ESMFold2
(biohub-esm-3.4.1.post1)Code (opens in a new tab)Paper (opens in a new tab)Docs (opens in a new tab)Run as batch
Predict protein and protein-complex structures from sequences with ESMFold2 confidence outputs.
Predict protein and protein-complex structures from sequences with ESMFold2 confidence outputs.
ESMFold2 is Biohub's protein structure prediction model built from ESMC protein language model representations and a diffusion-based structure generator. It predicts atom-level protein structures and protein complexes directly from amino acid sequence, returning confidence scores that help separate reliable folded regions from uncertain loops, termini, or interfaces.
ProteinIQ runs Biohub ESM release 3.4.1.post1 with pinned model checkpoints. To design new binders, use ESMFold2 Binder Design, which shares this release and bills by runtime.
Compared with the original ESMFold, ESMFold2 is aimed at higher-accuracy structure and complex prediction. It is still useful for sequence-only folding, but the confidence outputs are especially important when assessing multi-chain assemblies or deciding whether to spend time on downstream analysis.
Run ESMFold2 online by entering one or more protein chains as raw amino acid sequence, FASTA text, a FASTA file, or an RCSB sequence fetch. ProteinIQ predicts the structure with ESMFold2-Fast or the full ESMFold2 model, then returns an mmCIF structure, pLDDT, pTM, iPTM, and optional confidence matrix outputs.
| Input | Accepted format | Notes |
|---|---|---|
Protein | Raw sequence, FASTA text, .fasta, .fa, .fas, or .txt | The 20 standard one-letter amino acids, plus X for an unknown residue, which ESMFold2 models as UNK. |
Multiple chains | Up to 10 protein chains | Each chain is modeled as part of the same protein complex. |
RCSB fetch | PDB ID, for example 1UBQ | ProteinIQ fetches the protein sequence from RCSB and uses the sequence, not the deposited coordinates, as the prediction input. |
Job name | Text | Optional label for finding the prediction later. |
The total sequence length across all chains is capped at 2000 residues on ProteinIQ. For long multidomain proteins, running individual domains can produce more interpretable confidence patterns than folding the full construct in one job.
ProteinIQ supports protein-only ESMFold2 jobs in this release. ESMFold2 itself also folds DNA, RNA, ligands, modified residues, covalent bonds, MSAs and distogram conditioning, but these are not part of the online form yet.
| Setting | Default | Description |
|---|---|---|
Model variant | ESMFold2-Fast | ESMFold2-Fast is optimized for sequence-only inference. ESMFold2 full uses the larger model and may be preferable when accuracy matters more than runtime. |
Number of loops | 10 | Number of folding refinement loops. More loops can give the model more opportunity to refine difficult regions, with longer runtime. |
Sampling steps | 100 | Number of diffusion sampling steps used during structure generation. Higher values increase compute and may improve difficult predictions. |
Diffusion samples | 1 | Number of structures sampled from the same input, from 1 to 4. Each sample is returned in the model's own order, so compare their confidence to pick one. Inputs above about 1600 residues need fewer than 4 samples to fit in GPU memory. |
Random seed | Empty | Seeds input preparation and diffusion sampling. Set a value to reproduce a run; leave it empty for an unseeded run. |
Include PAE matrix | Off | Adds a JSON file with predicted aligned error values for domain and interface interpretation. |
Include pair-chain iPTM | Off | Adds a JSON file with pairwise chain-interface confidence for protein complexes when available. |
Include pooled pair embedding | Off | Adds a NumPy .npy array of the model's final pair representation averaged over one token axis, with one 256-value row per token. |
For a quick first pass, keep ESMFold2-Fast, Number of loops = 10, and Sampling steps = 100. Increase compute only when a prediction is scientifically important enough to justify a slower run or when comparing alternative chain definitions.
| Output | Format | Meaning |
|---|---|---|
Predicted structure | .cif | Native mmCIF coordinate file for the predicted protein or protein complex. |
pLDDT | JSON and structure metadata | Per-residue local confidence from 0 to 1, summarized as mean, minimum, and maximum pLDDT. The mmCIF B-factor column stores the same values multiplied by 100. |
pTM | Result table | Global fold confidence. Higher values indicate a more confident overall topology. |
iPTM | Result table | Interface confidence for multi-chain predictions. Higher values suggest more reliable chain placement. |
PAE matrix | Optional JSON | Predicted aligned error between residue pairs. Useful for domain orientation and interface uncertainty. |
Pair-chain iPTM | Optional JSON | Chain-by-chain interface confidence matrix for complex predictions. |
Pooled pair embedding | Optional .npy | Float32 array with one row per token, for use as a structure-aware representation in your own analysis. |
The structure is returned as mmCIF rather than PDB. mmCIF handles larger structures, chain identifiers, and modern structure metadata more reliably than legacy PDB format.
