ESMFold¶

Single-sequence protein structure prediction model that uses the ESM-2 language model backbone to predict 3D atomic coordinates without requiring multiple sequence alignments.

License: MIT · Molecules: protein, complex · Tasks: structure_prediction

Actions: fold · Variants: 1

At a glance¶

Use it when

You need to screen thousands to millions of protein sequences for structural quality and foldability, and speed matters more than maximum accuracy per prediction
You are performing initial structural triage before investing in more expensive methods -- use ESMFold to identify promising candidates, then run AlphaFold2 or Chai-1 on the top hits
You need rapid structural assessment of protein engineering variants to check whether mutations disrupt the native fold before wet-lab testing
You are annotating metagenomic or large genomic datasets where no MSA is readily available and computing MSAs for every sequence is prohibitively expensive
You need a fast structural checkpoint in an automated pipeline -- ESMFold's seconds-per-prediction throughput enables structure-aware filtering in high-throughput design workflows
You are working with single-chain, soluble, globular proteins where the protein family is well-represented in UniRef50 -- ESMFold accuracy approaches AlphaFold2 for these targets
You need a cost-effective structure prediction that runs on modest GPU hardware (A10G) without specialized infrastructure for MSA databases

Strengths

Fastest structure prediction in this catalog -- single-sequence inference in seconds without any MSA computation, enabling proteome-scale structural screening of millions of sequences
No external database dependencies -- the ESM-2 3B language model backbone has already internalized co-evolutionary signals from UniRef50, eliminating the MSA bottleneck entirely
Runs on a single A10G GPU with 16GB RAM, making it the most cost-effective structure prediction model in this catalog -- roughly 10-50x cheaper per prediction than diffusion-based methods
Per-residue pLDDT confidence scores reliably triage prediction quality: pLDDT > 70 correlates with TM-score > 0.8 for well-represented protein families, enabling automated quality filtering
Supports multi-chain complexes (up to 4 chains via colon separator) with a combined 768-residue limit, covering many biologically relevant homo- and hetero-dimers
MIT license with no usage restrictions -- suitable for commercial pipelines, clinical applications, and large-scale academic studies
Deterministic inference with fixed recycling (4 recycles by default) ensures reproducible results across runs, critical for screening campaigns and benchmarking
Proven at scale: the original paper predicted structures for over 600 million metagenomic protein sequences, demonstrating industrial-level throughput

Limitations

Lower accuracy than MSA-based methods (AlphaFold2, Chai-1, RF3) on proteins with distant homologs -- single-sequence approach misses explicit co-evolutionary coupling signals from alignments
Maximum total sequence length of 768 residues limits applicability for large multi-domain proteins and complexes -- many biologically important proteins exceed this threshold
Multi-chain complex predictions are significantly less accurate than single-chain, with interface contacts being particularly unreliable compared to dedicated complex predictors
Cannot model non-protein molecules -- no support for DNA, RNA, small-molecule ligands, or post-translational modifications that are critical for drug discovery workflows
Membrane protein predictions are systematically weaker due to under-representation in training data; transmembrane helical bundles are often incorrectly packed
Intrinsically disordered regions are assigned low pLDDT (correctly indicating uncertainty) but the predicted coordinates are meaningless -- not suitable for studying disordered protein ensembles
Relative domain orientations in multi-domain proteins are often incorrect, especially when connected by flexible linkers -- each domain may be individually correct but the assembly is wrong
Very short peptides (< 30 residues) lack sufficient context for the ESM-2 backbone to generate meaningful structural representations

Reach for something else when

You need the highest possible structural accuracy for a specific target -- use AlphaFold2 for MSA-based protein prediction or Chai-1/RF3 for multi-molecule complexes
You need to predict protein-ligand, protein-DNA, or protein-RNA complex structures -- ESMFold is protein-only; use Chai-1 or RF3 for multi-molecule complexes
Your protein exceeds 768 total residues or has multiple large domains with flexible interdomain linkers -- ESMFold cannot handle the length and inter-domain orientation will be unreliable
You are studying intrinsically disordered proteins or conformational ensembles -- ESMFold produces a single static structure that is not biologically meaningful for disordered regions
You need membrane protein structure with lipid bilayer context -- ESMFold under-represents membrane proteins and cannot model the membrane environment
You need high-accuracy protein-protein interface prediction -- dedicated complex predictors (Chai-1, RF3) are substantially more reliable for interface geometry
You need binding pose prediction for drug discovery -- ESMFold cannot dock ligands; use Chai-1 or RF3 instead

