Good morning. It is my pleasure to introduce Dr. Danny Macaulay, who will be presenting this year's William Shoemaker Honorary Lecture. Dr. Macaulay is a consultant and professor in intensive care medicine at the Regional Intensive Care Unit at the Royal Victoria Hospital and Queen's University of Belfast, and undertook his training in Belfast, Birmingham, London, and San Francisco. He's a program director of the MRC-NIHR Efficacy and Mechanism Evaluation Program, and has several research interests, including acute respiratory distress syndrome and clinical trials. Today, Dr. Macaulay will present ARDS from treating a syndrome to identifying modifiable traits. Please join me in welcoming Dr. Danny Macaulay. Thank you. Many thanks to MAS for the very kind and generous introduction and thanks to you and all the organizers for the invitation. It genuinely is an honor to present here today. I guess it's also worth highlighting at the start that as I tell the story of moving to modifiable traits from a syndromic approach, that this reflects a huge amount of work from a large number of people. So although I'm presenting today, it reflects many other people's work. So first off, just to highlight my disclosures, most of which are not relevant to the presentation today, but I'll just highlight the collaboration with Randox to develop a phenotyping assay that we've been working on to enable a precision medicine approach in ARDS. And I'll mention that as we go through the talk. So over the next several minutes, what I'd like to do is review where we are in terms of the current pharmacological approaches to ARDS and why I think there's a need for a more precision medicine approach. I'll talk about phenotypes or modifiable or treatable traits in ARDS and how we are approaching the ability to phenotype our patients in real time and then close with a discussion of some of the future clinical trials in precision medicine and ARDS. So here is my slide summarizing the current pharmacological therapies for ARDS. And as you can see, although we've been at this for over 50 years, we haven't made much progress. There is still controversy over the role of steroids, for example, described in the initial description by Ashbach. And really, we haven't made the progress that perhaps we should have. And I guess whenever we think of why that might be, one of the main reasons is that the syndrome of ARDS is very heterogeneous. And as we have moved from the Berlin definition into the new proposed global definition, it is possible that we might even be dealing with a more heterogeneous overall population. We can now include patients on the basis of an ultrasound rather than just chest x-ray and CT identified infiltrates. We include people who have oximetry as the basis for the diagnosis. And also, perhaps most importantly, as we see an emerging use of high flow nasal oxygen, we can now accept that patients meet the definition of ARDS if you're on high flow nasal oxygen. And I guess that presents the issue that we may become more heterogeneous. And I think there are some advantages to having a syndromic approach and perhaps they largely relate to more supportive therapies. And that's where we have made progress in terms of lung protective ventilation and prone positioning. But perhaps the syndromic definition is not the right approach for pharmacological therapies. And I think that's reflected in this slide, which is not a conclusive summary of all the trials that have shown no difference. But you can see here that there are many trials that have been undertaken in this heterogeneous population with limited progress. And perhaps the approach of reducing heterogeneity of the overall population might solve the lack of pharmacological therapies. And this cartoon really summarizes where we are. We have an overall population of patients who fulfill a syndromic definition, and that's the overall phenotype. But underneath that syndrome, there may be sub-phenotypes that are identifiable. And even within those sub-phenotypes, there may be different mechanisms or endotypes that contribute to the sub-phenotype. And it may be that as we can phenotype patients better, we may be able to target therapies more accurately. And I guess on the basis of that, I'd sort of like to sort of talk through how I think we need to progress in terms of precision medicine for ARDS. And we need to start initially to take an approach to discover if sub-phenotypes exist. And then if these phenotypes do exist, identify if we can identify them at the bed space, but also prospectively identify that these phenotypes exist. And that will hopefully introduce the opportunity to test either repurposed or new biological therapies to get to the individualized critical care medicine approach we want to see for our patients. So to do that, currently we need quite detailed statistical methods. And one such method is called latent class analysis. And this is a picture of Carl Pearson, who with one of his collaborators applied clustering techniques to measurements of shore crabs that his