Alicia Morgans: Hi, I'm so excited to be here at ESMO 2022 where I have the opportunity to speak with Dr. Felix Feng of UCSF. Thank you so much for being here with me.

Felix Feng: Thank you, Alicia, for having me.

Alicia Morgans: Wonderful. Felix, you have really been pioneering in so many ways in the field of prostate cancer and biomarkers and trying to help us understand who's going to respond versus who won't. One of the most recent things that you've been looking into is AI, so artificial intelligence, and using that to really enhance our understanding of pathology in prostate cancer. Can you tell us a little bit about that concept and then maybe where we're going in terms of the field and AI in prostate cancer?

Felix Feng: I'd love to, and I think I'll start with the AI part and then I'll talk about the pathology part. And I think, well, even within the field of prostate cancer, but especially outside the field of prostate cancer, artificial intelligence as a technology to help personalize decisions regarding patients, it's here. It's already here. And the reason why it's here is because there's so much information that's been gathered about prostate cancer and other diseases over the last few decades. And fundamentally, I think the principle of AI is to try to leverage all that information to help make better decisions, to allow patients more information in terms of what therapies may or may not benefit them, and also to help physicians guide those patients along those decision making processes. And the most valuable data sets out there are phase III clinical trials. And these, as you know, are large clinical trials that have randomized patients to one treatment versus another treatment.

And they're large enough where the results from these trials are considered to be very, very, what we call robust, very trustworthy, and also very believable as well. Now, when you look back on phase III clinical trials and you want to create, let's say, biomarkers to help personalize therapy, you have to use the data that's still relevant nowadays. And it turns out that pathology slides as they were prepared two decades ago, three decades ago, are similarly prepared nowadays. And so you can go back to samples that are very old. You can take a picture of those pathology slides. You can use artificial intelligence approaches to try to optimize for, let's say, the clinical outcome in the case of who should benefit from a treatment versus who shouldn't. And you can basically create these classifiers that are still robust using today's data sets. That's the premise.

Alicia Morgans: And it's so interesting, and I remember this concept from years ago where there was development of AI to really look at skin lesions, for example, and to understand which of these skin lesions are melanoma versus not. And so these algorithms can be developed and can really inform us with our human eyes in things that we can't even detect or see, which can be obviously very, very valuable. What are some of the studies that you've done with your team?

Felix Feng: Yeah, so there's actually a lot of AI that's been done on, let's say, diagnosis of cancers, which is what you're referring to about skin cancers. And my team is more interested in developing tools that don't necessarily diagnose cancer, but tell a patient who has cancer what treatments they should or shouldn't get. And so a study we did last year basically was a study where we took data from five phase III clinical trials of patients who had localized prostate cancer and have been treated with different approaches involving either hormone therapy and radiation, et cetera, and have been followed more importantly for 10 to 20 years on these trials. And we were able to develop an AI classifier that predicted what the chances that they would have metastasis at five years and 10 years down the road was.

And this was significantly better than just the current standard clinical variables that we as clinicians use to try to estimate prognoses for these patients. And so for example, even within patients who have high risk prostate cancer based on their PSA levels and their Gleason scores, we can use this classifier and really improve within this high risk area, we can improve the prediction of who has really, really bad disease versus who has not so bad disease. That was our first study, and it was really the first study that was validated in samples from five phase III clinical trials.

Alicia Morgans: Wow. And I think that's something that is, as you said, just so useful in figuring out who can benefit from these different treatments. And one of the things that I think your group is also looking into that is of high interest in prostate cancer, and actually I think across medicine, is understanding disparities and understanding whether different groups that may be different by race or ethnicity or something else, whether they may have different outcomes. And I think you've presented some data, your team has presented some data along those lines as well.

Felix Feng: Yeah, no, so we gave a presentation at the ASCO conference earlier this year as well. And one of the strengths of AI is that it works very well in the data sets that you can develop these classifiers on and validate them on. One of the weaknesses of AI is it only works really on the patient populations that you build the tool off of. And there are a lot of disparities in the field of prostate cancer and other cancer fields as well, where there are underrepresented patient populations in various studies. For example, there's been a lot of genomic tools created for prostate cancer, but if you look at the number of patients who are African American who've had genomic testing of sorts, it's really, really low relative to the percentage of these patients in the typical US population. And that is a major disparity, and that's one we should address.

