Download as pdf or txt
Download as pdf or txt
You are on page 1of 19

Send Orders for Reprints to reprints@benthamscience.

ae
1868 Current Topics in Medicinal Chemistry, 2015, 15, 1868-1886

Nano-QSPR Modelling of Carbon-Based Nanomaterials Properties

Maryam Salahinejad*

Application Radiation Research School, NSTIR, Tehran, Iran

Abstract: Evaluation of chemical and physical properties of nanomaterials is of critical importance in a broad variety of
nanotechnology researches. There is an increasing interest in computational methods capable of predicting properties of
new and modified nanomaterials in the absence of time-consuming and costly experimental studies. Quantitative Struc-
ture-Property Relationship (QSPR) approaches are progressive tools in modelling and prediction of many physicochemi-
cal properties of nanomaterials, which are also known as nano-QSPR. This review provides insight into the concepts,
challenges and applications of QSPR modelling of carbon-based nanomaterials. First, we try to provide a general over-
view of QSPR implications, by focusing on the difficulties and limitations on each step of the QSPR modelling of nano-
materials. Then follows with the most significant achievements of QSPR methods in modelling of carbon-based nanoma-
terials properties and their recent applications to generate predictive models. This review specifically addresses the QSPR
modelling of physicochemical properties of carbon-based nanomaterials including fullerenes, single-walled carbon nano-
tube (SWNT), multi-walled carbon nanotube (MWNT) and graphene.
Keywords: Nano-QSPR, Physiochemical properties, Fullerenes, SWNTs, MWNTs, Graphene.

1. INTRODUCTION graphite with two-dimensional properties. The honeycomb


lattice structure of graphene with a large delocalized π -
Carbon-based nanomaterials (CBNs) offer interesting
electron system gives rise to many interesting properties
physical and chemical characteristic structures with superior [11]. Carbon nanotubes (CNTs), can be described as gra-
and unique thermal, mechanical, optical and electrical prop-
phene sheets rolled into a cylindrical tube. SWNTs are hol-
erties. Considering the aforementioned advantageous proper-
low cylindrical nanostructure tubes made from a wrapped
ties, carbon-based nanomaterials have found wide range ap-
graphene sheet, while the MWNTs consist of multiple layers
plications in biomedical, technological and environmental
of graphene sheets rolled up to form polymeric structures of
research. CBNs are important candidate for synthetic scaf-
concentric tubes [10]. Based on the rolling vector of gra-
folds in tissue engineering [1], drug delivery systems and phene sheet which shows the arrangement of the sp2 carbon
cancer therapy [2], cellular sensors and imaging agents [3].
hexagons, the tubes can have the shape of armchair, zigzag
Due to their potential applications in improving structural
and or chiral (Fig. 1D, 1E and 1F respectively). The chiral
strength of materials they have been identified as promising
vector (Ch) defined as Ch =na1 + ma2, where a1 and a2 are
nanomaterials in the field of electronics [4], cosmetics [5],
graphene lattice vectors and n and m are integer numbers
polymers [6] and energy [7]. In recent years, many environ-
characteristic for nanotube of any helicity [12, 13]. The
mental researches have been done relying heavily on the paired value of the integer numbers (n,m), which character-
performance of CBNs with their increasing applications in
ized the atomic structure of a nanotube, plays an important
composite filters, high-flux membranes, antimicrobial
role in many physical and chemical (physicochemical) prop-
agents, nanoprobes and sensors [8].
erties of nanotubes. Due to different diameter, length, and
Various types of CBNs such as fullerenes, graphene, chirality during their growth, carbon nanostructures exhibit a
multi-walled carbon nanotubes (MWNT) and single-walled rich structural diversity, which has led to the discovery of
carbon nanotubes (SWNT) are shown in Fig. (1). Fullerenes physicochemical properties of CBNs difficult and fascinating
are a class of cage-like molecules of carbon, Cn, with fairly as well.
common structures of C60 and C70. Fullerene C60, consisting
Theoretical studies and computational methods of nanos-
of 60 sp2 hybridized carbon atoms arranged as 20 hexagonal
tructured materials can help to obtain their structural infor-
and 12 pentagonal rings, represents a special class of spheri-
mation and may result to a fast estimate of their properties or
cal quasi-aromatic systems [9]. The π -system of delocalized activities in the absence of expensive experiments. Quantita-
electrons in fullerene structure make them like an electron
tive Structure-Activity/Property Relationship (QSAR/QSPR)
deficient alkene which reacts readily with electron rich spe-
approaches, as one of the major computational molecular
cies [10]. Graphene is a single atomic layer of crystalline
modeling methodologies act as an effective means to predict
and estimate of activities or properties of compounds based
*Address correspondence to this author at the Application Radiation Re-
search School, NSTIR, Tehran, Iran;
on their structures [14]. Nano-QSAR or Quantitative Nanos-
Tel: +989125253143; Fax: +982188221166; tructure-Activity Relationships (QNAR), first introduced by
E-mails: salahinejad@gmail.com; salahinejad@yahoo.com Puzyn et al. [15] refers to the application of QSAR methods

1873-5294/15 $58.00+.00 © 2015 Bentham Science Publishers


Nano-QSPR Modelling of Carbon-Based Nanomaterials Properties Current Topics in Medicinal Chemistry, 2015, Vol. 15, No. 18 1869

A) B) C)

D) E) F)
Fig. (1). Carbon-based nano material structures. A) Fullerene C60, B) Geraphene, C) Multi-walled carbon nanotube (MWNT), and D) Arm-
chair, E) Zigzag and F) Chiral single-walled carbon nanotube (SWNT).

for construction of predictive models of biological activities tors and the property in question; and validating of reliabil-
of nanomaterials. There are some recently published reviews ity, robustness and prediction capabilities of QSPR models.
on QSAR modelling of nanomaterials, but most of those The workflow of the QSAR/QSPR paradigm is shown in
manuscripts focus on toxicity or biological activity of nano- Fig. (2).
materials especially nanoparticles [16-21]. The aim of this Some comprehensive reviews [14, 22-24] provide the in-
review is to summarize the most significant achievements of
depth representation of the main concepts involved in the
QSPR methods in modelling of carbon nanomaterials phys-
development of QSAR/QSPR models. Therefore, only a
icochemical properties and their recent applications to gener-
summary of the important elements of the QSPR modelling
ate predictive models.
process, which used in modelling of carbon nanomaterials
properties, is provided here. Instead, the difficulties and
2. GENERAL SCHEME OF QSPR challenges on each step of building a nano-QSPR model will
All physical, chemical, biological and technological be further investigated.
properties of a particular molecule depend on its molecular
structure and vary with it in a systematic way [22]. The es- 2.1. Data Collection
tablishment of correlations and quantitative prediction be-
Data collection or selection of an appropriate training set
tween diverse molecular properties and chemical structure is
is a key step in developing a reliable QSPR model. Katritzky
known as quantitative structure-activity/property relationship
et al. summarized some criteria which should be considered
(QSAR/QSPR). QSAR methods are considered the most
in a data set selection [22]. Although, there is no direct and
popular theoretical method for chemical researches world-
wide, and are extensively applied in the fields of medicinal, simple answer to the size of the data set in constructing a
environmental, agricultural and technological aspects of life QSPR model, but it should utilize as many as possible com-
[17, 23-25]. The major goal of the QSPR studies is to con- pounds. If the number of compounds in the training set is not
struct a statistical model that relates the property of interest large enough, the data may not span the complete property
and descriptive parameters (descriptors) derived from the space, hence the QSPR production of new similar com-
structure of the molecule. In QSPR methodology, a set of pounds can't be reliable. The structural diversity and activ-
descriptors (Xi) is connected to experimental property values ity/property value distribution of compounds in a dataset are
(Y) with a mathematical model (f), which defined as equa- very important for QSAR model development [26]. The
tion 1. range of the property values should cover a range of at least
1 logarithmic unit. One challenges in developing nano-
Y (property) =f (Xi) (1) QSAR models, is considering that the QSAR modelling has
The construction of QSAR/QSPR model typically com- been developed for structurally diverse small organic com-
prises of five main steps: collecting the experimental data pounds, but nanomaterials are not single discrete molecules
set; calculating descriptors from molecular structures; choos- like small organic molecules, rather are often poorly defined
ing informative descriptors that are deemed to be important mixtures or distributions of materials. Hence, nanomaterials
for explaining the desired property or activity; generating have multiple dimensional physicochemical properties [27]
linear or non-linear relationships between selected descrip- but have limited diversity in structure. In the case of carbon
1870 Current Topics in Medicinal Chemistry, 2015, Vol. 15, No. 18 Maryam Salahinejad

nanotubes the distribution of structures probably includes lated by mathematical formulas or computational algorithms
different lengths, diameters, and numbers of layers in the [30]. There are numerous programs that generally produce a
walls. Due to high structural complexity and the diversity of pool of different molecular descriptors involving: constitu-
nanomaterial properties it is difficult to develop quantitative tional, topological, geometric, electrostatic and quantum-
parameters capable of characterizing their structural and chemical, fragment-based and molecular fingerprint descrip-
chemical properties [28]. tors for discrete, small organic molecules, not materials. Un-
fortunately, such packages which generate hundreds or thou-
sands of descriptors for small organic molecules, have not
yet developed for nanomaterials. As discussed above, due to
high structural complexity and the diversity of nanomaterial
properties, nano-specific descriptors or nanodescriptors are
needed to be developed to represent the physicochemical
properties of nanomaterials for developing reliable QSAR
models. Unlike small organic compounds, there is no sys-
tematic way to describe the physicochemical, geometrical
and structural features of nanomaterials. One important chal-
lenge in nano-QSPR modelling is to identify the relevant
features and develop quantitative parameters describing
unique molecular properties of nanomaterials for modelling
[15].
Different experimental and theoretical descriptors are
proposed and developed to describe the carbon nanomateri-
als. Theoretical descriptors were derived from physics-based
computational methodologies, such as molecular dynamics
(MD) simulations [31] and quantum chemistry [32]. The MD
simulation method is based on Newton’s second law and has
provided detailed information at the microscopic level, in-
cluding atomic positions and velocities [33]. Quantum chem-
istry methods are based on approximate solutions of the
Schrödinger equation and use quantum mechanics to deal
with chemical problems e.g. structure, molecular properties
and stability during reactions [34]. Chemical graph theory
have been also utilized to calculate some topological descrip-
tors including topological indices such as Wiener, Padmakar-
Ivan [35], Szeged indices [36], Simplified Molecular-Input
Line-Entry System (SMILES) and International Chemical
Identifier (InChI) descriptors of CNTs and fullerenes [37-
39]. Experimental descriptors derived from spectros-
copy/microscopy characterization results of nanomaterials
e.g. NMR spectroscopy [40], scanning tunneling microscopy
[41] and scan electron microscopy [42] also can be used as
descriptors. Borders et al. [31] defined and developed 20
descriptors for mechanical property prediction of carbon
nanotubes by MD simulations. Table 1 displays some theo-
Fig. (2). Workflow of the QSAR/QSPR paradigm. retical descriptors derived from MD and DFT calculations or
graph theory for carbon nanotubes.
The reliability of the experimental property, as an impor-
tant issue, determines the prediction ability and accuracy of 2.3. Variable Selection
QSPR models [22]. Unfortunately, the most nano-QSAR
models suffer from lack of highquality experimental data The relevant descriptors only should be used in the
and sufficient data available for being able to build reliable QSPR models to avoid overfitting and to improve predic-
and predictive models. One important difficulty in extending tion ability of the models. A feature selection method
the traditional QSPR to nano-QSPR methodology would be should completely eliminate uninformative and intercorre-
possibility to develop a reliable and predictive model based lated descriptors to avoid overfitting of QSPR model and
on a limited number of experimental data on nanomaterials make easier interpretation of the model. Leaving unin-
[29]. formative descriptors in models can degrade the model sub-
stantially. As an acceptable guideline, “five or six data
2.2. Descriptors points per descriptor”[43] is required to develop a QSPR
model. Considering this rule and the lack of sufficiently
The molecular descriptors are numerical value that char- large data sets for nanomaterials, some preliminary knowl-
acterize various molecular properties, which can be meas- edge or using a variable selection procedure is required to
ured physicochemical properties or theoretical indexes calcu- select the best set of descriptors.
Nano-QSPR Modelling of Carbon-Based Nanomaterials Properties Current Topics in Medicinal Chemistry, 2015, Vol. 15, No. 18 1871

Table 1. Theoretical descriptors for carbon nanotubes.

