DeepMind’s New Genome Atlas Maps the Impact of 9 Billion Human Genetic Variations

The human genome contains approximately 3 billion DNA bases, yet only around 2% of them directly code for proteins. The remaining DNA is much harder to interpret. A new AI-generated atlas of the human genome from Google DeepMind could help researchers understand this complex biological code.

Some of the most common mutations in the human genome involve changes to individual nucleotides, or DNA letters. These variants contribute to differences between people, including differences in disease risk. In rare cases, a change to a single DNA letter can directly cause disease.

The AlphaGenome Atlas uses predictions from DeepMind’s AlphaGenome artificial-intelligence model to estimate the effects of approximately 9 billion possible single-letter changes across the human genome. The resource, which is available free of charge for non-commercial use, also includes more than 100 million short insertions and deletions observed in human DNA.

Researchers say the AlphaGenome Atlas could support the diagnosis of rare and unexplained diseases, improve understanding of common diseases and biological traits, and reveal hidden rules governing how DNA controls gene activity.

However, the atlas is not a replacement for laboratory experiments or the detailed analysis required to diagnose individual patients, says Martin Kircher, a bioinformatician at the Max Delbrück Center for Molecular Medicine in Berlin. “This is a convenient and generous way to scale up access to powerful models.”

Instant access to genome predictions

Since AlphaGenome was released, approximately 9,000 researchers have accessed its predictions through an application programming interface (API), according to DeepMind product manager Davi Hariharan. Using the API requires software development skills, however, which can be a barrier for some biologists.

To build the AlphaGenome Atlas, DeepMind calculated predictions for each of the three possible nucleotide substitutions at every DNA letter in the human genome. The resulting dataset contains approximately 1 petabyte of information. It also includes predictions for more than 100 million short insertions and deletions found in human genomes.

DeepMind said the project was inspired by its AlphaFold database, which contains more than 200 million predicted protein structures and has been accessed by millions of users.

“Removing friction also increases people’s curiosity to jump in,” says Žiga Avsec, who leads the AlphaGenome team. “Instant access feels like magic.”

Like AlphaGenome itself, the atlas provides thousands of predictions about the potential effects of genetic variants. These include effects on the tissues where nearby genes are expressed and changes to chromatin, the structure formed when DNA is packaged inside cells.

But Hariharan says that many API users repeatedly asked for a simpler way to interpret the results: “Can you give me one score to help me understand? Should I be concerned about this variant or should I dig deeper?”

To address this need, the Avsec team developed the AlphaGenome Variant Impact (AVI) score. This single measure estimates the predicted biological impact of a genetic variant. In a preprint, DeepMind and academic researchers report that AVI scores reliably distinguished disease-causing mutations from harmless variants in clinical genomics databases.¹

The AVI score and other AlphaGenome predictions also helped researchers at the Broad Institute in Cambridge, Massachusetts, prioritize non-coding mutations as possible causes of severe epilepsy.

Mafalda Díaz and Jonathan Fraser, computational biologists at the Center for Genome Regulation in Barcelona, Spain, said in an email to Nature that rare-disease researchers often rely on less computationally intensive models to interpret genetic variants. Applying models such as AlphaGenome across the entire human genome is not feasible for many research groups, they said. “By removing that computational barrier, the Atlas should become a valuable resource.”

Using AI to decode DNA motifs

Avsec is particularly interested in using the atlas to investigate short DNA sequences called motifs. These motifs occur throughout the human genome and can influence how genes are switched on or off.

Some DNA motifs help control the production of messenger RNA (mRNA), which carries instructions for making proteins. Other motifs attract transcription-factor proteins that regulate gene expression. However, scientists still have an incomplete understanding of how many non-coding DNA sequences function.

Using AlphaGenome predictions as a guide, the researchers mapped thousands of DNA motifs across the human genome and inferred their possible roles in different cell types. These roles include activating genes, repressing gene activity and changing how accessible DNA is to regulatory proteins.

The study “provides a searchable dictionary of non-coding DNA,” Julia Zeitlinger, a molecular biologist at the Stowers Institute for Medical Research in Kansas City, Missouri, and a co-author of the preprint, said at a press conference.