Welcome, everyone, and thank you for joining us today for this webinar. My name is Cathy Bowes. I'm the Marketing Manager here at Akoya Biosciences, and later I'll be joined by my colleague Anchen Zhang, who will facilitate the Q&A session. Today's webinar is part of our Spatial Insights Driving Drug Discovery series, which explores the impact of spatial biology from early-stage discovery to the clinic. Over the past couple of months, we've welcomed expert speakers who have shared insights on how spatial biology is reshaping our understanding of complex biological systems, optimizing therapeutic developments, and advancing precision medicine from the lab to the clinic. These webinars are available to watch on demand, so stay tuned until the end of this webinar, and I'll share more details on how you can access these and listen again. Today, we welcome Dr. Joe Yeong from the Institute of Molecular and Cell Biology in Singapore. Dr. Yeong will share his latest research, which is AI-powered profiling of trillions of immune cell distribution data points to reveal key predictors of cancer recurrence risk. D r. Yeong, it is a pleasure to have you presenting this webinar for us today. Thank you for joining us. If you're ready, I'd like to invite you to begin your presentation. Hi, good morning, everyone. Or maybe good afternoon, everyone. This is Joe Yeong from Singapore General Hospital. Today, it's my great honor to actually present to you some of our works, and particularly one of the papers that we recently published as a Nature cover story. As you can see, the title is Spatial Awareness: AI-Powered Profiling of Trillions of Immune Cell Distribution Data Points Reveal Key Predictors of Cancer Recurrence Risk. I guess today, as you can see in the title, I really want to focus on how do we, you know, like profile so many data points and make sense out of it, and make biological sense, make immunological sense, and make clinical sense. This is my disclosure. None of these are actually relevant to today's talk. You can see that today's outline, before I actually talk about this paper and also featuring a technology called spatial proteogenomics. I need to start the discussion from the AI component, because that is actually paving the way that you understand the initiative and what we have done in the paper. I will also end the presentation with our future works, which is like we call it 4D live imaging. It's actually a very interesting concept, which I would like to share with you. Yeah, please stay tuned. I often start my presentation to actually show this slide, you know, like H&E. Everybody here in the audience should actually know what is H&E, right? You know, like it is a gold standard for even today's pathological diagnosis, right? A lot of the labs are also using it for staining tissue. The field actually knows the technology already existed for 150 years, almost. If you look at across the entire medical field, there are not many technologies that actually existed or are still using it as a routine practice for more than 100 years. I believe, and a lot of us believe, that H&E will come to, you know, like a new version soon, soon. In the era of AI, you know, big data, cloud-based technology, and all this, we will have a new version we should call H&E 2.0. This is the terminology that we actually talked about a few years back. If you Google it now, it's actually quite a widely used terminology. I mean, I don't think I need to spend more time in the next two slides, particularly showing you what is spatial technology. This is RNA-level transcriptomics. You can do like 18,000 genes at the single-cell level. At the protein level, you can do like maybe 100 of them. This is also not difficult. Our group has been in the field for, you know, like more than 10 years. We published more than 130 papers in the field and optimized thousands of antibodies and translated at least three assays into the market and clinics. Especially, you know, like the key thing that we usually want to highlight is that we're actually doing it as a translational assay for immuno-oncology in a CLIA lab as early as 2019. Our group also, you know, like to facilitate, you know, the field and to, you know, we actually built a few tools. For example, this is a platform that we built. This is important because this is a platform to visualize and manage your spatial data. As you can see there, you know, if your spatial data, if you do not have a platform to visualize them, your image is usually, you know, costs you 200 gigabyte and 300 gigabyte, where you actually have no way to even open it without a special software or special talents, right? This is a very important platform to actually help us to actually at least provide a basic cloud-based technology for people to visualize their data. This is just using raw images. You can just scan the QR code to actually look at these particular images and the rest of the images. They can also provide some sort of automation of quantification. Yeah, I have said a thing about the spatial. I think it's the time to actually show you