🧬 GeneLM-Evo2

Genomic Intelligence Powered by Evo2

A full-stack genomic variant analysis platform leveraging the Evo2 DNA language model for zero-shot pathogenicity prediction

Features • Tech Stack • Quick Start • Architecture • API

---

✨ Features

🔬 Evo2-Powered Analysis 7B parameter DNA language model Zero-shot variant pathogenicity prediction ~95% AUROC on BRCA1 benchmark Real-time inference on H100 GPUs	🧭 Interactive Gene Browser Browse by chromosome or search by gene Interactive sequence viewer with nucleotide highlighting Click any base to trigger variant analysis Support for 24+ genome assemblies
🏥 ClinVar Integration Fetch clinically curated variants Compare Evo2 predictions vs clinical labels One-click analysis for SNVs Confidence scoring with delta-likelihood	⚡ Modern Tech Stack Next.js 15 with Turbopack React 19 + TailwindCSS 4 Modal serverless infrastructure UCSC & NCBI API integration

---

🛠️ Tech Stack

Frontend

Technology	Version	Purpose
Next.js	15	React framework with App Router
React	19	UI library
TailwindCSS	4	Utility-first CSS
shadcn/ui	Latest	Component library
Framer Motion	12	Animations
TypeScript	5.8	Type safety

Backend

Technology	Version	Purpose
Python	3.12	Runtime
Modal	Latest	Serverless GPU infrastructure
Evo2	7B	DNA language model
PyTorch	2.8	Deep learning framework
Flash Attention	2.8.3	Efficient attention
CUDA	12.6	GPU acceleration

External APIs

• UCSC Genome Browser API — Reference sequence data
• NCBI ClinVar API — Clinical variant annotations
• NCBI Gene API — Gene information and coordinates

---

🚀 Quick Start

Prerequisites

• Node.js 20+ and npm
• Python 3.12+
• Modal account (sign up)
• NVIDIA GPU with CUDA support (for local development, or use Modal's H100s)

Frontend Setup

# Navigate to frontend directory
cd genelm-frontend

# Install dependencies
npm install

# Create environment file
cp .env.example .env.local

# Configure your environment variables
# NEXT_PUBLIC_ANALYZE_SINGLE_VARIANT_BASE_URL=<your-modal-endpoint>

# Start development server
npm run dev

The frontend will be available at http://localhost:3000

Backend Setup

# Navigate to backend directory
cd genelm-backend

# Install Modal CLI
pip install modal

# Authenticate with Modal
modal setup

# Deploy the application
modal deploy main.py

# Or run locally for development
modal serve main.py

After deployment, Modal will provide an endpoint URL for the analyze_single_variant API.

---

🏗️ Architecture

┌─────────────────────────────────────────────────────────────────┐
│                         Frontend (Next.js)                       │
│  ┌─────────────┐  ┌─────────────┐  ┌─────────────────────────┐  │
│  │ Gene Browser│  │  Sequence   │  │   Variant Analysis      │  │
│  │  Component  │  │   Viewer    │  │      Dashboard          │  │
│  └──────┬──────┘  └──────┬──────┘  └───────────┬─────────────┘  │
│         │                │                      │                │
└─────────┼────────────────┼──────────────────────┼────────────────┘
          │                │                      │
          ▼                ▼                      ▼
   ┌──────────────┐ ┌──────────────┐    ┌────────────────┐
   │  NCBI Gene   │ │ UCSC Genome  │    │  Modal Backend │
   │     API      │ │     API      │    │    (H100)      │
   └──────────────┘ └──────────────┘    └───────┬────────┘
                                                │
                                                ▼
                                        ┌──────────────┐
                                        │    Evo2      │
                                        │  (7B Model)  │
                                        └──────────────┘

Data Flow

1. Gene Selection → User browses/searches genes via NCBI Gene API
2. Sequence Loading → Genomic sequence fetched from UCSC API
3. Variant Selection → User clicks nucleotide or selects ClinVar variant
4. Evo2 Analysis → Request sent to Modal backend with H100 GPU
5. Prediction → Model scores reference vs variant sequences
6. Results → Delta-likelihood score + pathogenicity prediction returned

---

📡 API Reference

Backend Endpoint

POST /analyze_single_variant

Analyze a single nucleotide variant using Evo2.

Request Body:

{
  "variant_pos": 43119628,
  "alt_allele": "G",
  "genome": "hg38",
  "chromosome": "chr17"
}

Response:

{
  "position": 43119628,
  "reference": "A",
  "variant": "G",
  "delta_score": -0.00234,
  "prediction": "Likely pathogenic",
  "confidence": 0.87
}

Field	Type	Description
position	int	Genomic position
reference	str	Reference allele
variant	str	Alternative allele
delta_score	float	Log-likelihood difference (ref - var)
prediction	str	"Likely pathogenic" or "Likely benign"
confidence	float	Confidence score (0-1)

---

📊 Performance

BRCA1 Benchmark

Metric	Value
AUROC	~95%
Model	Evo2 7B
Context Window	8,192 bp
Variants Tested	500 SNVs
Classification	LOF vs FUNC/INT

The model uses a threshold-based classification derived from Youden's J statistic optimization on the BRCA1 saturation mutagenesis dataset.

Inference Performance

Configuration	Latency
Modal H100 (cold start)	~30s
Modal H100 (warm)	~2-5s
Batch scoring (100 variants)	~60s

---

📁 Project Structure

GeneLM-Evo2/
├── genelm-frontend/          # Next.js frontend application
│   ├── src/
│   │   ├── app/              # App router pages
│   │   ├── components/       # React components
│   │   │   ├── gene-sequence.tsx
│   │   │   ├── known-variants.tsx
│   │   │   └── ui/           # shadcn/ui components
│   │   └── utils/            # API utilities
│   │       ├── variants-api.ts
│   │       ├── genome-api.ts
│   │       └── genes-api.ts
│   ├── package.json
│   └── tailwind.config.ts
│
├── genelm-backend/           # Modal serverless backend
│   ├── main.py               # Evo2 model & API endpoints
│   └── requirements.txt
│
└── README.md

---

🔮 Roadmap

• Batch variant analysis
• VCF file upload support
• Additional gene benchmarks (TP53, BRCA2)
• Variant effect visualization
• Export results to PDF/CSV
• Multi-model comparison (ESM, Nucleotide Transformer)

---

🙏 Acknowledgments

• Arc Institute — Evo2 DNA language model
• Modal — Serverless GPU infrastructure
• UCSC Genome Browser — Reference genome data
• NCBI ClinVar — Clinical variant database
• shadcn/ui — Beautiful UI components

---

📄 License

This project is licensed under the MIT License - see the LICENSE file for details.

---

Built with 🧬 by Jarvis Zhang