Matcha: Multi-Stage Riemannian Flow Matching for Accurate and Physically Valid Molecular Docking
Authors: Daria Frolova, Talgat Daulbaev, Egor Sevriugov, Sergei A. Nikolenko, Dmitry N. Ivankov, Ivan Oseledets, Marina A. Pak
Organizations: Ligand Pro, Moscow, Russia · Skolkovo Institute of Science and Technology, Artificial Intelligence Center, Moscow, Russia · Skolkovo Institute of Science and Technology, Center for Molecular and Cellular Biology, Moscow, Russia · Artificial Intelligence Research Institute, Moscow, Russia
Predicting a protein--ligand binding pose is central to structure-based drug design, yet current docking methods still struggle to combine speed, accuracy and physical plausibility. We introduce Matcha, which refines a ligand pose by three flow-matching stages in translation, rotation and torsion space, then minimizes, filters and ranks candidates with GNINA. Across the benchmarks considered, Matcha attains the strongest physical plausibility and runs about 28× faster than modern large-scale co-folding models. Code and benchmarking artifacts are available at https://github.com/LigandPro/Matcha and archived at https://doi.org/10.5281/zenodo.23568093. Model checkpoints are archived at https://doi.org/10.5281/zenodo.23569055.
Figures & tables
Figure 1: Matcha overview. (a) Multi-stage docking pipeline. Matcha uses three flow-matching stages to generate candidate ligand poses, followed by GNINA minimization, unsupervised plausibility filtering and GNINA -based pose selection. Rotatable bonds and their periods define the torsional component of the generative state space. (b) Architecture of the Matcha velocity model used in stages 1–3. Ligand atoms, protein residues and dedicated translation and rotation tokens are processed by a transformer with distance-aware attention biases, followed by specialized heads for translation, rotation and torsion prediction.
Figure 2: Blind ligand docking performance across benchmarks. (a) Astex diverse set ( n=85 ). (b) PoseBusters V2 benchmark set ( n=308 ). (c) PDBBind test set ( n=363 ). (d) DockGen dataset ( n=330 ). For Matcha , bars show the mean over five inference seeds, and the error bars are one standard deviation across those seeds. The interval to the left of each Matcha bar is the variability of the PB-valid rate; the interval to the right is the variability of the RMSD≤2\text{,}\mathrm{\text{\AA}}$$ rate.
Figure 3: Pocket-aware ligand docking performance. Docking success is evaluated with the true binding-site center provided, separating pose-generation accuracy from pocket-localization accuracy. (a) Astex diverse set. (b) PDBBind test set. (c) PoseBusters V2 . (d) DockGen . Success is reported as the percentage of predictions with symmetry-corrected heavy-atom RMSD ≤2\text{,}\mathrm{\text{\AA}}$$ and, where indicated, satisfying the PoseBusters validity criteria.
Figure 4: Average inference time on Astex and RMSD≤2\text{,}\mathrm{\text{\AA}}$$ & PB-valid success on PoseBusters V2 .
Figure 5: Ablation and sampling analyses of Matcha . Blind docking performance is evaluated after progressively varying generation stages, post-processing steps and the number of sampled poses. (a–c) Stage-wise performance on Astex , PDBBind and PoseBusters V2 , respectively. (d–f) Contribution of post-processing steps on the same benchmarks, respectively. (g) On Astex , full Matcha (blind and pocket-aware), the same poses ranked by the GNINA CNN score without validity filtering, and GNINA docking from Stage 1 or from a pocket center. (h) Sample-size dependence on PoseBusters V2 . Success is reported as the percentage of predictions with symmetry-corrected heavy-atom RMSD ≤2\text{,}\mathrm{\text{\AA}}$$ and, where indicated, satisfying the PoseBusters validity criteria.
Appendix figures & tables13 assets
Supplementary material from the paper’s appendix.
Appendix
Method
Time (s)
Matcha latte
8.5
Matcha
13.6
DiffDock-L
32
NeuralPlexer
65
FlowDock
39
GNINA (P2Rank) for one initial pose
18
Appendix
Table 1: Comparison of the average inference time for blind docking models (for Astex test set)
Figure 6: Blind ligand docking success rates with different pocket prediction methods on Astex Diverse set.
Figure 7: Blind ligand docking success rates with different pocket prediction methods on PDBBind test set.
Figure 8: Blind ligand docking success rates with different pocket prediction methods on PoseBusters V2 dataset.
Figure 9: Blind ligand docking success rates with different pocket prediction methods on DockGen dataset.
Figure 10: Comparison of full and pocket-based alignment strategies in blind docking for Astex Diverse set.
Figure 11: Comparison of full and pocket-based alignment strategies in blind docking for PDBBind test set.
Figure 12: Comparison of full and pocket-based alignment strategies in blind docking for PoseBusters V2 dataset.
Figure 13: Comparison of full and pocket-based alignment strategies in blind docking for DockGen dataset.
Figure 14: Blind ligand docking success rates on the PoseBusters V2 dataset for the full benchmark and the temporally held-out subset.
Figure 15: Effect of the number of Riemannian Euler integration steps. We evaluate the inference-time integration budget by varying the number of Euler steps used by Matcha . (a–c) Blind docking success rates as a function of the number of Euler steps on Astex , PDBBind and PoseBusters V2 , respectively. Success is reported as the percentage of predictions with symmetry-corrected heavy-atom RMSD ≤2\text{,}\mathrm{\text{\AA}}$$ and, where indicated, satisfying the PoseBusters validity criteria.
Figure 16: Effect of the Matcha post-processing module applied to the baseline methods. (a) Astex . (b) PoseBusters V2 . (c) PDBBind . (d) DockGen .
Figure 17: Speed–accuracy trade-off on PoseBusters V2 when every method receives the same Matcha post-processing module ( Minimized & scored success rate; see Figs. 16 – 16 ). Complexes per hour is computed on one NVIDIA A100 40GB GPU.
Department of Bioengineering and Therapeutic Sciences, Schools of Pharmacy and Medicine, University of California, San Francisco, San Francisco, CA, USA · UCSF–Stanford Center of Excellence in Regulatory Science and Innovation (CERSI), San Francisco, CA, USA