Welcome to the Non-Alcoholic Fatty Liver Disease Symposium 2021. This year, we will cover the topic of precision medicine in evaluation and treatment of MAF-LD. I am Silvia Zorcoian, a physician scientist from the University of Buenos Aires and the National Scientific and Technical Research Council of Argentina. This is my disclosure. Keystones of precision medicines are to predict the onset and to predict the disease progression. So, patients may be classified into high or low risk for disease onset and severity, and therapeutic interventions can be personalized. And most importantly, precision medicine must ensure health benefits across geography, ethnicity, age, and gender. The question is, what do we need to empower the transition from conventional to precision medicine? And the answer is, we need to understand the biological mechanisms of the initiation and progression of the disease. And the knowledge needed should be multiscale across all levels of biological organization and extremely heterogeneous in type, including understanding the perturbation of physiological processes, multiscale biological information across all levels from DNA, RNA, proteins, and metabolites. Finally, we do need to uncover the interconnections and the dynamics of the parts interaction to determine how the system functions or exhibits this function. Hence, it is essential to integrate information from the different levels with clinical data, which finally will be translated into efficient decision making for healthcare. In the case of NAFLD, the biological information gained through the years has to be integrated with clinical information and translated into efficient and reliable decision making. This process increases efficiency in clinical setting and improves diagnosis, which ultimately facilitates the disease monitoring and therapies and saves on cost of healthcare. Here we detected three topics that not only gain attention over the last years, but also may explain the pathophysiology of the disease and suggest new approaches to diagnosis and therapy. And these topics are genomics, single cell transcriptomics, and artificial intelligence. Then we prepare for you this amazing agenda with great speakers. Dr. Rothman will talk about the clinical utility of genomics for NAFLD and NASH. Dr. Henderson will illustrate single cell transcriptomic signatures to deconstruct NAFLD and NASH prognosis. Dr. Rinella will talk about precision medicine, must fit precisely fit for purpose. And finally, Dr. Garry will illustrate artificial intelligence for assisting NAFLD and NASH diagnosis. So thank you very much and enjoy this symposium. Good morning, everyone. I'm Yaron Rothman from the National Institutes of Health, the NIDDK of Bethesda, Maryland. I thank the organizers for inviting me to talk today. And I was asked to talk to you about the clinical utility of genomics for NAFLD and NASH. However, I think a better title might have been, should I genotype my NAFLD patient? Here are my disclosures. So this is the outline of the talk. I will talk about genetic findings that are associated with fatty liver disease and NASH. And then I'll spend the second half of the talk talking about how these are applicable or not to the individual patient along these specific topics of diagnosis, staging, prognosis and treatment. So we think about fatty liver disease in a simplified manner, typically as an overload of food intake that leads to accumulation of fat. In some subjects, we have progression to NASH, hepatitis, fibrosis, and the development of those outcomes as cirrhosis, hepatocellular carcinoma, and liver-related complications. And at least theoretically, genetic components could affect each and any part of this pathway. Today, there are several genes that have been associated with non-alcoholic fatty liver disease. Due to lack of time, I'll focus only on the top three today, because there's most data available on that. So the first finding, basically, I think that revolutionized the field of genetics in fatty liver disease, were these two papers published in 2008, which identified an association between PNPLA3 and fatty liver disease. This was a genome-wide association study from the Dallas Heart Study, and this is a GWAS looking at the association with ALT. And both of those found a target on chromosome 22, which mapped out to a gene called PNPLA3. These previous studies were in the general population. We, and many others, quickly found that the PNPLA3 is also associated with liver histology. For example, we found that the minor allele in this variant is associated with increased steatosis, inflammation, malary bodies as a mark of injury, and liver fibrosis. Not only is it associated with histology, but we actually, others have found that it's associated with real outcomes. For example, this study of a cohort of Italian patients found that having the risk allele G is actually associated with increased risk of developing hepatocellular carcinoma, whether you look at all patients with fatty liver disease, or just those at high risk with F3 or 4. Recently, it was actually shown that PNPLA3 variants can actually predict liver-related mortality in the general population. Here, the terminology is the protein, so the M is equivalent to the G allele in the SNP. Overall, over this scheme of fatty liver disease, it seems that the mutation in PNPLA3 is increasing the risk for everything, for accumulation of fat, for NASH, fibrosis, and hard outcomes. The second variant that was identified a few years later was this variant in the gene TM6-SF2, again identified through the Dallas Heart Study to be associated with liver fat content. For example, you can see here that homozygotes for the minor allele have a much higher average fat content than other genotypes, although you can also notice that this is a relatively rare occurrence. TM6-SF2 was also found to be associated with histological severity of anaphylactic steatosis, inflammation, ballooning, and fibrosis. It was identified to be associated with hepatocellular carcinoma in subjects with alcohol-associated liver disease, and this year was also associated with liver-related mortality in this nice paper. Overall, TM6-SF2 has the same pattern of association as PNPLA3, although the magnitude of effect is somewhat smaller. A third gene I want to mention is HSD17B13, my favorite target. This was identified initially to be associated with ALT in a genome-wide association study in 2011, and our group confirmed that the variant from that study was actually associated with liver histology. Interestingly, the minor allele is associated with increased steatosis, decreased inflammation, decreased ballooning, decreased malarie bodies, and a trend towards decreased fibrosis. This is actually not a mutation, not a coding variant, but the causative variant is actually this one, 72613567, which causes an altered splicing of the gene product and has the same pattern of association. This is just showing a confirmation. These are the associations in our histological group, which is derived from the National Clinical Research Network, where the minor allele is associated with more steatosis but decreased injury, protected from injury. This protection was confirmed in the UK Biobank database in collaboration with Liz Filiotis from the University of Michigan. This variant was identified simultaneously and independently by Nora Abel-Huston and the Regeneron group. Again, using a genome-wide association study and the same pattern of association, the trend for increased steatosis, whereas they show protection from NASH, from fibrosis, and also from hepatocellular carcinoma and cirrhosis. These genetic findings have since been replicated in many other studies. This is an example from an alcohol-associated liver disease, where the variant is protected from severity as well as from hepatocellular carcinoma in these patients. This is a cohort study from Denmark, where the variant is actually associated with decreased liver-related mortality, the mutation. In contrast to what I've shown you before, HST17 mutations in HST17B13 seem to be driving increased steatosis, but a decrease in NASH, fibrosis, and liver-related outcomes. As I mentioned, there are other variants that I will not discuss today. Consequently, these variants seem to be relatively independent of each other. There's some debate whether PNP-Lif3 and HST17B13 are truly independent. But overall, the question is then, if we genotype a subject for more than one, how can we synthesize that information? The tool or the way to do that was using genomic risk scores or GRS. There are several ways to do that. I'm showing one example. In this study, the authors just counted the number of risk alleles in these three genes. You can have anywhere between zero to six risk alleles. What they show is that with more risk alleles, there's an increase in average ALT in two different