Study Design and Participants

The objective was to characterize the transcriptomic response to sepsis caused by FP, including an analysis of temporal changes and a comparison with sepsis caused by CAP. The subjects were adult patients admitted to the ICU with sepsis as part of the UK Genomic Advances in Sepsis (GAinS) study (www.ukccg-gains.org) with predefined inclusion and exclusion criteria (described in the online supplement). FP was diagnosed at laparotomy as inflammation of the peritoneal membrane secondary to large bowel perforation and fecal contamination (18). CAP was defined as a febrile illness associated with cough, sputum production, breathlessness, leukocytosis, and radiological features of pneumonia acquired in the community or within 2 days of ICU admission (21, 22).

This was a prospective observational study. The transcriptomic response was investigated in peripheral blood leukocytes and compared between patients and over time, and with control subjects undergoing elective cardiac surgery (described in the online supplement). Discovery and validation cohorts were used to identify shared and specific gene expression patterns among patients with FP and patients with CAP (see Figure E1 in the online supplement).

Sample Collection

Blood samples (5 ml) were collected (VACUETTE ethylenediaminetetraacetic acid–coated tubes; Greiner Bio-One, Kremsmünster, Austria) on the first, third, and/or fifth day after ICU admission. The total blood leukocyte population was isolated using LeukoLOCK filters (Life Technologies, Carlsbad, CA) and then stabilized using RNAlater (Life Technologies), followed by total RNA extraction (described in the online supplement). Extensive, anonymized clinical information was recorded using an electronic case report form (5).

Microarray Analysis

Genome-wide gene expression was quantified using Illumina HumanHT-12 v4 Expression BeadChips (47,231 probes; Illumina, San Diego, CA) with sample processing, data preparation, background subtraction, transformation, and normalization using the vsn package (23) (described in the online supplement). Gene expression data are available through the ArrayExpress database (accession number E-MTAB-5273/E-MTAB-5274).

Statistical Analysis

Analysis was performed using the R statistical software package (R Foundation for Statistical Computing, Vienna, Austria) unless otherwise specified, with statistical tests, power calculations, differential gene expression, enrichment testing, predictive modeling, and cluster analysis described in the online supplement.

Transcriptomic Response to FP: Identification of SRSs

We quantified genome-wide gene expression for 117 adult patients with FP and 126 adult patients with CAP admitted to the ICU with sepsis, analyzing RNA from blood leukocytes rapidly isolated at the bedside. All patients showed evidence of organ dysfunction based on Sequential Organ Failure Assessment (SOFA) scores during ICU admission. These patients were recruited concurrently and processed in parallel as discovery (64 with FP, 73 with CAP) and validation cohorts (53 with FP, 53 with CAP; the 53 patients with CAP were previously published [5]). We also analyzed samples obtained from 10 subjects prior to elective cardiac surgery as nonseptic control subjects in the discovery cohort. Demographics and clinical covariates for the discovery set (FP, n = 64 patients; CAP, n = 73 patients; 221 samples; see Figure E1) are shown in Tables 1 and E1A.

We investigated whether sepsis response subgroups based on global patterns of gene expression were present in FP. The combined FP cohort (147 samples from 117 patients) had a 28-day mortality of 13%; the mean age was 66 years, and 50.4% of patients were male (Table 2, Table E1B). We applied two complementary approaches to this FP cohort: (1) hypothesis-free unsupervised hierarchical clustering based on observed patterns of variation in gene expression and (2) assignment based on expression of a specific seven-gene set previously shown to be predictive of SRS group membership in CAP (5).

We first analyzed variation in global gene expression (2,716 probes representing the top 10% most variably expressed genes) using agglomerative hierarchical clustering followed by consolidation of group membership using k-means. This approach revealed two distinct patient groups: SRS1_FP (n = 68 [46%]) and SRS2_FP (n = 79 [54%]) (Figures 1A and 1B). Following group assignment, 1,075 genes were found to be differentially expressed between these groups (Figure 1C, Table E1C), showing strong correlation with those previously observed in patients with CAP (5) (Pearson’s r = 0.791) (Figure E2). To determine the functional overlap with SRS groups identified in CAP, we performed pathway, function enrichment, and upstream regulator analysis (Figure 1D, Table E1D). Cell death, apoptosis, necrosis, T-cell activation, and endotoxin tolerance were significantly enriched biological functions in SRS1 for both sources of infection. We proceeded to test for association with outcome in FP and found that SRS1_FP group membership was associated with higher early mortality (14-d mortality log-rank test P = 0.0096; 18.8% vs. 4.3%; hazard ratio, 4.78; 95% confidence interval, 1.29–17.65) (Figure 1E). The effect of SRS group membership remained after inclusion of age, SOFA score, Charlson comorbidity index, and sex in a Cox proportional hazards model (P = 0.037; hazard ratio, 4.23; 95% confidence interval, 1.09–16.39).

