Machine learning-guided discovery and preliminary biological evaluation of novel GSK-3β inhibitory scaffolds with anti-inflammatory potential

This research presents an efficient machine learning-based computational pipeline aimed at the discovery of novel Glycogen Synthase Kinase-3β (GSK-3β) inhibitors. A Random Forest (RF) classification model was created using a carefully chosen set of confirmed inhibitors from 2016 to 2024. The optimized RF model was very strong and accurate (Accuracy: 0.89; AUC-ROC: 0.91 ± 0.02), making it a great first filter for finding chemicals in large spaces. We using this classifier on a new library (Asinex) to obtain the promising candidates, which then went through a series of steps, including ADMET filtering, molecular docking, and molecular dynamics (MD) simulations. The pipeline found two good candidates, RF03 and RF05. In vitro enzymatic assays demonstrated moderate GSK-3β target engagement, producing IC₅₀ values of 85.04 µM and 58.73 µM, respectively. The most important thing is that the lead hit, RF05, which has a non-classical, sterically hindered aromatic ester scaffold, worked very well at the cellular level. More biological tests showed that RF05 greatly lowers the levels of pro-inflammatory cytokines (TNF-α and IL-6) while keeping low cytotoxicity. We suggest that this unique chemotype's high lipophilicity and structural pre-organization make it easier for cells to pass through membranes and build up inside cells, turning a moderate biochemical affinity into a strong phenotypic response. This validated computational strategy significantly reduces experimental attrition. Compound RF05 is a potential lead scaffold that gives a structurally different base for the rational design of new anti-inflammatory drugs.