Integrating Machine Learning-Based Molecular Design with Experimental Validation for the Discovery of EGFR Inhibitors in Lung Cancer

The emergence of drug resistance and off-target toxicities in epidermal growth factor receptor (EGFR) targeted therapies underscores the urgent need for novel inhibitor scaffolds. This study integrates artificial intelligence-driven generative models with experimental validation to discover novel, selective EGFR inhibitors. Utilizing REINVENT4, a reinforcement learning-based generative framework, we performed a stage-wise, multi-objective optimization using a curated dataset of active EGFR inhibitors. The optimization was guided by a composite reward function incorporating docking scores, and quantitative estimates of drug-likeness (QED). Candidate molecules were subsequently evaluated using molecular dynamics (MD) simulations, synthesized, and subjected to in vitro kinase and cellular assays. The generative pipeline successfully converged on a promising N-(quinolin-5-yl) benzenesulfonamide scaffold. Among the synthesized candidates, Hit1 (4-pyridyl derivative) exhibited potent in vitro EGFR kinase inhibition (IC50 = 21.22 nM), comparable to the approved drug Gefitinib. MD simulations analyses revealed that hydrogen bond interactions with Lys745 and proper occupation of the Val726 hydrophobic cavity are critical for binding. Notably, Hit1 demonstrated robust, targeted anti-proliferative activity against EGFR-mutant non-small cell lung cancer (NSCLC) cells (PC9 and HCC827), while displaying greater than 1000-fold selectivity over wild-type EGFR cells (A549). Our findings validate the efficacy of a target-aware reinforcement learning approach for de novo drug design. The discovered quinoline-sulfonamide derivative represents a highly promising, synthetically tractable lead compound for the development of next-generation mutation-selective EGFR inhibitors.