Read the Full Video Transcript

Ashish Kamat: Hello, everybody, and welcome once again to UroToday's Bladder Cancer Center of Excellence. I'm Ashish Kamat, and it's my distinct pleasure to welcome once again to this forum someone who really needs no introduction, someone who's been with us many times before, Professor David McConkey from Hopkins Greenberg Institute of Bladder Cancer.

David, thank you so much for taking the time today to join us and share with our audience what you discussed, the workshop that you led at the recent Beacon Think Tank on artificial intelligence. Really looking forward to what you have to share with us.

David McConkey: Thank you, Ashish. It's a pleasure to see you and to work with you, and welcome to all of you out there. I hope you enjoy what we have to talk about today. So the session that we led at the Beacon Think Tank involved some description of the state of the art of AI, mostly related to AI pathology. And I was charged with describing what we are doing within mostly the Southwest Oncology Group's GU Committee, where I work with Seth Lerner and Josh Meeks on the bladder cancer angle, but also Amir Goldkorn and Monty Pal with both prostate and renal. There's a lot of activity going on right now involving collaborations between the NCTN cooperative groups and industry, as well as ongoing collaborations with academic centers. So overall, we're thinking about how to do this in a way that's impactful and also collaborative and open.

And Seth and I worked very closely with a lot of you on the Cancer Genome Atlas project, and I think we learned a lot of lessons about how one might be able to do this. So just to summarize some of the lessons that we learned from TCGA. One of the priorities, maybe the top priority, was to create an open access resource data for other people to use. We spent a lot of time worried about the quality of the data that went into the project. There was a lot of quality control, a lot of rigorous evaluation. We wanted to do our best to avoid any batch effects that would injure downstream discovery science. And overall, we wanted our biomarker studies to understand both to improve our understanding of cancer biology, but also to produce immediate clinical impact if possible. Recognizing that actually either one on its own can also be valuable.

So in parallel, the work that we're doing in the National Cooperative Trials Network has also been evolving. So the NCI has asked us to change what we were doing, say, back 15 years ago when we put a lot of discovery work straight into our clinical protocols. And now the emphasis is really on storing tissues for future use until trials are 75% accrued. And then these projects are rigorously evaluated by peer review. And the priority is to validate previous discoveries using very robust and locked-down platforms. Of course, the statistical plans have to be very robust and concrete funding plans should be in place. And ideally, when these studies are performed, we ask the investigators to do their science and submit their calls to our stats center, where the connections between the biomarkers and clinical outcomes are connected by our group statisticians. So we've recognized also very much over the past 10 years that partnerships with industry can really enhance the rigor in our studies.

So industry has locked-down platforms. You can think of sequencing platforms for clinical testing and mutation profiling, et cetera. They're really designed to inform clinical practice, and they adhere to all the right standards for that. Industry partnerships often provide financial support. Within the GU committee, we've benefited a lot from this. We're already working closely with industry partners both within SWOG and NRG and other cooperative groups. And we have plans to expand these relationships. The key to this relationship is that we need to make sure that we're not signing agreements that prevent others from using our data. So there are no exclusive agreements that we're creating between industry and SWOG or industry and the other cooperative groups in general that would prohibit other people from using the data. So we've been very active in working with industry to do discovery work and also validation work with artificial intelligence and digital pathology.

So our banking grant covers high-resolution scanning of H&E sections. And we've been doing this actually for a long time because oftentimes we have central pathology review, and oftentimes our tissues have to be marked so that macro dissections can occur. And so these scanned images are much easier to share than sending glass slides around. And they're also a lot easier to share than tissue sections in terms of the legal agreements that have to be in place. So whereas sending glass slides or other tissue derivatives to another center would require a materials transfer agreement, these scanned images only require data use agreements. And so we've been very active in this area. We've been collaborating both with academic groups and Bishoy Faltas at Cornell has just finished a study where he used the scanned H&Es from our COXEN clinical trial, and we're also acquiring a lot of new industrial partners. Three of them are listed there.

