AI tool helps avoid unnecessary prostate cancer treatment

More than 50,000 cases of prostate cancer are diagnosed in the UK every year. It’s second only to lung cancer as the leading cause of cancer mortality in men.

But most men diagnosed with the disease will die with it, rather than from it. Which means the quandary confronting patients when they receive a prostate cancer diagnosis at the moment is whether or not to aggressively treat the disease and run the risk of unnecessarily suffering a range of life-changing and serious side effects.

But now a new AI-based approach has helped scientists at Oxford University to unpick the genetic changes that occur in a tumour as prostate cancer develops, and identify a set of changes that single out cancers that are destined to be more aggressive.

This could help doctors provide much better advice to patients around how to manage their disease. I’ve been speaking to the lead researcher Dan Woodcock…

Dan - The current standard of care involves taking a biopsy and then performing a grading. So a pathologist looked down a microscope at what the cells look like, then assigned it a grade, and that grade is really good at identifying very high risk patients and very low risk patients. But the intermediate risk category, that doesn't really supply very much information on.

Chris - And I presume that most people probably do fall into that intermediate category then?

Dan - That's correct.

Chris - So we've got basically a big problem making accurate prognostications for people who want to know, do I really go for it and get this treated aggressively or can I afford to take this a bit more slowly?

Dan - Yeah, that's right. People with cancer often obviously err on the side of caution and opt for that treatment, which they may well not need.

Chris - So how have you approached this then? How have you tried to get at that middle ground?

Dan - Cancers don't really happen overnight. They start off as a normal cell, then genetic alteration, DNA mutation, will occur over time. More and more mutations will accumulate in that cell and then it'll go malignant and then it'll start growing uncontrollably and ultimately it will metastasise and then kill the patient. So we took DNA sequences from over 150 patients, tried to evaluate which mutations were in those patients at all these different stages, and then put that information back together and that basically gave us a better understanding of how the tumour evolves and what might distinguish aggressive disease or high risk disease from the more low risk ones.

Chris - The human body's got thousands of genes in it, though. So how do you look at all those moving parts all at once to try and spot the pattern as to what is evolving into what?

Dan - Well, we developed a kind of AI algorithm able to identify patterns of genetic alterations that are actually linked with the progression of the disease and ultimately the outcome.

Chris - Ah, so you can use AI to spot what's changing when, consistently?

Dan - Yep, that's right. So we use the type of interpretable AI so we can gain the information back from the computer about what's going on at the genetic level, and then use that information to further our study.

Chris - And what did your AI tell you?

Dan - The AI revealed there was two different routes of progression. One of these routes is a standard route of prostate cancer development, it's relatively slow growing, and after a number of years, perhaps decades, the cancer will get to a stage where it will metastasise and become a more dangerous form of the disease. What we also found is that there's an alternative evolutionary route, more like a fast track for more aggressive disease, so it reached the aggressive state a lot quicker. If you think of it like being stuck in a traffic jam on a motorway, you're very slow going, the motorway is full of traffic, but then you can see, okay, I can take this route off here, go through some of the side lanes and get to my destination much faster.

Chris - And so presumably if you know what those side roads are that take you on the fast track to a tumour, you could then apply that to, if you've got a biopsy from a patient, and say, well, is someone showing signs that they're on those side roads, therefore their disease is likely to become aggressive more quickly? And that would be useful to the patient because they'd know to go in and treat that disease more aggressively?

Dan - Yes, that's exactly right. So what the interpretable AI showed us was the features of those side roads, really, so what routes those side roads can occupy and what DNA mutations characterise those side roads. And then you're absolutely right, we can use that knowledge to then extrapolate what individual tumours patients have and then tell them what type of disease, what road they're going down, and then give them a better treatment outcome.

Chris - I was going to ask you, if you've now got almost like the map of what those side streets that take people to a more aggressive cancer are, does this mean you can put roadblocks on them? Does this show you possible therapeutic avenues to block those pathways in different ways or in better ways than we do already, or even that we can do already but we just don't think to use them in certain patients, so that we can optimise someone's care?

Dan - Yes, it does. So in patients where prostate cancer's already metastasised, so spread to other organs, we use something called hormone blockade to slow down the progress of that disease. The tumours like taking the fast track, they've already learned a different way to process these hormones. It makes sense that by using hormone blockade it will actually be even more effective on those tumours and that's something that we're looking into.