2026-04-20-pm-ai-spatial-transcriptomics-pathology

Date: 2026-04-20 Primary topic: agentic-ai Cross-referenced from: agentic-ai, bioinfo-tools, single-cell-spatial-omics Word count: ~12,615

Auto-generated captions

This transcript is Vimeo auto-generated English. Expect misheard names and technical terms. Use the ripgrep/search approach for exact quotes.

Right. Hi, everyone. So, thank you to everyone who came to our session, and I'll be very, very brief. The session, as you can see, is about AI and the meeting point between AI and spatial transcriptomics, to see if and to what extent we can predict spatial gene expression or any other interesting things that one can think of predicting directly from the histopathology slides. And we will have three talks. The first two talks will be by two amazing superstar scientists in this bridging field of AI, computational biology, and spatial transcriptomics. The first speaker will be Mingyao Li, Dr. Mingyao Li from UPenn. I'm sure many of you are aware of her cutting-edge work. The second speaker is James Zou from Stanford. A general AI leader and so on, but James regrettably is under the weather, but we have his talk recorded. Then the third speaker will be myself. And after that, Mingyao and me, we will be happy to take questions and have discussion. Also, if you can see the barcode, or how is it called? I don't know. The thing, the QR thing. You can scan it, and then you can put questions into that link, and presumably, I'm supposed to see them somewhere and answer. But if not, just raise your hand because I'm technologically challenged. Okay, guys? So without further ado, we will hopefully start on time. Dr. Li, please. Looking forward to your talk. Thank you so much for the very kind introduction, and thank you for the invitation. It's a great pleasure to be here. So today, I'm going to share our recent work on digital tissue twins, and in particular, how to turn histology into spatial omics at the population scale. Spatial omics technologies, including spatial transcriptomics and spatial proteomics, have transformed our understanding of tumor ecosystems. However, despite their power, the current spatial technologies remain expensive and slow, and this makes them difficult to scale across the full clinical enterprise. So for example, just mounting a small piece of tissue can easily cost you a few thousand dollars. But on the other hand, many of the spatial omics platforms allow the generation of a histology image from the same tissue slice. And oftentimes, these histology images are very high resolution, high quality, and importantly, these histology images are very inexpensive, and indeed, they are already routinely being generated for nearly every cancer patient in the clinic. So this highlights a fundamental data tradeoff issue in spatial biology. On one hand, spatial omics is molecular information-rich but is expensive and slow, difficult to scale. On the other hand, digital pathology is common, cheap, and fast, but contains no molecular information. So a central question that the field is facing right now is: how can we bridge this gap? And this central question has motivated our development of what I call the digital tissue twins, where our goal is to transform those routine H&E slides into spatially resolved and molecularly informed digital representation of human tissues at the population scale so that we can link spatial biology, AI, and clinical outcomes together to enable risk prediction, patient stratification, and eventually therapeutic decision-making. So in the last few years, my lab has been developing this AI-powered ecosystem to help us build what I call the digital tissue twins. In this ecosystem, we start from H&E stained histology images. We have developed the tools that can conduct the H&E QC and the cellular resolution tissue masking, and the cellular resolution histopathology image segmentation. And leveraging these processed histology images, we also have developed tools that can help researchers to automatically select the most informative samples and the regions of interest for the expensive spatial omics experiments. And after the spatial omics data are generated, we can further leverage these histology images to enhance the quality of the spatial omics data. And after all of this data preparation is being done, we will build AI models to virtually predict the spatial omics data at the population scale. In the meantime, we also perform multi-modal data integration so that we can have a comprehensive representation of every single tissue section in every single patient. And throughout this inference and integration process, we use rigorous statistical modeling to ensure that the AI models that we build are reliable and are trustworthy. So the core idea behind building digital tissue twins is teaching AI to see molecules within a histology image. And this is based on a simple but powerful observation that is similar histology implies a similar molecular profile. So this means that if we can ask the AI to learn this link, then the histology images can become proxies of the spatial molecular marvelous that people typically obtain in the expensive spatial omics experiment. So our work in this space started back in 2021. We were among the first to treat the histology as a primary modality in spatial omics analysis.And then later we developed the iSTAR, where we used a hierarchical vision transformer to extract image features out of a histology image. And this mimics how a pathologist examines a histology slide under the microscope. And after the image features are extracted, we further leverage these image features to enhance the spatial resolution of gene expression data to single cell level. So for example, medium does not provide a single cell resolution gene expression. By using iSTAR, we can spatially enhance the resolution to single cell level. So importantly, iSTAR demonstrated that we can now turn a routinely obtained H&E slide into single cell resolution spatial omics profiles. And building upon the success of iSTAR, we later developed the iSCALE, which utilized the virtual prediction idea to scale up spatial transcriptomics for larger sized tissues that far exceed the capture areas of conventional spatial omics platforms. And this framework is generic. It can be applied to any tissue types of any size as long as you can measure a H&E stained histology image. So basically, by doing this innovative computational analysis, we can turn a task that is experimentally impossible into something that is computationally feasible. So last year, we further leveraged the relationship between tissue morphology and molecular profiles to develop this new tool called i2Omics, where we wanted to use AI to assist biologists to automatically select the most informative regions of interest for the expensive spatial omics experiments. So by doing so, we ensure that you can target your expensive spatial omics experiment to the biologically most important tissue area. And since the development of these methods, my lab has been continuing trying to push the field forward. Where our eventual goal is to turn histology slides into single cell resolution spatial maps at the population scale. But when predicting gene expression directly from histology images, we're facing a fundamental challenge. The challenge is that the gene expression is highly dynamic. The dynamic range can substantially vary from one person to the other. So when we directly predict a gene expression from histology images alone, we don't really know the relevant ranks of the dynamic ranges across the different people we are trying to predict. And this leads to two fundamental issues that we have to consider. The first issue is that when we do gene expression prediction at the population scale, we wanted to make sure that the predicted gene expression magnitudes are comparable across the different people we are predicting. But if you only have the histology images as input, this dynamic range is not available. The second risk is that when you only use the histology images as input, you are risking the possibility of generating a lot of false positive predictions. And this is an issue that none of the existing methods that are addressing or even talk about. But I think in order to apply these virtual predictions in the clinical applications, it's very important to ensure that we are reducing the false positive predictions from all of these AI models. So to solve this question, we asked, "What if we just provided the AI models a little bit of extra information, such that the AI knows the relevant ranks of the dynamic ranges that we are trying to predict?" It turns out bulk RNA-seq, which is very inexpensive to obtain, can exactly provide this information we need. And this motivated us to develop this AURORA model, where we utilize bulk RNA-seq data as a constraint for the AI models when doing the predictions. And in this process, we are trying to mimic how a human look at a histology slide. We gradually zoom into the tissue, and we progressively enhance the spatial resolution of the gene expression prediction we make until we reach to the final single cell resolution. And throughout this process, we also generate cell type