Compugen Ltd. (CGEN) Earnings Call Transcript & Summary

Hello, everyone, and welcome to today's webinar. I'm Christina Jewell of Labroots, and I'll be your moderator for today's event. Today's presentation, Exploring the Immune-Tumor Microenvironment Using High-Resolution Single-Cell Spatial Transcriptomics, will be presented by Dr. Roy Granit, Director and Head of Computational Discovery at Compugen. For a complete biography on Dr. Granit, please visit the biography tab at the top of your screen. Today's educational web seminar is presented by Labroots and brought to you by Vizgen. To learn more, please visit our sponsor's website at vizgen.com. Now we encourage you to participate today by submitting any questions you may have during the presentation. [Operator Instructions] Dr. Granit, welcome. You may now begin your presentation.

Hello, everyone. Good morning, good afternoon, wherever you are. First, I want to thank Vizgen and Labroots for the opportunity to tell you a bit about the exciting research we're doing at Compugen. And specifically, today, I'm going to talk about how we explore the tumor microenvironment using high-resolution single-cell spatial transcriptomics. This is my legal statement. And at Compugen, we are a clinical stage cancer immunotherapy biotech company and a pioneering computational drug target discovery. Immunotherapy drugs called checkpoint inhibitors can shrink cancers and, in some cases, even eradicate the tumors altogether. These drugs, which include the well-known Keytruda from Merck, are prescribed to thousands of patients a year for multiple kinds of cancer indications, but they only work for a minority of them. And at Compugen, we're working to help cancer patients who are resistant to existing immunotherapy drugs. Our computational discovery platform is the engine that is fueling our competitive advantage and pipeline, and already has delivered multiple assets, which part of them are fully owned. We have clinical programs, including COM701, a first-in-class anti-PVRIG antibody; COM902, potential best-in-class anti-TIGIT antibody. We also have a validating strategic partnership with Gilead on our 503 asset and with AstraZeneca on the PD-1/TIGIT bispecific, in which the TIGIT component is derived from Compugen's COM902. AstraZeneca believes that this product can reach a peak sale of $5 billion when it reaches the market, hopefully, and it's clinically tested in Phase III trials right now, so there's great potential there. We also have multiple early-stage undisclosed assets, which we expect to feed our future pipeline, and an opportunity to deliver long-term value and additional drugs to the patients. So the way our discovery engine works is a sustainable, flexible-loop platform for the discovery of novel IO targets and target development. At the basis of this platform, we have the Compugen proprietary knowledge base, in which we gather data from different types of omics over the years and have generated a really unique database. From this database, we venture into a discovery project. We start with an hypothesis, then we move on to the tool selection or generation, gather the specific data types that we need to conduct our analysis, and then this feeds back into our platform, and it yields eventually new drug targets to be tested in the lab. When in the lab, we know how to test really rapidly and validate the targets coming from our pipeline and conduct various assays to validate the function. We also have a clinical aspect, where we test these drugs in the clinic, and all of the data coming from the validation and the clinic is driving back into our proprietary knowledge base, where it's enriching our capabilities to conduct even further discovery efforts. One of the pathways that we're very interested in Compugen is the DNAM-1 axis. In this axis, we have PVRIG and TIGIT. Both were discovered by our computational platform. These are inhibitory checkpoints that bind their ligands. PVRIG binds PVRL2, and DNAM binds PVR, while DNAM is an activating receptor, which binds them both, but with lower affinity. We are developing antibodies to inhibit these interactions in the hope that it can release DNAM to conduct its activatory role and to activate the T cells. PD-1 also acts to inhibit DNAM by dephosphorylating it, so we believe that the triple combination, anti-PD-1, anti-PVRIG, and anti-TIGIT might be the optimal way to activate the T cells. And we're currently looking into these combinations in clinical studies, and specifically in indications that currently do not respond to immunotherapies. We have some initial signs of success in the clinic. Here you can see an example of a patient treated with a combination of the triplet in ovarian tumors, and you can see that in some cases the patient responded and their tumors shrunk significantly, so we are very happy with these results, and we continue to develop these projects and to study them further. And focusing on the DNAM-axis, we developed this hypothesis where we have regions called lymph nodes or immune aggregates where T cells meet dendritic cells, and we believe that PVRIG is found on unique types of T cells called T-stem cell memory cells, and in these unique regions, they meet the dendritic cells which are supposed to activate them, but because of the interaction between PVRIG and PVRL2, this does not happen as it can to realize the full potential. We believe that the TSCM cells, and this is based on literature, they have a great potential to proliferate and to give rise to more active T cells that can migrate to the tumor, and also they have a self-renewal potential, so they can maintain many waves of T cells that can act against the tumor, and they are also being important in a response to immunotherapy. Once the T cells are activated, they can then migrate and act against the tumor, but in the tumor, they meet another inhibitory signal. The cancer cells also express PVRL2, and this interaction between PVRIG and PVRL2 hinders the activity of the T cells also in the tumor, so we believe that our treatment can also alleviate this inhibition and allow the T cells also to better act against the tumor. So we developed this hypothesis based on a lot of experimentation and a lot of data that we've collected, and because we wanted to further establish this and to validate our hypothesis in order to guide the next steps in our clinical studies, we needed really to adopt the latest innovations in spatial technologies in order to accomplish this task, which is not very simple. But luckily, the technologies have changed and developed over the years. If in the past we used to look at the tumors using bulk RNA-Seq, where all the cells are mixed together, we could not differentiate between them, in the last years, single cell allowed us to look at individual cells separated and gaze at their expression profiles. And now new technologies can also enable us to look at the spatial location