Google DeepMind's AlphaGenome Atlas Maps 9 Billion DNA Variants

On September 8, 2026, Google DeepMind released AlphaGenome Atlas, a precomputed, searchable catalog of molecular-effect predictions for every possible single-nucleotide variant (SNV) in the human genome — roughly 9 billion distinct changes. The dataset spans 1 petabyte, more than thirty times the size of the AlphaFold database, and is freely accessible to non-commercial researchers through a zero-code web portal at alphagenome.google/atlas.

AlphaGenome Atlas is the productization of AlphaGenome, a deep-learning model first published as a preprint in June 2025 and formally in Nature in January 2026 (vol. 649, pp. 1206–1218). Where the model required users to write code and submit queries one variant at a time, the Atlas precomputes every possible SNV en masse. As Pushmeet Kohli, VP Science at Google DeepMind, told Scientific American, it represents the first time any researcher can access a comprehensive map of the human genome and its variations by simply opening a browser.

What the Atlas predicts

For each of the ~9 billion SNVs, the Atlas predicts effects across 11 molecular modalities: gene expression, transcription initiation, RNA splicing (including splice-site usage and junction patterns), chromatin accessibility, histone modifications, transcription-factor binding, chromatin contact maps, and polyadenylation. These predictions span hundreds of human and mouse cell types and tissues, offering a multidimensional view of how a single base change ripples through the molecular machinery of the cell.

A key addition beyond the original model is the AVI (Aggregate Variant Impact) score, which integrates AlphaGenome's non-coding predictions with AlphaMissense's coding-region pathogenicity scores into a single ranking metric. AVI feature attributions decompose that score into interpretable biological contributions — splicing, chromatin accessibility, conservation — so researchers can quickly prioritize which variants merit deeper investigation. The Atlas also includes a de novo DNA motif collection of over 2,500 recurring sequence patterns across the genome.

From computation to lookup

The practical shift is stark. Previously, researchers wanting AlphaGenome predictions had to run computationally expensive queries; now they can look up any variant in seconds. Jonathan Sebat, a psychiatric geneticist at UC San Diego, described the change in Scientific American: researchers no longer need to compute anything — they can literally look up everything.

Early applications have already produced results. At the Broad Institute, a team used AVI scoring to prioritize variants in an unresolved epileptic encephalopathy case, identifying a key variant in the DNM1 gene that creates an erroneous splice site, causing abnormal protein extension — a prediction subsequently confirmed experimentally. At the University of Exeter, researchers applied the Atlas to whole-genome data from over 54,000 UK Biobank participants, discovering 22 percent more non-coding genetic associations and identifying 19 BMI-associated genetic regions previously undetected.

Julia Zeitlinger, a molecular biologist at the Stowers Institute for Medical Research, described the Atlas as a searchable dictionary for non-coding DNA — the vast regulatory regions of the genome that determine when and where genes are active, and which have remained largely opaque to interpretation.

Architecture and scale

AlphaGenome processes DNA sequences up to 1 megabase in length through a U-Net-inspired hierarchical architecture combining an encoder, a Transformer with cross-device communication for long-range dependencies, and a decoder. The model produces predictions at single-base-pair resolution for most outputs across 5,930 human genomic tracks. Training follows a two-stage pipeline: four-fold cross-validation generates fold-specific models and an all-fold teacher ensemble, followed by knowledge distillation into a single student model that runs inferences in under one second on an NVIDIA H100 GPU.

The Atlas also covers over 100 million short insertions and deletions (indels) observed in human genomes, extending its scope beyond single-base changes.

Context: from protein folding to genome reading

AlphaGenome Atlas arrives as the natural sequel to DeepMind's AlphaFold, which predicted the 3D structures of over 200 million proteins and earned the 2024 Nobel Prize in Chemistry for Demis Hassabis and John Jumper. Where AlphaFold addressed the protein layer, Atlas addresses the DNA regulatory layer — the genomic dark matter that determines how, when, and where those proteins are produced. DeepMind frames the two as complementary halves of a broader ambition to make biology computable.

The scale is unprecedented. At 1 PB, the Atlas is over thirty times the size of the AlphaFold database. Approximately 9,000 researchers had already been using AlphaGenome through its API before the Atlas launch, according to Nature News.

Limitations and caveats

The Atlas is not a clinical tool. DeepMind explicitly states that AlphaGenome has not been validated or approved for any clinical use. The accompanying paper for the Atlas itself is currently a preprint and has not yet undergone peer review, though the underlying AlphaGenome model was peer-reviewed and published in Nature.

The model was trained exclusively on human and mouse data; its generalizability to other species is unknown. Predictions remain computational inferences — not experimental measurements — and require wet-lab confirmation before any clinical or therapeutic decision. Martin Kircher, a bioinformatician at the Max Delbrück Centre for Molecular Medicine in Berlin, told Nature News that the Atlas is a useful and generous way to scale up access to a strong model, while Mafalda Dias and Jonathan Frazer of the Centre for Genomic Regulation in Barcelona noted that removing the computational barrier makes it a valuable resource.

Commercial licensing terms have not yet been disclosed; DeepMind says a commercial path through Google Cloud is forthcoming, which may shape how broadly the pharmaceutical and biotech industries adopt the tool.

Analysis

The significance of AlphaGenome Atlas lies less in a single algorithmic leap — the underlying model was already published — and more in the decision to precompute and democratize access. By converting a computationally expensive query into a free, instant lookup, DeepMind effectively removes the primary bottleneck between AI prediction and widespread scientific use. Whether this parallels the AlphaFold trajectory — where open access catalyzed thousands of downstream studies — will depend on how quickly researchers trust these predictions enough to act on them, and on whether the preprint's claims survive peer review. The early real-world results from Broad and Exeter suggest that trust is already forming in selected use cases. The open question is whether that extends from rare-disease diagnostics into the far messier world of common-disease genetics, where the signal-to-noise ratio is lower and the stakes of over-interpretation are higher.