Machine Learning– driven discovery and optimization of PI3Kδ inhibitors for diffuse large B-cell lymphoma

Diffuse large B-cell lymphoma (DLBCL) remains difficult to treat, as current PI3Kδ inhibitors are limited by toxicity and poor metabolic stability. To address these challenges, we applied a machine learning (ML)–guided design framework integrating predictive quantitative structure–activity relationship (QSAR) modelling, absorption, distribution, metabolism, excretion, and toxicity (ADMET) profiling, molecular docking, and molecular dynamics simulations. A curated dataset of 3268 compounds was reduced to 9 descriptors, supporting a robust Random Forest model that achieved strong predictive performance on an external test set (R2 ≈ 0.63; concordance correlation coefficient (CCC) ≈ 0.76), while maintaining high training set fit (R2 = 0.94). This indicates good generalization to unseen compounds. Scaffold prioritization nominated Compound 9, which showed strong docking affinity (–9.5 kcal/mol) and favourable pharmacokinetic predictions. Guided optimization produced analogues 9a–9c, of which 9b demonstrated superior absorption (96.6%), reduced CNS penetration, and stable hydrogen bonding with VAL828 in 100-ns molecular dynamics (MD) simulations (root-mean-square deviation [RMSD] ∼1.2–1.8 Å). Molecular mechanics–generalized Born surface area (MM-GBSA) analysis confirmed favourable binding free energy (–77.6 kcal/mol), with 9b achieving enhanced stability through balanced energetic contributions. Collectively, these results highlight Compound 9b as a computationally prioritized lead for further experimental evaluation and illustrate the power of machine learning–driven, simulation-validated pipelines in accelerating kinase inhibitor discovery for aggressive lymphomas. Experimental validation is needed to confirm these predictions.