DOI: 10.1002/minf.201000068

An Improved Weighted-Residue Profile Based Method ofUsing Protein–Ligand Interaction Information in IncreasingHits Selection from Virtual Screening: A Study on VirtualScreening of Human GPCR A2A Receptor AntagonistsMohammad Shahid,[a] Vinod Kasam,[a] and Martin Hofmann-Apitius *[a, b]

1 Introduction

Virtual screening has emerged as an important tool thatplays a critical role in the discovery of novel lead com-pounds. Among the various methodologies in this area, theuse of structure based design methods is increasing rapidly.Having experimentally determined the protein structure ofa potential drug target, virtual screening by moleculardocking, in combination with many other methodologies,has become an established method in lead identificationand optimization.[1] In the molecular docking approach, thedocking tools are used to predict the binding modes of thesmall molecule that is docked to a 3D structure of a recep-tor and to score the potentially best fitting molecule in theactive site pocket of the receptor.[2] However, the dockingprograms are generally more successful in predicting goodbinding modes as compared to correctly scoring the poten-tially good binders due to the limitations in their built-inscoring schemes.[3, 4] These limitations and challenges in thescoring functions and their failures in efficiently scoring thenear-native binding poses have been reported from time totime.[5,6] Various alternate strategies and methodologieshave been suggested with strong emphasis on post-dock-ing strategies.[7] Nevertheless, the docking programs arestill used as attractive tools for virtual compound screeningand their coupling with other tools such as molecular dy-namics simulations[8] or re-scoring by using MM-PBSA/GBSAmethods have resulted in further improvements in achiev-ing the goals.[9] Similarly, many other approaches such asutilizing structural interaction information in the com-

pounds screening process are gaining importance.[10, 11]

Such approaches have proved very useful in the improve-ment of efficiency of the in silico predictions.[7] Amongthese approaches, protein-ligand interaction fingerprintshave shown to enrich the screening results by rescoringthe docking poses to facilitate the filtering process.[11–14]

The main focus of this paper is to demonstrate an easy andmore efficient way of post-processing the virtual screeningresults with achieving gain in accuracy.

1.1 Protein-Ligand Interaction Fingerprints

Protein-ligand interaction fingerprints (PLIFs) are lineararrays of bit strings, which encode the presence or absenceof a particular interaction between the protein and ligandatoms in a protein-ligand complex, which can be used insimilarity search. Several studies have utilized the PLIFs in-formation in optimizing and improvement of virtual screen-

[a] M. Shahid, V. Kasam, M. Hofmann-ApitiusDepartment of Bioinformatics, Fraunhofer Institute for Algorithmsand Scientific Computing SCAISchloss Birlinghoven, 53754 Sankt Augustin, Germany*e-mail : martin.hofmann-apitius@scai.fraunhofer.de

[b] M. Hofmann-ApitiusBonn-Aachen International Center for Information Technology(B-IT), University of BonnDahlmannstrasse 2, 53113 Bonn, Germany

Abstract : The use of protein-ligand interaction informationhas been reported to improve and optimize the docking re-sults in virtual screening experiments. Here we propose animproved weighted-residue profile based method to profilethe protein-ligand interactions based on the available data-set of known actives and utilize this weighted residue pro-file information, together with the scoring function, as se-lection criteria to increase hit rates in virtual screening ex-periments. The generated fingerprint data is not directly

based on the protein-ligand complexes but taken from theavailable interaction data derived from the docking results.The ability of the method to recover the active compoundswas tested on two data sets of a compound library de-signed for antagonists of the A2A receptor. The resultsshow better hits enrichments by using the weighted-resi-due based profiles of protein-ligand interactions as com-pared to the normal binding energy based scoring schemesof the two docking programs.

Keywords: GPCR · Adenosine receptor A2A antagonists · Protein-ligand interaction fingerprints · Docking · Virtual screening

Mol. Inf. 2010, 29, 781 – 791 � 2010 Wiley-VCH Verlag GmbH & Co. KGaA, Weinheim 781

ing results by post-processing the docking poses.[15, 16] Themain objective of utilizing the PLIF information in screeninga large database of compounds is to speed up the filteringprocess of biologically active compounds by reducing theburden of visually analysing a large number of protein-ligand complexes to select the compounds with goodbinding modes. The filtering process becomes easy to sortout the compounds by identifying the key interactions ofthe compounds with receptor residues.

