George Netto, MD (Host): Hello, everyone. And welcome back to the podcast. As some of you may know, earlier this year, we launched in partnership with Elsevier, the Journal of Precision Pathology. Naturally, for this podcast, we will be occasionally selecting salient articles from the journal to bring to you.

Today, I chose an excellent article from the first issue of the journal, and I invited Professor Yuri Tolkach, the senior author on the study, to join us. Yuri is a Professor in Pathology in the University of Cologne in Germany, and the study is a multi-institutional study; includes, in addition to his own institution, collaborators from Kameda Medical Center in Japan and Nagasaki University in Japan.

The exciting study is about how to use computational profiling on H&E slides in adenocarcinoma to try to decipher some spatial parameter and single cell parameters. So, thank you, Yuri, for joining us today.

Yuri Tolkach, MD, PhD: Yeah. Thank you, Professor Netto, for inviting me. And I'm happy to tell a little bit more about our study.

Host: Wonderful. So, as we mentioned, you're trying to do away with, or to find a cheap way, right? Because single-cell spatial transcriptomics is very expensive, and we have these H&Es that we've been doing for centuries. And now, you're applying this technology or these models that you developed to see if we can get equally important data. So, why did you do this specific study to start with on lung adenocarcinoma?

Yuri Tolkach, MD, PhD: Yeah, I think, the main motivation behind this study, as you mentioned, we have a lot of information coded in H&E slides, and we truly believe that H&E slides and analysis based on H&E slides will stay mainstream for at least five, 10 years because of costs of other methods and, of course, the availability of this information. It's already there at our hands, and we can analyze it.

Still for pathologists, it's sometimes difficult to analyze complex patterns, especially quantitative parameters, but that's exactly the area where AI and algorithm can perform extremely good. And that's how we employ our single-cell algorithm. So, we try to quantify classical cell populations like lymphocytes, granulocytes, and so on. And we observe that there is a lot of prognostic value in these cells. So, we do several types of analysis in our paper to extract this information.

Host: Well, we'll try to cover as much as possible. It's very complex, a lot of informatics, bioinformatic and statistics, but your background as a mathematician I'm sure helps you, and we'll try to simplify it as much as possible to our audience. So, tell me a little bit about the cohorts. I touched that this is multinational. What cohorts did you start with, and what parameters did you measure? And we'll take it from there.

Yuri Tolkach, MD, PhD: Yeah. So, we have several cohorts. And I think you can enjoy in the field of computational pathology are some open source cohorts like TCGA, The Cancer Genome Atlas. It's also possible to use a PLCO cohort or the cohort of large screening study from United States. They were included in our study, but we also used the cohort of resectable adenocarcinoma cases, lung adenocarcinoma cases from our own department and also from our cooperation partners from Japan, from Nagasaki, and from Kameda Medical Center.

So, we used these cohorts in a specific way. It's always nice to have several cohorts because one of the cohorts you can use for exploration just to finding new parameters, and other cohorts you use to validate them. And I think this validation step is very important and why you need multiple cohorts.

I think the parameters actually are very basic, because we see clearly that from quantification of major cell populations such as lymphocytes, you can already extract a lot of information. Just to mention a starting point for prognostic use case in lung adenocarcinoma. So, classical grading is actually pretty strong, because we analyze different architecture patterns, and they already carry a lot of prognostic information.

However, we have these issues of intraobserver variability. That means when I interpret it as one pattern, another pathologist can see it in another way. And computational algorithms are also there to improve this objectivity of diagnosis. And basically, the cell population we analyze are, again, for example, lymphocytes.

As you know, lymphocytes, the main task of lymphocytes is to kill tumor cells. For example, one of parameters that was found to be especially relevant in our study is the distances between lymphocytes and tumor cells. That means when tumor cells are interesting for lymphocytes, lymphocytes come closer, sometimes go immediately in the neighborhood of the tumor cells, and that's a sign of good prognosis. For example, this parameter worked well in our study.

Host: Excellent. And clearly, you do it in different compartments, stromal, so you measure those cells' densities and percentage and distances among cells, immune cells and tumors. But you do that in the stromal compartment and also the density in the tumor compartment and the epithelial compartment.

So, yeah, you looked at all these. But beside lymphocytes, you know, you also looked at other immune cells, right? Like, neutrophils and the architecture that you mentioned. You used some of the informatic tools you use, you know, because you measured many parameters.

