Changing a single letter in our DNA can sometimes alter how a cell works, and an AI has now virtually examined every possibility.
Our DNA contains around 3 billion pairs of chemical letters. At each position, one letter can be replaced by any of the other three. That represents nearly 9 billion possible changes. Most have no noticeable consequences, but some can disrupt a gene or the way it is used.
Molecular model of the DNA double helix.
Credit: OpenStax Anatomy and Physiology — CC BY 4.0
Testing them one by one in the laboratory would be unrealistic. Google DeepMind therefore used its artificial intelligence model AlphaGenome to calculate the possible effects of these 9 billion changes in advance. The results are compiled in AlphaGenome Atlas, a database of around 1 petabyte, or one million gigabytes.
In practical terms, the tool estimates what a change in DNA could cause inside cells. For example, a mutation can alter the activity of a gene or disrupt the production of a protein. These clues can then help researchers choose which mutations to study in reality.
To make this process easier, each change also receives a score called AVI. A single number thus summarizes the likely significance of its effect. This step is particularly useful when a patient with a rare disease has thousands of genetic differences and researchers need to identify those that warrant further analysis.
The value of this work is not limited to genes that directly provide instructions for making proteins. These regions account for only about 2% of the genome. The remaining 98% notably helps control gene activity. AlphaGenome Atlas can also examine changes located in this vast portion of DNA.
Researchers have already compared some predictions with experiments. In the study of an unresolved rare disease, the tool notably helped identify a mutation linked to the DNM1 gene. The calculations indicated that it disrupted the way the cell assembles its instructions before producing a protein. Experimental tests subsequently confirmed this effect.
AlphaGenome Atlas remains primarily a research tool: its predictions replace neither a biological experiment nor a medical diagnosis. The database is now freely accessible to academic researchers through a web interface and an API. They can therefore start with a specific mutation, examine its possible effects, and then decide which ones to test first in the laboratory.