predictions at every location we examine. So by the end of this analysis, we are able to give you a single cell resolution spatial map that has a single cell resolution spatial gene expression data across the whole transcriptome. And we are also able to give you the cell type information at every single location in the tissue of interest. So now let me show you some of the evaluations that we have. So in order to establish the reliability of AURORA, we need to compare the predicted values versus the ground truth. So in this analysis, we leverage the high quality spatial gene expression data generated by my wonderful collaborators, Dr. Houmam Akhdar and Lihua Wang's labs at MD Anderson. So in this dataset, we have utilized twenty-three lung adenocarcinoma patients, and we have the medium gene expression data generated from all of them. We spatially enhanced the resolution of this data into single cell level by iSTAR. And in this analysis, we trained our virtual prediction model using eighteen of the samples, and we virtually predicted the gene expression to the remaining five. And this allowed us to compare the predicted values and the directly measured values in the original experiment. And the lung tissue is particularly interesting, because in the lung tissue, there is a lot of airspace. If you see positive predictions in the airspace location, it means that you have false positive predictions. So this dataset allows us to evaluate not only the Pearson correlation between the predicted and the observed, but also allowed us to evaluate the degree of false positive predictions. So looking at the top right corner of the plot, you can see that among all of the methods that we compared to, AURORA achieved the highest Pearson correlation and the lowest false positive rate. In contrast, many of the state-of-the-art methods that are available out there, they generate false positive predictions. So now let me show you how a false positive prediction looks like. So in this tissue slide, you can see that we have a pathologist annotated blood vessels, and at the bottom left, you can see the ground truth measurement of gene MTH11.And if you compare the predictions of different approaches with the ground truth, you can see that AURORA achieved a high correlation and a very low false positive rate. In contrast, two of the top competing methods, they generated a lot of false positive signals. So this is just one of the many genes that we predicted. In reality, if you wanted to utilize these AI models in the clinical applications, you need to ensure that your predictions are reliable, you are not simply predicting false positives. So now after establishing the reliability of AURORA, we next applied AURORA to generate the single-cell resolution spatial maps on population cohort studies. And here we utilized the data produced by TCGA, and we virtually predicted a single-cell resolution spatial maps on nearly 4,000 TCGA samples across different tumor types. And in this talk, I'm going to focus on lung adenocarcinoma, and we applied our virtual prediction model trained on 18 medium samples to this TCGA cohort. So in total, we predicted 885 TCGA lung adenocarcinoma slides. And at the bottom, you can see one of the tissue samples that we predicted. Although we don't really know the ground truth in TCGA, but we do know the underlying pathology. So if you look at the KRT8 expression that we predicted, you can see that the predicted expression pattern agreed very well with the known tumor pathology. So the real value of virtual prediction at the population scale lies in the possibility to identify important and clinically relevant spatial biomarkers for cancer. So to this end, we conducted a tumor spatial niche discovery. We utilized the cellular composition information we inferred from AURORA to do this tumor microenvironment analysis, and we identified quite a few interesting spatial biomarkers. So for example, we were able to detect TLS, which is well known to play an important role in antitumor response. And due to the high-resolution information we predict, we can also classify the TLS into early stage TLS and mature stage TLS. And we can further utilize the mature TLS that we identified to stratify the patients according to the areas of the mature TLS. And interestingly, we found that those patients that have a larger area of mature TLS, they tend to have a better survival. So after seeing this, we then asked a deeper question. We wanted to ask, does the location matter? So in this analysis, we looked at the co-localization between tumor and the TLS, and we found out that those patients that have higher degree of co-localization between tumor and the TLS, they tend to have better cancer survival. So this example demonstrated that by doing virtual predictions, we can even recover higher-order spatial structure in the tumor samples. And another interesting spatial biomarker we recovered in this analysis is what I call the spatial cuff. And in fact, this type of structure was first reported in the "Cancer Cell" paper by Dr. Linghua Wang's group. So in this paper, Dr. Linghua Wang's group conducted a very sophisticated interesting spatial analysis utilizing single-cell resolution spatial omics data generated from Xenium, CosMx, Merscope MC, and CODEX. And they discovered this structure called the Eisman cuff, where you can see that you have the tumor cells in the center, and they are surrounded by the cuffs. And this type of structure is conserved across different tumor types. And more importantly, this type of a spatial cuff leads to poor prognosis. So if you look at the top plot, you can see that using AURORA, we are able to recover and recapitulate this type of a spatial cuff pattern, and we further stratify our patients in the TCGA cohort according to how much this spatial cuff they possess. And we see that patients that have a higher degree of this spatial cuff tends to have a lower survival outcome. So this example demonstrated that those findings that you obtain from the expensive single-cell resolution spatial omics experiments can be recovered, recapitulated using this virtual prediction concept. So while transcriptomics is informative and powerful, but we know that biology is more than RNA. So in order to have a full picture of the tissue that we study, we need to understand the other omic modalities. So to solve this question, we recently developed a tool called MISL, which allowed us to simultaneously model multiple omic modalities together with histology image information for very high-resolution tissue segmentation. And using MISL, we were able to detect a very small structure called the high endothelial venule in the TLS automatically. Although MISL is powerful, currently it's only limited to the analysis of one single tissue section. But nowadays, in the real-world applications, we often see diverse spatial omics datasets generated from different tissue sections, either coming from the same patients or coming from diverse patients. And ideally, we wanted to have a technology that is able to simultaneously measure all of the potential omic modalities on the same tissue size. But realistically, this is almost impossible because as you're adding additional omic modalities onto the same tissue size, you cannot avoid to destroy the sample, meaning that the quality of the data you'll get is going to be degraded. So in order to solve this challenge, we recently developed another tool called Old Cozy, which allows us to use a sophisticated graph-based approach to integrate fragmented spatial omics data to virtually reconstruct what I call the complete spatial omics. So the goal is that by the end of this integration, we can do this virtual prediction and virtually predict missing omic modalities, meaning that we can turn a task that is experimentally impossible into something that is computationally feasibleSo to demonstrate the power of COZEE, let me show you an application again to this lung cancer study. So in this analysis, we virtually integrated three million cells obtained from four lung adenocarcinoma patients. And you can see that the first two patients, they have a histology image measure. They also have a whole transcriptomic gene expression. And the third patient has a histology image and the proteins. And the fourth patient only has a histology image information available. And by doing this integration and computational virtual prediction, we are able to identify a small cluster that is corresponding to macrophages. And the reason that we are able to infer that this cluster is corresponding to macrophages is because we're able to leverage the molecular information provided by the previous three samples. So after we identified the macrophages on patient four using their HE as input only, we visualized our results together with our pathologist collaborator, and the pathologist confirmed that the cluster we identified in patient four indeed corresponding to macrophages. So here are the virtual predictions we produced. You can see that our virtually predicted spatial gene expression profiles and the protein expression profiles agreed with the direct measurements very well. And using COZEE, we can further integrate data in the 3D context. So suppose