and we can start the session exactly on time. So I'd like to welcome everybody to this very topical and forward-looking session on artificial intelligence for the diagnosis and the treatment of liver fibrosis. We are fortunate to have five experts on the field who, of course, you'll hear from immediately. And each speaker will be speaking for 15 minutes, and then afterwards, all the speakers will be available at the front to answer questions. So please do have lots of questions, but please hold onto them until after all five speakers have given their presentations. And then I would also like to remind everybody to stay afterwards for the fibrosis SIG meeting, which we'll be taking in this room immediately after this session. So for the first speaker, I'd like to welcome Dr. Mary Ranella to the speaker's podium, please. And Dr. Ranella will be speaking about current methods and gaps in the staging of MASH fibrosis, where AI may help. So we're just at the starting point, so please. Thank you very much, Rajen. Thank you for the invitation. So over the next 15 minutes, what I'll do is try to explain to you what the main issues are with histological assessment of fibrosis and how we might minimize some of these with the use of AI. So advancing is a bit of an issue right now. There we go. OK, yeah, but I can't advance beyond that, which is not great. OK, these are my disclosures. They really are not relevant other than some work with histoindex. So by matter of outline, what we'll start with is outlining the challenges in the histological assessment of fibrosis, particularly in the context of MASH, the limitations of the ordinal scales that we use, the inter-observer variability that's been a problem with respect to the assessment of early efficacy in therapeutic compounds, how digital pathology might address the existing barriers that we have through improving the dynamic range, quantitative metric, providing a quantitative metric of collagen burden, and providing an unbiased or less biased approach. This has opportunities to assess disease progression and regression as well, novel patterns of response, possibly mechanism of action specific, and clinical outcomes. So I think we'd start with advanced fibrosis, actually cirrhosis in the context of viral hepatitis. So even in the context of viral hepatitis where you're able to completely eliminate the cause of inflammation and that results in fibrosis deposition, you can see on the left, in a patient treated on it with Entacavir, that after a year of having no viral load, you have still cirrhosis. So it takes actually five years or up to five years. That's where the follow-up IFC here is to really show very demonstrable change in fibrosis. And to the right, the same with hepatitis C after SVR. So asking MASH in this context to show a one-stage reduction in fibrosis, which is the FDA mandated endpoint for approval is very much a tall order. So collagen burden across stages is of course not linear and you can see that most of this big range is gonna come in the advanced fibrosis and that's really where the concentration is in clinical trials right now. So categorical scoring really cannot account for this substantial variation within stages. And here's an example of a patient with cirrhosis with a very dramatic reduction in collagen content, although nodules still persist and this technically is still cirrhosis over towards the right. We know the changes in collagen proportionate area are tracked with clinical events and that was demonstrated at least in this context in the CYM2-ZMAB and STELLAR4 data and also to progression to cirrhosis as well. So there are many challenges as I mentioned, the high dynamic range of fibrosis content, especially in advanced stages, limitations of biopsy size and quality, which actually can also impact machine learning and difficulty showing changes in cirrhosis missing potentially efficacious compounds. So with respect to quantification of fibrosis in mazzled, CPA was sort of the first thing that we started to see and it was actually really exciting because you could measure and then correlate with outcomes. However, there are also a lot of limitations. There were issues with artifact, there were issues with underestimation and overestimation, particularly in the context of cirrhosis where you really need that granularity. This was generally in fragmentation led to underestimation of fibrosis. So you need generous tissue and that's clearly not what we have a lot of the time. So how might AI-based methods help us overcome some of these barriers? So there are multiple phases that go from when a patient has a biopsy down to how this is analyzed and assessed. And so there are many ways where things can vary and be different from one system to another. So what I'd like to do is basically give you just a very high level view of the platforms that are most used right now in this context. So starting with PathAI, uses standard stains, predominantly mesons trichrome, has good reproducibility and reliability. It uses pathologist annotation and the output tracks nicely and I'll show you in a minute with CRN stage, which is the fibrosis scoring system we use in MeSH and provides color overlay highlighting important features. And then further provides a continuous fibrosis score as well. It's been used in clinical trials to corroborate histologic change that's been demonstrated. It's also been used to show identifying evidence of improvement in the absence of a categorical change in fibrosis and is seeking biomarker qualification. Pharmanest through FibroNest, which is the fibrosis measurement, again, uses standard stains, has good reproducibility and reliability, is trained on phenotypes. So, and this is a little bit different. So whereas other systems may use a pathologist annotate or a more sort of unbiased approach, this has a little bit of a different twist in the sense that it provides the phenotype and then it asks the machine to tell them where the differences are. And here you get a continuous scale fibrosis output and information on the quantitative, our quantitative comments on fibrosis traits individually to a very detailed level. Again, this has also been used to corroborate and identify areas where efficacy may have been missed. HISTOindex uses a slightly different or a very different technique, uses unstained deparaffinized slides. This is a proprietary second harmonic generation microscopy of digital path AI, has very good inter-machine and intra-machine agreement. You can see here that it's trained also with pathologists after the processing and then it's, that information is used to be, to sort of train the algorithm. The outputs are very detailed with respect to number, length, diameter, orientation, contour, cross-linking of collagen across the biopsy and the Q fibrosis provides a quantitative, continuous scale of collagen fiber parameters for the phenotypic assessment. Again, similar work has been done in the clinical trial space here. Biosalvia uses a little bit of a different approach. Again, uses a standard stain, but has a, uses sort of a deterministic approach to sort of predetermine rules or statements that then allow the assessment in a morphologic fashion. So, but different, no training is required with this. It just follows the rules, again, and with output you get a CPA perisinusoidal, perivascular, and other percentages on where fibrosis is and characteristics of the fiber length, width, et cetera. So, improved quantification of fibrosis is, I would argue, critical in this particular space because we are unable, as I showed you, with viral, if in viral hepatitis you can't show a one-stage improvement in a year, it's gonna be very hard for a disease that's multifactorial with drugs that aren't able to completely eliminate the driving force of the disease. So, here's one way in which, in this case, this is second harmonic generation imaging. This is a tool, again, that I described earlier. This can be used, as could others, to improve concordance. And in this particular study, you can see in the unassisted read versus the assisted read, there was an improvement from intra, and intra-observer agreement using this assistive technology. And really, all of these are assistive technologies, at least at this point. This is an example from Biosalvia's platform. And you can see here that there is good correlation with different types or locations of collagen here, perivascular and septal collagen, with good correlation with categorical staging. In this particular study, again, this uses PathAI. Again, a different approach using trichrome-based stains. This, in particular, this particular study looked at, used the data from Stellar 3, which was a solancertib in stage three patients, and Stellar 4, which is solancertib in F4 patients, cirrhotic patients, and ATLAS. And what this did is it, from there, derived a large training, validation, and test slide set to work with this algorithm, and then come up with multiple, actually, things that were helpful. This model was trained on CRN, which is what we use for MASH, and ISHOC as well, and then end-to-end encryption, the whole slide images of the trichrome stains. There's good correlation with pathologic scores, and you can see on the bottom right, the CRN score, and then the patients, and on the right, the AI-generated score, and you can see good correlation. So, here we have another method. Again, this is using FibroNest, and here you can see that phenotypic differentiation. So, on the bottom, advanced fibrosis, on the top, mild fibrosis, and you can see there's a detailed output on collagen fiber location, and then a color overlay to highlight differences, and so that's an example here. And then, to the right, towards the bottom, you can see a fibrosis severity score that is generated over each column, is a patient going through different phenotypic stages, and so you can develop sort of a quantitative score for this. So, this is a good example, I think, of something that's fairly recent that we're starting to think about. So, aldofermin is an FGF19 analog, and these data come from a 24-week biopsy study. In this trial, the primary input was a drop in steatosis, and you can see very handily, there was a very significant drop in steatosis, greater than 30% is thought to be meaningful from a histologic standpoint. However, from a fibrosis standpoint, even though there were more people who met the fibrosis endpoint, this was, of course, not powered for that, but there wasn't really a difference. And when this was looked at using, in this case, a histoindex, you can see in the top bar graph area, or the squiggly graph, I don't know how I would define that, but anyway, what that is, is it's identifying individual traits, and you can see that in the top, there were only seven that actually were significantly different in the treatment versus the placebo group, and then what they did is because