PatientOne: Precision Medicine Workflow

Comprehensive precision medicine workflow for Stage IV Ovarian Cancer using all MCP servers.

Overview

Quick references: PatientOne Profile | Platform Overview | DRY_RUN Mode | Cost Analysis

PatientOne demonstrates end-to-end precision medicine analysis integrating:

Clinical data (demographics, CA-125 trends)
Genomic variants (VCF, CNVs, TCGA comparison)
Multiomics (RNA/Protein/Phospho from PDX models)
Spatial transcriptomics (10x Visium, 900 spots)
Imaging (H&E histology, multiplex IF)
Perturbation prediction (GEARS treatment response modeling)

All synthetic data for demonstration purposes.

What makes PatientOne unique: Unlike traditional bioinformatics pipelines that analyze individual data types in isolation, PatientOne shows how AI can seamlessly integrate across all modalities through natural language — replacing weeks of glue code with conversational requests.

Data Integration Flow

flowchart LR
    subgraph Input["5 Data Modalities"]
        CLIN[Clinical<br/>Demographics<br/>CA-125<br/>Treatment Hx]
        GEN[Genomic<br/>VCF Mutations<br/>CNVs<br/>TCGA Compare]
        OMICS[Multi-Omics<br/>RNA-seq<br/>Proteomics<br/>Phospho]
        SPAT[Spatial<br/>10x Visium<br/>900 spots<br/>6 regions]
        IMG[Imaging<br/>H&E<br/>IF markers<br/>Cell seg]
    end

    subgraph Integration["AI Integration Layer"]
        CLAUDE[Claude Desktop<br/>MCP Orchestration]
    end

    subgraph Output["Precision Medicine Output"]
        RES[Resistance<br/>Mechanisms]
        TARGETS[Treatment<br/>Targets]
        TRIALS[Clinical<br/>Trials]
    end

    CLIN --> CLAUDE
    GEN --> CLAUDE
    OMICS --> CLAUDE
    SPAT --> CLAUDE
    IMG --> CLAUDE

    CLAUDE --> RES
    CLAUDE --> TARGETS
    CLAUDE --> TRIALS

    style CLAUDE fill:#fff4e1,stroke:#ff9800,stroke-width:3px
    style RES fill:#e8f5e9,stroke:#4caf50,stroke-width:2px
    style TARGETS fill:#e8f5e9,stroke:#4caf50,stroke-width:2px
    style TRIALS fill:#e8f5e9,stroke:#4caf50,stroke-width:2px

Research Use Only Disclaimer

CRITICAL: This workflow is for RESEARCH and EDUCATIONAL purposes only.

NOT clinically validated — Do not use for actual patient care decisions
NOT FDA-approved — Not a medical device or diagnostic tool
NOT a substitute for clinical judgment — Requires expert review
FOR demonstration — Shows feasibility of AI-orchestrated precision medicine
FOR research — Hypothesis generation and method development

All data is synthetic. Any resemblance to actual patients is coincidental.

See HIPAA compliance

Prerequisites

System Requirements

Python: 3.11+
Claude Desktop: Latest version (Download)
RAM: 16GB recommended
Disk: 50GB free space
OS: macOS, Linux, or Windows with WSL2

Setup Verification

Check Python version:

python3 --version  # Should show 3.11 or higher

Verify Claude Desktop configuration:

cat ~/Library/Application\ Support/Claude/claude_desktop_config.json
# Should show all MCP servers configured

Confirm data files exist:

ls -lh ../../data/patient-data/PAT001-OVC-2025/
# Should show 17 files (~3.2 MB total)

First-Time Setup

If you haven't installed the MCP servers yet:

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

# Install dependencies (5-10 min)
cd manual_testing
./install_dependencies.sh

# Configure Claude Desktop
cp docs/getting-started/desktop-configs/claude_desktop_config.json ~/Library/Application\ Support/Claude/claude_desktop_config.json

# Restart Claude Desktop
# Verify servers loaded (should see all servers in Claude Desktop)

# Test basic server connectivity
./verify_servers.sh

Running Modes: DRY_RUN vs Actual Data

PatientOne can run in two modes:

Mode	Purpose	Data Source	External APIs	Best For
DRY_RUN (default)	Demo & testing	Synthetic responses	None	Quick demo, CI/CD, learning
Actual Data	Real analysis	Your files	May connect	Production, research, clinical

Quick Mode Selection:

DRY_RUN mode (default): No setup needed, works immediately with synthetic data
Actual Data mode: Requires data files and configuration — see Data Modes Guide

Tip: Start with DRY_RUN mode to understand the workflow (5 min), then switch to actual data for real analysis.

