An integrated graph neural network and structure-based workflow for the discovery of allosteric PCSK9 inhibitors

Proprotein convertase subtilisin/kexin type 9 (PCSK9) plays a central role in cholesterol homeostasis by promoting the degradation of low-density lipoprotein receptors (LDLR), making it an attractive therapeutic target for cardiovascular disease. Although monoclonal antibodies and RNA-based therapies targeting PCSK9 have shown clinical success, developing orally available inhibitors remains a major challenge, largely because of the complex protein-protein interaction landscape of PCSK9. In this study, we present an integrated artificial intelligence-driven drug discovery framework that combines a graph neural network (GNN) model with structure-based molecular modeling to identify and characterize novel allosteric PCSK9 inhibitors. A curated dataset of 1,039 PCSK9 ligands with experimentally measured IC₅₀ values was compiled from multiple public databases and used to train an optimized message-passing GNN. The resulting model demonstrated excellent predictive performance (R² = 0.949 on the test set) and robust generalization within a well-defined applicability domain. To translate predictive insights into mechanistic understanding, top-ranked compounds were subjected to a multistage structural workflow including blind and targeted docking, pharmacophore modeling, molecular dynamics simulations, MM/GBSA binding free-energy calculations, and protein-protein interaction analysis of the PCSK9-LDLR interface. Among the identified candidates, Ligand40 exhibited the most favorable binding energetics, while Ligand42 demonstrated stable interaction patterns and enhanced dynamic restraint, supporting its prioritization for further investigation. Overall, this work demonstrates that coupling deep graph learning with physics-based simulations can facilitate the discovery of candidate PCSK9 modulators. The proposed framework is scalable and transferable, enabling AI-assisted discovery of small-molecule and peptide modulators for challenging protein-protein interaction targets.