there was such dramatic loss in fat, they did a correction factor for steatosis, and that actually is something that we're starting to think about, not just with steatosis, but with volume in general, which changes in response to therapeutics. So I think this is an interesting and different approach. So what AI can also do is identify novel paradigms, and this is an example of one. So if you talk to a pathologist, they'll definitely tell you, yeah, this kind of looks like a regressing lesion or a progressing lesion, but I think where we were missing, what we were missing was the ability to sort of more granularly quantify and describe these changes. And so this is an example here. If you look towards the bottom, these are two patients from Tropifexor trial. The top one is in the F2 patient, the bottom one is in an F3 patient. You can see portal and septal collagens highlighted there in yellow and blue respectively, and you can see that even though there's definitely a reduction in fibrosis, the stage did not change, and so these patients would be non-responders. I think that if you look at the changes in collagen content, you can see that there's definitely a signal. So what Arun Sanyal's group then did is that they actually did this in a murine model using their way to simulate regression, or progression regression was to use a high-fat, high-sugar diet compared to chow, and then they withdrew it after a period of time to sort of create those phenotypes, and what he was able to, or they were able to show is that there's definitely a signature of fibrosis progression or regression that sort of recapitulates what we see in the clinical space. This, I think, is an interesting finding as well. So this is using fibrinase, which remember as the phenotypic approach, and many pathologists will tell you, yes, there are differences between mash and alcohol-related liver disease. Sometimes they're not clearly articulated, and I think many people believe that they're really, it's hard to distinguish them. So what, in this particular project, the computer, the algorithm was told, this is mash or nash, and this is ash, and then output was the following, and so interestingly, they found that they were able to discriminate ALD from mash with a good, with an AROC of 0.96, and a sensitivity of, I think it's 86, I can't really read because I don't have my glasses, and with this cut point of 4.36. So this is obviously very, very preliminary, but it's interesting that you might be able to use AI to teach you things that you are otherwise not able to see, even pathologists. So where can AI help? More accurate categorization of fibrosis burden, obviously, ultimately breaking away from categorical fibrosis stages, more sensitive capture of early treatment responses, this is critical, this has killed a lot of therapeutics in this space. To identify and model patterns of progression and regression, again, also very important. AI-based methods should be routinely incorporated into clinical trials, and with histologic endpoints, and I really believe that this is going to be one of the ways that we're finally going to be able to achieve something. We have, actually, histoindex has been approved to have their platform as a primary endpoint in a phase two trial. And then lastly, for widespread implementation, what? Okay, good. Approaches in mass commercial development and approval by regulatory authorities is necessary, but it does look promising. And with that, I will close, and thank you for your attention, and acknowledge those that sent me figures. Thank you. Thank you very much, it's exactly on time. It's my pleasure now to introduce Professor Jakob Kauter. Hopefully, I pronounced that appropriately, who will be speaking to us, and the topic is an introduction to AI for histology and its application for the diagnosis of liver disease and liver fibrosis. And while Professor Kauter's getting ready, I just would like to point out, he's from the University of Dresden, but what's notable is he's a professor of digital medicine, which is certainly not a title that was around when I was in training, so welcome. Thank you very much. Hi, everybody, thank you for the kind invitation. Yeah, so actually, I'm a medical doctor. I'm practicing in GI medical oncology, did part of my training in hepatology, but here, today, I'm going to speak a little bit more about the technical details, because the technical background, we have just heard about basically all the applications in the field where we can use AI to analyze histopathology, slides for non-malignant liver diseases. So let's look at some of the underlying technology and also technological developments that could tell us what might be possible in a couple of years from now. These are my disclosures again. So let me start this session maybe by showing this iconic AI image. So this is an avocado chair, and this has become one of the most prominent images in the AI space. This was one of the first images that was generated by the model Dolly in 2021. This was an AI model released by OpenAI, so the company behind Chet-GPT, okay? And they trained a model on text and image pairs from the internet, on all of Wikipedia, Instagram, and basically wherever you can get images and text. And the model learned to understand all kinds of different concepts without being explicitly trained for these concepts. And in order to demonstrate that the model had understood these concepts, they came up with the avocado chair as an example, because there is no avocado chair on the internet, and this showed basically the capability of the model to piece together different things. But looking back at this from today's perspective, this actually looks pretty bad and really not so nice, right? So in 2022, the mid-journey AI model, this was the avocado chair output that we got there, and now we have even better models. So this is a liver researcher sitting in an avocado chair, which is the output of the Dolly 3 model, the latest model. So it's really, I hope this gives you an intuition just how fast the field is developing, right? And things that were completely unimaginable two, three years ago are now completely normal. And I think our job as physicians is to try to keep up with these developments and understand what is possible and how we can in the end use this for scientific discovery and also better tools for clinical routine. So the question is how can we apply this technology in hepatology and maybe let's take one step back from the avocado chair back to like simple classification systems. And we've basically heard it all in the previous talk. And yeah, it's pretty difficult for human pathologists to consistently grade what is going on in liver biopsies. And that's a huge problem, especially in clinical trials. And nowadays there's a bunch of tools available that can automate this and give you a very consistent readout. These tools use artificial neural networks, deep learning. And for example, this is one of the path AI examples. But actually there's several tools which were shown in the previous presentation by different companies who have brought these things to the market, which can be used for research use and in clinical trials. There's this very nice review in Journal of Hepatology, which was published recently, which points out all of this. So the technology for these like simple image classification tasks has arrived, is broadly available. You can use it. It's a commodity, right? It's just out there. We just have to use it. When it comes to the technical, the underlying technology, basically there's two different philosophies behind this. One is strongly supervised and the other weakly supervised deep learning. So strongly supervised means you basically obtain thousands or hundreds of thousands of pathologists expert annotations, right? So they circle cells and then, or other structures, and then you train a deep learning system to recapitulate what the pathologist has done. Okay. The alternative to that is weakly supervised deep learning, where you just take, for example, one biopsy and you give it just one information, a global label for the patient, for example, the fibrosis grade or even something like survival, right? Was the patient alive after 10 years? Yes or no. You can use all of these things as global labels and then train your deep learning systems on it. And this is what we call weakly supervised deep learning. And while strongly supervised deep learning is like very well established, I think there's still a lot of potential for weakly supervised deep learning to even exceed what pathologists can do right now. And let me give you some examples. I'm sorry, they're coming from the cancer field, so HCC, so it's maybe a slightly different application, but the general method, I hope, is clear. So this is a publication by my friend, Giulio Calderaro from Crete, to which we also contributed, where we showed together that RNA expression signatures in HCC can be predicted quite well from H&E slides of HCC with deep learning, right? So instead of running the transcriptomic analysis, deep learning can just approximate this information from H&E slides. So this prediction, the predictive power of this is not perfect, but it's often good enough, for example, to predict immunotherapy response. And in this case, the signatures that we looked at were gene expression signatures that are associated with immunotherapy response. And in a very recent follow-up study, which is not in this presentation, but which was published a couple of days ago in Lancet Oncology, Giulio's group showed that you can actually predict the response to atelzobev in HCC with deep learning from H&E slides using weakly supervised deep learning, okay? So these things can be used as biomarkers. Maybe even going back a little bit more to other types of GI cancer, so in colorectal cancer, my own group did a lot of work where we basically predicted molecular alterations based on H&E slides, such as microsatellite instability in colorectal cancer, which qualifies patients for immunotherapy in metastatic setting, but also increasingly in neoadjuvant setting. And basically, we can predict this to a certain extent, which can be useful in clinical application from pathology slides. And this is now an approved clinical device, right? So you can extract molecular biomarkers for patient, for treatment selection with deep learning from histopathology slides. One of the most challenging things in this field is that the technology is advancing so fast, so every single clinically approved device that uses deep learning is completely outdated now, because the approval process takes two or three years, right? And in these two or three years, the techniques have just evolved so much. So in this recent study, what we showed that by using two very new techniques, self-supervised learning and transformer neural networks, you basically get a much better predictive