# ESMFold

> Source: https://aiwiki.ai/wiki/esmfold
> Updated: 2026-07-28
> Categories: AI Models, AI for Science, Meta AI
> License: CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/) - attribute to "AI Wiki (aiwiki.ai)"
> Cite as: AI Wiki. "ESMFold." aiwiki.ai, 28 Jul 2026. https://aiwiki.ai/wiki/esmfold
> From AI Wiki (https://aiwiki.ai), the free encyclopedia of artificial intelligence. Reuse freely with attribution.

**ESMFold** is a protein structure prediction model developed by the [Meta AI](https://aiwiki.ai/wiki/meta_ai) Fundamental AI Research (FAIR) Protein Team. It predicts an all-atom three-dimensional protein structure directly from an amino-acid sequence. Unlike [AlphaFold](https://aiwiki.ai/wiki/alphafold) 2 and [RoseTTAFold](https://aiwiki.ai/wiki/rosettafold), ESMFold does not build a multiple sequence alignment (MSA), search a template library, or query an external sequence database during inference. It instead passes representations from the ESM-2 protein [masked language model](https://aiwiki.ai/wiki/masked_language_model) to a learned folding network.[^1]

The production ESMFold checkpoints use a 3-billion-parameter ESM-2 model, a 48-block Folding Trunk, and an eight-block structure module. The separate ESM-2 family ranges from 8 million to 15 billion parameters. The 15-billion-parameter model was central to the paper's scaling experiments, but it is not the language-model stem in `esmfold_v0` or `esmfold_v1`.[^2] This distinction matters because descriptions that call ESMFold itself a 15-billion-parameter folder conflate the ESM-2 research family with the released folding checkpoints.

ESMFold was first described in a bioRxiv preprint posted on July 21, 2022. The peer-reviewed article appeared in *Science* on March 17, 2023.[^1][^3] Meta released code and weights in 2022 and used the model to predict structures for more than 617 million metagenomic protein sequences in the first ESM Metagenomic Atlas. The original prediction run took about two weeks on a heterogeneous cluster of roughly 2,000 GPUs.[^1][^4]

## Overview

| Attribute | Value |
|---|---|
| Developer | Meta AI FAIR Protein Team |
| Primary task | Protein structure prediction from amino-acid sequence |
| Production language-model stem | ESM-2 3B, 36 transformer layers |
| Folding network | 48-block Folding Trunk plus eight-block structure module |
| Approximate checkpoint parameters | 3B in ESM-2 plus 690M in the folding network |
| Inference inputs | Protein sequence; no MSA, templates, or external database search |
| Main outputs | Atomic coordinates, pLDDT, pTM, and predicted aligned error |
| Paper model | `esmfold_v0` |
| Recommended original checkpoint | `esmfold_v1` |
| Initial preprint | July 21, 2022 |
| Journal publication | *Science* 379(6637), March 17, 2023 |
| Original code license | MIT |
| Original Atlas data license | CC BY 4.0 |

## Background

Protein language models apply sequence-modeling methods to strings of amino acids. Earlier FAIR work trained a deep [Transformer](https://aiwiki.ai/wiki/transformer) on 250 million protein sequences with a masked-token objective. That study found that representations learned by [unsupervised machine learning](https://aiwiki.ai/wiki/unsupervised_machine_learning) organized proteins by biological relationships and encoded information useful for secondary-structure, contact, and remote-homology tasks.[^5] In modern terminology, the training setup is also an example of [self-supervised learning](https://aiwiki.ai/wiki/self_supervised_learning): the sequence supplies its own labels because the model predicts residues that were hidden from its input.

High-accuracy structure predictors took a different route. AlphaFold 2 combines MSA-derived evolutionary features, structural templates, an Evoformer, and an equivariant structure module. Its CASP14 results established a major accuracy advance for [protein folding](https://aiwiki.ai/wiki/protein_folding) and structure prediction.[^6] RoseTTAFold processes sequence, pairwise, and coordinate information with a three-track network and also uses evolutionary information from MSAs.[^7] These methods can be highly accurate, but their complete pipelines include database searches and MSA construction before neural-network inference.

