Disorder Mechanisms Knowledge Base (dismech)

A curated knowledge base of disease pathophysiology, with structured evidence from the literature.

⚠️ Disclaimer: AI-curated, and not medical advice

• DisMech is AI-curated and AI-maintained. Most content is generated and maintained by AI curation agents, with human review as the pull-request gate. Automated validation guarantees that citations exist, quoted snippets are exact, and ontology terms are real — it does not guarantee scientific correctness.
• DisMech is not medical advice. The contents of this resource are not intended to inform medical diagnosis or treatment. Inclusion of any statement or approach in a DisMech page is purely the result of generative methods and human review applied to publicly accessible literature, data, and other curated biomedical knowledge resources. If you have a health concern, consult a qualified healthcare professional.

Full statement: docs/disclaimer.md · rationale: Design Decisions §11

Browse the Knowledge Base

View all disorders online

Each disorder page includes:

• Disease mechanisms and pathophysiology
• Clinical phenotypes with HPO term links
• Genetic factors and variants
• Treatment options with NCIT term links
• All claims backed by PubMed evidence

Related browsers:

• Curation history dashboard — explore curation, review, and audit activity by date, agent tool, model, and section.
• Monarch NAMs browser — explore New Approach Methodologies (NAMs) and research datasets represented in DisMech.

How It Works

Project Overview Slides

• Unlocking Disease Mechanisms: Agentic AI for Clinical Knowledge - TMC AI Keynote
• Assembling mechanistic evidence for rare disease surrogate endpoints: the role of AI and NAMs
• ICBO 2026

Source of Truth: YAML Files

The knowledge base is stored as structured YAML files in kb/disorders/. Each disorder has its own file (e.g., Asthma.yaml) containing:

name: Asthma
pathophysiology:
  - name: Airway Inflammation
    description: Chronic inflammation of the airways
    biological_processes:
      - preferred_term: inflammatory response
        term:
          id: GO:0006954
          label: inflammatory response
    evidence:
      - reference: PMID:12345678
        supports: SUPPORT
        snippet: "Exact quote from the paper"
        explanation: "Why this supports the claim"
phenotypes:
  - name: Wheezing
    frequency: FREQUENT
    phenotype_term:
      preferred_term: Wheezing
      term:
        id: HP:0030828
        label: Wheezing
treatments:
  - name: Bronchodilators
    description: Relax airway smooth muscle
    treatment_term:
      preferred_term: respiratory tract agent therapy
      term:
        id: NCIT:C15986
        label: Pharmacotherapy

Ontology Bindings

Entities are linked to authoritative ontologies:

• Phenotypes: Human Phenotype Ontology (HP)
• Cell types: Cell Ontology (CL)
• Biological processes: Gene Ontology (GO)
• Diseases: Mondo Disease Ontology (MONDO)
• Treatments: NCI Thesaurus (NCIT)
• Anatomy: Uberon (UBERON)

Evidence Requirements

All claims must cite PubMed references with exact quotes. This prevents misinformation and enables verification.

Validation Pipeline

Multiple layers of automated validation ensure data quality and prevent AI hallucinations:

1. Schema validation: Ensures correct YAML structure against the LinkML schema
2. Ontology term validation: Verifies term IDs exist and labels match authoritative sources (HPO, MONDO, GO, etc.)
3. Reference validation: Confirms that quoted snippets actually appear in cited PubMed abstracts
4. Compliance analysis: Measures coverage of recommended fields (descriptions, evidence, ontology terms)

# Run schema + term validation
just qc

# Validate a single file
just validate kb/disorders/Asthma.yaml

# Validate a KB entry's evidence snippets against the cited abstracts
just validate-kb-references kb/disorders/Asthma.yaml

# Check a deep-research report's citations (a different question -- see docs)
just validate-research-reference research/Asthma-deep-research-falcon.md

# Analyze compliance with recommended field coverage
just compliance-all

# Compliance with weighted scoring and threshold checks
just compliance-weighted

Why Reference Validation Matters

All evidence snippets must be exact quotes from paper abstracts, not paraphrases. The reference validator fetches abstracts from PubMed and checks that the quoted text appears verbatim. This catches:

• AI-generated paraphrases that don't match the actual paper
• Wrong PMIDs (e.g., a PMID that points to an unrelated paper)
• Fabricated citations

When validation fails, either fix the snippet to match the actual abstract or remove the evidence item.

QC Dashboard

Generate a visual dashboard showing compliance metrics across all disorder files:

just gen-dashboard

This creates dashboard/index.html with:

• Summary metrics (files analyzed, average compliance, violations)
• A dedicated capability_metrics.html report with aggregate coverage, evidence, ontology, mechanism, treatment, and compliance metrics
• Slot compliance comparison chart
• Detailed views of the 10 lowest-compliance files (priority curation targets)
• Full table of all files sorted by compliance
• A dedicated not_yet_curated.html report listing referenced MONDO diseases that do not yet have local DisMech pages
• A dedicated priority.html report ranking MONDO curation candidates with action recommendations and coverage summaries

View online: QC Dashboard or locally: open dashboard/index.html

Generate only the MONDO curation priority view:

just gen-priority-dashboard

Generate a local-only dashboard across all uncurated MONDO disease descendants:

just gen-priority-dashboard-all-mondo

This writes large artifacts under tmp/priority-dashboard-all-mondo/ so they remain gitignored and do not get committed to GitHub by accident.