performance for predictive biomarkers. So it makes sense to stay up to date and to use the latest technologies and to incorporate them into clinical biomarkers. Maybe my last example here is another study, which we recently did in liver cancer. We trained a deep learning system to differentiate HCC and cholangiocarcinoma, and then also applied it to combined HCC and cholangiocarcinoma tumors and showed that these predictions that we obtain in the end make sense and could allow us to recategorize these patients basically, which we validated clinically and also with spatial transcriptomics and some other like orthogonal methods. The field is moving towards multimodal deep learning methods. That means we are not just using one type of image anymore, but we can use agent E plus genomic information, agent E plus radiology, agent E plus immunohistochemistry and build one single deep learning model that integrates all of this and gives you one single prediction. And I think this is very exciting because in, again, sorry for the non-liver use case, but again, in colorectal cancer, we have seen that we can get better response prediction and outcome prediction by using such a multimodal approach. So what we should do altogether is to try to bring these technologies to relevant clinical questions in the hepatology field, right? So basically just please be aware that all these data that are obtained are a useful resource and even multimodal data, even text plus image data can be used to develop hopefully the next generation of AI tools that we can use to have better clinical endpoints in these studies. So last five minutes, maybe I will give you some very recent updates because the field is moving and the field is moving so fast that yeah, sometimes technological developments are much faster than we are usually used to in academic routines. So here's an example of the latest GPT-4 model. So CHET-GPT with the GPT-4 vision model now can interpret images, so this morning I just took a random H&E image of a steatotic liver and I uploaded it and I asked CHET-GPT what is this, right? And it gives you a perfect description, right? It can tell you that this is steatosis, that fat is in the cells and this is something for which just a few months ago we would have had to train a very like dedicated model using maybe many, many annotations and now yeah, all of this is obsolete. So I ask myself maybe all the work that we did in the last couple of years, maybe it's all obsolete because GPT-5 will be able to solve this out of the box. Who knows? The fact is that technology is advancing and we cannot ignore it and we have to then use this technology, evaluate it rigorously and incorporate it of course in clinical decision making in the end. This is not for clinical use at the moment, right? It's not regulatory. Do not use CHET-GPT for clinical decision making. It's not a medical device. It's at your own risk. It's just to play around basically but yeah, of course it's pretty cool that this technology is out there now. I will give you some more GPT examples. So yeah, GPT is a large, CHET-GPT basically has this GPT-4 model which is a large transformer model which can process images and text very well and this morning there was already an AI session at this conference which was overflowing and the overflow area was also overflowing so maybe not everybody saw these examples. Let me just give you some of these examples what is possible maybe beyond the image processing with these tools now. There is something was released just last week which is called GPT Builder where you basically can take text information such as just PDFs of guidelines and upload them and provide these as additional knowledge to the model, okay? So here just in five minutes basically I uploaded the ASLD guideline, the easel guideline for non-alcoholic fatty liver disease and the latest Delphi consensus statement on the new nomenclature, right? And so I provided these pieces of information to GPT and then I have like a custom CHET-GPT that has all this knowledge and I can ask it a question like give me the new definition or how should I treat a certain patient and basically you can and this yeah took me just a couple of minutes to build it's now online it's called Liver Care Guide and you can use it if you have the paid and only if you have the paid CHET-GPT subscription for $20 a month. I have no conflict of interest here, right? But you can basically use it and talk to the guidelines and ask them questions. So this is really, really cool. This was unimaginable a year ago and this is definitely going to change the way we work and if you have children who are going to school and they're already using this all the time, right? So if you need help ask your children and it's not perfect. So also this morning I asked CHET-GPT to generate some anatomic images and there are some misunderstandings here. So yeah I mean this is really what I got out this morning. It's there are it shows us the limitations and we should not be replaced like immediately which is nice to know. So my time is almost up and I just want to point out this is a photo of our research team in Dresden in Germany. It's about half of the people have a computer science background, half of the people are medical doctors who learn programming and are building these methods themselves. And so I think the young generation of physicians really some of them want to learn this and want to become experts in this field. So please everybody if you have young people like this in your teams give them the opportunity to learn these tools and then also the tools to rigorously evaluate these things in clinical routine and in according to the standards we need for basically any new tool in order to apply it in the clinics. So thank you very much for your attention. Thank you so much Professor Kader. So very exciting and perhaps a little scary. So it's my great pleasure to invite Professor Rohit Lumba to come and speak to us next and Professor Lumba will be speaking to us about applying AI for the diagnosis of liver fibrosis imaging elastography and multimodal approaches. Welcome. Thank you much Heather and Robert for this amazing symposium and thank you to all of you for being here. I think we're all really excited about the role of AI and how we can really change the care of our patients. I'll discuss briefly on imaging and potentially on multimodality assessment and where things may be going and what we can actually do today clinically. These are my disclosures and this is going to be the brief outline of my talk. Can we use AI to improve the diagnosis of a disease? I think this is something that we will be using in our clinical practice and this is one example which is currently in clinical practice and so this is where people looked at deep learning algorithms to develop a tool where basically you could assess the risk for diabetic retinopathy or even actually all or many retinal diseases and but this exam one is able to but the number of feeds that are going in about the order of hundred thousand or more images and I think this is something totally plausible and absolutely going to happen in liver disease and particularly in HCC detection and also lesion detection. It's already happening. This has happened in lung cancer in mammography but this is one example where this is a FDA approved tool now available for diagnosis and you can see you can optimize it by feeding the number of images and giving particular diagnosis and now you don't need a physician or ophthalmologist to do an exam because this test provides you a much better diagnostic accuracy compared to a exam by a physician. What about genetic diseases and this is an area I think we're going to see a lot of evolution because we have not done much in genetics and liver disease. We've got hemochromatosis where we do genetic testing then we have you know for phenotype we have alpha-1 antitrypsin but potentially we may be moving on to doing whole exome sequencing in select group of patients. Potential use could be in lean naphroletes, use could be in patients with cirrhosis that we are calling cryptogenic right now. It's possible there are rare variants that are inducing those diseases and once we identify those rare variants we might be looking at new treatments that are targeted to those rare diseases and this is just one example that if you had a series of undiagnosed individual with unknown problem including liver disease this could really lead to a diagnosis with just advanced genetic testing to up to 35 percent. There was a paper by Sylvia looking at a small group of patients you could potentially diagnose 19 percent of patients with rare problems that may be coming. I'll give you one example of that in hepatology that can happen tomorrow. This is the idea that if you have patients who have a lean naphrolete so somebody with a body mass index that's below 25 but they still have rip-roaring steatohepatitis stage 3 fibrosis why is that? What is driving their disease? It's possible that there's a genetic cause and we may want to you know consider that. Also it's possible that they may have other environmental exposures. It could be alcohol, it could be environmental toxins, but it's also possible that there's a clear genetic predisposition. So that might be a group of patients where we start identifying that the risk of having bridging fibrosis or cirrhosis in this population is so low that we might consider routinely doing whole exome sequencing and trying to identify you know rare variants. So you could really sub-classify patients with lean naphrolete and why are we picking lean naphrolete and not doing everybody who has fatty liver disease? The idea would be you want to enrich the cohort so you're not increasing the cost of care for individuals and that would be the idea. Moving on to common problems and how AI could be utilized to solve common problems. This is taking real world data from Optum, also NASH CRN data and you know developing a model from a highly phenotype cohort and taking that model into real world and utilizing that to identify cases who have a particular problem. In this case non-alcoholic steatohepatitis. So this is just one example but I think over time we may be able to improve our diagnostic abilities. These things will become important as we have new therapies come along. Just imagine a future that next year we have a new therapy for NASH and you are managing 100,000 patients in your health population or a system that you are in. How are you going to identify patients who need to be treated? It's not going to be by an individual physician or a hepatologist going through the records or relying on a primary care physician to identify you know which patients need to be referred. Potentially we could develop systems that could be utilized and applied in the entire population and we can even risk stratify patients that bring in the highest risk