Atlases have long guided us through uncharted territory. Now, an AI-generated atlas by Google DeepMind seeks to do the same for the vast landscape of our DNA.
Ever since the Human Genome Project, scientists have painstakingly traced the myriad DNA mutations that contribute to health and disease. But that quest has largely been stymied by the genome’s vast scale. Only two percent encodes the proteins that make our bodies work; the rest may control how genes are turned on or off or be junk left over from evolution.
With roughly nine billion possible DNA letter swaps, testing each one in the lab is impossible. Making sense of their interactions is an even tougher challenge. Yet these changes often contribute to differences in risk for cancer, dementia, and other medical scourges.
DeepMind’s new atlas could lend researchers a hand. Generated from the company’s AlphaGenome AI released last year, the searchable database predicts the effects of every possible DNA letter swap. Thousands of researchers have already experimented with AlphaGenome, but those studies required some coding prowess, raising the barrier to entry.
AlphaGenome Atlas may make the AI more accessible. Analysis of individual DNA changes, down to the level of specific tissues, is readily available through a web portal for non-commercial use. As the most comprehensive catalog of how genetic mutations might affect molecules in the body, it could help uncover the mutations underlying traits and illnesses. By charting the genome’s “dark matter”—regions that don’t encode proteins— it might also reveal hidden rules that direct gene activity. The details are described in a paper.
“This represents the first time that any researcher in the world can access a comprehensive map of the human genome and its variations by simply opening a browser,” said Pushmeet Kohli, DeepMind’s vice president of science, in a press briefing.
The Language of Life
With just four DNA letters—A, T, C, and G—our genomic instructions seem simple. But the actual genetic playbook is far more complex. After piecing together the first draft of the human genome at the turn of the century, scientists were surprised by how little of it guided protein manufacturing. A staggering 98 percent didn’t seem to do much, earning the nickname junk DNA.
Long overlooked, these non-coding sections have increasingly captured attention for their role in regulating gene expression. Some DNA snippets can even operate thousands of letters away from the genes they control, making their involvement tough to decipher.
Non-coding DNA is also highly dynamic. Some genetic chunks can be duplicated or cut out as cells divide. Others jump to distant locations, reverse their sequences, or elbow their way into protein-coding genes.
Single-letter swaps are among the most prevalent DNA mutation. These can be relatively harmless. But they also can lead to diseases such as sickle cell anemia or raise a person’s “bad cholesterol” levels, increasing the risk of heart attacks. Gene-editing clinical trials are already underway to tackle these problems. But engineering a safe and effective treatment requires knowing which DNA swaps to make, and that’s been a roadblock.
Here’s where AlphaGenome comes in. Formally released early this year, the AI works in three steps. First, it spots short patterns in DNA sequence. Then it shares that information across a larger region of the DNA strand, letting it connect local patterns to distant letters. Finally, AlphaGenome translates those patterns into predictions of downstream biological effects.
The AI is customizable for different projects, allowing researchers to home in on DNA changes related to their specific questions. But it can only be accessed through an automated programing interface (API) which requires writing code and makes the data harder to access.
“AlphaGenome is helpful for analyzing specific variants and has found widespread use in research, but we wanted to show researchers a big-picture view of variants across the entire genome,” wrote the DeepMind team in a blog post.
Genome Cartographer
The new atlas does away with much of the coding and analysis, allowing researchers to search for DNA variants across the genome to see their potential effects.
To build the database, the team computed predictions for all three possible swaps at every DNA letter—for example, changing A to T, C, or G—resulting in a whopping petabyte of data.
As with AlphaGenome itself, the atlas generates thousands of predictions about how DNA changes affect molecular processes in different tissues. These include what happens when a nearby gene is switched on or how changes in the shape of chromatin, the tightly folded form of DNA, alter its biological activity.
“Just as an atlas is a collection of maps, linking together features of the land like altitude and location, AlphaGenome Atlas charts the molecular effects of DNA variants across the genome,” wrote the team.
But interpreting the atlas takes more work. With billions of potential changes, which ones should researchers prioritize?
To help them navigate the most promising variants, the team also developed a single metric to measure their predicted effects. Called the AlphaGenome Variant Impact (AVI) score, it combines AlphaGenome with AlphaMissense, a model that predicts the effects of mutations in protein-coding regions. Together, these two tools help distinguish harmless mutations from those more likely to play a role in disease.
In collaboration with the Broad Institute, the score has already helped researchers find and prioritize a non-coding DNA variant that may contribute to severe epilepsy. Rare disease researchers, who often lack the funding and computing resources needed to run genomic AI models directly, could particularly benefit from the atlas.
“If somebody is studying a disease, and they don’t have any idea about what cell types to look for or what molecular processes are impacted, then starting with an AVI score…is a great starting point to help you prioritize variants and try to find that needle in the haystack,” said genomic lead and study author Žiga Avsec in a press conference.
Beyond tackling genetic diseases, the atlas could also help decode mysterious non-coding motifs, or snippets of DNA scattered across the genome. Some motifs control the production of messenger RNA, which carries genetic instructions to the cell’s protein-making factories. Others alter the activity of individual genes. But most remain poorly understood, if they have a function at all.
Linking these motifs to large health databases, such as the UK Biobank, could map the gene interactions and resulting proteins underlying height and other complex traits. The atlas could also help AI agents rapidly generate hypotheses for human collaborators to explore in the lab.
AlphaGenome Atlas isn’t meant to replace real-world experiments. And unlike AlphaFold, DeepMind’s protein structure-predicting AI that garnered a Nobel Prize, DeepMind needs to further boost its accuracy. But the atlas is shaping up to be a valuable guide for genomic explorers navigating the vast DNA landscape that makes us human.
