T xT Kl xTf - ICM

Target Specific Compound Identification using Support Vector Machine.
Dariusz Plewczynski 1,2* , Marcin von Grotthuss 1 , Stephane Spieser 3 , Leszek Rychewski 1 , Lucjan S. Wyrwicz 1 ,
Uwe Koch 3
1) BioInfoBank Institute, Limanowskiego 24A/16, 60-744 Poznan, Poland, Tel: +48-61-8653520, Fax:
+48-61-8643350, E-mail: darman@bioinfo.pl
2) Interdisciplinary Centre for Mathematical and Computational Modeling, University of Warsaw,
Warsaw, Poland
3) Istituto di Ricerche di Biologia Molecolare (IRBM) “P. Angeletti”, Merck&Co. Inc., Pomezia, Italy
* The correspondence should be addressed to Dariusz Plewczynski (darman@bioinfo.pl) and Uwe Koch
(uwe_koch@merck.com).
RUNNING TITLE: Target Specific Compound Identification by Support Vector Machine.
KEYWORDS:
1) Compound Identification,
2) Protein target specificity,
3) MDL Drug Data Report,
4) Machine-learning methods,
5) Atom Pairs,
6) Support Vector Machine.
ABREVIATIONS:
1. SVM – support vector machine,
2. AP – atom pairs,
3. MDDR - MDL Drug Data Report.
1

ABSTRACT
In many cases at the beginning of a HTS-campaign some information about active molecules is already
available. Often active compounds (such as substrate analogues, natural products, inhibitors of a related protein
or ligands published by a pharmaceutical company) have been identified in low-throughput validation studies of
the biochemical target. We would like to evaluate in how far support vector machine can be trained on those
compounds and used to classify a collection with unknown activity. This approach is aimed on reducing the
number of compounds to be tested against the given target. Our method predicts biological activity of chemical
compounds based only on the Atom Pairs (AP) two dimensional topological descriptors. The supervised
Support Vector Machine (SVM) method is trained here on compounds from the MDL drug data report
(MDDR) known to be active for specific protein target. For detailed analysis we have selected five different
biological targets: cyclooxygenase-2, dihydrofolatereductase, thrombin, HIV-reverse transcriptase and
antagonists of the estrogen receptor. The accuracy of compounds identification is estimated here using the
recall and precision values. The sensitivities for all types of targets are over 80% and the classification
performance reaches 100% for selected targets. The second application of our method address the problem
when at the beginning of a HTS-campaign no initial set of actives is known on a selected protein target. Then
the virtual high-throughput screening (vHTS) is applied in most cases by flexible docking procedure. The vHTS
experiment typically contain a large percentage of false positives that should be verified by costly and time-
consuming experimental follow-up assays. The subsequent use of our machine learning method improves the
speed (you do not have to perform the docking on all compounds of the database) and also the accuracy of
HTS hit lists (the enrichment factor).
INTRODUCTION
Genomic research provides an ever increasing number of potential drug targets. In the past large compound
collections were tested for a single target. Recently it became common practice in the applied research to screen
large collections of compound for potential activity of in vitro high throughput screening (HTS) model studies
to identify new lead compounds. However, with a larger number of drug targets often the question is raised how
to preselect chemical versus biological space more efficiently. Our aim of the present study is to present fast
and reliable in silico method that captures the essential features of inhibitor molecules.
High throughput screening (HTS) allows for the testing of millions of compounds for activity against the
chosen target. As a result a set of lead molecules with relatively high activity against the target is identified.
The number of these compounds can be up relatively high and therefore these compounds are usually subjected
to further prioritization based on the assessment of various molecular characteristics. Although, highly
2

