Toxicology Letters 232 (2015) 519–532

Contents lists available at ScienceDirect

Toxicology Letters
journal homepage: www.elsevier.com/locate/toxlet

OpenVirtualToxLab—A platform for generating and exchanging in silico

toxicity data
Angelo Vedani a,b, *, Max Dobler b , Zhenquan Hu a , Martin Smieško a
a
Department of Pharmaceutical Sciences, University of Basel, Klingelbergstrasse 50, 4056 Basel, Switzerland
b
Foundation Biographics Laboratory 3R, Klingelbergstrasse 50, 4056 Basel, Switzerland

A R T I C L E I N F O A B S T R A C T

Article history: The VirtualToxLab is an in silico technology for estimating the toxic potential—endocrine and metabolic
Received 3 July 2014 disruption, some aspects of carcinogenicity and cardiotoxicity—of drugs, chemicals and natural products.
Accepted 3 September 2014 The technology is based on an automated protocol that simulates and quantifies the binding of small
Available online 18 September 2014
molecules towards a series of currently 16 proteins, known or suspected to trigger adverse effects:
10 nuclear receptors (androgen, estrogen a, estrogen b, glucocorticoid, liver X, mineralocorticoid,
Keywords: peroxisome proliferator-activated receptor g, progesterone, thyroid a, thyroid b), four members of the
In silico toxicology
cytochrome P450 enzyme family (1A2, 2C9, 2D6, 3A4), a cytosolic transcription factor (aryl hydrocarbon
Protein-mediated adverse effects
3D simulations
receptor) and a potassium ion channel (hERG). The toxic potential of a compound—its ability to trigger
Mechanistic interpretation adverse effects—is derived from its computed binding affinities toward these very proteins: the
Open-access platform computationally demanding simulations are executed in client–server model on a Linux cluster of the
University of Basel. The graphical-user interface supports all computer platforms, allows building and
uploading molecular structures, inspecting and downloading the results and, most important,
rationalizing any prediction at the atomic level by interactively analyzing the binding mode of a
compound with its target protein(s) in real-time 3D. Access to the VirtualToxLab is available free of charge
for universities, governmental agencies, regulatory bodies and non-profit organizations.
ã 2014 The Authors. Published by Elsevier Ireland Ltd. This is an open access article under the CC BY-NC-
ND license (http://creativecommons.org/licenses/by-nc-nd/3.0/).

1. Introduction In silico techniques for the prediction of toxicological endpoints

In the light of the REACH (Registration, Evaluation, Authoriza-
tion and Restriction of Chemicals) initiative of the European Union,
computer-based approaches are explicitly sought after as they may
predict the toxicity of drugs and chemicals—both to avoid
unnecessary animal tests and to reduce the number of late-stage
failures in drug discovery & development. Adverse effects triggered
by small molecules are frequently associated with their binding to
so-called “off targets”—bioregulators involved in biosynthesis,
signal transduction, transport, storage, and metabolism. Among
others, those include nuclear receptors, enzymes of the cyto-
chrome P450 family and ion channels (Colborn et al., 1993; Dibb,
1995; Guillette et al., 1995; McLachlan and Arnold, 1996; Rihova,
1998; Fischer, 2000; Aronov, 2005; De Graaf et al., 2005; Crivori
and Poggesi, 2006).
* Corresponding author. Fax: +41 61 267 1552.
E-mail address: angelo.vedani@unibas.ch (A. Vedani).
are extremely appealing because of their expeditious return of
results and inexpensiveness (Muster et al., 2008). Computational
approaches are typically based on human data and can be applied
to hypothetical compounds, which is of great relevance for drug
discovery—both ecological and economical. They can be classified
into expert systems, QSAR (quantitative structure–activity rela-
tionships), protein modeling and ADME (adsorption, distribution,
metabolism, excretion) modeling. A large body of both review and
research articles exists for these technologies (see, for example,
Cronin et al., 2003; Veith, 2004; Helma, 2005; Piclin et al., 2006;
Simon-Hettich et al., 2006; Amini et al., 2007; Aronov et al., 2007;
Bender et al., 2007; Custer et al., 2007; Ecker and Chiba, 2007;
Ekins, 2007; Serafimova et al., 2007; Enoch et al., 2008; Kavlock
et al., 2008; Merlot, 2008; Pavan and Worth, 2008; Benfenati et al.,
2009; Green and Naven, 2009; Nigsch et al., 2009; Spreafico et al.,
2009; Valerio, 2009; Rossato et al., 2010; Cronin and Madden,
2010; Bars et al., 2011; Vuorinen et al., 2013; Gupta et al., 2013;
Roncaglioni et al., 2013; Shah and Greene, 2014; Toropov et al.,
2014; Schilter et al., 2014; Singh and Gupta, 2014; Ekins, 2014).

http://dx.doi.org/10.1016/j.toxlet.2014.09.004
0378-4274/ã 2014 The Authors. Published by Elsevier Ireland Ltd. This is an open access article under the CC BY-NC-ND license (http://creativecommons.org/licenses/by-nc-
nd/3.0/).
520 A. Vedani et al. / Toxicology Letters 232 (2015) 519–532

Fig. 1. Flow chart of the VirtualToxLab. *Software licensed from Schrödinger Inc., Portland, OR, **from the University of Minnesota and ***from the Biographics Laboratory 3R,
Basel. (For interpretation of the references to color in the text, the reader is referred to the web version of this article.)

Computational assessment of a compound’s toxicity should
always be discussed along with its ADME properties as those define
the bioavailability—a prerequisite for triggering a molecular
mechanism leading a toxic effect. Only when quantitatively
combining all aspects, one might be in a position to predict a
toxic endpoint. Otherwise one should employ the term “toxic
potential”, implying that other conditions must be met in order for
an adverse effect to manifest itself.
Developing and validating a three-dimensional model is very
laborious but would seem to be necessary when the molecular
mechanism triggering the adverse or toxic effect occurs via a
multifaceted molecular mechanism. Skin irritation, for example,
might be safely described by the physicochemical properties of a
compound. On the other hand, a complex endpoint such as
endocrine disruption may not be explainable even by a larger
ensemble of molecular descriptors.

Fig. 2. Directional force field employed in Cheetah (Vedani and Huhta 1990; Rossato et al., 2010).
 A. Vedani et al. / Toxicology Letters 232 (2015) 519–532 521

Fig. 3. 4D-Boltzmann sampling as employed in the VirtualToxLab. Left: a compound’s (here: genistein) representations in aqueous solution (software Aquarius, Biographics
Laboratory 3R, 2013). The ligand–solvent interactions are quantified using a directional force field (Vedani and Huhta, 1990; Rossato et al., 2010) and Boltzmann weighted to
obtain the interaction energy of the 4D ensemble. Right: sampling (software Alignator and Cheetah) and quantifying (software BzScore4D) of a ligand’s representation(s) at the
binding site of the target protein. The images were generated with the BioX software (Dobler, 2012).

