Neeraj Agarwal: Hi, my name is Dr. Neeraj Agarwal. I'm a Professor of Medicine and Director of Genitourinary program at the Huntsman Cancer Institute, University of Utah. It's such a huge pleasure and honor to have Dr. Jason Efstathiou, who is a Professor of Radiation Oncology in the Harvard Medical School, and a faculty in radiation oncology at Massachusetts General Brigham Cancer Program. Is that correct, Jason, with the new affiliation?

Jason Efstathiou: Neeraj, so great to be with you. Always a pleasure to be with UroToday. Yes, no, it's Mass General Brigham Cancer Institute.

Neeraj Agarwal: Fantastic. I think we are getting used to the new name.

So tell me, Jason, you have done so much in the field, but we'll focus on your presentation at the APCCC meeting. And especially we are interested in not only we, I think the whole field is interested in how you are using the genomics and the artificial intelligence in personalizing medicine in patients who are undergoing radiation therapy for prostate cancer.

Jason Efstathiou: Yeah, absolutely, Neeraj, happy to discuss that. I mean, I think one of the biggest challenges in prostate cancer is not that we lack effective treatments, is that we struggle to match the right treatment to the right patient. We have traditional clinical risk groups like NCCN that are helpful, but they're ultimately blunt tools and they're heterogeneous, and they may be getting things wrong even up to 25 to 40% of the time, and they group together patients with very different tumor biology. And so as a result, we may end up both overtreating some patients, exposing them to unnecessary toxicity, or undertreating others, and/or undertreating others where we miss potentially opportunities for cure.

So, we don't really have a treatment problem, but we have a selection problem. And that's really where genomics and AI are starting to make a meaningful difference as they refine risk stratification. Very importantly, these tools don't always agree, and that discordance may actually be where some clinical value lies. And the key question isn't whether genomics or AI is necessarily better, it's how do we integrate them to make smarter, more personalized treatment decisions for patients?

Neeraj Agarwal: So, let's start with the genomics first. The tools you have right now, what you're using them, how you are using them, and then we can talk about the AI. It is amazing to see the relative lack of awareness and knowledge among so many colleagues out there, including myself. So, we'll really like to glean from you about the specifics of genomics and the AI, the tools you are using to avoid, as you said, overtreatment and undertreatment.

Jason Efstathiou: Sure, yeah. I mean, what's fascinating about genomics and AI, they're actually approaching this problem from different angles. Right? So genomic classifiers like Decipher, are really measuring tumor biology, what the cancer is at a molecular level. And Decipher specifically is a 22 gene expression classifier. The AI tools like Artera, on the other hand, are leveraging digital pathology and clinical data to recognize patterns. Patterns sometimes invisible to humans, how the cancer behaves and how it responds to treatment. So, Decipher is really biology, Artera is phenotype and it's context specific and it's pattern recognition. So in a way, genomics tells us what the tumor is, and AI helps us understand what it's likely to do. And I think it's important to emphasize that they are not competing technologies, they're actually very complementary.

I'm happy, if you'd like, to talk about examples where these tools may look at avoiding overtreatment or avoiding undertreatment, if you'd like.

Neeraj Agarwal: Please. Absolutely, yeah.

Jason Efstathiou: Okay, yeah. So, I think one of the clearest areas is where these tools may help avoid overtreatment. So for example, in intermediate-risk disease, we know that ADT, androgen deprivation therapy can improve outcomes, but not for everybody. And AI-based models like Artera can identify a subset of patients who derive essentially no metastasis or survival benefit from ADT. And that's incredibly important because ADT carries real side effects as we know, as you well know, metabolic, cardiovascular, quality of life, etc.

So, there's a number of studies. For example, in RTOG/NRG 9408, Artera has been used to identify a biomarker negative group, where there is no distant MET or prostate cancer-specific benefit from ADT. An important nuance on that is it's not measuring biochemical failure, which is an independent endpoint, and I know is a complicated endpoint in prostate cancer. It's really important to know that these tools depend on the endpoint that are chosen. Right? And so, these tools like the AI tools are really optimizing for the endpoint that you choose. Similarly, in high-risk disease, there's emerging evidence that patients may not benefit from prolonged durations of ADT, and they could potentially be treated less intensively. So if you look at RTOG 9202 in a biomarker negative group, there may be no benefit from longer-term ADT, and this has major implications, huge quality of life impacts.

There's other tools out there that are genomic-based. For example, looking at radiation dose escalation, tools like PORTOS, which suggests that some patients may not benefit from higher doses or dose escalation of radiation, and therefore could potentially avoid additional toxicity. So, de-escalation is not just theoretical, it's clinically meaningful. And we're starting to identify patients who can safely get less treatment, and I think that's a big win.

Neeraj Agarwal: This is great. So patients can avoid long-term ADT using these tools, can avoid ADT at completely in intermediate-risk setting in some occasions. Is that correct, using these tools?

