Google's AI Predicts How a Small Change in Someone's DNA Affects Each Kind of Cell in the Body
Google DeepMind, Google's AI research lab, has released AlphaGenome 2, a program that predicts how a tiny spelling difference in a person's DNA changes the way a particular kind of cell behaves. Because the same DNA is read differently in a gut cell than in a blood cell, this could help researchers find which differences actually lead to disease, and where in the body.
Google DeepMind, Google's AI research lab, today released AlphaGenome 2, a program that predicts what happens when a person's DNA has a small spelling difference. DNA is the long instruction book inside nearly every cell, written in four chemical letters. People differ from one another by millions of single letters, and most of those differences sit not in the genes themselves but in nearby stretches that work like dimmer switches, turning genes up or down. The hard question is which of those differences matter.
Every cell carries the same instruction book, but a gut cell, a brain cell and an immune cell each read it differently. The first version of the program, released last year, predicted the effect of a DNA change averaged across a whole body tissue. The new one makes its prediction for a specific kind of cell, and even for the same cell in different states, such as an immune cell at rest versus one that has just been switched on by a vaccine. It learned from more than 107 million measurements of individual cells taken from 68 kinds of human and mouse tissue.
Google gives two examples. For inflammatory bowel disease, a painful illness of the gut, earlier lab research had found a single DNA letter that turns up a gene called ETS2 in one kind of immune cell in the gut. Without being told the answer, the program picked out the same letter and the same cell. In a second test, it showed that a DNA difference linked to severe COVID-19 has a stronger effect in immune cells that have been switched on than in resting ones.
On one test using human immune cells, Google says the new program found the DNA letter actually responsible for a change in gene activity 30% of the time, up from 16% for its earlier version. Overall, Google says it finds nearly twice as many DNA differences that change gene activity as the best earlier tools. Sarah Teichmann of the University of Cambridge, quoted by Google, said: "What makes AlphaGenome 2 exciting is how it adapts directly to new single-cell data without retraining."
Scientists can use it free over the internet for non-commercial research, and companies can request early access. Google says plainly that it is a research tool only and has not been tested or approved for diagnosing patients. The performance figures come from Google's own tests, with no outside review yet. And even the better score means the right answer is found in fewer than one case in three on that test. Reproducing a known finding is a useful check; the bigger test, as a writer at the AI news site The Neuron points out, is whether its predictions lead to new discoveries that hold up in the lab.
The new program is slower to run than the old one, so it does not replace the map Google published in September of every possible single-letter change in human DNA. It adds a closer look, cell by cell, on top of it.
Sources
- Google DeepMind — AlphaGenome 2 announcementdeepmind.google · primary
- The Neuron — what AlphaGenome 2 does and the limits of its test resultstheneuron.ai