patient in the first year, less risk patients in the following years. This is one example is a proof of concept study where you can really look at clinical features that are associated with steatosis, cirrhosis, features of portal hypertension, and use an AI-based system to really diagnose whether this particular lesion is HCC or it's an adenoma or a hemangioma. I think that's very easy because when we are sitting in our liver cancer group meetings we have specific features that identify that this lesion is a liver cancer and we can really feed those with hundreds and thousands of images and classify those models and also AI may be easily able to differentiate whether this is you know a HCC versus an adenoma versus a perfusion abnormality or regenerating nodule and this is already happening in the field. What about assessment of liver fat? Again this is one example where we use a deep learning method to take raw ultrasound waves and convert them into a fat signal. With this idea you could really quantify liver fat similar to what you do on MRI, proton density fat fraction, and you could get to a diagnostic accuracy of about 0.98. So you can now see that this is a routine ultrasound image that you know you may be doing in your patient, but it gives you raw data that we are not using. Currently, we take the image and actually using that to answer a clinical question, mild, moderate, severe steatosis. But if we see and get the entire raw data, you may be able to develop models that can automatically quantify a fact. So is this just fiction or this could happen? It's actually happening. So for example, Siemens ultrasound now has a UDFF function where actually our algorithms that are being modified are being utilized to quantify liver fat. So it's already happening. This is another example where using multiple features on a gadolinium enhanced MRI, as well as PDFF, so combining fat with more contrast enhanced features of MRI, you could potentially diagnose NASH and differentiate. Of course, these diagnostic accuracies are not where we would want to be, but the idea is these are small studies. If we could provide more enhanced data and more MR-based biomarkers, we may be able to improve these. And I'll give you some examples. AI applications in image processing are already happening. If you go to your radiologist, ask, can you do this automatically? I think that's exactly how it is happening. And across all our trials, the PDFF assessment is automated, and that's using AI to automate how much fat is there in the liver. Similarly, we're using automated assessment for two-dimensional MR elastography, as well as on certain equipments on shear wave elastography and ultrasound. And these are some of the AI-based innovations that are already in clinical practice. This is one example where you could have a CT scan or MRI, and you could automatically calculate the volume of the liver. Now, if you were manually doing it and you didn't have a code, it could take about one to one and a half hours for one volumetric assessment with an analyst. So you can really replace that, and why would that be important? If a patient is going for a resection of the liver, of a segment, or for HCC in a non-sterotic patient, it would be really important to understand how much volume would be conserved. So this is clinically useful information and being applied in many centers already. This is another paper that I looked at where they were able to utilize 2D MR cholangiopancreatography for diagnosis for PSC. And again, I think all of these tools will improve our precision for diagnosis, and also reduce variability in diagnosis. And potentially, we could assess treatment response as well. This is one example of a study where they looked at patients with advanced fibrosis and cirrhosis who had hepatitis C and were treated. And then they looked at another set of patients with cirrhosis and looked at features that could predict whether somebody has cirrhosis now or they're regressing. Some of the features they looked at was automatic spleen volume, liver segmental volume distribution. This is the idea that as liver disease progresses, the right lobe may actually shrink a little bit, and the left lobe may expand, which we know as we palpate a patient with cirrhosis. And so these potentially could be utilized and automatically calculated using an automated AI-based algorithm. This is again, potentially applicable in clinical practice as we scale it up. This is a paper that Winston Dunn is working along with collaborators, just looking at different clinical parameters, as well as liver stiffness-based parameters, to come up with an idea to differentiate stage three, stage four fibrosis, versus earlier degree of fibrosis. I think he'll be presenting this in this meeting. So this is one idea that this is a patient we followed over years to develop cirrhosis over time, and you can monitor their liver stiffness. This is on 2D-MRE, and this is a 2D-MRE, this is a 3D-MRE, and this is PDFF, where liver fat is going down over time. Now, currently we are assessing these as one modality at a time. And then this is a patient with cirrhosis, Werner-Wendt sleeve gastrectomy. You can see this patient, our liver stiffness close to six kilopascal. Within six months to a year, dropped down to 4.5 kilopascal. Patient still has cirrhosis, but this patient's risk for decompensation has significantly gone down. And this is now using AMRA-based modality. This is a company based in Sweden where you can look at whole body fibrosis fat quantification, visceral, as well as subcutaneous, and muscle volume as well. What if we combine all these features and with just a press of button, you get all of these estimates in one go automatically? You would think that that would be very helpful. That is already available today where we would be able to apply those and longitudinally monitor our patients. In this meeting over the last two days, I've heard a lot of people worrying about sarcopenia, frailty. This would be easy way for us to monitor our patients who are on therapies such as GLP-1 analogs. And we think a patient with cirrhosis exposed to it, are they losing muscle? How much would that be? I think this could be quantified. Other ways of looking into the liver, not just into stiffness, which we will look at further, is an idea that you can monitor treatment response using not only liver fat quantification, which is now routinely automated on AI-based algorithms, but also look at corrected T1. Another property, a physical property on MR that can tell us if you have improvements in liver disease severity over time. We have a study that's ongoing, it's called Goldmine, that is run across multiple centers in the world where we are doing baseline MR elastography, 2D, 3D, as well as MRI, PDFF, and saving these images for developing multiple biomarkers in future that could be automatically generated and potentially be predictors of future risk of disease. And patients are getting repeat MRIs every two years. These are some of the sites. And these are our outcomes, where we're looking at progression to cirrhosis as well as hepatic decompensation or liver-related morbidity and mortality. Idea would be that we would have collected serum plasma urine stool DNA in these patients at baseline and every two years. We have advanced imaging of their liver and that will be available along with cross-sectional abdominal imaging. And that data would be then linked with their genotype, phenotype, and metabolome. And then we will be able to potentially predict who is going to improve, who is going to develop a CVD risk versus a liver disease decompensation. This is already happening in our study. So patients at each of these sites get a hepatogram where we get a MRI stiffness estimate. This particular patient had 2.25 kilopascal. And these are automatically generated and a PDFF of 9.6%. So about 10% fat and liver stiffness is close to normal. So this patient has fatty liver. And this could be generated automatically without a manual evaluation. Why is that helpful? This will eliminate the need for, and you can scale it up and have results available to our patients faster. This is the idea that what if we could actually make these assessment within one minute? And this would really require better technology, but I think that is happening. Currently it takes about 20 minutes to generate a PDFF and MRE image. What about multi-parametric MRE assessment? There are other tools. Stiffness tells us something associated with fibrosis. But what if you have a patient who has improvement in inflammation and that too over short term? Could we use another MR-based biomarker to assess? This is a patient that underwent bariatric surgery. And you can see that there's improvement in another elastography parameter called loss modulus and a different frequency, 30 hertz. So doing exams at multiple frequency gives us a different property of the liver that could be utilized to assess inflammation. And the idea would be that these changes would not be addressed in the setting of fibrosis. And so these could be a test using loss modulus or another parameter called damping ratio. So this is what would be a future update on combining this PDFF as well as MRE-based platform, both on 2D as well as 3D. This is an idea about predicting decompensation just on liver stiffness on MR elastography, but you could potentially combine multiple MR-based biomarkers in one feed and predict future risk. This is a review that we wrote a year ago, combining multiple modalities and predicting patient's risk of current disease now and future prognosis as well as treatment response. And I think this would occur as the previous speaker just talked about. These are some of the key construct that are important as we are developing these AI-based tools. And I think many of these were mentioned in the previous talks, so I'll skip to the next one. I think it's really important to have inclusive data. If the feed itself is related to a particular set of patients that do not expand or do not allow for the entire spectrum of disease and patients to be studied, then the models would not be perfect. And we have to be mindful of bias. So the models developed in the United States may not work in China, may not work in other countries. And it's really important that we are able to develop models that are able to be validated in multiple populations. So in summary, AI-based application is already here, both in research and in precision medicine. It will have a role in undiagnosed liver disease assessment and AI-based methods routinely utilized in clinical care and will be in coming years in radiology and omics-based diagnostics. Automated reads are becoming standard of care in advanced imaging for liver fat and fibrosis quantification. Thank you. Thank you, Dr. Blumberg. Our next speaker, we have two more to go, and we're going to round this out. Maria Zetnik is local