Here we represent an easy and improved method ofusing the interaction information in a residue profile basedon knowledge of the receptor interactions with knownactive compounds. Weights are assigned to key interactionsin order to generate a weighted-residue profile; specific in-teractions, which are more frequently present or absent inthe set of active compounds, are represented in this profile.This approach is different from that shown in recent stud-ies[7, 12] where interactions are calculated from the protein-ligand complexes by applying geometric rules to determineinteraction fingerprints and use similarity measures to findoptimal interactions similar to that of a reference com-pound. Contrary to this, we use a simple weighted binaryfeature vector to identify a residue interaction that is eitherhaving a close contact interaction or hydrogen bonding in-teraction with the ligand. The weighted-residue profile isalso much simpler and very easy to calculate optimallyfrom the available data set of known active compounds ascompared to use complex functions to calculate weightedPLIF profiles as shown in the literature.[13, 14] The weightvalues are simply added up to the final energy basedscores already calculated by the docking programs for apredicted docking pose resulting in a so-called Z-score.Which means that the built-in scoring function is not com-pletely ignored but rather the individual key interactions’weight scores from the profile are used as an additionalpiece of information to rank the compounds. Moreover, thekey interactions in the residue profile can also be used as areference profile to which the similarity of other com-pounds in the screening library can be calculated as a Tani-moto-coefficient (TC).

Improvements to other approaches are that we do notonly consider the structural interactions but also take thesignificance of the interactions into account. Another im-provement in this method is that the residue profiles arecreated in a simpler way by the knowledge derived fromthe interactions of a set of known active compounds withthe receptor. This approach, which is based on a set ofmany active molecules, helps to solve the problem of se-lecting any single molecule as a best reference in the stud-ies of protein–ligand interactions. Moreover, different pro-files can be used for different docking tools in use whichcan be interactively optimized and further improved basedon the knowledge of the interactions of the known activecompounds with the receptor under study.

2 Methods and Materials

2.1 Preparation of the Residue Profile

An initial list of the residues in the active site pocket of thetarget receptor involved in their interaction with the boundligand is prepared. Such interaction information can beeasily obtained either from the available structural data ofthe X-ray structure (cocrystallized protein-ligand complex),LIGPLOT,[17] PDBSum[18] or any other tool that can determinethe residues that are involved in any kind of interactionwith the ligand in the bound state. Once the key residuesthat are involved in the interactions are identified, a simpleresidue profile is created from these interactions as shownin Table 1.

Residues in the profile are listed with their interactiontypes and their weight values. A residue profile is just aplain text file containing the residue names and the interac-tion types. As shown in Table 1, the weight values areequal in the initial profile, and are not taken into consider-ation at this moment.

2.2 Generation of Protein-Ligand Interaction Fingerprints

Protein-Ligand Interaction Fingerprints are generated fromthe interaction information of the top best docking resultsby the two widely used docking programs FlexX[20] and Au-toDock.[21] We used these tools successfully in our previousprojects.[22, 23] FlexX has actually been developed atFraunhofer Institute SCAI. Both, FlexX and AutoDock areamongst the most widely used docking tools.[24] FlexX isbased on an incremental base fragment construction and aconformation sampling algorithm while AutoDock is basedon Lamarckian genetic algorithm. FlexX can, by default,generate the interaction information of the predicted bind-ing poses along with the binding energy scores of thedocked conformations of the ligand. However, AutoDockdoes not produce such interaction information by defaultand therefore, AutoDockTools’[25] scripts were modified toproduce similar interaction information for the docked con-formation of the ligand like those of FlexX. The programsSKIFP and SKIFP2SVG, which we have written in C/C ++language, are used to generate and visualize respectively,the fingerprints from the protein-ligand interaction infor-mation of each predicted docked conformation.

Table 1. A sample initial residue profile created from the interac-tions in Figure 1. Interaction types and their initial weights arelisted in front of each residue.

