Identification of pediatric septic shock subclasses based on genome-wide expression profiling

The study protocol was approved by the individual Institutional Review Boards of each participating institution (N = 11 institutions). Children ≤10 years of age admitted to the pediatric intensive care unit and meeting published, pediatric-specific criteria for septic shock were eligible for the study [14]. Controls were recruited from the ambulatory departments of participating institutions using the following exclusion criteria: a recent febrile illness (within 2 weeks), recent use of anti-inflammatory medications (within 2 weeks), or any history of chronic or acute disease even remotely associated with inflammation. The median age (intra-quartile range (IQR)) for the control cohort (N = 32) was 1.6 years (0.2 to 3.7 years). There were 19 males and 13 females in the control cohort.

After obtaining informed consent from parents or legal guardians, blood samples were obtained within 24 hours of initial presentation to the pediatric intensive care unit with septic shock. Severity of illness was calculated using the pediatric risk of mortality (PRISM) III score [15]. Organ failure was defined using pediatric-specific criteria [14, 16]. Annotated clinical and laboratory data were collected daily while in the pediatric intensive care unit. Clinical, laboratory, and biological data were entered and stored using a web-based database developed locally.

The data and protocols described in this manuscript are compliant with the minimum information about a microarray experiment (MIAME) and are deposited in the National Center for Biotechnology Information Gene Expression Omnibus (GEO) under accession number GSE13904 (GEO, http://​www.​ncbi.​nlm.​nih.​gov/​geo/​). All of the controls and 67 of the patients with septic shock have been previously reported in analyses addressing completely different questions than that addressed in the current study [9, 11‐13]. An additional 31 patients in the septic shock cohort have not been previously reported. Total RNA was isolated from whole blood samples using the PaxGene™ blood RNA system (PreAnalytiX, Qiagen/Becton Dickson, Valencia, CA, USA) according the manufacturer's specifications. Microarray hybridization was performed by the Affymetrix Gene Chip Core facility at the Cincinnati Children's Hospital Research Foundation as previously described using the Human Genome U133 Plus 2.0 GeneChip (Affymetrix, Santa Clara, CA, USA) [9, 11‐13].

Analyses were performed using one patient sample per chip. Image files were captured using an Affymetrix GeneChip Scanner 3000. CEL files were subsequently preprocessed using robust multiple-array average (RMA) normalization and GeneSpring GX 7.3 software (Agilent Technologies, Palo Alto, CA, USA). All signal intensity-based data were used after RMA normalization, which specifically suppresses all but significant variation among lower intensity probe sets [17]. All chips representing patient samples were then normalized to the respective median values of the controls. Lists of differentially regulated genes were generated using a series of expression and statistical filters embedded in the GeneSpring GX 7.3 software. Further details regarding these filters will be provided in the Results section.

Gene lists of differentially expressed genes were primarily analyzed using the Ingenuity Pathways Analysis (IPA) application (Ingenuity Systems, Redwood City, CA, USA) that provides a tool for discovery of signaling pathways and gene networks within the uploaded gene lists as previously described [12, 18]. Adjunct analyses of gene lists were performed using the National Institutes of Health Database for Annotation, Visualization and Integrated Discovery (DAVID) [19]. Both applications are based on the established biomedical literature and use specific approaches to estimate significance (P values) based on non-redundant representations of the microarray chip and to convert the uploaded gene lists to gene lists containing a single value per gene. The P values provide an estimate of the probability that a given enrichment is present by chance alone and are derived using corrections for multiple comparisons.

The first step toward identification of septic shock subclasses involved derivation of an initial working list of genes differentially regulated between patients with septic shock (N = 98) and controls (N = 32). The initial working list was derived using an empiric, discovery-oriented expression filter designed to select genes that were increased or decreased ≥2-fold in at least 25%, but not more than 50% of patients with septic shock, relative to the median of controls. This expression filter yielded a working list of 6,099 genes that were subjected to unsupervised hierarchical clustering as shown in Figure 1. An a priori decision was made to identify the putative major septic shock subclasses based on the first- and second-order branching patterns of the condition tree (top of Figure 1). Using this strategy, Figure 1 suggested the existence of three major septic subclasses that we arbitrarily designated as subclasses A, B, and C.

These 6,934 genes were then subjected to unsupervised hierarchical clustering as shown in Figure 2. Based on the first- and second-order branching patterns of the condition tree (top of Figure 2), the differential patterns of gene expression demonstrate the existence of three major subclasses of patients with septic shock (subclasses A, B, and C). Of the patients designated subclass A in Figure 1, 75% belonged to subclass A in Figure 2; 82% of the patients designated subclass B in Figure 1 belonged to subclass B in Figure 2; 80% of the patients designated subclass C in Figure 1 belonged to subclass C in Figure 2.

Further evidence for the existence of subclasses A, B, and C was derived by conducting principal component analysis based on the above 6,934 genes and the individual patients in each of the subclasses indentified in Figure 2. Principal component analysis is a mathematical vector space transformation which allows for reduction of multidimensional data sets to lower dimensions (principle components) accounting for variability in the data set [20]. As shown in Figure 3, the principal component analysis based on three dimensions, accounting for 67.1% of the variance in gene expression, provided a high degree of separation between the three septic shock subclasses. Thus, identifiable subclasses of patients with septic shock exist based on genome-level expression patterns.