from Harvard, and she's going to discuss AI for the therapy of liver disease and liver fibrosis. Can you see your phone? Yes. It's a double-clock on you. Can you click Start? It'll take a second for this to catch up. OK. OK. Great. Thank you. It's OK. Yeah. You can actually go ahead. So you're down here, and it'll kick in. OK. Great. Good afternoon, everyone. Very excited to be here with you today. This is my first time at the liver meeting, as my group is focusing mainly on fundamental innovation of AI and how AI can help augment and accelerate biomedical research. So for centuries, the method of discovery, the fundamental practice of science that scientists use to understand our world logically and systematically, has remained largely the same. And now AI is changing that. It's been increasingly incorporated into experimental workflows in order to help with processing and acquisition of biomedical data sets of different modalities, leveraging those data sets in order to generate hypotheses that can then steer and guide downstream experimentation and simulation. As has been clear from the previous talks, AI for biomedical research is already here. In a recent poll we have run with Nature, we asked 1,600 researchers about their thoughts on the rising use of AI in science. And 27% of them responded that they think AI will be essential for their field of research in the next decade. Another 47% said that they think it will be very important for their field of research. AI is impacting biomedical research in two fundamental ways. First, it is changing the way science is being done, because we are increasingly relying on CGPT-style models to draft documents, review certain documents, help with ideation, generation of novel ideas. And second, it's being used to augment research, providing insights that traditionally could not be obtained using existing scientific methods. There are numerous examples of that, going from folding of proteins to providing diagnosis for rare genetic diseases with no label data, as well as using generative AI to optimize and design molecules. So the vision of our group is to really lay the foundations of AI that can help us enhance the understanding of medicine and design of safe and effective new therapies, eventually enabling AI to innovate and learn on its own. And we are approaching this problem by tackling four big questions, which relates to how AI can be used to design, refine, and generate novel therapeutic candidates. This primarily involves tackling and developing multi-modal generative models that can refine molecular structures. Once those are identified, we can predict and optimize protein and gene targets and see how they bind effectively to novel therapeutic candidates, find relevant disease, patient, and cell type context in which that therapeutic effect would be maximized, and finally, match therapies to patient benefits. And in this very brief talk, I want to share with you two vignettes that will touch upon two different areas of the cycle. So the first one is that of matching therapies to patient benefits. And here I want to outline a model called TXGNN, which stands for Therapeutics Graph Neural Network, which is a foundation-style model. Foundation-style models are models that are profoundly changing deep learning paradigm. Whereas in the past, we were designing and training a separate deep learning model for each task, for each disease and phenotype of interest, we now can train a single large model that has capability of handling many diseases at the same time. TXGNN is trained on a multi-modal knowledge graph that captures information across 17,000 disease phenotypes. And once trained, it can process a variety of therapeutic tasks and predict candidate indications and contraindications in a zero-shot manner. The way you can think of it is, from the user perspective, it's a model that takes as input a description of a therapeutic candidate, which can be not already approved drug or investigational compound currently in development, and a disease phenotype. And it will produce a score representing the likelihood of that drug being indicated or contraindicated together with an explanation that is human interpretable compact signature. This model is grounded in prior biomedical knowledge, which means that it is not suffering from the problem of hallucinations or making things up that some of the other models suffer from. Let me share some of the results of TXGNN. We found that TXGNN can find therapeutic candidates for many diseases for which currently there are no FDA-approved therapies. So look at scenario B, which corresponds to this orange or yellow bars on the right part of the plot, showing improved performance of TXGNN over state-of-the-art AI models in nominating drugs that could be repurposed for diseases that currently have no FDA-approved treatments. This is especially important for rare diseases of which there are over 7,000 of them, yet only 5% of rare diseases have any FDA-approved label. After extensive benchmarking of the model, we asked the model to make novel predictions, predictions that would connect drugs to diseases that have not yet been explored or tested by scientific community. And we did numerous evaluations of those novel predictions. First, we confirmed that if we take novel predictions, novel drug disease indications, candidate indications the model makes that were not present in the data the model was trained on, taking those predictions and comparing them against off-label prescription decisions made by clinicians in a large health care system showed that those novel predictions tend to be enriched at the level that is comparable to enrichment of FDA-approved indications, suggesting that there might be real potential in those novel predictions the model made. Second, we used what is known as temporal split, which also means that we train the model on all the data up to a certain point in time, say 2021, then ask the model to make predictions, wait for a year, see what are novel FDA approval in that year, and then ask, well, how many of those approvals made in the last year would the model predict correctly? And results were very promising. And third, what we have done is to teamed up with experimentalists and ask them to screen drugs in the order of how promising they are for a novel emerging pathogen predicted by our model versus what would be non-screening of compounds not guided by AI. And that led to a one order of magnitude improvement in the yield of compounds that inhibited viral growth relative to what we would expect if screening would not be done according to AI model. Of course, an important question that becomes if we have such a model and we can develop, implement, test it as AI scientists, how can we push it forward towards more responsible implementation so that our end users can use that model in their own research? So to work on that front, we adopted a clinician-centered design and designed an AI interface that exposes TXGNN predictions plus explanations of those predictions. And we evaluated the utility of TXGNN predictions by performing a usability study, comparing the use of the model and the efficiency of the work done by scientists when using the model versus not using the model at all. Across several metrics, including scientific consensus, user confidence, and agreement where scientists can use explanations provided by the model as a means to decide whether a prediction is trustworthy, relevant, potentially something of interest to follow up or not. So results showed that scientists, when using this tool, were able to be over 46% more accurate and 49% more confident than not using the tool and explanations that they provided. It also helped them decide whether a particular therapeutic hypothesis is relevant or not much faster than what would be possible have they not had the abilities to see explanations. In the last few minutes, I want to kind of change the gears a bit and go from this matching of therapeutic effects for approved FDA drugs to patients and kind of change gear from that to molecular biology setting, where I think there are also tremendous amounts of opportunities that new AI models create. And so in a large initiative that we are currently working on in my group in collaboration with Chan Zuckerberg initiative, we are very interested in potentially finding ways to build multimodal generative virtual cell simulators that are enabled by models trained on tens of millions of single cell profiles across healthy and diseased states and across different data set generated in single cell genomics over the last decade in order to enable AI scientists ask questions about the potential effect of certain perturbations before those perturbations are actually introduced in a cell. Let me illustrate that with an anecdote of what that actually entails and what's the critical capability that needs to be developed here. So let's forget about biology for a moment. Let's think about the word apple. So apple is a polysemic word. What does that mean? That means we can grow an apple or we can buy an apple. The meaning of the word apple here in these two examples is really resolved via sentence context. You can grow an apple, and it means that apple is a tree, or buy an apple, and apple would be referred to a company or stock. So in similar ways, one can use large single cell data sets to now resolve the function or therapeutic effects of genes. H2FX is pleiotropic gene. The role and the function that the gene play and its products play are resolved via context. But that context is not now the context given by sentence or paragraph or a piece of text, but it's given by several contexts in which that gene and its products are expressed and active. So imagine how great would it be if we would have very precise AI models that would be able to dynamically adjust their outputs to different biological contexts, where the model would not be making the same prediction for the same gene no matter what, but it would vary its prediction depending on the context in which that AI model operates. And so that's exactly what our model called PNaCl can do. PNaCl is a contextual graph neural network. What does that mean? It's a graph neural network because it's a type of deep learning model that can operate not only on grid-based tabular data, such as images or text, but it can operate on any interconnected relational graph network structure data. And it's contextual because it can dynamically adjust its outputs based on context in which it operates. PNaCl is trained on a large single cell transcriptomic data tabular sapiens, which is the reference human cell, single cell atlas, coupled with the human protein-protein interaction network that is specialized across 177 cell types, which we use to represent this biological context. And so what that means is that the model can provide representations or embeddings of predictions for individual