ESMFold asks whether a sufficiently large [protein language model](https://aiwiki.ai/wiki/language_model) can internalize enough evolutionary regularity during pretraining to support structure prediction from one sequence at inference time. Its answer is practical rather than absolute: ESMFold is faster and independent of external search infrastructure, but its average accuracy in the paper remained below the full AlphaFold 2 pipeline.[^1]

## Architecture

### ESM-2 protein language model

ESM-2 is an encoder-only transformer trained to reconstruct masked amino acids from their sequence context. The paper sampled sequences with even weighting across about 43 million UniRef50 clusters drawn from roughly 138 million UniRef90 sequences. Across training, the models saw about 65 million unique sequences. Fifteen percent of positions were selected by the masked-language-modeling objective.[^1]

The public ESM-2 family contains six principal sizes:[^2]

| Checkpoint family | Parameters | Layers | Hidden dimension |
|---|---:|---:|---:|
| ESM-2 8M | 8M | 6 | 320 |
| ESM-2 35M | 35M | 12 | 480 |
| ESM-2 150M | 150M | 30 | 640 |
| ESM-2 650M | 650M | 33 | 1,280 |
| ESM-2 3B | 3B | 36 | 2,560 |
| ESM-2 15B | 15B | 48 | 5,120 |

The released models use [rotary position embedding](https://aiwiki.ai/wiki/rope). In the Science study, larger ESM-2 models generally produced lower validation perplexity and representations from which more accurate contacts and structures could be projected. A structure-module-only head on the 15B model reached average TM-scores of 0.72 on the paper's CAMEO set and 0.55 on its CASP14 set. The full production system took a different design point: a smaller 3B stem followed by a much deeper folding network.[^1] The result supports a protein-specific example of [scaling laws](https://aiwiki.ai/wiki/scaling_laws), but it does not show that every protein or every downstream task improves monotonically at every model size.

### From ESM-2 representations to a folding network

The official implementation freezes the ESM-2 weights. It computes hidden states from every ESM-2 layer, learns a weighted combination of those states, and projects the result into the Folding Trunk's sequence representation. The public v1 configuration sets `use_esm_attn_map` to false, so the production Hugging Face configuration does not directly initialize its pair representation from ESM-2 attention maps.[^8][^9] Attention maps were important evidence in the paper's analysis of learned contacts, but that scientific probe should not be confused with the production v1 input path.

The Folding Trunk maintains a per-residue sequence state and a residue-pair state. Each of its 48 blocks lets information pass between these tracks through sequence attention, triangle updates, and pairwise operations. Relative residue positions are added to the pair state. This resembles parts of AlphaFold 2's Evoformer, but without an MSA dimension or an extra MSA stack.[^1][^8]

An eight-block structure module converts the trunk output into rigid frames, torsion angles, and atom coordinates. The original code uses the [OpenFold](https://aiwiki.ai/wiki/openfold) implementation of the AlphaFold-style structure module. It also produces pLDDT, pTM, and predicted aligned error outputs.[^8] The Science paper describes three recycling steps after the first pass. The public v1 configuration represents the corresponding maximum as four trunk passes, and the original command-line interface describes four as its default.[^1][^2][^9]

### Training the folding head

The folding head was trained on about 25,000 clusters covering approximately 325,000 experimentally determined structures, augmented with about 12 million structures predicted by AlphaFold 2. The paper model used a May 2020 structural cutoff so that its 194 CAMEO and 51 CASP14 evaluation structures were later than its training set. The folding network used the same broad family of losses as AlphaFold.[^1]

The public source freezes and detaches the ESM-2 stem rather than lightly fine-tuning it end to end. Trainable components include the mixture over ESM-2 layers, input projections, Folding Trunk, structure module, and confidence heads.[^8] The baseline claim that the production stem was lightly fine-tuned is therefore not retained.

