All right, welcome, everyone. I hope you're here for the granolopery lecture. We're going to be talking about novel techniques in the diagnosis of ILD, and thank you all very much for the interest. I'll be talking a little bit about optical coherence tomography. Show of hands, anyone heard of optical coherence tomography? Anyone using it in clinical practice? Good, it should be none because it's not clinically available. But research-wise, anyone doing any research in it? Okay, great. Well maybe I will find some interest in what we're doing here. I'm over at Mass General in Boston. So what we're going to do briefly is just review some of the limitations. I'm sure we're probably all quite well aware of sort of where we're at as far as limitations go in diagnosing and detecting fibrotic ILD, and explain how OCT can actually be quite uniquely suited to fill in some of those gaps. And then we'll just highlight some of the images of recent trials that we are involved in. So what I want to bring up here on several of these slides are pictures of histology, because I think really understanding the concept of OCT is a little bit of understanding ultrasound, but a lot of understanding histopathology. So as you know, early detection, therapeutic intervention, especially at the pre-symptomatic stage can be very, very challenging to detect, diagnose, understand progression, and how do we really track those changes. We obviously need some type of more precise assessment. That's why we're all here. We've got some really great ideas. Limitations-wise, high-res CT, it's great. Certainly come a long way from a decade ago in detecting some smaller changes. But it can still be a little bit difficult to detect any microscopic changes. And then as a result, any of our therapies can be a little bit difficult to detect any true blips in progress or being able to halt the progression. And then surgical lung biopsy is really still the gold standard. Yes, we've made some progress, and there is interest and clinical use in using cryobiopsies, which we use at our center and many centers across the world. Even in cryobiopsy, there's associated morbidity, mortalities, obviously, with surgical lung biopsy as well. It's as safe as we can get it, obviously, but if there was a way that perhaps we wouldn't need to subject our patients to any large pieces of tissue, that would be great, right? So enter optical coherence tomography. This is just a picture here. It's quite a small catheter, and the idea behind is it's a single fiber that's actually through a wound cable. And so when you actually take the pictures, you extend the catheter, and essentially this translucent sheath outside the bronchoscope, and we reach it out to the edge of the lung, and then the cable pulls back as the images are acquired. It's analogous to ultrasound. Those that use radial endobronchial ultrasound might feel it looks a little bit similar. I'll show you some cross images, and that's really sort of what you see as you march through your initial images. But again, I'm going to go back to the histopathology, because you kind of have to make that jump. And so it's not as simple as, oh, I've got the concepts of radial endobronchial ultrasound and my pictures down from that type of imaging, and then making the jump into optical coherence tomography. The similarities of the cross-sectional imaging are helpful to understand, but the difference in ultrasound compared to OCT is that you're using a different modality in order to acquire your images. With radial ultrasound, you are transmitting sound, and then you get scatter back, and the signal that you get back is essentially your 2D image. OCT, on the other hand, uses light. And if you think about thunder and lightning, the real advantage in using the light, well, there are a few advantages, actually, but one of the advantages is that light moves much faster than sound, and so you're able to get more resolution. You get essentially a microscopic-level, cellular-level view of images that are two to three millimeters outside your catheter. The other big advantage to OCT, because it's using light, it isn't scattered by things like air. As you know, with ultrasound, if you have any air in between the tissue you're trying to image and your instrument, you are going to have some type of artifact, and OCT is not subject to that artifact. So this is a nice picture that was brought up by one of our PhD candidates that is doing quite a bit of research in our group, but that's essentially the view that you get. Kind of seems like gobbledygook as you kind of look at it this way, but what's really nice about it, like I said, the cable is pulled back. As you extend it out to the edge, you get about a eight to 10-centimeter length of information. And so then what we do is essentially collate all of those 2D images into this 3D structure. And this is kind of where I said you kind of have to make that jump, because I think a lot of us who understand endobronchial ultrasound are kind of looking at those 2D images and wanting to have some type of similarities there, but in understanding what looks like normal lung and what looks like abnormal lung, there's that jump. But I'm going to go over a little bit on what that jump looks like for learners. So again, the features histopathology, you're going to have some spatial heterogeneity looking at areas of subplural fibrosis and honeycomb change. So keep those in mind as you look at some of the OCT images. So again, I'm going to go back, show the same picture. If you can kind of start seeing now what we're looking at, all of those little holes, many of them are actually just normal alveolar spaces. So what's the accuracy of this OCT technology? We published a recent perspective study in the Blue Journal last year, and we were looking at several sites. We started out with few. We're probably up to about eight sites per patient now, and that's distributed through all lobes of the lung. We'll typically review the high-res CT before we bring the patient