studies. Importantly, these are associated with much higher risk for cirrhosis and for hepatocellular carcinoma. The odds ratios are quite remarkable, even though the number of cases is quite small. When you're listening to this, I'm sure you're saying to yourself, wow, these are amazing p-values and amazing odd ratios. You're probably also impressed by my command of PowerPoint graphics. But should I genotype my patient? How does that apply to the individual patient? And when we discuss this, I want to think about the four things that we do as clinicians. Diagnosis, staging of the disease, talking about prognosis, and offering or selecting for therapy. So let's look at these one by one. If we look at a hypothetical patient, let's say we have a 53-year-old Hispanic male who is obese, is diabetic, treated with metformin, does not drink alcohol, has mildly elevated ALT with negative viral serologies. If we were together in the same room at Anheim, as we all hope to be, I would ask you to vote. What is the next test you want to order? Do you want to do an ultrasound of the liver, or do you want to genotype that patient for PNP Day 3? And I'm hoping that most of you would vote for an ultrasound. So we'll send the patient for an ultrasound, and the ultrasound demonstrates hyperechoic liver. And again, in the Anheim room, I would ask the audience, what does the patient have? And everyone would yell at me, non-alcoholic fatty liver disease. So if we also genotype the patient, does it help us? If we find that the patient is homozygote for the low-risk variant in PNP Day 3, does that lead us to think that the patient does not have NAFL? I don't think so. Let's try another patient. Let's say that we have a 32-year-old Caucasian female with the same chronically elevated ALT. This patient is lean, has no medical problems, does not take any medications, does not drink alcohol, has negative viral serologies, and her ultrasound is normal. Do we think that this patient has fatty liver disease? Probably not. If she's homozygote for the high-risk genetic variant, do we think she has fatty liver disease? Still probably not. Can we quantify the contribution of the genetics to diagnosis? So let's look at some data from this beautiful Genome-Wide Association study by Quentin Lansky and others that was published last year. So one of the secondary cohorts, this is a cohort of Italian subjects, and has about 40% of the cohort is with fatty liver disease, so that's roughly close to the population prevalence. The risk allele in PNP Day 3 was associated with the presence of fatty liver disease with very beautiful odds ratios for a genetic study and a p-value to die for. These are graphs not from the paper, but I extracted the data. So if you look at the allelic frequency between fatty liver and control, you see that there's a marked difference, but you're actually seeing that both alleles are quite common in the groups. And if you look at the genotypes, and you're looking at the homozygotes, heterozygotes, and homozygotes for the low-risk allele, you see that you have fatty liver patients and controls in both groups. If we actually look at what is the positive predictive value, we see that if all we know about the subject is that they're a homozygote for the high-risk allele, in this cohort there's a 61% chance that they have fatty liver, but also a 40% chance that they don't. And if they're homozygotes for the low-risk allele, there's only a 75% chance that they don't have fatty liver. So what I'm suggesting is that I don't think that genetics will be useful to diagnose fatty liver disease, especially since we have simpler, accessible, and cheaper tools. But, and there's always a but, I just would like to draw your attention to this beautiful review that was published just in October in Hepatology by Silvia Villarino and others. And genomics may be important when we're looking for non-common disorders. So they suggest that if you have a lean patient who has no visceral adiposity, and maybe has some other markers suggestive of rare disorders, this may be a patient worthy of genomic evaluation. And you can find things like lipodystrophy, mitochondrial disorders, hypobeta lipoproteinemia, and this may direct therapy. So overall, though, I think that it's not useful for diagnosis. What about staging? So can we assess the fibrosis? You've seen this graph before. This is from our own data. But then let's look at the predictive value of genomics for fibrosis. So if we look at what is the positive predictive value in this cohort of about 800 people with biopsy proven ash, we see that if you're homozygote for the high risk allele, you only have had 41% positive predictive value for advanced fibrosis. If you're homozygous for the low risk, you only have 71% negative predictive value. So these are not quite useful. Can we use them to improve on non-invasive tests on it? So I calculated the fib4 data from that study using the acceptable cutoffs. We see that the positive predictive value for the high threshold of fib4 is 80% in this cohort, and the negative predictive value for that would be 82%. So these are acceptable. Can we improve on this if we genotype the patient? So what happens now? Well, the positive predictive value for the high threshold does not change much. The negative predictive value for the low threshold does not change much. And the intermediates are still very, very intermediates. So I would argue that although genetic information may alter our pretest probability and may improve those tests overall in a population, their ROC curves may look a little bit better. I don't think they'll be useful for the individual patient. What about prognostication? So again, I've shown you this figure showing in a cohort of Italians with fatty liver disease the association of PNPLA3 genotype with development of hepatocellular carcinoma. So over the study period, which was a median of about seven years, people with heterozygous or homozygous for the risk allele had about 10% risk of developing HEC if they were F3F4. If you look at the entire population, this is actually somewhat lower. Again, these are data that I extracted from the paper. I tried to estimate the annual risk based on those, and you should take those with a huge grain of salt. But overall, if we look at the annualized risk for people with advanced fibrosis with F3F4, it seems to be about 1.1% to 1.2% if they carry at least one copy of the risk allele. Now, currently, ASLD guidelines recommend screening for paracellular carcinoma, all serotics. The new guidance from the AGA that was published earlier this year suggested also considering screening for people with F3F4. Now, can we use this data to suggest that people with low genetic risk should not be screened? Maybe, but this is really based on a small number of subjects, this data, and I think this merits further study. I can see a role for genetics in potentially selecting or amending HCC screening criteria, but this is not yet based on data. Finally, I want to talk about therapy, and when we think about the applicability of genomics to treatment in NAFLD or NASH, I can see several applications. For example, can we use genomics to choose who to treat? We want to take people with high genomic risk and treat them more aggressively, intervene earlier, maybe tell them that they should never drink alcohol. It makes sense, but this actually has to be studied formally. Alternatively, can we use genomics to choose how to treat? It is quite plausible that there will be genetic variants that will determine responsiveness to treatment. For example, there may be variants that will make a person more responsive to a GLP-1 agonist, to an FXR agonist, to an HCC inhibitor, and these may not necessarily be the same genetic determinants of disease progression. For example, think about the IL-28B story that we had in hepatitis C in the interferon era. However, for this, this is, I think, somewhat futuristic. First, we need to have therapies that are approved, then we need to have large studies of responsiveness in the genetic context. What about identifying new therapies for fatty liver disease? Well, here is where there's some exciting news. There is a clinical trial targeting PNPLA3 using an antisense oligonucleotide run by AstraZeneca, and this is already ongoing in humans. This is not specified, but my prediction is that the target population for this kind of treatment will be subjects that have the mutant, because the mutant is worse than the wild type, so we don't necessarily want to knock down the wild type, but we do want to knock down a harmful mutant. This is treatment based on genetic information, but also may require genotyping in order to select subjects, which is precision medicine. Similarly, there are two trials now to target HSD17B13 in humans, one by Alnylam and the other by Arrowhead. Both of those are using antisense oligonucleotides as well, and