of the cells. But why is that important? Why do we need to know the location of the cells in the tumor? So this is why. So if you look at a tumor and you enumerate the number of CD8 T cells in infiltrated and excluded tumors, sometimes you do not see any difference. The number of CD8 T cells remains the same, but if you look at the staining of the cells, you can see that they are spatially located differently, where in the infiltrated tumors, the cells can enter the tumor and act against it and infiltrate it, while in the excluded, the T cells cannot penetrate for various reasons into the tissue, and then they cannot act against the tumor and actually eradicate it. So this information is highly valuable. We've selected to use the MERFISH technology to study the DNAM-axis in the tumor microenvironment. And just to remind everyone what this technology is. So basically, MERFISH is Multiplexed Error-Robust Fluorescence in situ Hybridization, and in essence, it's a chain of sequential FISH reactions conducted in situ, and each of the probes allow us to encode a binary sequence in situ, which we can then later translate into the mRNA transcript that we set out to explore. And so, for each cell, we have a readout of different transcripts in situ, and also with a spatial localization of the cell. To conduct this study, we designed a gene panel of 350 curated marker genes selected computationally and based on our specific know-how, and we also use a stain for the membrane and markers to separate the cells and have the cell boundaries, and we looked into a fresh-frozen and FFPE colorectal and ovarian tumors. We initially wanted to use the fresh-frozen tumors, but when we started the project, it was COVID time, and we tried to shift the samples to Vizgen. And unfortunately, at the time, there weren't that many flights, and as you can see here on the left, actually, the samples got defrosted along the way, and we couldn't work with them. So we decided to look into the FFPE solution. Initially, we were skeptical, but then Vizgen showed us that actually, as you can see on the right, the quality remains the same, and we actually conducted a small study, and we saw that the number of transcripts that we can obtain using FFPE assay is much larger, so we decided to choose this solution and to proceed with that. Since this was a new technology, we first entered into a technology establishment phase where we wanted to see that we can actually use this technology. And we started with a very bird's eye view of the tissue, selected a few lineage markers, and we can see that it's nicely separated between the epithelial, stromal, and the endothelial in space, and we decided to dive even deeper. We separated the tumor into regions field of views, and we enumerated the number of probes in each of these regions, and then conducted a gene-by-gene spatial correlation analysis. And what we found was that the lineage markers of different cell types really clustered nicely together, as you can see here on the right, that the tumor markers, the myeloid T cells, and the rest of the [ populations ] really clustered together, indicating and supporting that this technology is valid and enabled us to detect transcript at the relevant location in a very accurate way, so we felt comfortable moving into the next step, which is single cell resolution of using MERFISH. So here you can see how it works. We start with a cell boundary stain, which enables us to segregate the cells into individual cells, and then you can see here in the middle that each transcript is colored differently. This is just cytocoloring, but basically it's the binary sequence that enables us to do this. And then for each cell we have basically the content of mRNA for the 360 genes, and then we can cluster this information using standard single cell tools, and separate the different clusters and annotate them, so we know in each cluster which cells we have, and then we can project this information into the tissue where we actually did the assay, and we can say where each of the different cells is located in space, so this is very interesting and useful for us. And going back to the question of infiltration and exclusion, we wanted to see in this tumor what happens in regards to the T cells. So what we saw with looking at the T cells, we saw that the T cells are really nicely infiltrating this specific tissue. We wanted to further validate known biology in order to establish this method. So we conducted the following analysis. We looked at an ovarian tumor, and we asked, let's find out which cells reside next to mature dendritic cells. The mature dendritic cells are known to express chemoattractions to [ lure ] T cells, so we assumed that we would find them in close proximity. And using this analysis, we probed which cells reside in 25-micron radius next to the dendritic cells. And as you can see on the right, in the left bar, you can see that the number of neighboring T cells is much higher than the expected normal distribution of any other cell types. So the T cells actually reside next to the dendritic cells. And as I mentioned before, the dendritic cells express the chemoattractant, and the T cells actually express the receptor for this, CXCR3. So, we can see that this interaction between the T cells and the dendritic cells actually occurs using this cell ligand analysis on the right. So the cells not only meet each other, but they also have the potential to communicate and to conduct the activation of the T cells because the costimulatory molecules are expressed on the dendritic cells, and so this really recapitulates the known biology in this respect. So, feeling comfortable with these tools, we wanted to dive into the DNAM-axis and to study it. We initiated the study by looking at colorectal tumors, and we found these nice samples where we had the immune aggregates regions of lymphocytes that reside in close proximity to the tumors, as you can see here, the tumor is on the right and the immune aggregate is on the left, and we've used MERFISH in order to classify the cells in this specific tissue. What we saw is that in the immune aggregate region, we can see a nice and unique mixture of cells of lymphocytes. We have B cells, we have T cells, we have plasma cells, all in very close proximity next to each other in very unique pattern that is characteristic of these regions. We also have unique endothelial cells, which are important for these structures. Then we dived even deeper into this region, and we can find in the immune aggregate regions dendritic cells that express PVRL2 in very close proximity to CD8 T cells that express PVRIG. So they possibly interact as we initially assumed. Looking into the tumor region, we can see different lineage markers, and we can see different known genes, such as PD-L1. And interestingly, we can see that PD-L1 is rather low in these kinds of tumor, and this is something that's often seen in