Alternatives

Model	Better when	Worse when
prostt5	ProstT5 derives Foldseek 3Di structural tokens from sequence three orders of magnitude faster than atomic-level folding, enabling rapid structural annotation and Foldseek structure search across large databases	ProstT5 produces a coarse structural alphabet (3Di) rather than full 3D atomic coordinates; ESMFold predicts complete atomic structures with per-residue pLDDT confidence
chai1	Multi-molecule co-folding (protein + DNA + RNA + ligand) with diffusion architecture; superior accuracy on heterogeneous biomolecular complexes and protein-ligand docking	Requires A100 80GB GPU with 64GB system RAM; significantly slower and more expensive; no advantage for simple single-chain protein structure prediction at scale
rf3	Newest structure prediction model with highest accuracy on multi-molecule benchmarks; supports templates, cyclic peptides, and diverse conditioning; diffusion sampling provides structural diversity	Requires A100 40GB GPU and is orders of magnitude slower; the additional accuracy is unnecessary for large-scale screening where ESMFold's speed-accuracy tradeoff is optimal

API & schema¶

Call an action with POST /api/v1/{slug}/{action} — the request envelope is {"items": [...], "params": {...}} and a success returns {"results": [...]}. See the HTTP API page for the base URL, error shape, and full contract.

`fold`¶

Call it

curl -X POST http://127.0.0.1:8000/api/v1/esmfold/fold \
  -H "Content-Type: application/json" \
  -d '{
  "items": [
    {
      "sequence": "MKTAYIAKQRQISFVKSHFSRQLEERLGLIE"
    }
  ]
}'

Request — ESMFoldRequest

Field	Type	Required	Constraints	Description
`items`	list[ESMFoldRequestItem]	yes	items 1–2	Batch of inputs to process in a single request. Up to 2 sequences per request.

Nested types

ESMFoldRequestItem

Field	Type	Required	Constraints	Description
`sequence`	string	yes	len 1–771	A protein sequence in single-letter amino-acid codes; for complexes, separate up to 4 chains with ":" (768 residues total).

Raw JSON Schema

{
  "$defs": {
    "ESMFoldRequestItem": {
      "additionalProperties": false,
      "properties": {
        "sequence": {
          "description": "A protein sequence in single-letter amino-acid codes; for complexes, separate up to 4 chains with \":\" (768 residues total).",
          "maxLength": 771,
          "minLength": 1,
          "title": "Sequence",
          "type": "string"
        }
      },
      "required": [
        "sequence"
      ],
      "title": "ESMFoldRequestItem",
      "type": "object"
    }
  },
  "additionalProperties": false,
  "properties": {
    "items": {
      "description": "Batch of inputs to process in a single request. Up to 2 sequences per request.",
      "items": {
        "$ref": "#/$defs/ESMFoldRequestItem"
      },
      "maxItems": 2,
      "minItems": 1,
      "title": "Items",
      "type": "array"
    }
  },
  "required": [
    "items"
  ],
  "title": "ESMFoldRequest",
  "type": "object"
}

Response — ESMFoldResponse

Field	Type	Required	Constraints	Description
`results`	list[ESMFoldResponseResult]	yes		Per-input results, returned in the same order as the request items.

Nested types

ESMFoldResponseResult

Field	Type	Required	Description
`pdb`	string	yes	Predicted structure in PDB format.
`mean_plddt`	number	yes	Mean per-residue pLDDT confidence score (0–100); higher values indicate more confident predictions.
`ptm`	number	yes	Predicted TM-score (pTM) for the overall structure (0–1).