the data from the Arrowhead study was actually shown recently at the ESL meeting, and again, I think that here people will have to be genotyped, and the target population is likely going to be people who carry the wild type, because the mutation in HSD17B13 is protective. This is the data from the Arrowhead trial that was posted in the ESL in the spring. These are only five patients, and what you're seeing is the response in their ALT after treatment with an oligonucleotide targeting HSD17B13. I think that we are not really going to use genetics for diagnosis. I don't think we will use them for staging. I think we may be using them for prognostication and selection of subjects for screening or for treatment, and the treatment we're getting there. To conclude, I've told you about genetic variants that are associated with the pathogenesis and the progression of NAFLD and NASH, and I think that genetic analysis may be useful in the future for evaluating the rare and unexplained patient, to risk stratify subjects for treatment and for screening, for precision medicine, and to help us identify new targets and new therapeutic targets. But to answer the question I posed at the beginning, should I genotype my NAFLD patient, I think at this time the answer is not yet, and with that, I thank you, and I hope to see some questions in the chat. Good afternoon, everyone. My name is Neil Henderson. I'm based in Edinburgh, and I'd like to thank the organizers very much for asking me to give this talk for the NAFLD SIG session. It's a real shame I can't see you all in person. Hopefully next year when things get back to normal. So what I'd like to discuss over the next 15, 20 minutes or so is how we've been using single cell genomics approaches in Edinburgh to try and help deconstruct and understand human NASH at the cellular and molecular level. Just a disclosure statement to begin with. So the fibrotic niche in the liver is in many ways similar to fibrotic niches in other fibrotic organs in that we have, you know, major players in terms of mesenchymal cells that become myofibroblasts and lay down scar, and the usual panoply of cell types. So immune cells, endothelial cells. And within the liver, we have some elegant clockwork down in the space of DISA, and this is a deliberately oversimplified schematic here showing the various different cell lineages. But what we know from a lot of work over the last 30, 40 years is that if you interfere or manipulate with really any of these cell lineages and mouse models, you can start to abrogate and manipulate fibrosis. But what perhaps is less clear is how liver, human liver fibrosis comes together in terms of these multi different cell lineages driving fibrosis. So in Edinburgh, we set up a few years ago, a single cell genomics pipeline. We're fortunate in Edinburgh in that we have the National Liver Transplant Unit. And so we have multiple transplants going on every week, and this has allowed us to procure human liver tissue for these types of studies fairly quickly and readily. We see a broad demographic of liver disease, like many transplant centers do throughout the world. And so we initially started with a single cell RNA sequencing pipeline, and you can see here we use the 10X chromium system. But we've also started to expand into other modalities, such as single nuclei RNA sequencing, spatial transcriptomics and multi-home sequencing, as I've listed bottom left. So the initial studies we did took five healthy livers and five cirrhotic livers. And this is the single cell census we did initially, about 135,000 cells, which we published in 2019. And these are TSNE dimensionality reduction plots, where for someone like myself, who's essentially a biologist, it allows me to try and get my head around this highly dimensional, complex data. And so what we started very deliberately with was just asking the question of what's the difference broadly between healthy and cirrhotic data sets? You can see in the middle there we've plotted that. And then we use clustering approaches to subdivide all these non-parenchymal cells into the different clusters and then manually annotate them by lineage using signatures. And we still use manual annotation of clusters. Other groups use automated approaches, but we find that it's still most helpful to use manual annotation to split them up into all the various cell lineages. And so in the next few slides, I'm going to use macrophages specifically as an exemplar cell type of how we took some of these lineages and started to try and work out whether there were certain macrophage populations that were enriched in cirrhosis and what they might be doing. So this is a TSNE plot of all the various different mononuclear phagocyte clusters. And what struck us very quickly was that there were two populations at informatics level, so this is gene level data, that were very much expanded in cirrhosis. And you can see here SAM1 and SAM2 were significantly expanded in cirrhosis. So we then wanted to ask the question of, OK, we've seen that at gene level, but can we validate that at protein level? And so what these plots show is John Wilson Kanemori in my lab, who's one of the bioinformaticians, created differential gene markers for this macrophage subpopulation. And John found TREM2 and CD9 to be highly upregulated at gene level in these macrophages. So then Prakash Ramakandran, who was leading on the wet lab side of the work on this project, then prospectively isolated single cells from healthy and cirrhotic human livers and did flow cytometry with these markers, TREM2 and CD9, to see if this macrophage population was truly expanded in cirrhosis. And what you can see here, and this is high end on the right-hand graph of both healthy and cirrhotic livers, that these TREM2, CD9 positive macrophages were indeed expanded in cirrhosis. So our next question we wanted to ask was, what's the topography of these TREM2 positive, CD9 positive macrophages in human liver fibrosis? And so this is just some multiplex amino fluorescence. We have collagen in green with a hepatocyte nodule in the middle with swathes of fibrotic septae around it. And what caught our eye was that these TREM2, CD9 positive macrophages did indeed seem to be enriched in the areas of scarring. And so what Prakash and Elena in the lab then did was they used various approaches to ask the question of, are these macrophage subpopulations truly enriched in the scar regions? And this is an image J approach where you can split the livers into nodules versus fibrotic septae. And in essence, this confirmed what had caught our eye initially, i.e. with quantitation that TREM2 positive, CD9 positive macrophages are upregulated in the fibrotic niche. Then switching lineage, this is endothelial subpopulations in the healthy and cirrhotic datasets. Again, using a very similar approach, clustering, subclustering, and then looking at what endothelial subpopulations might be expanded in cirrhosis. I'd like to draw your eye to the ACKR1 and PLVAT positive populations in the middle. These were very much expanded in cirrhosis. And again, taking differential gene markers that John had generated, the wet lab guys then did validation and interrogation in multiplex amino fluorescence. So you can see here ACKR1 and PLVAT used as markers in yellow and green. ERG is an endothelial nuclei marker that we find very useful for marking endothelia. And you can see that there was this expansion of ACKR1, PLVAT positive endothelia within the fibrotic septae. There's hepatocyte nodules to the left and the top right of this micrograph. And again, with quantitation and cirrhotic liver, this confirmed that these ACKR1, PLVAT positive endothelial subpopulations were indeed expanded in cirrhosis. We got to a stage where we were able to identify from the initial large dataset, SCAR-associated endothelia, SCAR-associated macrophages, and also SCAR-associated mesenchymal cells. And I've not had time to present the data on the mesenchymal cells today. However, we then wanted to ask the question, how do all these SCAR-associated nonparenchymal subpopulations actually come together to drive fibrosis within the fibrotic niche? And to try and get at that and understand that more, we used a ligand receptor algorithm called CellPhoneDB, and this was developed in Sarah Teichman's lab down at the Sanger in Cambridge. There's been a couple of iterations of this algorithm, as there often is with single-cell algorithms, because everybody just makes them better and better. And I've added the reference there at the bottom for people who are interested in looking at this. But we found this an incredibly helpful algorithm, because it allows you to add which ligands and receptors are interacting on all the various subpopulations within these datasets. So you can look in purely the healthy liver datasets, or you can use it to interrogate SCAR-associated subpopulations in