successful, this approach can not be applied simultaneously to the large number of drug targets emerging from
genomic research. One solution is to reduce the number of compounds to be tested to those with a high
probability of activity. There are many ways in which various computational methods can contribute in this
process. We focused here on the application of support vector machine (SVM) – i.e. supervised machine
learning approach. We evaluated our method in terms of its capability to recognize known ligands for five
divergent protein targets of the highest medicinal relevance, which already have been investigated in several
drug discovery programs.
Chemists have gathered expertise on features in molecular structures that are important for inhibition on
specific targets. Thus, even the 2D structure of ligand allows for some estimate of its activity for a given protein
target. We tried to describe this empirical knowledge about inhibitors in terms of computational prediction
model. By application of support vector machine algorithm we classify compounds that are active against this
biological target according to the commercially available MDL drug data report [1]. In the past various
machine-learning approaches have been used for a number of compound-based classification problems. For
example neural networks have been used as drug-likeness filters to distinguish drugs from non-drugs, to
classify compounds based on their ADME properties, toxicity and target specificity. In many of these
applications the use of parameters describing the compound’s topology gave satisfactory results. That is why
we have used AP two dimensional ligands descriptors to represent the variety of chemical space.
Target identification is a critical step following the discovery of small molecules. In [2] Nidhi et al.
provided an in silico method for predicting potential targets for compounds on the basis of chemical structure
alone. They used the multiple-category Laplacian-modified naive Bayesian model trained on extended-
connectivity fingerprints of compounds from 964 target classes in the WOMBAT (World Of Molecular
BioAcTivity) chemogenomics database. The algorithm was then tested by finding the three top most likely
protein targets for all MDDR (MDL Drug Database Report) database compounds [1]. On average, the correct
target was found 77% of the time for compounds from 10 MDDR activity classes with known targets [2]. The
support vector machine was used recently to describe high-throughput screening (HTS) data with great success
[3]. With carefully selected parameters, SVM models increased the hit rates significantly, and 50% of the active
compounds could be recovered by screening just 7% of the test set. The authors found that the size of the
training set played a significant role in the performance of the models, i.e. a training set with 10,000 member
compounds is likely the minimum size required to build a model with reasonable predictive power [3]. In other
work by using an in-house data set of small-molecule structures, encoded by Ghose-Crippen parameters,
several machine learning techniques were applied to distinguish between kinase inhibitors and other molecules
with no reported activity on any protein kinase [4]. They compared four approaches: support vector machines
(SVM), artificial neural networks (ANN), k nearest neighbor classification with GA-optimized feature selection
3

(GA/kNN), and recursive partitioning (RP). Support-vector machines, followed by the GA/kNN combination,
outperformed the other techniques when comparing the average of individual models. Similar to our approach
is presented in [5]. They have performed virtual screening using some very simple features, by employing the
number of atoms per element as molecular descriptors but without regard to any structural information
whatsoever. These atom counts are able to outperform virtual-affinity-based fingerprints and Unity fingerprints
in some activity classes. This fact can partly be explained by highly nonlinear structure-activity relationships,
which represent a severe limitation of the "similar property principle" in the context of bioactivity [5].
In our previous paper [6] we answered the following question: “How well do different classification methods
perform in selecting the ligands of a protein target out of large compound collections not used to train the
model?”. In this work support vector machines, random forest, artificial neural networks, k-nearest-neighbor
classification with genetic-algorithm-optimized feature selection, trend vectors, naive Bayesian classification,
and decision tree were used to divide databases into molecules predicted to be active and those predicted to be
inactive. Training and predicted activities were treated as binary. We reported significant differences in the
performance of the methods independent of the biological target and compound class. Different methods can
have different applications; some provide particularly high enrichment, others are strong in retrieving the
maximum number of actives. We also showed that these methods do surprisingly well in predicting recently
published ligands of a target on the basis of initial leads and that a combination of the results of different
methods in certain cases can improve results compared to the most consistent method. In the present paper we
focus our attention only on our novel SVM-based method and provide more in-depth description of the
methodology and results. We believed that this paper can help all readers to use our protocol for solving
similar HTS problems. Therefore we decided not to include a comparison of all current known approaches
thinking it is rather outside of the scope of this manuscript. Our results provide higher sensitivity and selectivity
comparing to other recently published methods (such as combination of SVM with naïve Bayesian trained on
Ghose-Crippen parameters and others [2-5]).
The list of hits generated by virtual high-throughput screening (vHTS) typically contain a large percentage of
false positives, making experimental follow-up assays necessary to distinguish active from inactive substances.
Here we would like to present another application of SVM based method aimed at improving the accuracy of
HTS hit lists by the subsequent use of machine learning method. The virtual screening procedure often is
performed on the large chemical libraries and selecting hits by statistical algorithms instead of time-costly
docking procedure is of great importance [7]. We address this problem by the case study on five protein targets:
HIV-reverse transcriptase, COX2, dihydrofolate reductase, estrogen receptor and thrombin in conjunction with
MDL Drug Data Report database [1]. The virtual HTS was performed with a set of different flexible docking
and scoring methods. Our results reveal that support vector machine algorithm is able to speed-up the vHTS
4