Raw JSON Schema

{
  "$defs": {
    "ESMFoldResponseResult": {
      "properties": {
        "pdb": {
          "description": "Predicted structure in PDB format.",
          "title": "Pdb",
          "type": "string"
        },
        "mean_plddt": {
          "description": "Mean per-residue pLDDT confidence score (0\u2013100); higher values indicate more confident predictions.",
          "title": "Mean Plddt",
          "type": "number"
        },
        "ptm": {
          "description": "Predicted TM-score (pTM) for the overall structure (0\u20131).",
          "title": "Ptm",
          "type": "number"
        }
      },
      "required": [
        "pdb",
        "mean_plddt",
        "ptm"
      ],
      "title": "ESMFoldResponseResult",
      "type": "object"
    }
  },
  "properties": {
    "results": {
      "description": "Per-input results, returned in the same order as the request items.",
      "items": {
        "$ref": "#/$defs/ESMFoldResponseResult"
      },
      "title": "Results",
      "type": "array"
    }
  },
  "required": [
    "results"
  ],
  "title": "ESMFoldResponse",
  "type": "object"
}

Usage¶

One-line summary: Single-sequence protein structure prediction model that uses the ESM-2 language model backbone to predict 3D atomic coordinates without requiring multiple sequence alignments.

Overview¶

ESMFold is a protein structure prediction model developed by Meta AI's Fundamental AI Research (FAIR) team. It predicts full-atom 3D protein structures directly from amino acid sequences, bypassing the multiple sequence alignment (MSA) step that makes AlphaFold2 and similar methods slow. ESMFold achieves this by coupling the ESM-2 protein language model (3B parameters) with a folding module derived from AlphaFold2's architecture.

The key advantage of ESMFold is speed: predictions take seconds rather than minutes or hours, making it suitable for large-scale structural screening and rapid prototyping. While its accuracy is somewhat lower than MSA-dependent methods, it provides reliable predictions for most well-characterized protein families.

ESMFold is described in the same paper as ESM-2: Lin et al., "Evolutionary-scale prediction of atomic-level protein structure with a language model," Science (2023). The model shares the ESM-2 language model backbone with the ESM-2 embedding model available in this catalog.

Architecture¶

Property	Value
Architecture	ESM-2 transformer encoder + AlphaFold2-derived folding trunk
Language model backbone	ESM-2 3B (36 layers, 2560 hidden dim)
Total parameters	~3B
Training data	UniRef50 (language model) + PDB + AlphaFold2 distillation (folding trunk)
Max sequence length	768 residues
Max chains	4 (concatenated with `:` separator)
License	MIT

See MODEL.md for detailed architecture description.

Capabilities & Limitations¶

CAN be used for: - Predicting 3D protein structures from single amino acid sequences - Multi-chain protein complex prediction (up to 4 chains, 768 residues total) - Rapid structural screening of large protein sets - Assessing fold confidence via pLDDT and pTM scores - Quick structural validation of protein engineering candidates

CANNOT be used for: - Protein-ligand complex prediction (use Chai-1 instead) - Nucleic acid structure prediction (use Chai-1 instead) - Sequences longer than 768 residues - Complexes with more than 4 chains - Generating protein embeddings (use ESM-2 encode action instead) - Sequence design or generation

Other considerations: - Accuracy is lower than MSA-dependent methods (AlphaFold2, Boltz) for most targets - Performance degrades for proteins with few homologs in UniRef50 - The model uses GPU memory snapshots for fast cold starts - Batch size is capped at 2 sequences per request - CUDA out-of-memory errors for long sequences raise a server error rather than returning a silent empty result

Confidence Metrics¶

Metric	Range	Interpretation
pLDDT (per-residue, reported as mean_plddt)	0-100	> 90: very high confidence; 70-90: confident; 50-70: low confidence; < 50: likely disordered or misfolded
pTM	0-1	> 0.8: high confidence in overall fold; 0.5-0.8: moderate confidence; < 0.5: fold topology may be incorrect

Usage Examples¶

# Single-chain structure prediction
from models.esmfold.schema import (
    ESMFoldRequest,
    ESMFoldRequestItem,
)

single_chain_request = ESMFoldRequest(
    items=[
        ESMFoldRequestItem(
            sequence="MKTVRQERLKSIVRILERSKEPVSGAQLAEELSVSRQVIVQDIAYLRSLGYNIVATPRGYVLAGG"
        ),
    ],
)

# Multi-chain complex prediction (chains separated by ":")
multi_chain_request = ESMFoldRequest(
    items=[
        ESMFoldRequestItem(
            sequence="MKTVRQERLKSIVRILERSKEPVSGAQ:LAEELSVSRQVIVQDIAYLRSLGYN"
        ),
    ],
)

Architecture & training¶

Architecture¶

Model Type & Innovation¶

ESMFold is an end-to-end protein structure prediction model that couples a large protein language model (ESM-2) with a structure prediction module inspired by AlphaFold2's folding trunk. The key innovation is that ESMFold requires only a single protein sequence as input -- no multiple sequence alignment (MSA) is needed. This eliminates the computationally expensive homology search step that dominates the wall-clock time of MSA-dependent methods like AlphaFold2 and RoseTTAFold.