Manifestations of toxicity are frequently mediated by regulato- step (blue borders) the compound’s behavior in aqueous solution is
ry macromolecules such as enzymes, receptors, ion channels or simulated. This includes the identification of the protonation and
DNA. These targets represent complex and flexible three-dimen- tautomeric state at physiological pH (software Epik: Schrödinger,
sional entities that attempt to optimize their interaction with a 2012), conformational sampling in (implicit) aqueous solution
small molecule (e.g., a xenobiotic) by adapting their 3D conforma- (software MacroModel: Schrödinger, 2012), calculation of atomic
tion, a mechanism referred to as “induced fit”. Protein-bound partial charges, both the solvent-accessible and polar surface area
solvent molecules are frequently involved in stabilizing small- (software AMSOL: Cramer and Truhlar, 1992) as well as the
molecule ligands or, upon release to the “bulk solvent” contribute simulation and quantification of the compound’s interactions with
favorably to the binding entropy. Accounting and quantifying these explicit solvent (in 50
50
50 Å water box: software Aquarius, cf.
effects belongs to the most challenging tasks in the computational Fig. 3). The second and central step (green borders) consists in the
sciences. identification of the compound’s potential binding mode(s) by
In this account, we present the more recent developments of simulating its 3D interaction with the protein (pharmacophore-
the VirtualToxLab—most noticeably, the change from multi- based pre-alignment: software Alignator/Dolina: Smieško, 2013;
dimensional QSAR (mQSAR) to 4D Boltzmann scoring for
computing binding affinities based on the three-dimensional
structure of protein–ligand complexes. By avoiding the training of
a model against a set of compounds with known effects (such as in
QSARs) but using an “ab initio” approach instead, the bias of a
prediction from any training set is removed as the changes in free
energy of ligand binding, DG, are computed by direct comparison
of a compound's “behavior” in aqueous solution (mimicking the
cytoplasm) with those at the target protein employing the same
directional force field. Moreover, the risk of extrapolation—
occurring when attempting to predict properties of compounds
not truly represented in a QSAR’s training set—is purged. The new
protocol has been validated with a total of 1288 test compounds
and employed to estimate the toxic potential of more than
2500 drugs, chemicals and natural products. All results are posted
on http://www.virtualtoxlab.org. We explicitly invite all interested
non-profit organizations to freely access/utilize the technology,
and share their results with the scientific community at http://
www.biograf.ch/data/projects/OpenVirtualToxLab.php.

2. Methods

The technology underlying the VirtualToxLab has recently been

described in great detail (Vedani et al., 2012). In this account, we
therefore focus on the most recent extensions and the freely
accessible platform—the OpenVirtualToxLab.
The flow chart of the VirtualToxLab is shown in Fig. 1. The
technology consists of two distinct modules: the user interface
(light green) and the server backend (light blue) which communi-
cate through an SSH protocol. The user interface features an
embedded 3D viewer for inspecting both input (compounds to be
uploaded) and output structures (resulting protein–ligand com-
plexes) and a 3D model builder to readily generate the three-
dimensional structure of any small molecule of interest. In a first
Fig. 4. Comparison of experimental (horizontal axis) and predicted binding
affinities (vertical axis) as obtained with the VirtualToxLab 5.0 featuring the novel
4D-Boltzmann scoring function. The solid diagonal line corresponds to a perfect
prediction, i.e., pKcalc = pKexp.1 The dashed lines mark predictions 1.0 log unit (a
factor 10) off the experimental value, the dotted lines 2.0 log units (a factor 100).
706 of the 1288 compounds (55%) are predicted within one, 1082 (84%) within two,
and 1232 (96%) within three orders of magnitude from the experiment. The
individual target proteins included 140 (AhR), 104 (AR), 106 (ERa), 96 (ERb), 110
(GR), 50 (hERG), 52 (LXR), 48 (MR), 95 (PPARg), 91 (PR), 82 (TRa), 82 (TRb), 52
(1A2), 85 (2C9), 55 (2D6) and 40 (3A4) test compounds, respectively.
1
pK = –log (K).
522 A. Vedani et al. / Toxicology Letters 232 (2015) 519–532

Table 1
Calculation of the toxic potential for bisphenol A following Eqs. (1)–(3).

Target Calculated affinity  e.s.d. (M) (VirtualToxLab) TPnormalized (Eq. (1)) we.s.d. TPindividual (Eq. (2))
AR 4.6
107  1.6
107 0.326 0.957 0.311
AhR 3.0
105  1.2
105 0.189 0.949 0.180
CYP 1A2 2.3
105  1.0
105 0.198 0.944 0.187
CYP 2C9 3.5
105  1.6
105 0.185 0.943 0.174
CYP 2D6 4.0
105  1.7
105 0.179 0.948 0.170
CYP 3A4 1.8
104  7.6
105 0.131 0.947 0.124
ERa 8.0
106  2.6
106 0.232 0.958 0.223
ERb 5.4
108  1.6
108 0.395 0.964 0.381
GR 1.3
106  5.0
107 0.292 0.951 0.278
hERG 1.0
105  4.4
106 0.225 0.945 0.213
LXR 5.4
105  2.4
105 0.170 0.945 0.161
MR 1.4
106  5.7
107 0.289 0.950 0.274
PPARg 7.1
106  3.0
106 0.236 0.947 0.224
PR 1.5
106  6.2
107 0.286 0.949 0.272
TRa 1.7
105  6.0
106 0.208 0.956 0.199
TRb 9.7
106  4.1
106 0.226 0.947 0.214

Overall toxic potential = 0.484 (Eq. (3); range from 0.0 to 1.0); main target = ERb.

full Monte-Carlo sampling: software Cheetah: Rossato et al., 2010;
Vedani et al., 2012). The last step (red borders) comprises the
quantification of the individual binding affinities (software
BzScore4D) and the estimation of the toxic potential therefrom
(Vedani et al., 2012). These pieces of information—along with the
3D structures of all protein–ligand complexes—are then made
available to the client via the user interface. All relevant data can be
downloaded and stored locally and, most important, removed
completely from the server. The VirtualToxLab servers (currently
featuring 512 cores) are hosted by the University of Basel and
located in a physically and electronically safe environment.
The flexible-docking protocol employed in Alignator and
Cheetah aims at identifying all potential binding modes at the
target protein, thereby specifically allowing for induced fit and
dynamic solvation (Vedani et al., 2012). The underlying force field
features directional terms for hydrogen bonds and metal–ligand
interactions, allows for metal–ligand charge transfer and includes
polarization terms (cf. Fig. 2; Vedani and Huhta, 1990; Rossato
et al., 2010). During the conformational sampling of a small
molecule in the binding pocket of a protein, a total of 6000
(optionally: 12,000 in double-sampling mode) different binding
poses are generated at each of the 16 currently employed proteins,

Fig. 5. Graphical user-interface of the VirtualToxLab. The embedded 3D viewer/model builder allow designing, submitting and analyzing any compound of interest. The
models are displayed in real-time 3D (cf. http://www.biograf.ch/data/projects/mechanism.php) which allows verifying any prediction by inspecting the binding mode at the
target protein(s) with atomic resolution. All images are generated with the embedded viewer or builder.
 A. Vedani et al. / Toxicology Letters 232 (2015) 519–532 523

Fig. 6. Binding of bisphenol A to the estrogen receptor b. In the 4D representation (stereo image; top panel) the individual poses are rendered employing Boltzmann-scaled
intensities, i.e., the higher its contribution to the binding affinity the more prominent (i.e., the more opaque) it is depicted. The ligand is depicted in licorice mode, the protein
and the solvent by lines. The four individual binding modes (middle panel) disclose the details, e.g., the hydrogen bonds, which are represented as yellow, dashed lines. All
images were generated with the novel VTLViewer4D (Dobler, 2014). The bottom panel shows the experimental electron density map of bisphenol A bound to the estrogen
receptor a (PDB code = 3UU7, resolution = 2.2 Å). The experimental pose corresponds to pose 1 of our computed structures. The image was generated with the ASTEX viewer
(accessed through the PDB at http://www.rcsb.org).

