Twenty years ago, the completion of the Human Genome Project gave science a text of three billion letters. It seemed the main task was done — the book of life had been read. In reality, the difficulties were only beginning: understanding what each substitution of a single letter means and how it changes the workings of cells turned out to be far harder than reading the sequence.
On 8 September 2026, DeepMind presented Atlas AlphaGenome — a database with a volume of about one petabyte. It contains precomputed molecular consequences of more than nine billion possible single-nucleotide substitutions in the reference human genome. That is thirty times larger than the volume of the AlphaFold database. Access to the resource is open free of charge for scientific purposes through a web portal, an API and integrations with other tools.
The AlphaGenome model, published in Nature earlier this year, is capable of analysing up to a million bases at a time and predicting the effect of a variant on chromatin accessibility, transcription factor binding, histone modifications, RNA splicing and gene expression levels. Atlas turns these predictions into a ready-made reference: instead of running heavy computations every time, researchers receive ready results and an additional integral score, AVI. It combines the assessments of AlphaGenome and AlphaMissense, helping to rank variants by potential impact in both coding and non-coding regions.
Its practical value has already been confirmed in real studies. In collaboration with the GREGoR Consortium and the Broad Institute, AVI helped identify a variant in the DNM1 gene in a child with undiagnosed epileptic encephalopathy: the model pointed to the creation of an abnormal splice site, which experiments later confirmed. In work with the UK Biobank, the use of Atlas predictions increased the number of detected non-coding associations by 22 percent and made it possible to narrow hundreds of candidates down to a few key regulatory variants affecting plasma protein levels and body mass index.
Atlas not only ranks variants but also identifies more than 2500 repetitive short DNA motifs — the "words" of the genome that determine where and how regulatory proteins work. This provides additional clues about the mechanisms through which a mutation can disrupt the normal functioning of a cell. At the same time, the developers and independent experts emphasise the limits: a high AVI score indicates a noticeable molecular effect, but it is not equivalent to proven pathogenicity. The predictions concern the molecular level, not final clinical outcomes, and require further experimental verification.
The resource lowers the barrier to entry for laboratories without enormous computing power and reduces duplication of calculations. As with AlphaFold, open access accelerates collective progress, allowing researchers to focus on interpretation and experiments rather than on re-running models.
