Lippincott Williams & Wilkins



Supplemental DataUnsupervised analysis of transcriptomics in bacterial sepsis across multiple datasets reveals three robust clustersTimothy E Sweeney, Tej D Azad, Michele Donato, Winston A Haynes, Thanneer M Perumal, Ricardo Henao, Jesús F Bermejo-Martin, Raquel Almansa, Eduardo Tamayo, Judith A Howrylak, Augustine Choi, Grant P Parnell, Benjamin Tang, Marshall Nichols, Christopher W Woods, Geoffrey S Ginsburg, Stephen F Kingsmore, Larsson Omberg, Lara M Mangravite, Hector R Wong, Ephraim L Tsalik, Raymond J Langley, Purvesh KhatriTable of Contents:Jargon-Free SummarySupplemental MethodsSupplemental ResultsSupplemental Tables 1-11Supplemental Figures 1-6Jargon-Free Summary Sepsis may not be a single disease, but rather a spectrum composed of several ‘endotypes’ (also known as clusters, or subclasses of disease). Previous groups have identified subclasses in relatively homogeneous clinical circumstances, but it is unknown whether general sepsis endotypes exist. We here hypothesized that there are sepsis clusters that exist broadly across patients with sepsis, and used transcriptomic data (gene expression microarray and RNAseq) from whole blood from a wide range of clinical settings to test this hypothesis. Several dozen transcriptomic studies of sepsis have been completed and publicly deposited for further re-use, but there are vast technical differences between them, including different sample processing methods, different microarray or RNAseq technologies, and different post-processing normalization methods. There are also biological/clinical differences, such as different patient demographics, enrollment criteria, and geographic locations. These biological and technical differences have previously made it impossible to pool data from disparate studies. We recently published a new bioinformatics method that relies on an assumption that healthy controls among different studies are largely the same (a strong assumption). Using this assumption, we are able to pool data in a bias-free manner (i.e., without assuming anything about the sepsis cases) from different studies into a single framework, and allow them to be analyzed as though they were gathered in a single large study. We thus gathered all transcriptomic studies of bacterial sepsis at hospital admission, and split them into all studies with healthy controls (which we used for discovering sepsis clusters) and those without healthy controls (which we used as validation for the clusters we found in discovery). Across the discovery data (700 patients from 14 datasets), we use advanced bioinformatics to determine that the transcriptomic data was ideally split into 3 clusters. We performed pathway analysis in the gene expression profiles of the subjects in the 3 clusters, and found that one cluster had a high innate immune / reduced adaptive immune signal (‘Inflammopathic’), one cluster had a reduced innate immune / high adaptive immune signal with low mortality (‘Adaptive’), and one cluster showed both clinical and molecular irregularities in the coagulation and complement systems (‘Coagulopathic’). Cluster membership was associated with significantly different age, shock status, clinical severity, white blood cell differerential, and mortality. However, we further characterized the effect of age, shock status, and illness severity on cluster membership, and showed that they explain very little of why patients are assigned to the given clusters. This suggests that cluster membership is not simply explained by obvious clinical variables. In order to ever have any clinical relevance, we need some way to determine cluster membership for any given new patient. In other words, we would need some diagnostic blood test that determines cluster membership that could be run when a patient presents with sepsis. We thus derived a 33-gene classifier in the discovery data that had an 83% accuracy in re-assigning discovery patients to their same clusters. We then applied this 33-gene classifier in 9 external, independent datasets (N=600), to retrospectively assign each of the 600 patients to one of the three clusters (Inflammopathic, Adaptive, or Coagulopathic). Having retrospectively assigned these patients to the three clusters, we then had to determine whether they recapitulated the same clinical and biological characteristics as the original Inflammopathic, Adaptive, and Coagulopathic groups. We showed that the same relative patterns of age, severity, shock, and mortality were found, on average, between the validation clusters and the discovery clusters. We also showed that the same pathways were generally activated among patients across cohorts assigned to the same cluster.Our analysis suggests, but does not prove, that there are three different sepsis subtypes (Inflammopathic, Adaptive, and Coagulopathic). These subtypes have significantly different clinical and molecular profiles. We have also produced a 33-gene classifier which is able to identify any new patient as belonging to one of these clusters. The idea of an endotype really only becomes clinically useful once it is coupled with an endotype-specific therapy. Thus, the next steps for this work will be not just confirming the clinical and molecular profiles of patients assigned using the 33-gene classifier, but also trying to identify treatment response patterns with the clusters.Supplemental methodsThe COMMUNAL algorithmA common clustering approach would apply a single clustering algorithm (e.g., k-means clustering) and a single validation metric (e.g., the gap statistic ADDIN EN.CITE <EndNote><Cite><Author>Tibshirani</Author><Year>2001</Year><IDText>Estimating the number of clusters in a data set via the gap statistic</IDText><DisplayText>(1)</DisplayText><record><dates><pub-dates><date>2001</date></pub-dates><year>2001</year></dates><keywords><keyword>clustering</keyword><keyword>groups</keyword><keyword>hierarchy</keyword><keyword>K-means</keyword><keyword>uniform distribution</keyword><keyword>Statistics &amp; Probability</keyword><keyword>STATISTICS &amp; PROBABILITY</keyword><keyword>Mathematics</keyword></keywords><isbn>1369-7412</isbn><work-type>Article</work-type><titles><title>Estimating the number of clusters in a data set via the gap statistic</title><secondary-title>Journal of the Royal Statistical Society Series B-Statistical Methodology</secondary-title></titles><pages>411-423</pages><contributors><authors><author>Tibshirani, R</author><author>Walther, G</author><author>Hastie, T</author></authors></contributors><language>English</language><added-date format="utc">1439405339</added-date><ref-type name="Journal Article">17</ref-type><auth-address>Tibshirani, R (reprint author), Stanford Univ, Dept Hlth Res &amp; Policy, Stanford, CA 94305 USA&#xD;Stanford Univ, Dept Hlth Res &amp; Policy, Stanford, CA 94305 USA&#xD;Stanford Univ, Dept Stat, Stanford, CA 94305 USA</auth-address><rec-number>813</rec-number><last-updated-date format="utc">1439405339</last-updated-date><accession-num>WOS:000168837200013</accession-num><electronic-resource-num>10.1111/1467-9868.00293</electronic-resource-num><volume>63</volume></record></Cite></EndNote>(1)) at a single number of variables (e.g., 1,000 genes, usually chosen arbitrarily) to determine clusters. However, this approach can lead to unstable, non-reproducible results ADDIN EN.CITE <EndNote><Cite><Author>Sweeney</Author><Year>2015</Year><IDText>Combined Mapping of Multiple clUsteriNg ALgorithms (COMMUNAL): A Robust Method for Selection of Cluster Number, K</IDText><DisplayText>(2)</DisplayText><record><urls><related-urls><url> Mapping of Multiple clUsteriNg ALgorithms (COMMUNAL): A Robust Method for Selection of Cluster Number, K</title><secondary-title>Sci Rep</secondary-title></titles><pages>16971</pages><contributors><authors><author>Sweeney, T. E.</author><author>Chen, A. C.</author><author>Gevaert, O.</author></authors></contributors><language>eng</language><added-date format="utc">1449857753</added-date><ref-type name="Journal Article">17</ref-type><dates><year>2015</year></dates><rec-number>966</rec-number><last-updated-date format="utc">1449857753</last-updated-date><accession-num>26581809</accession-num><electronic-resource-num>10.1038/srep16971</electronic-resource-num><volume>5</volume></record></Cite></EndNote>(2). Here we used the COmbined Mapping of Multiple clUsteriNg ALgorithms (COMMUNAL) method, which integrates data from multiple clustering algorithms and validity metrics across a range of included variables to identify the most robust number of clusters present in the data ADDIN EN.CITE <EndNote><Cite><Author>Sweeney</Author><Year>2015</Year><IDText>Combined Mapping of Multiple clUsteriNg ALgorithms (COMMUNAL): A Robust Method for Selection of Cluster Number, K</IDText><DisplayText>(2)</DisplayText><record><urls><related-urls><url> Mapping of Multiple clUsteriNg ALgorithms (COMMUNAL): A Robust Method for Selection of Cluster Number, K</title><secondary-title>Sci Rep</secondary-title></titles><pages>16971</pages><contributors><authors><author>Sweeney, T. E.</author><author>Chen, A. C.</author><author>Gevaert, O.</author></authors></contributors><language>eng</language><added-date format="utc">1449857753</added-date><ref-type name="Journal Article">17</ref-type><dates><year>2015</year></dates><rec-number>966</rec-number><last-updated-date format="utc">1449857753</last-updated-date><accession-num>26581809</accession-num><electronic-resource-num>10.1038/srep16971</electronic-resource-num><volume>5</volume></record></Cite></EndNote>(2).In unsupervised clustering, high-dimensional distance calculations between samples are used to identify sub-groupings in the data. It is thus important to include variables (here, genes) that are likely to be informative while minimizing non-informative variables, so as to increase the signal-to-noise ratio. In a typical single-dataset clustering algorithm, usually some measure of variance is used to rank variables. Across multiple co-clustered datasets, however, the potential for high variance due to inter-dataset technical differences means that this metric may be less useful. We thus ranked the top 5,000 genes across the discovery datasets using an algorithm that accounts for both within-dataset variance and between-dataset variance (measured via mean absolute deviance) ADDIN EN.CITE <EndNote><Cite><Author>Planey</Author><Year>2016</Year><IDText>CoINcIDE: A framework for discovery of patient subtypes across multiple datasets</IDText><DisplayText>(3)</DisplayText><record><dates><pub-dates><date>Mar</date></pub-dates><year>2016</year></dates><keywords><keyword>Algorithms</keyword><keyword>Breast Neoplasms</keyword><keyword>Cluster Analysis</keyword><keyword>Computational Biology</keyword><keyword>Computer Simulation</keyword><keyword>Datasets as Topic</keyword><keyword>Female</keyword><keyword>Gene Expression Profiling</keyword><keyword>Humans</keyword><keyword>Ovarian Neoplasms</keyword><keyword>Prognosis</keyword><keyword>ROC Curve</keyword><keyword>Software</keyword></keywords><urls><related-urls><url>: A framework for discovery of patient subtypes across multiple datasets</title><secondary-title>Genome Med</secondary-title></titles><pages>27</pages><number>1</number><contributors><authors><author>Planey, C. R.</author><author>Gevaert, O.</author></authors></contributors><language>eng</language><added-date format="utc">1481328847</added-date><ref-type name="Journal Article">17</ref-type><rec-number>1132</rec-number><last-updated-date format="utc">1481328847</last-updated-date><accession-num>26961683</accession-num><electronic-resource-num>10.1186/s13073-016-0281-4</electronic-resource-num><volume>8</volume></record></Cite></EndNote>(3). The algorithm works as follows: we first use median absolute deviation (MAD) to rank all genes within each dataset, such that genes with the largest MAD are ranked highest. The median overall ranking is calculated across datasets. However, since the distribution of clusters may be different in each dataset, this meta-ranking may down-weight informative genes from unevenly distributed datasets. Thus, the top 20 genes from each individual dataset are also included (this number is arbitrary but set as default by the original algorithm authors). A final meta-ranking algorithm incorporates the top individual and pooled gene rankings into a single list. Further details can be found in the original manuscript by Planey & Gevaert (Genome Med, 2016), and in the accompanying software package ‘Coincide’ (). These ranked genes were then progressively included in the COMMUNAL algorithm. Gene ontology testingTo validate whether the different clusters were indicative of different biology, we assigned each of the genes used in the final clustering to the cluster in which the gene had the highest absolute effect size using the Significance Analysis for Microarrays (SAM) method ADDIN EN.CITE <EndNote><Cite><Author>Tusher</Author><Year>2001</Year><IDText>Significance analysis of microarrays applied to the ionizing radiation response</IDText><DisplayText>(4)</DisplayText><record><dates><pub-dates><date>Apr</date></pub-dates><year>2001</year></dates><keywords><keyword>Apoptosis</keyword><keyword>Cell Cycle</keyword><keyword>DNA Damage</keyword><keyword>DNA Repair</keyword><keyword>Down-Regulation</keyword><keyword>Gene Expression Profiling</keyword><keyword>Gene Expression Regulation</keyword><keyword>Humans</keyword><keyword>Oligonucleotide Array Sequence Analysis</keyword><keyword>RNA, Messenger</keyword><keyword>Radiation, Ionizing</keyword><keyword>Reproducibility of Results</keyword><keyword>Statistics as Topic</keyword><keyword>Tumor Cells, Cultured</keyword><keyword>Up-Regulation</keyword></keywords><urls><related-urls><url> analysis of microarrays applied to the ionizing radiation response</title><secondary-title>Proc Natl Acad Sci U S A</secondary-title></titles><pages>5116-21</pages><number>9</number><contributors><authors><author>Tusher, V. G.