Structure-free, site-resolved contrastive learning extends small-molecule discovery beyond the reach of structure-based modeling

Abstract

Virtual screening asks which molecules, among an enormous space of drug-like chemistry, are worth synthesizing and testing against a protein target. Most modern methods answer by building and scoring an explicit three-dimensional pose through molecular docking, or the co-folding models that now approach experimental accuracy. Building these poses presumes a well-defined pocket, and the non-orthosteric, cryptic, and intrinsically disordered sites where unexplored ligandability lies offer none. There, these methods fail to generalize. Here we present Ptarmigan-1, a contrastive model that co-embeds the residues of a protein with candidate small molecules in a shared latent space, from sequence and two-dimensional chemistry alone, and without ever constructing a pose. Engagement reduces to the proximity of precomputed embeddings. Freed from the pose, Ptarmigan-1 trains directly on chemoproteomic and bioactivity data of mixed resolution, scores a compound in ten milliseconds rather than the tens of seconds a co-folding model demands, and resolves each prediction to the residues a compound engages. On well-folded, orthosteric targets it performs comparably to a collection of co-folding and docking models, and on covalent, cryptic, and disordered sites it matches or exceeds them. It localizes reversible and covalent inhibitors to the pockets they engage, even for targets withheld from training, and screens the entire human proteome against a library of 3.4 billion compounds in under a day. By decoupling molecular recognition from structure, Ptarmigan-1 recasts virtual screening as a reusable index that continuously improves as data accumulate.