278, 371–373, 412, 433, 448–450, 453–457, 459,
220, 382, 386, 395, 409, 412, 415, 418, 420, 429, 432, 460, 462, 470, 537, 576, 578
447, 520 Mycotoxins�� 368, 381, 383
I N
In chemico�� 65, 67, 69, 72, 73, Nano-formulation��347–361
453, 485–501 Nanotechnology�� 357, 360
Induction��90–95, 99, 100, 103, 222, 303, Network pharmacology��544
358, 359, 448, 489, 494–496, 499, 500 Neural networks�� 84, 120, 128–132, 154,
In silico��55, 60, 62, 65, 66, 72, 73, 142, 145, 435, 530, 532, 542
171, 199, 201, 215, 234, 235, 241, 242, 253, 255, 277, Non-testing methods (NTM)��86, 172, 199–215,
288, 291, 300, 303, 306, 307, 314, 320, 321, 323, 324, 448, 465, 467
340, 347, 384, 398, 407, 412, 418–424, 438, 447, 451, Nutraceuticals��367, 369–370, 372–374
454, 455, 468, 470, 498–501, 519–533, 537, 538, 541
In vitro��56, 59, 63, 65–67, 69, 72–74, 90, 102,
O
109, 111, 183, 224, 225, 233–239, 241, 242, 300, 315, OECD QSAR Toolbox��55–75, 455, 458,
316, 320, 321, 323, 336, 342, 347–349, 356–358, 377, 460–465, 522
379, 384, 385, 402, 404, 407, 418, 448–450, 453, 454, Omics�� 56, 384, 517, 538, 539
457, 459, 462, 465, 467, 476, 485–501, 521, 541, 544
Ion channels�� 238, 313–324 P
Ionic liquids (ILs)�� 425–427, 559–569
Patch clamp��315, 316, 319, 321
K Pesticides�� 129, 142, 220, 221, 224–226, 242,
368, 381, 382, 413, 415, 426, 430, 461, 464, 467, 469
Key event (KE)�� 62–64, 72, 132, 224–226, Pharmaceuticals�� 7, 80, 159, 233, 275,
485–488, 491–494, 498 291, 324, 330, 347, 367, 395, 407, 424, 469, 536
Pharmacovigilance��315–320
L Plant alkaloids��368
Lead identification��237–240 Point of departure�� 222, 227
Lead optimization��12, 182, 235, 237–240 Predictive modeling�� 81, 89, 91, 92, 97, 101, 128, 129,
Library design��275–284 141, 182, 253, 255, 258, 259, 262–264, 266–268, 288,
Ligand mobility��254–256, 258, 291, 300–304, 306, 309, 310, 384, 419, 434, 550, 579
259, 268–271 Predictive toxicology��73, 74, 127, 181–195, 357, 459
Computational Toxicology: Methods and Protocols
587
Index

Property space��236, 250–259, 271 Skin sensitization�� 60, 64, 65, 109, 110,
Psychotropic drug toxicity��576–578 485, 486, 488, 491–494, 496–501
Structure-activity��3, 6, 55, 56, 109, 119,
Q 120, 141, 171, 235, 242, 271, 288, 300, 319, 333, 357,
QSAR models��11, 58, 80, 114, 119, 141, 397, 412, 423, 433, 435, 448, 560
172, 174, 199, 235, 288, 357, 398, 452, 515, 520, 537, Support vector machines (SVM)�� 84, 101, 125, 126,
560, 574 130, 132, 158, 173, 475, 478, 479, 481, 482, 542, 543
Quantitative structure-activity relationship Systems biology��56, 407, 505–518,
(QSAR)��4, 55, 79, 119, 141, 171, 538–540, 552
172, 174, 199, 235, 271, 288, 328, 357, 397, 451, 482,
T
515, 520, 537, 560, 573
Toxicity�� 4, 56, 80, 110, 119, 171, 200, 220,
R 233, 246, 275, 300, 314, 327, 340, 347, 370, 380, 396,
Random forest (RF)�� 101, 127, 128, 130–132, 451, 475, 512, 519, 560, 574
173, 301, 305, 307, 331, 475, 498, 530, 542, 543, 552 Toxicity prediction�� 7, 130, 131, 133, 171,
Read-across�� 7, 55, 58–63, 66–68, 72, 109–111, 278, 356, 398, 417–421, 426, 432, 436, 438, 454–455,
171, 172, 174, 176–177, 199, 201–203, 207, 213, 237, 465, 505–518, 537, 542, 543, 545, 552
419, 436, 458, 465, 467, 520, 522–527, 530, 532 Toxicity ranking index of cations (TRIC)�� 560,
Registration, Evaluation, Authorisation and Restriction of 562–563, 566–569
Chemicals (REACH)��80, 86, 107–115, Toxicity testing models��536
129, 414, 416, 430, 451, 454, 465–467, 520, 521, 532 Toxicology�� 7, 56, 81, 119, 171, 181, 226, 234,
Reliability�� 5, 41, 63, 69, 71, 90, 112, 113, 247, 278, 298, 323, 340, 360, 370, 459, 535–553, 574
142–144, 158, 159, 161, 165, 167, 183, 186, 195, 200, Toxic potential��368, 370–380, 386,
201, 250, 255, 258, 261, 268, 270, 332, 356, 527, 528, 387, 421, 463, 574
530–532, 540, 564
V
Risk management��108, 405, 409, 410, 416, 432
Validation��13, 14, 16, 60, 66, 67, 72, 80,
S 86–88, 95–102, 125, 141–143, 176, 182–184, 188, 189,
Safe nutraceuticals�� 380, 382, 384 275–284, 288, 301, 307, 315, 321, 419, 423, 453, 458,
Safety��55, 56, 59, 112, 142, 167, 211, 219–227, 459, 476, 488, 490, 492–498, 520, 525, 526, 529, 530,
233–235, 238, 240, 242, 275, 313, 340, 347–361, 368, 575, 576, 578, 579
380, 382, 384, 387, 408–410, 412, 415, 416, 418, 462, Virtual screening�� 86, 236, 250, 277, 288,
463, 467, 468, 522, 536, 545 291, 302, 306, 309, 330, 337, 516, 541
Similarity searching�� 70, 238, 277
W
Simplified Molecular Input Line Entry System
(SMILES)��13, 17–19, 21–23, 46, 69, Waste management��405
84, 112, 183, 280, 281, 289–294, 296, 297, 302, 303, Weight of evidence (WoE)��63, 65–69, 109–112,
306, 478, 550, 551, 561, 567, 575, 579 115, 199–215, 237, 451, 465, 468, 520, 521, 533