Residue number Interaction type Weights

LEU167 Contact 1PHE168 H-Acceptor 1PHE168 Contact 1GLU169 H-Donor 1GLU169 H-Acceptor 1GLU169 Contact 1

782 www.molinf.com � 2010 Wiley-VCH Verlag GmbH & Co. KGaA, Weinheim Mol. Inf. 2010, 29, 781 – 791

Full Paper M. Shahid et al.

The initial list of residues is obtained from the availableinteraction data of the cocrystallized ligand bound in the X-ray structure of A2A receptor.[19] A data set of about 160known antagonists with experimentally determined bindingaffinities were collected from the literature,[26–29] which areactive against the human Adenosine receptor A2A (PDB-Code: 3EML), a G-protein coupled receptor (GPCR) subtypewhich is an important therapeutic target in neurodegenera-tive disorders.[30] These known active compounds are notall similar but comprised of diverse compound classes ofA2A antagonists including adenine derivatives togetherwith xanthine and non-xanthine derivatives. All the com-pound in the data set of known actives were dockedagainst A2A receptor using FlexX and AutoDock and inter-action information were extracted from the docking results.

The fingerprints were generated and visualised by SKIFPand SKIFP2SVG. The compounds were sorted on the basisof decreasing affinity values and not by the order of bind-ing energies or docking scores. The highly actives are dis-played at the top while the less actives at the bottom,which facilitates building an efficient interaction profile bythe identification of the key residues involved in the inter-action of highly active compounds for the A2A receptor(Figure 2). In Figure 2, the top listed (80 out of 160 com-pounds) represents the highly actives (lower Ki values)while the bottom listed compounds represents the less ac-tives (higher Ki values) antagonists of A2A receptor. Eachdark band in the pattern represents the presence of the in-teraction with a given residue for that compound.

Figure 1. Schematic representation of the interactions between A2A receptor residues and ZM241385 (the cocrystallized ligand) at theligand binding site. The dashed lines represent the close contacts and hydrogen bonding interactions of the ligand with the residues andthe semi-circles show the hydrophobic interactions. ZM241385 is displayed in ‘ball-and-stick’ model. Residues’ mutations that are reportedto disrupt agonist and/or antagonist binding are within blue squares.[19]

Mol. Inf. 2010, 29, 781 – 791 � 2010 Wiley-VCH Verlag GmbH & Co. KGaA, Weinheim www.molinf.com 783

Using Protein–Ligand Interaction Information in Increasing Hits Selection from Virtual Screening

2.3 Preparation of the Weighted Residue Profile

We have developed a variation of building a weighted resi-due profile. A weighted residue profile is prepared fromthe fingerprint patterns of interactions as given in Table 2.This method takes into account the activity data of theknown active compounds. The weights in the residue pro-file are assigned based on the fingerprint densities of thehighly active compounds which highlights that moreweights should be assigned to those residues whose inter-action bits are only present or darker in the top highlyactive compounds subset than the less actives. A finger-print pattern present in all of the complete data set wouldmean that the interaction is present in all of the com-pounds and therefore is not very significant in the activityand would constrain a higher weight value to be assignedto the respective residue in the profile (Figure 3).

2.4 Screening Data Sets

Two test data sets of about 1000 compounds in each setwere randomly selected from the focused library including80 highly active known compounds. The target-focused li-brary was designed for the antagonist study of A2A recep-tor using Pharmacophore models, Feature Trees[31] similaritysearching and scaffold hopping in the large combinatorialFragment Space.[32] This random database set is taken fromthe focused library instead of selecting compounds fromany random “drug-like” compound library and this ensuresto avoid artificial enrichments in recovering the known ac-tives amongst the library.[33] One data set contained com-pounds from ZINC[34] database while the other data setcontained the compounds filtered from the combinatorialFragment Space.

Figure 2. Fingerprint bits for the set of known active compounds by their interaction with the receptor residues given in Table 2. (A:docked with FlexX, B: docked with AutoDock). Compounds are listed on the basis of decreasing affinity, highly actives (lower Ki values) atthe top and less actives (greater Ki values) at the bottom.

784 www.molinf.com � 2010 Wiley-VCH Verlag GmbH & Co. KGaA, Weinheim Mol. Inf. 2010, 29, 781 – 791

Full Paper M. Shahid et al.