HTML Generation

YAML files are rendered to browsable HTML pages with clickable ontology term links.

uv run python -m dismech.render --all

CX2 Export For NDEx

Disorder pathographs can also be exported as CX2 networks for publication to NDEx.

# Write CX2 JSON to a file
uv run dismech-cx2 kb/disorders/Stargardt_Disease.yaml -o Stargardt_Disease.cx2.json

# Same flow via justfile target
just export-cx2 kb/disorders/Stargardt_Disease.yaml -o Stargardt_Disease.cx2.json

# Export all disorders to a directory, skipping those with no pathograph edges
just export-cx2-all -o /tmp/cx2

# Upload directly to NDEx using env vars
export NDEX_USERNAME=...
export NDEX_PASSWORD=...
uv run dismech-cx2 kb/disorders/Stargardt_Disease.yaml --ndex-upload

# Bulk upload all disorders to the NDEx test server
export NDEX_USERNAME=...
export NDEX_PASSWORD=...
just upload-cx2-test-all

dismech-cx2 uses the existing pathograph graph model, preserves node and edge metadata, and applies a deterministic layout so the uploaded network is immediately viewable in NDEx. Add --dot-layout if Graphviz and pydot are available and you want a Graphviz layout instead of the built-in layered layout. The default NDEx upload visibility is PRIVATE, and the just upload targets default the host to https://test.ndexbio.org. Production releases use the protected, manually triggered Publish NDEx release GitHub Actions workflow described in the operator documentation.

Agentic Curation Guide

This knowledge base is curated with Claude Code — an AI agent that knows the schema, validates ontology terms, and checks evidence against PubMed abstracts. There are two ways to start curating:

Route 1: Clone the repo and use Claude Code CLI

Best for power users who want full control, deep research providers, and iterative curation sessions.

Prerequisites: A Claude Pro/Max subscription and just installed (installation).

# 1. Clone and enter the repo
git clone https://github.com/monarch-initiative/dismech.git
cd dismech

# 2. Launch Claude Code
claude

# 3. Ask it anything — or run a curation skill directly
> Give me a tour of the dismech project
> What disorders are missing ontology terms?
> /curate Parkinson Disease

The /curate command triggers a full curation workflow: deep literature research, YAML generation, ontology term binding, evidence validation, and a ready-to-review PR.

You can also ask the agent open-ended questions, request targeted edits, or run QC:

> Add NCIT treatment terms to the Asthma entry
> Validate references in kb/disorders/Lupus.yaml
> Which files have the lowest compliance scores?

Optional — deep research providers:

For narrative-style deep research, set up an Edison Scientific API key:

export EDISON_API_KEY=your_key_here

For literature-search-style retrieval via Asta, request an API key from Allen AI Asta and export:

export ASTA_API_KEY=your_key_here

Then run:

just research-disorder asta Liver_Cirrhosis

Asta outputs are typically literature packets rather than polished narrative reports: they prioritize relevant papers, summaries, snippets, and metadata such as PMID/DOI.

Route 2: Use Claude Code on the web

Best for quick contributions — no local setup required.

1. Go to claude.ai/code
2. Open the monarch-initiative/dismech repository
3. Set up an environment with the following settings:
   • Allow access to external websites — needed for fetching PubMed abstracts, ClinicalTrials.gov data, and ontology lookups
   • Set EDISON_API_KEY — for deep literature research via Edison Scientific (sign up → Account → Profile → Create new token)
4. Interact with the agent in the web UI — ask questions, request edits, or run /curate
5. When you're done, click the Create pull request button to submit your changes for review

That's it. The web UI handles git branching, commits, and PR creation automatically.

What happens during curation

Regardless of which route you use, the agent follows the same workflow:

1. Research — gathers pathophysiology, phenotypes, treatments, and genetic factors from the literature
2. Draft — generates a structured YAML file with ontology term bindings and PubMed-backed evidence
3. Validate — runs schema validation, ontology term checks, and reference verification (just qc)
4. Submit — commits changes and opens a pull request for human review

For detailed contribution guidelines, see CONTRIBUTING.md.

Evidence recuration reports

Run just jev-audit to assess disease assertion/snippet pairs with Jev and produce CSV reports for recuration. Corpus assessment history and the weekly workflow live in dismech-evals. See the audit guide for inventory, cached reruns, report columns and the weekly GitHub Action.

Schema Documentation

• LinkML Schema Docs
• Schema source: src/dismech/schema/dismech.yaml