Google DeepMind launches AlphaGenome Atlas mapping 9 billion DNA variants

The 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 one petabyte of data, making it roughly 30 times the size of Google's AlphaFold database. In a blog post, DeepMind said the resource should dramatically speed up genetic research, as previously each variant had to be tested one at a time or in a laboratory, a process that would take several human lifetimes to cover all nine billion possible mutations.

Researchers can access the Atlas free for noncommercial use through a web portal, the existing AlphaGenome application programming interface, and as a skill in Google Antigravity. DeepMind said commercial access will later be offered through a Google Cloud service. The Atlas builds on the AlphaGenome model launched in January, which can take up to one million DNA letters and predict thousands of molecular measurements, including gene expression, chromatin accessibility, and RNA splicing. According to the journal Nature, AlphaGenome outperformed leading non-specialized models in 25 of 26 variant effect prediction benchmarks.

DeepMind also introduced the AlphaGenome Variant Impact score, which combines AlphaGenome predictions with outputs from AlphaMissense, a model for examining protein-altering variants. The score condenses predicted biological effects into a single number to help researchers prioritize variants for investigation.

Collaborators have tested the approach on disease and population genetics. Researchers at the University of Exeter used Atlas predictions to hunt for rare, non-coding variants affecting protein levels in human blood. By filtering candidates by predicted molecular effect, they found 22% more associations than the same analysis without the Atlas, and in one case narrowed a region of 526 candidates down to four. Gareth Hawkes, a Medical Research Council fellow at Exeter, said, “The human genome is a massive search space. We can use it to shrink the haystack.”

At the Stowers Institute for Medical Research, scientists used Atlas predictions to sort transcription factors by their function in different cell types, for example separating repressors that leave DNA accessible but block gene activation. Investigator Julia Zeitlinger said mapping these factors without the Atlas “would not have been possible” because the experimental work would be too laborious.

Despite the resource's promise, DeepMind's genomics lead Žiga Avsec cautioned that Atlas predictions should not be seen as a direct substitute for experimental evidence. AlphaGenome works well for certain variant types, such as those affecting promoters or splicing, but less well for others, especially enhancers. The predictions are also not as accurate as AlphaFold's protein structure predictions. A Nature paper describing the Atlas says it is a research tool that should only form part of a clinical diagnosis, with gaps in training data limiting its ability to understand indirect effects such as changes in regulatory protein levels. The predictions are “accurate enough to really point us in the right direction,” Avsec said, but remain a guide for researchers.