Matcha: Multi-Stage Riemannian Flow Matching for Accurate and Physically Valid Molecular Docking

TL;DR

Matcha tackles the protein–ligand docking problem by introducing a three-stage Riemannian flow matching pipeline that refines poses in $\mathbb{R}^3$, $\mathrm{SO}(3)$, and $\mathrm{SO}(2)$, complemented by a scoring model and unsupervised physical validity filtration. It achieves high docking accuracy and physical plausibility while delivering substantially faster inference than contemporary co-folding models (approximately $25\times$ faster). Across benchmarks such as Astex, PDBBind, and PoseBusters V2, Matcha attains state-of-the-art results on certain datasets and competitive performance on others, underscoring the value of integrating geometric flow with learned scoring and validity checks. While its generalization to unseen pockets lags compared to larger co-folding models, the approach provides a scalable, geometry-aware framework that balances accuracy, efficiency, and physical realism for structure-based drug discovery.

Abstract

Accurate prediction of protein-ligand binding poses is crucial for structure-based drug design, yet existing methods struggle to balance speed, accuracy, and physical plausibility. We introduce Matcha, a novel molecular docking pipeline that combines multi-stage flow matching with learned scoring and physical validity filtering. Our approach consists of three sequential stages applied consecutively to refine docking predictions, each implemented as a flow matching model operating on appropriate geometric spaces ($\mathbb{R}^3$, $\mathrm{SO}(3)$, and $\mathrm{SO}(2)$). We enhance the prediction quality through a dedicated scoring model and apply unsupervised physical validity filters to eliminate unrealistic poses. Compared to various approaches, Matcha demonstrates superior performance on Astex and PDBbind test sets in terms of docking success rate and physical plausibility. Moreover, our method works approximately 25 times faster than modern large-scale co-folding models. The model weights and inference code to reproduce our results are available at https://github.com/LigandPro/Matcha.

Figures (20)

• Figure 1: Matcha consists of three flow matching models that generate a set of ligand poses. After unsupervised physical validity filtration, all poses are scored with a separate model, and the prediction is the pose with the best score. We highlight rotatable bonds and their periods in the ligand on this figure.
• Figure 2: The architecture of the velocity model of Matcha (stages 1, 2, 3).
• Figure 3: Blind ligand docking success rates on Astex Diverse Set ($n=85$).
• Figure 4: Blind ligand docking success rates on PDBBind test set ($n=363$).
• Figure 5: Blind ligand docking success rates on PoseBusters V2 dataset ($n=308$).