Reinforcement learning for de novo RET inhibitor design: potency-focused optimization and polypharmacological multi-parameter optimization against resistance mutations
Surendra Kumar, Vinay Pogaku, Mi‐Hyun Kim · Journal of Cheminformatics · 2026
Despite their role as oncogenic drivers and predictive biomarkers, alterations in rearrangement during transfection (RET), a receptor tyrosine kinase (RTK), remain significant challenges due to off-target effects and reduced efficacy against emerging mutations, necessitating more selective and innovative drug design. Herein, we report a multi-layered framework encompassing both reinforcement learning (RL) training and post-processing phases, for the efficient de novo design of RET inhibitors striking a balance between structural novelty and predictive reliability. We evaluated two RL strategies: potency-focused optimization (PFO) and polypharmacological multi-parameter optimization (PMPO), with the latter integrating 131 predictive models for off-target selectivity and phenotypic activity. For the 23 kinase targets, pIC50 classification thresholds were defined using Youden's J statistic (mean J = 0.762, range 0.679–0.918). The 108 NSCLC phenotype models showed limited discriminative performance (mean AUC = 0.550). After applicability domain filtering, the increased mean AUC allowed the use of the phenotype models as a confidence-weighted filter in the generative workflow. Post-hoc similarity distribution analysis confirmed that the generated candidates represent structurally novel yet domain-consistent chemotypes, as evidenced by peak densities situated within the Tanimoto similarity range of 0.20–0.40 (generated vs known). Notably, PMPO significantly enhanced multi-parameter consistency (CV: 5.1% vs. 16.5%) and selectivity control (p < 0.0001) without compromising primary target potency. A central achievement of this work is the observation of independent scaffold convergence across diverse optimization settings, providing orthogonal validation of the identified quinoxaline core as a robust RET-inhibitory chemotype. After the common post-processing phases, including hinge-core analysis, scaffold novelty assessment, drug-likeness filtering, and drug-target affinity prediction, RL-1185 compounds were identified and prioritized as a lead molecule exhibiting high potency against multiple RET alterations (IC50: 0.78 nM for RETV804M, 142 nM for RETG810R), the novelty (Tanimoto coefficient: 0.579), and favorable synthetic tractability. Scientific contribution This work provides three key contributions: (1) Establishment of a multi-layered technical safeguard framework that ensures computational reliability and structural novelty through training constraints and post hoc validation. (2) A reward-design strategy integrating target potency, predicted off-target selectivity, and an exploratory phenotypic baseline constraint establishing a general RL-based inverse design framework in which predictive models both rank generated molecules, and shape the explored chemical space. (3) Experimental evaluation of computationally prioritized compounds identified a novel aminoquinoxaline scaffold with sub-nanomolar inhibitory activity against the RET wild-type and gatekeeper mutant (IC50 < 1 nM).