120 (240) thereof fully minimized and 12 (24) each are retained for
the quantification of the binding energy (Vedani et al., 2012). This
protocol is computationally demanding and results in 20–80 h of
CPU time for the estimation of the toxic potential of a single
compound. Our Linux cluster currently hosts 512 cores, allowing
for an average process rate of 300–400 compounds per day.
The sampling protocol has been previously described (Rossato
et al., 2010; Vedani et al., 2012). The calculation of the associated
binding affinity, however, has been completely redesigned. While
the previously employed mQSAR technology (the 6D-QSAR
software Quasar, cf. Vedani et al., 2000, 2005; Vedani and Dobler,
2002) represents the highest QSAR level, its predictive power is—as
in principle for any such approach—limited to compounds at least
similar to the ensemble of ligands represented in the underlying
training set. For that reason, this algorithm was augmented by a
direct (force-field based) estimation of the binding affinity,
obtained from a molecule's identified binding modes at a
particular protein. While this enlarged the applicability range
for predictions, it still suffered from the problematic of deriving
reliable energy values from force-field calculations. Particularly for
large and flexible or highly charged ligands, this sometimes yielded
less accurate binding affinities.
We therefore decided to employ a novel scheme by calculating
free energies of ligand binding, DG, from the difference in
interaction energies of a compound’s 4D representations in
aqueous solution (mimicking the cytoplasm) with those at the
target protein (Fig. 3). While binding affinities derived with this
protocol may not reach the accuracy previously obtained through
mQSAR, the underlying protocol does not require the training of
any data. Instead, the binding energy (and, therefrom, the binding
affinity) is obtained in an “ab initio”-type approach, solely
depending on the quality of the underlying force field. This would
seem of utmost importance for our aim, as the compounds
submitted to the VirtualToxLab by third-party users are typically
not similar to any of those in the (previously employed) training
set.
In case of the phytoestrogen genistein binding to the estrogen
receptor b (cf. Fig. 3), the ligand–solvent interaction is computed to
15.1 kcal/mol for the sampled 25 low-energy conformers. At the
protein, the corresponding quantity yields 25.8 kcal/mol for the
524 A. Vedani et al. / Toxicology Letters 232 (2015) 519–532

Fig. 7. Comparison of experimental (lime color) and predicted binding modes (colored by atom type) as obtained with the VirtualToxLab 5.0 employing the Alignator and Cheetah
software. Key hydrogen bonds are depicted as yellow, dashed lines. (A) diethylstilbestrol bound to the estrogen receptor a (PDB code: 2ERD, resolution = 2.0 Å); (B) genistein
bound to the estrogen receptor b (PDB code: 1QKM, resolution = 1.8 Å); (C) dexamethasone bound to the glucocorticoid receptor (PDB code: 1M2Z, resolution = 2.5 Å); (D)
progesterone bound to the progesterone receptor (PDB code: 1A28, resolution = 1.8 Å). The experimental structures lack hydrogen atoms, as they are typically not resolved in
protein structures determined by X-ray crystallography. Additional examples (for nuclear receptors) are given in Smieško (2013). All images were generated with the VMD
software (Humphrey et al., 1996). (For interpretation of the references to color in this figure legend, the reader is referred to the web version of this article.)

Fig. 8. Binding modes (4D ensembles) as identified by automated, flexible docking (software Alignator and Cheetah). The protein is depicted as ribbons, the small molecules as
stick models colored by atom. (A) diethylstilbestrol bound to the estrogen receptor a; (B) genistein bound to the estrogen receptor b; (C) dexamethasone bound to the
glucocorticoid receptor; (D) progesterone bound to the progesterone receptor. For clarity, only the energetically most favorable poses are shown. The images were generated
with the BioX software (Dobler, 2012).
 A. Vedani et al. / Toxicology Letters 232 (2015) 519–532
Fig. 9. Toxicityalerts as generated bytheVirtualToxLab 5.0. The complete listing for 2665 compounds is posted at http://www.biograf.ch/data/projects/virtualtoxlab_results.php.
12 lowest-energy poses. The energy difference between those two
states, 10.7 kcal/mol, corresponds to a binding affinity of
1.1
108 M which is close to the experimental value of
1.2
108 M. To challenge the technology, we have employed
1288 compounds—representing over 30 chemical classes—to test
the predictive power of our approach (Fig. 4). 706 of the
1288 compounds (55%) are predicted within one, 1082 (84%)
within two, and 1232 (96%) within three orders of magnitude from
the experiment. The predictive r2 is computed to 0.574, which
would seem to be moderate only, particularly when compared to
values previously obtained by mQSAR techniques for the very
16 target proteins where the individual predictive r2 ranged from
0.739 to 0.942 (Vedani et al., 2012; Vedani, 2014). In contrast
hereto, the binding affinities in the most recent version of the
VirtualToxLab are no longer derived from trained models but,
instead, computed directly from the difference of a compound’s
interaction with the solvent and the target protein. Hence, they are
independent from any training set and may be safely applied to any
type of molecule, albeit with a slightly reduced predictive power
when compared to trained (e.g., QSAR) models. Of course, this
holds only for compounds similar to the (formerly) trained
ligands—for all other classes of molecules, only a direct scoring
approach may generate reliable predictions.
525
Compounds predicted too weakly (Fig. 4: points below the
diagonal) can be explained by the fact that our sampling protocol
(cf. above) is by no means exhaustive and that for this particular
case, the correct binding pose could not be identified. Most of these
compounds bind to proteins with large binding pockets, such as
hERG, LXR, PPARg and CYP3A4. On the other hand, compounds
predicted too strongly (Fig. 4: points above the diagonal) might
trigger an induced fit that has been simulated but could not be
appropriately quantified. Other factors of uncertainty include
entropic effects and the quantification of protein-bound solvent
released upon ligand binding. A final source of inaccuracy may
stem from the sampling of a compound’s representations in
aqueous solution (software Aquarius). While currently the 25
energetically most favorable conformations (obtained from
conformational sampling employing an implicit solvent model;
software MacroModel), are optimized in explicit solvent, they may
not include all relevant representations. We modified the protocol
to include 100 conformers (requiring approximately 2–4 extra CPU
hours per compound) but, unfortunately, with only minimal
benefit.
The philosophy underlying the VirtualToxLab is to estimate the
toxic potential of a compound through the normalized individual
binding affinities towards a series of protein models known or
 526 A. Vedani et al. / Toxicology Letters 232 (2015) 519–532

Fig. 10. Binding of tetrahydrogestrinone to the androgen receptor as identified through automated, flexible docking (software Alignator and Cheetah). The ligand and key
amino-acid residues are depicted in licorice mode, the protein main-chain by cartoons and the inner surface (colored by depth). Protein-bound water molecules are shown as
blue beads. Hydrogen bonds are highlighted as yellow, dashed lines. The 3D (left/right stereo) representation allows for an optimal depth perception—particularly of the
protein cavities. The image was generated with the VMD software (Humphrey et al., 1996). (For interpretation of the references to color in this figure legend, the reader is
referred to the web version of this article.)

suspected to trigger adverse effects. The result is a value ranging 0.484 suggests a moderate risk, particularly with respect to
from 0.0 (none) to 1.0 (extreme), which may be interpreted as a binding to the estrogen receptor b. The VirtualToxLab estimates the
toxicity alert. In a first step, the individual binding affinities are binding affinity at 54 nM, which compares well with the
normalized for each individual target protein according to Eq. (1). experimental value of 90 nM. Apart from the estrogen receptor

affinity > 1:0
102 M ! affinitynorm ¼ 0:0 )