why, as a group like us, as pathologists, as immunologists, we actually come to the space of AI. A few years back, I would call it a serendipity that we actually found that multiplex IF, or spatial technology, you want to call it, the same slide you actually can scan H&E, the actual same slide, not the serial section, right? This is something that you already know now. When we first started to talk about it in 2017, 2018, there were not many people who really believed it. We believed it, and we actually jumped into AI because of this. I see this as a great future. Why? Because if you can do that, literally every single cell that you can see on the H&E, as a pathologist by training, you can recognize this is a tumor cell, this is immune cell, this is fibroblast, this is a vessel. That can be labeled with 18,000 genes at the RNA level, 100 proteins at the protein level, at a single-cell level, right? You kind of like really can provide, you know, like if you provide AI, an actual ground truth at a single-cell level, providing the morphology, the image of the single cell, and also the protein and RNA expression. You can do very good AI learning. When your input to the AI is single cell, your output of the AI is also single cell. It will be similar to an actual staining like this that you are seeing at the right-hand side. This is actually what we call virtual staining. It's another terminology that came out in the very beginning, one of the pioneers to actually come out of the terminology. The interesting part is that H&E, you can wash away. When you wash away, you can still stain your IHC or your IF to actually, on site, verify or validate your AI output. This is something that is, I would say, unprecedented, you know, like an opportunity to actually do something like this in the AI setting. The left-hand side is the expandability of the AI, which I will explain to you in the next few slides. I'll skip this one. Just that particular slide just to show you how do we do overlay and build the AI, right? The key here is that, sorry. The key here is that I want to show you the expandability of AI in a clinical sense and biological sense, right? When your output is single-cell and you can visualize them, you can actually look at them. When you have this kind of AI manifest to you in a top-down menu like this, you actually can provide your own domain knowledge to actually do a sanity check. For example, in this case, as an immunologist, you will know that PD-L1 and T-cell should be mutually exclusive, right? Your CD3 is actually a CD3 subset. Once your AI manifests the same thing on this particular patient sample, you kind of convince that, okay, at least this AI algorithm works for this particular patient. You can move to the downstream. It could be, you know, like downstream of your analysis or downstream of the diagnosis in the particular setting, right? This is something we feel, you know, like the human is always in the loop of this AI decision-making. This could be a safer choice, and this is actually something that you can actually provide input and say no to the AI if you don't want to use it and go back to the practice, right? This is another example for the pathologist particularly. Once you see this virtual staining on, like, for example, TTF1, this should be an adenocarcinoma for lung, or p40 positive, it could be a squamous cell carcinoma. You can look at the images and see whether the morphology makes any sense of it. You know, like then that is actually very easy to kind of like verify or do a sanity check for this AI. Okay, enough for, you know, talking about like a patient, single patient journey of the AI that we do. Now I use a population setting to actually tell you the impact of the AI. I guess I do not need to convince you that, you know, like in the biopharma setting, and today's webinar is actually on this topic, right? Clinical trial is something that, you know, as a patient, you may have difficulty to actually get into a very right trial for you, right? At the pharma end, they also have difficulty to actually recruit patients. In the middle, clinicians, pathologists, oncologists, we work very hard to actually recruit and complete this clinical trial. Often, they actually fail, unfortunately. Every year, more than 90% of the clinical trials actually fail, and the loss of $9 billion, right? We believe that there might be something we can do in the AI era to improve this. For example, you know, like if you want to recruit a clinical trial, usually you have an inclusion criteria. Nowadays, in the precision oncology era, it's often a test, right? You test for biomarker A. If the biomarker A is positive for this patient, you will recruit this patient. If the biomarker A positivity rate in your population is only 20%, meaning that you need to test 100 patients to actually recruit 20 of them, right? at best, right? This is very, you know, you know, like the resource consumed, right? Especially tie, machine tie, clinician tie, you know, like, and all these things. We believe that this usually leads to the failure. I mean, like if you recruit too slow, right? Every single patient has H&E, right? This is an ego standard of pathological diagnosis, right? If you can use H&E, apply H&E 2.0 to actually predict the biomarker A at first, right? Then you select the best possible positive patient to actually do the test. Very likely, the first 20 tests, you already get 20 positive, and you can, you know, move on, right? This is the general concept. Not only that, it becomes very efficient, and you save the 80 tests that you do not need to do for the resources, tie, and all these things. Also, there's another very important point, which is that if you do H&E 2.0, you do not need to be limited to 100 tests. You only need to do, you can actually go to 1,000 tests, right? 