Environmental Toxicology and Chemistry, Vol. 35, No. 11, pp.

2691–2697, 2016
# 2016 SETAC
Printed in the USA

MONTE CARLO–BASED QUANTITATIVE STRUCTURE–ACTIVITY RELATIONSHIP MODELS

FOR TOXICITY OF ORGANIC CHEMICALS TO DAPHNIA MAGNA

ALLA P. TOROPOVA,*y ANDREY A. TOROPOV,y ALEKSANDAR M. VESELINOVIC
,z JOVANA B. VESELINOVIC
,z

DANUTA LESZCZYNSKA,x and JERZY LESZCZYNSKIk
yIRCCS-Istituto di Ricerche Farmacologiche Mario Negri, Milan, Italy
zUniversity of Nis, Faculty of Medicine, Department of Chemistry, Nis, Serbia
xInterdisciplinary Nanotoxicity Center, Department of Civil and Environmental Engineering, Jackson State University, Jackson, Mississippi, USA
kInterdisciplinary Nanotoxicity Center, Department of Chemistry and Biochemistry, Jackson State University, Jackson, Mississippi, USA

(Submitted 29 March 2016; Returned for Revision 18 April 2016; Accepted 21 April 2016)

Abstract: Quantitative structure–activity relationships (QSARs) for toxicity of a large set of 758 organic compounds to Daphnia magna
were built up. The simplified molecular input-line entry system (SMILES) was used to represent the molecular structure. The Correlation
and Logic (CORAL) software was utilized as a tool to develop the QSAR models. These models are built up using the Monte Carlo
method and according to the principle “QSAR is a random event” if one checks a group of random distributions in the visible training set
and the invisible validation set. Three distributions of the data into the visible training, calibration, and invisible validation sets are
examined. The predictive potentials (i.e., statistical characteristics for the invisible validation set of the best model) are as follows: n ¼ 87,
r2 ¼ 0.8377, root mean square error ¼ 0.564. The mechanistic interpretations and the domain of applicability of built models are
suggested and discussed. Environ Toxicol Chem 2016;35:2691–2697. # 2016 SETAC

Keywords: Computational toxicology Ecological risk assessment Environmental toxicology Aquatic toxicology Organic
contaminant

