Simpatico

Virtual screening asks a simple question at an awkward scale: out of a billion purchasable compounds, which few thousand are worth ordering for a given protein target? Molecular docking answers it one protein-ligand pair at a time, which makes it accurate enough to trust and far too slow to run alone at library scale. The embedding-retrieval alternative, popularized by DrugCLIP, trains a model to place proteins and their likely binders near each other in a shared vector space so that candidates can be pulled out by nearest-neighbor search with no per-pair computation at all. Its weakness is compression: collapsing a whole pocket and a whole molecule into one vector each discards the partial, fragment-level complementarity that actually drives binding.

Simpatico, from Travis Wheeler's lab at the University of Arizona, keeps the retrieval paradigm and removes the bottleneck by embedding every heavy atom separately. A CLIP-style contrastive objective pulls protein-atom and ligand-atom embeddings together when the corresponding atoms are in contact in a crystal structure, so the index holds one vector per heavy atom rather than one per molecule. Screening a target means querying each pocket-atom embedding against that index and aggregating the retrieved atoms into a per-molecule score; because each query returns a fixed number of neighbors, wall-clock time grows sublinearly with library size.

The result is a model trained from scratch on roughly 19,000 crystal structures that outperforms both dense-retrieval and docking baselines on standard decoy benchmarks while running orders of magnitude faster. It was posted to bioRxiv in June 2025 with code, weights, and data released under a BSD-3-Clause license.

#Key Features

• Atom-level rather than molecule-level embeddings: Each heavy atom of a pocket or ligand gets its own vector, preserving partial and alternative binding compatibility that single-vector representations average away.
• Sublinear screening time: Retrieval against a FAISS index returns a fixed number of ligand atoms per protein-atom query, so query cost does not scale linearly with the number of molecules indexed.
• No pose estimation required: Scores come from embedding proximity alone, with no 3D ligand conformer generation, no docking calculation, and no structure prediction step in the loop.
• Trained without pretraining: The contrastive objective is learned directly from crystal-structure contacts rather than fine-tuned from a larger pretrained backbone, which keeps the training recipe self-contained.
• Rigorous homolog removal: Training complexes sharing high sequence identity with any benchmark target are dropped before evaluation, so reported enrichment is not inflated by memorized targets.

#Technical Details

The architecture is a graph neural network of residually stacked GATv2 attention blocks with SiLU nonlinearities, producing L2-normalized 64-dimensional embeddings so that inner products correspond to distances. Pocket graphs keep atoms within 6 Å of any ligand atom as surface atoms, trim anything more than 8 Å beyond them, connect atoms within 5 Å, and append virtual nodes per surface atom; ligand graphs use heavy atoms with one-hot species and hydrogen counts, connected out to three covalent bonds. Contrastive training pairs a protein atom with an interacting ligand atom (≤ 4 Å), draws intra-molecular negatives from atoms more than 6 Å away, and adds hard inter-molecular negatives chosen by a ranking curriculum.

Training draws on 19,443 PDBBind complexes spanning 3,876 distinct proteins, filtered per benchmark by MMseqs2 to remove homologs at 90% identity — leaving 15,271 complexes for the DUD-E evaluation. Simpatico reaches an enrichment factor at 1% of 30.37 on DUD-E's 102 targets and 23.38 on LIT-PCBA's 15 targets, against 20.44 and 15.59 for DrugCLIP, and exceeds AutoDock Vina, Gnina, and SurfDock across DEKOIS's 81 targets. Screening runs at roughly 14 seconds per million molecules, putting a 100-million-compound library at about 26 minutes on a single GPU.

#Applications

Simpatico is aimed at the first triage step of a small-molecule campaign: taking a protein structure with a defined pocket and reducing an ultra-large make-on-demand library to a shortlist small enough to dock properly or order outright. Because a compound library's atomic embeddings are computed once and reused across every subsequent target, the marginal cost of screening a new protein is low, which suits target panels, selectivity counter-screens, and repurposing sweeps.

#Impact

The entry demonstrates that the information bottleneck, not the retrieval paradigm, was limiting embedding-based screening — a finding that argues for finer-grained representations elsewhere in structure-based design. The authors are candid about what remains: an oracle analysis shows their simple summation aggregator captures only about 65% of the achievable performance on DUD-E and 16% on the harder LIT-PCBA, so scoring rather than retrieval is now the limiting step. Per-atom indexing also multiplies storage relative to per-molecule indexing, and the model returns no predicted pose, so downstream docking is still needed for structural interpretation. The work remains a preprint, and the repository is under active development.