Knowledge-graph attention network ChemFlow identifies modulators of EGFR

Knowledge-graph drug-target interaction (DTI) models usually assume that test compounds are already represented in the graph, which limits their use for external chemical libraries. Here we present ChemFlow, which addresses this gap by encoding drugs from molecular structure with GATv2 and proteins from fold-specific BioKG RotatE embeddings, preserving protein-side biological context without requiring compound-side graph connectivity. We evaluate ChemFlow with node-masked held-out-compound splits, where test drugs and all incident graph edges are removed before training. Across BioKG, Therapeutic Target Database (TTD) Approved, TTD Approved + Phase, and post-2020 validation, ChemFlow improves precision-recall performance relative to sequence-based, descriptor-based, and docking baselines, including under class-imbalance settings. We then use ChemFlow to screen MedChemExpress compounds against epidermal growth factor receptor (EGFR). Model scoring, docking, and molecular dynamics prioritize allitinib, SYHA-1813, and GNF-2. Cellular thermal shift assay (CETSA) and EGFR phosphorylation assays support direct target engagement and separate inhibitor-like from agonist-like signaling effects. ChemFlow therefore connects drug-inductive DTI prediction with experimentally testable compound prioritization.