QSAR strategy and experimental validation for the development of a GPCR focused library

An innovative approach has been developed in order to construct new GPCR focused libraries. Experimental binding data generated in house from 1939 diverse drug and druglike compounds on 40 GPCR targets were used to develop and validate a global GPCR QSAR model accounting for pharmacophore features related to a general GPCR-binding behavior. To this end, proprietary 3-D descriptors representing pharmacophore fingerprints of the various conformers of the molecules were used to encode compound structures in a numerical form. Statistical treatment of the data was based on two different approaches, linear regression and predictive neighborhood behavior, and synergy models relying on both these two independent approaches were also developed. The best QSAR model was selected on hand of its statistical parameters (R-2, RMS) and percentage of correctly predicted compounds on a randomly chosen validation set (20% of the compounds). A diverse GPCR library of 2,400 compounds was prepared by applying the global QSAR model on compounds already synthesized in house, as well as on virtual combinatorial compounds which were then synthesized if predicted to be potential GPCR binders by the model. The set of building blocks used to build combinatorial libraries has been enriched in original GPCR-like monomers, specially designed for this purpose according to medicinal chemistry know-how and literature knowledge. To validate our approach, 240 compounds (10%) of this library were randomly chosen and tested on 21 different amine and peptide GPCRs, together with 720 combinatorial compounds from an in house diversity-based hit-seeking library, as a reference. The experimental results on these 960 compounds were analyzed after pooling the compounds into those predicted as GPCR-active vs. inactive (360 and 600 compounds respectively). The average hit rate was found to be 5.5 fold higher for the GPCR predicted compounds and, furthermore, the global QSAR model was able to recognize not only classical templates but also original ones, allowing us to identify new GPCR chemotypes interacting with aminergic and peptidergic GPCRs.