The abstract is read at the publisher; the summary is JClub's.
AlphaFold 3 converts evolutionary signal from a small set of sufficiently divergent homologues into a compact latent geometry that decodes residue-pair distances, contacts, fold similarity, and structural confidence.
- Why it matters: It remains unclear how comparative sequence information is transformed into three-dimensional structural knowledge, and why AlphaFold 3 sometimes fails despite strong overall accuracy.
- What they did: Using mechanistic interpretability, the study analyzes single- and pair-representations across AlphaFold 3’s forward pass with representational-geometry analysis, linear probing, activation patching, and MSA perturbations.
- The result: Removing the MSA disrupts latent geometries, replacing it with evolutionarily similar homologues gives substantially less information, and only phylogenetically diverse subsampling largely restores prediction accuracy and representational coherence.
The findingWhy it mattersWhat they didThe result
AI summary of the abstract · may contain errors · not medical advice · read the paper.