OpenFold-3: Fold Anything!

A practical guide to running OpenFold-3 on Vici.bio.

Download

Summary

OpenFold-3 is the open-source version of AlphaFold-3, designed to provide high accuracy biomolecular structural predictions. This multimodal engine predicts all components simultaneously, thus predicting how molecules influence each other’s shapes as they bind instead of predicting each shape separately, and then trying to dock. This simultaneous 3D and interaction modelling more closely resembles biological processes and thus allows for higher accuracy predictions.

Without the restrictions experienced in AlphaFold-3, this developing predictor is the prefect model to use for unrestricted, biomolecular modelling and high-throughput structural research. While currently a preview release, it is already on par with AlphaFold-3 with RNA structures and remains competitive to other top performing molecular predictors.

How to use OpenFold-3 on Vici.bio

Using OpenFold-3 on Vici bio is designed to be intuitive. Paste your sequences into the molecules boxes and hit execute. By default, the system runs with optimized setting for highaccuracy, fast folding. An outline of the parameters is described below to allow for tailoring of the model to fit your projects specific needs.

Run OpenFold-3 on Solo or batch fold in Workflows .

Supported inputs

Proteins

One chain per entry. Repeat chains for stoichiometry.

  • Paste each chain as a separate FASTA sequence.
  • For multimers, duplicate chains to match stoichiometry.

DNA and RNA

Provide strings 5′ to 3′.

  • OpenFold-3 is already strong for RNA structures.
  • Use paired MSA when partners co-evolve.

Ligands, ions, cofactors

Model small molecule interactions directly.

  • Provide SMILES strings or CCD codes.
  • Works for metal ions like [Zn+2] and cofactors.

Glycans and modifications

Capture PTMs inside the structure.

  • Insert CCD codes at the modified site in sequence.
  • Useful for glycosylation and phosphorylation.

Assembly size

Scales well, but very large complexes are harder.

  • Accuracy can drop beyond about 5,000 amino acids.
  • Consider splitting assemblies when possible.

Advanced parameter controls

MSA strategy

Unpaired MSA

Independent homolog search per molecule.

  • Best when partners did not co-evolve.
  • Good for exploratory multi-component systems.

Paired MSA

Species-matched pairing for interfaces.

  • Critical for receptor-ligand and antibody-antigen.
  • Often improves interface placement (ipTM).

Both

Local folding plus co-evolution signals.

  • Balanced default for mixed uncertainty.
  • Useful when you care about both fold and binding.

None

Skip MSA completely.

  • Best for highly synthetic de novo designs.
  • May reduce interface reliability for natural complexes.

Sampling and reproducibility

Templates

Use known structures to guide detail.

  • Enable when a close, high-resolution template exists.
  • Useful for refining side-chain and ligand context.

Seeds and samples

Explore multiple plausible conformations.

  • Default is 5 seeds and 5 samples per seed.
  • Increase sampling for flexible molecules (antibodies, long RNA).

Mini case study: interpreting ipTM

High certainty

GO
ipTM > 0.80
Meaning: Confident binding pose and atomic details.
Recommendation: Safe for lead optimization and wet lab validation.

Best choice when you need a reliable interface geometry.

Structural sketch

KEEP
ipTM 0.60 to 0.80
Meaning: Fold and binding site likely correct, fine details uncertain.
Recommendation: Use for hypotheses. Consider templates and more sampling.

Good for direction, not final atomic decisions.

Unstable interface

DROP
ipTM < 0.50
Meaning: Missing context or interaction may not occur naturally.
Recommendation: Re-check MSA strategy and stoichiometry.

Treat as low confidence until inputs are fixed.

Important: OpenFold-3’s oracle performance can be higher than its internal ranking. Always inspect the top 5 samples, not only rank 1.

How it works

OpenFold-3’s key advantage is unified biomolecular prediction. It predicts the whole system together (proteins, DNA, RNA, ligands, and modifications), so partners can shape each other as the 3D structure is generated. This is an open-source reproduction of the AlphaFold-3 approach, designed to model binding in a way that matches how biology behaves, where conformations adjust during interaction.

Concepts that set it apart

  • Unified diffusion process: OpenFold-3 uses a diffusion model to generate 3D structures. During diffusion, components move together, and stable interactions can bias the system toward better-fitting poses.
  • Evolutionary guidance: MSAs provide evolutionary context that guides atom placements. This is especially important for RNA and for challenging interfaces.

Frequently asked Questions

References

  1. Abramson, J. et al. (2024) Accurate structure prediction of biomolecular interactions with AlphaFold 3. Nature 630, 493–500.
  2. Callaway, E. (2024) Who will make AlphaFold3 open source? Scientists race to crack AI model. Nature 630, 14–15.