ESM-2

ESM-2 is Meta's Fundamental AI Research protein language model family, trained on hundreds of millions of protein sequences with masked language modeling. Scaling unsupervised learning to 250M+ sequences let biological structure and function emerge from the language model alone: the 15B-parameter ESM-2, paired with a folding head as ESMFold, generates accurate 3D structure predictions end-to-end directly from a single amino acid sequence — no multiple-sequence alignments required.

Released in 2022 with open weights and code under the MIT license, ESM-2 outperformed all tested single-sequence protein language models on structure benchmarks and powered the ESM Metagenomic Atlas of 617 million predicted structures. It remains the workhorse protein LM behind countless downstream tools in drug discovery, enzyme engineering, and variant-effect prediction.