ESMFold

Introduced by Meta AI in late 2022 alongside the ESM-2 protein language model (Science, 2023), ESMFold demonstrated that a protein language model trained on millions of unaligned sequences implicitly encodes enough evolutionary and structural signal to fold proteins end-to-end. By feeding ESM-2 embeddings directly into a folding head instead of running an expensive multiple-sequence-alignment search, it runs roughly 60 times faster than AlphaFold2 — fast enough to fold the hundreds of millions of sequences in the ESM Metagenomic Atlas.

Code and weights are MIT-licensed via the facebookresearch/esm repository. The tradeoff for speed is accuracy: ESMFold trails AlphaFold2, especially on proteins where a deep alignment would help, and it outputs a single structure with slightly worse-calibrated confidence. It is the right tool for metagenomic-scale screening and fast triage, not for the highest-stakes single-target structure determination.