INTRODUCTION chemistry approaches using such methodology are referred to

Modern civilization has spawned a new and great concern
for nature because massive productions of hazardous chemicals
and pollutants have an enormous effect on the ecosystem. The
global production of chemicals has increased from 1 million tons
in 1930 to 400 million tons in the 21st century. In addition, every
year newly synthesized compounds are introduced into the
ecosystem. Aquatic ecological systems are important compo-
nents of the natural world. Their pollution could introduce
harmful and difficult-to-correct environmental outcomes.
Unfortunately, relatively few chemical compounds have been
subjected to adequate assessment for their perilous environmen-
tal properties [1]. Prevention is always a more efficient remedy
than restoration. Such an approach involves a priori evaluation
of environmental effects of considered classes of compounds.
For risk assessment, Daphnia magna, an important
freshwater invertebrate species in aquatic food webs, has
been used worldwide for many years as a representative test
species for the ecotoxicological evaluation of industrial
chemicals. Experimental investigations have been carried out
to address this issue; however, this approach is expensive and
time-consuming. An alternative involves development of
theoretical (computational) methods to estimate toxic end-
points. Such methods could be applied for substances, which
have not been examined experimentally. They use the available
experimental data on these endpoints for compounds belonging
to the same group of chemicals [2–11]. Computational
This article includes online-only Supplemental Data.
* Address correspondence to alla.toropova@marionegri.it
Published online 25 April 2016 in Wiley Online Library
(wileyonlinelibrary.com).
DOI: 10.1002/etc.3466
2691
as quantitative structure–property relationships (QSPRs) and
quantitative structure–activity relationships (QSARs) [12,13].
It is to be noted that the QSAR technique leads to a significant
reduction of the use of test animals; this reduction is
recommended by the Registration, Evaluation, Authorisation,
and Restriction of Chemicals (REACH) regulation [14].
In QSAR modeling defining molecular descriptors is 1 of the
most important steps. There are various molecular descriptors
calculated from different molecular features including 1 based
on a molecular graph [15]. An attractive alternative for the
representation of the molecular structure by graph is the
simplified molecular input-line entry system (SMILES) [16–18].
The SMILES notation can be used for calculating the optimal
descriptor, a molecular descriptor that depends on both the
molecular structure and the property under analysis but does not
explicitly depend on details from the molecular 3-dimensional
geometry. For this reason the development of QSAR models
where molecular descriptors based on SMILES notation are
used is an attractive direction of research in the field of QSPR
theory and applications because every QSPR model that
includes geometry-dependent molecular descriptors usually
includes a relatively difficult calculation of the optimum
molecular geometry, with high computational costs and long
development time.
The Correlation and Logic (CORAL) software can be used as
a computational tool to carry out calculations to build up the
QSPR/QSAR based on molecular descriptors from SMILES
notation. All QSAR models are based on the Monte Carlo
approach according to the principle “QSAR is a random event”
[19–26]. The aim of the present study was to test the predictive
potential of the CORAL models for prediction of toxicity of a
large group of organic compounds to D. magna.
2692 Environ Toxicol Chem 35, 2016
MATERIALS AND METHODS
Data
The toxicity data on D. magna for 758 compounds were
taken from the literature [27]. The negative logarithm of 48-h
lethal concentration (millimoles per liter) is examined as the
endpoint (–log 50% lethal concentration [pLC50]). The above-
mentioned compounds were distributed into the training set
(“builder of a model”), the invisible training set (“blocker
of overtraining”), the calibration set (“estimator of predictive
potential”), and the external validation set (final confirmation of
the predictive potential of a model). These distributions obey the
following principles: 1) they are random, and 2) the fractions of
the compounds considered for each set are approximately 40%,
approximately 40%, approximately 10%, and approximately
10% for the training, invisible training, calibration, and
validation sets, respectively.
Optimal descriptors
The version of optimal descriptors suggested in the
literature [28,29] was utilized in the present study:
DCW ðT  ; N  Þ ¼ W ðBONDÞ þ W ðNOSPÞ
þ W ðHALOÞ þ WðPAIRÞ
X X
þ W ð Sk Þ þ W ðSSk Þ
X With numerical data on the correlation weights of all features
þ W ðSSSk Þ
Figure 1. Graphical scheme of the definition of the T* and N* parameters.
A.P. Toropova et al.
In Equation 1, CW(x) is the correlation weight of SMILES
attribute x; BOND is a global SMILES descriptor which
is a mathematical function of the presence or absence of
double (¼), triple (#), and stereochemical (@) bonds; NOSP
is a global SMILES descriptor which is a mathematical
function of the presence or absence of the chemical elements
nitrogen (N), oxygen (O), sulfur (S), and phosphorus (P);
HALO is a global SMILES descriptor which is a mathematical
function of the presence or absence of the chemical elements
fluorine (F), chlorine (Cl), bromine (Br), and iodine (I) (i.e.,
halogens); and PAIR represents the simultaneous presence
of pairs of the above-mentioned (in parentheses) SMILES
elements.
It is possible to calculate the correlation weights, which
establish correlation between the optimal descriptor and the
desired endpoint. The calculation can be done by the Monte
Carlo method [19–29].
The T* and N* parameters are from the Monte Carlo
optimization procedure. The former represents the threshold
to classify all molecular features extracted from SMILES into
2 classes: rare (noise) and active. The model calculated with
correlation weights includes solely active features, whereas
correlation weights of rare features are defined as equal to 0.
The latter is the number of epochs of the Monte Carlo
ð1Þ optimization, which give the maximal correlation coefficient for
the calibration set (Figure 1) [30].
involved in building up the model together with defined values
QSAR models for toxicity to Daphnia magna Environ Toxicol Chem 35, 2016 2693

of T* and N*, a predictive model can be calculated using the potential of a model can be estimated via a defect of Fk,
training set: defect(Fk) [31,32]:

