DeepMind's 9-Billion-Mutation Atlas Changes the Rules of Genomics

A reference manual for the dark half of the genome

For twenty years, genomics has suffered from a chronic identity crisis. The Human Genome Project delivered the book in 2003. Scientists finally had the complete sequence of human DNA. What they did not have was the ability to read it.

That gap has been narrowing, slowly, for more than a decade. But the release of Google DeepMind’s AlphaGenome Atlas on Tuesday is a different order of magnitude. The database predicts the biological consequences of all nine billion possible single-letter substitutions across the human genome. It is, in practical terms, the first comprehensive map of how any given mutation might alter the machinery that switches genes on and off. Researchers can now query it through a browser instead of running weeks of lab experiments per variant.

Pushmeet Kohli, DeepMind’s vice president for research, put it plainly in a briefing call: for the first time, any scientist in the world can access a full map of human genetic variation “by simply opening a browser.” He also framed the release as finishing the unfinished business of the Human Genome Project. “We bought the book,” Kohli said, “but we did not understand how to read it.”

The non-coding problem, solved at scale

The biological stakes of this release hinge on a number most readers will find surprising: roughly two percent of the human genome codes for proteins. The other ninety-eight percent governs when and where those proteins get made. For decades, that regulatory landscape was largely unmapped terrain. Mutations in coding regions are relatively straightforward to interpret because they directly alter protein structure. Mutations in non-coding regions are far harder, because their effects are indirect, diffuse, and context-dependent. They switch genes up or down in specific tissues at specific times.

AlphaGenome, the underlying AI model DeepMind released last year, was built to predict exactly these kinds of regulatory effects. Atlas is what happens when you run it across an entire reference genome and compare every base to every possible alternative. The result is a catalogue of roughly 27,000 predictions per variant, spanning hundreds of human and mouse cell and tissue types. That includes how a mutation affects gene expression, RNA splicing, transcription factor binding, and chromatin accessibility.

To make sense of the volume, DeepMind introduced the AlphaGenome Variant Impact, or AVI, score. It combines AlphaGenome’s regulatory predictions with those from AlphaMissense, an earlier DeepMind model focused on protein-altering changes. An AVI score of ten places a variant in the top ten percent of impact; a score of thirty puts it in the top thousandth. Critically, the score breaks down which biological process drives the impact — splicing, gene expression, or protein change — so researchers can distinguish a mutation that disrupts a promoter from one that creates a cryptic splice site.

Early results suggest a diagnostic breakthrough

The beta-testing phase, reported in the paper now on bioRxiv, contains results that are difficult to dismiss as incremental.

At the Broad Institute, Laura Covill and Anne O’Donnell-Luria worked with the GREGoR Consortium, which investigates unexplained rare genetic disorders. They applied Atlas to a patient with epileptic encephalopathy whose condition had resisted diagnosis. The AVI score flagged a variant in the DNM1 gene, which is involved in synaptic function in brain cells. Sixty-nine percent of the score came from splicing effects: the model predicted the variant creates a false splice site in a brain-specific version of the gene, adding thirteen amino acids to the resulting protein. Because that protein variant is barely expressed in blood, earlier RNA sequencing of blood samples had yielded nothing. Lab experiments confirmed the prediction, and the variant was reclassified as likely pathogenic.

In a retrospective test on previously solved GREGoR cases, AVI placed the known causal variant among a patient’s top fifty candidates 29.5 percent of the time. The existing CADD ranking method achieved 12.5 percent. The difference is not marginal.

Gareth Hawkes at the University of Exeter ran Atlas against whole-genome data from more than 54,000 U.K. Biobank participants, hunting for rare non-coding variants that influence circulating protein levels. Filtering candidates by predicted molecular effect yielded 22 percent more associations than the same analysis without Atlas. In one instance, the candidate list shrank from 526 to four.

“The human genome is a massive search space,” Hawkes said. “We can use it to shrink the haystack.”

Julia Zeitlinger at the Stowers Institute used the atlas’s motif maps — a catalogue of more than 2,500 recurring short DNA sequences where transcription factors bind — to sort repressors by cell-type specificity. She noted that mapping thousands of such sites experimentally “would not have been possible,” and that four decades of laboratory work has validated only a tiny fraction of the motifs the model predicts.

Who wins, who loses, and what the access model means

DeepMind is making Atlas freely available to academic researchers through a dedicated website. Commercial access will come through Google Cloud licensing “soon,” according to Kohli. He did not disclose terms. Isomorphic Labs, Alphabet’s AI drug-discovery spin-off, already has access, though it too requires a commercial license.

This access architecture is significant. The scientific community benefits immediately from the free academic tier. Pharmaceutical companies and biotechs seeking to deploy the data for drug target identification will need to negotiate licensing. Whether the commercial terms are structured to favor large incumbents or to leave room for smaller firms remains unclear. The broader pattern here mirrors what has happened with AlphaFold, another DeepMind release: open for academic use, commercially gated, and inevitably reshaping competitive dynamics in fields built on biological data.

The implications for drug discovery are enormous but still theoretical. Most genetic variants linked to disease sit in non-coding regions, meaning they affect gene regulation rather than protein structure. If Atlas can reliably predict which regulatory mutations are pathogenic, it effectively turns a long-shot hunt into a guided search. That matters for rare diseases especially, where the diagnostic odyssey can span years and families. It also matters for common diseases with strong genetic components, where non-coding risk variants have been difficult to translate into drug targets.

Ewan Birney, director of EMBL’s European Bioinformatics Institute, is already working to integrate the AVI score into Ensembl’s Variant Effect Predictor, the annotation tool used by countless genomics labs worldwide. That integration is a signal that the data is being absorbed into the existing research infrastructure rather than treated as a standalone product.

Important caveats

DeepMind is not selling this as a replacement for laboratory evidence. Žigas Avsec, DeepMind’s genomics lead, said the model performs well for certain variant classes, such as those affecting splicing or promoters, but can miss others — particularly in enhancers. The predictions are “accurate enough to really point us in the right direction,” he said, but should not be treated as universal truth. They are less reliable than AlphaFold’s protein-structure predictions, another DeepMind milestone.

The paper also notes gaps in training data and limited ability to capture indirect effects, such as mutations that act through changes in the levels of regulatory proteins themselves. These are real limitations, not footnotes. A variant that looks benign in one cell type might be disruptive in another, and the model’s tissue coverage, while broad, is not exhaustive.

For clinical diagnosis, DeepMind explicitly states that Atlas predictions form only part of the evidence chain. They should guide downstream studies, not replace them.

Why this matters beyond the lab

The most compelling dimension of this release is its timing. The Human Genome Project celebrated its twentieth anniversary this year. Two decades after reading the sequence, we are finally getting a practical grammar for the language it encodes — especially the regulatory language that comprises the vast majority of the genome.

For patients with rare diseases, the timeline compression is tangible. A diagnosis that once took years of sequential testing could, in principle, be narrowed in weeks. For researchers studying complex diseases, the non-coding variants that population genetics has identified for years as associated with conditions ranging from diabetes to schizophrenia now have a predictive framework attached.

For the drug-discovery industry, the question is whether this data turns previously undruggable targets into tractable ones. Regulatory mutations are harder to target than protein misfolds, but understanding which regulatory changes drive disease is the first step toward designing interventions that modulate gene expression rather than block protein function.

The AlphaGenome Atlas does not answer every question. It does not replace experiments. It will not cure diseases on its own. But it does something that has been impossible until now: it makes the full landscape of human genetic variation navigable. That is not a small thing. It is the difference between wandering through a vast dark forest and finally having a map.