Scientist Stories: Emma Davenport, Stratifying Sepsis Patients through Transcriptomic Profiling

Hello, I'm Emma Davenport, and I'm a group leader at the Wellcome Sanger Institute. And today I'm going to share our work on stratifying sepsis patients through transcriptomic profiling. One of the main focuses of our research group is how we can use functional genomics to understand patient heterogeneity. And we're interested both in terms of heterogeneity within disease, but also in the way that individuals respond to treatment. There are a number of different data types that we can use under the umbrella of functional genomics to understand patient heterogeneity. The first of these is genetics, and we can use this to conduct genome-wide association studies to identify variants that are associated with different traits and diseases. We can also measure gene expression and look at differential expression between two different groups, so for example, between different subgroups within a disease or looking at pre- and post-treatment. And finally, we can combine these two data types together through expression quantitative trait locus or eQTL mapping to start thinking about how gene expression might be regulated by the underlying DNA. One of the diseases that we work on is sepsis. Sepsis is a life-threatening condition that occurs when the body's immune system is responding to infection, but it ends up injuring its own tissues and organs. It can be caused by a chest infection, an abdomen problem, or an infected cut or bite, but it's most frequently caused by a bacterial infection. Anyone can develop sepsis, and it kills 52,000 people each year in the UK, which is more than breast, bowel, and prostate cancer combined. Worldwide, a third of people with sepsis die. This figure is just highlighting some of the sepsis symptoms that a patient might experience. So, for example, shivering and fever, or feeling very cold, having having extreme pain or discomfort, having very pale or discoloured skin, being very sleepy or being difficult to rouse. The patient will feel incredibly unwell, so they might feel like they're dying, and finally, a shortness of breath. What you can see though from these symptoms is that there's no single symptom that can be used to diagnose sepsis, and they can be quite variable, and this can look quite similar to other conditions. So for example, sepsis can initially look like flu or a chest infection. And because we don't have a reliable diagnostic marker, this makes sepsis very challenging to treat and to study. In terms of current treatments, the underlying infection can be treated with antibiotics, and any organs that are failing can be supported. So for example, if someone has problems with their kidney, they can be put onto dialysis. But it's still very hard for us to target the immune response that we think is actually damaging the organs, and we don't have good drugs to able to do that. There's been many, many clinical trials over years that have been trying to identify different drugs that may be able to help with the sepsis treatment, but we think that the substantial variation across individuals in terms of the response to sepsis is making this very challenging. In the clinical trial setting, if a drug doesn't work, we don't know if that's because we're not focusing on the right group of patients, So it might be that it only works in a very small, in a much smaller subset of individuals. There are lots of different factors that can contribute to the host sepsis response. These can include environmental factors, such as the access to healthcare, the timing of any therapies that are given, and the types of treatments that are initiated. There's also pathogen factors, so how virulent the pathogen is, whether there's any microbial resistance, And finally, there's many host factors. So for example, there might be genetic predisposition, there might be epigenetics playing a role, comorbidities, or also social demographic, age, gender, and ethnicity factors. Given all these different factors that can play a role, we're particularly interested in using functional genomics to study sepsis for a number of different reasons. The first of these is that gene expression can capture both genetic and environmental influences. And as we saw from the figure before, we think that there's multiple different influences on the final host response. In terms of the genetics, genome-wide association studies have been conducted for many, many different traits and diseases. In particular, for infectious diseases, this has generally been looking at susceptibility to an outcome from infectious disease. From these studies, and particularly from the infectious disease studies, these have demonstrated that genetic variation may play a role in the heterogeneity of patients and their response to infection. For sepsis in particular, it hasn't been very successful for the genome-wide association studies. There's been a couple that have been conducted, but only a small number of regions across the genome have been identified. One of the reasons we think that this might be the case is, again, due to this incredible heterogeneity. So by grouping all of the patients together, we may we might not be looking at quite the right population for identifying the specific variants that might be contributing. And so it might be better to subset patients to then be able to do the analysis. But one of the other challenges for genome-wide association studies is that you need very large numbers, and sepsis patients are very, very difficult