The architecture consists of three stages: 1. ESM-2 language model backbone: A 36-layer, 3B-parameter transformer encoder (esm2_t36_3B_UR50D) processes the input sequence and produces per-residue embeddings. 2. Folding trunk: A series of structure module blocks (adapted from OpenFold/AlphaFold2) transform the language model embeddings into single and pairwise representations using invariant point attention (IPA). The trunk performs iterative refinement ("recycling") to progressively improve the predicted structure. 3. Structure module: Converts the refined representations into 3D atomic coordinates, producing a full-atom protein structure with per-residue confidence scores (pLDDT) and a predicted TM-score (pTM).

The central insight is that the ESM-2 language model, trained purely on evolutionary sequence data, already captures sufficient structural information in its internal representations to predict 3D structure without explicit evolutionary covariance signals from MSAs.

Parameters & Layers¶

Component	Details
Language model	ESM-2 (esm2_t36_3B_UR50D), 36 layers, 3B parameters
Folding trunk	Structure module blocks with invariant point attention
Total parameters	~3B (dominated by the ESM-2 backbone)
Hidden dimensions	2560 (ESM-2 backbone)
Attention heads	40 (ESM-2 backbone)
Recycling iterations	4 (default in BioLM implementation via `num_recycles=4`)
Output	Full-atom PDB coordinates, pLDDT, pTM

Training Data¶

Property	Details
Language model pre-training	UniRef50 (same as ESM-2)
Structure supervision	Experimentally determined structures from the Protein Data Bank (PDB)
Distillation	AlphaFold2-predicted structures from the AlphaFold Protein Structure Database

The language model backbone (ESM-2) is pre-trained on UniRef50, covering proteins from all domains of life. The folding trunk is then trained on experimental PDB structures, with additional distillation from AlphaFold2 predictions to increase the diversity of training structures.

Known biases: - Structures in PDB are biased toward well-ordered globular proteins that crystallize easily - Membrane proteins and intrinsically disordered regions are under-represented - Multi-chain complexes are included in training, but performance degrades with chain count

Loss Function & Objective¶

The folding trunk is trained with a combination of losses adapted from AlphaFold2:

FAPE loss (Frame Aligned Point Error): Measures the error in predicted atomic positions under local reference frames, capturing both backbone and side-chain accuracy
Distogram loss: Cross-entropy on predicted inter-residue distance distributions
pLDDT head loss: Supervised training of the per-residue confidence score
pTM head loss: Supervised training of the predicted TM-score

Tokenization / Input Processing¶

Property	Details
Tokenizer	Character-level (one token per amino acid)
Vocabulary	Standard amino acid alphabet + extended characters
Max sequence length	768 residues (BioLM implementation)
Multi-chain support	Chains concatenated with `:` separator (up to 4 chains)
Special processing	Chunk size set to 768 for memory-efficient attention

Multi-chain complexes (e.g., homodimers, heterodimers) are specified by separating chains with the : character. The BioLM implementation supports up to 4 chains (max_n_multimers = 4) with a total length of 768 residues plus up to 3 separator characters (771 total characters maximum).

Performance & Benchmarks¶

Published Benchmarks¶

ESMFold was evaluated on CAMEO and CASP14 targets, comparing single-sequence prediction against MSA-dependent methods.

CAMEO (Continuous Automated Model Evaluation)¶

Model	GDT-TS	TM-score	Method
ESMFold	-	0.75	Single-sequence
AlphaFold2	-	0.88	MSA-based
RoseTTAFold	-	0.76	MSA-based

Key Findings (Lin et al., Science 2023)¶

ESMFold achieves competitive accuracy with MSA-based methods for proteins with high evolutionary coverage
On single-domain proteins, ESMFold pLDDT > 70 correlates with TM-score > 0.8 relative to experimental structures
Prediction speed is approximately 60x faster than AlphaFold2 due to elimination of MSA search
Accuracy degrades for sequences with few homologs in UniRef50 (low evolutionary coverage)

