Learning the language of phylogeny with MSA Transformer

Area of Science:

• Computational biology
• Bioinformatics
• Evolutionary biology

Background:

• Classical phylogenetics assumes independence between sites, potentially missing complex interactions like epistasis.
• Protein language models, such as MSA Transformer, can learn dependencies within protein sequences from multiple sequence alignments (MSAs).

Purpose of the Study:

• To investigate if MSA Transformer captures evolutionary distance and reflects epistasis in protein evolution without explicit training on these features.
• To evaluate the utility of MSA Transformer's representations for phylogenetic inference.

Main Methods:

• Systematic shuffling of natural and simulated MSAs to test the model's reliance on column-wise conservation.
• Reconstructing phylogenetic trees using internal embeddings from MSA Transformer.
• Comparing reconstructed trees with those from maximum likelihood inference.
• Applying the method to RNA-dependent RNA polymerase and nucleo-cytoplasmic large DNA virus domains.

Main Results:

• MSA Transformer effectively utilizes column-wise conservation to discern phylogenetic relationships.
• Phylogenetic trees reconstructed from MSA Transformer embeddings show high consistency with maximum likelihood methods.
• Established and novel evolutionary relationships were identified in viral protein families.

Conclusions:

• MSA Transformer's representations capture evolutionary signals, complementing traditional phylogenetic inference.
• This approach offers a powerful tool for more accurate reconstruction of protein family evolutionary histories.