2.5 Virtual Screening by Molecular Docking

Docking calculations were carried out using FlexX (version3.1.4) and AutoDock4 (version 4.2.2) for both data sets. De-fault parameter sets were used in FlexX dockings while La-marckian genetic algorithm default parameters were usedin AutoDock dockings. Rigid receptor docking calculationswere used in both docking programs. Docking jobs’ datamanagement, execution and post-processing of the resultswere performed on the compute clusters using our well-established virtual screening framework.[22] The top 10 solu-tions predicted for each docking were retained for furtheranalysis using the interaction fingerprints yielding a de-scriptive set of receptor residues interactions for each com-pound. At the stage of post processing the docking results,Z-scores and TC-MIF (Molecular Interaction Fingerprintswith Tanimoto-Coefficient metric) values were calculatedfor each of the top 10 solutions using the weighted-residueprofiles as shown in Table 2 by the SKIFP program and Fin-gerPrintLib[7] libraries. The best solutions out of the top 10

with the highest Z-score and highest TC-MIF values wereretained for each compound.

2.6 Rescoring by Z-score

Predicted docked poses (docking solutions) are rescored byZ-score, which is just an addition of the weight values tothe docking score calculated by the docking program. Theweight values are taken from the weighted protein residuesprofile and simply added to the docking score of a predict-ed docking solution. Suppose, a docking solution at rank 3or 5 has a docking score less than the top 1st ranked solu-tion, but has a better binding mode and the required inter-actions are matched with those specified in the residueprofile. Simply adding the weight values into its dockingscore will prioritize and re-rank this lower ranked dockingsolution. The pharmacophores incorporated into and deter-mined by other methods and docking tools work in a dif-ferent way. Using pharmacophoric constrains might oftenlead to no docking solutions if the strict pharmacophoricconstrains are not fulfilled by the predicted docking solu-tion. The Z-score is using the pharmacophoric features toprioritize the docking solutions based on the significant in-teractions matching with those specified in the residue pro-file.

3 Results and Discussion

Docking results were ranked and filtered on the basis ofthe following three criteria; 1. on the basis of highestenergy/binding energy scores of FlexX and AutoDock, 2. onthe basis of highest Z-score by using PLIFs with weighted-residue profiles for the top 10 docking solutions, and 3. onthe basis of TC-MIF values calculated for the top 10 dock-ing solutions with using the cocrystallized ligand as a refer-ence. Database enrichment is calculated from the percentactives recovered from the percent database screened. Re-sults quantified from both screening data sets have shownthat Z-score increases hit rates in the top 5 % to 15 % of

Figure 3. Weighted-residues profiles comparison used with FlexX and AutoDock dockings. Weight values are assigned to each interactiontype based on the fingerprint patterns presented in Figure 2 for the two docking tools. With slight differences in the weight values for cer-tain interaction types, both profiles indicate significance of key interactions in the residue profiles.

Table 2. Weighted residue rrofiles for FlexX and AutoDock.

Residues Interactiontype

FlexXweights

AutoDockweights

LEU167 Contact 15 20PHE168 H-Acceptor 1 1PHE168 Contact 1 1GLU169 H-Donor 30 30GLU169 H-Acceptor 20 30GLU169 Contact 30 5TRP246 Contact 5 15LEU249 Contact 5 5HIS250 Contact 10 20ILE252 Contact – 5ASN253 H-Acceptor 30 20ASN253 H-Donor 30 30HIS264 Contact 15 20ALA265 Contact 15 20LEU267 Contact 5 20MET270 Contact 20 5ILE274 Contact 5 5HIS278 Contact 5 5

Mol. Inf. 2010, 29, 781 – 791 � 2010 Wiley-VCH Verlag GmbH & Co. KGaA, Weinheim www.molinf.com 785

Using Protein–Ligand Interaction Information in Increasing Hits Selection from Virtual Screening

the screened database in both docking programs for alldata sets (Figure 4 and Figure 5). Performance measures ofthe three filtering criteria were also quantified by ROC (Re-ceiver Operating Characteristics) curves as shown inFigure 6. Z-score has shown similar performance, in termsof active compounds recovery in the top ranked com-pounds to TC-MIF based filtering in two data sets, and thusimproves the recall ratio in the top ranking results. Furthermore, the binding poses of the top ranked compounds aremore similar to the binding pose of the cocrystallizedligand in the case of Z-score and TC-MIF than the energybased scoring. This indicates the usefulness of this ap-proach to rerank the compounds with good bindingmodes without taking the cocrystallized ligand as a refer-ence.

