Google DeepMind releases AlphaGenome Atlas, a map of predictions for 9 billion single-letter DNA changes

Google DeepMind releases AlphaGenome Atlas, a map of predictions for 9 billion single-letter DNA changes

Previously, AlphaGenome predicted the effects of one selected DNA change at a time. The new genome-wide map provides precomputed predictions that help researchers select variants and then test their predicted molecular effects experimentally.

Each person's genome contains millions of differences from the reference sequence. Researchers need to identify the few that may be associated with a disease or trait. Changes in regulatory regions of DNA are particularly difficult to assess: these regions determine when and in which cells a gene is active.

On September 8, Google DeepMind released AlphaGenome Atlas, a database of precomputed predictions for the effects of around nine billion possible single-letter DNA changes. Researchers can use the map to select variants without running the model separately for each candidate.

Atlas assigns each variant an AVI score. It combines predictions about processes that regulate gene activity, potential effects on proteins, and how well the corresponding region has been conserved across species. The contributions of individual features show what drove a high score, such as a predicted change in splicing, gene activity, or a protein. Splicing is the processing of a gene's RNA copy in which segments are joined together; a new splice site changes the composition of that copy.

In the team's technical paper, a high AVI score placed a variant in an intron of the DNM1 gene near the top of the list. An intron is a region usually removed from the RNA copy. This gene is associated with epileptic encephalopathy, a brain disorder involving epileptic seizures and impaired development. The model attributed the variant's high score to a predicted change in splicing in neuronal RNA. The authors then changed 265 DNA letters individually upstream of a DNM1 region and tested an artificial copy of this fragment in five types of laboratory cells. Twelve intronic variants extended an RNA segment while preserving the three-letter grouping of the genetic code; these included three previously reported likely pathogenic variants.

Using UK Biobank data from 54 189 participants, with blood levels of 2 028 proteins, the authors grouped rare noncoding variants by the process they were predicted to alter. After applying a statistical procedure to distinguish independent signals, they identified 728 associations between these variant groups and protein levels, compared with 595 in the baseline analysis, an increase of 22%.

AlphaGenome takes a DNA sequence as input. The three-dimensional folding of DNA, the cell's state, and its environment also influence gene activity, so predictions are tested experimentally in cells.

The map supports a practical workflow: select a variant, identify the process it is predicted to disrupt, and test that prediction experimentally.