Diffusion-based structure prediction can sometimes model conformational ensembles

Summary

Protein structure prediction methods reliant on diffusion can sometimes predict conformational heterogeneity (1). Authors of Diffold had to reweigh the PDB by clustering by TM-score. Prior efforts by others were not successful: the conformational diversity of EigenFold was more representative of error than actual dynamics (2), whereas authors of AlphaFold3 found that it predicted the same (incorrect) conformation of E3 ubiquitin ligases, rather than an ensemble of states (3).

See also

• Biomolecular diffusion models cannot reproduce the equilibrium dynamics of the simulations they are trained on

1.

Fan J, Li Z, Alcaide E, Ke G, Huang H, E W. Accurate Conformation Sampling via Protein Structural Diffusion. Journal of Chemical Information and Modeling. 2024;64(22):8414–26. Available from: https://doi.org/10.1021/acs.jcim.4c00928

2.

Jing B, Erives E, Pao-Huang P, Corso G, Berger B, Jaakkola T. EigenFold: Generative Protein Structure Prediction with Diffusion Models. 2023; Available from: https://arxiv.org/abs/2304.02198

3.

Abramson J, Adler J, Dunger J, Evans R, Green T, Pritzel A, et al. Accurate structure prediction of biomolecular interactions with AlphaFold 3. Nature. 2024;630(8016):493–500. Available from: https://doi.org/10.1038/s41586-024-07487-w