Researchers from DeepMind, a subsidiary of Google, which specializes in artificial intelligence, revealed on Tuesday a tool that predicts whether genetic mutations will cause diseases or not, in a useful achievement in many fields, most notably research related to rare diseases.
Vice President of Research at DeepMind, Pushmeet Kohli, said during a press conference that this innovation is “a new step in understanding the effects that artificial intelligence technology has on natural sciences.”
The tool dealt with mutations called “missense” because they represent a change in only one letter of the genetic code, which makes them capable of changing the function of proteins, knowing that every cell accomplishes its function with the help of proteins that constantly provide their instructions.
An individual has about 9,000 of these mutations, most of which are benign, according to the company, but some of them can cause diseases, including cystic fibrosis.
To date, four million of these mutations have been observed in humans, but only 2% of them are classified as either disease-causing or benign.
In total, there were 71 million possible mutations, which DeepMind's Alpha Messence tool reviewed and was able to predict 89% of them.
The tool gave each mutation a score between 0 and 1 to indicate its likelihood of being disease-causing. The result was that 57% of the mutations were likely to be benign and 32% to be disease-causing. The nature of the remaining mutations is uncertain.
The database was made available to the public and to various scientists, while a study was published on Tuesday in the prestigious journal Science to present this data.
"Superior performance"
"Alpha Messenger" outperformed existing tools, according to what experts Joseph Marsh and Sarah Tishman confirmed in an article also published in "Science" magazine.
Scientist Jin Qing from DeepMind explained at the press conference that the tool “does not determine the type of disease that mutations will cause,” adding, “But we believe that our predictions are useful for raising the rate of diagnosis of rare diseases, and perhaps to help us find new genes associated with diseases.”
The researchers indicated that this could indirectly lead to new treatments, but they warned against using the tool exclusively to make diagnoses.
The tool was trained on a database of DNA from both humans and primates, allowing it to see which genetic mutations were dominant.
“The tool learns what a protein sequence normally looks like, and when it is presented with a sequence with a mutation, it can tell whether that is a cause for concern or not,” Cheng said.
He added, "It is somewhat similar to human language. If we replace one word with another in an English sentence, a person who is fluent in English will understand whether this change affects the meaning of the sentence or not."
The AlphaMessence tool was based on AlphaFold, another machine learning program that DeepMind introduced more than two years ago, which made it possible to publish the largest protein database.