Support AF3 JSON inputs (ligands) in folds; don't treat them as proteins (#41)

[DimaMolod]

Fixes #41.

Problem

Running a mixed fold like <UNIPROTID>+ligand.json:80 (AF2 .pkl protein features + an AF3 ligand.json, AF3 backend) failed: a snakemake -n dry-run scheduled download_uniprot ligand and create_features ligand.pkl.xz, i.e. the .json token was treated as a protein to fetch/build features for.

Root cause was two layers deep:

1. The parser's FoldDataset stripped the .json extension during normalization, so ligand.json arrived here as ligand — fixed in Preserve .json tokens in FoldDataset normalization (0.5.1) alphapulldown-input-parser#4 (0.5.1).
2. This workflow then derived its download_uniprot / create_features / inference-input targets without distinguishing JSON inputs from proteins.

Standalone AlphaPulldown already supports this (CLI uses parse_fold; backend has a golden-tested protein+ligand path) — the gap was only in the Snakemake wrapper, as noted in the issue.

Changes

workflow/rules/common.smk (pure, unit-tested):

• is_json_input(name) — detect *.json tokens.
• split_fold_inputs(fold) — partition a fold into protein chains vs direct AF3 JSON inputs.
• format_af3_requested_fold(fold) — moved here from the Snakefile and made JSON-aware: proteins → <base>_af3_input.json; *.json passed through unchanged (previously it produced ligand.json_af3_input.json).
• chain_residue_count reads JSON inputs (ligand-only → 0).

workflow/Snakefile:

• Feature targets (kept_proteins / required_feature_paths), lookup_features, and the length filter now go through split_fold_inputs. JSON inputs are required as the JSON file itself (provided via feature_directory → symlink_features) and are never downloaded or generated.
• symlink_features made idempotent (skips already-correct links, handles broken links).

Pin: requires the parser's .json-preserving normalization → alphapulldown-input-parser>=0.5.1 (env, CI, README).

Verification

snakemake -n on P12345+ligand.json:80:

	before	after
jobs	download_uniprot, create_features, symlink_features, structure_inference	symlink_features, structure_inference
ligand	downloaded + create_features ligand.pkl.xz	symlinked ligand.json, used directly

• AF2 features + AF3 backend: inference inputs P12345.pkl.xz, ligand.json.
• Pure AF3: --input P12345_af3_input.json+ligand.json:80 (ligand not double-suffixed).

Unit tests added for all new helpers; full suite green (25 passed).

Note on ordering

The runtime fix needs parser 0.5.1 on PyPI (KosinskiLab/alphapulldown-input-parser#4). CI here pins >=0.5.1, so it will go green once 0.5.1 is published. The added unit tests cover the workflow helpers directly and pass regardless.

🤖 Generated with Claude Code