You used something called K-mean clustering. Can you explain to a simple mind what is that tool? Because more and more pathologists will be reading papers that have these meta clusters, cellular neighborhood, K-mean clustering. So, it'll be important for us to get familiar with the language.

Yuri Tolkach, MD, PhD: So, I think the concept of, for example, K-means clustering is pretty simple. Why do we need that? For example, we analyze lymphocytes, we analyze neutrophilic and eosinophilic granulocyte. They carry negative prognosis. That means when we have neutrophilic granulocyte, then it's mostly more aggressive tumors. And then, we analyze connective tissue cells. They are also prognostically relevant. And then, you have, like, multiple parameters that are significant. And the problem is: how do you integrate all these parameters in one single scoring? And I think that K-means clustering allows us to take multiple parameters from single tumors and to cluster these tumors into subgroups.

And I think it's basically just finding the cases with similar patterns in the tumor microenvironment. And K-means, it's like a simple but powerful method to find these different clusters, which also show their prognostic value. But now, we don't work with 10 single parameters. Now, we work with such a higher level clusters or subgroups of patients. And I think it's more interesting for clinical practice because you don't want, for example, as oncologist to work with dozens of parameters. You need one classification system that integrates everything you can find in lower level single parameters.

And you also mentioned one specific aspect of the study called meta clusters, and I think the meta clusters is another perspective on the tumor microenvironment where we take one cell, for example, tumor cell, and we look at the neighbors of tumor cell. And these neighbors, they build such meta clusters, such patterns, that can be also decoded from this information and provide additional value. It's like a method already to have a deep look into the tumor and sometimes very insightful for deciphering of tumor biology, not only for extracting prognostic value.

Host: Wow. Yeah, I had to ask a question about the meta clusters and neighborhoods since we chose that picture and put it on the cover of the first issue. So, it'd be embarrassing if I didn't understand exactly what that is. Wonderful. So, many of these things, it's funny, as a pathologist, as a morphologist, which you and I are too, but the difference is you're a pathologist with AI in his brain, is that these are the things we look at.

You know, we have clusters and neighborhoods that help us a lot of time in diagnosis, in classification. But having the computational power just now make it more objective, and you can cluster how many parameters together you're looking at, and look at it in subgroups of patients and hopefully come up with something.

What do you think, you know, the strongest findings that you found in terms of—cause we know in lung cancer, we do have already a lot of predictive biomarkers, molecular, and in addition to the traditional clinical pathologic. How do you think, you know, some of your findings may add to that? You mentioned lymphocytes as a good prognosticator and neutrophils as negative. Can you expand a little bit about that?

Yuri Tolkach, MD, PhD: Yeah, I think there are two sides about the findings of this study. First of all, I think the findings can be already used as an alternative grading system. Of course, prone validation, I think the validation is the pain point of all AI algorithms. You have to perform not only retrospective, but also prospective validation and multicentric validation. That's where we are heading up right now with our algorithms. But also, you mentioned the molecular status. We know for lung adenocarcinoma, molecular driver status, EGFR, KRAS, STK11, it's a very important high-level denominator. And, of course, we need to look at all these parameters we found in context of molecular alteration.

And here, our main hypothesis is that, for example, we take all patients with EGFR alteration, and our parameters should be able to allow to stratify this EGFR-mutated cancers into different subtypes. And exactly these subtypes can show us, for example, if we can count with response to the targeted therapy against EGFR or not. So, that's our hypothesis and also hope a little bit that we can use these parameters to translate them into predictive tools, not only prognostic tools.

Host: When we were talking about the material and method or what your study design, you mentioned in the paper that you wanted it to move away from a black box kind of foundation model, where you don't know what—so, how is the model you're building and that you utilize in this is more interpretable?

And you touch upon the generalizability with having different cohorts, validating some and make sure that it's still valid in others. But for the interpretability, which is a big deal for us pathologists, is people like you create models, and we have no clue what's going on, data in and output out. I think that wasn't your intention here. You were intentional about making it interpretable. So, can you share an example of how if we use this model, we can figure out, "Oh, that's why it's classifying this tumor a certain way?"

Yuri Tolkach, MD, PhD: Yeah, I think that's a very important aspect. And I think that we always work, develop algorithms with an assumption that if we develop a new prognostic or predictive tool, we need to learn something new about tumor. And that's exactly what interpretable features allow, and that is something that is very difficult with foundational models.