</author><author>Tibshirani, R.</author><author>Chu, G.</author></authors></contributors><language>eng</language><added-date format="utc">1387591713</added-date><ref-type name="Journal Article">17</ref-type><rec-number>169</rec-number><last-updated-date format="utc">1402432042</last-updated-date><accession-num>11309499</accession-num><electronic-resource-num>10.1073/pnas.091062498</electronic-resource-num><volume>98</volume></record></Cite></EndNote>(4). Since these genes had a generally high variance across samples, higher differential expression of a gene within a given cluster suggests it contributes to that cluster’s identity. Gene ontology (GO) enrichment was performed using ToppGene for the resulting gene lists ADDIN EN.CITE <EndNote><Cite><Author>Chen</Author><Year>2009</Year><RecNum>0</RecNum><IDText>ToppGene Suite for gene list enrichment analysis and candidate gene prioritization</IDText><DisplayText>(5)</DisplayText><record><dates><pub-dates><date>Jul</date></pub-dates><year>2009</year></dates><keywords><keyword>Animals</keyword><keyword>Disease</keyword><keyword>Genes</keyword><keyword>Humans</keyword><keyword>Internet</keyword><keyword>Mice</keyword><keyword>Protein Interaction Mapping</keyword><keyword>Proteins</keyword><keyword>Software</keyword></keywords><urls><related-urls><url> Suite for gene list enrichment analysis and candidate gene prioritization</title><secondary-title>Nucleic Acids Res</secondary-title></titles><pages>W305-11</pages><number>Web Server issue</number><contributors><authors><author>Chen, J.</author><author>Bardes, E. E.</author><author>Aronow, B. J.</author><author>Jegga, A. G.</author></authors></contributors><edition>2009/05/22</edition><language>eng</language><added-date format="utc">1485307725</added-date><ref-type name="Journal Article">17</ref-type><rec-number>1188</rec-number><last-updated-date format="utc">1485307725</last-updated-date><accession-num>19465376</accession-num><electronic-resource-num>10.1093/nar/gkp427</electronic-resource-num><volume>37</volume></record></Cite></EndNote>(5). A Benjamini-Hochberg corrected p-value smaller than 0.05 was used as the significance threshold. Cluster classifier and application in validation datasetsExternal validation is a key component of any exercise in clustering. However, in validation, it is important to switch to a supervised method (classification) rather than continuing to simply used unsupervised clustering in new validation datasets. There are two primary reasons for this. First, a de novo clustering does not produce labels. If we ran a clustering on each new dataset, and it produced 3 clusters (call them A, B, C), we would have no way of matching the new clusters to the discovery (Inflammopathic/Coagulopathic/Adaptive) clusters. We would instead have to rely on trying to ‘pattern match’ the closest phenotypic and molecular profiles (e.g. C = Inflammopathic, A = Coagulopathic, B = Adaptive) but this clearly introduces a large bias. The classifier, on the other hand, directly produces a label; thus we can directly ask whether a validation sample classified as ‘Inflammopathic’ matches the discovery ‘Inflammopathic’ phenotypic and molecular profile, which is a more relevant clinical question. The second reason to derive a classifier is that without one, there is no way to assign a single new patient to a sample in a clinical setting. This is because clustering relies on the presence of an entire cohort to establish relative distances between samples, and so can only be done retrospectively. In contrast, classification can determine a subtype assignment for a single patient prospectively. If, for instance, it was necessary to identify to which cluster a patient belongs when he or she were admitted to hospital, a validated classifier would be necessary. We thus built a gene-expression-based classifier for the resulting clusters using a two-step process in a 1-vs.-all, round-robin fashion for all clusters using all genes. First, we used SAM examining all genes (not just the genes used in clustering) to find genes statistically significantly associate with a given cluster. We then used a greedy forward search to find a gene set that maximally separated the given cluster from all other clusters ADDIN EN.CITE <EndNote><Cite><Author>Sweeney</Author><Year>2015</Year><IDText>A comprehensive time-course-based multicohort analysis of sepsis and sterile inflammation reveals a robust diagnostic gene set</IDText><DisplayText>(6)</DisplayText><record><dates><pub-dates><date>May</date></pub-dates><year>2015</year></dates><urls><related-urls><url> comprehensive time-course-based multicohort analysis of sepsis and sterile inflammation reveals a robust diagnostic gene set</title><secondary-title>Sci Transl Med</secondary-title></titles><pages>287ra71</pages><number>287</number><contributors><authors><author>Sweeney, T. E.</author><author>Shidham, A.</author><author>Wong, H. R.</author><author>Khatri, P.</author></authors></contributors><language>eng</language><added-date format="utc">1432159430</added-date><ref-type name="Journal Article">17</ref-type><rec-number>755</rec-number><last-updated-date format="utc">1432159430</last-updated-date><accession-num>25972003</accession-num><electronic-resource-num>10.1126/scitranslmed.aaa5993</electronic-resource-num><volume>7</volume></record></Cite></EndNote>(6). If there are K clusters, such a method would produce K scores; we thus fit a multiclass logistic regression model to the K scores as the final classifier using the R package nnet. Thus, to apply the classifier to an external dataset, one would need to calculate each cluster’s score, and then apply the multiclass logistic regression model to the set of scores to get an assigned outcome (see main Methods). We tested the classifier in the discovery data using leave-one-out cross validation to estimate its accuracy in validation applications. Testing of the validation datasetsWe applied the classifier to the validation datasets, and for each validation dataset, demographic and phenotypic data for each assigned cluster were calculated. Since these datasets were tested separately, the data are presented both as means of output from each validation dataset and as a pooled output. We next examined whether the cluster assignments in the validation clusters were exhibiting the same biology as their matching clusters in the discovery data. We first scaled each gene within its local dataset, and then took the mean of each gene within each assigned cluster within each dataset. This left us with a vector of mean differences for each gene within each cluster. We then correlated these mean difference vectors across the discovery clusters and all validation clusters, and plotted the results as a heatmap.We also took a pathways-based approach to confirm the consistency of the biology between discovery and validation clusters. Within each validation cohort, we used the same SAM method as above to assign overrepresented GO pathways to each validation cluster. We then tested every GO pathway that was found to be significant in discovery clusters in every validation cluster, and plotted the resulting significance levels as a heatmap, with row (GO) order determined by significance level in the discovery clusters.E-value sensitivity analysisIn order to address the possibility of unmeasured confounding in sample assignment, we performed a sensitivity analysis based on the ‘E-value’ ADDIN EN.CITE <EndNote><Cite><Author>VanderWeele</Author><Year>2017</Year><IDText>Sensitivity Analysis in Observational Research: Introducing the E-Value</IDText><DisplayText>(7)</DisplayText><record><dates><pub-dates><date>Jul</date></pub-dates><year>2017</year></dates><urls><related-urls><url> Analysis in Observational Research: Introducing the E-Value</title><secondary-title>Ann Intern Med</secondary-title></titles><contributors><authors><author>VanderWeele, T. J.</author><author>Ding, P.</author></authors></contributors><edition>2017/07/11</edition><language>eng</language><added-date format="utc">1502343665</added-date><ref-type name="Journal Article">17</ref-type><rec-number>1202</rec-number><last-updated-date format="utc">1502343665</last-updated-date><accession-num>28693043</accession-num><electronic-resource-num>10.7326/M16-2607</electronic-resource-num></record></Cite></EndNote>(7). In this method, a given risk ratio (RR) is used to determine an ‘E-value’, which is the effect size that an unmeasured confounder would have to have on both the explanatory variable and the outcome variable in order to explain away the observed RR. In order to put the E-value range into context, we then test the RRs of the already measured potentially confounding variables (here age, shock, and severity) for both the explanatory variable (cluster assignment) and the outcome variable (mortality). The resulting RRs are then compared to the calculated E-values to determine how much greater of an effect size a potential confounder would have to have in order to explain away the observed effect. In this application, it helps test the relationship between cluster assignment and mortality.Cluster ClassifierThe different datasets encompassed a broad range of microarray types, so we built a two-stage method of classification wherein a generative model (regression) is run on the outputs from parameter-free algorithms (differential gene expression), thereby overcoming technical differences between microarrays. Thus, there are two stages to the classifier: the first produces three cluster-specific scores by looking at differential gene expression. Each of three cluster-specific scores is calculated by computing the geometric mean of the ‘up’ genes for the given cluster, and subtracting the geometric mean of the ‘down’ genes. Thus, for example, the ‘Inflammopathic’ score is calculated as: (ARG1*LCN2*LTF*OLFM4)^(1/4) – (HLADMB). In the second stage, a multiclass regression algorithm takes each of these three scores for each sample and produces a final prediction. This two-stage process was necessary to utilize the full breadth of data across a wide range of microarray types. Clinical parametersOne key clinical variable is clinical severity, as measured by a standardized score. Since each of the different datasets used different clinical severity scores (e.g. APACHE II, SOFA, SAPS, PRISM), we could not pool these scores across datasets. Instead, for each dataset, we calculated the mean clinical severity score, and then labelled patients as either ‘high clinical severity’ (greater than the mean) or ‘low clinical severity’ (less than the mean). The percent of ‘high clinical severity’ patients within each cluster within each dataset was then calculated as a way of testing for clinical severity across the different datasets.Immunosuppression status was available for two datasets (GSE63042 and GSE66099). In each case the category was binary (either immunosuppressed or not) as retrospectively recorded by the enrolling team. For GSE66099 we do not have the exact criteria; for GSE63042, the composite category included absolute neutropenia, AIDS, chronic immunosuppressants or corticosteroids, chemotherapy, or ‘other immunosuppressants’. The categories are thus believed to be heterogeneous.General methodsAll analyses were conducted in the R statistical computing language, version 3.1.1. Categorical data were tested with Chi-square or Fisher Exact, and continuous data were tested with ANOVA. Significance was set at P<0.05 unless otherwise specified.Supplemental ResultsDiscovery cluster assignmentAt 500 genes there was an 84% agreement between the K-means and PAM algorithms in assigning samples to the 3 clusters. We removed the 16% of samples (N=112) with disagreeing assignments as ‘unclustered’, while the remaining samples were assigned to discovery clusters (Supplemental Table 1). There were varying distributions of clusters across datasets (Supplemental Table 2), which was expected given the varying enrollment criteria of each dataset. In order to evaluate if the 500-gene subset was capturing most of the variance across all measured genes in the discovery data, we performed principal components analysis (PCA). Both PCAs showed clear separation among the three clusters, with the ‘unclustered’ samples distributed among the three clusters (Supplemental Figure 4). A heatmap of the 500 genes used in clustering also showed clear differences between the clusters, as expected (Supplemental Figure 5).Clinical coagulopathy in the Coagulopathic clusterTo investigate whether the Coagulopathic cluster had functional evidence of coagulopathy, we studied whether standard measures of coagulopathy were differentially distributed across the three clusters. In the only cohort for which we had access to these data (pediatric ICU, GSE66099), disseminated intravascular coagulation (DIC) was significantly associated with the Coagulopathic cluster (P<0.05, Supplemental Table 9). In another dataset (adults, CAPSOD, GSE63042), the intersection of thrombocytopenia (platelets <100,000) and prolonged INR (>1.3) was also significantly associated with the Coagulopathic cluster (P<0.05, Table 4), though neither parameter on its own was significantly associated with cluster type (Supplemental Table 10). These findings suggest that the Coagulopathic cluster may be associated with advanced forms of coagulopathy such as DIC, but not thrombocytopenia alone. Comparison to previously established sepsis endotypesTwo groups have previously performed clustering using sepsis transcriptomic profiles. Wong et al. (discovery dataset GSE66099) derived three endotypes of pediatric sepsis, and have since validated two: Endotype A (higher mortality, with adaptive immune suppression and decreased glucocorticoid receptor signaling) and Endotype B (lower mortality) PEVuZE5vdGU+PENpdGU+PEF1dGhvcj5Xb25nPC9BdXRob3I+PFllYXI+MjAwOTwvWWVhcj48SURU

