Generative AI-Based Drug Design: Target-Aware Molecular Generation for EGFR Type I Inhibitors with Multiplatform Docking Validation.

The development of selective kinase inhibitors remains an ongoing challenge in the drug development process, particularly regarding resistance mutations in the epidermal growth factor receptor (EGFR). EGFR resistance mutations, particularly T790 M and C797S variants, pose significant challenges in oncology therapeutics, necessitating novel approaches to identify resistance-circumventing inhibitors. In this study, a target-conditioned generation framework is proposed, which incorporates ATP-binding pocket information into structural generation through structure-aware conditioning. A three-layer, bidirectional Long Short-Term Memory network is conditioned on 20 essential binding site residues from the EGFR crystal structure (PDB 1M17). Training data comprised 6,038 Type I EGFR inhibitors from the ChEMBL database, filtered for drug-like properties (pChEMBL ≥ 5.0, Lipinski compliance). From 1000 generated sequences, 82.6% were chemically valid, with 79.5% ECFP4 novelty versus the training set. Five prioritized candidates demonstrated predicted binding scores of -7.7 to -8.4 kcal/mol in AutoDock Vina, used here for computational candidate ranking, compared to -7.57 kcal/mol for the erlotinib redocking reference under the same protocol; all five candidates exhibited zero Lipinski violations. Cavity-based cross-checking using CB-Dock2 confirmed that all five molecules consistently targeted the ATP-binding pocket (C2 cavity), supporting target-site specificity of predicted binding modes (intertool score consistency