Genome Informatics 17(2): 239{247 (2006) 239

A Large-Scale Computational Approach to Drug

Repositioning

Yvonne Y. Li Jianghong An

yli@bcgsc.ca jan@bcgsc.ca

Steven J.M. Jones

sjones@bcgsc.ca

Canada's Michael Smith Genome Sciences Centre, 570 West 7th Avenue, Vancouver,British Columbia, V5Z 4S6, Canada

Abstract

We have developed a computational pipeline for the prediction of protein-small molecule in-teractions and have applied it to the drug repositioning problem through a large-scale analysis ofknown drug targets and small molecule drugs. Our pipeline combines forward and inverse docking,the latter of which is a twist on the conventional docking procedure used in drug discovery: insteadof docking many compounds against a speci�c target to look for potential inhibitors, one compoundis docked against many proteins to search for potential targets. We collected an extensive set of1,055 approved small molecule drugs and 1,548 drug target binding pockets (representing 78 uniquehuman protein therapeutic targets) and performed a large-scale docking using ICM software to bothvalidate our method and predict novel protein-drug interactions. For the 37 known protein-druginteractions in our data set that have a known structure complex, all docked conformations werewithin 2.0�A of the solved conformation, and 30 of these had a docking score passing the typicalICM score threshold. Out of the 237 known protein-drug interactions annotated by DrugBank,74 passed the score threshold, and 52 showed the drug docking to another protein with a betterdocking score than to its known target. These protein targets are implicated in human diseases, sonovel protein-drug interactions discovered represent potential novel indications for the drugs. Ourresults highlight the promising nature of the inverse docking method for identifying potential noveltherapeutic uses for existing drugs.

Keywords: drug repositioning, computational drug discovery, inverse docking, docking

1 Introduction

An integral part of recent computational drug discovery research focuses on the high-throughputscreening of chemical databases to �nd inhibitors of speci�c protein targets. However, drugs designedto interact speci�cally with a certain target frequently also interact with other proteins. These o�-target interactions may cause harmful drug side e�ects; however, they may also lead to new therapeuticopportunities. Understanding the potential o�-target interactions of existing drugs is of major interestto pharmaceutical research, not only for providing insight into drug side e�ects, but also for discoveringnovel therapeutic uses of drugs. Finding new indications for existing drugs, also known as drugrepositioning, represents an e�cient approach to drug discovery, since existing drugs already haveclinical history and thus require much less time and money to develop into a drug speci�c for the newdisease [3].

Molecular docking is a computational method used to predict how a ligand molecule interacts witha protein binding site, both in terms of binding conformation and binding a�nity. This method iswell established as a virtual screening method in drug discovery [13], where typically a large com-pound database is docked against a speci�c protein binding site, in order to discover novel inhibitors

240 Li et al.

of that target. The inverse scenario, whereby a speci�c small molecule is docked to a panel of pro-tein targets - so called `inverse docking' - has been little studied. This method was �rst used byChen and Zhi [9] in 2001 to predict potential o�-target protein-drug interactions. Since then, a fewstudies have used it for predicting unanticipated kinase targets of several kinase inhibitors [17]. Thecombination of forward and inverse docking has been used in cross-docking studies for investigatingenzyme-metabolite selectivity [15] and used in creating protein-ligand docking a�nity matrices forimproving the enrichment factor of virtual screening [11].

However, inverse docking is also suited to computational drug repositioning analysis. By dockingan existing drug to a panel of known therapeutic targets, the potential of the drug to bind targetsother than its intended can be assessed, thus leading to potential drug therapies for other diseases.Such a method would be particularly valuable for improving the e�ciency of current drug discoverypipelines. Here, we present a computational approach that combines forward and inverse docking forthe drug repositioning problem.

2 Method and Results

2.1 Molecular Docking Pipeline

The automated molecular docking pipeline was implemented using the ICM software package [1] andis illustrated in Figure 1.

