Read the Full Video Transcript

Ashish Kamat: Hello everyone, and welcome to UroToday's Bladder Cancer Center of Excellence. I'm Ashish Kamat, and joining me today in the UroToday Studios is Yair Lotan, who really needs no introduction. Yair, you've done so much work in bladder cancer over the years, especially in the world of markers when it comes to bladder cancer. And here at the AUA in San Antonio 2024, you're presenting data on the Valar Lab's AI-based assessment tool, for us to understand BCG unresponsive disease, who's going to respond, who's not? Share with us the highlights of that presentation.

Yair Lotan: I think you are very well aware that when you see a patient who has high-risk bladder cancer, you look at the risk factors, their stage, their grade, maybe the size of the tumor, multifocality if you have variant histology. And then you give them an average risk of response to BCG. And right now, that's our gold standard. We don't have good alternatives, and unfortunately, we have a shortage. And so, we're all trying to weigh the likely benefits for the patient using these clinical factors. And people have tried to use these clinical factors to try to create risk scores so that we can identify who's at higher risk, who's not. The EORTC has a risk score. There are a bunch of organizations that put together risk scores. And they have a moderate level of predictability, but we're not using all the data that is currently available to us.

The pathologist does give us a grade and we do have a stage, but there's a lot of information that we just don't know that we're missing. And one of the opportunities now with artificial intelligence is to try to help identify additional independent risk factors that will help us guide treatment, guide prognosis, and also give additional information about how likely the patient is to recur, and potentially not respond to therapy. At this time, there are a lot of clinical trials looking at combination therapies, looking at novel therapies. And so, if you could identify those patients at higher risk, for example, you might be able to offer them maybe a little bit more intensive therapy. Or if possibly they're a good patient group, maybe you can de-intensify. So we give maintenance therapy, but if we don't have a lot of BCG, maybe we give the maintenance, especially the long-term maintenance, to those patients at higher risk.

What we do, and this was a multi-center collaboration, which is really the only way to do good science nowadays, we took almost a thousand patients who were BCG-naive. And what we did was we took the pathologic slides, we scanned them, and then we taught an AI algorithm to look at the pathology, to look at certain features, vascularity, size, morphology of cells, and to try to identify those patients at higher risk for failure of treatment, for recurrence, for BCG unresponsiveness. And the most important thing is that we then took two-thirds of the patients and we validated this, because really you don't want to have overfitting with an algorithm. And so what we found was, we could identify a group that was more than twice as likely to recur, more than three times as likely to progress. And this was independent of other risk factors such as stage and grade and multifocality.

And the algorithm is built not only on the AI algorithm but also for recurrence on multifocality, and for progression on stage T-one patients, for example, do worse than CIS and TA patients. But the interesting thing is that this algorithm was independent of all other clinical information, which really demonstrates its power. And I would have to say that one of the nice things about artificial intelligence is that the more you give it information, the more it learns. So this is the first iteration, it's in a large cohort, 900 patients, 300 training, 600 validation. But I can only imagine what the power will be if you had 5,000, 10,000 patients. But even right now, it definitely is showing some potential for clinical utility.

Ashish Kamat: No, that's a great summary. Thank you for that. So if you actually drill down and you look at the marker and you look at this AI-based marker, it helps you identify which patients are not going to have a good response to BCG by a factor of two to three-fold. How would you use this today in your practice? And again, I know you've done a lot of work with markers across the board, we've been on multiple projects together, ignore all of those for the moment. And you have H&E slides, you have all the clinical parameters in your mind, how do you use this or how would you recommend using this test?

Yair Lotan: Yeah, so I think right now, quite honestly, if you sat in front of a patient, I think patients want to get some information about how likely they are to recur. We're not at a point where we can say, "We're not going to treat you with BCG," because we don't have an alternative. But there's a bridge trial looking at GEMDOCE versus BCG. There are at least three trials of combination therapy of BCG plus a checkpoint. Now all these therapies, for example, checkpoint inhibitors have toxicity. If they do show a benefit, you might want to say, "Well, for which patient should I have combination therapy compared to normal therapy?" And I think that will help eventually drive decision-making in terms of therapeutics. Right now, it's mostly information.

And I think it's good to be able to prepare a patient, even if you think that they should get BCG, about how likely they are to need cystectomy or other treatments. Because as you know, these are difficult discussions for patients whether or not to remove their bladder when they have non-invasive disease. And the earlier you have a good indication of whether that's likely or not, it can help steer the patient to at least start considering other options.

Ashish Kamat: No, I'm glad you brought that up, right, because we want to counsel our patients. We want to give them all the information they can get. And this is another great tool that allows us to do that with the patients. Now, again, I don't pretend to follow AI as you do. I don't pretend to understand it as much as my son, for example, a computer science major. But talk to me a little bit about how these markers were developed, and your sense as to... You mentioned that if you feed it more data, it will get smarter. So will the test evolve, and will the test in a year be different from what you presented today? Or is it just the next iteration of the same test?

