AlphaVX

Batch-score genetic variants against AlphaGenome and generate interpretable effect reports — gene expression, splicing, chromatin accessibility, histone marks, and transcription factor binding — in one command.

alphavx score variants.vcf -o results/

---

What It Does

AlphaGenome (DeepMind, Jan 2026) is a deep learning model that predicts how DNA mutations affect gene regulation across thousands of functional genomic tracks. It analyzes up to 1 million bases of context and matches or exceeds the strongest models in 25 of 26 variant-effect benchmarks.

AlphaVX makes it practical to use:

• VCF batch scoring — Score all variants from a standard VCF file in one command
• Resilient execution — Automatic retry with exponential backoff, result caching, and resumable runs
• Multi-modal scoring — Expression (RNA-seq), splicing, chromatin (ATAC/DNase), histone marks (ChIP), TF binding — all scored automatically
• Interpretable output — Per-variant report with significance flags, tissue-specific effects, and quantile rankings against common variants
• Gene panel filtering — Focus on clinically relevant genes instead of scoring everything

---

Quick Start

Install

pip install alphavx

Or from source:

git clone https://github.com/RenX86/AlphaVX
cd AlphaVX
pip install -e .

Get an API Key

1. Sign up at deepmind.google.com/science/alphagenome
2. Set your key:

   export ALPHAVX_API_KEY="your-key-here"
   

Run

# Score all variants in a VCF
alphavx score input.vcf -o results/

# Score with gene panel filter
alphavx score input.vcf --genes BRCA1,TP53,CFTR -o results/

# Score a single variant
alphavx query chr17:7674220:G\>A

# Generate HTML report from cached results
alphavx report results/

---

Output

results/
├── scores.csv            # Full scoring table (all variants × all modalities)
├── significant.csv       # Variants exceeding significance thresholds
├── report.html           # Interactive summary report
├── cache/                # Cached API responses (resumable)
└── plots/
    ├── summary_heatmap.png
    └── per_variant/
        ├── chr17_7674220_G_A.png
        └── ...

scores.csv columns

Column	Description
chrom	Chromosome
pos	Position (1-based)
ref / alt	Reference and alternate alleles
gene	Nearest gene
modality	Scored output type (RNA_SEQ, SPLICE_SITES, DNASE, etc.)
tissue	Biosample / tissue name
raw_score	AlphaGenome raw effect score
quantile_score	Percentile rank vs common variants (±0.999990 max)
significant	Boolean flag (quantile > 0.995)

---

How It Works

VCF file
   ↓
Parse variants (chrom, pos, ref, alt)
   ↓
Check cache → skip already-scored variants
   ↓
For each variant:
   → Build 1MB interval around variant
   → Score against all recommended variant scorers
   → Tidy scores, match gene strand
   ↓
Aggregate results → flag significant hits
   ↓
Generate report (CSV + HTML + plots)

---

Configuration

Copy alphavx.yaml and customize:

api:
  key_env: ALPHAVX_API_KEY
  max_retries: 3
  retry_delay: 5.0

scoring:
  modalities:
    - RNA_SEQ
    - SPLICE_SITES
    - SPLICE_SITE_USAGE
    - SPLICE_JUNCTIONS
    - DNASE
    - ATAC
    - CHIP_HISTONE
    - CHIP_TF
  quantile_threshold: 0.995
  sequence_length: 1048576

output:
  format: [csv, html]
  plots: true
  cache: true

---

Requirements

• Python ≥ 3.10
• AlphaGenome API key (sign up)
• Internet connection (API-based, no local GPU needed)

---

Tech Stack

Component	Choice
AlphaGenome SDK	alphagenome>=0.6.1
CLI	Typer + Rich
Caching	SQLite (zero-dependency, resumable)
Reports	Self-contained HTML
Plots	Matplotlib + Seaborn
Packaging	pyproject.toml + Hatch

---

Development

git clone https://github.com/RenX86/AlphaVX
cd AlphaVX
pip install -e ".[dev]"
pytest

---

License

Apache 2.0

---