ZXh0PklkZW50aWZpY2F0aW9uIG9mIHBlZGlhdHJpYyBzZXB0aWMgc2hvY2sgc3ViY2xhc3NlcyBi

YXNlZCBvbiBnZW5vbWUtd2lkZSBleHByZXNzaW9uIHByb2ZpbGluZzwvSURUZXh0PjxEaXNwbGF5

VGV4dD4oOS0xMSk8L0Rpc3BsYXlUZXh0PjxyZWNvcmQ+PGtleXdvcmRzPjxrZXl3b3JkPkJsb29k

IENlbGxzPC9rZXl3b3JkPjxrZXl3b3JkPkNoaWxkPC9rZXl3b3JkPjxrZXl3b3JkPkNoaWxkLCBQ

cmVzY2hvb2w8L2tleXdvcmQ+PGtleXdvcmQ+Q2x1c3RlciBBbmFseXNpczwva2V5d29yZD48a2V5

d29yZD5EYXRhIEludGVycHJldGF0aW9uLCBTdGF0aXN0aWNhbDwva2V5d29yZD48a2V5d29yZD5G

ZW1hbGU8L2tleXdvcmQ+PGtleXdvcmQ+R2VuZSBFeHByZXNzaW9uIFByb2ZpbGluZzwva2V5d29y

ZD48a2V5d29yZD5IdW1hbnM8L2tleXdvcmQ+PGtleXdvcmQ+SW5mYW50PC9rZXl3b3JkPjxrZXl3

b3JkPk1hbGU8L2tleXdvcmQ+PGtleXdvcmQ+T2xpZ29udWNsZW90aWRlIEFycmF5IFNlcXVlbmNl

IEFuYWx5c2lzPC9rZXl3b3JkPjxrZXl3b3JkPlJOQTwva2V5d29yZD48a2V5d29yZD5TaG9jaywg

U2VwdGljPC9rZXl3b3JkPjwva2V5d29yZHM+PHVybHM+PHJlbGF0ZWQtdXJscz48dXJsPmh0dHA6

Ly93d3cubmNiaS5ubG0ubmloLmdvdi9wdWJtZWQvMTk2MjQ4MDk8L3VybD48L3JlbGF0ZWQtdXJs

cz48L3VybHM+PGlzYm4+MTc0MS03MDE1PC9pc2JuPjxjdXN0b20yPlBNQzI3MjA5ODc8L2N1c3Rv

bTI+PHRpdGxlcz48dGl0bGU+SWRlbnRpZmljYXRpb24gb2YgcGVkaWF0cmljIHNlcHRpYyBzaG9j

ayBzdWJjbGFzc2VzIGJhc2VkIG9uIGdlbm9tZS13aWRlIGV4cHJlc3Npb24gcHJvZmlsaW5nPC90

aXRsZT48c2Vjb25kYXJ5LXRpdGxlPkJNQyBNZWQ8L3NlY29uZGFyeS10aXRsZT48L3RpdGxlcz48

cGFnZXM+MzQ8L3BhZ2VzPjxjb250cmlidXRvcnM+PGF1dGhvcnM+PGF1dGhvcj5Xb25nLCBILiBS

LjwvYXV0aG9yPjxhdXRob3I+Q3ZpamFub3ZpY2gsIE4uPC9hdXRob3I+PGF1dGhvcj5MaW4sIFIu

PC9hdXRob3I+PGF1dGhvcj5BbGxlbiwgRy4gTC48L2F1dGhvcj48YXV0aG9yPlRob21hcywgTi4g

Si48L2F1dGhvcj48YXV0aG9yPldpbGxzb24sIEQuIEYuPC9hdXRob3I+PGF1dGhvcj5GcmVpc2h0

YXQsIFIuIEouPC9hdXRob3I+PGF1dGhvcj5BbmFzLCBOLjwvYXV0aG9yPjxhdXRob3I+TWV5ZXIs

IEsuPC9hdXRob3I+PGF1dGhvcj5DaGVjY2hpYSwgUC4gQS48L2F1dGhvcj48YXV0aG9yPk1vbmFj

bywgTS48L2F1dGhvcj48YXV0aG9yPk9kb20sIEsuPC9hdXRob3I+PGF1dGhvcj5TaGFubGV5LCBU

LiBQLjwvYXV0aG9yPjwvYXV0aG9ycz48L2NvbnRyaWJ1dG9ycz48bGFuZ3VhZ2U+ZW5nPC9sYW5n

dWFnZT48YWRkZWQtZGF0ZSBmb3JtYXQ9InV0YyI+MTM4OTIzNjQzNDwvYWRkZWQtZGF0ZT48cmVm

LXR5cGUgbmFtZT0iSm91cm5hbCBBcnRpY2xlIj4xNzwvcmVmLXR5cGU+PGRhdGVzPjx5ZWFyPjIw

MDk8L3llYXI+PC9kYXRlcz48cmVjLW51bWJlcj4xNzQ8L3JlYy1udW1iZXI+PGxhc3QtdXBkYXRl

ZC1kYXRlIGZvcm1hdD0idXRjIj4xMzg5MjM2NDM0PC9sYXN0LXVwZGF0ZWQtZGF0ZT48YWNjZXNz

aW9uLW51bT4xOTYyNDgwOTwvYWNjZXNzaW9uLW51bT48ZWxlY3Ryb25pYy1yZXNvdXJjZS1udW0+

MTAuMTE4Ni8xNzQxLTcwMTUtNy0zNDwvZWxlY3Ryb25pYy1yZXNvdXJjZS1udW0+PHZvbHVtZT43

PC92b2x1bWU+PC9yZWNvcmQ+PC9DaXRlPjxDaXRlPjxBdXRob3I+V29uZzwvQXV0aG9yPjxZZWFy

PjIwMTE8L1llYXI+PElEVGV4dD5WYWxpZGF0aW9uIG9mIGEgZ2VuZSBleHByZXNzaW9uLWJhc2Vk

IHN1YmNsYXNzaWZpY2F0aW9uIHN0cmF0ZWd5IGZvciBwZWRpYXRyaWMgc2VwdGljIHNob2NrPC9J

RFRleHQ+PHJlY29yZD48ZGF0ZXM+PHB1Yi1kYXRlcz48ZGF0ZT5Ob3Y8L2RhdGU+PC9wdWItZGF0

ZXM+PHllYXI+MjAxMTwveWVhcj48L2RhdGVzPjxrZXl3b3Jkcz48a2V5d29yZD5DaGlsZCwgUHJl

c2Nob29sPC9rZXl3b3JkPjxrZXl3b3JkPkZlbWFsZTwva2V5d29yZD48a2V5d29yZD5HZW5lIEV4

cHJlc3Npb248L2tleXdvcmQ+PGtleXdvcmQ+SHVtYW5zPC9rZXl3b3JkPjxrZXl3b3JkPkluZmFu

dDwva2V5d29yZD48a2V5d29yZD5JbnRlbnNpdmUgQ2FyZSBVbml0cywgUGVkaWF0cmljPC9rZXl3

b3JkPjxrZXl3b3JkPk1hbGU8L2tleXdvcmQ+PGtleXdvcmQ+UHJvZ25vc2lzPC9rZXl3b3JkPjxr

ZXl3b3JkPlByb3NwZWN0aXZlIFN0dWRpZXM8L2tleXdvcmQ+PGtleXdvcmQ+UHJvdGVpbiBBcnJh

eSBBbmFseXNpczwva2V5d29yZD48a2V5d29yZD5SaXNrIEZhY3RvcnM8L2tleXdvcmQ+PGtleXdv

cmQ+U2V2ZXJpdHkgb2YgSWxsbmVzcyBJbmRleDwva2V5d29yZD48a2V5d29yZD5TaG9jaywgU2Vw

dGljPC9rZXl3b3JkPjwva2V5d29yZHM+PHVybHM+PHJlbGF0ZWQtdXJscz48dXJsPmh0dHA6Ly93

d3cubmNiaS5ubG0ubmloLmdvdi9wdWJtZWQvMjE3MDU4ODU8L3VybD48L3JlbGF0ZWQtdXJscz48

L3VybHM+PGlzYm4+MTUzMC0wMjkzPC9pc2JuPjxjdXN0b20yPlBNQzMxOTY3NzY8L2N1c3RvbTI+

PHRpdGxlcz48dGl0bGU+VmFsaWRhdGlvbiBvZiBhIGdlbmUgZXhwcmVzc2lvbi1iYXNlZCBzdWJj

bGFzc2lmaWNhdGlvbiBzdHJhdGVneSBmb3IgcGVkaWF0cmljIHNlcHRpYyBzaG9jazwvdGl0bGU+

PHNlY29uZGFyeS10aXRsZT5Dcml0IENhcmUgTWVkPC9zZWNvbmRhcnktdGl0bGU+PC90aXRsZXM+

PHBhZ2VzPjI1MTEtNzwvcGFnZXM+PG51bWJlcj4xMTwvbnVtYmVyPjxjb250cmlidXRvcnM+PGF1

dGhvcnM+PGF1dGhvcj5Xb25nLCBILiBSLjwvYXV0aG9yPjxhdXRob3I+Q3ZpamFub3ZpY2gsIE4u

IFouPC9hdXRob3I+PGF1dGhvcj5BbGxlbiwgRy4gTC48L2F1dGhvcj48YXV0aG9yPlRob21hcywg

Ti4gSi48L2F1dGhvcj48YXV0aG9yPkZyZWlzaHRhdCwgUi4gSi48L2F1dGhvcj48YXV0aG9yPkFu

YXMsIE4uPC9hdXRob3I+PGF1dGhvcj5NZXllciwgSy48L2F1dGhvcj48YXV0aG9yPkNoZWNjaGlh

LCBQLiBBLjwvYXV0aG9yPjxhdXRob3I+TGluLCBSLjwvYXV0aG9yPjxhdXRob3I+U2hhbmxleSwg

VC4gUC48L2F1dGhvcj48YXV0aG9yPkJpZ2hhbSwgTS4gVC48L2F1dGhvcj48YXV0aG9yPldoZWVs