1,000 AI H&E 2.0 costs you very little. Everybody has H&E. Your 20 patients that are recruited from that are actually the best of 1,000 instead of the best of 100. They are likely to be the responder, and the clinical trial is likely to be successful. This is a concept that we actually apply to our own drugs. You know, this is something called EBC-129, and it's actually a Singapore homemade kind of ADC drug. It's going to read out in ASCO soon. I mean, I believe when this webinar or, you know, broadcast could be already read out. Yeah, we actually, as a group like us, we are not only in the discovery team of this particular drug making. We also were the ones that developed the IHC-based companion diagnostic assay to move this entire asset actually into FDA IND filing and subsequently Phase 1 clinical trial. Now it's just concluding Phase 1b clinical trial to moving to Phase 2. Not only the companion diagnostic assay, we also built the H&E 2.0 version of this, as you can see in here, as a virtual staining. You can see that the virtual staining of the serial section looks very similar. You can wash away the H&E, as I told you, right? You use H&E to predict the virtual staining, and you can wash away, you can stain the companion diagnostic assay, and then build them into a virtual staining-like kind of images. You can see that if you're following my cursor, my mouse, these four images look very similar. At the cohort level, the correlation AUC can be more than 90%. We are very confident that this is going to actually move forward. Just to conclude at the AI portion, right, this is actually a single-cell levels of prediction. It's a safer choice because if you reach 90% accuracy, you actually have, in a 10-cell setting, you miss one cell. Not at the case level, but in the usual classical AI pathology where you actually miss at the case level, right? This is visualizable, so it's actually explainable, and actually humans are always in the loop. This is the first part of my talk about AI to actually set a scene and to set a scene of the spatial and to move forward to actually talk about the paper that we particularly want to talk about today. I actually, you know, like with the title of Spatial Medicine, Spatial Awareness, right? As you all know, especially in this audience of Akoya Webinar, you actually should know that the spatial proteomics is actually named as the method of the year last year. One of the particular papers that highlight in Nature is this paper about autoimmune disease. Our group subsequently published a similar concept of spatial medicines for cancer. You can see that this is generally the summary of this is that in the recurrence, we can actually predict the recurrence of liver cancer. The key will be the NK cell, particularly subtype of the NK cell that you can find in the tumor. Some of this actually, if the one that not recurrence, they will see this NK cell going into the tumor. It's making sense, right? They go into the tumor, so these patients should have better prognosis, and then they are not unlikely to recurrence in the next five years. If you don't see these particular NK cells going into the tumor, or not even in the margin, the so-called stroma tumor interface, these patients are very likely to be a bad prognostics and actually likely to recurrence in the next five years. In fact, the liver cancer recurrence rate is as high as 70%. The left-hand side of the patients are actually the majority of them. I talk about the summary in the very beginning because I want to focus today's talk on the strategy and the gist of the paper, but not like really telling you step by step and every single figure that's already in the publication, right? That would be a little bit boring, I feel. Again, you know, like I told you the summary, and here I told you the initiative that we actually have for this particular paper. We started the work in 2021, and then at the time, the high-end spatial technology, I would say that, you know, like the Hyperion, right? Only started to actually come out and become quite widely accessible. We were thinking that we want to make use of all this technology. Then, how do we scope spatial story is something that at that time I had been thinking. When we talk about spatial, at that time, my first feelings is that I want to do something even can explain the spatial