Most computational protein-stability metrics correlate poorly with experimental dG

Summary

Most zero-shot computational protein-stability metrics correlate only modestly with experimental folding $\Delta G$. In a megascale small-domain dataset, buried hydrophobic nonpolar surface area reached a Spearman correlation of 0.60, while the best tested zero-shot sequence-likelihood model reached 0.43 and lost most of its signal after controlling for high buried surface area (1). By contrast, dedicated neural networks trained to predict absolute protein stability (SaProt$\Delta G$ and ESM3$\Delta G$) reach correlations of approximately 0.88.

See also

• Protein folding neural networks cannot predict protein stability
• Structure-based methods outperform sequence-based methods on protein stability prediction of point mutants, but not full sequences

1.

Cho Y, Tsuboyama K, Litberg TJ, Jung MD, Obisesan A, Wang Q, et al. Accurate protein stability prediction for small domains using mega-scale experiments. openRxiv; 2026. Available from: https://doi.org/10.64898/2026.05.19.726285