Read the Full Video Transcript

Ashish Kamat: Hello and welcome to UroToday's Bladder Cancer Center of Excellence. I'm Ashish Kamat Kama and it's a pleasure to welcome, once again, to our forum Professor Josh Meeks, who has been with us a few times before talking about different things in bladder cancer. Josh, it's a real pleasure to welcome you today to talk to us about your recent work in risk prediction and, essentially, the large body of work that you did that was published recently when it comes to identifying certain phenotypes and genotypes that correlate with response to immunotherapy. So, congratulations on the paper, excellent work, and thanks for taking the time to share it with our audience.

Joshua Meeks: Thanks for having us, Ashish. It's great to be here to talk about this. Again, really great to come together and talk about some of the science here. The main question that we were interested in with our project was, right when the immunotherapy wave was coming to hit bladder cancer, it was clear that there is some benefit. And at that point, I have data here from two large randomized trials, it kind of shows some of the power of immunotherapy. That when you compare it head to head with chemotherapy, and again, these are two trials, two different drugs, but when you compare the chemotherapy arms, and again, this is metastatic disease versus immunotherapy, the response rate is about the same, but it's a consistent flat line at about 30%.

This started us asking the question, is there any way that you could potentially identify biomarkers for identifying that 30%? Because, again, we don't really have biomarkers for chemotherapy, but we thought for immunotherapy it was such a powerful tool, is there any way to potentially to look at that? And so, where we really started working was going back to the neoadjuvant paradigm that we are benefited with bladder cancer by having. As you know, we have patients that come to see us with a bladder tumor, they're a muscle invasive tumor. As surgeons, we're able to sample that, then they go on three cycles of neoadjuvant therapy, three to six, really depending on where you are and what they're giving, and then we do a surgery where we're able to look at an assay of how that therapy worked. And so, this paradigm, a lot of fields have tried to set this up to do research, but this is actually what we do as standard of care as urologists in bladder cancer.

It lets us, number one, ask, are there biomarkers present in that pre-treatment sample that help predict response to the therapy? And then you can ask right away, you can get feedback, did it work? And you're either identifying response or not. If it didn't work, you can then go on to say, what are these predictive biomarkers for resistance? So this is our paradigm, and we said, can we use this for immunotherapy?

But the real challenge was that, in the US, there really weren't any such trials going forward at that point with neoadjuvant immunotherapy. And so that's kind of when Andrea Necchi's clinical trial was published. This is PURE-01, this is so well known to our community, published in 2018. Andrea has really been given a ton of credit, and deservedly so for this trial. 143 total patients that got three cycles of pembrolizumab prior to cystectomy. The response rate was around 40%, the rate of down-staging was just over 50%, and the 1-year recurrence free survival was 87%.

We used that as our discovery cohort. We started collaborating with Andrea right around the publication of this paper and said, "Is there anything that you would want to do from a research perspective?" He published some biomarker data on that, but we said, "If you'd be interested in taking a deeper dive and really looking at this in much greater detail, we'd love to do that as a collaboration." And we were so fortunate that Andrea was willing to work with us.

After we did the work on that, what came from it, we said, "Well, this is interesting, but we have to validate it." And the only other cohort that had all been done, this was ABACUS, again, not done in the US. This was a platinum ineligible trial. As you know, it's two cycles of atezolizumab, very similar response, maybe a little less because they got fewer cycles and they're platinum ineligible. Tom Powles was the PI here, and again, was willing to work with us because he saw thought there may be a potential for us to all work together. Again, this is kind of the question we started framing. What's going on with checkpoint immunotherapy and can we identify biomarkers of response?

Now, this is really just an incredibly important slide because it kind of shows all the people across the world who worked on this project. Most of the science was originally initiated by Gordon Robertson. He was our collaborator. Gordon's been given credit for the molecular work of the TCGA analysis. When Gordon and I started working together on this project, it became, how do we pull all this information together to try to figure out who was responding and who wouldn't? Obviously Andrea and Tom are critical. Claurice and Mauro were part of the bioinformatics team, and we had, again, teams across the country that really helped process all these samples.

When we first started looking at it and we compared them, we said, "Is there any signatures that predict response of both PURE-01 and ABACUS?" And we couldn't find anything. There was just no strong signals that, across trials, predicted biomarker response. So then Gordon said, "Well, maybe it's not that simple. Maybe it's just not one thing. Maybe there's a lot of complexity there." And so the way that he approached complexity was to do consensus-based clustering and try to say, "Are there like tumors that go together that could somehow be associated with response?" There's a lot of ways to do this clustering, but one way to do it is to do the clustering based on expression profiling, but then be able to say, "Is there any of these that are associated with clinical outcomes?"

