AI Model Unites Sequences and Structures to Map Protein Evolution

An international research consortium has developed a novel protein language model that successfully unites amino acid sequences and three-dimensional structures into a single, cohesive mapping framework. The breakthrough, published in the Proceedings of the National Academy of Sciences, was spearheaded by researchers from the University of Haifa, Tel Aviv University, and the Earth-Life Science Institute at the Institute of Science Tokyo. Led by Professor Rachel Kolodny, Guy Yanai, Professor Nir Ben-Tal, Gabriel Axel, and Associate Professor Liam M. Longo, the team addressed a longstanding computational challenge in evolutionary biochemistry: the historical separation of protein sequence and structure analysis. Traditional protein language models treat sequences and structures as distinct data streams, often placing the same protein in different regions of a conceptual protein map. The newly developed Contrastive Learning Sequence-Structure model, or CLSS, employs contrastive learning to align both data types within a shared numerical embedding space. During training, the model processes complete sequences and their corresponding structural conformations, learning to assign nearly identical vector representations to related pairs while segregating unrelated ones. This approach generates a unified protein universe map where neighboring points reflect close evolutionary and functional relationships, regardless of whether sequence or structure is input. Validation tests demonstrate that CLSS embeddings accurately mirror established expert-curated classifications, including ECOD and CATH, without explicit training on those hierarchies. The model also outperforms current state-of-the-art architectures in classification benchmarks, confirming the computational advantage of integrated sequence-structure analysis. Notably, CLSS successfully positions short protein fragments alongside complete sequences and structures, enabling researchers to trace how ancient, modular protein domains have been repurposed and recombined over billions of years of evolution. By overlaying biological properties onto the unified maps, the team identified distinct spatial distributions linked to specific functions, such as the concentration of organic cofactor-binding proteins in defined regions and the broader dispersion of metal-binding enzymes. These patterns illustrate how global protein mapping can reveal large-scale evolutionary trajectories that conventional methods struggle to detect. The development establishes a new foundation for exploring protein diversity, function, and ancestry. Researchers anticipate that unified sequence-structure representations will streamline database queries, accelerate rational protein engineering, and enable more accurate reconstructions of ancestral protein landscapes. By treating sequence and structure as interdependent rather than isolated variables, the CLSS model offers a scalable computational lens for decoding the molecular evolution that has shaped life over nearly four billion years.