cold tumors. Looking at PVRL2, the ligand of PVRIG, we can see that it's much more abundant than PD-L1 gene. So this is supporting the DNAM1 hypothesis and encouraging us that it might work in noninflamed tumors, where PD-L1 normally -- PD1 doesn't really work. We then further characterize the tumor, and we separated the tumor into the tumor border and into the interior of the tumor, and we profile the CD8 T cells, the different T cells and different lymphocytes. And what you can see is that in the tumor border, we can see more naive and TSCM-like cells before entering the tumor, and we can also see some endothelial cells that's found in very close proximity to these cells in the perivascular niches. While in the tumor itself, we can see enrichment of exhausted cells that have already seen antigen. So this is really interesting. This is the way that possibly the cells arrive from the immune aggregates via these blood vessels, and then they can penetrate into the tumor and become exhausted. This was indeed in line and was previously described in the literature, but it was really amazing to see it in such high resolution and establish further this hypothesis. Looking at another sample, which contained immune aggregates, we decided to dive deeper into it. And here you can see another example of the unique organization of cells in space. You can see that we have a very organized B cell compartment next to T-regulatory cells, and the TSCM are also in these immune aggregate regions. While the exhausted cells are found predominantly outside of these immune aggregate regions, possibly after they've been activated and have seen some antigen presented by dendritic cells, they went out and are migrating towards the tumor. We then compared again the composition of these regions, and we can see in another way, in a more systematic way, the differential populations between the regions with the more naive and stem-like cells in the immune aggregates, and in the tumor [ more exhausted ] cells, and of course, epithelial cells and macrophages. We quantify this systematically across several immune aggregates and tumor regions, and we can see that, indeed, TSCM cells localized to immune aggregates, while exhausted cells localized to the tumor region. Looking at specific genes in our pathway and other known checkpoints, we can see that PVRIG is found more predominantly in the immune aggregate regions, while other immune checkpoints that are known are found more predominantly in the exhausted and the tumor regions, and this is a unique property of PVRIG. Using another technique, we wanted to look at the different regions using another technique to locate cell-cell interaction, and we decided to use a cell niche approach. What's unique about this approach is that it's unbiased. Generally, you ask the question, how do cells cluster in space, and this combines cells of different types, different lineages, and if there is a unique pattern, it's called a niche, and we can characterize the different niches. So we took this ovarian tumor where we had 3 different regions. We have a really exhausted infiltrated region, as you can see in the blue. We have a noninflamed region with less T cells, and we have some region in red that is inflamed, but not exhausted. So we conducted the niche analysis. We got several niches, and we characterized them. So you can see in the bottom left, we have a region where we have less inflammation. We have SPP macrophage enrichment, which are known to be suppressive macrophages. So this makes sense. In the middle, we can see that we have a niche composed of CXCL9, more inflammatory macrophages, and exhausted and infected T cells. So this is a nice way to detect this combination of cells in a unique surrounding. And in the right, we can see that we can also, again, detect the immune aggregates using this niche approach, and we can see B cells and mature dendritic cells found in close proximity. Going back to our specific questions, we asked where we can find interaction between PVRIG and NECTIN2, which is another name for PVRL2, and we conducted this comparison between TSCM cells and activated or mreg dendritic cells, and what we can find is that this interaction occurs mainly in the immune aggregates region we found using the niche approach. So it supports our initial hypothesis. And when we asked, where can we find the expression in close proximity between T cells that express PVRIG and tumor cells that express PVRL2, we found this predominantly in the effector and more exhausted region, as we described in our initial scheme. So the cells reach here, and then they have this second inhibitory interaction. So to summarize, using this tool, we were able to show that the immune aggregates regions are found in tumors that contain TSCM, early differentiated T cells, and mature DCs in a unique way. We found that the DNAM members are preferentially present in these immune aggregate regions versus the tumor, where we have more exhausted genes, and we found that PVRIG positive T cells interact with PVRL2 dendritic cells in the immune aggregate regions, while PVRL2 positive tumor epithelia is found more in exhausted regions and more in the tumor regions. So actually, it was very exciting to see that our initial hypothesis actually holds out in reality, and using this unique high-resolution method, and it was really encouraging and really amazing to see it. So, to summarize, in general, we can see that spatial transcriptomics offers a really unique opportunity to study the tumor microenvironment at a single cell resolution level. We can actually recapitulate known biology using this method, for example, the association between CXCL10 mature DCs and CXCL3 T cells. We established that TSCM cells localized to immune aggregate regions, while exhausted cells localized to the tumor region. And our preliminary data suggests that PVRIG show unique expression in the immune aggregate region over other checkpoints. Hence, we believe that PVRIG blockade might enhance TSCM activation by the dendritic cells in the immune aggregate regions or otherwise in different regions, resulting in dendritic cells expansion and differentiation. And this mechanism has really potential to lead to really a lot of ways of expansion of T cells that can act against the tumor, also with less inflamed indications, as we saw, where we don't have PD-L1, but we do have PVRL2. Therefore, we believe that PVRIG blockade has potential to address what we started, the resistance to immunotherapy, and to help patients in indications that currently do not respond to immunotherapy. I would like to thank our colleagues at Compugen that conducted this study and assisted with advice and with different aspects. And of course, the Vizgen team for being very collaborative, being very open, and giving us this early access to technologies and for developing this method. So thanks, everyone. And now we'll move on to questions.