to recruit. This is a very severe illness, and patients are frequently admitted to the intensive care unit. And this is an environment which is very challenging to then recruit patients into trials. So this is another reason why we think that using functional genomics, so for example, looking at gene expression, is particularly interesting in sepsis because we can conduct these studies on much smaller cohort sizes. From all the GWAS that have been conducted, both in infectious disease and more generally for traits and diseases, the majority of associations that have been identified involve non-coding variation. So this suggests a role for gene expression, again, providing further evidence as to why this is interesting to study in sepsis. There have been some interesting studies that have been conducted from cells isolated from healthy individuals that have then been stimulated, so for example, with LPS. And these have highlighted that the gene expression response to stimulation can be modulated by genetic variants, again supporting this idea that genetics is influencing the way people respond to an infection. And finally, there's been substantial expression studies of pediatric sepsis. And this has found 3 subclasses of patients, but this hasn't been well replicated in adults. Adults. So there's really an opportunity here for us to explore what's going on in the adult patients. So for these reasons, the first part of this talk is focusing on how we've used functional genomics to understand the substantial individual variation observed in response to sepsis. The Genomic Advances in Sepsis, or GAIN, study was set up by Julian Knight's group at the University of Oxford a number of years ago. And here, we recruited patients to explore inter-individual heterogeneity in the transcriptomic response sepsis. We've recruited patients from over 30 intensive care units from around the UK, and we were focusing on patients that had either community-acquired pneumonia or faecal peritonitis as the cause of the sepsis. These are the two major causes of sepsis in the UK. For each of these patients, samples were taken on day 1, day 3, and/or day 5 after they were admitted to the ICU, and these samples included leukocytes for cell— for gene expression and buffy coat for genotyping data generation. So just a little bit more information about the cells that we're going to be looking at in terms of the gene expression. We've been using Leukoloc filters as a way of collecting whole blood and processing it very rapidly at the bedside. So as I mentioned, it's very challenging to conduct some research studies in the ICU because obviously these patients are very severely ill and there's not necessarily access to a lot of equipment for processing blood. So these Leukoloc filters enabled us to take the blood, the research nurses could then pass the blood through the filter, and this allowed us to stabilize the gene expression profile very rapidly to really get a snapshot in the leukocyte cell fraction of what's happening at the gene expression level at this very critical time in someone's illness. So I'm going to start by describing a derivation cohort that we initially generated for 265 patients with sepsis due to community-acquired pneumonia. And for these patients, we generated gene expression using one of the Illumina microarrays. For this cohort, we started off by taking the top 10% of probes that were most variable across our cohort, and we conducted unsupervised hierarchical clustering analysis, and this revealed two distinct clusters, which we have designated as sepsis response signature, or SRS1 and SRS2. We started exploring what was different between these two groups, and so we initially looked at differential gene expression. Having looked at all of the probes across the genome, we found around 3,000 were differentially expressed between SRS1 and SRS2, with the majority being downregulated in SRS1. To explore a little bit more about what these two groups are potentially representing, we conducted pathway enrichment, and from the pathways and the disease phenotypes and the regulators that we identified as being enriched, it is highlighted that this SRS1 group is identifying individuals with an immunodeficiency. Suppressed phenotype. An even more striking result, though, is actually the association between these SRS groups and mortality. This has always been a very challenging area in terms of sepsis research to try and understand why it is that some individuals survive and some individuals don't. It's very difficult from clinical data alone to predict which individuals are going to make a good recovery and which individuals, unfortunately, won't survive.. And so this finding was very exciting for us that we were able to observe that the SRS1 group, so those that have this immunosuppressed phenotype, also have a poorer outcome compared to the SRS2, and that's highlighted here in the survival plot with the SRS1 individuals with the blue survival line. So having identified this in our derivation cohort, we were then very keen to try and replicate it in a validation And so we continued to recruit patients through our intensive care units, and we came up with a 7-gene set that we, which we then were able to use to call SRS in additional cohorts. So these 7 genes are basically representing that big transcriptomic differences that we're seeing between the 2 groups. So we recruited just over 100 additional new patients to the study, and we used this 7-gene set to predict whether an individual was in SRS1 or SRS2. And then when