Pathogen-specific host response in critically ill patients with blood stream infections: a nested case-control study

Abstract

Background: Knowledge of the contribution of the pathogen to the heterogeneity of the host response to infection is limited. We aimed to compare the host response in critically ill patients with a bloodstream infection (BSI).

Methods: RNA profiles were determined in blood obtained between one day before and after a positive blood culture. Differential expression and pathway analyses were performed on independent patients' samples by RNA sequencing (discovery) or microarray (validation). Additional patients were included for the discovery and validation of transcriptome classifiers of pathogen-specific BSIs. Twenty biomarkers reflecting key host response pathways were measured in blood.

Findings: We included 341 patients, among which 255 with BSI, 25 with viral infection and 61 non-infectious controls. The cultured pathogen explained 41·8% of the blood transcriptomic variance in patients with BSI. Gene set enrichment analysis showed a global resemblance between monomicrobial BSIs caused by Streptococcus, Staphylococcus aureus and Escherichia coli, which were clearly different from BSI caused by coagulase-negative staphylococci or Enterococcus. BSI by Streptococcus was associated with the highest number of differentially expressed genes, indicating strong innate and adaptive immune activation. An eight-gene streptococcal classifier performed well across different Streptococcus species, and was validated in external cohorts. Plasma biomarker profiling showed that E. coli BSI was associated with the strongest response in the cytokine and systemic inflammation domain, and S. aureus BSI with the strongest endothelial cell activation.

Interpretation: The causative pathogen explains a substantial part of the heterogeneity of the host response in critically ill patients with BSI.

Funding: Center for Translational Molecular Medicine and the European Commission.

Keywords: Bacteraemia; Biomarkers; Host response; Intensive care; Transcriptome.

Fig. 1

Patients with different monomicrobial BSIs exhibit distinct blood transcriptomic profiles. a) Venn–Euler plot showing the number of validated DEGs for each group with a positive blood culture with a single pathogen, compared to non-infectious controls. DEGs were discovered using RNAseq and validated in an independent cohort using U219 microarray; a DEG was defined by having an absolute value of Hedges' g > 0·8 in both cohorts and the direction of regulation being concordant. Not all intersections are shown due to the imperfect solution of the eulerr package; as a consequence 94 DEGs are not depicted. b) Heatmap showing the clustering of BSI groups based on pathway enrichment scores compared to non-infectious controls. Pathways from the four highest levels of the Reactome database that were significant for at least one monomicrobial group (versus non-infectious controls) are shown, i.e., if BH-adjusted-p < 0·05 in both discovery (RNAseq) and validation (U219) cohorts, and the direction of regulation was concordant between them (same sign of normalized enrichment score (NES)). In this way 344 pathways were significant and clustering was based on the mean of the two NES scores from discovery (RNAseq) and validation (U219) cohorts. c) Targeted pathway analysis of immune system pathways. Pathways are shown in a nested manner reflecting the Reactome hierarchical structure. Pathway significance was calculated separately for discovery and validation cohorts; then these two p values were combined using Fisher's method to which BH-adjustment was applied for multiple testing. Pathway direction (up or down regulation) was determined by the mean of two NES values which were calculated separately for discovery and validation. If the pathway direction in the two cohorts were discordant (in opposite direction) the pathway was deemed not significant (NS) and assigned a zero NES. d) Heatmaps showing expression levels of genes in selected pathways from adaptive, innate and cytokine signalling in the immune system. For each gene z-scores were calculated separately for both discovery and validation cohorts. For each pathway the DEGs with the 10 largest F-values (from one-way ANOVA on z-scores are shown, after assessing homogeneity of variance across groups using Levene's test). CoNS, coagulase-negative staphylococci; EC, E. coli; ENT, Enterocococcus; NI, non-infectious (controls); SA, S. aureus; STR, Streptococcus.

Fig. 2

Association between transcriptomic endotypes and BSI group. (a) Distribution of the four MARS endotypes, (b) the three SRS endotypes, and (c) the Inflammopathic, Adaptive, and Coagulopathic endotypes for each of the five BSI groups, assessed by combining the RNAseq and U219 cohorts. Fisher's exact test was used to test for association between BSI group and endotype. p values indicate the probability of observing the data or something more extreme assuming the null hypothesis of no association between BSI group and endotype is true. CoNS, coagulase-negative staphylococci; EC, E. coli; ENT, Enterocococcus; SA, S. aureus; STR, Streptococcus.

