Summary
Predicted protein structures are less designable than experimental models in ways that suggest distortions at the local but not global level (1). This was observed by designing predicted models and attempting to forward-fold them, and could explain why training inverse folding methods exclusively on predicted models worsens performance relative to including experimental structures.
Figures
Figure 2C from (1)
1.
Lin Y, Lee M, Zhang Z, AlQuraishi M. Out of Many, One: Designing and Scaffolding Proteins at the Scale of the Structural Universe with Genie 2. 2024; Available from: https://arxiv.org/abs/2405.15489