pLC50 ¼C0 þ C1  DCWðT  ; N  Þ ð2Þ PT ðFk Þ  PC ðFk Þ

defect ðFk Þ ¼ ð3Þ
N T ðFk Þ þ N c ðFk Þ
The next step involves validation of the developed model
calculated with Equation 2. Its predictive potential should be In Equation 3, PT(Fk) and PC(Fk) are probabilities of attribute Fk
checked with the validation set. in the training and the calibration set, respectively; NT(Fk) and
One can use 2 versions of the optimization for the NC(Fk) are prevalence (frequency) of attribute Fk in the training
correlation weights [28,29]. First, the balance of correlation, set and the calibration set, respectively. The defect(Fk) ¼ 1,
where the training set is arranged into 2 groups: the compounds if NC(Fk) ¼ 0.
from the first group are utilized to calculate correlation The defect of SMILES, d(SMILES), can be estimated via
weights, and the compounds included in the second group are defects of Fk in the SMILES:
utilized for control of the absence of the overtraining (i.e.,
the ideal model for visible substances is accompanied by X
dðSMILESÞ ¼ defect ðFk Þ ! min ð4Þ
a poor model for substances which are not involved in the Fk 2SMILES
calculation of the optimal correlation weights). Second,
the traditional model, where the training set does not include The defect of a distribution into the training, invisible training,
invisible “passive” part: in fact, the invisible training set has calibration, and validation sets can be estimated via the sum of
considerably distinctive influence on the model in comparison defects of SMILES from the training and calibration sets:
with the case where these compounds are directly involved
in the training set. In the present study, a comparison of these X
2 cases is carried out. dðDistributionÞ ¼ dðSMILESÞ ð5Þ
SMILES 2Train & Calib
The approach can be carried out utilizing the CORAL
software. A detailed description of the practical use of the
Computational experiments have shown that the described
CORAL software is available in the literature [28].
models have preferable predictive potential if the d(Distribution)
Domain of applicability is minimal.
The statistically robust domain of applicability can be
Various distributions of the data into the training, invisible
introduced via inequality
training, calibration, and validation sets result in different
prevalence rates of features in the training and calibration sets.
The measure of influence of a feature, Fk, for possible predictive dðSMILESÞ < 2  dðSMILESÞ ð6Þ

Table 1. Statistical characteristics of quantitative structure–activity relationship models for 3 distributions of data into the training, invisible training, calibration,
and validation setsa

Distribution Set n r2 q2 s F R2m DR2m

1 Balance of correlations Training 288 0.7399 0.7359 0.786 813

Invisible training 284 0.7729 0.7697 0.770 960
Calibration 87 0.7712 0.7519 0.711 286 0.6768 0.1672
Validation 99* 0.7805* 0.668*
Traditional scheme Training 572 0.7815 0.7799 0.717 2039
Calibration 87 0.7256 0.7011 0.835 225 0.6225 0.0721
Validation 99 0.6605 0.870
2 Balance of correlations Training 320 0.7443 0.7407 0.786 926
Invisible training 279 0.7460 0.7424 0.756 814
Calibration 75 0.7979 0.7850 0.682 288 0.6578 0.1860
Validation 84* 0.8229* 0.6276*
Traditional scheme Training 599 0.7738 0.7722 0.722 2042
Calibration 75 0.7979 0.7857 0.683 288 0.6616 0.1847
Validation 84 0.7203 0.800
3 Balance of correlations Training 307 0.7394 0.7357 0.802 866
Invisible training 282 0.7205 0.7165 0.796 722
Calibration 82 0.9053 0.8999 0.527 765 0.8633 0.0205
Validation 87* 0.8377* 0.564*
Traditional scheme Training 589 0.7764 0.7748 0.726
Calibration 82 0.8746 0.8661 0.590 0.8205 0.0081
Validation 87 0.7426 0.787
a
n is the number of compounds in a set; r2 is the correlation coefficient between experimental and calculated –log 50% lethal concentration; q2 is the leave-one-
root mean squared error; F is the Fischer F ratio; R2m and DR2m are measures of predictability [33]
out cross-validated r2; sqisffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
  
r ðx; yÞ ¼ r  1 
2 2  r2  r2 
m 0
r2 ðx; yÞ þr2 ðy; xÞ
R2m ¼ m 2
m

DR2m ¼ r2m ðx; yÞ  r2m ðy; xÞ

where x and y are vectors of predicted and experimental endpoint values, respectively; r20 is the correlation coefficient between experimental and predicted values
calculated without intercept.
*indicates significance.
2694 Environ Toxicol Chem 35, 2016 A.P. Toropova et al.