you have zero tissue sections generated from a 3D tumor block, but in reality, it's very possible that you use different spatial omics technologies using different tissue sections. So using COZEE, we're able to integrate everything together into the same space. And now you can visualize the multi-omic information in the real 3D tumor volume, and you can see some structures that cannot be seen while you study the tissue sections in 2D. So, for example, we were able to visualize the complex TIAs network, and in addition, you can study the spatial relationship between blood vessels and the tumor. And in this analysis, in fact, we did very, very comprehensive evaluations. We considered almost all of the possible omic modalities, ranging from histone modifications to chromatin accessibility, gene expression, and protein abundance, and metabolites. And we also considered almost all of the possible existing commercialized spatial omics platforms out there, and we also considered many different tissue types and disease contexts. If you want to learn more about COZEE, you are welcome to read our recent archive paper. So to scale up this biological insight to the broader field, we recently developed the Foundation model. So here, we utilized the rich dataset produced by Dr. Kadaris and Dr. Wong's labs. And we build our Foundation model using 36 million cells generated from 54 tissue sections. And in this dataset, we have a rich spatial annotation information obtained by the pathologist, and we also have whole transcriptomic gene expression data. We also have protein expression data across 35 proteins. So by building this Foundation model, now the research community can simply fit in your data into our Foundation model, and we are able to give you comprehensive multi-modal information at a single-cell resolution. We're able to give you the pathologist annotation. We're able to give you the whole transcriptomic gene expression information. We're able to give you the whole 35 protein spatial proteomics information, all at the single-cell resolution level. So now let me show you the application to the TCCCA cohort using COZEE Foundation. So you can see that by applying the histology images in the TCCCA cohort into COZEE Foundation, we're now able to transfer the pathologist annotation in our training dataset onto TCCCA. So now we have the annotation available. Additionally, we are able to produce a whole transcript from spatial gene expression and the protein expression for 35 proteins at a single-cell resolution. And because the resolution is so high, this allows us to automatically detect the tumor boundary. So here are four representative samples in the TCCCA. You can see that the boundary we detected is very fine-grained, and this analysis is very fast. Imagine if you ask a pathologist to manually annotate the tumor boundary in the TCCCA cohort, how long that is going to take. And this analysis is totally unsupervised. It's super fast. And this tumor boundary analysis further allowed us to conduct a tumor interface analysis. So, for example, we can look at the cell type compositions along the tumor edge, and we found that those samples that have a higher proportion of lymphocytes along the tumor edge, they tend to have a better survival. And furthermore, we conducted this spatial gradient analysis. We looked from the tumor edge into the tumor core, and we gradually examined the proportions of lymphocytes from the tumor edge down to the tumor core. And this allowed us to examine which patients have a higher degree of lymphocyte infiltration as compared to the other. And interestingly, we found out that those patients that have a higher degree of lymphocyte infiltration into the tumor core, they tend to have a better cancer survival. So, so far I have demonstrated that virtual predictions can help us produce a population scale spatial omics datasets, and this can allow us to automatically detect the tumor boundary, tumor edge, tumor interface, et cetera, and can help us identify potential spatial biomarkers. But in order to utilize these virtual prediction tools in the clinical setting, we wanted to make sure that the results we produce are reliable. And the clinicians need to know where the model is certain, where the model is just simply guessing. So in order to address this real challenge, we developed another tool called Utopia, where we wanted to use principle statistical approaches to help people understand why we are confident about our virtual predictionsSo Utopia is based on a statistical technique called conformal learning, which means that our model, our approach, is a prediction model agnostic, meaning that this approach can be applied to any virtual prediction models. And using Utopia, we can give you a statistically calibrated confidence score depending on the spatial resolution you are considering and depending on the biological target you are predicting. So here we consider the different spatial resolutions ranging from lower resolution to higher single-cell resolution, and we also consider the different biological entities that we predict, ranging from a single gene to a meta-gene, from by the combination of multiple genes, and from a specific cell type to a more broadly defined cell class. So here I want to show you a real-world application. So in this analysis, we focused on the virtual prediction problem in a sample that has gastric cancer. We built our virtual prediction model using the molecular information obtained from the two red boxes, the regions of interest we selected. And after we built our virtual prediction model, we applied this virtual prediction model to the rest of the tissue, only using the underlying historic image information as input. So now you can see that across the different spatial resolutions that we predict, lower resolutions are much easier to obtain a higher confidence than a higher resolution. This makes a lot of sense. And we also see that predicting a single gene like CD4 is a much harder task than predicting a meta-gene like a TIS signature. And we also found that a continuous variable, so the gene expression, is much harder to predict than a categorical variable like a cell type. So basically, by doing this analysis, we are able to give you a principle, a statistical framework that can tell you where your virtual predictions are reliable, where your virtual predictions are not reliable. So this means that when we utilize virtual predictions in the clinical setting, we know where is the tissue region that we should conduct the real spatial omics experiment. So to summarize, spatial omics is a powerful technology, but currently, spatial omics is still too expensive to scale up in the real-world applications. And in this talk, I have demonstrated to you that by using innovative AI tools, we can now turn histopathology data into digital tissue twins, where every H&E slide can now become a spatially resolved and molecularly informed tissue map. And this allows us to scale spatial omics across diverse patient populations, and the resulting data will enable us to do risk prediction, patient stratification, and also allows us to bridge spatial biology, AI, and therapeutic decision making together. So before I finish, I wanted to thank the following members of my lab who did much of the work I presented today. I also wanted to thank my wonderful collaborators and my funding support. Thank you very much for your attention. Happy to take questions. Yeah. Please, questions at the end. Okay. Yeah. Thank you. Thank you so much for this amazing talk. And could you please put on the James Xu's recording? Thank you. Hi, everyone. I apologize that I'm not able to attend AACR in person this year, but I'm happy to present some of the work we've been doing around generative and agentic AI for spatial omics. So this presentation will have two parts. In the first part, I want to present this AI method that we developed called Rosie, which is an approach to generate in silico spatial proteomic images from H&E. In the second part, I want to explore some of the applications in using these kind of spatial and other single-cell omics data to accelerate therapeutic development. And we'll do this with AI agents, with the virtual biotech. So there's been tremendous interest in generating very rich spatial omics information from tissue samples, and spatial proteomics that leverages multiplex immunofluorescence imaging has emerged as a particularly promising approach. So methods like Codex can simultaneously image 50 to 100 different protein markers in parallel on the same tissue sections. And this enables us then to be able to phenotype many more different cell types at higher resolution, and also be able to understand some of the better tumor microenvironments and cell-cell interactions. However, generating such high dimensional single-cell spatial omics data is quite expensive, right? So here's sort of a comparison across different methods. So with H&E images, the molecular information we collect is relatively sparse. It's mostly capturing morphologies, but the benefit is that it's very cheap and also can be collected in a few hours. However, if we move towards this highly multiplexed immunofluorescence imaging like Codex, which