Read the Full Video Transcript

Zachary Klaassen: Hi. My name is Zach Klaassen. I'm a urologic oncologist at the Georgia Cancer Center. And I am delighted to be joined on UroToday by Dr. Dan Spratt, who is a radiation oncologist and chair of the department, University Hospitals Seidman Cancer Center, Case Western Reserve University. Dan, thanks so much for joining us on UroToday.

Daniel Spratt: Thank you so much for having me. It's always a pleasure.

Zachary Klaassen: So we're going to be discussing some GU ASCO 2025 data you presented. I really like this study because it's taking two awesome biomarkers that we use every day, looking into the comparison and the complementary aspect of risk stratification using Decipher as well as digital pathology artificial intelligence in nearly 10,000 localized prostate cancer patients. So I'm excited for you to share your slides and go through some of the points from GU ASCO.

Daniel Spratt: All right. Happy to. All right. Thank you so much. So exactly. So we performed a very large study in about 10,000 patients that had the Decipher or the 22-gene genomic classifier with either localized prostate cancer and biopsy or at time of prostatectomy and determined all these patients had digital pathology of the samples that were analyzed. And we used multiple available AI models from the digital path to see what's the synergy or added value of using these together.

So these are my disclosures. And so as you had mentioned, the risk stratification in prostate cancer right now in 2025, if you look at the NCCN guidelines, we have a couple of what's called advanced tools, one of them being this 22-gene genomic classifier—it's a prognostic biomarker—as well as there's a digital pathology-based AI multimodal biomarker. And both of these can be used after biopsy. And the genomic biomarker Decipher can also right now be used from a prostatectomy specimen.

And they've both been shown in many trials post hoc to be prognostic of outcomes like distant metastasis. But as I said, it's really unknown. What do we do in clinic? Is there one—everyone asks, what's better? Which do we use? Do we use both? And so that was the genesis of this study. So as you see here, what we did is these are patients that had commercial testing with the Decipher test. And part of this, we also obtained their traditional NCCN variables, like their Gleason score, T stage, PSA, et cetera.

But also, these slides are also digitized from these patients. And so through a fairly standard, but optimized process to obtain digital pathology images, we then put and trained two models. And these are models that one is called GigaPath. One here is ABMIL. And these models both were tested. These importantly to note are not the models that are part of the company, ArteraAI. That's its own proprietary model.

We don't have what that model is to be able to truly test and integrate. But these are prognostic optimized models that are available. And basically, then these get trained and optimized for time to distant mets. And so this is using—we've successfully linked real world data through claims in many thousands of patients, so we can actually have outcomes data for this real world cohort. And as you look here—because we trained.

So we have a training set for both post-biopsy as well as post-prostatectomy with those samples. And then probably more importantly is the validation results. So if we focus—I mean, it's a large study. But on the validation, what you see is what we expect of who's getting this genomic test. Median age is 67. Quite a bit of intermediate risk prostate cancer, about a third, a low risk prostate cancer, and a little less as you get into the high risk group.

And you can see the Decipher distribution here—about half being low Decipher score. And post-prostatectomy, you can see here, as expected, a little higher risk cohort of who's being tested post-prostatectomy. And you can see here to the right how this is—really, how many events were in this data. So there's not tremendous—outcomes in prostate cancer are good despite the large size.

Given the follow-up here, you're talking about four years of median follow-up.

And so the distant mets rates overall are on the low side, but still enough events to be able to evaluate this. And this is the biopsy cohort. This is the surgical cohort. Again, more events here because, again, it's a higher risk cohort. So to dive into the results, again, the left is the results from the biopsy cohort, the validation set. And on the right is the prostatectomy. So what you're seeing here is a five-year AUC for distant metastasis.

And so when we start at the bottom, our traditional NCCN risk groups, an AUC of 0.5 would be a coin flip here. So it's better than a coin flip, but it's about 0.67. Just the Decipher, the GC test, by itself is superior to NCCN, as been reported previously. In this specific cohort, it actually doesn't appear adding even the NCCN variables to the GC test really helped it at all. And then when you start looking at these two different AI digital path models here.

Again, just the digital path models, especially this other one which we'll mainly focus on, also beat out our NCCN risk group. So when you combine them all together, it had a very good performance, about the same as the GC test. But when I show you the multivariable results, really, it appears that both the digital path and the genomic classifier results were independently adding prognostic value. It's a little more obvious, again, the event rates are higher post-prostatectomy that we have because it's sort of a higher risk cohort.

Similar, you can see here in this case, the best discrimination really came from when you added all of them together, the digital path model with GC with either of these digital path models, this ABMIL or this GigaPath here. So we're talking very good AUCs, 0.84. And so when you look at both the univariable, but really the multivariable models here, when you put in a lot of our traditional variables—so again, this is the biopsy model age, the GC score, the NCCN risk groups, and this digital path using this model here—what you'll see is actually the only significant variables here in a multivariable model are the GC score and this digital path model.

