AI-driven generative framework integrating ML-QSAR and fragment learning for isoform-selective PI3K inhibitor design

Isoform-selective inhibition of class I phosphoinositide 3-kinases (PI3Ks) remains a major challenge in oncology and immune-mediated diseases, where dysregulated PI3K signaling drives tumor progression, therapeutic resistance, and aberrant immune activation. Efforts to achieve precise isoform selectivity are constrained by the high structural similarity among the catalytic subunits α, β, δ, and γ. In this regard, we developed an artificial intelligence (AI)-driven integrative framework that combines machine learning-based quantitative structure-activity relationship (ML-QSAR) modeling, fragment-level selectivity profiling, and reinforcement-learning generative chemistry to design isoform-selective PI3K inhibitors. Curated ChEMBL datasets were used to train independent XGBoost models for each isoform, achieving strong predictive performance (R² = 0.76-0.82; RMSE = 0.48-0.51) and interpretable SHapley Additive exPlanations (SHAP)-based feature attribution. Fragment analysis identified isoform-specific structural motifs that were used to guide targeted molecular exploration with the FREED +  + reinforcement-learning algorithm. The framework generated over 10,000 unique compounds, and molecular docking analysis showed favorable binding energies (- 7.9 to - 9.7 kcal/mol) and interactions consistent with known isoform-selective inhibitors. Generated molecules also exhibited suitable drug-likeness and synthetic accessibility, highlighting their potential as viable lead compounds. Collectively, this study demonstrates how combining predictive ML models with fragment-aware generative AI enables rapid discovery of selectivity-optimized PI3K inhibitors. The proposed pipeline is generalizable to other multi-isoform targets and establishes a scalable AI methodology for next-generation rational drug design in precision oncology and immune-modulating drug development.