���������

��������

�������

��������������������

�����

������

�������

�����������������

������

����������

������

���������������

������

������

�����

������

��������������

���������������

��!�

�����������������

������

"������

Figure 1: The molecular docking pipeline.

Given a list of Protein Data Bank (PDB) identi�ers, the protein 3D structures are retrieved fromPDB and are �rst required to be X-ray crystal structures with a minimum resolution of 2.5�A. Thestructures are then separated into chains as annotated in PDB. For structures with multiple chains, itschains are grouped into a set of non-redundant sequences, based on PDB's chain redundancy analysisat the 95% sequence identity level [5].

The PDB structures are prepared for docking by adding and optimizing hydrogens and by removingwater molecules, metal ions, and other solvent molecules. Pockets are predicted in the structures usingthe PocketFinder algorithm, which is based on a transformation of the van der Waals potential [2]. Thisalgorithm has been shown to have excellent predictive capability for human pockets: when applied toa large and comprehensive data set of 17,126 human protein pockets, PocketFinder correctly predicted96.8% of holo pockets and 95% of apo pockets [2].

Predicted pockets are then further �ltered for quality. First, they are de�ned at a cut-o� ofminimum 150 �A3 volume. Second, when multiple pockets are predicted in a protein, the two largest

A Large-Scale Computational Approach to Drug Repositioning 241

are retained. Though this will inevitably omit real druggable sites, the PocketFinder method hasshown that its two largest predicted pockets may cover as many as 92.7% of real binding sites [2].Third, pockets that have residues within 3.5�A with occupancy values less than 1 or temperature factors(B-factors) greater than 60 are removed - these are non-stringent cut-o�s used only to detect the mostunreliable residues. Finally, the protein's observed PDB sequence is aligned to its known SwissProtsequence and pockets that have residues within 3.5�A which are engineered mutations or missing inthe observed sequence are removed. An example of predicted pockets for the Hemoglobin alpha chainprotein is shown in Figure 2. For each pocket, the receptor is de�ned as a box with a 3.5�A marginaround the outermost points of the pocket.

The small molecule database is prepared as a multi-object sdf �le and docked to each pocket. ICMperforms grid-based rigid-receptor exible-ligand docking through a modi�ed Monte-Carlo searchingprocedure and a rigorous empirical scoring function. ICM has performed well in recent assessmentsof popular docking packages (including, Glide, GOLD, Autodock, DOCK) [6, 8], and has shown top-ranking pose prediction accuracies of 45% [16] and 50% [8] in studies using small protein benchmarksets. After docking, the scores and conformations of high scoring compounds are gathered for furtheranalysis. For proteins with multiple pockets, the docking score of the best scoring pocket is used torepresent the protein-small molecule interaction. These time-intensive pocket prediction, preparation,and docking steps are rendered feasible by using a 400-processor cluster, on which we have ICMlicensed for 200 processors.

2.2 Data Set: Binding Pocket Database and Small Molecule Drug Database

We drew known protein-drug interactions from DrugBank [18], a manually curated database holdingdetailed information on drugs and protein targets. DrugBank contains therapeutically useful interac-tions for over 4,000 drugs, including FDA-approved, experimental, biotech, and nutriceutical drugs,corresponding to over 6,000 drug targets. For our analysis, we �rst focused on the 1,055 FDA approveddrugs and their corresponding 510 human proteins targets.

447 SwissProt identi�ers of therapeutic targets corresponding to 1,055 known small molecule drugswere retrieved from DrugBank. 1,345 X-ray PDB structures with resolution better than 2.5�A wereobtained for 165 of these proteins for which 246 known drugs corresponded. This resulted in 1,836 non-redundant PDB chains with at least one chain per structure. PocketFinder predicted 2,272 pocketswhen the two largest pockets over 150�A3 were retained. After the B-factor and occupancy �lteringstep, 2,082 pockets remained. After removing pockets with nearby engineered mutations or gaps,1,548 pockets remained, representing 78 unique human proteins.

