Target Preference Maps: A machine learning model generalizing transferable drug-receptor interactions and guiding drug discovery

Abstract

Modern AI models can decode the genomic landscape and protein structure world. Yet, they fail to generalize to one of the most important fields: small-molecule drug discovery. Since the late 1970s, the advent of macromolecular crystallography inspired the notion that structural knowledge alone could enable a "lock-and-key" approach to drug design. However, drug discovery continues to depend on costly, resource-intensive, and largely serendipitous screening campaigns that probe only an infinitesimal fraction of the drug-like chemical space. Despite some successful cases, our understanding of, and reasoning from, non-bonded interaction chemistry remains limited for general applicability. Furthermore, though structural databases contain hundreds of thousands of entries, a strong historical bias pervades protein-drug structures, hindering reliable advances through AI scaling. Here, we present a machine-learning framework that learns atom-type-specific spatial preference maps from local protein microenvironments in protein-ligand structures. By excluding ligand topology from the model input and learning from local atom-level environments, the framework is designed to reduce dependence on whole-ligand memorization and to capture transferable interaction preferences. The resulting maps recover chemically meaningful interaction patterns, including cases involving bridging waters and metal-dependent environments. The model was validated using retrospective and prospective real-world data in drug optimization when targeting a challenging protein-protein interface. This shows that the method can provide interpretable workflows to guide molecule optimization and provide input for downstream generative or docking workflows.