ZXIsIEQuIFMuPC9hdXRob3I+PGF1dGhvcj5Eb3VnaHR5LCBMLiBBLjwvYXV0aG9yPjxhdXRob3I+

VGVndG1leWVyLCBLLjwvYXV0aG9yPjxhdXRob3I+UG95bnRlciwgUy4gRS48L2F1dGhvcj48YXV0

aG9yPkthcGxhbiwgSi4gTS48L2F1dGhvcj48YXV0aG9yPkNoaW1hLCBSLiBTLjwvYXV0aG9yPjxh

dXRob3I+U3RhbGV0cywgRS48L2F1dGhvcj48YXV0aG9yPkJhc3UsIFIuIEsuPC9hdXRob3I+PGF1

dGhvcj5WYXJpc2NvLCBCLiBNLjwvYXV0aG9yPjxhdXRob3I+QmFyciwgRi4gRS48L2F1dGhvcj48

L2F1dGhvcnM+PC9jb250cmlidXRvcnM+PGxhbmd1YWdlPmVuZzwvbGFuZ3VhZ2U+PGFkZGVkLWRh

dGUgZm9ybWF0PSJ1dGMiPjEzODkyMzY0MzQ8L2FkZGVkLWRhdGU+PHJlZi10eXBlIG5hbWU9Ikpv

dXJuYWwgQXJ0aWNsZSI+MTc8L3JlZi10eXBlPjxyZWMtbnVtYmVyPjE3MjwvcmVjLW51bWJlcj48

bGFzdC11cGRhdGVkLWRhdGUgZm9ybWF0PSJ1dGMiPjEzODkyMzY0MzQ8L2xhc3QtdXBkYXRlZC1k

YXRlPjxhY2Nlc3Npb24tbnVtPjIxNzA1ODg1PC9hY2Nlc3Npb24tbnVtPjxlbGVjdHJvbmljLXJl

c291cmNlLW51bT4xMC4xMDk3L0NDTS4wYjAxM2UzMTgyMjU3Njc1PC9lbGVjdHJvbmljLXJlc291

cmNlLW51bT48dm9sdW1lPjM5PC92b2x1bWU+PC9yZWNvcmQ+PC9DaXRlPjxDaXRlPjxBdXRob3I+

V29uZzwvQXV0aG9yPjxZZWFyPjIwMTU8L1llYXI+PElEVGV4dD5EZXZlbG9waW5nIGEgY2xpbmlj

YWxseSBmZWFzaWJsZSBwZXJzb25hbGl6ZWQgbWVkaWNpbmUgYXBwcm9hY2ggdG8gcGVkaWF0cmlj

IHNlcHRpYyBzaG9jazwvSURUZXh0PjxyZWNvcmQ+PGRhdGVzPjxwdWItZGF0ZXM+PGRhdGU+RmVi

PC9kYXRlPjwvcHViLWRhdGVzPjx5ZWFyPjIwMTU8L3llYXI+PC9kYXRlcz48a2V5d29yZHM+PGtl

eXdvcmQ+Q2hpbGQ8L2tleXdvcmQ+PGtleXdvcmQ+Q2hpbGQsIFByZXNjaG9vbDwva2V5d29yZD48

a2V5d29yZD5GZWFzaWJpbGl0eSBTdHVkaWVzPC9rZXl3b3JkPjxrZXl3b3JkPkZlbWFsZTwva2V5

d29yZD48a2V5d29yZD5HZW5lIEV4cHJlc3Npb24gUmVndWxhdGlvbjwva2V5d29yZD48a2V5d29y

ZD5HbHVjb2NvcnRpY29pZHM8L2tleXdvcmQ+PGtleXdvcmQ+SHVtYW5zPC9rZXl3b3JkPjxrZXl3

b3JkPkluZGl2aWR1YWxpemVkIE1lZGljaW5lPC9rZXl3b3JkPjxrZXl3b3JkPkluZmFudDwva2V5

d29yZD48a2V5d29yZD5JbnRlbnNpdmUgQ2FyZSBVbml0cywgUGVkaWF0cmljPC9rZXl3b3JkPjxr

ZXl3b3JkPk1hbGU8L2tleXdvcmQ+PGtleXdvcmQ+TWF0aGVtYXRpY2FsIENvbXB1dGluZzwva2V5

d29yZD48a2V5d29yZD5PZGRzIFJhdGlvPC9rZXl3b3JkPjxrZXl3b3JkPlBoZW5vdHlwZTwva2V5

d29yZD48a2V5d29yZD5Qcm9zcGVjdGl2ZSBTdHVkaWVzPC9rZXl3b3JkPjxrZXl3b3JkPlJlcHJv

ZHVjaWJpbGl0eSBvZiBSZXN1bHRzPC9rZXl3b3JkPjxrZXl3b3JkPlNldmVyaXR5IG9mIElsbG5l

c3MgSW5kZXg8L2tleXdvcmQ+PGtleXdvcmQ+U2hvY2ssIFNlcHRpYzwva2V5d29yZD48a2V5d29y

ZD5TaWduYWwgVHJhbnNkdWN0aW9uPC9rZXl3b3JkPjwva2V5d29yZHM+PHVybHM+PHJlbGF0ZWQt

dXJscz48dXJsPmh0dHA6Ly93d3cubmNiaS5ubG0ubmloLmdvdi9wdWJtZWQvMjU0ODk4ODE8L3Vy

bD48L3JlbGF0ZWQtdXJscz48L3VybHM+PGlzYm4+MTUzNS00OTcwPC9pc2JuPjxjdXN0b20yPlBN

QzQzNTE1ODA8L2N1c3RvbTI+PHRpdGxlcz48dGl0bGU+RGV2ZWxvcGluZyBhIGNsaW5pY2FsbHkg

ZmVhc2libGUgcGVyc29uYWxpemVkIG1lZGljaW5lIGFwcHJvYWNoIHRvIHBlZGlhdHJpYyBzZXB0

aWMgc2hvY2s8L3RpdGxlPjxzZWNvbmRhcnktdGl0bGU+QW0gSiBSZXNwaXIgQ3JpdCBDYXJlIE1l

ZDwvc2Vjb25kYXJ5LXRpdGxlPjwvdGl0bGVzPjxwYWdlcz4zMDktMTU8L3BhZ2VzPjxudW1iZXI+

MzwvbnVtYmVyPjxjb250cmlidXRvcnM+PGF1dGhvcnM+PGF1dGhvcj5Xb25nLCBILiBSLjwvYXV0

aG9yPjxhdXRob3I+Q3ZpamFub3ZpY2gsIE4uIFouPC9hdXRob3I+PGF1dGhvcj5BbmFzLCBOLjwv

YXV0aG9yPjxhdXRob3I+QWxsZW4sIEcuIEwuPC9hdXRob3I+PGF1dGhvcj5UaG9tYXMsIE4uIEou

PC9hdXRob3I+PGF1dGhvcj5CaWdoYW0sIE0uIFQuPC9hdXRob3I+PGF1dGhvcj5XZWlzcywgUy4g

TC48L2F1dGhvcj48YXV0aG9yPkZpdHpnZXJhbGQsIEouPC9hdXRob3I+PGF1dGhvcj5DaGVjY2hp

YSwgUC4gQS48L2F1dGhvcj48YXV0aG9yPk1leWVyLCBLLjwvYXV0aG9yPjxhdXRob3I+U2hhbmxl

eSwgVC4gUC48L2F1dGhvcj48YXV0aG9yPlF1YXNuZXksIE0uPC9hdXRob3I+PGF1dGhvcj5IYWxs

LCBNLjwvYXV0aG9yPjxhdXRob3I+R2VkZWl0LCBSLjwvYXV0aG9yPjxhdXRob3I+RnJlaXNodGF0

LCBSLiBKLjwvYXV0aG9yPjxhdXRob3I+Tm93YWssIEouPC9hdXRob3I+PGF1dGhvcj5TaGVraGFy

LCBSLiBTLjwvYXV0aG9yPjxhdXRob3I+R2VydHosIFMuPC9hdXRob3I+PGF1dGhvcj5EYXdzb24s

IEUuPC9hdXRob3I+PGF1dGhvcj5Ib3dhcmQsIEsuPC9hdXRob3I+PGF1dGhvcj5IYXJtb24sIEsu

PC9hdXRob3I+PGF1dGhvcj5CZWNrbWFuLCBFLjwvYXV0aG9yPjxhdXRob3I+RnJhbmssIEUuPC9h

dXRob3I+PGF1dGhvcj5MaW5kc2VsbCwgQy4gSi48L2F1dGhvcj48L2F1dGhvcnM+PC9jb250cmli

dXRvcnM+PGxhbmd1YWdlPmVuZzwvbGFuZ3VhZ2U+PGFkZGVkLWRhdGUgZm9ybWF0PSJ1dGMiPjE0

Mzc1ODk3OTk8L2FkZGVkLWRhdGU+PHJlZi10eXBlIG5hbWU9IkpvdXJuYWwgQXJ0aWNsZSI+MTc8

L3JlZi10eXBlPjxyZWMtbnVtYmVyPjc5MzwvcmVjLW51bWJlcj48bGFzdC11cGRhdGVkLWRhdGUg

Zm9ybWF0PSJ1dGMiPjE0Mzc1ODk3OTk8L2xhc3QtdXBkYXRlZC1kYXRlPjxhY2Nlc3Npb24tbnVt

PjI1NDg5ODgxPC9hY2Nlc3Npb24tbnVtPjxlbGVjdHJvbmljLXJlc291cmNlLW51bT4xMC4xMTY0

L3JjY20uMjAxNDEwLTE4NjRPQzwvZWxlY3Ryb25pYy1yZXNvdXJjZS1udW0+PHZvbHVtZT4xOTE8

L3ZvbHVtZT48L3JlY29yZD48L0NpdGU+PC9FbmROb3RlPgB=

ADDIN EN.CITE PEVuZE5vdGU+PENpdGU+PEF1dGhvcj5Xb25nPC9BdXRob3I+PFllYXI+MjAwOTwvWWVhcj48SURU