Computational Toxicology: Methods and Protocols

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Computational Toxicology: Methods and Protocols

Uploaded by

Copyright:

Available Formats

Methods in

Molecular Biology 1800

Orazio Nicolotti Editor

For further volumes:

Methods and Protocols

ISSN 1064-3745 ISSN 1940-6029 (electronic)

Library of Congress Control Number: 2018943843

© Springer Science+Business Media, LLC, part of Springer Nature 2018

Printed on acid-free paper

To Chiara and Maria Giovanna

Bari, Italy Orazio Nicolotti

Part I Where We Are and Where We Are Going To

Part II Molecular and Data Modeling

Part III Impact in Drug Discovery and Development

Part IV Predicting Human Health Toxicology Endpoints

25 Predictive Systems Toxicology���������������������������������������������������������������������������� 535

Antreas Afantitis • NovaMechanics Ltd, Nicosia, Cyprus

Domenico Gadaleta • IRCCS–Istituto di Ricerche Farmacologiche Mario Negri,

Ovanes G. Mekenyan • Laboratory of Mathematical Chemistry (LMC), As. Zlatarov

Anita Sosnowska • Laboratory of Environmental Chemometrics, Faculty of Chemistry,

Where We Are and Where We Are Going To

Molecular Descriptors for Structure–Activity Applications:

Many toxicological applications rely on the principle that the bio-

Francesca Grisoni and Viviana Consonni contributed equally to this chapter.

any alteration in the chemical constitution (ΔC), there corresponds

where P is the biological/physicochemical property of a com-

structure–activity relationship) or QSPR (quantitative structure–

(b) 1-Dimensional (1D). According to this representation, mol-

(e) 4-Dimensional (4D). In addition to the molecular geometry,

molecular structure in a series of binary digits (bits) that represent

rolling hills, rugged canyons [91] or can be an intermediate sce-

most of the modeling and descriptor-based protocols. In addition,

For the guided example, we considered two QSAR models to pre-

2.1 Endpoint Cytochromes P450 (CYP) are a family of monooxygenase enzymes

set for model selection. The final models were recalibrated on

compounds (i.e., the nearest neighbors) with known experimental

Fitting Validation All data

Mol5 3-bromophenyl nicotinate

Mol6 2-(9H-xanthen-9-yl)benzoic acid

native software exist, such as E-Dragon [111] (see Note 2).

ID SMILES Class Class code nBO nBM C% …

Fig. 6 Items provided in the MATLAB file (“CYP3A4_dataset.mat”), available at

In this section, we present a step-by-step tutorial to apply the previ-

3.1 Structure To make molecules machine-readable and calculate molecular

molecular graph that also includes hydrogens is called H-filled

(e) The topological distance, which is the length of the shortest

the listed neighbors are arranged anticlockwise, while “@@”

ID nBM nBO C% SIC3 ATSC4p ATSC6i nROH NaasC

data management. SMILES can be easily generated starting from

information about its molecular geometry and overall topology.

ment of individual entities in space and the spatial relationships

3.2.3 Indices Indices of neighborhood symmetry are topological informa-

with a through-bond distance from the considered vertex

SICm is normalized according to a function of the number of

where X is the numerical matrix containing the descriptor values,

data-pretreatment and Euclidean distance calculation [132], as

In the case of new molecules to be predicted, their descriptors

the jth descriptor in the training set of molecules (as in Eq. 11),

3.4.2 Distance Molecular similarity is often expressed through a distance measure

Original data Range-scaled data

*Training set parameters used for range-scaling

Scaled descriptor values

ID 2D Structure C% SIC3 ATSC4p ATSC6i nROH NaaSC

Bari, Italy Orazio Nicolotti

Part I Where We Are and Where We Are Going To

Part II Molecular and Data Modeling

Part III Impact in Drug Discovery and Development

Part IV Predicting Human Health Toxicology Endpoints

25 Predictive Systems Toxicology�� 535

where X_scal_new is the MATLAB array containing the range-