## Evaluation

### Accuracy

The paper evaluated ESMFold on 194 CAMEO structures from April to June 2022 and 51 publicly released CASP14 structures selected for competition difficulty. It reported the following average TM-scores:[^1]

| Method in the Science comparison | CAMEO TM-score | CASP14 TM-score | Inference information |
|---|---:|---:|---|
| AlphaFold 2 with released search protocol | 0.88 | 0.85 | Sequence, MSA, and templates |
| ESMFold paper model | 0.83 | 0.68 | Sequence only |
| RoseTTAFold | 0.82 | Not reported in this comparison | Sequence and MSA |

These values are tied to the paper's target selections and protocols. They should not be combined with values from unrelated datasets or later model versions as though they were one controlled benchmark. The paper also tested AlphaFold 2 and RoseTTAFold after removing their MSAs, but described that setup as artificial because AlphaFold 2 had not been trained for single-sequence use.[^1]

ESMFold's performance varied with how well ESM-2 modeled a sequence. The paper reported a negative correlation between ESM-2 perplexity and TM-score within both evaluation sets. On 18 CASP14 proteins for which ESM-2 perplexity was below 7, ESMFold and AlphaFold 2 differed by less than 0.03 in average TM-score, with no target differing by more than 0.1. This was a selected subset, not the average result over CASP14.[^1]

### Speed

On one NVIDIA V100 GPU, the paper measured 14.2 seconds for a 384-residue ESMFold prediction, about six times faster than one AlphaFold 2 neural-network model. For shorter proteins, the forward-pass speedup reached about 60 times. The paper treated MSA search as an additional end-to-end advantage: the high-sensitivity search protocols used by the published AlphaFold 2 and RoseTTAFold pipelines could take more than ten minutes before folding began.[^1]

Hardware, sequence length, batching, recycling, precision, and attention chunking all affect runtime. The 60 times figure is therefore a benchmark maximum for shorter sequences, not a universal speed ratio for every ESMFold and AlphaFold deployment.

### Confidence

ESMFold reports pLDDT as a per-residue estimate of local structural accuracy and pTM as a whole-structure confidence estimate. In common PDB output, pLDDT uses a 0 to 100 scale. A pLDDT above 90 generally indicates very high local confidence, while a value above 70 often corresponds to a locally correct backbone. Values below 50 can occur in flexible or intrinsically disordered regions, but they can also mean that the model lacks enough information. High pLDDT across separate domains does not establish that their relative orientation is correct.[^10]

The ESMFold paper found its confidence estimates to be calibrated on the held-out CAMEO set. For predictions with normalized pLDDT above 0.7, average LDDT approached the AlphaFold 2 result on that set. This is a statistical relationship, not experimental validation of any individual prediction.[^1]

## Checkpoints and implementations

### `esmfold_v0` and `esmfold_v1`

The original Meta repository distinguishes two full folding checkpoints:[^2]

- `esmfold_v0` uses the 3B ESM-2 stem and 48 folding blocks. It is the model used for the paper's experiments and the first Atlas release. Its PDB training cutoff was May 2020.
- `esmfold_v1` uses the same 3B stem and 48-block layout. Meta recommends it over v0, and the March 2023 Atlas update used it.

The repository also exposes structure-module-only ablation checkpoints at multiple ESM-2 sizes. Its documentation explicitly says those research ablations are not recommended for normal structure prediction.[^2]

Meta released the original implementation through `facebookresearch/esm` under the MIT license. The repository was archived as read-only on August 1, 2024, but the files and checkpoint URLs remain available. The paper's v0 code and weights were also archived on Zenodo.[^2][^11]

### Hugging Face Transformers

[Hugging Face](https://aiwiki.ai/wiki/hugging_face) added ESMFold as `EsmForProteinFolding`, with `facebook/esmfold_v1` as the public v1 checkpoint. The port uses parts of OpenFold and provides a conventional `from_pretrained` [inference path](https://aiwiki.ai/wiki/inference_path). It preserves the basic 3B ESM-2, 48-block trunk, and eight-block structure-module configuration.[^9] The integration is an implementation of ESMFold, not evidence that ESMFold itself uses every ESM-2 model size.

