Researchers investigating AlphaFold 3 have identified a computational principle explaining how the artificial intelligence model transforms comparative evolutionary sequence data into three-dimensional biomolecular structures. Using mechanistic interpretability techniques, Jonathan Feldman and Jeffrey Skolnick analyzed single- and pair-representations across the network forward pass. Their findings reveal that the system does not derive structural coordinates directly from raw sequence or multiple-sequence alignment depth, but instead relies on sequence variation to constrain geometric priors.
To trace this transformation, the authors combined representational-geometry analysis, linear probing, activation patching, and deliberate perturbations of the multiple-sequence alignment and its phylogenetic diversity. The investigations demonstrated that a small number of sufficiently divergent homologues provides enough comparative evolutionary information to largely recover the structural predictions obtained using a complete alignment.
Inside the model architecture, the pairformer module maps this evolutionary signal into a compact latent space. Within these internal representations, residue-pair distances, contacts, fold similarity, and structural confidence metrics become progressively decodable. Completely removing the multiple-sequence alignment disrupts these latent geometries, highlighting the structural role played by input sequence divergence.
Alignment substitution tests further illuminated the underlying mechanics of structural inference. Supplying evolutionarily similar homologues yielded substantially less structural information, while evolutionarily unrelated sequences failed to rescue prediction accuracy despite maintaining formal alignment inputs. In contrast, subsampling phylogenetically diverse sequences largely restored both prediction accuracy and representational coherence.
These experimental results indicate that AlphaFold 3 operates by constraining learned geometric priors and structural hypotheses through sequence variation observed in the multiple-sequence alignment. The researchers noted that this analytical framework helps clarify the model predictive power while shedding light on characteristic failure modes, with direct implications for protein design, conformational diversity, out-of-distribution generalization, and confidence interpretation.
Subject of Research: Mechanistic interpretability of structural inference in AlphaFold 3
Article Title: AlphaInterp: Revealing how AlphaFold 3 processes structure, confidence, and evolutionary information
Article References: Feldman, J., & Skolnick, J. (2026). AlphaInterp: Revealing how AlphaFold 3 processes structure, confidence, and evolutionary information. PLOS Computational Biology, 22(11), e1014981. https://doi.org/10.1371/journal.pcbi.1014981
Image Credits: AI Generated
DOI: 10.1371/journal.pcbi.1014981
Keywords: AlphaFold 3, mechanistic interpretability, protein structure prediction, pairformer, multiple-sequence alignment, phylogenetic diversity, representational geometry, linear probing, activation patching, structural confidence, computational biology, geometric priors, homologues, biomolecular assemblies, latent representations, protein design, conformational diversity
News Source: Drew Townsend. (November 19, 2026). Mechanistic Interpretability Shows How AlphaFold 3 Extracts Protein Structural Hypotheses. Scienmag.