ZXh0PklkZW50aWZpY2F0aW9uIG9mIHBlZGlhdHJpYyBzZXB0aWMgc2hvY2sgc3ViY2xhc3NlcyBi

YXNlZCBvbiBnZW5vbWUtd2lkZSBleHByZXNzaW9uIHByb2ZpbGluZzwvSURUZXh0PjxEaXNwbGF5

VGV4dD4oOS0xMSk8L0Rpc3BsYXlUZXh0PjxyZWNvcmQ+PGtleXdvcmRzPjxrZXl3b3JkPkJsb29k

IENlbGxzPC9rZXl3b3JkPjxrZXl3b3JkPkNoaWxkPC9rZXl3b3JkPjxrZXl3b3JkPkNoaWxkLCBQ

cmVzY2hvb2w8L2tleXdvcmQ+PGtleXdvcmQ+Q2x1c3RlciBBbmFseXNpczwva2V5d29yZD48a2V5

d29yZD5EYXRhIEludGVycHJldGF0aW9uLCBTdGF0aXN0aWNhbDwva2V5d29yZD48a2V5d29yZD5G

ZW1hbGU8L2tleXdvcmQ+PGtleXdvcmQ+R2VuZSBFeHByZXNzaW9uIFByb2ZpbGluZzwva2V5d29y

ZD48a2V5d29yZD5IdW1hbnM8L2tleXdvcmQ+PGtleXdvcmQ+SW5mYW50PC9rZXl3b3JkPjxrZXl3

b3JkPk1hbGU8L2tleXdvcmQ+PGtleXdvcmQ+T2xpZ29udWNsZW90aWRlIEFycmF5IFNlcXVlbmNl

IEFuYWx5c2lzPC9rZXl3b3JkPjxrZXl3b3JkPlJOQTwva2V5d29yZD48a2V5d29yZD5TaG9jaywg

U2VwdGljPC9rZXl3b3JkPjwva2V5d29yZHM+PHVybHM+PHJlbGF0ZWQtdXJscz48dXJsPmh0dHA6

Ly93d3cubmNiaS5ubG0ubmloLmdvdi9wdWJtZWQvMTk2MjQ4MDk8L3VybD48L3JlbGF0ZWQtdXJs

cz48L3VybHM+PGlzYm4+MTc0MS03MDE1PC9pc2JuPjxjdXN0b20yPlBNQzI3MjA5ODc8L2N1c3Rv

bTI+PHRpdGxlcz48dGl0bGU+SWRlbnRpZmljYXRpb24gb2YgcGVkaWF0cmljIHNlcHRpYyBzaG9j

ayBzdWJjbGFzc2VzIGJhc2VkIG9uIGdlbm9tZS13aWRlIGV4cHJlc3Npb24gcHJvZmlsaW5nPC90

aXRsZT48c2Vjb25kYXJ5LXRpdGxlPkJNQyBNZWQ8L3NlY29uZGFyeS10aXRsZT48L3RpdGxlcz48

cGFnZXM+MzQ8L3BhZ2VzPjxjb250cmlidXRvcnM+PGF1dGhvcnM+PGF1dGhvcj5Xb25nLCBILiBS

LjwvYXV0aG9yPjxhdXRob3I+Q3ZpamFub3ZpY2gsIE4uPC9hdXRob3I+PGF1dGhvcj5MaW4sIFIu

PC9hdXRob3I+PGF1dGhvcj5BbGxlbiwgRy4gTC48L2F1dGhvcj48YXV0aG9yPlRob21hcywgTi4g

Si48L2F1dGhvcj48YXV0aG9yPldpbGxzb24sIEQuIEYuPC9hdXRob3I+PGF1dGhvcj5GcmVpc2h0

YXQsIFIuIEouPC9hdXRob3I+PGF1dGhvcj5BbmFzLCBOLjwvYXV0aG9yPjxhdXRob3I+TWV5ZXIs

IEsuPC9hdXRob3I+PGF1dGhvcj5DaGVjY2hpYSwgUC4gQS48L2F1dGhvcj48YXV0aG9yPk1vbmFj

bywgTS48L2F1dGhvcj48YXV0aG9yPk9kb20sIEsuPC9hdXRob3I+PGF1dGhvcj5TaGFubGV5LCBU

LiBQLjwvYXV0aG9yPjwvYXV0aG9ycz48L2NvbnRyaWJ1dG9ycz48bGFuZ3VhZ2U+ZW5nPC9sYW5n

dWFnZT48YWRkZWQtZGF0ZSBmb3JtYXQ9InV0YyI+MTM4OTIzNjQzNDwvYWRkZWQtZGF0ZT48cmVm

LXR5cGUgbmFtZT0iSm91cm5hbCBBcnRpY2xlIj4xNzwvcmVmLXR5cGU+PGRhdGVzPjx5ZWFyPjIw

MDk8L3llYXI+PC9kYXRlcz48cmVjLW51bWJlcj4xNzQ8L3JlYy1udW1iZXI+PGxhc3QtdXBkYXRl

ZC1kYXRlIGZvcm1hdD0idXRjIj4xMzg5MjM2NDM0PC9sYXN0LXVwZGF0ZWQtZGF0ZT48YWNjZXNz

aW9uLW51bT4xOTYyNDgwOTwvYWNjZXNzaW9uLW51bT48ZWxlY3Ryb25pYy1yZXNvdXJjZS1udW0+

MTAuMTE4Ni8xNzQxLTcwMTUtNy0zNDwvZWxlY3Ryb25pYy1yZXNvdXJjZS1udW0+PHZvbHVtZT43

PC92b2x1bWU+PC9yZWNvcmQ+PC9DaXRlPjxDaXRlPjxBdXRob3I+V29uZzwvQXV0aG9yPjxZZWFy

PjIwMTE8L1llYXI+PElEVGV4dD5WYWxpZGF0aW9uIG9mIGEgZ2VuZSBleHByZXNzaW9uLWJhc2Vk

IHN1YmNsYXNzaWZpY2F0aW9uIHN0cmF0ZWd5IGZvciBwZWRpYXRyaWMgc2VwdGljIHNob2NrPC9J

RFRleHQ+PHJlY29yZD48ZGF0ZXM+PHB1Yi1kYXRlcz48ZGF0ZT5Ob3Y8L2RhdGU+PC9wdWItZGF0

ZXM+PHllYXI+MjAxMTwveWVhcj48L2RhdGVzPjxrZXl3b3Jkcz48a2V5d29yZD5DaGlsZCwgUHJl

c2Nob29sPC9rZXl3b3JkPjxrZXl3b3JkPkZlbWFsZTwva2V5d29yZD48a2V5d29yZD5HZW5lIEV4

cHJlc3Npb248L2tleXdvcmQ+PGtleXdvcmQ+SHVtYW5zPC9rZXl3b3JkPjxrZXl3b3JkPkluZmFu

dDwva2V5d29yZD48a2V5d29yZD5JbnRlbnNpdmUgQ2FyZSBVbml0cywgUGVkaWF0cmljPC9rZXl3

b3JkPjxrZXl3b3JkPk1hbGU8L2tleXdvcmQ+PGtleXdvcmQ+UHJvZ25vc2lzPC9rZXl3b3JkPjxr

ZXl3b3JkPlByb3NwZWN0aXZlIFN0dWRpZXM8L2tleXdvcmQ+PGtleXdvcmQ+UHJvdGVpbiBBcnJh

eSBBbmFseXNpczwva2V5d29yZD48a2V5d29yZD5SaXNrIEZhY3RvcnM8L2tleXdvcmQ+PGtleXdv

cmQ+U2V2ZXJpdHkgb2YgSWxsbmVzcyBJbmRleDwva2V5d29yZD48a2V5d29yZD5TaG9jaywgU2Vw

dGljPC9rZXl3b3JkPjwva2V5d29yZHM+PHVybHM+PHJlbGF0ZWQtdXJscz48dXJsPmh0dHA6Ly93

d3cubmNiaS5ubG0ubmloLmdvdi9wdWJtZWQvMjE3MDU4ODU8L3VybD48L3JlbGF0ZWQtdXJscz48

L3VybHM+PGlzYm4+MTUzMC0wMjkzPC9pc2JuPjxjdXN0b20yPlBNQzMxOTY3NzY8L2N1c3RvbTI+

PHRpdGxlcz48dGl0bGU+VmFsaWRhdGlvbiBvZiBhIGdlbmUgZXhwcmVzc2lvbi1iYXNlZCBzdWJj

bGFzc2lmaWNhdGlvbiBzdHJhdGVneSBmb3IgcGVkaWF0cmljIHNlcHRpYyBzaG9jazwvdGl0bGU+

PHNlY29uZGFyeS10aXRsZT5Dcml0IENhcmUgTWVkPC9zZWNvbmRhcnktdGl0bGU+PC90aXRsZXM+

PHBhZ2VzPjI1MTEtNzwvcGFnZXM+PG51bWJlcj4xMTwvbnVtYmVyPjxjb250cmlidXRvcnM+PGF1

dGhvcnM+PGF1dGhvcj5Xb25nLCBILiBSLjwvYXV0aG9yPjxhdXRob3I+Q3ZpamFub3ZpY2gsIE4u

IFouPC9hdXRob3I+PGF1dGhvcj5BbGxlbiwgRy4gTC48L2F1dGhvcj48YXV0aG9yPlRob21hcywg

Ti4gSi48L2F1dGhvcj48YXV0aG9yPkZyZWlzaHRhdCwgUi4gSi48L2F1dGhvcj48YXV0aG9yPkFu

YXMsIE4uPC9hdXRob3I+PGF1dGhvcj5NZXllciwgSy48L2F1dGhvcj48YXV0aG9yPkNoZWNjaGlh

LCBQLiBBLjwvYXV0aG9yPjxhdXRob3I+TGluLCBSLjwvYXV0aG9yPjxhdXRob3I+U2hhbmxleSwg

VC4gUC48L2F1dGhvcj48YXV0aG9yPkJpZ2hhbSwgTS4gVC48L2F1dGhvcj48YXV0aG9yPldoZWVs

ZXIsIEQuIFMuPC9hdXRob3I+PGF1dGhvcj5Eb3VnaHR5LCBMLiBBLjwvYXV0aG9yPjxhdXRob3I+

VGVndG1leWVyLCBLLjwvYXV0aG9yPjxhdXRob3I+UG95bnRlciwgUy4gRS48L2F1dGhvcj48YXV0

aG9yPkthcGxhbiwgSi4gTS48L2F1dGhvcj48YXV0aG9yPkNoaW1hLCBSLiBTLjwvYXV0aG9yPjxh

dXRob3I+U3RhbGV0cywgRS48L2F1dGhvcj48YXV0aG9yPkJhc3UsIFIuIEsuPC9hdXRob3I+PGF1

dGhvcj5WYXJpc2NvLCBCLiBNLjwvYXV0aG9yPjxhdXRob3I+QmFyciwgRi4gRS48L2F1dGhvcj48

L2F1dGhvcnM+PC9jb250cmlidXRvcnM+PGxhbmd1YWdlPmVuZzwvbGFuZ3VhZ2U+PGFkZGVkLWRh

dGUgZm9ybWF0PSJ1dGMiPjEzODkyMzY0MzQ8L2FkZGVkLWRhdGU+PHJlZi10eXBlIG5hbWU9Ikpv

dXJuYWwgQXJ0aWNsZSI+MTc8L3JlZi10eXBlPjxyZWMtbnVtYmVyPjE3MjwvcmVjLW51bWJlcj48

bGFzdC11cGRhdGVkLWRhdGUgZm9ybWF0PSJ1dGMiPjEzODkyMzY0MzQ8L2xhc3QtdXBkYXRlZC1k

YXRlPjxhY2Nlc3Npb24tbnVtPjIxNzA1ODg1PC9hY2Nlc3Npb24tbnVtPjxlbGVjdHJvbmljLXJl

c291cmNlLW51bT4xMC4xMDk3L0NDTS4wYjAxM2UzMTgyMjU3Njc1PC9lbGVjdHJvbmljLXJlc291

cmNlLW51bT48dm9sdW1lPjM5PC92b2x1bWU+PC9yZWNvcmQ+PC9DaXRlPjxDaXRlPjxBdXRob3I+

V29uZzwvQXV0aG9yPjxZZWFyPjIwMTU8L1llYXI+PElEVGV4dD5EZXZlbG9waW5nIGEgY2xpbmlj

YWxseSBmZWFzaWJsZSBwZXJzb25hbGl6ZWQgbWVkaWNpbmUgYXBwcm9hY2ggdG8gcGVkaWF0cmlj

IHNlcHRpYyBzaG9jazwvSURUZXh0PjxyZWNvcmQ+PGRhdGVzPjxwdWItZGF0ZXM+PGRhdGU+RmVi

PC9kYXRlPjwvcHViLWRhdGVzPjx5ZWFyPjIwMTU8L3llYXI+PC9kYXRlcz48a2V5d29yZHM+PGtl

eXdvcmQ+Q2hpbGQ8L2tleXdvcmQ+PGtleXdvcmQ+Q2hpbGQsIFByZXNjaG9vbDwva2V5d29yZD48

a2V5d29yZD5GZWFzaWJpbGl0eSBTdHVkaWVzPC9rZXl3b3JkPjxrZXl3b3JkPkZlbWFsZTwva2V5

d29yZD48a2V5d29yZD5HZW5lIEV4cHJlc3Npb24gUmVndWxhdGlvbjwva2V5d29yZD48a2V5d29y

ZD5HbHVjb2NvcnRpY29pZHM8L2tleXdvcmQ+PGtleXdvcmQ+SHVtYW5zPC9rZXl3b3JkPjxrZXl3

b3JkPkluZGl2aWR1YWxpemVkIE1lZGljaW5lPC9rZXl3b3JkPjxrZXl3b3JkPkluZmFudDwva2V5

d29yZD48a2V5d29yZD5JbnRlbnNpdmUgQ2FyZSBVbml0cywgUGVkaWF0cmljPC9rZXl3b3JkPjxr

ZXl3b3JkPk1hbGU8L2tleXdvcmQ+PGtleXdvcmQ+TWF0aGVtYXRpY2FsIENvbXB1dGluZzwva2V5

d29yZD48a2V5d29yZD5PZGRzIFJhdGlvPC9rZXl3b3JkPjxrZXl3b3JkPlBoZW5vdHlwZTwva2V5

d29yZD48a2V5d29yZD5Qcm9zcGVjdGl2ZSBTdHVkaWVzPC9rZXl3b3JkPjxrZXl3b3JkPlJlcHJv

