esmfold

ESMFold is a sequence-to-structure predictor designed to infer full atomic-level protein folds directly from a single amino acid sequence, without relying on multiple sequence alignments. It is built on transformer protein language models that were scaled up to 15 billion parameters and trained on a mixture of experimental and high-quality predicted protein structures. As the language models grew, information about atomic-resolution structure emerged naturally in their learned representations, which made it possible to translate a sequence straight into coordinates rather than searching evolutionary databases first.

In practice, ESMFold delivers structure predictions at speeds comparable to or faster than alignment-based systems while reaching near-equivalent accuracy on standard benchmarks, which opened the door to proteome-scale applications. The work behind it produced the ESM Metagenomic Atlas, a public database of structures predicted for more than 600 million metagenomic proteins, demonstrating the model's value for environmental and microbiome research. More recently, the broader ESM research line has continued to evolve with efforts such as ESMC and ESMFold2, signaling ongoing investment in larger protein foundation models that can serve both structural biology and downstream protein engineering tasks.