is now able to capture at single-cell resolution, 50 plus different biomarkers, but the cost is much more expensive, can be thousands of dollars per slide and also takes multiple days to collect. So there's this really several orders of magnitude increase in cost and complexity when we go from H&E to this highly multiplexed spatial proteomics.So our goal here is to see, can we actually use this H&E kind of images to help us to generate in silico CODEX, in silico spatial omics? And this is analogous in some ways to the ability of AI to try to colorize black and white images into these more vivid colors. Right, so in our case, we view the H&E image as sort of analogous to these black and white images, and we want to essentially use AI to add these different colors to it so that it looks more like this very rich kind of spatial omics, like a CODEX data. In order to do this, we need to build a trainer model that can impute CODEX proteomics data from H&E images. And this requires collecting a large paired data set where we have both H&E images and CODEX measurements that are measured on the exact same tissue sections. So to do this, we collaborated with our collaborators at Enable Medicine, and we generated, at the time, the largest paired H&E and CODEX data sets. So this is the data set we collected for ROSI, which has over 2,000 samples covering 30 million cells from 16 different disease types. And this is all done at sort of a high resolution at the individual pixel level. And importantly here, the CODEX and H&E images are coming from the same tissue, so they're perfectly aligned. And this is orders of magnitude larger than the data sets that we used in previous work from our group, where we were able to also do some of these imputations, but it was done on a much more limited set. So leveraging this large scale paired data set, we developed ROSI, and here's sort of a brief overview of how it works. So it takes input, the standard H&E images, and the algorithm will first crop those images into these individual patches. Each patch will be 128 by 128 pixels. And then we have a deep learning model that takes that patch, right, and then try to predict and to impute the proteomics profiles for each of these 50 or more biomarkers at a much smaller, higher resolution of eight by eight pixels. And the benefit of doing this is that this enables the model to be able to capture information from a broader, larger context in order to predict and impute the biomarkers at a smaller window. And we showed in our paper that this is a more effective approach. And after that imputation, then the algorithm would take these imputed eight by eight patches and stitch them all back together into this whole slide level, generated spatial proteomic image. That's the in silico CODEX output from the ROSI algorithm. This in silico CODEX then will be perfectly aligned with the original H&E sample. So with this approach, ROSI is able to predict in silico CODEX for up to 50 different biomarkers. So here's an example demonstration where we have these two H&E images that were given as input into the algorithm. And in the second column, I show the generated in silico data that's imputed by ROSI algorithm. And this can take maybe less than $1, less than an hour of time. A few minutes. And the third column corresponds to where we took multiple days and thousands of dollars to perform the actual experiment to do this CODEX multiplex immunofluorescent imaging. And comparing the ROSI output with the experimental outcome, which is the ground truth, we can see that the ROSI output actually captures a lot of the key structures and cell types as you do from the actual experiments. And this is orders of magnitude, I think, over 1,000 times faster and cheaper. And in order to show that the results are quite representative across different samples, what I'm showing here is actually the different percentiles of different input histology H&E images, where we rank these different images by the Pearson correlation between their imputed CODEX with the actual experimentally measured CODEX. So 99th percentile corresponds to the samples where the ROSI performed better, and the 25th percentile corresponds to samples where our algorithm performed relatively worse. And what we can see here is that even if we look at the medium sample or the 25th sample where the algorithm performed relatively worse, the output from the in silico CODEX actually still matches a lot of the fine grain structures that you see from the actual spatial CODEX measurements. So this gives us more confidence that the algorithm can be quite reliable. We also perform additional assessments on data on H&E images from other locations. So here are data from three different external sites, and here the ROSI algorithm is also able to impute and generate some of these fine grain structures that we saw in these out-of-distribution external validation sites. And one benefit and one application of this approach was in silico CODEX of ROSI, is that we can actually then infer more granular tissue structures directly from the H&E images. So here we're looking at three common kinds of tissue structures. So in dark blue, it's the stroma, light blue is the more tumor-like epithelium, and orange captures these more endothelium structures. And on the first column are the ground truth that we get from when we actually do the expensive spatial proteomics experimental measurements. We can capture these structures very accurately.And the second column corresponds to if we use Rosie to first impute the spatial proteomics and then do the same kind of segmentations. We can see here we often capture very similar, very precise imputed structures compared to the ground truth. We also did a comparison where we took the original H&E image itself and extract morphology features from the H&E image, and then trained a deep learning model to directly predict these structures from the H&E image morphologies, and that actually performed much worse. So I think this is kind of an interesting comparison that demonstrates the benefit of this Rosie approach to first impute the high multiplex proteomics structures and then use those proteomics information then to classify the different tissue structures. So that ends up being richer and more robust than directly trying to classify the tissue structures from the morphology information itself. And then finally, one of the benefits of Rosie is that it's really able to make these imputations at a single pixel resolution, which corresponds to around 25 microns. So this enables the algorithm then to capture a lot of the sub-single-cell resolution, spatial structures of different proteins. So this enables richer kinds of sub-single-cell or subcellular analysis. So just to summarize, so with Rosie, we demonstrate that from H&E images, we can actually use AI to infer protein expressions, cell types, structures that typically required much more expensive multiplex immunofluorescent imaging assays. And this demonstrates the power and the promise of AI to help us to bridge the gap between these more common and routinely available assays and these much more experimental research-level assays. And all of this can help us to enable precision pathology. So with Rosie, we see how with these kind of generative AI approaches, we can enable us to make these spatial omics information more broadly accessible by imputing them from H&E images. In the second part, I want to highlight some examples of how we can then take these kinds of single-cell and spatial omics assays and help us with therapeutic developments. In a recent work, we created what we call a virtual biotech, which is a large organization of different AI agents that can help us with drug discovery and development. In this virtual biotech, it's a team of tens of thousands of AI agents, which is led by a chief scientist officer, or CSO agent, and working under the CSO agent, we have these different divisions, which corresponds to the standard cross-functional divisions of a typical drug discovery organization. So, for example, we have a division of agents that focuses on target identification, another division of agents that look at safety, clinical developments, and so on. All of these divisions are operationalized by teams of AI agents with different specializations for specialized roles. For example, under the target ID division, we actually have an AI agent who specializes in doing single-cell analysis, another agent that's very good at doing all the statistical genetics analysis, and also agents that specialize in doing these more spatial analysis. And behind the scenes, to enable these agents to carry this very in-depth analysis, we created a universe of different resources and tools that enables agents then to do analysis from the raw data. So this corresponds to hundreds of different MCPs that we've collected. So, for example, we have created MCP so that enables the agents to directly access and download single-cell atlas data that are relevant for their analysis. They can also access all of these different histology H&E images as well as data from genetics, from clinical trials, and from chemistry. And the way that all of this work together is that the user might have a