Use OpenFold3 Online
Commercially Available OpenFold3 No-Code Web Server
OpenFold3-preview2: Open-Source AlphaFold3 Reproduction for Biomolecular Structure Prediction
OpenFold3-preview2 (OF3p2) is the second prerelease of OpenFold3, an open system for general biomolecular structure prediction. Its technical report, written by the OpenFold3 Team (a core team of 16 plus about 70 listed contributors from institutions including Columbia University, Lawrence Livermore National Laboratory, Novo Nordisk, AWS, Seoul National University, the Chan Zuckerberg Initiative, Absci, SandboxAQ, AMD, NVIDIA, the University of Bristol, the OpenFold Consortium and Pfizer), describes it as the most performant academic reproduction of AlphaFold3 (AF3) to date and as trainable from scratch.
The stated goal is an open-source AF3 with performance parity across all molecular modalities. The authors say that no reproduction, including OF3p2, yet matches AF3 everywhere, and the report is candid about where the gaps remain.
The team states that it is releasing model weights, training and inference code, benchmarking code and distillation datasets. The report itself does not state a license or give repository URLs.
Property | Detail |
|---|---|
Task | General biomolecular structure prediction (AF3-style) |
Inputs | Sequences with MSAs (subsampled to 1,024 sequences), structural templates, MSA seeds |
Outputs | 3D structures with ipTM, pTM, PAE and pLDDT confidence scores |
Default inference | 10 trunk recycles, 5 MSA seeds, 5 diffusion samples per seed (25 samples) |
Training | 155,000 steps on 256 NVIDIA H100 GPUs |
How OpenFold3-preview2 Works and Is Trained
The model follows the AF3 recipe: a trunk with recycling, a diffusion-based sampler and a confidence head that predicts PAE and pLDDT, with training hyperparameters taken from the AF3 supplementary information.
Protein self-distillation: a monomer set built from clustered MGnify sequences (2025_03 release), reaching 13M sequences after filtering MSAs to Neff above 4.0, with structures predicted with and without templates and the higher-confidence one kept.
RNA self-distillation: about 125,000 structures from Rfam v15.1 cluster representatives, filtered to average PDE below 2.
Training schedule: 131,500 initial steps plus two fine-tuning stages of 8,000 and 15,500 steps; unlike AF3 there is no third stage, and the PAE head is trained from the start.
Bug fixes versus OF3p: corrections to the template mask, a missing filter for templates with many unresolved aligned residues, and the pLDDT inclusion radius.
OpenFold3-preview2 Benchmark Comparisons
The report compares OF3p2 with Protenix-v1, Boltz-1, Boltz-2, Chai-1 and AF3. Numerical metrics appear only in its figures, so the summary below is qualitative.
Benchmark | Finding |
|---|---|
Protein-small molecule (Runs N' Poses) | Similar to Protenix v1.0 on most similarity bins; AF3 remains best overall; Boltz-2 has a slight edge on high-similarity bins |
Biological polymers (FoldBench) | Comparable to AF3 and Protenix-v1; on par or ahead on protein monomers and protein-RNA, slightly behind on nucleic acid monomers, protein-protein and protein-DNA |
Antibody-antigen | Performance scales with MSA seed count, at a lower rate than AF3; stronger oracle scaling than Protenix-v1, though Protenix-v1 ranks better |
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 OpenFold3 on Tamarind Bio
Protein-ligand complexes: predict how small molecules sit in a target, evaluated in the report on Runs N' Poses.
Protein and nucleic acid assemblies: model protein-RNA, protein-DNA and protein-protein complexes, evaluated on FoldBench.
Antibody-antigen modeling: sample many MSA seeds to improve candidate quality, as studied in the report.
Open baseline: compare an open AF3-style model with other co-folding tools on your own targets.
How to Use OpenFold3 on Tamarind Bio
Log in and open the tool: sign in at tamarind.bio and select OpenFold.
Provide sequences: enter the chains of your complex; the model uses sequences with MSAs.
Add other molecules: include ligands or nucleic acid chains where the tool exposes them.
Set parameters: the report's settings are 10 recycles, 5 MSA seeds and 5 diffusion samples per seed; adjust seeds and samples if exposed.
Run the job: submit and wait for the predicted structures.
Download outputs: retrieve structures and confidence scores (ipTM, pTM, PAE, pLDDT).
Interpret: rank samples by confidence; for ligand contacts the report ranks by average ipTM between ligand and contacting chains.
Use downstream: take selected models into analysis, design or docking workflows.
Things to Keep in Mind
No reproduction, including OF3p2, fully matches AF3 across all modalities.
Ranked performance lags AF3 more than oracle performance, so confidence-based selection is a bottleneck.
Scaling with MSA seed count is weaker than AF3, and Boltz-2 keeps an edge on high-similarity ranked bins.
Sporadic inference or assessment failures occur across models.
RNA distillation structures made on MI300A APUs with the rocBLASLt backend have reduced chemical validity; the released set uses regenerated structures, and NVIDIA GPUs are unaffected.
Source: The OpenFold3 Team, "OpenFold3-preview2 Technical Report", portal.openfold.io (no date or DOI stated on the fetched page).