OpenBind
v0.5.1Open protein-ligand co-folding with the OpenBind-0 parameters for OpenFold3
OpenBind-0 (OB0) is the first open co-folding model from the OpenBind consortium, a UK project based at Diamond Light Source that is building a large open dataset of protein-ligand structures for drug discovery. OB0 runs on the OpenFold3 architecture and predicts the 3D structure of a protein together with its ligand from a sequence and a SMILES string. It returns ranked complex structures with pLDDT, pTM and ipTM confidence scores.
The ProteinIQ webserver runs OpenBind-0 online without a local installation or GPU. It uses the released OpenFold3 0.5.1 code with the OB0 parameter file of3-ob-2025-06-30-174k.pt, which is also the default model behind the OpenFold3 page. The code, weights, training data and training recipes are released under Apache 2.0, so predictions can be used in commercial work.
A typical run takes a target protein and one or more ligands, optionally with a pocket constraint that names the binding-site residues. Ions, cofactors, extra protein chains, DNA and RNA can be added to the same complex.
Classical docking tools such as AutoDock Vina place a ligand into a fixed receptor structure and rank poses with a scoring function. OpenBind-0 predicts the protein and the ligand together. Starting from random atom positions, its diffusion module denoises the whole complex, so the binding site can take a different shape for each ligand and no receptor structure is needed.
OB0 was trained on PDB structures released up to 30 June 2025. The first release does not yet include OpenBind's own crystallography data; that data is meant for later models in the series. The OpenBind team reports OB0 results on a preview of the PLINDER benchmark and on new crystal structures from its own targets. On the enterovirus A71 2A protease, OB0 reaches 92.2% top-25 and 73.8% top-1 success. On the FatA target, top-25 success is 28.2%, compared with 14.5% for OpenFold3 preview 2 and 39.4% for Protenix-v1. No model passed 10% on the dengue and Zika RdRp targets.
OpenBind uses the same pricing as OpenFold3, starting at 50 credits per job. The quote depends on the total number of protein, DNA and RNA residues, the number of diffusion samples and the number of model seeds. Ligands and ions do not change the price. The exact quote is calculated before submission.
Credits for one model seed:
| Residues | 1 diffusion sample | 5 diffusion samples (default) | 10 diffusion samples |
|---|---|---|---|
| 150 | 50 | 67 | 103 |
| 300 | 61 | 108 | 168 |
| 500 | 87 | 155 | 240 |
| 1,000 | 141 | 251 | 389 |
| 1,500 | 243 | 433 | 671 |
| 2,500 | 427 | 762 | 1,181 |
Three model seeds cost 1.6 times as much: 173 credits for 300 residues with 5 diffusion samples, or 401 credits for 1,000 residues. Complexes above 1,200 residues run on larger GPUs, which raises the rate per residue. Generate MSA, Use templates, the pocket constraint and the output format do not change the price.
| Input | Formats | Limit |
|---|---|---|
| Protein | FASTA sequence, PDB or ENT file, or a PDB ID fetched from RCSB | Up to 10 |
| Ligand | SMILES, SDF, MOL or MOL2 file, or a PubChem compound | Up to 10 |
| Ion | CCD code such as ZN, MG, CA or FE | Up to 10 |
| DNA | FASTA sequence or PDB file | Up to 10 |
| RNA | FASTA sequence or PDB file | Up to 10 |
Every job needs at least one protein, DNA or RNA chain, and the whole complex can have up to 3,500 residues. Each ligand entry holds one molecule, so add a separate ligand for each compound or copy. Chain IDs (A, B, C and so on) are assigned in the order the molecules are added and are shown on each card; the pocket constraint uses them.
Protein, DNA and RNA cards can mark a chain as cyclic and can list residue modifications as a position and a CCD code, for example position 3 as SEP for phosphoserine.
| Parameter | Type | Default | Description |
|---|---|---|---|
Model seeds | integer | 1 | Number of random seeds, from 1 to 5. Each seed produces its own samples. |
Diffusion samples | integer | 5 | Structures generated per model seed, from 1 to 10. |
Random seed | string | empty | Fixed seed for repeatable predictions. Empty uses a random seed. |
Output format | enum | CIF | Structure files as CIF or PDB. |
| Parameter | Type | Default | Description |
|---|---|---|---|
Generate MSA | boolean | true | Builds a multiple sequence alignment with the ColabFold MMseqs2 server. Adds a few minutes. |
Use templates | boolean | true | Adds structural templates from the ColabFold template search. Has no effect when MSA is off. |
Pocket constraint | string | optional | One binding-site constraint for one ligand or ion, in the format below. |
The pocket constraint guides one ligand or ion toward residues you name, written as ligand_chain|chain:residue,chain:residue|max_distance. The distance in Å is optional and defaults to 4.0. B|A:199,A:200,A:236,A:253,A:256|4.0 asks for ligand chain B within 4 Å of those five residues of protein chain A. Residue numbers are positions in the submitted sequence, starting at 1, not PDB author numbering. With a constraint, OpenFold3 runs an extra pocket proposal and refinement step for that ligand.
