The molecular fallout of every possible single-letter edit in the human genome is now a public resource, precomputed at a scale that dwarfs earlier biology databases.
AlphaGenome Atlas stores the predictions of DeepMind’s AlphaGenome model for roughly 9 billion single-nucleotide variants in the hg38 reference genome, along with more than 100 million insertions and deletions seen in gnomAD, UK Biobank and All of Us cohorts.
Size gives a sense of the ambition. At roughly a petabyte the resource runs 30-plus times larger than the AlphaFold Database, with each variant averaging 27,000 experiment-specific predictions spanning hundreds of human and mouse cell types.
A companion score, AVI, compresses the model outputs with conservation and loss-of-function evidence into a single number per variant, then attributes that number to mechanisms such as splicing or gene expression.
DeepMind says AVI covers coding DNA, about 2 percent of the genome, and the non-coding majority that regulates it, posting its largest accuracy gains on disease-linked non-coding variants.
Early adoption comes from rare-disease work. A Broad Institute team inside the GREGoR consortium scored unsolved cases with AVI and surfaced a deep intronic DNM1 variant that creates a brain-specific cryptic splice site, extending the protein by 13 amino acids, a prediction confirmed by experimental screens.