in, and so we've got some predetermined sites where we want to get some type of information. Things on your CT where your radiologist is pointing to this and saying, are you getting early interstitial changes? Is this already bronchiolectasis, honeycomb change? And so we're imaging all of that. And what we really looked at here was sensitivity and specificity for matching it against the diagnosis, the histopathological diagnosis of UIP. The blindedness was occurring both on the person who was reading the OCT, but also on the people who were looking at the pathology. And so many of these were surgical lung biopsies. There are some cryobiopsies mixed in with that as well. This is an example of a lung that has, quote unquote, preserved architecture. So all of that signal, those are pretty much just alveolar spaces. Contrast that to a typical, quote unquote, typical UIP picture, and you can see there's quite a bit more signal. And you're able to pick out some differences here on, for example, honeycomb change. And I think us, as first looking at it, I still look at it with my pathologist as we're taking the images. She's very good at picking that out and saying, oh, I see honeycomb change. I see normal alveoli. I'm still probably on that learning curve, I would say. Probably I do not all of those cases, right? I think we selected about 25, 27 people. I probably did about a quarter of those. So I'm learning some of the differences, what looks like mucus, those types of things. But I'd say I'm very much still on the learning curve. This is another example of something that's a little different, airway-centered fibrosis. And we can also see things like branch points, which is going to become important in some of our later research, what we're working on right now. And those are essentially the images we get in real time. So I'm not moving the catheter, it's the cable that's moving back. You basically extend the catheter out, you park it as far out as you can get. We kind of get a general sense on what looks like or what feels like a short airway segment. We try to get the longest airway segments. Right now, we're not using any fluoroscopy confirmation for how far out we are, but we get general senses of how much alveoli and smaller airways, apart from the conducting airways, how far we're getting out there. So the results of this study showed very excellent concordance with the pathologic diagnosis of UIP and the ability to diagnose it using the endobronchial OCT. And then, remember, I talked a little bit on what our learning curve really is. The people that were trained for, pathologists who were trained for reading the OCT images once they were all acquired, underwent a three-hour training session and they were able to, half of it was done with just training and then half of it was done with testing with all of the different types of patterns of ILD. And then they were asked to independently evaluate some type of test of one of the interstitial lung diseases. And they were tested on whether they were diagnosing UIP or a non-UIP pattern. And they all did extremely, extremely well. Obviously, these are only three people that were tested, and the validity would obviously have to be tested on a larger population of people. But two out of the three, at the end of that training, were essentially making the same diagnosis that they were with other people, the primary investigator who's making the diagnosis. So now we're looking at, can we use this to look at disease progression? And the pictures here are really just to show you how intricate we need to be in making the map. Because, like I said, right now we're not using any fluoroscopy for trial purposes, fluoroscopy confirmation. And it can be a little difficult. We're not going to get CAT scans intraoperatively. Seems a little aggressive for some of these patients. So what we're essentially doing, because the catheter is somewhat stiff, we are taking pictures of exactly the way the bronchoscope is oriented, and then exactly what division the catheter is going out of. So I've got some obvious markers. You can see there, that's a distal BI. We can pretty much identify that, and you can identify your ALP images. And then I take another step where the next branch, I'll take another picture. So I can essentially leave myself picture breadcrumbs, so that when we go back in six months, we're able to go down that airway again. And it's actually worked quite well. This is an example of an airway where we, at the top, did a baseline image, and then down at the bottom. And you can actually match up the branch points. And this is just an example of, all this is unpublished data right now, but this is an example of an airway that we believe we matched up. And in the top image, there's really not much in the way of any early changes. And the bottom, we felt we were able to get a detection of early honeycombing. And that is pretty much everything I wanted to go over. Any questions? Do we have time, or do we just get the end? Okay. So we'll just do questions at the end, so save them, and we'll move on. Thank you. There you go. And then, how do we go to you? There you go. There you go. Perfect. Cool. Thank you. All right. Thank you very much. That's really, really impressive technology. So I'm going to be talking to you today about deep learning-based analysis of chest CT. So hopefully we can kind of come to an understanding about what this is, and how maybe in the future we'll be using it to diagnose and give prognosis to our patients. So I'm Brad Bemis. I am Associate Professor of Medicine at Loyola in Chicago. And I'm the co-director of the Interstitial Lung Disease Program there. My only disclosure is that I'm on the Speaker's Bureau for Berenger Ingelheim. So today we'll talk about a couple different things. So we'll learn about how a CT scan can help us to obtain a little bit more of a precise diagnosis by the use of artificial intelligence. We'll learn about how to predict prognosis based on these HRCT results, and then can