To further validate our findings, we adopted a second approach, namely assignment of SRS group membership in this FP cohort using the seven-gene set (DYRK2, CCNB1IP1, TDRD9, ZAP70, ARL14EP, MDC1, and ADGRE3) we previously established in patients with CAP (5) (Figure E3). This showed strong concordance with the results obtained by unsupervised analysis (misclassification rate, 21.1%; area under the curve [AUC], 0.923) (Figure 1F). The groups defined using the seven-gene set showed a significant difference in early mortality rates (log-rank test P = 0.030 for patients with FP), a differential gene expression signature strongly correlated with that seen in the original CAP analysis (r = 0.845), as well as pathway enrichment comparable to findings in patients with CAP. We also derived a six-gene set (CD163, ZDHHC19, MME, FAM89A, ZBP1, and B3GNT2) from the FP data predictive of SRS_FP group membership, with a 4.1% misclassification rate based on internal leave-one-out cross-validation (AUC, 0.975) (Figure 1F). When applied to the original CAP cohort (5), this six-gene set again performed well (misclassification rate, 27.9%; AUC, 0.931).

A poor outcome subgroup has previously been reported in children with sepsis (24). We investigated the similarity between pediatric endotype A and SRS1 by comparative differential gene expression analysis. Although there was some overlap in pathway enrichment (e.g., T- and B-cell receptor signaling), the gene expression patterns that distinguish SRS groups were not enriched in the pediatric endotype contrast (Figure E4).

Temporal Changes in SRSs

Serial sampling on Day 1, Day 3, and Day 5 following ICU admission provides the opportunity to investigate the relationship between SRS group membership and disease progression. We found that 11 (46%) of 24 patients with FP with serial samples moved between groups over time; 10 moved from SRS1 to SRS2, only 1 of whom died (Figure 1G). Thirteen patients (54%) remained in the same SRS group; three of seven patients remaining in SRS1 died, compared with no deaths among the six remaining in SRS2. This movement between SRS groups involves large changes in gene expression, illustrated by CD163, encoding an innate immune sensor for bacteria, and one of the six gene classifiers for SRS_FP (Figures 1H and E5).

Influence of Source of Infection on the Transcriptomic Response

We proceeded to further characterize and compare gene expression in patients with FP and patients with CAP to determine the relative importance of shared and specific features of the transcriptomic response. As expected, in the discovery cohort, gastrointestinal comorbidities were more common in patients with FP, whereas patients with CAP had more respiratory comorbidities, higher respiratory rates and oxygenation requirements, and higher lymphocyte counts (Table 1 and Table E1A). Although mortality did not differ significantly, patients with CAP had higher Acute Physiology and Chronic Healthy Evaluation II and SOFA scores than patients with FP (Table 1).

To understand the relationships between source of infection, SRS group, and heterogeneity in the septic response, we performed principal component analysis using the top 10% most variably expressed transcripts. Nonseptic control subjects clustered together and were clearly distinct from septic patients (Figure 2A). Among the combined population of patients with sepsis, there was clear segregation based on SRS group but not source of infection (Figure 2A).

To further elucidate the drivers of variation in the septic response, we calculated correlations between the observed variance in global gene expression for the sepsis samples (the first six principal components of variance, representing approximately 50% of the variance in the data), a comprehensive set of clinical covariates, and SRS group (Figure 2B, Table E1E). This showed that variation in gene expression was most strongly correlated with SRS group membership (PC1 r = 0.77; false discovery rate [FDR], <2.2 × 10⁻¹⁶). More modest correlation was seen with source of infection (PC1 r = −0.417; FDR, 6.12 × 10⁻⁹), bicarbonate levels (PC1 r = 0.35; FDR, 2.2 × 10⁻⁸), and neutrophil count (PC1 r = −0.29; FDR, 2.0 × 10⁻³). When the dataset was divided into CAP and FP samples and the analysis repeated, we found that SRS remained the most significant correlate (Figure E6).

When we compared all patients in the discovery cohort (FP, n = 64; CAP, n = 73) with nonseptic control subjects (n = 10), we found that sepsis was associated with differential regulation of a large proportion of the transcribed genome (18.0% of transcripts assayed; 4,881 probes; >1.5-fold change [FC]; FDR, <0.05) (Table E1F). Pathway analysis showed that eukaryotic initiation factor 2 signaling was the top canonical pathway, consistent with the reported role for this translational initiation factor in response to viral and bacterial infection (25), and predicted key upstream regulators IL-13, ATB1, TGFB1, and IL-2 (Table E1G). We used the same methodology to compare gene expression in trauma patients with healthy subjects, using the previously published genomic response to trauma dataset (26) but restricting this to samples from the same time frame as the sepsis cohort (5 d following injury). We found that, of the genes differentially expressed in the sepsis response and measured in the trauma cohort, the majority (n = 1,884 [64%]) were also differentially expressed in trauma (Figure 2C). Commonality was seen with eukaryotic initiation factor 2 signaling and inflammation-related pathways enriched in both sepsis and trauma responses, whereas the inflammatory response to organismal injury was specific to trauma and tumor necrosis factor receptor 1 (TNFR1) signaling was specific to sepsis (Figure 2C, Table E1H).

