Improving Quantitative Structure-Activity Relationships through Multiobjective Optimization

A multiobjective optimization algorithm was proposed for the automated integration of structure- and ligand-based molecular design. Driven by a genetic algorithm, the herein proposed approach enabled the detection of a number of trade-off QSAR models accounting simultaneously for two independent objectives. The first was biased toward best regressions among docking scores and biological affinities; the second minimized the atom displacements from a properly established crystal-based binding topology. Based on the concept of dominance, 3D QSAR equivalent models profiled the Pareto frontier and were, thus, designated as nondominated solutions of the search space. K-means clustering was, then, operated to select a representative subset of the available trade-off models. These were effectively subjected to GRID/GOLPE analyses for quantitatively featuring molecular determinants of ligand binding affinity. More specifically, it was demonstrated that a) diverse binding conformations occurred on the basis of the ligand ability to profitably contact different part of protein binding site; b) enzyme selectivity was better approached and interpreted by combining diverse equivalent models; and c) trade-off models were successful and even better than docking virtual screening, in retrieving at high sensitivity active hits from a large pool of chemically similar decoys. The approach was tested on a large series, very well-known to QSAR practitioners, of 3-amidinophenylalanine inhibitors of thrombin and trypsin, two serine proteases having rather different biological actions despite a high sequence similarity.