Groundbreaking AI by DeepMind accurately predicts harmful genetic mutations in humans

Receive free Artificial intelligence updates

We’ll send you a myFT Daily Digest email rounding up the latest Artificial intelligence news every morning.

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

​The AI tool, called AlphaMissense, assessed all 71mn ‘‘missense” mutations, in which a single letter of the human genetic code changes. Of these, 32 per cent were classified as likely to be pathogenic, 57 per cent benign and the rest uncertain. The findings were published on Tuesday in the journal Science.

Illustrating the resources being poured into AI for life sciences, Meta chief executive Mark Zuckerberg announced on Tuesday that the philanthropic group he founded with his wife Priscilla Chan, the Chan Zuckerberg Initiative, would build “one of the largest computing systems dedicated to non-profit life sciences”. It will focus on using AI to model what happens in living cells.

Human experts have so far discovered the clinical effect of just 0.1 per cent of missense variants, which change the structure of proteins, the body’s main working molecules. “Experiments to uncover disease-causing mutations are expensive and laborious,” said Žiga Avsec, a researcher on the project that was based at DeepMind’s London headquarters.

“Every protein is unique and each experiment has to be designed separately which can take months,” Avsec said. “By using AI predictions, researchers can get a preview of results for thousands of proteins at a time, which can help to prioritise resources and accelerate more complex studies.”

“We should emphasise that the predictions were never really intended to be used for clinical diagnosis alone,” said Jun Cheng, also a researcher on the project. “They should always be used along with other evidence. However, we do think that our predictions will help to increase the diagnosis rate of rare disease and also potentially to help us find new disease-causing genes.”

The UK government’s Genomics England tested the tool’s predictions against its own extensive records of genetic variants causing rare diseases, and was impressed by the results, said Ellen Thomas, deputy chief medical officer.

“We were not involved in generating the tool or in providing data to train it, so we could give an independent assessment,” Thomas said. “It is completely different from the tools we already use. I think it’s a great advance and we’ve been pleased to be involved in just the final stages of thinking about using the tool.”

Thomas said she expected AlphaMissense to be used in healthcare as “a co-pilot for clinical scientists, flagging which variants they should be focusing on so that they can do their jobs more efficiently”.

DeepMind built on its AlphaFold tool, which predicts protein structure, to develop AlphaMissense. The AI tool also learnt from a vast amount of biological evidence about the characteristics of mutations in humans and other primates that make a genetic variant pathogenic or benign.

The company — founded as a specialist AI developer in 2010 and bought by Google in 2014 — has made the tool “freely available to the scientific community”. Its predictions will be incorporated into the widely used Ensembl Variant Effect Predictor run by the European Bioinformatics Institute in Cambridge.

AlphaMissense has limitations, Avsec said. The most important is that its predictions of pathogenicity “are made in a general sense and do not tell us the biophysical nature of what a variant does”. Those insights might emerge more clearly as the tool is developed further, he added.

Sarah Teichmann, head of cellular genetics at the Wellcome Sanger Institute in Cambridge, who was not involved in the research, said that while individual missense mutations were an important cause of disease, other clinically significant changes in DNA lay beyond the scope of the tool.

“We shouldn’t exaggerate and say this is going to solve everything,” she said. “But it is a real advance to have such a powerful interpretive AI integrating so much genomic data.”