Choosing the Optimal Rigid Receptor for Docking and Scoring in the CSAR 2013/2014 Experiment

Abstract

The 2013/2014 Community Structure–Activity Resource (CSAR) challenge was designed to prospectively validate advancement in the field of docking and scoring receptor–small molecule interactions. Purely computational methods have been found to be quite limiting. Thus, the challenges assessed methods that combined both experimental data and computational approaches. Here, we describe our contribution to solve three important challenges in rational drug discovery: rank-ordering protein primary sequences based on affinity to a compound, determining close-to-native bound conformations out of a set of decoy poses, and rank-ordering sets of congeneric compounds based on affinity to a given protein. We showed that the most significant contribution to a meaningful enrichment of native-like models was the identification of the best receptor structure for docking and scoring. Depending on the target, the optimal receptor for cross-docking and scoring was identified by a self-consistent docking approach that used the Vina scoring function, by aligning compounds to the closest cocrystal or by selecting the cocrystal receptor with the largest pocket. For tRNA (m1G37) methyltransferase (TRMD), ranking a set of 31 congeneric binding compounds cross-docked to the optimal receptor resulted in a R2 = 0.67; whereas, using any other of the 13 receptor structures led to almost no enrichment of native-like complex structures. Furthermore, although redocking predicted lower RMSDs relative to the bound structures, the ranking based on multiple receptor structures did not improve the correlation coefficient. Our predictions highlight the role of rational structure-based modeling in maximizing the outcome of virtual screening, as well as limitations scoring multiple receptors.

Optimal strategies for virtual screening of induced-fit and flexible target in the 2015 D3R Grand Challenge

“…As reported, our approach of using “close” methods for pose predictions, where all available crystallographic information (mostly co-crystals) is used, yielded the most accurate poses in the community-wide experiment for both targets. Consistent with our earlier work [29], align-close was particularly robust for both flexible targets. Although dock-close predicted better-docked poses to the rather tight pocket of HSP90, the same method failed predicting good poses in the much larger binding site of MAP4K4.…”

“…We participated in the 2013/14 CSAR challenge that involved rank-ordering compounds to homology models of the receptors with a given protein primary sequence, identifying close-to-native bound conformations out of a set of decoy poses, and rank-ordering the affinity of sets of congeneric compounds to a given protein. Our predictions were among the best in the field [29, 30]. We showed that the most significant contribution to a meaningful enrichment of native-like models was the identification of the best receptor structure for docking and scoring.…”

“…For affinity ranking, the choice of methods seems to be more dependent on the binding pocket features. Consistent with MAP4K4, another kinase from CSAR 2013/14 competition, SYK, had “cross” methods as the best ranking method [29], whereas for TRMD, a target with a rigid and buried pocket perhaps resembling HSP90, dock-cross performed the best [29]. These results suggest a preference of method selection in different pocket types, but the specific features that dominate the selection of the optimal method remains to be determined.…”

“…These strategies included methods that utilize all available receptor/ligand co-crystals (“close”), all available ligands and a single holo -receptor structure (“min-cross”) and only a single receptor/ligand co-crystal (“cross”). As in the 2013/14 CSAR competition [29], we found that the method that predicted the best docking poses was not the same as the ones that predicted the best ranking of active compounds. Similarly, different methods were shown to predict the optimal ranking of active compounds for HSP90 and MAP4K4, i.e., “close” and “cross”, respectively.…”

Combined Approach of Patch-Surfer and PL-PatchSurfer for Protein–Ligand Binding Prediction in CSAR 2013 and 2014

“…Values of Baumgartner et al . 49 were computed from Figure 6 of their paper. Values of Matiny et a l. 46 were computed from R 2 values in Table 3 of their paper.…”