Evolutionary-scale prediction of atomic level protein structure with a language model

The work demonstrates that directly inferring protein structure from primary sequence with a large language model enables an order-of-magnitude speedup in high-resolution structure prediction. Leveraging the insight that language models learn evolutionary patterns across millions of sequences, the authors train models up to 15 billion parameters, the largest protein language model to date. As the models scale, they learn information that supports prediction of a protein's three-dimensional structure at the resolution of individual atoms.

The approach is up to 60x faster than the prior state of the art while maintaining resolution and accuracy, making structure prediction feasible at evolutionary scale. Building on this speed, the authors present the ESM Metagenomic Atlas, the first large-scale structural characterization of metagenomic proteins with more than 617 million structures, including over 225 million high-confidence predictions and millions of structures novel relative to experimentally determined ones, opening a view into poorly understood proteins.