Lead Finder's performance in virtual screening was assessed on a set of 34 protein targets. These structures were chosen on the basis of:

• protein relevance for drug discovery research;
• availability of high-quality 3D-structure of a protein;
• availability of sufficient number of well-characterized active ligands;

For each protein target a set of active ligands was extracted from the following public sources: PDB database¹, KiBank², active ligands exposed by Surflex developers 3. A set of 1904 decoy ligands used in all current virtual screening studies was borrowed from the original Surflex publication⁴, available for download from Surflex developers site³.

Benchmarking experiments consisted of creating a test library of compounds (obtained by mixing active ligands with the set of decoy compounds), docking each compound from the library to each particular target, and rank-ordering compounds according to the calculated VS-score. For quantitative characterization of virtual screening efficiency two well recognized parameters were used: area under the so-called receiver operating curve (ROC) and enrichment factor (EF). Results of virtual screening experiments are presented in the table below.

Lead Finder has been able to achieve significant enrichments for almost all protein targets. The most outstanding enrichments were obtained for oligopeptide-binding protein (OppA), orotidine-5´-phosphate decarboxylase (OPD), protein tyrosine phosphatase 1B (PTP1B), beta-secretase, HIV-1 protease, factor Xa, peroxisome proliferator activated receptor gamma (PPAR-γ). Analysis of docked ligand poses revealed that for OppA and OPD Lead Finder correctly found multiple hydrogen bonds and complementary electrostatic interactions of charged groups, which yielded high enrichments for those targets. For PPAR-γ, HIV-1 protease and beta-secretase the basis for impressive results was in correct detection of crucial hydrogen bonds and hydrophobic contacts between active ligands and corresponding proteins. Interestingly, that PPAR-γ was traditionally recognized as a difficult target^5,6,7 and even two protein models were used for virtual screening studies of PPAR-γ⁵, however Lead Finder was able to achieve excellent results using a single protein structure model. Successful docking of positively charged fragment of factor Xa inhibitors gave rise to high enrichment for the corresponding target.

Nuclear receptors demonstrated from moderate (for glucocorticoid receptor (GR) and progesterone receptor (PR)) to high (for estrogen receptor and mineralocorticoid receptor) enrichments. Close analysis of GR and PR reveals that their ligand binding cavities are too spacious and can easily accommodate ligands of diverse shape and size, giving rise to false positives.

Thymidylate synthase also revealed quite modest enrichment. In this case we attribute results to insufficiently specific binding of its native ligands (predicted average binding energy over active ligands comprised -8.7 kcal/mol only), which is probably stipulated by relatively shallow and surface exposed binding site of the enzyme. Probably, similar situation takes place for neuraminidase, which binding site represents extended funnel; ligands have too much freedom to slide along protein surface, which dilutes the specificity of binding. As in the case with TS, native ligands of neuraminidase reveal quite modest binding energy (-7.3 on average).

Finally, protein kinases deserve special attention as they are intensively studied as drug targets nowadays⁸. Currently, 7 protein kinases were assessed by Lead Finder in virtual screening experiments. Overall, kinases revealed good enrichment from ROC = 0.86 for fibroblast growth factor receptor kinase (FGFR) and vascular endothelial growth factor receptor kinase 2 (VEGFR) to ROC = 0.95 for epidermal growth factor receptor kinase (EGFR) and ROC = 0.96 for tyrosine kinase C-SRC (SRC). Close examination of predicted poses for active kinase ligands revealed common feature determining the degree of virtual screening success: targets, for which active ligands succeeded forming crucial correlated hydrogen bonds with hinge fragment of a kinase, demonstrated higher enrichment than targets for which these hydrogen bonds were observed less frequently. We also found that relative mobility of N- and C-terminal lobes of a catalytic kinase domain influences accessibility of correct ligand-binding pose. For this reason evaluation of multiple protein conformers is suggested to achieve high enrichment results in case of kinases.

---

• H. M. Berman, J. Westbrook, Z. Feng, G. Gilliland, T. N. Bhat, H. Weissig, I. N. Shindyalov, P. E. Bourne: The Protein Data Bank. Nucleic Acids Research, 28 pp. 235-242 (2000).
• Zhang, J.-W., Aizawa, M., Amari, S., Iwasawa, Y., Nakano  T. and Nakata, K.: Development of KiBank, a database supporting structure-based drug design. Computational Biology and Chemistry, 28 (5-6), 401-407 (2004) [ Abstract ].
• http://www.jainlab.org/downloads.html
• Tuan  A. Pham and Ajay  N. Jain Parameter Estimation for Scoring Protein-Ligand Interactions Using Negative Training Data  J. Med. Chem. 2006, 49, 5856-5868[ Abstract ].
• R. A. Friesner, J. L. Banks, R. B. Murphy, T. A. Halgren, J. J. Klicic, D. T. Mainz, M. P. Repasky, E. H. Knoll, M. Shelley, J. K. Perry, D. E. Shaw, P. Francis, P. S. Shenkin, Extra Precision Glide: Docking and Scoring Ltdorporating a Model of Hydrophobic Enclosure for Protein-Ligand Complexes, J. Med. Chem. 2006, 49, 6177-6196[ Abstract ].
• Gregory  L. Warren, C. Webster Andrews, Anna-Maria Capelli, Brian Clarke, Judith LaLonde, Millard  H. Lambert, Mika Lindvall, Neysa Nevins, Simon  F. Semus, Stefan Senger, Giovanna Tedesco, Ian  D. Wall, James  M. Woolven, Catherine  E. Peishoff, and Martha  S. Head A Critical Assessment of Docking Programs and Scoring Functions  J. Med. Chem. 2006, 49, 5912-5931[ Abstract ].
• Niu Huang, Brian  K. Shoichet, and John  J. Irwin Benchmarking Sets for Molecular Docking  J. Med. Chem. 2006, 49, 6789-6801[ Abstract ].
• John  P. Overington, Bissan Al-Lazikani and Andrew  L. Hopkins How many drug targets are there? Nat Rev Drug Disc, 2006, 5, 993-996. [ Abstract ].