Researchers at Google DeepMind, an artificial intelligence company owned by the US giant, unveiled on Tuesday a tool that predicts whether genetic mutations are potentially pathogenic or not, an advance that could particularly help in research into rare diseases.
This is “a new step in recognizing the impact that artificial intelligence is having on the natural sciences,” Ziga Avsec, vice president of research at Google DeepMind, said at a press conference.
The tool examined so-called “missense” mutations, in which a single letter of the genetic code is affected and the function of proteins can be changed. However, every cell fulfills its function with the help of proteins that constantly transmit their instructions.
A person carries around 9,000 of these mutations, most of which are harmless, according to the company, but some of which can lead to diseases, such as cystic fibrosis.
To date, four million of these mutations have been observed in humans, but only 2% of them have been classified as pathogenic or benign.
And in total there are 71 million such possible mutations. They were verified by Google DeepMind tool called AlphaMissense, which was able to comment on 89% of them.
He assigned each mutation a value between 0 and 1, which indicates the risk that it is pathogenic, i.e. the cause of a disease. Results: 57% were classified as probably benign and 32% as probably pathogenic, the rest remained uncertain.
The database was published and made available to all scientists. A corresponding study was published on Tuesday in the renowned journal Science.
AlphaMissense shows “superior performance” compared to existing tools, wrote experts Joseph Marsh and Sarah Teichmann in an article also published in Science.
The tool “does not say what disease it will cause,” Jun Cheng, a scientist at Google DeepMind, clarified during the press conference. “But we believe our predictions can help increase the diagnosis rate of rare diseases and may also help us find new genes involved in diseases.”
Indirectly, this could lead to the development of new treatment methods, say the researchers, who, however, warn against using AlphaMissense alone to make a diagnosis.
The tool was trained on a database of human and primate DNA to detect which genetic mutations are prevalent.
“It learns what a protein sequence normally looks like, and when presented with a sequence with a mutation, it can tell whether it looks worrying or not,” Jun Cheng explained.
“It’s a bit like human language: if you replace a word in an English sentence, a person who knows English will immediately recognize whether that change affects the meaning of the sentence or not,” he said.
AlphaMissense relied on AlphaFold, another machine learning program introduced by Google DeepMind more than two years ago, which helped publish the largest protein database.