procedure by limiting the set of ligands to be docked on a target and provide also the better enrichment of the
HTS hit rate.
COMPUTATIONAL METHODS
Data sets and AP descriptors
Five diverse protein targets were tested: human cyclooxygenase-2 [8], dihydrofolate reductase [9, 10], thrombin
[11, 12], antiestrogen [13] and HIV reverse transcriptase [14]. The datasets used for training and testing is
comprised of both active and inactive compounds from the subset of the MDL drug data report [1]. All
compounds are clinically tested or already launched on the market. Inhibitors and non-active compounds for
those targets were used for training supervised machine-learning algorithm. For additional tests we have
selected also compounds that are now biologically tested for inhibition of targets, and are not yet on the marker
or clinical tests.
The entire pool of compounds for cyclooxygenase-2 target contains 112 inhibitors and 10452 inactive
compounds divided randomly into two subsets: training (75 inhibitors, 2106 inactive ones) and testing
(respectively 37 and 8346). In the case of dihydrofolate we have 28 inhibitors (divided into 17 for training and
11 for testing) and 10529 inactive ones (2149 and 8380). For reverse transcriptase we have selected 114
inhibitors (79 and 35) and 10450 inactive compounds (2130 and 8320), and for thrombin we have 112
inhibitors (77 and 35) and 10459 inactive ones (2036 and 8423). For the last target we have collected 34
inhibitors (22 and 12) and 11580 inactive compounds (2528 for training and 9052 for testing). For additional
test we have selected molecules biologically tested for inhibition of cyclooxygenase-2 (792 molecules),
dihydrofolate (154), thrombin (1066), reverse transcriptase (597) and antiestrogen (256).
We have utilized the regular atom pair AP descriptors [15] due to their proven success in classifying
compounds, ease of use and interpretability. To encode the molecule structures we employed the MIX tools
script [16], which counts for each atom pair the number of covalent bonds that join them. Thus for each
compound it yields a binary vector with 1 for all present types of atom pairs, and 0 for those that are absent in a
molecule. In addition to those we have tested also a larger set of additional 6 descriptors such as: TT (regular
topological torsion), DP (pairs using sq types), DT (torsions using SQ types), DRUGBITS (substructures) and
ROF6 set of descriptors [16]. When including in the training those additional descriptors we have not observed
any significant improvement of results.
Classification and model validation
5

We trained support vector machine algorithm for each type of a target. First, we have created the dataset of
compounds with experimentally verified activity (positive instances). Then we have built the dataset of inactive
compounds (negative instances). The negative instances are chosen randomly from launched or preclinical
compounds that have no experimentally verified activity for the selected type of the target. These two datasets
(positive and negative instances) are projected then as sets of points into a multidimensional space using AP
two dimensional descriptors described in the previous subsection.
In is well known that some machine-learning methods have difficulties handling unbalanced training sets, i.e.
when the number of positive instances is substantially different from the number of negatives. Therefore for
each target we selected randomly 1/3 of negatives for training and 2/3 for testing, whereas 2/3 of positive
instances for training and 1/3 for testing. The selections were repeated few times for all targets, yet no
significant differences were observed between those various selections. We present here the average results for
training and testing phase of experiments. Models are derived for each training set independently and used for
prediction of activity of the compounds in the test set.
There are many ways to present the performance of a SVM classifier. We use here accuracy E, precision P and
recall R values, together with confusion tables. Their definitions are given below:
fp + fn
E = * 100%
,
tp + fp + tn + fn
tp
R = * 100%
, [Eq. 1]
tp + fn
tp
P = * 100%
,
tp + fp
where tp is the number of true positives, fp is the number of false positives, tn is the number of true negatives
and fn is the number of false negatives. The classification error E provides an overall error measure, whereas
recall R measures the percentage of correct predictions (the probability of correct prediction), and precision P
gives the percentage of observed positives that are correctly predicted (the measure of the reliability of positive
instances prediction).
The Support Vector Machine (SVM) Method.
SVM is an effective statistical learning method [17-19] with good performance yet easier to implement
then neural networks. It was successfully applied to various problems including text classification [20], image
recognition tasks [21], bioinformatics [22, 23] and medical applications [24, 25]. The SVM approach has been
6