The main purpose of this study is not to compare the re-sults of different docking tools or to compare the method-ologies of using protein-ligand interaction fingerprints. Themolecular interaction fingerprint method (TC-MIF) is usedto evaluate the efficiency of the current approach and totest the ability of this method for use in postprocessingand filtering results of virtual screening (Figure 7).

As a cross validation test, when equal weight values areassigned to each interaction type in the weighted-residueprofiles, the results are not improved in terms of hits en-richment. This indicates that certain residue’s interactionsare more important than others and must be signified byassigning a higher weight value to them. Similarly, inter-changing the profile of FlexX with AutoDock and vice versa(Table 2 and Figure 3) do not produce satisfactory results ascompared to using the respective profile of the dockingtool, although the results in terms of database enrichmentare still better than the energy based scores.

Figure 4. Z-score comparison with FlexX energy scores and TC-MIFbased rankings for two data sets. Z-score is calculated from theweighted-residue profile and TC-MIF values are calculated frommolecular interaction fingerprints of the docked poses in the pro-tein complex.

Figure 5. Z-score comparison with AutoDock binding energyscores and TC-MIF based rankings for two data sets. Z-score is cal-culated from the weighted-residue profile and TC-MIF values arecalculated from molecular interaction fingerprints of the dockedposes in the protein complex.

786 www.molinf.com � 2010 Wiley-VCH Verlag GmbH & Co. KGaA, Weinheim Mol. Inf. 2010, 29, 781 – 791

Full Paper M. Shahid et al.

3.1 Application of the Method on Benchmark Data Sets

To test the general applicability of the method, one othertarget MAOB and four more targets from standard bench-

mark datasets were used from DUD, a directory of usefuldecoys for benchmarking virtual screening.[35] The targetswere randomly picked from the database. All the targetstructures and the compound databases were energy mini-

Figure 6. Performance measures of the three filtering criteria used in screening two data sets with FlexX. In both data sets, Z-score hasgreater recall of the active compounds than TC-MIF and Energy based rankings in the top 5–15 % of the database screened as also shownin the enrichment plots.

Mol. Inf. 2010, 29, 781 – 791 � 2010 Wiley-VCH Verlag GmbH & Co. KGaA, Weinheim www.molinf.com 787

Using Protein–Ligand Interaction Information in Increasing Hits Selection from Virtual Screening

mized and prepared for docking study with FlexX. The crys-tal ligands bound in the receptor complex were redockedto reproduce the similar crystal structure’s binding modes.The known active compounds of the respective target pro-

teins were used to generate the target specific residue pro-files. Afterwards, the decoy databases were screenedagainst the respective targets and the results of virtualscreening were ranked on the basis of docking score as

Figure 7. Performance measures of the three filtering criteria used in screening two data sets with AutoDock. In both data sets, Z-scorehas greater recall of the active compounds than TC-MIF and Energy based rankings in the top 5–15 % of the database screened as alsoshown in the enrichment plots.

788 www.molinf.com � 2010 Wiley-VCH Verlag GmbH & Co. KGaA, Weinheim Mol. Inf. 2010, 29, 781 – 791

Full Paper M. Shahid et al.

well as Z-score. Improvements in terms of recovering theknown ligands from the decoy database were shown in theresults obtained by ranking with Z-score as presented inFigure 8.

3.2 Binding Mode Analysis

The results ranked on the basis of three criteria have alsoshown that top ranking compounds have good bindingmodes that are similar to the reported binding mode ofthe reference compound ZM241385 (the cocrystallizedcompound in the crystal structure of the receptor) which isa selective A2A receptor antagonist.[36] For comparison, the

Figure 8. Improvements in recovering known ligands from the database of decoys by ranking with Z-score. Top left: MAOB (MonoamineOxidase-B), top right: ACE (Angiotensin-converting enzyme), bottom left: ACHE (Acetylcholine esterase) and bottom right: INHA (Enoyl ACPreductase).

