Google DeepMind has mapped the biological impact of 9 billion possible mutations in the human genome, launching a new artificial intelligence tool that could transform our understanding of genetic disease and the mysterious non-coding regions of our DNA.
The system, called the AlphaGenome Atlas, systematically examines billions of potential genetic variations to predict their consequences within the human body. The development marks a significant step in interpreting the vast complexity of the human genome, a structure that scientists have yet to fully decode despite decades of international research.
The human genetic instruction manual is written in a chemical alphabet of four nucleotide bases: adenine (A), cytosine (C), guanine (G), and thymine (T). This sequence runs to roughly 3 billion letters in total, tightly coiled within the chromosomes of nearly every living human cell.
While most human DNA is identical from person to person, occasional variations occur. The most common type of genetic variation is a single-letter substitution, known as a point mutation, where one nucleotide base is swapped for another. These minute differences form the basis of many biological variations between individuals, influencing everything from benign traits to our susceptibility to various illnesses. In certain cases, a single incorrect letter in the genetic code can directly cause devastating rare genetic diseases.
To build the new atlas, DeepMind evaluated the effects of replacing every single letter in the 3 billion-letter human genome with each of the three alternative letters. This comprehensive calculation, accounting for three possible changes at every single position, resulted in a total of 9 billion possible mutations.
The atlas relies on predictions generated by the AlphaGenome artificial intelligence model, which the London-based Alphabet subsidiary introduced last year. The resulting repository of genetic predictions is vast, producing a dataset that reaches one petabyte in size, equivalent to the storage capacity of roughly a million standard laptop computers.
Decoding the dark matter of DNA
One of the most profound applications for the AlphaGenome Atlas is unravelling the mysteries of the 98 percent of the human genome that does not directly code for the production of proteins.
For many years, these vast stretches of non-coding DNA were often dismissed by geneticists as functional junk. However, modern biology has revealed that these regions are far from useless. Instead, they contain intricate regulatory mechanisms that act as the control panels for the genome, determining exactly when, where, and how actively specific genes are expressed.
Understanding these regulatory functions is significantly more difficult than deciphering traditional protein-coding genes. Using the artificial intelligence model's predictions, the DeepMind team has mapped thousands of short DNA sequences across the genome, known as motifs.
These motifs serve various critical functions within the cell. Some are responsible for controlling the production of messenger RNA, the crucial molecule that carries genetic instructions from the DNA out to the cell's protein-making machinery. Other motifs act as landing pads, attracting transcription factors, which are specialised proteins that bind to DNA and regulate gene activity.
Through the predictions generated by the system, researchers can now investigate whether these motifs activate genes in different types of human cells, or whether they alter the physical accessibility of the DNA structure itself. Because of this, researchers consider the AlphaGenome Atlas to be a searchable dictionary for non-coding DNA.
The system is also designed to be cross-referenced with large-scale genomic and healthcare databases. This will allow scientists to investigate the connections between specific genetic variants and biological traits or diseases. For example, a researcher scanning the genetic data of thousands of individuals could use the atlas to more easily identify rare DNA changes that correlate with characteristics such as height or specific protein levels in the blood.
Evaluating biological impact
The AlphaGenome model is capable of forecasting thousands of different biological outcomes triggered by a single DNA modification. These predictions range from determining which specific tissues a gene will become active in, all the way to predicting how a mutation will affect the complex, three-dimensional physical structure that DNA forms when folded inside a cell.
To make this overwhelming volume of data manageable, the new atlas gathers these predictions into an accessible format for the scientific community. A central feature of this effort is the introduction of a new metric called the AlphaGenome Variant Impact (AVI) score.
The AVI score distils the biological importance of a genetic variant down to a single numerical value. In initial testing, the AVI score successfully distinguished known disease-causing mutations from harmless genetic variations when applied to existing clinical genome databases.
Breaking down access barriers
Since the AlphaGenome model was first made available, DeepMind reports that approximately 9,000 researchers have accessed its predictions through an application programming interface. However, querying this system typically requires scientists to write their own software code to retrieve and analyse the data.
This technical requirement presented a significant barrier to entry, particularly for biologists and medical researchers who may lack advanced programming expertise. A primary motivation behind the creation of the AlphaGenome Atlas was to eliminate this obstacle entirely.
By pre-computing the predictions for all 9 billion possible genomic changes and assembling them into a central, searchable interface, DeepMind has granted researchers direct, immediate access to the data. This approach closely mirrors DeepMind's strategy with its AlphaFold database, a separate artificial intelligence system that solved the decades-old challenge of predicting the three-dimensional structures of proteins.
DeepMind previously pre-calculated hundreds of millions of protein structures and made them freely available, creating a resource that the company notes is now used by millions of people worldwide.
The future of AI-driven research
Looking ahead, the AlphaGenome Atlas is expected to become heavily utilised by autonomous artificial intelligence agents. These agents, which are software programs designed to navigate complex biological databases and research tools with increasing autonomy, could leverage resources like the new atlas to conduct research at an unprecedented scale.
These agents have the potential to automatically investigate genetic variants, identify plausible links to diseases, and highlight the most critical regions of the genome that warrant closer inspection by human experts.
In this respect, DeepMind positions the AlphaGenome Atlas not as a standalone system that will single-handedly solve the entirety of the human genome. Instead, it is presented as a massive, large-scale research tool designed to guide scientists through a labyrinth of billions of genetic variants, helping them pinpoint exactly where they should focus their attention next.