BioLM Verification Results¶

The BioLM implementation loads official pre-trained weights via esm.pretrained.esmfold_v1(). Verification is performed against golden reference outputs stored in R2:

Test Case	Metric	Threshold	Status
Single-chain prediction	RMSD	< 0.5 Angstroms	PASS
Single-chain prediction	Relative tolerance (pLDDT, pTM)	1e-1	PASS
Multi-chain prediction	RMSD	< 0.5 Angstroms	PASS
Multi-chain prediction	Relative tolerance (pLDDT, pTM)	1e-1	PASS

Comparison to Alternatives¶

Model	MSA Required	Speed	Ligands	Multi-chain	When to prefer
ESMFold (this)	No	Fast (~seconds)	No	Limited (up to 4 chains)	Rapid prototyping, large-scale screening
AlphaFold2	Yes	Slow (~minutes-hours)	No	Yes	Maximum accuracy when speed is not critical
Boltz	Optional	Moderate	Yes (SMILES)	Yes	Protein-ligand complexes, diverse biomolecules
Chai-1	Optional	Moderate	Yes (SMILES)	Yes	Alternative to Boltz for complex prediction

Strengths & Limitations¶

Pros¶

No MSA required -- single-sequence input makes predictions dramatically faster than AlphaFold2
End-to-end differentiable -- language model and folding module trained jointly
Supports multi-chain complexes (up to 4 chains in BioLM)
Provides calibrated confidence metrics (pLDDT, pTM) for assessing prediction reliability
MIT licensed with no restrictions on commercial use
Shares the ESM-2 backbone, benefiting from the same evolutionary representations

Cons¶

Accuracy lower than MSA-dependent methods (AlphaFold2, Boltz) for most proteins
Performance degrades significantly for sequences with few homologs (orphan proteins)
Limited to 768 residues per prediction in BioLM (shorter than AlphaFold2's limits)
No ligand or small molecule support
Multi-chain capability is basic compared to dedicated complex prediction methods (Boltz, Chai-1)
No explicit handling of post-translational modifications

Known Failure Modes¶

Low-homology proteins: Sequences with few detectable homologs in UniRef50 produce unreliable structures (low pLDDT, low pTM)
Intrinsically disordered regions: These regions will have low pLDDT but may be modeled as extended or compact structures that do not reflect their biological disorder
Large multi-chain complexes: Accuracy decreases with more than 2 chains, and memory usage increases quadratically
CUDA out of memory: Very long sequences or sequences near the 768-residue limit can exceed GPU memory. The implementation raises a ModelExecutionError so the failure is visible to the caller rather than being silently swallowed
Membrane proteins: Under-represented in training data; transmembrane helical bundles may be poorly predicted

Implementation Details¶

Inference Pipeline¶

Request
  |-- 1. Validate sequences (alphabet, length, batch size, chain count)
  |-- 2. Extract sequences from request items
  |-- 3. Batch sequences by token count (max 1024 tokens per batch)
  |-- 4. For each batch:
  |     |-- Forward pass: model.infer(batch, num_recycles=4)
  |     |-- Convert outputs to PDB format: model.output_to_pdb(outputs)
  |     |-- Extract per-sequence confidence scores:
  |     |     |-- mean_plddt: average per-residue confidence
  |     |     \-- ptm: predicted TM-score
  |     \-- Handle CUDA OOM: raise ModelExecutionError (propagates as typed server error)
  \-- 5. Return ESMFoldResponse with results list

Memory & Compute Profile¶

Input	GPU Memory (approx)	Inference Time (approx)	Notes
Single chain, ~100 residues	~8 GB	~5-10s	Well within A10G limits
Single chain, ~500 residues	~12 GB	~15-30s	Attention scales O(n^2)
Multi-chain, ~768 residues total	~20 GB	~30-60s	Near memory limit

The BioLM deployment uses an A10G GPU (24 GB VRAM) with 16 GB system RAM and 4 CPU cores.

Determinism & Reproducibility¶

Setting	Value
`torch.manual_seed`	42
`torch.cuda.manual_seed_all`	42
`torch.no_grad`	Yes (inference)
cuDNN deterministic	Not explicitly set
cuDNN benchmark	Not explicitly disabled

The model produces reproducible outputs on the same GPU architecture. Small numerical differences may occur across different GPU types due to floating-point operation ordering differences in CUDA kernels.

