Back to AI Research

AI Research

CoMoFlow gives researchers control over simulated protein motions

Key Takeaways

  • CoMoFlow uses collective motion to guide protein backbone generation.
  • The authors compare it with flow baselines on ATLAS and designed protein ensembles.
  • CoMoFlow conditions protein backbone generation on collective motion.
  • The authors report closer matches to molecular dynamics ensembles on ATLAS and de novo proteins.
  • The CoMoFlow preprint describes a way to generate several possible shapes of a protein backbone while controlling which motions the model emphasizes.

The CoMoFlow preprint describes a way to generate several possible shapes of a protein backbone while controlling which motions the model emphasizes. The authors use collective modes, which describe coordinated changes across parts of a structure. They then use those modes to guide a flow-based generator. The preprint was submitted to arXiv on November 6, 2026.

Why generate an ensemble of protein shapes?

A single static structure shows one arrangement of the protein. An ensemble describes other arrangements it may take. The paper argues that an ensemble generator should capture how multiple parts move together as well as variation in each part. A generator that matches local spread can still give a misleading picture if it misses those coordinated changes.

How does CoMoFlow expose the motion?

The method constructs protein-specific collective modes from a reference equilibrium structure. It learns how strongly those modes occur in molecular dynamics ensembles, then samples coefficients to express a desired variation. These values condition the generator and shift its starting distribution through a step the authors call prior recentering. Flow matching then turns the conditioned starting point into a generated backbone. Making the modes explicit gives a researcher a way to vary a targeted motion rather than relying only on uncontrolled samples.

What did the evaluation establish?

The authors evaluated the generated ensembles on ATLAS and de novo proteins. They report improvements over flow-based baselines in reconstructing the simulated conformational distribution and coordinated structural variation. They also demonstrate direct control over targeted changes in the generated structures.
These results compare generated backbones with molecular dynamics references. They do not establish that a proposed conformation has been measured in a physical sample. For researchers choosing a generator, the useful comparison is whether it captures the specific coordinated motion they need, along with the overall distribution.

Comments