Try PatientOne in 5 Minutes

Option 1: Quick Demo (Single Test)

Run Test 1 to see clinical + genomic integration:

Open Claude Desktop
Copy/paste the prompt from: test-prompts/DRY_RUN/test-1-clinical-genomic.md
Expected output:

Patient demographics (Sarah Anderson, 58yo, Stage IV HGSOC)
CA-125 trajectory showing initial response then resistance
Key mutations: TP53 R175H, PIK3CA E545K, PTEN LOH
TCGA subtype: C1 Immunoreactive
BRCA1 germline mutation implications

Duration: 5-10 minutes

Option 2: Complete Analysis (All Tests)

Run all modular tests sequentially. See test-prompts/README.md for the full test index (10 DRY_RUN + 4 SYNTHETIC_DATA tests) and prerequisites.

Cost: ~$1 for all DRY_RUN tests (tokens only). See Cost Analysis for real-data costs.

MCP Server Orchestration

How MCP Servers Contribute

Workflow Stage	MCP Servers Engaged	Tools Used	Output
1. Clinical Retrieval	Epic	`query_patient_records`, `search_diagnoses`	Demographics, CA-125 trends, ICD-10 codes
2. Genomic Analysis	FGbio, TCGA	`validate_fastq`, `query_gene_annotations`, `compare_to_cohort`, `get_mutation_data`	VCF variants, CNV profile, TCGA subtype
3. Multiomics Integration	MultiOmics	`integrate_omics_data`, `calculate_stouffer_meta`, `create_multiomics_heatmap`	Resistance gene signatures, pathway activation
4. Spatial Processing	SpatialTools, DeepCell	`filter_quality`, `split_by_region`, `align_spatial_data`, `segment_cells`	Spatial expression maps, tissue segmentation
5. Histology Analysis	OpenImageData, DeepCell	`fetch_histology_image`, `register_image_to_spatial`, `classify_cell_states`	Cell counts, phenotype distributions

Imaging Modality Reference

Understanding the difference between imaging types is critical for correct analysis:

Image Type	Microscopy Mode	Staining Method	Analysis Server(s)	Use Case
H&E	Brightfield	Chromogenic (Hematoxylin=blue nuclei, Eosin=pink cytoplasm)	OpenImageData	Tissue architecture, morphology, cellularity assessment
IF (single-plex)	Fluorescence	Single fluorescent antibody	OpenImageData + DeepCell	Protein marker quantification (CD8, Ki67, etc.)
MxIF (multiplex)	Fluorescence	Multiple fluorophores (2-7 colors)	OpenImageData + DeepCell	Cell phenotyping, protein co-localization, co-expression analysis
Spatial RNA-seq	N/A (sequencing)	Tabular CSV data (no images)	SpatialTools only	Gene expression patterns across tissue

Key Differences:

H&E: Brightfield microscopy with colored (chromogenic) stains — NOT fluorescence
IF/MxIF: Fluorescence microscopy with fluorescent antibodies — requires different analysis
Spatial data: No images, just CSV files with coordinates and expression values

What is MxIF? MxIF (Multiplexed Immunofluorescence) enables imaging of multiple protein markers (2-7+) on a single tissue section through repeated rounds of staining, imaging, dye inactivation, and background subtraction.

The Patient One workflow uses the open-source DeepCell-TF library (https://github.com/vanvalenlab/deepcell-tf) for AI-based cell segmentation in MxIF images.