used also in analysis of gene expression data [26], classification of microarrays data [27], to infer gene
functional classification [28-30] and for protein analysis [31-33].
Most of those tasks have the property of sparse instance vectors. The SVM approach has the ability to
construct predictive models with the large generalization power even in the case of large dimensionality of the
data when the number of observation available for training is low. SVM always seeks a globally optimized
solution and avoids over-fitting, so the large number of features (as in our binary representation of ligands
topology) is allowed. The SVMlight implementation done by Thorsten Joachims [34] is used in the field of
bioinformatics [35]. The crucial idea behind is a sparse instance vectors property to obtain compact and
efficient representation.
The output of the training phase is a classification function i.e. a model. It consists from the set of D
support vectors Tj andα i , which are nonzero, positive real numbers. Those constants are obtained from
optimization procedure (quadratic programming QP problem) used to find the maximal margin hyperplane. The
number of free parameters of the QP problem is equal to the number of all instances in the training dataset. The
non-zero parameters α i describe the strength of this particular i-th support vector in the decision function.
SVM chooses as support vectors those points that lie closest to the separating hyperplane. The kernel function
is used to define the feature space after nonlinear mapping function from the embedding space. The mapping
function Ω need not be explicitly defined because in the kernel function is used only the inner product of it. The
kernel function is a positive define function reflecting the similarity between an input sample and the set of
support vectors Ti. In most cases three types of kernels are used: the linear, polynomial or radial basis.
The reliability of a classification of a ligand [36] as an active one is given by the cost function:
where ( T T )
i
f
[ ] ∑ ( { [ ] } { } )
= i D
( T x ) = liα
iK
Ω T x , Ω Ti
, [Eq. 1]
i=
1
K , is the proper kernel function that defines the feature space, Ω is a nonlinear mapping function
from embedding space T into the feature space, and li are known a priori class labels for support vectors. We
use l = + 1 for positive cases and l = −1for
negative ones. The kernel function is a positive define function
i
i
reflecting the similarity between an input sample and the set of support vectors Ti. The non-zero parameters α i
describe the strength of this particular i-th support vector in the decision function. SVM chooses as support
vectors those points that lie closest to the separating hyperplane. The mapping function Ω need not be explicitly
defined because in the kernel function is used only the inner product of it.
RESULTS AND DISCUSION
7

The major goal of this study was to test to what extent the supervised support vector machine method is capable
of learning and predicting the target specific inhibition likelihood for chemical compounds based on MDL drug
data report [1]. We have prepared for each of 5 different protein targets 2 datasets: for training and testing. Each
include compounds which are know to inhibit the protein target (2/3 of all available positives for training, 1/3
for testing) and those known not to be active for the selected target (1/3 of all available negatives for training,
2/3 for testing). Support vector machine algorithm was trained on the first data set and tested on the second one.
In the following paragraphs we discuss the classification results obtained for each of biological targets and
present its performance using confusion tables for training and testing datasets. In addition we provide also the
overall value for classification error and precision/recall values. In general the SVM models yield the successful
classification of compounds for all targets (see Table I for confusion tables for training and testing datasets). It
provides robust and reliable models for all types of protein targets. The SVM algorithm turn to be the method
of choice for any practical purpose: it is very fast, efficient and robust.
We performed two additional in silico experiments. The first one uses for training all available in MDL drug
data report active compounds annotated as preclinical or launched. The testing is done on potential inhibitors
that are annotated as biologically tested. The second experiment trains the SVM method on oldest 1/3
compounds known to be inhibitors, and test their accuracy on the rest: 2/3 newest developed and patented
compounds. The performance of machine-learning models in both experiments for each type of protein target is
described by the recall R and the precision P. The recall R value measures the percentage of correct predictions,
whereas precision P gives the percentage of observed positives that are correctly predicted. These measures of
accuracy are calculated separately for each type of protein target and presented in Tables II (the first
experiment) and Table III (the second experiment). The typical recall value is around 60%, and the precision P
is close to 100% for all targets (the first experiment). The results for the second experiment are slightly worst,
which is caused by the discovery of novel drug classes that are presented in the newest 2/3 compounds.
In the Table IV we present results on enrichment studies of virtual high-throughput screening. We trained SVM
algorithm on the set of first 10% of the best scoring ligands from the docking and scoring experiments for each
protein target. The subset of MDL drug data report inhibitors including both active and non-active compounds
for those targets were docked on protein targets using various docking methods (FLOG [37], GLIDE [38, 39],
FRED [40], Dock [41], Autodock [42, 43] and ICM [44]) followed in some cases by the scoring (SS,
ChemScore or internal ICM score). The 10% of best docked compounds were then used for training supervised
machine-learning algorithms. The classification models were tested then on the rest of active compounds (data
not shown). Our results support possibility to train machine learning algorithms on docking and scoring results.
Such trained models can be later applied to large databases in order to select ligands for further experimental
verification. This procedure will allow for faster selection of compounds in virtual HTS experiments even in
8