consider how we're going to apply these tools going forward, both in clinical practice, but also in clinical trials. So quick audience response. So pull out your smartphones and open up that chest app. And so kind of an interesting question. So what role does chest CT play in your practice? So do you use a chest CT just to diagnose ILD, to use it to diagnose and then regularly follow up patients with interstitial lung disease every one to two years perhaps? Or just use it intermittently with symptomatic worsening, symptomatic decline rather? So should open for you guys here. All right, perfect. That's a really interesting split. We got about a 50-50 on whether we're getting it regularly or whether we're using it intermittently. And I think that's a great capture of where we are right now today in the science of interstitial lung disease. So chest CT undoubtedly has a major role in our management and diagnosis of interstitial lung disease. It is a central tenant of our guidelines for the diagnosis of ILD, okay? It clearly has that role already, but what roles might it play in the future with quantification and prognosis related to interstitial lung disease? We already use it a little bit, right? So you guys were just in the previous session about progressive pulmonary fibrosis. We're using CT to determine which patients are progressing already based on the in-build trial. If somebody had progressive disease based on a worsening amount of fibrosis on their CT, they were eligible to be enrolled in this trial, and that leads to management changes for our patients and potentially new diagnosis of progressive pulmonary fibrosis. But HRCT can be a more useful tool. So we don't always have access to a multidisciplinary discussion, or we don't always have access to a kind of expert pulmonary fibrosis foundation care center network. We see big swaths of the country where this just doesn't exist, where maybe there's less than expert thoracic radiologists that are able to help diagnose our patients with UIP or some version of UIP that might lead to a change in prognosis. And even if we do have an expert thoracic radiologist, we know that maybe that's not the most useful tool, that there's not perfect agreement between one thoracic radiologist versus the other. This clinical trial looked at 96 thoracic radiologists that had varying levels of experience from over 20 years of experience to even just fellows that were looking at these high-resolution CTs, and they looked at 150 CT scans that had different diagnoses of IPF, NSIP, autoimmune disease, or chronic HP. And they were asked to say, hey, can you tell us, is this definite, possible, or inconsistent with UIP? And ultimately, our agreement wasn't that great. So using a capital statistic, we know that one is perfect and zero is poor, and maybe we're okay with 0.4 or so, which is about a moderate amount of agreement. And we can see that even amongst the thoracic radiologists with over 20 years of experience, the capital was only 0.48, so not ideal. And so between thoracic radiologists, maybe they're not even perfect all the time. So what if we add computers and add artificial intelligence to the scenario, and we can see how maybe Terminator 2 does with diagnosing our UIP. And so we have this CT scan that's maybe not clearly one diagnosis or another, and what if we feed it through artificial intelligence? So artificial intelligence is kind of this broad terminology that we kind of all understand a little bit, but maybe not all the way, that there's each of these different layers of ways that we can use artificial intelligence to look at data and to give us some answers. You can see on the right here, this is one of the early studies in artificial intelligence recognizing zip codes based on handwriting in Buffalo, New York, that the AI was able to correctly identify which numbers these look at. I don't even know if I could tell you what numbers those are, but an AI was able to correctly identify that through use of kind of a complex neural network. And so we have kind of an umbrella term of AI with kind of a subcategory machine learning, and then deep learning is what we're gonna be talking a little bit more about today, where deep learning uses these complex neural networks in order to kind of replicate a human brain, so to speak, and identify different patterns within the data that we feed it in order to answer questions for us. So to think about that a little more. So these are two very cute dogs, good boys. And so we might tell a machine that a dog is something with fur, it has four legs, it has a tail, it has floppy ears, maybe, and maybe they aren't always floppy. And so we tell a computer to say, okay, well, this is what we want you to define as a dog. And in machine learning, kind of that little bit term just above deep learning, the machine will look for those characteristics. It'll say, okay, well, you've told me what fur is, let me go look and see if there's fur, let me see if there's legs, tails, and floppy ears. But what if you get kind of the outlier, right? So we have this outlier here that does pass as a dog, but it doesn't quite fit all of those characteristics, and maybe it doesn't look exactly the same. And so that's where deep learning comes in. So you feed artificial intelligence each of these pictures and say, this is a dog, the AI will train itself. It'll be kind of something we call unsupervised learning. It'll try to identify different characteristics within the data that we feed it through these complex neural networks of perceived patterns that maybe we as human beings can't directly define. We can't say, well, this is a dog because, but it'll learn those things just due to the way that we learn these things, and we can identify that as a dog. And so these complex neural networks look somewhat like this in a very simplified pattern. We have this input layer where we will tell the machine this is a dog, okay? And then there's these hidden layers where the machine does its own work, assigns different