Google’s AlphaGenome Atlas Maps 9 Billion DNA Variants

Google DeepMind has turned a difficult genomics problem into a searchable database: AlphaGenome Atlas contains predictions for the molecular effects of about 9 billion possible single-letter changes in human DNA. The one-petabyte resource is designed to help researchers identify which mutations deserve laboratory investigation, particularly in the large portion of the genome that does not directly encode proteins.

AlphaGenome Atlas starts with the part of DNA scientists understand least

The human genome contains roughly 3 billion DNA letters, but only about 2% directly encode proteins. The remaining sequence is often described as non-coding DNA, although much of it helps regulate when, where and how genes operate. A single change in one of these regions can alter gene activity without changing a protein's underlying sequence, making such variants difficult to interpret.

AlphaGenome was developed to predict those regulatory effects from DNA sequence. AlphaGenome Atlas takes that model a step further by calculating predictions in advance across the human reference genome. Instead of asking the model to analyse one mutation at a time, researchers can search a prepared catalogue and inspect predicted effects almost immediately.

The atlas calculates three possible substitutions at every DNA letter

Each position in the reference genome contains one of four DNA bases: A, C, G or T. At any position, there are three alternative single-letter substitutions. Across roughly 3 billion positions, that produces around 9 billion possible single-nucleotide variants, meaning mutations involving just one DNA letter.

DeepMind used AlphaGenome to pre-compute predictions for these variants, producing a dataset of about one petabyte. That is roughly 1,000 terabytes of information, large enough that storing and searching the results efficiently becomes part of the engineering challenge itself. The practical benefit is that researchers do not need to repeatedly run the underlying model for every candidate mutation they want to investigate.

A single score gives researchers a way to prioritize mutations

Searching billions of predictions would still be difficult if every result required the researcher to interpret dozens of separate measurements. AlphaGenome Atlas therefore introduces the AlphaGenome Variant Impact score, or AVI score, which combines predicted effects across coding and non-coding regions into a ranking that can help researchers decide which variants deserve closer examination.

The score is not a diagnosis and it does not establish that a mutation causes disease. It is better understood as a filtering mechanism. If researchers have thousands of possible variants from a patient's genome or a population study, a ranking system can help move the most promising candidates toward experimental testing while lower-priority candidates receive less immediate attention.

The useful part is what happens after the prediction

The Atlas is already being used in research settings to narrow down genetic questions. DeepMind says researchers at the Broad Institute used the score to prioritize variants in rare-disease investigations, including a variant in the DNM1 gene that was predicted to create an abnormal splice site. Splicing is the process cells use to remove and join sections of RNA before producing mature messenger RNA, so a prediction that a mutation disrupts splicing can provide a specific biological mechanism for researchers to investigate.

That distinction matters. AlphaGenome Atlas does not replace laboratory work; it helps decide where laboratory work should begin. Researchers can use the predictions to select variants for experiments, compare candidate mutations, and investigate regulatory mechanisms that would otherwise require much more manual screening.

Why the non-coding genome makes this harder

Traditional genetic analysis has often been easier when a mutation changes a protein-coding sequence because the connection between the DNA change and the resulting protein can be relatively direct. Regulatory DNA is different. A mutation can affect gene expression, transcription-factor binding, chromatin accessibility or RNA splicing without changing the protein-coding sequence itself.

AlphaGenome was designed to model several of these biological signals together. Its published research describes predictions covering gene expression, transcription initiation, chromatin accessibility, histone modifications, transcription-factor binding, chromatin contact maps and splicing-related measurements. Combining these signals gives the Atlas more context than a system that evaluates only whether a DNA sequence resembles a known coding mutation.

The model is powerful, but its predictions still need biology

The biggest limitation is also the easiest one to misunderstand: a predicted effect is not the same thing as an experimentally demonstrated effect. AlphaGenome's published research showed strong performance across many variant-effect benchmarks, but the model still has limitations around cell-specific behaviour, complex genetic backgrounds and some long-range regulatory effects.

Real human genomes also contain more than isolated single-letter substitutions. They include insertions, deletions, structural changes and combinations of variants that can interact with one another. A model that evaluates one change against a reference genome cannot automatically explain every combination found in an actual person's genome. The Atlas does include predictions involving more than 100 million observed insertions and deletions, but that does not remove the broader challenge of interpreting complex personal genomes.

A one-petabyte database changes the research workflow

The most interesting part of AlphaGenome Atlas may therefore be less about the raw number of predictions and more about access. A researcher who previously needed to write software, run a computational model and wait for results can instead search a prepared resource through a research interface. That lowers the technical barrier between a biological question and a computational prediction.

There is a useful precedent in DeepMind's AlphaFold Database, which made hundreds of millions of predicted protein structures accessible to researchers. AlphaGenome Atlas applies a similar idea to genetic variation, but the underlying problem is different: instead of predicting the three-dimensional structure of proteins, it attempts to map how changes in DNA may alter the regulatory machinery that controls cells.

What AlphaGenome Atlas cannot tell researchers yet

The Atlas cannot by itself prove that a mutation causes a disease, determine how a patient will respond to treatment, or replace clinical genetic interpretation. Its predictions are computational evidence that can guide further investigation. Experimental validation remains necessary when researchers need to establish what a particular variant actually does in a biological system.

That limitation is important because genetic regulation depends on biological context. The same sequence change can behave differently depending on the cell type, developmental state and surrounding molecular environment. A genome-wide prediction system can identify promising signals, but biology still has the final word.

The next step is turning a map into tested biology

AlphaGenome Atlas gives researchers something they previously did not have at this scale: a precomputed prediction for essentially every possible single-letter substitution across the human reference genome. Its value will ultimately be measured not by the size of its database, but by whether those predictions help researchers find mutations, regulatory mechanisms and disease pathways that can be confirmed experimentally.

That makes the Atlas less like a finished catalogue of genetic truth and more like a very large research map. The difficult work now moves to the next stage: choosing the predictions that matter, testing them in cells and organisms, and determining which computational signals consistently correspond to real biological effects.