Everything else falls out of the model effectively. You can see here NCCN high or very high versus low almost as significant. But really, it's striking that just two variables that are based—these are just from basically the information from your biopsy—pretty much contain the vast majority of the prognostic information. And it's even more clear here post-prostatectomy that, again, on multivariable analysis, the only variables that are significantly associated with distant metastasis are the digital path as well as the genomic classifier.

And they're both independently associated. So it's not that one contains all the information. The other one no longer carries any value. So I think that the takeaways to me is that the 22-gene genomic biomarker, which has plenty of validation in many studies, is independently prognostic for distant metastasis, which we've known for years. But it's also independent even after you include multiple or different of these digital pathology AI models, again, at least the models we tested in this study. But similarly, digital path AI models, the ones we tested, are also independently prognostic for distant mets after inclusion of the GC score, which is the first time this has been shown.

And really, the combination of both of these tools with the clinical variables resulted in what appears to be the strongest prognostic model collectively, when you look at all of our analysis for distant mets. So these results, I would say, suggest that one biomarker doesn't fully capture all of the information that the other one does. Both seem to yield the best results. And in our study, we were able to train an integrated model of both the digital path and the genomic biomarker. And that may be one of the reasons that yielded such good results. So thank you so much.

Zachary Klaassen: That's a great presentation, Dan. I think walking through that was really helpful. I was selfishly excited to cover this at GU ASCO for UroToday in the written form, but also to talk to you because we use both of these in the clinic. And we've often thought this is two different biomarkers likely capturing other things. And you've shown that clearly, they both probably are. And the sum of the two parts is better than each by themselves. So I guess, my question is, where does this place the clinical parameters that we've used for decades? Are we now moving to another generation where the combination of these two biomarkers are going to be what we should be using?

Daniel Spratt: Yeah. I think that in time, I really do believe the value, at least, obviously, when we're more fortunate to be able to have these tests of things like the Gleason score, are going to eventually disappear. And as you know, there's so much heterogeneity in who reads it. And the objective nature—you send this, whether it's for a genomic test or whether you send it for this digital path, it becomes an objective score. And you get the same result every single time from that same sample. And the prognostic information, as you see it, just striking that even our NCCN risk groups, which use PSA data or T stage, perform worse than just a single one of these tests on your biopsy.

So I do believe it's the future. But we also—one of the struggles, we know these are more prognostic. And even the NCCN guidelines already acknowledges that. The challenges, especially I would almost say for the active surveillance discussion, as well as a radiation oncologist when we add hormones and how long, it's really challenging because our trials enrolled patients based on those historical kind of variables. So new trials, as you well know, are underway. And so the future is definitely exciting.

Zachary Klaassen: No, absolutely. I think we've looked at both Decipher and Artera mostly. Looking at other disease spaces, as you potentially move a combination of these forward with clinical trials, where do you see the biggest benefit? What disease space, in your opinion, is where we really need this combination?

Daniel Spratt: Yeah. When I think of what is the decision that men want personalized the most, I would say—I mean, of course, active surveillance is important. I think, though, that's one area our clinical tools, while not perfect, they're pretty good. And so where maybe they'll help is—as you see, people keep creeping more and more into that intermediate risk space. Maybe these are the tools that are going to help us really have confidence in who can and can't.

I think that men—what do they hate more than anything? It's hormone therapy. And of course, if it can help them live longer, feel better, it's absolutely what we want to recommend for them. But we know that we're overtreating men. We just haven't had the tools to personalize it well enough. And so I think to me, similar to breast cancer with some of the tests they have that personalize use of chemo, I think systemic therapies are often something that—whether it's the potential financial toxicity, or the physical fear, or morbidity—I think that using these to personalize is going to be the biggest bang for our buck.

Zachary Klaassen: Yeah, for sure. I think from a high-risk standpoint, long-term versus maybe shorter-term ADT. We know the cumulative effects of ADT. If we can get somebody a lower dose, lower time frame, and get the same results, that's ultimately what we're looking for, right?

Daniel Spratt: 100%. Absolutely.

Zachary Klaassen: You gave us some great take-home messages on the last slide. Maybe just a thought or two to round out our discussion.

Daniel Spratt: Yeah. I think just a caveat—I always like to be critical of our own work—is just so people know, we did not evaluate the Artera test because that's proprietary. We don't have that model. So that'll be something exciting. I'm sure people—it's just as commercial use that they'll be understanding because that model may be better than just the models we had used.

But I think that this is definitely going to be the future because what's unknown is, can genomics capture everything? I'm not sure. Can digital path capture everything? I'm not sure. And so it's exciting. And we just got to figure out—because we already in clinic struggle how to use one of them—using both. So that's really upon us as both researchers and clinicians to be able to provide those answers to help people take care of their patients.

Zachary Klaassen: Yeah. Great discussion, Dan. Always appreciate your time on UroToday.

Daniel Spratt: All right. Thanks for having me.