2 10 ½logð1:0
102 Þ  logðaffinityÞ
1:0
10 M  1:0
10 M ! affinitynorm ¼ (1)
½logð1:0
1010 Þ  logð1:0
102 Þ
10
affinity < 1:0
10 M ! affinitynorm ¼ 1:0

Next, the individual toxic potential, TPindividual, is calculated, again
for each individual target protein (Eq. (2)):
TPindividual ¼ affinitynormalized
weightstandard deviation
with weightstandard deviation = 1.0–0.125
(standard deviation/affin-
ity); standard deviation over the 12 (24) models and therein:
0.125 = 1/DpKmin,max (DpKmin,max = 8.0: affinity range from
1.0
102 M to 1.0
1010 M).
Therefrom, the overall toxic potential (TPoverall) is determined
as follows: first, the 16 TPindividual are ranked by their value. Then,
their contribution to the TPoverall is summed up according to Eq. (3).
X
16  
TPoverall ¼ 1:0  TPoverall;current
TPindividual;n
Wsuper family
n¼1
with wsuper family = 1.0/n (n: nth member of a super family).
To avoid substantial TPs resulting from high affinities to
evolutionary similar protein targets (e.g., ERa and ERb), a
correcting weight, wsuper family, is applied. It decreases the
contribution for the nth member to the TP. Currently, five super
families are recognized: nuclear receptors I (AR, ERa, ERb, GR, MR,
PR), nuclear receptors II (LXR, PPARg, TRa, TRb), cytochrome
P450 enzymes (1A2, 2C9, 2D6, 3A4) and two singular families (AhR,
hERG).
Table 1 represents an extract from the log file and shows how
the toxic potential (TP) is calculated. From the normalized binding
affinity (affnorm) using the weights reflecting the standard
deviation (we.s.d.), the individual toxic potential (TPind) is obtained
for each of the 16 target proteins. After ranking the contributions
and using Eq. (3), the overall TP is calculated. The example shows
how the toxic potential for bisphenol A (a polymer additive present
in many products of our daily life) is computed. The overall value of
b, the compound would also seem to bind moderately to the
androgen receptor (460 nM), the glucocorticoid receptor (1.3 mM),
the mineralocorticoid receptor (1.4 mM), and the estrogen receptor
(2)
a (8.0 mM). The graphical-user interface allowing to up/download
data and to inspect/visualize results is 3D and 4D shown in Fig. 5.
Fig. 6 shows the 4D representation of bisphenol A binding to the
estrogen receptor b. The calculated binding affinity of 54 nM
compares acceptably with the experimental value of 90 nM. The
most prominent pose contributes 79.2% to the binding affinity, the
second one 13.1% with the remaining poses contributing 7.7%.
Multiple binding modes of small molecules binding to proteins
have also been experimentally identified (see, for example, Pineda-
Sanabria et al., 2011; Wang et al., 2013). This suggests that a 4D
representation might be preferred over a 3D approach. The
(3) computational expense, although significant, would seem to be
justified because the biologically relevant pose might be missed
when simply selecting the energetically most favorable binding
mode. Even experimental techniques (e.g., X-ray crystallography)
might not always identify the bioactive conformation, particularly
if the crystallization conditions (pH, buffer, temperature) are
different from those at the physiological state.
3. Results and discussion
Predicting the binding affinity of a small molecule towards a
protein first requires the binding mode being correctly and
accurately identified. To test our algorithm (cf. Vedani et al.,
2012; Rossato et al., 2010), we have applied the docking protocol
implemented in the VirtualToxLab (i.e., software Alignator and
Cheetah) to molecular systems for which the binding mode has
been identified by means of X-ray crystallography. Fig. 7 compares
the lowest-energy conformer as obtained through automated,
flexible docking (software Alignator and Cheetah) with the
 A. Vedani et al. / Toxicology Letters 232 (2015) 519–532 527

Fig. 11. Probing the kinetic stability of the danazol–androgen and estrogen receptor b complexes, respectively, by means of 5.0
109 s molecular-dynamics simulations. Top
panel: danazol binding to the androgen (left) and estrogen receptor b (right). The ligand is depicted licori-ce mode, the protein by lines, and hydrogen bonds are shown as yellow
dashed lines. The images were generated with the VMD software (Humphrey et al.,1996). Bottom panel: molecular-dynamical behavior of the protein–ligand complexes. The total
ligand–protein interaction energies (recorded at 1.0
1011 s intervals) are shown as thick lines (AR: blue, ERb: pink), those with key hydrogen bonds by thin lines. While the
thermodynamic binding affinity for the AR would seem to be in agreement with the molecular-dynamics simulations, that for the ERb is not, suggesting a binding affinity in the mid
nM range. Details are given in the text. (For interpretation of the references to color in this figure legend, the reader is referred to the web version of this article.)

experimental X-ray crystal structure. While the rms agreement is
clearly within 1.0 Å (for B and D even within 0.5 Å), this is not
necessarily sufficient for calculating the binding affinity within a
factor of 10 (corresponding to 1.4 kcal/mol in a protein–ligand
complex featuring a total energy ranging form 5000 to 10,000 kcal/
mol) as the conformational flexibility of the ligand at the binding
site is not accounted for. Such conformational and rotational
flexibility has been verified, for example, through solution – NMR
techniques for M2WJ 332 binding to an artificial 13-base pair
construct (Wang et al., 2013).
In earlier accounts (Vedani et al., 2000, 2005; Vedani and
Dobler, 2002) we have demonstrated that a 4D representation
including all (Boltzmann weighted) feasible poses can provide
more accurate estimations of the associated binding affinities.
Fig. 8 shows the corresponding 4D ensembles for the very
compounds: diethylstilbestrol bound to the estrogen receptor a,
genistein bound to the estrogen receptor b, dexamethasone bound
to the glucocorticoid receptor and progesterone bound to the
progesterone receptor. The individual poses are Boltzmann-
weighted, i.e., only the energetically most favorable binding
modes contribute significantly to the binding energy.
Using the VirtualToxLab, we have estimated the toxic potential
(endocrine and metabolic disruption, some aspects of carcinoge-
nicity and cardiotoxicity) for over 2500 compounds—drugs,
chemicals and natural products—and posted the results on
http://www.virtualtoxlab.org. The aim of the technology is to
generate toxicity alerts, i.e., ranking the tested compounds in three
groups: toxic potential (TP)  0.3 (low), 0.3 < TP  0.6 (moderate)
TP > 0.6 (high). Fig. 9 shows the toxic potential for a selection of
compounds.
More informative than the toxic potential itself is the
underlying binding-energy profile (cf. Table 1 for bisphenol A),
as it provides specific information at which target protein an
elevated binding affinity—potentially triggering an adverse effect—
528 A. Vedani et al. / Toxicology Letters 232 (2015) 519–532

Fig. 12. Bindingof the perfumeodorant(R,R)-galaxolide tothe progesterone receptor. The ligand is colored byatom,the protein’s main chain representedbyribbons. The lipophilic
residues lining the binding pocket are depicted by their volume (yellow surface). The hydrogen bond to Gln725 is indicated a yellow, dashed line. The 3D image was generated with
the VMD software (Humphrey et al., 1996). (For interpretation of the references to color in this figure legend, the reader is referred to the web version of this article.)