Thank you, Dr. Granit, for that excellent presentation. It is now time for our live Q&A portion of our webinar. So to our audience, if you haven't submitted any questions, please do so now. Simply type them into the Ask a Question box, click Submit, and we will answer as many of your questions as we have time for. Okay, Dr. Granit, let's dive in here. Our first question, can you discuss how you have managed cell segmentation?

Okay, so a good question. Indeed, this is one of the challenges that are present in using these technologies, especially if you want to focus on the single cell resolution. And I can say first that this is still something that is unresolved in the entire industry or the entire field. It's something that's very challenging to do because you have different cell types of different shapes, and you have different aspects of the tissue that make this hard. But basically, there's been advantages that have become available over the years. So, we started this project about 3 years ago. And back at the time, it was pretty hard to do this. And in the initial versions of the segmentation we obtained, it really was a challenge. We saw a lot of oddly-shaped cells that don't make sense biologically. And we had to develop our own code to screen these polygons and to make sure these are indeed cells. And it was a true challenge. Nowadays, when we receive the data from Vizgen, from MERFISH, they're using the more latest algorithms like Cellpose 2. So it's improved greatly. So it's less of a challenge. But I can say that we still do some quality control over the segmentation. And when we see that we have issues, or we don't have to use a single cell resolution, we can use things like divide the tissue into polygons, and then you don't have to rely on the segmentation. You can just look at different regions and compare them. So it all depends on the question, but it's something we developed some solutions still working to improve.