Descriptors Symbol Definition

Area A Cross-sectional area of nanotube (Å2)

Chiral angle CA Chiral angle (radians)

Average radius RA Average of first three end rings radii for relaxed CNT (Å)

Theoretical radius RT Theoretical radius of perfect nanotube (Å)

Aspect ratio AR Ratio of the length to the diameter

% Single defects %SD Percentage of single defects

% Double defects %DD Percentage of double defects

% Methyl groups %M Percentage of surface functionalized methyl

Number of Missing C’s CM Number of missing due to a vacancy defect

Number of Methyl groups MN Number of methyl functional groups

Number of Total C’s CT Total number of carbons in the test case

MN/CT MN/CT Ratio of methyl groups to total number of carbons

CM/CT CM/CT Ratio of missing carbons to total number of carbons

Number of Single defects #SD Number of single defect types

Number of Double defects #DD Number of double defect types

Number of Nonsp2 C’s CN2 Number of nonsp2 hybridized carbons

CN2/CT CN2/CT Ratio of nonsp2 hybridized carbons to total number of carbons

Surface area SP Total surface area of nanotube (uses average radius)

Defect surface area SD Surface area of defects

Ratio of SD/SP SD/SP Defect area to total surface area

HOMO HOMO Highest occupied molecular orbital

LUMO LUMO Lowest unoccupied molecular orbital

Energy gaps E Energy differences between HOMO and LUMO

Ionization potential IP Ionization potentian

Electrophilicity Index Ω Electrophilicity index

Chemical hardness η Chemical hardness

Electron affinity EA Electron affinity

Electronegativity Χ Electronegativity

Wiener index W Wiener index

Padmakar-Ivan index PI Padmakar-Ivan index

Szeged index Sz Szeged index

There are statistical techniques available for proper de- sign (FFD) [51]. Sparse feature selection methods as novel
scriptor selection based on filtering and wrapper techniques. mathematical techniques have been recently applied for find-
In filtering method, no model is built, and statistical proper- ing and extracting a small number of the most significant
ties of the features are evaluated using some criteria. Wrap- features without overfitting the model [52, 53]. Sparse meth-
per approaches, which are based on constructing and evaluat- ods employed Bayesian statistics with a laplacian prior [54],
ing a series of models to select a subset of descriptors, in- are efficient tool for supervised feature selection of descrip-
cludes sequential backward feature elimination, sequential tors, specifically, when dealing with data sets with the large
feature forward selection [44], genetic algorithm [45-47], number of features compared to small sample sizes [55]. For
enhanced replacement method (ERM) [48, 49], successive more details, refer to the reviews published in variable re-
projection algorithm (SPA) [50] and fractional factorial de- duction methods for QSAR [56-58].
1872 Current Topics in Medicinal Chemistry, 2015, Vol. 15, No. 18 Maryam Salahinejad

2.4. Model Construction error in calibration (RMSEC) and the root-mean-square error
in prediction (RMSEP) also used as alternative terms for
The main step of the QSPR methodology is to derive a
error in calibration and prediction sets respectively. The pre-
model that mapping between the property and the values of
dictive power of the QSPR models is often quantified in
the selected features. In many cases, linear models as the
terms of predictive squared correlation coefficient (R2p) and
basis of QSPR method from its beginning can generate good SEP [76]. A valid model with high generalization ability has
models. Linear models are easily interpretable and suffi-
squared correlation coefficient of calibration (R2c) and SEC
ciently accurate for small datasets of similar compounds
values close to those of the model validation of R2p and SEP
[24]. Linear multivariate statistical methods, such as multiple
respectively. Generally, higher values of Q2, R2c and R2p and
linear regression (MLR), principal component regression
lower values of the SEC, RMSEC, SEP and RMSEP are de-
(PCR) [59] and partial least-squares regression (PLS) [60,
sired to develop a reliable QSAR model.
61] are widely used in QSPR model buildings. However, the
structure-property relationship, in some cases, may be non- Besides of robustness, accuracy and predictive ability of
linear in nature and nonlinear approaches are required to QSPR models, they should have a defined domain of appli-
model properly such properties. Artificial Neural Networks cability. The applicability domain (AD) of a QSPR model is
(ANN) [62, 63], the k-Nearest Neighbor (KNN) [64, 65] and a theoretical region of chemical space defined by the struc-
Support Vector Machines (SVM) [66, 67] are quite popular tural diversity of the compounds in the data set and the mod-
nonlinear techniques in QSPR analysis. However, nonlinear elled response, for which a given QSPR should make reliable
models seem to have the drawbacks of tendency to overfit predictions [77]. The AD determines the boundaries of the
the data, more complicated estimation and interpretation in model developed to find how well it will work for new com-
ascertaining the descriptors which are most significant in the pounds. Puzyn et al. argued that it was impossible to develop
resulting model [68]. Ridge regression for linear models and a universal model for all nanosystems and that it is practi-
Bayesian regularization for neural network models can be cally impossible to establish one QSPR model for a very
employed to control the complexity (nonlinearity) of the wide applicability domain in the case of classical chemicals.
QSPR models. Bayesian regularized neural nets automati- The plot of the standardized residuals versus the leverage,
cally choose the optimum complexity of a QSAR model, Williams plot, is often used to visualize the applicability
achieving the best balance between bias and variance [53, domain. Fig. (3) displayed the applicability domain of a
54, 69]. QSPR model of SWNTs dispersibility in different organic
solvents using a William plot [78].
2.5. Model Validation
Model validation, as an important and crucial step, is re- 3. NANO-QSPR MODELLING OF PHYSICOCHEMI-
quired to ensure reliability of the models developed in QSPR CAL PROPERTIES OF CBNS
analysis. Development and validation of nano-QSPR models 3.1. Fullerene’s Solubility
is a difficult task in the lack of high-quality and sufficient
quantity of experimental data for nanomaterials. Unfortu- Solubility is an important physicochemical factor for
nately, some nano-QSPR reported using very small data sets pharmaceutical, environmental, and industrial applications of
(less than 25 data), were validated insufficient according to organic molecules [79]. Early knowledge of solubility is
common QSPR modeling practices such as OECD principles crucial for estimating transport properties of organic mole-
[70]. In the absence of enough data, the internal validation cules in drug development [79, 80] and is also important for
methods are utilized more to validate the QSPR models. environmental fate estimation, ecotoxicity and separation
Leave-one-out cross-validated squared correlation coeffi- science. Unfortunately, solubility is also one of the most
cient (Q2) as one of the most popular validation criteria [71], difficult molecular properties to predict using computational
N-fold cross-validation or leave-many-out [72], permutations models and a myriad of computational methods have been
of the data (randomization test or scrambling) that represent used to estimate aqueous solubility directly from molecular
the model robustness [72] are common internal validation structures [80]. A number of QSPR approaches have been
methods. A reliable and predictive QSPR model should be reported to develop reliable models for accurate prediction of
statistically significant and robust, be validated by making solubility of different organic compounds [80-82].
accurate predictions for external data set that not used in the
Different QSPR approaches including linear solvation
model development and have a defined domain of applicabil-
energy relationship (LSER) [83-86], neural networks and
ity [14, 71]. Models relying only on training and internal test
support vector machine [87-90], SMILES (simplified mo-
sets cannot reliably prove their robustness. Furthermore, the
lecular input line entry system) and InChI based optimal
separation of the external validation set should be independ-
descriptors [37-39, 91], and others two dimensional (2D)
ent of any modelling process. Validation of an external set is
QSAR methods [92-96] have been applied to predict the
one of the most widely used methods of correlation testing.
solubility of C60 in different organic solvents. Table 2 sum-
The purpose of the external validation is to evaluate how
marizes the characteristics of QSPR approaches for C60 solu-
well obtained model generalizes objects whose data have not
bility in different organic solvents.
taken part in the process of model development [73, 74].
Standard error of calibration (SEC) and standard error of Linear Solvation Energy Relationship (LSER), originally
prediction (SEP), as robust estimators, were used to assess developed by Kamlet and Taft [97], have been extensively
the predictive quality of the derived QSPR models since applied to study of nanomaterial properties prediction. LSER
these are independent on the number of parameters in the methods are specific subclass of thermodynamic relation-
model and size of the training set [75]. The root-mean-square ships known as linear free energy relationships (LFERs),
Nano-QSPR Modelling of Carbon-Based Nanomaterials Properties Current Topics in Medicinal Chemistry, 2015, Vol. 15, No. 18 1873

2.5

2.0

1.5
Train set
Test set
1.0

Studentized Residuals
0.5

0.0

-0.5

-1.0

-1.5

-2.0

-2.5
0.00 0.05 0.10 0.15 0.20 0.25 0.30 0.35 0.40 0.45 0.50 0.55 0.60

Leverage

Fig. (3). The Williams plot for QSPR model of SWNTs dispersibility in different organic solvents. Dotted lines represent ±2 studentized re-
sidual, dash line represents a warning leverage (h* ≈ 0.52).

which use solvatochromic parameters to describe the de- dependent variables by Abraham et al. [103]. Abraham sol-
pendency of physicochemical properties of a molecule to vation equation, consisting of the solute excess molar refrac-
solute-solvent interactions. LSER models typically express tivity, the solute dipolarity/polarisability, the overall or
in five main descriptors, including the solute size, polar- summation hydrogen-bond acidity and basicity, and the
ity/polarizability, hydrogen bond acidity, hydrogen bond McGowan characteristic volume, showed a good correlation
basicity and the excess molar refraction [98-100]. A most with C60 solubility in 20 investigated solvents. It indicated
common and widely accepted model of the LSER, repre- that C60 is not completely nonpolar, has no hydrogen bond
sented by Abraham [101] is expressed as follows: acidity but is a slightly stronger hydrogen basicity than com-
pounds such as acetophenone and behaves more like a poly-
SP = = c + eE + sS + aA + bB + vV (2) alkene than an aromatic system. Makitra et al. examined a
where SP is any free energy related property, c is the regres- two-parameter equation to correlate the solubility of C60 in
sion constant, E is the excess molar refraction, S is the po- 47 organic solvents [104]. The required descriptor were the
larizability/dipolarity, A is the hydrogen bond donating abil- solvent refractive index function, f(n)=(n2-1)/(n2-2) and the
ity, B is the hydrogen bond accepting ability, V is the mo- solvent cohesive energy density, δ 2. This simple equation
lecular volume, and the e, s, a, b, and v are the regression gave an correlation coefficient of 0.95 and RMSEC of 0.48.
coefficients. Multiple linear regression analysis is often util- In the next study, they used a larger dataset of 89 solvents
ized to fit the coefficients and building the model. These and by adding another descriptor, the molar polarization
descriptors are determined using experimental techniques, function f(ε)= (ε-1)/(ε+ 2), to the previous model and remov-
therefore LSER is an experiment-based QSAR/QSPR ap- ing eight outliers, a correlation coefficient of 0.87 with and
proach. These useful parameters are developed and exten- RMSEC value of 0.47 were achieved [105]. They simply
sively used to obtain QSPR models and predict chemical concluded that the solubility of C60 increases with increasing
properties of nanomaterials. solvent polarizability and decreases with increasing solvent
polarity and cohesive energy density.
LSER analysis was used extensively in the early reports
on fullerene solubility modelling. Sivaraman et al. reported Marcus et al. regarded the solubility of C60 in various
the first attempt to explain the trends in C60 solubility and solvents at two different temperatures of 298 and 303 K and
showed the correlation between experimental fullerene’s treated it by multivariate stepwise linear regression applied
solubility and calculated solubility parameter for 15 different as the linear solvation energy approach [106, 107]. They
organic solvents at 303 K [86]. Ruoff et al. also investigated constructed a four-variable in 55 solvents with R2c of 0.86
C60 solubility trends in a larger data set of 47 organic sol- and SEC of 0.4 at of 298 K, and a three-variable model in 20
vents, and examined the relation between C60 solubility and solvents with R2c of 0.72 and SEC of 0.52 at 303K [106].
various solvent properties such as index of refraction, dielec- The next study, they used larger data set of 113 and 32 solu-
bility data items for 298 K and 303 K respectively [107]. The
tric constant, molecular size, Hildebrand solubility parame-
QSPR model with four variables and 95 data set (18 outliers)
ter, and H-bonding strength [102]. These studies showed no
showed R2c =0.99 with SEC of 0.41 at 298K. With three
systematic way to explain the C60 solubility with no predic-
variables, the results gave an R2c of 0.99 and SEC=0.39 at
tive model, indicated just certain trends in solubility.
303K, leaving seven outliers and 25 data set. It indicated that
The solubility of C60 were transformed into octanol-water increasing molar volume and solvent polarity diminished the
partition coefficient (log P) and gas-solvent partition coeffi- solubility of C60, whereas electron pair donation ability and
cient or Ostwald solubility coefficient (logL) and used as polarizability enhanced solubility.
1874 Current Topics in Medicinal Chemistry, 2015, Vol. 15, No. 18 Maryam Salahinejad

Table 2. Summary of QSPR approaches for C60 solubility in different organic solvents.

Dataset Model Performance Descriptors Statistical Method Ref.

