Life Sciences MCP

FastMCP wrappers for essential life sciences APIs and datasets. A microservices-based approach to accelerate scientific research by providing MCP server access to biological databases, gene nomenclature services, protein interaction networks, and drug-target databases.

Vision

Enable AI agents to seamlessly query the world's most important life sciences databases through the Model Context Protocol (MCP), accelerating drug discovery, drug repurposing, and biomedical research.

Current Status: 12 MCP servers operational covering genes (HGNC, Ensembl, Entrez), proteins (UniProt, STRING, BioGRID), compounds (ChEMBL, PubChem), pharmacology (IUPHAR/GtoPdb), targets (Open Targets), pathways (WikiPathways), and clinical trials (ClinicalTrials.gov).

The Modern Drug Discovery Stack (2025)

The current best practice in computational drug discovery is an integrative approach using programmatic APIs across multiple data layers:

┌─────────────────────────────────────────────────────────────┐
│                    DRUG REPURPOSING STACK                   │
├─────────────────────────────────────────────────────────────┤
│  Layer 1: Chemical/Drug                                      │
│  ChEMBL → PubChem → DrugBank                                │
├─────────────────────────────────────────────────────────────┤
│  Layer 2: Target/Protein                                     │
│  UniProt → STRING → IUPHAR/GtoPdb → STITCH                  │
├─────────────────────────────────────────────────────────────┤
│  Layer 3: Gene/Genomics                                      │
│  HGNC → Ensembl → NCBI/Entrez                               │
├─────────────────────────────────────────────────────────────┤
│  Layer 4: Disease/Phenotype                                  │
│  OMIM → Orphanet → Open Targets                             │
├─────────────────────────────────────────────────────────────┤
│  Layer 5: Knowledge Integration                              │
│  Open Targets Platform (aggregates all above)               │
└─────────────────────────────────────────────────────────────┘

Key Insight: Open Targets as a "Meta-API"

Open Targets Platform is emerging as the single most valuable API for drug discovery because it:

• Aggregates ChEMBL, UniProt, Ensembl, and disease databases into a unified interface
• Provides a modern GraphQL API for flexible querying
• Maps targets to diseases with genetic and experimental evidence
• Enables drug repurposing by connecting approved drugs to new indications
• Already does the integration work you'd otherwise do manually

This makes Open Targets an excellent starting point for AI-driven drug discovery workflows.

Planned MCP Servers

Tier 0: Strategic Priority (Drug Discovery Core)

Tier 1: Foundation (Gene/Protein Layer)

Tier 2: Pharmacology & Interactions

Tier 3: Pathways & Clinical Trials

Tier 4: Genomics & Identifiers

Summary

Completion Status:

• ✅ 12 servers operational - HGNC, UniProt, ChEMBL, Open Targets, STRING, BioGRID, IUPHAR/GtoPdb, PubChem, Ensembl, Entrez, WikiPathways, ClinicalTrials.gov
• ⛔ 1 server blocked - DrugBank (requires commercial API key)
• 🔜 4 servers planned - STITCH, KEGG, OMIM, Orphanet

Test Coverage:

• Total integration tests: 500+ passing
• Total unit tests: 100+ passing
• Coverage: All 12 operational servers have comprehensive test suites
• Gateway server: 34+ MCP tools from 12 databases

Agentic Architecture (Team of Tools)

We are building a Team of Agents where each specialized tool plays a role in the scientific reasoning loop.

graph TD
    subgraph "Reasoning Layer (Strategies)"
        Literature[Literature Agent]
        Validator[Validation Agent]
    end

    subgraph "Structured Truth Layer (Life Sciences MCP - 12 Operational)"
        HGNC["HGNC ✅<br/>(Gene Identity)"]
        UniProt["UniProt ✅<br/>(Protein Function)"]
        OpenTargets["Open Targets ✅<br/>(Disease)"]
        ChEMBL["ChEMBL ✅<br/>(Compounds)"]
        STRING["STRING ✅<br/>(Interactions)"]
        WikiPathways["WikiPathways ✅<br/>(Pathways)"]
        ClinicalTrials["ClinicalTrials.gov ✅<br/>(Trials)"]
    end

    subgraph "Unstructured Knowledge Layer"
        PubMed["(PubMed/BioRxiv)"]
        FullText["(PDFs/Figures)"]
    end

    Literature -->|Reads| PubMed
    Literature -->|Extracts Claims| FullText
    Literature -->|Queries| Validator

    Validator -->|Grounds Gene Terms| HGNC
    Validator -->|Validates Proteins| UniProt
    Validator -->|Checks Disease Evidence| OpenTargets
    Validator -->|Finds Compounds| ChEMBL
    Validator -->|Discovers Interactions| STRING
    Validator -->|Analyzes Pathways| WikiPathways
    Validator -->|Finds Clinical Trials| ClinicalTrials

    style Validator fill:#e1f5fe,stroke:#01579b
    style Literature fill:#f3e5f5,stroke:#4a148c

The "Structured Truth Layer"

This repository (lifesciences-research) acts as the Grounding Engine. When a Literature Agent reads a paper and claims "Drug X targets Protein Y," it uses this MCP to:

1. Resolve "Protein Y" to a precise UniProt ID (resolving synonyms).
2. Validate if "Drug X" actually binds to "Protein Y" in ChEMBL/OpenTargets.
3. Harden the unstructured text into a structured Knowledge Graph.