colleague had identified and was perceived to be a single population of crabs. But by applying observed data, Pearson was able to show that in fact, within this overall population, there were two distinct populations. And this is thought to be the initial description of the use of clustering techniques to identify hidden or latent subgroups. And essentially what latent class analysis does is it takes an unsupervised approach. And with the development of computing power, we've been able to really bring this into the clinical arena over the last several years. And it takes measured variables to identify these hidden or latent subgroups. And these classes aren't informed by our predefined concepts of sepsis or severity of illness. It takes a data driven approach. And also importantly, it's agnostic to outcomes. And this allows us to move from a heterogeneous population to a more homogenous subgroup. And really, Carolyn Kalfi, who's shown here, has led the way in this field. And this is her seminal paper published in Lancet Respiratory Medicine now 10 years ago. And in a secondary analysis of two ARDS net trials, she was able to show that there were two distinct subphenotypes within the overall population. These were termed hyperinflammatory and hypoinflammatory. Now, I think Carolyn would acknowledge that that's probably not the best descriptive terms that were used, because it probably represents very inflamed and inflamed, but in a different way. But these are the labels that we have stuck with. And you can see that the hyperinflammatory phenotype is characterized by higher levels of cytokines, where in contrast, there's lower levels of bicarbonate protein C and systolic blood pressure. And these very clearly separate the two groups. It is interesting, again, that whenever you look at sort of the classical ways in which we have separated patients, for example, sepsis, you can't differentiate the groups based on the presence of sepsis or other etiological risk factors. So these are hidden, so to speak. But even more importantly than identifying that these two phenotypes existed, Carolyn was able to show in this work that there was a substantially worse outcome in patients who were in the hyperinflammatory subphenotype. And you can see here mortality was substantially higher and there were fewer ventilator-free days and organ failure-free days in patients in the hyperinflammatory phenotype. So this was the initial description. Carolyn has went on to show this in multiple other cohorts, as have other investigators. So this isn't just a single finding. This is a very robust finding across a range of other observational and clinical trials. And importantly, we looked at this in the trial that I did of simvastatin in ARDS. And that was a UK trial, very different from the population that might be recruited in the US ARDS net trials. And yet these findings were consistent that these phenotypes existed. And again, just looking at the consistency of the outcomes across a range of these trials, you can see very consistently that the hyperinflammatory phenotype have a very substantially worse outcome both in adults, but also in pediatric populations as well, and also in observational cohorts. So it does appear that these subphenotypes are robust. This is another piece of work from Louis Boss. And interestingly, in this study, he took a different clustering approach. And he took a range of 20 biomarkers, looking at inflammation, coagulation markers, and endothelial biomarkers. And again, he was able to find two distinct subphenotypes. These were described as reactive and non-inflammed. But interesting, whenever you look at those two phenotypes, they overlay very closely with the hyper and hypoinflammatory. So again, additional evidence of the robustness of this finding of the two clusters. But what I think is even more interesting is that underneath the two clusters, there are different mechanisms. And if you remember to the original slide where I showed that underneath the subphenotypes, there might be different mechanisms that result. And in this transcriptomic analysis, Louis Boss and colleagues were able to show that in the hyperinflammatory phenotype, that the mechanisms that were important were predominantly neutrophil-driven and oxidative phosphorylation pathways, while in contrast, in the hypoinflammatory phenotype, it was mostly MAP kinases pathways that were activated. And again, as we'll talk a wee bit about simvastatin and ARDS, there was a suggestion that those mechanisms that were upregulated in the hyperinflammatory phenotype might be modulated by simvastatin, suggesting a biological rationale for why we might see heterogeneity of treatment effect in patients with hyperinflammatory phenotypes compared to hypo if treated with simvastatin. So I think that is really robust that these are present in multiple cohorts using different techniques and underpinned by different mechanisms. And then, as I mentioned, there was a trial that we conducted where we randomized patients with ARDS to receive simvastatin 80 milligrams