And so one thing we did was we used samples where 20% of the patients give or take were African American. We used that to develop our AI classifiers. And then we also used similar populations to validate. And when we look at the performance of our AI classifier in, let's say, Caucasian patients versus African American patients, I think the take home is that the classifier performs equally well. And that's important because I think you want a tool that performs well broadly across patient populations, across heterogeneous groups of patients as well. And so the more inclusive your data set is in which you build the classifier, the better it'll perform when you try to validate that classifier and more importantly when you try to apply it to patients as well. And one of the most important stories in the AI field is a story of, I don't want to use the word but a story of failure.

And so there was a very prominent company who I won't name that developed an AI classifier to identify diabetic retinopathy, and they developed it in a US population, and they tried to apply it to clinics in Thailand. And it turns out it didn't work very well. And I think the lesson there is that, well, they probably didn't have enough of the Thai population in their original training set. And an AI only works in the situation which you build the AI tool. And so for us, it comes down to who are people who get prostate cancer and trying to basically make sure that the tool is built to reflect that heterogeneity of patients.

Alicia Morgans: Well, I'm glad that you and the team are doing that and certainly addressing that disparity from the beginnings of the integration of this technology within this disease at least because there are a lot of disparities very, very evident throughout the disease state, but there's also evidence that if we can get drug to patients, that we can overcome a lot of these differences that we see.

Felix Feng: Absolutely.

Alicia Morgans: Really to be commended for you and the team. As we wrap up, what would your final message be about AI and prostate cancer? Anything else you'd like to share with the audience?

Felix Feng: I think we're just scratching the surface of how we can personalize therapy. There are very important treatment decisions for which we don't have great tools. For example, in intermediate risk prostate cancer patients who should get hormone therapy and who shouldn't. And Dan Spratt worked with us to create an AI classifier and validate it in a phase three trial that does specifically that. It takes a patient population where we know there's a subset that benefit from hormone therapy and there's a larger subset that don't, and we've been able to identify those subsets. Then, in high risk prostate cancer patients the question is who should get two years of hormone therapy versus who should get a few months? And we're actively working on developing a classifier in that space for that patient population. For patients with even more aggressive disease, the question is who should get abiraterone added onto, let's say, two years of hormone therapy?

And the STAMPEDE trial that's being discussed at the ESMO conference this year is actually the ideal forum to be able to answer that question. And Gert Attard is working with us to develop a classifier of who should get abiraterone and who shouldn't for very high risk prostate cancer. And down the road, you can imagine there's questions that you can answer about other situations in the surgical setting in the context of who should get treatment versus no treatment. And I think that the days of having a biomarker, a genomic biomarker that you try to apply across all these different clinical contexts, those days are eventually going to fade in the sense that each specific clinical context like intermediate risk prostate cancer, or high risk prostate cancer or metastatic disease should probably have different classifiers personalized to those patients. You and I are here in Paris at the ESMO Conference right now, and one of the big studies being presented is the RADICALS Trial.

The RADICALS Trial is basically for patients who've had surgery, who should get, at least the part of the RADICALS Trial that's being presented now is who should get zero months versus six months versus 24 months of hormone therapy? And I think the major conclusion is that the vast majority of patients don't need two years of hormone therapy, but that there is a subset that really does need. And we'll see this in the presentation where you need to treat 17 patients with 24 months of hormone therapy to benefit one in terms of metastasis free survival. That means that 16 out of 17 patients in which we would give two years of hormone therapy to don't need that two years or need something shorter. And when you see a scenario like that, you really recognize that we can actually better personalize therapy.

And I think let's say pathology, AI based approaches is one manner in which we can do this, but ultimately I think it comes down to getting the right treatment to the right patient at the right time. And what that means is we don't want to undertreat patients, but we also don't want to overtreat patients. And I think within all the scenarios in which we treat prostate cancer patients, we fundamentally overtreat some, we undertreat some, and some actually get the right treatment, and hopefully we can expand the percentage of patients that really do get the right treatment. That's a long term goal.

Alicia Morgans: And I think that's fantastic. I think we essentially are partially turning on the lights to see what we need to see in terms of understanding the treatments for patients and these approaches to AI as they continue to work for us and enlighten us, I feel like we'll be able to turn on the bright lights and really understand for you, for you, for you, how do we best take care of you and match that treatment, as you said, with the right treatment to the right patient at the right time. Wonderful, wonderful explanation for all of this, and congratulations to you and your team. You are to be commended for the work that you're doing across the entire spectrum of patients who have prostate cancer, and we look forward to your continued contributions.

Felix Feng: Thank you, Alicia.