cases where no initial information about a set of active compounds for selected protein target is known. The
random scores, from training of SVM on randomly selected subset of 1000 compounds, for recall and precision
are equal to 50% and the most methods are able to gain recall/precision up to 70%.
Our results are in close agreement with other comparative studies [4, 7, 36, 45, 46]. The SVM method is fast
and reliable machine-learning method that outperforms other types of algorithms. It is also well suited for the
classification of small molecules using 2D descriptors with respect to their potential inhibition on selected
target classes. The selection of molecular descriptors should be done in accordance with the balance between
general and detailed level of description. The MIX tools descriptors [16] are useful for the classification of
compounds by SVM with respect to their potential for inhibition of selected protein targets. The subsequent
SVM training on results of docking experiment allows for speed-up vHTS procedure and to enrich the hit list.
Our aim was to present possible in silico application of machine learning methods to the HTS data. The number
of HTS hits can become large requiring some type of prioritization. The set of active compounds often contains
a substantial number of false positives. False negatives are also important but difficult to identify
experimentally. Application of a machine learning model can help to identify true positives and help to select
compounds for retesting. In still another application the results of the HTS itself can be used for training and the
model used to identify new compounds not present in the screening collection to be synthesized or bought from
a commercial source.
These tools can be also valuable whenever a large data set of molecules is to be screened in order to select
structures that have a higher likelihood of being inhibitors. Such molecules are often desired in order to enrich
in-house target-specific libraries of pharmaceutical companies. An empirically derived in silico method can also
help to set priorities within the list of accessible in-house molecules to be tested experimentally. Similar
approach can be used to enrich the initial set of patented molecules by including compounds from the large
commercial collections that are predicted to be active for the same protein target family.
ACKNOWLEDGMENTS
This work was supported by EC BioSapiens (LHSG-CT-2003-503265) and EC SEPSDA (SP22-CT-2004-
003831) 6FP projects as well as the Polish Ministry of Education and Science (PBZ-MNiI-2/1/2005 and
2P05A00130). MvG and LSW would like to thank the Foundation for Polish Science for the fellowship.
9

Table I. SVM classification results for selected 5 targets on launched and preclinical inhibitors from
MDDR database.
Protein Target Training dataset Testing dataset
COX2 predicted 0 1 predicted 0 1
10
Recall/
Precision
Recall/
Precision
Training set Testing set
observed 0 2106 0 0 8308 38 92% 73%
observed 1 6 69 1 10 27 100% 42%
DH 0 1 0 1
observed 0 2151 0 0 8390 9 100% 73%
observed 1 0 17 1 3 8 100% 47%
TH 0 1 0 1
observed 0 2032 0 0 8401 22 98% 74%
observed 1 1 47 1 9 26 100% 54%
RT 0 1 0 1
observed 0 2130 0 0 8277 43 100% 31%
observed 1 0 54 1 24 11 100% 20%
AE 0 1 0 1
observed 0 2351 2 0 9028 24 100% 42%
observed 1 0 22 1 7 5 92% 17%
The SVM classification performance on the set of preclinical or launched inhibitors from MDDR database is
described here using the confusion tables. Columns represent observed in experiments class of a compound for
each of targets (active/inactive) whereas rows represent the prediction results. The list of protein targets
include: cyclooxygenase-2 (COX2), dihydrofolate (DH), thrombin (TH), reverse transcriptase (RT) and
antiestrogen (AE). First we present results on training datasets with 2/3 available positives (preclinical or
launched inhibitors for the selected target) and 1/3 of negatives (randomly selected subset of preclinical or
launched inhibitors knowing not to inhibit selected target). On the right there are results of SVM models on
testing datasets containing 1/3 of positives and 2/3 of available negatives not used in training of SVM method.
The last two columns present the precision and recall values on both training and testing datasets.