daily for up to 28 days compared to placebo with the primary outcome of ventilator-free days. And in that study, overall, there was no difference between the two groups. However, in this really nice secondary analysis that Carolyn conducted in collaboration with our team, we were able to show that although there was no difference in outcomes in patients who were randomized to receive simvastatin who were in the hypoinflammatory phenotype, in contrast, there was a really significant improvement in outcome in patients who were in the hyperinflammatory phenotype who received simvastatin compared to placebo. Now, it's really important to mention that this was an unplanned post hoc analysis. So I think we need to take it as hypothesis generating rather than something that we should change practice at the minute, but promising. And there's been a range of other secondary analyses in multiple ARDS trials and COVID trials that have shown heterogeneity of treatment effect based on these two inflammatory phenotypes to ventilatory therapies such as PEEP, fluid strategies, and also corticosteroids. So certainly, multiple findings supporting this heterogeneity by treatment effect. And then I guess it's worthwhile just pausing to think about, well, if these phenotypes exist in ARDS, perhaps they might exist in our other syndromic conditions. And it may be that we move away from the syndromes of ARDS, acute hypoxia, disparity failure, and sepsis to thinking more about, can we identify these modifiable traits? So perhaps as patients are admitted to the intensive care needing ventilation and pressers, independently of any syndromic label, perhaps underneath those labels, there might actually be the concept that these treatable or modifiable traits exist beyond ARDS. And again, we have really nice new data emerging from Pratik Sinha and Carlin's group to suggest that that might be the case. So this is a paper which looked at two observational cohorts and two randomized control trials. And again, as you can see here, there's a very consistent signal of finding the hyperinflammatory and hypoinflammatory phenotypes, with the hyperinflammatory phenotype, again, having much worse outcomes. But perhaps what is even more interesting, again, is the idea of heterogeneity of treatment effect. And in the VAST trial, there was no heterogeneity of treatment effect, according to the phenotypes. But in the secondary analysis of the PROYAS trial, which studied activated protein C, as you will all know, you can see very clearly that there is a heterogeneity of treatment effect with a potential benefit in patients in the hyperinflammatory phenotype who received activated protein C. So certainly suggesting that these modifiable traits may not be limited to ARDS, but also in sepsis and potentially in patients who don't fit a predefined syndromic label. So that's where we've got to in terms of the approach. We've identified these phenotypes. We've shown that there's heterogeneity of treatment effect. But the limitation is that we need fairly complicated statistical modeling techniques to identify the patients. And that's a limitation for translating this to the bedside. So what Pratik Singha and Carolyn Kalthi have gone on to do then is actually say, well, is there a way to define a parsimonious model, so a model that uses a much more limited number of clinical characteristics and biomarkers to sort of identify these phenotypes without the need for this complex statistical modeling? And the short answer is that is possible. So if you look at the graph here, you can see that there are multiple models that have a very strong capacity to identify hyper and hypoinflammatory with the area under the curve that you can see ranging from 0.92 to 0.95. And the one that's highlighted uses bicarbonate soluble TNFR1 and IL-6. And the reason that's highlighted is we've gone on to try and look at this in a prospective study. So this is the model that is possible to deliver at the bedside. But you'll note that the biomarkers that are needed to measure that are soluble TNFR1 and IL-6. And that's a limitation because most of our labs aren't able to measure those biomarkers in a timely fashion. So the next part of the story was, well, can we measure or develop an assay that might measure those at the bedside? And this was the work that I mentioned at the outset with a company called Randox and using a pre-existing device that they had for a toxicology assay called the multistat device. They were able to use that hardware in collaboration with a new assay to measure IL-6 and soluble TNFR1, and then combining that with the arterial blood gas bicarbonate and using the model that I showed in the previous slide, it is possible to phenotype patients potentially at the bedside. And this is some initial preliminary data that we have. And in this setting, we took samples from the HARP2 trial and we compared the point of care assay in a lab environment compared to the gold standard ELISA measurements. And again, with those approaches, we're able to show pretty convincingly that it