Mol. Inf. 2010, 29, 781 – 791 � 2010 Wiley-VCH Verlag GmbH & Co. KGaA, Weinheim www.molinf.com 789

Using Protein–Ligand Interaction Information in Increasing Hits Selection from Virtual Screening

rank assigned to the reference compound ZM241385 bythe three criteria in 1000 database compounds dockedagainst A2A receptor are shown in Table 3 which highlightsthe increase in efficiency of binding mode as well. Thesame cocrystallized compound has been ranked among thetop 1 % of the database screened in the case of ranking byZ-score and TC-MIF in all data sets.

4 Summary and Conclusions

Docking scores, so far, are standard ways to sort out andselect the best hits while analysing virtual screening results.The inability of scoring functions to accurately score welland rank order drug candidates in a virtual screening cam-paign is currently the main limitation of docking tools.Thus, by taking the available knowledge of the proteinstructure, selection of the candidate compounds from thedocking results require a lot of human post-processingwhich is much time consuming. To address this problem,we have developed a simple and interactive way of gener-ating protein-ligand interaction information from the dock-ing results. This information can be utilized efficiently inthe analysis of large-scale structure-based virtual screeningresults to prioritize candidate compounds with good bind-ing modes. The Z-score, which is based on the weighted-residue profile, is obtained for a set of important residuesinteractions and then added to the docking score for a pre-dicted docking solution/pose. In this study, Z-score hasdemonstrated sufficient increase in the enrichment ofactive compounds against the randomly selected 1000compound data set. The data sets were obtained from thealready prepared focussed library of A2A receptor. Com-pared to docking-score alone (energy based), the additiveZ-score methodology performed very well in retrieving theactive compounds against the database of decoys. In con-trast to the protein-ligand interaction fingerprints that wereproposed previously, the methodology proposed in thispaper is simple and can be easily implemented in an exist-ing virtual screening workflow, which further makes thismethod easily adaptable by structural biologists and me-dicinal chemists. Though the importance of PLIF in rationaldrug discovery was identified long ago[11, 13] and severalother studies have been published,[4, 12] the current studydiffers by deriving PLIFs from the generated docking re-

sults, rather than from the explicit use of protein-ligandcomplexes. Furthermore, fingerprint bits are residue basedand not determined from the atom to atom contacts in themolecule complex. Weights in the weighted residue profilesare also determined from the initial fingerprints data of theknown highly active compounds instead of using anymathematical function to calculate weights as presented inReference.[14]

As we have demonstrated in this paper, the use of PLIFinformation can significantly enrich the number of positivehits in a virtual screening experiment performed with twodifferent docking tools. The amount of time spent and ef-forts involved in filtering by manual post-processing will beeffectively reduced by integrating the current approach inthe virtual screening pipeline. We therefore propose to in-tegrate this approach at the crucial stage of analysing andfiltering the results of virtual screening by molecular dock-ing.

Acknowledgements

We are very much thankful to Didier Rognan, the author ofFingerPrintLib[7] for providing us the executable librariesand OpenEye[37] for granting us the evaluation license ofOEChem Toolkit libraries.

References

[1] A. N. Jain, Curr. Opin. Drug Discov. Devel. 2004, 7, 396 – 403.[2] D. B. Kitchen, et al. , Nat. Rev. Drug Discov. 2004, 3, 935 – 949.[3] N. Moitessier, et al. , Br. J. Pharmacol. 2008, 153, 7 – 26.[4] S. C. Brewerton, Curr. Opin. Drug Discov. Devel. 2008, 11, 356 –

364.[5] H. F. Velec, H. Gohlke, G. Klebe, J. Med. Chem. 2005, 48, 6296 –

6303.[6] P. Pfeffer, G. Neudert, G. Klebe, Chem. Cent. J. 2008, 2, 16 – 16.[7] G. Marcou, D. Rognan, J. Chem. Inf. Model. 2007, 47, 195 – 207.[8] N. Okimoto, et al. , PLoS Comput. Biol. 2009, 5.[9] G. Degliesposti, et al. , ChemMedChem 2009, 4, 1164 – 1173.