And our next question, why did you choose the Vizgen MERSCOPE over other technologies?

Yes. So again, we started this project 3 years ago. Back at the time, all of these spatial technologies were really not mature. There weren't that many technologies. And Vizgen were, I think, the first to have a mature commercial-grade technology that we were able to use. So this was, indeed, a relevant aspect for us. And then another aspect that we had is that we have a really complex question to ask, interaction between cells. And for this, we needed really high-resolution single-cell transcriptomics, and other approaches didn't really allow this. So this was also a consideration. And maybe third option, the reason is that because MERFISH is relying on probes and FISH, we believe this technology allows for detecting really rare transcripts. And we're often asking questions about such rare transcripts. Well, methods that rely on poly(A) capture often are biased towards genes that are more highly expressed. So you lose some of the genes of interest. And maybe one of the more practical reasons also is that Vizgen are really cooperative. They really give us access to the latest technologies. So with the company, it was really great to have a partner to work with, while we couldn't find that in many other vendors.

And Dr. Granit, this question has come in several versions. So I'm going to combine it here. Why have you focused on PRVIG and not TGIT?

Okay, yes. So first, I would like to say that both of these genes came out in our computational platform. We were [ back-to-back ] to discover TIGIT in 2009. And we first discovered PRVIG and took it into the clinic. So basically, both of these are of interest to us. We were the first, actually, to take PRVIG into clinical studies. And this is why we have a unique interest in this gene. But we believe that the entire pathway is relevant, as I've shown in one of the slides. We believe that we need, actually, the triplet to inhibit PD-1, PRVIG, and TIGIT in order to really efficiently stimulate the cells. And we can see this, actually, in the clinic, when we, in some indications, where patients receive treatment, anti-PD-1 or anti-TGIT, and do not respond, while in the triplet, and even some patients that got previous lines with these treatments and did not respond and then came to us and got the treatment of the triplet, they saw some initial signs of response. So we believe this really makes sense and really will allow the T cells to be fully operational. And also, as you can see, some tumors do not express PD-1, but they do express NECTIN2, PVRL2, the ligand of PVRIG. So in these specific tumors, we believe that the effect of adding the inhibition for PVRIG will be significant.

And we have several questions coming in asking about your quality controls. What quality controls do you use for cell segmentation results?

Yes, for cell segmentation. So we've developed some of our own proprietary methods. We examine the raw polygons. So we define some of the logic. Because it's proprietary, I can't really reveal. But basically, it's something relating to how the cells look like and what's the right shape, what's the right order of the polygons. Then, of course, we know that we have different lineages, different genes that should be expressed in certain cells. So, we have a set of lineage markers we can check to make sense to see if the cells actually express the right transcript or not. Sometimes, we currently cannot solve issues of spillover. Spillover is a case where you have a transcript that is found very adjacent between 2 cells, and then it might appear to be in 1 cell, but actually it's derived from the second cell. So, in this case, we can't really solve the issue. There are some algorithms coming out to try and solve it, but in this case, we just discard them. So if you see a mixture of markers or oddly shaped cells, then we just remove them.