at the macroscopical level, not microscopical level necessarily. Meaning that we first do not talk about the spatial relationship under the microscope. Let's discuss some of the spatial relationship even without the microscope, right? That's why we talk about tumor center, we talk about invasive front of the tumor, which is the stroma tumor interface, and also the adjacent stroma. We asked the question, you know, like, can we actually find an immune cell subset that you can really see them kind of like penetrating into the tumor center from the stroma, right? This is number one. Number two is that you find this particular subset, and they correlate to a particular clinical outcome. In liver cancer, as I told you, the question that a lot of people are asking is that, can you find this high-risk patient that they are likely to recur in the next five years, which is 70%, right? How do you find this 70% versus another 30% and have a high confidence about this? You can do something for the 70%, right? You can, you know, treat, you can find an adjuvant treatment for them so they can delay or prevent their relapse. That is the clinical levels of question. We have used other spatial technology to profile in this level. I will tell you in the next few slides how do we profile, and that is actually the key of the study. Then how do we apply the AI strategy of all these trillions of data points, right? Ultimately, we zoom into about five genes or five markers, right? Because if you want to do some sort of clinical translation, you cannot just do too many, right? Five is kind of like a magic number that at the time we build that you can package into a multiplex IHC panel or IF panel to roll it out potentially clinically implementable, right? In this paper, we also talk about proteomics, which is a mass spec technology. I'll talk about it a little bit later on. As a Nature paper, obviously, you cannot stop there. You have to validate this in the in vivo setting, ex vivo setting. And there is generally the entire story of this paper. Kind of long story short, one of the end points of this paper figure is that we use all this technology, find the five genes, and put it into an immune scoring system. This immune scoring system to predict the risk of recurrence for liver cancer, the accuracy is as high as 0.82. That is actually superior to all the prognostic factors that are already known in the market in a clinical setting. Even combined, all of them, it's like they only have 0.72. There is including tumor size, you know, like staging, BCLC, TNM, and the vascular invasion. Obviously, the other two graphs, you can imagine that these are the graphs that reviewers ask for, you know, like, oh, you know, you have something like this. Can you make sure this is not solely attributed to NK cell alone or subtype of NK cell? And then whether you really need multiple regions of the tumor, including tumor center, invasive front, and all these things. Do you really need them? If you do not have all the regions and all the markers, does it really work? We have answered all these things and ensured that you at least need one or two markers, and you at least need one or two regions to actually build a superior model. Here come the key things that I really want to show you, the AI strategy. Nowadays, I will call it an AI era, right? In the AI era, everybody will talk about, you know, like, yeah, you know, how to use applied AI. I think I have heard from our post-grad student talk about, you know, oh, I think I have an experiment, a lot of data put it in, ultimately there's no P-value, doesn't work. Maybe we can pass to our AI team, and then, you know, they can do some sort of ML, DL magic to actually make it work. I think that is kind of like a concept that a lot of biologists thought about AI, right? I think we need to have a strategy, not really just putting all the data points together and hope, hope, hope, and finger crossed for some sort of magical things will happen, right? In this paper, I already told you the general concept of the story. Let me break down into numbers and try to scare you away, okay? As you can see, all these numbers are actually in a very conservative level because we actually have more than 300 patients. Yes, we only have three tissue locations. We have obviously more than 100,000 single cells every single tissue. You have more than 18,000 plex because only RNAs, you already have 18,000, right? You are not alone. You have to talk about protein level. We have more than one technology, right? Your clinical pathological parameters also have more than 100. We just use the baseline of this number, and then we dice everything together. This is the variables that you actually need to deal with, right? I was speaking, you know, like when the calculator showed me these numbers, I actually do not know how to read this. I actually put this number in ChatGPT, and ChatGPT told me this is 162 trillion and 1.62 times 10 to the 14, right? This is a number that, you know, very difficult to imagine. We designed a strategy, and in the very beginning, a lot of