Researchers at Google DeepMind are using artificial intelligence to predict whether mutations in human genes are likely to be harmful, in one of the first examples of artificial intelligence technology helping to speed the diagnosis of diseases caused by genetic variants.

The artificial intelligence tool, called AlphaMissense, evaluated all 71 million "missense" mutations, in which a single letter in the human genetic code is changed. Of these, 32% were classified as possibly pathogenic, 57% were benign, and the remainder were indeterminate. The findings were published Tuesday in the journal Science.

Meta CEO Mark Zuckerberg announced on Tuesday that the Chan Zuckerberg Initiative, the philanthropic organization he co-founded with his wife Priscilla Chan, will build "one of the largest computing systems dedicated to nonprofit life sciences," illustrating the investment of resources in artificial intelligence in the life sciences field. It will focus on using artificial intelligence to simulate what happens in living cells.

Human experts have so far discovered only 0.1% of the clinical impact of missense variants that alter the structure of proteins, the body's main working molecules. "Experiments to discover disease-causing variants are expensive and laborious," said Žiga Avsec, a project researcher at DeepMind's London headquarters. "Each protein is unique and each experiment must be designed individually, which can take months. By using AI predictions, researchers can preview results for thousands of proteins at once, which helps prioritize resources and speed up more complex studies."

"We should emphasize that these predictions were never really intended to be used for clinical diagnosis alone," said Cheng Jun, also a researcher on the project. "They should always be used together with other evidence. However, we do think that our predictions will help improve the diagnosis of rare diseases and may also help us find new disease-causing genes."

AlphaMissense predictions show mutations in two protein structures (see another image). Red is harmful, blue is benign, gray is uncertain

The British government's Genomics England tested the tool's predictions against its own extensive record of genetic variants that cause rare diseases, and the results were impressive, deputy chief medical officer Ellen Thomas said.

"We were not involved in the development of the tool and did not provide the data to train it so we could independently evaluate it," Thomas said. "It's completely different from tools we're already using. I think it's a huge step forward and we're excited to be involved in the final stages of considering using this tool." She hopes AlphaMissense will be used in healthcare to be "a co-pilot for clinical scientists, flagging the variants they should be concerned about so they can do their jobs more efficiently."

DeepMind developed AlphaMissense based on its AlphaFold tool for predicting protein structures. The AI ​​tool also learned from a wealth of biological evidence about the signatures of genetic mutations in humans and other primates that make genetic variants pathogenic or benign.

The company, founded in 2010 as a professional artificial intelligence developer and acquired by Google in 2014, has made the tool "free to the scientific community." Its predictions will be incorporated into the widely used EnsemblVariantEffectPredictor run by the European Bioinformatics Institute in Cambridge.

AlphaMissense also has limitations, Avsec said. Most importantly, its predictions of pathogenicity "are only general and do not tell us anything about the biophysical properties of the variant." He added that these insights may emerge more clearly as the tool is further developed.

Sarah Teichmann, head of cytogenetics at the Wellcome Sanger Institute in Cambridge, who was not involved in the study, said that while individual missense mutations are important causes of disease, other clinically significant changes in DNA are beyond the scope of the tool.

"We shouldn't exaggerate and say this is going to solve everything," she said. "But having such powerful interpretive AI to integrate so much genomic data is really an improvement."