proteins, not only in a global sense, but in a cell-type specific manner. This creates many new opportunities for multimodal deep learning, for transfer learning across context, for contextualized predictions. Let me just mention one of these three opportunities where we were interested in seeing whether the model is able to start nominating potential therapeutic targets and study the genomic effects of drugs in a cell-type specific manner. And so the way we approach that is by taking the large PNaCl model, that is pre-trained AI model, and then fine-tune it towards this downstream prediction task, where the model was asked to try to distinguish between targets that were investigated for the disease or have advanced to different stages of clinical development for a disease, versus those that did not. And so I have here an example for RNA-inflammatory disease. I would be very happy to discuss opportunities for liver diseases, where we showed that the model was able to pinpoint certain specific cell types that it found most informative and predictive of therapeutic targets that are currently being investigated for novel therapies in RNA-inflammatory disease. And this was done in collaborations with our partners from Mass General and Brigham and Women. And it can even do that kind of characterization at the individual gene level by asking for a particular gene for which prediction is made, what are the cell types that are most likely druggable for that gene if we want to target it. There are many other uses that PNaCl can perform. This is kind of the advantage of foundational-style models that have multiple capabilities beyond those that the model was explicitly trained on. Here on the left, I'm showing results how PNaCl can nominate therapeutic targets in cell type-specific manner much more accurately, relative to alternative methods based on attention-based deep learning models and based on various network science techniques. Here on the right, I'm showing that what's possible to do is to take these contextual embeddings and combine them with other established data modalities. So here, specifically, we looked at PD-L1 protein interactions and B7-1 and CTLA-4 protein interactions, which are critical interactions in immuno-oncology. And we wanted to understand whether taking structural embeddings of how these proteins look like and bind to each other and augmenting those structural embeddings that capture the shape of proteins with these contextual embeddings that capture the function of those proteins in immune cells can help us better distinguish between binders versus non-binders. And the answer to that was, yes, absolutely. What does that indicate? In order to really improve the models and make predictions that are much more precise, bringing contextual information and combining it with existing modalities can really help. OK, so let me wrap up. How does all this come together here? So I think what's incredibly important is to build these meeting points between, on one end, AI scientists who can develop novel fundamental algorithms based on transformers, self-supervised learning, zero-shot learning, and biomedical scientists who really know what are the bottlenecks and opportunities that are at the cusp of benefiting from ML. So we founded Therapeutic Commons, which is an open science global initiative that promotes the development of new algorithms and advancing therapeutic science. It comes with ready-to-use AI tasks, curated benchmarks across all stages of drug discovery and development, from early drug design to late drug repurposing and pharmacovigilance. Leaderboard, so we can see for a given task what is the best model currently available. Open source large language models so that you can directly use in your research and you can understand fully how they work. And those with no coding experience can interact with these models directly through interactive human AI feedback loops. There are many capabilities that the Commons have, from predictive AI to generative AI, where it can predict molecular properties and help with admin prediction early on, to generative modeling, where we can design protein pockets such that they maximize the binding affinity between ligands and protein structures in the same way as we have seen before for the avocado chair. But now think about replacing avocado chair with molecular structures, such as proteins and ligands. And so, of course, that would not be possible with a team of fantastic students that I'm able to work with. And all of our resources and datums and models are publicly available. Thank you very much. Thank you so much. So our final speaker is Sonia McParland from University of Toronto. She is going to talk to us about spatial transcriptomics for the understanding, diagnosis, and treatment of liver fibrosis. And I just want to remind the speakers that when Dr. McParland is done, if you guys can please come to the stage, and we'll have the panel discussion open for questions. It'll take a second for it to communicate. So it's already going here. So it should kick in in a second. Great. I know. It seems like it's not going to work. Awesome. OK, thank you so much. So I'm really happy to be here today and follow these wonderful speakers and really talk about how we're using spatial transcriptomics to understand liver fibrosis and multimodal approaches to layer HME onto transcriptomic data. So this kind of work follows our global effort, really, to map the liver in the healthy liver across a lifespan. And just before I get into some of the approaches that we're using, I really wanted to share some of the protocols that we're using. And so they are open source, and they're available online. And so any of the data that I talk about today, any of the approaches I talk about today, we're really committed to open science and sharing these approaches. And so please reach out if there's anything that we can give you to advance your understanding of liver fibrosis. So really, some of the interesting or some of the challenges that we've come up with when modeling or profiling the healthy liver was that, number one, fibroblast and mesenchymal cells are very hard to enzymatically release. Enzymes were very dissociation sensitive. And there were biases due to the approach used for tissue disruption. So really, this is where spatial transcriptomics with a histological context comes into play. And so what we found in healthy liver, and this was work that Tallulah talked about last year, Tallulah Andrews, was that if we use spatial transcriptomics, we can really take away some of the bias in profiling the function and the zonation of hepatocytes. And so this is data from the 10x genomics platform. And each of these, I'll just remind everyone that each of these spots are mini-bulk RNA seqs with 55 microns in diameter. So it's an unbiased approach to looking at individual cells within the repeating units of the liver. And so what we're able to see is that, in general, looking at the function of hepatocytes, really a spatial transcriptomics approach, as opposed to a single cell or single nucleus RNA sequencing approach, is the most appropriate way to profile a hepatocyte. And so that led us really to start our examination of single cell profiling and spatial profiling of liver disease. So the disease that we really started looking at was primary sclerosine cholangitis, which I don't need to remind this audience, is a very progressive immune-mediated cholestatic liver disease. And it has a lot of patchy pathology. So with this in mind, our workflow involves looking at intra-hepatic heterogeneity. And right now, we're using the pathologist to annotate these populations and the histology. So we don't use any automated approaches yet, but certainly that is an opportunity for collaboration. We use immunohistochemical characterization, as well as we take samples for single nucleus RNA sequencing, single cell RNA sequencing, as well as spatial transcriptomics. And the idea is that we can find and validate biological signatures and select early therapeutic targets. And so this is just an example. We take samples from explanted tissues. And I'll talk about also our early stage biopsies. And what we do is we take a sample for histology. We take a sample for single nuc. We dissociate cells for downstream validation once we've already done the transcriptomics. And then we take OCT-embedded tissue, as well as our FFPE tissue for spatial transcriptomics. And so in our efforts to look at primary sclerosine cholangitis, we first performed single cell RNA sequencing and single nucleus RNA sequencing, as well as spatial transcriptomics. And what I'm going to point out is our single cell RNA-seq map. And again, we had originally generated a healthy map using single cell RNA sequencing. So these are from neurologically deceased donors. We also compared this to PBC and two samples of patients from PBC and eight samples of patients from PSC. And what we saw was that while there were shared disease patterns between PSC and PBC, in particular in PSC, there were expanded myeloid populations that had fibrosis and immunosuppressive related signatures. So our big question was, as we want to target the lesions and the disease patterns in PSC, we really wanted to understand the spatial organization, in particular, of immune cells, including myeloid cells in the PSC liver. And it's because of the fact that myeloid cells, from a nanoparticle perspective, are easily targeted and they can be reprogrammed and they are plastic. So our first examination really involved using spatial transcriptomics, as I showed earlier, for the healthy liver. So these are mini-bulk RNA-seqs with 55 microns in context. And so what we wanted to do really was focus in on looking at the histologically apparent lesions and looking at transcripts that were associated with different parts of those lesions. And so when we performed our unbiased analysis, we saw that the clusters were really associated. As opposed to in healthy liver, they were associated with those repeating units in the periportal central venous axis. The cluster distribution in the disease was very focused or very centered around the PSC scars. So if we go back to our question and our hypothesis that we can use macrophages, potentially, to reprogram the tissue microenvironment in the liver, and we landmark our tissue with our central vein, as well as our lesions, and we look at markers such as MARCO, which is a marker of more tissue resident macrophages, we find that the MARCO transcripts are on the outside of the lesions. Whereas if we look at, again, landmarking our tissue and looking at a gene called lysozyme, which is more recently recruited macrophages, we find those recently recruited macrophages, which are red is highly expressed, black is excluded, they're highly expressed within the lesion. It suggests that there's potentially those more recently recruited macrophages