The 1,055 approved small molecule drugs with sdf representation were downloaded from DrugBank.Known protein-drug interactions used to assess docking results were obtained from DrugBank's Ap-proved DrugCards (�le dated 2005/6/27). 237 of them are represented by proteins and drugs in ourdata set.

2.3 Re-Docking Known Protein-Drug Complexes

Of the 237 known protein-drug interactions in the data set, 37 interactions (representing 19 uniqueproteins) have already been solved in a PDB structure complex. We �rst veri�ed whether theseinteractions could be predicted using our pipeline. The advantage of this benchmark data set is thatboth the binding site and binding conformations for the ligand are known. We docked cognate ligandsto predicted protein pockets for these 37 interactions. For this smaller data set, we performed a morerigorous pocket prediction in that pockets containing metal ions were prepared with the metal ionpresent. In addition, for multi-chain proteins, we also predicted inter-chain pockets (one example isshown in Figure 2). There were 101 pockets in total.

The results are summarized in Table 1, listing for each interaction the docking score and the rootmean square deviation (RMSD) value between the docked ligand and the PDB conformation. All 37

242 Li et al.

Figure 2: Left: The two pockets predicted for PDB entry 1R1X (Hemoglobin alpha chain) overlapwell with known ligand positions. Right: The binding pocket (darker glob) of PDB entry 1TCO iscomprised of three separately colored chains.

complexes exhibit a RMSD value less than 2.0�A which is the standard cuto� used to de�ne a successfulmolecular docking. 31 of the 37 complexes exhibit an RMSD value less than 1.0�A which is a much morerigorous measure of successful docking [8]. A couple of docked examples in Figure 3 illustrate that evenlarge exible ligands are docked correctly. The low RMSD values also show that despite retaining twopockets for each protein, the best scoring ligand conformation always docked to the correct pocket.However, though the RMSDs were all well predicted, docking scores in only 29 of the 37 complexeswere smaller (better) than the default ICM score cuto� (which is -32). Further investigation showsthe poorly scoring complexes all involve pockets with a metal ion present. It should be noted thatwe also tried docking without the metal in the pocket, and though the scores were more reasonable(around -25), the RMSDs were greater than 2.0�A. Thus it seems, at the moment, we are not able topredict reliable scores for pockets containing metals.

Overall, these results show that the docking and scoring algorithms are quite accurate predictorsof actual protein-drug binding when the protein is already in a correct conformation.

Table 1: Docking score and RMSD values for the re-docking of the 37 structure complexes.

PDB PDB RMSD Dock PDB PDB RMSD Dock PDB PDB RMSD Dockprotein ligand (�A) score protein ligand (�A) score protein ligand (�A) score1K74 REA 0.8 -71.0 1BRP RTL 0.5 -48.7 1UHO VDN 1.3 -35.7

