FlexiFlow: decomposable flow matching for generation of flexible molecular ensemble
Riccardo Tedoldi, Ola Engkvist, Patrick Bryant, Hossein Azizpour, Jon Paul Janet & Alessandro Tibo · 2025
Read FlexiFlow: decomposable flow matching for generation of flexible molecular ensemble by Riccardo Tedoldi, Ola Engkvist, Patrick Bryant, Hossein Azizpour, Jon Paul Janet & Alessandro Tibo (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2511-17249.txt.
Introduction
Riccardo Tedoldi and five co-authors present FlexiFlow, a generative model for drug discovery that designs a molecule and its shapes together. A molecule's three-dimensional conformations govern its properties and how tightly it can bind a protein target, yet leading design models produce only a single conformation. FlexiFlow extends flow matching so it samples a molecule jointly with an ensemble of its low-energy shapes, while respecting the symmetries the physics demands.
On the QM9 and GEOM Drugs benchmarks, the model generated valid, unstrained, novel molecules that matched the training distribution closely. Its conformational ensembles covered the relevant shapes about as well as established physics-based methods, at a fraction of the computing time. The model also transferred to designing ligands for a given protein pocket, even when training data lacked conformations for that task.
Catalog record
| CREATOR | Riccardo Tedoldi, Ola Engkvist, Patrick Bryant, Hossein Azizpour, Jon Paul Janet & Alessandro Tibo |
|---|---|
| YEAR | 2025 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 85 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2511-17249
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
FlexiFlow: decomposable flow matching for generation of flexible molecular ensemble Riccardo Tedoldi, Ola Engkvist, Patrick Bryant, Hossein Azizpour, Jon Paul Janet & Alessandro Tibo Abstract Sampling useful three-dimensional molecular structures along with their most favorable conformations is a key challenge in drug discovery. Current state-of-the-art 3D de-novo design flow matching or diffusion-based models are limited to generating a single conformation. However, the conformational landscape of a molecule determines its observable properties and how tightly it is able to bind to a given protein target. By generating a representative set of low-energy conformers, we can more directly assess these properties and potentially improve the ability to generate molecules with desired thermodynamic observables. Towards this aim, we propose FlexiFlow , a novel architecture that extends flow-matching models, allowing for the joint sampling of molecules along with multiple conformations while preserving both equivariance and permutation invariance. We demonstrate the effectiveness of our approach on the QM9 and GEOM Drugs datasets, achieving state-of-the-art results in molecular generation tasks. Our results show that FlexiFlow can generate valid, unstrained, unique, and novel molecules with high fidelity to the training data distribution, while also capturing the conformational diversity of molecules. Moreover, we show that our model can generate conformational ensembles that provide similar coverage to state-of-the-art physics-based methods at a fraction of the inference time. Finally, FlexiFlow can be successfully transferred to the protein-conditioned ligand generation task, even when the dataset contains only static pockets without accompanying conformations. 1 Introduction Flow matching Lipman et al. (2023) and diffusion models ( Ho et al., 2020 ) now deliver state-of-the-art generation across images, audio, and 3D shapes ( Yang et al., 2024 ) , enabled by strong theory and flexible density modeling. Recent work sharpens flow matching for higher fidelity ( Domingo-Enrich et al., 2025 ) , compositional generation ( Skreta et al., 2025 ) , and faster, more stable, or guided sampling ( Liu et al., 2025 ) . Along these lines, we introduce a conditional flow decomposition framework that enables the joint generation of molecular graphs and multiple conformers. Figure 1: From noise samples, our model generates both molecular graphs and their conformational ensembles. Recent diffusion and flow-matching models have advanced de novo molecular generation ( Hoogeboom et al., 2022 ; Schneuing et al., 2024 ) , protein design ( Watson et al., 2023 ; Anand and Achim, 2022 ) , and protein structure prediction ( Jumper et al., 2021 ; Ingraham et al., 2019 ) , leveraging E(3)/SE(3)-equivariant architectures to capture 3D symmetries. Their application to unconditional 3D small-molecule generation is highly promising ( Vignac et al., 2023 ; Le et al., 2024 ; Irwin et al., 2025 ) , reaching benchmark ceilings on QM9 ( Ramakrishnan et al., 2014 ) and GEOM Drugs ( Axelrod and Gómez-Bombarelli, 2022 ) . However, current models typically produce only a single conformer per molecule, limiting conformational diversity critical for drug discovery, as different conformations can exhibit significantly different biological activities and properties. For a specific target protein, sampling multiple ligand conformations to match distinct interaction patterns is crucial for optimizing binding affinity and selectivity of a drug candidate ( Leach et al., 2009 ) . To address this limitation of current models, we propose a novel framework that extents flow matching to jointly generate molecular graphs and representative sets of conformations. To this end, we introduce FlexiFlow, a novel architecture that uses this paradigm to handle two sets of coordinates. FlexiFlow preserves permutation invariance on categorical and equivariance on two sets of coordinates features, moreover, it leverages this decomposition to design novel molecular structures along with their 3D conformers (illustration in Figure 1 ). Our framework shows promising results in generating valid, unique, and novel molecular structures with realistic conformations. Unlike other approaches, like Bolzmann generators ( Noé et al., 2019 ; Diez et al., 2025 ) , FlexiFlow efficiently generates high-quality molecular structures and conformations while not requiring additional MD data for training and avoiding major computational and transferability limitations. The method can be successfully applied to other data modalities, where we present results on MNIST in Appendix E . We summarize the key contributions as follows: • We propose a novel framework leveraging conditional independence to decompose the flow matching objective, enabling simultaneous generation of a graph with a single reference conformation along with an arbitrary set of conformers. • We introduce a new architecture, FlexiFlow, that handles equivariance on two coordinate sets for 3D molecular generation. • We conduct extensive experiments on benchmark datasets, demonstrating that FlexiFlow achieves state-of-the-art performance in 3D molecular generation tasks. • We compare FlexiFlow generated conformers with those generated by traditional physics-based methods (CREST ( Pracht et al., 2020 ; Pracht et al., 2024 ) ), showing that FlexiFlow generates high-quality conformers at a fraction of the computational cost. • We show that the FlexiFlow framework transfers effectively to tasks that lack conformation datasets. In particular, we investigate protein conditioning: given a specific static pocket, we generate molecular graphs, each with multiple conformations. This demonstrates the potential of our method for real drug-discovery applications. 2 Related Works Flow matching unifies score-based diffusion and ODE-based generative modeling by learning vector fields that map simple priors to complex data distributions ( Lipman et al., 2024 ) . Subsequent work has improved its scalability ( Wildberger et al., 2023 ) and sample quality ( Gat et al., 2024 ) . Recent 3D molecular design methods fall into two groups: (1) generate molecular graphs then infer bonds from coordinates ( Hoogeboom et al., 2022 ) , and (2) jointly generate the graph and its bonds ( Irwin et al., 2025 ) . EDM by ( Hoogeboom et al., 2022 ) addresses this challenge using a diffusion model for 3D molecular structure generation. Other methods (GCDM ( Morehead and Cheng, 2024 ) , GFMDiff ( Xu et al., 2024 ) , EquiFM ( Song et al., 2023 ) , GeomLDM ( Xu et al., 2023 ) , GeomBFM ( Song et al., 2024 ) , MUDiff ( Hua et al., 2023 ) ) often produced unstable structures, but newer architectures and training strategies (FlowMol ( Dunn and Koes, 2024 ) , MiDi ( Vignac et al., 2023 ) and EQGAT-diff ( Le et al., 2024 ) ) have rapidly improved. Conditional flow matching (CFM) is now a leading approach ( Song et al., 2023 ; Campbell et al., 2024 ) , with recent work improving efficiency and generation quality Tabasco ( Vonessen et al., 2025 ) , FlowMol3 ( Dunn and Koes, 2025 ) , and SemlaFlow ( Irwin et al., 2025 ) . Conformers are critical when conditioning on a protein pocket ( Peng et al., 2022 ; Dong et al., 2024 ) , as they shape molecular bioactivity and properties. Conformers can be generated using physics-based tools such as CREST ( Pracht et al., 2024 ) which searches for low-energy conformational minima are computationally expensive. More recent approaches, such as Adjoint Sampling ( Havens et al., 2025 ) , generate conformers via flow matching while keeping the molecular graph fixed. This substantially reduces the computational cost required to produce conformers that closely resemble minimum-energy states. We extend these advances with a novel conditional