where dðSMILESÞ is the average defect of SMILES over the

training and calibration sets.
Mechanistic interpretation
The developed QSAR models allow for mechanical
interpretation of the studied phenomena. Having the numerical
data on the correlation weights of features in several runs of the
Monte Carlo optimization, one can extract 3 categories of these
features [28–32]: 1) features which have a positive value of
the correlation weight in all runs (these are promoters of the
endpoint increase); 2) features which have a negative value of
the correlation weight in all runs (these are promoters of the
endpoint decrease); and 3) features which have both negative
and positive values of the correlation weight in different runs of
the optimization. These are features with an unclear role (one
cannot classify these features as promoters of an increase or a
decrease of the endpoint).

RESULTS

QSAR models
The QSAR models for toxicity to D. magna, which are
calculated with the balance of correlations for 3 distributions
of data into the training, invisible training, calibration, and
validation sets, are as follows:

pLC50 ¼ 0:8701 ð  0:0084Þ þ 0:09209 ð 0:0002Þ

ð7Þ
 DCWð3;10Þ

pLC50 ¼ 0:0020 ð 0:0099Þ þ 0:09745Þ ð 0:0002Þ

ð8Þ
 DCWð3;10Þ

pLC50 ¼ 0:0016 ð 0:0097Þ þ 0:09505 ð 0:0002Þ

ð9Þ
 DCWð3;10Þ

However, if one does not use the balance of correlation
approach, the developed models are different. The QSAR
models for toxicity to D. magna, which are calculated with the
traditional scheme (without the invisible training set) for 3
distributions of the data into the training, calibration, and
validation sets, are as follows:
pLC50 ¼ 2:087 ð  0:0028Þ þ 0:08818 ð 0:0001Þ
ð10Þ
 DCWð3;10Þ
pLC50 ¼ 2:301 ð  0:0026Þ þ 0:08368 ð 0:0001Þ
ð11Þ
 DC Wð3;10Þ
pLC50 ¼ 2:538 ð  0:0023Þ þ 0:09843 ð 0:0001Þ
ð12Þ
 DCWð3;10Þ
Figure 2. Graphical representation of correlations between experimental
and calculated –log 50% lethal concentration (pLC50) values.
3, respectively. This changes when the traditional scheme of
distribution is used. In such a case the numbers of outliers are
14, 12, and 5 for distributions 1, 2, and 3, respectively. Thus,
distribution 3 seems to be preferable for both versions of the
Monte Carlo optimization (balance of correlations and traditional
scheme). In fact, the criterion expressed as inequality 6 is an
indicator of compounds, which have rare, untypical molecular
features. In other words, these compounds are suspected to be
outliers. However, the data on the number of these suspected
compounds give an additional possibility to estimate different
distributions of the data into the training, invisible training,
calibration, and validation sets: “if the number of the above-
mentioned outliers is smaller, then distribution is better.”

Table 1 contains the statistical characteristics of the models Mechanistic interpretation

calculated with Equations 7 to 12. All of these models are
satisfactory according to statistical criteria [33] represented
in Table 1. However, it is to be noted that the balance of
correlations approach gives a better prediction for all 3 random
distributions. Figure 2 represents the models calculated with
Equations 7 to 9 graphically.
Table 2 contains the lists of SMILES attributes, which are
revealed to be stable promoters of the pLC50 increase or
decrease. The data on prevalence of these promoters in the training,
invisible training, and calibration sets is displayed in Table 3.
DISCUSSION

Domain of applicability Comparison of statistical quality of QSAR for D. magna

The numbers of outliers according to inequality 6 for the Table 4 contains the statistical characteristics of QSAR
balance of correlations are 10, 9, and 7 for distributions 1, 2, and models for D. magna available in the literature [34–37]. This
QSAR models for toxicity to Daphnia magna Environ Toxicol Chem 35, 2016 2695

