Google Publishes an AI Map Predicting the Effect of Every Possible One-Letter Change in Human DNA

Google DeepMind today opened a free website for researchers that predicts what would happen if any single letter of the human genetic code were swapped for another: about 9 billion possible changes in all. It could help scientists find the hidden genetic causes of rare diseases, but it is a research tool, not a medical test.

Google's AI research lab, Google DeepMind, today released AlphaGenome Atlas, a free website that lets researchers look up an AI program's predictions for every possible one-letter change to the human genetic code.

Some background. Our DNA is a long string of chemical letters, and a change in a single letter can sometimes cause disease. The hard part is telling which changes matter. Only about 2% of our DNA spells out the proteins our bodies are built from. The other 98% works more like a set of switches and dimmers that decide when and how strongly genes turn on, and that is where most of the genetic differences linked to traits and diseases turn up. Those are especially hard to read.

DeepMind already had an AI program, AlphaGenome, that predicts what a DNA change does to those switches. Now it has run that program on every possible single-letter swap in the human genome, about 9 billion of them, and stored the answers. Each change gets predictions across hundreds of human and mouse cell types and tissues, plus one summary score of how much impact it is likely to have.

DeepMind shared early examples from outside researchers. One team used it to spot a DNA change tied to a severe form of epilepsy that earlier analyses had missed. Another found 22% more links between stretches of DNA and body traits in a large British health study than they had found without it.

The website is free for academic research, with a paid version for companies planned on Google's cloud. DeepMind states plainly that it "has not been validated for, and is not approved for, any clinical use" and is not a substitute for medical advice. Carl de Boer, a genetics researcher at the University of British Columbia who was not involved, told IEEE Spectrum that AlphaGenome is the field's leading program of its kind, but warned that the single summary score "is probably going to be easily misinterpreted," since many diseases involve many genetic changes at once and predictions still need checking in the lab.