Nc=20 SEC=0.4 5 LSER [103]

Nc=47 R=0.95, RMSEC=0.48 2 MLR [104]

Nc=81 R=0.93, RMSEC=0.47 3 MLR [105]


2
Nc=55(298K) R c=0.86, SEC=0.4 4
LSER [106]
Nc=20(303K) R2c=0.87, SEC=0.52 3

Nc=120(298K) R2c=0.99, SEC=0.42 4


LSER [107]
Nc=25(303K) R2c=0.99, SEC=0.39 3

Nc=126 SEC =0.45 5 ANN [89]

Nc=76 RSMEC= 0.42, RMSEP=0.50 9 MLR


[87]
Np=20 RSMEC= 0.25, RMSEP=0.35 ANN

R2c=0.91, R2 p=0.91,
Nc=92 SVM
RSMEC= 0.10, RMSEP=0.15 5 [90]
Np=30 MLR
R2c=0.89, RSMEC= 0.13

Nc=100 R2c=0.91, Q2=0.89, RSMEC= 0.12 MLR


5 [88]
Np=24 R2c=0.94, Q2=0.94, RSMEC=0.07 ANN

Nc=45 (Aliphatics) R2c=0.91, SEC=0.26 4 MLR


[95]
Nc=25(Aromatics) R2c=0.94, SEC=0.21 5 MLR

Nc=29 R2c=0.87, Q2=0.79, RSMEC= 0.16


5 MLR [112]
Np=7 R2p=0.81, RSMEP= 0.22

Nc=25 R2c=0.81, SEC=3.6 a


1 MLR [96]
Np=11 R2p =0.79, SEC=4.67a

Nc=92 R2c=0.86, SEC=0.4, Q2=0.85


1 MLR [39]
Np=30 R2p =0.89, SEP=0.4

Nc=92 R2c=0.93, SEC=0.27, Q2=0.93


1 MLR [38]
Np=28 R2p =0.91, SEP =0.33,

Nc=92 R2c=0.95, SEC=0.25, Q2=0.94


1 MLR [37]
Np=30 R2p=0.93, SEP=0.26

Nc=92 R2c=0.86, SEC=0.41, Q2=0.84


4 MLR [94]
Np=30 R2p =0.9, SEP=0.35

R2c=0.88; Q2=0.72, SEC=0.39


5 SVM
Nc=92 R2p =0.87, SEP=0.43
[114]
Np=30 R2c=0.86; Q2=0.77, SEC=0.42 b
6 PLS
R2p =0.72, SEP=0.67

Nc=61 R2c=0.85; Q2=0.79, RSMEC=0.38


4 MLR [108]
Np=16 R2p =0.87, RMSEP=0.37

Nc=15 ( n-Heptane) R2c=0.90, RSMEC=0.18, Q2=0.84 3


MLR [93]
Nc=15 ( 1-octanol) R2c=0.7, RSMEC=0.08, Q2=0.93 3

Nc=18 R2c=0.76, SEC=17.6a, Q2=0.72


1 MLR [113]
Np=9 R2p =0.92, SEP=12.5a
a
S was used instead of logS, b: number of latent variable
Nc: number of data in training set, Np: number of data in prediction set, R: correlation coefficient, R2c: squared correlation coefficient of training set, R2p: squared correlation coeffi-
cient of prediction set, Q2: leave-one-out cross-validated squared correlation coefficient, SEC: standard error of calibration, SEP: standard error of prediction, RMSEC: root-mean-
square error in training, RMSEP: root-mean-square error in prediction.
Nano-QSPR Modelling of Carbon-Based Nanomaterials Properties Current Topics in Medicinal Chemistry, 2015, Vol. 15, No. 18 1875

A LSER model, based on solvent empirical parameters ume, polarizability parameter, LUMO energy, saturated sur-
for the prediction and description of the C60 solubility in face, and average polarizability, calculated using semi em-
various solvents was reported [108]. The authors collected pirical quantum-chemical methods, were chosen as relevant
‘‘Equilibrium and Kinetic’’, ‘‘Multiparameter’’ and ‘‘Spec- molecular properties and inputs of a back-propagation neural
troscopic’’ parameters for 79 organic solvents from the lit- network. Without model validation in the prediction of an
erature and applied them to build a four-variable multilinear external or test set, and with excluding some outliers the
QSPR model. The presented experimental-based QSPR model showed the SEC of 0.45. The results showed that in a
model yielded an R2c of 0.85 and SEC of 0.38, with remov- large number of solvents the solubility of C60 decreases with
ing two outliers outside the AD of the model. Besides the increasing molar volumes of the solvents and increases with
problem that the empirical descriptors are restricted to those their polarizability and saturated surface areas.
solvents with available experimental data, the reported ex- Danauskas and Jurs developed ANN and MLR models
perimental-based QSPR model showed no improvement in
predicting fullerene solubility in 96 diverse solvents [87]
the statistical performance when compared to other reported
using four types of descriptors including: topological, elec-
MLR models (Table 2).
tronic, geometric and geometric/electronic hybrids. The
Solubility parameter is an important quantity for predict- feed-forward computational neural networks model consist-
ing solubility relations and have been widely used to select ing of a 9-3-1 architecture, having RMSEC and RMSEP of
proper solvent, predict the compatibility of materials and 0.255 and 0.346 respectively. Randomization test was per-
characterize the surface interactions in industry or laboratory formed to show the absence of chance correlations between
applications. The Hildebrand solubility parameters and Han- independent and dependent variables. As authors claimed,
sen solubility parameters (HSP) are common empirical the nonlinear ANN model, that yielded lower error, 15.3%
methods for the prediction of solubility. The Hildebrand and 33.7% decrease in RMSEC and RMSEP respectively,
solubility parameter is defined as the square root of the cohe- gave a good prediction for an external test set.
sive energy density indicates the relative solvency behavior
Least-squares support vector machine (LSSVM) was util-
of a specific solvent. The cohesive energy density of a liquid
ized to develop C60 solubility models in 128 diverse organic
is directly proportional to the heat of vaporization which is a
solvents on the basis of CODESSA software consisting of
direct reflection of the degree of van der Waals forces hold-
constitutional, topological, geometrical, electrostatic, and
ing the molecules of the liquid together [109]. Hansen pa-
quantum-chemical descriptors of compounds reported by Liu
rameters divide the total Hildebrand value into three differ- et al. [90]. The data set was separated into a training set of
ent components as equation 3:
92 compounds and a test set of 30 compounds. Furthermore,
()2=()2 + ()2+ ()2 (3) a process of leave-one-out cross-validation of the training set
was performed. Linear and nonlinear solubility generated
where, δ t =total Hildebrand parameter; δ d =dispersion force
models gave satisfactory prediction results. The R2c was
component, δ p=polar component and δ h=hydrogen bonding
parameter. Hansen solubility parameters are widely used to 0.892 and 0.903, and RMSEC was 0.126 and 0.116 for
LSSVM and MLR models respectively. LSSVM nonlinear
correlate and predict the behavior of solvents.
models produced slightly better models with good predictive
Hansen and Smith [83] used Hansen solubility parame- ability than did the linear regression. SVM techniques are a
ters [110] to correlate the C60 solubility in a database of 89 powerful approach to modelling but they are quite complex
organic solvents and 30 polymers. The Hansen solubility and required to optimize the kernel parameters.
parameters δd, δp, and δh, were estimated for C60 as 19.7, 2.9,
Feed forward neural network (FFNN) approach was also
and 2.77 MPa1/2, respectively. These values were taken as a
used to consider the nonlinear behavior of different molecu-
three-dimensional coordinate system of the centre of the
lar descriptors for prediction of solubility of C60 in 124 or-
sphere containing good solvents in the HSP space. A simple
ganic solvents. Genetic algorithm was used to extract the
composite affinity parameter, Relative Energy Difference
statistically effective molecular descriptors. The proposed
(RED), was defined to classify organic solvent. Good sol-
vents showed the RED value less than 1 and progressively FNN model consisting of a 5-6-1 structure, having an R2c of
0.94 and RMSEC of 0.267. The results of FNN were better
poorer solvents had higher RED values. It concluded that
in comparison with the GA-MLR equation. Although the
both the good solvents and the compatible polymers for C60
data set divided to training and test set, no statistical results
solubility were polycyclic hydrocarbons and aromatic hydro-
reported for prediction ability of the model obtained [88].
carbons with halogen substitutions.
Huang explored the application of solution model of Sivaraman et al. proposed several regression equations
for each individual class of solvents as well as for combina-
Thomas and Eckert [111] with consideration of dispersion,
tions of these sets [95]. The models were based on topologi-
polar, acidity, and basicity interactions was used to correlate
cal indices, polarizability parameter and indicator variables
the solubility of C60 in 54 solvents [84]. The author found
denoting the structural details of the solvent molecules. The
that fullerene had dispersion and polar components near
values of R2c for individual subsets ranged between 0.91for
those that characterize aromatics, a small acidity, and a zero
basicity component and suggested that the model was able to Alkans to 0.99 for Cycloalkanes with the corresponding val-
ues SEC from 0.04 (Cycloalkanes) to 0.22 (Alkyl halides).
correlate solubility data ranging from alkanes to alcohols.
Although, the models suggested for solvents subsets showed
Kiss et al. applied artificial neural network (ANN) ap- good statistical performance and were better than the model
proach to introduce nonlinearity in QSPR model to predict for the combined data set, they suffered from small numbers
the solubility of C60 in 126 organic solvents [89]. Molar vol- of solvents and lack of an external validation set, therefore,
1876 Current Topics in Medicinal Chemistry, 2015, Vol. 15, No. 18 Maryam Salahinejad

they had limited application to particular families of sol- R2p=0.90. Their study revealed a correlation of HOMO en-
vents. It indicated that the variation in neither polarizability ergy (HOMO–LUMO gap), heteroatom fragments and geo-
parameter nor structural parameters encodes the changes in metrical parameters with solubility of C60.
solvent-solute interaction due to changes in substituent, their
In contrast with the other papers, where the solubility of
combination, and substitution pattern in the carbon skeleton.
C60 in different organic solvents was modelled, the last two
Another study was reported the QSPR modeling of C60 solu-
bility for an individual class of solvents [112]. The authors works of Table 2, represented QSPR models for C60 solu-
claimed that they developed the models in various organic bility in one solvent. Two general QSPR models were de-
solvents, but the models were constructed only for 36 ben- veloped by Martin et al. [93], describing the solubility of
zene derivatives solvents. The presented models, with R2 c 14 polycyclic aromatic hydrocarbons (PAH) and C60 in two
and R2p less than 0.87 and 0.81 respectively, showed less different condensed media, an average polarity solvent with
statistical performance compared to models obtained by an OH group (1-Octanol) and a nonpolar solvent (n-
Sivaraman et al. [95] for C60 solubility in aromatic solvents Heptane). The CODESSA program descriptors [92], in-
(Table 2). cluding constitutional, topological and geometrical descrip-
tors and quantum chemical descriptors were calculated. The
Toropov and coworkers created some QSPR models of proposed three-parameter QSPR models were derived with
the solubility of C60 in different organic solvent using Heuristic and best multilinear regression (BMLR) descrip-
SMILES [38, 39, 96] and InChI [37] based optimal descrip- tor forward selection approaches. The solubility QSPR
tors. Optimal descriptors correlation weights (DCW) are model in 1-octanol gave an R2 c and Q2 of 0.96 and o.93
calculated from SMILES or InChI text-based descriptions of respectively, while the solubility model in n-heptane
a molecule. At the first attempt, they created a one-variable showed less statistical performance with an R2c of 0.9 and
QSPR model of C60 solubility using a DCW calculated from Q2 of 0.83. The reliability of QSPR models was assessed
rules of the SMILES fragments for a data set of 36 substi- via leave-one-out and leave-50%-out methods and both
tuted benzene solvents [96]. DCW defined as the production showed statistically good performance. However, the reli-
of SMILES fragments which optimized by Monte Carlo ability in predictions for C60 solubility has been questioned
method and expressed as: by Puzyn et al. [15] due to structural difference between
DCW= k=1NCW(Sk) (4) PAHs and the fullerene, as well as the large difference in
observed solubilities of fullerene compared to the poly-
where Sk is as a rule one character fragment of the SMILES cyclic aromatic compounds. The next one, reported by
notation. The obtained model statistically characterized by Toropova's group, described the QSPR models between the
R2c=0.81 with SEC of 3.60, and for the training set and R2p molecular structure of C60 and C70 fullerene derivatives and
of 0.79 with SEP=4.67, without transformation the solubility their solubility in chlorobenzene [113]. The SMILES-based
to log unit. They agin used the DCW to produce one-variable optimal descriptors calculated using CORAL freeware [91]
QSPR model of C60 solubility in 122 different organic sol- gave a reasonable prediction of the solubility for both C60
vents, while, here, the DCW defined as summation of and C70 fullerene derivatives.
SMILES fragments optimized by Monte Carlo method [39].
Statistical characteristics of the model (Table 2), showed All of the papers appeared in the literature used 2D-
slightly better than their previous work for a subclass of sol- QSPR modelling of solubility of C60 , and in our knowl-
vents of substituted benzene. They also tried to build nonlin- edge, there is no publication in 3D-QSPR of C60 solubility
ear SMILES based models using the DCW2 or DCW0.5 but in various organic solvents. Recently, the 3D-QSPR model-
the statistical quality of the models of the logS was not im- ling has been examined as an efficient tool for predicting
proved. At the next attempt, they modified the DCW as solubility of fullerene C60 [114] and dispersibility of
combinations of the SMILES attributes. The multiplicative SWNTs [78] in different organic solvents. 3D-QSPR refers
SMILES based optimal descriptors technique provided im- to use of force field calculations to compute spatial proper-
provement over the previous models for predictions of ties of three-dimensional structure (3D) of compounds.
solubility of fullerene in organic solvents [38]. Molecular interaction fields (MIF) between chemical
probes and every atom of a molecule at surrounding points
In the similar study, they used DCW (InChI) as the corre- of a predefined grid box [115, 116]. GRid-INdependent
lation weight of the InChI attribute to construct QSPR mod- Descriptors (GRIND) are alignment independent three di-
els in various organic solvents [37] which defined as mensional descriptors derived from 3D maps of interaction
DWCInChI=WIk, where the Ik was InChI attribute and W(Ik) energies between the molecule and chemical probes based
was the correlation weight for the Ik. It concluded that the on GRID molecular interaction fields (MIF) [117]. GRIND
solubility model obtained from InChI-based optimal descrip- descriptors are calculated as a function of the distance in-
tors are better than models based on optimal descriptors stead of the position of each grid point and therefore insen-
which were calculating with SMILES-based descriptor [37].
sitive to the position and orientation of the molecular struc-
The same research group utilized Ab initio quantum- ture in the space to overcome the alignment problem. MIF
chemical descriptors to build QSPR model of C60 solubility are generated around the molecule by probing different
on the same data set of 122 organic solvents. Genetic algo- chemical probes such as DRY, N1 , O, and TIP probes. The
rithm and multiple regression analysis (MLRA) was used to DRY probe represents hydrophobic interactions, N1 (am-
construct one-, two-, three- four- and five-variable QSPR ide) and O (carbonyl) probes represent hydrogen bond do-
models of C60 solubility [94]. The best performance was ac- nor and acceptor groups, respectively and the TIP probe
complished by the four-variable MLRA model with represents a shape-field.
Nano-QSPR Modelling of Carbon-Based Nanomaterials Properties Current Topics in Medicinal Chemistry, 2015, Vol. 15, No. 18 1877

