Reproducing the family-curation coverage

Overview · Gene index · Family index

The curated records live in interpro/panther/PTHR*/PTHR*-review.yaml. This directory holds the scoped input snapshots, source extracts, index builders and validation reports. The scripts reproduce retrieval, joins and reports; they do not generate biological judgments.

From the repository root:

UV_NO_SYNC=1 uv run python projects/PROTNLM_EVALUATION/family-curation/build_mapping.py
UV_NO_SYNC=1 uv run python projects/PROTNLM_EVALUATION/family-curation/build_index.py
UV_NO_SYNC=1 uv run python projects/PROTNLM_EVALUATION/family-curation/unassigned-checks.py
UV_NO_SYNC=1 uv run python projects/PROTNLM_EVALUATION/family-curation/validate_families.py

• scope.csv freezes 288 cohort memberships, representing 282 distinct accessions. build_scope.py can rebuild selection from the explicit cohort files and available gene-review identifiers; compare its output before using it with newer cohorts.
• uniprot-records.jsonl.gz preserves exact-accession responses, request URLs, retrieval times and errors. uniprot-retries.jsonl.gz resolves six failed initial requests without altering the initial snapshot. All 282 records have successful responses; no accession redirects were silently substituted.
• baseline-membership.csv preserves the applicable membership-index observations from the start of this pass. build_mapping.py uses that frozen baseline and the successful UniProt responses. No gene-symbol inference or canonical bridge is used to create an exact-record PANTHER assignment.
• fetch_records.py refuses to overwrite a snapshot. Use a new dated bundle for a refresh. fetch_family_sources.py preserves cached PANTHER metadata and raw integrated InterPro responses under family-sources/; three integrated entries return HTTP 410. Such responses are absence of usable source content, not negative biological evidence.
• mapping.csv and family-groups.json cover the 210 exact-assignment families. additional-context-families.json records the additional SPCS2 family separately. coverage.csv joins every exact input to either a family review or an individual domain/family assessment.
• unmapped-sources.jsonl.gz, unassigned-results.json and unassigned-checks.py preserve independent checks of the 18 inputs without exact-record PANTHER assignment, including shared MOD identifiers and sequence relationships.
• Family source extracts retain reviewed status, evidence codes and generated-description flags. extract_root_sources.py regenerates one batch of deterministic source extracts. dossiers/ contains working source leads, which are not independent experimental validation. Raw sources remain available for checking the excerpts.
• decisions.jsonl and root-term-decisions.json mirror the final root-batch family judgments. The authored family YAMLs are authoritative; assignments.json and the batch result files delimit work and checks.
• membership-update.json, the batch-C result file and additional-context-families.json document additions to the shared accession-to-PANTHER index. Only observed exact-record assignments were added. Canonical family assignments were not projected onto short benchmark inputs.
• validation-all.json reports schema, family membership, anchored residues, source paths and literal quotations for all 211 reviews. validation-go.json and its frozen QuickGO responses cover ontology identifiers in the 210 exact-assignment families. PTHR13085-go.json checks the additional SPCS2 term. family-gene-crosscheck.txt records comparison with the available gene-review corpus.

Family summaries and term assessments describe the current biological interpretation. Curation-session provenance is recorded separately under history/other/PTHR*/ and history/projects/PROTNLM_EVALUATION/.