### Multiple chains

ESMFold was developed as a single-sequence predictor and was not trained on protein complexes. The Science paper nevertheless evaluated it on 2,978 recent multimeric complexes and found the same qualitative DockQ category as AlphaFold-Multimer for 53.2 percent of chain pairs.[^1] The public v1 implementation also accepts multiple chains separated by a colon, inserts a glycine linker internally, and offsets residue indices between chains.[^2][^8]

This interface means that "monomer only" is no longer an accurate description of the released code. It does not make ESMFold a complex-trained replacement for AlphaFold-Multimer, and its interface does not accept ligands, nucleic acids, ions, or post-translational modifications.

## ESM Metagenomic Atlas

The first ESM Metagenomic Atlas applied ESMFold to more than 617 million sequences from MGnify90. The paper included sequences from 20 to 1,024 residues, covering 99 percent of that MGnify90 collection. The run produced about 365 million predictions above the paper's "good confidence" threshold of normalized mean pLDDT above 0.5 and pTM above 0.5. About 225 million met both values above 0.7, and about 113 million met the paper's very-high-confidence threshold.[^1]

The authors used Foldseek to compare a random sample of one million high-confidence predictions with experimental structures in the Protein Data Bank. At a TM-score threshold of 0.5, 12.6 percent of that sample had no structural match. The result is an estimate on a filtered sample, not evidence that every unmatched prediction represents a new fold.[^1]

Meta launched Atlas v0 on November 1, 2022. A March 2023 update associated with MGnify `2023_02` raised the site's stated total from 617 million to 772 million structures and added precomputed ESM-2 embeddings.[^4][^12] The original Atlas website provides structure viewing, sequence and structure search, bulk-download information, and a hosted fold-sequence interface. Its documentation distinguishes local pLDDT from global pTM when filtering predictions.[^12]

The Atlas is an application of ESMFold at scale, while its embeddings are outputs of ESM-2. Treating all Atlas features as capabilities of the folding head would blur that distinction.

## Uses

### High-throughput structural hypotheses

ESMFold is useful when a project has many protein sequences and the database-search stage of an MSA-based pipeline would dominate cost or operations. The Atlas demonstrates this use at metagenomic scale. On smaller projects, researchers can use ESMFold predictions to prioritize sequences, compare predicted topologies, or prepare hypotheses for experimental follow-up. A predicted structure remains a model and should be checked against confidence outputs, alternative predictors, biochemical evidence, and experiments when the conclusion is consequential.

The same speed can make ESMFold useful inside iterative protein-design workflows. A 2022 protein-programming preprint from FAIR used ESMFold as part of a system that generated and evaluated designed proteins.[^13] Such use does not make ESMFold an experimental oracle: confidence and agreement with a designed backbone are filters, not proof that a candidate expresses, folds, binds, or functions in a cell.

### Downstream modeling

Predicted coordinates can be inputs to structural search, visualization, site annotation, and other [AI for Science](https://aiwiki.ai/wiki/ai_for_science) workflows. They can also inform early computational work in [AI in drug discovery](https://aiwiki.ai/wiki/ai_in_drug_discovery). ESMFold itself does not predict a ligand pose or binding affinity, however, so a protein-only output should not be described as a drug-discovery result.

ESM-2 embeddings are separately used for sequence representation, contact analysis, and other transfer-learning tasks. Those are capabilities of the underlying [large language model](https://aiwiki.ai/wiki/large_language_model), not outputs of ESMFold's coordinate head. Likewise, zero-shot variant scoring is associated with ESM sequence models such as ESM-1v and ESM-2, not with the folding checkpoint as such.[^2][^5]

## Limitations

The main tradeoff is visible in the paper's own controlled comparison. ESMFold avoided MSA and template search and ran much faster, but AlphaFold 2 had higher average TM-score on both CAMEO and CASP14.[^1] ESMFold should therefore be chosen for its input assumptions and throughput, not because it universally matches the best MSA-based accuracy.

