Google DeepMind's AI tool could pinpoint our genetic faults

Google DeepMind has showed that its tool AlphaFold could predict the structure of any protein from its genetic sequence alone.

It’s becoming increasingly clear that scientists are only scratching the surface of what artificial intelligence can teach us about human biology and disease. Case in point: new work by Alphabet's AI subsidiary, Google DeepMind, that promises to help winnow down the genetic causes of disease.

Recall that it's only been a year since DeepMind researchers knocked the socks off the scientific world by unleashing a database with the predicted structure of more than 200 million proteins — kind of like a family photo album of nearly every protein found in every organism on Earth. Now, they’ve built off that work to develop a tool that can pluck harmful genome mutations from thousands of minute-but-benign ones. The tool, called AlphaMissense, is not the quantum leap we saw last year. But it’s a harbinger for other positive advances that could spill out of AlphaFold and other machine-learning efforts.

Last year, Google DeepMind showed that its tool AlphaFold could predict the structure of any protein from its genetic sequence alone. It was important caveats as one of the most important scientific advances of the year, albeit with some important caveats. Researchers have busy plumbing it — to inform drug-discovery efforts, improve vaccine and drug design, and fill in some of the many blanks about the protein universe.

The magnitude of the advance might not have registered with the public, DeepMind’s CEO Demis Hassabis told The Verge, but the ability to translate a string of letters into a complex, three-dimensional form has had “the most unequivocally biggest beneficial effects so far in AI on the world.” He noted that, “Every Big Pharma company is using it to advance their drug discovery programs. I’ve had multiple, dozens, of Nobel Prize-winner-level biologists and chemists talk to me about how they’re using AlphaFold.” (The tool is free.)

Now, DeepMind has tweaked AlphaFold to analyse those sequences for an entirely different task: identifying which single-letter changes in the genome might cause harm, and which are likely to be benign. The result is AlphaMissense, the details of which it revealed in a paper this week in the prestigious journal Science.

It’s not the first tool designed to identify relevant mutations from the long list of minor errors in our genomes, and certainly others will come after it. But it’s a welcome improvement over what’s out there — and one that might help solve genetic mysteries and even clear a path to developing therapies for rare diseases.

AlphaMissense tries to overcome another longstanding problem: Although genome sequencing has become far more accessible, researchers often encounter mutations that defy interpretation — amino acids are swapped in a ways never seen before, or a mutation hasn’t been sufficiently studied to determine if it’s pathogenic or not.

Families looking for answers to an abnormality are told, “we did your genetic testing and found a ‘variant of uncertain significance’,” says Timothy Yu, a neurologist at Boston Children’s Hospital. While Yu is skeptical that AlphaMissense could get rid of all uncertainty, “it should help,” he says.

DeepMind scientists stressed on a call with media that AlphaMissense isn’t offering an alternative to a genetic diagnosis. Rather, the tool is meant to be an additional filter when a researcher is confronted with a long list of possible mutations, allowing doctors like Yu to get an answer faster. Early intervention is critical to making a difference in patients’ lives, particularly in an era where other technologies make it increasingly possible to quickly design a bespoke drug to address a problematic mutation.

AlphaMissense isn’t the only recent AlphaFold advance worth watching. Another notable, less splashy development came earlier this month. Two
separate papers, published simultaneously in Nature, offered a way to organise AlphaFold's 200 million protein structures into something akin to a family tree. By connecting even the most distant protein relatives, that effort that could help scientists glean broader insights about human biology.

For example, by studying the structure of similar proteins, Martin Steinegger, a computational biologist at Seoul National University, was able to draw links between bacteria and proteins related to human immunity, while another team separately identified never-before-seen protein shapes. The hope is their efforts could help scientists understand the vast percentage of proteins whose function remains a mystery.

There’s been plenty of worrying about the damage AI could do, but discoveries like these show how AI can accelerate progress on the most vexing questions about human biology and health — and motivate scientists to keep pushing to see what the machines can help us do better next.

Lisa Jarvis is a Bloomberg Opinion columnist covering biotech, health care and the pharmaceutical industry. Views are personal, and do not represent the stand of this publication.

Credit: Bloomberg