specific query, and the query could be, "Help me develop a therapeutic strategy for a particular kind of disease target," for example. And then the CSO agent would take that query and work with his chief of staff to come up with a plan for the analysis. And then it would delegate that plan to the different divisions of agents. All these agents then would proceed in parallel. Maybe one agent will look at the histology images, another agent will look at the genetics data. They can do all of these analysis in parallel. We have a reviewer agent that reviews all of these intermediate results. And then finally, all of these results will be synthesized by the CSO agent, then presented back to the human user. So I want to illustrate then a case study to demonstrate the workings of this virtual biotech. So a specific question we asked the biotech agents is to help us to design therapeutic strategies for CD276, also called B7H3, as a target in lung cancer. So this is a transmembrane checkpoint inhibitor protein, which has emerged in recent years as a very exciting target. So the first thing that the virtual biotech agent did is that it started to look at some genetics evidence and actually found that there's really not very strong genetic evidence around this target. But then the agents reasoned that that's okay because this is a target that's more relevant in tumor cells, whereas a lot of the genetics is looking at the germline mutations. So then the agents pivoted and decided to look at single-cell data. So it first pulled single-cell data from different cell atlases from tissues that are relevant to lung cancer. And by analyzing those single-cell data, it found that this target is particularly highly expressed in tumor-associated fibroblastsRight? And based on that, the agents came up with a hypothesis that perhaps these tumor-associated fibroblasts are expressing this checkpoint inhibitor, which then has the effect of suppressing the immune response in their surrounding cells. So to test that hypothesis, the agent next decided to perform a cell-cell communication analysis. By analyzing receptor ligand pairs, the agent actually found that indeed, these tumor-associated fibroblasts have a lot of communications with all the different major immune cell types. But here is where the agent reasons that, okay, to really confirm that in cell-cell communication, this will be really useful to actually analyze spatial data to really sort of be able to visualize these patterns of these immune-suppressing effects. And this is where the spatial analysis becomes really useful. So to do this, the agent actually found and downloaded these spatial transcriptomic datasets from Visium. And by analyzing this dataset, the agents actually showed that indeed in areas where these fibroblasts are expressing this P7H3 protein, the surrounding neighborhoods are systematically depleted of different immune cell types. So there's quite a significant effect. So this actually really supports the agent's hypothesis that indeed that these fibroblasts are expressing this target, and then which has the effect of suppressing the immune response in their neighborhoods. And then finally, the agents are able to then connect all of this together with human clinical cohorts. So we look at cohorts of patients from lung cancer that are expressing this target either at a high level or a low level. And so by stratifying this target expression, the agents actually found that patient cohorts that express these targets highly are less responsive to different treatments, as measured both in terms of overall survival and also disease-specific survival. So all of these different pieces of evidence come together, and it's all conducted autonomously by the AI agents in the virtual biotech to form a mechanistic story of how this target in fibroblasts may have this kind of immune-suppressing effects, which then leads to worse outcomes for patients. Now, based on all of that information and by analyzing all of that data, especially the spatial omics data, the agents actually then designed a therapeutic strategy for CD276 in lung cancer. So from its analysis, the agent said, "Okay, I recommend designing an antibody-drug conjugate, or ADC, or a monoclonal antibody that's targeting CD276." And this antibody then will be linked to a chemotherapy payload, DXd. And this is designed and proposed by the agents based on information available to it before January 2025. And what's really quite interesting is that several months later, in August 2025, we have the FDA had actually approved-- gave a breakthrough designation to Merck. So Merck had actually independently converged onto a similar strategy of ADC with the same antibody targeting CD276. It's actually linked to the same chemotherapy payload, DXd, and they showed quite promising results, which received this breakthrough designation from the FDA. I think this provides actually a very nice and exciting external independent validation for how a therapeutic strategy from the agents which came from leveraging all of these single-cell spatial analysis can then be supported by clinical study results from third parties. So another thing that the agent did more broadly that's quite exciting is it has helped us to curate data from all these different clinical trials. So here we ask the agents to basically curate clinical trial outcomes because all of these clinical trials are registered, but often the outcomes of what happened in these clinical trials are spread out across different sources, different papers, press releases, sometimes company websites. It will be very useful if we have a centralized database where all these clinical trial informations are collected. So here the virtual biotech spun out thirty-seven thousand clinical trial agents. Each agent is responsible for annotating the outcomes of one trial. So that agent then will go for that trial to read all the things that happened in that trial and look up all the papers, the news articles, websites to find out what are all the endpoints and what happened to the endpoints. If the trial had stopped, why did it stop? Are there any side effects, adverse events, and so on. So this creates a really nicely curated database of clinical trial outcomes for fifty-six thousand trials. And then based on that, we ask the agent, so can you find single-cell features that's based on these molecular features that are actually strongly predictive of clinical trial success? So that would be if we have features that can predict trial success, then that's often a holy grail for the drug development because then it can help them to prioritize different drug targets. So here the agents actually came up with two interesting single-cell features based on single-cell transcriptomics information. The first is that the agent proposed this tau cell type specificity score, which measures if you look at the drug target, how selective is the expression of that drug target across different cell types. The second feature measures the bimodality, basically how on and off is that drug target across different cell types.The agents reasoned that if this drug target has a very distinct on/off switch pattern, then it's more controllable and potentially a better drug target. And then by collecting these single-cell molecular features with actual clinical trial outcomes, the agent found that indeed, drug targets that have these more cell type-specific expressions are 48% more likely to reach market, right? And overall, if we look at both the tau score and the bimodality score, they're highly correlated with better outcomes across phase two, phase three, phase four. They're associated with better chances of meeting the clinical primary and secondary endpoints, and these trials are less likely to be withdrawn for various reasons. So I think this is quite exciting, in particular because these single-cell features are actually orthogonal to genetic features. So people have known for some time that drug targets with genetic evidence are more likely to be successful. And here the agents actually showed that even among drug targets that have no genetic support, if they do have these single-cell evidence supporting it, then those are more likely to be successful. So I think this is an example of how these agents then can help to integrate single-cell data and then help us to prioritize different drug targets. So I'll just stop here. I hope to show these two examples illustrate how this generative and agentic AI can both enable spatial omics analysis and also leverage this analysis to help us to improve our therapeutic developments. I want to especially thank Eric Wu, who led the ROSI work, and Harrison Jiang, who led the virtual biotech work. Thank you very much for your time. Okay, I'm just timing myself. Awesome. So I think the most useful agent that we can possibly build today would be a very simple agent, actually, that just signs data-sharing agreements between institutes. Just imagine that we can cut the half a year nonsense lawyering up of data-sharing agreements by two agents that do that in one day. That would be a tremendous contribution to cancer research. Okay. So back to spatial biology. I want to talk about towards fast and accessible precision