Medical Literature Knowledge Graph

For a while now I’ve been telling anybody who’d listen that knowledge graphs are a good way to provide grounded information to language models. In recent months I’ve been slowly building a graph with a bunch of PubMed papers on cancer research, and finally I’m starting to see results worth talking about.

This conversation demonstrates:

• Natural language querying of a medical knowledge graph
• GraphQL API for structured entity and relationship retrieval
• Semantic search capabilities across biomedical entities
• Relationship traversal to discover connections between entities
• Provisional vs canonical entity distinction based on standardization status
• Usage tracking and confidence scores for entity mentions

What really is interesting about this is that (at least to me, and I am very much not a doctor), it sounds pretty intelligent. Nothing it’s saying is crazy or stupid.

If you want to view the graph and click on nodes and things, go to http://104.131.171.175/graph-viz/.

---

User: show me the first 15 biomarkers please

Assistant: I'll query for the first 15 biomarkers in the knowledge graph.

{
  biomarkers: entities(filter: { entityType: "biomarker" }, limit: 15) {
    items {
      entityId
      entityType
      name
      status
      confidence
      usageCount
      source
      canonicalUrl
    }
    total
  }
}

First 15 Biomarkers (out of 81 total)

1. CA-125

   • Status: provisional
   • Confidence: 0.45, Usage: 2
   • Source: PMC12757429

2. JC-1 fluorescent probe

   • Status: provisional
   • Confidence: 0.475, Usage: 1
   • Source: PMC12757429

3. BCR Shannon diversity

   • Status: provisional
   • Confidence: 0.35, Usage: 2
   • Source: PMC12757429

4. ...

5. DDR-IF score

   • Status: provisional
   • Confidence: 0.475, Usage: 21
   • Source: PMC12757875

6. ...

7. STING activation score

   • Status: provisional
   • Confidence: 0.475, Usage: 1
   • Source: PMC12757875

Observations:

• All 15 biomarkers shown are provisional (none have been promoted to canonical status yet)
• They come from 2 medical papers (PMC12757429 and PMC12757875)
• Usage counts range from 1 to 21 (DDR-IF score being the most used)
• Confidence scores range from 0.35 to 0.475
• Many relate to cancer research (STING pathway, TCGA cohort, CA-125, etc.)

Query 3: DDR-IF Score Investigation

User: DDR-IF score is discussed (possibly for the first time?) in this paper: https://pubmed.ncbi.nlm.nih.gov/41487593/ - curious that it has no canonical ID

Assistant: Excellent detective work! You're absolutely right. The DDR-IF score has no canonical ID because it's a novel composite biomarker introduced in that very recent paper (PMC12757875, which corresponds to PubMed ID 41487593).

Why DDR-IF score is provisional (not canonical):

Novel Research Contribution: The paper introduces the DDR-IF score as a new composite biomarker that integrates:

• DNA Damage Response (DDR) alterations (HRD component)
• Immune Infiltration (IF) features (TME/tumor microenvironment)
• STING activation (immune signaling)

From the relationships extracted:

1. Purpose: Predicts PARP-ICB (PARP inhibitor + immune checkpoint blockade) response in ES-SCLC (extensive-stage small cell lung cancer)

   • AUROC = 0.81 (strong predictive power)

2. Clinical utility: Identifies tumors most susceptible to PARP-ICB combination therapy

3. Cut-off: ≥0.65 defines "DDR-IF-high" status

Why no canonical ID exists:

• This is cutting-edge research (2026 paper based on the creation date in properties)
• Not yet incorporated into standardized biomarker databases like:
  • MeSH (Medical Subject Headings)
  • LOINC (Logical Observation Identifiers Names and Codes)
  • NCBI Gene (for molecular markers)
• May require validation in multiple studies before databases assign canonical identifiers
• It's a composite score (not a single molecule/gene), which makes canonicalization more complex

The knowledge graph correctly identified it as provisional with decent usage (21 mentions across 3 papers), showing it's gaining traction but not yet standardized in biomedical ontologies.

Query 4: Understanding STING Activation

User: dumb question, what is STING activation?

Assistant: Not a dumb question at all! STING is actually pretty complex immunology. Let me search the knowledge graph for more details.

STING = Stimulator of Interferon Genes

What it does:

STING is part of your innate immune system - your body's first-line defense against threats. Think of it as an alarm system inside cells.

How it works:

1. Detection: When DNA appears in the wrong place inside a cell (e.g., from viruses, bacteria, or damaged cancer cells), a sensor called cGAS detects it

2. Activation: cGAS triggers STING protein, which gets phosphorylated (p-STING in the biomarker list)

3. Alarm response: Activated STING causes the cell to produce:

   • Interferons (signaling proteins that alert the immune system)
   • Inflammatory cytokines (recruit immune cells to attack)

Why it matters in cancer:

The double-edged sword:

• Good news: Cancer cells with damaged DNA can trigger STING → immune system attacks tumor
• Bad news: Chronic activation can lead to "cGAS-STING exhaustion" where the pathway stops working (like a fire alarm that gets ignored)

In the DDR-IF score context:

• Tumors with DNA damage (from PARP inhibitors) → release DNA fragments → activate STING
• STING activation → recruits immune cells → makes immunotherapy (checkpoint blockade) work better
• The DDR-IF score measures STING activation to predict which tumors will respond to the PARP + immunotherapy combination

Clinical relevance:

• "STING-desert phenotypes" = tumors with no STING activity (won't respond well to immunotherapy)
• "STING agonists" = drugs being developed to artificially activate this pathway

So STING activation is basically your cell's burglar alarm going off when it detects suspicious DNA, calling in the immune system cavalry!