ZHVjaWJpbGl0eSBvZiBSZXN1bHRzPC9rZXl3b3JkPjxrZXl3b3JkPlNldmVyaXR5IG9mIElsbG5l

c3MgSW5kZXg8L2tleXdvcmQ+PGtleXdvcmQ+U2hvY2ssIFNlcHRpYzwva2V5d29yZD48a2V5d29y

ZD5TaWduYWwgVHJhbnNkdWN0aW9uPC9rZXl3b3JkPjwva2V5d29yZHM+PHVybHM+PHJlbGF0ZWQt

dXJscz48dXJsPmh0dHA6Ly93d3cubmNiaS5ubG0ubmloLmdvdi9wdWJtZWQvMjU0ODk4ODE8L3Vy

bD48L3JlbGF0ZWQtdXJscz48L3VybHM+PGlzYm4+MTUzNS00OTcwPC9pc2JuPjxjdXN0b20yPlBN

QzQzNTE1ODA8L2N1c3RvbTI+PHRpdGxlcz48dGl0bGU+RGV2ZWxvcGluZyBhIGNsaW5pY2FsbHkg

ZmVhc2libGUgcGVyc29uYWxpemVkIG1lZGljaW5lIGFwcHJvYWNoIHRvIHBlZGlhdHJpYyBzZXB0

aWMgc2hvY2s8L3RpdGxlPjxzZWNvbmRhcnktdGl0bGU+QW0gSiBSZXNwaXIgQ3JpdCBDYXJlIE1l

ZDwvc2Vjb25kYXJ5LXRpdGxlPjwvdGl0bGVzPjxwYWdlcz4zMDktMTU8L3BhZ2VzPjxudW1iZXI+

MzwvbnVtYmVyPjxjb250cmlidXRvcnM+PGF1dGhvcnM+PGF1dGhvcj5Xb25nLCBILiBSLjwvYXV0

aG9yPjxhdXRob3I+Q3ZpamFub3ZpY2gsIE4uIFouPC9hdXRob3I+PGF1dGhvcj5BbmFzLCBOLjwv

YXV0aG9yPjxhdXRob3I+QWxsZW4sIEcuIEwuPC9hdXRob3I+PGF1dGhvcj5UaG9tYXMsIE4uIEou

PC9hdXRob3I+PGF1dGhvcj5CaWdoYW0sIE0uIFQuPC9hdXRob3I+PGF1dGhvcj5XZWlzcywgUy4g

TC48L2F1dGhvcj48YXV0aG9yPkZpdHpnZXJhbGQsIEouPC9hdXRob3I+PGF1dGhvcj5DaGVjY2hp

YSwgUC4gQS48L2F1dGhvcj48YXV0aG9yPk1leWVyLCBLLjwvYXV0aG9yPjxhdXRob3I+U2hhbmxl

eSwgVC4gUC48L2F1dGhvcj48YXV0aG9yPlF1YXNuZXksIE0uPC9hdXRob3I+PGF1dGhvcj5IYWxs

LCBNLjwvYXV0aG9yPjxhdXRob3I+R2VkZWl0LCBSLjwvYXV0aG9yPjxhdXRob3I+RnJlaXNodGF0

LCBSLiBKLjwvYXV0aG9yPjxhdXRob3I+Tm93YWssIEouPC9hdXRob3I+PGF1dGhvcj5TaGVraGFy

LCBSLiBTLjwvYXV0aG9yPjxhdXRob3I+R2VydHosIFMuPC9hdXRob3I+PGF1dGhvcj5EYXdzb24s

IEUuPC9hdXRob3I+PGF1dGhvcj5Ib3dhcmQsIEsuPC9hdXRob3I+PGF1dGhvcj5IYXJtb24sIEsu

PC9hdXRob3I+PGF1dGhvcj5CZWNrbWFuLCBFLjwvYXV0aG9yPjxhdXRob3I+RnJhbmssIEUuPC9h

dXRob3I+PGF1dGhvcj5MaW5kc2VsbCwgQy4gSi48L2F1dGhvcj48L2F1dGhvcnM+PC9jb250cmli

dXRvcnM+PGxhbmd1YWdlPmVuZzwvbGFuZ3VhZ2U+PGFkZGVkLWRhdGUgZm9ybWF0PSJ1dGMiPjE0

Mzc1ODk3OTk8L2FkZGVkLWRhdGU+PHJlZi10eXBlIG5hbWU9IkpvdXJuYWwgQXJ0aWNsZSI+MTc8

L3JlZi10eXBlPjxyZWMtbnVtYmVyPjc5MzwvcmVjLW51bWJlcj48bGFzdC11cGRhdGVkLWRhdGUg

Zm9ybWF0PSJ1dGMiPjE0Mzc1ODk3OTk8L2xhc3QtdXBkYXRlZC1kYXRlPjxhY2Nlc3Npb24tbnVt

PjI1NDg5ODgxPC9hY2Nlc3Npb24tbnVtPjxlbGVjdHJvbmljLXJlc291cmNlLW51bT4xMC4xMTY0

L3JjY20uMjAxNDEwLTE4NjRPQzwvZWxlY3Ryb25pYy1yZXNvdXJjZS1udW0+PHZvbHVtZT4xOTE8

L3ZvbHVtZT48L3JlY29yZD48L0NpdGU+PC9FbmROb3RlPgB=

ADDIN EN.CITE.DATA (9-11). Davenport et al. (validation dataset EMTAB-4421.51) derived two clusters of adult sepsis: sepsis response signature (SRS) 1 (higher mortality, with endotoxin signaling, T-cell repression, and NF-kB activation) and SRS 2 (lower mortality, with T-cell activation and interferon signaling)PEVuZE5vdGU+PENpdGU+PEF1dGhvcj5CdXJuaGFtPC9BdXRob3I+PFllYXI+MjAxNjwvWWVhcj48

SURUZXh0PlNoYXJlZCBhbmQgRGlzdGluY3QgQXNwZWN0cyBvZiB0aGUgU2Vwc2lzIFRyYW5zY3Jp

cHRvbWljIFJlc3BvbnNlIHRvIEZlY2FsIFBlcml0b25pdGlzIGFuZCBQbmV1bW9uaWE8L0lEVGV4

dD48RGlzcGxheVRleHQ+KDEyLCAxMyk8L0Rpc3BsYXlUZXh0PjxyZWNvcmQ+PGRhdGVzPjxwdWIt

ZGF0ZXM+PGRhdGU+RGVjPC9kYXRlPjwvcHViLWRhdGVzPjx5ZWFyPjIwMTY8L3llYXI+PC9kYXRl

cz48dXJscz48cmVsYXRlZC11cmxzPjx1cmw+aHR0cHM6Ly93d3cubmNiaS5ubG0ubmloLmdvdi9w

dWJtZWQvMjgwMzYyMzM8L3VybD48L3JlbGF0ZWQtdXJscz48L3VybHM+PGlzYm4+MTUzNS00OTcw

PC9pc2JuPjx0aXRsZXM+PHRpdGxlPlNoYXJlZCBhbmQgRGlzdGluY3QgQXNwZWN0cyBvZiB0aGUg

U2Vwc2lzIFRyYW5zY3JpcHRvbWljIFJlc3BvbnNlIHRvIEZlY2FsIFBlcml0b25pdGlzIGFuZCBQ

bmV1bW9uaWE8L3RpdGxlPjxzZWNvbmRhcnktdGl0bGU+QW0gSiBSZXNwaXIgQ3JpdCBDYXJlIE1l

ZDwvc2Vjb25kYXJ5LXRpdGxlPjwvdGl0bGVzPjxjb250cmlidXRvcnM+PGF1dGhvcnM+PGF1dGhv

cj5CdXJuaGFtLCBLLiBMLjwvYXV0aG9yPjxhdXRob3I+RGF2ZW5wb3J0LCBFLiBFLjwvYXV0aG9y

PjxhdXRob3I+UmFkaGFrcmlzaG5hbiwgSi48L2F1dGhvcj48YXV0aG9yPkh1bWJ1cmcsIFAuPC9h

dXRob3I+PGF1dGhvcj5Hb3Jkb24sIEEuIEMuPC9hdXRob3I+PGF1dGhvcj5IdXR0b24sIFAuPC9h

dXRob3I+PGF1dGhvcj5Tdm9yZW4tSmFiYWxlcmEsIEUuPC9hdXRob3I+PGF1dGhvcj5HYXJyYXJk

LCBDLjwvYXV0aG9yPjxhdXRob3I+SGlsbCwgQS4gVi48L2F1dGhvcj48YXV0aG9yPkhpbmRzLCBD

LiBKLjwvYXV0aG9yPjxhdXRob3I+S25pZ2h0LCBKLiBDLjwvYXV0aG9yPjwvYXV0aG9ycz48L2Nv

bnRyaWJ1dG9ycz48ZWRpdGlvbj4yMDE2LzEyLzMwPC9lZGl0aW9uPjxsYW5ndWFnZT5lbmc8L2xh

bmd1YWdlPjxhZGRlZC1kYXRlIGZvcm1hdD0idXRjIj4xNDg1MzA1NTkyPC9hZGRlZC1kYXRlPjxy

ZWYtdHlwZSBuYW1lPSJKb3VybmFsIEFydGljbGUiPjE3PC9yZWYtdHlwZT48cmVjLW51bWJlcj4x

MTg3PC9yZWMtbnVtYmVyPjxsYXN0LXVwZGF0ZWQtZGF0ZSBmb3JtYXQ9InV0YyI+MTQ4NTMwNTU5

MjwvbGFzdC11cGRhdGVkLWRhdGU+PGFjY2Vzc2lvbi1udW0+MjgwMzYyMzM8L2FjY2Vzc2lvbi1u

dW0+PGVsZWN0cm9uaWMtcmVzb3VyY2UtbnVtPjEwLjExNjQvcmNjbS4yMDE2MDgtMTY4NU9DPC9l

bGVjdHJvbmljLXJlc291cmNlLW51bT48L3JlY29yZD48L0NpdGU+PENpdGU+PEF1dGhvcj5EYXZl

bnBvcnQ8L0F1dGhvcj48WWVhcj4yMDE2PC9ZZWFyPjxJRFRleHQ+R2Vub21pYyBsYW5kc2NhcGUg

b2YgdGhlIGluZGl2aWR1YWwgaG9zdCByZXNwb25zZSBhbmQgb3V0Y29tZXMgaW4gc2Vwc2lzOiBh

IHByb3NwZWN0aXZlIGNvaG9ydCBzdHVkeTwvSURUZXh0PjxyZWNvcmQ+PGRhdGVzPjxwdWItZGF0

ZXM+PGRhdGU+RmViPC9kYXRlPjwvcHViLWRhdGVzPjx5ZWFyPjIwMTY8L3llYXI+PC9kYXRlcz48

dXJscz48cmVsYXRlZC11cmxzPjx1cmw+aHR0cDovL3d3dy5uY2JpLm5sbS5uaWguZ292L3B1Ym1l

ZC8yNjkxNzQzNDwvdXJsPjwvcmVsYXRlZC11cmxzPjwvdXJscz48aXNibj4yMjEzLTI2MTk8L2lz

Ym4+PGN1c3RvbTI+UE1DNDgyMDY2NzwvY3VzdG9tMj48dGl0bGVzPjx0aXRsZT5HZW5vbWljIGxh

bmRzY2FwZSBvZiB0aGUgaW5kaXZpZHVhbCBob3N0IHJlc3BvbnNlIGFuZCBvdXRjb21lcyBpbiBz

ZXBzaXM6IGEgcHJvc3BlY3RpdmUgY29ob3J0IHN0dWR5PC90aXRsZT48c2Vjb25kYXJ5LXRpdGxl

PkxhbmNldCBSZXNwaXIgTWVkPC9zZWNvbmRhcnktdGl0bGU+PC90aXRsZXM+PGNvbnRyaWJ1dG9y

cz48YXV0aG9ycz48YXV0aG9yPkRhdmVucG9ydCwgRS4gRS48L2F1dGhvcj48YXV0aG9yPkJ1cm5o

YW0sIEsuIEwuPC9hdXRob3I+PGF1dGhvcj5SYWRoYWtyaXNobmFuLCBKLjwvYXV0aG9yPjxhdXRo

b3I+SHVtYnVyZywgUC48L2F1dGhvcj48YXV0aG9yPkh1dHRvbiwgUC48L2F1dGhvcj48YXV0aG9y

Pk1pbGxzLCBULiBDLjwvYXV0aG9yPjxhdXRob3I+UmF1dGFuZW4sIEEuPC9hdXRob3I+PGF1dGhv

cj5Hb3Jkb24sIEEuIEMuPC9hdXRob3I+PGF1dGhvcj5HYXJyYXJkLCBDLjwvYXV0aG9yPjxhdXRo

b3I+SGlsbCwgQS4gVi48L2F1dGhvcj48YXV0aG9yPkhpbmRzLCBDLiBKLjwvYXV0aG9yPjxhdXRo

b3I+S25pZ2h0LCBKLiBDLjwvYXV0aG9yPjwvYXV0aG9ycz48L2NvbnRyaWJ1dG9ycz48bGFuZ3Vh

Z2U+RU5HPC9sYW5ndWFnZT48YWRkZWQtZGF0ZSBmb3JtYXQ9InV0YyI+MTQ2MjQwMDkyOTwvYWRk

ZWQtZGF0ZT48cmVmLXR5cGUgbmFtZT0iSm91cm5hbCBBcnRpY2xlIj4xNzwvcmVmLXR5cGU+PHJl

Yy1udW1iZXI+MTAyMzwvcmVjLW51bWJlcj48bGFzdC11cGRhdGVkLWRhdGUgZm9ybWF0PSJ1dGMi

PjE0NjI0MDA5Mjk8L2xhc3QtdXBkYXRlZC1kYXRlPjxhY2Nlc3Npb24tbnVtPjI2OTE3NDM0PC9h

Y2Nlc3Npb24tbnVtPjxlbGVjdHJvbmljLXJlc291cmNlLW51bT4xMC4xMDE2L1MyMjEzLTI2MDAo

MTYpMDAwNDYtMTwvZWxlY3Ryb25pYy1yZXNvdXJjZS1udW0+PC9yZWNvcmQ+PC9DaXRlPjwvRW5k

Tm90ZT5=

ADDIN EN.CITE PEVuZE5vdGU+PENpdGU+PEF1dGhvcj5CdXJuaGFtPC9BdXRob3I+PFllYXI+MjAxNjwvWWVhcj48

