Summary
AlphaFold3-generation methods like Boltz and Chai exclusively incorporate MSA information into the pair representation, in contrast with Evoformer-based methods that reason over the information in the folding trunk (1). Mechanistic probing of AlphaFold3 likewise found that structural signal becomes concentrated and increasingly linearly decodable in the pair representation as the Pairformer proceeds (2).
1.
Abramson J, Adler J, Dunger J, Evans R, Green T, Pritzel A, et al. Accurate structure prediction of biomolecular interactions with AlphaFold 3. Nature. 2024;630(8016):493–500. Available from: https://doi.org/10.1038/s41586-024-07487-w
2.
Feldman J, Skolnick J. AlphaInterp: Mechanistic Interpretability of AlphaFold 3 Reveals How Evolutionary Information Shapes Protein Structure Prediction. openRxiv; 2026. Available from: https://doi.org/10.64898/2026.04.22.720175