we plot our survival curve again, you can see that we're able to confirm that this SRS1 group had this much poorer outcome compared to individuals in the SRS2. We wanted just to make sure that those 7 genes that we were using really were capturing the full sort of transcriptomic response that was being represented by the SRS groups, and so we conducted differential gene expression in this new validation cohort and again looked at pathways and phenotypes that were enriched for the differential expression. On this plot, we've now got some lighter purple colored bars, which is the validation cohort, and you can see that again very similar pathways and functions are being identified as enriched in the differential expression in this validation cohort as what was originally found in the discovery cohort. Having identified these SRS groups and been able to validate them within our cohort, there are a number of questions that we were then really interested in starting to address. The first of these is, having seen SRS group in community-acquired pneumonia, are we also able to identify them in other diagnoses of sepsis? The second is starting to explore whether SRS group may be associated with response to treatment. And this will be particularly informative for trying to understand why clinical trials clinical trials have failed so much over time. And the final one is thinking about how the host genetics might be contributing to SRS group membership and whether people are predestined to be in one group or the other. So to start with the first question, we, in addition to our community-acquired pneumonia patients, we also recruited faecal peritonitis patients to the GAINES cohort. So these are patients where there's been some sort of perforation in the bowel, which means that they have bacteria sort of leaking, and so this is a very different cause of sepsis. In 117 patients that were recruited to the GAIN study that had sepsis as a result of faecal peritonitis, we again used our 7-gene set to assign SRS1 or SRS2 group membership. And again, from this survival curve, you can see that this SRS1 group are consistently having a much poorer outcome, particularly in the early sort of first couple of weeks after the sample was taken. So moving on to the second question, this is obviously a very challenging question to address, but our collaborators had run a clinical trial which gave us an opportunity to start exploring this. So here we're looking to see whether or not there's an interaction between the SRS group and treatment with hydrocortisone. So the VANISH trial was set up and enrolled just over 400 patients into this 2x2 double-blind randomized trial. There were 2 study drugs that were used. So if the maximum infusion rate was reached for the first study drug, then patients went on to be randomized into the study drug 2, which was either hydrocortisone or steroids, which is a common treatment for sepsis, or a placebo. So 175 of these patients had blood samples collected, which enabled RNA to be extracted and gene expression to be profiled through microarray on the day of enrollment. There was no difference in terms of Study Drug 1, but there was an interesting finding for Study Drug 2. So, first of all, focusing just on the placebo patients for Study Drug 2, if we take all of the patients on the left-hand side who were assigned to SRS1, these individuals show a much poorer outcome than those that are in SRS2. So this, again, is a very reassuring finding that we're consistently seeing that these SRS groups are able to pick up this difference in mortality. And this cohort in particular is slightly different to the GAINES cohort in that it involved all-cause sepsis. So again, we're expanding our cohort in which SRS groups are being able to be identified. If we then look at what happens when we focus on the patients that were then given the hydrocortisone, on the left-hand side, you can see there's very little difference in survival between the placebo and the hydrocortisone for the SRS2 group. But on the right side, you can see that there is a difference dependent on whether or not the individual was given the drug. So this is a very small study, and it needs to be replicated. However, it's a really exciting proof of principle for us that this is one way in which the SRS groups could be really informative for having more of a translational impact. As I mentioned earlier, clinical trials have been very unsuccessful in the past for sepsis, and we think part of that is that sepsis is being treated as one condition and maybe we need to be sub-grouping patients. This is one way that we can explore doing that. And this particular finding is suggesting that this is something that we should be continuing to pursue. And that's something that we're hoping to do over the coming years. So to start thinking about the third question, which is how much genetics is potentially influencing SRS group, this is an area that my group is particularly interested in, and we've been using eQTL mapping and EQTL interactions to start exploring this. Expression quantitative trait locus, or EQTL mapping, is one way in which we can determine the contribution of genetic variation to the heterogeneity that we're observing in gene expression across our sepsis patients. This little cartoon here is showing you what an EQTL looks like. So the idea here is that we're looking for a linear association between a genetic variant or a single nucleotide polymorphism in the DNA and changes in the expression of a gene. So as you can see, it might be that you