SURUZXh0PlNoYXJlZCBhbmQgRGlzdGluY3QgQXNwZWN0cyBvZiB0aGUgU2Vwc2lzIFRyYW5zY3Jp

cHRvbWljIFJlc3BvbnNlIHRvIEZlY2FsIFBlcml0b25pdGlzIGFuZCBQbmV1bW9uaWE8L0lEVGV4

dD48RGlzcGxheVRleHQ+KDEyLCAxMyk8L0Rpc3BsYXlUZXh0PjxyZWNvcmQ+PGRhdGVzPjxwdWIt

ZGF0ZXM+PGRhdGU+RGVjPC9kYXRlPjwvcHViLWRhdGVzPjx5ZWFyPjIwMTY8L3llYXI+PC9kYXRl

cz48dXJscz48cmVsYXRlZC11cmxzPjx1cmw+aHR0cHM6Ly93d3cubmNiaS5ubG0ubmloLmdvdi9w

dWJtZWQvMjgwMzYyMzM8L3VybD48L3JlbGF0ZWQtdXJscz48L3VybHM+PGlzYm4+MTUzNS00OTcw

PC9pc2JuPjx0aXRsZXM+PHRpdGxlPlNoYXJlZCBhbmQgRGlzdGluY3QgQXNwZWN0cyBvZiB0aGUg

U2Vwc2lzIFRyYW5zY3JpcHRvbWljIFJlc3BvbnNlIHRvIEZlY2FsIFBlcml0b25pdGlzIGFuZCBQ

bmV1bW9uaWE8L3RpdGxlPjxzZWNvbmRhcnktdGl0bGU+QW0gSiBSZXNwaXIgQ3JpdCBDYXJlIE1l

ZDwvc2Vjb25kYXJ5LXRpdGxlPjwvdGl0bGVzPjxjb250cmlidXRvcnM+PGF1dGhvcnM+PGF1dGhv

cj5CdXJuaGFtLCBLLiBMLjwvYXV0aG9yPjxhdXRob3I+RGF2ZW5wb3J0LCBFLiBFLjwvYXV0aG9y

PjxhdXRob3I+UmFkaGFrcmlzaG5hbiwgSi48L2F1dGhvcj48YXV0aG9yPkh1bWJ1cmcsIFAuPC9h

dXRob3I+PGF1dGhvcj5Hb3Jkb24sIEEuIEMuPC9hdXRob3I+PGF1dGhvcj5IdXR0b24sIFAuPC9h

dXRob3I+PGF1dGhvcj5Tdm9yZW4tSmFiYWxlcmEsIEUuPC9hdXRob3I+PGF1dGhvcj5HYXJyYXJk

LCBDLjwvYXV0aG9yPjxhdXRob3I+SGlsbCwgQS4gVi48L2F1dGhvcj48YXV0aG9yPkhpbmRzLCBD

LiBKLjwvYXV0aG9yPjxhdXRob3I+S25pZ2h0LCBKLiBDLjwvYXV0aG9yPjwvYXV0aG9ycz48L2Nv

bnRyaWJ1dG9ycz48ZWRpdGlvbj4yMDE2LzEyLzMwPC9lZGl0aW9uPjxsYW5ndWFnZT5lbmc8L2xh

bmd1YWdlPjxhZGRlZC1kYXRlIGZvcm1hdD0idXRjIj4xNDg1MzA1NTkyPC9hZGRlZC1kYXRlPjxy

ZWYtdHlwZSBuYW1lPSJKb3VybmFsIEFydGljbGUiPjE3PC9yZWYtdHlwZT48cmVjLW51bWJlcj4x

MTg3PC9yZWMtbnVtYmVyPjxsYXN0LXVwZGF0ZWQtZGF0ZSBmb3JtYXQ9InV0YyI+MTQ4NTMwNTU5

MjwvbGFzdC11cGRhdGVkLWRhdGU+PGFjY2Vzc2lvbi1udW0+MjgwMzYyMzM8L2FjY2Vzc2lvbi1u

dW0+PGVsZWN0cm9uaWMtcmVzb3VyY2UtbnVtPjEwLjExNjQvcmNjbS4yMDE2MDgtMTY4NU9DPC9l

bGVjdHJvbmljLXJlc291cmNlLW51bT48L3JlY29yZD48L0NpdGU+PENpdGU+PEF1dGhvcj5EYXZl

bnBvcnQ8L0F1dGhvcj48WWVhcj4yMDE2PC9ZZWFyPjxJRFRleHQ+R2Vub21pYyBsYW5kc2NhcGUg

b2YgdGhlIGluZGl2aWR1YWwgaG9zdCByZXNwb25zZSBhbmQgb3V0Y29tZXMgaW4gc2Vwc2lzOiBh

IHByb3NwZWN0aXZlIGNvaG9ydCBzdHVkeTwvSURUZXh0PjxyZWNvcmQ+PGRhdGVzPjxwdWItZGF0

ZXM+PGRhdGU+RmViPC9kYXRlPjwvcHViLWRhdGVzPjx5ZWFyPjIwMTY8L3llYXI+PC9kYXRlcz48

dXJscz48cmVsYXRlZC11cmxzPjx1cmw+aHR0cDovL3d3dy5uY2JpLm5sbS5uaWguZ292L3B1Ym1l

ZC8yNjkxNzQzNDwvdXJsPjwvcmVsYXRlZC11cmxzPjwvdXJscz48aXNibj4yMjEzLTI2MTk8L2lz

Ym4+PGN1c3RvbTI+UE1DNDgyMDY2NzwvY3VzdG9tMj48dGl0bGVzPjx0aXRsZT5HZW5vbWljIGxh

bmRzY2FwZSBvZiB0aGUgaW5kaXZpZHVhbCBob3N0IHJlc3BvbnNlIGFuZCBvdXRjb21lcyBpbiBz

ZXBzaXM6IGEgcHJvc3BlY3RpdmUgY29ob3J0IHN0dWR5PC90aXRsZT48c2Vjb25kYXJ5LXRpdGxl

PkxhbmNldCBSZXNwaXIgTWVkPC9zZWNvbmRhcnktdGl0bGU+PC90aXRsZXM+PGNvbnRyaWJ1dG9y

cz48YXV0aG9ycz48YXV0aG9yPkRhdmVucG9ydCwgRS4gRS48L2F1dGhvcj48YXV0aG9yPkJ1cm5o

YW0sIEsuIEwuPC9hdXRob3I+PGF1dGhvcj5SYWRoYWtyaXNobmFuLCBKLjwvYXV0aG9yPjxhdXRo

b3I+SHVtYnVyZywgUC48L2F1dGhvcj48YXV0aG9yPkh1dHRvbiwgUC48L2F1dGhvcj48YXV0aG9y

Pk1pbGxzLCBULiBDLjwvYXV0aG9yPjxhdXRob3I+UmF1dGFuZW4sIEEuPC9hdXRob3I+PGF1dGhv

cj5Hb3Jkb24sIEEuIEMuPC9hdXRob3I+PGF1dGhvcj5HYXJyYXJkLCBDLjwvYXV0aG9yPjxhdXRo

b3I+SGlsbCwgQS4gVi48L2F1dGhvcj48YXV0aG9yPkhpbmRzLCBDLiBKLjwvYXV0aG9yPjxhdXRo

b3I+S25pZ2h0LCBKLiBDLjwvYXV0aG9yPjwvYXV0aG9ycz48L2NvbnRyaWJ1dG9ycz48bGFuZ3Vh

Z2U+RU5HPC9sYW5ndWFnZT48YWRkZWQtZGF0ZSBmb3JtYXQ9InV0YyI+MTQ2MjQwMDkyOTwvYWRk

ZWQtZGF0ZT48cmVmLXR5cGUgbmFtZT0iSm91cm5hbCBBcnRpY2xlIj4xNzwvcmVmLXR5cGU+PHJl

Yy1udW1iZXI+MTAyMzwvcmVjLW51bWJlcj48bGFzdC11cGRhdGVkLWRhdGUgZm9ybWF0PSJ1dGMi

PjE0NjI0MDA5Mjk8L2xhc3QtdXBkYXRlZC1kYXRlPjxhY2Nlc3Npb24tbnVtPjI2OTE3NDM0PC9h

Y2Nlc3Npb24tbnVtPjxlbGVjdHJvbmljLXJlc291cmNlLW51bT4xMC4xMDE2L1MyMjEzLTI2MDAo

MTYpMDAwNDYtMTwvZWxlY3Ryb25pYy1yZXNvdXJjZS1udW0+PC9yZWNvcmQ+PC9DaXRlPjwvRW5k

Tm90ZT5=

ADDIN EN.CITE.DATA (12, 13). For each subject in these two cohorts, we compared our cluster assignments to the previously published assignments (Supplemental Table 10). Most samples (60 of 69) in the Inflammopathic cluster were Wong et al. Endotype B. However, the converse was not true: an additional 22 Endotype B samples were in the Adaptive cluster and 13 were in the Coagulopathic cluster. Consistent with these findings, Goodman and Kruskal’s lambda showed unidirectional significance with our clusters’ ability to explain the Wong clusters, but not visa-versa. The correlations were stronger in the Davenport et al. clusters where the Inflammopathic and Adaptive clusters largely represented SRS 1 and 2, respectively. The lambda for our clusters and the Davenport clusters was bidirectionally significant. These results suggest that Endotype B and SRS 1 may also represent the Inflammopathic cluster, while SRS 2 may represent the Adaptive cluster.Supplemental References ADDIN EN.REFLIST 1. Tibshirani R, Walther G, Hastie T. Estimating the number of clusters in a data set via the gap statistic. Journal of the Royal Statistical Society Series B-Statistical Methodology 2001; 63: 411-423.2. Sweeney TE, Chen AC, Gevaert O. Combined Mapping of Multiple clUsteriNg ALgorithms (COMMUNAL): A Robust Method for Selection of Cluster Number, K. Sci Rep 2015; 5: 16971.3. Planey CR, Gevaert O. CoINcIDE: A framework for discovery of patient subtypes across multiple datasets. Genome Med 2016; 8: 27.4. Tusher VG, Tibshirani R, Chu G. Significance analysis of microarrays applied to the ionizing radiation response. Proc Natl Acad Sci U S A 2001; 98: 5116-5121.5. Chen J, Bardes EE, Aronow BJ, Jegga AG. ToppGene Suite for gene list enrichment analysis and candidate gene prioritization. Nucleic Acids Res 2009; 37: W305-311.6. Sweeney TE, Shidham A, Wong HR, Khatri P. A comprehensive time-course-based multicohort analysis of sepsis and sterile inflammation reveals a robust diagnostic gene set. Sci Transl Med 2015; 7: 287ra271.7. VanderWeele TJ, Ding P. Sensitivity Analysis in Observational Research: Introducing the E-Value. Ann Intern Med 2017.Supplemental Table 1: Sample-level clustering assignments in the discovery and validation datasets. Attached as a separate Excel spreadsheet (Supplemental Digital Content 2, ). Supplemental Table 2: Clustering in the discovery data broken out by individual dataset (A) and microarray type (B). A: DatasetsInflammopathicAdaptiveCoagulopathicUnclusteredEMEXP35679201EMTAB15482753317GSE117555010GSE13015 gpl6106824511GSE13015 gpl69472814GSE203462310GSE287503241GSE3334124405GSE4001221324GSE4058601911GSE570651272116GSE6568228194212GSE6609976583123GSE695281443917Total190247151112B: Microarray TypeInflammopathicAdaptiveCoagulopathicUnclusteredGPL969201GPL57096675740GPL57124405GPL6106824511GPL624401911GPL694762448GPL103322753317GPL105581443917GPL1366728194212Total190247151112Supplemental Table 3: Significant GO terms from the 500 genes used in deriving the discovery cluster assignments. Attached as a separate Excel spreadsheet (Supplemental Digital Content 3, ). Supplemental Table 4: Demographic and clinical variables according to discovery clusters, with unclustered samples added as a group. Not all variables were available for all samples, so the totals are not always consistent. Statistically significant differences between groups (age, shock, and mortality) are the same as seen when unclustered samples are excluded (Table 2). However, the addition of unclustered samples now results in a statistically significant difference in frequency of Gram negative infections, which is higher in these subjects. Inflammo-pathicAdaptiveCoagulo-pathicUnclusteredP value (chisq / ANOVA)Total Samples Assigned175219108112Male (percent)58.459.461.560.90.952(N, gender)16618013592Age (years, +/- sd)34.8 (32.1)38.5 (28.7)49.7 (29.4)45.1 (28.9)0.000152Age < 18 years (percent)16.817.615.911.60.660Age > 70 years (percent)27.720.036.424.40.01604(N, age)15516513286WBC count (+/- sd)18.02 (16.18)13.83 (10.64)12.87 (13.3)16.56 (10.36)0.29Neutrophils (+/- sd)59.67 (18.31)61.14 (16.42)58.15 (23.1)48.27 (19.4)0.15Bands (+/- sd)17.04 (12.77)11.58 (11.57)6.75 (6.13)15.87 (12.22)0.0048Lymphocytes (+/- sd)15.89 (13.8)20.17 (12.71)27.05 (23.16)29.87 (20.52)0.0068Monocytes (+/- sd)6.07 (4.33)6.19 (3.82)6.6 (6.66)4.33 (3.83)0.5(N, differential)55362115Immunosuppressed (percent)5.88.911.510.50.78(N, immune status)69452619Gram negative (percent)46.248.451.469.80.033(N, Gram status)911573753Shock (percent)73.032.262.241.71.30E-09(N, shock status)1001524548High Clinical Severity (percent)50.832.456.350.00.005(N, clinical severity)124102876830-day mortality (percent)29.88.125.424.22.97E-05(N, survivor status)1241607162Supplemental Table 5. Age, shock and severity as cluster predictors in discovery data. A. Multiple regression was run on each cluster type (vs. all others, plus unclustered) to determine whether age and shock (and their interaction) could predict cluster assignment. Shown are P-values for each covariate, along with final model R2 values and unexplained residual variance. Only samples with sample-level age, shock, and severity data were used here. B. Risk ratios (RRs) for mortality by cluster, with resulting E-values. C. Age, shock and severity are shown to have lower RRs for mortality and for cluster assignment than the E-values in (B), showing that a confounding variable explaining the clusters would have to have a substantially greater effect size than age, shock, or severity. We can thus conclude that the observed risk ratios of