oncology, and I'll aim to do it in 20 minutes so we'll have time for discussion. So these are my disclosures, and this is published stuff that I will not talk about, and I just want to explain what is this dividing line. So in this published work, my lab over the last two or three years have developed ways to infer molecular omics data bulk, gene expression, methylation, what have you, directly from the whole slide images and then use that for clinical translational applications. So, for example, when we infer transcriptomics, we showed that we can use the inferred bulk transcriptomics from the whole slide images to infer patient survival and response almost at the level that you could do from the sequencing data. And when we infer methylation, we focused on brain tumors where methylation is used-- Sequencing methylation is used as the gold standard for classifying brain tumors at a high resolution. We have shown that we can actually do that at a level that is practically indistinguishable from the sequence methylation. We can do it from the inferred methylation. And you can see publications from different years just to show you that in a span of one year or one year and a half, we can double and triple the accuracy of the clinical value that we can bring to the patients on the same tasks. And that's whole bulk inference clinical application from whole slide images. And the reason that we can triple, quadruple our accuracy is because we are leaning on the advancements from other groups that are building better and better foundation models, and we also have better biological insights and clinical insights and so on. But what I want to focus today is on these types of analysis where we move from inferring bulk molecular data from the whole slides to inferring spatial omics data as Mingyao has shown so nicely. And our goal is to learn to teach the machines to learn to infer spatial expression, spatial proteomics at any spot in the tumor, at any cell in the tumor, automatically segment and annotate the cells, and use that to define and study and validate a conceptually new typeOf clinical survival and response biomarkers that are spatially grounded. And that's a whole new arena that requires a lot of thought and a lot to study, and we are just making our first steps, but I think that in the next two, three years, it will be transformative and hopefully, reach the clinic sometime soon. So the first thing that I want to share with you is this paper, which is to appear soon, which has been led by Eldad Schulman and Emma Campanolo from my lab with many collaborators, where we taught machines to learn visium data in breast cancer by training on visium data. And we then created maps, inferred spatial visium maps of the whole breast cancer TCGA compendium, which is 1,100 patient samples at a detailed spatial resolution, and then we used that to characterize the TME of these patients, find survival predictions, treatment response predictions, and so on. And just to show you a bit on that before I move to the really unpublished data. So you can see here overall we can robustly predict about 4,000 genes. Here you can see the histograms of the correlations between the measured and inferred genes across the spots in each sample, and we did extensive validations, and we can predict about 4,000 genes. The spatial transcriptomics of visium of 4,000 genes. So out of the 10 or 15,000, we can predict only 4,000. But it's important to notice that these are not random genes. These are highly enriched in the main cancer hallmark pathways. It's interesting when you think about it, with the inferable predictable genes, that those are evident from the slides are actually those that are belonging to the cancer drivers and hallmarks. And as you know, one picture is worth a thousand words. Here are three key genes in breast cancer, ELK2, KI67, and MYC. And on the left side of each pair, these are samples from patients. This is the measured spatial visium transcriptomics, and this is the inferred. Here is another patient sample, the measured and the inferred, and so on. And you can look at it, each of you, and get a sense of how good it is by yourself. And of course, we are not perfect, but you can form your own impression. And just to show one last slide from that topic because I really want to leave ample time for discussion. So this is two AUC curves, patient survival. This is a study, Samut et al. by Carlos Caldas' group, published in "Nature" about two and a half years ago, where for each breast cancer patient, they did whole exome sequencing, RNA sequencing. They also built direct predictors from the whole slide images and also from the clinical and lab data. And on the green AUC curves is what they could get, actually, what they publish in the "Nature" paper for these patients, the AUCs for the response to treatments, chemotherapy or chemotherapy and trastuzumab. And the blue is what we do. So the green is for about $10,000 per patient and many weeks of work, and the purple is what we do just from the image by spatial characterization of the tumor microenvironment, and you can see for yourselves. But where I want to take you now is to completely new and unpublished data, where we are performing an attempt to infer spatial transcriptomics to automatically segment each cell, infer its spatial transcriptomics on a slide, and then-- Oh, sorry. Before that, I want to take you to a foray where we infer spatial tumor hotness. So what is tumor hotness? We have a paper on bioRxiv where we found a signature of the tumor immune activation. Again, one number for each slide, okay? And that's the tumor hotness, okay? And we find that this one scholar, the tumor hotness, across many different cancer types, this one scholar is actually, in our hands, the best predictor that we have of checkpoint immunotherapy response across different cancer types, different clinical trials. We never trained on them. So we just look at the slide, we estimate the tumor hotness, one scholar from zero to one, and this is the kinds of AUCs and odds ratios you get. So the mean odds ratio that you get across different test cohorts is five. Well, if you aim to build a direct predictor without automatically inferring this hotness signature and so on, you get much, much worse results. So it shows you the power of inferring molecular data. But this was, again, one hotness scholar, okay?For a whole slide. But we went on, and this is available on BioRxiv, and we'll soon submit it. But we went further, and we built, and when I say we, it's Tai Huang, we built predictors of the tumor immune activation, the tumor hotness at each spot on the H&E. So it's a pan-cancer predictor trained on 12 cancer types and 208 slides and so on. And it achieves accuracies of larger than 0.8 in inferring the tumor immune activation in its spot. And if you think about it, it opens very interesting opportunities to systematically study the true cellular modulators of tumor hotness. And here again, you can see some pictures, and you can see a head and neck image, the H&E slide. This is the matching Visium slide. So since we have the hotness signature, we can infer the signature on the measured H&E. We call that the actual spatial hotness. And here is the totally inferred hotness just from the H&E slide. And you can see, again, you form your own impression from these few examples on how accurate or inaccurate we are. And just to show you that, so we have taken, since this is a pan-cancer hotness predictor, okay, so we take the good old TCGA data. And, as you know, thousands of papers have built survival, Cox survival predictors based on the measured RNA. So that invests thousands of dollars in the bulk to do it for each patient and what have you. And here you can see how actually, quite surprisingly, the level of immune hotness, one scholar has inferred from this spatial analysis and summed up, I'm not going into the details, actually, the spatial hotness of a tumor in most of the different tumor types is much better predictor of patient survival than the RNA sequence data, the thousands of papers that use that. Okay, so now I'm coming to our efforts to develop these spatial predictors at the single cell, and I'll show you ongoing work in inferring Visium HD, inferring Codex data, and annotating cells based on these inferences. So again, one picture, just to give you an idea. This is a colon cancer data set or no colon, published last year in "Nature Genetics." Here is the original picture, the slide, the Visium HD of the authors, and you can see the different cell types that the authors have annotated. And here is the inferred annotations of the cells based on the inferred Visium data. This costs $10,000, and you have to wait weeks. This, you buy my fellows a cup of coffee, and you get it the next day. And you can decide where to invest your money, right? It's a free country still. And this is another slide. This is a normal colon. And you can see we are not perfect, right? Even my fellows are not perfect. So you can see here there is a measured, annotated something, red cell something, and we miss it here and so on. And if you look carefully in this slide, you can see that there is some false positive made. I don't see them, but trust