How to Use OpenFold3 Online for Biomolecular Complex Prediction

Introduction

OpenFold3 is an open, AlphaFold3-class system for predicting biomolecular structures across proteins, nucleic acids, ligands, and mixed complexes. Instead of treating folding and docking as separate steps, OpenFold3 is built around multimodal co-folding: the model receives every molecular component together and predicts a single joint structure.

On Neurosnap, you can run OpenFold3 online without installing model code, configuring sequence databases, or managing GPU inference. This tutorial walks through a small, well-characterized protein-ligand example so you can see how to set up inputs, choose practical options, and interpret the output, including the real confidence metrics from the run shown here.

The OpenFold3 (AlphaFold3) service page on Neurosnap, where the job is configured and submitted.

What OpenFold3 Does

OpenFold3 follows the AlphaFold3-style shift from single-protein structure prediction toward general biomolecular interaction modeling. AlphaFold3 introduced a diffusion-based framework for predicting the structures of complexes that can include proteins, nucleic acids, small molecules, ions, and modified residues [1]. OpenFold3 reproduces that general modeling paradigm in an open, trainable implementation [2,3].

The practical difference matters. In older workflows, a protein structure might be predicted first, then a ligand or partner molecule would be docked afterward. In an AlphaFold3-class workflow, the model reasons over the full molecular system at once. That does not make every prediction correct, but it gives the model a direct path to use sequence context, molecular identity, and pairwise interactions together.

OpenFold3 is best treated as a structural hypothesis generator. It can produce plausible models of a complex, but the final interpretation still depends on confidence metrics, visual inspection, comparison across samples, and follow-up validation.

When to Use OpenFold3 Online

OpenFold3 is most useful when the scientific question depends on how multiple molecules fit together, not only on the fold of a single protein chain. Common use cases include predicting a protein-protein complex, modeling a protein with a small molecule, testing a protein-RNA or protein-DNA assembly, comparing several candidate ligands or binding partners, and generating structural hypotheses before docking, rescoring, molecular dynamics, or experimental follow-up.

You can open the service directly here:

https://neurosnap.ai/service/OpenFold3%20%28AlphaFold3%29

Example: Carbonic Anhydrase II With Acetazolamide

For this walkthrough we use human carbonic anhydrase II with acetazolamide, a compact and well-characterized protein-ligand system. The RCSB entry 3HS4 reports a 1.1 Å crystal structure of human carbonic anhydrase II complexed with acetazolamide, where the ligand binds directly to the active-site zinc and contacts active-site residues [4].

The example is a good first OpenFold3 tutorial because it has one protein chain, one small molecule, a known experimental reference structure, a clear binding-site question, and a manageable input size.

Use the chain A protein sequence from PDB 3HS4:

Use acetazolamide as the ligand. PubChem lists acetazolamide as compound CID 1986, with formula C4H6N4O3S2 and canonical SMILES CC(=O)Nc1nnc(s1)S(N)(=O)=O:

Configure the OpenFold3 Job

On the OpenFold3 submission page, configure the job as a small protein-ligand prediction.

The most important setup detail is that the protein and ligand are submitted in the same job. OpenFold3 is not being used here as a standalone protein-folding step followed by docking; it is being asked to predict the joint protein-ligand structure.

For this tutorial we use 10 diffusion samples. Increasing the number of samples raises the chance of seeing a good candidate, but it also increases runtime and cost. The OpenFold3-preview2 technical report uses multi-sample inference in its benchmarking, a useful reminder that sampling strategy affects practical performance [2]. For reference, the run shown here completed in about 9 minutes.

Reading the Results

After the job finishes, start with the ranked structures and the interactive viewer.

For this example, the first thing to check is whether acetazolamide is positioned near the carbonic anhydrase active site. The experimental 3HS4 structure is a useful reference because the ligand is known to bind the zinc-containing active site [4]. A plausible prediction should place the ligand in a chemically reasonable pocket rather than floating away from the protein surface.

