A systems immunology approach to investigate cytokine responses to viruses and bacteria and their association with disease

Abstract

Patterns of human immune responses to viruses and bacteria and how this impacts risk of infections or onset/exacerbation of chronic respiratory diseases are poorly understood. In a population-based birth cohort, we measured peripheral blood mononuclear cell responses (28 cytokines) to respiratory viruses and bacteria, Toll-like receptor ligands and phytohemagglutinin, in 307 children. Cytokine responses were highly variable with > 1000-fold differences between children. Machine learning revealed clear distinction between virus-associated and bacteria-associated stimuli. Cytokines clustered into three functional groups (anti-viral, pro-inflammatory and T-cell derived). To investigate mechanisms potentially explaining such variable responses, we investigated cytokine Quantitative Trait Loci (cQTLs) of IL-6 responses to bacteria and identified nine (eight novel) loci. Our integrative approach describing stimuli, cytokines and children as variables revealed robust immunologically and microbiologically plausible clustering, providing a framework for a greater understanding of host-responses to infection, including novel genetic associations with respiratory disease.

Figure 1

Cytokine response patterns according to stimulus. Fold induction (response/media) of a panel of 28 innate, inflammatory and immune cytokines from PBMCs; each dot represents a child. (a) Viral stimuli: RSV, RV1B, RV16, polyIC, R848 and CpGA; (b) T-cell stimulus: PHA. (c) Bacterial stimuli: Strpn, LTA, Hin, LPS, PAM, PGN, FSL and Fla. This figure was generated using beeswarm package in R (version 3.6.3, https://www.R-project.org/).

Figure 2

Hierarchical clustering (HC) of cytokine responses. (a) Heatmap of the mean level of response for each cytokine-stimulus pair (307 children, media-normalised). (b) Hierarchical tree for stimuli. (c) Hierarchical tree for cytokines. Values on nodes indicate probability values (%) of the found clusters appearing in bootstrap resampling. Green numbers are estimates from ordinary bootstrap probabilities (BP); red numbers are “approximately unbiased” (AU) estimates from multi-scale bootstrap resampling, which are less biased estimates than BP (see “Methods” section). Plot (a) was generated using seaborn package in Python (version 3.8, https://www.python.org/). Plots (b,c) were generated using pvclust package in R (version 3.6.3, https://www.R-project.org/).

Figure 3

Hierarchical clustering for pro-inflammatory cytokine data. (a) HC with heatmap on pro-inflammatory cytokine responses to all stimuli. (b) Pairwise Pearson correlation coefficients of production of pro-inflammatory cytokines in response to bacterial ligands. MCP4 and IL-17 were weakly induced and so are excluded from this panel. (c) Scatter plot with IL-6 responses to LPS versus IL-6 responses to Hin, and Pearson correlation coefficient between LPS and Hin. This figure was generated using seaborn package in Python (version 3.8, https://www.python.org/).

Figure 4

Association of rs8028121 with cytokine-stimulus pairs. The cQTL rs8028121 was associated with IL-6 production in response to bacterial stimuli (x-axis = 1) more strongly than IL-6 production in response to viral stimuli (x-axis = 2), other pro-inflammatory cytokine responses to bacterial stimuli (x-axis = 3), virus-induced cytokine responses to viral stimuli (x-axis = 4), pro-inflammatory cytokine responses to PHA (x-axis = 5), and virus-induced cytokine responses to PHA (x-axis = 6).