Caching Behavior¶

Response caching is handled by the serving infrastructure outside the model container. The model container itself is stateless; cache key and eviction policy are determined by the deployment layer, not by this code.

Versions & Changelog¶

Version	Date	Changes
v1	2024-11-06	Initial ESMFold implementation with fold action
v1 (updated)	2024-12-23	Added multi-chain support with `:` separator
v1 (updated)	2026-03-14	Migrated to declarative download system and source layer setup

Biology¶

Molecule Coverage¶

Primary Molecule Type(s)¶

ESMFold is designed for protein structure prediction from amino acid sequences. It accepts single-chain proteins and small multi-chain complexes (up to 4 chains concatenated with : separators) with a maximum total length of 768 residues.

The model is trained on structures from the Protein Data Bank (PDB) and AlphaFold2 distillation data, covering proteins from all domains of life. Performance characteristics vary by protein type:

Globular, soluble proteins: Best performance. These represent the majority of PDB training data, and the ESM-2 backbone has strong representations for this category.
Single-domain proteins: Excellent accuracy when the protein has detectable homologs in UniRef50. For well-covered families, pLDDT > 70 typically corresponds to TM-score > 0.8 relative to experimental structures.
Multi-domain proteins: Each domain may be predicted well individually, but relative domain orientations can be unreliable, especially for flexible linker regions.
Multi-chain complexes: Supported but with degraded accuracy compared to single chains. Performance decreases with the number of chains.
Membrane proteins: Under-represented in training data. Transmembrane regions may be poorly predicted, especially for multi-pass helical bundles.
Intrinsically disordered regions: The model will assign low pLDDT scores to these regions (correctly reflecting uncertainty), but the predicted coordinates do not represent the biological ensemble of conformations.

Cross-Applicability¶

Molecule Type	Applicability	Evidence	Caveats
Globular proteins	High	Core training and evaluation target (CAMEO, CASP14)	Best performance category
Enzymes	High	Active site geometry generally well-predicted for globular enzymes	Catalytic mechanism and substrate binding not modeled
Antibodies	Moderate	Can fold individual Fv domains	CDR loop conformations may be inaccurate; dedicated antibody structure predictors (e.g., ABodyBuilder2) may outperform
Protein complexes	Moderate	Supports up to 4 chains via `:` separator	Interface contacts are less reliable than dedicated complex predictors (Chai-1)
Peptides	Low	Very short sequences (< 30 residues) provide limited context for the ESM-2 backbone	Structure prediction for short peptides is inherently difficult; consider experimental NMR data
DNA/RNA-binding proteins	Moderate	Protein structure itself can be predicted	Nucleic acid partners are not modeled

Biological Problems Addressed¶

Rapid Protein Structure Prediction¶

Problem: Determining a protein's three-dimensional structure is essential for understanding its function, designing drugs that target it, and engineering proteins with desired properties. Experimental methods (X-ray crystallography, cryo-EM, NMR spectroscopy) are slow, expensive, and not always feasible. Computational methods like AlphaFold2 provide high accuracy but require multiple sequence alignments (MSAs) that take minutes to hours to compute, creating a bottleneck for high-throughput applications.

How ESMFold helps: ESMFold predicts protein structure from a single amino acid sequence in seconds, bypassing the MSA computation entirely. The ESM-2 language model backbone has already learned structural patterns from evolutionary data during pre-training, so explicit MSA signals are not needed for many proteins.

Biological meaning: The output is a full-atom 3D structure in PDB format, with per-residue confidence scores (pLDDT) indicating which parts of the structure are reliable. A mean pLDDT above 70 generally indicates a usable fold. The predicted TM-score (pTM) provides a global assessment of whether the overall fold topology is correct.

Practical considerations: ESMFold trades some accuracy relative to AlphaFold2 for dramatically faster inference. This makes it ideal for screening large protein sets where rapid structure assessment is more important than maximum accuracy.

Large-Scale Structural Screening¶

Problem: Genome sequencing projects produce millions of predicted protein sequences, but structural characterization lags far behind. Metagenomic studies routinely identify hundreds of thousands of novel protein families with no experimental structure. Understanding even the rough fold of these proteins can provide functional insights.