So, these are two large paradigms for training of AI algorithms. With foundational models, it's very quick. You even sometimes don't need a real pathology domain knowledge. But the problem is you don't understand why the model decided if it would be a good prognosis or bad prognosis, and so on. And interpretable features always allow to understand and to get more insights why the decision was made. So, it's a very pathology style of developing the models. But I think that it's just also a right way to do that because we need to understand what the AI is recommending us.

Host: So, you would describe like the neighborhood, the meta neighborhood, the meta clusters, because you're deciding, okay, what is the neighbors of the epithelial cells? Is it a fibroblast? And how distant? Is it a lymphocyte? And how distant? And neutrophil... That's some of the interpretability elements in your model, you think?

Yuri Tolkach, MD, PhD: Yeah, these elements allow full interpretability. Of course, when parameters are going to be a bit more complex, we have also very complex parameters that characterize the organization of cells in the microenvironment. But I think we also look at this through the concept of intra-tumoral heterogeneity because we want to understand this dynamics within the tumor, why there is firstly a less aggressive tumor and why the tumor are going to be more aggressive through acquisition of some molecular genetic alterations. That's something we can see or decipher with our parameters. So, we see there is something happening, and we find what exactly is changing on the level of neighborhoods, and then we can either understand that immediately or we can say, "Okay," for example, we don't understand it at such high level, so we have only lymphocytes or we have only fibroblasts, and then we plan additional studies, for example, multiplexed imaging or spatial transcriptomics to exactly tell, for example, it was this kind of lymphocytes or it was this type of fibroblasts. So, at least it shows us where the relevant parts that can be studied in more details with more advanced methods.

Host: That's, very important. So, it's also positioning you for digging deeper and with the most more expensive technology figure out what's happening there. That's great. Where do you think a lot of these—first, is this going to be an open source model? Like, if you develop these models, other institutions can use?

Yuri Tolkach, MD, PhD: Yeah. So, we have a dedicated manuscript under review for single-cell algorithm. And we plan to release one version of this single-cell algorithm that can identify all the cells, make it open source, and, we hope it will be usable by all academic research groups. And I think it's supposed to be a valuable contribution, because this algorithm seems to be relatively precise for detecting these major cell populations.

Host: I think that's the way to go. And, of course, we all need resources to fund future research. So, somehow we can have it both ways, where the model is open source, open for the public to use it, but there is some way we can license certain models to support our own, you know, group because nothing is free, right you have a lot of engineers and a lot of tissues and scanning.

By the way, if you add all the whole slide images in the five cohorts, there was almost 2800, 2791 slides from almost 1300 patients. So, that's another issue, is it's whole slide images. It's not a region of interest, which usually some of these expensive technologies and transcriptomic force you to use. So, that's another strength of the study.

I failed to mention that the study, clearly, many of the authors from beside your collaborators, I wanted to point out Professor Reinhard Büttner, it's so good to see him now going also AI and molecular working with people like you. But he's done so much for understanding of molecular, understanding of lung cancer. So, he's another senior co-author here. What are the next steps for this? What do you think you're going to do next?

Yuri Tolkach, MD, PhD: So, I think there are two main directions that I already touched a little bit. So from one side, we try to plan prospective studies that also one of the limitations of many development studies for AI tools. So, we plan a real prospective usage of the algorithms, maybe non-interventional, just following some patients, along their disease trajectories with these algorithms.

And from the other perspective, as I mentioned, we want to bring the molecular driver status and tumor morphology, but not just to describe, to really try to understand if we can use morphological phenotype as addition to molecular status to predict response to therapy.

Host: So, these will be the turrets. And hopefully, we'll have you back to discuss the findings of a prospective study like that. Yuri, thank you very much first for submitting the article, to our new journal, and we're really appreciative; and, second, for accepting to be a guest here.

And I look forward to seeing you in person actually next week or 10 days from now. You're going to be lecturing at the course that we're putting together for the IAP on—guess what? On AI and computational. So, I look forward and keep up the great work. You've done tremendous. You and I share the interest in prostate cancer, but I see you're branching out, lung and others. So, wonderful.

Yuri Tolkach, MD, PhD: Yeah, my pleasure. Thank you for inviting me.