Quick Start

# Install dependencies
uv sync --extra dev

# =============================================================================
# Run Individual MCP Servers
# =============================================================================

# Tier 0: Drug Discovery Core
uv run fastmcp run src/lifesciences_mcp/servers/chembl.py        # ChEMBL compounds & bioactivity (✅ 112 tests)
uv run fastmcp run src/lifesciences_mcp/servers/opentargets.py   # Target-disease associations (✅ 9 tests)
uv run fastmcp run src/lifesciences_mcp/servers/drugbank.py      # Drug interactions (⛔ requires API key)

# Tier 1: Gene/Protein Foundation
uv run fastmcp run src/lifesciences_mcp/servers/hgnc.py          # Gene nomenclature (✅ 21 tests)
uv run fastmcp run src/lifesciences_mcp/servers/uniprot.py       # Protein search & lookup (✅ 29 tests)
uv run fastmcp run src/lifesciences_mcp/servers/string.py        # Protein-protein interactions (✅ 12 tests)
uv run fastmcp run src/lifesciences_mcp/servers/biogrid.py       # Genetic/protein interactions (✅ 11 tests)

# Tier 2: Pharmacology & Interactions
uv run fastmcp run src/lifesciences_mcp/servers/iuphar.py        # Pharmacological targets (✅ 59 tests)
uv run fastmcp run src/lifesciences_mcp/servers/pubchem.py       # Chemical structures (✅ 100 tests)

# Tier 3: Pathways & Clinical Trials
uv run fastmcp run src/lifesciences_mcp/servers/wikipathways.py  # Biological pathways (✅ 4 tools)
uv run fastmcp run src/lifesciences_mcp/servers/clinicaltrials.py # Clinical trials (✅ 3 tools, 13 tests)

# Tier 4: Genomics & Identifiers
uv run fastmcp run src/lifesciences_mcp/servers/ensembl.py       # Genomic annotations (✅ 86 tests)
uv run fastmcp run src/lifesciences_mcp/servers/entrez.py        # NCBI gene database (✅ 58 tests)

# =============================================================================
# Run Tests
# =============================================================================

# Run all tests
uv run pytest tests/ -v

# Run integration tests only
uv run pytest -m integration -v

# Test specific server
uv run pytest tests/integration/test_hgnc_api.py -v -m integration         # 7 tests ✅
uv run pytest tests/integration/test_uniprot_api.py -v -m integration      # 12 tests ✅
uv run pytest tests/integration/test_chembl_api.py -v -m integration       # 50+ tests ✅
uv run pytest tests/integration/test_opentargets_api.py -v -m integration  # 9 tests ✅
uv run pytest tests/integration/test_drugbank_api.py -v -m integration     # 7 tests (⛔ skipped without API key)
uv run pytest tests/integration/test_string_api.py -v -m integration       # 12 tests ✅
uv run pytest tests/integration/test_biogrid_api.py -v -m integration      # 11 tests ✅
uv run pytest tests/integration/test_iuphar_api.py -v -m integration       # 48 tests ✅
uv run pytest tests/integration/test_pubchem_api.py -v -m integration      # 19 tests ✅
uv run pytest tests/integration/test_ensembl_api.py -v -m integration      # 24 tests ✅
uv run pytest tests/integration/test_entrez_api.py -v -m integration       # 20 tests ✅
uv run pytest tests/integration/test_wikipathways_api.py -v -m integration # Integration tests ✅
uv run pytest tests/unit/test_clinicaltrials_client.py -v                  # 13 unit tests ✅

Example Usage

HGNC Server (Gene Nomenclature)

from lifesciences_mcp.clients import HGNCClient

async with HGNCClient() as client:
    # Fuzzy search for genes
    results = await client.search_genes("BRCA")
    # Returns: PaginationEnvelope[SearchCandidate]

    # Strict lookup by HGNC CURIE
    gene = await client.get_gene("HGNC:1100")  # BRCA1
    # Returns: Gene with cross_references to UniProt, Ensembl, OMIM, etc.

