AI Has Mapped Every Potentially Problematic Spot in the Human Genome — All at Once.
On September 9, Google DeepMind officially launched AlphaGenome Atlas, a searchable prediction database covering the entire human genome. It represents an unprecedented engineering feat in the history of life sciences: calculating all 9 billion possible single-base mutations across the human genome.
So, what does 9 billion actually mean?
The human genome contains roughly 3 billion base positions, and each position can undergo three possible substitutions, resulting in approximately 9 billion possible single-nucleotide substitutions. Experimentally validating each of these mutations one by one would be virtually impossible. AlphaGenome Atlas has not only calculated them all, but also generated around 27,000 prediction metrics for every mutation, assessing its potential effects on gene expression, RNA splicing, chromatin accessibility, and 3D DNA structure across hundreds of human and mouse cell types.
Nine billion mutations multiplied by 27,000 metrics produces more than 240 trillion predicted values. Together, the dataset occupies approximately 1 petabyte (PB)—around 30 times the size of the AlphaFold Database.
DeepMind CEO Demis Hassabis reflected that AlphaFold mapped the universe of proteins, while AlphaGenome Atlas is beginning to map the human genome.
Most importantly, the atlas is designed to be highly accessible to researchers. There is no need to set up a local environment or write a single line of code. Researchers can simply enter a genomic position in a browser and receive the results within seconds.
DNA is often described as the language of life, written using four bases: A, T, C, and G. Humanity first visualized the DNA double helix in 1953. In 2003, the Human Genome Project was completed after 13 years and approximately $2.7 billion in investment, allowing scientists to read all 3 billion letters of the human genome.
But reading the genome is far from understanding it.
Only around 2% of those 3 billion letters directly encode proteins. The remaining 98% may not encode proteins, but they play critical roles in controlling when, where, and how strongly genes are expressed. In fact, many disease- and trait-associated variants are found precisely in these non-coding regions. Until now, determining what happens when just one of these letters changes has been extraordinarily difficult.
AlphaGenome changes that equation.
The model processes up to 1 million DNA bases at a time, using convolutional layers to capture local patterns and Transformers to model long-range dependencies. It can directly predict thousands of epigenetic and transcriptional tracks at single-base resolution.
To evaluate a variant, researchers simply run the original and mutated sequences through the model and compare the results. The predicted effects can be generated in roughly a second.
But even at one second per variant, processing all 9 billion possibilities sequentially would still take around 285 years.
What Atlas does is essentially compress those 285 years of computation into a precomputed, searchable reference table.
The system's potential was demonstrated even before its official release.
A team led by Anne O’Donnell-Luria and Laura Covill at the Broad Institute encountered a child who had suffered seizures since infancy, accompanied by spasms, developmental delay, and low muscle tone. Previous genetic analyses had failed to establish a diagnosis, leaving the case unresolved for years.
This time, the researchers used Atlas's AVI score to re-rank all candidate variants identified in the child.
The highest-ranked candidate was a G-to-A mutation in an intronic region of the DNM1 gene on chromosome 9. Atlas reconstructed the likely molecular mechanism: the mutation created an aberrant splice site, resulting in abnormal extension of the translated protein.
Subsequent laboratory experiments fully confirmed the mechanism predicted by the AI, helping solve the molecular mystery behind the child's epileptic encephalopathy.
In another application, researchers at the University of Exeter applied Atlas to whole-genome data from more than 54,000 UK Biobank participants. By grouping rare variants according to their predicted molecular effects, the approach identified 22% more non-coding genetic associations.
Atlas is built as a multi-layered system.
At the foundation are the raw predictions, covering approximately 27,000 molecular effects. The next layer provides a unified scoring framework, combining AlphaGenome with AlphaMissense to produce an intuitive AVI pathogenicity score.
Above that is a transparent attribution layer, allowing scientists to understand why a mutation receives a high score—whether it disrupts RNA splicing, alters chromatin accessibility, or affects another regulatory mechanism.
At the top sits a genome-wide vocabulary of 2,601 frequently occurring DNA patterns, with approximately 253 billion instances annotated across the genome.
Work that once required multiple research teams and years of experiments can now be narrowed down to a handful of clicks.
Looking back at DeepMind's trajectory in biological computing:
2020: AlphaFold cracked the protein-folding problem.
2022: The AlphaFold Database opened to the global scientific community, providing more than 200 million predicted protein structures.
2023: AlphaMissense assessed the pathogenicity of 71 million protein mutations.
2025: AlphaGenome advanced the understanding of non-coding regulatory regions.
Today: AlphaGenome Atlas has mapped the possible effects of every single-base substitution across the human genome.
In just five years, DeepMind has moved from understanding the physical structure of individual proteins to systematically mapping the potential consequences of every possible single-base change in the human genome.
From reading life, to understanding life, to predicting life.
The next era of computational biology may already be underway.
