Modification for alphafold3 backend

[Qrouger]

No description provided.

[DimaMolod]

The problem I see is that when AF2-specific flags are passed to the AF3 backend, they’re silently dropped in run_multier_jobs.py. This can mislead users into thinking the options were applied when they weren’t. For instance, passing --num_cycle (AF2) instead of the correct AF3 flag --num_recycles results in the argument being ignored with no warning. To prevent this confusion, we recently added the _validate_flags_for_backend check in run_structure_prediction.py, so unsupported flags are now detected and raise exception.

[Qrouger]

Yes, but at the moment run_multimer_jobs.py forward all of its flags, even if they are empty, to run_structure_prediction.py and _validate_flags_for_backend raised error.
ValueError: The following flags are not supported by backend 'alphafold3': ['compress_result_pickles', 'models_to_relax', 'msa_depth_scan', 'multimeric_template', 'num_cycle', 'num_predictions_per_model', 'pair_msa', 'remove_keys_from_pickles', 'remove_result_pickles', 'use_ap_style', 'use_gpu_relax']

Solution is probably to modify run_multimer_jobs.py, so that it does not pass empty flags to run_structure_prediction.py and modify flags (--num_cycle to --num_recycles just for AF3) to keep only one same flag in command line.

[DimaMolod]

I think use_ap_style flag is not used in the alphafold3_backend.py, as it is used earlier in run_structure_prediction.py

[Qrouger]

Yes, use_ap_style is not used by alphafold3_backend.py, but run_structure_prediction.py still expects this flag in order to correctly construct the output directory. However, _validate_flags_for_backend raises an error if this flag is passed to the Alphafold3 backend.

Since _validate_flags_for_backend runs at the very beginning of the main function in run_structure_prediction.py, the output directory defaults to something like ./result/ instead of the expected ./result/prot1_and_prot2/.

[DimaMolod]

Thank you @Qrouger , these are very helpful fixes!