3D Molecule Generation from Rigid Motifs via SE(3) Flows
Organizations: School of Computation, Information and Technology, Technical University of Munich, Germany · Munich Center for Machine Learning (MCML), Munich, Germany · Munich Data Science Institute (MDSI), Technical University of Munich, Germany
Abstract
Three-dimensional molecular structure generation is typically performed at the level of individual atoms, yet molecular graph generation techniques often consider fragments as their structural units. Building on the advances in frame-based protein structure generation, we extend these fragmentation ideas to 3D, treating general molecules as sets of rigid-body motifs. Utilising this representation, we employ SE(3)-equivariant generative modelling for de novo 3D molecule generation from rigid motifs. In our evaluations, we observe results competitive with state-of-the-art bond-implicit methods, surpassing them in atom stability on GEOM-Drugs, while yielding a 2x to 10x reduction in generation steps and 3.5x compression in molecular representations compared to standard atom-based methods. We release our code at https://github.com/rpoletukhin/MotiFlow.
Figures & tables
| Method | Steps | Stability ( ) | Val. / Uniq. ( ) | ||
| A, % | M, % | V, % | V U, % | ||
| Training Data | – | ||||
| EDM Hoogeboom et al. [ 2022 ] | |||||
| EDM-Bridge Wu et al. (2022) | |||||
| GeoLDM Xu et al. (2023) | 98.9_{\text{\tiny\pm.1}} | 89.4_{\text{\tiny\pm.5}} | 93.8_{\text{\tiny\pm.4}} | 92.7_{\text{\tiny\pm.5}} | |
| EquiFM Song et al. (2023) | 98.9_{\text{\tiny\pm.1}} | 88.3_{\text{\tiny\pm.3}} | 94.7_{\text{\tiny\pm.4}} | 93.5_{\text{\tiny\pm.3}} | |
Appendix figures & tables15 assets
Supplementary material from the paper’s appendix.
Appendix
| Approach | De novo fragment-based generation in 3D | Deterministic fragment-to-atom lift in 3D |
| Fragment-based graph generation Jin et al. [2018] , Jin et al. [2020] , Maziarz et al. [2022] , Geng et al. [2023] | ||
| HierDiff Qiang et al. [2023] | ||
| SigmaDock Prat et al. [2026] | ||
| Pocket-conditioned autoregressive methods Ghorbani et al. [2023] , Li et al. [2024] | ||
| Decomposition-guided pocket methods Guan et al. [2023] , Zhou et al. [2024] | ||
| MotiFlow |
| Method | Steps | Stability ( ) | Val. / Uniq. ( ) | TV ( ) | Str. En. ( ) | SA ( ) | QED ( ) | |||
| A, % | M, % | V, % | V U, % | A | B | |||||
| Training Data | – | – | – | |||||||
| Bond-explicit methods (for reference; not directly comparable) | ||||||||||
| SemlaFlow Irwin et al. [2025] | – | – | – | – | – | – | ||||
| FlowMol- CTMC Dunn and Koes [ 2024b ] | – | – | – | – | – | – | – | |||
| EQGAT-diff Le et al. [2024] | – | – | – | – | – | – | ||||
| Method | Steps | Stability | V C | TV ( ) | SA ( ) | QED ( ) |
| A, % ( ) | % ( ) | A | ||||
| Training Data | – | – | ||||
| Bond-explicit methods (for reference; not directly comparable) | ||||||
| SemlaFlow Irwin et al. [2025] | – | – | – | – | ||
| EQGAT-diff Le et al. [2024] | – | – | – | – | ||
| JODO Huang et al. [2024] | – | – | – | – | ||
| Temperature Schedule | Atom Stable, % | Molecule Stable, % | Valid, % | , % |
| (Constant) | 87.2_{\text{\tiny\pm0.2}} | |||
| (Original) | 99.1_{\text{\tiny\pm0.1}} | 92.6_{\text{\tiny\pm0.5}} | 95.3_{\text{\tiny\pm0.6}} | 86.3_{\text{\tiny\pm0.9}} |
| 99.0_{\text{\tiny\pm0.0}} | 91.4_{\text{\tiny\pm0.3}} | 94.6_{\text{\tiny\pm0.3}} | ||
| 98.5_{\text{\tiny\pm0.0}} | 87.5_{\text{\tiny\pm0.3}} | 92.7_{\text{\tiny\pm0.5}} | 90.4_{\text{\tiny\pm0.7}} |
| Method | Steps | Time, s ( ) |
| EDM Hoogeboom et al. [2022] | ||
| END Cornet et al. [2024] | ||
| Schedule | Atom Stability, % | Validity , % | Connectivity , % | , % |
| (1) | 86.2_{\text{\tiny\pm.3}} | 72.4_{\text{\tiny\pm.9}} | 62.4_{\text{\tiny\pm1.0}} | |
