AlphaInterp: Revealing how AlphaFold 3 processes structure, confidence, and evolutionary information
How does evolutionary information become structural knowledge? AlphaFold 3 predicts protein structures and biomolecular assemblies with remarkable accuracy, yet how it transforms comparative sequence information into three-dimensional structure remains poorly understood. Here, we use mechanistic interpretability to follow this transformation through AlphaFold 3 and identify a computational principle underlying its structural inference. By extracting single- and pair-representations across the model’s forward pass and combining representational-geometry analysis, linear probing, activation patching, and perturbation of the multiple-sequence alignment (MSA) and its phylogenetic diversity, we show that AlphaFold 3 does not simply infer structure from raw sequence nor from MSA depth. Instead, a small number of sufficiently divergent homologues provides enough comparative information to largely recover the structure predicted using the full MSA. The pairformer transforms this evolutionary signal into a compact latent space in which residue-pair distances, contacts, fold similarity, and structural confidence become progressively decodable. Removing the MSA disrupts these latent geometries. Replacing it with evolutionarily similar homologues provides substantially less information, while evolutionarily unrelated sequences fail to rescue prediction despite preserving the formal alignment input. In contrast, phylogenetically diverse MSA subsampling largely restores both prediction accuracy and representational coherence. These results suggest that AlphaFold 3 has learned to infer structure by constraining its geometric priors and structural hypotheses by the sequence variation of the MSA. This framework helps explain both the model’s remarkable predictive power and its characteristic failure modes, with direct implications for protein design, conformational diversity, out-of-distribution generalization, and the interpretation of model confidence.
Authors
- Jeffrey Skolnick (ORCID: https://orcid.org/0000-0002-1877-4958)
- Jonathan Feldman (ORCID: https://orcid.org/0000-0003-4130-6447)
Institutions
- Georgia Institute of Technology (US)
Publication Details
- Journal
- PLoS Computational Biology
- Published
- 2026-11-19
- DOI
- https://doi.org/10.1371/journal.pcbi.1014981
- Citations
- 11
- Primary Topic
- Protein Structure and Dynamics
- Type
- article
- Field-Weighted Citation Impact
- 12.31
Funders
- National Institute of General Medical Sciences