NexVar: Advanced Variant Analysis Platform

Paper: Genome modeling and design across all domains of life with NexVar
Repository: Eldergenix/NexVar

Project Abstract

NexVar represents a paradigm shift in computational genomics, democratizing access to state-of-the-art foundation models for variant effect prediction (VEP). This platform integrates the NexVar DNA language model—trained on 8.8 trillion tokens from all domains of life—into a cohesive, full-stack application.

By coupling High-Performance Computing (HPC) backends with an accessible, research-grade frontend, this project aims to impose the predictive power of 1-million-context-length models directly onto the clinical research workflow.

System Capabilities

The architecture allows researchers to:

• Predict Pathogenicity: Utilize the NexVar model to classify SNVs as Likely Pathogenic or Likely Benign based on evolutionary likelihood intervals.
• Analyze Long-Range Context: Leverage the StripedHyena 2 architecture to understand mutation effects within windows of up to 1 million base pairs (backend capability).
• Validate via ClinVar: Benchmark AI predictions against established clinical assertions from NCBI ClinVar.
• Visual Genome Analysis: Interactively browse chromosomes and gene regions (hg38/hg19 assemblies).

Architecture Components

The platform consists of two primary subsystems:

1. The Inference Engine (Backend)

Located in nexvar-backend/

• Core: Python-based API wrapping the NexVar library.
• Compute: Deployed on Modal Serverless GPU infrastructure (H100).
• Function: Handles model loading (7B/40B parameters), tokenization, and forward-pass inference.

2. The Research Workspace (Frontend)

Located in nexvar-frontend/

• Core: Next.js / TypeScript web application.
• Interface: Scientific dashboard for gene search, variant input, and results visualization.
• Deployment: Hosted on Railway for high availability.

Getting Started

Prerequisites

• Computational: Access to Modal (for backend) and Railway (for frontend).
• Local: Python 3.10+ and Node.js 18+.

Installation Strategy

Phase 1: Deploying the Inference Engine

The backend must be operational first to provide API services.

cd nexvar-backend
pip install -r requirements.txt
modal deploy main.py
# >> Save the output URL.

Phase 2: Launching the Workspace

The frontend consumes the inference API.

cd nexvar-frontend
npm install
npm run dev

Citation

If this platform aids your research, please cite the foundational work:

Brixi, G., et al. (2025). Genome modeling and design across all domains of life with NexVar. bioRxiv. doi:10.1101/2025.02.18.638918