| (2) | 92.7_{\text{\tiny\pm.0}} | 77.9_{\text{\tiny\pm.3}} | 74.8_{\text{\tiny\pm0.4}} | |
| MotiFlow (Baseline) | 95.0_{\text{\tiny\pm.0}} | 93.1_{\text{\tiny\pm.2}} |
| Method | Steps | Composition | Substructure |
| Matching, ( ) | Tanimoto Sim. ( ) | ||
| cEDM Bao et al. [2023] | – | .671_{\text{\tiny\pm.004}} | |
| EEGSDE Bao et al. [2023] | – | .750_{\text{\tiny\pm.003}} | |
| cEDM * Cornet et al. [2024] | 69.6_{\text{\tiny\pm0.6}} | .601_{\text{\tiny\pm.000}} | |
| 73.0_{\text{\tiny\pm0.6}} | .640_{\text{\tiny\pm.002}} | ||
| 74.1_{\text{\tiny\pm1.4}} | .663_{\text{\tiny\pm.002}} |
| Dataset | Total Atoms | Heavy Atoms | ||||
| Mean | Median | Max | Mean | Median | Max | |
| GEOM-Drugs | ||||||
| QMugs | ||||||
| Dataset | Fragmentation | Cut-off | Fragments per molecule | Max. size | ||
| Mean | Median | Max | ||||
| GEOM-Drugs | No Rings | – | ||||
| Planar Rings | ||||||
| Planar Rings | ||||||
| Planar Rings | ||||||
| QMugs | No Rings | – | ||||
| Fragment. Strategy | Steps | QMugs | GEOM-Drugs | ||||||||
| Stabil. | V C | SA ( ) | QED ( ) | TV ( ) | Stabil. | V C | SA ( ) | QED ( ) | TV ( ) | ||
| A, % ( ) | % ( ) | A | A, % ( ) | % ( ) | A | ||||||
| No Rings | 86.5_{\text{\tiny\pm0.6}} | 83.0_{\text{\tiny\pm.2}} | 0.74_{\text{\tiny\pm.00}} | 0.51_{\text{\tiny\pm.00}} | 5.12_{\text{\tiny\pm.06}} | 89.1_{\text{\tiny\pm0.7}} | 80.9_{\text{\tiny\pm.2}} | 0.71_{\text{\tiny\pm.00}} | 0.69_{\text{\tiny\pm.00}} | 3.57_{\text{\tiny\pm.05}} | |
| 85.5_{\text{\tiny\pm1.8}} | 85.6_{\text{\tiny\pm.3}} | 0.74_{\text{\tiny\pm.00}} | 0.50_{\text{\tiny\pm.00}} | 4.98_{\text{\tiny\pm.06}} | 86.7_{\text{\tiny\pm1.8}} | 82.9_{\text{\tiny\pm.4}} | 0.71_{\text{\tiny\pm.00}} | 0.69_{\text{\tiny\pm.00}} | 3.36_{\text{\tiny\pm.08}} | ||
| Planar Rings | 95.1_{\text{\tiny\pm0.1}} | 82.2_{\text{\tiny\pm.8}} | 0.72_{\text{\tiny\pm.00}} | 0.60_{\text{\tiny\pm.00}} | 8.56_{\text{\tiny\pm.04}} | 94.9_{\text{\tiny\pm0.0}} | 81.3_{\text{\tiny\pm.4}} | 0.70_{\text{\tiny\pm.00}} | 0.71_{\text{\tiny\pm.00}} | 4.47_{\text{\tiny\pm.02}} | |
| 95.5_{\text{\tiny\pm0.1}} | 84.0_{\text{\tiny\pm.5}} | 0.72_{\text{\tiny\pm.00}} | 0.60_{\text{\tiny\pm.00}} | 8.08_{\text{\tiny\pm.17}} | 95.2_{\text{\tiny\pm0.1}} | 82.1_{\text{\tiny\pm.3}} | 0.70_{\text{\tiny\pm.00}} | 0.71_{\text{\tiny\pm.00}} | 4.24_{\text{\tiny\pm.14}} | ||
| Method | Steps | Stability ( ) | Val. / Uniq. ( ) | TV ( ) | Str. En. ( ) | SA ( ) | QED ( ) | |||
| A, % | M, % | V, % | V U, % | A | B | |||||
| Data | – | – | – | |||||||
| Uniform Prior | 98.7_{\text{\tiny\pm.1}} | 90.8_{\text{\tiny\pm.1}} | 95.2_{\text{\tiny\pm.5}} | 89.2_{\text{\tiny\pm.5}} | 1.42_{\text{\tiny\pm.06}} | 12.2_{\text{\tiny\pm.6}} | 14.3_{\text{\tiny\pm.5}} | 0.70_{\text{\tiny\pm.00}} | 0.47_{\text{\tiny\pm.00}} | |
| 98.7_{\text{\tiny\pm.1}} | 90.4_{\text{\tiny\pm.4}} | 93.5_{\text{\tiny\pm.3}} | 87.5_{\text{\tiny\pm.4}} | 1.38_{\text{\tiny\pm.09}} | 15.2_{\text{\tiny\pm.2}} | 10.7_{\text{\tiny\pm.1}} | 0.70_{\text{\tiny\pm.00}} | 0.47_{\text{\tiny\pm.00}} | ||
| No | 98.6_{\text{\tiny\pm.0}} | 89.5_{\text{\tiny\pm.2}} | 95.5_{\text{\tiny\pm.1}} | 88.7_{\text{\tiny\pm.5}} | 2.35_{\text{\tiny\pm.05}} | 4.9_{\text{\tiny\pm.5}} | 16.5_{\text{\tiny\pm.2}} | 0.70_{\text{\tiny\pm.00}} | 0.47_{\text{\tiny\pm.00}} | |
| 98.9_{\text{\tiny\pm.0}} | 90.8_{\text{\tiny\pm.2}} | 94.5_{\text{\tiny\pm.1}} | 88.0_{\text{\tiny\pm.4}} | 2.35_{\text{\tiny\pm.08}} | 6.4_{\text{\tiny\pm.5}} | 11.8_{\text{\tiny\pm.1}} | 0.70_{\text{\tiny\pm.00}} | 0.47_{\text{\tiny\pm.00}} | ||
3D Molecular Generation
3D molecular generation creates molecules by modeling their chemical structures and three-dimensional geometries.
See all 61 papers