the disease datasets. And so we use this to try to start to understand, for example, how do SCAR-associated endothelial subpopulations interact with SCAR-associated macrophage populations or mesenchymal cells, and so on. So it generates a lot of data, as you might imagine, and a lot of outputs, but what I would say is there's no substitute for the human eye, and so we pored over these datasets. And what was reassuring was there was quite a lot of known prior biology that we found, but also some new interactions that there perhaps wasn't much literature on at all. And so I'm going to just discuss in this slide some of the validation approaches we've used on the wet lab side to try and make sense and also confirm and interrogate the ligand-receptor interactions that we'd seen in informatics data. So what we've done here, top left, is we've taken PLVAP-positive endothelia, prospectively isolated with flow cytometry and cell sorting from cirrhotic human livers, and you can see here that in informatics data, we saw that JAGAD1, a notch ligand, was upregulated at gene level on the PLVAP-positive scar-associated endothelia, and we've confirmed that with flow cytometry there with quantitation on the right-hand graph. You can then go into the tissue and look at topography. So we've used PDGFR-alpha to mark myofibroblasts here, and you can see that DLL4, which is a notch ligand, is present on the endothelia there, with notch 3, the receptor for DLL4, present on the PDGFR-alpha-positive mesenchyme. Then bottom left, we've prospectively isolated quiescent human stellate cells also with PLVAP-positive endothelia, which were the scar-associated endothelia, and when you co-culture those two primary cell types, what we found was that the stellates upregulated their collagen 1 gene expression, and then when we added DBZ, which is a notch inhibitor, to that co-culture system, we could inhibit collagen 1 gene expression. And then on the right, taking quiescent stellate cells, when you knock down notch 3 with siRNA, that inhibited collagen 1 gene expression in the primary human stellate cells. So this is just an example of the various ways that one can use to try and interrogate the single-cell data. Obviously, you can also use preclinical rodent models. The only caveat there, I would say, is that they sometimes don't, or commonly don't, recapitulate some of the populations we see in the human data sets for all the reasons we all know. Short-term rodent models versus, you know, a patient who may have had NASH for 20 years. So there are clearly advantages, but also limitations of the preclinical models as well. And so you can start to build up these what we call interactome maps, which in essence are derived from the cell phone type analysis with also wet lab validation layered on. And you can start to build up these fibrotic interactome maps of how the different SCAR associated populations are talking to each other and interacting with each other. And, you know, we very much hope, along with many others in the field who are using this type of approach, that this will allow us to get sufficient resolution on the system to allow very targeted approaches to antifibrotic therapies. So in the last five minutes or so, I'd like to talk about some of the new directions in the single cell genomics field. And multimodal is a very trendy term just now and for good reason, because the range of readouts that you can derive from a single cell these days is absolutely incredible. Moving rapidly on over the last few years from the initial work where there was single cell transcriptomes, you can now measure single cell epigenomes, single cell proteomics is beginning to gain traction. I don't think it's quite there yet, but it's nearly there. And spatial transcriptomics. And not only that, you can get multiple readouts from the same cell. So really, this field is really beginning to blossom a lot in terms of the different modalities that you can use to try and understand your disease of interest. And obviously in NASH liver fibrosis, these approaches are being used a lot now, but across the board in different types of organ fibrosis, they're being deployed. And if people are interested, we summarised some of this work in Nature last year and some of this review already be out of date. I'm well aware of that, but for people who are interested, you might be interested to read that review. So in Edinburgh, we've been using two different further techniques to single whole cell transcriptome sequencing. We've been using single nuclei RNA sequencing and spatial transcriptomics. The big advantage of single nuclei RNA sequencing for us is that this has opened up hepatocyte data. Before with our wedge biopsies, we couldn't do perfusion digest and get hepatocytes out cleanly or easily, and the yields were very low. So we've used single nuclei RNA sequencing to really open up that, and that's been very helpful. The other advantage of single nuclei RNA sequencing is that you can use it on frozen tissue, which opens up biobanks as well. And spatial transcriptomics, we've been using the 10X Visium system, and this allows in situ on-slide sequencing when we make single cell suspensions for single nuclei or single cell RNA-seq. Obviously that disrupts all the native interactions, whereas with on-slide sequencing like this, with an array of barcodes, it preserves that spatial resolution. And so this is a really exciting technology that's coming on leaps and bounds, and you can use this on frozen tissue, but also there's kits coming out for FFPE sections, which will open this up further. So in Edinburgh, over the last couple of years, we've been building up a data set along with collaborators using single nuclei RNA sequencing to generate big data sets of both healthy human liver, but also as fibrosis and steatosis increases through the F and S grades, and also looking at NAFLD explants with this approach as well. And I would like to highlight Ross and Bas and John for all their hard work in my group on this. In terms of spatial transcriptomics, as I mentioned earlier, we're using the 10X Visium system. This is just some early work we've been doing with the Visium system in NASH cirrhosis. The top picture is a heat map of end features or numbers of genes, and you can see that you have variable amounts of genes across the topography of the liver. And then bottom left, you can split this data up into hepatocyte-rich spots and non-parenchymal cells in the scar areas. And this is quite reassuring in that the hepatocytes or the hepatocyte signatures are within the areas you know are hepatocytes because you can link it back to the original H and E. And similarly, the non-parenchymal signatures are in the fibrotic scarred regions. And Andreas, one of the bioinformaticians in my group, has been spearheading our spatial analysis along with Ross in the wet lab side. And when we apply hepatocyte signatures, we see that these correlate with the parenchymal nodules of hepatocytes. So this is reassuring in terms of that's the topography we would expect. And when we overlay signatures for scar-associated mesenchymal cells, these are very much inhabiting the fibrotic septae and not the hepatocyte nodules. So we've been asking some simple questions initially just to reassure ourselves that the outputs from the spatial transcriptomics data are making sense. One of the approaches we're currently also taking is integrating and pairing single nuclei RNA sequencing data with spatial data from the same liver sample. And that's been very helpful to try and leverage as much information as possible from these readouts. So we have two Nash cirrhotic samples on the left, which overlay fairly well in the UMAP despite being from different patients. And then we use an algorithm called CellAssign on the right to understand, just to say, these are 55 micrometer spots on the array for spatial transcriptomics. And we're using this to understand which spots are most enriched for certain lineages. And so obviously each spot has multiple different cell types within it. In the middle of a parenchymal nodule, it's going to be mainly hepatocytes, but there will obviously, of course, be some endothelial and stellate cells. Whereas if you're in the scar, it's going to be far more mesenchymal cells than hepatocytes or very few hepatocytes at all. And so it's that combination of trying to use prior biology to help understand the readouts from this type of data. And I would say this whole area is rapidly expanding. And I think we're all learning as a community how to best interrogate spatial transcriptomics data in this context. So just to summarize, single cell genomics methodologies really are driving a resolution revolution. What it allows is a very powerful unbiased exploration of cell states and types at single cell level. And so we're really seeing across the liver community some really nice unexpected insights