mortality due to cluster assignment could be explained away by an unmeasured confounder that was associated with both cluster assignment and mortality by an effect size greater than 2.04-6.16 (depending on cluster), but that neither shock, high clinical severity, or age > 70 years has an effect size of this magnitude.AInflammopathicAdaptiveCoagulopathicUnclusteredN with age & shock & severity data86903738intercept0.82740.02260.2610.216age0.34370.43670.60.739shock0.23030.25290.7310.713severity0.76220.97560.8010.926age*shock0.05540.77940.1780.469age*severity0.32020.23090.8750.648shock*severity0.61190.55120.9910.885age*shock*severity0.80520.48590.9280.496Adjusted R-squared0.0970.0640.007-0.0001Percent of variance unexplained90.393.699.3100BRR mortalityE-value mortalityInflammopathic1.903.21Adaptive0.306.16Coagulopathic1.352.04CRR mortalityRR InflammopathicRR AdaptiveRR CoagulopathicShock0.972.780.491.70High severity1.491.170.541.46Age > 70 years1.261.040.681.55Supplemental Table 6: The classifier for cluster assignments. (A) Genes that make up the three subparts of the classifier score. (B) Coefficients for regression classifier; each ‘score’ refers to the geometric mean difference of ‘Up’ minus ‘Down’ genes for each class defined in (A). The scores are multiplied through the given regression coefficients to form final predicted probabilities of cluster assignment. (C) Two-way matrix for re-assignment of discovery clusters comparing “True” assignment from the original COMMUNAL clustering to predicted assignment by the cluster classifier. AInflammopathicAdaptiveCoagulopathicUpDownUpDownUpDownARG1HLA-DMBYKT6GADD45AKCNMB4RHBDF2LCN2PDE4BCD24CRISP2ZCCHC4LTFTWISTNBS100A12HTRA1YKT6OLFM4BTN2A2STX1APPLDDX6ZBTB33SENP5PSMB9RAPGEF1CAMK4DTX2TMEM19RELBSLC12A7TP53BP1PLEKHO1SLC25A22FRS2BInterceptInflammopathic ScoreAdaptive ScoreCoagulopathic ScoreAdaptive-0.617-3.1507.1871.594Coagulopathic0.494-1.3820.4801.569 TruePredictedCInflammopathicAdaptiveCoagulopathicInflammopathic156441Adaptive423613Coagulopathic30797Supplemental Table 7: Breakdown of cluster assignments in the validation cohorts. Each subject was assigned to the most likely sepsis cluster based on the 33-gene cluster classifier. InflammopathicAdaptiveCoagulopathicEMEXP3850987EMTAB4421.51637837GSE1047414146GSE28658231GSE3270718219GSE63042344723GSE63990253312GSE66890183014GSE74224253019Supplemental Table 8: Demographic and outcomes data shown per cluster per validation dataset. Not all variables were available for all datasets, so each variable is shown only for the datasets that included the given variable. Note that no ‘unclustered’ samples are present because the classifier does not output an ‘unclustered’ class. Male, PercentInflammopathicAdaptiveCoagulopathicDataset NEMEXP385011.1%62.5%42.9%24EMTAB4421.5157.1%75.6%59.5%178GSE1047435.7%61.5%66.7%34GSE3270733.3%57.1%44.4%48GSE6304267.6%57.4%52.2%104GSE6689055.6%53.3%64.3%62GSE7422450.0%46.4%78.9%74mean (sd)44.4 (17.6)59.1 (8.4)58.4 (12.0)Age, mean (years)InflammopathicAdaptiveCoagulopathicDataset NEMEXP38501.5740.84381.36924EMTAB4421.5159.5966.2366.7178GSE1047465.3648.9260.533GSE3270759.0654.3359.7848GSE6304262.9453.8763.13104GSE668906159.0373.2162GSE7422459.55960.2174mean (sd)52.7 (14.7)48.8 (13.6)55.0 (15.2)Gram Negative, PercentInflammopathicAdaptiveCoagulopathicDataset NEMEXP3850100.0%100.0%100.0%24GSE28658100.0%100.0%100.0%6GSE7422443.8%58.3%30.0%74mean (sd)81.2 (25.5)86.1 (19.6)76.6 (33.0)Septic Shock, PercentInflammopathicAdaptiveCoagulopathicDataset NGSE6689083.3%43.3%42.9%62GSE7422460.0%30.0%47.4%74mean (sd)71.7 (11.7)36.7 (6.7)45.1 (0.0)High Clinical Severity, PercentInflammopathicAdaptiveCoagulopathicDataset NEMEXP385077.8%25.0%28.6%24EMTAB4421.5144.4%26.9%29.7%178GSE1047442.9%35.7%50.0%33GSE3270744.4%38.1%66.7%48GSE6304247.1%17.0%52.2%104GSE6689033.3%63.3%28.6%62mean (sd)48.3 (13.8)34.3 (14.7)42.6 (14.6)MortalityInflammopathicAdaptiveCoagulopathicDataset NEMEXP385022.2%25.0%14.3%24EMTAB4421.5133.3%21.8%24.3%178GSE1047428.6%38.5%33.3%33GSE3270738.9%28.6%44.4%48GSE6304229.4%17.0%43.5%104GSE639908.0%6.1%16.7%70GSE6689038.9%7.4%41.7%62mean (sd)28.5 (10.0)20.6 (10.7)31.2 (11.9)Supplemental Table 9: Association of clinical coagulopathy with the Coagulopathic cluster. (A) GSE66099 (Discovery dataset, pediatric), disseminated intravascular coagulation (DIC) by cluster type. (B) GSE63042 (Validation dataset, adults) intersection of thrombocytopenia (platelets <100,000) and prolonged INR (>1.3) by cluster type. Association p-value tested with Fisher Exact test. AGSE66099, Discovery: DICNoYesPercent YesP ValueInflammopathic641215.80.0078Adaptive5358.6Coagulopathic201135.5Unclustered18521.7BGSE63042 (Validation):Platelet count < 100,000 & INR > 1.3NoYesPercent YesP ValueInflammopathic2926.50.034Adaptive3725.1Coagulopathic16627.3Supplemental Table 10: Neither thrombocytopenia nor prolonged INR were significantly associated with cluster type, though INR>1.3 showed a trend towards significance in the Coagulopathic group. Association p-value tested with Fisher Exact test. Aplatelet count < 100KGSE63042 (Validation)NoYesPercent YesP ValueInflammopathic276180.3352Adaptive35513Coagulopathic16627BINR > 1.3GSE63042 (Validation)NoYesPercent YesP ValueInflammopathic1013570.0873Adaptive101152Coagulopathic11091Supplemental Table 11: Comparison of cluster assignments to previously published clusters from (A) Wong et al. and (B) Davenport et al. Goodman and Kruskal’s lambda for dependence was performed in both directions. Non-overlap of the 95% CI with 0 is considered significant. AWongEndotypesABClustersInflammopathic960Adaptive2322Coagulopathic1313Unclustered712Excluding ‘unclustered’: Lambda, (Wong|Core): 0.19, 95% CI 0.06-0.33Including ‘unclustered’: Lambda, (Wong|Core): 0.15, 95% CI 0.04-0.27Excluding ‘unclustered’: Lambda, (Core|Wong): 0.02, 95% CI 0-0.38Including ‘unclustered’: Lambda, (Core|Wong): 0.02, 95% CI 0-0.33BDavenportSRS Groups12ClustersInflammopathic612Adaptive1265Coagulopathic2017Lambda, (Davenport|Core):0.49, 95%CI 0.37-0.61Lambda, (Core|Davenport):0.63, 95%CI 0.51 – 0.76Supplemental Figure 1: Principal components analysis of the discovery datasets pre- and post-COCONUT. Prior to COCONUT co-normalization, the discovery datasets are entirely separated by technical batch effects. These technical effects are removed post-COCONUT, as evidenced by a general overlapping of the discovery datasets in the first two principle components.Supplemental Figure 2: Output from the two consensus clustering algorithms using K-means (A, B) and partitioning around medioids (C,D). (A,C) Cumulative density functions of consensus assignments by number of clusters. (B,D) Consenus mappings by cluster. 1=Inflammopathic, 2=Adaptive, 3=Coagulopathic.-228600112395ABCDABCDSupplemental Figure 3: COMMUNAL map of cluster optimality. X-axis shows number of clusters, Y-axis shows number of included genes, Z-axis shows mean validity score (higher is better). Red and blue dots show automatically assigned optima at each number of included genes. COMMUNAL automatically chose the following 5 validity measures: gap statistic, connectivity, average silhouette width, g3 metric, Pearson’s gamma. The resulting map shows the mean of standardized values of each validity measure across the entire tested space. Stable optima at K=3 clusters are seen over most of the tested space, indicating strong, consistent biological signal at this number of clusters. Red arrow shows chosen clustering (stable K [3] at lowest number of genes [500]).Supplemental Figure 4: Principal Components Analysis (PCA) of the discovery clustering results (including the 16% of samples that went unclustered in the final analysis, in gold) using either all 8,946 genes present in the COCONUT conormalized data, or only the 500 genes actually used in the clustering analysis. PCA is an unsupervised dimensionality reduction technique that allows for the visualization of high-dimensional data. Here we show that the cluster assignments that we recovered in an unsupervised manner are clearly separated in high-dimensional space, as demonstrated by the first three principal components. Adaptive samples appear to separate from Inflammopathic and Coagulopathic samples along PCs 1 and 2, while PC3 largely separates the Inflammopathic and Coagulopathic samples.-389614194503InflammopathicAdaptiveCoagulopathicUnclustered00InflammopathicAdaptiveCoagulopathicUnclustered-16510667385InflammopathicAdaptiveCoagulopathic00InflammopathicAdaptiveCoagulopathicSupplemental Figure 5: Heatmap of the 500 genes included in the clustering analysis for the discovery clusters, with hierarchical clustering of the genes solely for visualization. 22332953294380AdaptiveAdaptive5803901653844InflammopathicInflammopathic0100901500Supplemental Figure 6: Comparison of raw predicted probability of cluster assignment in the discovery data. Histograms of probability show clear decision by the model for Adaptive, but Inflammopathic and Coagulopathic have less predicted certainty. 38736104162425CoagulopathicCoagulopathic ................
................

In order to avoid copyright disputes, this page is only a partial summary.

Google Online Preview   Download