When to use DeepCell in PatientOne Workflow:

MxIF/IF images requiring cell segmentation and quantification (CD8, Ki67, TP53/Ki67/DAPI multiplex)
NOT for H&E images (used for visual morphology assessment only in this workflow)
NOT for tabular spatial data (CSV files) — use SpatialTools instead

Test Descriptions

TEST_1: Clinical + Genomic Analysis

Servers: Epic, FGbio, TCGA Files: 3 (patient_demographics.json, lab_results.json, somatic_variants.vcf)

What it does:

Retrieves patient demographics and treatment history
Analyzes CA-125 tumor marker trajectory
Identifies somatic mutations and CNVs
Compares to TCGA ovarian cancer cohort
Determines molecular subtype

Key Findings:

Platinum-resistant disease (8-month recurrence)
TP53/PIK3CA/PTEN driver mutations
C1 immunoreactive subtype
BRCA1 germline mutation — HRD-positive

TEST_2: Multi-Omics Resistance Analysis

Servers: MultiOmics Files: 4 (pdx_rna_seq.csv, pdx_proteomics.csv, pdx_phosphoproteomics.csv, sample_metadata.csv)

What it does:

Integrates RNA/Protein/Phospho data from 15 PDX samples
Compares resistant vs sensitive samples (7 vs 8)
Performs Stouffer's meta-analysis with FDR correction
Identifies dysregulated pathways

Key Findings:

PI3K/AKT/mTOR pathway activation in resistant samples
PIK3CA, AKT1, mTOR, RPS6KB1 upregulated (p < 0.001)
Drug efflux: ABCB1 (MDR1) overexpression
Anti-apoptotic: BCL2L1 upregulation

TEST_3: Spatial Transcriptomics

Servers: SpatialTools Files: 3 (visium_gene_expression.csv, visium_spatial_coordinates.csv, visium_region_annotations.csv)

What it does:

Processes 10x Visium spatial RNA-seq tabular data (900 spots, 31 genes)
Identifies 6 tissue regions (tumor_core, proliferative, interface, stroma, etc.)
Maps spatial expression patterns from CSV files
Quantifies immune cell distribution
Generates visualizations: Spatial heatmaps, gene expression matrices, autocorrelation plots

Note: Uses only tabular CSV data, not images. DeepCell is NOT needed for this test.

Key Findings:

Immune exclusion phenotype (CD8+ low in tumor core)
High proliferation in tumor_proliferative region (Ki67+, PCNA+)
Thick stromal barrier separating immune cells from tumor
Spatial heterogeneity in resistance markers

TEST_4: Histology & Imaging

Servers: OpenImageData (H&E + MxIF), DeepCell (MxIF segmentation only) Files: 4 TIFF images used in test (7 available: H&E brightfield, IF single-markers, multiplex IF)

Test Files:

PAT001_tumor_HE_20x.tiff - H&E brightfield (openimagedata ONLY)
PAT001_tumor_IF_CD8.tiff - IF fluorescence (openimagedata + deepcell)
PAT001_tumor_IF_KI67.tiff - IF fluorescence (openimagedata + deepcell)
PAT001_tumor_multiplex_IF_TP53_KI67_DAPI.tiff - MxIF 3-channel (openimagedata + deepcell)

Key Findings:

Tumor cellularity: 70-80%
Ki67 proliferation index: 45-55% (HIGH)
CD8+ T cell density: 5-15 cells/mm2 (LOW, mostly peripheral)
CD3+ overall: 30-50 cells/mm2 (moderate T cells, but not cytotoxic)

TEST_5: Integration & Recommendations

Servers: All servers (synthesis) Files: None (builds on previous tests)

What it does:

Synthesizes findings across all 5 modalities
Integrates molecular, spatial, and clinical insights
Identifies actionable treatment targets
Generates precision medicine recommendations

Key Recommendations:

Primary: PI3K inhibitor (Alpelisib) targeting PIK3CA E545K mutation
Secondary: Anti-PD-1 immunotherapy to overcome immune exclusion
Tertiary: PARP inhibitor consideration (BRCA1 mutation, but PIK3CA pathway may limit efficacy)
Clinical trial: NCT03602859 (alpelisib + paclitaxel in ovarian cancer)

TEST_6: Clinician-in-the-Loop (CitL) Review

Servers: patient-report Files: draft_report.json (generated from TEST_1-5)

What it does:

Generates draft report with quality gates (4 automated checks)
Clinician validates 10 molecular findings (CONFIRM/UNCERTAIN/INCORRECT)
Assesses NCCN + institutional guideline compliance
Reviews quality flags, makes decision: APPROVE / REVISE / REJECT
Creates HIPAA-compliant audit trail with digital signature (10-year retention)

See citl-quick-test.md for the hands-on CitL test guide.