me. So we make mistakes, and we are not perfect in both. And now I want to take you to predicting Codex data, what James has shown, except that he's not here, so I can say, but much better. Okay. That's the price of staying at home. So this is a pan-cancer Codex predictor. And you can see on the Y-axis, the correlation between the measured and inferred Codex level of each biomarker across each individual cell. And these are the top ones. There's other ones that I'm not showing that we do less. But overall, we do mean correlations of 0.4 to 0.6, depending on the indication. These are results in cross-validation in three cancer types on which we trained, but then we have an ensemble pan-cancer predictor. And these are actually test results on independent data sets. So you can pick your favorite markers, and again, you can see what one can do. So again, just to quickly show, here is, let's say you're interested in PD-L1 or CD833. Here is the original H&E. Here is the measured matching sample of measured PD-L1, and here is the inferred PD-L1 in that sample. And here is CD48, the H&E, the measured CD48, and the inferred CD48. And this is for two other markers, CD45, for example, and KI67. And these are not particularly good examples. These are really typical examples, and you can see the overall correlations. And based on thatYou can now, since we segment the cells and we have the markers inferred and so on, we can annotate the cells. We can't do it yet at a very high resolution. We can do it on a resolution of five or six cell types, and that still isn't as great as we should. But you can imagine that we can infer not also the markers, we can having predictors of Visium HD or Xenium HD, we can combine predictors. So we are going to improve that in the near future very much. But even so, we can infer tumor cells very accurately and other cell types, quite interesting. And I'll show you the importance and application of that. And what is important, as Mingyao mentioned, we have ways to estimate the confidence of our predictions. So if we are ready to give up and leave out some of the cells and focus on the more confident cells, then we can have very confident predictors. As I always tell my fellows, less is more. So just to show you, since we can predict cancer cells so accurately and annotate them, automatically, you give us a slide and we pinpoint with 97% accuracy or more every tumor cell on the slide. Just imagine that we put it on the pathologist's desktop. He puts on a slide, immediately gets all the tumor cells or what have you. So what can you do now the dream is happening? Okay. You can annotate confidently a large subset of the cells which you segment in each slide, just the H&E in each tumor. So you can now start to systematically look for all the triads, quadruples, quantiples-- I don't know, my English is not good enough. All the spatial relations between different cell types and associate them, and that's what important, with patient survival and treatment response, because you can apply to large cohorts of hundreds of thousands of patients. All you need is love. All you need is the H&E slides, right? So that's what, in a study led by Amos Temer and Tai Huang, they can systematically scan, this is ongoing work, all the triads. Very interestingly, we see that this is in breast cancer, that triads that involve lymphocytes and endothelial cells, lymphocytes emigrating from vascular blood vessels and so on, are highly associated with either increased hazard ratios in survival in breast cancer or decreased ones. So actually, the lymphocyte endothelial is associated with a decreased hazard ratio. And when you look at triads that are associated in chemotherapy and trastuzumab response, you find again that that triad of lymphocytes and endothelial cells is associated with this five times odds ratio. This is clinically very meaningful, if ever validated, of patients' response to trastuzumab. So in summary, to keep in time, we are developing a comprehensive suit of AI predictors at the single cell, at the Visium, at the transcriptomic level, at the bulk level, at the transcriptomic level, and at the single-cell level that predicts anything you want to know about what's happening in your slides. It's really a dream come true that is unfolding as we do it. And that gives rise to the possibility to questions that I couldn't think we could study one or two years ago. What are the mediators of tumor hotness and tumor immune activation at the cellular level? And a new type of spatially grounded biomarkers, and possibly a way to democratize precision oncology to really transform its accessibility to patients in many underserved regions in our nation and across the world. In principle, it could be developed in other disorders. And just at 1:00 pm, I was asked by the editors of Cancer Discovery, you can scan this thing, and at 1:00 pm, they said they published a paper, a prospective paper that we wrote in Cancer Discovery that further elaborates and describes this vision, which is close to my heart, of democratizing precision oncology. This is our lab from one or two years ago. This is my close friend, Ken Aldape, who was a partner in crime for all the methylation and brain work, and many of our fellows. This is actually the garden in our house, and now you may understand why I'm not moving in St. Louis and staying in Bethesda. So, thank you all so much. And Mingyao, would you like to join me, please, to answer questions? And I think there are questions on the right. And anyone can also ask us. So, thank you so much. We have 12 minutes. So I am not sure I know how to operate thatSo let's start from questions for the audience. Is that okay? Anyone wants to ask? Everyone is running away on the most interesting-- Because you need to go to other places. And yes, please. Yes. Hi, Eitan. Hey. Fantastic work as always. And I promised I would try to ask you a question that would bring on some more discussion and perhaps be a little controversial. So when we think about the evolution of tumors under therapeutic pressure, we have to sometimes disregard some of the preconceived assumptions about established gene networks. That certain patterns of gene upregulation and downregulation that are important in the establishment of a tumor are likely going to change under therapeutic pressure. In your spatial prediction models, how do you account for the likely heterogeneity and perturbation of predicted gene expression based on morphology when you've trained on a lot of untreated samples? Thank you so much. Mingqiao, would you like to answer, please? Or you want me or whatever. Please. I can try to answer, but I don't think I have a good solution for that, because as you can see in my training data, I don't have patients that have response to a therapy. So I think if you don't have the patient to have a response directly measured, I think the information you learn is also limited. As I mentioned in my Aurora work, we see that the dynamic range is a substantial variable from one person to the other. I think the same thing is true when a person is undergoing a therapy. Maybe the dynamic range is different. So if we have some type of training data available, I think that the model should be able to see that. But right now, if we don't have such training data, I think the ability to do that will be limited. But that is my guess. I don't really know. Right. And I think that, just as a quick follow-up, and then I'll step down, I would be interested in understanding or predicting mechanisms of resistance. And I think that's a challenge right now with any morphology-based approach because there's a near infinite combination of interactions between tissue acquired and whatever spatial data you have available and, of course, the therapies that are being tested. I feel like this is a very big challenge for our field, and I'm looking forward to great advances in this. And I'm curious whether you have any other thoughts on this very challenging issue. I- Go ahead. Okay. I can answer this question because I thought a lot about this. I don't think virtual prediction is able to solve everything. The only purpose, for me at least, why I do virtual prediction is I wanted to generate a virtual population cohort so that I can identify the important tissue regions for us to focus on. Because it's just not possible to generate the real spatial omics data for thousands of patients. However, we can use the virtual prediction as a start to localize which regions in the tissues are really important. And once we localize the proxy of the regions, we can conduct the real spatial omics experiment and investigate what is really going on. That is a method about doing virtual prediction. Yeah. Please. Thank you so much, Adam. Please. Hello. It's just a general question. So you showed that you did the enrichment analysis on the high correlation genes, but did you do the enrichment analysis on the low correlation genes as well? So which genes are badly predicted in general? All the genes which are not well predicted. Very good. No, I mean by default, we point to the main hallmarks, cancer hallmarks, that are well-predicted, but no. But I'll tell you what. There's some disappointing news here because