If the MSA server fails, the job is retried once without an MSA and the result carries a warning, because predictions without an MSA are usually less accurate.
The results open in three tabs. Viewer shows the ranked structures in 3D, Data lists each prediction with its scores, and Files holds every file for download.
| File | Contents |
|---|---|
prediction_rank01.cif | Best-ranked complex. Further files (rank02, rank03 and so on) follow the ranking order. |
| Scores per prediction | Ranking score, average pLDDT, pTM and ipTM from each sample's aggregated confidence file. |
One job returns Model seeds × Diffusion samples structures, for example 5 with the defaults. Files are .pdb instead of .cif when the output format is PDB. Per-atom pLDDT is stored in the B-factor column, so most viewers can color the structure by confidence.
Predictions are sorted by OpenFold3's own sample ranking score, which combines the confidence terms. The top-ranked structure is a reasonable first choice, but compare the top few before you trust a binding pose.
ipTM estimates how well the interfaces between chains are predicted, including the protein-ligand interface. For a binding question it is the most useful single number. A high pTM with a low ipTM means the protein fold is confident but the ligand placement is not. See ipTM vs pTM for thresholds and how the two scores relate.
Average pLDDT summarizes the whole complex, so a well-folded protein can score high even if the pocket is uncertain. Color the structure by pLDDT in the viewer and check the residues that contact the ligand. Values above 70 there support the pose; values below 50 suggest a disordered or poorly predicted site. The pLDDT guide explains the bands.
When most diffusion samples and seeds put the ligand in the same pocket with the same orientation, the pose is more believable than any single score suggests. If the ligand jumps between sites across samples, treat the binding mode as unknown and consider a pocket constraint or a docking run for comparison. Structure prediction confidence covers how to combine these signals.
Co-folding models can return ligands with distorted geometry or flipped stereocenters, even in confident poses. The OpenBind team reports that inference-time chemical steering raised the share of poses that are both accurate and chemically valid from 48% to 61% in its benchmarks. Steering is not part of the released OpenFold3 0.5.1 code yet, so this webserver runs OB0 without it. Check ligand bond lengths, chirality and clashes before using a pose, for example with PoseBusters.
OpenBind-0 predicts structure only. Confidence scores are not binding affinities and should not be used to rank compounds by potency. Boltz-2 predicts an affinity value alongside the complex.
The T4 lysozyme L99A with toluene example co-folds the 164-residue cavity mutant with toluene (Cc1ccccc1) and the default settings. It returns 5 ranked structures for 71 credits. To judge the result, check that the toluene sits in the buried cavity created by the L99A mutation, that it lands there in most of the 5 samples, that ipTM is high, and that the cavity residues around it are above 70 pLDDT. The crystal structure 7L39 shows the experimental pose for comparison. The TEM-1 beta-lactamase, allosteric pocket example shows the same check with a pocket constraint that names the allosteric site.
Both pages run the same OpenFold3 0.5.1 code with the same OpenBind-0 parameters, settings and prices, so the same input gives the same kind of result on either page. The OpenBind page starts from a protein-ligand layout with ligand examples; the OpenFold3 page puts protein, DNA and RNA complexes first. Other co-folding options for protein-ligand work include Boltz-2 for affinity, Chai-1 and Protenix. For a fixed receptor and many ligands, DiffDock or AutoDock Vina are faster.
OB0 is trained on structures released before July 2025 and does not yet include OpenBind's own data. Its accuracy drops on targets unlike the PDB, such as the viral RdRp sites in the OpenBind benchmark, and on flexible binding sites. Predictions for very large or highly flexible ligands, covalent ligands and membrane proteins are less reliable. The OpenBind team describes OB0 as a first release, with more models to follow as its dataset grows.
Open-source AlphaFold3-based structure prediction for proteins, ligands, DNA, and RNA
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.
Open-source biomolecular structure prediction with Protenix
Predict protein-ligand binding poses and confidence scores without a predefined binding site.
Validate generated and docked ligand poses with structure, chemistry, and reference-pose checks.