might be expected (cf. also the fingerprinting display mode in
Fig. 5). The VirtualToxLab interface allows exporting the individual
binding affinities into a csv file and, hence, to compute a
customized toxicity alert. Most important, our technology allows
rationalizing a given binding affinity by inspection of the
associated protein–ligand complexes in real-time 3D using the
embedded 3D/4D viewer or, after exporting the coordinates in PDB
format, with any other software of choice. Fig. 10 shows the
computed binding mode of the anabolic steroid tetrahydrogestri-
none to the androgen receptor. The associated binding affinity of
32 nM compares reasonably well with the experimental value of
8.5 nM.
As the docking and scoring algorithms within the VirtualToxLab
are based solely on thermodynamic considerations, it is suggested
to probe the kinetic stability of the protein–ligand complex of
interest by means of molecular-dynamical simulations. If the key
interactions (hydrogen bonds, salt bridges, binding to metal ions,
hydrophobic contacts) remain stable throughout a decent simula-
tion time (t  5.0
109 s), the “thermodynamic findings” may be
accepted; otherwise, the underlying affinity and the associated
toxic potential must be reduced accordingly. Anabolic steroids, for
example, are expected to bind strongly to the androgen receptor,
but less significantly to the estrogen receptors a and b. This was,
for example, found for stanozolol (AR = 6.1 nM, ERb = 350 nM,
corresponding to a selectivity factor of 57) but not for danazol (AR:
14 nM, ERb = 4.8 nM; selectivity factor < 1.0). We therefore simu-
lated the dynamic behavior of both the danazol–androgen and
estrogen receptor b complexes for 5.0
109 s each using the
Desmond software (Bowers et al., 2006) as implemented in the
VirtualDesignLab (Eid et al., 2013). The results are illustrated in
Fig. 11. For the danazol–androgen complex, the total ligand–
protein interaction energy (sampled at 1.0
1011 s intervals: blue
line) is varying between 37.5 and 54.3 kcal/mol (thick blue line),
the average being 47.8 kcal/mol. The key hydrogen bonds—
necessary to trigger an agonistic effect—with Asn705 (thin red line:
average energy = 5.5 kcal/mol), Thr877 (thin green line: 3.9
kcal/mol) and Arg752 (thin cyan line: 3.6 kcal/mol) are stable
throughout the entire simulation. In contrast hereto, the ligand–
protein interaction energy for the danazol–estrogen receptor b
complex varies between 24.6 and 41.6 kcal/mol (thick pink
line), the average being 33.6 kcal/mol. The key hydrogen bonds—
necessary to trigger an agonistic effect—with His475 (thin black
line: average energy = 0.4 kcal/mol), Glu305 (thin yellow line:
1.3 kcal/mol) and Arg346 (thin blue line: 0.1 kcal/mol) are never
really established throughout the entire simulation. This suggests
that the “kinetic” binding affinity is significantly lower (e.g., a factor
of 100) and the selectivity factor appropriate as expected.
A compound’s bioavailability is a prerequisite for its binding to a
target protein. Various molecular descriptors (e.g., lipophilicity,
solvent-accessible and polar surface areas, H-bond donors and

Fig. 13. Binding of the parent compound zearalenone (left) and one of its metabolites b-zearalanol (right) to the estrogen receptor b. The ligands and key amino-acid residues
are colored by atom, the inner protein surface by depth gradient. Hydrogen bonds are indicated as yellow dashed lines. While both compounds engage in a hydrogen bond
with Glu305 (visible at the bottom), b-zearalanol is additionally stabilized by a hydrogen bond with His475 (top right), resulting in a 28-fold increased computed binding
affinity: 29 nM versus 820 nM. Both images were generated with the VMD software (Humphrey et al., 1996). (For interpretation of the references to color in this figure legend,
the reader is referred to the web version of this article.)
 A. Vedani et al. / Toxicology Letters 232 (2015) 519–532 529

Fig. 14. (A) Binding of E121 (citrus red 2) to the aryl hydrocarbon receptor; (B) DHCMT to the androgen receptor; (C) drospirenone to the progesterone receptor; (D) bisphenol
A to the estrogen receptor b. The ligands and key amino-acid residues are colored by atom, the inner protein surface by depth gradient. Hydrogen bonds are indicated as
yellow dashed lines. All images were generated with the VMD software (Humphrey et al., 1996). (For interpretation of the references to color in this figure legend, the reader is
referred to the web version of this article.)

acceptors, rotatable bonds, membrane permeability) have been
found to be closely associated with the oral (Lipinski, 2000; Veber
et al., 2002) as well as transdermal (Lian et al., 2008) availability of
a compound. Values for such descriptors may nowadays be readily
computed by freely available tools found in the Internet. The
binding of the perfume odorant galaxolide to the progesterone
receptor may serve as an example. The calculated binding affinity
of 560 nM seems to be somewhat worrying. Visual inspection
(Fig. 12) justifies the prediction as the binding mode shows a well-
defined H-bond with the side-chain amide of Gln725 and the
hydrophobic portion of the molecule properly accommodated in
the lipophilic part of the binding pocket. The averaged computed
logP of galaxolide of 4.8  0.9 (software ALOGPS; Tetko and
Tanchuk, 2002) along with its low polar surface area (PSA < 10 Å2)
and absence of any rotatable bonds and H-bond donors suggests,
that galaxolide could be available in systemic circulation if ingested
or applied to the skin. The compound has been indeed detected in
the plasma of healthy young adults using body-lotion cosmetics in
concentrations up to 4.1 mg/L (Hutter et al., 2005). However, even
at such a concentration, galaxolide should not substantially
interfere with the endogenous ligand progesterone (Kexp = 3.7 nM,
Kcalc = 22 nM).
False-positive predictions may, thus, occur in all cases where
the kinetic stability of a protein–ligand complex is lower than the
thermodynamic and—probably more relevant—when the ADME
predisposition is unfavorable. We therefore plan to augment our
technology with a series of corresponding pre-filters in the near
future.
False-negative predictions may occur for at least three reasons.
Firstly (and most frequently), when the adverse effect of a
compound is triggered mechanisms other than those currently
tested in the VirtualToxLab. Examples include Ochratoxin A (OcA), a
well known mycotoxin which does not significantly bind to any of
our target proteins and is associated with a toxic potential of
0.519 suggesting only a moderate toxicity. While the toxic
mechanism of OcA has not yet been fully disclosed (see, for
example, Sorrenti et al., 2013), a critical step of the toxic pathway is
the long residence time of OcA at the plasma protein serum
albumin. Secondly, a toxic response may be triggered by a
metabolite rather than by the parent compound. While our
technology does not automatically generate feasible metabolites
(several pieces of third-party software have been developed for
this very purpose), at least primary metabolites should always be
tested along with a parent compound. In an earlier study, we have
analyzed the activity of cyclo-diBA (a condensation product of
glycidyl ether and bisphenol A) metabolites—a compound that is
unintentionally formed as by-product during the coating of food
cans and, due to its lipophilic character, migrates from the epoxy
resin of the coating into the fatty tissues e.g., of canned fish
(Biedermann et al., 2013). Another example includes the metab-
olites of the mycotoxin zearalenone, which are known to display
estrogenic activity (see, for example, Takemura et al., 2007;
Metzler et al., 2010). While the VirtualToxLab suggests a toxic
potential of 0.409 for the parent compound, one of its metabolites,
b-zearalanol, is estimated at 0.504. Fig. 13 compares the identified
binding modes for the parent compound zearalenone and its
metabolite b-zearalanol. Another reason for a false-negative
530 A. Vedani et al. / Toxicology Letters 232 (2015) 519–532

Table 2
Comparison of calculated and experimental binding affinities for selected targets/compounds.