weights to the likelihood that this characteristic or that characteristic is a dog as well. And then it'll ultimately spit out in the output layer, dog or not dog, right? Okay, and so then we, is that correct? Is it not correct? And the machine will go back and forth through processing forward and backward propagation, and will identify over time, kind of learn over time what are the characteristics of a dog. And so similarly, we'll stop talking about dogs, unfortunately, right now. We can start talking about UIP. And so similarly, we can feed a machine and feed artificial intelligence this input image, and we define it as saying, okay, well, this is UIP, or IPF, or something like that. And then it will go through its hidden layers to define kind of what are those characteristics. Again, unsupervised, we're not telling it to look for honeycombing or anything like that, and then it will spit out a diagnosis of ILD or not ILD. And so one of the kind of first papers that really proved this concept by Dr. Walsh in 2018, kind of identified that this computer software can provide decision support in classifying a CAT scan as being UIP. Okay, and so these HRCT scans, they collected from a couple referral centers, they were classified according to our ATS criteria, and then based on computing restraints, which were present in 2018, not that long ago, they couldn't feed the whole CT scan to the artificial intelligence, and so they made these kind of these mosaics, these HRCT scans made up of four different slices, or montages, rather, of each CT scan, and then gave it to the computer using a TensorFlow deep learning algorithm. And so what they found was that the algorithm was pretty good and pretty comparable to the thoracic radiologist that made these diagnoses. So we can see up here our AUC for the receiver operator curve, as well as our specificity and sensitivity between the algorithm and the radiologist were very similar. Okay, and ultimately this does look similarly to our patients, too. This is a Kaplan-Meier curve looking at AI diagnosed UIP versus radiologist diagnosed, and we can see that whether the radiologist described it or whether the AI described it, they had very similar outcomes and very similar prognosis. You can see hazard ratios for each were very similar around three times increased risk of mortality for those who had UIP compared to those who didn't. So the AI was just as good as a thoracic radiologist in telling us there's UIP and there's a certain prognosis associated with it. This is an example of such of what the AI identifies as being the important parts of the CT scan. Again, we're not training it, but it identifies certain areas that maybe we would correspond to honeycombing, but maybe some other areas that maybe aren't quite clearly honeycombing that the AI identified as being important parts that made this CT scan consistent with UIP. And actually just published this month in the Chess Journal, Dr. Brett et al, looked at a very similar study, but this time used histopathologic diagnosis. This is really important, right? So we looked at, okay, well, our gold standard is thoracic radiologist said this is UIP. Now we're looking at pathology. So this path showed us that this is UIP. So they prepared the CT scans, fed it into this EfficientNet B3 deep learning computer and said, okay, this is UIP or this is not UIP based on histopath standards. And they tested this against what our thoracic radiologist might say. And you can see this is enriched for people who have these kind of alternative diagnosis and kind of unclassifiable types of ILD, right? So we have definite UIP, probable, indeterminate, and alternative diagnosis. So it was enriched for the toughest cases that we might even see in our patients. The receiver operator curves are here. You can see AI did just as good visually, but if we look at the kind of receiver operator curve, the area under the curve was even better for deep learning or AI based reads on these CTs compared to human beings. So this is kind of termed a superhuman result. These are kind of above and beyond what a thoracic radiologist might be able to classify. This is kind of an example, kind of an interesting case where the thoracic radiologist said this doesn't look like UIP really at all. And the AI said this is UIP and it was confirmed by PATH as being UIP related fibrosis. So again, histopath standard, they used bigger amounts of data because computing power has gotten a lot better these days than it was in 2018. And so this is a really interesting study just published. So CT plays a role in disease behavior and prognosis as well. This is a patient of mine who had CT scanned three years apart. And I don't know if I'm looking at this, I could tell the difference between these two. Maybe there's slightly more reticulation or something like that. And maybe a radiologist would say there's slight progression, but ultimately can we do a better job than comparing? We've all, I'll pull this up, we've all done this in clinic, right? Pull one CT scan up and then the other CT scan up and scroll back and forth. You're like, yeah, it kind of looks the same. It looks all right. And then maybe there's a better way, right? Than just looking at a CT scan for a pulmonologist in clinic who admittedly is not a thoracic trained radiologist. So there's a lot of ways that we can look at this. So kind of dating back to the early 2000s, we looked at different amounts of Hounsfield unit expression in these patients. And we could see that, okay, this is a normal CT scan that has a nice big peak here and we're leftward shifted. So to these negative Hounsfield units versus a patient who has UIP, maybe has a smaller peak and it's shifted a little bit more rightward. And this is important because this does correlate with FVC. You can see on our table five that we've copied and pasted here that things like kurtosis, the peakiness and the skewness, the rightward deflection do correlate with changes in FVC. So it is a real finding. And ultimately this