Jason Efstathiou: Well, I think it's really important here, Neeraj, and maybe we can get to this a little later, but the limitations of these tools. I think we have to use the word, may, quite a bit here, because I think prospective validation and clinical trials is super needed in biomarker work. Right? And we can touch on those limitations, but these models are often trained on specific endpoints, like distant METs or survival, and that influences what they're optimized for. So for example, an AI model might not show benefit from ADT in terms of metastasis, but we know that ADT still impacts biochemical recurrence. You know?

There's also issues with thresholds, right? All these biomarkers, what we're doing is turning continuous biology into binary decisions, and many of these analyses are still retrospective, and some of the cutoffs may have been arbitrary, and they're based on post-hoc analyses. And that's why ongoing trials that are looking at prospectively validating tools like this, like NRG-GU009, or GU010, which have embedded Decipher into escalation, de-escalation decisions and intermediate and high-risk disease, are so important. These ongoing prospective trials are so critical to really define how best to use these tools in practice.

Neeraj Agarwal: So, how do you tackle those situations when the genomic classifiers do not agree with the AI tools?

Jason Efstathiou: Yeah. Well, so you're raising the very important issue of discordance, right? So this is, I think, where it gets especially interesting, and I would argue even most clinically important, is when the tools disagree.

What we're seeing is that genomic classifiers and AI models often have very modest correlation. Okay? Concordance is like 60%, varies widely, so that means up to 40% discordance in studies, small studies that have looked at this. And so in some cases, you may have a patient with a high genomic risk score and a low AI-based risk score, or the reverse, and so really quite weak correlation. And actually in some studies, near zero correlation in metastatic patients. Right? So what is this telling us? This is telling us these aren't redundant tools. They're capturing different biology, different endpoints, different training biases. And ultimately, that's not noise, that's signal, right? It tells us that these tools are capturing different aspects of tumor biology.

And so, we can think of a real-world example, right? A patient with favorable intermediate-risk disease where Decipher suggests a relatively high risk of metastasis, but the AI model suggests a very low risk and no benefit from ADT. Right? These are the patients that keep us up at night, and they're also the patients where these tools have the greatest potential value. And they force us to think more, I would say, deeply about the biology and the goals of treatment. Right? So I think we have to avoid thinking of it just as a problem, but more as an opportunity. Right? So, instead of kind of... I think it means we're talking about benefit, we're talking about harms, we're talking about uncertainty with the patient, and I think that's a very important discussion to be having.

Neeraj Agarwal: So, how do you reconcile in the clinic? This is for our colleagues out there who are getting discordant results from the genomic classifier versus AI. Where do you lean to?

Jason Efstathiou: Yeah, I think that's really important. I mean, the discordance, again, likely reflects different biology. Right? And so, I think that what we have to think about is that in those cases, it's not about choosing one test over another, per se. Right? It's more about integrating the information along with clinical judgment and having that conversation with the patient. So, I don't really average the score, right? It's one more piece of information and it's one more tool, along with the PSA and the MRI and the PET scan and the Gleason grade and the grade group. And we really need to view it that way and embrace that discussion until we have prospective data to guide us.

Neeraj Agarwal: So ultimately, it comes down to the old saying, "We are clinicians and these are tools to help us help our patients," is that correct?

Jason Efstathiou: I think that's absolutely correct. I mean, you start with clinical risk, you choose a biomarker for the decision at hand. Okay? Don't overextend it into other scenarios. You translate it into what the absolute benefit is, and you act only with total confidence in that biomarker where the evidence fits. Right? So I think the future really is not genomics versus AI. The future is integration, combining genomic data, AI-based pathology, advanced imaging like MRI and PSMA PET, and clinical factors into a more complete picture of the patient's disease.

So, what we want to move towards is more of a true multimodal risk stratification, right? Precision oncology in that way. It's not competition between these tools, it's convergence. And early data suggests that combining these approaches may improve prediction in certain settings. And again, prospective trials like NRG-GU009, 010, will be key in validating certain situations, but combined genomic classifying AI sometimes modestly improves prediction in some cohorts, not all cohorts, in some cohorts, but that is the way I think we need to move is more towards convergence. And ultimately, the goal is precision where we tailor treatment not just to what the tumor looks like, but what it actually is and what it's actually doing.

Neeraj Agarwal: Well, that was a wonderful discussion. Thank you, Jason, for sharing your insights into how we use genomic classifiers and artificial intelligence-based tools to help our patients better.

Jason Efstathiou: Thanks so much, Neeraj. Again, I think the key here is prospective validation. We got to do the trials, right? And in understanding how these tools complement each other and how to integrate them thoughtfully into clinical decision-making. Again, we're moving away from treating prostate cancer based on just its initial appearance and clinical factors, towards its biology. All these tools are giving further insight into that, and I think that's a major step forward. It's a pleasure to be with you, as always.

Neeraj Agarwal: Thank you very much.