Target Compound Experimental affinity (M) Calculated affinity (M) Factor offa
AR Bisphenol A 7.5
105 3.9
108 19
DDT 1.8
105 1.7
106 10
Diethylstilbestrol 1.4
105 1.3
106 10
17b-Estradiol 4.1
107 6.0
108 5.8
Genistein 8.5
105 1.1
105 6.9
Tetrahydrogestrinone 8.4
109 3.2
108 2.8

ERa Bisphenol A 1.2
105 5.9
106 1.0

Coumestrol 1.0
107 5.7
108 0.7
Diethylstilbestrol 2.3
1010 9.7
109 41
17b-Estradiol 9.1
1010 3.7
108 40
Genistein 2.1
107 5.7
107 1.8
Kaempferol 3.1
106 4.2
107 6.3

ERb Bisphenol A 9.3
108 5.3
108 0.7

Coumestrol 2.5
108 9.7
109 1.6
Diethylstilbestrol 6.4
109 9.7
1010 5.5
17b-Estradiol 3.9
109 4.0
109 0.1
Ethinylestradiol 3.4
109 7.6
109 1.2
Genistein 1.2
108 1.1
108 0.1

AhR Benzo[a]anthracene 3.5
107 1.3
107 1.7

Benzo[a]pyrene 3.6
107 1.3
107 1.8
Dibenzo[ah]anthracene 1.6
108 9.1
108 4.7
TCDD 1.0
108 2.1
108 1.1

GR Dexamethasone 3.0
109 1.8
109 0.6

Prednisolone 3.2
108 4.1
109 6.7

hERG Astemizole 9.1
1010 1.6
108 17

Cisapride 5.6
107 1.9
106 2.3

LiverX Podocarpic acid benzyl ester 6.3
107 7.6
107 0.2
MR Dexamethasone 1.0
109 5.7
109 4.7
PPARg Farglitazar 1.2
109 4.0
1010 2.0
PR Progesterone 3.7
109 4.0
109 0.6
TRa Desamino-triiodo-L-thryoinine 5.6
109 6.4
1010 7.8
TRb Triiodo-L-thryoinine 9.3
1011 5.5
1010 4.9
1A2 1,4-Dichloronaphtaline 2.3
106 8.5
106 2.7
2C9 Losartan 8.1
106 6.8
106 0.2
2D6 Bupivacaine 7.8
107 2.7
106 2.5
3A4 Miconazole 4.8
107 4.9
107 0.1
a
Deviation (factor off) = Kexp/kcalc 1.0 for Kexp > kcalc; factor = Kcalc/kexp 1.0 for Kexp < kcalc.
The “1.0” correction term is necessary, as for Kexp = kcalc, the ratio is 1.0 but the deviation 0.0.