TEST_7-9: End-to-End Tests

Test 7: Single-prompt E2E covering 6 servers
Test 8: Test 7 + PubMed, ClinicalTrials.gov, bioRxiv connectors
Test 9: Focused E2E with Seqera nf-core pipeline discovery

Data Assets

All synthetic patient data located in: /data/patient-data/PAT001-OVC-2025/

File Inventory (18 files, ~3.2 MB total)

Modality	Files	Size	Content Description
Clinical	2	10.7 KB	`patient_demographics.json`, `lab_results.json`
Genomic	1	2.3 KB	`somatic_variants.vcf` (12 key variants)
Multiomics	4	505 KB	`pdx_rna_seq.csv` (1K genes), `pdx_proteomics.csv` (500), `pdx_phosphoproteomics.csv` (300), `sample_metadata.csv`
Spatial	4	315 KB	`visium_gene_expression.csv` (900 spots x 31 genes), `visium_spatial_coordinates.csv`, `visium_region_annotations.csv`
Imaging	7	2.2 MB	H&E + IF (DAPI, CD3, CD8, Ki67, PanCK) + multiplex
TOTAL	18	~3.2 MB	Complete precision medicine dataset

Key Findings from PatientOne Analysis

1. Molecular Resistance Mechanisms

From Multiomics Integration (MCP-MultiOmics):

PI3K/AKT/mTOR pathway activation in carboplatin-resistant PDX samples
Upregulated genes/proteins: PIK3CA, AKT1, mTOR, RPS6KB1 (Stouffer's combined p < 0.001)
Drug efflux: ABCB1 (MDR1) overexpression (log2FC = 2.8, FDR < 0.01)
Anti-apoptotic: BCL2L1 upregulation

From Genomic Analysis (MCP-FGbio + MCP-TCGA):

PIK3CA E545K activating mutation (allele frequency 38%)
TP53 R175H hotspot mutation (loss of function)
PTEN loss of heterozygosity (tumor suppressor inactivation)
TCGA subtype: C1 Immunoreactive (immune infiltration expected, but...)

2. Tumor Microenvironment

From Spatial Transcriptomics (MCP-SpatialTools):

6 distinct spatial regions identified
Immune exclusion phenotype: CD8+ T cells enriched at tumor periphery, sparse in core
Proliferation gradient: Ki67/PCNA high in tumor_proliferative region
Stroma barrier: Thick stromal band separating immune cells from tumor

From Histology Imaging (MCP-OpenImageData + MCP-DeepCell):

Tumor cellularity: 70-80%
Ki67 proliferation index: 45-55% (high)
CD8+ T cell density: 5-15 cells/mm2 (LOW, mostly peripheral)
CD3+ overall: 30-50 cells/mm2 (moderate T cell presence, but not cytotoxic)

3. Clinical-Molecular Integration

From Clinical Data (MCP-Epic):

CA-125 response pattern: Initial deep response (1456 -> 22 U/mL) followed by resistance (-> 389 U/mL)
BRCA1 germline mutation: HRD-positive — PARP inhibitor candidate, BUT PIK3CA pathway may confer resistance
Platinum-free interval: 8 months — platinum-resistant category

4. MTB-Ready Treatment Recommendations

All recommendations are presented for Molecular Tumor Board review using AMP/ASCO/CAP evidence tiers. Clinician validation is required before clinical use.

Primary Target: PI3K/AKT Pathway (Tier 1 — FDA-approved biomarker)

Consider: Alpelisib (PIK3CA inhibitor) given E545K mutation
Clinical trial: NCT03602859 (alpelisib + paclitaxel in ovarian cancer)

Secondary Target: Immune Checkpoint (Tier 2 — evidence from other tumor types)

Consider: Anti-PD-1 (pembrolizumab, nivolumab) to overcome immune exclusion

PARP Inhibitor Re-consideration (Tier 1 — FDA-approved for BRCA1+ ovarian)

Given BRCA1 mutation + HRD score 42, PARP inhibitor (olaparib, niraparib) remains option
Caution: PIK3CA pathway activation may limit efficacy

Bias Audit Summary

Audit Date: 2026-01-12 | Risk Level: MEDIUM (acceptable with mitigations)

The PatientOne workflow has undergone comprehensive bias auditing. Key findings:

