DeepMind releases AlphaGenome Atlas with 9 billion DNA predictions

Google DeepMind announced Tuesday the release of AlphaGenome Atlas, a comprehensive database predicting the biological effects of all 9 billion possible single-letter mutations in human DNA. The resource is now available free to academic researchers worldwide, with commercial licensing through Google Cloud coming soon.

The database represents a significant leap beyond the 2003 Human Genome Project, which mapped the complete DNA sequence but left researchers without tools to interpret what variations mean. "As the saying goes, we bought the book," said Pushmeet Kohli, DeepMind's vice president for research, "but we did not understand how to read it."

Why it matters

Until now, testing genetic variants required running models one mutation at a time or conducting laboratory experiments—work that would have taken many human lifetimes to complete for all possible variations. AlphaGenome Atlas eliminates this bottleneck, giving any researcher with a browser instant access to predictions that previously didn't exist. This acceleration could fundamentally change the pace of genetic disease research and drug discovery.

How the system works

DeepMind built the Atlas by running its AlphaGenome AI model across a reference human genome, comparing each DNA base against three possible alternatives. Each variant connects to roughly 27,000 individual predictions about effects on gene expression and protein manufacturing instructions across hundreds of cell and tissue types in humans and mice.

The system also analyzed over 100 million insertions and deletions from population databases including the U.K. Biobank and the U.S. National Institutes of Health's All of Us database.

To make these predictions actionable, DeepMind created the AlphaGenome Variant Impact (AVI) score. A score of 10 places a variant among the 10% most impactful in the genome, while a score of 30 identifies it as one in a thousand. Each score breaks down whether impact comes from splicing, gene expression, or protein changes.

Addressing the non-coding genome

Protein-coding DNA represents just 2% of the human genome. The remaining 98% controls when and where genes activate—a portion scientists have struggled to interpret. "AlphaMissense looks at proteins," explained Žiga Avsec, DeepMind's genomics lead. "With AlphaGenome, we are focusing on the regulatory part of the genome."

The Atlas also catalogs more than 2,500 recurring DNA sequences that transcription factors bind to throughout the genome.

Early validation from researchers

Beta testers reported promising results. Researchers at the Broad Institute used the AVI score to solve a previously unexplained case of epileptic encephalopathy, identifying a variant in the DNM1 gene that earlier blood tests had missed. Laboratory experiments confirmed the prediction.

In retrospective testing on solved cases, AVI placed the known causal variant among a patient's top 50 candidates 29.5% of the time, compared to 12.5% for CADD, an existing ranking method.

Gareth Hawkes at the University of Exeter applied Atlas to data from over 54,000 U.K. Biobank participants, finding 22% more associations than analysis without the tool.

Limitations acknowledged

DeepMind emphasized that Atlas predictions don't replace laboratory experiments. Avsec noted the model works well for variants 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 shouldn't be treated as "universal truth."

The research paper describes Atlas as one component in clinical diagnosis evidence chains, acknowledging gaps in training data and limited ability to capture indirect effects.

These details were first reported by Fortune.