Automated pharmacophore hypothesis generator for SILCS-Pharm: ph4gen

Here we present ph4gen, a tool for the automated generation of pharmacophore hypotheses developed for use with the site identification by ligand competitive saturation (SILCS) technology. It builds upon the previously established SILCS-Pharm approach which generates the pharmacophore features from a given macromolecular structure, and SILCS-Hotspots which identifies potentially-druggable binding sites. The tool examines a set of pharmacophore features, typically 15-20, and ranks all permutations of hypotheses of a desired length, with default of 4 to 5 features per hypothesis. A diverse data set of seven proteins (namely, BACE1, CDK2, JNK1, P38, PTPT1B, Thrombin, and TYK2) and 336 ligands was used to train the logistic regression model which ranks these hypotheses and recalls at least one experimental pharmacophore hypothesis in four of seven cases in the top 10. The model was tested on a set of four separate proteins (namely FXR, HDM2, HSP90, and TRMD) and 354 ligands and recalls at least one experimental pharmacophore hypothesis in the top 5 in all four test systems, with mean precision and recall of 0.23 ± 0.06 and 0.15 ± 0.04, respectively. We show that the distribution of hypotheses generated is similar to that of the validated hypotheses, and that even in cases of near-misses, the pharmacophore hypotheses have similar chemical features to validated ones. The tool can also be tuned to provide greater or lesser chemical diversity when utilized in virtual screening. Ultimately, ph4gen is a useful tool as part of the SILCS workflow to identify druggable binding sites, generate drug-like pharmacophore hypotheses, and rank in silico hits for experimental validation.