FlexiFlow

FlexiFlow is a flow-matching generative model for 3D molecular design that treats a molecule not as a single rigid structure but as a flexible ensemble of conformations. Most generative models for 3D small molecules sample one conformer per molecule, collapsing the conformational diversity that governs how a compound actually behaves when it binds a target. FlexiFlow instead jointly samples the molecular graph together with multiple low-energy conformations, capturing the flexibility that is central to molecular recognition and thermodynamics.

The central idea is a decomposable flow-matching formulation that extends continuous flow matching to generate molecules with several conformers at once, while preserving the physical symmetries the problem demands: equivariance to rotations and translations and invariance to atom permutations. By learning the full conformational ensemble, FlexiFlow aims to bridge the gap between fast deep-learning generators and slower physics-based conformer sampling methods.

FlexiFlow was developed by Riccardo Tedoldi, Ola Engkvist, Patrick Bryant, Hossein Azizpour, Jon Paul Janet, and Alessandro Tibo, in a collaboration led by AstraZeneca's Molecular AI group together with KTH Royal Institute of Technology, Chalmers University of Technology, Science for Life Laboratory, and Stockholm University. It was posted to arXiv in November 2025.

#Key Features

• Joint molecule and ensemble generation: FlexiFlow simultaneously samples a molecular graph and multiple 3D conformations, modeling a molecule as a flexible ensemble rather than a single fixed structure.
• Decomposable flow matching: A decomposable flow-matching formulation extends continuous flow matching to the multi-conformer setting, letting the model generate several coupled conformations in a single framework.
• Symmetry-aware design: The architecture maintains equivariance to rotations and translations and invariance to atom permutations, respecting the physical and graph symmetries inherent to 3D molecules.
• Pretrained and transferable: The pretrained model transfers to protein-conditioned ligand generation, extending de novo design from unconditioned generation toward structure-based drug design.
• Fast ensemble sampling: The approach performs comparably to physics-based conformer-generation methods while offering substantially faster inference.

#Technical Details

FlexiFlow is a flow-matching architecture operating on 3D molecular representations, built to be equivariant to rigid-body transformations and invariant to atom permutations. Its decomposable flow-matching objective is designed so that a single model can jointly generate a molecule and a set of distinct conformations rather than denoising one conformer at a time. The model is pretrained for de novo 3D molecular generation and is reported to achieve state-of-the-art results on the standard QM9 and GEOM Drugs benchmarks, generating valid and novel molecules that capture conformational diversity. On ensemble quality, the authors report performance comparable to physics-based conformer generation while running considerably faster, and they show that the pretrained model transfers to protein-conditioned ligand generation. As the work is a November 2025 preprint, exact parameter counts, training-set sizes, and detailed hyperparameters await the full publication.

#Applications

FlexiFlow targets computational chemists and drug-discovery teams who need realistic 3D molecules together with their accessible conformations. Because binding, selectivity, and physicochemical properties depend on a molecule's conformational ensemble rather than a single pose, generating ensembles directly can support de novo design, virtual library construction, and downstream docking or scoring workflows. The demonstrated transfer to protein-conditioned ligand generation positions FlexiFlow for structure-based drug design, where candidate ligands must be proposed within the context of a target binding site.

#Impact

FlexiFlow contributes to the rapidly growing area of 3D molecular generative modeling by reframing generation around conformational ensembles, an aspect that single-conformer generators largely overlook. Its reported state-of-the-art results on QM9 and GEOM Drugs, comparable ensemble quality to physics-based methods at faster speed, and transfer to protein-conditioned settings suggest a flexible foundation for structure-based design. As a recent preprint from an industry-academia collaboration, its practical advantages will become clearer as the community benchmarks it and as code and pretrained weights are released. The authors state that code and weights are to be released upon full publication; they are not yet publicly available, and no model license has been specified.