TL;DR
Google DeepMind has released a precomputed database covering the likely biological effect of all nine billion single-base substitutions possible in human DNA, free to academic researchers. Previously such questions were settled variant by variant, in a laboratory. The UK connections are substantial and mostly absent from the headlines: the validation work runs through UK Biobank, a University of Exeter researcher ran one of the trials, and EMBL’s European Bioinformatics Institute intends to build the output into tooling that geneticists everywhere already use.
What it contains
AlphaGenome Atlas was produced by running DeepMind’s existing AlphaGenome model across a reference genome and comparing each base against its three alternatives. Every variant carries roughly 27,000 individual predictions spanning hundreds of human and mouse cell and tissue types. Insertions and deletions numbering beyond 100 million, drawn from population databases, were scored too.
The accompanying summary metric, the AlphaGenome Variant Impact score, is designed for practical triage: a 10 places a variant in the most impactful tenth of the genome, a 30 in the top thousandth, and the figure is decomposed to show whether protein change, expression or splicing is doing the work. That breakdown addresses a real gap — protein-coding sequence accounts for just 2%, and the other 98%, which governs when genes switch on, has proved far harder to read.
Pushmeet Kohli, who heads DeepMind’s AI for science work, framed the release against the Human Genome Project: the book was bought in 2003, he said, but not yet readable.
Early results, and the caveats
Broad Institute researchers used the score to revisit unsolved rare disease cases and reclassified a variant as likely pathogenic in an unsolved epileptic encephalopathy case, once bench work confirmed what the model had predicted. Exeter’s MRC fellow Gareth Hawkes applied it to whole-genome data covering more than 54,000 UK Biobank participants, lifting the association count by 22% against an unaided run — and in one instance cutting 526 candidate variants to four.
DeepMind is unusually direct about limits. Genomics lead Žiga Avsec said accuracy falls short of what AlphaFold achieved for protein structure, that enhancer variants in particular get missed, and that researchers should not treat the predictions as settled truth.
Looking forward
Non-commercial access opens now; commercial licensing arrives via Google Cloud, and sister company Isomorphic Labs will need a licence like anyone else. For the UK the immediate value sits with academic and NHS-adjacent research groups, though it lands the same week a minister linked Palantir mistrust to rising NHS data opt-outs — a reminder that health data goodwill is finite and separately spent.