1FM6 REA 0.4 -70.1 3ERD DES 0.3 -48.4 1FKJ FK5 0.8 -35.5

1FM9 REA 0.5 -69.7 1SQN NDR 0.4 -45.9 1BKF FK5 0.9 -35.1

1FBY REA 0.6 -67.9 1HWI MMM2 0.5 -45.7 1MMK BH4 0.4 -33.1

1XZX T3 0.5 -60.6 1RBP RTL 0.4 -44.0 1KW0 BH4 0.5 -29.8

1HWK MMM2 0.7 -58.8 1Z95 MMMM 0.4 -42.8 1XOZ CIA 0.7 -25.1

2LBD REA 0.4 -58.0 1XP0 VDN 1.3 -39.6 1XLX CIO 1.5 -19.4

1T46 STI 0.4 -56.4 1FKB RAP 0.7 -39.6 1UZF AMCO 0.5 -14.9

3LBD REA 0.4 -55.9 1NB9 RBF 0.8 -39.0 1J8U THB 1.7 -12.5

1HWJ MMM2 0.4 -54.7 1TBF VIA 0.8 -38.9 1CIL ETS 0.6 -9.9

1TCO FK5 0.8 -52.9 1FKF FK5 0.7 -36.7 1A42 BZO 1.7 -7.7

1A28 ASTR 0.3 -50.4 1HWL FBI 0.5 -36.0 1AZM AZM 1.4 -7.7

1M2Z DEX 0.4 -49.4

A Large-Scale Computational Approach to Drug Repositioning 243

Figure 3: Left: 1HWK (HMDH HUMAN) is shown with complexed ligand Atorvastatin and its dockedconformation. Right: The ligand Alitretinoin of 2LBD (RARG2 HUMAN) is shown in its predictedpocket along with the docked conformation.

2.4 Large-Scale Docking

A large scale docking using ICM was carried out for the 1,055 approved drugs against the 1,548pockets. Proteins represented by multiple PDB structures (i.e. multiple pockets) use the best scoringdrug-pocket interaction as the protein-drug docking score. Results are illustrated in Figure 4 with the1,055 drugs along the horizontal and 31 of the 78 proteins along the vertical.

Though ICM recommends a docking score cut-o� of -32, this value should ideally be tailored toeach receptor. For example, a study using ICM docking to EGFR discovered novel inhibitors with acut-o� value of -28 [7]. We chose to apply a simple receptor-speci�c threshold system. For proteinsbelonging to the 37 known complexes in the re-docking analysis, we used those docking scores ascut-o�s. If protein is known to bind several drugs, as is often the case in our data set, the cut-o�was the poorest known-drug-score (but still better than the default -32). These cut-o�s are the mostreliable since it is known that the drug docked to the correct binding site with a correct conformation.For proteins that are not in a known structure complex but showed a good docking score (better than-32) with a known interacting drug, we used this docking score as the cut-o�. These cut-o�s are lessreliable since without a 3D structure, we cannot be absolutely sure that the drug docked to the correctprotein pocket with a correct conformation. The remaining 47 proteins of the 78 did not have anyknown interactions predicted with a good score and are not shown in Figure 4.

Of the 237 known protein-drug interactions, 74 show a docking score better than -32. In such cases,it is interesting to examine how a protein ranks when the drug is docked to all proteins (`ProteinRank'or inverse docking rank) as well as how the drug ranks when all drugs are docked to the protein(`DrugRank' or docking rank). Thus, ProteinRank ranges from 1 to 78 whereas DrugRank rangesfrom 1 to 1,055. The ranks for a few of the 74 good-scoring known interactions are shown in Table 2,illustrating several di�erent rank combinations.

When both ranks are high, the protein and drug are speci�c to each other. A ProteinRank that isnot 1 suggests that the drug may bind to another target as well as the primary target. A DrugRankthat is not 1 suggests that other drugs may have potential to bind this protein. In total, 52 of the74 interactions do not have a ProteinRank of 1, and are thus interesting for further investigation.

244 Li et al.

Figure 4: Map of large-scale docking results, with speci�c scoring thresholds applied for each protein.The 1,055 drugs are represented on the horizontal and the 31 proteins having a good docking scorewith a known drug are on the vertical. Protein-drug interactions whose docking score does not surpasstheir designated cut-o� are shown as black and those that do are shown in gray (where brighter shadesof gray represent better scores).

For example, in the case of Acitretin and Retinoic acid receptor gamma-1 (RARG1 HUMAN) fortreatment of severe adult psoriasis, some top targets for Acitretin are shown in Table 3. DrugBankannotated 7 targets, 2 of which do not have a solved human structure, leaving 5 targets in our inputdata set. All �ve targets (shaded rows) show good docking scores with this drug, despite that thereare no solved structures of this drug in complex with any protein. The top ranking target is serumalbumin, and though it was not annotated in DrugBank as a drug target, it was noted that in thesystemic circulation, over 98% of Acitretin is bound to lipoproteins or albumin in the blood [10]. Theremaining targets are novel potential targets of this drug, and should be the �rst to be experimentallytested for a laboratory interested in Acitretin repositioning. In Table 2, it can be seen that Acitretinis the 9th ranked drug for RARG1 HUMAN. There are thus 8 other known drugs that have potentialto be repositioned for severe psoriasis.