So what kinds of things can you do with AIPath? So I've just summarized a few of my favorites here. So David Berman at Queens University did a nice study using AIPath to do digital grading. Several groups have done predictions of molecular subtypes, and I'm speaking here about basal luminal and not the histologic subtypes that pathologists are familiar with. Josh Levy has spoken to us a couple of times, most recently at the IBCG meeting that Ashish led about... Actually, I think it was at the Beacon Think Tank, wasn't it? Where he talked about the digital cytology. He also presented that at the ACR meeting back in May. Bishoy Faltas and our own Alex Barris have got AIPath predictions for response to chemotherapy.

And so here are just a few examples of these studies. So here's the study by David Berman, where you can find the paper in Laboratory Investigation 2023. This is one from Europe where they used the TCGAs and scanned H&Es to develop an AI classifier for molecular subtypes. You can see these are actually the MD Anderson subtypes listed here. And then Josh Levy's digital cytology platform he shared with us several times recently really looks promising in terms of aiding identification of abnormal cells in urine. These are a couple of examples of patients where longitudinal measurements were made with this platform. And then finally, this is Alex Barris' kind of summary scheme from his Cell Reports Medicine paper where he presented his own AIPath algorithm for predicting response to neoadjuvant cisplatin-based chemotherapy. I'd be anxious to help Alex validate that study within the S1314 COXEN scanned H&Es.

So what is the value in these industry partnerships? Well, there's clear value to patients. If we can help them validate AIPath algorithms to immediately improve clinical outcomes, that's great. There's value to investigators. Those immediately involved in these projects get high-impact papers, they get grant funding, and they may be able to serve as industry consultants. Value to the scientific community is a bit more complicated. So many of these tools are black-box tools that we really can't understand very well and they don't really tell us much about the biology. We can't check the math. But there are tools available where that kind of information can be obtained. And so some examples of where there's clear value would be, for instance, in developing tools to understand how to identify histologic subtypes and also perhaps looking at different cell types for spatial profiling of tissues. That would be a lot simpler and a lot less expensive than doing spatial transcriptomics.

So for us, the best-case scenario for these black-box situations is we at SWOG would like to validate existing biomarkers. So one example of that is ArteraAI's prostate cancer risk stratification tool, which was actually trained under the leadership of Felix Feng and others on NRG Cooperative Group radiation cohorts and is already locked down. We have plans to share scanned H&Es from one of our clinical trials, the S1216 trial, with Artera. That will only require a data use agreement. We'll ask them to perform their risk stratification, and then Cathy Tangen and others in our SWOG stats team will connect these calls with clinical outcomes.

So here's the published classifier. It's in New England Journal of Medicine Evidence 2023. You can see the reference at the bottom. So you can see the overall benefit from adding ADT to radiation therapy in this cohort was fairly modest. But if you see if they applied their predictive model to this cohort and separated them into benefit versus no benefit based on the path classifier, those who were predicted to be positive, you can see comparing this blue curve to the red one, had significant increase in benefit, whereas those where the model predictor would be negative exhibited no benefit at all. So this is the kind of classifier that we'd love to validate using our own scanned H&Es.

We're also interested in doing discovery though. So Seth Lerner's leading a project with Artera, where he's going to look at both our surgical trial S1011 and also the COXEN trial S1314 where they'll develop new algorithms. It's just a little bit more complicated for us to confirm that the classifiers are robust. So helpful to have plans in place for validation studies later. What else can these companies add?

Well, if we didn't have support for slide scanning, they could help us do that and also perhaps store the images. If the platform is biologically interpretable, and I mentioned different cell types or different histologic variants, then that can certainly inform hypothesis generation of future discovery efforts. We've also considered and talked with some of them about providing back some financial support to the NCTN to help us pay for other things. For example, like the stats team, they've been involved in a lot of these projects and could use additional support.

So we think that artificial intelligence tools are very quickly going to complement conventional pathology. And in fact, in some cases, they might even replace them. We heard Hikmat Al-Ahmadie mention perhaps doing that during Ashish's IBCG meeting recently to identify histologic subtypes. Collaborations between the NCTN and industry can certainly produce significant synergy in bringing these tools to practice. The challenge is to develop strategies that help us within the cooperative groups maximize rigor and reproducibility so that we preserve the integrity of our brand. So preferably that would be involving blinded validation, but again, we're open to doing discovery efforts to enhance impact for patients.

And finally, we think that these H&E scanned images and other AI-related data sets that we're generating should be as open access as possible. Recognizing that in some cases there's going to be some identification possible with RNA-seq and DNA-seq where we're going to have to have controlled access. We want to try to model ourselves after the TCGA as much as possible. So thank you for your attention.

