Rethinking molecular evolution through protein language model embeddings

Abstract

Protein language models compress protein sequences into high-dimensional embeddings that capture biochemical, structural, and functional constraints without explicit supervision. We highlight that these embeddings encode rich evolutionary information, enabling new geometry-based views of homology, divergence, and convergence, and calling for a synthesis between classical molecular evolution and systematic evolutionary embedding analysis.

Keywords: deep learning; evolutionary bioinformatics; genomics; machine learning; phylogenetics; protein language models.