The VolSurf approach generates 3D maps of molecular SVM analysis resulted in a model with five latent variables
interaction field (MIF) based on probing the grid around a and an R2c=0.88; Q2=0.72 and R2p=0.87. The SEC and SEP
target molecule with different chemical probes, and converts of SPA-SVM model were 0.39 and 0.43, respectively.
the calculated MIFs to simple and interoperable molecular
The five most significant VolSurf descriptors selected
descriptors [118]. By using various probes such as hydrogen
by SPA procedure were: critical packing parameter (CP),
bonding interactions, hydrophobic interactions or donor
group with specific cutoff values for energy were utilized to that represents the concentration of hydrophobic interac-
produce different molecular properties. These descriptors tions on the molecular surface, polarizability (POL), flexi-
quantify the polarity, shape, size and hydrophobicity of bility parameter (Flex), capacity factor (CD1), which repre-
molecules, including molecular volume and surface, hydro- sents the ratio of the hydrophobic volume over the total
philic and hydrophobic regions. molecular surface, and intrinsic solubility (SOLY) which
calculated from thermodynamic solubility model and de-
GRIND and VolSurf descriptors were successfully em- fined as aqueous solubility in Mol/liter at 25o C. Fig. (4)
ployed to develop predictive 3D-QSPR model of C60 solubil- showed the VolSurf GRID fields obtained with OH2 and
ity in 132 different organic solvents [114]. Two variable se- DRY probes for a strong solvent, 1-Phenylnaphthalene.
lection techniques namely successive projections algorithm Fig. (5) displayed the VolSurf GRID fields obtained with
(SPA) [119] and genetic algorithm (GA) [120] were used to OH2 , DRY and N1 probes for Tetrahydrofuran as a very
extract the more informative Volsurf descriptors and build weak solvent for C60 . As concluded from Figs. (4 and 5),
more predictive models. the VolSurf GRID fields from DRY probe had no expres-
The data set divided into three different sets of training, sion for the Tetrahydrofuran (Fig. 5B), while a large hy-
test and external sets. Four solvents with indistinctive value drophobic region located around 1-Phenylnaphthalene as a
of C60 solubility (log S<−8), also considered as second ex- strong solvent for C60 (Fig. 4B). GRID fields obtained with
ternal set. SVM and PLS were used to construct several N1 probe had a large area on Tetrahydrofuran but its not
QSPR models. The SVM model based on Volsurf descrip- any region around 1-Phenylnaphthalene, indicating the in-
tors extracted using the SPA method showed the best statis- significant role of hydrogen bond donor of solvent in C60
tical performance in C60 solubility modelling. The SPA- solubility.

(A) (B)
Fig. (4). VolSurf GRID fields from A) OH2 probe and (B) DRY probe at energy level of -1kcal/mol of 1-Phenylnaphthalene as a strong sol-
vent for C60 (log S= -1.9).

(A) (B) (C)


Fig. (5). VolSurf GRID fields from A) OH2 probe, (B) DRY and (C) N1 probe at energy level of -1kcal/mol of Tetrahydrofuran as a very
weak solvent for C60 (log S<-8).
1878 Current Topics in Medicinal Chemistry, 2015, Vol. 15, No. 18 Maryam Salahinejad

The GRIND descriptors were related to C60 solubility in and sign of the enthalpy of mixing (ΔHmix) as discussed in
different organic solvents by means of PLS analysis with six details elsewhere [122-124]. The interactions between sol-
latent variables. The GIND model with the FFD variable vent molecules and CNTs are very important to successfully
selection method gave an R2c =0.86; Q2=0.77 and R2p=0.72 disperse a nanotube in a solvent. There is not numerous and
with the SEC and SEP of 0.42 and 0.67 respectively. Fig. (6) sufficient experimental data on dispersion of carbon nano-
depicted the PLS coefficient plot of chemical descriptors, tube in organic solvents, hence a few QSPR models reported
which were directly or inversely correlate to the C60 solubil- on the dispersibility modeling of carbon nanotubes. Most
ity. The largest peaks of PLS coefficients were related to studies explored the dependency of solubility parameters and
DRY and TIP probes, which represent hydrophobicity, shape CNTs dispersibility.
and size of the molecules, whereas N and O probes showed
no significant peaks. A deeper inspection of PLS plot re- Ham et al. explored the dispersions of SWNTs in various
vealed that the variable of DRY-TIP: distance 7.2-7.6 Å, organic solvents and aqueous surfactant emulsions and man-
explained a positive and the largest impact on C60 solubility aged to correlate the degree of dispersion state with Hansen
and variable of DRY-TIP at distance of 2.8-3.2 Å, showed solubility parameters [125]. It concluded that organic sol-
an inverse relationship between this variable and C60 solubil- vents with certain dispersive component (δd) values dis-
ity. This indicated that adjacent of hydrophobic groups, in persed SWNTs better compared to solvents with high polar
the distance about 3 Å, could reduce the interaction between component (δp) values or hydrogen-bonding component (δh)
the solvent molecules and fullerene with spherical surface values. The dispersion states of the nanotubes in various or-
area. The next two important GRIND descriptors related to ganic solvents were classified into three groups as “dis-
DRY-DRY blocks at distance of 1.6-2Å and 6-6.4Å respec- persed,” “swollen,” and “sedimented”. Nanotubes were pre-
tively. It is well known that DRY probe interacts with differ- cipitated in the solvents with high polar component (δp) val-
ent types of π systems (aromatic or vinyl type) but have not ues or hydrogen-bonding component (δh) values. Surfactants
high affinity to aliphatic moieties. Thus, variables related to dispersed SWNTs if they had hydrophobic alkyl chains equal
DRY probe had small region in the weak solvent for C60 to and longer than a decyl group, regardless of their func-
solubility (Fig. 5B). Validation, reliability and robustness of tionality or their hydrophilic heads. Application of the Han-
models obtained checked by the prediction of external test sen solubility parameters theory to several types of carbon
sets, leave-one-out Q2 and progressive scrambling approach. nanotubes in common organic solvents was studied by
Detriche et al. [126]. The top four solvents for CNTs dis-
The results obtained from Volsurf and GRIND 3D-QSPR persibility had 19<δd<24 MPa1/2, 5<δp<8 MPa1/2, and 3<δh<5
methods confirmed that the hydrophobicity and the shape MPa1/2. They showed that the mixture of two good solvents
parameters (steric effects) of solvent molecules had an effec- can give better dispersibility than a single solvent.
tive impact on C60 solubility in different organic solvents.
Bergin et al. measured the dispersibility of single-walled
3.2. CNTs Dispersion carbon nanotubes in a range of organic solvents and explored
the nanotube dispersibility as a function of the Hansen solu-
Geckeler and Premkumar recommended to use the dis- bility parameters of the solvents [127]. A new set of solubil-
persion term when dealing with carbon nanotubes [121], due ity parameters was suggested, based on the surface energies
to ability of CNTs in formation of colloids with respect to of nanotubes and solvent. They claimed the level of SWNTs
their diameters and lengths. However, both solubilization dispersibility was more sensitive to the dispersive Hansen
and dispersion processes strongly depend on the magnitude parameter than the polar or H-bonding Hansen parameter.

Fig. (6). PLS coefficient plot of GRIND descriptors of organic solvents for C60 solubility.
Nano-QSPR Modelling of Carbon-Based Nanomaterials Properties Current Topics in Medicinal Chemistry, 2015, Vol. 15, No. 18 1879

However, they mentioned that the role of polar and H- the best statistical performance with R2c = 0.99, Q2 = 0.93,
bonding solubility parameters on the dispersion of SWNTs RMSEC of 0.072. The ERM-PLS model satisfied a set of
could not be neglected. rigorous validation criteria and performed well in the predic-
tion of an external test set with an R2P of 0.98 and RMSEP=
Cheng et al. reported a systematic study of the effect of a
0.072. Table 3 shows the most important GRIND descriptors
series of organic solvents on the dispersion of as-produced
involved in ERM-PLS modelling of SWNTs dispersibility in
SWNTs. The dispersibility of SWNTs in different solvents
different organic solvents with their PLS coefficients. The
was quantitatively assessed in terms of the absorption and
DRY-N1 at 6.4-6.8 Å had the highest coefficient with posi-
scattering parameters. Their correlation with the Hildebrand
tive correlation on SWNTs dispersibility. This variable is
solubility parameters and three-dimensional Hansen solubil- strongly expressed in solvents with donating groups in aro-
ity parameters was investigated with the aim of establishing matic backbone and shows week value in aliphatic backbone.
structure-property relationships governing the solubilizing This concludes that a distance of about 6.6Å between hydro-
process [128]. Interestingly, in contrast to the Ham and co- phobic moieties and hydrogen donor groups of solvent's
workers, Cheng et al. concluded that the dispersion compo- structure has the most favorable influence on the SWNTs
nent (δd) seems not to be the dominant factor for dispersing dispersibility. Fig. (7) shows the graphical display of GRIND
SWNTs, while polar component (δp) and hydrogen-bonding variable DRY-N1: distance 6.4-6.8Å for a good solvent (N-
component (δh) are important factors for obtaining SWNT Cyclohexyl-pyrrolidinone) and a weak solvent (benzyl alco-
dispersions. In a follow-up study, they explored the correla- hol) in the dispersion of SWNTs. The TIP probe, which indi-
tions between the dispersion limit of SWNTs and solvent cate the shape and size of the molecules, comprised four
solubility parameters, including the Hildebrand solubility variables having significant impact on the model, whereas
parameter and three-dimensional Hansen solubility parame- the O probe representing hydrogen bond acceptor showed no
ters [129]. It was identified that degree of SWNTs dispersion important peaks. The N1-TIP variables indicate the favorable
depends critically on sample preparation conditions and in position of a hydrogen bond donor site or a polar group with
particular sonication time. It was found that SWNTs easily respect to the size and shape of a molecule. Remarkably, this
dispersed in solvents with Hildebrand solubility parameter variable showed an inverse effect at short distance of ap-
ranged from 22-24 MPa1/2 and Hansen polarity component proximately 3.6Å and but had a positive effect on distance
(δp)12-14 MPa1/2 and no clear correlation between disper- longer than 11 Å, on SWNTs dispersibility (Table 3). The
sion limits and the dispersion force (δd) or hydrogen bonding inverse effect of DRY-TIP variable at distance: 7.6-8 Å, in-
force (δh) were evident. One limitation of the aforementioned dicated that adjacent of a π system (aromatic backbone) and
studies is that the models are not well-represented and there- a bulky group in the distance about 7.8 Å, could reduce the
fore cannot be reproduced to predict dispersibility of CNTs interaction between the solvent molecules and SWNTs. It
for new organic solvents. concluded that, the hydrophobic interactions in the small
Some studies to investigate the relationship between solvent molecule with polar or bulky groups were restricted
chiral vectors and the water solubility (logS) and partition and caused a reduction of the SWNTs dispersibility in the
coefficient (logP) of SWNTS were reported by Torrens [130, solvent. Although the GRIND variables defined as distances
131]. Toropova et al. performed a QSPR study on the data between molecular features, but the importance of TIP, DRY
reported by Torrens and represented a simple two-variable and N1 probes, in a preliminary analysis, suggested that the
models of water solubility and octanol water partition coeffi- presence and orientation of hydrophobic and polar groups
cient of SWNTs based on chiral vectors of nanotubes as (hydrogen bond donor group) as well as the size and shape
structural descriptors [132]. The models obtained showed a of the solvent molecules were crucial for the SWNTs dis-
great performance. Unfortunately, in these models in addi- persibility. The analysis of the applicability domain of the
tion to use very small data sets (eight training and eight pre- ERM-PLS model based on GRIND descriptors is displayed
diction set), the water solubility and partition coefficient in Fig. (3). All the training and validation compounds were
which were derived from computational models, was used to within the chemical domain, indicating that there was no
build models actually building a model of a model [17] not a chemical outlier or structure influential compound outliers
model of experimental data. and the predictivity of the model was reliable.

We applied GRIND methodology for modeling of Table 3. Most important GRIND descriptors involved in
SWNTs dispersibility in a series of 29 different organic sol-
ERM-PLS modelling of SWNTs dispersibility in dif-
vents [78] which their experimental values were reported by
ferent organic solvents.
Bergin et al. [122]. Log-transformed of maximum concentra-
tion of SWNTs in organic solvents (logCmax) was used in the
following QSPR modellings. The data set split into a training Variable Distance (Å) PLS-Coefficient
of 22 solvents and a prediction set of 7 solvents. Five vari-
able selection techniques including fractional factorial design DRY-N1 6.4-6.8 0.91
(FFD) [51], stepwise multiple linear regression [133],
successive projection algorithm (SPA) [50], genetic algo- N1-TIP 10.8-11.2 0.62
rithm (GA) [47] and enhanced replacement method (ERM)
TIP-TIP 2.4-2.8 -0.32
[48] were used to extract the more informative factors from
GRIND descriptors and generate more predictive models. DRY-TIP 7.6-8 -0.49
Various PLS models were constructed based on selected
variables. ERM-PLS model, with six latent variable, gave N1-TIP 3.6-4 -0.60
1880 Current Topics in Medicinal Chemistry, 2015, Vol. 15, No. 18 Maryam Salahinejad

Fig. (7). Graphical display of GRIND variable DRY-N1: distance 6.4-6.8 Å for A: a good solvent (N-Cyclohexyl-pyrrolidinone) and B: a
weak solvent (benzyl alcohol) in dispersion of SWNTs.

