TL;DR: Google DeepMind precomputed predicted molecular effects for all 9 billion possible single-letter changes in the human genome and released them as AlphaGenome Atlas with a single-number impact score.

Google DeepMind introduced AlphaGenome Atlas on September 8, 2026, as a platform holding predictions for the effects of 9 billion single-nucleotide variants, or every single-letter change possible in the human genome, according to Google DeepMind. The company described the collection as the most comprehensive catalogue of how genetic mutations affect molecular biology. Academic researchers can explore it through a free-to-use website portal, according to Google DeepMind. Ars Technica reported the announcement on September 9, 2026, framing it as an attempt to predict the consequences of every possible single-base variant.

What Google released on September 8

AlphaGenome Atlas starts from AlphaGenome, an artificial intelligence model that can predict how genetic variants affect biological processes, according to Google DeepMind. That model was useful for examining particular variants and had found widespread use in research, but the team wanted to show a genome-wide view, according to Google DeepMind. The new step was to precompute those predictions at scale and make the result directly browsable, according to Google DeepMind.

The Verge reported that the platform contains a predictive map of every possible DNA letter change in the human genome. The same report said Atlas holds forecasts for how each of the nine billion variants could affect the body at a molecular level, such as changing how much of a particular protein is produced. Google DeepMind said the resource charts molecular effects across the genome in the manner of an atlas linking landscape features.

Access was opened through three routes, according to Google DeepMind. Those routes are an intuitive website portal, an AlphaGenome programming interface, and a skill in Google Antigravity. The Verge reported that scientists can explore the predictions through a web portal, through the AlphaGenome interface, and as a skill in the agentic development platform Antigravity. The company said Atlas was available to researchers for noncommercial use through its website starting September 8, with commercial use on Google Cloud to follow soon, according to The Verge.

Why 9 billion is the number

Human DNA uses four chemical letters, usually shortened to A, C, G and T, and the human genome contains roughly three billion letter pairs, according to The Verge. Those letters carry instructions for how, when and where genes are switched on and off and to what degree, according to The Verge. Because each position can change to the three letters absent from the reference, the total search space is about nine billion single-letter substitutions, according to The Verge.

Ars Technica explained the arithmetic in similar terms. The human genome is about 3 billion bases long, so testing the other three DNA bases missing at each reference position means sending a total of 9 billion cases through AlphaGenome software, according to Ars Technica. The process moves position by position: if the first base on chromosome 1 is an A, the system evaluates replacement with G, C or T, then repeats at the next base, according to Ars Technica.

Testing that many changes in the laboratory is practically impossible, according to Google DeepMind. In a press briefing, DeepMind genomics lead Ziga Avsec acknowledged that the underlying AlphaGenome model had already been released and said building the genome-wide catalog took additional time, according to The Verge. “Basically it took us some time to really precompute and also analyze this many variants because the space is so big,” he said, according to The Verge.

The hard part is non-coding DNA

AlphaGenome is designed to identify potential functions of non-coding DNA, which does not encode proteins but makes up the vast majority of the human genome, according to Ars Technica. Part of that non-coding DNA controls the protein-coding portion by indicating where and when cells should make messenger RNAs and how those RNAs should be processed into mature protein-coding forms, according to Ars Technica. Another part looks like leftover viruses and other molecular parasites, according to Ars Technica.

The protein-coding share is small. The portion of the human genome that encodes proteins is less than 3 percent, according to Ars Technica. Google DeepMind put the coding share at 2%, with the remaining 98% in the non-coding category that orchestrates gene activity and houses most trait-associated variants. The Verge reported that Atlas extends predictions across the genome, including the large stretches that do not directly code for proteins but can control how genes behave.

Non-coding DNA contains a mixture of functional elements. It includes centromeres, which help ensure chromosomes are divided evenly between cells, and caps that protect chromosome ends, according to Ars Technica. It also includes regulatory DNA that controls gene activity, signals that help determine what should and should not be included in mature messenger RNAs, and sequences that help control how DNA is packaged inside the cell, according to Ars Technica.

