Nature Methods Биотехнология

Transforming molecular evolutionary analyses with protein foundation models

Kumar, Sudhir, Alibutud, Rohan, Allard, John B., Mowlaei, Mohammad E., Shi, Xinghua
Springer Nature · 2026-11-18 · DOI 10.1038/s41592-026-03304-6

Аннотация

Transformer-based protein foundation models (pFMs) are emerging as powerful tools for analyzing protein sequences. Their adoption in evolutionary analysis requires articulating their evolutionary foundations and their relationship to traditional models and workflows. We synthesize research showing that pFMs trained on millions of natural protein sequences capture residue–residue dependencies, enabling inference of functional constraints, mutational pathogenicity and evolutionary relationships. We introduce primary products of pFM-based analysis and demonstrate that they are informed by neutral evolutionary patterns, without incorporating substitution matrices, phylogenies or even multiple sequence alignments. We suggest that analyzing individual sequences with pFMs will streamline many evolutionary and functional analyses, including those with few homologs. Development and integration of pFMs with traditional approaches will advance molecular evolutionary investigations from phenomenological modeling of primary sequence differences to models of residue dependencies. We discuss limitations, computational considerations and open questions regarding the routine use of pFMs in protein evolutionary research. This Perspective traces the history of evolutionary analysis from early substitution matrices to coevolutionary models and proposes that protein foundation models complement classical statistical models by enabling single-sequence evolutionary analysis and will shift the field from modeling independent residue differences to studying the evolution of residue dependencies.

Читать у издателя →