CoCoFold2

COMPUTATIONAL STRUCTURAL BIOLOGY

CoCoFold2Scalable latent refinement of diffusion-based protein structure predictions from limited-particle cryo-EM data

Junwen Liao·Mingxu Hu·Chenglong Bao

LangTaoSha Preprint Server · 2026

Preprint Code User Guide

A frozen structural prior, guided by experimental particles.

From predictions to particle-guided structures

CoCoFold2 connects a frozen Protenix-v1 diffusion prior to cryo-EM particle observations. It optimizes a target-specific latent perturbation while keeping network weights fixed.

CoCoFold2 overview: limited-particle observations guide latent refinement through a frozen diffusion prior.
Method overview. Deterministic latent refinement connects predicted atomic structures to experimental particles through differentiable rendering. View full-size figure ↗

Frozen diffusion prior

Refine the target-specific latent representation without fine-tuning the pretrained network.

Particle observations

Use upstream particle poses and CTF estimates to connect structures to experimental images.

Component-parallel refinement

Distribute component caches across GPUs using Independent or Contextual conditioning.

Refining large assemblies

Component-parallel refinement supports structural refinement across multiple GPUs. The preprint presents Independent and Contextual conditioning strategies and structural examples for MSP-1 and TRPM8.

Component conditioning strategies and initial and refined structures for MSP-1, PDB 6ZBH, and TRPM8, PDB 6O77.
Component conditioning and structural examples. MSP-1 (PDB 6ZBH) and TRPM8 (PDB 6O77). See the preprint for the experimental conditions, evaluation and limitations. View full-size figure ↗

CoCoFold2 requires a sufficiently accurate initial prior and informative experimental observations. Particle poses and CTF parameters are supplied upstream; CoCoFold2 does not estimate them.

Try CoCoFold2

Start with a small example, then adapt the workflow to your data. The user guide includes input preparation, parameters, saved outputs and restart instructions.

01Install the software ↗ 02Run the minimal example ↗ 03Refine across two GPUs ↗

Citation

If you use CoCoFold2, please cite the preprint:

@article{liao2026cocofold2,
  title = {CoCoFold2: scalable latent refinement of diffusion-based protein structure predictions from limited-particle cryo-EM data},
  author = {Liao, Junwen and Hu, Mingxu and Bao, Chenglong},
  journal = {LangTaoSha Preprint Server},
  year = {2026},
  doi = {10.65215/LTSpreprints.2026.09.15.000338},
  url = {https://doi.org/10.65215/LTSpreprints.2026.09.15.000338}
}