Precision Medicine MCP Platform

Dedicated to PatientOne -- a dear friend who passed from High-Grade Serous Ovarian Carcinoma in 2025.

The Problem

Standard HGSOC workup (BRCA1/2, HRD panel, CT imaging) generates no immunotherapy hypotheses. Manual multi-modal analysis across genomics, spatial transcriptomics, imaging, and clinical data takes an estimated 40 hours and $6,000-9,000 per patient -- making integrated analysis clinically impractical.

The Platform

A 19-server MCP architecture orchestrated by AI (Claude + Gemini) executes a 5-stage pipeline:

flowchart LR
    A["1 Data<br/>Acquisition"] --> B["2 Spatial<br/>Deconvolution"]
    B --> C["3 Target<br/>Profiling"]
    C --> D["4 Causal<br/>Inference"]
    D --> E["5 Report"]

    subgraph servers [" "]
        direction TB
        S1["EHR · GEO · TCGA"]
        S2["Spatial · DeepCell · CIBERSORTx"]
        S3["OpenTargets · Neoantigen"]
        S4["Perturbation · Quantum"]
        S5["Patient Report"]
    end

    A --- S1
    B --- S2
    C --- S3
    D --- S4
    E --- S5

    AI(["AI Orchestrator<br/>Claude + Gemini"]) -.-> A
    AI -.-> B
    AI -.-> C
    AI -.-> D
    AI -.-> E

Architecture at a glance

                  +--------------------------------------+
                  |           CLIENT LAYER               |
                  |  Claude Desktop / Hospital EHR       |
                  |  Adapter / Research Notebook         |
                  +----------------+-----------------+
                                   |
                         MCP (FastMCP >= 2.13)
                                   |
   +---------------------------------------------------------------+
   |                                                               |
   |  DATA ACQUISITION      ANALYSIS & INFERENCE      REPORTING   |
   |                                                               |
   |  mockepic              spatialtools    (16)      patient-     |
   |  epic                  multiomics     (10)       report (5)   |
   |  geodownload           perturbation    (8)                    |
   |  mocktcga              quantum-fidelity(6)                    |
   |  genomic-results       opentargets     (6)                    |
   |  fgbio                 neoantigen      (6)                    |
   |                        cibersortx      (5)                    |
   |  [7 servers]           openimagedata   (5)       [1 server]   |
   |                        deepcell        (3)                    |
   |                        cell-classify   (3)                    |
   |                        cardiometabolic (5)                    |
   |                                                               |
   |                        [11 servers]                           |
   +---------------------------------------------------------------+
                     19 custom servers, 104 tools

All tools accessible via natural language. Every AI result requires clinician APPROVE/REVISE/REJECT. HIPAA-compliant. See Server Registry.

The Results

The platform surfaces clinically actionable findings that standard workup cannot reach — validated across three independent use cases:

The same 19-server architecture runs all three. No disease-specific code changes between use cases.

Validated results — PAT001 (HGSOC)

Try It

# Clone and explore
git clone https://github.com/lynnlangit/precision-medicine-mcp.git
cd precision-medicine-mcp

# Run tests for any server (DRY_RUN mode, no external deps needed)
cd servers/mcp-multiomics && uv run pytest -v

# Or use Claude Code to explore interactively
claude

All servers default to DRY_RUN mode (mock responses, no API keys needed) for quick validation. Set *_DRY_RUN=false to use synthetic patient data for end-to-end testing.

Learn More

Video: 5-minute demo | Paper: Why MCP for Healthcare | External connectors: Setup guide

Known limitations

• DRY_RUN mode returns synthetic data — not for clinical decisions. Set *_DRY_RUN=false with real data for validated results.
• GEARS model trained on synthetic GSE184880 subset — retrain on real TCGA data before clinical use.
• Quantum server falls back to CPU on non-CUDA hardware (Apple Silicon, cloud VMs without GPU). Results are identical; training is slower.

Apache 2.0 | Python 3.11+ | FastMCP >= 2.13 | uv for package management