Table 2. List of molecular features extracted from simplified molecular input-line entry system (SMILES), which are promoters of increase (or decrease) for
–log 50% lethal concentration (pLC50)

No. Feature, Fk Comment Impact for pLC50

1 BOND10000000 Presence of double bonds Increase

2 BOND00000000 Absence of double, triple, and stereochemical bonds Increase
3 NOSP01000000 Presence of oxygen together with absence of nitrogen, sulfur, and phosphorus Increase
4 þþþþN─ ─ ─B25 555 Simultaneous presence in a molecule of nitrogen and double bond (s) Increase
5 S ... Presence of sulfur atom (s) Increase
6 þþþþS─ ─ ─ B25 555 Simultaneous presence of Sulfur together with double bond (s) Decrease
7 C . . . C . . . (. . .) Branching at pair of carbon atoms in sp3 state Decrease
8 n ... Presence of nitrogen in sp2 state Decrease

provides a good summary of the already executed approaches.
One can compare the previously published data with the
statistical characteristics of the models developed in the present
study. The data presented in Table 4 clearly indicate that the
QSAR models calculated with Equations 7 to 9 are comparable
with models suggested in the literature.
Advantages of the CORAL models
There are various benefits of application of the CORAL
models. The approach described in the present study applies
the QSAR model using solely experimental values of pLC50
together with data on the molecular structure. There is no
need to use information on physicochemical parameters,
3-dimensional representation of the molecular systems, and
quantum mechanics descriptors for the considered compounds.
The method applied in the present study delivers QSAR models
in accordance with Organisation for Economic Co-operation
and Development principles [38].
Disadvantages of the CORAL models
Though CORAL models are efficient and reliable, there are
also some drawbacks of such approaches: 1) there are SMILES
attributes (this is related to SSk and SSSk involved in Equation 1)
which have no physical interpretation, and 2) the Monte Carlo
calculations take considerable time to execute, especially if the
number of compounds is large (e.g., n > 500).
Availability of examined data
The Supplemental Data contain technical details: 1) SM1
contains correlation weights for calculations with Equations 7
to 9; 2) SM2 contains experimental and predicted pLC50
values calculated with Equations 7 to 9; 3) SM3 contains the
correlation weights obtained in 3 probes of the Monte Carlo
optimization where one can find the stable promoters of the
pLC50 increase together with promoters of the pLC50 decrease;
and 4) SM4 contains 3 distributions into the training (i.e., the
training, invisible training, and calibration sets) and external
validation sets which were examined in the present study
(these distributions can be checked with the CORAL software);
and 5) SM5 contains the CORAL method used to build the
described models.
CONCLUSIONS
Using the CORAL approach, predictions of toxicity of 758
organic compounds to D. magna were carried out. All QSAR

Table 3. Correlation weights and prevalence of molecular features Fk extracted from simplified molecular input-line entry system (SMILES), which are
promoters of increase (or decrease) for –log 50% lethal concentration

Feature, Fk Distribution CW(Fk) in probe 1 CW(Fk) in probe 2 CW(Fk) in probe 3 NT NIT NC Defect(Fk)

BOND10000000 1 4.50212 4.25292 2.50098 163 138 47 0.0001

2 4.49822 7.49671 4.99822 178 146 29 0.0008
3 5.00259 3.50238 3.99991 163 158 34 0.0006
BOND00000000 1 7.75332 9.49539 10.50317 115 139 39 0.0003
2 9.99812 13.75023 9.24921 133 123 45 0.0010
3 8.24893 9.50242 8.00085 137 114 45 0.0006
NOSP01000000 1 2.50275 3.99775 3.49717 105 115 27 0.0004
2 3.25394 3.24730 2.50415 121 95 34 0.0005
3 3.49710 3.50430 3.49783 112 105 37 0.0006
þþþþN─ ─ ─B25
555 1 0.99642 3.49920 2.50496 76 59 21 0.0002
2 4.49525 4.50001 1.49646 83 65 10 0.0014
3 1.00461 0.99860 1.75335 67 74 15 0.0004
S ... 1 6.00091 5.49845 6.49918 44 36 13 0.0001
2 5.75436 5.74986 4.50392 47 40 2 0.0025
3 5.49772 5.25079 3.49748 40 40 9 0.0004
þþþþS─ ─ ─B25
555 1 –3.75458 –4.49712 –3.49840 38 29 11 0.0001
2 –1.75444 –8.00161 –6.50246 40 34 1 0.0027
3 –2.75126 –4.75226 –4.75001 32 32 9 0.0001
C ... C ... (...) 1 –0.00259 –0.24797 –1.24937 87 88 27 0.0001
2 –1.24737 –1.24876 –0.25035 95 80 23 0.0001
3 –0.75343 –1.24625 –2.25130 89 92 19 0.0005
n ... 1 –2.50344 –2.75050 –3.99880 12 17 3 0.0005
2 –0.75458 –1.24965 –0.99687 13 15 3 0.0000
3 –2.24872 –3.00371 –4.24579 15 15 3 0.0007

