# Knowledge Map

This repository builds the microbiome + imaging phenotype portion of the SanoMap knowledge graph. All extracted artifacts remain organized around the same direct-evidence policy that MINERVA uses: only verified figure edges are promoted to `CORRELATES_WITH`, only text-derived phenotype-to-disease edges are admitted as `ASSOCIATED_WITH`, and direct text subject-to-phenotype candidates or shared-disease bridge matches stay in audit-only lanes.

The following Mermaid diagram makes the current schema explicit:

```mermaid
flowchart LR
  M[Microbe] -->|CORRELATES_WITH| RF[RadiomicFeature]
  M -->|CORRELATES_WITH| BCF[BodyCompositionFeature]
  MS[MicrobialSignature] -->|CORRELATES_WITH| RF
  MS -->|CORRELATES_WITH| BCF
  M -. audit-only text axis .-> RF
  M -. audit-only text axis .-> BCF
  MS -. audit-only text axis .-> RF
  MS -. audit-only text axis .-> BCF
  RF -->|ASSOCIATED_WITH| D[Disease]
  BCF -->|ASSOCIATED_WITH| D
  RF -->|MEASURED_AT| BL[BodyLocation]
  BCF -->|MEASURED_AT| BL
  RF -->|ACQUIRED_VIA| IM[ImagingModality]
  BCF -->|ACQUIRED_VIA| IM
  IM -->|REPRESENTED_BY| IR[ImageRef]
  M -->|POSITIVELY_CORRELATED_WITH|NEGATIVELY_CORRELATED_WITH| D
  BH["Bridge Hypotheses (audit-only)"]
  BH -. audit-only .-> D
```

Each edge is grounded in a particular pipeline stage:

- `Microbe`/`MicrobialSignature` candidates come from the MINERVA-aligned NER + cleanup pipeline.
- `RadiomicFeature` and `BodyCompositionFeature` nodes originate in the radiomics text extractor plus vision verification stages.
- `Disease` mentions are filtered through the shared span cleanup helper before entering the relation extractor.
- `BodyLocation` nodes (liver, lung, colon, abdomen, etc.) are extracted from text mentions that carry a `body_location` field. They link to phenotype features via `MEASURED_AT`.
- `ImagingModality` nodes (CT, MRI, PET, DXA, etc.) are extracted from text mentions that carry a `modality` field, with DICOM codes where applicable. They link to phenotype features via `ACQUIRED_VIA` and to representative figures via `REPRESENTED_BY`.
- `ImageRef` nodes are first-class graph nodes produced from verified Vision Track figures. Each node stores the PMCID, figure ID, panel ID, topology (heatmap/forest_plot), and image path. They complete the professor's four-part chain: `Disease ← Feature → BodyLocation / ImagingModality → ImageRef`.
- `POSITIVELY_CORRELATED_WITH|NEGATIVELY_CORRELATED_WITH` edges are direct signed microbe-disease edges derived from MINERVA-aligned relation extraction with self-consistency, separate from the phenotype axis.
- `PhenotypeAxisCandidate` rows capture direct text subject-to-phenotype evidence candidates and remain explicitly marked `not_for_graph_ingestion`.
- `BridgeHypothesis` rows are retained in audit artifacts but explicitly marked as `not_for_graph_ingestion` so they never become asserted edges.