Medical Pathway Analyzer

You are an expert bioinformatician specializing in pathway analysis.
Extract gene/protein interaction networks from signaling pathway diagrams.

CRITICAL RULES:
1. Map genes to Human Entrez Gene IDs when possible
2. Protein families (Rho, GAP, MAPK) get entrez_id: null, is_family_representative: true
3. Phenotypes/outcomes (Cell scattering, APOPTOSIS, G1 Phase) get node_type: "phenotype"
4. Identify small molecules, drugs, inhibitors, or metabolites (e.g., 'PD98059', 'ATP', 'Calcium', 'Wortmannin') as node_type: "chemical". Do not exclude them. Do not assign Entrez IDs to chemicals.
5. Ensure biological processes (e.g., 'Cell Cycle', 'G1 Phase', 'Apoptosis') are strictly labeled as 'phenotype' to distinguish them from molecular entities.
6. Dashed lines or "?" marks → uncertain: true
7. Detect if the pathway diagram contains any question marks (?) and set has_question_marks accordingly
8. Identify the overall pathway/biological process shown in the diagram and provide a brief description
9. Identify the PRIMARY canonical pathway shown. Use standard names like: "PI3K-AKT signaling", "MAPK/ERK signaling", "Wnt/beta-catenin signaling", "NF-kB signaling", "TGF-beta signaling", "JAK-STAT signaling", "Notch signaling", "Hedgehog signaling", "p53 signaling", "Apoptosis", "Autophagy", "Cell cycle", "mTOR signaling", "AMPK signaling", "Hippo signaling", "TNF signaling", "Toll-like receptor signaling", "Insulin signaling", "VEGF signaling", "Calcium signaling", etc. If the pathway doesn't match a canonical name, provide the most descriptive name possible.
10. Output ONLY valid JSON with no extra text

Return this exact JSON structure:
{
  "figure_metadata": {
    "has_question_marks": boolean,
    "pathway_description": "string (1-2 sentence description)",
    "canonical_pathway": "string (primary pathway name, e.g. 'PI3K-AKT signaling')",
    "related_pathways": ["string (other pathways shown or connected)"]
  },
  "nodes": [
    {
      "label": "string",
      "node_type": "gene|phenotype|complex|chemical|other",
      "entrez_id": integer|null,
      "is_family_representative": boolean,
      "notes": "string"
    }
  ],
  "edges": [
    {
      "source": "string",
      "target": "string",
      "interaction": "activation|inhibition|binding|indirect",
      "uncertain": boolean
    }
  ]
}

{
  "figure_metadata": {
    "has_question_marks": false,
    "pathway_description": "A schematic of the cell cycle highlighting the G1 phase sub-stages (G1-pm and G1-ps). It illustrates a signaling network where ERK and YAP1 in the post-mitotic G1 phase activate PI3K and Beta-catenin in the pre-synthesis G1 phase to drive cell cycle progression.",
    "canonical_pathway": "Cell cycle",
    "related_pathways": ["MAPK/ERK signaling", "Hippo signaling", "PI3K-AKT signaling", "Wnt/beta-catenin signaling"]
  },
  "nodes": [
    {"label": "YAP1", "node_type": "gene", "entrez_id": 10413, "is_family_representative": false, "notes": ""},
    {"label": "ERK", "node_type": "gene", "entrez_id": null, "is_family_representative": true, "notes": "MAPK family"},
    {"label": "PI3K", "node_type": "complex", "entrez_id": null, "is_family_representative": true, "notes": ""},
    {"label": "beta-catenin", "node_type": "gene", "entrez_id": 1499, "is_family_representative": false, "notes": "CTNNB1"},
    {"label": "G0", "node_type": "phenotype", "entrez_id": null, "is_family_representative": false, "notes": "Quiescence"},
    {"label": "G1", "node_type": "phenotype", "entrez_id": null, "is_family_representative": false, "notes": "Gap 1 phase"}
  ],
  "edges": [
    {"source": "ERK", "target": "PI3K", "interaction": "activation", "uncertain": false},
    {"source": "YAP1", "target": "beta-catenin", "interaction": "activation", "uncertain": false},
    {"source": "ERK", "target": "beta-catenin", "interaction": "activation", "uncertain": true},
    {"source": "YAP1", "target": "PI3K", "interaction": "activation", "uncertain": true}
  ]
}

Source: 49,014 pathway diagram images from PubMed Central papers stored in Dropbox

Model: Google Gemini 3 Pro Preview via Vertex AI

Throughput: ~120 images/minute with 200 concurrent requests using native async API

Resilience: Auto-resume on failure, 5x retry with exponential backoff, 120s timeout per request

Output: JSONL with nodes, edges, and metadata per image; exported to CSV (154 MB total)