5 checks passed: Insurance status, geographic location, race/ethnicity coding, spatial algorithms, PDX models
3 biases detected and mitigated: Euro-centric BRCA variant databases (MEDIUM), GTEx reference ranges 85% European (MEDIUM), generic cell type references (LOW)
Fairness metrics: All within acceptable thresholds (<10% disparity)
No proxy features used (geographic, socioeconomic data excluded)

Full details: Ethics & Bias Framework

Troubleshooting

Issue: "MCP servers not found"

Cause: Claude Desktop config not loaded or servers not installed

Fix:

# Verify config exists
cat ~/Library/Application\ Support/Claude/claude_desktop_config.json

# If missing, copy template
cp docs/getting-started/desktop-configs/claude_desktop_config.json ~/Library/Application\ Support/Claude/

# Restart Claude Desktop

Issue: "Cannot find data files"

Cause: Incorrect file paths or data not present

Fix:

# Verify data exists
ls -lh data/patient-data/PAT001-OVC-2025/
# Should show 17 files

# Check absolute path in prompt matches your system
pwd  # Note current directory
# Update file paths in prompts to match your installation

Issue: "Server returned error / DRY_RUN warnings"

Cause: Servers are in DRY_RUN mode (expected behavior for testing)

Explanation:

All MCP servers are configured with DRY_RUN=true by default
This prevents actual external API calls while demonstrating tool orchestration
Servers return realistic synthetic responses

To use your own data:

See the Data Modes Guide for complete instructions

Issue: "Context limit exceeded"

Cause: Trying to run all tests in single prompt

Fix:

Run tests individually (Test 1 through Test 5)
Each test designed to fit within Claude Desktop context limits
Do NOT combine multiple tests in one prompt
Clear conversation history between tests if needed

Issue: "Missing Python packages"

Cause: Virtual environments not set up correctly

Fix:

cd manual_testing
./install_dependencies.sh

# Verify each server's venv
for server in ../servers/mcp-*/; do
    echo "Checking $server"
    $server/venv/bin/python --version
done

Expected Outputs

For Each Test

Claude Desktop will generate:

Data Summary: Key statistics from loaded files
Tool Execution: MCP server calls with results
Analysis: Interpretation and synthesis
Findings: Bullet-point key discoveries

Final Integrated Output (After TEST_5)

Comprehensive report including:

Executive Summary: Patient profile and precision medicine strategy
Molecular Profile: Genomic alterations, pathway dysregulation
Microenvironment: Spatial distribution, immune landscape
Resistance Mechanisms: Multi-omics signatures
Treatment Plan: Evidence-based recommendations with rationale

Support

Issues: https://github.com/lynnlangit/precision-medicine-mcp/issues Documentation: https://github.com/lynnlangit/precision-medicine-mcp

Last Updated: 2026-05-12 Testing Status: 10 DRY_RUN tests + 4 SYNTHETIC_DATA tests validated Data: 100% synthetic for demonstration purposes

FilesExpand file tree

README.md

Latest commit

History

README.md

File metadata and controls

PatientOne: Precision Medicine Workflow

Overview

Data Integration Flow

Research Use Only Disclaimer

Prerequisites

System Requirements

Setup Verification

First-Time Setup

Running Modes: DRY_RUN vs Actual Data

Try PatientOne in 5 Minutes

Option 1: Quick Demo (Single Test)

Option 2: Complete Analysis (All Tests)

MCP Server Orchestration

How MCP Servers Contribute

Imaging Modality Reference

Test Descriptions

TEST_1: Clinical + Genomic Analysis

TEST_2: Multi-Omics Resistance Analysis

TEST_3: Spatial Transcriptomics

TEST_4: Histology & Imaging

TEST_5: Integration & Recommendations

TEST_6: Clinician-in-the-Loop (CitL) Review

TEST_7-9: End-to-End Tests

Data Assets

File Inventory (18 files, ~3.2 MB total)

Key Findings from PatientOne Analysis

1. Molecular Resistance Mechanisms

2. Tumor Microenvironment

3. Clinical-Molecular Integration

4. MTB-Ready Treatment Recommendations

Bias Audit Summary

Troubleshooting

Issue: "MCP servers not found"

Issue: "Cannot find data files"

Issue: "Server returned error / DRY_RUN warnings"

Issue: "Context limit exceeded"

Issue: "Missing Python packages"

Expected Outputs

For Each Test

Final Integrated Output (After TEST_5)

Related Documentation

Support