[10] M. D. Kelly, R. L. Mancera, J. Chem. Inf. Model. 2004, 44, 1942 –1951.

[11] Z. Deng, C. Chauqui, J. Sing, J. Med. Chem. 2004, 47, 337 – 344.[12] V. I. P�rez-Nueno, et al. , J. Chem. Inf. Model. 2009, 49, 1245 –

1260.[13] C. Chuaqui, Z. Deng, J. Singh, J. Med. Chem. 2005, 48, 121 –

133.[14] R. K. Nandigam, et al. , J. Chem. Inf. Model. 2009, 49, 1185 –

1192.[15] A. Kumar, M. I. Siddiqi, J. Mol. Graph. Model. 2008, 27, 476 –

488.[16] S. Renner, et al. , J. Chem. Inf. Model. 2008, 48, 319 – 332.[17] A. C. Wallace, R. A. Laskowski, J. M. Thornton, Protein Eng.

1995, 8, 127 – 134.[18] R. A. Laskowski, Nucleic Acids Res. 2001, 29, 221 – 222.[19] V. P. Jaakola, et al. , Science, 2008, 322, 1211 – 1217.[20] M. Rarey, et al. , J. Mol. Biol. 1996, 261, 470 – 489.[21] G. M. Morris, et al. , J. Comput. Chem. 1998, 19, 1639 – 1662.

Table 3. Reference compound ZM241384 ranked in two screeningdata sets by three criteria.

Criteria FlexX dockings AutoDock dockings

Data set 1 Data set 2 Data set 1 Data set 2

Energy score 15 32 208 230Z-score 1 3 9 13TC-MIF 3 2 1 1

790 www.molinf.com � 2010 Wiley-VCH Verlag GmbH & Co. KGaA, Weinheim Mol. Inf. 2010, 29, 781 – 791

Full Paper M. Shahid et al.

[22] M. Shahid, et al. , Stud. Health Technol. Inform. 2009, 147, 212 –221.

[23] V. Kasam, et al. , J. Chem. Inf. Model. 2007, 47, 1818 – 1828.[24] S. F. Sousa, P. A. Fernandes, M. J. Ramos, Proteins, 2006, 65,

15 – 26.[25] M. F. Sanner, J. Mol. Graphics Mod. 1999, 17, 75 – 61.[26] G. Strappaghetti, et al. , Bioorg. Med. Chem. 2001, 3, 575 – 583.[27] C. E. Mueller, et al. , J. Med. Chem. 2002, 45, 3440 – 3450.[28] M. Mantri, et al. , J. Med. Chem. 2008, 51, 4449 – 4455.[29] M. H. Holschbach, et al. , Eur. J. Med. Chem. 2006, 41, 7 – 15.[30] D. Blum, et al. , The Lancet Neurology, 2003, 2, 366 – 374.[31] M. Rarey, J. S. Dixon, J. Comput. Aided Mol. Des. 1998, 12, 471 –

490.[32] U. Lessel, et al. , J. Chem. Inf. Model. 2009, 49, 270 – 279.

[33] M. L. Verdonk, et al. , J. Chem. Inf. Comput. Sci. 2004, 44, 793 –806.

[34] J. J. Irwin, B. K. Shoichet, J. Chem. Inf. Model. 2005, 45, 177 –182.

[35] N. Huang, B. K. Shoichet, J. J. Irwin, J. Med. Chem. 2006, 49,6789 – 6801.

[36] S. M. Poucher, et al. , Br. J. Pharmacol. 1995, 115, 1096 – 1102.[37] OpenEye, OpenEye : Scientific Software, http://www.eyesopen.

com, cited: 2010 1 – 3.

Received: June 25, 2010Accepted: September 3, 2010

Published online: October 27, 2010

Mol. Inf. 2010, 29, 781 – 791 � 2010 Wiley-VCH Verlag GmbH & Co. KGaA, Weinheim www.molinf.com 791

Using Protein–Ligand Interaction Information in Increasing Hits Selection from Virtual Screening