Forecasting cancer - what's over the horizon?

28 Sep 2018

by ecancer reporter Will Davies

The sounds of the shipping forecast will offer soothing sounds to any radio-listener in the UK, a fixture of British airways for 151 years. While weather prediction technology has improved exponentially over the last century-and-half, accuracy of long-range predictions is only slightly more hit than it is miss, and many remain guarded in putting much trust in them...

While, at first, it sounds like a cryptic crossword read aloud, to a trained ear there is a wealth of information within to let you know what's about to happen, and where.

Lundy Fastnet Irish sea. Westerly 3 or 4, backing southerly. Good, occasionally moderate.
German Bight. Southwest 5 to 7, occasionally gale 8. Moderate or rough. Mainly fair. Good.
Thyroid.  Right local. Resectable. Good.

Which it is why the equivocation of prognostic models to chart tumour development strikes me as a bold comparison. I mentioned as much to Prof Trevor Graham, Barts Cancer Institute at Queen Mary University of London, whose work on characterising tumour evolution is described as such.

“Yeah, it's always amusing” he replied. “I've talked about it in public events a couple of times recently and talking about how wonderful weather forecasting is always brings a laugh. But I think it's actually true.  Weather forecasting is one of the best ways that we have to predict the future. I think one of the other forecasting success stories is in finance, although that’s had some high profile failures...”

Prof Grahams research, published in Nature Genetics, is part of an ongoing collaboration between his lab and those of Andrea Sottoriva (Centre for Evolution and Cancer, Institute of Cancer Research, London, UK) and Chris Barnes (UCL Genetics Institute, University College London, London, UK), whose interdisciplinary approach combines biology, computing and mathematics to develop models with which tumour evolution might be predicted, similarly to barometric weather stations feeding raw data to the Met Office for analysis.

“We're an unusual lab for the institute, because we are half theory-based. Half the people in the lab are from backgrounds like mathematics , theoretical evolutionary biology, physics, computer science, things like that. It’s that kind of quantitative understanding that’s at the heart of the research that we do. And so we’re interested in using these mathematical, and evolutionary tools to make sense of the cancer genome.”

“We can say with high confidence what the weather is likely to be in the next few days, and the reason the forecasting works is they combine fantastic data; we measure around the country, in all the right places, the temperature, the humidity and those kind of things and... But that's only half a story. The other half of the story, which is critical, is the mathematical models that tell us how to make sense of all that data.”

“If we had either part alone, we can't make the forecast. If we only have measurements with no ability to make a forecast, imagine what a terrible forecast it would be to say: it's sunny today, therefore, it’ll be sunny next week. Correlation alone is not enough, and it's the fact we have the mathematical models that tell us how the atmosphere changes over time that let us make a forecast.”

If you substitute “changes in atmosphere” to changes in tumour microenvironment, proteome or transcriptome, that is where the team’s modelling research sits.

“That’s the gap that our work is trying to fill; to make the mathematical descriptions of how a tumour changes over space and time. And be able to seed of all this data that we already have into those models to make predictions about the future.”

By assembling computational models of cancer cell divergence into subclonal types, and delving into the data from high-throughput sequencing of tumour samples, the team is able to measure the internal clone dynamics of a tumour. This means that the possible branches on a tumour’s evolutionary tree can, in theory, be determined, and sorted for likelihood and fitness advantage.

In the models, the earliest-emerging subclones, and those with the largest selection advantage over the original cells, produced detectable deviations in tumour composition. Applying a Bayesian framework to produce estimates of the parameters by which a synthetic ‘tumour’ could evolve was used to validate the predictive capacity of the model in examples of increasingly complex tumours.

“This work is in its infancy. But we hope to develop it in the future to really enable personalised medicine but for the individual tumour, or we can think about how best to act, when to give various treatments, say, based on a prediction of how the individual tumour is going to change in the coming days, months, years.”

As the team continues to refine the mathematical models, and apply it to broader tumour types, other labs around the world are turning to machine learning to empower their algorithms to improve themselves. Prof Graham was optimistic of integrating these big data techniques with the work already done by the team, and extending their forecasting capabilities far and wide.

“I think a commonality between all of these big data approaches is to find hidden patterns in the data and our approach is to try to learn the underlying mathematical rules that produce the patterns that we see. And then, once we know what those rules are, we can start to do forecasting and project forward about what the future will look like, and I think things like AI will help us find more of these rules more quickly, the nuances around actually how the rules work.”