Mixing features from AF2 and AF3 (ligands related)

[epkumpu]

Hi!

I am running the snakemake based pipeline with precomputed (AF2-style) features from your database and AF3 backend for inference. This works fine with protein only. I would now like to include ligand molecules. As I understand it, ligands for the inference can only come from AF3-style features (meaning data pipeline should be AF3), but if I try to use AF3 for data pipeline, it no longer recognizes the feature database's AF2-format feature files and runs MSAs again for the protein sequences. I would like to avoid this if possible, so that I can take advantage of the feature database.

Any thoughts on if it is going to be possible to use the AF2-style features for proteins and AF3-style features for ligands when using AF3 inference backend or is this going to hit a wall somewhere? On the surface it seems like this should be possible as we can use both feature types separately for AF3 inference backend.

Cheers,
EP

[DimaMolod]

Hi,

Yes, this should be possible.

The key is to separate feature generation from AF3 inference input preparation. You should not need to run the AF3 data pipeline for protein chains if you already have AF2/AlphaPulldown .pkl features.

The AF3 backend can mix AlphaPulldown protein objects loaded from AF2-style .pkl files with AF3 JSON inputs in the same job. It converts the protein features into AF3 protein-chain inputs, adds the AF3 JSON chains, writes one combined AF3 input JSON, and skips the data pipeline, so protein MSAs should not be re-run.

So the practical route is to keep using the AF2-style feature database for proteins, and provide/generate an AF3 JSON for the ligand, then use both in the same fold specification, e.g.:

ProteinA+ligand.json

where ProteinA.pkl comes from the AF2 feature database and ligand.json is the AF3-format input for the ligand.
One caveat: create_individual_features.py currently cannot create ligand AF3 JSONs from FASTA. Its AF3 FASTA path only recognizes DNA, RNA, and protein/peptide descriptions. Therefore the ligand currently needs to be provided as an AF3 JSON input directly, e.g. using ccdCodes or smiles.

Hope this helps,
d

[epkumpu]

I tested AlphaPulldownSnakemake 2.1.8 + AlphaPulldown 2.4.0.

Fold line:
+ligand.json:80

feature_directory in config.yaml points to the folder containing the AF2-style database .pkl files AND another folder with ligand.json

config.yaml has the following:
create_feature_arguments:
--data_pipeline=alphafold2
structure_inference_arguments:
--fold_backend=alphafold3

But Snakemake dry-run still schedules:
download_uniprot ligand
create_features ligand.pkl.xz

So the Snakemake workflow seems to treat the JSON ligand as a protein token before AlphaPulldown starts working on it. Any pointers?

Cheers,
EP

[DimaMolod]

Hi EP,

Thanks for testing. I checked the AlphaPulldownSnakemake workflow, and I think you are right: this currently works in standalone AlphaPulldown, but not through the AlphaPulldownSnakemake wrapper.

For now, the easiest workaround is to run AlphaPulldown directly with run_structure_prediction.py for this mixed case. Use the AF2 feature database for the protein .pkl files, and include the folder containing ligand.json in --features_directory.

For example, conceptually:

--input ProteinA+ligand.json
--features_directory /path/to/af2_features,/path/to/ligand_json_folder
--fold_backend alphafold3

Supporting this in AlphaPulldownSnakemake would require a workflow change so JSON inputs are treated as AF3 JSON inputs, not as proteins that need feature generation.

I will mark this issue with the bug label and transfer it to the AlphaPulldownSnakemake repo as related.