Yair Lotan: I think because of the fact that this is using pathologic information, this was developed with pathologists annotating slides and looking at various characteristics to teach the algorithm. So I think a lot of people look at artificial intelligence and they say, "It's a black box." It's not a black box. It's data that you're feeding into a computer, and somebody has to annotate degree of vascularity, morphology of cells, density of cells, architecture, things like that. And these are discrete fields that are entered. There are thousands per slide. And so we take all of these things and we say, "Okay, high grade versus low grade." But there are clearly differences within high grade. There are differences between low grade. So I think what happens is, just like facial recognition, they look at a lot of features. We look, and I don't know, we look at... We don't even know what our brain does to recognize a person from another person, but it's taking a lot of different issues whether or not symmetry and the spacing of the eyes, the color of the eye.

But we recognize people and we just assume that we do, but we don't understand the full process. Here, we're teaching the algorithm based on a lot of different features that our pathologists... They can't sit there and look at a thousand different features and measure every cell. But clearly, the computer can be taught some patterns. Should it get better? Yes. I mean, my understanding of AI is that the more you teach it, the better it will learn, and maybe it'll incorporate new features. I think there's also an opportunity eventually to add other molecular features that we don't get from an H&E slide. You could stain it for different characteristics. You incorporate profiling FGFR mutation, for example, before you give an FGFR drug. Not just to assess, is it positive or negative? Because you can have 5% of cells expressing, and it'd be positive; you can have 95%, and likely there would be differences. But down the road, I think this could be made more sophisticated. But what I'm really impressed with is how good it is without any of those other added features.

Ashish Kamat: And to take your analogy further, here at the AUA, if I don't recognize someone by their facial features, I look at their name tag. So here we can use this, and then, like you said, do molecular markers, and that could be the tag that makes the big difference. The way I look at this, and again, it's incremental value in helping us counsel patients and identify who's going to do better with treatment. Clearly, like you said, it's not like we can tell patients they need a radical cystectomy, right? Because they can still try BCG, but it raises our level of suspicion, our monitoring of these patients. We will be a little bit more careful with a particular patient. Now, this is not the topic of your abstract, obviously, but the technology allows us to potentially do this. I think one of the big roles that something like this has to play, and share your thoughts with me, is maybe in the intermediate risk patients, figuring out who needs a TRBT and who can go on active surveillance? What are your thoughts?

Yair Lotan: Now, you led an effort at the IBCG to create what I think is a very clinically useful algorithm based on risk factors. And so you include risk factors, how many recurrences the patient had, when's the last time the recurrence, have they had prior treatment, size, etc. We're actually exploring the use of an artificial intelligence algorithm in intermediate-risk patients also. Now, I am curious to see whether or not low-grade cells, which tend to maybe have a very similar appearance, whether or not an algorithm can actually identify differences. But we've actually put together a cohort of 150 patients, and other institutions are going to contribute to that. And we actually just started a prospective study. So I think there's definitely room for improvement for all of us.

And as you say, patients can benefit from adjuvant therapies, but they also have the risk of side effects, there's inconvenience, there's cost. And so identifying which patients are most likely to benefit is going to be a useful tool. And I think one of the things that distinguishes us, you and I both work with a lot of different companies that create molecular markers, genetics, and there are quite a few commercially available, is that sometimes there's not enough tissue. And they have to get the tissue and there's a turnaround time and there's quite an expense. And here we're just using scans potentially of an H&E or people sending slides. But even if you have a small amount of tissue, you're not going to have enough DNA or RNA to run an analysis. So in many ways, this is a user-friendly approach, because you can't run out of the source, for example.

Ashish Kamat: Again, a very important point, because with all the molecular markers we're doing, there has to be quality control. We went to frozen tissue, then we went back to paraffin. That's not something that the... I hate to use the word average, but the community urologist, I guess, would really want to invest in. Whereas this is something you could scan, send over, have it analyzed, get the results. So it is definitely a lot more provider-friendly and obviously patient-friendly, because we don't need to insist that the patient has another biopsy. So really, again, thank you so much for taking the time and sharing with us. In closing, just for the folks that are listening, your high-level conclusions from the study that you presented, and where you think it's headed next?

Yair Lotan: So obviously we did a large study with validation. There's actually an early access program. I think that we have to recognize that more data is going to be important. We're doing a prospective study right now that's multicenter. I think if we can validate these findings, I think that there will be a way to incorporate this into practice, at least initially, just to help give you prognostic information. But I think as some of these clinical trials come out, and if they do show positive results, I think we're all going to be keen on trying to select which patients should be included, because there's going to be an extra cost and toxicity. So I see this being a short-term situation, not a long-term situation, where people will say, "I really want to incorporate this." And as you know, there's already an AI algorithm in the guidelines to help select patients for radiation for prostate cancer. So I think the field is moving much more quickly at trying to adopt some of these approaches.

Ashish Kamat: Right. Yair, thank you as always for taking the time, and enjoy the rest of the AUA.

Yair Lotan: I appreciate it. You too.