knowledge-graph-tools

Files (2)

• SKILL.md
• knowledge_graph.py · template/

SKILL.md (first lines)

# Knowledge Graph Tools

Use this skill for building and querying biomedical relationship graphs from public drug-discovery APIs, not for clinical decision-making.

Typical triggers:
- build a knowledge graph seeded from a disease, a drug, or a target list
- find shortest paths from a drug to a disease through intermediate targets and pathways
- identify hub targets that bridge multiple disease areas or drug mechanisms
- expand the neighborhood around a protein to see connected drugs, diseases, and pathways
- merge OpenTargets, ChEMBL, STRING, and Reactome data into one queryable graph

## Working Rules

1. Every node gets a typed label: `drug`, `target`, `disease`, or `pathway`.
2. Every edge records its source database and, where available, an evidence score.
3. Use canonical identifiers: Ensembl for targets, ChEMBL for drugs, EFO for diseases, Reactome stable IDs for pathways.
4. Hub analysis reflects database connectivity, not biological importance; well-studied proteins dominate.
5. Shortest-path hypotheses are topological leads, not validated biology.
6. Do not claim causal or therapeutic conclusions from graph structure alone.

## Environment Check

```bash
which python3 || true
python3 - <<'PY'
mods = ["networkx", "requests"]
for name in mods:
    try:
        __import__(name)
        print(f"{name}: ok")
    except Exception as exc:
        print(f"{name}: missing ({exc})")
PY
```

If networkx or requests is missing, say so immediately. If network access is blocked, only the `query` mode on pre-built GraphML files will work.

## Bundled Assets

- `templates/knowledge_graph.py`

## Build: Disease-Centric Graph

Use `templates/knowledge_graph.py --mode build --seed-type disease` for:
- fetching disease-associated targets from OpenTargets
- fetching known drugs for those targets from OpenTargets
- adding protein-protein interactions from STRING
- adding pathway membership from Reactome
- assembling typed nodes and edges into a single graph

Quick start:

```bash
python3 templates/knowledge_graph.py \
  --mode build \
  --seed-type disease \
  --seed "Crohn's disease" \
  --max-targets 30 \
  --include-string \
  --include-reactome \
  --output kg/crohn_graph.graphml \
  --summary kg/crohn_summary.json
```

Deliverables:
- GraphML file with typed nodes (`entity_type`) and typed edges (`relation`, `source_db`, `score`)
- summary JSON with node/edge counts by type, top hubs, and data sources queried

## Build: Drug-Centric Graph

Use `--seed-type drug` to start from a drug and expand through its targets:

```bash
python3 templates/knowledge_graph.py \
  --mode build \
  --seed-type drug \
  --seed "imatinib" \
  --max-targets 20 \
  --include-string \
  --include-reactome \
  --output kg/imatinib_graph.graphml \
  --summary kg/imatinib_summary.json
```

## Query: Shortest Path

Use `--mode query --query-type shortest-path` on an existing GraphML file:

```bash
python3 templates/knowledge_graph.py \
  --mode query \
  --input kg/crohn_graph.graphml \
  --query-type shortest-path \
  --from-node "CHEMBL941" \
  --to-node "EFO_0000384" \
  --summary kg/path_result.json
```

Deliverables:
- summary JSON with path length, node sequence, and edge relations for each step

## Query: Hub Analysis

Use `--mode query --query-type hubs`:

```bash
python3 templates/knowledge_graph.py \
  --mode query \
  --input kg/crohn_graph.graphml \
  --query-type hubs \
  --top-k 20 \
  --summary kg/hub_targets.json
```

Deliverables:
- summary JSON with top-K nodes ranked by degree and betweenness centrality, with entity type

## Query: Neighborhood Expansion

From DrugClaw/DrugClaw, Apache-2.0.

More from DrugClaw/DrugClaw