Virtual Screening of Kinase Based Drugs: Statistical LearningTowards Drug Repositioning

Abstract

Kinases are phosphate catalysing enzymes that have traditionally proved difficult to target against ligands, and hence inefficacious in drug development. There are two colluding reasons for this. First is the issue of specificity. The homogeneity that exists between the kinase ATP-binding pockets makes it a non-realisable target to develop compounds that would inhibit only one out of 538 protein kinases encoded by the human genome, without inhibiting some of the others. Second, producing compounds with the required efficacy to rival the millimolar ATP concentrations present in cells is stoichiometrically inefficient. This study uses a recently propounded computational strategy based on Structure-Based Virtual Screening (SBVS) that was previously benchmarked on 999 DUD-E protein decoys (Chattopadhyay et al, Int Sc. Comp. Life Sciences 2022), to rank potential ligands, or by extension rank kinase-ligand pairs, identifying best matching ligand:kinase docking pairs. The results of the SBVS campaign employing several computational algorithms reveal variations in the preferred top hits. To address this, we introduce a novel consensus scoring algorithm by sampling statistics across four independent statistical universality classes, statistically combining docking scores from ten docking programs (DOCK, Quick Vina-W, Vina Carb, PLANTS, Autodock, QuickVina2, QuickVina21, Smina, Autodock Vina and VinaXB) to create a holistic SBVS formulation that can identify active ligands for any target. Our results demonstrate that CS provides improved ligand:kinase docking fidelity when compared to individual docking platforms, requiring only a small number of docking combinations, and can serve as a viable and thrifty alternative to expensive docking platforms.