DeepMind releases atlas scoring 9 billion genome variants

Google DeepMind is the research organization within Google that builds AI models. It started out as an independent company founded in London in 2010, was acquired by Google in 2014, and took its current name in 2023 after merging with another Google AI unit. This announcement comes from DeepMind's science-focused team, placing it in the same lineage as AlphaFold, which predicts protein structures, and WeatherNext, which forecasts weather.

Human DNA consists of roughly 3 billion base pairs, but only about 2% of that codes for proteins. The remaining 98%, known as the non-coding region, has long remained largely uncharted territory for scientists. DeepMind's earlier AlphaGenome model showed how a single-letter change in that non-coding region could disrupt molecular processes like protein production, but it stopped short of mapping the genome as a whole.

To put it simply, AlphaGenome is an AI model that predicts what happens when a single letter in DNA changes, and the new Atlas is the full set of results from running that model in advance on all 9 billion possible variants in the human genome. Researchers no longer need to run the model themselves in their own labs — they can just search for the variant they're curious about and see the prediction right away.

To make this massive set of predictions manageable, the Atlas introduces a new metric called AlphaGenome Variant Impact, or AVI score. It rolls predictions for both coding and non-coding regions into a single number, so researchers can quickly gauge which variants deserve a closer look instead of combing through thousands of individual data points.

In fact, this score has already produced results at several research sites. At the Broad Institute, a team led by Laura Coville used the AVI score to prioritize candidate variants in a long-unsolved rare-disease case, and pinpointed a single variant that creates an incorrect splice site in the DNM1 gene. A splice site is where a cell cuts out unnecessary sections of a gene and stitches together the parts it needs; when that site is disrupted, the resulting protein comes out wrong. The prediction reportedly became decisive evidence in solving that case.

The second case is bigger in scale. Dr. Gareth Hawkes applied the Atlas to data from more than 54,000 UK Biobank participants, grouping rare non-coding variants — ones that are hard to detect because of statistical noise — by their predicted molecular effects. That approach turned up 22% more genetic associations in non-coding regions than before, and by focusing on the top 1% of variants by impact, the team identified 19 new genetic regions linked to body mass index, or BMI.

The Atlas is available starting today through a web portal that requires no coding knowledge to use. DeepMind describes this as part of an effort to lower the barrier to entry so that any clinical researcher or biologist worldwide can use it — though it's worth noting that the database provides research-grade predictions, not a validated tool ready for direct use in clinical diagnosis or treatment decisions.

Until now, geneticists had to design experiments and spend months or years just to confirm the effect of a single variant. The Atlas cuts that timeline down by collapsing the first step of the process — deciding which variant to investigate — into a single search. As the Broad Institute and Biobank cases show, this approach is already producing concrete results in both rare-disease diagnosis and complex-trait research, and that's the real headline of this release.