Other limitations follow from the released input and output design:

- ESMFold v1 accepts amino-acid chains, not ligands, DNA, RNA, ions, or modified residues.
- Its multichain mode is available in code, but the model was not trained on protein complexes.
- pLDDT is local confidence. It cannot by itself establish domain orientation, interface correctness, biological assembly, or function.[^10]
- The model returns a structural prediction rather than an experimentally observed conformational ensemble. Flexible states, alternative folds, membrane context, binding partners, and assay conditions may change the biologically relevant structure.
- Sequence and pair attention increase memory use with protein length. The original implementation exposes attention chunking and CPU offload for long sequences, with a speed cost.[^2]

These points are especially important when predictions are used in protein engineering or downstream [diffusion model](https://aiwiki.ai/wiki/diffusion_model) pipelines. A fast internal score can pass along a folding model's blind spots.

## Release history and successors

| Date | Event |
|---|---|
| July 21, 2022 | Initial ESM-2 and ESMFold bioRxiv preprint posted |
| November 1, 2022 | Meta launched Atlas v0 and released ESMFold code and weights |
| March 17, 2023 | Peer-reviewed Science article and Atlas `v2023_02` update |
| August 1, 2024 | Original `facebookresearch/esm` repository archived as read-only |
| 2025 | [EvolutionaryScale](https://aiwiki.ai/wiki/evolutionaryscale) authors published [ESM3](https://aiwiki.ai/wiki/esm3), a generative model over protein sequence, structure, and function |
| May 2026 | Biohub released ESMFold2 in a separate ESM codebase; the associated preprint followed in June |

ESM3 is related by research lineage but is not a version of the original ESMFold architecture. It is a multimodal generative [language model](https://aiwiki.ai/wiki/language_model) for protein sequence, structure, and function.[^14] [AlphaFold 3](https://aiwiki.ai/wiki/alphafold_3), developed by [Google DeepMind](https://aiwiki.ai/wiki/google_deepmind) and Isomorphic Labs, likewise addresses a broader task than ESMFold by modeling complexes containing proteins, nucleic acids, small molecules, ions, and modified residues with a substantially different architecture.[^15]

Biohub's 2026 ESM repository describes ESMFold2 as a diffusion-based successor built on ESMC, with optional MSA input and support for broader biomolecular inputs. It is a separate model and should not be used to retroactively describe the capabilities or benchmarks of ESMFold v0 or v1.[^16]

## See also

- [Protein structure prediction](https://aiwiki.ai/wiki/protein_structure_prediction)
- [AlphaFold](https://aiwiki.ai/wiki/alphafold)
- [RoseTTAFold](https://aiwiki.ai/wiki/rosettafold)
- [ESM3](https://aiwiki.ai/wiki/esm3)
- [Hugging Face Transformers](https://aiwiki.ai/wiki/transformers_library)
- [Unsupervised learning](https://aiwiki.ai/wiki/unsupervised_learning)

