DeepMind Releases AlphaGenome Atlas: A Map of the Effect of Every Single-Letter DNA Substitution

Google DeepMind has published AlphaGenome Atlas, a "predictive map of every possible single-letter substitution in the human genome." The genome contains about 3.1 billion base pairs, and each position has three possible alternative nucleotides. That gives an order of magnitude of roughly 9.3 billion hypothetical mutations. This is our arithmetic, not a figure from the release. The description of the Two Minute Papers video, where the news came from, contains only the title of DeepMind's post and links to it, so details of the methodology and metrics need to be checked in the primary source.

What Atlas is and where it came from

Atlas is built on AlphaGenome, a DeepMind model the company announced in 2025. According to that announcement, the model takes up to 1 million base pairs as input and predicts molecular properties of DNA at single-letter resolution: gene expression levels, chromatin accessibility, and splicing. To assess a mutation, the model compares its predictions for the original and the altered sequence. The difference is the presumed effect.

Atlas, judging by its name, turns this process into a ready-made reference. Researchers don't need to run the model on every variant. They look up a substitution in the precomputed map and see which regulatory signals it is predicted to shift.

Why precomputation makes sense

Running a megabase-scale window through a large neural network for billions of variants requires a serious GPU budget. No lab, let alone a clinic, would do this for every query. Computing once and distributing the results is the more economical option. This is our conclusion from the scale of the task; in the materials available to us, DeepMind does not state the cost of the computation.

The task itself is an important one. Only a small part of the genome, about 2%, codes for proteins. The remaining variants lie in regulatory regions, and their effect on a person is usually unclear. Clinical genetics has a separate category for such findings: "variants of uncertain significance." A map of predicted effects offers a way to prioritize candidates before spending money on laboratory validation.

• Developer: Google DeepMind
• Scope: single-letter substitutions (SNVs) in the human genome
• Scale: ~3.1 billion positions × 3 alternatives ≈ 9.3 billion variants (estimate)
• Base model: AlphaGenome, context of up to 1 million base pairs, single-letter resolution (per the 2025 announcement)
• Prediction types: gene expression, chromatin, splicing, and other regulatory signals
• Format: a precomputed map, not a model you have to run yourself

Limitations

A prediction remains a prediction. The model is trained on experimental data, a significant share of which comes from cell lines and tissues, and for an individual patient a shift in signal does not equal disease. Atlas is not a diagnostic tool. For now it looks like a filter for researchers. The metrics DeepMind reports for validating the map, and the datasets it was tested on, should be checked in the blog and accompanying materials.

How to use it

Geneticists and bioinformaticians can reasonably use Atlas as a preliminary filter. Take a list of variants from sequencing or a GWAS, sort it by predicted regulatory effect, and manually check only the top of the list. Cross-check the results against databases such as ClinVar and against experimental data.

For biotech teams, the map is useful for finding regulatory targets. ML engineers should look closely at the pattern itself: when the input space is finite but enormous and the model is expensive at inference, it pays to compute the whole atlas once and release it as data. Before using it, check the license and access terms: in 2025, AlphaGenome's were limited to non-commercial use, and Atlas's may differ.