does correlate with an increase in mortality. When you have a kind of a more normal looking peak and a more normal leftward shift, you have a better survival compared to if you have a more rightward looking peak or a smaller peak or a rightward looking curve. And so this has obviously been improved upon and one of the kind of first models looking at machine learning was this AMFM type model called Adaptive Multiple Feature Method. This was looking at, again, identifying what's a tail, what's fur, what's floppy ears. We had expert radiologists assigned to honeycombs, ground glass, reticulation, emphysema. We taught the machine how to do that and then it came back out and spit out a score. And so looking at the primary endpoint of 10% decline of FVC, hospitalization or death, ultimately the AMFM model did predict a worsening outcome. So you can see the hazard ratio increased for mortality in those patients. Again, the Kaplan-Meier curve there too. Kind of more getting into the deep learning model, though, Caliper is a tool that's been used and again, these are all kind of not ready for prime time, but getting close. This computer-aided lung informatics model that looked at each pixel, identified features of ground glass, reticular, honeycomb change, attenuation, but more importantly, pulled out this pulmonary vessel volume which correlated with mortality. This is kind of a new finding that the AI was able to diagnose that humans hadn't done before. This Caliper tool has been used as an endpoint in clinical trials too. So we've kind of identified this, when you have a worsening score of fibrosis based on this model, this identifies progression and we're gonna power a trial based on whether you're having progression on this deep learning model. And you can see these are, in the panel A here, these are GAP scores, so differing GAP scores obviously correlate with survival and this is Caliper scores that mirror that almost perfectly. Kind of the truly unsupervised learning model that's out there right now, data-driven textural analysis, looked at regions of interest of normal lung versus UIP lung. These again are identified by the model, not by humans. And this DTA score, rather, corresponds to a decline of FEC and DLCO. A 5% change in the DTA score is correlated with a 5% decline in forced vital capacity. And again, similarly, you can see that changes in DTA score correlate with an increasing or worsening mortality in our patients with interstitial lung disease. So what role is deep learning gonna play? So I think HRCT is definitely an untapped resource. The smart man once said, we should use the tools that we already have at our disposal rather than getting new tools for us, right? So we already have our CT scan and what if we leverage that to identify our patients that are progressing, identify our patients that maybe can be diagnosed with UIP earlier than we can based on the visual tools that humans can perceive. So it will lead to earlier recognition of UIP and ultimately might play a role in giving our patients a prognosis as well. And probably in the near future, will be a primary outcome in clinical trials defining treatments for interstitial lung diseases with a progressive phenotype as well. All right, thank you very much. That's my talk. This is Dr. Newton. Thank you. So let me pull it up here. So I'm actually gonna talk about some new tools. Yeah, so thank you all for coming today. My name's Chad Newton. I'm assistant professor at UT Southwestern in Dallas. And today I'm gonna talk about using omics or omics-based technologies in the diagnosis and management of patients with interstitial lung disease. So learning objectives today from my talk, we're first gonna discuss the use of OMICS and how it's used in clinical trials. So learning objectives today from my talk, we're first gonna define omics. What do we actually mean when you hear the word omics? And then I'm gonna spend the majority of the time actually talking about how we could potentially apply some of these omics-based technologies in caring or diagnosing or managing our patients with interstitial lung disease across a few different domains. So omics, first of all, what is it? So really it's a high-throughput biochemical assay. It's a tool, right? It's an assay that can comprehensively and simultaneously measure lots and lots and lots of molecules, all right? So there's several different types of omics-based platforms. So you have epigenomics, genomics, transcriptomics, metabolomics. And recently we have these single-cell technologies that are able to really dig into a specific cell within an organ system to figure out their profiles compared to other cells in the same organ system. And so this really provides potentially kind of this holistic and potentially unbiased view of an entire biologic system, especially when you start layering some of these technologies in with each other, like epigenomics layered with genomics and transcriptomics, you really start getting, in theory, a full picture of what's going on within the cell or within the organ itself. So what does this actually look like in practice? Well, first you have to have some sort of biologic sample, right? So in patients with interstitial lung disease, we have lung tissue that we could sample, we have blood samples, we have BAL fluid, we have excelled condensate, right? So there's lots of potential biologic samples. Preferentially, these should be coming from patients, though, right? If we're wanting to develop additional knowledge toward how we're gonna care for patients, we need samples coming from patients. Once we have those, we isolate stuff, right? We isolate RNA, DNA, proteins, metabolites, and then we put it through these omics-based platforms, and it generates kind of this huge cloud of data, right? This very complex, multidimensional data set that is really too complicated for the human brain to be able to look through and identify associations. So we use these high-capacity computing systems that not only identify but can