Clarify AF3 refold settings in paper/notebook: MSA usage and template usage (VHH/mAb vs de novo binders)

[cytokineking]

I’m trying to reproduce the AF3 “refold” validation step as done in the paper, but it’s unclear if you used MSAs or templates or if you ran everything in single-sequence mode. The notebook only states “Refolding is performed without a template” and recommends “20 seeds × 5 models”, but doesn’t show the AF3/AF3Score invocation or fold-input construction.

Could you clarify, for the paper runs:

1. If you refolded without a PDB template, did you still use an MSA (i.e., run the AF3 data pipeline / sequence search or provide a separately computed MSA)?
2. If you used an MSA, was it target-only, or (for VHH/mAb) did you also use antibody MSAs (and paired heavy/light MSAs where applicable)?
3. If you did not use MSAs, did you run truly single-sequence refolds for all chains (target + antibody/binder)?

If you can share an example command or fold-input JSON snippet for the refold step used in the paper (including how “20 seeds × 5 models” maps to seeds/samples), that would be extremely helpful.

[MacroHongZ]

This is a fold-input JSON example（5 seed）. For AF3 refold in VHH/mAb and binders, all chains use MAS without template.

[cytokineking]

Can you give more detail on how you calculated MSAs for VHH and mAbs? Would it be possible to share the exact MSA files used with the bundled VHH and mAb frameworks? This would be extremely useful and make using the example frameworks much easier.

[MacroHongZ]

https://github.com/google-deepmind/alphafold3

1. Use run_alphafold.py to run only the data pipeline, completing MSA generation and template search.
2. Remove the template information for all chain from the JSON file.
3. Use run_alphafold.py to run only the inference step.

[cytokineking]

ok so just to confirm, do you use the same MSA for all antibodies/VHH developed from a given framework? Or do you calculate new MSAs for each of the antibodies/VHH that are subject to the final refold?