PETIMOT: learning protein motions with AI from sparse experimental data

What is it about?

Proteins are not static objects — they constantly move and deform to carry out their biological functions, whether binding to other molecules, transmitting signals, or catalyzing chemical reactions. Understanding these motions is essential but remains a major challenge, since experimental techniques (X-ray crystallography, cryo-EM) usually only capture a single frozen snapshot of a protein, while classical physics-based simulations (molecular dynamics) are extremely computationally expensive. In this work, we introduce PETIMOT, a new neural network that learns to directly predict the most likely directions of motion for a protein from its sequence and 3D structure. Unlike recent generative approaches that require massive amounts of simulation data, PETIMOT is trained only on experimental structures already available in the Protein Data Bank (PDB), taking advantage of the multiple conformations deposited for similar proteins. The model combines deep learning techniques that respect the geometric symmetries of molecules (SE(3)-equivariance) with representations derived from pre-trained protein language models.

Featured Image

Why is it important?

Speed and efficiency: PETIMOT is about 2.75 times faster than classical normal-mode analysis (NMA), and thousands of times faster than state-of-the-art generative models (AlphaFlow, BioEmu). Better accuracy on functional motions: the model outperforms both traditional physics-based approaches (NMA) and modern generative approaches based on diffusion or flow-matching, particularly for capturing large-amplitude and slow conformational changes, which are often the most biologically relevant. Generalization: PETIMOT performs well even on proteins very different from those used in training, and also generalizes to molecular dynamics data (ATLAS) without any retraining. Interpretability: by working in a space of linear motions (rather than generating complex conformational ensembles), the results are easier to interpret and apply directly by structural biologists. Practical applications: this type of tool can accelerate the study of protein functional mechanisms, help guide heterogeneous cryo-EM reconstruction, and support protein engineering and drug development projects.

Perspectives