Protein conformational ensemble generator conditioned on backbone geometry alone, sampling MD-like dynamics without MSAs or a folding model.
No providers recorded yet. Browse all providers
BBFlow is a flow matching model that generates conformational ensembles of protein backbones, developed by the Gräter group across the Max Planck Institute for Polymer Research, the Heidelberg Institute for Theoretical Studies, and Heidelberg University, with collaborators at Stockholm University and the Karlsruhe Institute of Technology. It was posted in February 2025 and published at NeurIPS 2025.
Proteins are not single structures, and the ensembles molecular dynamics produces are often what a question actually requires — how flexible a loop is, whether a cryptic pocket opens, how two domains move together. MD is expensive enough that generative emulators have become an active alternative. The dominant approach, exemplified by AlphaFlow, fine-tunes a pretrained folding model and consumes evolutionary information as multiple sequence alignments or protein language model weights. That inheritance is costly at inference and carries a bias: for sequences with shallow alignments, and for de novo designed proteins with no evolutionary history at all, the conditioning signal is missing.
BBFlow drops the folding model entirely. Instead of predicting the fold from sequence, it takes the equilibrium backbone structure as input and generates fluctuations around it. The result is a model small enough to train from scratch in a few GPU days that runs an order of magnitude faster than its predecessor and works precisely where evolutionary conditioning fails.
BBFlow.from_tag('latest') and samples through a Python API, a bbflow_sample CLI, or a Colab tutorial.A protein backbone is represented as a sequence of SE(3) frames, one rotation and translation per residue, and flow matching is formulated on the SE(3)^N manifold with geodesic interpolation paths. The network follows the GAFL architecture, an extension of FrameDiff and FrameFlow in which invariant point attention is replaced by Clifford frame attention; hyperparameters are inherited from those models. Conditioning enters as a distance embedding of the equilibrium structure, applied both to the learned vector field and to the prior distribution.
Training and evaluation use ATLAS, a curated set of three 100 ns all-atom MD trajectories for 1,390 structurally diverse proteins, under the same 1,265 / 39 / 82 split AlphaFlow used. Ensembles are scored on RMSF error and correlation, pairwise RMSD, dynamic cross-correlation matrices, a Wasserstein distance over principal components, and transient contact accuracy. On the ATLAS test set BBFlow matches AlphaFlow with templates on RMSF and principal components, beats it on pairwise RMSD and median RMSF — AlphaFlow-T systematically over-stabilizes and samples too close to the equilibrium structure — and outperforms template-free AlphaFlow, ESMFlow-T, two distilled variants, and ConfDiff on nearly every metric. Inference, measured per conformation of the 302-residue protein 7c45A on an A100-80GB, is about 40 times faster than AlphaFlow-T at length 300, with the margin widening for larger proteins. Ensemble generators targeting other distributions, such as BioEmu-1, are evaluated separately because they do not reproduce a fixed MD protocol.
The natural use is annotating structures with dynamics at a scale MD cannot reach: screening candidate proteins for a target motion, flagging flexible loops or hinge regions before a design decision, or generating starting conformations for targeted simulations. Because it needs only a backbone, it slots directly downstream of a structure predictor or a design pipeline. Protein designers benefit most — a de novo backbone can be checked for unintended flexibility before synthesis, exactly the case where alignment-based ensemble methods have nothing to work with. The multi-chain capability extends this to oligomeric assemblies.
BBFlow's argument is a useful negative result: neither pretrained folding weights nor evolutionary sequence information are needed to reproduce the conformational fluctuations seen in 300 ns of MD. That reframes ensemble generation as a structure-to-structure task rather than a folding task, and the efficiency gains follow from the smaller problem being solved. The code is released under the MIT license with a pretrained checkpoint, the preprocessed ATLAS dataset, and training scripts, though it depends on the separate GAFL and FrameFlow packages. Scope is bounded by the training distribution: the model emulates ensembles matching the temperature, force field, and 100 ns simulation length of ATLAS rather than the full Boltzmann distribution, and it generates backbones only, without side chains.
Much of this page is generated or calculated automatically. Flag anything that looks off and we will re-run it.