Last year, the Nobel prize for chemistry honoured the use of AI to predict the structures of proteins. It was a sign of just how far and fast the technology is driving progress in that field. Our understanding of some fundamental questions in biology is advancing rapidly. The question is, how can we make similar progress when it comes to the prevention of disease and the expansion of good health?

Both parts are important. Disease and sickness have been the main focus of healthcare, for obvious reasons. But as the discipline’s very name suggests, it could and should also focus on the care of health: on improving our levels of “wellness” and extending our “healthspan”, the number of years for which we can expect to enjoy happy and active lives. Some years ago, my colleague Leroy Hood and I conducted the first big studies of thousands of people from the point of view of wellness. We monitored genomes (the set of genetic material that each of us carries), proteomes (all the proteins that the genome encodes) and metabolomes (all the small-molecule chemicals that a person creates).

Since then, similar efforts have arisen elsewhere, such as the UK Biobank, the Emirati Genome Programme in the United Arab Emirates and the Institute for Healthier Living Abu Dhabi, where I am an advisor. But right now we don’t have nearly enough data about how health changes across a person’s life — and data is the essential input to AI systems, which could help us to uncover patterns that are crucial to our good health. With these systems, you can do a very different style of data analysis by constructing “digital twins” — digital models of individuals that monitor the person’s health trajectory so that they can receive personalised health programmes.

To that end, my lab is building something I call the N of 1 Analyzer, “n” being the abbreviation in statistics for “sample size” — so the name means that it’s analysing just one person, not the large numbers more typical of scientific studies. The point of the N of 1 Analyzer is to discover how much I can say about an individual’s health and the treatments likely to help them most, based on individual measurements — something that has not been doable to any great extent until now. It will do this by using AI to do “deep reasoning” on the details of what’s going on in a person’s body.

What is deep reasoning? Historically, we have used statistics to study biological information or to discover biological markers for disease. With AI, however, we can go further, and exploit everything we know about what we’re measuring, not just the measured data itself. Large language models can interpret our knowledge about what a protein, metabolite or gene does in the body, and this can be combined with the results of statistical analyses of proteomes, metabolomes, epigenomes and genomes. This context-aware data analysis is what I call deep reasoning and unlocks profound new ways of looking at health.

Consider how it changes the way we think about biomarkers, one of the central concepts in modern medicine. As soon as I start looking for a biomarker for a disease, I’ve made a huge assumption: I’ve assumed that the disease always follows a trajectory where the biomarker is involved. But diseases can actually follow a variety of trajectories, and unless these are being monitored, it’s not clear what they look like. With deep reasoning, the large language model has context about disease processes that it draws from countless scientific papers, which it can use to reason about the changes that a particular individual is showing. Result: real personalised medicine.

Or consider gene variants that give you high or low risk for a disease. Let’s take as an example the gene APOE, whose variant APOE4 gives you high risk for Alzheimer’s, whereas APOE2 gives you low risk. But this difference only manifests in later life — around your late sixties. Then, people with the APOE4 variant can struggle to remember their children, while those with APOE2 are fine. Why? My favoured explanation is that it’s because the APOE gene ultimately dictates how fast cholesterol is transported out of certain brain cells. If the rate of transport is slow, as it is in people with APOE4, your neurons will die when short of oxygen — and your ability to supply oxygen to your brain declines as you get older.¹ So now we have a mechanism that explains why APOE4 matters when you’re older but not when you’re young.

If that is indeed the mechanism, I can use that information to understand my personal resilience to the threat of Alzheimer’s and to take protective measures. For instance, among people who use saunas four to seven times a week the rate of dementia drops by 65 per cent² and the rate of cardiovascular disease by 50 per cent.³ The same is true for other conditions. For diabetes, for example, you would want to check three things that are involved in the early stages of the disease: how well your beta cells are working, how well you control glucose levels and where in your body you are depositing fat.

There are limitations to this digital-twin approach. Any model of a real system is an approximation. We will not be able to be able to represent the full complexity of human biology any time soon: there’s still an incredible amount that is unknown. There’s also a whole philosophical discussion as to the ways in which the analysis amounts to actual “reasoning”. I think it’s not — it’s actually an approximation based on human language — but the salient point is that it provides different, potentially complementary insights to a statistical average.

Looking ahead, there is an opportunity to develop a new kind of science based on data from digital twins. To date, most personalised medicine has catered to the rich, who can afford the extensive physical testing needed to build up personalised data. But there are now comparatively cheap ways to get personalised data. Large language models can set up personalised systems without the need for physical testing, and we can push AI hallucinations in this space to near zero. You can tell the system about yourself — and we can put privacy safeguards in place so that information remains with you and doesn’t go elsewhere — and get advice about, for instance, insomnia: what could help me sleep better? How important is it? Should I worry about it? Or, my grandpa suffered from Alzheimer’s: should I worry about that?

A good next step might be to focus on densely phenotyping a million people — that is, recording their detailed individual characteristics — and taking measurements every few months so that we can understand the trajectories of diseases. Over the next year or so my group will be working with various groups at that scale, although not yet gathering all the deep, longitudinal information that’s theoretically available. But that can still be valuable. Under a project called Personalized Analytics for Transforming Health Care (PATH) from the US Advanced Research Project Agency for Health, monitoring people with diabetes has revealed that there are maybe three dominant paths in there. That’s a fundamental and significant feature that we didn’t understand before.

Making that knowledge more widely useable need not be expensive: perhaps using home testing kits based around disposable paper-based devices and mobile phones. Some biomarkers may correlate reliably with a feature that can be measured by phone — a measurement or statistical analysis of features on a face, say. As people become used to such tracking, and more appreciative of its value, they will derive more and more benefit from it — and so will researchers, since that means they will have access to the long-term datasets needed to take the next leap with healthcare AI systems.

A decade from now, we can anticipate that on the one hand there will be huge amounts of data streaming in from people taking charge of their own wellness, and on the other huge amounts of information about fundamental processes and mechanisms unearthed by automated experimental systems. Governments should understand these opportunities, because few other entities have the funding and staying power that it will take to do this work. Already the UK, the United Arab Emirates, China and, to a limited degree, the US are putting money into this area. There’s a race on right now, an important one, to develop tools to realise the potential in a big way — and ensure that healthcare really does come to mean taking care of health.