is possible to phenotype using this point of care device. The limitation of this work was that this was done in a controlled laboratory conditions rather than in the wild type situation of our health care systems. So the next step then was to try and undertake an observational cohort. And in this, we have recently completed this study where we have undertaken a prospective multicenter observational trial called the FIND trial. And really, the two key objectives of this study was first to identify prospectively that the ARDS subphenotypes existed. Many of the previous examples have been secondary analysis of previously completed cohorts. But also ask that in the health care setting where we have clinicians using this assay that it was possible to validate the point of care assay for phenotype allocation. We were also fortunate enough to conduct an exploratory cohort of patients with COVID-19. And I'll show some of that data. But also, we've included patients with acute hypoxic respiratory failure who don't meet the ARDS definition. Again, to get at this idea that it's not about the syndromic definition, but it's whether or not the phenotypes exist underneath the overall population. And this is work that Tomas led. And this was a very small study in the middle of the COVID pandemic. But what we were able to show in this study of 39 patients was that we could phenotype patients at the bedside conducted by clinicians. And we showed that that matched very carefully to the phenotypes that had been described previously in the other studies. So some tantalizing suggestion that this was actually deployable in the clinical arena. And then this is some preliminary analysis of the FINE study that we've recently completed. This was an interim analysis. So we haven't looked at the outcomes. I can't tell you if the phenotypes that we see have the same outcomes as has previously been described yet. But what I can show is that whenever you look at phenotype allocation using the point of care device in the clinical setting compared to the phenotype allocation with the gold standard ELISA R&D assays, you really see very impressive correlation. So I think we're now on the cusp of being able to phenotype patients at the bedside. And I think this will hopefully drive interest, not just from the company that we currently work with, but many other diagnostic companies to develop similar assays. So I focused very much on the inflammatory phenotypes to date in the talk. But it is worth just pausing to sort of make the point that it is somewhat naive to suggest that that is the answer for all of the subpopulations that might exist underneath the overall population of ARDS. So I think the inflammatory phenotypes are our most robust findings to date. But it is not plausible to imagine that there won't be other phenotypes that are present within the overall population. And this is a really nice paper by David Maslov, where he talked about the different approaches, whether it be proteomics, transcriptomics, that might subsequently inform the phenotyping of our critically ill patients. So very much moving towards identifying these modifiable traits. And I just thought it would be interesting to show a couple of potential other examples. So this is another protein biomarker. And again, this is a secondary analysis of the HARP2 trial. And interestingly, in this study, we find a very similar finding using IL-18 as we had done with the inflammatory phenotypes. If you were randomized to simvastatin and you were IL-18 low, there was no difference between the two groups. However, if you were IL-18 high and were randomized to receive simvastatin, you had a significantly improved outcome. And this might get at the concept that there might be an inflammasome subphenotype within the overall population, given that IL-18 is a very potent biomarker of inflammasome activation. So there will almost certainly be other protein biomarkers. But there's also likely to be transcriptomic analysis. And this is work from Gillian Knight and Tony Gordon in the UK. And in this work, they undertook a secondary analysis of the VANISH trial, which randomized patients to hydrocortisone or placebo. And what they found was that there were two phenotypes based on the transcriptomic analysis, SRS1 and SRS2. And interestingly, whenever patients were randomized to receive hydrocortisone in this trial, there was a differential treatment effect with SRS2 having a better outcome. So suggestion that there might be secondary analyses using transcriptomic analysis that might be useful. And then finally, as an acknowledgment to one of my other collaborators, Ewan Gallagher in Toronto, is that we shouldn't forget physiology as a potential way to phenotype patients. And this was a secondary analysis of our REST trial. And in the REST trial, we randomized patients with acute hypoxic respiratory failure to lung protective ventilation or lower tidal volume ventilation, aiming for a tidal volume of three mils per kilogram, facilitated by extracorporeal CO2 removal. But interestingly, although we