are driving lesion expansion, whereas the MARCO positive macrophages are potentially part of the repair process. So when we associate or when we take multiple genes from tissue repair macrophages versus recently recruited macrophages and generate a gene list, again, we see that the more inflammatory or more recently recruited macrophages that have higher inflammatory potential are found within the lesion as opposed to outside the lesions. So one of the questions we had, and again, we wanted to use spatial transcriptomics for this, was to characterize PSC lesions at different stages of progression. So our big question is whether some of these defects that we're seeing in late disease and some of this distribution of myeloid cells and immune cells in late disease, do we see them in early disease? Does this represent a target that's separate from just end-stage liver disease? And this is wonderful work we're doing in collaboration with Dr. Amanda Ricciuto at SickKids, as well as Dr. Kenan Hirschfeld, Amy Gilvary, Mark Cottrell, as well as Blaine Syed. And so what we're able to do in this approach is we take biopsies, and again, we mark off the regions of interest, including the relatively healthy central venous area and the areas of biliary proliferation. And what we do is we, again, look at particular genes and their expression. And again, this is just looking in an unbiased approach, but we're focusing in on the marcopositive macrophages. And what we see is those marcopositive macrophages that are more tissue repair macrophages are actually found in red around that central venous area, whereas they're relatively excluded from the areas of more biliary proliferation and inflammation. Then again, if we look at lysozyme, again, a gene that really marks these recently recruited macrophages, we find, again, that they're enriched in the lesions, and they're less present in the relatively healthy tissue. However, the concern, though, is that these high resolution, the spatial transcriptomics is really only capturing, say, 10 to 30 cells in each of those 55 micron spots. So it means that a cell-cell interaction may not be inferred from these mini-bulk RNA-seqs, and we really have to look with increasing resolution. And so what we're using is the Xenium platform from 10X Genomics, which gives us single-cell resolution and allows us to layer multimodal data, including H and E, on top of the transcriptomic data. So here you see H and E from a patient with primary sclerosine cholangitis at the time of explant. And what we do with our pathologists is we mark off our lesional area or area of fibrosis, and then we line up our H and E with the DAPI within that tissue. And then we look, again, here we look at 500 genes, as opposed to an unbiased approach. But what you can see here are genes in yellow. This is lysozyme again, and in blue, that's marco. So what you can see is, as we saw in the spatial transcriptomics using Visium, the lysozyme is found within the more lesional area, and then that marco is found in the more tissue or the more healthy tissue. And so what we're able to do with this information is also then go to the individual cell populations and look, as opposed to with many bulk RNA-seqs, we can look at what these T-cells are interacting within the tissue. So what we can see is, in green, we see these more recently recruited macrophages, and then we can look at T-cells within these lesions and see whether the macrophages are interacting with those T-cells. And so when we look, then, at these populations and then select genes associated with T-cells, including CD4 and CD8 T-cells, then we can look at the per-cell, the per-gene localization, and we can look at this with respect to the lesions. So what you can see is that these T-cells and macrophages are found in the lesional area compared to MARCO, which you'll see next is found more here in the relatively healthy tissue. And so then we can take that data and treat it like a single-cell data set and then annotate it based on our previous single-cell data set. So what we can do is then import these annotations into our Xenium browser, and then these annotations now have a liver context that was informed by our pilot single-cell RNA-seq. And so what we're able to do then is perform cell-cell interaction analysis. So what we can see, which is an improvement from these bulk RNA-seqs, is we can predict the cells that are interacting in these tissues by the ligands and receptors they're expressing. And we can also see the cells that are found around those cells within the tissue niche. And so what we're doing is we're looking also at the functional capacity of these populations that are predicted to be dysfunctional in the tissue by our spatial transcriptomics. So what we do is we take the populations, or we extract the populations from the healthy and diseased tissue. Here, this is healthy tissue, NDD tissue, as well as PBC and PSC tissue. And we can stimulate with lipopolysaccharide, and we actually see that macrophages that are found within these tissue lesions in PSC are less able to secrete TNF-alpha. And so in the last few minutes, I really want to just show how we can kind of close the loop and use single-cell and spatial transcriptomics to start to understand how to treat the diseased liver. And so what we're using is 3D human tissue culture. So this is the precision-cut liver slices. And this is where we take human and mouse tissue. We agarose embed, and we slice and culture these in the air-liquid interface. And what we're able to do is look at the viability of these populations using ATP, as well as histology. And we're able to use H&E to see that there's intact hepatocyte nuclei, but we're also able to then pull out the nuclei and look at these cells by single-nucleus RNA sequencing. And we can see whether there are cells that are proliferating, so expressing SNC4 as well as top 2A. And we can see that these populations are not only viable, but they're actually proliferating. And so as proof of concept, we treat these slices with interleukin-4. And again, we just want to show that this is a system we can perturb over the course of five days. And what we see is when we perturb the system with interleukin-4, we can promote proliferative pathways, and we promote the enrichment of tissue repair macrophages, as you can see by the enhanced expression of a tissue repair macrophage score. So it suggests that we can use unbiased approaches or spatial transgenomics to understand perturbations in these tissues, as long as they can take place within five days. And what we've also seen is that we're able to do this, and the data, the unpublished data I just showed you was for the mouse system, and it was all optimized in mouse. But what we can do is perform this on healthy human tissue, as well as tissue from patients with PSC at the time of explant. So it means we can perturb that system and potentially read that out by either xenium, so the multimodal approach, layering the spatial transcriptomics onto the HNE, and have our pathologists kind of grade the changes, both histologically as well as transcriptomically. So in summary, we are seeing through single-cell and spatial transcriptomics that we can really profile not just the cells that are making up the disease niche, but how they're interacting with each other on a single-cell level. And we're seeing that in liver inflammation driven by PSC. Both activated and dysfunctional myeloid cells are making up that interface between the lesions and the healthy tissue, and they represent a target for immunomodulation. And so I'd like to acknowledge our absolutely wonderful team, including the students, the postdoc, Dr. Tallulah Andrews, who's now a brilliant new PI in University of Western Ontario and was always looking for collaborators, our patient partners at PSC Partners Canada, and our surgeon colleagues. Thank you. Thank you, Sonia. Thank you so much. The speakers can quickly come to the stage. If you have questions, please line up at the microphones. Identify yourself prior to asking your question. All the speakers are here available to answer any questions. And I'll remind people before they leave, too, we're having our liver fibrosis SIG business meeting. So we hope that those of you that are on the SIG can join us for that. Oh, Dr. Costolari, you're first. Hello. So, Enes Costolari from Mayo Clinic. So I have a question that maybe can be asked to all of you, a very high level question, and naive. So we know that we can use AI now for pathology images to predict diagnosis, et cetera. Now the question is, is it possible to use AI to integrate all these single cell spatial omics, et cetera, datas, in order to create something like CHAT-GPT, for example, and to ask them what is the pathway or what molecules go wrong in a particular disease or in liver diseases? Yeah, very broad question, but any thoughts on the future about this? Sure. Maybe I can try and answer. So yes, definitely we are seeing AI systems that can integrate different types of data, genomic data, histology, radiology. But right now, we still need a couple of hundred patients to train on, if we train, build something ourselves. So this doesn't seem to be feasible for anything single cell or spatial omics, right? So that's why I'm not aware of any studies that have integrated this. But in principle, the capabilities are there. So my suggestion would be if you have substantial data sets of a couple of hundred patients where you have different types of data, even text and genomics and images, then yes, AI can integrate it. But for data sets which are much smaller, we are still very limited at the moment. If I can add to that, so I think in single cell genomics, the approach will be similar as we're seeing in CHAT-GPT model. So CHAT-GPT is also not trained on individual patient level data. It's just trained on large amounts of web scale data. So for single cell genomics, what we are expecting is these meta aggregators now appearing that connect dozens of single cell molecular atlases that were developed over the last few years in a unified, harmonized data representations. The largest one currently available is by CZI called Cell Exchange, which has 50 million cells for each cell high dimensional readout. So the number of tokens that are available through that resource is comparable to the size of the data set that was used to train GPT 3.5 Turbine models. What that means is that it's just a matter of time, and the cell simulators that I mentioned are of kind of projects of that nature that aim to develop large pre-trained single cell multimodal models. And those then would be fine tuned on small amounts of patient level data. That's how broadly we're thinking about the approach here. Thank you everyone for these great talks. I think these seem very interesting tools, but my question to you is how far are we from being able to predict matrix cross-linking? I