In another study, we developed some QSPR models, able relationship between Hildebrand solubility parameter, Han-
to correlate and predict the dispersibility of SWNTs at the sen solubility parameters and surface tension and graphene
same organic solvents data set based on different kind of 2D dispersibility. It concluded the good solvents for graphene
descriptors [134]. Three new amidine solvents reported by should had Hansen parameters of δd=18.0, δ P=9.3 and δ h=7.7
Barman et al. [135] were used as an external set. A feed for- MPa1/2, and by a Hildebrand solubility parameter of δ t23
ward neural network trained by back-propagation algorithm MPa1/2. The authors suggested that the level of dispersibility
and MLR were applied to model constructions. The best of graphene was more sensitive to the dispersive Hansen
model with four-variable MLR model was defined as fol- solubility parameter, but surprisingly, nonzero values of the
lows: polar and H-bonding Hansen parameters are required for a
solvent to disperse graphene.
log(Cmax) = 2.861+0.402(GCUT_SLOGP_3)-
2.785(AM1_Eele)-3.283(diameter)+0.685(b_rotR) (5) The only QSPR study on graphene dispersion in different
2 2
R c = 0.88, RMSEC= 0.14, Q =0.82, R2P=0.07, RMSEP=0.08 solvents has been reported by Yousefinejad and Hem-
mateenejad [140]. They used the data reported by Hernandez
GCUT_SLOGP calculates log P based on atomic contri- et al. [139] and performed MLR analysis with stepwise se-
butions and a modified graph distance [136], b_rotR and lection of variables to construct two linear QSPR models. In
AM1_Eele variables were a fraction of rotatable bonds and the first model, a three-variable MLR model, based on sol-
electronic energy respectively. The model obtained eluci- vent empirical parameters, gave an R2c=0.85 with SEC=0.17
dated the contributions of steric (diameter), polarizability and with seven solvents collected as prediction set the model
(GCUT_SLOGP), molecular flexibility (b_rotR) and electro- yilded R2p=0.9. The solvent polarity/polarizability, the
static (AM1_Eele) interactions to the dispersibility of maximum absorption band of copper(II) N,N,N',N'-
SWNTs in organic solvents. The ANN models taking into tetramethylethylenediamine acetylacetonate complex in pure
accounts both linearity and non-linearity in model construc- solvent and 19
F-NMR shielding of para-Nitro
tion, yielded a slightly more predictive models than MLR f1uorobenzenes as a shielding probe in a solvent, extracted
analysis. from the literature, were included in MLR model. The sec-
ond model with four theoretical descriptors, derived from
3.3. Graphene Dispersion Dragon software and DFT calculations, yielded an R2c 0f
Graphenes are supplied in granular form, while the avail- 0.81with SEC of 0.18 and R2p of 0.76. Although the model
ability of processable graphene in single-layer sheets are a based on the solvent empirical scales showed better results in
prerequisite for exploiting their optimal properties in most comparison to the model based on the theoretical structural
applications. Micromechanical synthesis of mono layer gra- parameters, but both models confirmed the significant role of
phene is extremely low yielded techniques and not applica- dispersive interactions compared to H-bonding and polar
ble for large scales [137], hence the exfoliate graphite into interactions on the dispersibility of graphene in organic sol-
monolayer sheets in liquid phase directly or disperse gra- vents.
phene powder in aqueous solution or organic solvents [138]
are of particular importance. 3.4. Adsorption onto CNTs
Hernandez et al. measured the dispersibility of graphene Adsorption is a surface phenomenon in which a sub-
in 40 various organic solvents [139]. They also explored the stance, in gas or liquid phase, accumulates on a solid surface.
Nano-QSPR Modelling of Carbon-Based Nanomaterials Properties Current Topics in Medicinal Chemistry, 2015, Vol. 15, No. 18 1881

The large surface area and chemical properties of CNTs ability was the second most predominant factor with a nega-
make them as a strong adsorption capacity for a variety of tive correlation which indicates that the sorbent surface has a
synthetic organic contaminants. Adsorption mechanisms weaker tendency to donate protons than water at the liquid-
mainly govern by the effects of properties of both CNTs and solid interface.
organic chemicals along with environmental conditions and QSPR and LSER techniques were used to develop pre-
it is also influenced by CNTs suspendability, which strongly
dictive models for adsorption of organic contaminants onto
affects their mobility, exposure and risk in the environment
MWNTs, reported by Abul et al. [146]. Adsorption data for
[141]. Therefore, understanding the adsorption phenomena
29 of low molecular weight aromatics by MWCNTs with
of organic contaminants to CNTs is of great importance for
less than 5% oxygen content were used as training set to
environmental applications of CNTs as special adsorbents
generate MLR models. QSPR model with three molecular
and assessment of the fate and transport of toxic chemicals connectivity indices gave an R2 of 0.88 and LSER equation
and organic contaminants with CNTs in the environment.
with four solvatochromic descriptors yielded an R2 of 0.83.
The adsorption of some polar and nonpolar organic com- The validities of the developed QSPR and LSER equations
pounds with varied physical-chemical properties to graph- were evaluated using 30 aromatics of the data reported by
eme, one SWNT and two MWNTs was evaluated [142]. Xia et al. [145]. The higher error value the LSER model in-
Three adsorbents groups, including: nitro aromatics (polar dicated that the QSPR approach has the stronger prediction
group), nonpolar aromatics (benzene, methylated, and chlo- ability than LSER approach. The combination of training
rinated benzenes) and nonpolar aliphatics (cyclohexane) ex- and validation data for all 59 compounds which were used to
hibited different adsorption affinity on CBNs. The adsorp- obtain a combined LSER equation gave an R2 of 0.84. They
tion affinity was correlated poorly with hydrophobicity but also generated LSER equations at different concentrations
increased in the order of nonpolar aliphatic<nonpolar aro- (infinite concentrations, 1% and 10% of adsorbate aqueous
matics<nitro aromatics, and within the group of nitro aro- solubility). In the next study, Apul et al. conducted three
matics, the adsorption affinity increased with the number of different learning approaches, MLR, ANN and SVM, to de-
nitro functional groups. The strong adsorptive interaction velop QSPR models for adsorption data for 39 aromatic
between carbon nanotubes and nitro aromatics was due to compounds onto MWCNTs [147]. The same validation set
the π-π electron-donor-acceptor (EDA) interaction between of 30 aromatics was used. In the model development proc-
nitro aromatic molecules (electron acceptors), while the ess, for eliminating the surface area effect of MWCNTs the
strong π-electron-dependent polarizable interactions between adsorption coefficients were normalized by specific surface
CNTs and nonpolar aromatics and cyclohexane might have area and the log unit of surface area normalized adsorption
contributed to the adsorption of nonpolar aromatics to carbon coefficients (log KSA) was used. The number of nitrogen and
nanotubes. oxygen atoms in rings, Kow (with positive correlation), the
dipole moment of the molecule and the number of hydrogen
Yang et al. reported the first systematic study on poly-
bonds (negative correlation) are the dominant variables in
cyclic aromatic hydrocarbons (PAHs) sorption by various
carbon nanomaterials [143]. They investigated the adsorp- QSPR modelling of adsorption coefficient of aromatics onto
MWNTs. A five-variable MLR model yielded an R2c of 0.76
tion of naphthalene, phenanthrene and pyrene onto six car-
and R2p of 0.59. ANN and SVM models resulted in better
bon nanomaterials, including fullerenes, SWNTs and
agreement between predicted and measured values with R2 of
MWNTs. They presented a single parameter linear relation-
0.83 and 0.78 respectively, showed the nonlinear relation
ship between the adsorbed volume capacity and the specific
between logKSA and physicochemical properties.
surface area. The next study by the same authors, the aque-
ous adsorption of a series of 13 phenols and anilines by In a recently published paper, Zhao et al. introduced a
MWNTs were examined and employed LSER to investigate modified model of concentration-dependent polyparameter
the adsorptive behaviors of nondissociated species [144]. A linear free energy relationships (pp-LFERs) model to de-
single variable linear model based on the chemical hydro- scribe the compound-CNTs interactions and to predict sorp-
gen-bonding donor parameter (Rm) values of organic chemi- tion behavior of chemicals on CNTs in a wide range of con-
cals was presented. They concluded that besides the van der centrations [148]. This modified model combines LFERs
Waals force, H-bonding interactions from solutes as hydro- parameters with aqueous equilibrium concentrations of com-
gen bonding donors and followed by π-electron polarizabil- pounds in a wide range of concentrations, from -5 to 0 of log
ity, may play important roles on the adsorption of phenols Ce/Cs, where Ce represents the equilibrium concentration and
and anilines by CNTs in the aqueous environment. Cs is the solubility of the compounds, in MWCNT15. A new
model was therefore obtained as follows:
Xia et al. successfully applied LSER approach in predict-
ing a mixture of 28 aromatics, to depict relative contributions logki=f1+ f2Ri + f3Pi + f4 Ai+ f5Bi+ f6Vi i=1,2,3,…n (6)
of molecular interactions on CNT adsorption [145].
Where f1, f2, f3, f4, f5, and f6 are the quadratic functions of
logk = −1.33 + 0.043R + 1.75P − 0.37A− 2.78B+ 4.18V (5) log(Ce/Cs) and describe the relative contribution of interac-
tion strength from molecular force of lone-pair electrons,
n=28, R2c=0.93, F=63, Q2 =0.89, R2p=0.78
dipolarity/polarizability, hydrogen-bond acidity, hydrogen-
External validation was conducted using 12 different bond basicity and hydrophobic interactions, respectively.
compounds, suggesting satisfactory prediction for the exter- The LFER equations were represented at various concentra-
nal validation compounds. According to their LSER equation tions for MWCNT resulted in models with R2>0.8, RMSEC<
(5), hydrophobicity had the strongest contribution to adsorp- 0.5 and Q2>0.7. They concluded that at low Ce
tion, represented by the V term. Hydrogen bond donating (<0.000753mg/L) π-π stacking interaction was the most sig-
1882 Current Topics in Medicinal Chemistry, 2015, Vol. 15, No. 18 Maryam Salahinejad

nificant and the hydrogen-bond basicity interaction as the a model [71]. As mentioned earlier, the model with three or
less dominant interaction while at Ce>0.000753mg/L, hydro- four data points per variable encountes a meaningless chance
phobic interaction became the most dominant interaction. correlations [43]. When too many variables are involved in
MLR analysis, the model complexity problem are occurred
Xia et al. proposed a biological surface adsorption index
(BSAI) to characterize surface adsorption energy of nanoma- and the model is susceptible to overfitting. Overfitted models
generally perform poorly when are used to make predictions,
terials [27]. It is assumed that the adsorption properties of
although they have shown low error in training set [150].
nanomaterials are governed by five fundamental forces of
DFT simulations are carried out to determine preferential
molecular interactions in biological processes, including
interaction modes, binding conformations, and underlying
Coulomb force (charged particles), London dispersion (hy-
driving forces of the molecule-SWCN systems. They re-
drophobic interactions), hydrogen-bond acidity and basicity,
dipolarity/polarizability and the lone-pair electrons. A set of sulted the preference of the bridge configuration in mole-
cule-SWCN interaction mode and the main effect of π–π
small molecule probes was allowed to competitively adsorb
stacking in the adsorption of simple benzene derivatives on
onto a set of nanoparticles and the adsorption coefficients of
SWNTs, while the substituent play a secondary effect on the
the probe molecules were measured and used to create a set
adsorption process. The predicted DFT energies do not show
of nanodescriptors representing the contributions and relative
a strictly positive or negative correlation with experimental
strengths of each molecular interaction. This method suc-
cessfully predicted the adsorption of coefficients set of 28 adsorptions.
simple benzene derivatives, as molecule probes, onto the
MWCNTs. Abraham descriptors were used to demonstrate 3.4. Mechanical Properties of CNTs
the BSAI approach and an LSER equation was developed The extraordinary mechanical properties of CNTs, e.g.
with R2=0.93 and Q2 of 0.89. The nanodescriptors of high elastic modulus and tensile strength make them an ideal
MWCNTs representing the relative contributions of the four choice for mechanical reinforcement of polymers, metal ma-
types of molecular interactions. The lipophilicity is a strong trices, and ceramic for improved stiffness and strength [151].
contributor with positive impact on adsorption coefficients Numerous experimental studies have been performed to de-
and Hydrogen-bond basicity is the second most significant termine the mechanical properties of CNTs, but experimental
factor with negative impact, suggests that the sorbent surface measurements is a tedious task which shows highly different
has a weaker tendency to donate protons (to the probe com- results due to differences in the experimental methods used.
pounds) than water at the liquid–solid interface. External Hence the simulating CNTs with quantum mechanics are
validation was conducted using 12 different compounds, widely used to predict their mechanical properties at infinite
given a R2p of 0.78. In a follow-up study, the authors deter- length [151]. A number of QSPR studies were reported on
mined the surface adsorption forces of 16 different nanoma- computed physicochemical properties of CBNs. The proper-
terials including SWNTs, C60, MWCNT-OH, l-MWCNT, ties derived from computational methods, usually quantum
MWCNT8nm-COOH, MWCNT15nm-COOH, MWCNT- mechanical (QM) techniques, were used as independent
40nm-COOH, carbon hollow sphere and graphene oxide
variable to build QSPR models, in fact building a model of a
[27]. The same probe compounds of 28 benzene derivatives
model not a model of experimental data. However the ob-
were used. The R2 of MLR analysis ranged from 0.84 for
tained QSPR models based on these purely computational
nC60(OH)20 to 0.95 for MWCNT40nm-COOH. When the
data set, in the absence of reliable experimental data set, can
five-dimensional information of the nanodescriptors was
give primary estimation on desired property and would be
reduced to two dimensions, the 16 nanomaterials were clas-
straightforward by adding experimental data and recreating
sified roughly into three clusters according to their surface
adsorption properties. The studied nanomaterials roughly the mathematical models.
clustered as strong adsorption nanomaterials including car- Young’s modulus (Y) is a measure of stiffness and de-
bon hollow spheres, carbon nanotubes and their derivatives, fined as the ratio of axial stress to axial strain over ranges of
medium adsorption nanomaterials including C60 (powder), stress in which Hooke’s law holds true. The Possion's ratio is
polar derivative of fullerenes and graphene oxides, and weak the ratio of the lateral contraction strain to longitudinal ex-
adsorption nanomaterials such as oxide and metal nanoparti- tension strain when a material is stretched elastically uniaxi-
cles. ally [152].
In a recently research, Liu et al. employed QSPR model- Borders et al. developed 20 descriptors from molecular
ing and DFT simulations to explore the structure-dynamics- dynamics (MD) simulations to build QSPR for Young’s
adsorption relationship of simple benzene derivatives with modulus and Poisson’s ratio in two separate analyses: va-
CNTs [149]. The same data set of Xia et al. [145] and their cancy only and vacancy plus methyl functionalization [31].
adsorption activities were used to QSPR model construction. The data set contained 78 SWNTs covering a 10×range of
Eight selected descriptors, through the stepwise method, was radii, all three chiralities, surface vacancies, surface func-
used to generate the MLR-based QSAR model. The QSAR tionalization, and multiple chiral angles. Many QSPR mod-
model exhibits a R2c of 0.986, Q2 of 0.968 with SE of 0.109. els were presented with different descriptors and at different
With eight descriptors involved in MLR model of 25 data, SWNTs sizes. Each model was trained with 80% of the data
the proposed MLR-QSPR model constructed with almost and 20% used as an external test. For vacancy only test cases
three data points per descriptor without any validation of CN2/CT (number of non-sp2 hybridized carbons per the total
prediction set and tests on chance correlations. Q2 is not the carbons) and chiral angle was identified as critical descrip-
sufficient criteria to examine the model validation and the tors for both Young’s modulus and Poisson’s ratio. The two
high value of Q2 can not guarantee the high predict ability of variable QSPR model gave R2c values of 0.92 and 0.87 for
Nano-QSPR Modelling of Carbon-Based Nanomaterials Properties Current Topics in Medicinal Chemistry, 2015, Vol. 15, No. 18 1883