And so when we looked at this, we actually found that there were three subtypes that were associated with less response than the average, and two that were associated with a better response. And so I think that's the key element that says there may be up to three different ways that you can have a poor response and two ways you have a good response. Again, when the patient's in front of them, you don't really know, but this tells you that it's a lot more complex and maybe describing some of the biology better.

Now, we tried to compare this to PD-L1 expression, and it all correlated so that the good responsive subtypes had higher levels of PD-1 and the poor responsive ones had actually lower levels of PD-1. So that was our first internal check that we actually may have something here. The response data correlated with clinical outcomes. This is recurrence-free survival that was actually published after we started doing this work. And sure enough, the poor response of subtypes each had the higher number of clinical recurrences. But I think the part that we're the most excited about is, if you're one of the subtypes that actually had a good response, in that group, only 1 out of 15 or even 1 out of 19 of these patients had a clinical recurrence at 2 years, which far surpasses what we see with the chemotherapy data.

This is where we think, going forward, this may actually have power. It's not a plan to replace chemotherapy, but if you can identify those who have a great response to neoadjuvant immunotherapy, that may be a better treatment for certain people. And that kind of gets us to precision, which is what we've all been interested in doing in bladder cancer.

When we compared the biology of the different subtypes about what makes them up, we started doing that not with individual genes, but with gene pathways. And so we try to highlight some of that data here. I think if you look at that box, that box shows some of the most important pathways that have been described by Mariathasan et al in their atezolizumab paper published several years ago. So CD8, immune signatures, and antigen presentation. You can see that in the poorly responsive S1 subtypes, all of those pathways are repressed, whereas in the best acting subtypes, the S3, those are very high levels. Again, if you just compare good-acting and bad-acting expression, they're almost mirror images of each other. Again, that's internally consistent.

What I think is the most interesting part of the data is when you just look at the poor-acting responders, S1 and S4. These are patients that don't respond to immunotherapy but have very different gene expression profiles. So for example, in the S1s, these are high levels of FGFR3 expression, which is obviously a known target in bladder cancer, followed with low levels of immune activation. On the other side, S2 tumors has high levels of cell cycle proliferation histones, mismatch repair, nucleotide, excision repair. So very, very different gene expression profiles kind of consistent with the fact that you can not respond to immunotherapy for a number of different mechanisms, and identifying which those are going to help us better treat these patients. So you could imagine if you have an erdafitinib or an FGF receptor inhibitor, treating an S1 tumor would make sense, but if it's an S4, there probably isn't going to be much, much response there. And that's where maybe chemotherapy is a better target for them.

We said, "Well, we want to look at these subtypes in other cohorts." To do that we had to build a classifier, and that's where we sort of leaned on Claurice, who's really become just famous and bladder cancer for building these subtype classifiers. She did consensus MIBC, she did our prior T1 classifier. So she built a classifier based on our expression profile from PURE01. We were then able to put the ABACUS data in there after Tom gave us access to the samples. And this is sort of what that looks like. Again, internal consistency when we look at PD-L1, realizing that they're different PD-L1 immunohistochemistries, they're different trials, but we see relative expression to be pretty similar. So for example, S1 at the lowest, S3 at the highest. There's some differences, and that may actually matter.

This is comparing the clinical outcomes of the two trials for the same tumors. Again, there's a lot of markers. We have a huge supplement to the paper where we did a lot of comparison to validate that these really are the same kind of tumors using a bunch of different methods. Again, we use gene expression, we use mutation data. But when we did that, I think we pretty cleanly could show that the S1s in one trial were very similar to the S1s in the other. And in general, the response is somewhat consistent, but there's a couple outliers. So for example, the S2 and S3 tumors you see here, they had very different responses. So these patients did much better with PURE-01 than atezolizumab.

I think this is exciting to us because we think of these drugs as being potentially the same. We don't have data that they're different, but I think the trial data would suggest there may be some difference. And some of that may be trial design, but we will this concept of, again, there may be better drugs targeting either the tumor microenvironment or what two cells are there, there may be better activity for better patients. Again, we'd love to be able to say, "This drug is going to work better for this patient going forward." Again, this is something that we really want to validate in future things.

As a final level of validation, we actually looked at IMvigor010, this was the adjuvant trial. Again, the length of this work took us forever to get out, and during the time this project was published, and so we were granted access to get to look at this. What we found was that in patients that got atezolizumab in the adjuvant setting, they actually had a better response to the more favorable subtypes compared to the relatively poor response to the poor responsive subtypes. That went away in the cohort of patients that were in the control group.

