Single and Multi-Objective Optimization of De Novo Drug Design for Sophisticated Biomarkers, RET Alterations

Biomarker based oncology therapeutics leverage molecular signatures to identify patient populations most likely to benefit from a specific treatment, thereby increasing clinical success rates. Despite the significant merits of these biomarkers, the high rate of mutations across various therapeutic markers continues to impede innovative drug design. RET is a complex therapeutic biomarker, characterized by high alteration rates that complicate the development of targeted therapies in oncology. Herein, this study provides a systematic evaluation of single-objective (SO) and multi-objective (MO) reinforcement learning strategies aimed at the de novo design of inhibitors targeting complicated RET malignancies. Both approaches employed the identical downstream selection logic comprising hinge-core analysis, scaffold novelty assessment, drug-likeness filtering, and drug-target affinity (DTA) predictions using SMPLIP-Score and ChargeNET models. The SO approach optimized solely for RET-MUT potency, generating compounds with favorable synthesizability profiles (Tanimoto similarity 0.4–0.6). The MO approach incorporated 23 off-target kinase classification models and 108 NSCLC focused phenotype models; 85% of MO candidates simultaneously achieved high RET potency (pIC50 ≥ 7.0), favorable selectivity (off-target probability ≤ 0.3), and predicted phenotypic activity (NSCLC probability ≥ 0.3). Consequently, while SO optimization achieved rapid output with acceptable multi-parameter performance (off-target selectivity: 0.686 ± 0.126; phenotype: 0.945 ± 0.093), MO optimization provided enhanced selectivity consistency (0.555 ± 0.084, 3.2-fold reduced variability) without sacrificing target potency (mean pIC50 7.03 vs 7.6 for SO). The downstream selection and experimental validation further yielded novel amino quinoxaline derivatives to show the high potency against multiple RET alterations (IC ₅₀ : 0.78 nM for RET ^V804M , 142 nM for RET ^G810R , 0.53 nM for RET ^I788N ), representing an unseen scaffold for RET-driven cancers. Scientific Contribution This work provides three key contributions: (1) the first systematic comparison of single-objective versus multi-objective reinforcement learning for kinase inhibitor discovery using identical evaluation workflows; (2) integration of 23 off-target kinase and 108 phenotype classification models into generative molecular design, enabling prospective selectivity optimization; and (3) experimental validation of a novel amino quinoxaline scaffold with sub-nanomolar activity against clinically relevant RET resistance mutations (V804M, G810R, I788N).