Open structure prediction after AlphaFold 3

AlphaFold 3 was a landmark: for the first time a single model could predict the joint structure of proteins together with nucleic acids, ions and ligands. But its model weights are restricted to non-commercial use and cannot be redistributed. For companies — and for anyone who values fully reproducible pipelines — that restriction created an opening, and the open-source community filled it fast.

The open alternatives

Three projects now cover most of what teams need:

• Boltz-2 is MIT-licensed and, uniquely, predicts binding affinity alongside structure. It is the first deep-learning model to approach physics-based free-energy perturbation accuracy while running roughly a thousand times faster — which makes in-silico screening practical.
• Protenix, from ByteDance, is an Apache-2.0 reproduction of AlphaFold 3. Protenix-v2 posts large gains on hard antibody–antigen complexes and adds training-free geometric guidance during diffusion.
• OpenFold3 aims to be a faithful, trainable reproduction with long-term community maintenance, and is notable as the only open model that matches AlphaFold 3 on monomeric RNA.

Choosing between them

If you need commercial rights and affinity prediction, start with Boltz-2. If you want maximum accuracy on antibody complexes, evaluate Protenix. If you care about training your own variant, OpenFold3 is built for that.

All three accept the same conceptual input — sequences plus optional MSAs and ligands — so it is worth benchmarking two of them on your own targets before committing.

Don't forget search

Prediction is only half the story. Once you have structures, Foldseek lets you search hundreds of millions of them in seconds, turning the AlphaFold database into a practical discovery tool. The combination of open prediction and open search is what makes modern structural bioinformatics so productive.