UniProt Server (Protein Search & Lookup)

from lifesciences_mcp.clients import UniProtClient

async with UniProtClient() as client:
    # Phase 1: Fuzzy search for proteins
    results = await client.search_proteins("p53 tumor suppressor", page_size=10)
    # Returns: PaginationEnvelope[ProteinSearchCandidate]

    # Get top candidate
    top_candidate = results.items[0]
    print(f"{top_candidate.id}: {top_candidate.name} ({top_candidate.organism})")
    # Output: UniProtKB:P04637: Cellular tumor antigen p53 (Homo sapiens)

    # Phase 2: Strict lookup with complete protein record
    protein = await client.get_protein(top_candidate.id)
    # Returns: Protein with cross_references to HGNC, Ensembl, RefSeq, PDB, OMIM, etc.
    print(f"Function: {protein.function[:100]}...")
    print(f"Cross-refs: HGNC:{protein.cross_references.hgnc}, Ensembl:{protein.cross_references.ensembl_transcript}")

PubChem Server (Chemical Compound Search & Lookup)

from lifesciences_mcp.clients import PubChemClient

async with PubChemClient() as client:
    # Phase 1: Fuzzy search for compounds
    results = await client.search_compounds("aspirin", page_size=10)
    # Returns: PaginationEnvelope[PubChemSearchCandidate]

    # Get top candidate
    top_candidate = results.items[0]
    print(f"{top_candidate.id}: {top_candidate.name} ({top_candidate.molecular_formula})")
    # Output: PubChem:CID2244: Aspirin (C9H8O4)

    # Phase 2: Strict lookup with complete compound record
    compound = await client.get_compound(top_candidate.id)
    # Returns: PubChemCompound with SMILES, InChI, cross_references
    print(f"SMILES: {compound.canonical_smiles}")
    print(f"InChI: {compound.inchi[:50]}...")
    print(f"Cross-refs: ChEMBL:{compound.cross_references.get('chembl')}, DrugBank:{compound.cross_references.get('drugbank')}")

    # Token-efficient slim mode
    compound_slim = await client.get_compound("PubChem:CID2244", slim=True)
    # Returns only: id, name, molecular_formula (~20 tokens vs ~115-300)

WikiPathways Server (Biological Pathways)

from lifesciences_mcp.clients import WikiPathwaysClient

async with WikiPathwaysClient() as client:
    # Phase 1: Search for pathways
    results = await client.search_pathways("EGFR signaling", species="Homo sapiens")
    print(f"Found {len(results.items)} pathways")

    # Get pathway details
    pathway = await client.get_pathway(results.items[0].id)
    print(f"Pathway: {pathway.name}")
    print(f"Components: {pathway.component_counts.genes} genes, {pathway.component_counts.metabolites} metabolites")

    # Find pathways for a specific gene
    gene_pathways = await client.get_pathways_for_gene("EGFR", species="Homo sapiens")
    print(f"EGFR appears in {len(gene_pathways.items)} pathways")

    # Get pathway components (graph structure)
    components = await client.get_pathway_components(pathway.id)
    print(f"Data nodes: {len(components.data_nodes)}")
    print(f"Interactions: {len(components.interactions)}")

ClinicalTrials.gov Server (Clinical Trials)

from lifesciences_mcp.clients import ClinicalTrialsClient

async with ClinicalTrialsClient() as client:
    # Phase 1: Search clinical trials
    results = await client.search_trials(
        query="cancer immunotherapy",
        condition="lung cancer",
        phase="PHASE3",
        status="RECRUITING"
    )

    # Get trial details
    trial = await client.get_trial(results.items[0].id)
    print(f"Trial: {trial.title}")
    print(f"Phase: {trial.phase}")
    print(f"Status: {trial.status}")
    print(f"Enrollment: {trial.enrollment}")

    # Get trial locations
    locations = await client.get_trial_locations(trial.id)
    print(f"Trial sites: {len(locations)}")
    for loc in locations[:3]:
        print(f"  - {loc.facility_name}, {loc.city}, {loc.state}")

MCP Tool Interface

All servers expose their functionality as MCP tools:

# HGNC tools
await mcp.call_tool("search_genes", {"query": "BRCA", "page_size": 5})
await mcp.call_tool("get_gene", {"hgnc_id": "HGNC:1100"})

# UniProt tools
await mcp.call_tool("search_proteins", {"query": "insulin", "page_size": 10})
await mcp.call_tool("get_protein", {"uniprot_id": "UniProtKB:P04637", "slim": False})

# PubChem tools
await mcp.call_tool("search_compounds", {"query": "aspirin", "page_size": 10})
await mcp.call_tool("get_compound", {"pubchem_id": "PubChem:CID2244", "slim": False})