these data points and strategies come from the GeoMx data. Just to recap, we actually have three regions, right? You know, the tumor center, invasive front, and stroma, right? Also, every single point has given you 18,000 genes, right? Also have some protein in here, right? We particularly look at the NK cell rich region because we already, in the first figure, we already know that NK cells are important. We zoom into the NK cell to do this. Yeah, when you have three regions, you know, stroma, invasive front, and tumor center, and how do you analyze this data, right? At that time, we feel that a pattern recognition could be a way to go. When you have three, if you're following my mouse, this is the key figure. When you have three locations and there are only eight patterns that you can go, you know, like you can either go flat or you can go up and up, like increasing or decreasing, or you can go down then up. It looks like a Nike shape. Internally, we actually called this particular project Nike Project for a very long time until we submitted the first version to Nature, right? This is one particular, I would say, I'm proud of this invention, the idea that we have like to actually put these three locations into a pattern. Our first question is obviously asking, you know, the patient has high risk versus low risk, do they have different patterns? The different particular different patterns, what are the genes associated to that pattern, right? Are they particularly for one or two types of immune cells, and how do we package these data points and then to feed into the AI setting, right? That is actually a general concept. Once you know that we are comparing pattern to pattern, you can see here, there are only four patterns that actually P-value is significantly different between the high risk versus low risk. They are actually accountable for about 241 genes. We put the 241 genes factoring in the three components and then fit it into a machine learning models and ask the AI to rank them, you know, like, okay, based on accuracy. This is the graph that you can see down there. You can see the AI actually ranked, and then all of them, and then all the five genes that we talk about in this paper are actually in the first six or seven. We excluded some of them because those we cannot find a good antibody. Because, as I told you, we want to package them into a multiplex IF panel to validate this finding. The antibody is important. You can see exactly what we have done for this one. That is how we move on to the next step. This is the first part of the AI that I want to share with you. There is another second part of the AI that I really want to share with you again, right? Once you actually settle, you know, like how you get from trillions of things and then come to actually only five genes, right? Then you package them into a multiplex IF panel, and then obviously you can quantify them, and then you come up with a way to actually quantitate them in the three regions. How do you move on, right? Because obviously people talk about immune scoring, right? Immune scoring needs to have a scoring system, right? Usually it's like from zero to one kind of thing or zero to 100 kind of things, right? As a pathologist, you will know that the scoring system is very difficult to use when it is at a gray area. For example, if the cutoff is 0.5, so 0.49 versus 0.51, which ones are positive, which ones are negative, right? This is another thing that we would like to actually apply AI to actually avoid this or solve this problem or dilemma, right? You can see the diagram down there that we feed the AI in the invasive front and tumor center data and then of these five genes and try and find a best weightage system to actually use our own demographic and population to show that the majority of the patient scoring will be sitting at the extreme side of the scoring system. You know, in this case, it's zero to one, and the majority of them, you can see, are actually between zero to 0.2 and also between 0.8 to one. Not many of them are actually sitting in the middle. Yeah, in the middle is here, right? If you have something like that, when you apply in a real-life setting, it's likely that they are not going to run into dilemma in the frequent levels. There is something that we have the AI to actually decide the weightage of the biomarker to compose a scoring system. The next few slides is really, you know, like the so-called expandable spatial, right? Because yes, now we found biomarker, we found immune cells, and then that correlates to clinical outcome. We even have already a scoring system. Is it by chance or does it really have some immunological sense and mechanism? We actually do ex vivo co-culture things. Basically, it's 3D bioprinting the cells at the single cell level and observing the NK cell to actually penetrate the tumor in the ex vivo setting and the live imaging. We really see that if you actually knock down and knock in SPON2 to actually modulate the speed and also the ability to actually penetrate the tumor and kill the tumor cells. Also, interferon