Running AF3 using run_multimer_jobs.py

[ncoudray]

Hi -

I have just been updating from 2.0.3 to 2.4.0 but I am a bit confused on how to run alphafold3 with the command lines.

First, I ran

create_individual_features.py \
  --description_file=description.csv \
  --fasta_paths=Data/two_targets.fasta \
  --path_to_mmt=template/ \
  --data_dir=/scratch4/datasets/alphafold3 \
  --save_msa_files=True \
  --output_dir=Output/msa \
  --use_precomputed_msas=True \
  --max_template_date=2050-01-01 \
  --data_pipeline=alphafold3 \
  --skip_existing=False

which worked and generated two files:
Output/msa/4Y25_af3_input.json
and
Output/msa/5T10_af3_input.json

I then tried to run the second command using something like:

run_multimer_jobs.py  \
  --mode=custom \
  --num_cycle=3 \
  --num_predictions_per_model=1 \
  --output_path=Output/models \
  --data_dir=/scratch4/datasets/alphafold3 \
  --protein_lists=Data/protein_list5.txt \
  --monomer_objects_dir=Output/msa \
  --multimeric_template=True \
  --msa_depth=128 \
  --fold_backend=alphafold3 \
  --use_ap_style=True \
  --job_index=$SLURM_ARRAY_TASK_ID

But get this error:

Traceback (most recent call last):
  File "/vast/hpc/x86_64-linux/alphapulldown/2.0.3/miniconda/bin/run_structure_prediction.py", line 354, in <module>
    app.run(main)
  File "/vast/hpc/x86_64-linux/alphapulldown/2.0.3/miniconda/lib/python3.10/site-packages/absl/app.py", line 316, in run
    _run_main(main, args)
  File "/vast/hpc/x86_64-linux/alphapulldown/2.0.3/miniconda/lib/python3.10/site-packages/absl/app.py", line 261, in _run_main
    sys.exit(main(argv))
  File "/vast/hpc/x86_64-linux/alphapulldown/2.0.3/miniconda/bin/run_structure_prediction.py", line 324, in main
    parsed_input = parse_fold(FLAGS.input, FLAGS.features_directory, FLAGS.protein_delimiter)
  File "/vast/hpc/x86_64-linux/alphapulldown/2.0.3/miniconda/lib/python3.10/site-packages/alphapulldown/utils/modelling_setup.py", line 85, in parse_fold
    raise FileNotFoundError(
FileNotFoundError: ['4Y25', '5T10'] not found in ['Output/msa']

For Data/protein_list5.txt, I have
4Y25;5T10

I tried also
4Y25_af3_input.json;5T10_af3_input.json

and some other intermediate combinations (4Y25_af3_input;5T10_af3_input or 4Y25_af3;5T10_af3) but all gave me similar errors

It must be something I missed - can anyone assist?

Thanks,
Best,
Nicolas

[DimaMolod]

Hi Nicolas,

Thanks for the detailed report. You are very close; there are two separate things happening here.

First, for AlphaFold3 feature files the protein list needs to contain the JSON filename, including the .json suffix. So in your case Data/protein_list5.txt should be:

4Y25_af3_input.json;5T10_af3_input.json

Bare names such as 4Y25;5T10 are interpreted as AlphaFold2-style feature names, so AlphaPulldown looks for pickle files such as 4Y25.pkl and 5T10.pkl. It will not automatically map 4Y25 to 4Y25_af3_input.json.

Second, the traceback shows that the job is still running the old 2.0.3 installation:

/vast/hpc/x86_64-linux/alphapulldown/2.0.3/miniconda/bin/run_structure_prediction.py

That explains why even the JSON form gave a similar error. Please make sure the environment used by the SLURM job is the 2.4.0 environment and that it has the current parser installed:

which run_multimer_jobs.py
which run_structure_prediction.py
python -c "from importlib.metadata import version; print(version('alphapulldown')); print(version('alphapulldown-input-parser'))"

You should see AlphaPulldown 2.4.0 and alphapulldown-input-parser>=0.4.0, and the script paths should not point to the 2.0.3 installation.

For the second command, something like this should be enough:

run_multimer_jobs.py \
  --mode=custom \
  --num_cycle=3 \
  --output_path=Output/models \
  --data_dir=/scratch4/datasets/alphafold3 \
  --protein_lists=Data/protein_list5.txt \
  --monomer_objects_dir=Output/msa \
  --fold_backend=alphafold3 \
  --use_ap_style=True \
  --job_index=$SLURM_ARRAY_TASK_ID

--monomer_objects_dir=Output/msa is still the right option name here; for AF3 it points to the directory containing the generated JSON feature files. The AF2-specific options such as --num_predictions_per_model, --multimeric_template, and --msa_depth are not needed for the AF3 backend in the structure prediction step. Any templates found during AF3 feature generation should already be represented in the generated *_af3_input.json files.

I also updated the wiki to make this clearer: AF3 entries in protein_list.txt should keep the generated *_af3_input.json filename rather than using the bare sequence ID.

[DimaMolod]

Hi @ncoudray
just checking if removal of --multimeric_template=True and --msa_depth=128 solved your issue?

Cheers,
Dima

[ncoudray]

Hi @DimaMolod

Thanks so much for the feedback - Sorry it's taking me some time to get back - I think it's ok but not 100% sure, so I'll get back to you in a few days about it (it seems to have gone past this error but is now giving me another one which is linked to an incomplete installation - I am just waiting for the admin to update the installation before giving it another try)

I'll keep you updated -
Thanks,
Cheers,
Nicolas