Soumyajit Roy: Thank you, Dr. Agarwal, for this opportunity. I appreciate it.
Neeraj Agarwal: So before we come to your ASCO 2026 annual meeting presentation on validation of your PSA NADIR model, I'd like to give our viewers some background here. So we know in metastatic hormone-sensitive prostate cancer, now called as metastatic androgen pathway modulation sensitive prostate cancer, so I'll use the word MAPMS for now, henceforth. We know that PSA nadir of less than 0.2 predict much improved survival compared to those patients who do not achieve a nadir of less than 0.2. And this is important because of the difference in the magnitude of survival these patients have.
Really, trials have shown hazard ratio varying from 0.3 to 0.4, really massive difference in survival. And then brings up the question of, how can we intervene early in these patients, so escalate their treatment? So if they're getting ADT plus ARPI, can we add something further? So a TRIPLE-SWITCH trial is going on. Or PEACE-6 trial is going on, where we can add lutetium to the ADT plus ARPI backbone.
On the other side, if we know who is going to achieve a PSA nadir of less than 0.2 at baseline, it will be great for prognostication and counseling and help with patient's anxiety or improve patient's journey, as we heard. And we can also enrich or select patients for given clinical trials. So there are so many ways we can use this value if we can know who will achieve a PSA nadir of less than 0.2 versus not at baseline. The problem is, there's nothing at the baseline right now which can predict. We use clinical factors, we use our own hunch to predict, but really there's no way. So first of all, you have developed this PSA NADIR model we have used in our clinic, and it is working out very well. So before I get to that, congratulations for developing the model.
Soumyajit Roy: Thank you. Thank you.
Neeraj Agarwal: This is a long introduction, but I really wanted our viewers to know about the background first. So PSA NADIR model, what are the components of this nadir model?
Soumyajit Roy: So very simple components that you can get in the clinic, like their age, their body mass index, their hemoglobin, then their baseline PSA, basically the pre-ARPI PSA, and then the volume of disease by charted definition, the metastatic stage at presentation, and then presence or absence of visceral metastasis. That's it. All the things that you can get through imaging or lab, and that's all.
Neeraj Agarwal: And when I used the PSA NADIR model in my clinic, it took me about five minutes to use this, because all those data points which you just mentioned are already available in my patient's records. Even my nurse, actually, or medical assistant, can take those values and enter in the ...
Soumyajit Roy: App.
Neeraj Agarwal: ... online app, yes. How did you interrogate and validate that model?
Soumyajit Roy: So that's a great question. So I'm thankful to all my mentors, including you, Dr. Saad, Dr. Spratt, that I had access to data sets like ARASENS, TITAN, and LATITUDE. And then Dr. Sweeney, Dr. Davis from ANZUP Group also helped me by getting access to the ENZAMET dataset. And then what I did is I first trained and internally tested that model in ARPI treated patients from LATITUDE, TITAN, and ARASENS. And then I externally validated that model into the enzalutamide treated patients from the ENZAMET cohort. And I found that the model was discriminating the responders versus non-responders pretty well, with an area under ROC curve of 0.82. And the model was well calibrated, which means the observed frequencies of PSA response was aligning well with the predicted frequency of PSA response.
Neeraj Agarwal: Many of us are not a statistician like you are. You are a radiation oncologist and you have masters in statistics and bioinformatics. For those who are not, what is the meaning of area under the curve of 0.82?
Soumyajit Roy: So area under the curve is basically, you are plotting at each threshold, or each threshold of predicted probability of PSA response, you are plotting the true positives rate and the false positive rate. And you get a curve by plotting that, and you are basically measuring the area under that curve. Now, the best model in the world should have an AUC of one. The worst model should not have anything more than 0.5. So between 0.5 to one. If it is a good model, it should have somewhere between 0.75 to one, in that range. Which says that in this case, the model did fairly well, with an AUC of 0.82.
Neeraj Agarwal: Thank you very much. Sure. Let's keep going.