Table 2: The ProteinRank and DrugRank for several known protein-drug interactions. Proteins aredenoted by their SwissProt identi�ers.

Protein Drug ICM Score Protein Rank Drug Rank

RARG1 HUMAN Alitretinoin -59.1 1 1KIT HUMAN Imatinib -49.0 1 1PRGR HUMAN Medroxyprogesterone -47.8 1 14RETBP HUMAN Vitamin A -52.6 3 1RARG1 HUMAN Acitretin -46.3 2 9

When both ranks are 1, there is no drug in the database that docks better to the protein target,and there is no other protein to which the drug docks better. This means that the protein-drugcombination is ideal within the scope of the data set, and no better target can be found for the drug.This situation may still be interesting for drug repositioning in two scenarios: �rst, if more proteinsare included in the data set, the rankings may change; second, ICM docking does not purport to rankresults by binding a�nity - thus, docking scores lower than a known interaction (though they shouldstill be close to or better than the default -32 score) may also re ect potential binding interactions.

For example, imatinib is a well-known inhibitor of the Bcr-Abl fusion protein for the treatment ofchronic myeloid leukemia. In our docking analysis, the ABL-imatinib interaction was docked with ascore of -31.6, not passing the score threshold. However, closer inspection of the human ABL structure,

A Large-Scale Computational Approach to Drug Repositioning 245

Table 3: Inverse docking results for Acitretin. Proteins are denoted by their SwissProt identi�ers.Known targets are shaded. The remaining proteins are candidates for further drug repositioninganalysis.

Rank Protein Docking score to Acitretin

1 ALBU HUMAN -482 RARG1 HUMAN -463 DYR HUMAN -464 PTN1 HUMAN -455 RARA HUMAN -426 NOS3 HUMAN -427 HBB HUMAN -408 GSHR HUMAN -399 RXRA HUMAN -3910 THB1 HUMAN -3916 FA10 HUMAN -3617 RARG2 HUMAN -3618 RARB HUMAN -35

1OPL, shows that it is bound to a di�erent inhibitor, and is in a conformation not ideal for bindingimatinib due to steric clashes. We then added mouse ABL PDB structures (with 98% sequence identityto human ABL) into our pipeline, and they showed an ABL-imatinib interaction score of -58.4. This ishigher than the KIT-imatinib interaction in our current analysis and thus mouse-ABL would becomethe top ranking target for imatinib instead of KIT. Imatinib is known to inhibit c-KIT, a proteinkinase linked to gastrointestinal stromal tumor, and was shown to be e�ective for this cancer [12].Thus, this KIT-imatinib interaction was correctly predicted in our docking analysis.

3 Discussion & Conclusion

Conceptually, inverse docking models a more realistic biological environment - once a drug enters thebody or speci�c cells, which of the multitude of di�erent proteins will it bind to? Inverse dockingmethods have potential in many important areas of drug discovery, such as for predicting the un-wanted side e�ects for known and candidate drugs, determining druggable protein targets, and as weparticularly address in this study, predicting potential bene�cial therapeutic uses for known drugs. Toour knowledge, this represents the �rst time that inverse docking has been used for drug repositioninganalysis. Given the tremendous time and cost of drug discovery today, last estimated at 12 yearsand $900 million US to discover and develop one new drug [14], inverse docking may allow for moree�cient drug discovery pipelines to be implemented. However, there are still many challenges to beovercome for this type of analysis, which have been especially noted during our study.