How ESMFold helps: The speed of ESMFold (seconds per prediction) makes it feasible to predict structures for entire proteomes or metagenomes. The authors of the ESMFold paper used it to predict structures for over 600 million metagenomic protein sequences, demonstrating the scalability of the approach.

Biological meaning: Even low-confidence predictions (pLDDT 50–70) can reveal fold-level similarity to known structures, enabling functional annotation of previously uncharacterized proteins. High-confidence predictions (pLDDT > 70) can be used as starting points for more accurate methods or experimental validation.

Structure-Guided Protein Engineering¶

Problem: Protein engineers often need rapid structural feedback when designing mutations or evaluating variant libraries. Traditional structure prediction is too slow for iterative design cycles where hundreds of variants need structural assessment.

How ESMFold helps: The fast inference time allows researchers to predict structures for mutant libraries, evaluate structural stability of designed sequences, and filter candidates before expensive experimental testing.

Biological meaning: Comparing predicted structures of wild-type and mutant proteins can reveal whether mutations disrupt the protein fold, alter active site geometry, or create steric clashes. Large changes in pLDDT or pTM between wild-type and mutant suggest the mutation may be destabilizing.

Applied Use Cases¶

ESMFold has been widely adopted for rapid structure prediction workflows. Selected published applications:

Protein-peptide docking (Zalewski et al., 2025): ESMFold-predicted structures used for protein-peptide docking without MSAs, achieving 20–28% acceptable-quality poses using an efficient polyglycine linker approach.
Binding-site prediction (DeepProSite, Fang et al., 2023): ESMFold-predicted structures fed into a topology-aware graph transformer for protein binding-site prediction, outperforming sequence-only baselines.
Protein-engineering ranking (APPRAISE, Ding et al., 2024): ESMFold structures used to rapidly rank engineered protein variants (AAVs, nanobodies, miniproteins) by target-binding propensity before wet-lab testing.
Antimicrobial peptide classification (Cordoves-Delgado et al., 2024): ESMFold-predicted structures as graph representations for antimicrobial peptide classification, outperforming 20 state-of-the-art methods on a 67,058-peptide dataset.
Joint structure and fitness prediction (SPIRED-Fitness, 2024): ESMFold used as a benchmark for a new end-to-end structure-and-fitness prediction framework, demonstrating 5-fold inference acceleration while maintaining comparable accuracy.

Predecessor Models¶

ESM-1b (Rives et al., 2021): The predecessor protein language model. ESMFold uses the improved ESM-2 backbone rather than ESM-1b.
AlphaFold2 (Jumper et al., 2021): The MSA-dependent structure prediction method whose folding module architecture inspired ESMFold's structure prediction component. AlphaFold2 remains more accurate but is much slower due to MSA computation.

Complementary Models¶

ESMFold is closely related to other models in this catalog:

ESM-2: ESMFold uses the ESM-2 3B language model as its backbone. ESM-2 embeddings (encode action) capture the same evolutionary representations that ESMFold uses for structure prediction. Users who need embeddings rather than structures should use ESM-2 directly.
ThermoMPNN: Predicts stability changes (ddG) from structure. Combining ESMFold structure predictions with ThermoMPNN stability estimates provides a more complete characterization of engineered protein variants.

Typical multi-model workflows: 1. Use ESMFold for rapid structure prediction of a candidate set, then use Chai-1 for high-accuracy prediction of the top candidates 2. Use ESM-2 log_prob to score variant effects, then use ESMFold to assess structural impact of top-ranked variants 3. Use ESMFold to generate initial structure models, then use MPNN for sequence design on the predicted structures

Alternative Models¶

Alternative	Advantage Over ESMFold	Disadvantage vs ESMFold
AlphaFold2	Higher accuracy, especially for multi-domain proteins	Requires MSA, much slower inference
Chai-1	Handles ligands, nucleic acids, and diverse complexes; higher accuracy for multi-chain	Slower inference, more resource-intensive

When to choose ESMFold: Use ESMFold when you need fast, single-sequence protein structure predictions for screening, prototyping, or large-scale analysis where speed matters more than maximum accuracy. It is the fastest structure prediction option in this catalog.

When to choose alternatives: Use Chai-1 when you need predictions for protein-ligand complexes, RNA/DNA-containing complexes, or when maximum structural accuracy is required. Use AlphaFold2 when you have MSAs available and need the highest accuracy for single-chain or multi-chain protein structures.