a lot of the important immune genes that we would have loved to see in the predicted pathways are not necessarily inferable at the spatial resolution. And that is because they don't have a priori, a spatial variation, a large spatial variation. So think of the extreme case. If a gene has zero spatial variation, you cannot predict it. There's nothing to predict. There's no variation for the machines to predict upon. So there is no systematic comparison there then? It's like there is specific pathways that would be well predicted. For example, G2M checkpoint, like with the KI67, it's pretty well predicted. But let's say, I don't know, some other pathway like ME targets? So we know what is predictable and what is not, and that's the important thing, okay? Know your limitations, know where you are good, and work with that. Yeah. Think positive. Yeah. And just one suggestion, too. I would just reconsider the "path to hotness" name because I just Google and I'm afraid my IT team will ping me back. You afraid what? Sorry. So I Google "path to hotness" on Google. Yes. Google corrected for other stuff. Google corrected for? Like, just corrected. Like, it spelled it all other things. Okay. Thank you. Okay. Thank you so much. Yep. Thank you. Hey, how's it going? Kevin Boehm from Sloan Kettering. Great talks. My question is along the same lines around, I understand that you can identify cancer hallmarks, but to what extent can you identify subclonal determining phenotypic states? Do you want to answer, Mingqiao?Please. I don't think I can answer this question. Okay. So I do not know the answer. That remains to be studied. I think to understand the spatial organization of subclonal, you really need to be able to infer sequencing data on a spatial level, mutations, copy numbers, what have you. You can do it from spatial expression to some extent, but I'm not a big believer. So I think it's a very important question, but I don't think yet we have the means to answer it. Sweet. Thanks. Yeah. Jerry Lee from NCI. Nice to see you. That's a great talk. But the question is for Dr. Li. I may miss this. My question's about the models you developed. Were you able to distinguish different tumor cell populations with different genetic or epigenetic background without additional training data? So currently, the model is not able to distinguish that because, as you have seen in my talk, not all of the training data are obtained from spatial transcriptomics. So I do not have information on the epigenetic profiles. But all of the prediction models, I think, are very generic in nature. If we have spatial omics data that have the epigenetic information, we can do such training as well. So my lab has been collaborating with Dr. Yanxiang Dong at Penn, and Dr. Dong is an expert in developing spatial epigenetic technologies. And we have been applying our virtual prediction models to his datasets, and it works equally well. So we were able to computationally predict chromatin accessibility and epigenetic marks. But it's not at the population scale because the training data we have is still very limited at this point. Thanks. Thank you. If there's not, maybe we can still figure out. We have three minutes to answer questions. Where do I see the questions? They should be right on the monitor. But it's not. Oh, these are the questions. Oh my God, so many questions. Oh, God. Oh, here is something for you because I see TLSs. Oh, okay. I'm saved. So how do you know the detected aggregate of B and the T cells is TLS without IHC and IF for validation on the same sample? That is a good question. I don't think the calling of a TLS is accurate. I think a better name for it should be a lymphoid aggregate. That's another question. Or that's James? That is for me. Okay. COSY. So what is the resolution in map at which COSY operates? The resolution is eight micron. It's close to a single-cell resolution. Next one is also for me. Yes, go ahead. Can you extend your analysis to set pathological samples such as fine nodule aspirates and bone marrow peripheral smears? I have not tried that, so I don't have an answer for that question because I don't have access to this type of data. Did you consider these checkpoints? I have not. So yeah, we focused on PD-1, PD-L1, and then there is a question, and I think this will be the last one. Which spatial transcriptomic platforms is better to make an AI model? Would you like to? So I have experience working with all sorts of spatial omics platforms. And among all of the platforms I have tried so far, I prefer Xenium the best. I second that. You second that? Yeah. I second that, and I just want to say that now 10x has released an all-gene Xenium platform, which I think should be amazing, and I'm nothing with 10x or anything, which I think could be extremely helpful. But in general, we found Xenium 5,000 to be a very powerful platform, and now with this new whole gene transcriptomic Xenium, that could be transformative. You want to see if there's any other question? I think there's one question for you, Doctor. Oh, for me? Yeah. Okay. How is the tumor hotness quantity defined? Okay. The tumor hotness or the tumor immune activation, we didn't invent the term hotness. Everyone talks about hot and cold tumors in the literature, right? So it's inferred by a level of enrichment of some immune activation signature, which we derived, and it's in this bio RV paper, and I didn't have time to explain. And the level of enrichment is normalized and calibrated to a scholar value between zero and one. Yes, please. Hi. Last question. So, just one comment and question. I think you said that predicting spatial expressions from H&E is cheaper than doing molecular testing, but then on the other end, to produce that data, to even model that further- To produce that data? Produce that data- Yeah ... using Xenium and Visium, it requires a lot of cost structure in place, funding, et cetera. How do you anticipate this whole technology of capturing the spatial transcriptomics or proteomics, first using Xenium or Visium, and then making models on top of it, being clinically translated in a cheaper way compared to just doing molecular testing? No, that's the whole idea, right? You only generate this training data once in principle. Well, you generate two free training data sets because you want to test and validate it, but then you're good to go for eternity just from the H&E slides. That's the whole point. But would you be okay with just having it trained once and not trained again? That's okay. And- Of course, you can do that in a year or two or whatever, and it may be in different ethnic groups and things like that, but that's very cheap. You do it just once, and then you can run on thousands or ten thousands or whatever of patient slides, and you can get results in real-time with almost zero cost. There's no question that if you can do it, it's extremely valuable. We need to carefully understand the different population, different scanning machines, prospectively, whatever, how well we can do it. But if we can do it, that's huge. Right. So to scale the data to that level, to capture enough heterogeneity in each and every aspect that you mentioned, that requires a good amount of funding and- Of course ... over the time it will pay off, but then how do we justify that to our community? I think it's a no-brainer. The value is tremendous. This is transformative. I'm not bothered at all by the ability to justify it. I'm just-- Carefully, we need to test how good it will work. That's the thing. That we still need to study very carefully. Okay. Thank you. Oh, thank you. Last question, please. Hi, this is Ray from Inside. Very nice talk. So, I know when we're trying to train the AI model, we always make sure that the sample has a very good QC. But in the real clinical settings, so the sample QC is always at issue. So I want to know, have you had an understanding how the QC can affect- Yeah ... the accuracy or the prediction? 100%. I got you. I'm so sorry. Just because of the time. Would you like to answer first, or you want me to pitch in? Sure. I can answer this question because I thought a lot about that. If you remember my slide on the ecosystem of where I try to build, we start from QC of the pathology slides. That step is very important. And thank you so much, and I can only second and say that we also developed careful QC tools and measures. We show it in our upcoming Path to Space paper, which is-- earlier versions are on bioRxiv, but the full version will come out in cell, I don't know, in two months or something. And if you look there, we have whole sections about QC measures, both on the slide quality scanning level and on the transcriptomics level and how they influence the quality of the prediction. Okay, so we did an in-depth study of that question. Thank you so much because it's very important. Thank you. So in reality, when patient slides come in, we will run QC and only if it's reasonable, we will aim to make predictions. If not, we'll say, "Guys, we are sorry. We can't trust our predictions on that and just do the sequencing or what have you." Okay? That's helpful. Thank you. Thank you so much. Thank you everyone. We are past the time and thanks to Dr. Zhu and yeah, Mingye