Table II. The classification accuracy for SVM method on selected 5 targets for biologically tested
inhibitors of MDL drug data report.
Protein Target Number
of
positives/
negatives
cyclooxygenase-2
10408
dihydrofolate
10492
thrombin
10418
reverse transcriptase
10406
antiestrogen
12164
11
Number of
biological
testing
compounds
793
110 69%
95%
28 154 71%
98%
111 1070 62%
98%
113 596 52%
96%
34 255 40%
96%
Recall/
Precision
of the best
method
MDL drug data reports inhibitor and non-active compounds for selected five targets. SVM algorithm is trained
here on the whole set of preclinical or launched inhibitors for five protein targets. The classification models
were then tested then on biological tested on selected biological target according to MDDR compounds
(excluding those that are preclinical or already launched on the market).
The first column presents the protein target name. The second column gives numbers of positives and negatives
instances used in training. The third column shows the number of compounds annotated by MDL drug data
report for being biologically tested on selected protein target. The recall and precision values for the SVM
method are in the fourth column.

Table III. The classification accuracy for support vector machine method on selected 5 targets for
patented compounds from MDL drug data report.
Protein Target Number
of training
negatives/
positives
oldest 1/3
Cyclooxygenase-2
Number
of testing
negatives/
positives
newest 2/3
12
Recall/
Precision
on the
training
dataset
Recall/
Precision
on the
testing
dataset
2106 8347 94% 64%
31 81 100% 66%
dihydrofolate 2151 8390 100% 50%
12 16 100% 44%
thrombin 2034 8425 100% 33%
34 76 100% 66%
reverse 2130 8323 100% 46%
transcriptase 54 59 100% 33%
antiestrogen 2353 9700 100% 9%
6 22 100% 15%
The supervised SVM method classification performance trained using AP descriptors. The list of protein targets
include: cyclooxygenase-2 (COX2), dihydrofolate (DH), thrombin (TH), reverse transcriptase (RT) and
antiestrogen (AE). The oldest one-third (first four columns) or the oldest two-thirds (last four columns) of a
subset of MDL drug data report inhibitors and non-active compounds for those targets were used for training
supervised machine-learning algorithms. The classification models were tested then on rest of patented
compounds.
The first column presents the protein target name. The second column gives numbers of positives and negatives
instances used in training (1/3 oldest patented compounds). The third column shows the number of 2/3 newest
compounds annotated by MDL drug data report for inhibition on selected protein target and patented. The recall
and precision values for the SVM method on the training datasets is included in the fourth column. The fifth
column presents the precision and recall values on the testing dataset.

Table IV. The classification accuracy for support vector machine method on selected 5 targets for top
10% of best docked compounds.
Protein Target #positives in
MDDR db
FLOG
docking
results
Best Docking
Methods
Cyclooxygenase-2 98 1 GLIDE
FRED
13
Best Scoring
Methods
SS
ChemScore
dihydrofolate 27 6 ICM ICM 10
thrombin 99 46 ICM ICM 99
reverse transcriptase 108 0 ICM ICM 10
antiestrogen 32 17 GLIDE
FRED
GlideScore
ChemScore
Best
D&S
Results
60
63
Recall/
Precision
on the
training
dataset
93,20%
98,97%
64,74%
62,43%
72,64%
68,28%
79,29%
80,16%
74,00%
24
23 71,36%
The supervised SVM method classification performance trained using AP descriptors. The list of protein targets
include: cyclooxygenase-2 (COX2), dihydrofolate (DH), thrombin (TH), reverse transcriptase (RT) and
antiestrogen (AE). The subset of MDL drug data report inhibitors including both active and non-active
compounds for those targets were docked on protein targets using various docking methods. The 10% of best
docked compounds were then used for training supervised machine-learning algorithms. The classification
models were tested then on the rest of active compounds.
The first column presents the protein target name. The second column gives numbers of active compounds
found in MDL data drug report database for selected protein target. The number of negatives used for docking
experiment was fixed for all targets and equal to 10000. The third column shows the results of FLOG fast and
flexible docking procedure i.e. the number of active compounds found in first 10% of the ordered by the FLOG
docking score ligands. The fourth and fifth column present the best docking and scoring method name. The
sixth column presents the number of active compounds found in first 10% of the list of ligands ordered by the
best docking program followed by the scoring. The set of 10% of the best scoring compounds was then used to
train support vector machine (SVM). The recall and precision values for the training is presented in the seventh
column.

