Summary
Confidence metrics from structural modeling and design neural networks, namely pLDDT, ipTM, Rosetta scores, ProteinMPNN likelihoods, and ESM-2 log-likelihoods, do not correlate with binding for de novo binders (1,2). For pLDDT is also broadly true of protein stability. (3) showed that metrics like ipTM can quickly respond to even small changes in the conditioning representations used to drive diffusion-based structure prediction, showing how brittle such metrics are with respect to input sequences and MSAs.
Figures
Ref (1)
See also
- Protein folding neural networks cannot predict protein stability
- Protein structure prediction and design metrics don’t correlate with expression probability
- Sequence- and structure-derived ML quality metrics from ML do not correlate with each other
1.
Li Q, Vlachos EN, Bryant P. Design of linear and cyclic peptide binders of different lengths from protein sequence information. openRxiv; 2024. Available from: https://doi.org/10.1101/2024.06.20.599739
2.
Kosonocky CW, Abel AM, Feller AL, Cifuentes Rieffer AE, Woolley PR, Lála J, et al. Validation and analysis of 12,000 AI-driven CAR-T designs in the Bits to Binders competition. openRxiv; 2026. Available from: https://doi.org/10.64898/2026.03.03.709355
3.
Maddipatla A, Rzayev A, Pegoraro M, Pacesa M, Schanda P, Marx A, et al. Inference-time optimization for experiment-grounded protein ensemble generation. 2026; Available from: https://arxiv.org/abs/2602.24007