Life Sciences Graph Builder

Orchestrate multi-API graph construction using the Fuzzy-to-Fact protocol.

Architecture

┌─────────────────────────────────────────────────────────────────────────┐
│                         GRAPH CONSTRUCTION KIT                          │
├─────────────────────────────────────────────────────────────────────────┤
│  TIER 1: MCP TOOLS (Verified Nodes)                                     │
│  ├── HGNC: search_genes, get_gene                                       │
│  ├── UniProt: search_proteins, get_protein                              │
│  ├── ChEMBL: search_compounds, get_compound                             │
│  ├── STRING: search_proteins, get_interactions                          │
│  ├── Open Targets: search_targets, get_associations                     │
│  └── WikiPathways: get_pathways_for_gene, get_pathway_components        │
├─────────────────────────────────────────────────────────────────────────┤
│  TIER 2: CURL COMMANDS (Relationship Edges)                             │
│  ├── ChEMBL /mechanism: Drug → Target                                   │
│  ├── ChEMBL /drug_indication: Drug → Disease                            │
│  ├── ChEMBL /activity: Drug → Target (with Ki/IC50)                     │
│  ├── Ensembl /homology: Gene → Orthologs                                │
│  ├── STRING /enrichment: Protein Set → GO/KEGG terms                    │
│  └── NCBI elink: Gene → PubMed                                          │
├─────────────────────────────────────────────────────────────────────────┤
│  TIER 3: GRAPHITI (Persistence)                                         │
│  └── add_memory: Persist validated subgraph as JSON episode             │
└─────────────────────────────────────────────────────────────────────────┘

Workflow: Fuzzy-to-Fact Protocol

Phase 1: Anchor Node (Naming)

Resolve fuzzy user input to canonical identifier.

# MCP: HGNC
result = hgnc.search_genes("p53")
gene = hgnc.get_gene("HGNC:11998")  # → cross_references: UniProt, Ensembl, Entrez

Phase 2: Enrich Node (Functional)

Decorate node with metadata and cross-references.

# MCP: UniProt
protein = uniprot.get_protein("UniProtKB:P04637")
# → function text reveals interactors: BAX, BCL2, FAS

Phase 3: Expand Edges (Interactions)

Build adjacency list from interaction databases.

# MCP: STRING
interactions = string.get_interactions("STRING:9606.ENSP00000269305")
# → MDM2 (0.999), SIRT1 (0.999), ATM (0.995)

# Curl: Open Targets (gene-disease)
curl -s -X POST "https://api.platform.opentargets.org/api/v4/graphql" \
  -H "Content-Type: application/json" \
  -d '{"query": "{ target(ensemblId: \"ENSG00000141510\") { associatedDiseases(page: {size: 5}) { rows { disease { name } score } } } }"}'

Phase 4: Target Traversal (Pharma)

Follow edges to actionable targets.

# MCP: HGNC (resolve downstream effector)
bcl2 = hgnc.search_genes("BCL2")  # → HGNC:990

# MCP: ChEMBL (find inhibitors)
venetoclax = chembl.search_compounds("Venetoclax")  # → CHEMBL:3137309

# Curl: ChEMBL mechanism (Drug → Target edge)
curl -s "https://www.ebi.ac.uk/chembl/api/data/mechanism?molecule_chembl_id=CHEMBL3137309&format=json" \
  | jq '.mechanisms[] | {action: .action_type, target: .target_chembl_id}'
# → INHIBITOR → CHEMBL4860 (BCL2)

Phase 5: Persist Graph

Store validated subgraph in Graphiti.

# MCP: Graphiti
graphiti.add_memory(
    name="TP53-BCL2-Venetoclax pathway",
    episode_body=json.dumps({
        "nodes": [
            {"id": "HGNC:11998", "type": "Gene", "symbol": "TP53"},
            {"id": "HGNC:990", "type": "Gene", "symbol": "BCL2"},
            {"id": "CHEMBL:3137309", "type": "Compound", "name": "Venetoclax"}
        ],
        "edges": [
            {"source": "HGNC:11998", "target": "HGNC:990", "type": "REGULATES"},
            {"source": "CHEMBL:3137309", "target": "HGNC:990", "type": "INHIBITOR"}
        ]
    }),
    source="json",
    group_id="drug-repurposing"
)

Quick Edge Discovery Commands

Example: Drug Repurposing Graph

Build a complete subgraph for drug repurposing analysis:

# Step 1: Anchor - Resolve gene
# MCP: hgnc.search_genes("TP53") → HGNC:11998

# Step 2: Get protein context
# MCP: uniprot.get_protein("UniProtKB:P04637")
# → function mentions BCL2

# Step 3: Find BCL2 inhibitors
curl -s "https://www.ebi.ac.uk/chembl/api/data/mechanism?target_chembl_id=CHEMBL4860&format=json" \
  | jq '.mechanisms[] | {drug: .molecule_chembl_id, action: .action_type}'

# Step 4: Get drug indications
curl -s "https://www.ebi.ac.uk/chembl/api/data/drug_indication?molecule_chembl_id=CHEMBL3137309&format=json" \
  | jq '.drug_indications[:3][] | {disease: .mesh_heading, phase: .max_phase_for_ind}'

# Step 5: Find clinical trials
curl -s "https://clinicaltrials.gov/api/v2/studies?query.intr=venetoclax&filter.overallStatus=RECRUITING&pageSize=3&format=json" \
  | jq '.studies[] | {nct: .protocolSection.identificationModule.nctId}'

# Step 6: Persist to Graphiti
# MCP: graphiti.add_memory(...)

Node Types (Canonical CURIEs)

Edge Types

Query Best Practices

Gene Discovery (Human-Centric)

• Default to species=9606 (human) for gene/protein searches
• Use page_size=10 for exploration, page_size=50 for batch operations
• Use slim=True for batch operations to reduce token usage
• Only use organism=null for comparative genomics across species

Drug Discovery vs Repurposing

• Drug repurposing: Use max_phase≥2 (clinical validation, shorter approval path)
• General discovery: No phase filter (include preclinical tools, mechanism probes)
• Check mechanisms before bioactivity data

Clinical Landscape

• Default status=RECRUITING for active research
• Use phase filter only for specific analysis:
  • PHASE3+ for commercialization analysis
  • PHASE1/2 for early pipeline
  • No filter for full landscape

See Also

• lifesciences-genomics: Ensembl, NCBI, HGNC endpoints
• lifesciences-proteomics: UniProt, STRING, BioGRID endpoints
• lifesciences-pharmacology: ChEMBL, PubChem, IUPHAR endpoints
• lifesciences-clinical: Open Targets, ClinicalTrials.gov endpoints