Google DeepMind's latest AI tool promises a new era for genetics research, but experts warn to take its predictions with caution

Google DeepMind has revealed a new artificial intelligence (AI) tool that can predict the molecular impacts of tweaking any one of the billions of letters that make up our genetic code.

On Sept. 8, Google's AI powerhouse announced a freely available database called AlphaGenome Atlas, which contains data on 9 billion possible changes to the human genetic code and estimates how these changes will affect different tissues and cellular processes. What's more, the tool spits out a simple score to help scientists quickly weigh the knock-on impacts of these changes.

How a sequence of 3 billion or so letters of chemical code is packaged, delivered and read to create a human life is a major question in genetics research. Last year, DeepMind announced a tool that could propel researchers toward an answer. The tool, called AlphaGenome, predicts how changes to a DNA sequence alter proteins, cells and, ultimately, the human experience.

Now, one year later, DeepMind has released AlphaGenome Atlas, which the tech giant hopes will make AlphaGenome accessible to more researchers.

"It looks like a great resource," Greg Findlay, a group leader at the Francis Crick Institute in London who isn't involved with AlphaGenome Atlas, told Live Science. However, other experts said that, although the tool is a step forward, it still can't answer all of the big questions in genetics.

"It's not this holy grail," Tuuli Lappalainen, a professor in genomics at KTH Royal Institute of Technology in Stockholm and a senior associate faculty member at the New York Genome Center who isn’t involved with AlphaGenome Atlas, told Live Science.

Reading the future of genetic variants

Lappalainen's lab has been using AlphaGenome to study the effects of variation in genomes over the past year. She pointed out that this technology is not new. However, according to Google DeepMind, AlphaGenome outperforms previous models in the resolution at which it can predict the impact of changes to the genome. Around 98% of our DNA is noncoding, meaning cellular machinery doesn't directly read it to make the proteins that keep our cells functioning.

Geneticists now appreciate that this code, once thought to be "junk" DNA, is instead needed to manage how and when to read the parts of the genome that do encode our molecular building blocks, or coding genes. However, evolution hasn't produced an ordered genome. Instead, some regulatory instructions are distributed widely throughout the code.

AlphaGenome allowed researchers to explore how changes in a single pair of DNA "letters" could influence up to 1 million pairs of surrounding code ‪—‬ a far greater spread than previous models were capable of and one that captured much of each gene's regulatory network.

But this power came at a price. Until the release of Atlas, AlphaGenome required users to have enough bioinformatics experience to access and use the models' automated programming interface ‪—‬ an area of expertise not all geneticists have. What's more, calculating the effects of each change was a strenuous workout for academics’ computing resources.

To create AlphaGenome Atlas, DeepMind researchers computed every possible base ("letter") change and made the resulting data ‪—‬ totaling 1 petabyte, or 1 million gigabytes, of information ‪—‬ freely available. This takes the load off academics' overworked personal computers. What's more, the technology is now wrapped up in an easy-to-access, user-friendly web portal.

"You really do not need to be an expert in these methods to be able to go there and look something up in a browser," Lappalainen said.

Another simplification is the AlphaGenome Variant Impact score, a metric that predicts how much biological effect a change will have. In an accompanying preprint paper, DeepMind showed that the score could separate disease-linked mutations from harmless ones in a clinical dataset. While the score's simplicity might make it harder to interpret in certain cases, Lappalainen said, it would be useful for researchers who want to investigate a list of gene variants.

Clouds in the crystal ball?

The tool still isn't a perfect predictor, however. A Sept. 11 preprint study from a research team led by Katie Pollard, director of the Gladstone Institute of Data Science and Biotechnology and a professor at the University of California, San Francisco, suggested that while AlphaGenome was adept at finding causal mutations, it persistently underestimated their impact.

Pollard's team found that the model couldn't always link changes in regulatory elements to the genes they controlled, especially when they were not close together on the genome. "My perhaps naive hope would be that people take these predictions with an appropriate grain of salt," Lappalainen said.

Despite these limitations, Lappalainen said advances like AlphaGenome Atlas are part of a wider move toward more collaborative, data-led genomics. But she warned that the field would need to give equal focus to "wet-lab" experiments, which work out the accuracy of variant predictions.

Although these tests are tedious and consume much more time than a search on AlphaGenome Atlas' interface, she said, they are essential for generating the data that AI models need to improve their predictions.

"We're still very much data-limited in biology, and that data needs to be created," Lappalainen said.