Nicholas D. James, PhD, FRCP, MBBS, on Using AI to Identify Benefit From Prostate Cancer Therapy

It's a great pleasure to be here to present some more data from the STAMPEDE trial. So the particular data we presented today relates to men included in the two abiraterone comparisons in the trial, abiraterone on its own and abiraterone combined with enzalutamide. We presented the primary results of this a few years ago and what they show is that for men with non-metastatic disease, the addition of two years of abiraterone with or without the second drug improves metastasis-free survival and overall survival substantially. So metastasis-free survival has improved, hazard ratio of 0.5—a 50% improvement—and risk of death is reduced by 40%, hazard ratio 0.6, so big effect. However, a lot of the men actually do fine on the control arm. So these are non-metastatic patients, they don't all relapse and significant proportions are cured. So those patients are therefore not benefiting from having an extra drug. They can only be harmed because they were going to do fine without it. So what we've done is we have access to the digitized H&E slide archive and we've collaborated for a long time with Artera and they've developed an artificial intelligence tool which was initially trained on patients with low-risk prostate cancer having radiotherapy with or without hormones in various settings. And the tool is approved in the United States and recently approved in the UK for purpose of identifying patients who are benefiting from hormone therapy. Now in this particular case, all of the patients are getting standard hormone therapy, androgen deprivation therapy, and we're adding a second hormone therapy on top of it. So we wanted to know two things. One is, would the tool behave as a prognostic tool in this population with much higher-risk prostate cancer? And the second thing is, could it identify the patients who are benefiting from abiraterone? So firstly, if you run the MMAI tool, as it's called, over men with advanced prostate cancer, either locally advanced or metastatic, they all come out with high scores compared with the patients with lower-stage prostate cancer, kind of as you'd expect. What we then did was we then did some modeling to say was there a cut point that we could put onto the scores that separated men into different prognostic groups. And the answer was that if we separated the patients at the top quartile—so we took the top quartile off and said these were high scores—and said everybody else was a low score at the bottom three quartiles, that did indeed separate, gave us the maximum separation within our prognostic groups. So either high-risk localized or node-positive. Even more interesting though is when we took the men in the top quartile and looked at whether there was benefit or not from abiraterone. That group defined almost all the patients getting the benefit from abiraterone. So the men in the lowest three-quarters were deriving very little benefit from having abiraterone added on top. So not only was it giving us valuable prognostic information, but it was also identifying patients who would benefit from having the second drug added on top. So this is really important because the test is already available, it's already approved—not for this particular indication—but it means we would anticipate that these data should extend the use into this very important population. Now in the UK, this is of particular relevance to us because at the moment abiraterone is not available in England for men with this stage prostate cancer, even though there's a halving of the risk of spread. And so we're going back to NHS England to present these data to them in the next few weeks saying, well, we now have a way of identifying a way of not needing to even consider funding the drug for three-quarters of the men with the disease, and you'll still get almost all of the survival benefit by just giving the drug to a quarter of them. So it's got very important and immediate practical implications. But also the thing we're really excited about is that we've previously done various other analyses on these same samples in the lab looking at RNA expression, profiling various gene mutations and things. And that will split the same group into prognostic groups, but it does not predict benefit from abiraterone. So the AI is identifying some aspect of the biology that we have not been able to identify so far. So we will of course be going back to these samples to interrogate to try and figure out what the underlying biology is that the AI tool is finding. And so that potentially opens up routes to, we hope, new druggable targets. So we're super excited about it, both the immediate practical implications for patient care, but also for the new research avenues it opens up.