Google DeepMind Publishes Atlas of 9 Billion DNA Variants for Research

Google DeepMind has published AlphaGenome Atlas, a database that predicts the biological effects of roughly 9 billion possible single-letter changes to human DNA. It promises a faster way to narrow genetic searches, but the company says the resource is not validated for clinical use or diagnosis.

A lookup table for every one-letter swap

Rather than requiring a researcher to score one DNA change at a time, the Atlas stores predictions in advance. DeepMind created it by running AlphaGenome across a reference human genome and comparing every reference base with each of its three possible alternatives.

The result is an approximately 1-petabyte dataset. Each potential variant is linked to an average of about 27,000 predictions, including predicted effects on gene expression and transcription across hundreds of human and mouse cell and tissue types.

What the Atlas adds

• Predictions spanning both protein-coding and non-coding DNA regions.
• An AlphaGenome Variant Impact, or AVI, score that combines gene-regulation and protein-altering predictions.
• Maps of more than 2,500 short DNA motifs that transcription factors bind to.

One score to rank a crowded field

The Atlas’s shortcut is its AVI score, which combines AlphaGenome’s predictions about gene regulation with AlphaMissense predictions for protein-altering mutations. Researchers can use the score to rank variants instead of sorting through the underlying predictions individually.

In a retrospective test on previously solved GREGoR rare-disease cases, AVI put the known causal variant in a patient’s top 50 candidates 29.5% of the time, compared with 12.5% for CADD, an existing ranking method. That comparison measures prioritization in solved cases, not a clinical diagnosis.

Promising filters still need experiments

In an analysis of whole-genome data from more than 54,000 U.K. Biobank participants, filtering variants by Atlas-predicted molecular effects produced 22% more associations than the same analysis without Atlas. In one region, the process reduced 526 candidates to four.

A rare-disease case involving the DNM1 gene illustrates the intended workflow: researchers used AVI to prioritize a variant, then laboratory experiments confirmed the prediction. DeepMind also cautioned that AlphaGenome can miss some variant classes, particularly enhancer effects.

The Atlas is available for non-commercial research through a website, while commercial access through Google Cloud licensing is planned. The launch gives researchers a shared starting point for deciding what to investigate, while leaving biological confirmation to follow-up work.