AF3 backend in pulldown mode overwrites output files instead of creating per-pair directories

[abiadak]

Hi,

when using the AlphaFold3 backend in pulldown mode, I’ve noticed that output files are overwritten in the same output directory, instead of being organized into separate subdirectories per protein pair (as is done with the AF2 backend).

For example, the following files are generated in the same output directory and get overwritten across different pairs/runs:
This makes it difficult to run multiple pulldown predictions in a single output path, since results from different protein pairs overwrite each other:

Q99LX0_feature_metadata_2024-09-20.json
Q99MN1_feature_metadata_2024-09-23.json
Q9D1B9_feature_metadata_2024-09-20.json
Q9D1C3_feature_metadata_2024-09-23.json
ranked_0_confidences.json
ranked_0_data.json
ranked_0_model.cif
ranked_0_summary_confidences.json
ranking_scores.csv
seed-1617292632_sample-0
seed-1617292632_sample-1
seed-1617292632_sample-2
seed-1617292632_sample-3
seed-1617292632_sample-4
seed-3171458788_sample-0
seed-3171458788_sample-1
seed-3171458788_sample-2
seed-3171458788_sample-3
seed-3171458788_sample-4

Is this the expected behavior for the AF3 backend, or is there a recommended way (or planned change) to have one output subdirectory per protein pair, similar to the AF2 workflow?

Thanks for your help.

[DimaMolod]

Hi,
I think you are doing something wrong. The pipeline should create a subdirectory for each complex and it does so when running the test predictions:
https://github.com/KosinskiLab/AlphaPulldownSnakemake/blob/main/config/config.yaml
Maybe there is something wrong with your output directory?

[abiadak]

Hi,

I’m not using the Snakemake pipeline. Instead, I’m running the run_multimer_jobs.py script directly.

In this setup, when using the AlphaFold3 backend in pulldown mode, the output directory is not automatically split into per-complex subdirectories, and this files end up being written to the same path:

ranked_0_confidences.json
ranked_0_data.json
ranked_0_model.cif
ranked_0_summary_confidences.json
ranking_scores.csv

The workaround I’ve used so far is to run each complex independently, explicitly setting the output directory to what would normally be the automatically created one (e.g. ID1_and_ID2). This avoids file overwriting, but requires manual handling on the user side.

[DimaMolod]

Sorry, now I get it.
If you want to predict via run_multimer_jobs.py, just add this flag to the script:

run_multimer_jobs.py --use_ap_style <the rest of your flags>

that should help.