ESMFold on the platform

ESMFold predicts a protein's 3D structure directly from its amino acid sequence — no multiple sequence alignment, no database lookup. Reach for it when your target isn't in a database: novel, engineered, or mutated sequences. If the protein is already known by UniProt ID, fetch it with AlphaFold instead (faster, no GPU); if an experimental structure has been deposited, download it from the PDB. See what you can achieve for the full set.

Try it

What ESMFold does

ESMFold uses the ESM-2 language model to predict atomic-level protein structures from sequence alone, much faster than MSA-based methods. It is not the most accurate predictor available — its authors report an average TM-score of 0.83 on CAMEO and 0.68 on CASP14 against AlphaFold2's 0.88 and 0.85 on the same sets. The trade is speed and MSA independence for accuracy, and how ESMFold results compare to the bench states the size of that trade with citations.

Citations

Lin Z, Akin H, Rao R, Hie B, Zhu Z, Lu W, Smetanin N, Verkuil R, Kabeli O, Shmueli Y, dos Santos Costa A, Fazel-Zarandi M, Sercu T, Candido S, Rives A. Evolutionary-scale prediction of atomic-level protein structure with a language model. Science. 2023 Mar 17;379(6637):1123-1130. doi: 10.1126/science.ade2574

Key achievements

• Speed: Order of magnitude faster than MSA-based methods at inference time
• No MSA required: ESMFold does not require a multiple sequence alignment (MSA) step, making it fully "standalone"
• Scale: Speed enables high-throughput structural genomics at scale

Core capabilities

• Single-sequence prediction: Predicts structures from primary amino acid sequences alone—no evolutionary information needed
• End-to-end pipeline: Uses token embeddings from the pre-trained ESM-2 language model directly
• Confidence estimation: Provides per-residue confidence scores (pLDDT) for reliability assessment
• Accuracy depends on language-model perplexity: proteins the language model finds unsurprising score close to AlphaFold2; accuracy falls as perplexity rises (Pearson r = −0.55 with TM-score on CAMEO, −0.67 on CASP14, author-reported)

Technical approach

ESMFold is built upon a BERT-like transformer architecture trained using masked residue prediction. The model learns to predict hidden amino acids based on surrounding context, developing deep understanding of protein sequence-structure relationships.

The model predicts:

• Backbone coordinates: Cα, C, N, and O atom positions
• Side chain orientations: χ angles for amino acid side chains
• Inter-residue distances: Predicted distance and angle distributions
• Confidence metrics: Per-residue and per-structure reliability scores

When to use ESMFold

Primary use cases

Structure prediction for novel proteins

• De novo designed proteins: Validate computational designs before experimental characterization
• Synthetic biology constructs: Predict structures of engineered proteins and fusion constructs
• Mutant analysis: Assess structural impacts of point mutations or domain insertions

High-throughput structural genomics

• Target selection: Prioritize proteins for experimental structure determination
• Virtual screening: Generate structural models for drug design campaigns
• Functional annotation: Predict structures to infer protein function from sequence

Early-stage drug discovery

• Target validation: Generate initial models for novel therapeutic targets
• Antibody modeling: nanobody and antibody frameworks only. ⚠︎ On a 75-nanobody benchmark the framework came out at 0.6–0.7 Å median RMSD but CDR3 at 3.3 Å, and on 200 antibody–antigen complexes ESMFold produced zero successful predictions — see how ESMFold results compare to the bench
• Protein-protein interaction studies: ⊘ Not supported in practice. ESMFold v1 is single-chain, reaches acceptable DockQ on only 41% of dimers, and scored 0/200 on antibody–antigen complexes. Use Boltz-2 or Chai-1 for complexes

Webapp integration and workflow

Input methods in .bloom*

Direct sequence entry:

• Paste amino acid sequences in FASTA or raw format
• Real-time validation with sequence length and composition checks
• Automatic detection of standard vs. non-standard amino acids
• Preview of sequence properties (length, molecular weight, isoelectric point)

UniProt integration:

