Yair Lotan: Great. Thank you so much. I'm happy to present this collaborative effort. As you mentioned, this is an effort with Valar Labs and one that has other institutions collaborating as well. I'll start off really talking a little bit about the AI approach. And I think a lot of us think of AI as a black box and sometimes it is in terms of feeding a computer a lot of information and letting it come up with all its own conclusions. But the approach that we're using with histopathology is actually a little different. The first thing we actually do is actually do a lot of education for the computer. We're trying to teach it what are tumor cells, what are immune cells, what are fibroblasts. And we actually have a platform that's trained with more than 500,000 pathologist annotations using a lot of tissue. And the first thing we do is actually try to teach it so that the computer to recognize what is cancer, what's not cancer, and then afterwards try to inform it regarding prognosis. So, this is just a very generic slide, but it shows the way we try to break down a histology slide for the computer. Once we have the platform, which can look at morphologic features accurately, then we can give them information about prognosis, what happened to the patient, and then it can try to learn from the patients who did well and then did poorly. And then you finally have a validation with an independent cohort to try to see whether or not the learning that AI did is actually valid.
In other words, did it learn a pattern and can it predict future outcomes? This is just a summary slide of what was done previously, non-muscle-invasive bladder cancer. Initially, we looked at patients who are BCG-naive and predicted response to BCG. We also looked at differential treatments such as Gem/Doce and BCG, and we're able to show that the algorithm was able to identify patients who might do poorly with BCG and might do better with Gem/Doce. This still has to be validated, but we also looked at likelihood of recurrence and progression in patients who are BCG-exposed and high risk. And this test is now commercially available. What I presented at SUO was initial effort at looking at prediction in a muscle-invasive cohort, and we started with tissue from TURBT specimens and specifically looking at the TCGA. And the TCGA is an important dataset that has been extensively discussed for a lot of discovery in terms of genetics for cancer. But one of the requirements in the initial TCGA was that the patients did not get neoadjuvant chemotherapy. And so when we're able to look at TURBT specimens on patients who didn't get chemotherapy, which gives you a better idea of whether or not you're going to be able to predict a response purely based on response to surgery. And we developed an identifier of patients who did poorly versus did well.
And then we validated on a cohort from UT Southwestern who similarly did not get chemotherapy and subsequently had cystectomy. And what we found is that we're able to identify a cohort using a biomarker that controlled for age and carcinoma in situ. There's only so much information clinically you can get in a patient who had a TURBT with muscle-invasive disease. But we're able to find those with more than a threefold higher risk of recurrence, greater than a fourfold risk of dying of bladder cancer and greater than a threefold risk of dying of any cause. And you can see a large differentiation from patients who had a low risk from a high risk. And these were patients who just had cystectomy upfront, but subsequently had other treatments. And so I think it's quite interesting that we're able to use histopathology from TURBT to predict outcomes after cystectomy. And this is without looking at any genetics and just looking at a histopathology slide.
So, I think the conclusion from this particular study is that we can stratify clinical T2 muscle-invasive bladder cancer patients, and future efforts are going to be to evaluate prediction in patients who also get neoadjuvant chemotherapy. And the thought is that we might be able to identify those patients who might not need care, may not need systemic treatments and might do well with local therapy alone, and also maybe we can identify subsequently patients who might require additional treatment and not local therapy alone. So, just another potential tool that might help us both prognosticate for patients. I think a lot more work we'll have to do to look at prediction of response, but I think this might be an important piece.
Ashish Kamat: Thanks, Yair, so much. Again, you've done a lot of work in this in the whole biomarker sphere. I think AI is the next biomarker. And as you said, it's machine learning. It's not unsupervised AI, where the machine is just figuring out what to do with its training and going back. You and I have obviously talked about this, but just share with us a little bit some of your thoughts on where this would fit in with, like you said, the clinical parameters, ctDNA, with all of those new tools and prognosticators that we have, where do you think this would fit in? Would it be like a initial triage or an initial screening, or do you think this would be more of a composite piece of the entire puzzle?
Yair Lotan: So philosophically, I think there are a couple of different ways to think about bladder cancer. There's been a lot of efforts as of late to get away from cystectomy altogether and try just to do systemic therapy, EV-Pembro, Gem-Cis, Durva, and skip the bladder removal part. And so, a lot of the trial efforts have come around, giving systemic therapy, and hopefully being able to preserve the bladder. But we know that that really is not going to be uniform for all patients. And in fact, we don't even know the natural history in a patient who does have a complete response. How long will that last? If you're 50 or 55, will that last for the next 30, 40 years of your life, or will you have future recurrences and chances of dying? On the other hand, we know that all these systemic therapies have a lot of toxicity. And so if you're a young patient with muscle-invasive disease, maybe just removing your bladder will be the best treatment. But now the question is, should you throw in systemic therapies that might lead to future kidney problems, future neuropathies, maybe other systemic complications from immune therapies?
