Google DeepMind has announced AlphaGenome Atlas, a computational catalogue containing predicted molecular effects for approximately 9 billion possible single-letter changes in human DNA. The resource is intended to help academic researchers prioritise genetic variants for further investigation, including variants in non-coding DNA that can influence how genes are regulated.

The Atlas precomputes thousands of molecular-effect predictions for each single-nucleotide variant (SNV) across aspects of gene regulation and across hundreds of human and mouse cell types and tissues. An SNV is a change involving one nucleotide, or “letter”, in a DNA sequence.

The company says the Atlas is available for non-commercial use through a website portal and the AlphaGenome API. Commercial use through Google Cloud is planned, but the announcement does not specify a date or the terms.

Contents

What AlphaGenome Atlas contains

The Atlas is designed as a precomputed reference rather than a tool that requires researchers to run a new prediction for every variant. For each of approximately 9 billion single-nucleotide changes, it provides predictions related to molecular processes involved in gene regulation.

Coding DNA contains instructions used to make proteins, while non-coding DNA can regulate when and where genes are active. The announcement describes coding regions as approximately 2% of the genome and non-coding regions as the remaining 98%.

That distinction matters because the source describes non-coding regions as containing most trait-associated variants. A variant in non-coding DNA may affect a regulatory sequence, for example, changing the amount of RNA or protein produced from a gene. A catalogue that assigns predicted molecular consequences to these variants could help researchers decide which ones merit laboratory testing.

The resource also includes more than 2,500 recurring DNA sequence motifs and their locations. The announcement does not provide the complete list of molecular outputs or the accuracy of each output across different tissues, cell types and variant classes.

Google DeepMind describes the Atlas as a 1-petabyte dataset, more than 30 times larger than the AlphaFold Database. The size figure does not provide the benchmark metrics needed to assess prediction quality.

How the variant-impact score works

AlphaGenome Atlas includes an AlphaGenome Variant Impact, or AVI, score. The score combines predictions from AlphaGenome and AlphaMissense into a single number for each variant, allowing variants to be ranked for further analysis.

The score is accompanied by feature attributions. These are intended to indicate which predicted biological features contributed to a variant’s score. The announcement gives chromatin accessibility, RNA splicing and gene expression as examples.

RNA splicing is a step in gene expression in which segments of an RNA transcript are joined and removed to produce mature RNA. A variant that disrupts a splice site can change the resulting transcript and, in some cases, alter the protein made from it.

These outputs can make a prioritisation score more interpretable than a single unexplained ranking. However, an AVI score remains a model-generated prediction. It is not, by itself, experimental evidence that a variant causes a disease or affects a trait.

Google DeepMind says its testing found “best-in-class” AVI performance across multiple variant-pathogenicity and rare-disease benchmarks. The supplied announcement does not include the benchmark datasets, numerical performance metrics, confidence intervals or direct comparisons needed to independently assess that claim.

Reported research uses

The announcement describes several uses by external collaborators, and the findings below are reported by Google DeepMind rather than independently established by the supplied evidence.

In work with the GREGoR Consortium, Google DeepMind reports that researchers investigating unsolved rare disease identified a DNM1 variant associated with epileptic encephalopathy. According to the company, AlphaGenome predicted that the variant created an incorrect splice site, and experimental screens validated that prediction. This is an example of a prediction being followed by laboratory testing, not evidence that the full Atlas has been experimentally validated.

Google DeepMind reports that a University of Exeter researcher applied the Atlas to whole-genome data from more than 54,000 UK Biobank participants. The announcement says that grouping rare variants by their predicted molecular effects produced 22% more non-coding genetic associations.

According to Google DeepMind, the same analysis identified regulatory variants associated with protein levels, including variants related to PLA2G7 and EGLN1. The company also reports that a BMI-focused analysis identified 19 genetic regions among the 1% of non-coding variants predicted to be most impactful.

These reported examples suggest possible uses in variant prioritisation and genetic-association research. They do not, based on the information supplied, establish that the predictions improve clinical diagnosis, prevent disease or lead to an effective treatment. The announcement also does not provide enough methodological detail to independently evaluate the reported 22% increase or the 19 BMI-associated regions.

What the Atlas does not establish

AlphaGenome and AlphaGenome Atlas have not been validated for, and are not approved for, clinical use, according to Google DeepMind’s medical disclaimer. The resource should therefore be understood as a research tool rather than a diagnostic system.

The announcement describes a computational resource, model benchmarking and selected collaborator analyses, not a single clinical trial. The UK Biobank work involved more than 54,000 participants, but the supplied evidence does not give a formal study design, comparator or effect estimates for that analysis. Other reported uses do not include sample sizes in the announcement.

The Atlas also does not eliminate the need for laboratory investigation. A prediction that a variant affects splicing, gene expression or another molecular feature can help researchers select experiments, but the biological effect still needs to be tested. Even experimental confirmation of a molecular effect would not automatically establish the variant’s contribution to a patient’s disease or to a population-level trait.

The announcement’s claims about broad coverage, benchmark performance and research impact come from Google DeepMind. The source does not provide the full methods or independent evidence needed to assess how prediction quality varies across tissues, cell types, variant classes and rare-disease contexts.

For now, AlphaGenome Atlas is best characterised as a large, searchable prediction resource for academic genomics research. Its practical value will depend on how accurately its scores identify biologically important variants and whether independent studies show that those predictions improve downstream experiments or genetic interpretation.

Sources