gamma is a very important kind of killing cytokine in this mechanism. We also apply spatial analysis and found that the SPON2-positive NK cell, they are always in the proximity of the interferon gamma-positive CD8 T cell. It looks like this is the combination that they actually do the killing in this particular context. I mean, the next few slides, I will split relatively fast because this is the part that's already published. I wouldn't want to really focus on it. This is actually kindly done by our Chinese colleagues as well. This particular figure is just to show you that we know that NK cells interact with the CD8 T cells in the spatial proximity, and we just want to use a Transwell, you know, like experiment to show that they really need a physical contact rather than just a cytokine-cytokine exchange. Last but not least, we also expand this and demonstrate this in a real-world setting by knocking down the SPON2-SPON2, you know, like, and this is actually particularly dragging the entire revision is that they're asking us to actually knock down the NK cell-specific SPON2 and showing that whether this is really working. Fortunately, we can also demonstrate that. We believe that this is also opening another door for us that potentially this could be a new way of looking at NK cell therapy. As you can see, this entire paper concept is really like the spatial medicine concept similar to what we started talking about in the very beginnings, right? You have a scoring system, and then you can actually stratify the patient into high risk and low risk. As you can see, this immune scoring is based on NK cells and CD8 T cells as a baseline knowledge, right, and a mechanism, right? It is making sense to actually treat them with immunotherapy. We did, and we really found that the patient actually, that they have high TIS, they usually respond to immunotherapy. This becomes like they really converted from a high-risk patient usually having bad prognosis into becoming a good prognosis patient. This is something that we are very excited. Last but not least of this particular paper, I just want to also show you one particular technology that we actually adopted in this paper, but we don't, it's not a big part of the paper, so a lot of people kind of like miss this when they read. If you go to the figure two, then we have this spatial mass spec analysis that some of the people call this kind of technology Deep Visual Proteomics, but I don't know whether this is actually a widely used terminology but similar concept. Basically, just find anywhere that you can actually get the pieces out from the tissue, very small at that time. This is four years ago, it's 200 micrometers, but nowadays, I think we can do like a lot less than that. You take the tissue out piece by piece and then sent to mass spec. In this paper, obviously, we talk about validation of the NK cell, validation of the five genes, but in reality, we actually find even more exciting stuff, right? Two years ago, we published this paper in Proteomics, it's a mass spec signature journal. Because when we wanted to do that, the initiative is to actually validate at the protein levels our GeoMx data, right? GeoMx is also like pieces by pieces kind of concept, right? We found that if you talk to the mass spec people, they'll tell you that this is, you know, like impossible. You must fight junk and the noise is too high. We find a way, figure out a way to actually optimize it to actually detect like thousands of proteins from those. Once you actually solve the technical problem, a lot of these proteins that you detect, they are actually unannotated or even, you know, like PTM in the setting, right? In fact, in the immunological field now, people call it dark proteomics. This is something that our groups, I will say another serendipity findings from this particular study. We are quite focused on this as we call it spatial dark proteomics. Yeah, it's something that we are working very hard about it. Stay tuned on it. To kind of like summarize on this particular paper, we feel that my dream is that now, you know, this work is a four-year work, right? Obviously, you do multi, you know, like non-biased profiling, AI, and also do the ex vivo, in vivo, and all these things. You find kind of like find a way to treat this patient a better way, maybe, right? Can we do this in a fairly fast way in the near future and to actually do all this? Obviously, need the power by AI and do this in like two weeks turnaround time. You can really, you know, like apply this to a patient, you know, like, so that is something that I will call your dream, but I think it's actually very likely to happen. Last but not least, for all the concept that I showed you to you on the AI setting, we also apply all these things to actually like to help some rural and remote area. We call it Global Health Project. We do in Africa and also maybe in Southeast Asia. I can't go into details, but this is something I just want to show you. We have done some works in Africa with colleagues and like-minded friends. This is a very interesting