ESMFold

ESMFold predicts all-atom protein structure from a single amino acid sequence, with no multiple sequence alignment and no template search anywhere in the pipeline. It was developed by the protein team at Meta AI and published by Lin et al. in Science in 2023 alongside ESM-2, the protein language model family whose representations it reads.

The design follows from a finding about ESM-2: as those models scale, an atomic-resolution picture of structure emerges in their internal states purely from masked language modeling over evolutionary sequence data. If the picture is already there, a folding network need not rebuild evolutionary context from an alignment at inference time — it can read it out of the language model instead. ESMFold does exactly that, training a folding head on experimental structures on top of a frozen 3B-parameter ESM-2 encoder.

Dropping the alignment changes the economics of folding. The homology search behind a high-sensitivity MSA can take over ten minutes per sequence in the published AlphaFold 2 and RoseTTAFold pipelines. ESMFold pays none of it, which made it practical to fold an entire metagenomic database: the ESM Metagenomic Atlas, 617 million sequences from MGnify90.

#Key Features

• No alignment at inference: Structure comes from the primary sequence alone, eliminating the homology search and the MSA-conditioned attention machinery of conventional folding pipelines.
• Simplified folding trunk: Folding blocks alternate updates to a sequence representation and a pairwise representation, then hand off to an equivariant transformer structure module with three recycling steps.
• Calibrated confidence: Predictions above pLDDT 0.9 reach a median all-atom RMSD95 of 1.42 Å on CAMEO, and on metagenomic sequences pLDDT correlates with agreement to AlphaFold 2 at Pearson r = 0.79.
• Incidental complex prediction: Never trained on complexes, ESMFold still assigns the same qualitative DockQ category as AlphaFold-Multimer for 53.2% of chain pairs across 2,978 recent multimeric structures from the PDB.
• ESM Metagenomic Atlas: 617 million predicted structures released as an open resource, searchable by sequence and by structure and downloadable in bulk.

#Technical Details

ESMFold pairs a folding head of roughly 690 million parameters with the 3B ESM-2 encoder — the largest in that family that still permits inference on a single GPU — whose weights stay frozen throughout training. A learned weighted sum across encoder layers seeds the sequence state, and 48 folding blocks refine it alongside a pairwise state before the structure module emits coordinates and per-residue confidence. The released checkpoints are esmfold_v1 (esmfold_3B_v1.pt) and the earlier esmfold_v0. Training used AlphaFold 2's loss formulation on roughly 25,000 clusters covering about 325,000 experimentally determined PDB structures, plus a distillation set of about 12 million AlphaFold 2 predictions.

Evaluation applies a May 2020 temporal cutoff, so test structures are held out. ESMFold averages a TM-score of 0.83 on a 194-structure CAMEO test set and 0.68 on the 51 CASP14 targets, against 0.88 and 0.85 for AlphaFold 2 with full MSAs and templates, and 0.82 for RoseTTAFold on CAMEO. Ablating the alignment drops both of those below ESMFold, and removing the language model costs more than removing the folding trunk: with an eight-block trunk, CAMEO LDDT falls from 0.74 to 0.58 without the language model and to 0.66 with the trunk deleted. Accuracy tracks language model perplexity at Pearson -0.55 on CAMEO and -0.67 on CASP14. On a single NVIDIA V100 a 384-residue protein folds in 14.2 seconds, six times faster than one AlphaFold 2 model on the forward pass alone and up to roughly 60x on shorter sequences, before the alignment search is counted.

#Applications

ESMFold is built for structural characterization at throughput. Metagenomic and environmental sequencing projects routinely produce millions of sequences from organisms with no sequenced relatives, where MSA construction is either infeasible or returns nothing useful; ESMFold gives those datasets a structural triage pass that alignment-based predictors cannot. The same property makes it a common refolding step in protein design loops, where a sequence generated by an inverse-folding model such as ProteinMPNN has no evolutionary family by construction and self-consistency has to be scored thousands of times per campaign.

#Impact

The ESM Metagenomic Atlas made structural biology tractable at metagenomic scale: in a sample of a million high-confidence structures, 12.6% had no experimentally determined counterpart in the PDB above TM-score 0.5. ESMFold itself became infrastructure for the protein design community, where its speed and its indifference to evolutionary depth make it a natural scoring step in generate-and-filter loops. Its clearest limitation is accuracy on hard targets: it trails AlphaFold 2 on CASP14, and is weakest exactly where the language model finds a sequence surprising. ESMFlow, the companion model in the AlphaFlow paper, adds conformational ensembles on top of ESMFold, and the lineage continued at Biohub with ESMFold2, which folds full biomolecular complexes from ESMC representations.