Google DeepMind has released a predictive map charting the effects of all 9 billion possible single-letter DNA mutations across the human genome, Nature reported. The resource, named the AlphaGenome Atlas, is freely available for non-commercial research.
The AlphaGenome Atlas spans one petabyte of data. DeepMind built the resource by computing predictions for each of the three possible nucleotide substitutions at every position across the 3-billion-letter human genome. The dataset also catalogs more than 100 million short insertions and deletions observed in human populations. The project expands on DeepMind's AlphaGenome model, which the London-based company introduced last year.
Around 9,000 researchers had accessed AlphaGenome predictions through an application programming interface before this release, according to DeepMind product manager Dhavi Hariharan. Using the interface required researchers to write software code, which presented a barrier for some laboratory biologists. DeepMind structured the new atlas after its AlphaFold database, which hosts more than 200 million protein structure predictions.
To assist researchers sorting through the volume of data, the development team created the AlphaGenome Variant Impact (AVI) score. The single numerical score indicates the likelihood that a variant alters human biology. In a preprint paper, researchers reported that the AVI score separated disease-causing mutations from benign variants in a clinical genomics database. Investigators at the Broad Institute in Cambridge, Massachusetts, used the score to identify a non-coding mutation as a candidate cause of severe epilepsy.
External specialists cautioned that computational predictions do not eliminate laboratory work. Martin Kircher, a bioinformatician at the Max Delbrück Centre for Molecular Medicine in Berlin, said the atlas cannot replace bench experiments or case-by-case medical assessments, though he noted it expands access to a powerful model. Mafalda Dias and Jonathan Frazer, computational biologists at the Centre for Genomic Regulation in Barcelona, told Nature that running full-genome models was previously unfeasible for most rare-disease teams due to heavy computational requirements.
DeepMind's team also used the atlas to analyze short functional sequences known as motifs throughout the human genome. These segments regulate messenger RNA production or bind transcription factors that control gene expression. Julia Zeitlinger, a molecular biologist at the Stowers Institute for Medical Research and preprint co-author, said mapping these motifs and inferring their behavior across cell types provides a searchable dictionary for non-coding DNA.
