DeepMind Unveils AlphaGenome Atlas to Predict the Effects of 9 Billion DNA Variants

The effort to decode the human genome is currently caught between two incompatible realities. On one side, computational biologists argue that the sheer scale of DNA—roughly 3 billion base pairs, yielding 9 billion possible single-letter mutations—requires artificial intelligence to predict which changes matter, because testing them all physically is impossible. On the other side, translational researchers maintain that advanced AI models are too computationally expensive for everyday laboratory use, and that a prediction remains useless until it is validated in a physical clinical trial.[2][3]

Google DeepMind has attempted to bypass this tension entirely by removing the AI model from the user's workflow. On September 8, 2026, the company released the AlphaGenome Atlas, a 1-petabyte database that pre-calculates the molecular impact of every possible single-letter DNA change. Instead of running a computationally intensive model to test a variant, researchers can now simply look up the answer in a web browser.[1][2][3]

Historically, scientific research has focused heavily on the 2% of the genome that directly codes for proteins. The remaining 98%, often referred to as non-coding regions, regulates how and when those genes are turned on or off. Understanding this regulatory architecture is critical for tracing the origins of complex traits and diseases, but the scale of the non-coding genome has made systematic analysis prohibitively difficult.[2]

DeepMind previously addressed this with AlphaGenome, an AI model capable of predicting how genetic variants disrupt biological processes. "Our AlphaGenome model has already shown how single changes in these non-coding DNA regions can disrupt molecular processes like protein production, but the bigger picture remained unclear," wrote Pushmeet Kohli, DeepMind's vice president of science, and Žiga Avsec, genomics initiative lead, in the release announcement. Furthermore, running the model on individual variants required significant computational resources and coding expertise, creating a barrier to entry for many rare-disease researchers and smaller laboratories.[2]

To build the Atlas, DeepMind ran the AlphaGenome model across all 9 billion possible single-nucleotide variants in advance, calculating the effects of all 3 possible letter swaps at every position—for example, changing an A to a T, C, or G. The database also includes predictions for more than 100 million short insertions and deletions. By pre-computing the entire field of possibilities, the company shifted the computational cost from the individual researcher's budget to its own infrastructure.[3]

Navigating a 1-petabyte dataset of raw biological predictions presents its own operational challenge. To prevent researchers from drowning in data, the Atlas introduces 1 unified metric: the AlphaGenome Variant Impact (AVI) score. This single number condenses predictions from 2 separate systems—AlphaGenome, which analyzes non-coding regions, and AlphaMissense, DeepMind's model for protein-altering variants.[2][3]

The AVI score provides a ranked starting point for scientists working through large genetic datasets. Rather than sifting through thousands of disparate metrics across hundreds of cell types, laboratory teams can use the score to instantly prioritize the most promising variants for experimental follow-up. The tool is designed to act as a hypothesis generator, narrowing the field of possibilities before teams commit laboratory time and capital.[2][3]

The Atlas is already being deployed in active research. At the Broad Institute, scientists have used the AVI score to prioritize variants in unsolved rare disease cases. In one instance, the tool highlighted a critical non-coding variant in the DNM1 gene, predicting that the mutation created an incorrect splice site associated with severe epilepsy.[2]

DeepMind has made the AlphaGenome Atlas freely available for non-commercial academic research through an interactive web portal, eliminating the need for coding skills to query the data. Commercial availability through Google Cloud is planned for a later date, signaling a potential shift in how hyperscalers monetize AI—by charging for access to massive, pre-computed scientific datasets rather than just the underlying models.[2][3]

The release fundamentally changes the sequence of genomic research. The immediate challenge for life-sciences leaders is no longer generating a genome-scale list of hypotheses, but building reproducible, high-throughput processes for deciding which of those AI-generated leads deserve physical validation in the lab.[3]