MutAIon - LLM based Variant Analysis WebApp

This project implements a web application for variant effect prediction, leveraging the Evo2 large language model to determine the pathogenicity of single nucleotide variants (SNVs). It includes:

• Python Backend: A FastAPI service running on an H100 GPU via Modal serverless infrastructure.

• AI Model: Evo2 for classifying SNVs as pathogenic or benign.

• Database Integration: Fetches ClinVar classifications for comparison.

• Frontend: Built with Next.js, React, TypeScript, Tailwind CSS, and Shadcn UI (T3 Stack).

• Features:

  • Genome assembly selection
  • Chromosome browsing and gene search (e.g., BRCA1)
  • Reference genome display
  • Mutation input and prediction
  • Comparison with ClinVar data

Architecture

Backend

• Framework: FastAPI
• Deployment: Modal serverless GPU (H100)
• Endpoint: /predict accepts gene, position, reference and alternative alleles, returns pathogenicity score and label.
• Dependencies: evo2, modal, fastapi, uvicorn, sqlalchemy, requests

Frontend

• Framework: Next.js with TypeScript

• Styling: Tailwind CSS

• Components: Shadcn UI (Buttons, Inputs, Cards)

• Pages:

  • / Assembly selection
  • /browse/[assembly] Chromosome listing
  • /gene/[geneId] Gene details and sequence

• API Integration: Uses React Query to call FastAPI endpoints.

Setup

Prerequisites

• Node.js >= 18
• Python 3.10+
• Modal CLI and account

Backend Installation

cd backend
pip install -r requirements.txt
modal deploy

Frontend Installation

cd frontend
npm install
npm run dev

Usage

1. Launch backend via Modal.
2. Start frontend in development mode.
3. Navigate to http://localhost:3000, select your genome assembly.
4. Browse or search for genes, input mutations or select existing variations.
5. View Evo2 predictions alongside ClinVar classifications.

Contributing

Contributions are welcome! Please open issues or submit pull requests.

License

MIT License