Much of it has no apparent present-day role. Most non-coding DNA is junk, including remains of ancient viral infections, DNA-level parasites and genes inactivated by mutation, according to Ars Technica. Distinguishing useful from non-useful sequences has remained an ongoing challenge for biologists, according to Ars Technica.

Why prediction is difficult

Proteins that interact with DNA are often not highly selective about the sequences they attach to, potentially binding at random across the genome and tolerating some mutation, according to Ars Technica. Many of those proteins are also specific to cell type, with different DNA-binding populations in liver cells, nerve cells, immune cells and other lineages, according to Ars Technica. In many cases, the presence of many different protein binding sites in a compact space matters more than any single site, according to Ars Technica.

Researchers previously built software tools that identify individual sites of interest in non-coding DNA, according to Ars Technica. Google’s approach treats the task as a probability and context problem suited to AI, and the company developed the AlphaGenome AI system to evaluate sequences for potential function, according to Ars Technica. AlphaGenome was trained using public databases of human and mouse genomes, allowing it to learn patterns between DNA changes and biological processes, according to The Verge.

The system remains bounded in scope. It is currently limited to sequences from mice and humans, according to Ars Technica. It was also trained on only a limited number of cell types that biologists have studied exhaustively, according to Ars Technica. Even so, it can help when a researcher studying a gene finds changes in non-coding regions and needs a hint about likely significance plus a hypothesis about mechanism, according to Ars Technica. Its predictions are generally as good as or better than those from specialized software tools, according to Ars Technica.

What AlphaGenome tries to forecast

For specialists, Ars Technica listed the molecular readouts AlphaGenome attempts to identify as “gene expression, transcription initiation, chromatin accessibility, histone modifications, transcription factor binding, chromatin contact maps, splice site usage, and splice junction coordinates and strength.” Google DeepMind said Atlas contains thousands of molecular effect predictions for each variant, covering multiple aspects of gene regulation across hundreds of human and mouse cell types and tissues.

Those detailed forecasts are the starting point for the other Atlas layers, according to Google DeepMind. The layers allow researchers to link variants directly to the functional DNA sequences they disrupt, according to Google DeepMind. The company also compiled more than 2,500 recurrent DNA sequences, described as the words of the genome, along with their locations, according to Google DeepMind.

The project builds on two earlier DeepMind efforts, according to The Verge. One is AlphaGenome, an AI model unveiled the prior year to help scientists identify genetic drivers of disease. The other is AlphaMissense, an earlier tool focused on predicting which small mutations might alter proteins, according to The Verge. Atlas extends beyond protein-altering changes to cover regulatory regions as well, according to The Verge.

The AVI score and feature breakdown

To help scientists find the most consequential changes, Google DeepMind released the AlphaGenome Variant Impact score, known as AVI, according to Google DeepMind. The AVI combines AlphaGenome with AlphaMissense, the company’s model for predicting effects of protein-altering DNA variants, and condenses both models’ predictions into a single number, according to Google DeepMind. “Now, researchers can rapidly rank variants and interpret their molecular effects at the same time,” the company said, according to The Verge.

The single-number design is intended for rapid ranking. AVI helps researchers score and order variants by potential impact, according to Google DeepMind. Crucially, it applies to both coding and non-coding regions, according to Google DeepMind. That breadth matters because non-coding DNA orchestrates gene activity and contains most trait-associated variants, according to Google DeepMind.

Each AVI value is linked to an explanation. Every AVI score connects to distinct biological features driving it, such as aspects of gene regulation predicted by AlphaGenome or the protein-impact score from AlphaMissense, according to Google DeepMind. Google DeepMind also calculated AVI feature attributions that highlight which molecular processes, such as RNA splicing or gene expression, are predicted to be most disrupted by each variant.

