Summary
Diffusion- and flow matching-based protein backbone design methods show fold-level novelty by recycling memorized subdomain structures (1).
Details
Across eight diffusion and flow-matching generators, 80.2-98.2% of outputs contained at least one domain alignable to CATH S40 at an aligned-length TM threshold of 0.5; requiring every domain to match reduced the range to 28.4-87.6%. These retrieval rates establish local alignability, not that every matched domain was memorized during training. RetFold, a retrieval-only baseline, reached 96.0% any-domain and 83.8% all-domain retrieval at roughly two orders of magnitude lower computational cost.
Figures
Figure 3 from (1)
See also
- Protein backbone diffusion models undersample loop-rich and alpha-beta domains and functional motifs
- Nearly half of CATH domains do not pass designability filters used to evaluate protein backbone design performance
1.
Xu T, Zhang Y, He M, Shen L, Liu Z, Ying T, et al. Is Retrieval All You Need? Assessment and Emergence of Novelty in Protein Structure Generation. 2026. Available from: https://arxiv.org/abs/2608.10598