We've looked at this a little bit more, and what is driving this mostly is the ctDNA-positive patients. In those who are ctDNA negative, again, you don't really see differences. So what's driving that distance between the subtypes really comes down to what we see in the ctDNA-positive patients, which kind of makes sense, right? They're the ones who we think at the most benefit.

We were lucky enough to have mutation data, and so we aligned the mutation data with the gene expression profile. It kind of tells us a lot about how these mutations may be acting. Not directly. I think that still needs to be figured out. But for example, we found KRAS, FGF receptor 3, and KMT2C, all those mutations, regardless of subtypes, are associated with immune suppression and a bunch of immune suppressed pathways. Interestingly, those mutations were overrepresented in our unfavorable subtypes. As a comparison, ATR, which is A DNA damage sensor, which is kind of a part of the DNA damage pathway, mutations in that gene actually led to increased pathways of immune recruitment and immune activation. And that was overrepresented in our more favorable S2 subtypes. So that's kind of an interesting comparison.

I think where a lot of physicians would want to say is, "What was the point of this exercise?" I'd say that the goal going forward is to try to utilize this data that we've made available. We've given it to a whole bunch of folks, anyone who really asked for it, we're happy to provide it to them. But the whole point going forward is, can we use this data to pick better therapies for our patients? One thing that we are interested in, is there anything that we could leverage? So we looked at gene expression profiling, specifically regulon maps.

Regulons are big pathways regulated by master gene transcription factors. One that we found that was highly active in our resistance subtype was this one called KDM5B. And when you look at the genes targeted by KDM5B, there's some activated genes that are shown there in red, but really there's mostly repressed genes. And then most of them are involved in immune activation. So the concept here is that, in S1, you have a lot of KDM5B, and most of the genes that it's targeting by repression are immune responsive genes. So, potentially, if you could inhibit them, then you could potentially activate the immune system and wake these tumors up. And, potentially, that's where the pembrolizumab or any kind of checkpoint therapy would work. And the cool thing about this is we don't really know how it interacts with the FGF receptor 3. About a third of these patients have a mutation, but overall we don't really know how KDM5B and FGF receptor 3 kind of come together and we're trying to figure all that out.

When we looked at this, we see inverse activity between KDM5B regulon and the immune infiltration. So the higher the KDM5B regulon activity, the lower the amount of immune activation. And that was found in our data. But we went back and looked at the TCGA, we sat on the same correlation. So across bladder cancer, KDM5B appears to be repressing the immune microenvironment.

Now, the challenge is that it's hard to figure out a drug for this. There is a commercially available drug. There's a clinical trial that was done once in lung cancer but nothing in bladder cancer. This is C70. So probably it's not tolerated well or not been used. But interestingly, in cell lines, when you give it to cells, the inhibitor, you get activation of immune pathways. The interesting thing, again, we said, well, what about Erda? Because erdafitinib is available, it does work in patients, and you actually find that many of these genes overlap with what erdafitinib is doing. So, while C70 is not available, Erda is, and I think a lot of people think about that drug working to repress cellular proliferation, repressing the oncogene. But there's a lot of activity that's non-cell autonomous, that works by immune activation. And we're seeing that here. Much of it overlaps with KDM5B activity.

And so, going forward, I think what we've shown is that there's a bunch of different mechanisms here. We'd like to try to look at this prospectively in clinical trials. We're interested in trying to see if you can prospectively identify people that may respond better to immunotherapy. I think that's really where this needs to go, is that, can you pick patients in the neoadjuvant or adjuvant setting that could have a better response? And alternatively, can you identify better therapies for patients that you can then use to combine with immunotherapy to give them better outcomes?

Ashish Kamat: Great. Thanks so much, Josh. I mean this is a truly Herculean effect, and as you said, it's team science, but you led the team and congratulations on this body of work. A couple of questions if I could ask you. You did the clustering for this based on the response, or was it an unsupervised clustering and then you did further refinement to identify the response markers? I wasn't quite sure looking at the paper how you guys did that part.

Joshua Meeks: There's no perfect way to do the clustering. When you do the clustering, you can look at, is there two subtypes, three subtypes, four subtypes, five subtypes? I think the black box or the art of clustering is how sharply the boxes come together and how the clustering works based on the different statistical algorithms. There's no perfect for that, right? If you talk to a number of our colleagues, Dave McConkey, people who've done this for a long time and generate these clusters, there's no perfect. So you can try and generate silhouette plots and things to say what looks better than another, but there's a lot of different ways to do it and there's no perfect.