quantify associations between different molecules and pathways, and even between those molecules and our phenotype of interest. And in doing that, we actually generate knowledge regarding pathobiology, so mechanisms of disease. And through that same process, we can use that information gained from the mechanism of disease to generate what we're calling novel biomarkers, so actually seeing molecules that help inform our decision-making process. And how that looks is, if you think about interstitial lung disease or pulmonary fibrosis, it's really a collection of heterogeneous disorders, right? All these patients have scarring or inflammation in their lungs, but they kind of look different, right? They're all over the place. And so you can use these omics-based platforms to really deconvolute this heterogeneity and form smaller groups of patients, or called endotypes, that are based on a really distinct biologic process, right? So now we've taken this heterogeneous group and we've broken them down into similar, smaller groups, and then we can use that information to decide how to manage them or how to treat them. So we can, if we start thinking about biomarkers, we have to think about how we could potentially use them, right? Biomarkers is almost like a loaded word. Everyone throws it around, but what does it actually mean and what are we gonna do? So if you think about it, there's not gonna be one biomarker that really tells us exactly how to treat or manage a patient, right? It's probably gonna be a series of markers or even dozens or hundreds of markers that we're gonna use for different reasons in caring for our patients. So in those that actually don't have interstitial lung disease yet, right, we can find biomarkers to help us figure out their susceptibility risk or even detect disease earlier than what we could with our standard practices. In people that have established disease, we could use biomarkers to help classify their diagnosis, right, do they have IPF, do they have HP, do they have something else I didn't even think of? What is their progression risk, what's their prognosis, and what medications could I use or not use in patients with this specific endotype? So the entire purpose of this, or this kind of well-rounded multiple domain of biomarkers, in theory, gives us the ability to practice what's called personalized medicine or precision medicine, where we're taking individual patients with this very heterogeneous kind of collection of patients and using very directed treatment algorithms for that one person that's sitting in front of you based on their underlying biology. It's kind of a grandiose idea, but in theory, that's what omics and other platforms like these CT algorithms might help us to do in the future. So let's actually talk about some of these domains. So the first we're gonna talk about is ILD susceptibility, so using omics or omics-based platforms to determine predisposition risk. So really, this all boils down to genomics, right? So the Human Genome Project started a couple of decades ago. The entire goal of that was to sequence the human genome so that we could figure out what diseases someone is gonna develop when they get older or as they progress through life, right? And then we're gonna use that information to come up with ways to treat them. And although it's probably a lot more nuanced than that, some of those tenets still hold true. So if you look at interstitial lung disease, over the last two decades, we have found a multitude of genomic variation that really starts to nail down some of these endotypes of patients with ILD. So they kind of come in two flavors. We have common variants, we have rare variants. Common variants are common, right? So people without interstitial lung disease actually have these risk alleles in the common variants. And then you have rare variants that are not present in a general population. They're really only present in families or patients with disease. So we can use some of this information to start trying to dig down into susceptibility risk, specifically in looking at the rare variants. So this is a study of about 950 patients with IPF and familial pulmonary fibrosis where they did whole genome sequencing. They found about 14% of this population actually had a pathogenic rare variant. And when you looked at the familial pulmonary fibrosis subset, 25% of those patients harbored one of these variants. So a quarter of patients with a family history actually may have one of these variants. So using that information, if we had kind of a hypothetical family here, right? And we have this female here coming to your clinic and saying, I've been diagnosed with pulmonary fibrosis. My dad and my sister both died from pulmonary fibrosis and now my son is wondering what is his risk of getting ILD? So we can use what we know about genomics to start answering some of those questions. So we could do sequencing for her. And let's just say we found a pathogenic variant in this TERT gene, which is the most common rare variant gene to be implicated in familial pulmonary fibrosis. We could actually test the other potentially at-risk individuals, so blood relatives from our patient. And if they are found to have that same TERT variant, then we can tell them that they are at higher risk of developing disease at some point in their life, right? So this is already available to us. So we can do genetic testing in our patients with familial pulmonary fibrosis and their relatives. And this can help refine susceptibility risk for individual patients. So it's almost like you're providing personalized medicine, at least for this family. So now let's talk a little bit about ILD classification. So again, looking at genomics first, I'm showing here kind of an aggregate of data from lots of studies looking at the MUC5B common variant. And so on the x-axis here, we have the risk allele carrier frequency. And on the y-axis, you have a bunch of ILD diagnosis. What I'm showing here is that across ILD diagnosis, you can find individual