A selectivity-aware machine-learning workflow for prioritizing CDK2-biased kinase inhibitor candidates

Abstract Cyclin-dependent kinase 2 (CDK2) is a therapeutic target of interest, but selective inhibition remains challenging because ATP-competitive inhibitors often engage related CDK family members. Here, we developed a selectivity-aware computational workflow to prioritize CDK2-biased kinase inhibitor candidates using curated ChEMBL bioactivity data, target-specific machine-learning models, scaffold-guided analog generation, applicability-domain filtering, medicinal-chemistry triage, control-compound benchmarking, molecular docking, and MM-GBSA rescoring. IC 50 data for CDK1, CDK2, CDK4, CDK6, CDK7, and CDK9 yielded 7,067 compound-target records and 4,879 standardized compounds. Extra Trees models trained with RDKit descriptors and Morgan fingerprints provided useful multi-target activity prediction, with the CDK2 model achieving MAE = 0.523, RMSE = 0.699, and R 2 = 0.607 in random-split evaluation. Predicted CDK2 selectivity margin recovered measured CDK2-biased compounds more effectively than predicted CDK2 potency alone, including ROC-AUC = 0.921 and average precision = 0.846 in retrospective enrichment and ROC-AUC = 0.888 ± 0.061 in scaffold-held-out enrichment. Focused sulfonamide analog generation and filtering prioritized 221 CDK2-biased candidates, of which six showed CDK2-favored docking in both SP and XP modes. MM-GBSA rescoring further identified one compound with convergent CDK2 preference across SP docking, XP docking, and post-docking energetic rescoring. This workflow provides a reproducible strategy for enriching CDK2-biased analogs for experimental evaluation.