AI-driven drug design for head and neck squamous cell carcinoma using GNN and structure-based simulations

Abstract

Head and neck squamous cell carcinoma (HNSCC) remains a major global health burden with limited targeted therapeutic options. Here we present an integrated artificial intelligence (AI)-driven drug discovery framework combining graph neural networks (GAT, GCN, AttentiveFP and GIN), classical machine learning models (Random Forest, XGBoost and SVM), and structure-based simulations to identify potential inhibitors associated with HNSCC. Specifically, the GCN model converts SMILES strings into molecular graphs, utilizing atom features (e.g., atomic number, hybridization) and bond features (e.g., bond type, conjugation) as inputs to learn graph-level representations. Using curated bioactivity datasets from ChEMBL and PubChem, we trained seven predictive models and constructed a consensus ensemble framework that achieved high precision, recall, and an AUROC exceeding 0.90 on independent validation datasets. Based on ensemble consensus scores from seven predictive models, large-scale virtual screening of the ChEMBL and DrugBank libraries yielded 36,771 and 391 high-confidence hits, respectively, with chemical space and scaffold analyses revealing substantial structural diversity and significant overlap with DrugBank (sharing 85.9% of its scaffolds), demonstrating effective coverage of drug-relevant chemical space. Through HNSCC-associated gene screening, PPI network analysis, MCC-based hub gene ranking, and machine learning-based feature selection, PIK3CA was prioritized as a key therapeutic target for subsequent validation. Molecular docking and 200-ns molecular dynamics simulations further suggested stable binding interactions for several top candidates, including CHEMBL4454174, CHEMBL4458803, and CHEMBL4591345. In addition, clinical tissue validation showed increased PI3Kα, p-AKT, and p-S6 expression in HNSCC tumor tissues compared with adjacent tissues, supporting activation of the PI3Kα/AKT/S6 signaling axis. Functional validation using siRNA-mediated PIK3CA knockdown further indicated reduced PI3Kα expression, attenuated AKT/S6 signaling, and impaired cell viability, proliferation, and migration in HNSCC cells. These findings highlight the potential of integrating AI-based prediction, structural validation, and biological validation to accelerate the discovery of candidate therapeutics for precision oncology.