Use OpenBind Online

OpenBind-0: Open Co-Folding Model for Protein-Ligand Structure Prediction

OpenBind-0 (OB0) is the first in a planned series of open co-folding models for protein-ligand complexes, released by the OpenBind Consortium, which is based at Diamond Light Source in Harwell, UK. OB0 is built on the OpenFold3 architecture and adds inference-time chemical steering to improve the physical validity of predicted ligand poses. According to the release post, the training data, code, model weights and training recipes are publicly available under the Apache 2.0 license, so the model can be retrained or fine-tuned on your own data.

Structure-based drug design depends on knowing how small molecules sit in protein binding sites, yet the public pool of experimental protein-ligand structures is limited. OpenBind is a UK consortium that generates structural and affinity data and releases it openly, and its data is intended to support active learning, where model predictions help choose the next experiments.

OB0 was trained on Protein Data Bank data up to June 30, 2025, with no OpenBind data in the training set, which makes the newly released fragment-to-hit structures a clean prospective test of how well co-folding models generalize.

How OpenBind-0 Predicts Protein-Ligand Complexes

OB0 uses a nearly final OpenFold3 architecture to co-fold a protein with a ligand and output a complex structure. The main change at inference time is chemical steering, which nudges sampling toward ligand geometries that pass physical validity checks.

• Open foundation: the model is built on OpenFold3, with weights, code and training recipes released.

• Chemical steering: on the similarity benchmark, joint success (accurate pose plus PoseBusters validity) rose from 48% to 61% with steering.

• Date-split training: a 2025 PDB cutoff lets the authors compare against OpenFold3-preview2 (2021 cutoff), AlphaFold3, ESMFold2, OpenDDE, Boltz-2 and Protenix-v1-20250630.

OpenBind-0 Benchmark Results on Fragment and Hit Targets

On a similarity-binned benchmark in the style of Runs N' Poses, OB0 performs comparably to AlphaFold3 and other models, and accuracy rises as test complexes resemble the training data. Gains from four extra years of PDB data are uneven: OB0 mainly outperforms OpenFold3-preview2 when test complexes move into a higher similarity bin.

The authors also tested prospective fragment-to-hit data released with the post: 717 ligand-bound structures across three targets.

What the failures show

For FatA, models usually find the pocket but misplace the ligand. For RdRp, failures include the wrong pocket and an incorrect pocket conformation.

What is Tamarind Bio?

Tamarind Bio is a no-code bioinformatics platform built to give life scientists and researchers access to powerful computational tools. Many cutting-edge machine learning models are hard to deploy and use. Tamarind provides an intuitive, web-based environment that removes the complexity of high-performance computing, software dependencies and command-line interfaces.

The platform is designed for biologists, chemists and other researchers who may not have a background in programming or cloud infrastructure but want to run models on their own data. Key features include:

• A user-friendly graphical interface for setting up and launching experiments

• A robust API for integration into existing research pipelines

• An automated system for managing and scaling computational resources

Tamarind treats information and data security as a top priority, as detailed in its Trust Center and Terms of Service.

Accelerating Discovery with OpenBind-0 on Tamarind Bio

• Pose prediction for hits: predict how fragments and follow-on compounds bind a target, particularly when a similar complex exists in the PDB.

• Physically valid poses: use chemical steering to favor ligand geometries that pass PoseBusters-style checks.

• Open, retrainable workflows: because weights and recipes are public, the model can be fine-tuned on in-house structures.

• Model benchmarking: compare OB0 predictions with other co-folding tools on your own targets.

How to Use OpenBind-0 on Tamarind Bio

1. Log in and open the tool: sign in at tamarind.bio and select OpenBind.

2. Provide the protein: enter the target protein sequence, as co-folding models take sequences rather than a pre-built pocket.

3. Provide the ligand: supply the small molecule, for example as a SMILES string (if exposed).

4. Set parameters: choose the number of samples and whether chemical steering is enabled, if exposed.

5. Run the job: submit and wait for the predicted complexes.

6. Download outputs: retrieve the predicted structures and any confidence outputs.

7. Interpret: inspect the pocket and ligand placement; the release post reports success at both top-1 and top-25, so review several ranked poses.

8. Use downstream: feed poses into docking refinement, design or analysis workflows.

Things to Keep in Mind

• Success was very low for RdRp and moderate for FatA, and the RdRp structures are dissimilar to the training data.

• Fine-tuning on fragment data helps some targets (EV-A71 2A, modestly FatA) but barely helps RdRp; generalizable fine-tuning protocols still need work.

• Flexible binding sites remain a persistent challenge for co-folding methods.

• The benchmark uses a preview version of PLINDER.

Source: OpenBind Consortium, "OpenBind-0: Advancing Open Molecular Structure Prediction" (blog post, openbind.uk, August 21, 2026), openbind.uk. Code and release: github.com/aqlaboratory/openfold-3; blog scripts and data: OpenBind-0-model-release-info.