patients who have the risk allele in the MUC5B SNP. So it tells you if you have a patient sitting in front of you, checking for this SNP or any other common variant isn't really gonna help you for sure put them into a bin, right? Because they could still fall into any of these other categories. The same is true for these rare variants that I spoke of a while ago, the telomere rare variants and the surfactant variants. You can find patients across a wide variety of interstitial lung diseases that harbor these variants. So using genomics alone, or at least a single variant, or the variants that we know of now, doesn't really help us put patients into a specific ILD diagnosis bin. So what about transcriptomics? This is a heat map from a study from almost 20 years ago, and I love it. So this, Moises Selman and colleagues actually took lung biopsy specimens and did an RNA microarray transcriptomic analysis. And we're able to easily differentiate IPF from HP. So it showed us, almost 20 years ago, that if we look at lung tissue and RNA-seq signature, or RNA signature from lungs, we can differentiate at least some ILD diagnostic criteria. But we don't do surgical lung biopsies on that many patients, right? A lot of people now either don't want it or don't qualify for a surgical lung biopsy. So now we have something, though, called the Invisia classifier. And this is based on small tissue samples. So this is transbronchial biopsies. And this classifier uses an RNA-seq machine learning algorithm to identify a subset of genes that readily differentiate usual interstitial pneumonia from not usual interstitial pneumonia. So it's binary. It's either present or it's absent. And importantly, when you layer this in with high-resolution CT patterns, the test characteristics are actually quite favorable. So you have an AUC of about 0.86, and then you have a sensitivity and specificity of over 90%. So the caveat here, though, is that Invisia is able to improve our confidence in a UIP classification, specifically in patients where we suspect they have IPF. So if we're suspecting they have another cause of ILD that could have UIP, we're still not exactly sure how this performs in that subset, right? So in a specific subset or endotype of patients, this can be helpful. And because of it, it's actually been approved by the FDA and is clinically available now. So what about looking at ILD risk, or progression risk and prognosis? So again, going back to the genomics, we've known for a while now that rare variants in these telomere-related genes across a variety of interstitial lung diseases are associated with worse survival, and it's also been shown by multiple groups. So the presence of one of these rare variants in a telomere gene kind of tells you that the patient is at risk, or has a higher mortality risk. The same is true if you extrapolate that to telomere length. So looking at leukocyte telomere length, shorter leukocyte telomere length, which is sometimes caused by these pathogenic rare variants has been associated with worse outcomes, poor survival. Again, in a wide variety of interstitial lung diseases by several different groups in patients with different ethnic backgrounds. So this seems to be a fairly consistent finding from a genomic standpoint in assessing progression risk and prognosis. These same variants have also been associated with rapid decline in lung function. And looking at telomere length either by qPCR or by whole genome sequencing, it's also associated with rapid progression in ILD. So there's kind of a common theme here, but the issue is when we start trying to translate this into clinical practice, right, from the research realm to actually using it at bedside with our patients, we run into some questions, right? Who should we be testing this for? What populations, right? Is it every single person who has a little bit of scarring in their lungs, we should be measuring these things? It's probably not cost effective. If we do measure them, what thresholds are actually, should we be making our decisions based on, right? Is it a telomere length less than 10th percentile? Is it a quartile thing? We still don't quite know. And even if we knew that, what are those test characteristics of those thresholds, right? Can we actually sink our teeth into a positive predictive value, a negative predictive value, so that we know what to do once we get that information? So this is an active area of research, and because of it, we still are looking for some of these genomic markers to kind of make it to bedside, but just not quite there yet. So there's a transcriptomic analysis that have looked at survival also. So this is the Yale group, several years ago, took peripheral blood cells, measured the, or extracted RNA, and did RNA microarrays, and found a 52-gene prognostic signature that was associated with survival in patients with IPF. And then importantly, they took that signature and they replicated it in six cohorts, and it replicated perfectly, or almost perfectly, as perfectly as these things do, right? So it looks like this is a consistent signature, at least in IPF patients, across several different cohorts. But the question is, again, who should we be using? Is this something that we could potentially do for everyone? And we don't quite know yet, so it's not available clinically. But I think something that's even more exciting when you look at transcriptomic analysis is in gauging progression risk in disease activity. So this is a study published last year, where they took the Comet clinical trial cohort, which collected blood at multiple time points throughout a year during the trial. And the investigators took RNA from peripheral blood cells at baseline and at time four months, and they looked at the change in a gene expression profile over those four months, and they correlated that to near-term FEC decline. So this is really a measure of IPF activity. We were gonna measure something in their blood that's changing over four months, and that's gonna predict