into liver biology and disease. It's a very rapidly evolving field, as I mentioned. You can now measure multiple omic readouts from the same single cell, which is just incredible. And these technologies, we hope as a community, should very much help drive a new year of precision medicine in the treatment of liver disease. And I genuinely think it's a very exciting time to be working in hepatology in this area. I'd like to acknowledge many people I've listed on the left, the members of my lab group who have been absolutely driving on all of this work in all the various myriad ways. And many people have been working very, very hard on these types of approaches in our group. I'd like to thank the patients in the liver transplant unit for kindly donating their bits of their explant liver tissue, because if they didn't do that, none of this would be possible. I'd like to thank the transplant coordinators and surgeons for being absolutely integral to our tissue procurement pipeline. And I'd also like to thank multiple collaborators, Sarah Teichman and John Marioni in Cambridge, for helping us get going with the informatics analysis approaches, and Quentin, Jackie and Robert Schwabe, for all their help with our ongoing projects in this area. I'd also like to draw your eye to two web browsers we have that are publicly available that allow browsing of our data sets. We're very keen on making our data as accessible as possible. It's a very expensive area in terms of experiments, so we want everyone to maximise what we can get out of these expensive data sets. And finally, I'd like to thank all the funders who funded this work and without whom we couldn't do all these various different experimental approaches. And thank you for your time, and I'd be very happy to take any questions. Thank you. I'd like to thank Dr. Sokoyan and her SOVA for the invitation to speak today at the SIG Symposium. My presentation is entitled Precision Medicine Precisely Fit for Purpose. These are my disclosures. So far today, we've heard about the clinical utility of genomics for NAFLD and NASH and how single-cell transcriptomic signatures might be used in the future to deconstruct NAFLD and NASH prognosis. Precision medicine is an emerging approach for disease treatment and prevention that takes into account individual variability in genes, environment, and lifestyle for each person. Over the next 20 minutes, I'll discuss the role of precision medicine now and in the not-too-distant future for the management of NAFLD. Mortality in the context of NAFLD is most decisively linked to fibrosis, and that is reiterated in this analysis of all the biopsies performed in Sweden from 1966 to 2017, amounting to 10,568 biopsies with a median follow-up of 14.2 years. In this cohort, there were 4,333 deaths during the period of observation. And what's most notable about the study is that even simple steatosis was associated with an increase in mortality, as was non-sorotic NASH, of which both extrapedic cancer and cirrhosis were the most common, again emphasizing that both of these, steatosis and non-fibrotic NASH, should be addressed to hopefully prevent long-term morbidity and mortality. Progression to cirrhosis is less predictable in NASH than in other chronic liver diseases. NASH is best characterized by repetitive bouts of injury and repair. Over time, various factors that can be common or unique among individuals, such as aging, lifestyle and nutrition, extent of hepatic and extrapedic comorbidities, alcohol, etc., lead to metabolic and inflammatory stress that then results in the accumulation of fibrosis. At any point along the process, reparative processes can have beneficial impact. The earlier the balance tips in favor of restorative processes, the higher the likelihood of preventing disease progression, decompensation, and death. For the vast majority of patients with NAFLD, disease will improve or even resolve completely over time with successful lifestyle change. This approach combats a disease at a high point in the causal pathway and should be attempted and supported as a cornerstone of therapy. This illustrates some of the targets, not meant to be complete, that are being explored. Many have failed, and those that haven't have produced fairly modest impact on NASH, related at least in part to the significant heterogeneity of disease between individuals. The bound milieu of the individual with NASH is complex, and it becomes difficult to separate the impact of genes from that of the environment. In this picture, you can see Archie Manning, a fellow at the University of Michigan, father of the pro football player Eli, and Peyton, both highly celebrated pro football players, as many of you know. In contrast, you can see here Lionel Messi and his brother Matias, who have less similarities despite similar genetics. In the background, you can see a gene chip of a patient with NASH, and you can see how many genes may be over- or under-expressed, which is illustrative of this complex disease. Recent estimates suggest that 21,000 protein-encoding genes from the human genome and specific influence of genetic polymorphisms is becoming increasingly clear. In 2008, Stefano Romero, working with Helen Hobbs, identified the association between a gene variant of PNPLA3 in hepatic fat from the Dallas Heart Study, explaining some of the ethnic variability of the disease. Since then, other gene variants have been identified to play a role in modulating fatty liver disease, as discussed by the previous speaker. While we are not ready to guide therapy using genetic scores, significant progress has been made. One can envision how one could combine genetic information, weighted metabolic comorbid risk, in addition to environmental factors, to model trajectory of disease. In this hypothetical model, the number of inherited NAFLD risk gene variants, which can be expressed as weighted genetic risk score on the Z-axis. The Y-axis presents the number of exogenous risk factors for NAFLD. The interaction of all the presented factors over time, represented by T on the X-axis, results in the NAFLD phenotype. The color code illustrates susceptibility in red and healthy state in green, and the final disease stage is indicated by the white circle. The number of exogenous and genetic risk factors in NAFLD at the time during which they can exert damage determine individual's trajectory and progression over time. The inherited predisposition generally remains stable over time, whereas the exposome or the external factors can change to modulate the trajectory in carriers of NAFLD-associated gene variants. Such a model could facilitate risk stratification, outcome prediction, targeted screening, and allow for proper individualization of therapy. While the presence of genetic risk is seemingly not modifiable, the association of these variants with fibrosis are largely dependent on the extent of steatosis, and thus could be exacerbated in the context of obesity. Furthermore, a recent study suggests that increased ingestion of certain dietary micronutrients may be protective as evidenced by reducing the reduced odds of having greater than or equal to F2 disease in the context of the G allele of PNPLA3. For example, increased intake of N3 PUFAs, isoflavones, methionine, and choline were associated with a reduced risk of having F2 or higher disease, whereas carbohydrate intake was associated with more disease. Here you can see examples of general nutrients that could be modified in somebody's diet to potentially favorably impact NASH histology in the context of PNPLA3 at-risk genotype or just in general. Here you can see the potential for differences in PNPLA3 genotypes with respect to response to lifestyle intervention. On the left is an example. The left graph shows dietician-led lifestyle modifications and more intensive program, and you can see the CC, CG, and GG are genotypes noted below. You can see that there is more intraepatic triglyceride reduction in those with GG genotype. This is also present in routine care, which basically was no intervention, however, not statistically significantly. Based on these, albeit small data, each G allele has a 3 percent more absolute reduction in intraepatic triglycerides. For example, 6 percent more reduction in intraepatic triglycerides in homozygotes compared to CC homozygotes. So, if NAVLD is associated with the presence of PNPLA3 at-risk allele at an individual level, is it more amenable to changes in macronutrient content or lifestyle intervention, recommending diet and exercise is the right thing to do? This would be true regardless. It is important, though, to note that we should have caution. Caution is needed when you translate a genetic biomarker data that may have meaning at a population level into an individual risk prediction. In this particular example, a genetic test predicting a hypothetical condition with variants carrying modest, as