Important implications arise from the limitations of current molecular docking methods. First,docking is heavily dependent on the input protein structure. A strong point of this study was tocarefully �lter for the quality of pockets prior to docking. For example, poorly resolved structurescannot be reliably used in docking as the protein atom positions are already imprecise. Similarly,pockets having nearby residues with low occupancy values or high temperature factor values are alsounreliable for docking. Pockets with nearby engineered mutations are not representative of wild-typebinding pockets, and pockets with gaps in their nearby observed sequence may be missing key residues.

Second, docking methods today do not easily handle protein exibility, presence of water and metalions, and accurate scoring. Protein exibility is partially compensated for when multiple structures

246 Li et al.

have been solved for the same protein in PDB. However, this a very small corner of protein conforma-tional space, so it is expected that the predictive ability of inverse docking will be much improved withthe development of scalable exible-receptor exible-ligand docking methods. For protein-drug com-plexes with metals, our docking accurately predicted drug binding conformations but failed to predictreasonable scores. We have tried to minimize scoring issues by only further analyzing proteins thatdocked to known drugs with good scores, since these proteins are more likely to be in conformationsamenable for small molecule inhibition. These proteins are also more likely to be scored appropriately.We are currently evaluating di�erent scoring methods to better represent metal ions. In the future,we plan to explore exible-protein exible-ligand docking with ICM and with other docking software.

One major challenge in inverse docking is comparing docking scores across di�erent pockets, asthe score threshold should ideally be speci�c to each receptor. In a best case scenario, a large set ofknown inhibitors with known binding a�nities would be docked to each protein pocket, to determinean appropriate docking score cut-o� for expected small molecule binders. Such a task, however, wouldrequire an immense amount of literature search, and is often impossible as many protein pockets havenot yet been characterized using diverse ligands. We have used a simple initial approximation by usingthe docking scores of known protein-drug complexes from PDB and interactions from DrugBank as areference point for our receptor-speci�c docking score cut-o�s. In addition, we limit our investigationof novel protein-drug interactions to only the dataset where known protein-drug interactions couldbe reproduced. We are currently gathering more binding a�nity information from the literature tobetter tailor the docking score cut-o�s.

Finally, our study is limited by the number of known proteins and drugs in the input data set.Having only 78 unique protein targets limits the utility of ProteinRank, as some of top-ranking targetsare likely to change when more proteins are added to the data set. By using only human proteinstructures in our analysis, many informative structures were omitted - for example, the mouse ABLstructures. Thus, we are currently incorporating structures from other organisms aside from human,when their sequence identity to the human protein is above 95%. In addition, we are including knownprotein targets for experimental drugs as annotated in DrugBank - for example the LCK protein whichis a therapeutic target critical in T-cell receptor signaling [4] - but no drugs have yet been approvedfor it.

Despite the limitations discussed, inverse docking opens a new avenue for computational drug dis-covery, with the potential to discover more about both adverse e�ects and novel therapeutic targetsof existing drugs. Our analysis has begun to address some limitations by applying stringent pocketquality �lters, requiring targets to have a well-docked known drug before further analysis, and intro-ducing receptor-speci�c docking score cut-o�s. By docking 1,055 known drugs to 1,548 pockets of 78proteins we have found 52 interactions with potential pharmacological relevance, thus showing theability of the inverse docking method as a tool for drug repositioning. The analysis has also pointedout many limitations of a large scale inverse docking analysis, which we are currently addressing, andwe anticipate that an improved pipeline will be able to better predict known and novel protein-druginteractions. It is di�cult to measure the accuracy of inverse docking, as previously unknown inter-actions may be false positives or novel predictions. However, the space of experimentally testing allprotein-drug combinations is vast, and large-scale inverse docking provides a computational �lter toselect the most likely interactions to experimentally test.

References

[1] Abagyan, R., Totrov, M., and Kuznetsov, D., ICM: A new method for protein modeling anddesign: Applications to docking and structure prediction from the distorted native conformation,J. Comput. Chem., 15:488{506, 1994.

