Scientists have introduced CoMoFlow, an algorithm that improves how computational models recreate the variety of shapes that proteins can take during simulations. The method builds special collective motion patterns unique to each protein from its equilibrium structure. It then learns how often each pattern appears in a set of long molecular‑dynamics simulations. When generating new protein shapes, CoMoFlow samples pattern coefficients and feeds them into a flow‑based generator, guiding the synthetic structures toward realistic, coordinated movements.

Experiments on two benchmark datasets – the ATLAS protein set and a collection of de‑novo proteins – show CoMoFlow outperforms other strong flow‑based techniques. Across most evaluation metrics, it achieves higher correlation in the dynamic cross‑correlation matrix (DCCM), lower Wasserstein‑2 distances for principal‑component analysis (PCA), and more accurate root‑mean‑square fluctuation (RMSF) predictions. These gains indicate that CoMoFlow captures both the distribution of conformational states and the subtle coordinated variations that arise in real protein dynamics.

CoMoFlow’s design allows researchers to control specific protein motions. By selecting particular coefficients, users can steer the model toward desired structural changes while preserving physical plausibility. However, the current implementation is limited to proteins with fewer than 730 residues, as longer chains were not included in the training or evaluation sets.

The study demonstrates that learning protein‑specific collective patterns and their statistical distribution can significantly enhance the fidelity of generative models for protein conformations, while also providing a framework for targeted conformational manipulation.