have a SNP or a variant in the promoter of a gene, and for example, the transcription factor might bind more strongly to the G allele, which results in increased amount of RNA that's expressed and therefore increased expression of the gene. On the right-hand side, you can see that the box plots are then highlighting this linear association that we have between the genotype of the SNP of interest and the expression of the gene. We started off by applying our eQTL mapping approach in the derivation cohort. As I mentioned, we had 265 patients initially, and 240 of these had both gene expression and genotyping data available. We used a linear model to find associations between the SNP genotype and expression of the gene, and in this situation we focused on cis-eQTL, where the SNP is in close proximity to the gene. Looking at all of the genes across the genome, we found just under 4,000 that showed evidence of an eQTL. Going back to some of our differential expression work from this initial cohort, we were able to find a number of networks that were enriched for genes that would differentially express between SRS1 and SRS2. This includes this particular network, which is a hypoxia network, and you can see there's a couple of genes in the center, so particularly HIF1-alpha and EPAS1, that are both very central to the response. This particular figure is colored by the differential expression of these genes between SRS1 and SRS2. But if we take exactly the same network again, and this time we color them by whether or not that gene shows evidence of an eQTL, you can see that we're getting a good number of these genes showing evidence of genetic regulation, and particularly again, these very central genes that are important to the network of HIF1-alpha and EPS1.. So this is starting to provide us with some evidence of how genetics could be influencing this huge transcriptomic response that we're seeing between SRS1 and SRS2 group. If individuals are predisposed to have a slightly higher or lower level of expression of some of these key genes in these networks, that's one way in which that then there could be this sort of signal transduction through all the different pathways and having this up or down regulation of gene expression. So as well as being able to identify eQTLs, we also know that the environment can perturb some eQTL effects, and there's been lots of studies now that have been done to measure gene expression and look at eQTL effects that are modulated dependent on tissue or cell type or following stimulation. One such study was conducted by Fairfax et al. back in 2014.. And this particular study isolated monocytes from healthy individuals and then looked at how genotype can modulate the gene expression response to different stimuli. So here in this particular figure, there's monocytes in the naive state, monocytes that have been stimulated for 2 hours of LPS, for 24 hours of LPS, or for 24 hours with interferon gamma. What this highlights is that some SNP gene pairs, so some eQTLs, can have quite differing effects depending on the stimulated state. So you can see here in the naive monocytes, there's actually a negative eQTL effect. There's no eQTL effect after 2 hours stimulation with LPS. There's a positive eQTL after 24 hours with LPS, and then there's no eQTL again with the interferon gamma. So what we're really interested in is whether or not some eQTL effects can be modulated by the SRS group, so whether or not an individual is in SRS1 or SRS2. Going back to our little cartoon again of what an eQTL is, it might be that an eQTL is active in one particular SRS group, so that's the blue one at the top with SRS1, and then in SRS2, for example, maybe the transcription factor is no longer active, which means that then there isn't this difference in expression when the G allele is present in the promoter, and therefore we lose this eQTL effect between SRS1 and SRS2. This difference in eQTL effect can be identified by modeling an interaction, and our hypothesis is that if these eQTL effects are modulated by SRS group, which we're defining here as in vivo SRS-eQTL interactions, this might be able to help us to identify regulatory variants that are functional in the specific SRS groups and also start thinking about if there are particular regulatory drivers that are driving groups of differential expression, and really whether these can then be targets for modulating the immune response to sepsis. Over time, we've continued to recruit individuals to our study, and we've continued to increase our gene expression and genotyping datasets. So this little section here is now focusing on just under 500 patients for which we have both gene expression microarray data and genotyping data available. The first step in our in vivo eQTL interaction approach is to identify cis-eQTLs, so this is where the SNP is in close proximity to the gene of interest across the genome using all the available samples that we have, and in this case we're including multiple samples per patient if they had additional samples collected while in the ICU. So across the genome we found 2,000 probes that showed evidence of an eQTL. And then for each of these eQTL genes, we tested the most significantly associated SNP for an interaction with our SRS groups. Of the 2,000 or so eQTLs, we found 278 of these show evidence of an interaction. This is an example of one of the interactions that we identified. So this is the TAF1C gene. On the x-axis, we have the genotype of our SNP of interest, and on the y-axis, we have expression of the gene. Each point