Young’s modulus, and Poisson’s ratio respectively. Further executed modelling will continue to be a promising and in-
analysis and literature findings indicated the effect of chiral deed a valuable tool in the most areas of nano research.
angle is negligible at larger CNT radii for both properties. In
the analysis of vacancy plus methyl functionalization, the CONFLICT OF INTEREST
critical descriptors were CN2/CT, chiral angle, and MN/CT
(number of methyl groups per total carbons). The 3- The authors confirm that this article content has no con-
descriptor models showed R2c values of 0.92 and R2P of 0.93 flict of interest.
for Young’s modulus and R2c values of 0.91 and R2P of 0.81
for Poisson’s ratio. The computational model was qualified ACKNOWLEDGEMENTS
and quantified against three separate sets of experimental Declared none.
data.
Mohammadinasab and Goodarzi performed density func- REFERENCES
tional theory (DFT) study to calculate the energy and elec- [1] Ku, S. H.; Lee, M.; Park, C. B. Carbon‐Based Nanomaterials for
tronic properties of zigzag carbon nanotubes. The relation- Tissue Engineering. Adv. Healthc. Mater., 2013, 2, 244-260.
ship between some of the topological indices such as Wie- [2] Lim, D. J.; Sim, M.; Oh, L.; Lim, K.; Park, H. Carbon-based drug
ner, Padmakar-Ivan and Szeged in contrast to the electric delivery carriers for cancer therapy. Arch. Pharm. Res., 2014, 37,
43-52.
moments and energy (kJmol−1) and quantum chemical de- [3] Cha, C.; Shin, S. R.; Annabi, N.; Dokmeci, M. R.;
scriptors such as HOMO, LUMO, the most positive net Khademhosseini, A. Carbon-based nanomaterials: multifunctional
atomic charges (q+) and the most negative net atomic charges materials for biomedical engineering. ACS Nano, 2013, 7, 2891-
(q-) of CNTs were presented [153]. Taherpour used a four 2897.
data set of properties derived from MD calculations, to es- [4] Avouris, P.; Chen, Z.; Perebeinos, V. Carbon-based electronics.
Nat. Nanotech., 2007, 2, 605-615.
tablish one variable linear and nonlinear models between the [5] Raj, S.; Jose, S.; Sumod, U.; Sabitha, M. Nanotechnology in
number of carbon atoms of the fullerenes (n) with the struc- cosmetics: Opportunities and challenges. J. Pharm. Bioallied Sci.,
tural mechanical properties. The relationship between n with 2012, 4, 186.
diameter, volume, molecular energy, endurance load and [6] Ramanathan, T.; Abdala, A.; Stankovich, S.; Dikin, D.; Herrera-
Alonso, M.; Piner, R.; Adamson, D.; Schniepp, H.; Chen, X.;
compressive stiffness of empty fullerenes (n = 20, 60, 80, Ruoff, R. Functionalized graphene sheets for polymer
and 180) were constructed and the equations obtained were nanocomposites. Nat. Nanotech., 2008, 3, 327-331.
extended for other twenty fullerenes (Cn, n from 24 to 300) [7] Ng, Y. H.; Ikeda, S.; Matsumura, M.; Amal, R. A perspective on
[154]. In a similar study, the relationship between Cn index fabricating carbon-based nanomaterials by photocatalysis and their
and fullerenes polarizabilities on the basis of monopole- applications. Energy Environ. Sci., 2012, 5, 9307-9318.
[8] Mauter, M. S.; Elimelech, M. Environmental applications of
dipole interactions theorem were presented [155]. carbon-based nanomaterials. Environ. Sci. Tech., 2008, 42, 5843-
Unfortunately, the aforementioned studies in this section 5859.
[9] Haddon, R. C.; Brus, L. E.; Raghavachari, K., Electronic structure
used the properties derived from computational methods to and bonding in icosahedral C60. Chem. Phys. Lett., 1986, 125, 459.
build QSPR models. Although such computational models [10] Dinadayalane, T.; Leszczynski, J. Fundamental Structural,
are good for qualitative explanations, but experimental data Electronic, and Chemical Properties of Carbon Nanostructures:
are required for evaluating and verifying these models to Graphene, Fullerenes, Carbon Nanotubes, and Their Derivatives. In
Handbook of Computational Chemistry, Leszczynski, J., Ed.
provide a quantitatively accurate prediction. Springer Netherlands: 2014; pp 793-867.
[11] Geim, A. K.; Novoselov, K.S. The rise of graphene. Nat. Mater.,
CONCLUSION 2007, 6, 183-191.
[12] Galano, A., Carbon nanotubes: promising agents against free
This paper reviewed the applicability of the QSPR meth- radicals. Nanoscale, 2010, 2, 373-380.
odologies for the prediction of the diverse properties of car- [13] Terrones, M.; Terrones, H., The carbon nanocosmos: novel
bon-based nanomaterials. Different steps in the construction materials for the twenty-first century. Philos. Trans. A Math. Phys.
Eng. Sci., 2003, 361, 2789-2806.
of QSPR models along with the important challenges in ex- [14] Tropsha, A. Best Practices for QSAR Model Development,
tending the traditional QSPR to nano-QSPR methodology Validation, and Exploitation. Mol. Inform., 2010, 29, 476-488.
was described. Then the most significant achievements of [15] Puzyn, T.; Leszczynska, D.; Leszczynski, J., Toward the
QSPR methods in modelling of carbon-based nanomaterials Development of “Nano-QSARs”: Advances and Challenges. Small,
2009, 5, 2494-2509.
properties and their recent applications to generate predictive [16] Clark, K. A.; White, R. H.; Silbergeld, E. K. Predictive models for
models were presented. There is an increasing interest in the nanotoxicology: current challenges and future opportunities. Regul.
use of well-established nano-QSPR models, providing suc- Toxicol. Pharmacol., 2011, 59, 361-363.
cessful results in a rapid and accurate prediction of nanoma- [17] Le, T.; Epa, V. C.; Burden, F. R.; Winkler, D.A. Quantitative
terials properties reducing the time and cost of the experi- structure–property relationship modeling of diverse materials
properties. Chem. Rev., 2012, 112, 2889-2919.
mental work. However, the future progress of nano-QSPR [18] Tantra, R.; Oksel, C.; Puzyn, T.; Wang, J.; Robinson, K. N.; Wang,
requires the development of adequate and relevant quantita- X. Z.; Ma, C. Y.; Wilkins, T. Nano (Q) SAR: Challenges, pitfalls
tive parameters describing unique molecular properties of and perspectives. Nanotoxicology, 2014, 1-7.
nanomaterials. For most nanomaterials properties, no suffi- [19] Epa, V.; Winkler, D.; Tran, L. Chapter 5 - Computational
Approaches. In Adverse Effects of Engineered Nanomaterials,
cient data to develop reliable QSPR models are available. Shvedova, B. F. P. A., Ed. Academic Press: Boston, 2012; pp 85-
Nano-QSPR models obtained from a small data set may suf- 96.
fer from the impossibility of making accurate predictions for [20] Worth, A. In Computational nanotoxicology–towards a structure-
external data set, regardless of their good performance in activity based paradigm for investigationg the activity of
internal validations. However, nano-QSAR, with properly nanoparticles, Icon Workshop. Towards Predicting Nano-Bio
Interactions, Zurich, Switzerland, 2007.
1884 Current Topics in Medicinal Chemistry, 2015, Vol. 15, No. 18 Maryam Salahinejad