• Direct UniProt accession lookup (e.g., "P12345")
• Automatic fetching of canonical sequences
• Isoform selection for proteins with multiple variants
• Integration with UniProt annotations and features

Previous job integration:

• Reuse sequences from other tools (blast searches, sequence analysis)
• Automatic extraction from multi-sequence outputs
• Provenance tracking maintains connection to original data source

Advanced job configuration

Parameter optimization:

• Num recycles: Control refinement iterations (1-10, default 3)
• Chunk size: Handle very long sequences (>1,000 residues)
• CPU offloading: Memory management for large proteins

Job management:

• Custom job naming for better organization
• Project-based quota tracking and management
• Integration with .bloom*'s job queuing and monitoring system

Interactive visualization

Built-in 3D viewer:

• Real-time structure prediction progress
• Confidence-colored models (pLDDT scores)
• Interactive inspection of predicted domains and secondary structure
• Export options for external visualization tools

Quality assessment:

• Per-residue confidence visualization
• Predicted aligned error (PAE) matrices
• Secondary structure predictions
• Domain boundary identification

Integration with .bloom* ecosystem

Upstream data sources

Sequence databases:

• Direct UniProt and PDB sequence lookup
• Integration with BLAST search results
• Import from custom sequence libraries

Previous analyses:

• Reuse sequences from motif searches
• Integration with sequence clustering results
• Connection to phylogenetic analysis outputs

Downstream applications

Structure-based drug design:

• Direct integration with DiffDock for molecular docking
• Pocket analysis and binding site identification
• Virtual screening preparation

Protein engineering:

• Model validation for computational design
• Assessment of stability and folding
• Analysis of protein-protein interfaces — ⊘ use a co-folding model, not ESMFold (see translatability)

Functional studies:

• Domain identification and characterization
• Active site prediction and validation
• Evolutionary conservation mapping

Workflow automation

Batch processing:

• Multiple sequence submission in single jobs
• Parallel processing for high-throughput studies
• Integration with .bloom*'s batch job management

Project organization:

• Automatic linking to research projects
• Sharing capabilities for collaborative studies
• Export options for publications and presentations

Best paired with

• Molecular dynamics: Refine ESMFold models with simulation
• Docking studies: ESMFold structures can be docked into (DiffDock, AutoDock), but a predicted receptor costs pose accuracy. On the PDBBind test set, DiffDock's top-1 success (RMSD < 2 Å) fell from 35.0% on crystal receptors to 21.7% on esmfold_v1 receptors, and search-based docking collapsed to 2–10%. Details and citation in how ESMFold results compare to the bench
• Experimental validation: Guide NMR or X-ray crystallography experiments
• Disorder prediction: Identify regions needing special handling

Scope and limitations

Supported inputs

Sequence requirements

• Single protein chains: Multi-chain complexes not supported
• Standard amino acids: 20 canonical amino acids only
• Length range: 10-2,500 amino acids (optimal performance: 50-500 residues)
• Sequence formats: Raw amino acid sequence or FASTA format

Input validation

• Invalid characters: Non-standard amino acids rejected
• Length limits: Sequences above 2,500 residues are rejected; confidence may drop in some regions above ~1,000 residues
• Gaps and unknowns: X characters treated as glycine for prediction

Prediction limitations

Accuracy considerations

• Novel folds: Lower accuracy for proteins with novel tertiary structures
• Disordered regions: Poor performance on intrinsically disordered proteins
• Low complexity: Struggles with low-complexity sequences (poly-A, etc.)
• Transmembrane proteins: Limited accuracy for membrane-spanning regions

Confidence interpretation

• pLDDT scores: Range from 0-100, with >90 indicating high confidence
• Domain boundaries: Confidence often drops at domain interfaces
• Flexible loops: Lower confidence for loops and linkers
• Structured cores: Highest confidence for α-helices and β-sheets

Performance characteristics

Speed and scaling

• Typical runtime: 10-60 seconds for average protein (50-200 residues)
• Scaling: Roughly linear with sequence length
• Hardware requirements: GPU acceleration recommended for best performance
• Batch processing: Supports multiple sequences in single job

Deterministic behavior

• Reproducible results: Same sequence always produces identical structure
• Version consistency: Results stable across model versions for same input
• Parameter sensitivity: Small changes in parameters rarely affect major structure

Quality assessment and validation

Confidence metrics (pLDDT)

Per-residue confidence scores indicating local structural accuracy:

These are per-residue confidences, not a global guarantee — see how ESMFold results compare to the bench for what each band has been measured to support.

