The bottleneck in reading the human genome has never been sequencing it β it has been working out which of the roughly 9 billion possible single-letter changes actually do anything. Google DeepMind's answer, released on September 8, is to stop guessing one variant at a time and precompute all of them: AlphaGenome Atlas is a one-petabyte catalogue of predicted molecular effects, free for academic use and more than 30 times the size of the AlphaFold Database.
Key takeaways
- AlphaGenome Atlas contains predicted molecular effects for about 9 billion single-nucleotide variants, covering both coding and non-coding regions of the human genome.
- A new AlphaGenome Variant Impact (AVI) score condenses AlphaGenome and AlphaMissense predictions into one number so researchers can rank variants before running experiments.
- Early collaborators used it to pin down a DNM1 variant tied to epileptic encephalopathy and to surface 22% more non-coding trait associations in UK Biobank data.
Why this is a release at all
AlphaGenome itself shipped last year and is already in research use, a point DeepMind genomics lead Ziga Avsec acknowledged in a press briefing covered by The Verge, where he attributed the year-long gap to the raw cost of precomputing and analysing a variant space that large. The engineering work, in other words, was the product: an on-demand inference problem became a lookup table, which is a different thing to hand a biologist.
That distinction is what makes the AlphaFold comparison load-bearing. The 2022 database expansion took protein structures from roughly 190,000 experimentally determined entries to more than 200 million predictions, and the reason it reshaped the life sciences was accessibility rather than novelty β researchers with no coding background could finally browse it.
What a single variant lookup returns
Each of the 9 billion entries carries thousands of molecular effect predictions across hundreds of human and mouse cell types and tissues, which the AVI score compresses into one rankable number by folding in AlphaMissense's protein-impact predictions alongside AlphaGenome's regulatory ones. Feature attributions then decompose that score back into chromatin accessibility, splicing and conservation, so a high rank comes with a mechanism attached rather than just a warning light.
A fourth resource maps more than 2,500 recurrent DNA motifs β the short sequences that transcription factors bind. DeepMind reports best-in-class AVI performance on several variant pathogenicity and rare disease benchmarks, and stresses that the score works in non-coding DNA, the 98% of the genome that houses most trait-associated variants and where interpretation has always been weakest.
What collaborators found before launch
Working with the GREGoR Consortium, Broad Institute researchers Laura Covill and Anne O'Donnell-Luria used AVI to re-rank rare disease candidates that earlier analyses had passed over, and surfaced a DNM1 variant β a gene tied to epileptic encephalopathy β that the model predicted would create a faulty splice site and an abnormally extended protein. Experimental screens confirmed the prediction and turned up neighbouring variants behaving the same way.
A second test ran at population scale. Gareth Hawkes of the University of Exeter applied the Atlas to whole-genome data from over 54,000 UK Biobank participants and, by grouping rare variants according to predicted molecular effect, pulled 22% more non-coding associations out of the statistical noise β among them regulatory variants governing PLA2G7 and EGLN1. Narrowing a body-mass-index analysis to the most impactful 1% of non-coding variants left 19 genetic regions worth targeting experimentally.
Availability and limits
Access runs through the website portal, the AlphaGenome API and a skill inside Google Antigravity today for non-commercial use, with Google Cloud availability for commercial users promised later. DeepMind is blunt that this is a machine learning output and not clinical guidance β AlphaGenome has not been validated or approved for any clinical use, and every hit still needs wet-lab confirmation.
The release also fits a pattern. Google DeepMind's science group has been converting research models into operational infrastructure at a steady clip, most recently when WeatherNext bought forecasters an extra day of cyclone warning, and it arrives as co-founder Demis Hassabis steps back from running the lab to concentrate on scientific research and drug discovery at Isomorphic Labs.
FAQ
Is AlphaGenome Atlas free to use?
Yes, for non-commercial academic research. Access runs through a free web portal and the AlphaGenome API. Commercial use will be offered on Google Cloud at a later date.
What is the AVI score?
The AlphaGenome Variant Impact score combines AlphaGenome's regulatory predictions with AlphaMissense's protein-impact predictions into a single ranking number. It ships with feature attributions showing which molecular processes drive the score, and DeepMind reports best-in-class results across several variant pathogenicity and rare disease benchmarks.
Can doctors use it to diagnose patients?
No. DeepMind states that AlphaGenome has not been validated for clinical use and that Atlas output is not a substitute for professional medical advice, diagnosis, or treatment. It is a research prioritisation tool whose outputs still require experimental validation.