prediction may lie in the fact that our sampling of the ligand at the
protein's binding site while extensive (cf. above) is not exhaustive.
Thus, the correct binding mode may simply not been have
generated within the 6000–12,000 trials. Finally, molecules that
trigger a substantial induced fit (i.e., including changes in the
protein's main-chain conformation) are currently beyond our
computational time scale.
We have applied the VirtualToxLab technology to a series of
2564 compounds. The results are posted on http://www.virtual-
toxlab.org. Fig. 14 shows four selected examples. According to our
calculations, E121 (citrus red 2; a food dye, classified as class 2B
carcinogen) displays an affinity of 420 nM towards the AhR, a
protein known to trigger chloracne and related diseases (see, for
example, Okey et al., 1994). The overall toxic potential is estimated
at 0.471, indicating a moderate risk at exposure or intoxication
(orange skins). Dehydrochloromethyltestosterone (DHCMT) has
been systematically applied to athletes in the former German
Democratic Republic with tragic consequences for some. The
VirtualToxLab calculates the binding affinity of DHCMT toward the
AR to 11 nM and estimates the toxic potential to 0.545. In the past,
drospirenone, an oral contraceptive has frequently been in the
news. According to our simulations, it not only binds to the
progesterone receptor (36 nM) and the estrogen receptor b
(310 nM) but also to the GR (43 nM). The overall toxic potential
is estimated at 0.640, which should be definitely interpreted as a
toxic alert. Bisphenol A, a known endocrine disrupter would
mainly seem to bind to the estrogen receptor b (54 nM; exp. = 93
nM); substantial affinities are also computed toward the GR
(1.3 mM) and the estrogen receptor a (8.0 mM). The overall toxic
potential is estimated to 0.484, suggesting a moderate risk,
particularly at prolonged exposure Table 2.
4. Conclusions
The VirtualToxLab—an in silico technology developed at the
Biographics Laboratory 3R, Basel—allows predicting the toxic
potential (endocrine and metabolic disruption, some aspects of
carcinogenicity and cardiotoxicity) of drugs, chemicals and natural
products. It is based on an automated protocol that simulates and
quantifies the binding of any small molecule towards a series of
16 proteins known or suspected to trigger adverse effects: the
androgen, aryl hydrocarbon, estrogen a, estrogen b, glucocorti-
coid, hERG, liver X, mineralocorticoid, progesterone, thyroid a,
thyroid b and the peroxisome proliferator-activated receptor g as
well as the enzymes cytochrome P450 1A2, 2C9, 2D6 and 3A4. The
toxic potential is derived from the binding affinities to these
16 target proteins, ranges from 0.0 (none) to 1.0 (extreme) and may
be interpreted as a toxic alert. The three-dimensional structure of
 A. Vedani et al. / Toxicology Letters 232 (2015) 519–532
compounds to be tested can generated using the embedded model
builder or imported from external sources. The results can be
inspected in real-time 3D or downloaded (coordinates of the
protein–ligand complexes in PDB format) and analyzed by
third-party software. The graphical user interface supports all
major operating systems (Mac OS X, Linux, Windows). The
calculation of the toxic potential of a compound depends on its
size and conformational flexibility. For a complete profile (16 target
proteins), this typically requires 20–80 h on a single processor of a
Unix cluster of the University of Basel, which is currently capable to
process 300–400 compounds/day in total. The user can choose to
make his/her simulation results (i.e., affinities, 3D binding modes,
toxic potentials) selectively visible to other users of the platform,
thus facilitating the dissemination of in silico toxicological results
of general interest. Emerging results and news on the technology
are continuously posted on http://www.virtualtoxlab.org. The
Open VirtualToxLab is freely accessible to universities, governmen-
tal or regulatory bodies, and non-profit organizations. On-line
registration (and 3D viewer libraries) are available at http://www.
biograf.ch/data/projects/OpenVirtualToxLab.php.
Conflict of interest
The authors declare that there are no conflicts of interest.
Transparency document
The Transparency document associated with this article can be
found in the online version.
Acknowledgements
The underlying research had been made possible by grants of
the Swiss National Science Foundation(#05321-135678), and the
Jacques en Dolly Gazan Foundation, Zug/Switzerland, which are
both gratefully acknowledged.
References
Amini, A., Mugleton, S.H., Lohdi, H., Sternberg, M.J.E., 2007. A novel logic-based
approach for quantitative toxicology prediction. J. Chem. Inf. Model. 47, 998–
1006.
Aronov, A.M., Balakin, K.V., Kiselyov, A., Varma-O’Brien, S., Ekins, S., 2007.
Application of QSAR to ion channels. In: Ekins, S. (Ed.), Computational
Toxicology—Risk Assessment for Pharmaceutical and Environmental Chemicals.
John Wiley & Sons, New York, USA, pp. 353–390.
Aronov, A.M., 2005. Predictive in silico models for hERG channel blockers. Drug
Discov. Today 10, 149–155.
Bars, R., Broeckaert, F., Fegert, I., Gross, M., Hallmark, N., Kedwards, T., Lewis, D.,
O'Hagan, S., Panter, G.H., Weltje, L., Weyers, A., Wheeler, J.R., Galay-Burgos, M.,
2011. Science based guidance for the assessment of endocrine-disrupting
properties of chemicals. Reg. Toxicol. Pharmacol. 59, 37–46.
Bender, A., Scheiber, J., Glick, M., Davies, J.W., Azzaoui, K., Hamon, J., Urban, L.,
Whitebread, S., Jenkins, J.L., 2007. Analysis of pharmacology data and the
prediction of adverse drug reactions and off-target effects from chemical
structure. ChemMedChem 2, 861–873.
Benfenati, E., Begigni, R., DeMarini, D.M., Helma, C., Kirkland, D., Martin, T.M.,
Mazzatorta, P., Quédraogo-Arras, G., Richard, A.M., Schilter, B., Schoonen, W.G.E.
J., Snyder, R.D., Yang, C., 2009. Predictive models for carcinogenicity and
mutagenicity: frameworks state-of-the-art, and perspectives. J. Env. Sci. Health
Part C. Environ Carcinog. Ecotoxicol. Rev. 27, 57–90.
Biedermann, S., Zurfluh, M., Grob, K., Vedani, A., Brüschweiler, B.J., 2013. Migration
of cyclo-diBA from coatings into canned food: method of analysis:
concentration determined in a survey and in silico hazard profiling. Food Chem.
Tox. 58, 107–115.
Bowers, K.J., Chow, E., Xu, H., Dror, R.O., Eastwood, M.P., Gregersen, B.A., Klepeis, J.L.,
Kolossváry, I., Moraes, M.A., Sacerdoti, F.D., Salmon, J.K., Shan, Y., Shaw, D.E.,
2006. Proceedings of the 2006 ACM/IEEE Conference on Supercomputing
Table of Contents. Tampa, FloridaACM, New York, NY, USA.
Colborn, T., vom Saal, F.S., Soto, A.M., 1993. Developmental effects of endocrine-
disrupting chemicals in wildlife and humans. Env. Health Perspect. 101, 378–
384.
531
Cramer, C.J., Truhlar, D.G., 1992. AM1-SM2 and PM3-SM3 parametrized SCF
solvation models for free energies in aqueous solution. J. Comp. -Aided Mol. Des.
6, 629–666.
Crivori, P., Poggesi, I., 2006. Computational approaches for predicting CYP-related
metabolism properties in the screening of new drugs. Eur. J. Med. Chem. 41,
795–808.
Cronin, M.T.D., Madden, J.C., 2010. In Silico Toxicology. RSC Publishing, Cambridge,
UK.
Cronin, M.T.D., Dearden, J.C., Walker, J.D., Worth, A.P., 2003. Quantitative structure–
activity relationships for human health effects: commonalities with other
endpoints. Env. Toxicol. Chem. 22, 1829–1843.
Custer, L.L., Kreatsoulas, C., Durham, S.K., 2007. Predicitve mutagenicity computer
models. In: Ekins, S. (Ed.), Computational Toxicology—Risk Assessment for
Pharmaceutical and Environmental Chemicals. John Wiley & Sons, New York,
USA, pp. 391–401.
De Graaf, C., Vermeulen, N.P.E., Feenstra, K.A., 2005. Cytochrome P450 in silico: an
integrative approach. J. Med. Chem. 48, 2725–2755.
Dibb, S., 1995. Swimming in a sea of estrogens, chemical hormone disrupters.
Ecologist 25, 27–31.
Dobler, M., 2012. BioX—a versatile molecular-modeling software. http://www.
biograf.ch/index.php?id=software (accessed 4.09.14.).
Dobler, M., 2014. Software is part of the VirtualToxLab distribution.
Ecker, G.F., Chiba, P., 2007. QSAR studies on drug transporters involved in toxicology.
In: Ekins, S. (Ed.), Computational Toxicology—Risk Assessment for
Pharmaceutical and Environmental Chemicals. John Wiley & Sons, New York,
USA, pp. 295–314.
Eid, S., Zalewski, A., Smieško, M., Ernst, B., Vedani, A., 2013. A molecular-modeling
toolbox aimed at bridging the gap between medicinal chemistry and
computational sciences. Int. J. Mol. Sci. 14, 684–700.