A Large-Scale Computational Approach to Drug Repositioning 247

[2] An, J., Totrov, M., and Abagyan, R., Pocketome via comprehensive identi�cation and classi�ca-tion of ligand binding envelopes, Mol. Cell Proteomics, 4(6):752{761, 2005.

[3] Ashburn, T.T. and Thor, K.B., Drug repositioning: Identifying and developing new uses forexisting drugs, Nat. Rev. Drug Discov., 3(8):673{683, 2004.

[4] Bolen, J.B. and Brugge, J.S., Leukocyte protein tyrosine kinases: Potential targets for drugdiscovery, Annu. Rev. Immunol, 15:371{404, 1997.

[5] Bourne, P.E., Addess, K.J., Bluhm, W.F., Chen, L., Deshpande, N., Feng, Z., Fleri, W., Green,R., Merino-Ott, J.C., Townsend-Merino, W., Weissig, H., Westbrook, J., and Berman, H.M., Thedistribution and query systems of the RCSB Protein Data Bank, Nucleic Acids Res., 32:D223{D225, 2004.

[6] Bursulaya, B.D., Totrov, M., Abagyan, R., and Brooks, C.L. 3rd., Comparative study of severalalgorithms for exible ligand docking, J. Comput. Aided Mol. Des., 17(11):755{763, 2003.

[7] Cavasotto, C.N., Ortiz, M.A., Abagyan, R.A., Piedra�ta, F.J., In silico identi�cation of novelEGFR inhibitors with antiproliferative activity against cancer cells, Bioorg. Med. Chem. Lett.,16(7):1969{1974, 2006.

[8] Chen, H., Lyne, P.D., Giordanetto, F., Lovell, T., and Li, J., On evaluating molecular-dockingmethods for pose prediction and enrichment factors, J. Chem. Inf. Model., 46(1):401{415, 2006.

[9] Chen, Y.Z. and Zhi, D.G., Ligand-protein inverse docking and its potential use in the computersearch of protein targets of a small molecule, Proteins, 43(2):217{226, 2001.

[10] Chien, D.S., Sandri, R.B., and Tang-Liu, D.S., Systemic pharmacokinetics of acitretin, etreti-nate, isotretinoin, and acetylenic retinoids in guinea pigs and obese rats, Drug Metab. Dispos.,20(2):211{217, 1992.

[11] Frantz, S., Drug discovery: Playing dirty, Nature, 437(7061):942{943, 2005.

[12] Fukunishi, Y., Mikami, Y., Kubota, S., and Nakamura, H., Multiple target screening method forrobust and accurate in silico ligand screening, J. Mol. Graph. Model., 25(1):61{70, 2006.

[13] Jorgensen, W.L., The many roles of computation in drug discovery, Science, 303(5665):1813{1818.2004.

[14] Kola, I. and Landis, J., Can the pharmaceutical industry reduce attrition rates, Nat. Rev. DrugDiscov., 3(8):711{715, 2004.

[15] Macchiarulo, A., Nobeli, I., and Thornton, J.M., Ligand selectivity and competition betweenenzymes in silico, Nat. Biotechnol., 22(8):1039{1045, 2004.

[16] Perola, E., Walters, W.P., and Charifson, P.S., A detailed comparison of current docking andscoring methods on systems of pharmaceutical relevance, Proteins, 56(2):235{249, 2004.

[17] Rockey, W.M. and Elcock, A.H., Progress toward virtual screening for drug side e�ects, Proteins,48(4):664{671, 2005.

[18] Wishart, D.S., Knox, C., Guo, A.C., Shrivastava, S., Hassanali, M., Stothard, P., Chang, Z., andWoolsey, J., DrugBank: A comprehensive resource for in silico drug discovery and exploration,Nucleic Acids Res., 34:D668{D672, 2006.