here is representing one of the samples that was collected from the sepsis patients in the ICU, and they're colored in red if that sepsis sample was defined as being SRS1, and they're colored in blue if they're defined as SRS2. What this plot shows is that individuals in the SRS1 group, so the red box plots, have a much steeper gradient, so a much stronger eQTL effect than those in the SRS2. On the right-hand side, this is just exactly the same data again, but this time we've got all of the SRS1 groups on the left and all of the SRS2 groups on the right. Each point now is coloured by the genotype of the SNP, and this is just to further highlight the difference in variance between these two groups. So there's a much greater difference between the genotype groups in the SRS Group 1 individuals compared to Group 2. So this is suggestive that there's something interesting going on functionally with that SNP, but only in the context of SRS1. And so again, starting to piece together additional pieces of information and evidence as to what the differences are between these two groups and whether it's genetics that is playing a role in determining whether people are going to be in one group or another. We've got a cohort now of over 1,000 individuals that have been recruited. And as I mentioned, we've got microarray data for around 500 of these. So we've recently increased our transcriptomic profiling, and this time we've switched to RNA sequencing, which is giving us some additional information to what the microarray has available. We have two particular questions that we're interested in answering initially with this new RNA sequencing dataset. The first of these is how well correlated the microarray and RNA sequencing transcriptomic signatures are. So to be able to answer this, we have generated RNA sequencing for 132 of the patients that were in our initial derivation cohort that had been measured with a microarray. We then want to know if we can use our same 7 gene set from the microarray to assign SRS group in the RNA sequencing data. And for this, we've generated RNA-seq data for an additional 534 newly recruited patients. So to look at the first question, here are the 7 genes that we've been using to define our SRS group. On the x-axis, we have microarray normalized expression, and on the y-axis, we have the RNA-seq normalized expression Each point is one of the 132 individuals that have both microarray and RNA sequencing data available. And as you can see, we're getting very good correlation across the 7 genes in terms of the expression using the 2 different types of technology. So we use this 7-gene set to then call SRS group in our new patients that we have RNA sequencing data for. And when we look at the survival curve, we again, we can see that those with this SRS1 group have much poorer outcome, so that SRS1 is associated with this early mortality compared to SRS2. So this is very reassuring that again we can translate our 7-gene set onto a new technology that can be used to assay the transcriptome, and again in our second half of our GAINES cohort that we are consistently seeing this difference in survival between these two groups.. So there's lots of directions in which we want to take this research. So now we have this large RNA sequencing data, we want to map both eQTLs and their interactions using this larger cohort. We want to start exploring if there are any master regulators, so for example transcription factors that are potentially controlling large groups of these eQTL interactions and the differentially expressed genes, and whether these can help identify the components of the immune system that we should be targeting for then intervening in terms of drug targets for treating sepsis. We also want to know if the eQTL interaction genes fall in the same pathways as differentially expressed genes, or whether they're able to offer us some independent and additional information that can help understand what's really driving these two different SRS groups that we've identified. And finally, we want to think about how we can use these eQTL interactions to better understand the contribution of genetics to SRS group membership. The question still remains, are people predestined to go into one SRS group or the other? How much do they change over time? And how much can we modulate these groups? And all of these are going to be incredibly important for really being able to translate this finding into the clinic and to improve the outcome of sepsis patients. So to conclude, SRS groups are the major predictor of transcriptomic variation in sepsis, regardless of diagnosis. SRS1 identifies a group of patients with an immunosuppressed phenotype and a much greater early mortality rate. These SRS groups have highlighted that they could be a potential opportunity for patient stratification, so for example, in the clinical trial setting, with the ultimate aim of being able to identify the right group of patients to give them the right drug at the right time. And finally, our eQTL and our eQTL interaction work highlights the sort of informativeness of these sorts of approaches for trying to understand how the host genetics is contributing to SRS group membership in these patients. So, this type of study wouldn't be possible without a fantastic team of collaborators to help with all of this work, and I'd just like to particularly thank Katie Burnham and Julian Knight, and of course the research nurses and the patients, without whom any of this work would not be possible. Thank you very much for your attention.