AlphaGenome Atlas: AI’s map of 9 billion DNA changes

AlphaGenome Atlas takes on a nearly impossible task: estimating what might happen if you change one DNA letter to another at any position in the human genome. There are about 3 billion positions in the genome, and each one can be swapped for three alternative letters. That adds up to roughly 9 billion possible single-nucleotide changes.

That is what Google DeepMind has now precomputed. Rather than testing these variants one by one in the lab, the company has released a resource that estimates their likely molecular consequences in advance — how a variant might affect gene activity, splicing, chromatin accessibility, and other regulatory processes. It is not a catalogue of experimentally proven effects. It is a very large predictive map.

The release matters because the Atlas is built on AlphaGenome, the model described in a Nature paper published in January 2026. In that study, the authors showed that the model can take long stretches of DNA — up to 1 million base pairs — and predict many functional genomic readouts at high resolution.

What AlphaGenome actually does

The long-standing challenge in genomics is that only about 2% of human DNA directly encodes proteins. The remaining 98% mostly acts as a regulatory control system. These regions do not contain the protein blueprint itself, but they help determine when, where, and how strongly genes are used. A large share of hard-to-interpret human variants sits in this noncoding portion of the genome.

AlphaGenome is designed to read both layers at once: the blueprint and the control panel. In the paper, the model predicts gene expression, transcription initiation, chromatin accessibility, histone modifications, transcription factor binding, chromatin contact maps, and splicing-related outputs. Across variant-effect benchmarks, it matched or outperformed the strongest external methods in 25 of 26 evaluations.

In practical terms, many tools can tell you where a variant is located. AlphaGenome tries to tell you what that variant is likely to do to cellular regulation.

Why the Atlas matters

A model can be queried variant by variant, but the Atlas goes a step further by precomputing results across the full space of possible single-letter substitutions. The result is a dataset of about 1 petabyte. Each variant is paired not only with predicted molecular effects, but also with a summary score called the AlphaGenome Variant Impact, or AVI, score. That score is meant to help researchers prioritize which variants deserve attention first.

The practical value is easy to see. In official DeepMind examples, Broad Institute researchers used these predictions to investigate a rare genetic case and highlighted a DNM1 variant that was later backed up experimentally. In another example, researchers at the University of Exeter analyzed data from more than 54,000 UK Biobank participants and used the Atlas to shrink the candidate list dramatically: in one region, they narrowed 526 raw candidates down to just 4.

That is where a resource like this becomes powerful. When the candidate space is enormous, it can act as a hypothesis filter rather than a final answer.

What the Atlas does not do

The key limitation is that AlphaGenome Atlas does not prove a variant’s effect by itself. These are computational predictions, not direct experimental observations. A high score does not automatically mean a variant causes disease, nor does it guarantee the same effect in every biological context.

So the best way to think about the Atlas is as a very strong first-pass filter. It helps researchers reduce the search space and focus experiments on the variants most likely to matter. No laboratory can test 9 billion mutations one by one, but with AlphaGenome Atlas, scientists now have a map that makes that search much more targeted.