METAL
Add us on GoogleMETAL

DeepMind scores all 9 billion possible human genome variants

AlphaGenome Atlas precomputes every single-base change in the genome as a 1-petabyte dataset, now open to researchers worldwide

DeepMind scores all 9 billion possible human genome variants

Image: generated by METAL AI

Summary

  • Google DeepMind has released the AlphaGenome Atlas, which precomputes the impact of roughly 9 billion possible single-base variants across the human genome.
  • Each variant carries an AlphaGenome Variant Impact (AVI) score that condenses its risk into a single number, and the full dataset runs to about 1 petabyte.
  • It covers both coding and non-coding regions and is open to researchers worldwide, but DeepMind stresses it has not been validated for clinical diagnosis or treatment.
AlphaGenome Atlas: Understanding the human genome

Google DeepMind has released AlphaGenome Atlas, a single dataset that precomputes the effects of every possible single-base variant in the human genome. That means every one-letter change that could occur anywhere in human DNA — about 9 billion of them in total. The whole dataset runs to roughly 1 petabyte, and each variant comes tagged with an AlphaGenome Variant Impact score, or AVI score, indicating how harmful it's likely to be. The atlas is freely available to researchers around the world, covering variants in both protein-coding regions and non-coding regions of the genome.

Starting from a grid representing all 9 billion possible single-base variants in the human genome. These variants pass through the AlphaGenome model, where each one is turned into a dense bundle of nearly 10,000 complex predictions. That bundle is then compressed down into a single circle — one AVI score. Finally, a researcher can simply look up that compressed score to immediately gauge risk.Starting from a grid representing all 9 billion possible single-base variants in the human genome. These variants pass through the AlphaGenome model, where each one is turned into a dense bundle of nearly 10,000 complex predictions. That bundle is then compressed down into a single circle — one AVI score. Finally, a researcher can simply look up that compressed score to immediately gauge risk.
Image: Generated by METAL AI

To put that in perspective: human DNA is a sequence of roughly 3 billion base pairs made up of four letters — A, T, G, and C. Count every possible way a single one of those letters could be swapped for another, and you get 9 billion variants. Until now, there was no practical way to check whether any given one of those variants actually causes disease. The AVI score solves part of that problem by compressing AlphaGenome's complex predictions into one number, letting researchers immediately gauge which variants are likely to matter.

AlphaGenome is a model that takes a stretch of genome sequence as input and predicts what happens inside a cell if a mutation occurs in that stretch. Gareth Hawkes, a lecturer specializing in whole-genome sequence analysis at the University of Exeter, notes in a video that even spending just one second looking at each base pair would take decades to get through the entire genome. Sam Bryen, a genomic analyst at the Population Genomics Center, similarly points out that scientists still don't fully understand how the instructions packed into those 3 billion base pairs interact with one another.

The challenge was that AlphaGenome's raw output for each variant is a dense bundle of nearly 10,000 numbers. The team condensed that into the single AVI metric, scaled so that more impactful variants get higher scores. That, they say, narrows a vast pool of candidates down to a much smaller set, making it far easier to zero in on variants that actually matter. AlphaGenome Atlas is both the petabyte-scale database holding these scores and a genome browser — a web interface that lets biologists without coding skills look up variant impacts by chromosome or cell type.

Google DeepMind is careful to note that these scores are not a substitute for professional medical advice, diagnosis, or treatment, and have not been validated or approved for clinical use. The team says that following this atlas release, it's now working on tying together specialized models like AlphaFold and AlphaGenome into a single agentic system, aimed at helping scientists tackle bigger problems facing humanity. More background is available in DeepMind's introductory post on the AlphaGenome Atlas.

Google DeepMind opened an AI campus in Korea this past April, expanding its collaboration with local industry, academia, and research institutions. Large-scale predictive datasets like this one, released for free, are the kind of resource Korean genomics and biotech research teams can query immediately, with no separate agreement required.

What actually changes here is speed. Previously, figuring out what a single variant found in a specific patient actually meant required a research team to run AlphaGenome themselves or dig through the relevant literature. Now that all 9 billion variants are precomputed, a single score lookup is enough to establish priority. That said, the score itself still isn't a diagnosis — moving to actual clinical use still requires a separate validation step.

Comments

Related stories