Protein conformational motion prediction from a single structure, using an SE(3)-equivariant GNN trained on ensembles mined from the PDB.
No providers recorded yet. Browse all providers
Structure prediction has largely solved the question of what shape a protein holds; it has not solved the question of how that shape moves. Functional behaviour — hinge closures, allosteric transitions, fold switches — lives in the motion, and the dominant computational answers are expensive. Molecular dynamics costs enormous compute, generative ensemble models such as AlphaFlow and BioEmu-1 sample approximate equilibrium distributions at high inference cost, and physics-based normal mode analysis depends heavily on the topology of the input structure.
PETIMOT (Protein sEquence and sTructure-based Inference of MOTions), from Sorbonne University and Université Grenoble Alpes, reframes the problem. Instead of sampling a distribution of conformations, it directly predicts a small linear subspace of motion — the leading eigenvectors of the positional covariance matrix computed across collections of experimental structures of the same protein or its close homologs. Those collections are sparse, noisy, and biased, but they are real, and a handful of linear modes explains most of the variance within them. Predicting a subspace rather than a distribution buys interpretability, linear-time scaling, and inference measured in seconds.
The model is an SE(3)-equivariant graph neural network initialised from protein language model embeddings, trained end-to-end against motions extracted from the Protein Data Bank with a loss designed for the geometry of the task rather than borrowed from coordinate regression.
Node embeddings come from ProstT5, a structure-aware protein language model that outperformed the larger ESM Cambrian 600M in ablation. Edge features encode SE(3)-invariant relative backbone frames. Fifteen message-passing blocks jointly update residue embeddings and motion vectors, with each message computed in the local frame of the central residue; the message MLP has one hidden layer of size 256. Training data was built from roughly 750,000 PDB chains clustered at 80% identity, yielding 7,335 conformational collections augmented to 36,675 samples via five reference conformations each. On the test824 split, PETIMOT reaches a 43.6% success rate against approximately 31% for AlphaFlow and BioEmu and 25% for normal mode analysis, with lower global subspace error (0.73 versus 0.77-0.78) and higher RMSF correlation (0.59 versus 0.51-0.52). It runs in under 16 seconds per protein where the generative baselines take over 38 hours. On the iMod open-closed transition benchmark the cross-validated model succeeds on 86% of cases; on 400 ATLAS molecular dynamics trajectories it reaches 60% without retraining.
PETIMOT suits any workflow that needs a fast, interpretable first guess at how a protein will move: prioritising cryptic pocket searches for drug discovery, seeding enhanced sampling with data-derived collective variables, generating starting ensembles for docking, or interpreting a newly solved structure in terms of its accessible deformations. Its speed makes proteome-scale screening tractable in a way that generative ensemble models are not.
The work argues that experimental structure collections, despite their sparsity, carry enough signal to learn transferable protein dynamics without simulation — a substantially cheaper data regime than the trend toward massive molecular dynamics corpora. The published version appeared in Acta Crystallographica Section D in 2026. The code is GPL-3.0, and the artifacts sit in a separate, published CC BY 4.0 Figshare record that holds considerably more than the checkpoint: the trained weights, the ground-truth motion subspaces, the train/test splits, PETIMOT's own predictions, and roughly 1.8 GB of AlphaFlow, BioEmu, ESMFlow, and normal-mode-analysis baseline predictions. Every number in the head-to-head comparison can therefore be recomputed by a third party without rerunning the baselines. The authors are candid about limits: the linear formulation cannot capture highly non-linear loop deformations, absolute success rates in the 35-45% range reflect both model error and the incomplete, crystallography-biased sampling of functional states in the PDB, and predicted motions may correspond to real states that have simply never been resolved.
Much of this page is generated or calculated automatically. Flag anything that looks off and we will re-run it.