Ranked model: Rank 1 (best) Rank 2 Rank 3 Rank 4 Rank 5 Rank 6 Rank 7 Rank 8 Rank 9 Rank 10

Then review the confidence metrics. For AlphaFold3-class outputs, do not reduce interpretation to a single rank.

pLDDT is a local confidence score. High pLDDT suggests the model is confident in local geometry, while low pLDDT highlights flexible or uncertain regions. For protein-ligand predictions, high protein pLDDT alone does not prove the ligand pose is correct.

pTM summarizes global fold confidence. It is most useful for asking whether the overall structure is internally coherent.

ipTM, or chain-pair interface confidence, is the metric that matters most for interactions. In protein-ligand or protein-protein modeling, interface confidence helps decide whether the predicted contact geometry is likely to be meaningful.

PAE and PDE-style error views help identify which parts of the system are mutually well-positioned and which relative placements are uncertain. AlphaFold3 introduced predicted distance error as another way to reason about coordinate uncertainty [1].

What the Metrics Show for This Run

The 10 diffusion samples for this job were remarkably consistent, which is a good sign: the model converges on essentially the same high-confidence pose every time rather than producing scattered alternatives. The top-ranked model (rank_1.cif) scored an average pLDDT of 97.5, a pTM of 0.98, and an interface ipTM of 0.97, with no steric clashes and no disordered regions flagged.

One detail is worth reading carefully. The per-chain pTM for the protein (chain A) is about 0.98, but the per-chain pTM for the ligand (chain B) is only about 0.19. That low ligand value is expected and is not a red flag: pTM is a fold-level metric and is not informative for a single small molecule. The meaningful number for the binding question is the chain-pair interface confidence (ipTM ≈ 0.97), which says the model is confident about how the ligand sits against the protein, not just about the protein fold on its own. This is exactly why interface confidence should be read separately from whole-protein confidence.

Practical Interpretation

For a protein-ligand OpenFold3 result, use this triage order:

1. Check that the protein fold is reasonable.
2. Check that the ligand is near a plausible binding pocket.
3. Check interface confidence, not just whole-protein confidence.
4. Compare multiple diffusion samples or ranked models.
5. Treat the result as a hypothesis unless it agrees with known chemistry, experimental structures, or independent scoring.

Ligand predictions deserve extra caution. The OpenFold3-preview2 technical report notes that protein-small molecule performance can differ between oracle evaluation and ranked evaluation: good samples may be generated, but the model may not always rank the best pose first [2]. In practice, inspect more than the top model when a ligand pose matters. In this run the samples agreed closely, but that consistency is itself information, not every system will behave so cleanly.

For carbonic anhydrase II, the key sanity check is whether the sulfonamide-bearing ligand sits near the zinc-containing active site. If a high-ranking model placed acetazolamide far from the active site, that model should not be treated as a successful binding prediction just because the protein structure looks confident.

Practical Tips

Run more than one sample when the interface matters. Diffusion-based models sample structural hypotheses, so a single output can miss a better candidate that appears in another sample.

Use known structures when available. If your system has a PDB reference, compare the predicted complex against it. If no reference exists, compare against related structures, conserved active-site geometry, or known interaction constraints.

Do not overinterpret a confident monomer. A protein chain can have high local confidence while the ligand or partner placement remains uncertain, which is precisely why the per-chain and interface metrics should be read separately.

Use OpenFold3 as part of a workflow. For ligand projects, consider follow-up docking, rescoring, molecular dynamics, or experimental validation. For protein complexes, inspect interface residues, run orthogonal predictors, and compare alternative oligomeric states.

Conclusion

OpenFold3 on Neurosnap gives you a direct way to test AlphaFold3-class co-folding online. Start with a small, interpretable system; submit all molecular components together; then inspect the ranked outputs through both structure and confidence.

The key habit is to separate three questions:

1. Did the model predict a reasonable local structure?
2. Did it place the molecules relative to each other in a plausible way?
3. Does the confidence support the specific interface or ligand pose you care about?

If the answer to all three is yes, as it was for the carbonic anhydrase II and acetazolamide complex here, the output is a strong structural hypothesis. If one answer is no, the result may still be informative, but it should be treated as a starting point for further modeling rather than a final structure.

Sources

1. Accurate structure prediction of biomolecular interactions with AlphaFold 3 — Abramson, Adler, Dunger, et al.; 2024; Nature.
2. OpenFold3-preview2 Technical Report — The OpenFold3 Team; 2026.
3. OpenFold3 GitHub repository
4. PDB 3HS4: Human carbonic anhydrase II complexed with acetazolamide
5. RCSB FASTA for PDB 3HS4
6. PubChem CID 1986: Acetazolamide
7. OpenFold3 on Neurosnap