Biological Background¶

Protein structure prediction is one of the grand challenges of computational biology. Proteins are linear chains of amino acids that fold into specific three-dimensional structures dictated by their sequence. The 3D structure determines the protein's function: enzymes have precisely shaped active sites, receptors have complementary binding surfaces, and structural proteins form organized assemblies. Understanding protein structure is therefore central to drug design, enzyme engineering, and fundamental biological research.

The protein folding problem: Given an amino acid sequence, predicting the 3D structure the protein adopts under physiological conditions has been studied for over 50 years. The breakthrough came with AlphaFold2 (2020), which achieved near-experimental accuracy by combining multiple sequence alignments (MSAs) -- evolutionary profiles showing which amino acids co-vary at different positions -- with a neural network architecture. However, computing MSAs requires searching large sequence databases, which takes minutes to hours per protein.

Single-sequence structure prediction: ESMFold demonstrated that protein language models trained on evolutionary sequence data already learn enough structural information to predict 3D structure without explicit MSA input. The language model's internal representations capture the same co-evolutionary signals that MSAs provide, but compressed into a single forward pass. This is analogous to how a human expert who has studied thousands of protein structures can make reasonable fold predictions from sequence alone.

Key terminology: - pLDDT (predicted Local Distance Difference Test): A per-residue confidence score (0-100) estimating how accurately the local structure around each residue is predicted. Values above 70 generally indicate reliable predictions. - pTM (predicted TM-score): A global score (0-1) estimating the overall fold correctness. Values above 0.5 indicate a broadly correct fold topology; values above 0.8 indicate high-confidence predictions. - PDB format: The standard text format for representing 3D molecular structures, specifying atomic coordinates, chain identifiers, and metadata. - MSA (Multiple Sequence Alignment): An alignment of a query protein against evolutionarily related sequences from databases like UniRef. Co-varying positions in the MSA reveal structural contacts. - Recycling: Iterative refinement of structure predictions by feeding outputs back through the model. ESMFold uses 4 recycles by default to progressively improve the predicted structure.

Sources & license¶

License: MIT (text)

Papers

Evolutionary-scale prediction of atomic-level protein structure with a language model — Science, 2023 · DOI arXiv

Source repositories

github: https://github.com/facebookresearch/esm

Cite

@article{lin2023evolutionary,
  title={Evolutionary-scale prediction of atomic-level protein structure with a language model},
  author={Lin, Zeming and Akin, Halil and Rao, Roshan and Hie, Brian and Zhu, Zhongkai and Lu, Wenting and Smetanin, Nikita and Verkuil, Robert and Kabeli, Ori and Shmueli, Yaniv and dos Santos Costa, Allan and Fazel-Zarandi, Maryam and Sercu, Tom and Candido, Salvatore and Rives, Alexander},
  journal={Science},
  volume={379},
  number={6637},
  pages={1123--1130},
  year={2023},
  doi={10.1126/science.ade2574}
}

ESMFold¶

At a glance¶

API & schema¶

fold¶

Usage¶

Overview¶

Architecture¶

Capabilities & Limitations¶

Confidence Metrics¶

Usage Examples¶

Architecture & training¶

Architecture¶

Model Type & Innovation¶

Parameters & Layers¶

Training Data¶

Loss Function & Objective¶

Tokenization / Input Processing¶

Performance & Benchmarks¶

Published Benchmarks¶

CAMEO (Continuous Automated Model Evaluation)¶

Key Findings (Lin et al., Science 2023)¶

BioLM Verification Results¶

Comparison to Alternatives¶

Strengths & Limitations¶

Pros¶

Cons¶

Known Failure Modes¶

Implementation Details¶

Inference Pipeline¶

Memory & Compute Profile¶

Determinism & Reproducibility¶

Caching Behavior¶

Versions & Changelog¶

Biology¶

Molecule Coverage¶

Primary Molecule Type(s)¶

Cross-Applicability¶

Biological Problems Addressed¶

Rapid Protein Structure Prediction¶

Large-Scale Structural Screening¶

Structure-Guided Protein Engineering¶

Applied Use Cases¶

Related Models¶

Predecessor Models¶

Complementary Models¶

Alternative Models¶

Biological Background¶

Sources & license¶

`fold`¶