The DNA of every cell in the human body is the same, but the same mutation may cause trouble only in cardiac muscle cells and nothing in immune cells.
Google DeepMind’s newly released AlphaGenome 2 aims to solve this problem: predict the impact of DNA sequences and mutations on a specific cell and in a certain state.
Previous generations could only look at the entire tissue, so mutations that worked in only one type of cell were easily missed. The new version was trained using a map of 68 tissues and more than 107 million single cells, covering both humans and mice.
Effectively, on a map containing 28 types of immune cells, it found 30% of the mutations that would change gene activity, nearly double the 16% found in the previous generation. It also replicated in computer tests a recent finding: A mutation linked to inflammatory bowel disease activates only one gene in certain intestinal immune cells, leaving other cells largely unaffected.
More practical for the laboratory: Researchers only need to take the data of dozens of their own cells, such as immune cells after vaccination, and feed it directly to it for analysis, without the need for retraining. It can also infer missing parts of experimental data.
Predictive indicators include gene expression, chromatin accessibility, DNA methylation, and more.
The model is now open to non-commercial research through the AlphaGenome API. DeepMind also makes it clear that it is not approved for clinical or diagnostic use. From "which cell has the problem" to "how to treat it", there is still a long period of wet experiments in between.