didn't see any overall difference in the primary outcome in the overall population, whenever you look at ventilatory ratio, you can see that there is heterogeneity of treatment effect with patients with a higher ventilatory ratio, perhaps reflecting a higher dead space, having a better outcome whenever they receive lower tidal volume ventilation facilitated by extracorporeal CO2 removal. So I guess that's just to emphasize the concept that there are multiple different ways that we will likely be able to phenotype our patients and move to a precision medicine approach as we move forward. So just to summarize, and this is a nice cartoon from a paper by Jeremy Beichler as to where we are. So we have discovered these sub-phenotypes exist. We have reproduced them in multiple cohorts. We've undertaken some reverse translation to define the mechanisms that sit below these phenotypes. We've identified some potential therapies based on the heterogeneity of treatment effect and the secondary analyses that are presented. And now we can discover these phenotypes at the bedside using point of care devices that are emerging. So the next step now is really to test the question, are these phenotypes meaningful in terms of developing therapies for our patients with ARDS? And this cartoon suggests how a potential platform trial might look. And that brings me to introduce the PANTHER platform trial. And you can see the website here. And the PANTHER trial is a really ambitious trial that I'm glad to be part of. It's truly international and collaborative. And if people are interested in being involved, please do get in contact. It's going to focus on the phase 2 space. I think too frequently, we have run towards doing phase 3 trials. So I think what we want to try to do in this trial is identify therapies that are the most promising to move to definitive phase 3. It's going to be precision medicine, as I'll explain, and it's going to be a platform trial. And hopefully, over the next several years, we'll have this trial up and running and established. And it will continue for many years to find therapies. What that means for a platform trial is that it's going to take potentially many years and it's running. And that's why one of the key aspects of this is embedding early career research engagement. And again, I'm really keen to speak to any early career researchers who would like to get involved, because it's really important that if we are going to do this, it'll need to be managed over the next number of years. And that will require early career researcher involvement. We're very keen to work with industry. And as an academic, maybe whenever I started, I thought the industry was a bit like the dark side and you couldn't talk to industry. But in fact, what is ever more clear is that the aims of industry align with what we as academics and clinicians want to do, i.e. find new therapies for patients. So it's really important to us that we collaborate with commercial companies. And the reason for that is also to make the platform sustaining. We're going to need academic funding to get the trial up and running. But if this is going to be a platform that continues in the long term, we'll need to mitigate the risk to academic funders by bringing in commercial funding. And another key part of it will be to embed biobanking. And I think that's really critically important. We wouldn't have found any of this exciting new progress if we hadn't have embedded biobanking within our trials. And again, a lot of that work has been driven by Michael Mathay, a sort of a long-term mentor of mine. And I think as we do these trials, we need to make sure that we embed biobanking to understand other potential mechanisms. And then I think something that's really important is that we involve patients and the public. And that is critical. We need to make sure that we are answering questions that are important to patients and that the outcomes that we're looking at are important to patients and that we're not exposing patients to excessive burden. So there's some of the key factors in the PANTHER trial. I'll come back to our intervention prioritization in the subsequent slide. The best part is this is fun, and it's a huge international collaboration. This slide just shows some of the people that are involved. There's several sort of screenshots of multiple people. But this involves investigators and clinicians from North America, Australia, Europe, Japan. So it's really great to see the sort of opportunities that this international collaboration presents. And just really to summarize where we are, the plan is to recruit patients with ARDS who have either hyper or hypo inflammatory phenotypes. We plan to initially study simvastatin or baracitinib, and they're the two that we're going to start with. We'll compare those to usual care, so this is going to be an open-label design. And the primary outcome is 20-a-day organ support free days, which I'll come back to. So these are the eligibility criteria. So we're going to recruit patients with ARDS based on the global