Rethinking molecular evolution through protein language model embeddings

Area of Science:

• Computational biology
• Bioinformatics
• Evolutionary biology

Background:

• Protein language models (PLMs) generate high-dimensional embeddings from protein sequences.
• These embeddings implicitly learn biochemical, structural, and functional constraints without supervised labels.

Purpose of the Study:

• To demonstrate that protein embeddings encode significant evolutionary information.
• To introduce novel geometry-based methods for analyzing protein evolution using these embeddings.
• To advocate for integrating evolutionary embedding analysis with classical molecular evolution.

Main Methods:

• Analysis of protein sequence embeddings generated by unsupervised language models.
• Application of geometric and topological methods to visualize and quantify evolutionary relationships.
• Comparison of embedding-based evolutionary insights with traditional phylogenetic approaches.

Main Results:

• Protein embeddings reveal rich evolutionary signals, including homology, divergence, and convergence.
• Geometric analysis of embeddings provides new perspectives on evolutionary trajectories.
• Embeddings capture complex evolutionary patterns not readily apparent through conventional methods.

Conclusions:

• Protein language model embeddings are a powerful resource for evolutionary analysis.
• A synthesis of evolutionary embedding analysis and molecular evolution is proposed.
• This approach offers novel tools for understanding protein evolution and diversification.