DeepMind opens AI prediction database for genetic variants — Live Science
Google DeepMind has introduced the free AlphaGenome Atlas database, which contains predictions about the consequences of 9 billion possible changes in the human genome. The tool assesses how replacing individual DNA “letters” may affect different tissues and cellular processes, and also provides a summary measure of a variant’s potential biological effect, Live Science reports.
Access to predictions without complex calculations
The Atlas was created on the basis of the AlphaGenome model, which DeepMind presented a year earlier. It models how changes in the DNA sequence may affect proteins, cells and, ultimately, the manifestations of these changes in the human body. According to DeepMind, the model has a higher resolution for predicting the effects of genetic changes than previous systems.
About 98% of human DNA does not directly encode proteins, but is involved in regulating when and how genes are read. AlphaGenome makes it possible to analyze how a change in one DNA base pair may affect a surrounding region up to 1 million base pairs long. According to the source, this covers a significant part of an individual gene’s regulatory network.
Before the Atlas appeared, working with AlphaGenome required bioinformatics skills and computing resources to use the model’s application programming interface. For the Atlas, DeepMind precomputed possible base changes and opened access to 1 petabyte of data through a web portal. Tuuli Lappalainen, professor of genomics at the Royal Institute of Technology in Stockholm, noted that users do not necessarily need to be specialists in such methods to search the data in a browser.
More current news is available on the UA.News Telegram channel Telegram.
Limitations of the model
The Atlas also offers the AlphaGenome Variant Impact metric, which is intended to assess the scale of a genetic change’s biological impact. In an accompanying preprint publication, DeepMind showed that this metric could distinguish disease-associated mutations from harmless ones in a clinical dataset.
At the same time, researchers warn that the model’s predictions are not definitive. A September 11 preprint prepared by a group led by Katie Pollard, director of the Gladstone Institute of Data Science and Biotechnology, indicates that AlphaGenome is good at finding causal mutations but systematically underestimates the scale of their impact. The model also could not always link changes in regulatory elements to the genes they control, especially when such regions are located far apart in the genome.
Lappalainen stressed that the predictions should be treated with caution and verified through laboratory experiments. According to her, it is precisely such studies that provide the data needed to further improve artificial intelligence models.
