Simpatico: accurate and ultra-fast virtual drug screening with atomic embeddings

Abstract Virtual screening, the in-silico assessment of large libraries of small molecules for binding to a therapeutic protein target, is a critical early step in drug discovery. The dominant approach, molecular docking, requires a separate calculation for each protein-molecule pair, and is too slow to apply alone at the billion-compound scale of modern compound libraries. A recent embedding-retrieval paradigm, exemplified by DrugCLIP, addresses this bottleneck by training deep models to map proteins and small molecules into a shared embedding space, such that proteins are co-located with their likely binding partners; candidate ligands can then be retrieved directly by nearest-neighbors search, with no per-pair calculation. However, current embedding-retrieval methods collapse each protein and each ligand into a single embedding, creating an information bottleneck that limits the representation of partial or alternative binding compatibility. We present simpatico, an embedding-retrieval virtual screening tool that instead produces a unique embedding for each atom in a protein pocket or ligand. A CLIP-style contrastive objective trains these atomic embeddings so that protein-ligand atom pairs known to interact are nearby in embedding space. To screen a protein target, each protein-atom embedding is used as a query against a vector database of precomputed small-molecule atomic embeddings, returning the closest atoms in the library; a simple aggregation step assigns a binding score to each candidate molecule containing retrieved atoms. Where prior retrieval-based methods index one vector per ligand, simpatico indexes one per heavy atom; query time grows sublinearly in library size. On challenging decoy benchmarks, simpatico achieves state-of-the-art predictive accuracy, outperforming recent dense-retrieval methods despite training on only ∼15,000 protein-ligand complexes from PDBBind, with no pretraining and no 3D ligand pose estimation. Simpatico also exceeds the accuracy of physics-based docking and deep-learning-augmented docking methods, is competitive with diffusion-based docking, and runs orders of magnitude faster than all three. Simpatico is open source software; all code, weights, and data may be accessed at https://github.com/TravisWheelerLab/Simpatico. Competing Interest Statement The authors have declared no competing interest. Footnotes This revision represents a substantial update to the underlying model, which now implements the CLIP loss function and a novel hard mining strategy. We have also implemented additional experiments and substantially revised the manuscript text.