The poster that really caught my eye here is looking at the work that you all have done and looking at a predictor, a genomic predictor, for patients who would respond to BCG or perhaps not respond to BCG. So look forward to kind of giving us a summary of that, but then I think importantly where we go next. So Jacob, I'll turn it over to you.
Jacob Lang: Our aim was to develop a genomic score to predict early BCG failure in patients with high grade non-muscle invasive disease. We took a look at 64 patients who underwent index or restaging TURBT at MSK. They received five of six doses of induction BCG and had at least one follow-up cystoscopy at our institution.
We then performed MSK impact, a 505 gene panel that's FDA authorized on each patient, and then used a machine learning model. So we used three machine learning models and then selected the best performing one in order to select genes from this panel that were significantly associated with protection or prediction of early BCG failure.
So of the genes that we selected, the most important marker for response to BCG was tumor mutational burden. We found that that had the highest weight and then the only significant predictor of risk was a TP53 mutation. All other variables in our model were protective. What we found was that we then normalized these to an NGS score of zero to 100 and then put them into tertiles with low, intermediate, or high risk NGS score.
So we found that low NGS score was significantly associated with a better recurrence for survival than those with intermediate or high risk disease. We found that the low, intermediate, and high risk NGS scores were clinically balanced across groups, so similar rates of high grade T1, CIS and AUA risk classification. And then we also did a COX regression, which showed that low NGS was the only significant predictor of response to BCG.
Sam Chang: Of success with BCG; is that right?
Jacob Lang: Exactly, of success with BCG. No clinical variables were significantly predictive in that model. We then did a AUC ROC curve and found that the AUA risk classification performed at about 0.54 for the AUC. And when we added our NGS score, it improved discriminatory accuracy by about 22%. So that was 0.66.
This is a preliminary study. We had 64 patients, 24 failures, and 40 long-term responders to BCG. And our next steps, we've expanded our strict BCG failure definition. So we're now including patients with BCG unresponsive disease. We're also including patients who had cystectomy or progressed to muscle invasive disease in order to kind of make our findings more robust.
Sam Chang: Yeah, that makes sense.
Jacob Lang: Yes. And then we've also expanded to about 80 patients in the failure group and 80 in the long-term responder group.
Sam Chang: So Gene, when you look at this cohort of patients, this was after the five out of six that received induction and then an evaluation with cystoscopy. So did all those patients get a TUR or was this a cystocytology to see if they failed quote unquote versus not? Tell me how you evaluated that initial cystoscopy.
Eugene Pietzak: Yeah. So I think this is a clinical cohort of patients essentially, and it's retrospectively designed and developed. So these patients all went under standard clinical assessment, which is basically an office cystoscopy, usually with narrow band imaging, urinary cytology, there wasn't any mandate biopsy, so it was only for cause or suspicion.
Sam Chang: Right. So p53 tumor mutational burden that was high, not very good, or at least those patients were less likely to have a response to BCG. Is that-
Jacob Lang: Tumor mutational burden was actually a predictor of response.
Sam Chang: Of response, so-
Jacob Lang: So that was the strongest predictor of response, which is in line with other studies.
Sam Chang: Yeah, and then what about p53?
Jacob Lang: P53 was a predictor of failure in this group, yeah.
Sam Chang: Of failure. So p53 positive, less likely to respond, which you kind of ... So let's conjecture. Okay. So p53 mutational change, et cetera, would you think then these would be more likely to respond to chemotherapy, Gene? Just throwing out conjecture, because these will be so important as we develop more non BCG or BCG alternatives, or for those patients who don't respond to BCG, it'd be great to know kind of your thought process behind what's next or what should-
Eugene Pietzak: Yeah, I think that's a great question. And obviously, we're all looking for sort of biomarkers to guide more personalized treatment approaches. I think p53 is more likely to be prognostic. You have that, that's just bad biology in general, [inaudible 00:05:52] this case. Tumor mutational burden, I think is very interesting. That's an area that we're really trying to hone in on with BCG response because as you know, I'm sure the audience knows, that's a very robust potential biomarker for response to immune checkpoint blockade. And so we know that BCG is a non-specific immunotherapy and so there's been this prevailing hypothesis for BCG specific T-cell immunity, but there's also this potential emerging story of tumor specific immunity that sort of is provoked by the BCG treatment.
Sam Chang: So as you all have been really some of the leaders in terms of helping trying to determine those patients who ultimately have success with BCG versus those who don't, how would this fit in with some of the work that you all have done, Gene, looking at the tumor microenvironment of those that have responded to BCG versus those, maybe it's not the cancer cells themselves, but the area around ... How are we going to put this all together, either one of you? I know we're real early, obviously, in trying to determine that there's so many different factors at play. Kind of tell me your thought process in terms of where we go next, or what's going to be influential, or all the above.
Eugene Pietzak: I think that's a really great question and that's really where this is all sort of moving towards and it's more of a multimodal type of approach where we can't just look at one thing. And that's why Jacob, who did a phenomenal job, I mean, he was a junior resident at Cornell and just looking for a project and just really ran with this, and now he's moving into his research here, and I'm really excited to see what great work's going to come from the upcoming dataset
Sam Chang: So what next? Yeah.
Jacob Lang: Yeah. So we're working on a transcriptomics project as well in relation to this looking at different failure responses. The idea is to come up with, like you said, kind of a combined signature that we can use at the outset at diagnosis for patients who have high grade disease, so we can hopefully eventually identify those who need escalation of care rather than just BCG.
Sam Chang: When you looked at risk factors, did you look at molecular classification as well if these tend to be more luminal versus basal? I don't know in terms of the different factors you looked at because so many things we're now trying to balance.
Jacob Lang: In this data set we did not, but in another dataset we are looking at that. I will say that some of the mutations in this dataset are in line with the changes for luminal versus basal subtypes, which will be important in our combined data set as well, and we're also going to look at them head to head eventually and compare.
Sam Chang: So as you gather this data, is the goal maybe to have one type of algorithm or are we going to need to look at multiple things and try to balance? Kind of tell me what you envision five, 10 years from now when you take over Dr. Pietzak's job exactly.
Jacob Lang: For ease of use, I would say we're looking to make a combined signature down the line. Yeah, that would be obviously I think most clinically useful and something that would be readily applicable to clinics.
Sam Chang: No, we have so many new treatment alternatives and options, but we're still stuck in high risk, intermediate risk, et cetera. So to really have a better idea for personalized medicine, this buzzword that we're getting better and better at, but if we have a better idea, specifically the true risk and prediction of response to different therapies for each individual patient and/or their tumors, because that's the other thing we haven't even talked about within the bladder, we're going to have tumor heterogeneity within of what's going to respond, what's not. I think the more information we get just from kind of the work that you all have done will be increasingly important as we attempt to determine the best treatments for these patients.
So Jacob and Eugene, it's always great to see Eugene. Very nice to see you, Jacob. Good luck with things as you go through your residency, and we'll look forward to future projects. And as you finish up your research year, look forward to spending some more time with you and telling us what you have found next.
Jacob Lang: Awesome. Thanks for having us.
Eugene Pietzak: Thanks, Sam. Much appreciated.