To cure disease, AI needs more of our data

Machines have already outsmarted humans at playing chess, identifying birdsong and predicting complex protein structures. But when it comes to the really clever and intuitive stuff, like original scientific research, we humans like to think that we still have the advantage.

We may need to think again. At the RAAIS artificial intelligence conference in London earlier this month, Daniel Cohen, president of the Canadian drug discovery company Valence Labs discussed the tantalising, if slightly unnerving, possibility of “autonomous scientific discovery”. Trained on specialist data, sophisticated AI models might soon be able to generate hypotheses, design and run experiments, learn from the results and rinse and repeat 24/7. “Our mission is to industrialise scientific discovery,” he said.

You do not need to talk to people in computational biology for long to understand their excitement about AI. The AI research company Google DeepMind has even spun off a separate company, Isomorphic Labs, to exploit this domain after its AlphaFold program modelled 200mn protein structures.

The promise is that computational biology can help advance scientific research, accelerate drug discovery and improve patient outcomes. Machines have a number of advantages over their flesh-and-blood researcher and lab assistant counterparts. For one thing they do not need to sleep, deal with colds, hangovers or messy relationships.

“I am so encouraged by the pace at which the field is moving,” Christina Curtis, professor of genetics and biomedical data science at the Stanford University School of Medicine, tells me. “This is changing how we understand disease, how we detect malignancy and how we treat and intercept it.”

Curtis was the senior author of a paper, published in Science last month, that explores the heritability of malignancy in various subsets of cancer. Using machine learning techniques, the researchers parsed thousands of genomes from individuals with pre-invasive and invasive breast tumours to explore differences in their immunological response to the disease. They found that the way tumour cells evolved in individuals was “sculpted” by the germ line genome they inherited at conception.

Such research might lead to earlier detection and more personalised treatments, improving the chances of survival. “More than 50 per cent of cancer diagnoses are stage 4 or beyond. We are getting information too late to aid decision making,” Curtis says. “Ideally, we can do this more pre-emptively.”

There are two big constraints. The first is that “genetics provides hints not answers”, according to one industry executive. Machines have flagged plenty of targets for drug development, but few successful products have been released. Even if the technology does lead to scientific breakthroughs, it takes many years to win regulatory approval for new drugs. 

Thore Graepel, the global lead for computational science at Altos Labs, previously helped develop the AlphaGo program at Google DeepMind. AlphaGo’s defeat of the world’s strongest player at the ancient game of Go was seen as a mind-blowing breakthrough in machine intelligence. But Graepel told the RAAIS conference that the biological complexities he now confronts in cell rejuvenation were “orders of magnitude” greater. “I have never seen so much complexity with so little data,” he said.

The second constraint is data sparsity. Curtis argues that patient data is like “liquid gold” for researchers but we do not yet have the mechanisms to capture it routinely. Of most use would be to combine a patient’s genetic information with longitudinal health data gathered throughout their treatments and lives.

Reorienting healthcare systems towards early monitoring and prevention and away from late diagnosis and treatment will require a monumental transformation of cumbersome organisations. But Britain’s Labour party, which appears poised to win next week’s general election, promises to accelerate this transformation in the National Health Service. Labour’s manifesto pledges to create a “Fit For the Future” fund to double the number of CT and MRI scanners to detect early-stage cancers.

Voters are rightly sceptical of politicians making big promises. But the strains on public finances in ageing societies may soon leave governments with no option but to follow this route. As the Dutch philosopher Desiderius Erasmus supposedly told us five centuries ago: “Prevention is better than cure.” To that end, AI may be among our greatest assets.

john.thornhill@ft.com