Optional rosetta repacking and fastrelax

[mosayebi]

Use the --rosetta_relax switch to relax the predicted structure in PyRosetta and resolve potential clashes. PyRosetta installation and a valid license are required.

Two relaxed structures will be saved in the prediction folder for each structure: one with side-chain repacking only, and the other with FastRelax applied. The corresponding Rosetta energy values will be saved in a CSV file. Relaxation is performed in parallel across --relax_cores CPU cores.

If your structure contains a ligand, use CCD ligands in the Boltz input and select CIF as the output format. This ensures the atom/residue naming is compatible with Rosetta conventions.

[poinco-gogo]

i'm afraid that PyRosetta is not free to use for commercial (this) and is not compatible with Boltz's license (MIT)

[hmms117]

You could consider the amber relax used in AF2 and ColabFold. I believe that would be ok.

[mosayebi]

You’re right that PyRosetta is only free for academic use, and purchasing a license is needed for commercial use. However, for those who already have a PyRosetta license, incorporating Rosetta relaxation can be beneficial. Like other ML-based structure prediction models, Boltz-1 often produces structures with resolvable clashes. Adding a relaxation step at the end of the pipeline helps refine these structures.

I agree that Amber-based relaxation is a great alternative. It would be fantastic to have it integrated into Boltz-1 as well. Until someone implements that, the PyRosetta relaxation is a useful option for those who can use it.