Dr Andrea Sottoriva introduces REVOLVER and considers the wider application of Artificial Intelligence in cancer research and care.
Developed by scientists at The Institute of Cancer Research (ICR) and the University of Edinburgh, UK, REVOLVER (Repeated Evolution of Cancer) is a machine learning technique with the ability to identify patterns in DNA mutation within cancers and forecast future genetic changes. The new method, which could be used to shape more effective treatments, is just one example of the application of Artificial Intelligence in cancer research and care, which could one day transform the way cancer is diagnosed, managed and treated.
To find out more about these possibilities, Health Europa spoke to Dr Andrea Sottoriva, a team leader in evolutionary genomics and modelling at ICR and the REVOLVER study lead. Here, he discusses how the new technique can be used to estimate the trajectory of tumour development and what remains to be discovered about the way machines learn.
How does REVOLVER exploit Artificial Intelligence to predict tumour growth?
We started by analysing genetic data from patient tumours. The data that come from studying the DNA of cancer cells from tumour samples are usually quite complex and noisy. It is very hard to make sense of them. We exploit a type of Artificial Intelligence called transfer learning, which is a type of machine learning. Transfer learning is an algorithm that tries to find hidden patterns in the data that are of interest.
We already know that tumours evolve through sequences of evolutionary steps. If we can find groups of patients with similar sequences that determine the evolution of their tumours, then we can exploit that knowledge in new patients. For example, we can identify sequences of evolutionary steps in a subset of lung cancers, which means that if a new patient comes along who has only an initial portion of such a sequence, we can make a prediction about what we expect to see next.
In which areas do you expect Artificial Intelligence in cancer research and care to have the most impact?
The first two areas targeted by machine learning and Artificial Intelligence researchers were computer vision and text and speech recognition. That began in the 1970s, so there’s decades of work on these two things. The majority of current machine learning approaches are therefore based on image analysis – either clinical or radiological images such as CT scans or pathological images such as microscopy stains. This is useful in order to simplify the work of a pathologist, for example. Instead of a human having to go through thousands and thousands of cells on his own, a machine can automatically recognise patterns in those cells and help with diagnosis.
More recently there has also been a focus on the analysis of quantitative data – that is, not just images but also health records or genetic data. That requires something of a different approach because it’s not about vision and it’s not about speech recognition. It’s about identifying patterns in a way that humans can’t. I don’t think what we are trying to teach these machines has an equivalent in humans.
I think in the future we will see more and more of this automatic image recognition but also more hybrid approaches like REVOLVER. We give a training set to a machine, ask it to look at, say, 10,000 images (in this case genetic data), tell it how to classify those images, and then it learns how to do that itself.
But we don’t just show the machine the genetic data, because we already know that cancers evolve, so we can interpret the data beforehand. Our approach integrates the kind of model-based approach you find in physics with the type of tools more usually seen in biology.
In physics, a theory makes a prediction and that allows you to analyse the data. In the case of REVOLVER, we still have the theory, but the noise and complexity of the data are too high for a solely physics-based approach. We need to find patterns in the data even though they are pre-processed.
What’s needed to unlock the full potential of Artificial Intelligence in cancer research and care?
There are some non-trivial algorithmic problems left to solve, which will require new methods and not simply more of the same. We also need better data. In the case of genetic data, for example, we need larger scale, higher resolution datasets. That’s the problem with machines: they learn from huge datasets. We talk a lot about big data in medicine, but very often we don’t actually have big data. Google has two million images, but we don’t have two million genetic profiles of cancers because fortunately we don’t have enough patients for any given cancer type.
We also need more basic research on applications – in other words, Artificial Intelligence that is not just classic computer vision or recognition.
Another really big obstacle is that we don’t yet know what a machine has learned. We can certainly teach a machine to do something interesting or classify images well, but we are still to learn the mechanisms behind that. That’s something of a black box, so that knowledge needs to be extracted somehow.
Looking forwards, where will your own research lie in this space?
We are specifically working on this overarching question of how to predict tumour evolution. At the moment we are building a set of methods and testing them. We try to tackle this problem from different angles; some of them involve Artificial Intelligence, some of them are just different types of mathematical modelling applied to the data.
Once we have built and validated this knowledge of what we can predict and how we can predict it, we will look at taking it into the clinic. We will start thinking about new types of clinical trial designs, for example, where we integrate predictions into the design of the clinical trial itself.
Dr Andrea Sottoriva
Team Leader, Evolutionary Genomics and Modelling
The Institute of Cancer Research
This article will appear in issue 7 of Health Europa Quarterly, which will be published in November 2018.