Ekins, S., 2014. Progress in computational toxicology. J. Pharmacol. Toxicol. Methods
69, 115–140.
Ekins, S., 2007. Application of QSAR to enzymes involved in toxicology. In: Ekins, S.
(Ed.), Computational Toxicology—Risk Assessment for Pharmaceutical and
Environmental Chemicals. John Wiley & Sons, New York, USA, pp. 277–294.
Enoch, S.J., Cronin, M.T.D., Schultz, T.W., Madden, J.C., 2008. Quantitative and
mechanistic read across for predicting the skin sensitization potential of
alkenes acting via Michael addition. Chem. Res. Toxicol. 21, 513–520.
Fischer, B., 2000. Receptor-mediated effects of chlorinated hydrocarbons.
Andrologia 32, 279–283.
Green, N., Naven, R., 2009. Early toxicity screening strategies. Curr. Opin. Drug
Discov. Dev. 12, 90–97.
Guillette, L.J., Crain, D.A., Rooney, A.A., Pickford, D.B., 1995. Organization versus
activation: the role of endocrine disrupting contaminants EDCs during
embryonic development in wildlife. Env. Health Perspect. 103 (Suppl. 7), 157–
164.
Gupta, S., Kapoor, P., Chaudhari, K., Gautam, A., Kumar, R., Raghava, G.P.S., 2013. In
silico approach for predicting toxicity of peptides and proteins. PLoS One 8,
e73957.
Helma, C., 2005. In silico toxicology: the state-of-the-art and strategies to predict
human health effects. Curr. Opin. Drug Discov. Dev. 8, 27–31.
Humphrey, W., Dalke, A., Schulten, K., 1996. VMD – visual molecular dynamics. J.
Mol. Graph. 33–38.
Hutter, H.P., Wallner, P., Moshammer, H., Hartl, W., Sattelberger, R., Lorbeer, G.,
Kundi, M., 2005. Blood concentrations of polycyclic musks in healthy young
adults. Chemosphere 59 (4), 487–492.
Kavlock, R.J., Ankley, G., Blancato, J., Breen, M., Conolly, R., Dix, D., Houck, K., Hubal,
E., Judson, R., Rabinovitz, J., Richard, A., Setzer, R.W., Shah, I., Villeneuve, D.,
Weber, E., 2008. Computational toxicology – a state of the science mini review.
Toxicol. Sci. 103, 14–27.
Lian, G., Chen, L., Han, L., 2008. An evaluation of mathematical models for predicting
skin permeability. J. Pharm. Sci. 97, 584–598.
Lipinski, C.A., 2000. Drug-like properties and the causes of poor solubility and poor
permeability. J. Pharm. Toxicol. Methods 44, 235–249.
McLachlan, J.A., Arnold, S.F., 1996. Environmental estrogens. Am. Sci. 84, 452–461.
Merlot, C., 2008. In silico methods for early toxicity assessment. Curr. Opin. Drug
Discov. Dev. 11, 80–85.
Metzler, M., Pfeiffer, E., Hildebrand, A.A., 2010. Zearalenone and its metabolites as
endocrine disrupting chemicals. World Mycotoxin J. 3, 385–401.
Muster, W., Breidenbach, A., Fischer, H., Kirchner, S., Müller, L., Pähler, A., 2008.
Computational toxicology in drug development. Drug Discov. 13, 303–310.
Nigsch, F., Macaluso, N.J.M., Mitchell, J.B.O., Zmuidinavicus, D., 2009. Computational
toxicology: an overview of the sources of data and modeling methods. Expert
Opin. Drug Metab. Toxicol. 5, 1–14.
Okey, A.B., Riddick, D.S., Harper, P.A., 1994. The Ah receptor: mediator of the toxicity
of 2, 3,7,8-tetrachlorodibenzo-p-dioxin and related compounds. Toxicol. Lett.
70, 1–22.
Pavan, M., Worth, A., 2008. Publicly accessible QSAR software tools developed by the
Joint Research Centre. SAR QSAR Env. Res. 19, 785–799.
Piclin, N., Pintore, M., Wechman, C., Roncaglioni, A., Benfenati, E., Chretien, J.R.,
2006. Ecotoxicity prediction by adaptive fuzzy-partitioning: comparing
descriptors computed on 2D and 3D structures. SAR QSAR Env. Res. 17, 225–251.
Pineda-Sanabria, S., Robertson, I.M., Sykes, B.D., 2011. Structure of trans-resveratrol
in complex with the cardiac regulatory protein troponin C. Biochemistry 50,
1309–1320.
Rihova, B., 1998. Receptor-mediated targeted drug or toxin delivery. Adv. Drug Deliv.
Rev. 29, 273–289.
532 A. Vedani et al. / Toxicology Letters 232 (2015) 519–532
Roncaglioni, A., Toropov, A.A., Toropova, A.P., Benfenati, E., 2013. In silico methods to
predict drug toxicity. Curr. Opin. Pharmacol. 13, 802–806.
Rossato, G., Ernst, B., Smieško, M., Spreafico, M., Vedani, A., 2010. Probing small-
molecule binding to cytochrome P450 2D6 and 2C9: an in silico protocol for
generating toxicity alerts. ChemMedChem 5, 2088–2101.
Schilter, B., Benigni, R., Boobis, A., Chiodini, A., Cockburn, A., Cronin, M.T.D., Lo
Piparo, E., Modi, S., Thiel, A., Worth, A., 2014. Establishing the leval of safety
concern for chemicals in food without the need for toxicity testing. Reg. Toxicol.
Pharmacol. 68, 275–296.
Serafimova, R., Todorov, M., Pavlov, T., Kotov, S., Jacob, E., Aptula, A., Mekenyan, O.,
2007. Identification of the structural requirements of mutagenicity by
incorporating molecular flexibility and metabolic activation of chemicals. II.
General ames mutagenicity model. Chem. Res. Toxicol. 20, 662–676.
Shah, F., Greene, N., 2014. Analysis of Pfizer compounds in EPA’s ToxCast chemical-
assay space. Chem. Res. Toxicol. 27, 86–98.
Schrödinger, 2012. Software Epik and MacroModel. Schrödinger, LLC, New York,
USA.
Simon-Hettich, B., Rothfuss, A., Steger-Hartmann, T., 2006. Use of computer-assisted
prediction of toxic effects of chemical substances. Toxicology 22, 156–162.
Singh, K.P., Gupta, S., 2014. In silico prediction of toxicity of non-congeneric
industrial chemicals using ensemble-learning based modeling approaches.
Toxicol. Appl. Pharmacol. 275, 198–212.
Smieško, M., 2013. Dolina—docking based on a local induced-fit algorithm:
application toward small-molecule binding to nuclear receptors. J. Chem. Inf.
Model. 53, 1415–1423.
Sorrenti, V., Di Giacomo, C., Aqcuaviva, R., Barbagallo, I., Bognanno, N., Galvano, F.,
2013. Toxicity of ochratoxin A and its modulation by antioxidants: a review.
Toxins 5, 1742–1766.
Spreafico, M., Smieško, M., Lill, M.A., Ernst, B., Vedani, A., 2009. Mixed-model QSAR
at the glucocorticoid receptor: predicting the binding mode and affinity of
psychotropic drugs. ChemMedChem 4, 100–109.
Takemura, H., Shim, J.Y., Sayama, K., Tsubura, A., Zhu, B.T., Shimoi, K., 2007.
Characterization of the estrogenic activities of zearalenone and zeranol in vivo
and in vitro. J. Steroid Biochem. Mol. Biol. 103, 170–177.
Tetko, I.V., Tanchuk, V.Y., 2002. Application of associative neural networks for
prediction of lipophilicity in ALOGPS 2.1. J. Chem. Inf. Comput. Sci. 42, 1136–
1145.
Toropov, A.A., Toropova, A.P., Raska Jr., I., Leszcynska, Leszcynski, D., 2014.
Comprehension of drug toxicity: software and databases. Comput. Biol. Med 45,
20–25.
Valerio Jr., L.G., 2009. In silico toxicology for the pharmaceutical sciences. Toxicol.
Appl. Pharmacol. 241, 356–370.
Vedani, A. 2014. VirtualToxLab—user and reference manual. Version 5.0. http://
www.biograf.ch/downloads/VirtualToxLab.pdf (accessed 04.09.14.).
Vedani, A., Dobler, M., Smieško, M., 2012. VirtualToxLab—a platform for estimating
the toxic potential of drugs, chemicals and natural products. Toxicol. Appl.
Pharmacol. 261, 142–153.
Vedani, Dobler, M., Lill, M.A., 2005. Combining protein modeling and 6D-QSAR –
simulating the binding of structurally diverse ligands to the estrogen receptor. J.
Med. Chem. 48, 3700–3703.
Vedani, A., Dobler, M., 2002. 5D-QSAR: the key for simulating induced fit? J. Med.
Chem. 45, 2139–2149.
Vedani, A., Briem, H., Dobler, M., Dollinger, K., McMasters, D.R., 2000. Multiple
conformation and protonation-state representation in 4D-QSAR: the
neurokinin-1 receptor system. J. Med. Chem. 43, 4416–4427.
Vedani, A., Huhta, D.W., 1990. A new force field for modeling metalloproteins. J. Am.
Chem. Soc. 112, 4759–4767.
Veith, G.D., 2004. On the nature: evolution and future of quantitative structure–
activity relationships (QSAR) in toxicology. SAR QSAR Env. Res. 15, 323–330.
Veber, D.F., Johnson, S.R., Cheng, H.Y., Smith, B.R., Ward, K.W., Kopple, K.D., 2002.
Molecular properties that influence the oral bioavailability of drug candidates. J.
Med. Chem. 45, 2615–2623.
Vuorinen, A., Odermatt, A., Schuster, D., 2013. In silico methods in the discovery of
endocrine-disrupting chemicals. J. Steroid Biochem. Mol. Biol. 137, 18–26.
Wang, J., Wu, Y., Ma, C., Fiorin, G., Wang, J., Pinto, L.H., Lamb, R.A., Klein, M.L.,
Degrado, W.F., 2013. Structure and inhibition of the drug-resistant S31N mutant
of the M2 ion channel of influenza A virus. Proc. Natl. Acad. Sci. U. S. A. 110, 1315–
1321.
Further reading
http://www.virtualtoxlab.org (last accessed on September 4, 2014).
http://www.biograf.ch/data/projects/OpenVirtualToxLab.php (last accessed on
September 4, 2014).
http://www.biograf.ch/data/projects/mechanism.php (last accessed on September
4, 2014).
http://www.biograf.ch/data/projects/virtualtoxlab_results.php (last accessed on
September 4, 2014).
http://www.rcsb.org (last accessed on September 4, 2014).