The Center for Data Innovation spoke to John Cassidy, co-founder and chief executive officer of Cambridge Cancer Genomics (CCG), a British company combining liquid biopsies with artificial intelligence (AI) to support improvements in cancer treatment. Cassidy talked about how cancers change over time, and how AI can help doctors to keep on top of tumors that become resistant to previously effective treatments.
This interview has been lightly edited for clarity.
Nick Wallace: What’s a liquid biopsy, and where does AI come in?
John Cassidy: Cancers are really heterogeneous. They’re all very, very different. The core tenet of cancer biology for the last 30 or 40 years has been trying to understand what particular flavor of cancer you have, so that we can better treat it. Traditionally, this is done by taking what’s called a solid biopsy. This is a sample of the solid tumor tissue, and then you do various stainings and histological analyses to understand what kind of cancer you have.
Unfortunately, the process of taking a solid biopsy is not always possible. So what a liquid biopsy tries to do is take information that’s present in a patient’s blood sample, and try to infer the same sort of information that you would get from a solid biopsy.
In particular, what we’re interested in is next-generation sequencing of liquid biopsies, so that means that these are pieces of DNA which are shed into the bloodstream by the solid tumor, and we can then detect the bits of DNA in the blood and try and understand what the cancer is doing.
This is where AI comes in. Cancers are not static beasts. They’re constantly evolving, constantly changing in response to therapy and other factors. What we do at CCG is take several liquid biopsies over time and then use machine learning to understand the trajectory of your tumor’s evolution, and understand what therapies may be useful to treat your tumor in the future.
Wallace: How does this method help patients, compared to conventional methods?
Cassidy: Our company is founded on this core tenet that tumors are ever-evolving beasts. The same therapy that’s useful on day one may not be the therapy that’s useful on day 60. What we try to understand is how your tumor is evolving and how different drugs are appropriate to use at different times throughout your tumor evolution, to best combat it. So yeah, sure, you’re on drug A at the start of your therapy, and it seems to be working—it’s all about understanding the day that it stops working, and on that day, what is the best drug to give you. That way, patients can avoid being on really toxic chemotherapies for long, long periods of time, when they’re just not having any use.
Wallace: How does one go about training an AI system to carry out this kind of work?
Cassidy: The data that you get from next-generation sequencing can be very difficult to use for machine learning and artificial intelligence. That is because you have a really sparse, high-dimensional data set to work with. Imagine that we have 20,000 genes. In order to get a square matrix of data which is most amenable for machine learning, you need 20,000 patients on the other axis, which is very, very difficult. So a lot of the practical steps we take are in reducing this dimensionality and reducing the sparsity of the data, creating one-hot vectors, and trying to make genomic data amenable to lots of machine learning techniques.
There’s a lot of different places we go to to get data. There’s lots of large public repositories or government-funded repositories like The Cancer Genome Atlas (TCGA), and the International Cancer Genome Consortium (ICGC), where people have uploaded sequencing data from solid tumors primarily. We use a lot of that data to train risk metrics and things like that, and to try to understand what mutations are important and what genes are important for what features. When it comes to taking the sequential liquid biopsy data, that is almost entirely our own data. We collect blood samples from patients every month or every two weeks, we collect the plasma from those blood samples, and we use next-generation sequencing to read-off the genomic information from the blood sample.
Wallace: We’ve all heard a lot about healthcare innovations involving AI, but there still seems to be quite a gap between what’s possible and what’s available to patients. The future, in other words, is not evenly distributed. What are the obstacles to wider use of AI in frontline medicine?
Cassidy: There are a few obstacles. We’re in a bit of a hype-cycle of artificial intelligence just now. My major concern is that some AI program will be rolled out nationwide and the UK’s National Health Service (NHS) will adopt it and everything, and it won’t live up to expectations, and policymakers and researchers will turn off AI, they won’t like it anymore, and they’ll forget about it for ten years.
I think in order to avoid that, we really need to stop thinking about artificial intelligence as this catch-all that’s going to solve all problems, and start thinking about the fact that all an artificial neural network does is make a decision that a human can make, given the right amount of data. We need to be really careful in the promise that we’re putting on artificial intelligence.
Secondary to that, there are some problems in regulation, and there are some problems in uptake in hospitals. Regulations, for example: the U.S. Food and Drug Administration (FDA) takes a really dim view of what they call “black box technique”—these are techniques where we can’t understand the steps that an algorithm went through to arrive at an answer. Again, if we think about neural networks, any machine learning research out there will know that it’s pretty difficult, in some cases, to understand why the algorithm came to the conclusion it came to. The FDA, in its current regulatory environment, doesn’t like that kind of thing. Which is fine. It’s something we need to work through as a society.
The other problem is delay in up-take in hospitals. Some of that delay is necessary, and it comes from doctors not trusting technologies until they’ve been used. Which is great, that’s what we want our doctors to do. The problem is that a lot of doctors, I think, fear what AI is going to do to their jobs. For that reason they can be a bit overly reluctant to work with companies and researchers who are working with machine learning and artificial intelligence. For the foreseeable future, AI is only ever going to be a kind of diagnostics augmentation platform, where it helps doctors make better decisions, but doctors are still required. It doesn’t replace doctors, it’s just a tool to make their decisions more accurate. If doctors can get into that mindset, the barriers will be reduced somewhat, I think.
Wallace: Most of the big burning questions about AI are about the future. So’s this one. Looking at your field, where is it going? What effect is AI going to have on the way we experience or think about cancer?
Cassidy: I hope that AI helps us to stop thinking about cancer as a static disease and start thinking about it as a dynamic, chronic disease, that we cannot necessarily cure, but we can manage over a long, long period of time, until some other problem takes people away. I think globally, in medicine as a whole, in the near or short term, we’re going to see lots of decision-making tools built using artificial intelligence or machine learning that help doctors do their job better. It’s about taking the wisdom of the crowd—hundreds and hundreds of doctors—before somebody, and presenting that information to a junior doctor and saying, “this is what the experience and the AI have come up with, now use this information as another parameter in your decision-making processes. Take it with a pinch of salt—take it in the same way you would take any other screening tariff— but use it to help you make a better decision.
I think in the distant future, or hopefully the not-too-distant future, we’ll start to see whole new applications of AI, in terms of drug repurposing, treating each patient as an individual, and encompassing lots and lots of different data points about that patient.
I was at a conference in Singapore last week, where people were very excited about the possibility of lots of different data sources, and the idea that smartphones were going to solve everything. Which raises a very important point: 30 years from now, the patient that comes in to see the doctor is going to come with terabytes and terabytes of data, and we have to understand how to integrate that properly, and how to understand that data properly, in order to make better decisions.