EXPLAINABLE AI FOR DECODING HALOGENATED SYZYGIUM FLAVONOID BINDING IN PI3KSALPHA

Dongare Tanvi Sopan*, Lokhande Rahul Prakash, Gaikwad Sakshi Rajesh, Dere Aditya Sampat, Gadge Shrutika Pralhad, Datkhil Parth Dnyaneshwar, Dhaygude Saurabh Hanumant, Gadekar Sainath Vijaysingh

The Phosphoinositide 3-kinase alpha (PI3K$alpha$) isoform is a primary driver of oncogenic signaling in breast cancer, yet the development of potent, isoform-selective inhibitors remains a significant structural challenge. While natural flavonoids from Syzygium species offer a promising chemical scaffold, their clinical utility is often limited by moderate binding affinities. In this study, we utilize semi-synthetic halogenation to enhance the potency of these scaffolds and introduce an Explainable AI (XAI) framework to decode the underlying binding mechanisms. A library of halogenated flavonoids was docked against the PI3K$alpha$ catalytic domain, and the resulting 3D poses were converted into high-dimensional Protein-Ligand Interaction Fingerprints (PLIP). An XGBoost regressor was trained on these fingerprints to predict binding energies, while SHAP (SHapley Additive exPlanations) was applied to the model to provide a quantitative, residue-level interpretation of the AI's decision-making process. Our results identify [Insert Lead Compound Name] as a superior inhibitor, with SHAP analysis demonstrating that the model’s high-affinity predictions were primarily driven by the formation of critical halogen bonds at the TYR158 residue. By transitioning from traditional "black-box" scoring to an interpretable XAI approach, this study not only identifies novel anti-cancer leads but also provides a mechanistic roadmap for the rational design of halogen-enhanced therapeutics.