Skip to content

BACKLOG.md

Latest commit

 

History

History
61 lines (54 loc) · 4.2 KB

BACKLOG.md

File metadata and controls

61 lines (54 loc) · 4.2 KB
Raw

Backlog

Open items, not yet scheduled as a plan. When one of these gets picked up, it goes through the normal brainstorming → spec → plan → subagent-driven-development cycle like everything else in this repo — this file is just the "what's left" list.

Smaller open items

All previously reviewed and explicitly parked as non-blocking — listed here so they don't get lost, not because any of them are urgent.

  • phiexplorer/dereference/chain.py:97 — taxid_to_name_map() uses org["full_name"] (bracket access) where the rest of the file uses .get(). Cosmetic; would only matter if an organism entry is ever missing full_name, which the current fixture and design don't exercise. (Flagged in the core-toolkit plan's final review.)
  • Test file structure inconsistency — tests/dereference/__init__.py exists but tests/extract/ and tests/reports/ don't have one. Harmless (pytest doesn't need it, test basenames are unique across the suite) but worth normalizing one way or the other.
  • phiexplorer/extract/_collect.py's INFECTIVE_ABILITY_TERMS fallback path and the wild-type allele filter have zero test coverage — the fixture's data always short-circuits both (every infective_ability extension already has a rangeDisplayName, and both fixture alleles are "deletion", never "wild type"). Two small hand-built test cases would close it. (Flagged in the extract-consolidation plan's final review.)
  • _organism_slug() in phiexplorer/reports/generate.py doesn't neutralize path-hostile characters (e.g. a / in a scientific name would break the output path). Hardening, not a live bug — PHI-base's actual fungal/bacterial organism names don't produce this today. (Flagged in the reports-layering plan's final review.)
  • No dataset-wide phenotype-extraction helper — extract_protein_phenotypes() takes one organism at a time; a dataset-wide analysis (e.g. "phenotype counts across all organisms") has to loop over chain.all_organisms() and concatenate results by hand, as done in Data-analysis-MU/2026-08-20-all-organisms-high-level-phenotype-counts.md. A extract_all_organisms_phenotypes()-style wrapper would close this.
  • 29 UniProt IDs are curated under two or more different organism entries in the v5.3 export (e.g. the effector avrB/P13835 under both Pseudomonas syringae and Pseudomonas savastanoi; SIX3/Q2A0P1 under four different fungal genera). Confirmed by inspecting the raw duplicate rows — looks like shared/ambiguous strain assignment or genuine cross-referenced orthologs in PHI-base's own curation, not an extraction bug, but any dataset-wide aggregation by UniProt ID must group/dedupe across organisms or it will overcount (see the same analysis note above). Not something to "fix" in phiexplorer; flagged here so future dataset-wide code doesn't rediscover it the hard way.
  • reports/ isn't benchmark-validated against real data the way extract/ is via phiexplorer/smoke.py. The reports-layering plan's manual real-data verification step did confirm the report files reproduce the correct numbers, but that wasn't turned into a permanent regression check. A smoke.py extension asserting actual output-file row counts (not just the DataFrame the writer wraps) would close this properly.

Possible future extraction dimensions

Not committed to — these are "Common Extensions" the ported prior-art workflow doc (PHI_ANALYSIS_WORKFLOW.md, in James Seager's original PHI5-zenodo-datamining project) named as future directions, never built here. Worth a fresh brainstorming pass each if they turn out to matter:

  • Cross-species comparative analysis — organism-wise summary tables, comparing phenotype distributions or host ranges across multiple organisms in one call, rather than one organism at a time.
  • Temporal analysis — curation trends over time via metadata.accepted_timestamp and publication year, tracking how curation of a topic has grown.
  • Effector protein cross-referencing — the current is_effector() rules are keyword + GO-term based; the original workflow doc also suggested cross-referencing against UniProt's own secreted-protein predictions, not yet implemented.