Ship the cofactor list with the pathway artifacts, derived not typed

[adamjohnwright]

Why

A consumer deciding whether a perturbation may travel through ATP needs to know which nodes are ATP — and that knowledge has to travel with the networks. It didn't, so this repo and DeltaSignal each kept a copy, and the copies drifted.

Audited against Release97, 6 of the 13 hand-written ids in _COFACTOR_STIDS are wrong:

stable id	comment says	Release97 actually is
R-ALL-29438	PPi	GTP [cytosol]
R-ALL-29390	Pi variant	PXLP [cytosol] (pyridoxal 5'-phosphate)
R-ALL-29360	ADP variant	NAD+ [cytosol]
R-ALL-217093	NADP+	absent from the release
R-ALL-110114	NADPH	absent from the release
R-ALL-29986	NAD+	absent from the release

Nothing caught any of it, because the list is only ever read here.

What this does

Derives the list. get_cofactor_species() returns every SimpleEntity whose ReferenceMolecule carries one of 28 curated ChEBI identifiers — 253 species at Release97, every compartment variant included. ChEBI identity is the only stable handle: a name query misses compartment variants and matches by substring ("phosphate" pulls in PXLP), and stable ids go stale between releases. This list managed both failure modes.

Ships it with the artifacts. export_cofactors() writes cofactors.csv into each pathway directory:

stable_id,molecule,chebi_id,name,in_network,reactome_release
R-ALL-113592,ATP,30616,ATP [cytosol],1,97

Pull a pathway out of S3 and it carries its own answer, pinned to the release it came from. Every known cofactor is listed rather than only the present ones, so a consumer can distinguish an empty intersection from a bundle predating the file — and the file is written even when a pathway contains none.

What this deliberately does not do

The networks are unchanged. This adds an artifact. The existing _COFACTOR_STIDS bridge suppression is left exactly as it is, defects and all: correcting those six entries changes which bridges are suppressed, which changes the networks, which needs its own A/B. That is a separate PR.

Testing

• 2 new tests in tests/test_provenance_exports.py covering the all-listed/flag-present contract and the write-even-when-empty case
• mypy src/ clean, ruff check src/ bin/ clean
• pytest tests/ — 977 passed. One pre-existing failure (test_main_edges_proportional_to_best_matches[Class_I_MHC…]) reproduces on main unchanged and is unrelated.

🤖 Generated with Claude Code