noted there, by 1.5 odds ratio, sizable as denoted by 10 odds ratio, and large as denoted by an odds ratio of 50, showing false positive fractions at 80 percent sensitivity, represented by the upper black dotted line, and then the fractions below of false positive frequency are 75 percent, 25 percent, and less than 10 percent, respectively. So, this graph demonstrates that very large odd ratios are needed to provide acceptably low false fractions in this sort of a situation applying these data to an individual risk stratification level. So, in this very elegant paper published almost 10 years ago, the authors initially demonstrated that inflammasome-deficient mice fed an MCD diet for 24 days had increased susceptibility to liver injury compared to wild-type mice, which you can see as demonstrated by AST, ALT, and the NAFLD activity score. These wild-type mice were then co-housed with the inflammasome-deficient mice for four weeks prior to the introduction of MCD diet. The transfer of microbiota, which occurred through coprophagia, resulted in a significant exacerbation of NASH in the wild-type cage mates, as denoted by W2…I'm sorry, wild-type WT in parentheses ASC negative slash negative, as compared to singly housed age- and gender-matched wild-type controls. Nearly 10 years ago, this very elegant paper was published demonstrating that inflammasome-deficient mice fed an MCD diet for 24 days had increased susceptibility to liver injury compared to wild-type mice. Wild-type mice were then co-housed with inflammasome-deficient mice for four weeks prior to the introduction of MCD diet. Then, transfer of microbiota, which occurred through coprophagia, resulted in significant exacerbation of NASH in the wild-type cage mates, as you can see in the middle panel, and exacerbation of NASH, as compared to the singly housed age- and gender-matched wild-type controls noted on the left. In co-housed wild-type mice, disease severity reached comparable levels to that of co-housed inflammasome-deficient mice. So, this demonstrated the ability to transmit susceptibility to develop NASH through the microbiome. The human microbiome is ostensibly more complex and impacted by numerous environmental factors. I encourage everyone to attend this year's Leanne Schiff State of the Art Lecture given by Bernd Schnabel on the role of the microbiome in liver disease. In this paper, the authors define a gut microbiome signature for NAFLD cirrhosis determined from metagenomic and metabolomic characterizations of stoon microbiomes from a twin study and family cohort of NAFLD cirrhosis probands and their first-degree relatives. In panel B to the left, you can see inverse Simpson alpha diversity scores that highlighted significant decreases in the richness of gut microbiota from the NAFLD cirrhosis group, and that's all the way to the left, depicted in orange compared to the non-NAFLD group depicted in blue. Significant alpha diversity differences were also observed between non-NAFLD control groups and NAFLD cirrhosis patients in two other cohorts in the middle panel from China and the far right panel from Italy, respectively. In the next panel over, you can see stacked bar plots depicting class-level differences in gut microbiome composition between NAFLD cirrhosis and non-NAFLD control groups. And then far to the right, receiver operating curve for the model including 19 discriminatory species in the UCSD proband cohort that included 27 NAFLD cirrhotic patients and 54 non-NAFLD control stool samples. This demonstrated high accuracy and the ability to distinguish patients with and without cirrhosis based on microbiome. This table shows data from a small study that determined the safety, tolerability, and impact of mucosal and stool microbiota and brain function in hepatic encephalopathy after capsule or fecal microbial transplant in a randomized single-blind placebo-controlled clinical trial of cirrhotic patients with recurrent hepatic encephalopathy. Although not the primary outcome of this study, differences in the occurrence of SAEs and specifically episodes of hepatic encephalopathy or episodes of hepatic encephalopathy that required an ER visit or an admission were more common in those in the placebo group compared to those that had FMT. It's important to note that all of these patients could have been on Rifaximin and they were all on a PPI, but nevertheless, this certainly suggests that there's a signal and certainly worth study in a larger group. In another study, also from Jaz Bazzage's group, you can see pre- versus post-FMT correlation network differences. It showed changes centered around the encephalopathy and serum IL-6. You can see in the bottom two bar graphs that IL-T dropped from before to after FMT compared to in the placebo group where there was no significant change, and importantly, there was cognitive improvement related to reduction in some of the factors measured by the encephalopathy. What does the future hold for harnessing the microbiome to treat liver disease? Increased insight into impact of diet, genetics, and environment on the microbiome, better functional characterization of the microbiome, and individualized microbial cocktails targeted to specific dysbiosis or stages of liver disease. The growth hormone axis directly impacts many physiologic processes that modulate the development of NASH. Impaired growth hormone secretion or action could promote visceral fat accumulation, decreased de novo lipogenesis, and worsened insulin resistance in addition to the adverse effects on mitochondrial function and activation of hepatic stellate cells, thereby leading to the accumulation of fibrosis. This is a study of 160 patients with medically complicated obesity, 10 percent of whom were normal and the rest represented the full spectrum of NASH with an enrichment in earlier stage disease. Even so, you can see that there is a variation in growth hormone levels for the different histology groups, which are depicted on the bottom left-hand panel. Compared to the obese normal, all patients with non-alcoholic steatohepatitis with fibrosis stage 2 and 3 had circulating levels of growth hormone that were less than 0.45 nanograms per mL, which is within criteria for the diagnosis of growth hormone deficiency. Now, it's important to note, though, that there are numerous issues with accurate measurement of growth hormone given its pulsatile nature, and it's difficult to measure accurately in a routine clinical setting. Tesamoralin is a growth hormone-releasing hormone analog, and it increases endogenous pulsatile growth hormone production and reduces visceral fat in patients with HIV. It's FDA-approved for the treatment of HIV lipodystrophy, but is being studied in the context of fatty liver disease. Tesamoralin has recently been found to modulate various pathways potentially relevant to NAFLD in the context of HIV that may also be applicable to non-HIV-related NAFLD. In this randomized controlled trial of 54 patients with HIV and truncal obesity, tesamoralin 2 mg sub-Q was given for six months. Here you can see on the left differential effects on visceral adipose tissue and subcutaneous adipose tissue with tesamoralin producing a significant reduction in visceral adipose tissue without a change in subcutaneous adipose tissue and placebo with no effect. And then moving towards the right, a significant reduction in visceral…sorry, excuse me, in fat mass with an increase in lean body mass. The opposite effects were seen in the placebo group. And then far to the right, you can see a significant reduction in absolute liver fat content in tesamoralin compared to placebo. This is a 12-month double-blind treatment trial during which participants were randomized one is to one to receive tesamoralin 2 mg a day or placebo, and then followed for six months with an open-label tesamoralin arm. The pre-specified primary analysis here was looking at changes between tesamoralin and placebo. You can see over to the left absolute and relative changes in hepatic steatosis. There were several pre-specified secondary endpoints, including histological assessment of hepatic fibrosis using fibrosis stage. You can see here to the right that there was a reduction in progression of fibrosis in those treated with tesamoralin. Again, a cautionary note that this was, excuse me, a small number of patients, but nevertheless is providing further enthusiasm for an ongoing study in non-HIV NASH, which is a 12-month randomized controlled trial with a six-month open-label extension, just like this one in patients with non-HIV fatty liver disease, and also looking at cardiovascular risk profiles in these patients. In a large multi-ethnic cross-sectional study, actually in several low testosterone levels are independently predictive of fatty liver in men. Based on this, these data, LPCN 1144 was used in the treatment of non-cirrhotic NASH. This is an oral prodrug of endogenous