Early cytokine signatures and clinical phenotypes discriminate persistent from resolving MRSA bacteremia

Abstract

Background: Staphylococcus aureus bacteremia (SAB) is a prevalent life-threatening infection often caused by methicillin-resistant S. aureus (MRSA). Up to 30% of SAB patients fail to clear infection even with gold-standard anti-MRSA antibiotics. This phenomenon is termed antibiotic-persistent MRSA bacteremia (APMB). The mechanisms driving APMB are complex and involve host phenotypes significantly impacting the immune response. Thus, defining early immune signatures and clinical phenotypes that differentiate APMB from antibiotic resolving (AR)MB could aid therapeutic success.

Methods: We assessed 38 circulating cytokines and chemokines using affinity proteomics in 74 matched pairs of vancomycin-treated SAB cases identified as ARMB or APMB after 5 days of blood culture.

Results: Unsupervised hierarchical clustering segregated APMB from ARMB based on differential levels of IL-10, IL-12p40, IL-13, CCL4, and TGFα. Additionally, CXCL1, CCL22 and IL-17A significantly differed between APMB and ARMB when correlated with diabetes, dialysis, metastatic infection, or cardiac vegetation. Combining immune signatures with these relevant clinical phenotypes sharply increased accuracy of discriminating APMB outcome to 79.1% via logistic regression modeling. Finally, classification-regression tree analysis revealed explicit analyte thresholds associated with APMB outcome at presentation especially in patients with metastatic infection.

Conclusions: Collectively, this study identifies previously unrecognized cytokine and chemokine signatures that distinguish APMB and ARMB at presentation and in the context of host clinical characteristics associated with increased disease severity. Validation of a biomarker signature that accurately predicts outcomes could guide early therapeutic strategies and interventions to reduce risks of persistent SAB that are associated with worsened morbidity and mortality.

Keywords: Staphylococcus aureus; Clinical phenotypes; Cytokine signatures; MRSA; Persistence.

Fig. 1

Global cytokine and chemokine expression differentiate MRSA infection outcomes. 38 cytokines, chemokines, and growth factors were measured in peripheral blood of MRSA-infected patients at time of diagnosis by multiplex Luminex bead assay. Heat map shows expression of 22 cytokines that were detected above the minimum level of detection in more than 75% of patient samples and included in the analysis. Columns represent individual patients, rows represent individual cytokines, and colors represent normalized mean cytokine concentration values (blue = low, red = high). Infection outcome classified as “APMB” (turquoise) or “ARMB” (pink). Rows and columns are ordered based on results of unsupervised hierarchical clustering, with dendrograms for the cytokine and patient clusters shown on the horizontal and vertical axes, respectively. Groups 1–3 on X and Y axes represent the main patient clusters (X) and cytokine clusters (Y) as identified by unsupervised hierarchical clustering

Fig. 2

Principal circulating cytokines and chemokines are significantly elevated during persistent infection. Concentrations of five analytes trending (p < 0.1) or significantly different (p < 0.05) between ARMB and APMB patients in the combined cohort are shown. Significance determined via two-way unpaired t-test

Fig. 3

Principal cytokines chemokines alone are insufficient to discriminate APMB outcome. A Logistic regression model for combined training cohort. ROC curve built using normalized (ln) values of 5 cytokines shown in Fig. 2. B Logistic regression model for combined training cohort. ROC curve built using normalized (ln) values of 5 cytokines shown in Figs. 2 and 3 additional cytokines (CXCL1, CCL22, IL-17A) identified by Lasso regression

Fig. 4

Outcome of MRSA infection is correlated with the interactions between circulating analytes and clinical phenotypes. Pearson correlation analyses were performed to evaluate the relationship between original and Lasso-identified predictive cytokines and chemokines and selected clinical variables. Correlations are shown from strongly negative (dark red) to strongly positive (dark blue) between individual cytokines and chemokines (normalized (ln) values) and clinical variables. Strength of correlation is shown by darkness of box. Numbers in individual boxes show correlation value between 2 parameters. Any significant correlations between parameters are marked with the following to demonstrate level of significance: * = p < 0.05, ** = p < 0.01, *** = p < 0.001. A Correlation matrix of cytokines and clinical variables in full combined cohort. B, C Additional correlation matrices separated by ARMB (B) vs. APMB (C) infection outcome

Fig. 5

Inclusion of clinical phenotypes increases accuracy of discriminating APMB outcome and mortality. A Logistic regression model predicting persistence for training cohort. ROC curve built using normalized (ln) values of 5 cytokines/chemokines and 4 selected clinical variables. B Logistic regression model predicting persistence for training cohort. ROC curve built using normalized (ln) values of 8 cytokines/chemokines and 4 selected clinical variables. C Logistic regression model predicting mortality for training cohort. ROC curve generated using normalized (ln) values of 8 selected cytokines/chemokines and 4 selected clinical variables

Fig. 6

Classification-regression tree (CART) analysis of persistence parameters. CART analysis was performed for 8 cytokines/chemokines and 4 clinical phenotypes from logistic regression modeling to determine analyte thresholds associated with APMB outcome. “Yes” answers to threshold limits (ex. IL10 < 8.2pg/mL) branch left from title while “No” answers to threshold limits (i.e. IL10 > 8.2pg/mL) branch right. Cytokine/chemokine thresholds determining persistent outcome are shown in raw pg/mL values. Some analytes (i.e.IL-10) appear multiple times, and their thresholds may change based on previous nodes (see figure for specific threshold values)