ESMFold2 confidence values are not experimental validation. They are model-estimated reliability signals, useful for triage and for deciding which parts of a prediction deserve closer inspection.
| Metric | Range | Best use |
|---|---|---|
pLDDT | 0 to 1 | Local residue-level confidence. High pLDDT supports backbone-level interpretation; low pLDDT often marks flexible or disordered regions. |
pTM | 0 to 1 | Overall fold confidence. Higher pTM indicates a more reliable global topology. |
iPTM | 0 to 1 | Interface confidence for complexes. Low iPTM means chain placement should be treated cautiously even if individual chains have high pLDDT. |
PAE | Error-like matrix | Domain and interface uncertainty. Low error between two regions supports a confident relative orientation; high error suggests the regions may move independently. |
Typical pLDDT interpretation follows the same convention used by AlphaFold-style structure predictors. The thresholds below use the 0 to 100 scale of the mmCIF B-factors; divide by 100 for the values in the result table.
| pLDDT | Interpretation |
|---|---|
| > 90 | Very high confidence. Backbone and many side-chain placements are likely reliable. |
| 70 to 90 | Useful model. Backbone geometry is often reliable, but details may need checking. |
| 50 to 70 | Low confidence. Use for rough hypotheses, not residue-level conclusions. |
| < 50 | Very low confidence. The region may be disordered, flexible, or outside the model's reliable regime. |
For complexes, pLDDT and iPTM answer different questions. A heterodimer can have confident individual chains and still have an uncertain interface. In that case, the model may know each fold but not how the chains pack together. Pair-chain iPTM and PAE are the best outputs to inspect before using the interface in docking, mutagenesis planning, or binder design.
Biohub's ESMFold2 starts from ESMC, a 6 billion-parameter protein language model trained on billions of protein sequences. The language model learns statistical regularities from protein evolution, then ESMFold2 uses those representations to guide a diffusion-based structure prediction model.
The diffusion component samples atomic structure rather than producing a single deterministic coordinate set in one pass. Sampling steps controls how many denoising steps are used, while Number of loops controls iterative refinement. ProteinIQ returns one sample by default. With more Diffusion samples, the model draws several structures from the same input, which helps show how stable a prediction is.
For multi-chain inputs, each sequence is passed as a separate protein chain in one structure prediction request. The returned mmCIF contains the predicted assembly, while pTM, iPTM, PAE, and pair-chain iPTM help identify whether the predicted interface is strong enough to use downstream.
| Tool | Best fit | Tradeoff |
|---|---|---|
| ESMFold2 | Protein-only structure prediction and protein-protein complexes from sequence | Does not accept ligands, nucleic acids, or MSA input in the ProteinIQ form. |
| ESMFold | Fast single-sequence structure prediction with familiar pLDDT output | Lower accuracy target than ESMFold2 for difficult complex prediction. |
| AlphaFold 2 | MSA-assisted protein folding and established AlphaFold-style workflows | MSA generation adds runtime and can be less convenient for designed or orphan proteins. |
| Boltz-2 | Protein, ligand, DNA, RNA, and affinity-aware complex prediction | Better fit for mixed biomolecular complexes, not the simplest protein-only prediction. |
| Chai-1 | Multi-component protein, nucleic acid, and ligand structure prediction | Useful for broad biomolecular inputs, but not focused on ESMFold2's protein-language-model route. |
ESMFold2 is a strong first choice when the input is protein-only and the key question is whether one or more chains can form a plausible folded structure. For protein-ligand binding or protein-nucleic acid assemblies, Boltz-2, Chai-1, OpenFold 3, or RosettaFold3 are more appropriate.
After prediction, common follow-up checks include MolProbity for geometry validation, DSSP for secondary structure assignment, USAlign for comparison against an experimental structure, and PDB Fixer when a downstream simulation or docking tool needs standardized coordinates.
Predict protein structures and complexes with single-sequence or MSA-assisted folding.
Open-source structure prediction for proteins, nucleic acids, and ligands
Generate protein conformational ensembles from sequence or structure inputs.
Predict paired antibody variable-domain structures from heavy and light chain sequences.
Predict biomolecular complex structures and binding affinities for proteins, ligands, DNA, and RNA.
Predict 3D structures for proteins, ligands, DNA, RNA, and multi-component complexes.
Predict protein structures and multimers directly from sequences without an MSA.
Predict immune receptor structures for antibodies, nanobodies, and T-cell receptors.
Predict controllable biomolecular structures for proteins, ligands, DNA, RNA, and complexes.
Low-memory Boltz inference for biomolecular structure and affinity prediction