Fig. 3

Discovery and validation of transcriptomic classifiers for BSIs. a) Variance partitioning (in the monomicrobial BSI cohort) revealed that the cultured microbe explained the most transcriptomic variance compared to other clinical variables. Violin plots show the distribution of percentages of the variance explained by each explanatory variable over all 13,294 genes (protein coding common between three platforms); the gene which explained the most by each explanatory variable is labelled. Boxplots inside the violin plots show the median gene value (black bar) and interquartile range (white box). Genes shown in black circles are outliers, defined as points falling more than 1·5 times the interquartile range (IQR) from the first or third quartile. The table below shows the overall transcriptomic variance explained, both as a fraction of total transcriptomic variance including residual variance, and as a fraction of the summed explained transcriptomic variance excluding residual variance. b) Variance partitioning focussed on pathogen by dichotomizing the cultured microbe variable into indicator variables, one for each monomicrobial group (non-infectious control group as baseline). Note that the regression model for this analysis included all the same baseline characteristics and platform as in panel A, but only the microbe variables are shown. This analysis indicates that Streptococcus infections explain the most transcriptomic variance. Boxplots inside the violin plots show the median gene value (black bar) and interquartile range (white box), Genes shown in black circles are outliers, defined as points falling more than 1·5 times the interquartile range (IQR) from the first or third quartile. c) Gene set size selection using ULTRA with repeated cross-validation within the discovery cohort and Matthews correlation coefficient (MCC) to identify the optimal classification performance. The gene set with the maximum MCC was selected as the critical set size. After repeated cross-validation a mean value for each set size was calculated (shown as dots), then a natural spline model was fit to these data points to determine the overall maximum MCC, shown by the vertical red line. The natural spline model was fitted with 3 internal knots placed in the default positions (at the 25th, 50th, and 75th percentiles of the data). The shaded band is a 95% confidence interval of the spline model. d) Receiver operating curves (ROCs) showing critical model classification performance in the hold-out validation cohort for each of the BSI groups. Number of genes in each classifier is shown in parentheses prior to the letter “G”. The 95% CIs of the AUCs, estimated by DeLong's method, are shown in the second parentheses. After the critical set-size was determined, this number of genes was used to build the critical model using the discovery data, this final critical model was then assessed in the hold-out validation. Note that this entire analysis was performed on the full BC-taken patient cohort which included viral and polymicrobial infections as controls. e) Distribution of critical model scores, used for the final classification, in the hold-out validation for each of the BSI groups. CoNS, coagulase-negative staphylococci; EC, E. coli; ENT, Enterocococcus; SA, S. aureus; STR, Streptococcus.

Fig. 4

Performance of the STR8G Streptococcus classifier. a) ROC plot showing STR8G classifier performance in the discovery (black) and hold-out validation (magenta) cohorts. The 95% CIs of the AUCs, estimated by DeLong's method, are shown in parentheses. The red “X” represents the optimal threshold for the STR8G classifier as defined by the Youden index in the discovery cohort. b) Pie chart showing breakdown of Streptococcus BSIs into species subgroups, including the Lancefield groupings (groups A, B and G), in the combined discovery and validation cohort. c) Boxplots showing the STR8G classifier score across Streptococcus species subgroups. The “Other streptococci” group included group B, group G and unspecified streptococci. The red dashed line represents the optimal threshold defined by the Youden index in the discovery cohort. d) STR8G classifier performance assessed in an external cohort of patients with community-acquired pneumonia admitted to the ward (not ICU). The red dashed line represents the optimal threshold defined by the Youden index in the original discovery cohort. OC: other cultures (i.e., other than blood culture). e) STR8G classifier performance assessed in an external cohort of patients with sepsis due to severe community-acquired pneumonia in the ICU (SEPCELL study). This data was taken from a randomized controlled trial and performance of STR8G is shown here for whole blood transcriptome samples (RNAseq) taken within 18 h of ICU admission (day 0), prior the initiation of intervention or placebo. The red dashed line represents the optimal threshold defined by the Youden index in the original discovery cohort. OC: other cultures (i.e., other than blood culture). f) Estimation of the STR8G classifier trajectory over time in the cases (BC + for Streptococcus) of the SEPCELL external cohort. The blue line depicts the mean trajectory over time with a 95% CI of the mean (shaded band). This was calculated using LOESS regression to estimate a trajectory per patient with corresponding standard errors, from which a random value was generated to calculate a mean value across patients; this was repeated 1000 times to generate the overall trajectory, and the 95% CI calculated from the 2·5th and 97·5th percentiles. The grey lines represent individual patients. The red dashed line represents the optimal threshold defined by the Youden index in the original discovery cohort.

Fig. 5

Plasma biomarker host response profiles. Comparison of plasma biomarkers indicative of “Cytokine release and systemic inflammatory responses” (panels a and b) and “Endothelial cell and procoagulant responses” (panels c and d), in patients with bacteraemia stratified according to the causative pathogen. Data are presented as principal component analysis (PCA) plots (panels a and c, left). Ellipses in PCA plots represent patient data points for each BSI group (not shown here for clarity). Arrows in PCA plots indicate direction and extent of correlation of plasma markers with loadings of PCA components. The boxplots (panels a and c, right) show the difference in the first (PC1) and second component (PC2) loadings between groups (asterisks indicate differences between groups compared to the overall-mean of the PC; ∗p < 0·05, ∗∗p < 0·01, ∗∗∗p < 0·001, ∗∗∗∗p < 0·0001). Exact p values for 5a PC1 (NI, p = 0·0002; EC, p = 0·0083; SA, p = 0·018), PC2 (NI p = 0·031; SA, p = 0·0043). Exact p values for 5c PC1 (NI, p < 0·0001; EC, p = 0·0064; SA, p = 0·0024). For heatmaps (panels b and d) each group was compared to non-infectious controls and Hedges' g was used to evaluate the difference in plasma levels between these two groups. Ang, angiopoietin; CD, cluster of differentiation; IL, interleukin; MMP, matrix metalloproteinase; PT, prothrombin time; RA, receptor antagonist; s, soluble; TREM, triggering receptor expressed on myeloid cells; CoNS, coagulase-negative staphylococci; EC, E. coli; ENT, Enterocococcus; SA, S. aureus; STR, Streptococcus.