# WikiPathways tools
await mcp.call_tool("search_pathways", {"query": "EGFR signaling", "species": "Homo sapiens"})
await mcp.call_tool("get_pathway", {"pathway_id": "WP:WP4868"})
await mcp.call_tool("get_pathways_for_gene", {"gene_symbol": "EGFR", "species": "Homo sapiens"})
await mcp.call_tool("get_pathway_components", {"pathway_id": "WP:WP4868"})

# ClinicalTrials.gov tools
await mcp.call_tool("search_trials", {
    "query": "cancer immunotherapy",
    "condition": "lung cancer",
    "phase": "PHASE3",
    "status": "RECRUITING"
})
await mcp.call_tool("get_trial", {"nct_id": "NCT:00461032"})
await mcp.call_tool("get_trial_locations", {"nct_id": "NCT:00461032"})

Architecture

New to this project? Read Platform Engineering for AI-Augmented Development first to understand our approach to AI-assisted development.

For binding technical specifications, see ADR-001 v1.2.

Design Principles

• Microservices: One MCP server per API/database for modularity
• Async-first: All tools use async/await for network calls
• Pydantic models: Strong typing for API responses
• Caching: Redis or in-memory caching for frequent lookups
• Rate limiting: Respect upstream API rate limits
• identifier.org URIs: Standard URI format for biological identifiers

Data Standards

Following patterns from nsclc-pathways:

• identifier.org URIs: http://identifiers.org/hgnc/1100 for BRCA1
• JSON-LD: Linked data format for semantic interoperability
• GraphML: Network export format for visualization tools

Configuration

Environment Variables

Most life sciences APIs are public and don't require authentication. However, two servers require API keys:

# Optional - BioGRID (free registration)
BIOGRID_API_KEY=your-key-here  # Get from https://thebiogrid.org/

# Optional NCBI (free registration)
NCBI_API_KEY=your-key-here # Get from https://account.ncbi.nlm.nih.gov/settings/

# Optional - DrugBank (commercial license required)
DRUGBANK_API_KEY=your-key-here  # Get from https://go.drugbank.com/

Note:

• BioGRID: Free API key available with registration at https://thebiogrid.org/
• NCBI: Free API key available with registration at https://account.ncbi.nlm.nih.gov/settings/
• DrugBank: Requires commercial license. DrugBank server is excluded from the gateway server and requires manual setup.
• All other 10 servers work without authentication

Development

# Install with dev dependencies
uv sync --extra dev

# Run tests
uv run pytest tests/ -v

# Lint and format
uv run ruff check --fix . && uv run ruff format .

# Type checking
uv run pyright

Testing with FastMCP

import pytest
from fastmcp import Client

@pytest.fixture
async def client():
    from lifesciences_mcp.hgnc import mcp
    async with Client(mcp) as client:
        yield client

async def test_get_gene_info(client):
    result = await client.call_tool("get_gene_info", {"symbol": "BRCA1"})
    assert result["hgnc_id"] == "HGNC:1100"

Quality Assurance

We maintain a comprehensive list of Test Scenarios covering data model validation, error handling, and edge cases.

Example: Search Candidate Validation

See docs/test_scenarios.md for the full list.

🧠 Intelligence Included: Pre-Configured Agent Skills

This repository includes a .claude directory containing optimized system prompts and skill definitions used to generate our research outputs.

• Clinical Trials Skill: Specialized instructions for navigating ClinicalTrials.gov, filtering by phase/status, and extracting inclusion criteria.
• Genomics Skill: Best practices for resolving gene symbols to Ensembl/HGNC IDs before querying.
• Graph Builder Skill: Instructions for constructing Neo4j knowledge graphs from unstructured literature.

🔬 Research & Validation

We use these tools to perform real-world analysis. All outputs are validated for factual accuracy.

References

Upstream APIs

• Open Targets Platform Documentation
• ChEMBL Web Services
• HGNC REST API
• UniProt REST API
• STRING API

Research

• Data-driven Drug Repurposing Strategies (2025)
• AI in Drug Repurposing (2025)
• Open Targets Drug Index

Related Projects and Showcases

Showcases:

• NSCLC Drug Repurposing Showcase (docs/showcases/nsclc-drug-repurposing/) - Complete end-to-end workflow demonstrating WikiPathways and ClinicalTrials.gov integration for non-small cell lung cancer research

Related Projects:

• nsclc-pathways - NSCLC signaling pathway analysis (original inspiration for WikiPathways integration)
• kg_rememberall - Knowledge graph construction from text
• FastMCP Documentation

Architecture Documentation:

• Architecture - Complete architecture analysis with 13,505 lines of code across 56 Python modules
• ADR-001 v1.2 - Binding architecture specification (Fuzzy-to-Fact protocol)
• Component Inventory - Detailed component reference
• API Reference - Usage guide with examples

License

MIT

Project Tracking

• Linear Project: Life Sciences MCP Server
• Discovery Issue: AGE-65