REFERENCES
1. MDL, MDL Drug Data Report (2004). Coverage: 1988-present; updated monthly. Focus: Drugs
launched or under development, as referenced in the patent literature, conference proceedings, and
other sources; descriptions of therapeutic action and biological activity; tracking of compounds
through development phases. Size: 132726 molecules,129459 models. Updates add approximately
10,000 new compounds per year. 2004.
2. Nidhi, et al., Prediction of biological targets for compounds using multiple-category Bayesian models
trained on chemogenomics databases. J Chem Inf Model, 2006. 46(3): p. 1124-33.
3. Fang, J., et al., Support vector machines in HTS data mining: Type I MetAPs inhibition study. J Biomol
Screen, 2006. 11(2): p. 138-44.
4. Briem, H. and J. Gunther, Classifying "kinase inhibitor-likeness" by using machine-learning methods.
Chembiochem, 2005. 6(3): p. 558-66.
5. Bender, A. and R.C. Glen, A discussion of measures of enrichment in virtual screening: comparing the
information content of descriptors with increasing levels of sophistication. J Chem Inf Model, 2005.
45(5): p. 1369-75.
6. Plewczynski, D., S.A. Spieser, and U. Koch, Assessing different classification methods for virtual
screening. J Chem Inf Model, 2006. 46(3): p. 1098-106.
7. Jenkins, J.L., R.Y. Kao, and R. Shapiro, Virtual screening to enrich hit lists from high-throughput
screening: a case study on small-molecule inhibitors of angiogenin. Proteins, 2003. 50(1): p. 81-93.
8. Kalgutkar, A.S. and Z. Zhao, Discovery and design of selective cyclooxygenase-2 inhibitors as nonulcerogenic,
anti-inflammatory drugs with potential utility as anti-cancer agents. Curr Drug Targets,
2001. 2(1): p. 79-106.
9. Dicker, A.P., et al., Identification and characterization of a mutation in the dihydrofolate reductase
gene from the methotrexate-resistant Chinese hamster ovary cell line Pro-3 MtxRIII. J Biol Chem,
1990. 265(14): p. 8317-21.
10. Schweitzer, B.I., A.P. Dicker, and J.R. Bertino, Dihydrofolate reductase as a therapeutic target. Faseb
J, 1990. 4(8): p. 2441-52.
11. Ambler, J., et al., The discovery of orally available thrombin inhibitors: studies towards the
optimisation of CGH1668. Bioorg Med Chem Lett, 1998. 8(24): p. 3583-8.
12. Menear, K., Progress towards the discovery of orally active thrombin inhibitors. Curr Med Chem, 1998.
5(6): p. 457-68.
13. Gustafsson, J.A., Therapeutic potential of selective estrogen receptor modulators. Curr Opin Chem
Biol, 1998. 2(4): p. 508-11.
14. Castro, H.C., et al., HIV-1 reverse transcriptase: a therapeutical target in the spotlight. Curr Med
Chem, 2006. 13(3): p. 313-24.
15. Sheridan, R.P., The centroid approximation for mixtures: calculating similarity and deriving structure-activity
relationships. J Chem Inf Comput Sci, 2000. 40(6): p. 1456-69.
16. Miller, M.D., R.P. Sheridan, and S.K. Kearsley, SQ: a program for rapidly producing
pharmacophorically relevent molecular superpositions. J Med Chem, 1999. 42(9): p. 1505-14.
17. Vapnik, V.N., The nature of statistical learning theory. Vol. xv. 1995, New York: Springer.
18. Vapnik, V.N., Statistical learning theory. Adaptive and learning systems for signal processing,
communications, and control. Vol. xxiv. 1998, New York: Wiley. 736.
19. Cristianini, N. and J. Shawe-Taylor, An introduction to support vector machines : and other kernelbased
learning methods. 2000, Cambridge, U.K. ; New York: Cambridge University Press. xiii, 189.
20. Joachims, T., Learning to classify text using support vector machines. Kluwer international series in
engineering and computer science SECS. Vol. xvi. 2002, Boston: Kluwer Academic Publishers. 205.
14