CW ¼ correlation weight; defect(Fk) ¼ defect of a feature, Fk, calculated with Equation 3; NT, NIT, and NC ¼ numbers of SMILES attribute x in the training,
invisible training, and calibration sets, respectively.
2696 Environ Toxicol Chem 35, 2016
Table 4. Comparison of the statistical quality of quantitative structure–activity relationship models for toxicity to Daphnia magna
Training set
No. n r2 s n
1 — 0.740–0.768 0.79–1.00
2 222 0.738
3 97 0.77 0.39
4 149 0.70 1.04
5 307 0.739 0.80
a
Equation 9.
models are based on SMILES notation optimal descriptors
and were developed with application of the Monte Carlo
method. The predictive potential of the applied approach was
tested with 3 random splits into the subtraining, calibration, test,
and validation sets and with different statistical methods.
All models considered in the present study are characterized
by the following features: 1) every time, the best statistical
characteristics for the calibration set are accompanied by
satisfactory predictive potential (the statistical characteristics
for the external validation set), and 2) the balance of correlation
approach gives better predictions than the traditional scheme.
Both features demonstrate that Monte Carlo method–based
modeling incorporated in CORAL software is a very promising
computational method in QSAR studies for risk assessment
related to toxicity of organic chemicals to D. magna. The
SMILES attributes (both local and global), which are promoters
of toxicity increase or decrease were identified and defined.
These structural features are related to organic chemical
toxicity, and their identification helped to improve the
understanding of organic chemical toxicity toward D. magna.
The Monte Carlo calculations described in the present study
can be reproduced using the CORAL software.
Supplemental Data—The Supplemental Data are available on the Wiley
Online Library at DOI:10.1002/etc.3466.
Acknowledgment—A.A. Toropov and A.P. Toropova thank the European
Commission project PeptiCAPS (project 686141). A.M. Veselinovi
c and
J.B. Veselinovi
c acknowledge support from the Ministry of Education
and Science, the Republic of Serbia (project 43012). D. Leszczynska and
J. Leszczynska acknowledge support from the National Science Foundation
(NSF/CREST HRD-0833178) and EPSCoR (362492-190200-01/NSFEPS-
090378).
Data availability—Data were taken from Zhang et al. [27]. In addition, the
Supplemental Data contain the data as Excel files.
REFERENCES
1. Mackay D, Hubbarde J, Webster E. 2003. The role of QSARs and fate
models in chemical hazard and risk assessment. QSAR Comb Sci
22:106–112.
2. Furtula B, Gutman I. 2011. Relation between second and third
geometric-arithmetic indices of trees. J Chemom 25:87–91.
3. Afantitis A, Melagraki G, Koutentis PA, Sarimveis H, Kollias G. 2011.
Ligand-based virtual screening procedure for the prediction and the
identification of novel b-amyloid aggregation inhibitors using Kohonen
maps and counterpropagation artificial neural networks. Eur J Med
Chem 46:497–508.