Table 1 provides the demographic and clinical characteristics of the three septic shock subclasses identified in Figure 2. Patients in subclass A had a significantly higher illness severity level (PRISM III score), a greater degree of organ failure, and a higher mortality rate, compared with patients in subclasses B and C. Patients in subclass A also had a significantly higher incidence of documented Gram-positive bacterial infection, compared with patients in subclass C, and were significantly younger, compared with patients in subclass B. A significantly greater proportion of patients in subclass B received hydrocortisone for cardiovascular shock compared with subclass C. None of the other clinical characteristics listed in Table 1 were significantly different between the three septic subclasses. Thus, the three septic shock classes identified through differential genome-wide expression patterns have significant differences in clinically relevant phenotypes.

To derive further biological information from the 6,934 genes depicted in Figure 2, we next sought to identify coordinately regulated gene clusters by conducting K-means clustering based on a maximum return of 10 clusters [21]. Figure 4 illustrates how the K-means clustering algorithm arranged the 6,934 genes into 10 clusters of coordinately regulated genes. All of the 6,934 genes are represented in 1 of the 10 clusters, with no genes unclassified.

The gene lists corresponding to each of the 10 clusters depicted in Figure 4 were individually uploaded to the IPA application and the analytical output was focused on enrichment for genes corresponding to signaling and metabolic pathways. Table 2 provides the top five (based on P values) signaling and metabolic pathways represented in each cluster. All clusters were enriched for signaling and metabolic pathways potentially relevant to the pathobiology of septic shock. Thus, the 10 K-means clusters of coordinately regulated genes that define septic shock subclasses A, B, and C, are biologically plausible in that they are broadly enriched for signaling and metabolic pathways relevant to the pathobiology of septic shock.

The above data demonstrate the existence of septic shock subclasses, in very broad terms, based on more than 6,000 differentially regulated genes and 10 K-means clusters consisting of between 300 and 1,100 genes. Herein we sought to refine the gene expression patterns that best distinguish the three septic shock subclasses. We assumed that the genes corresponding to the individual signaling and metabolic pathways listed in Table 2 likely have the most biological significance with regard to differentiating the three septic shock subclasses. Accordingly, we extracted the individual genes corresponding to these pathways (307 total genes). These 307 genes were then subjected to a leave one out cross-validation procedure using the support vector machines algorithm and the Fisher's exact test method of gene selection [22]. This cross-validation procedure yielded 89 correct subclass predictions (subclass A, B, or C) out of 98 (91%).

From the 307 genes used in the cross-validation procedure, we then extracted the top 100 genes based on subclass prediction strength. These top 100 most predictive genes were then uploaded to the IPA application and the analytical output was again focused on enrichment for genes corresponding to signaling and metabolic pathways. The top five signaling pathways derived from this analysis are shown in Table 3. All of these signaling pathways are relevant to the pathobiology of septic shock and are particularly relevant to the adaptive immune system.

Forty-four of the top 100 predictive genes corresponded to the top five signaling pathways shown in Table 3. These 44 genes (listed in Table 4) were subjected to hierarchical clustering based on the median expression values for each septic shock subclass, as shown in Figure 5. The gene expression pattern depicted in Figure 5 demonstrates that septic shock subclass A patients had generalized repression of these 44 genes, relative to subclasses B and C. Thus, patients in septic shock subclass A, having a clinical phenotype characterized by higher mortality, higher illness severity, and a higher degree of organ failure, are also characterized by repression of genes corresponding to key signaling pathways of the adaptive immune system. Importantly, the median absolute lymphocyte counts per mm³ (IQR) were not significantly different across the three subclasses: subclass A 1,585 (788–2,854); subclass B 1,530 (601–2,947); and subclass C 2,610 (1,329–4095).

Our previous studies demonstrated that pediatric septic shock is broadly characterized by large-scale repression of genes that either depend on normal zinc homeostasis for normal function or directly participate in zinc homeostasis [9, 11‐13]. In the current analysis, we determined if zinc biology-related genes were differentially regulated across the three septic shock subclasses. The 10 K-means clusters depicted in Figure 4 were interrogated for enrichment of zinc biology-related annotations ('zinc', 'zinc finger', 'zinc ion binding', 'metal binding', and 'metal ion binding') by uploading the individual gene cluster lists to the DAVID database. Cluster 8 was most significantly enriched for these functional annotations with P values ranging from 2.3E-5 to 9.7E-10 (data not shown). The genes corresponding to these zinc biology-related functional annotations were identified (181 genes) and subjected to hierarchical clustering based on the median expression values for each septic shock subclass, as shown in Figure 6. The gene expression pattern depicted in Figure 6 demonstrates that septic shock subclass A patients had generalized repression of these 181 zinc biology-related genes, relative to subclasses B and C. Thus, our previous observations regarding large-scale repression of zinc biology-related genes appears to be a relatively unique feature of patients in septic shock subclass A.