21. Vojtech, F. and H. Vaclay, Vojtech, F., Vaclay, H., An iterative algorithm learning the maximal margin
classifier. Pattern Recognition, 2003. 36(9): p. 1985-1996.
22. Kim, H. and H. Park, Protein secondary structure prediction based on an improved support vector
machines approach. Protein Eng, 2003. 16(8): p. 553-60.
23. Minakuchi, Y., K. Satou, and A. Konagaya. Prediction of protein-protein interaction sites using supprot
vector machnes. in International conference on mathematics and engineering techniques in medicine
and biological sciences. 2003.
24. Valentini, G., Gene expression data analysis of human lymphoma using support vector machines and
output coding ensembles. Artif Intell Med, 2002. 26(3): p. 281-304.
25. Guyon, I., et al., Gene selection for cancer classification using support vector machines. Mach. Learn.,
2002. 46: p. 389-422.
26. Brown, M.P., et al., Knowledge-based analysis of microarray gene expression data by using support
vector machines. Proc Natl Acad Sci U S A, 2000. 97(1): p. 262-7.
27. Furey, T.S., et al., Support vector machine classification and validation of cancer tissue samples using
microarray expression data. Bioinformatics, 2000. 16(10): p. 906-14.
28. Krishnan, V.G. and D.R. Westhead, A comparative study of machine-learning methods to predict the
effects of single nucleotide polymorphisms on protein function. Bioinformatics, 2003. 19(17): p. 2199-
209.
29. Pavlidis, P., et al. Gene functional classification from heterogeneous data. in 5th International
Conference on Computational Molecular Biology. 2001. Montreal, Canada: ACM Press.
30. Zien, A., et al., Engineering support vector machine kernels that recognize translation initiation sites.
Bioinformatics, 2000. 16(9): p. 799-807.
31. Jaakkola, T., M. Diekhans, and D. Haussler, A discriminative framework for detecting remote protein
homologies. J Comput Biol, 2000. 7(1-2): p. 95-114.
32. Hua, S. and Z. Sun, A novel method of protein secondary structure prediction with high segment
overlap measure: support vector machine approach. J Mol Biol, 2001. 308(2): p. 397-407.
33. Ding, C.H. and I. Dubchak, Multi-class protein fold recognition using support vector machines and
neural networks. Bioinformatics, 2001. 17(4): p. 349-58.
34. Schölkopf, B., C.J.C. Burges, and A.J. Smola, Advances in kernel methods : support vector learning.
Vol. vii. 1999, Cambridge, Mass.: MIT Press. 376.
35. Zavaljevski, N., F.J. Stevens, and J. Reifman, Support vector machines with selective kernel scaling for
protein classification and identification of key amino acid positions. Bioinformatics, 2002. 18(5): p.
689-96.
36. Burbidge, R., et al., Drug design by machine learning: support vector machines for pharmaceutical
data analysis. Comput Chem, 2001. 26(1): p. 5-14.
37. Miller, M.D., et al., FLOG: a system to select 'quasi-flexible' ligands complementary to a receptor of
known three-dimensional structure. J Comput Aided Mol Des, 1994. 8(2): p. 153-74.
38. Friesner, R.A., et al., Glide: a new approach for rapid, accurate docking and scoring. 1. Method and
assessment of docking accuracy. J Med Chem, 2004. 47(7): p. 1739-49.
39. Halgren, T.A., et al., Glide: a new approach for rapid, accurate docking and scoring. 2. Enrichment
factors in database screening. J Med Chem, 2004. 47(7): p. 1750-9.
40. Miteva, M.A., et al., Fast structure-based virtual ligand screening combining FRED, DOCK, and
Surflex. J Med Chem, 2005. 48(19): p. 6012-22.
41. Ewing, T.J., et al., DOCK 4.0: search strategies for automated molecular docking of flexible molecule
databases. J Comput Aided Mol Des, 2001. 15(5): p. 411-28.
42. Buzko, O.V., A.C. Bishop, and K.M. Shokat, Modified AutoDock for accurate docking of protein kinase
inhibitors. J Comput Aided Mol Des, 2002. 16(2): p. 113-27.
43. Vaque, M., et al., BDT: an easy-to-use front-end application for automation of massive docking tasks
and complex docking strategies with AutoDock. Bioinformatics, 2006.
15

44. Fernandez-Recio, J., M. Totrov, and R. Abagyan, ICM-DISCO docking by global energy optimization
with fully flexible side-chains. Proteins, 2003. 52(1): p. 113-7.
45. Byvatov, E., et al., Comparison of support vector machine and artificial neural network systems for
drug/nondrug classification. J Chem Inf Comput Sci, 2003. 43(6): p. 1882-9.
46. Glick, M., et al., Enrichment of high-throughput screening data with increasing levels of noise using
support vector machines, recursive partitioning, and laplacian-modified naive bayesian classifiers. J
Chem Inf Model, 2006. 46(1): p. 193-200.
16