Google DeepMind said testing showed AVI delivered best-in-class performance across many variant pathogenicity and rare-disease benchmarks. The company also said extensive community collaboration guided the design of AlphaGenome Atlas, according to Google DeepMind.

Scale: a 1-petabyte catalogue

Applying AlphaGenome predictions to billions of possible variants produced a massive dataset that Google described as roughly 1 petabyte in size, according to The Verge. Google DeepMind said AlphaGenome Atlas is a 1-petabyte dataset more than 30 times larger than the AlphaFold Database.

The comparison reaches back to 2022, when expansion of the AlphaFold Database grew available 3D-structure information from around 190,000 experimental structures to more than 200 million structure predictions, covering nearly all catalogued proteins known to science, according to Google DeepMind. That database offered a portal usable by researchers with no coding experience, with intuitive visualizations and support for large-scale protein-structure analysis, according to Google DeepMind. It quickly became an important resource across the life sciences, according to Google DeepMind.

Atlas follows the same accessibility idea for variant effects. Google DeepMind said it aimed to make predictions more accessible and give scientists an intuitive way to explore a vast dataset. Together, the molecular forecasts, AVI scores, attributions and motif collection are meant to support tasks from rapid variant ranking to detailed functional analysis, according to Google DeepMind.

Early use in rare-disease and trait work

Google DeepMind said trusted external collaborators had already used AlphaGenome Atlas to identify and experimentally verify key variants in unsolved rare-disease research. The same collaborators also found rare variants associated with common traits, according to Google DeepMind. The company presented those cases as early evidence that the precomputed resource could speed follow-up experiments, according to Google DeepMind.

The broader context is a longstanding genetics bottleneck. Changes to individual DNA letters can be harmless, contribute to ordinary differences between people, or play a role in disease, and a major challenge is determining which changes matter and how, according to The Verge. DNA letters contain instructions for how, when and where genes are switched on and off and to what degree, according to The Verge. Atlas is intended to help researchers focus on mutations that warrant closer attention, according to The Verge.

Atlas is the latest in a series of Google AI efforts aimed at science and medicine, at a time when DeepMind co-founder Demis Hassabis was stepping back from running the AI laboratory to concentrate on scientific research, including leading drug-discovery spinoff Isomorphic Labs, according to The Verge. The company’s best-known prior work in the area is AlphaFold, the protein-structure prediction model that earned Hassabis and John Jumper the 2024 Nobel Prize in Chemistry, according to The Verge. The company has also developed AI tools for weather prediction, computing and mathematics solutions, and an agentic co-scientist for researchers, according to The Verge.

What the catalogue does not mean

Ars Technica cautioned that almost nobody carries the reference genome used as the baseline. Each person likely differs from it at millions of bases and most people probably carry combinations of insertions, deletions, duplications and flipped sequences, according to Ars Technica. Many of those differences have no consequence because they fall in sequences without functional significance, according to Ars Technica. At the same time, many non-functional sequences are identical across humans because insufficient time has passed since a common ancestor for changes to accumulate, according to Ars Technica.

Only a small fraction of the changes Google evaluated are therefore likely both to appear in a real human genome and to have functional significance, according to Ars Technica. The practical value, in that account, is that the work pre-calculates mutations researchers might encounter, so a variant found by genome sequencing comes with an immediate sense of possible consequences, according to Ars Technica. It also lets researchers search genome-wide for changes with a particular type of impact, according to Ars Technica.

Ars Technica also noted a limitation tied to training data. Much of the information could have been obtained by consulting the ENCODE dataset that was among the model’s training sources, according to Ars Technica. The larger opportunity would be reliable analysis where no training data exist, such as Neanderthal and Denisovan genomes or cell types lacking ENCODE data, according to Ars Technica. Whether the system is ready for that step was not yet clear, according to Ars Technica. The same report said it would remain unclear what AlphaGenome adds beyond its training data until biologists use it heavily.

Sources referenced