Machine learning helps doctors stay ahead of the flu
In the northern hemisphere, flu season is once again upon us. Each year, as the virus circulates, millions are infected and hundreds of thousands of people die from the infection, especially the more vulnerable. Thankfully there is an international vaccine programme which uses samples collected from the opposite side of the world during their flu season to try to stay one step ahead of how the virus will evolve and change. Cambridge University’s Professor Derek Smith, an expert on the evolution of diseases like the flu, sits on the panel that designs those vaccines…
Derek - Flu is a respiratory virus. It kills, worldwide, about half a million people a year. It infects about 10% of the world's population each year.
Chris - Why do we have a flu season? What drives the fact that it always comes at a certain time of the year, give or take?
Derek - Well, we don't know. In places like the UK in the northern hemisphere where there's clear winters, it happens in the wintertime. But flu happens in Bangkok where it's 37 degrees year round, pretty much. In tropical regions like that, it usually happens during the rainy season. Now whether or not it's the amount of moisture in the air, whether it's the fact that during the rain and during the winter we're all in closed rooms together and we can transmit more easily, whether we don't see as much sunlight, it's really unclear and it's been an outstanding question for many, many years. Nobody knows yet.
Chris - And why do we keep catching it? Because if I catch measles, which I luckily haven't, I'd be immune for life. Why is that not the case with flu?
Derek - It's not the case with flu because the flu virus has adopted a lifestyle where it can change over time to evolve, to get mutations that can escape the immunity that we get. And typically we all get flu once every five or 10 years.
Chris - When it changes, what does it change about itself to mean that I can keep catching it?
Derek - It changes in a way that escapes the antibodies that we've built up against earlier strains in a very similar way to the antibodies that we all had against the early Covid strains, either from being infected or vaccinated. Those new variants as well, they've escaped that prior immunity, and flu does the same thing. There are critical amino acid substitutions on its surface in the area where the antibodies care about that can make it so those antibodies we have can't bind there anymore and neutralise the virus. So these new variants can reinfect us because our old antibodies don't work anymore
Chris - And that's why we need a vaccine every year.
Derek - That's exactly why we need a vaccine every year. And in many years, the variance of flu that we put in the vaccine, means they have to be changed to track the evolution of the virus.
Chris - You are one of the people who sits on the panel helping to decide what goes into those vaccines. How do you make that decision?
Derek - There's a fantastic worldwide network. I think it's about 150 countries worldwide now, taking throat swabs from people who look like they might have flu and, if it is, checking to see how different it is from earlier variants. This happens in those countries and it also happens by these viruses being sent to one of five laboratories around the world that do very careful analysis. One in Melbourne, Australia, Beijing, China, Tokyo, Japan, Atlanta in the US, and in London here in the UK. Those viruses are tested - as well as seeing how different they are from each other - to test how well our immunity will work against those viruses. And if our current immunity won't work well against the new viruses that are evolving, then we have to update the strain of flu that is in the vaccine.
Chris - What arrives in our winter is going to be six months out of phase with what was doing the rounds in the southern hemisphere. How do you know that what turns up here six months later from Australia is going to be what you think it is?
Derek - Yeah, so this is really the million dollar question because when we make the choice of which strain to put in the vaccine... for example, the strains that people are getting vaccinated with now to protect them against the flu, that will probably come in the next few months. We chose that strain in February this year because enough vaccines had to be made to vaccinate everybody. Currently I think there are about 700 million doses of vaccine made each year. We do the best job that can be done in February to figure out what's going to happen the next year. But it's not a perfect science yet.
... and they expect to be right about 40-70% of the time. But how can we improve on that? Well Simon Fraser University's Jessica Stockdale thinks we might be able to use machine learning to spot patterns in the evolving genetic code of flu that predict the next move the virus will make. As she shows in a paper this week in the journal Science Advances, she's used genetic data from epidemics in previous years to train her system, and then tested it using what we know the virus did more recently to gauge its accuracy. She was hitting the bullseye up to 95% of the time; used alongside existing approaches like those Derek mentioned, it could dramatically reinforce our present vaccination initiative...
Jessica - We take an approach which we can describe as building something like a family tree of our influenza viruses. So we collected about 30,000 public influenza sequences from between 1980 to 2020, and we built this phylogenetic tree, this family tree of those influenza viruses, and used the the structure of that tree, which tracks the mutations that are accruing in flu over time, to look at how the different flu viruses are related to one another to try and predict what was going to grow moving forwards. So our machine learning model looks at which flu viruses are around right now, it calculates various statistics on that tree; how fast are a small family of flu viruses growing? How large is this family? What's the shape to try and classify? Yes or no - will that group of flu viruses be a big problem for us in the next year?
Chris - I guess it's a bit like watching the flu dance floor and seeing what moves it makes, and if you know the dance, you can say, well, when it does that move, it tends to follow it with that move. And so you are doing that for the structure of the virus and therefore you've got some chance of making a guesstimate as to what it is likely to do next.
Jessica - Yes, exactly. It's a great analogy. And it's difficult because flu viruses don't always do the dance moves that we expect that they're going to do. But if we look over the whole globe of all of the different patterns that they're making, there are general patterns that emerge, we might hope, and that's what we're trying to predict.
Chris - Does it work though, Jessica? When you've taught your model using the past, is there not a risk that you end up answering the question that you wanted it to answer rather than what's genuinely going to happen? Can you test it going forward to make sure it's robust?
Jessica - Yes, it's difficult, but that's what we've been trying to do. We did five different experiments on the last five years of our data from 2016 to 2020. We tried to make a prediction, assuming that we did not know what happened. In the end, we masked from ourselves the truth and used that as a method to test our approach. And we found our model to be around 75 to 95% accurate at predicting what was going to be successful the next year.
Chris - The WHO reckon they get it right about 70% of the time, so that would put you at least as good as the current endeavour and possibly better.
Jessica - Yeah. So there's a couple of different ways of measuring success, but we found our approach to be pretty similar to that found by the WHO, despite the fact that we have some more limited data. We're limited to public data only, so that's why we hope that our approach would be a useful addition to the toolbox, I would say, in trying to do this vaccine strain selection.
Chris - So what's the computer doing that we can't with our current WHO panels and so on achieve, or we can but we can't do it as well as your computer system. What's the missing link that you are plugging into?
Jessica - I would say that a human or a person could do what the machine learning model is doing. It's just able to do these calculations a lot faster than us, and there are certainly things that the WHO are taking into account when they pick vaccine strains that we would still want to do, such as, 'is the strain we're selecting even viable for vaccine selection?' But this computer is able to be a helper for us to have this really fast thinking that can add into our own human intelligence that we're using to pick these strengths.