how they're gonna behave over the next eight months. And they were able to find 25 genes that correlated with FEC decline. And I think one of the most interesting parts of this study is that the majority of these pathways are in kind of inflammatory and immune pathways, as you would expect, right? We're looking at peripheral blood cells. But some of these pathways are actually fibrogenic and fibrosis pathways, which is fascinating, because if you're assessing a blood cell, and yet you're seeing fibrotic pathways that are activated from the lung, you know, it kind of starts to really open up a whole nother set of potential analyses and studies where we can use the blood to kind of peer into what we think is a lung-dominant disease. We can also use proteomics to gauge progression risk. And now we're stepping away from IPF and actually looking at the other huge group of patients called non-IPF ILD patients, right? These are HP patients, CTD, unclassifiable. So Justin Oldham's group actually took plasma from about 600 patients, and they were able to identify 31 proteins that were associated with progression over the next year. 17 of those proteins validated in a replication cohort. And so they used those 17 proteins to generate a regression-based risk score that dichotomized patients in low risk and high risk. And importantly, they published the test characteristics, which again, we need that to actually make decisions and determine whether this test is useful or not, right? So what they showed is that you had a sensitivity and a negative predictive value of about 90%, which means if you had a non-IPF ILD patient sitting in front of you, and in theory, we could do something like this for them, if you ended up categorizing them in a low-risk group, they had about a 10% chance of progressing over the next year. So that's information that we could use if we were to translate this to the bedside to actually make decisions, right? This is a very low-risk patient. Over the next 12 months, they're unlikely to progress. Maybe I'm not gonna start meds right now. Maybe I'm gonna keep following them. Maybe I'm not gonna push the doses of meds up and make them have side effects if their risk of progression is low. So these are ways in which we could potentially translate some of these discoveries to the bedside. So the last domain I'm gonna talk about here is medication selection, which I think is something that's really emerging over the last couple of years. So there was a study published by Justin Oldham's and Imran Nath's group several years ago that looked at the pharmacogenomic interaction in IPF patients between N-acetylcysteine and this TALiP SNP. And they found that patients who had this TT genotype of the TALiP SNP actually had favorable outcomes when exposed to N-acetylcysteine, but the other groups, the other genotypes didn't. And so importantly, that's actually Spergo, called the PRECISIONS clinical trial that's actively enrolling now. So we're gonna know over the next maybe couple of years whether using a genetic kind of enrichment process for clinical trials is actually able to demonstrate efficacy of medications that we thought weren't effective based on prior studies looking at the heterogeneous big population of IPF patients. And even if this trial is negative, right, it still is a proof of concept that we could do this. We can enroll patients like this, even though IPF is a rare disease, we can look at subgroups of those patients and study them in clinical trials. The other medication selection part from a genomic standpoint is that we actually published a couple of years ago using IPF patients that those, IPF patients with telomere length less than 10th percentile, so short telomere length that were exposed to immunosuppressive medications actually had worse outcomes. So we showed this harmful pharmacogenomic interaction, which we think is biologically relevant, right, but it's probably not clinically relevant. We don't treat IPF patients with immunosuppressive medications anymore. But Deji Adigan-Soye from Chicago actually looked at HP patients and mycophenolate exposure stratified by leukocyte telomere length and showed differential response to mycophenolate. So this is, again, tells us that we can use genomics and genomic information to try to figure out what medications we should initially start or at least try in our patients with ILD. So we've talked about a few or several different kind of omics-based analyses, but I want to stress to you that we actually, some of this is available now, right? We can use some of this information. So like I said before, we can use genetics and genomics to health-informed ILD susceptibility risk in selected populations, specifically in familial pulmonary fibrosis kindred, right? Not necessarily everyone with ILD, but in that subgroup of patients, this is informative. We can also use the Invisia classifier, right? To help improve our confidence in a UIP diagnosis in someone that we are suspecting of having IPF, all right? So these are available to us now. What's not necessarily available to us is how could we potentially use omics for ILD progression and for prognosis and even for medication selection. But over the next few years, I think we're gonna start having a little bit more momentum in this direction where we're gonna be able to know whether we have signatures, we can use signatures, these are validated or not validated, and we get test characteristics to know how to actually incorporate those into our clinical practice. And so I think over the next few years, it's gonna be fairly exciting in ILD and hopefully we'll end up being able to use these machine learning algorithms and deep learning algorithms on CT and OCT to really push the field forward. So with that, I thank you and I think we'll take questions for everyone here. Thank you.

optical coherence tomography

omics-based technologies

interstitial lung disease

fibrotic ILD

disease progression

biomarkers

prognosis

telomere-related genes

blood samples