Summary
Removing MSAs collapses the coherent latent geometry learned by AlphaFold3, even for familiar sequences. Low-confidence predictions occupy a poorly organized region rather than the compact pair-space manifold produced when comparative evolutionary context is available (1).
See also
- MSA-conditioned AlphaFold3 representations compress as the Pairformer proceeds
- High-confidence predictions from protein language models co-cluster together in embedding space and correlate with performance on variant effect prediction tasks
1.
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