## References

[^1]: Zeming Lin et al. ["Evolutionary-scale prediction of atomic-level protein structure with a language model."](https://doi.org/10.1126/science.ade2574) *Science* 379, no. 6637 (2023): 1123-1130. Accessible article PDF mirror used for line-level verification: https://yiheng-zhu.github.io/Yiheng/papers/5/ESMFold_Science_2023.pdf
[^2]: Meta FAIR. ["facebookresearch/esm: Evolutionary Scale Modeling."](https://github.com/facebookresearch/esm) Archived GitHub repository, accessed July 28, 2026. The README identifies the 3B ESM-2 stem, 48 folding blocks, approximately 690M folding-network parameters, v0 and v1 roles, model-family sizes, multichain input, default recycling, licenses, and repository status. See also the repository's ["ESM Metagenomic Atlas"](https://github.com/facebookresearch/esm/blob/main/scripts/atlas/README.md) notes for the v1 Atlas update.
[^3]: Zeming Lin et al. ["Language models of protein sequences at the scale of evolution enable accurate structure prediction."](https://www.biorxiv.org/content/10.1101/2022.07.20.500902v1) bioRxiv 2022.07.20.500902, posted July 21, 2022.
[^4]: Meta AI. ["ESM Metagenomic Atlas: The first view of the 'dark matter' of the protein universe."](https://ai.meta.com/blog/protein-folding-esmfold-metagenomics/) November 1, 2022.
[^5]: Alexander Rives et al. ["Biological structure and function emerge from scaling unsupervised learning to 250 million protein sequences."](https://pmc.ncbi.nlm.nih.gov/articles/PMC8053943/) *Proceedings of the National Academy of Sciences* 118, no. 15 (2021): e2016239118. DOI: https://doi.org/10.1073/pnas.2016239118
[^6]: John Jumper et al. ["Highly accurate protein structure prediction with AlphaFold."](https://www.nature.com/articles/s41586-021-03819-2) *Nature* 596 (2021): 583-589. DOI: https://doi.org/10.1038/s41586-021-03819-2
[^7]: Minkyung Baek et al. ["Accurate prediction of protein structures and interactions using a three-track neural network."](https://doi.org/10.1126/science.abj8754) *Science* 373, no. 6557 (2021): 871-876.
[^8]: Meta FAIR. ["ESMFold implementation."](https://github.com/facebookresearch/esm/blob/main/esm/esmfold/v1/esmfold.py) See also the ["Folding Trunk configuration and implementation"](https://github.com/facebookresearch/esm/blob/main/esm/esmfold/v1/trunk.py) and ["checkpoint loader"](https://github.com/facebookresearch/esm/blob/main/esm/esmfold/v1/pretrained.py), archived GitHub source, accessed July 28, 2026.
[^9]: Hugging Face. ["ESM model documentation."](https://huggingface.co/docs/transformers/model_doc/esm) See also the explicit [`facebook/esmfold_v1` configuration](https://huggingface.co/facebook/esmfold_v1/blob/main/config.json), accessed July 28, 2026.
[^10]: EMBL-EBI Training. ["pLDDT: Understanding local confidence."](https://www.ebi.ac.uk/training/online/courses/alphafold/inputs-and-outputs/evaluating-alphafolds-predicted-structures-using-confidence-scores/plddt-understanding-local-confidence/) AlphaFold practical guide, accessed July 28, 2026.
[^11]: Zeming Lin et al. ["ESM-2 and ESMFold-v0 Model Code and Weights."](https://doi.org/10.5281/zenodo.7566741) Zenodo, 2023.
[^12]: Meta AI. ["About the ESM Metagenomic Atlas."](https://esmatlas.com/about) Updated March 2023, accessed July 28, 2026. See also EMBL-EBI, ["MGnify just got bigger: new data, new AI opportunities,"](https://www.ebi.ac.uk/about/news/updates-from-data-resources/mgnify-meta-ai/) March 17, 2023.
[^13]: Brian Hie et al. ["A high-level programming language for generative protein design."](https://doi.org/10.1101/2022.12.21.521526) bioRxiv 2022.12.21.521526, 2022.
[^14]: Thomas Hayes et al. ["Simulating 500 million years of evolution with a language model."](https://doi.org/10.1126/science.ads0018) *Science* (2025).
[^15]: Josh Abramson et al. ["Accurate structure prediction of biomolecular interactions with AlphaFold 3."](https://www.nature.com/articles/s41586-024-07487-w) *Nature* 630 (2024): 493-500. DOI: https://doi.org/10.1038/s41586-024-07487-w
[^16]: Biohub. ["A world model of protein biology: ESMC, ESMFold2, and ESM Atlas."](https://github.com/Biohub/esm) GitHub repository, accessed July 28, 2026. The repository cites the associated [2026 preprint](https://www.biorxiv.org/content/10.64898/2026.06.03.729735). Release timing is corroborated by Ewen Callaway and Miryam Naddaf, ["Move over, AlphaFold: open-source model predicts shape of 1 billion proteins,"](https://doi.org/10.1038/d41586-026-01686-3) *Nature*, May 27, 2026.