Ashish Kamat: Thanks so much, David. As always, I mean you covered everything in such a very succinct manner, very comprehensive manner, that it really is exciting to hear what's going on. One of the things that struck me is that when people talk about AI, machine learning AI, whatever you want to call it, people just kind of jump right in and they don't truly take the time to figure out what you seem to have figured out, which is where are the gaps, where are the potential pitfalls, and where can this actually be, I hate to say harmful, but not beneficial to the scientific or the patient community. I really like the way you outlined those issues.

For folks that are listening, and again, our audience is global, right? It's not just here in the U.S., but it's overseas as well. Could you share some of your messages or caveats to folks that think, okay, let me take what's available, ChatGPT, Gemini, and try to apply it on my dataset? Some words of caution to folks that are trying to do that and come up with homegrown AI platforms?

David McConkey: Well, again, I think that there are a lot of pitfalls we've all heard about with ChatGPT and other natural language processing systems that create a lot of artifacts. And the same is true with scientific data. We do our best to check the math where we can, but oftentimes with these black-box algorithms, that proves to be impossible. So I think my recommendation would be to make sure that you're not doing this without the involvement of really well-trained experts, people who are facile with this. And also to do some peer review of the plan before you launch the plan. I think a lot of investigators don't take advantage of outside colleagues to review the experimental plan in place.

And finally, put in a validation step up front. So we've got a lot of one-off prediction models for neoadjuvant chemotherapy, no offense to anybody who's done this, but taking those and then testing them in a situation where you don't have access to the clinical outcomes, and somebody else connects the calls to the clinical outcomes, we think is pretty important wherever possible. But again, this is an evolving field. Everything is evolving. The scanned images may make a big difference on what platform you use for some algorithms more than others. Each of these details needs to be taken into consideration. And again, reproducibility is the key. If you can reproduce your results in a blinded manner on an independent validation cohort, that to me is the best we can do at this point.

Bishoy's got some interpretability built into his, which helps, but still, even with the interpretability, we don't really know exactly what the machine is seeing.

Ashish Kamat: Right. And again, the things that you mentioned are obviously things that we are aspiring to. I mean, that'd be great if we have predictive and prognostic models and predictive capability to help tailor treatments and drugs, et cetera, to patients. But one thing that really interested me when I heard you talk about this, and I've heard Josh Levy and others, and I think this might be something that could help the global bladder cancer community, is the simplicity of digital cytology, right? Because as you've heard, and as we all hear, cytology is almost a lost art. We have given up on it, and we've given up on it because it is not as reliable. And I know we're running out of time, but in short, could you share with us a little bit about how you might see central digital cytology helping folks say in remote areas in Africa, for example, that don't have a cytopathologist?

David McConkey: I think it's going to be disruptive in a good way. I think that you probably wouldn't even need somebody to be a central reviewer. You would just program it into a computer and the person could do it locally, maybe even on a PC. I think Josh is really hooked into something important there. I think also Hikmat's discussion of the need for a histologic subtype classifier. You've talked to me a lot. You've trained me a lot about how important percentage micropapillary might be. It's hard for us to reach agreement about what that looks like among different pathologists. If a computer could do that for us, just imagine all the impact we could generate based on what we know about the aggressive nature of the subtype. So I think there are some clear examples. I'm also interested in combining platforms. So Josh didn't talk about this very much in his talk, but Bishoy has combined gene expression profiling with AIPath. I think it's just the sky's the limit. You could think of combining it with DNA, RNA, spatial, radiomics, that this can really snowball in a hurry in a good way.

Ashish Kamat: Yeah, and I'm sure you would've covered it if you had more time, but one of the areas that, for example, excites me is using image analysis and AI in real time when we're actually doing, for example, cystoscopy and having a machine look at it and say, that is a small cell carcinoma based on all the data. You don't even need a biopsy maybe, right? Who knows? The future has got so much promise. Fingers crossed.

David McConkey: Another great example, Ashish. Really good one. Yeah.

Ashish Kamat: So like I said, David, I could always chat with you forever, but we don't have that much time. So thank you for taking the time once again, and thank you for today for allowing us to do this.

David McConkey: It's always a pleasure, Ashish. Any time.