Structural tokens improve zero-shot variant effect prediction in ESM3, but only when structures are not computationally derived

Summary

Structural tokens improved zero-shot variant effect prediction ESM3 inference only when using the experimental WT structure (1). Computational models obtained using Rosetta or short MD simulations substantially lowered the Spearman correlation, despite the resulting embeddings being almost identical.

Figures

Figures from (1)

See also

• SaProt
• Training inverse folding and diffusion models exclusively on predicted protein structures worsens performance due to how locally perfect they are

1.

Loux T, Wang D, Shakhnovich EI. More Structure, Less Accuracy: ESM3’s Binding Prediction Paradox. openRxiv; 2024. Available from: https://doi.org/10.1101/2024.12.09.627585