Now, Dr. Granit, have you tried alternatives to cell segmentation, like more recent deep learning methods, such as CompSeq-based, or have Cellpose 2 provided the most optimal results? I'm dealing with a similar issue in my current project.

So, yes. It's a really good question. We saw these methods in a recent conference, and we really want to test them. There was a package that can run several, I forgot its name, but we really plan to do this. It's something that we haven't gotten to accomplish it, but it's something that we're eager to try. So right now we [indiscernible] it's good or not, but based on what we saw in the conference, it looks promising.

Very good. And how many gene and transcripts per cell do we detect? Is it enough for confident integration with scRNA-seq data?

Yes. So one of the reasons that we chose MERFISH, and we compared different methods even recently, is the fact that you get this nice depth of detection in individual cells. Of course, it depends also on the size of the panel. Right now we use the panel of 350 genes. So the larger the panel, I suppose you will have more transcript in the cell. And it really depends on the cell types. Some cells that are more secretory might have more transcripts because they need to manufacture some protein, while others will have lower levels. But basically, we can see something. And we also, as part of quality control, we see several dozens of transcripts per cell. And we've done this exercise. We took single cell data, and then we tried to map it with the intersection between the genes. And it came out relatively good. I believe that in the future we will be using a broader panel. And then this task will become easier because we'll have more intersection between the 2 types of cells. So that's something we aim to improve. But it's doable also with the current panel, but not every cell can do this.

Very good. And Dr. Roy, Xenium will offer a panel of 5,000 genes soon. If you would start your project now, would you still go for MERFISH?

So good question. So we're currently evaluating our next steps. And actually, we've looked into Xenium. We're still not final on the decision, but it seems for us that at least for now we aim to try the 1,000 panel from MERFISH because of all what I've mentioned before, the good cooperative work, the fact that we can gain access and really work closely with Vizgen on some custom genes that we're interested in. And really, one of the main advantages also is the ability of planning, designing a whole 1,000 gene panel. In Compugen, we're focused on studying novel genes. So I believe that in Xenium, you start with an established panel, then you can only add a few genes, if I'm not mistaken. So for us, it's good to have the ability to design the panel from scratch. So for now, I think we'll stick with Vizgen. But we're open. We keep looking at the relevant technologies. And definitely, we aim to increase the number of genes we're detecting because it's really important.

Very good. And we have time for one final question. What is next for your research?

Yes. So first of all, I want to say that it's been really exciting to do this current research. It felt like we were like the people that initially developed the microscopes and were able to gaze at things previously humans have not been able to see. So of course, people have seen cells before, but not with such high resolution and with the ability to look at such a variety of genes. So it was really exciting and amazing to do this. And I believe that basically we were the first to do this in the tumor microenvironment in humans. So, first, we're really happy with this opportunity. As for the future plan, so we plan to, of course, increase the size of the panel. We have some additional targets coming up from our early pipeline we want to investigate using this method. And so this is our plan. We can't reveal the specifics, but we plan to do additional studies, broader studies, and try to work on questions of how do we [ integrate ] between samples? How do you ask questions across different samples? Yes, so this is basically our plan.

Excellent. And as we wrap, do you have any final comments for our audience?

So, first, I want to thank the audience for participating. I had a really nice time talking to you guys. I want to thank Vizgen and Labroots for the opportunity. And I hope as a community, we'll be able to take this type of analysis further. I know there's great collaboration on generating tools, and I think as a community, we should work together to solve all of these technical issues and help science actually fulfill its goal. So thanks a lot.

A big thank you again to Dr. Roy Granit for his time today and his important research. We'd also like to thank Labroots and our sponsor, Vizgen, for underwriting today's educational webcast. Now, before we go, I do also want to thank our audience for joining us today and for their interesting questions. Questions we did not have time for today, and additionally, those that are submitted during the on-demand period, they will be addressed by our speakers via the email address you provided at the time of registration. Today's webcast can be viewed on demand, and Labroots will alert you via email when it's available for replay. We encourage you to share that email with your colleagues who may have missed today's live event. Thank you, friends, for joining us. That's all for now. Take care, everyone. Have a great day.