Deep learning-assisted virtual screening of a large chemical library for selective GSK3β inhibitors.

Glycogen synthase kinase-3β (GSK3β) is a serine/threonine kinase involved in neurodegenerative, neuropsychiatric, and oncological disorders. The development of selective inhibitors remains challenging due to high kinase conservation, off-target effects, and suboptimal pharmacokinetic properties. A curated dataset of GSK3β molecules was obtained from PubChem, and molecular descriptors were calculated using PaDEL. GSK3BMTPred, a multitask deep neural network (DNN) model, was developed for simultaneous prediction of inhibitor classification and inhibitory potency. The optimized model achieved a training accuracy of 0.9809 for the classification task and a training coefficient of determination (R2) of 0.8692 for the regression task. The multitask framework improved generalization by learning shared molecular representations across both tasks, supported by SHapley Additive exPlanations (SHAP)-based interpretation of key physicochemical features. The model was integrated into a virtual screening workflow followed by molecular docking, molecular dynamics (MD) simulations, and Mechanics/Generalized Born Surface Area (MM/GBSA) analysis. This approach identified compounds showing stable interactions with key Adenosine Triphosphate (ATP)-binding residues and favorable predicted absorption, distribution, metabolism, excretion, and toxicity (ADMET) properties, including blood-brain barrier (BBB) permeability and GSK3α selectivity. Mp7 was identified as a selective GSK3β inhibitor, whereas Mp8 demonstrated dual activity toward both GSK3β and GSK3α isoforms. Overall, this study presents a scalable computational framework for the discovery and prioritization of selective GSK3β inhibitors. The GSK3BMTPred model is freely available at: https://github.com/PGlab-NIPER/GSK3BMTPred.git .