think this is important for fibrosis resolution. So I mean, we could see the fibrosis in multiple different ways. We could understand or identify the composition of the matrix doing transcriptomics, matrixomics. But at the end of the day, how cross-linked that matrix is, is going to be critical for treating these patients. So can we determine the extent of cross-linking? I think this is an important point. I think the, from a clinical perspective, we are thinking of not even looking at the liver. We're looking at pathology by trying to predict non-invasively so we don't actually ever have to do a liver biopsy. And so I think that would be the approach that, you know, we're taking in terms of, but I think if you're interested in understanding pathogenetic mechanisms, then it will be important to understand that in the lab in terms of responses that you're seeing and how fibrosis regresses. So it'll be important. So I think from a patient viewpoint, I would, you know, go towards non-invasive modalities for predicting prognosis and also treatment response. But understanding how a particular fibrosis lesion regresses in the lab, I think for those modalities would be important. Electron microscopy would be something that could potentially give us clues as to what is happening over time. And serial sampling, that can definitely happen in the setting if you have good animal models for disease progression and regression, but that would be one way of doing it. We could even do, in the setting of a trial, it could be, you know, N of 9 or 10, where you could have fresh tissue with electron microscopy, and then you could generate images to really assess how this cross-linking resolution is happening. Robert Schwabe, great talks by everyone. So my question was already partially answered, but I'm going to ask it kind of again. And it goes a little bit to Sonia, but maybe everybody else. So a lot of labs don't have the capability like yours to analyze the single-cell data. So how long do we need to wait, and what is needed to have an AI-based analysis of single-cell data where you don't need the coding capability in the lab? Yeah, that's an excellent question. So I mean, a lot of the bioinformaticians are actually sharing their R Markdown files. There's a lot of sharing of the data, and then the students kind of modify it. So I think it's getting easier and easier for the students that are interested in doing that to kind of modify other code. So I think the more there's kind of open sharing, but also, yeah, having ChatGTP help write the code, I think there are people that are doing that, so that would probably speed it up as well. But I think sharing the code as openly as possible, which is what's happening now in the community, is going to speed that up too. Hi. So I was thinking about AI. It's generally trained off of what's publicly available, and if we think about any medical or scientific literature, the populations are generally not diverse. There are biases within the racial population or sexes. Do you think that that's going to feed into any disproportionate biases that may already exist within healthcare for different socioeconomic statuses? Maybe I can answer. Yes, that is a massive issue, and there's many, many studies that are showing that AI is not just reproducing the biases that are present in the training data set, but also it's exponentially inflating these biases. And we know, especially in healthcare data, this is a massive, massive issue, so we have to be really, really mindful of this. And there are technical ways to address this, but probably none of them is perfect, so we always have to be very, very, very mindful. I think this is an important issue, and this is part of the caveats as to, you know, when you see the absolute truth, it may not be, because it's just, you know, going from the lens of what was fed into the model. This leads to a major bias, but it also could lead to a reverse bias, because what happens is that we may identify that you ask a question that, okay, we want to have somebody who could do mathematics and chemistry, and so then, you know, the AI-based model, and this has happened, would predict, and then you may actually end up identifying a man for that particular position. Then you remove that and say, no, we really don't want that assessed. What will happen is it will go to the next level of bias that's fed into the model. So I think understanding the biases built into the model and the data that's fed is critical, and the key thing is that nothing is perfect, but we can always improve, would be the idea. Female Speaker Just to add on that, the metadata that's provided with the transcriptomic data needs to be very rich as well, and finding algorithms that can actually pull out the ancestry, and the other thing is that the funders are actually identifying this, and they're saying if you want to have funding for a data set, 25 percent of it, it needs to be representative of at least your population, so the more we can go back to biobanks and pull out representative samples, the better we'll be to answer that question. Male Speaker Thank you so much. I enjoyed the talk very much. My question is to our friend from Germany. It was very interesting to see that when you uploaded that histological image to CharGBT, it was able to give you a description. Do you think that because simply CharGBT was able to go to the metadata associated with the image, or CharGBT was able actually to analyze the image, go and compare it to the massive data? Male Speaker Good point. I removed that. I took a screenshot, and I just named it image. Excellent point, but no. Male Speaker Very interesting, very impressive. Thank you so much. Female Speaker Thank you. I have a question for Dr. Lumba. So based on your experience for the next clinical trial point of view, do you think so far the AI-based machine learning or this is digital reading could be integrated to the consensus of pathology reading for the endpoint of the clinical trial? Is that feasible for now? Dr. Lumba Yeah, I think it's totally feasible because you can improve precision of assessment. The only thing would be that we would may still actually go back to showing improvement in one state of fibrosis, so that may still be imprecise because that's where the current outcome assessment is, you know, in the regulatory guidance. But I think in future we could show with series of trial that the AI-based assessment is providing you even more precise estimate that is linked to improvement of disease as well as worsening of disease. Currently, the data we have from natural history studies is worsening of disease and how that's associated. But that regression, if we can point out with the AI-based assessment, and there are multiple of those are available, I think this would be an efficient way of running clinical trials in future. Ranjan Mukherjee Hi. Hello. Thank you for all the nice talks. I am Ranjan Mukherjee from University of Pittsburgh. So my question is directed to Dr. Sonia. So my question may be a bit naive, but I wanted to ask you, is it possible to use the spatial transcriptomics to identify subpopulation of macrophages like M1, M2? So is it possible? Dr. Sonia Chattopadhyay Yeah, so you can use spatial, the Xenium platform because in general you have about 377 genes that are part of the multi-tissue panel and then you have 100, you can have up to 100 genes now and apparently in the future up to 1,000 genes that are specifically guided by the populations you want to subdivide. And so that, and they get titered based on your pilot single cell dataset. So absolutely you can do it with that. As well with the Xenium, the Visium, you deconvolve your mini-bulk RNA-seqs with your single cell dataset. So you can tell the difference between, and so the more, those more tissue repair ones are the more M2-like and the more ones, the ones that were more inside the lesions were more M1-like from both approaches. Okay. Thank you. Thank you everybody. Maybe I'll just finish up with just one question. So for the folks more involved in the development of AI, is there any sort of theoretical or conceptual reason why the human judgment needs to be the reference standard for AI? Actually currently it is, right, because AI is in its infancy. But can you see a, is there any theoretical reason why it always needs to be, well, the gold standard is human judgment as opposed to anything else? No. So even chat-GBT is trained in a self-supervised way, so on all the text which they scrape from the internet without like pre-imposing any other knowledge. So there, I'm not aware of any like theoretical limit except the availability of large unlabeled training datasets that would limit that to human, maybe you have another perspective? I think it's important here to distinguish between what, how we train AI models versus how we use them. So for training AI models, the current prevailing paradigm is that of self-supervised learning, which effectively allow us to leverage large amounts of unlabeled data without requiring manual annotations to train AI models. And so there are lots of studies showing that this is the kind of paradigm that will likely scale up to very large data, very promising. But on the other end, the use of AI models, especially in high-stakes decision-making domains such as healthcare or law or policing and other areas where decisions are being made that can affect individuals' life, in that regard then the question of autonomous AI, where AI would be making predictions that are kind of independent of human intervention, is particularly challenging because it touches questions related to liability and all other considerations. And that is where the human judgment part is ultimately incredibly important. But for training AI models, we really have these new algorithmic approaches based on self-supervised learning that can allow us to train models without large amounts of human-labeled data. Thank you. I completely agree with, you know, what's being said, but how we might apply in clinical practice as the tool this year or next year would be that it will be a tool as an adjunct to clinical judgment. So I may see a patient, you get the whole MRI, the MRI would have an automated read. I click on it, it'll say, potentially HCC, high likelihood, and it can give me a likelihood ratio. It could be 92 percent hepatocellular carcinoma in segment two clinical correlation needed. And then I still assess whether it is or not. But that's where I think how it will come to our clinical practice. Thank you so much, everybody. Thank you so much. This concludes the session. Congratulations to all of the speakers. This was so very well attended, and all the discussion was very stimulating.

artificial intelligence

diagnosing

treating

liver fibrosis

histological assessment

NASH

HCC

genetic factors

liver fat

stiffness

MRI imaging

biomedical research

spatial transcriptomics