[21] Winkler, D. A.; Mombelli, E.; Pietroiusti, A.; Tran, L.; Worth, A.; nanotubes in an embedded vertical array. Appl. Phys.. Let., 2002,
Fadeel, B.; McCall, M. J. Applying quantitative structure–activity 81, 910-912.
relationship approaches to nanotoxicology: Current status and [43] Topliss, J. G.; Edwards, R. P., Chance factors in studies of
future potential. Toxicology, 2013, 313, 15-23. quantitative structure-activity relationships. J. Med. Chem., 1979,
[22] Katritzky, A. R.; Kuanar, M.; Slavov, S.; Hall, C. D.; Karelson, M.; 22, 1238-1244.
Kahn, I.; Dobchev, D. A. Quantitative correlation of physical and [44] Gupta, N. "Feature Selection" From QSARWorld--A Strand Life
chemical properties with chemical structure: utility for prediction. Sciences Web Resource. http://www.qsarworld.com/qsar-ml-
Chem. Rev., 2010, 110, 5714-5789. feature-selection.php.
[23] Berhanu, W. M.; Pillai, G. G.; Oliferenko, A. A.; Katritzky, A. R. [45] Gao, H.; Lajiness, M. S.; Drie, J.V. Enhancement of binary QSAR
Quantitative structure–activity/property relationships: The analysis by a GA-based variable selection method. J. Mol. Graph.
ubiquitous links between cause and effect. ChemPlusChem, 2012, Model., 2002, 20, 259-268.
77, 507-517. [46] Sun, M.; Chen, J.; Cai, J.; Cao, M.; Yin, S.; Ji, M. Simultaneously
[24] Dudek, A. Z.; Arodz, T.; Galvez, J. Computational methods in Optimized Support Vector Regression Combined With Genetic
developing quantitative structure-activity relationships (QSAR): a Algorithm for QSAR Analysis of KDR/VEGFR-2 Inhibitors.
review. Comb. Chem. High Throughput Screening, 2006, 9, 213- Chem. Bio. Drug Des., 2010, 75, 494-505.
228. [47] Leardi, R., Application of genetic algorithm-PLS for feature
[25] Katritzky, A. R.; Lobanov, V. S.; Karelson, M. QSPR: the selection in spectral data sets. J. Chemometr., 2000, 14, 643-655.
correlation and quantitative prediction of chemical and physical [48] Mercader, A. G.; Duchowicz, P. R.; Fernández, F. M.; Castro, E.
properties from structure. Chem. Soc. Rev., 1995, 24, 279-287. A., Replacement Method and Enhanced Replacement Method
[26] Golbraikh, A.; Tropsha, A. Predictive QSAR modeling based on Versus the Genetic Algorithm Approach for the Selection of
diversity sampling of experimental datasets for the training and test Molecular Descriptors in QSPR/QSAR Theories. J. Chem. Inf.
set selection. Mol. Dversity, 2000, 5, 231-243. Model., 2010, 50, 1542-1548.
[27] Xia, X. R.; Monteiro-Riviere, N. A.; Mathur, S.; Song, X.; Xiao, [49] Mercader, A. G.; Duchowicz, P. R.; Fernández, F. M.; Castro, E.A.
L.; Oldenberg, S. J.; Fadeel, B.; Riviere, J. E., Mapping the surface Advances in the Replacement and Enhanced Replacement Method
adsorption forces of nanomaterials in biological systems. ACS in QSAR and QSPR Theories. J. Chem. Inf. Model., 2011, 51,
Nano, 2011, 5, 9074-9081. 1575-1581.
[28] Fourches, D.; Pu, D.; Tassa, C.; Weissleder, R.; Shaw, S. Y.; [50] Arauro, M. C. U.; Saldanha, T. C. B.; Galvao, R. K. H.; Yoneyama,
Mumper, R. J.; Tropsha, A., Quantitative nanostructure− activity T.; Chame, H. C.; Visani, V. The successive projections algorithm
relationship modeling. ACS Nano, 2010, 4, 5703-5712. for variable selection in spectroscopic multicomponent analysis.
[29] Rasulev, B.; Gajewicz, A.; Puzyn, T.; Leszczynska, D.; Chemometr. Intell. Lab. Syst., 2001, 57, 65-73.
Leszczynski, J. Nano-QSAR: Advances and Challenges. In [51] Baroni, M.; Costantino, G.; Cruciani, G.; Riganelli, D.; Valigi, R.;
Towards Efficient Designing of Safe Nanomaterials: Innovative Clementi, S. Generating Optimal Linear PLS Estimations
Merge of Computational Approaches and Experimental (GOLPE): An Advanced Chemometric Tool for Handling
Techniques, 2012; p 220. 3DQSAR Problems. Quant. Struct-Act Rel., 1993, 12, 9-20.
[30] Todeschini, R.; Consonni, V., Handbook of molecular descriptors. [52] Bi, J.; Bennett, K.; Embrechts, M.; Breneman, C.; Song, M.,
John Wiley & Sons: 2008. Dimensionality reduction via sparse support vector machines. J.
[31] Borders, T. L.; Fonseca, A. F.; Zhang, H.; Cho, K.; Rusinko III, A. Mach. Learn. Res., 2003, 3, 1229-1243.
Developing Descriptors To Predict Mechanical Properties of [53] Burden, F.; Winkler, D., Optimal sparse descriptor selection for
Nanotubes. J. Chem. Inf. Model., 2013, 53, 773-782. QSAR using bayesian methods. QSAR Comb Sci., 2009, 28, 645-
[32] Barnard, A. S., How can ab initio simulations address risks in 653.
nanotech? Nat. Nanotech., 2009, 4, 332-335. [54] Burden, F. R.; Winkler, D. A., An Optimal Self‐Pruning Neural
[33] Lindahl, E., Molecular Dynamics Simulations. In Molecular Network and Nonlinear Descriptor Selection in QSAR. QSAR
Modeling of Proteins, Kukol, A., Ed. Humana Press: 2008; Vol. Comb Sci., 2009, 28, 1092-1097.
443, pp 3-23. [55] Lin, D.; Zhang, J.; Li, J.; Calhoun, V. D.; Deng, H.-W.; Wang, Y.-
[34] Poater, A.; Saliner, A. G.; Solà, M.; Cavallo, L.; Worth, A. P. P., Group sparse canonical correlation analysis for genomic data
Computational methods to predict the reactivity of nanoparticles integration. BMC Bioinformatics, 2013, 14, 245.
through structure-property relationships. Expert Opin. Drug Deliv., [56] Goodarzi, M.; Dejaegher, B.; Heyden, Y. V., Feature selection
2010, 7, 295-305. methods in QSAR studies. J. AOAC Int., 2012, 95, 636-651.
[35] Ashrafi, A. R.; Loghman, A. PI index of zig-zag polyhex [57] Gonzalez, M. P.; Teran, C.; Saíz-Urra, L.; Teijeira, M. Variable
nanotubes. MATCH Commun. Math. Comput. Chem., 2006, 55, selection methods in QSAR: an overview. Cur. Top. Med. Chem.,
447-452. 2008, 8, 1606-1627.
[36] Ashrafi, A.; Ghorbani, M.; Jalali, M., The vertex PI and Szeged [58] Eklund, M.; Norinder, U.; Boyer, S.; Carlsson, L. Choosing feature
indices of an infinite family of fullerenes. J. Theor. Comput. selection and learning algorithms in QSAR. J. Chem. Inf. Model.,
Chem., 2008, 7, 221-231. 2014, 54, 837-843.
[37] Toropov, A. A.; Toropova, A. P.; Benfenati, E.; Leszczynska, D.; [59] Næs, T.; Martens, H. Principal component regression in NIR
Leszczynski, J., Additive InChI-based optimal descriptors: QSPR analysis: Viewpoints, background details and selection of
modeling of fullerene C60 solubility in organic solvents. J. Math. components. J. Chemometr., 1988, 2, 155-167.
Chem., 2009, 46, 1232-1251. [60] Geladi, P.; Kowalski, B. Partial leastsquares regression: A tutorial.
[38] Toropov, A. A.; Rasulev, B. F.; Leszczynska, D.; Leszczynski, J. Anal. Chim. Acta, 1986, 185, 1-17.
Multiplicative SMILES-based optimal descriptors: QSPR modeling [61] Wold, S.; Sjostrom, M.; Eriksson, L., PLS-regression: a basic tool
of fullerene C60 solubility in organic solvents. Chem. Phys. Lett., of chemometrics. Chemometr. Intell. Lab. Syst., 2001, 58, 109-130.
2008, 457, 332-336. [62] Despagne, F. e. e.; Massart, D. L., Neural networks in multivariate
[39] Toropov, A. A.; Rasulev, B. F.; Leszczynska, D.; Leszczynski, J. calibration. Analyst, 1998, 123, 157R-178R.
Additive SMILES based optimal descriptors: QSPR modeling of [63] Cheng, B.; Titterington, D. M., Neural Networks: A Review from a
fullerene C60 solubility in organic solvents. Chem. Phys. Lett., Statistical Perspective. Statist. Sci., 1994, 9, 2-30.
2007, 444, 209-214. [64] Hart, P.; Cover, T., Nearest neighbor pattern classification. IEEE
[40] Abou-Hamad, E.; Babaa, M. R.; Bouhrara, M.; Kim, Y.; Saih, Y.; Trans. Inf. Theory, 1988, 13, 21-27.
Dennler, S.; Mauri, F.; Basset, J. M.; Goze-Bac, C.; Wagberg, T. [65] Ajmani, S.; Jadhav, K.; Kulkarni, S. A., Three-dimensional QSAR
Structural properties of carbon nanotubes derived from 13C NMR. using the k-nearest neighbor method and its interpretation. J. Chem.
Phys. Rev. B, 2011, 84, 165417. Inf. Model., 2006, 46, 24-31.
[41] Odom, T. W.; Huang, J. L.; Kim, P.; Lieber, C. M. Structure and [66] Cortes, C. V., V., Support-vector networks. Mach. Learn., 1995,
electronic properties of carbon nanotubes. J. Phys. Chem. B, 2000, 20, 273-297.
104, 2794-2809. [67] Ivanciuc, O. Applications of Support Vector Machines in
[42] Li, J.; Stevens, R.; Delzeit, L.; Ng, H. T.; Cassell, A.; Han, J.; Chemistry. In: Reviews in Computational Chemistry. Wiley-VCH,
Meyyappan, M. Electronic properties of multiwalled carbon Weinheim: 2007; Vol. 23.
Nano-QSPR Modelling of Carbon-Based Nanomaterials Properties Current Topics in Medicinal Chemistry, 2015, Vol. 15, No. 18 1885

[68] Liu, P.; Long, W., Current mathematical methods used in review of CODESSA applications. Collect. Czech. Chem.
QSAR/QSPR studies. Int. J. Mol. Sci., 2009, 10, 1978-1998. Commun., 1999, 64, 1551-1571.
[69] Burden, F. R.; Winkler, D. A., Robust QSAR models using [93] Martin, D.; Maran, U.; Sild, S.; Karelson, M., QSPR modeling of
Bayesian regularized neural networks. J. Med. Chem., 1999, 42, solubility of polyaromatic hydrocarbons and fullerene in 1-octanol
3183-3187. and n-heptane. J. Phys. Chem. B, 2007, 111, 9853-9857.
[70] Gramatica, P., Principles of QSAR models validation: internal and [94] Petrova, T.; Rasulev, B. F.; Toropov, A. A.; Leszczynska, D.;
external. QSAR Comb Sci., 2007, 26, 694-701. Leszczynski, J., Improved model for fullerene C60 solubility in
[71] Golbraikh, A.; Tropsha, A., Beware of q2! Journal of Molecular organic solvents based on quantum-chemical and topological
Graphics and Modelling, 2002, 20, 269-276. descriptors. J. Nanopart. Res., 2011, 13, 3235-3247.
[72] Lindgren, F.; Hansen, B.; Karcher, W.; Sjöström, M.; Eriksson, L., [95] Sivaraman, N.; Srinivasan, T.; Vasudeva Rao, P.; Natarajan, R.
Model validation by permutation tests: Applications to variable QSPR modeling for solubility of fullerene (C60) in organic solvents.
selection. J. Chemometr., 1996, 10, 521-532. J. Chem. Inf. Comput. Sci., 2001, 41, 1067-1074.
[73] Schü ü rmann, G.; Ebert, R.-U.; Chen, J.; Wang, B.; Kü hne, [96] Toropov, A. A.; Leszczynska, D.; Leszczynski, J., QSPR study on
R., External Validation and Prediction Employing the Predictive solubility of fullerene C60 in organic solvents using optimal
Squared Correlation Coefficient -Test Set Activity Mean vs descriptors calculated with SMILES. Chem. Phys. Lett., 2007, 441,
Training Set Activity Mean. J. Chem. Inf. Model., 2008, 48, 2140- 119-122.
2145. [97] Abboud, J. L.; Kamlet, M. J.; Taft, R., Regarding a generalized
[74] He, L.; Jurs, P. C., Assessing the reliability of a QSAR model's scale of solvent polarities. J. Am. Chem. Soc., 1977, 99, 8325-8327.
predictions. J. Mol. Graph. Model., 2005, 23, 503-523. [98] Kamlet, M. J.; Abboud, J. L. M.; Abraham, M. H.; Taft, R. W.,
[75] Tarasova, A.; Burden, F.; Gasteiger, J.; Winkler, D. A., Robust Linear solvation energy relationships. 23. A comprehensive
modelling of solubility in supercritical carbon dioxide using collection of the solvatochromic parameters, .pi.*, .alpha., and
Bayesian methods. J. Mol. Graph. Model., 2010, 28, 593-597. .beta., and some methods for simplifying the generalized
[76] Consonni, V.; Ballabio, D.; Todeschini, R., Comments on the solvatochromic equation. Journal of Organic Chemistry, 1983, 48,
Definition of the Q2 Parameter for QSAR Validation. J. Chem. Inf. 2877-2887.
Model., 2009, 49, 1669-1678. [99] Lee, L.-H. Correlation between Lewis acid-base surface interaction
[77] Gramatica, P., Principles of QSAR models validation: internal and components and linear solvation energy relationship
external. QSAR Combi. Sci., 2007, 694-701. solvatochromic α and β parameters. Langmuir, 1996, 12, 1681-
[78] Rofouei, M.; Salahinejad, M.; Ghasemi, J. B., An Alignment 1687.
Independent 3D-QSAR Modeling of Dispersibility of Single- [100] Marcus, Y.; Kamlet, M.; Taft, R., Linear solvation energy
walled Carbon Nanotubes in Different Organic Solvents. Fuller. relationships: standard molar Gibbs free energies and enthalpies of
Nanotub. Car. N., 2014, 22, 605-617. transfer of ions from water into nonaqueous solvents. J. Phys.
[79] Savjani, K. T.; Gajjar, A. K.; Savjani, J. K., Drug solubility: Chem., 1988, 92, 3613-3622.
importance and enhancement techniques. ISRN Pharm., 2012, [101] Abraham, M. H.; Ibrahim, A.; Zissimos, A. M., Determination of
2012. sets of solute descriptors from chromatographic measurements. J.
[80] Salahinejad, M.; Le, T. C.; Winkler, D. A., Aqueous solubility Chromatogr. A, 2004, 1037, 29-47.
prediction: do crystal lattice interactions help? Mol. Pharm., 2013, [102] Ruoff, R.; Tse, D. S.; Malhotra, R.; Lorents, D.C. Solubility of
10, 2757-2766. fullerene (C60) in a variety of solvents. J. Phys. Chem., 1993, 97,
[81] Mitchell, B. E.; Jurs, P.C. Prediction of aqueous solubility of 3379-3383.
organic compounds from molecular structure. J. Chem. Inf. [103] Abraham, M. H.; Green, C. E.; Acree Jr, W. E., Correlation and
Comput. Sci., 1998, 38, 489-496. prediction of the solubility of Buckminsterfullerene in organic
[82] Wang, J.; Hou, T. Recent advances on aqueous solubility solvents; estimation of some physicochemical properties. J. Chem.
prediction. Comb. Chem. High Throughput Screening, 2011, 14, Soc., Perkin Trans. 2., 2000, 281-286.
328-338. [104] Flunt, R.; Makitra, R.; Makohon, O.; Pyrih, J., Quantitative
[83] Hansen, C. M.; Smith, A. L., Using Hansen solubility parameters to Treatment of Fullerene C60 Solubility Data. Org. React. Tartu,
correlate solubility of C60 fullerene in organic solvents and in 1997, 31, 21-26.
polymers. Carbon, 2004, 42, 1591-1597. [105] Makitra, R. G.; Pristanskii, R. E.; Flyunt, R. I., Solvent Effects on
[84] Huang, J.-C., Multiparameter solubility model of fullerene C60. the Solubility of C60 fullerene. Rus. J. Gen. Chem., 2003, 73, 1227-
Fluid Phase Equilibria, 2005, 237, 186-192. 1232.
[85] Kamlet, M. J.; Abboud, J. L. M.; Abraham, M. H.; Taft, R. W., [106] Marcus, Y., Solubilities of buckminsterfullerene and sulfur
Linear solvation energy relationships. 23. A comprehensive hexafluoride in various solvents. J. Phys. Chem. B, 1997, 101,
collection of the solvatochromic parameters, .pi.*, .alpha., and 8617-8623.
.beta., and some methods for simplifying the generalized [107] Marcus, Y.; Smith, A. L.; Korobov, M.; Mirakyan, A.; Avramenko,
solvatochromic equation. J. Org. Chem., 1983, 48, 2877-2887. N.; Stukalin, E., Solubility of C60 fullerene. J. Phys. Chem. B, 2001,
[86] Sivaraman, N.; Dhamodaran, R.; Kaliappan, I.; Srinivasan, T.; Rao, 105, 2499-2506.
P.V.; Mathews, C. Solubility of C60 in organic solvents. J. Org. [108] Yousefinejad, S.; Honarasa, F.; Abbasitabar, F.; Arianezhad, Z.,
Chem., 1992, 57, 6077-6079. New LSER Model Based on Solvent Empirical Parameters for the
[87] Danauskas, S. M.; Jurs, P. C., Prediction of C60 solubilities from Prediction and Description of the Solubility of
solvent molecular structures. J. Chem. Inf. Comput. Sci., 2001, 41, Buckminsterfullerene in Various Solvents. J. Sol. Chem., 2013, 42,
419-424. 1620-1632.
[88] Gharagheizi, F.; Alamdari, R.F. A Molecular‐Based Model for [109] Barton, A. F., CRC handbook of solubility parameters and other
Prediction of Solubility of C60 Fullerene in Various Solvents. cohesion parameters. CRC press: 1991.
Fuller. Nanotub. Car. N., 2008, 16, 40-57. [110] Hansen, C. M., 50 Years with solubility parameters-past and future.
[89] Kiss, I. Z.; Mándi, G.; Beck, M.T. Artificial neural network Prog. Org. Coat., 2004, 51, 77-84.
approach to predict the solubility of C60 in various solvents. J. Phys. [111] Thomas, E. R.; Newman, B. A.; Nicolaides, G. L.; Eckert, C. A.,
Chem. A, 2000, 104, 8081-8088. Limiting activity coefficients from differential ebulliometry. J.
[90] Liu, H.; Yao, X.; Zhang, R.; Liu, M.; Hu, Z.; Fan, B. Accurate Chem. Eng. Data, 1982, 27, 233-240.
quantitative structure-property relationship model to predict the [112] Pourbasheer, E.; Riahi, S.; Ganjali, M. R.; Norouzi, P., Prediction
solubility of C60 in various solvents based on a novel approach of solubility of fullerene C60 in various organic solvents by genetic
using a least-squares support vector machine. J. Phys. Chem. B, algorithm-multiple linear regression. Fuller. Nanotub. Car. N.,
2005, 109, 20565-20571. 2011, 19, 585-598.
[91] Benfenati, E.; Toropov, A. A.; Toropova, A. P.; Manganaro, A.; [113] Toropova, A. P.; Toropov, A. A.; Benfenati, E.; Gini, G.;
Gonella Diaza, R. CORAL software: QSAR for anticancer agents. Leszczynska, D.; Leszczynski, J., CORAL: QSPR models for
Chem. Bio. Drug Des., 2011, 77, 471-476. solubility of [C60] and [C70] fullerene derivatives. Mol. Dversity,
[92] Karelson, M.; Maran, U.; Wang, Y.; Katritzky, A.R. QSPR and 2011, 15, 249-256.
QSAR models derived using large molecular descriptor spaces. A [114] Ghasemi, J. B.; Salahinejad, M.; Rofouei, M., Alignment
Independent 3D-QSAR Modeling of Fullerene (C60) Solubility in
1886 Current Topics in Medicinal Chemistry, 2015, Vol. 15, No. 18 Maryam Salahinejad