Soumyajit Roy: And so that's what AUC says. And as I mentioned, the calibration was pretty good. So we looked at some of the exploratory endpoints, like conditional survival, depending on the PSA response. And we found that even after adjustment for some of the co-variables, like volume of disease by charted definition, this model could actually predict overall survival. And then it could also tell us whether in ENZAMET trial there was an heterogeneity of treatment effect from docetaxel on overall survival based on the predicted probability from this model on a continuous scale, not dichotomized. Again, these are all exploratory and I would be very cautious in interpreting them, but I just wanted to put up what we found.
Neeraj Agarwal: And how about your ASCO 2026 presentation?
Soumyajit Roy: So when I validated this, at that point in time it came to my mind, maybe we need another round or maybe many more rounds of retrospective validation. Of course, the dream would be to prospectively validate it. So I found that a Japanese cohort did a real world validation, and they found a very good performance of the model. So then thankfully I got access to this ARCHES cohort. Again, another enzalutamide treated cohort, and I validated my model there again with an AUC of 0.80 and well calibration.
Neeraj Agarwal: So ARCHES, for our viewers, is a phase-three trial which led to approval of enzalutamide in patients with metastatic APMS or metastatic hormone sensory prostate cancer. And enzalutamide became one of the standard of care options. And for you, it was very obvious, if you're looking for a external model, to get a database which is very similar to your previous data sets.
Soumyajit Roy: Correct.
Neeraj Agarwal: So ARCHES is, again, 1,000+ patient database where patients got enzalutamide as the experimental therapy lead and which showed improved overall survival. So in the ASCO 2026 meeting, you validated your PSA NADIR model in the ARCHES trial. And remind me, what was the AUC again?
Soumyajit Roy: Yes. So in 558 patients from the enzalutamide treated patients from the ARCHES cohort, the AUC was 0.80, very similar to what we found in ENZAMET. Also, there were some questions that if a patient has got six months of prior ADT, which was allowed in ARCHES, then the patient's baseline PSA is automatically going to drop to 0.2. So what's the value of this model at that point in time? So what we did is we actually did a sensitivity analysis by restricting our model and applying that to a cohort that had less than three months of prior ADT. And we found that even in that cohort, the AUC was 0.79. And again, the model was as well calibrated as it was in the original 558 patient cohort.
Neeraj Agarwal: So that's great news. The model still applies to those patients who we are seeing in the clinic who may have started androgen deprivation therapy through their local urologist or other specialist, and they're coming to see us. Model still is valid for them.
Soumyajit Roy: That is correct.
Neeraj Agarwal: How do I access to this model? I have access to this model, but for all other friends and colleagues who are listening to this podcast, how can you access this model?
Soumyajit Roy: So the original Nature communications paper had the model link, but unfortunately I found out somehow that the link is broken. I'm happy to share that link with this video today again, and that way all our viewers and audience can have access to that model. And if anyone is interested, as I mentioned, the Hiroshima cohort, they validated the model on their own, because the Nature communication paper has all the locked model coefficients. So someone can access the model that way as well.
Neeraj Agarwal: So if somebody needs access to that model, they can definitely look at this link in this video or approach, contact you?
Soumyajit Roy: Correct.
Neeraj Agarwal: Your email?
Soumyajit Roy: So my email is my first name, my last name, number eight, @gmail.com.
Neeraj Agarwal: You have done a fantastic job of creating this PSA NADIR model, which predict the PSA response, or PSA nadir, six months after starting the ARPI treatment for patients with metastatic APMS or hormone-sensitive prostate cancer at baseline. And I'm regularly using this model. It was really helping me with prognostication and treatment selection. And I'm really hoping colleagues worldwide will be able to use this model and make their decision making more personalized early on.
Soumyajit Roy: Thank you. Thank you, Dr. Agarwal. I appreciate all the help from my mentors.
Neeraj Agarwal: Thank you, Soum.
Soumyajit Roy: Thank you.