The short answer

Google DeepMind has published what it calls AlphaGenome Atlas: a "predictive map of every possible DNA letter change in the human genome." The catalogue covers 9 billion variants and the dataset runs to roughly 1 petabyte.

But a distinction has to be set down at the outset: this is not a catalogue of measurements but of predictions. Every row in it is a model's estimate that "if this letter changes, this is what happens at the molecular level." This piece explains both what the map is and why that distinction decides everything.

First the basics: what is a genome, what is a variant?

DNA is written in four chemical "letters" — A, C, G and T. The human genome contains roughly 3 billion letter pairs. Those letters carry the instructions that make life work: how, when, where and to what degree genes are switched on and off.

A change to a single letter is called a variant. A variant can be one of three things: entirely harmless, a contributor to ordinary differences between people, or a factor playing a role in disease. One of the field's central difficulties is exactly this: telling which change matters, and how.

Why is the scale so hard? At each of 3 billion positions the letter can become one of three other things. That means roughly 9 billion possible single-letter substitutions. Measuring all of them experimentally is impossible — which is where prediction comes in.

Why 9 billion? Understanding the number

The figure looks arbitrary at first, but it can be computed directly. There are 3 billion positions in the genome, and each holds one of four letters. If you want to change that letter, three options remain. Three times three billion is nine billion.

The arithmetic also shows the catalogue's limit: Atlas covers single-letter substitutions. Deleting a letter, inserting one, duplicating a section, or chromosome-level rearrangements all sit outside those 9 billion. Some genetic diseases arise from precisely those kinds of change.

So "every possible change" technically means "every possible single-letter substitution." The scope is still vast, but knowing the boundary is what lets you understand why something you were looking for is not in the catalogue.

What does "molecular effect" mean?

What Atlas predicts is not disease but a molecular outcome. That difference is the most easily skipped part of the story.

The model says: if this letter changes, the expression of a particular gene shifts this way, or this much less of that protein is produced. What it does not say is whether that will cause disease in a person. A molecular change may produce no observable consequence at all; the body often has compensating mechanisms. Or the consequence may appear only at a certain age, in combination with a particular environmental factor.

That chain — letter → molecular effect → cellular consequence → tissues → disease — has four steps, and Atlas predicts only the first. When a headline says "AI has mapped the causes of disease," what has been skipped are the three steps in between.

What does Atlas contain?

Atlas holds predictions for how each of those 9 billion variants could affect the body at a molecular level — for instance changing how much of a particular protein is produced. The researchers describe it as "the most comprehensive catalogue of how genetic mutations affect molecular biology."

There are three ways in: a web portal, a skill inside Google's agentic development platform Antigravity, and the AlphaGenome interface.

To stop researchers drowning in billions of possibilities, Google is also releasing a Variant Impact Score (AVI). Drawing on the company's other models for predicting the effects of DNA changes, it lets researchers rapidly rank variants and interpret their molecular effects at the same time.

The real advance: non-coding regions

Atlas builds on a model called AlphaGenome, which DeepMind introduced last year to help identify the genetic drivers of disease. An earlier tool, AlphaMissense, focused on predicting whether small mutations might alter proteins.

Atlas differs in scope: it extends predictions across the genome, including the vast stretches that do not directly code for proteins. That distinction matters, because most of the genome carries no protein recipe; instead it controls how genes behave. These regions, long dismissed as unimportant, are now known to play a decisive role in disease — and they are exactly the hardest to study experimentally.

What was the model trained on?

AlphaGenome was trained on public databases of human and mouse genomes, which let it learn patterns between DNA changes and biological processes. Applying those predictions to billions of possible variants produced what Google says is a dataset of roughly 1 petabyte.

DeepMind's genomics lead, Ziga Avsec, acknowledges the underlying model — AlphaGenome — had already been released, but says turning its capabilities into a genome-wide catalogue took time: "Basically it took us some time to really precompute and also analyze this many variants because the space is so big."

How reliable is this?

The most important reading rule here: a prediction is not a measurement. None of the 9 billion rows in Atlas was measured in a laboratory; each is an estimate produced by a model from patterns it learned.

That has practical consequences. The model probably produces more reliable predictions for regions well represented in its training data, and less reliable ones for understudied regions — yet both kinds of row look identical in the catalogue. The training data also comes from public human and mouse genomes, and population representation in those databases has historically been uneven, weighted towards samples of European ancestry. A prediction map carries the biases of the data it was built on.

So Atlas's function is not to be a catalogue of answers but to narrow down which questions are worth asking. Experimental validation is still needed; Atlas says where to point the experiment.

What is a prediction map good for in practice?

The concrete use runs like this. A patient has a condition of unknown cause and their genome is sequenced. The result yields thousands of variants specific to that person — most entirely harmless. The clinical geneticist's problem is choosing which one to look at.

Today that choice is made by consulting databases of known disease genes and by asking whether the variant disrupts a protein. Variants in non-coding regions mostly fall into the "significance unknown" box. Atlas's promise is to shrink that box: it offers a molecular prediction and a priority score for those variants too.

The second use is on the research side. If a genomic region associated with a disease is known but not which variant inside it is responsible, testing them all experimentally is expensive. Atlas ranks which variant to start with. The gain is not new knowledge but a shorter experimental queue.

When should you be sceptical?

Three questions help when reading announcements like this:

  • Prediction or measurement? Was the catalogue produced in a lab or by a model? For Atlas the answer is clear: by a model.
  • Where is the validation? Is the model's accuracy on known outcomes published separately, or is only the scope announced?
  • Which population? For someone outside the groups represented in the training data, the reliability of a prediction is not the same.

These three questions are not specific to Atlas; they apply to every large scientific catalogue produced with AI. The value of a resource is also measured by how clearly it states its limits — and DeepMind's genomics lead says plainly that the catalogue consists of precomputed predictions.

Who can use it?

  • Researchers can access it from today for noncommercial use through Google's website.
  • Access for commercial use will open on Google Cloud "soon."
  • The Variant Impact Score makes it possible to prioritise among billions of possibilities.
  • Being offered as a skill inside Antigravity means it can plug directly into agentic workflows.

In the wider picture

Atlas is the latest in Google's efforts to attack core problems in science and medicine with AI. Its best-known work in the area is AlphaFold, the protein-structure prediction model that won Demis Hassabis and John Jumper the 2024 Nobel Prize in Chemistry. The company has also built tools for weather forecasting, for finding new solutions in computing and mathematics, and an agentic "co-scientist."

The timing is meaningful too: DeepMind cofounder Demis Hassabis is stepping back from running the lab to focus on scientific research, including leading the drug-discovery spinoff Isomorphic Labs.

In summary

What matters about AlphaGenome Atlas is not a single discovery but scale: a question that could previously be asked one at a time has now been asked 9 billion times at once, with the answers made searchable.

But the map itself is not a discovery; it is where discovery starts. The phrase "a map of every possible DNA letter change" is accurate; the missing adjective is predicted. That adjective is already in the original sentence, and it is exactly the word that tends to fall out of coverage like this.