testosterone under development for the treatment of NASH. In this particular study, 56 adult male subjects with F1 to F3 fibrosis were randomized to either liposine alone, which is noted as treatment A, or in conjunction with D-alpha-tocopherol, which is noted as treatment B, compared to placebo with respect to safety and efficacy in a 36-week randomized trial. These data here are the 12-week data, and you can see a reduction in ALT and AST both in all subjects together, in addition to the eugonatal subjects, which is interesting. About 30% or so, 30% to 50% of the subjects in this trial were eugonatal. Here you can see over to the right, absolute changes in liver fat at week 12, again, in all patients on the top, and then in eugonatal patients on the bottom, so intriguing results that this can be effective even in patients that are eugonatal. In conclusion, nearly all individuals with NAFLD benefit from lifestyle intervention, but this needs to be supported by a multidisciplinary approach. Genetics are furthering our understanding of disease and may help provide population-level risk stratification, but not yet suitable for application at an individual level. Leveraging the microbiome has very high potential, but it's heavily influenced by many factors and more precision is needed before we can move this technology to the bedside. Multiple-targeted treatment begins with understanding the disease of the individual and individual disease drivers, including underappreciated endocrine drivers. And lastly, the identification of hyper-responders or non-responders from ongoing large trials will help us to develop a more individualized approach in the future. With that, I thank you for your attention. I'd like to thank the program chairs and the NASH SIG for the invitation to speak on this topic. Let us first start by defining some terms. AI refers to the ability of a machine to communicate, reason, and operate independently in both familiar and novel scenarios in a manner similar to a human. Machine learning refers to algorithms as statistical models that learn from label training data from which they are able to recognize and infer patterns. Now, commonly, machine learning and artificial intelligence are used interchangeably, and AI can be seen as an all-encompassing term that covers many of the techniques under artificial intelligence, such as machine learning in its different forms and other techniques such as natural language processing. There are several potential applications for AI in NAFLD and NASH. These include the identification of patients with NAFLD, NASH, or fibrosis in electronic healthcare systems or on different imaging modalities or looking at metabolomic or lipidomic data. Emerging applications include the prediction of response to therapy in clinical trials and prediction of outcomes. The most critical and unmet need is, in fact, the optimization and maximization of data obtained from liver biopsy analysis, which has been the focus of our research and will be the focus of this talk. So whether a researcher is studying the pathogenesis of NAFLD or looking at transcriptomics or genomics of NAFLD or doing a randomized trial, we all start with imaging, but we end up with interpretation of the liver biopsy. As you know, NAFLD is broadly divided into three major phenotypes, NAFLD, borderline NASH, and NASH, with NASH being a composite and complex phenotype to diagnose. It has three necessary features for diagnosis, which are steatosis, lobular inflammation, and ballooning, but it also has two additional important features, the portal inflammation and fibrosis, which has considerable prognostic and clinical importance. So as long as liver biopsy is used for phenotyping NAFLD and is considered the gold standard, we have to understand the factors that influence its sample quality and diagnostic yield. All these factors listed on this slide, such as length, sample processing, size, and core and web, all of them indeed, in fact, impact the yield of the biopsy. However, AI currently is unable to improve any of these, and these have not been the focus of AI technology. In fact, AI main application is in the field of biopsy sample analysis, and that indeed will be the focus of the rest of the talk. So let's first talk about the current state of histological phenotyping in NAFLD. The current state is characterized by being manual and semi-quantitative. Here you see the NASH CRN scoring system, the most widely used system for NAFLD histological phenotyping. As you can see for each lesion, there are limited bins and a limited scale that ranges from zero to four, and when you use the NAFLD activity score, which is a composite score of three lesions, it has a range of zero to eight. The limitation of this limited semi-quantitative scale could be seen in a hypothetical situation where two patients participating in randomized trials are evaluated. Patient A loses 5 percent of that at the end of the study and is moved from grade two to grade one, whereas patient B loses 20 percent of their macrosteatosis and remains in the same grade. Another limitation relates to reproducibility of the histological assessment. We have looked at this early on with our studies and noted that the intra-observer agreement was best for detecting and quantifying steatosis and fibrosis, but even the intra-observer agreement begins to drop when it comes to the cardinal lesions, lobular inflammation and ballooning, and it becomes in the fair range and becomes modest for the diagnostic classification of NAFLD. The same pattern holds true when you look at the intra-observer agreement and you see agreement is best between pathologists on steatosis and fibrosis, but it really is in the poor range for lobular inflammation, ballooning, and diagnostic classification. These findings were shown by many other groups, as summarized in this slide, and as you could see, overall agreement between pathologists is best on steatosis-grade and fibrosis staging, but it's really poor and not adequate for identifying and quantifying inflammation, ballooning, and diagnostic classification. To solve this challenge, we first tried two interventions. The first was educational, where we used the classical slides kindly provided by David De Kleiner, showing classical images of NAFLD lesions, such as ballooning and inflammation, and we went over the diagnostic criteria for each. We also provided a scoring sheet that has simplified written diagnostic criteria for each of the lesions and NAFLD classification. We then repeated the experiment with the two academic pathologists and noted non-significant numerical improvement in most of the NAFLD lesions for the intra-observer agreement, except for ballooning, which was significant. The rest of the improvement was not significant, but for the intra-observer agreement, in fact, we did not observe significant improvement, and in some cases, we witnessed deterioration. This study provided us with the impetus to consider AI and machine learning as a potential tool to solve these problems with biopsy analysis. If one were to use machine learning and AI to detect and quantify lesions in NAFLD, what are the top priorities? I think those are the cardinal NAFLD lesions, which include macrosteatosis, lung inflammation, hepatocyte ballooning, and fibrosis, but it's also important to detect and quantify portal inflammation because of its prognostic significance, as well as because it is the most common type of inflammation in pediatric NAFLD. Another important thing to pay attention to is that quantification of collagen area, a proportionate area, is not adequate by itself for assessing fibrosis. This has to include the fibrosis architectural pattern as well. So if we were to quantify automatically these lesions, there are overall, in broad terms, three main approaches. First is machine learning or AI-based. Others are algorithm-based. And then the third category is other techniques or methodology that are falling out of favor and are not seeing much development. So with machine learning-based methodology, what's quantified is the NAFLD lesions themselves directly, direct identification. They use the standard stains in the practice, and beside the digital scanner and the software, no additional equipment is necessary, whereas for the algorithm-based tools, they quantify correlates with NAFLD lesions. They don't quantify the lesions themselves. They can use unstained or stained slides, and in addition to the software, they could use special equipment such as a second harmonic generation microscopy or two-photon excitation fluorescence microscopy. Here you see the general approach to developing machine learning models. It all starts with the scientist's team, which includes computer scientists, pathologists, and hepatologists, which selects a panel of liver biopsies that are then digitally scanned. The final images are provided to pathologists, who