4. Afantitis A, Melagraki G, Sarimveis H, Koutentis PA, Igglessi-
Markopoulou O, Kollias G. 2010. A combined LS-SVM & MLR QSAR
workflow for predicting the inhibition of CXCR3 receptor by
quinazolinone analogs. Mol Divers 14:225–235.
5. Duchowicz PR, Comelli NC, Ortiz EV, Castro EA. 2012. QSAR study
for carcinogenicity in a large set of organic compounds. Curr Drug Saf
7:282–288.
A.P. Toropova et al.
Validation set
r2 s Reference
— — — 34
75 0.721 35
— — 0.34–0.44 36
89 0.768 0.88 37
87 0.838 0.564 Present studya
6. Ghaedi A. 2015. Predicting the cytotoxicity of ionic liquids using
QSAR model based on SMILES optimal descriptors. J Mol Liq
208:269–279.
7. Li Q, Ding X, Si H, Gao H. 2014. QSAR model based on SMILES of
inhibitory rate of 2,3-diarylpropenoic acids on AKR1C3. Chemometr
Intell Lab Syst 139:132–138.
8. Masand VH, Toropov AA, Toropova AP, Mahajan DT. 2014. QSAR
models for anti-malarial activity of 4-aminoquinolines. Curr Comput
Aided Drug Des 10:75–82.
9. Scotti L, Lima EO, da Silva MS, Ishiki H, Lima IO, Pereira FO,
MendonSca FJB Jr, Scotti MT. 2014. Docking and PLS studies on a set of
thiophenes RNA polymerase inhibitors against Staphylococcus aureus.
Curr Top Med Chem 14:64–80.
10. Scotti L, Ishiki H, Ferreira MJP, Francisco JBM Jr, De P Emerenciano
V, Silva MS, Scotti MT. 2012. In silico methods applied in food
chemistry: A short review with bitter and mutagenic compounds. Lett
Drug Des Discov 9:527–534.
11. Speck-Planche A, Kleandrova VV, Luan F, Cordeiro MNDs. 2015.
Computational modeling in nanomedicine: Prediction of multiple
antibacterial profiles of nanoparticles using a quantitative structure–
activity relationship perturbation model. Nanomedicine 10:193–204.
12. Torrens F, Castellano G. 2014. QSPR prediction of chromatographic
retention times of pesticides: Partition and fractal indices. J Environ Sci
Health B 49:400–407.
13. Torrens F, Castellano G. 2012. QSPR prediction of retention times of
phenylurea herbicides by biological plastic evolution. Curr Drug Saf
7:262–268.
14. van der Jagt K, Munn S, Torslov J, de Bruijn J, eds. 2004. Alternative
approaches can reduce the use of test animals under REACH.
Addendum to: Assessment of additional testing needs under REACH
effects of (Q)SARS, risk based testing and voluntary industry
initiatives. IHCP report EUR 21405 EN. Joint Research Centre
Institute for Health and Consumer Protection, European Commission,
Ispra, Italy.
15. Ivanciuc O. 2013. Chemical graphs, molecular matrices and topological
indices in chemoinformatics and quantitative structure-activity rela-
tionships. Curr Comput Aided Drug Des 9:153–163.
16. Weininger D. 1988. SMILES, a chemical language and information
system. 1. Introduction to methodology and encoding rules. J Chem Inf
Comput Sci 28:31–36.
17. Weininger D, Weininger A, Weininger JL. 1989. SMILES. 2.
Algorithm for generation of unique SMILES notation. J Chem Inf
Comput Sci 29:97–101.
18. Weininger D. 1990. SMILES. 3. Depict. Graphical depiction of
chemical structures. J Chem Inf Comput Sci 30:237–243.