journey. Our first paper that talking about Africa challenge and all these things is actually under revision in NEJM AI. Talking about journal, I would like to also call for papers as the editorial board members. We actually have an AI section in these two journals. Please, you know, like submit to us. We are very happy to actually talk about and promote how to really supply AI in this field and how do we make use of AI to facilitate our findings and validate our findings and also to push the translation of the entire field. I believe, I promise this is really my last slide. I also take this opportunity to advertise our event, which is actually a SITC partner, I would say the only official partner event in Asia. It's called World Immunotherapy Council Asia Pacific. You can find this on the website. If you want to register, you can scan this QR code. It's going to be in July, Singapore. We run this every year, last year in China, this year in Singapore. We will announce next year very soon, the location somewhere not very far from Singapore. The takeaway of today, as again, you know, like talk about AI in the very beginning, set a scene. Then the particularly Nature cover story, I break them into three components. You know, the AI strategy is very important. Also, expandable spatial is important as well. It cannot be just by chance findings. Also, I share with you our spatial mass spec. This is something that we are really excited now to move forward. Oh, I did not show you the future improved TAT, still do not have time. It is okay. I do not think it is very important. There is actually some of the future work. Yeah, it is okay. I think I missed that part, but yeah, next time maybe. So many people need to thank, you know, like pathologists, oncologists, AI scientists around the world and our my group, which is in the center of this one. I welcome any questions. Thank you very much. Thank you, Joe. Thanks for a great presentation and a very interesting story. Let's move to the Q&A section. Thank you. Yes, please. Yeah, we got some questions about our paper you present in your second part. The first one is, what's our biggest takeaway of the paper? Yeah, thanks. I think as I, you know, summarize and also I hope our friends and colleagues here get my point that I repeatedly talk about two things, right, in this paper. First is the AI strategy. I think nowadays the AI and also our big data era, especially in the spatial technology setting, our data are very, very big. When we want to apply AI to our data, we really need to first sit down, you know, be patient and don't, you know, jump into the data too soon. Let's think about a scope, a structure, and strategy of how do we want to feed the data to the AI and answer particular clinical questions or biological questions. I think this is important because not many of us have a huge computing power to do huge numbers of computing for this, right? Unless you have like strong working relationships with Microsoft or Google, these kind of big boys, right? If you are an academic lab, this is actually one of the limited bandwidth which you need to face at reality. I believe that providing some sort of strategy like I showed to you and then to actually convert all this data play into a better recognition. Then you try to filter them into only a few measurables and only feed the AI with this measurable. You can validate them subsequently, either using another models or either using another experiment. That leads to my second point, which is the so-called expandable spatial, right? Because nowadays, I think in the spatial era, I think also you can see that a lot of people, you know, like give you some sort of spatial signature or spatial architecture that, oh, this well might be actually predicting or correlate to one of the clinical outcomes or something like that. I feel a lot of the time, the so-called architecture or signature may not necessarily do very, very expandable in the immunological way, right? Some of these are even a little bit counterintuitive. I'm not saying that counterintuitive is wrong. It's more like I feel this paper, I mean, like in the spatial community, I talk a lot about this. I think like, for example, among the Jedi, the spatial Jedis, I talk a lot about like, I particularly like this paper a little bit, not only not because like this is my paper, but it's like this paper features a concept that you have a spatial discovery finding, and then you actually validate them using different technology, spatial technology, and you validate them from RNA to protein, and in the protein also more than one technology. Also subsequently, you actually kind of like validate them into ex vivo and also in vivo setting and try to make sense that the AI-powered discovery in the very beginning is not a noise. It is not by chance. It is something that makes some sense, right? At least can be explained even just partially immunologically or biologically. I think this is really the two points that I want to actually highlight for this two papers. This one paper, sorry. Thanks. Thank you. Thank