When we compared patients with CAP versus control subjects and patients with FP versus control subjects, we observed that gene expression patterns were highly correlated, demonstrating that most sepsis response pathways are common and independent of source of infection (Figures 2D, 3A, and 3B; Tables E1I and E1J). To further investigate differential gene expression between patients with sepsis due to FP or CAP, we compared the first available samples for each patient following ICU admission (FP, n = 64; CAP, n = 73). We found 310 probes (263 genes) differentially expressed between the two sources of infection (FC, >1.5; FDR, <0.05) (Figure 3C, Table E1K), significantly more than expected by chance based on permutation analysis (Figure E7). We noted significant differences in total and differential cell counts between CAP and FP. The inclusion of differential cell counts in the regression model had minimal influence, with a similar list of differentially regulated genes and FCs strongly correlated with the original analysis (Figure E8, Table E1K). IFN signaling (P = 6.1 × 10⁻¹⁰) and antigen presentation (P = 1.6 × 10⁻⁸) were the most significant enriched canonical pathways, upregulated in patients with CAP compared with FP (Figure 3D, Table E1L). Enriched networks involved IFN-α/β and the antimicrobial/inflammatory response (P = 1 × 10⁻³⁹) (Figure 3E). IFIT1, IFIT2, and IFIT3 (IFN-induced antiviral genes), BTN3A3 (major histocompatibility complex class I gene), IFIH1 (sensor of viral nucleic acids), and OAS2 (IFN-induced antiviral enzyme) were all upregulated in patients with CAP. IFNs, endotoxins, and TNF were found to be significant upstream regulators (Figure 3D, Table E1L).

We confirmed our findings in a validation set comprising 53 patients with FP and 53 patients with CAP prospectively recruited to the UK GAinS study (Figure E1). We found SRS group, source of infection, cell count, and day of sampling were strongly associated with the first six principal components of gene expression; genes differentially expressed between FP/CAP correlated with those found in the discovery cohort (Pearson’s r = 0.83; P < 2.2 × 10⁻¹⁶) (Table E1M). Pathway analysis demonstrated high concordance of enriched pathways and functions (Figure 3D, Table E1N). We found that a gene set comprising EPHB1, NQO2, ARG1, HMGB2, FGL2, and GPR162 was predictive of source of infection (FP or CAP) for the discovery cohort. We applied a sparse regression method to show that, for the discovery cohort, the misclassification rate was 29.9% with internal leave-one-out validation (AUC, 0.760); for the validation cohort, the test error was 28.3% (AUC, 0.798) (Figure E9). There was no difference in the proportion of different sampling days following ICU admission between FP and CAP samples in the analyzed cohorts (chi-square P = 0.350) (Figure E10).

Given a prominent viral signature among genes differentially expressed between patients with FP and patients with CAP, we investigated if this came from a subset of patients with CAP with a viral infection. Only 6 of 73 patients with CAP in the discovery cohort had a confirmed viral infection; we therefore analyzed the transcriptomic differences between patients with and without confirmed viral infection in our larger, previously published CAP cohort (25 vs. 240 patients) (5). We identified 39 differentially expressed genes (FDR, <0.05; FC, >1.5) (Figure 3F, Table E1O), including IFI27, an IFN-α–inducible protein reported as a marker of influenza infection (27), and LAMP3, a dendritic cell glycoprotein induced by influenza A infection (28). Pathway analysis showed the most significant enrichment for pattern recognition receptors, IFN signaling, and IFN regulatory factors (Table E1P). We then compared this gene set to the genes differentially expressed between patients with FP and patients with CAP and found significant enrichment of 24 of 39 genes (P <  1 × 10⁻⁴).

Temporal Changes in Gene Expression

Interpretation of gene expression patterns in sepsis may be complicated by dynamic differences in the host response over time (10, 29). We explored temporal changes in gene expression for patients with FP and patients with CAP from whom serial samples following ICU admission were available. We first sought to identify genes that varied over time using a repeated measures regression model. We found that 714 genes were significantly differentially expressed in patients with FP sepsis between Days 1 and 5, compared with only 80 genes in patients with CAP (FDR, <0.05; FC, >1.5) (Figure 4A, Table E1Q). We further analyzed our data using a multivariate empirical Bayes model that ranks genes based on differential expression for longitudinal series involving multiple biological conditions (30), restricting this analysis to patients from whom samples were available at all three time points. Notably, temporal changes in gene expression were again more pronounced in patients with FP (Table E1R), and the specific genes whose expression changed over time differed between FP and CAP (Figures 4B and C).

In patients with FP, it was possible to estimate with reasonable accuracy the time of onset of the acute event. We calculated days from disease onset for each FP sample and used this to investigate temporal changes. Using a linear model with limma analysis, we found that 140 genes showed significant changes in expression over time (FDR, <0.05) (Table E1S), including AAK1 and SNN (Figure 4D). Pathway analysis was significant for genes involved in phagosome formation; natural killer cell and IL-3 signaling with evidence of enrichment for specific biological functions, notably cellular degranulation, chemotaxis, leukocyte activation, macrophage adhesion, phagocytosis, and bacterial infection (Table E1T).