Boltz-2 intermediate representations can predict protein-protein binding affinity

Summary

Intermediate Boltz-2 representations can support a separate predictor of protein-protein $\Delta G$ and $\Delta \Delta G$ even though the native affinity module transfers poorly (1). PreFold-dG aggregates residue-level embeddings with inter-residue-distance weighting and predicts $\Delta G$ directly. It achieved state-of-the-art results on established benchmarks and remained robust on independent test sets; ablations indicated that all intermediate representations contributed, with different importance for $\Delta G$ and $\Delta \Delta G$.

See also

• The Boltz-2 affinity module cannot be effectively repurposed for PPI affinity prediction
• Structure prediction uncertainty metrics as energy functions

1.

Park S, Yim S, Kim D, Yoo K, Hwang D, Lee K, et al. PreFold-dG: estimating binding affinity of protein-protein interaction from intermediate representations of protein folding model. Bioinformatics. 2026;42(Supplement_2):btag489. Available from: https://doi.org/10.1093/bioinformatics/btag489