Novel semi-automated methodology for developing highly predictive QSAR models: application for development of QSAR models for insect repellent amides

Abstract

Conventional 3D-QSAR models are built using global minimum conformations or quantum-mechanics based geometry-optimized conformations as bioactive conformers. QSAR models developed using the global minima as bioactive conformers, employing the GFA, PLS and G/PLS methodologies, gave good non-validated r(2) (0.898, 0.868 and 0.922) and performed well on an internal validation test with leave-one-out correlation q(2) (LOO) (0.902, 0.726 and 0.924), leave-10%-out correlation q(2) (L10O) (0.874, 0.728 and 0.883) and leave-20%-out q(2) (L20O) (0.811, 0.716 and 0.907). However, they showed poor predictive ability on an external data set with best predictive r(2) (Pred-r(2)) of 0.349, 0.139 and 0.204 respectively. A novel methodology to mine bioactive conformers, from clusters of conformations with good 3D-spatial representation around pharmacophoric moiety, furnishes highly predictive 3D-QSAR models. The best QSAR model (model A) showed r(2) of 0.989, q(2) (LOO) of 0.989, q(2) (L10O) of 0.980, q(2) (L20O) of 0.963 and Pred-r(2) on eight test compounds of 0.845. The methodology is based on mimicking the multi-way Partial Least Squares (PLS) technique by performing several automated sequential PLS analyses. The poses/shapes of the mined bioactive conformers provide valuable insight into the mechanism of action of the insect repellents. All of the repetitive tasks were automated using Tcl-based Cerius2 scripts.