Different Organic Solvents. Fuller. Nanotub. Car. N., 2013, 21, [135] Barman, S. N.; Pan, D.; Vosgueritchian, M.; Zoombelt, A. P.;
367-380. Galli, G.; Bao, Z. Dispersion of single walled carbon nanotubes in
[115] Cruciani, G.; Carosati, E.; Clementi, S., 25 - Three-Dimensional amidine solvents. Nanotechnology, 2012, 23, 344011.
Quantitative Structure-Property Relationships. In The Practice of [136] Wildman, S. A.; Crippen, G. M. Prediction of Physicochemical
Medicinal Chemistry (Second Edition), Camille, G. W., Ed. Parameters by Atomic Contributions. J. Chem. Inf. Comput. Sci.,
Academic Press: London, 2003; pp 405-416. 1999, 39, 868-873.
[116] Langer, T.; Bryant, S. D., Chapter 29 - 3D Quantitative Structure– [137] Li, X.; Cai, W.; Jung, I. H.; An, J. H.; Yang, D.; Velamakanni, A.;
Property Relationships. In The Practice of Medicinal Chemistry Piner, R.; Colombo, L.; Ruoff, R. S. Synthesis, characterization,
(Third Edition), Camille Georges, W., Ed. Academic Press: New and properties of large-area graphene films. ECS Trans., 2009, 19,
York, 2008; pp 587-604. 41-52.
[117] Pastor, M.; Cruciani, G.; McLay, I.; Pickett, S.; Clementi, S., [138] Wang, J., Preparation of Graphene Dispersion: A Review. Curr.
GRid-INdependent descriptors (GRIND): a novel class of Phys. Chem., 2013, 3, 255-268.
alignment-independent three-dimensional molecular descriptors. J. [139] Hernandez, Y.; Lotya, M.; Rickard, D.; Bergin, S. D.; Coleman, J.
Med. Chem., 2000, 43, 3233-3243. N., Measurement of multicomponent solubility parameters for
[118] Cruciani, G.; Crivori, P.; Carrupt, P.-A.; Testa, B., Molecular fields graphene facilitates solvent discovery. Langmuir, 2009, 26, 3208-
in quantitative structure–permeation relationships: the VolSurf 3213.
approach. J. Mol. Struct., 2000, 503, 17-30. [140] Yousefinejad, S.; Hemmateenejad, B., A chemometrics approach to
[119] Galvao, R. K. H.; Araujo, M. C. U.; Fragoso, W. D.; Silva, E. C.; predict the dispersibility of graphene in various liquid phases using
Jose, G. E.; Soares, S. F. C.; Paiva, H.M. A variable elimination theoretical descriptors and solvent empirical parameters. Colloids
method to improve the parsimony of MLR models using the Surf. A Physicochem. Eng. Asp., 2014, 441, 766-775.
successive projections algorithm. Chemometr. Intell. Lab. Syst., [141] Pan, B.; Xing, B., Adsorption mechanisms of organic chemicals on
2008, 92, 83-91. carbon nanotubes. Environ. Sci. Tech., 2008, 42, 9005-9013.
[120] Leardi, R.; Boggia, R.; Terrile, M., Genetic algorithms as a strategy [142] Chen, W.; Duan, L.; Zhu, D., Adsorption of polar and nonpolar
for feature selection. J. Chemometr., 1992, 6, 267-281. organic chemicals to carbon nanotubes. Environ. Sci. Tech., 2007,
[121] Geckeler, K.; Premkumar, T., Carbon nanotubes: are they dispersed 41, 8295-8300.
or dissolved in liquids? Nanoscale Res. Lett., 2011, 6, 136. [143] Yang, K.; Zhu, L.; Xing, B. Adsorption of polycyclic aromatic
[122] Bergin, S. D.; Sun, Z.; Rickard, D.; Streich, P. V.; Hamilton, J. P.; hydrocarbons by carbon nanomaterials. Environ. Sci. Tech., 2006,
Coleman, J. N., Multicomponent Solubility Parameters for Single- 40, 1855-1861.
Walled Carbon Nanotube−Solvent Mixtures. ACS Nano, 2009, 3, [144] Yang, K.; Wu, W.; Jing, Q.; Zhu, L., Aqueous adsorption of
2340-2350. aniline, phenol, and their substitutes by multi-walled carbon
[123] Giordani, S.; Bergin, S. D.; Nicolosi, V.; Lebedkin, S.; Kappes, M. nanotubes. Environ. Sci. Tech., 2008, 42, 7931-7936.
M.; Blau, W. J.; Coleman, J. N., Debundling of single-walled [145] Xia, X.-R.; Monteiro-Riviere, N. A.; Riviere, J. E., An index for
nanotubes by dilution: observation of large populations of characterization of nanomaterials in biological systems. Nat.
individual nanotubes in amide solvent dispersions. J. Phys. Chem. Nanotech., 2010, 5, 671-675.
B, 2006, 110, 15708-15718. [146] Apul, O. G.; Wang, Q.; Shao, T.; Rieck, J. R.; Karanfil, T.,
[124] Landi, B. J.; Ruf, H. J.; Worman, J. J.; Raffaelle, R. P., Effects of Predictive model development for adsorption of aromatic
Alkyl Amide Solvents on the Dispersion of Single-Wall Carbon contaminants by multi-walled carbon nanotubes. Environ. Sci.
Nanotubes. J. Phys. Chem. B, 2004, 108, 17089-17095. Tech., 2013, 47, 2295-2303.
[125] Ham, H. T.; Choi, Y. S.; Chung, I. J., An explanation of dispersion [147] Apul, O. G.; Xuan, P.; Luo, F.; Karanfil, T., Development of a 3D
states of single-walled carbon nanotubes in solvents and aqueous QSPR model for adsorption of aromatic compounds by carbon
surfactant solutions using solubility parameters. J. Colloid nanotubes: comparison of multiple linear regression, artificial
Interface Sci., 2005, 286, 216-223. neural network and support vector machine. RSC Adv., 2013, 3,
[126] Detriche, S.; Zorzini, G.; Colomer, J. F.; Fonseca, A.; Nagy, J. B., 23924-23934.
Application of the Hansen Solubility Parameters Theory to Carbon [148] Zhao, Q.; Yang, K.; Li, W.; Xing, B. Concentration-dependent
Nanotubes. J. Nanosci. Nanotechnol., 2008, 8, 6082-6092. polyparameter linear free energy relationships to predict organic
[127] Bergin, S. D.; Sun, Z.; Rickard, D.; Streich, P. V.; Hamilton, J. P.; compound sorption on carbon nanotubes. Sci. Rep., 2014, 4.
Coleman, J. N. Multicomponent solubility parameters for single- [149] Liu, Y.; Zhang, J.; Chen, X.; Zheng, J.; Wang, G.; Liang, G.,
walled carbon nanotube− solvent mixtures. Acs Nano, 2009, 3, Insights into the adsorption of simple benzene derivatives on
2340-2350. carbon nanotubes. RSC Adv., 2014, 4, 58036-58046.
[128] Cheng, Q.; Debnath, S.; Gregan, E.; Byrne, H. J., Effect of Solvent [150] Katritzky, A. R.; Dobchev, D. A.; Slavov, S.; Karelson, M.
Solubility Parameters on the Dispersion of Single-Walled Carbon Legitimate utilization of large descriptor pools for QSPR/QSAR
Nanotubes. J. Phys. Chem. C, 2008, 112, 20154-20158. models. J. Chem. Inf. Model., 2008, 48, 2207-2213.
[129] Cheng, Q.; Debnath, S.; O’Neill, L.; Hedderman, T. G.; Gregan, E.; [151] Shokrieh, M. M.; Rafiee, R., A review of the mechanical properties
Byrne, H. J. Systematic study of the dispersion of SWNTs in of isolated carbon nanotubes and carbon nanotube composites.
organic solvents. J. Phys. Chem. C, 2010, 114, 4857-4863. Mech Compos Mater, 2010, 46, 155-172.
[130] Torrens, F., Calculations of organic-solvent dispersions of single- [152] Greaves, G. N.; Greer, A.; Lakes, R.; Rouxel, T. Poisson's ratio and
wall carbon nanotubes. Int. J. Quantum Chem., 2006, 106, 712- modern materials. Nat. Mater., 2011, 10, 823-837.
718. [153] Goodarzi, M.; Mohammadinasab, E., Theoretical Investigation of
[131] Torrens, F., Partition of solvents and co-solvents of nanotubes: Relationship between Quantum Chemical Descriptors, Topological
proteins and cyclopyranoses. Front. Drug Des. Discovery, 2005, 1, Indices, Energy and Electric Moments of Zig-zag Polyhex Carbon
231-266. Nanotubes TUHC6 [2p, q] with Various Circumference [2p] and
[132] Toropov, A. A.; Leszczynska, D.; Leszczynski, J., Predicting water Fixed Lengths. Fuller. Nanotub. Car. N., 2013, 21, 102-112.
solubility and octanol water partition coefficient for carbon [154] Taherpour, A. Quantitative relationship study of mechanical
nanotubes based on the chiral vector. Comput. Bio. Chem., 2007, structure properties of empty fullerenes. Fuller. Nanotub. Car. N.,
31, 127-128. 2008, 16, 196-205.
[133] Wiegand, R. E., Performance of using multiple stepwise algorithms [155] Taherpour, A.; Mahdizadeh, N. Theoretical and Quantitative
for variable selection. Stat. Med., 2010, 29, 1647-1659. Structural Relationship Study on Fullerenes Polarizabilities on the
[134] Salahinejad, M.; Zolfonoun, E., QSAR studies of the dispersion of basis of Monopole-Dipole Interactions Theorem. Orient. J. Chem.,
SWNTs in different organic solvents. J. Nanopart. Res., 2013, 15, 2012, 28, 247.
1-9.

Received: December 09, 2014 Revised: February 25, 2015 Accepted: March 24, 2015

You might also like