Matcha

Matcha is a molecular docking model that predicts how a small-molecule ligand binds to a protein, producing poses that are both accurate and physically realistic. Docking is a foundational task in structure-based drug discovery, but existing deep learning approaches face a persistent trade-off: large co-folding models can be accurate yet are computationally expensive, while faster methods often generate poses with steric clashes or distorted geometry that violate basic physical constraints.

Matcha addresses this with a multi-stage Riemannian flow-matching pipeline that decomposes the docking problem across three geometric spaces. Separate flow-matching stages operate on R^3 for translational placement of the ligand, SO(3) for its rotational orientation, and SO(2) for internal torsional degrees of freedom. Handling each type of geometric variable on its natural manifold, in sequence, lets the model refine poses efficiently and accurately. A final GNINA-based energy minimization step and unsupervised physical-validity filters remove implausible poses.

Developed by researchers at the AIRI Institute and the Skolkovo Institute of Science and Technology and released as a preprint in October 2025, Matcha is distributed with pretrained weights and a command-line interface under a non-commercial license. It positions itself as a fast, physically grounded alternative to heavyweight co-folding docking systems.

#Key Features

• Multi-stage geometric flow matching: Matcha runs sequential flow-matching stages on R^3, SO(3), and SO(2), treating translation, rotation, and torsion on their natural manifolds.
• Physical validity by design: GNINA energy minimization and unsupervised validity filters remove poses with clashes or unrealistic geometry, improving physical plausibility over pure neural predictors.
• Large speedup over co-folding: It runs roughly 31x faster than modern large-scale co-folding models while remaining competitive on accuracy.
• Open weights and CLI: Pretrained checkpoints and inference tooling are released, enabling direct use in docking workflows.

#Technical Details

Matcha frames docking as a Riemannian flow-matching problem solved in stages: a translational flow on R^3 positions the ligand in the pocket, a rotational flow on SO(3) orients it, and a torsional flow on SO(2) adjusts internal rotatable bonds. The learned poses are then refined by GNINA energy minimization and screened by unsupervised physical-validity filters. The authors evaluate on standard docking benchmarks including Astex, PoseBusters, and PDBBind test splits (with DockGen among the referenced datasets), reporting superior docking success rate and physical plausibility alongside roughly 31x faster inference than large co-folding models. Pretrained checkpoints are provided through the Hugging Face repository, and the model is released under a non-commercial license.

#Applications

Matcha is aimed at computational chemists and drug-discovery researchers who need to predict protein-ligand binding poses at scale. Its speed makes it practical for high-throughput virtual screening and pose prediction across large compound libraries, while its physical-validity filtering makes the resulting poses more directly usable for downstream analysis, structure-based optimization, and prioritization of candidate ligands.

#Impact

Matcha shows that decomposing docking into staged flow-matching processes on distinct geometric manifolds, combined with explicit physical-validity filtering, can deliver both speed and realism, addressing a common failure mode of fast neural docking methods that produce physically implausible poses. Its large speedup over co-folding models makes it attractive for high-throughput settings. As a recent preprint awaiting peer review and released under a non-commercial license, its accuracy on novel targets and its comparison to the newest co-folding systems remain to be established.