Google DeepMind has done to the human genome what it once did to the protein universe: run the model on everything in advance and hand researchers the answer key.
On September 8, the lab released AlphaGenome Atlas, a database holding predicted molecular effects for all roughly 9 billion possible single-nucleotide variants in the human genome — every DNA letter swapped for every alternative across the roughly 3 billion positions in the reference sequence, plus over 100 million short indels. The result occupies about 1 petabyte, more than 30 times the size of the AlphaFold Database that reshaped structural biology in 2022.
The Atlas is not a new model. It is the old model, exhaustively pre-run. AlphaGenome, the sequence-to-function system DeepMind released in June 2025 and detailed in Nature in January 2026, takes up to a million base pairs of DNA as input and predicts gene expression, RNA splicing, chromatin accessibility and 3D genome contacts across hundreds of human and mouse tissues and cell types. Using it meant writing code, provisioning GPUs, and deciding which variants were worth the compute. Roughly 9,000 researchers had touched the API since launch, according to DeepMind product manager Dhavi Hariharan, cited by Nature. The Atlas drops that per-query cost to zero and puts it behind a browser window.
“When we started thinking about this project, it seemed impossible to do that computationally,” said Ziga Avsec, DeepMind's genomics lead.
One number for the whole genome
Alongside the raw predictions, DeepMind released the AlphaGenome Variant Impact score, or AVI — a single composite number per variant fusing AlphaGenome's regulatory predictions with pathogenicity estimates from AlphaMissense, plus evolutionary conservation and protein-coding features. Its selling point is coverage: AlphaMissense scored only the protein-coding 2 percent of the genome, while AVI ranks variants across the non-coding 98 percent where most trait-associated variation sits.
On solved rare-disease cases from the GREGoR Consortium, DeepMind reports AVI placed the known causal variant inside the top 50 candidates 29.5 percent of the time, against 12.5 percent for CADD, the field's incumbent genome-wide scorer — roughly a 2.4x improvement. Each score decomposes into feature attributions — splicing, expression, chromatin accessibility, conservation, protein impact — so a high score arrives with a testable mechanism attached.
Collaborators supplied the demonstrations. At the Broad Institute, Laura Covill and Anne O'Donnell-Luria used AVI to re-prioritize variants conventional filters had discarded in unsolved epilepsy cases, surfacing a non-coding change in DNM1 — a gene tied to epileptic encephalopathy — that AlphaGenome predicted would create an aberrant splice site and abnormally extend the protein. Lab screens reproduced the effect. At the University of Exeter, MRC fellow Gareth Hawkes applied Atlas features to whole-genome data from 54,189 UK Biobank participants and recovered 22 percent more rare non-coding associations with circulating protein levels, plus 19 genomic regions linked to body mass index.
Both Covill and O'Donnell-Luria are coauthors on the Atlas manuscript. These are collaborator results, not independent replication, and the paper peer-review status is unconfirmed.
Analysis: precomputation as a product, and the gap it does not close
The interesting move is strategic rather than scientific. AlphaGenome already existed; what DeepMind shipped is the decision to spend the compute once, centrally, and distribute the output as a static artifact. That is the AlphaFold Database playbook applied to gene regulation. It changes who the user is: a precomputed atlas is legible to a bench biologist with a browser tab, not just a bioinformatics core. It is also distribution: free for non-commercial academic use from day one, with commercial access coming on Google Cloud, plus an API and a Google Antigravity skill.
The rare-disease bottleneck is where the value is meant to land, and it is worth being precise about which one moved. Diagnostic odysseys stall because thousands of candidate variants survive filtering and nobody knows which three to take to the bench. AVI reorders that queue, and 29.5 against 12.5 percent is a real triage improvement. But it converts a computational constraint into an experimental one. Labs can now generate more plausible hypotheses than they have cells, reagents and hours to test. That is a better problem than the old one, but still a problem.
The harder gap is between a molecular prediction and clinical evidence. AVI says a variant is predicted to disrupt splicing in a given cell type. It does not say the variant causes a patient's disease. DeepMind is unambiguous: AlphaGenome has not been validated for and is not approved for any clinical use, and the Atlas is not a substitute for professional medical advice. Every high-scoring hit remains a hypothesis requiring laboratory validation and the full apparatus of clinical variant interpretation before it informs anything.
Outside genomicists have flagged where misreading is likeliest. Carl de Boer of the University of British Columbia, who helped build a comparative evaluation framework for such models, praised the resource while warning about its use. “It has a clear use, but it also is probably going to be easily misinterpreted,” de Boer told IEEE Spectrum. “We're talking about a very complex system and there's a lot of moving parts.” The limits are specific: the 1-megabase window truncates long-range enhancer effects; AVI scores variants one at a time, capturing none of the polygenic interaction behind common disease; and collapsing thousands of tracks into one number can let conservation dominate while the real disruption sits elsewhere. An independent CRISPRi benchmark from Cold Spring Harbor, updated August 3, found AlphaGenome posting the highest correlations tested but systematically understating effect sizes — in K562 cells, heavily represented in training, and on the base model rather than the Atlas.
What to watch: whether independent groups reproduce the AVI rare-disease numbers on cohorts DeepMind did not touch; whether the manuscript clears peer review intact; whether GREGoR and similar consortia fold AVI into standard prioritization pipelines; and how the commercial Google Cloud tier gets priced. The precomputation is finished. The validation has barely started.
“It has a clear use, but it also is probably going to be easily misinterpreted.”— Carl de Boer, Genomicist, University of British Columbia