Predicted aligned error (PAE)

Measures expected position error between residue pairs:

• PAE matrix: NxN matrix showing predicted error between all residue pairs
• Domain identification: High PAE between domains indicates separate structural units
• Rigid body movements: Low PAE within domains, high between domains

Quality control checks

Structural validation

• Stereochemistry: Check bond lengths, angles, and Ramachandran plot
• Packing quality: Assess side chain packing and cavity identification
• Secondary structure: Verify predicted helices and sheets match sequence

Biological plausibility

• Hydrophobic core: Buried hydrophobic residues should form core
• Surface properties: Polar residues should face solvent
• Domain organization: Multi-domain proteins should show clear boundaries

Integration and workflow

Pre-ESMFold preparation

Sequence curation

• Signal peptides: Remove N-terminal signal sequences if present
• Propeptide removal: Remove propeptides that aren't part of mature protein
• PTM consideration: Account for disulfide bonds, glycosylation sites
• Isoform selection: Choose most biologically relevant isoform

Parameter optimization

• Num recycles: Default is 3; increase up to 10 for marginally better accuracy at the cost of runtime
• Chunk size: Adjust for very long sequences (>500 residues)
• CPU offloading: Enable for memory-constrained environments

Post-ESMFold refinement

Structure optimization

• Energy minimization: Reduce steric clashes and optimize geometry
• Side chain refinement: Optimize rotamer positions
• Loop modeling: Rebuild low-confidence loop regions

Validation and analysis

• Quality metrics: Calculate MolProbity scores, clash analysis
• Comparison tools: Align with known structures using TM-align or FATCAT
• Functional assessment: Check active site geometry, binding pockets

Integration patterns

Drug discovery workflow

1. Target identification → ESMFold prediction
2. Pocket analysis → Lead identification
3. Virtual screening → Hit optimization

Structural biology

1. Sequence analysis → ESMFold prediction
2. Model validation → Experimental planning
3. Structure determination → Model refinement

Common pitfalls and best practices

Input-related issues

• Avoid truncated sequences: Use full-length proteins when possible
• Check sequence quality: Ensure no sequencing errors or frameshifts
• Consider isoforms: Different splice variants may have different structures

Interpretation errors

• Don't over-trust high confidence: Even pLDDT >90 can have domain-level errors
• Validate functionally important regions: Focus confidence assessment on active sites
• Use multiple models: Compare predictions from different tools when possible

Performance optimization

• Batch similar sizes: Group sequences of similar length for efficient processing
• GPU utilization: Use GPU instances for faster predictions
• Memory management: Enable CPU offloading for large proteins

Chat Usage

Example Prompts

• "Fold this protein sequence with ESMFold: MKTAYIAKQRQISFVKSHFSRQDILDLWIYHTQGYFP"
• "Predict the structure of EGFR kinase domain (P00533, residues 696-1022)"
• "Fetch human insulin from UniProt and fold it with ESMFold"

Workflow Chaining

• Feeds into: DiffDock (docking), P2Rank (pocket prediction), PDB2PQR (electrostatics)
• Accepts from: UniProt (sequence fetch), BLAST (hit sequences), ESM-3 (embeddings)

AlphaFold DB Fetch

For proteins already in the AlphaFold Protein Structure Database, you can fetch pre-computed structures instead of running prediction:

• "Get the AlphaFold structure for P12345"
• Pre-computed structures use the same 3D viewer and pLDDT coloring as ESMFold predictions
• Use ESMFold when: no AlphaFold structure exists, you need a specific isoform, or you want to fold a novel/engineered sequence

FAQ

Read next