In silico classification and virtual screening of maleimide derivatives using projection to latent structures discriminant analysis (PLS-DA) and hybrid docking

Abstract

In silico screening algorithms are frequently included in drug discovery programs because a significant number of drug candidates have been detected through structure and ligand-based algorithms. In the current work 337 maleimide derivatives that are inhibitors and non-inhibitors of GSK-3α/β were successfully investigated by means of a projection to latent structures discriminant analysis and hybrid docking. These models developed with Dragon (M1) and OpenEye (M2) descriptors are statistically robust (training set M1: R ²_X  = 0.677, R ²_Y  = 0.976, Q ²_Y  = 0.970; M2: R ²_X  = 0.651, R ²_Y  = 0.835, Q ²_Y  = 0.830) and suitably predictive according to Golbraikh–Tropsha external validation criteria (test set M1: R ² = 0.949; M2: R ² = 0.835). The models appropriately explained the structural differences between active and inactive compounds in terms of graph topology, substitutional pattern, and molecular flexibility, and predicted false negatives in PubChem assay 1650. The model M2 showed 73.88 % correct external prediction against 264 active maleimides downloaded from ChEMBL. An evaluation of the key interactions with GSK-3β binding site residues was simulated by hybrid docking. A new virtual screening methodology involving equation M2 and hybrid docking was applied to 9,042 maleimide derivatives extracted from PubChem. The model M2 predicted 1,327 active compounds that were subsequently docked into the GSK-3β ATP binding site. Finally 648 compounds were established as hits after the exclusion of previously detected active maleimides. The structural diversity of the new compounds is high demonstrating that the scaffold hopping ability of the current approach is noticeable.

Graphical abstract