Google LLC’s DeepMind research unit said today it has used its AlphaGenome artificial intelligence model to predict the biological consequences of the more than 9 billion possible single-letter changes to human DNA.
It’s making the resulting database available free to researchers globally via the AlphaGenome Atlas, released today. The AlphaGenome Atlas is a precomputed catalogue of how each substitution of a single DNA base is likely to impact the machinery that switches genes on and off. It contains more than 1 petabyte of data, making it about 30-times the size of Google’s AlphaFold database.
In a blog post, DeepMind said it should help to dramatically speed up genetic research. Previously, researchers had to do this manually by running a model against each variant, one at a time, or test the variants in a laboratory. But this was an incredibly slow process, and would have taken several human lifetimes to discover the consequences of all of the 9 billion possible single-letter mutations.
With the Atlas, the work of biologists and medical researchers should be made much easier, aiding their understanding of genetic diseases and speeding up research into possible cures. Google is making it free for noncommercial researchers through a web portal, and also through the existing AlphaGenome application programming interface and as a skill in Google Antigravity. Later, DeepMind says, it will offer commercial access too, via a Google Cloud service.
The AlphaGenome Atlas builds on the launch of DeepMind’s AlphaGenome model in January. With AlphaGenome, researchers can take up to 1 million DNA letters and predict thousands of molecular measurements, including gene expression, chromatin accessibility and RNA splicing. The journal Nature reported that AlphaGenome outperformed leading nonspecialized models in 25 of 26 variant effect prediction benchmarks.
Now, DeepMind is transforming AlphaGenome’s predictions into a comprehensive resource that’s much easier to access. Previously, researchers would have had to submit variants through an API, which requires some level of coding knowledge.
That explains why only about 9,000 researchers had used the API since AlphaGenome was launched. With the AlphaGenome Atlas, DeepMind is essentially making all of AlphaGenome’s outputs available at once, together with a browser-based interface for researchers to explore them.
DeepMind also introduced what it calls the AlphaGenome Variant Impact score, or AVI score, which combines AlphaGenome’s predictions with the outputs of AlphaMissense, a model for examining protein-altering variants. The idea is to condense the predicted biological effects of a variant into a single number, which shows researchers which ones they should prioritize in their investigations.
Collaborators have tested this approach on a number of questions around disease and population genetics. Researchers at the University of Exeter used the Atlas’ predictions to hunt for rare, noncoding variants that affect the level of proteins in human blood. By filtering candidates by their predicted molecular effect, they yielded 22% more associations than the same analysis ran without using the Atlas. In one case, they narrowed a region of 526 candidates down to just four.
“The human genome is a massive search space,” said Gareth Hawkes, a Medical Research Council fellow at the University of Exeter. “We can use it to shrink the haystack.”
Meanwhile, researchers at the Stowers Institute for Medical Research used Atlas predictions to sort transcription factors based on what they do in different cell types. For instance, they separated the repressors that leave DNA still accessible but block a gene from switching on. Without Atlas, mapping these factors “would not have been possible,” said investigator Julia Zeitlinger, because doing the same work through experiments is just too laborious.
While AlphaGenome Atlas should dramatically speed up biological research, DeepMind’s genomics lead Žiga Avsec warned that its predictions shouldn’t be considered a direct substitute for experimental evidence. While AlphaGenome works well for some types of variant, such as those affecting promoters or splicing, it doesn’t do so well with others, especially enhancers. He noted that the Atlas predictions aren’t as accurate as what DeepMind’s AlphaFold model achieved in protein structure prediction.
In a paper published in Nature, DeepMind describes AlphaGenome Atlas as a research tool that should only be used to form part of a clinical diagnosis. It notes there are still gaps in its training data that limit its ability to understand indirect effects, such as through changes in the level of regulatory proteins.
For that reason, the AlphaGenome predictions should be seen as more of a guide for researchers. They are “accurate enough to really point us in the right direction for downstream studies,” Avsec said. But they should “not be treated as the universal truth.”
