AlphaGenome Atlas: Google DeepMind's map with predictions for 9 billion human DNA variants

Google DeepMind has unveiled AlphaGenome Atlas, a platform bringing together precomputed predictions of the molecular effect of 9 billion single-nucleotide variants, that is, practically every single-"letter" change that can occur in the human genome. The aim is to solve a practical problem: testing the effect of each of those variants in the lab is unfeasible, so the company has opted to precompute at scale the predictions of its previously published AlphaGenome model and turn them into an open, searchable resource.

AlphaGenome is the artificial intelligence model that already made it possible to analyse how a specific genetic variant affects biological processes such as gene regulation. What Atlas adds is scale: instead of having to query the model variant by variant, it provides from the outset a complete map of the genome with thousands of molecular effect predictions for each variant, covering aspects of gene regulation across hundreds of human and mouse cell types and tissues. The resulting dataset takes up one petabyte, more than 30 times the size of the AlphaFold Database, the reference DeepMind had already built for protein structures (which in its 2022 expansion went from some 190,000 experimentally obtained structures to more than 200 million predicted structures).

To help researchers prioritise which variants are worth studying, Atlas introduces the AlphaGenome Variant Impact score (AVI), a single score that combines AlphaGenome's predictions with those of AlphaMissense, DeepMind's model specialising in protein-altering variants. The combination makes it possible to score and rank variants both in the coding regions of the genome (barely 2% of the total, the part that produces proteins) and in the non-coding regions (the remaining 98%, which regulates gene activity and where most variants associated with traits and diseases are concentrated). According to DeepMind, the AVI score achieves state-of-the-art results on several variant pathogenicity and rare disease benchmarks. In addition, each score is broken down into "feature attributions", which indicate the specific biological processes —such as RNA splicing, gene expression, chromatin accessibility or evolutionary conservation— that explain the detected impact, making it easier to interpret why a variant is relevant and not just how much. The package is rounded out with a catalogue of more than 2,500 recurring DNA motifs, short sequences that act as the "words" of the genome and that help, for example, to locate transcription factor binding sites.

AlphaGenome Atlas is already available free of charge for academic research through a web portal designed to be explored without any need to code, as well as via the AlphaGenome API and as a "skill" integrated into Google Antigravity, the company's agentic development platform. DeepMind also announces that access for commercial use will come to Google Cloud shortly, adding to the existing availability of the base AlphaGenome model on GitHub and in Model Garden.

The article details three real use cases with external collaborators that illustrate the kind of findings the resource enables. In the field of rare diseases, researchers from the GREGoR consortium, among them Laura Covill and Anne O'Donnell-Luria of the Broad Institute, used the AVI score to prioritise previously overlooked variants and located a variant in the DNM1 gene, strongly associated with epileptic encephalopathy; AlphaGenome's predictions revealed the exact mechanism, the creation of an incorrect splice site that causes an anomalous extension of the resulting protein, a finding later validated through experimental assays that also turned up nearby variants with similar effects.

In population genetics, Gareth Hawkes, a Medical Research Council researcher at the University of Exeter, applied Atlas to whole-genome data from more than 54,000 UK Biobank participants. By grouping rare variants according to their predicted molecular effect, he managed to identify 22% more genetic associations in non-coding regions that would otherwise be hidden by statistical noise. That approach allowed him to pinpoint specific regulatory variants linked to levels of proteins such as PLA2G7 (related to ageing) and EGLN1 (a cellular oxygen sensor). Extending the method to hundreds of millions of non-coding variants and focusing on the 1% that Atlas predicts to be most impactful, Hawkes identified 19 genetic regions potentially related to body mass index.

In more basic molecular biology, Julia Zeitlinger and Melanie Weilert, of the Stowers Institute for Medical Research, used Atlas to distinguish which transcription factors only modify DNA accessibility and which are also capable of switching genes on or off, an example of how the motif catalogue helps interpret regulatory mechanisms.

DeepMind frames Atlas as a continuation of the strategy it already followed with AlphaFold: turning a predictive model that is powerful but costly to query variant by variant into a precomputed database accessible through an intuitive portal, an approach that in its day gave a notable boost to structural biology research. The company explicitly presents this launch as a starting point rather than a final result: as models such as AlphaGenome improve, it expects genome maps to become more complete and accurate, and it envisages integrating these resources into broader agent-based scientific workflows, such as those offered by Antigravity.

The article itself includes an important warning that should be passed on without qualification: the information generated by AlphaGenome Atlas is not a substitute for professional medical advice, does not constitute diagnosis or treatment, and the model has not been validated or approved for any clinical use. It is therefore a basic and translational research tool whose value is for now measured in publications, candidate prioritisation and ongoing experimental validation, not in immediate clinical applications.