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

Abstract

Background: Septic shock is a heterogeneous syndrome within which probably exist several biological subclasses. Discovery and identification of septic shock subclasses could provide the foundation for the design of more specifically targeted therapies. Herein we tested the hypothesis that pediatric septic shock subclasses can be discovered through genome-wide expression profiling.

Methods: Genome-wide expression profiling was conducted using whole blood-derived RNA from 98 children with septic shock, followed by a series of bioinformatic approaches targeted at subclass discovery and characterization.

Results: Three putative subclasses (subclasses A, B, and C) were initially identified based on an empiric, discovery-oriented expression filter and unsupervised hierarchical clustering. Statistical comparison of the three putative subclasses (analysis of variance, Bonferonni correction, P < 0.05) identified 6,934 differentially regulated genes. K-means clustering of these 6,934 genes generated 10 coordinately regulated gene clusters corresponding to multiple signaling and metabolic pathways, all of which were differentially regulated across the three subclasses. Leave one out cross-validation procedures indentified 100 genes having the strongest predictive values for subclass identification. Forty-four of these 100 genes corresponded to signaling pathways relevant to the adaptive immune system and glucocorticoid receptor signaling, the majority of which were repressed in subclass A patients. Subclass A patients were also characterized by repression of genes corresponding to zinc-related biology. Phenotypic analyses revealed that subclass A patients were younger, had a higher illness severity, and a higher mortality rate than patients in subclasses B and C.

Conclusion: Genome-wide expression profiling can identify pediatric septic shock subclasses having clinically relevant phenotypes.

Figure 1

Unsupervised hierarchical clustering of 98 patients with septic shock (horizontal dimension) and 6,099 genes (vertical dimension) derived from a discovery-oriented filtering approach. Both the condition tree (patient clustering) and the gene tree are based on the Pearson correlation similarity measurement. The first- and second-order branching patterns of the condition tree were used to identify the putative septic shock classes and are colored for illustrative purposes based on three major putative septic shock subclasses.

Figure 2

Unsupervised hierarchical clustering of 98 patients with septic shock (horizontal dimension) and 6,934 genes (vertical dimension) derived from a three group analysis of variance. Both the condition tree (patient clustering) and the gene tree are based on the Pearson correlation similarity measurement. The first- and second-order branching patterns of the condition tree are colored for illustrative purposes based on septic shock subclasses A, B, and C.

Figure 3

Three-dimensional principal component analysis (mean centering and scaling) based on the 6,934 genes illustrated in Figure 2. Individual patients are plotted based on their respective positions along the three axes derived from principal component analysis. Patient subclassifications are indicated by color.

Figure 4

K-means clustering of 98 patients with septic shock (horizontal dimension) and the 6,934 genes (vertical dimension) shown in Figure 2. The K-means clustering algorithm is based on 100 iterations, the Pearson correlation similarity measurement, and a maximum return of 10 clusters. The first- and second-order branching patterns of the condition trees are colored for illustrative purposes based on septic shock subclasses A, B, and C.

Figure 5

Hierarchical clustering of the 44 genes shown in Table 4. Each gene is colored by the median expression values for each of the respective septic shock subclasses, as labeled at the bottom of the figure.

Figure 6

Hierarchical clustering of the 181 genes corresponding to zinc biology-related functional annotations and derived from K-means cluster 8 shown in Figure 4. Each gene is colored by the median expression values for each of the respective septic shock subclasses, as labeled at the bottom of the figure.