Prognostic Power of Multi-Modal Artificial Intelligence Biomarker in CHAARTED Trial Subset – Mark Markowski

Zach Klaassen: Hi. My name is Zach Klaassen. I’m a urologic oncologist at the Georgia Cancer Center in Augusta, Georgia. We are in Chicago at ASCO 2024. I’m delighted to be joined by Dr. Mark Markowski, GU medical oncologist at Johns Hopkins University. Mark, thanks so much for joining us today.
Mark Markowski: Thanks for having me.

Zach Klaassen: We’ve heard a ton about AI over the last two or three years. It’s infiltrating medicine, specifically the ArteraAI Prostate Test. Just as a way of background, we’re going to talk about the data that you presented and charted, but how did it come to be in terms of where we’ve seen this start and where it is now?

Mark Markowski: Yeah, absolutely. Artera has been around for a number of years. And I believe back in 2022, they explored their multimodal artificial intelligence prostate test in patients with localized prostate cancer. They looked at several of the large NRG trials as their training cohort to really educate their model and establish their biomarker, and then they used another NRG trial to validate the cohort. They were able to show, based on their biomarker test, that they can predict those patients who had local disease who needed androgen deprivation therapy.

And so really, it was the first AI test that actually predicted responses to treatment. We used that model here to apply to the CHAARTED study.

Zach Klaassen: Even just a couple of weeks ago at AUA, we’re seeing it now in the post-prostatectomy setting too. So it’s continuing to move all over the disease space and showing some really cool results.

Mark Markowski: Right. And it’s going to continue to evolve. So as the model gets better and more data goes into the model, it should be better able to predict outcomes.

Zach Klaassen: Absolutely. So just walk us through, at a high level, the CHAARTED study. It seems crazy to say that’s an old study now, and it is; it’s in the ancient times of 2015, 2016. But just maybe a high-level study design for our listeners to set the stage.

Mark Markowski: Right. And it even goes back further than that. It’s an interesting history lesson. In 2004, we got docetaxel in the castration-resistant space, the first chemotherapy showing survival. And then in 2005, as part of ECOG, now ECOG-ACRIN, they began designing the CHAARTED study. In 2006, it opened, they enrolled about 800 patients. It was a randomized Phase 3 study in patients with metastatic, essentially hormone-naive prostate cancer. Everybody was getting ADT, and it was plus or minus docetaxel. So half of the patients got docetaxel, 75 milligrams per meter squared, every three weeks for up to six cycles. And the primary endpoint was overall survival.

It took several years to accrue the study. And then finally, about a decade ago, we saw the results, that chemotherapy prolonged survival. And as you started to dig deeper into the data, we found that high-volume patients were really the ones deriving benefit. And that became practice-changing almost immediately when that paper came out.

Zach Klaassen: Yeah, absolutely. In that context, great background, how does the MMAI model incorporate into that? There are a few different high, low volume … Maybe just walk our listeners through how you guys set that up.

Mark Markowski: Yeah. We talked about it before. This model was designed for patients with localized disease and it predicted response to ADT. So they took the same model and then applied it to the CHAARTED data. Essentially what the model does is take digital histopathology slides, run them through their algorithm, and then add clinical data to that. It generates essentially a biomarker score: high, intermediate, and low. We took those scores and applied them to overall survival in the cohort.

I think the top-line findings from our study were that the model was able to prognosticate survival in the CHAARTED patients that we had. We had about 450 patients, so just over half of the patients had imaging available. So it was a pretty sizable cohort. We showed that the model can independently predict overall survival.

To your point, everyone wants to know about high-volume disease. So we did a multivariable analysis that included the AI score, volume status, and stage at diagnosis. If you had local disease and then became metastatic versus de novo metastatic, those were all incorporated. Even with all those variables, the model was still able to independently predict survival, which is great. So the model works; even though it was designed for local disease, just taking the older algorithm and applying it to the metastatic setting, there seemed to be a prognostication there.

Zach Klaassen: The big question is, we have all these options for hormone-sensitive prostate cancer. What’s the next step for this? Do we do it in other trials to see if we should give this patient chemo, should we give him ARPI? What’s the next step for the model in this disease space?

Mark Markowski: I think that’s the million-dollar question: how do you take this technology? For you and me treating patients with metastatic prostate cancer, can we look at histopathology slides from diagnosis or metastatic biopsies, run them through the algorithm, and see if it can tell us which way we should be treating these patients?

I think that’s kind of the next frontier here. So the model is for localized prostate cancer, but as you revise the model and use maybe the CHAARTED study or STAMPEDE even to re-educate the model, can you start to be able to better predict responses to certain therapies? When we did it in our study, again, the model was validated for local disease; it did not predict for response to chemotherapy. And again, that wasn’t how it was designed, but I think with more data, it’s likely that we’re going to get there at some point.

Zach Klaassen: That’s great. Obviously not ready for primetime, but let’s just take metastatic hormone-sensitive prostate cancer. If we had a good biomarker on the horizon, how are we counseling our patients? We have high and low-volume disease. Let’s fast-forward two or three years, how do we counsel our patients with this information?

Mark Markowski: Yeah, I mean, I think that’s a great question. If we can get a clinical-grade test where we can put it into the algorithm, get a risk score, like a decipher score, and stratify these patients as high, intermediate, or low risk, then that’s going to be part of the counseling. If you are high-risk, high-volume, de novo metastatic disease, then you’re going to get a certain kind of treatment.

Moving ahead, we have triplet therapy that’s here for metastatic hormone-sensitive disease. Who needs triplet therapy? Can we de-escalate therapy?

Zach Klaassen: That’s exactly what I was going to say.

Mark Markowski: So if you have high-volume, de novo mets, and right now we’re looking at triplet therapy, perhaps with AI technology, maybe you don’t need that. Maybe you can get away with ADT or Abi for those patients and save chemo for later. So I think it’s not only seeing what they need but also what they don’t.

Zach Klaassen: Correct. No, that’s well said. Because I think as we think more about de-escalation, you could see this playing into it and saying, “If you are biomarker low-volume, maybe we treat you for 12 months and then give you a break.” We don’t know. I mean, this is the beauty of these conversations.

Mark Markowski: Right. But I think we’re going to get that information. There are big data sets out there. We used CHAARTED but also STAMPEDE, and there was data at ESMO last year with STAMPEDE and the Artera test with similar findings. Can we use those big data sets, re-educate the AI algorithm, and apply it to a validation cohort? I think it’s going to be fascinating in the years to come.

Zach Klaassen: For sure. Great conversation. Maybe a couple of take-home messages for our listeners?

Mark Markowski: Well, I think it’s important to know that AI technology is here and it’s being applied to different cancers, not only prostate cancer. I think there was an oral presentation in breast cancer using ArteraAI technology. So I think we’re at the very surface of what AI can do, and the more information we provide, the better outputs we will get.

It may end up being part of our clinical treatment paradigm. You get a score, you get a stage, you get a Gleason score, and we put that all together to figure out what’s best for you. We’ve thought about individualized treatment as maybe genomics and molecular, but I think there’s a role for artificial intelligence here, and we’ll see what happens.

Zach Klaassen: Yeah, it’s great. Mark, thanks so much for your time and expertise.

Mark Markowski: My pleasure.