Species phylogenetics using intraprotein residue dependence inferred with protein foundation models

RAM-NJ phylogenies inferred from esm-pFM residue attention matrices were generally as accurate as SAM-NJ phylogenies from msat-pFM attention matrices across established species relationships.

• Why it matters: Conventional phylogenetics relies on multiple sequence alignments and independent-site substitution models, so it is unclear whether residue dependence learned by protein foundation models can recover species relationships without those assumptions.
• What they did: The study compared RAM-NJ versus SAM-NJ phylogenies across different proteins and taxon sets, using Neighbor-Joining on pairwise distances estimated from pFM-derived RAMs and SAMs.
• The result: pFM-based analyses recovered species relationships without conventional substitution models, and RAMs did not require MSA construction, though pFM phylogenies were less accurate than standard maximum likelihood approaches.