19. Zivkovi
c JV, Truti
c NV, Veselinovi
c JB, Nikoli
c GM, Veselinovi
c
AM. 2015. Monte Carlo method based QSAR modeling of maleimide
derivatives as glycogen synthase kinase-3b inhibitors. Comput Biol
Med 64:276–282.

20. Veselinovi
c JB, Nikoli
c GM, Truti
c NV, Zivkovi
c JV, Veselinovi
c
AM. 2015. Monte Carlo QSAR models for predicting organophosphate
inhibition of acetylcholinesterase. SAR QSAR Environ Res 26:449–
460.

21. Veselinovi
c AM, Veselinovi
c JB, Zivkovi
c JV, Nikoli
c GM. 2015.
Application of smiles notation based optimal descriptors in drug
discovery and design. Curr Top Med Chem 15:1768–1779.
22. Achary PGR. 2014. QSPR modelling of dielectric constants of p-
conjugated organic compounds by means of the CORAL software. SAR
QSAR Environ Res 25:507–526.
23. Worachartcheewan A, Nantasenamat C, Isarankura-Na-Ayudhya C,
Prachayasittikul V. 2014. QSAR study of H1N1 neuraminidase
inhibitors from influenza a virus. Lett Drug Des Discov 11:420–427.
QSAR models for toxicity to Daphnia magna
24. Achary PGR. 2014. Simplified molecular input line entry system-based
optimal descriptors: QSAR modelling for voltage-gated potassium
channel subunit Kv7.2. SAR QSAR Environ Res 25:73–90.
25. Garc
ıa J, Duchowicz PR, Rozas MF, Caram JA, Mir
ıfico MV,
Fern
andez FM, Castro EA. 2011. A comparative QSAR on 1,2,5-
thiadiazolidin-3-one 1,1-dioxide compounds as selective inhibitors of
human serine proteinases. J Mol Graph Model 31:10–19.
26. Mullen LMA, Duchowicz PR, Castro EA. 2011. QSAR treatment on
a new class of triphenylmethyl-containing compounds as potent
anticancer agents. Chemometr Intell Lab Syst 107:269–275.
27. Zhang X, Qin W, He J, Wen Y, Su L, Sheng L, Zhao Y. 2013.
Discrimination of excess toxicity from narcotic effect: Comparison of
toxicity of class-based organic chemicals to Daphnia magna and
Tetrahymena pyriformis. Chemosphere 93:397–407.
28. Toropova AP, Toropov AA, Benfenati E, Leszczynska D, Leszczynski
J. 2015. QSAR model as a random event: A case of rat toxicity. Bioorg
Med Chem 23:1223–1230.
29. Toropova AP, Toropov AA, Benfenati E, Gini G, Leszczynska D,
Leszczynski J. 2011. CORAL: Quantitative structure-activity relation-
ship models for estimating toxicity of organic compounds in rats.
J Comput Chem 32:2727–2733.
30. Toropova AP, Toropov AA, Veselinovi
c JB, Veselinovi
c AM. 2015.
QSAR as a random event: A case of NOAEL. Environ Sci Pollut Res Int
22:8264–8271.
31. Toropova AP, Toropov AA, Rallo R, Leszczynska D, Leszczynski J.
2015. Optimal descriptor as a translator of eclectic data into prediction
Environ Toxicol Chem 35, 2016 2697
of cytotoxicity for metal oxide nanoparticles under different conditions.
Ecotoxicol Environ Saf 112:39–45.
32. Toropova MA, Toropov AA, Raska I, Raskov
a M. 2015. Searching
therapeutic agents for treatment of Alzheimer disease using the Monte
Carlo method. Comput Biol Med 64:148–154.
33. Ojha PK, Mitra I, Das RN, Roy K. 2011. Further exploring rm2 metrics
for validation of QSPR models. Chemometr Intell Lab Syst 107:194–
205.
34. Vikas R. 2015. Exploring the role of quantum chemical descriptors in
modeling acute toxicity of diverse chemicals to Daphnia magna. J Mol
Graph Model 61:89–101.
35. Kar S, Roy K. 2010. QSAR modeling of toxicity of diverse organic
chemicals to Daphnia magna using 2D and 3D descriptors. J Hazard
Mater 177:344–351.
36. Cassani S, Kovarich S, Papa E, Roy PP, van der Wal L, Gramatica P.
2013. Daphnia and fish toxicity of (benzo)triazoles: Validated QSAR
models, and interspecies quantitative activity–activity modeling.
J Hazard Mater 258–259:50–60.
37. Toropova AP, Toropov AA, Martyanov SE, Benfenati E, Gini G,
Leszczynska D, Leszczynski J. 2012. CORAL: QSAR modeling of
toxicity of organic chemicals toward Daphnia magna. Chemometr
Intell Lab Syst 110:177–181.
38. Organisation for Economic Co-operation and Development. 2007.
Guidance document on the validation of (quantitative)structure–
activity relationship [(Q)SAR] models. Series on Testing and
Assessment, No. 69. ENV/JM/MONO(2007)2. Paris, France.