Evidence Model¶
Purpose¶
The Evidence Workspace assembles multi-source biomedical evidence into deterministic claims and ranked hypotheses for a research question and disease context.
Maturity: BETA — suitable for exploratory research, not clinical decision support.
Core types¶
Defined in pipeline/evidence_workspace/schemas.py:
| Type | Role |
|---|---|
ResearchRequest |
Normalized user query + disease context |
EvidenceRecord |
Raw retrieval from a source adapter |
Claim |
Deterministic assertion extracted from evidence |
RankedCandidate |
Scored drug/target with explainable factors |
EvidenceDossier |
Aggregated output bundle |
GraphExplanation |
KG-backed rationale |
Citation |
Bibliographic / API reference |
Source adapters¶
Adapters under pipeline/evidence_workspace/sources/ and related modules:
| Source | Status |
|---|---|
| PubMed (Entrez) | STABLE |
| ClinicalTrials.gov v2 | STABLE |
| GWAS Catalog | BETA |
| FDA labels (openFDA) | BETA |
| Open Targets | BETA |
| GTEx, ChEMBL, bioRxiv | BETA (live connectors) |
Each adapter reports SourceStatus (fresh/stale/error).
Claim generation¶
- Adapters fetch normalized records.
- Deterministic extractors produce claims (no LLM required).
- Optional LLM enrichment (
OPENAI_API_KEY) — EXPERIMENTAL, disabled by default in secure deployments.
Deduplication & ranking¶
deduplicate_evidence()merges near-duplicate records- Ranking combines evidence strength, graph proximity, and configured weights
- Outputs include limitations and coverage warnings
Integrity controls¶
| Control | Implementation |
|---|---|
| Disease context isolation | normalize_request(), coverage checks |
| Provenance attachment | Shared build_provenance() |
| Secret-free metadata | Fingerprints exclude volatile fields |
| Research disclaimer | API + report templates |
Relationship to disease KG¶
Evidence claims may reference entities in the per-disease JSON KG. The universal biomed store provides ontology-resolved identifiers for cross-disease comparison.
Not implemented¶
- FHIR Evidence resource export (PLANNED)
- Phenopacket bundles (PLANNED)
- Real-world evidence ingestion (NOT_IMPLEMENTED)