So when we did this clustering, one of the ways we looked at it is, how the sharp are the clusters, how well do they come together? That's mostly measured by either how the silhouettes go or how the genes cluster together. But another thing is to say, are there any of these that matter clinically? When you look at the clinical outcomes, are any of them better than another? And so that's sort of how we elected to go forward with this system, is that the expression profiling was all done based on gene expression profile. But you have a lot of variability in how you pick the genes, how many genes in the statistical analysis you use, and then how does it matter clinically. So that putting all that together, that takes a lot of time in the beginning to say, "This is the algorithm we're going to go forward with." And that's maybe why, for example, you don't see the TCGA like that. The TCGA mostly is based on gene expression profile.

In that case, this is based on that together. But again, I think what pulls it out of it and why we do so much validation is that it's got to matter. It's got to matter both clinically, but then it's also got to matter from the biology of it. That's why we did so much validation with PD-L1, for example, to say, "Yeah, I think this is clustering along the lines of what matters from an immunologic perspective." And it kind of slants how you overall look at the data.

Ashish Kamat: Actually, you answered my next question because I was going to ask you if it was done in a way that was similar to TCGA and why you all found differences. And, of course, you answered that already. So, you found two clusters that were sensitive and then three that were not. And then within those three that were not sensitive, was there a predominant FGFR pathway beyond what you showed with the KDM, or was it in that one cluster specific?

Joshua Meeks: Yeah, so, really, that FGF receptor that was up in higher expression was really being driven in one cluster. Again, we, and I've kind of talked about before, is that it's not just the mutation. So the mutation's there, but there's a number of tumors in that group which has very similar activation of FGF receptor 3. So those pathways are up and downstream pathways are up, but they don't have an FGF receptor alteration.

Again, we're pretty excited about that because, obviously as the tumor progresses, you get fewer luminal cancers, you get fewer FGF alterations. In the early stage disease, like here in muscle invasive disease, there's a lot of patients who have luminal cancers that have of regulation of FGF receptor 3. So that's one of the things we're really excited about, is, can you use an RNA-based biomarker to direct therapy? Whereas right now all of our precision for TKIs, for FGF receptor 3 is based on mutation. So it's an exciting area that I think we're trying to figure out. You'd love it if there were more patients where the therapy could be effective based on their expression profile rather than just a single mutation.

Ashish Kamat: Of course. No, of course. Just based on the design of the studies and Andrea and Tom's patient cohorts, you've focused on the neoadjuvant patients, but have you also done any work that you could share with us on whether your paradigm and your clusters would hold true for the metastatic cohort or an even earlier cohort?

Joshua Meeks: Yeah, we just started talking today about the metastatic space. I think that the only challenge I'd say with that, Ashish, as you know, is that the time from the biopsy until the patient gets the drug can be far. And so the real great thing that we're lucky with is this, is that surgeons take these tissues that get preserved right away, the patient gets therapy right away, then you have an outcome. I think one of the concerns I've had about metastasis, like biomarkers, is the long time. In many cases, a patient has a cystectomy and then they have progression to metastasis with no biopsy potentially within years. So we're just starting to look at that now to see if we can use it. There's a couple of cohorts available that we're interested in. We're also really interested, for example, in CheckMate 274, and can we look at that. That biomarker data is on its way. We'd love to look at that, for example. Again, we think that that may be an option to say, number one, do you need drug? And two, will you respond to it?

Ashish Kamat: That brings me to another question mean is your platform or the platform that you guys use such that you could use this one for example, CTCs? And how much tissue do you need?

Joshua Meeks: It's an RNA-Seq. It's all RNA-Seq based data. So if you can do an RNA-Seq on that, I guess that's possible. I think at some point we'd like to start talking to having this done commercially. Because right now it's all research. It should be good in that sense. But we don't do that. We have enough on our plates, but there are people who do a very nice job of making the assay reproducible, and we'd like to do that. It's going to have to happen if there's any kind of potential use outside of our institution. So I think that's another piece of this that really needs to move forward, is, number one, validation, and two, can you do this across platforms? So there's a lot of work that really needs to be done before this could really stand up.

Ashish Kamat: Well, I think if anyone can tackle it, it's you, Josh, so I'm rooting for you.

Joshua Meeks: Oh, thank you.

Ashish Kamat: I thank you again for taking the time, spending it with us, and always a pleasure.

Joshua Meeks: Thank you so much for having me. Great to talk.