in turn provide annotations, and the annotations are then used to develop models and undergo internal validation. A common step is correlation of the quantification of the model with pathologist scores, and the final step is typically external validation of the model performance using unlabeled data. With supervised machine learning, which we used in our research, the machine learning classifier is built using labeled learning data as well as feature vector for each lesion. The feature vector is a composite of the lesion attributes, which include things such as size, color, shape, texture, surrounding structures, et cetera. Then the classifier is exposed to the lesions of interest on the slide, and then it makes the predictions, and then we calculate the model performance. Compare that to deep machine learning and neural networks, which relies on the use of nodes, which are mathematical models that find the best attributes of the region, and then it connects in a manner similar to human neurons with other nodes. This connection that are similar to the neurons is behind the nomenclature of neural networks. These networks could be composed of different layers that could be in the thousands of nodes, and some of these are laid deeply, and that's where the name deep machine learning networks come from. Importantly, how each node identifies the features is unknown, which is a factor referred to as the black box. Compare that to the annotation attributes or feature vector with supervised machine learning, where the attributes contributing to the classifier and identification of each lesion are known. Here you see the process for identifying the central vein. Different filters and different thresholds and scaled representations are applied, and different features of the lesion of interest are used to identify it. So following that, internal testing or validation is performed. Here you could see the two internal validation methods we used, which are essentially tenfold across validation and the one-off approach. Here you see an example of hepatocyte ballooning, where the model is exposed to the digital image. It applies a tile probability, then different thresholds, then it yields the results as to what it thinks is hepatocyte ballooning on this image. The next step is that the image is divided into tiles of equal size, and then each tile is classified as ballooning or no ballooning, and then the percent ballooning on the slide is calculated by dividing the total area of the ballooning tiles by the total tissue area, and the same approach is applied to different navel lesions. When we first started this research, we thought it was foundational to identify liver microscopic anatomies, and all these appear to be white regions on the H and E stains. For example, central vein, which identifies the lobular or central area of the lobule, and the portal area identified by the portal vein, portal artery, and bile ducts. These are foundational tasks to classify steatosis, inflammation, or fibrosis location in the lobule. Our classifier showed excellent ROC across the board, but its precision was best for identifying the bile ducts. It has good sensitivity for identifying the bile ducts and also good precision and recalls for identifying the portal vein. Those are two microscopic landmarks of the portal veins, of the portal tracts. We next developed a classifier for macrosteatosis, and as you can see here, the percent of steatosis calculated by the classifier correlated really well with the grade of steatosis given by the pathologist. It achieved an R-square of 93% for the correlation. The classifier would look at the image and label the area, as you see here, as macrosteatosis, and it will also give a precise number, for example, 25.6%, as compared to the semi-quantitative 0 to 3 scores. We next developed a classifier for ballooning, and as you can see, the association of the percent ballooning with the average of pathologist grades was fair, not great, and the R-square was only 49%. The model has excellent ROC for discriminating ballooning from non-ballooning on the slide, and really excellent precision for identifying the balloon cells, but only fair sensitivity for detecting them. We next developed a model for lobular inflammation, and as you can see, the percent lobular inflammation did not correlate well with the average of the pathologist grade, and the R-square was only 17%. Despite that, the discrimination of lobular inflammation from areas that are not lobular inflammation had an ROC of over 90%, but really with modest precision and poor sensitivity as well. These challenges will be seen later in the other slides with other groups as well. We next developed a model to assess fibrosis automatically, and as you can see, the collagen proportionate area calculated by this model correlated very well with both study pathologists and better with one study pathologist than the other. An important note to be made here is that for any collagen proportionate area, if you take, for example, 17%, that could represent stage 2, stage 3, or stage 4 fibrosis. That's why a collagen proportionate area alone is inadequate to assess fibrosis globally. Therefore, we developed a classifier that can identify the architectural type of fibrosis in these biopsies, and as you can see, the performance of our classifier was excellent with ROC over 90% for distinguishing normal patterns of fibrosis as well as advanced fibrosis, and its performance was modest to good for fibrosis levels in the middle. Now other groups have also used machine learning to look at NAFLD histology and automated. In this example from a study from the UK, you could see that the percent inflammation showed correlation with the average of the pathologist's grade, and same thing for ballooning. The correlation with the pathologist's grade was excellent for steatosis but poor for inflammation and modest for ballooning and fibrosis. The AU ROC for CPA for advanced fibrosis was 82% in that study. Deep machine learning was recently used to automatically assess NASH histology. As could be seen here, the automatically assessed steatosis showed good correlation with the pathologist's grades, whereas deep machine learning assessed and quantified liver inflammation showed poor correlation, whereas the correlation was modest for ballooning and fibrosis measurement per NASH CRN system, and the system had excellent correlation with the fibrosis as measured by the iShack fibrosis system. The study also assessed the accuracy of the deep learning model on unseen held-out annotations, and as you could see, for hepatocyte ballooning, the accuracy was 67%, for steatosis, 80%, for lobular inflammation, 74%, and for portal inflammation, 81%. For fibrosis, the accuracy was performed on the trichrome images, and that reached 90%. The same study evaluated the performance of the deep machine learning quantified features of NASH as a predictor of heart outcomes in participants in the STELR 3 and 4 randomized trials and compared it to performance of semi-quantitative annotation or quantification by central pathology. As you could see, machine learning had slightly better numerical, but not statistically significant, better performance than central pathology, but both of them had overall, at most, modest performance for predicting heart outcomes. In this study, an automated algorithm, Q-Fibs, relying on second harmonic generation microscopy was developed for correlates with NAFLD lesions, and as you could see, the Q-Fibs had excellent discriminatory ability for steatosis and good to excellent discriminatory ability with AUROC of 80% to 90% for fibrosis and good ROC for ballooning and inflammation. The correlation with pathologist scores was excellent for Q-steatosis and fibrosis, but was only modest for Q-inflammation and ballooning. So to summarize, AI has the promise to improve the current state of NAFLD identification and phenotyping. AI-based tools can provide continuous, accurate, and precise quantitative data from liver biopsies. The prediction and measurement of treatment response and prediction of outcomes are evolving tasks for AI in NAFLD, and considerable effort is still needed for further development and validation of AI-based methods using representative cohorts. With that, I'd like to thank and acknowledge all the collaborators who made our research possible, and I'd like to thank you for your attention. Thank you. This talk concludes the NAFLD Special Interest Group Program. I would like to take this opportunity to thank all the speakers for their excellent and very informative presentations. I hope you all enjoyed the program. Thank you for joining us.

Non-Alcoholic Fatty Liver Disease Symposium 2021

Precision medicine

NAFLD

NASH

Dr. Zorcoian

Disease onset prediction

Tailored treatments

Genomics

Genetic variants

Single-cell genomics

Novel techniques

Machine learning and AI