Machine learning reveals hidden dimensions of functional similarity in proteins - Commentary

Sep 29, 2026

Upload by dawugda

Machine learning is used to uncover hidden molecular determinants of functional similarity and phenotypic convergence in proteins. Large language models trained on protein sequences generate embeddings that capture biochemical and evolutionary properties, allowing proteins with similar functions to cluster even when traditional site-by-site sequence analysis fails. The discussion contrasts prior methods relying on sequence alignment and profile similarity with embedding-based approaches that detect convergence without requiring detectable amino-acid substitutions or high sequence identity, illuminating connections between genotype, structure, and function.

LanguageEnglish

CategoryResearch & Report

FormatPDF

Pages3

Reading Time8 min

AccessFree to download

The traditional view of molecular convergence

Protein language models: decoding molecular convergence

Embeddings and functional clustering

Beyond site-level sequence similarity

COMMENTARY OPEN ACCESS Machine learning reveals hidden dimensions of functional similarity in proteins Noor Youssefa, b, 1 , Sarah Gureva, b , and Debora S. Marksa, b Large language models trained on biological sequences, rather than natural language, are transforming biology, from predicting human genetic disease (1, 2) to the design of new-to-nature proteins (3–5) . In this issue of PNAS, Cao et al. (6) extend these applications to detect the molecular underpinnings of phenotypic convergence by decoding patterns invisible to traditional sequence analysis approaches (Fig. 1) . The Traditional View of Molecular Convergence Evolution often arrives at strikingly similar solutions to common environmental challenges. Bats and toothed whales independently evolved sonar-like echolocation systems for navigating in hard-to-see environments (7). Distantly related lineages of fish developed antifreeze proteins to survive polar seas (8). Flowering plants from disparate families converged on crassulacean acid metabolism (CAM) as a water-use efficient adaptation of photosynthesis (9) . These phenotypic convergences are encoded by genomic changes and therefore act as natural experiments that can illuminate the mapping between genotype and phenotype. For decades, the molecular basis of convergent evolution has been investigated primarily through the lens of site-by-site convergence: the emergence of the same amino acid change independently at homologous positions across different lineages (10) . These approaches have uncovered parallel substitutions in proteins underling echolocation in bats and cetaceans (11), and the repeated evolution of antifreeze glycoproteins in polar fishes (8) . Yet, this approach is inherently limited since proteins are not simple strings of independent amino acids but intricate three-dimensional structures where distant residues interact, where function emerges from collective properties, and where multiple Author affiliations: a Department of Systems Biology, Harvard Medical School, Boston, MA 02115; and bBroad Institute of Harvard and Massachusetts Institute of Technology, Cambridge, MA 02139 Author contributions: N.Y., S.G., and D.S.M. analyzed data; and N.Y. wrote the paper. Competing interest statement: D.S.M. is a cofounder of Seismic Therapeutic. D.S.M. is an advisor for Dyno Therapeutics, Octant, Jura Bio, Tectonic Therapeutic, and Genentech. Copyright © 2026 the Author(s) . Published by PNAS. This open access article is distributed under Creative Commons Attribution-NonCommercial-NoDerivatives License 4.0 (CC BY-NC-ND) . See companion article,“Language models reveal a complex sequence basis for adaptive convergent evolution of protein functions,” 10.1073/pnas.2418254122 . 1To whom correspondence may be addressed. Email: [noor_youssef@hms.harvard.edu](noor_youssef@hms.harvard.edu). Published January 2, 2026. Fig. 1. Detecting molecular convergence using protein language model embeddings. (A) Protein language models are deep neural networks trained on millions of protein sequences without explicit structural or functional supervision. The resulting numerical representations (embeddings) capture biochemical and evolutionary properties, enabling proteins with similar functions to cluster together in high-dimensional embedding space. (B) Traditional approaches detect molecular convergence by aligning homologous proteins and identifying identical amino acid substitutions that arise independently in distinct lineages. In contrast, protein language model embeddings can reveal functional convergence even in the absence of site-level sequence similarity. (C) Using protein language model embeddings, Cao et al. identify candidate genes underlying the convergent evolution of echolocation in bats and whales. molecular routes may lead to the same function. Tools like PSI-BLAST (12) and HMMER (13) began to address this by identifying functional relationships through shared sequence profiles, r

FAQs

What limitation of traditional convergence analysis does the article address?

Traditional approaches largely focus on site-by-site amino-acid changes at homologous positions, which is limited because protein function depends on three-dimensional structure and interactions among distant residues.

How do protein language model embeddings help detect functional convergence?

Embeddings map proteins into a high-dimensional space where proteins with similar functions cluster together, enabling detection of convergence even when there is little or no site-level sequence similarity.

What training objective allows protein language models to learn protein “grammar”?

Protein language models use self-supervised learning such as masked amino-acid prediction or next-residue prediction to learn patterns of amino acids that correlate with structural and functional properties.

Sep 29, 2026

Upload by dawugda