Raw folding-model confidence is not a general protein stability predictor

Summary

Raw confidence from protein folding neural networks is not a general predictor of protein stability. Mutation-induced ddG values show little or no correlation with changes in pLDDT (1), and high-pLDDT designs can still unfold during Hamiltonian Replica-exchange molecular dynamics (2). The networks nonetheless contain some stability-related signal: AlphaFold2 pLDDT correlates with stability within restricted fold families (3), and ESMFold pLDDT moderately discriminates experimentally successful from unsuccessful monomer designs (4). Dedicated downstream models can also predict ddG from AlphaFold structures about as accurately as from experimental structures (5). RMSD itself does not correlate with stability, although (6) showed that a custom strain score was able to predict this with some success.

Figures

Ref (1)

See also

• High-pLDDT designs can be insoluble
• Protein structure prediction and design confidence metrics do not correlate with binding affinity
• ESMFold pLDDT weakly correlates with intra-family differences in stability and experimental success, but not differences in cooperativity

1.

Pak MA, Markhieva KA, Novikova MS, Petrov DS, Vorobyev IS, Maksimova ES, et al. Using AlphaFold to predict the impact of single mutations on protein stability and function. PLOS ONE. 2023;18(3):e0282689. Available from: https://doi.org/10.1371/journal.pone.0282689

2.

Aina A, Hsueh SCC, Gibbs E, Peng X, Cashman NR, Plotkin SS. De Novo Design of a β-Helix Tau Protein Scaffold: An Oligomer-Selective Vaccine Immunogen Candidate for Alzheimer’s Disease. ACS Chemical Neuroscience. 2023;14(15):2603–17. Available from: https://doi.org/10.1021/acschemneuro.3c00007

3.

Ferrari ÁJR, Dixit SM, Thibeault J, Garcia M, Houliston S, Ludwig RW, et al. Large-scale discovery, analysis and design of protein energy landscapes. Nature. 2026;654(8120):1108–18. Available from: https://doi.org/10.1038/s41586-026-10465-z

4.

Garcia M, Dixit SM, Rocklin GJ. Evaluating zero-shot prediction of monomeric protein design success by AlphaFold, ESMFold, and ProteinMPNN. Protein Science. 2026;35(2):e70453. Available from: https://doi.org/10.1002/pro.70453

5.

Diaz DJ, Gong C, Ouyang-Zhang J, Loy JM, Wells J, Yang D, et al. Stability Oracle: a structure-based graph-transformer framework for identifying stabilizing mutations. Nature Communications. 2024;15(1):6170. Available from: https://doi.org/10.1038/s41467-024-49780-2

6.

McBride JM, Polev K, Abdirasulov A, Reinharz V, Grzybowski BA, Tlusty T. AlphaFold2 Can Predict Single-Mutation Effects. Physical Review Letters. 2023;131(21). Available from: https://doi.org/10.1103/physrevlett.131.218401