CoMoFlow Gives Protein Shape Generation a Set of Motion Controls

Beyond a single protein pose

Understanding protein flexibility means describing both the shapes a protein can adopt and how different regions move together. Molecular dynamics, or MD, simulations explore these variations by calculating molecular motion through many small time steps. That detail makes extensive sampling computationally expensive.

Generative AI offers a different approach: learn from simulated collections of structures, then generate new conformations without calculating an entire MD trajectory. But matching individual shapes does not necessarily mean accurately recovering coordinated variation across the collection.

In the CoMoFlow preprint, Yusong Wang and colleagues introduce an explicit representation of those coordinated changes. Their model generates protein backbone ensembles—collections of structural snapshots—and lets users steer selected patterns of movement.

Learning movements and their combinations

CoMoFlow starts from an equilibrium reference structure and its residue identities. During training, MD structures teach the model how proteins vary around their references. For an unseen test protein, generation uses only the input structure and residue identities, not that protein’s MD ensemble.

The model predicts eight protein-specific collective modes. Each mode describes a pattern of displacement across the backbone: a coordinated structural change rather than an independent adjustment to each residue. These learned modes are not simply classical harmonic normal modes calculated from a spring-like protein model.

Coefficients determine how strongly each mode contributes. CoMoFlow learns a probability distribution over these coefficients using a mixture of two Gaussian distributions. Importantly, it models dependencies between coefficients: geometrically distinct movement patterns can still tend to occur together.

A sampled combination guides a flow-matching generator, which progressively transforms an initial configuration into a protein conformation. The collective displacement both supplies information to the generator and shifts the starting backbone positions. The nonlinear generator handles the resulting structure, including residue orientations; the output is not merely the reference plus a linear displacement.

Training rewards accurate structures, useful collective modes, an appropriate coefficient distribution and agreement with the simulated relationships between coefficients.

What the benchmarks showed

The researchers trained on 1,269 natural proteins from ATLAS and evaluated on 82 held-out ATLAS proteins. A separate benchmark contained 100 computationally designed proteins used only for testing. Each protein’s reference simulations comprised three independent 100-nanosecond runs.

For ordinary ensemble evaluation, CoMoFlow generated 250 conformations per test protein. The measurements examined local flexibility, structural spread, coordinated residue movements, distribution matching and transient contacts.

Compared with BBFlow, a structure-conditioned baseline, CoMoFlow improved agreement with simulated residue-motion correlation patterns from 0.87 to 0.93 on ATLAS and from 0.83 to 0.89 on designed proteins. It also reduced errors in flexibility and overall conformational spread.

However, it did not lead every comparison. Some competing methods recovered transient contacts better and achieved higher flexibility correlations on designed proteins. Its advantage was therefore strongest in particular aspects of ensemble reconstruction, not universal superiority.

The authors also grouped simulated conformations into structurally distinct, densely populated regions, retaining only groups reproduced in multiple simulation replicas. Relative to BBFlow, CoMoFlow increased coverage of these regions from 0.45 to 0.62 on ATLAS and from 0.32 to 0.59 on designed proteins. It also better matched their relative populations. These groups were structural basins, not demonstrated long-lived kinetic states.

Steering a structural change

To test control, the researchers changed one mode coefficient while holding the others fixed. Generated structures responded consistently, although their changes were generally smaller than prescribed and other modes were not completely unaffected.

They then searched coefficient space for conformations that opened or closed the distance between selected protein regions. Targets came from the extremes of the MD distance distributions.

With the same budget of 600 generator calls per target, passive CoMoFlow sampling reached opening and closing targets for 32% and 38% of ATLAS test proteins. An iterative coefficient search raised those rates to 65% and 68%. Passive BBFlow sampling achieved 24% and 28%. This demonstrates useful search controls under that protocol, not guaranteed access to arbitrary protein shapes.

Where the evidence stops

All experiments were computational. Agreement with finite MD datasets does not establish complete sampling of real protein behavior, and generating an ensemble does not recover the timing or pathways of molecular transitions.

The model represents backbones rather than complete all-atom structures. Controlled outputs retained broadly stable local geometry in the reported checks, but those checks do not establish full physical or biological validity. Accuracy also deteriorated when input reference structures departed further from the original equilibrium reference.

The work is a preprint. Its reproducibility statement promises public code upon acceptance rather than reporting an accompanying release. The supported conclusion is narrower than a replacement for molecular simulation: explicitly learned collective movements improved several simulation-based benchmarks and provided a practical way to steer generated backbone changes.