you, great answer. The next one is, I think you covered a little bit in your study, but maybe it is worth to address here. What are you working on as a follow-up study of this paper? Thanks. Yeah. Yeah, as in the presentation, I think we have a few follow-up studies from this paper. I mean, the spatial mass spec is really a serendipity, and then we really like the technology that we actually have this idea when we do this paper. This number one is technical level. I think as usual, I would like to highlight more on the clinical levels or things like that. For example, we are working on H&E 2.0 version or even H&E 3.0 version of this scoring system. Stay tuned about this. Number three is that we also, because you can see in this setting, in this paper, we talk about we have a scoring system, yes. Still, you know, like the scoring system is based on the quantification of the images. That is the part that you actually, if you are seeing my screen now, the left-hand side of the screen, right? How do you quantitate the images in a fairly automation setting, especially if you fix the panel, if you fix the markers or fix the channels, and can you put it into a cloud-based software that will generate the report straight away? That is something that we want to work as a generalizable setting, right? Not only for this five marker, not only for this paper, but we would like to do this in a general setting to actually facilitate the entire multiplex IF IHC clinical translation. We believe that this is not something that is very far from the future. I believe also this is something that a few of our colleagues in the world are working on as well. Thank you. Since we still have time to cover the next question, the next question here is, anything you want to share to early career scientists from your cover story? Oh, the early career scientists. Yes. In long story cut short, in the simplified way, I would say working hard, never give up, and also have your huge sincerity at your heart, right? You know, this is, I told you, this is a four-year work. We started the works in 2021. After two years, we submit to Nature and Nature review it and give us a chance to revision. The first revision is the hardest one because it looks like the tone looks like okay, but they ask for a lot, a lot of experiment, especially the NKSL specific knockdown. After I discussed this with our Chinese colleague, and then we feel that, yeah, this is something that we need like at least one year to do it, but we don't know whether we, how do we tell the editor about this, right? I think, you know, like I feel sincerity is your biggest weapon. We actually openly tell the editor, right? We just honest and transparent tell them that we need one year to do this. Can you wait for us? If you can wait for us, we will do this. Then the editors actually call for a video conference with us, you know, a Zoom call with us to discuss how do we want to do that and why do we want to do that. Then does it make sense, and do we have confidence. Subsequently, we also go into multiple Zoom calls with them actually. Yeah, I think it's just that, you know, like you want to do something, you tell people that you want to do something, and then somebody will help you, right? Also, I think working very hard, I think everybody will tell you this, but I think never give up as well, you know, like sometimes, yeah, I think the size is difficult, right? Academic life is difficult, especially in the current situation, right? I think, you know, if you really like science and you really love to do what you are doing and just hang in there, and I believe, you know, like just do what you like, and then the science contribution to science and the contribution to society, and one day you will get recognized. Yeah. Thanks for giving a good suggestion to an early career scientist. Also, thank you again, gave a great presentation. Today we'll end the webinar here. Thanks for all the attendees. Thanks for your attention. Hopefully you find today's webinar is valuable. Thank you all. Thanks very much to Dr. Yeong for his presentation and to Anchen for facilitating the Q&A. If you submitted a question which was not addressed, we will get a response to you after the webinar. Today's presentation was just one session in the Spatial Insights Driving Drug Discovery series. If you missed early sessions, then fear not, these are all available to view on demand. Simply visit our website at akoyabio.com, navigate to the resources section, and there you'll find all the details of the webinars, including talk titles, abstracts, and links so you can listen again. That's all we have time for today. Before you leave, a quick ask if we may, as you leave the webinar, a short survey will appear in your Zoom window. We appreciate it if you could take a couple of moments just to complete that survey, as your feedback will help optimize the webinar experiences that we bring to you. That's all for today, and I'll now close the webinar. Thanks again for attending, and I hope you can join us on our next webinar. Thanks all, and goodbye for now.