Meta's ESMFold Predicts Protein Structures 60x Faster Without Database Lookups
Meta's single-sequence protein folding model keeps climbing Hugging Face downloads, offering AlphaFold2-level accuracy at up to 60x the speed with no MSA required.
Takeaways
- Meta's ESMFold v1 passes 2.18M downloads on Hugging Face, still trending.
- Predicts atomic-level protein structure directly from sequence, no MSA needed.
- Up to 60x faster than AlphaFold2 at comparable accuracy for well-understood sequences.
- Powers the ESM Metagenomic Atlas with 600M+ predicted structures.
- Ships MIT-licensed via
transformerswithEsmForProteinFolding. - Strong on de novo and engineered proteins where evolutionary homologs do not exist.
Meta's ESMFold v1 has quietly crossed 2.18 million downloads on Hugging Face, and it keeps trending for a reason: the model collapses protein structure prediction into a single forward pass through a language model, with no evolutionary database lookups required. If you have ever folded a protein with AlphaFold2, you know the MSA step alone can take minutes to hours per sequence. ESMFold skips it entirely.
One sequence in, atomic coordinates out
The core idea: a large protein language model can absorb the evolutionary signal that MSA-based pipelines have to construct on the fly. The original Transformer protein language models came from Alexander Rives and colleagues at Meta's Fundamental AI Research team in the paper "Biological structure and function emerge from scaling unsupervised learning to 250 million protein sequences." Scale that model up, bolt on a folding head, and you get end-to-end structure prediction from a raw amino acid string.
Scaling the language model to 15 billion parameters lets the authors study how model size affects the structural knowledge that emerges from pure sequence pretraining. Accuracy tracks the model's grasp of a given sequence: for low-perplexity inputs, ESMFold reaches parity with AlphaFold2 and RoseTTAFold.