Reactome Curator Tool - LLM-Powered Gene Pathway Annotation

An intelligent gene pathway annotation system that leverages Large Language Models and multi-agent frameworks to assist Reactome curators in annotating genes and their pathway involvement based on scientific literature.

🚀 Features

• Intelligent Literature Mining: Automated PubMed abstract retrieval and analysis
• Multi-Agent Architecture: CrewAI-powered specialist agents for different annotation tasks
• Reactome Integration: Direct integration with Reactome Neo4j database and data models
• Evidence-Based Annotation: Literature-supported pathway predictions with confidence scoring
• Full-Text Analysis: PDF paper processing for deeper information extraction
• REST API: Complete API for programmatic access and integration
• Interactive Chat Interface: Chainlit-powered conversational interface

🏗️ Architecture

Traditional Single-Agent Pipeline

• GenePathwayAnnotator: Core annotation engine with PubMed integration
• Literature Processing: Automated abstract retrieval, embedding, and similarity scoring
• Pathway Enrichment: Statistical analysis of protein-protein interactions
• Reactome Modeling: Direct pathway instance generation

New Multi-Agent Framework (CrewAI)

• ReactomeCurator: Converts structured data to Reactome instances
• LiteratureExtractor: Processes papers and extracts molecular information
• Reviewer: Domain expert validation and quality assessment
• QualityChecker: Technical compliance and consistency validation

📦 Installation

Prerequisites

• Python 3.10+
• Neo4j database (Reactome instance)
• MongoDB (for PubMed caching)
• OpenAI API access

Dependencies

# Install base dependencies
pip install -r requirements.txt

# Install CrewAI for multi-agent framework
pip install crewai crewai-tools

Note: At the local mac, use the paperqa env, which has installed all dependencies.

Environment Setup

Create a .env file:

OPENAI_API_KEY=your_openai_api_key
PUBMED_API_KEY=your_ncbi_api_key
REACTOME_NEO4J_URI=bolt://localhost:7687
REACTOME_NEO4J_USER=neo4j
REACTOME_NEO4J_PWD=your_password
REACTOME_NEO4J_DATABASE=reactome
PUBMED_MONGO_URI=mongodb://localhost:27017
PUBMED_MONGO_DB=pubmed_cache
PUBMED_MONGO_COLLECTION=abstracts
PDF_PAPERS_FOLDER=./data/papers

🔧 Usage

Multi-Agent Annotation (CrewAI)

from reactome_llm.CrewAILiteratureAnnotator import CrewAILiteratureAnnotator, AnnotationRequest
from reactome_llm.GenePathwayAnnotator import GenePathwayAnnotator

# Initialize
annotator = GenePathwayAnnotator() 
crewai = CrewAILiteratureAnnotator(annotator)

# Create annotation request
request = AnnotationRequest(
    gene="NTN1",
    papers=["25391454", "22982992", "23467207"],
    quality_threshold=0.7,
    enable_full_text=False
)

# Run multi-agent annotation
result = await crewai.annotate_literature(request)
print(f"Quality score: {result.quality_scores}")

Traditional Single-Agent Annotation

from reactome_llm.GenePathwayAnnotator import GenePathwayAnnotator

annotator = GenePathwayAnnotator()
result = await annotator.write_summary_for_gene_annotation("NTN1")

REST API

Start the server:

flask --app reactome_llm/ReactomeLLMRestAPI run --debug

Endpoints

Multi-Agent Annotation:

curl -X POST http://localhost:5000/crewai/annotate \
  -H "Content-Type: application/json" \
  -d '{
    "queryGene": "NTN1",
    "numberOfPubmed": 8,
    "qualityThreshold": 0.7,
    "targetPathways": ["Axon guidance"],
    "enableFullText": false
  }'

Traditional Annotation:

curl -X POST http://localhost:5000/annotate \
  -H "Content-Type: application/json" \
  -d '{
    "queryGene": "NTN1",
    "numberOfPubmed": 8,
    "cosineSimilarityCutoff": 0.38,
    "llmScoreCutoff": 3
  }'

System Status:

curl http://localhost:5000/crewai/status

📊 Quality Assessment

The multi-agent framework provides comprehensive quality metrics:

• Biological Accuracy (0-1): Correctness of molecular mechanisms
• Evidence Support (0-1): Strength of literature backing
• Mechanistic Consistency (0-1): Alignment with known biology
• Integration Quality (0-1): Compatibility with existing data

Quality Thresholds

• Approve: Score ≥ 0.7, no critical issues
• Requires Revision: Score 0.5-0.7, minor issues
• Reject: Score < 0.5, major inaccuracies

🧪 Testing

Run the validation suite:

python test_crewai_framework.py

Run example workflows:

python examples/crewai_annotation_examples.py

🗂️ File Structure

reactome_llm/
├── CrewAILiteratureAnnotator.py    # Main multi-agent orchestrator
├── ReactomeAgents.py               # Specialized agent definitions  
├── ReactomeTasks.py                # Task definitions for each phase
├── ReactomeTools.py                # Agent-specific tools
├── GenePathwayAnnotator.py         # Core annotation engine
├── ReactomeLLMRestAPI.py           # REST API with both approaches
├── ReactomeNeo4jUtils.py           # Neo4j database utilities
├── ReactomePubMed.py               # PubMed integration
└── README_CrewAI.md               # Detailed CrewAI documentation

examples/
└── crewai_annotation_examples.py  # Usage examples

test/
└── test_crewai_framework.py       # Validation tests

🔗 Data Sources

• Reactome Database: Neo4j graph database with pathway knowledge
• PubMed: Literature abstracts via NCBI E-utilities API
• IntAct: Protein-protein interaction data
• BioGRID: Molecular interaction database
• MongoDB: Local caching of PubMed abstracts

🚀 Deployment

Server Deployment

To deploy to production server (curator.reactome.org):

1. Zip the reactome_llm folder
2. Transfer and unzip on server
3. Configure .env with production settings
4. Run using the shell script:

./run_llm.sh

To stop the application:

ps aux | grep llm
kill <process_id>

Database Migration

MongoDB databases are generated locally and migrated:

# Export from local
mongodump --db your_database_name --out /path/to/backup

# Import to server  
mongorestore --db your_database_name /path/to/backup/your_database_name

📚 Documentation

• CrewAI Framework Guide - Detailed multi-agent documentation
• API Reference - Complete REST API documentation
• Examples - Usage examples and tutorials

🤝 Contributing

This tool is part of the Reactome project. For contributions:

1. Follow existing code patterns
2. Add tests for new functionality
3. Update documentation
4. Ensure compatibility with both single and multi-agent approaches

📄 License

This project is part of the Reactome curation tools and follows the same licensing terms.