So, the very different approach might be to say, who might benefit just from local therapy and no systemic therapy? And so I think this might be a tool that might help answer that specific question. And alternatively, yes, it might be something that we would be able to add onto a serum-based marker or a tissue-based molecular marker or genetic test that would just be another piece of information that would come into us to say, "Yes, you're a better actor," or, "No, you might have less aggressive disease." So, I think it's still early, but I'm thinking that it might be an easy tool relative to some of the other more expensive and more tissue-consuming type of approaches.
Ashish Kamat: Yeah, Yair, and again, the whole issue of the clinical complete response and how it correlates with PCR and how it correlates with outcomes, bladder sparing is clearly something that's at the forefront, but that's a slightly different question. Is it or isn't it? Because wouldn't you have to retrain the entire algorithm to now be able to predict for a CCR that is a durable CCR? Or am I misunderstanding this algorithm and you can take this signature and just apply it there?
Yair Lotan: I think fundamentally you can ask, when you have a patient who just gets diagnosed with muscle-invasive bladder cancer and has a TURBT, you as a physician and as a patient, you want to know, is this a more aggressive cancer or less aggressive cancer? Now, you might say, what tools do I have to take a patient with a T2 muscle-invasive bladder cancer? Compare it to another patient. And the tools might be ctDNA, it might be urine test, it might be RNA or DNA from the tumor, and maybe this will be another piece of information. Now, the question is, now that we have categorized this tumor, so this might be a tool to help categorize that tumor, then the question is, how do we best treat this patient? And the question might be, should this patient get systemic therapy and will that be enough?
Or the question could be, should this patient just get local therapy and will that be enough? So, I think we haven't finalized the decision on what the best approach is in general for muscle-invasive bladder cancer, but at some point we probably will be able to give a better idea for which patients might need more than just local therapy because they have higher risk for micrometastatic disease, or maybe they're more likely to have non-organ-confined disease. And maybe we identify patients who are least likely to have non-organ-confined disease, and they might benefit from local therapy alone.
Ashish Kamat: Yeah, no, absolutely. I mean, I couldn't agree with what you just said more, but just drilling deeper into this. So, this is a signature that predicts for outcomes factoring in a radical cystectomy though, right?
Yair Lotan: Correct.
Ashish Kamat: And I guess it was more of a basic question I'm asking, would you or would you not need to retrain it to now identify those patients taking out radical cystectomy? Or could you use the same markers? How does it work on the training side or how has it worked? Has this signature been trained for patients who don't get a cystectomy as well, or do you think that'd be a new training, a new marker?
Yair Lotan: Yeah, I think it would very much be training a new marker because ... And we are looking and going to present at the AUA data on patients who did get neoadjuvant chemotherapy. But again, it's very different to identify somebody who has worse disease that might not do as well with neoadjuvant therapy than to predict whether or not you are going to respond to such a therapy and to which therapy. Now, there's going to be several different potential options, platinum-based, EV-Pembro, cisplatin plus immunotherapy. So, I think that would require a lot more data to train it than just taking out the bladder and saying, "Did that TURBT predict better or worse disease?"
Ashish Kamat: Yeah, no, absolutely. And I think one of the things that you've alluded to a little bit is that if and when this gets validated, and it can be done on digitized slides, I think hopefully it'll be something that'll be readily available, high-throughput, and much less cost to the healthcare system than having to do very expensive assays with blood draws and genomic typing and then things of that nature. So potentially, it could be a driver of reducing cost burden for our patients as well.
Yair Lotan: I completely agree. I mean, we all recognize that asking for a genetic test means the company has to ask for the slides, the slides have to be sent, the tissue then has to be analyzed, and then we have to get it back. So, it's two to three weeks of waiting at the least, and then of course there's the expense of running all these assays. But I'm not sure that I'm ready to say that this is ready to going to replace these tests, but a lot of those genetic tests also have to be validated in terms of what their exact role is. I mean, we don't have any validated biomarker that says we should treat a patient differently. I mean, PD-L1 testing didn't pan out for the checkpoint inhibitors. Maybe FGFR testing is the only one that has been validated in terms of telling you how to treat a patient.
Ashish Kamat: Absolutely. Yair, as always, thank you for taking the time. Enjoy chatting with you. Take care.
Yair Lotan: I appreciate it. Thanks so much.