definition. So we're going to include patients who are on high-flow nasal oxygen and non-invasive respiratory support. We're going to sort of take a pragmatic approach and try not to exclude patients to ensure that we have generalizability. So we're going to treat patients early because we're sort of conscious that we want to modify some of the host inflammatory response. So patients within 48 hours of the diagnosis of ARDS. We don't want to recruit patients who have no potential to benefit. So therefore, if there's a planned treatment withdrawal within 24 hours, we would exclude those patients. And because the primary outcome includes organ support, both in terms of ventilation and pressure support, we exclude patients with pre-existing domiciliary ventilation. And that'll be at the platform level. And then the next level down, we will have domain-specific intervention exclusions, for example, in terms of simvastatin and baricitinib. And this just highlights the intervention selection process. And again, there's a website that you can submit your suggestions to. And I think many of us will recognize that during COVID, the process by which we decided that we would test some therapies maybe had something to be questioned. And we tested a whole variety of therapies that perhaps might not have been robust if we had gone through a process of selection. So what we've done is put together a process where a member of the trial team, an academic partner, or a commercial partner can submit a suggestion via the website. We then have a trial team that makes up the intervention selection committee. And we're very much looking for proof of concept that there might be a beneficial signal. But then we're going to add in an additional step where we have an independent group of experts not associated with the trial who will say, yes, we agree or we disagree. This is what should be prioritized. And then we'll take that to the trial management group to decide what we bring into the trial. And we hope that that system will be robust in terms of selecting therapies. So that's quite important because I said although I'm keen as part of the group to support collaboration with industry, we're not just going to say we'll take anything that might have funding attached to it. We're very much driven by a combination of evidence that something works and a commercial partner if that's the case. So this is the primary outcome. And I mentioned this organ support free days up to day 28. And that's an ordinal scale, which includes mortality scored as the worst outcome of minus 1, and then a continuous measurement of days alive and not requiring organ support. And that's important to patients in our discussions. But also it's very efficient in terms of statistical design. But it's also important to point out that from REMAP-CAP, we know this early measure of outcome translates into 180-day outcome. So it's quite a robust outcome for both patients and methodological reasons. And here's a list of the secondary outcomes that I'll give you an opportunity to read. But again, they're the key secondary outcomes that we have been identified in association with our patient groups. So this is a slide that summarizes the design. It's a Bayesian adaptive design. And I come from a frequentist background, I have to admit, and I'm being sort of slowly dragged into the Bayesian arena. But again, I think we have seen on many occasions the idea that we have a frequentist design where A is compared with B, and we get a fixed sample size where whenever you get to the end of the trial, the p-value is not 0.07, and we don't know if the answer is strong enough to change practice. So with a Bayesian design, you have pre-planned looks at the data to the point where whenever the data reaches a predefined statistical trigger of efficacy or futility, then the therapy stops and graduates into a phase 3 trial or is stopped for futility. And we have defined our triggers based on extensive modeling to get the best par and type 1 error. As I mentioned, there's no fixed sample size, and that's quite difficult for a funder to understand whenever we're saying, well, it could be 2,000, but it's more likely to be 1,000. But ultimately, what we have is the expected sample size to answer for simvastatin and baricitinib in the two phenotypes. But we also have a cap sample size. If the effect size is so minimal that you don't hit a trigger, it's probably not of sufficient importance to graduate into a phase 3. So that's our approach for sample size. So with that, I will conclude. And hopefully, I have convinced that phenotypes exist in ARDS and that these may represent treatable traits. And we're at the cusp where these phenotypes may be identified at the bedside. And that allows us to undertake precision medicine trials, of which PANTHER is one. And hopefully, in the next few years, I'll get invited back to present some results from the PANTHER trial. Thank you.

ARDS

precision medicine

sub-phenotypes

PANTHER platform trial

latent class analysis

biobanking

Bayesian adaptive design