Google DeepMind releases AlphaGenome Atlas with 9 billion DNA-variant predictions

• AlphaGenome Atlas makes predicted molecular effects for all 9 billion possible single-nucleotide changes in the human genome searchable, Google’s Atlas announcement says, with free access for academic research.
• The catalog is a one-petabyte precomputation, and Google DeepMind’s dataset post puts it at more than 30 times the size of the AlphaFold Database.
• The AlphaGenome Variant Impact score, or AVI, combines AlphaGenome, AlphaMissense and other signals to rank variants and expose likely molecular damage, according to Google DeepMind’s launch thread.
• The first public interface is a no-code website for clinical researchers and biologists, which Google’s portal announcement says is available now.

Only about 2% of the human genome encodes proteins, Google’s product post notes, while the rest contains much less understood regulatory sequence. The official AlphaGenome research repository documents a JAX model that works at single-base resolution on DNA sequences up to one million base pairs long.

The nine-billion count covers the three possible single-letter substitutions at each of roughly three billion positions in the human genome. Google says in its product post that AlphaGenome pre-calculated each variant’s regulatory impact, producing a one-petabyte corpus.

Atlas serves that fixed prediction corpus through a search interface. The scale is substantial: Google DeepMind’s dataset post describes it as more than 30 times the AlphaFold Database.

AVI is the Atlas’s single prioritization score. Google DeepMind’s launch thread says it combines AlphaGenome, AlphaMissense and other features, with two intended jobs:

• Prioritization: rank mutations from low to high predicted impact.
• Mechanism: surface possible damage pathways, including disrupted gene switches or RNA-splicing instructions.

Google’s product post says the score spans coding and non-coding predictions, reducing a large collection of model outputs to one sortable signal.

Google describes two early applications on its detailed Atlas page:

• Rare disease: A Broad Institute team used AVI to prioritize variants in an unsolved case. Google says the score highlighted a DNM1 variant predicted to create an incorrect splice site, providing supporting evidence that helped solve the case.
• Complex traits: In data from more than 54,000 UK Biobank participants, grouping variants by predicted molecular effect produced 22% more non-coding associations. Google says restricting analysis to the top 1% of predicted-impact variants identified 19 body-mass-index-linked regions for follow-up.

The browser is only one delivery surface. Google DeepMind’s distribution post lists the AlphaGenome API and a Google Antigravity skill, and says a Google Cloud route is coming.

Nature’s independent report describes the current release as free for non-commercial use and reports that roughly 9,000 researchers had already accessed AlphaGenome predictions through its API. Martin Kircher, a bioinformatician at the Max Delbrück Centre for Molecular Medicine, told Nature that the atlas cannot replace experiments or account for the specifics of an individual disease case.