Skip to content

LLM-PathwayCurator

LLM-PathwayCurator Enrichment interpretations → audited, decision-grade pathway claims.

LLM-PathwayCurator is a quality-assurance (QA) layer for enrichment analysis (EA) interpretation.
It does not introduce a new enrichment statistic. Instead, it turns EA outputs (ranked term lists) into evidence-linked, typed claims and assigns PASS/ABSTAIN/FAIL via a mechanical audit suite.

Core promise: we abstain when claims are unstable, under-supported, contradictory, or context-nonspecific.

EvidenceTable → distill → modules → claims → audits → report

Why this exists (the practical pain)

Enrichment tools return ranked term lists. In practice, interpretation becomes non-reproducible because:

  • representative terms are ambiguous under the study context
  • gene support is opaque → cherry-picking risk
  • related terms share / bridge evidence in non-obvious ways
  • there is no mechanical stop condition for fragile narratives

LLM-PathwayCurator converts “plausible narratives” into auditable decision objects.

What you get

  • EvidenceTable: term × supporting-genes contract (works for ORA and rank-based EA)
  • Evidence distillation: supporting-gene perturbations → stability proxies (survival-like scores)
  • Evidence modules: factorization of the term–gene graph (shared vs distinct evidence)
  • Typed claims (JSON): schema-bounded, evidence-linked (no free text required)
  • Mechanical audits: predefined gates → PASS/ABSTAIN/FAIL + reason codes
  • Decision-grade report: audit log + provenance + reproducible outputs

Quick start

pip install llm-pathway-curator

llm-pathway-curator run \
  --sample-card examples/demo/sample_card.json \
  --evidence-table examples/demo/evidence_table.tsv \
  --out out/demo/

Key outputs: - audit_log.tsv (PASS/ABSTAIN/FAIL + reason codes) - report.md, report.jsonl (decision objects) - distilled.tsv, modules.tsv, term_modules.tsv, term_gene_edges.tsv - run_meta.json (pinned parameters + provenance)

Next

Notes

  • LLM is proposal-only (optional): representative selection + typing.
  • Acceptance is never delegated: PASS/ABSTAIN/FAIL is decided by mechanical audits.
  • Counterfactual stress tests are internal (e.g., context swap, evidence dropout): no external knowledge required.