Pharmacogenomics Drug Risk Analysis
Gene profile + drug → personalized risk score + CPIC clinical recommendation

Link: http://geneai.tech/

1.3 million emergency room admissions per year in the U.S. are caused by adverse drug reactions. Most drugs are prescribed based on symptoms alone — not genetics. Patients with certain gene variants metabolize drugs too quickly, too slowly, or not at all, leading to toxicity or treatment failure.

CPIC pharmacogenomic guidelines exist, but they're buried in academic tables clinicians don't have time to look up mid-appointment.

GeneAI is an API + UI that takes a patient's gene profile and a drug name, and returns:

• Per-gene activity scores from a trained Set Transformer model
• CPIC clinical recommendation text (e.g., "Avoid codeine use" or "Use standard dosing")
• Plain-English explanation suitable for patients
• Supports both structured JSON and natural language input

Two prediction paths:

1. Natural Language — User query → OpenAI parses to structured JSON → model pipeline
2. Structured — Direct JSON {"genes": [...], "drug": "..."} → model pipeline

Model pipeline:

• Gene embeddings via ESM-2 protein language model
• Drug embeddings via Morgan fingerprints from SMILES
• Target flags from DrugBank drug-gene interactions
• Cross-Attention Set Transformer fuses multi-gene + drug features → per-gene risk scores

Post-prediction, CPIC recommendation text is looked up and optionally enriched by GPT-4o-mini.

├── api/                 # FastAPI backend
│   ├── routes/          # health, drugs, genes, predict, explain
│   ├── services/        # data_service, model_service, openai_service
│   ├── models/          # request/response schemas
│   └── static/          # docs + demo HTML pages
├── model/               # Set Transformer model
│   ├── model_api.py     # PharmaRiskModel interface
│   ├── model.py         # PharmaSetTransformer architecture
│   ├── set_transformer.pt
│   └── embeddings/      # gene_embeddings.pkl, drug_embeddings.pkl, target_flags.pkl
├── pipeline/            # Data pipeline + processed CPIC data
├── frontend/            # React + Three.js UI (GeneAI app)
└── requirements.txt     # Python dependencies

geneai.tech — UI live geneai.tech/docs — Documentation geneai.tech/demo — Interactive demo

# Clone
git clone https://github.com/Topupchips/HackIllinoisWinningIdea.git
cd HackIllinoisWinningIdea

# Install Python dependencies
pip install -r requirements.txt

# (Optional) Set OpenAI key for natural language + explain
export OPENAI_API_KEY=sk-...

# Run API
python -m uvicorn api.main:app --reload --port 8000

# Run frontend (separate terminal)
cd frontend && npm install && npm start

• http://localhost:3000 — GeneAI React app
• http://localhost:8000/docs — Documentation
• http://localhost:8000/docs/api — Interactive API explorer
• http://localhost:8000/demo — Live endpoint demo

All list endpoints support ?search=, ?page=, ?limit= and return fuzzy "did you mean" suggestions on no results.

Health check:

curl http://localhost:8000/health

{"status": "healthy", "model_loaded": true, "data_loaded": true, "drug_count": 323, "gene_count": 17}

Predict risk (structured):

curl -X POST http://localhost:8000/predict \
  -H "Content-Type: application/json" \
  -d '{"genes": [{"name": "CYP2D6", "phenotype": "Poor Metabolizer"}], "drug": "codeine"}'

[
  {
    "gene": "CYP2D6",
    "activity_level": 0.0,
    "medicine": "codeine",
    "text": "Avoid codeine use because of possibility of diminished analgesia."
  }
]

Predict risk (natural language):

curl -X POST http://localhost:8000/predict/natural \
  -H "Content-Type: application/json" \
  -d '{"query": "I am a CYP2D6 poor metabolizer taking codeine"}'

Explain risk:

curl -X POST http://localhost:8000/explain \
  -H "Content-Type: application/json" \
  -d '{"drug": "codeine", "risk_score": 8.5, "gene_contributions": {"CYP2D6": 0.85}}'

Required by HackIllinois rules.

What we built ourselves:

• Data extraction and cleaning pipeline (CPIC SQL, DrugBank XML, PubChem API)
• Risk score engineering and labeling logic
• Model architecture design (Cross-Attention Set Transformer)
• Training pipeline and feature engineering
• API design, endpoint logic, and service layer
• React + Three.js frontend

• Sanjavan Ghodasara
• Suhaan Khan
• Keming (Charles) Hu
• Ansh Tuli

MIT

Built for HackIllinois 2026