Protein language models trained on multiple sequence alignments of homologous proteins
successfully capture coevolution between amino acids in structural contact: this is one of the ingredients of the success of AlphaFold. We have used such models, especially MSA
Transformer, to generate new protein sequences from given protein families, and to predict which proteins interact among the members of two protein families.
Despite their successes, a drawback of models based on multiple sequence alignments is
that sequence alignment can be imperfect. Thus, we developed ProtMamba, a
homology-aware but alignment-free protein language model, which is able to generate new
protein sequences from given protein families.
Beyond the amino-acid scale, coevolution also exists between genes that in a genome. To
capture it, we trained ProteomeLM on complete proteomes spanning the tree of life. This
model allows quick and precise scans of whole protein interaction networks.

The talk is part of the Computational Approaches to Early Evolution Workshop (April 20-24, 2026)