Genomic Circuitry Underlying Immunological Response to Pediatric Acute Respiratory Infection

SUMMARY

Acute respiratory tract viral infections (ARTIs) cause significant morbidity and mortality. CD8 T cells are fundamental to host responses, but transcriptional alterations underlying anti-viral mechanisms and links to clinical characteristics remain unclear. CD8 T cell transcriptional circuitry in acutely ill pediatric patients with influenza-like illness was distinct for different viral pathogens. Although changes included expected upregulation of interferon-stimulated genes (ISGs), transcriptional downregulation was prominent upon exposure to innate immune signals in early IFV infection. Network analysis linked changes to severity of infection, asthma, sex, and age. An influenza pediatric signature (IPS) distinguished acute influenza from other ARTIs and outperformed other influenza prediction gene lists. The IPS allowed a deeper investigation of the connection between transcriptional alterations and clinical characteristics of acute illness, including age-based differences in circuits connecting the STAT1/2 pathway to ISGs. A CD8 T cell-focused systems immunology approach in pediatrics identified age-based alterations in ARTI host response pathways.

Graphical abstract

INTRODUCTION

Acute respiratory tract infections (ARTIs) are the leading worldwide cause of death in early childhood (Bryce et al., 2005; Williams et al., 2002) and are responsible for 75% of acute illnesses in young children (Cebey-López et al., 2015). Lower respiratory tract infections (LRTIs, including bronchiolitis and viral and bacterial pneumonia) are the leading cause of pediatric hospital admission for infectious disease (Goto et al., 2016). Viral infections are the primary etiology of most ARTIs (Gooskens et al., 2014; Huijskens et al., 2012). Influenza virus alone results in 50,000–430,000 hospitalizations and 3,000–49,000 deaths annually in the US (Wong et al., 2013). There have been detailed studies of immune responses in the setting of influenza vaccination (Herati et al., 2014; Tsang et al., 2014) and challenge (Davenport et al., 2015, 2016; Memoli et al., 2016; Wilkinson et al., 2012). However, our understanding of the variation in outcome after natural infection and in the host factors and gene circuits underlying diverse clinical responses remains incomplete. These issues are even less well understood in pediatric populations, where previous viral exposure or vaccination may play a larger role. It remains unclear how age influences clinical and immunological outcomes. Moreover, little information exists on global systemic changes in specific key antiviral cell types, such as CD8 T cells, that might differ in early life due to developmental changes and immunological exposure.

Host factors, including obesity, age, asthma status, and sex, may shape the inflammatory response and clinical outcome during ARTIs. Moreover, there is increasing evidence that such host factors can also influence adaptive immunity, including T cell responses. Although obesity in adults (Fezeu et al., 2011) and pediatric patients has been correlated with higher morbidity and mortality with the H1N1 influenza epidemic, obese patients do not generally have worse outcomes with other infections (Martino et al., 2011; Ray et al., 2005) (at least in adults [Ross et al., 2016]). Sex has also been shown to modify influenza infection (Klein et al., 2012; Robinson et al., 2011). Moreover, there is evidence of poor vaccine responses in older adults (Herati et al., 2014; McElhaney et al., 2016; Osterholm et al., 2012) and increased influenza morbidity and mortality in the elderly (Thompson et al., 2003) and children (Brownstein et al., 2005). We lack an understanding of the mechanistic link between core host immune pathways, clinical characteristics, and outcomes of individual patients.

Distinguishing different viral infections with whole blood or peripheral blood mononuclear cell (PBMC) transcriptomics has been a major goal (Bermejo-Martin et al., 2010; Herberg et al., 2013), and some datasets have been refined into signatures capable of identifying acute influenza (Berdal et al., 2011; Davenport et al., 2015; Ioannidis et al., 2012; Woods et al., 2013) or distinguishing between bacterial versus viral infections (Herberg et al., 2016; Parnell et al., 2012; Suarez et al., 2015; Tsalik et al., 2016), between different viral infections (Mejias et al., 2013; Statnikov et al., 2010; Zaas et al., 2009; Zhai et al., 2015), or both (Hu et al., 2013; Ramilo et al., 2007; Sweeney et al., 2016; Tsalik et al., 2016). Many studies highlight the role of interferon-stimulated genes (ISGs) (Zhai et al., 2015) as key upregulated genes consistent with the important role of interferon and STAT1/2 signaling in viral infection. Human influenza virus (IFV) studies to date have transcriptionally profiled PBMCs or whole blood. A key gap in knowledge remains about the core anti-influenza host transcriptomic response in CD8 T cells, a critical antiviral adaptive immune cell type. In addition, many existing studies have focused on upregulated genes, with less attention given to pathways that are downregulated.

Despite studies in influenza and other causes of ARTIs, many gaps remain in our understanding of how variability in the host at a macroscopic level (sex, age, asthma status) and microscopic level (transcriptional analysis of CD8⁺ T cell responses) affect responses to influenza. CD8 T cells are critical early antiviral immune cells, responding in non-immune hosts by rapidly sensing innate inflammatory signals in preparation for possible antigenic stimulation. In immune hosts, these cells can act as early as innate immune cells by “sounding the alarm” and producing or provoking production of key antiviral cues (Iwasaki and Medzhitov, 2015; Masopust and Schenkel, 2013). Thus, our analyses focused on the impact of the innate immune response in conditioning the CD8 T cell pool in acute IFV.

Our results identify distinct transcriptional signatures of different ARTIs and connect transcriptional signatures in influenza infection to clinical traits. Moreover, we developed a robust predictive signature of up- and downregulated genes. This signature has high precision and specificity to distinguish influenza infection from other viral and non-viral infections from infancy through adulthood. Parsing this signature in pediatric subjects revealed a dichotomy in the activation of ISGs in CD8 T cells in younger versus older children that was associated with serological evidence of previous influenza exposure. Together, these studies used a cell-type-focused, transcriptional-driven systems immunology approach to not only link signatures of ARTIs to clinical metadata, but also to develop a highly specific influenza infection signature that provides insights into the underlying biology of pediatric IFV infection.

RESULTS

Study Design

Participants with influenza-like illness (ILI; fever and cough in viral respiratory season) were recruited from The Children’s Hospital of Philadelphia (CHOP) Emergency Department (ED) during the summer and fall of 2009 and healthy controls (HCs) were recruited in the orthopedic pre-surgical clinic (Figure S1). Patients with ILI were clinically tested, and 29 patients with ARTIs (influenza [IFV], human rhinovirus [HRV], respiratory syncytial virus [RSV], and each of these with co-infections; Table 1) were identified along with 18 HC patients. For each patient and HC, CD8 T cells were sorted and subjected to transcriptional profiling.

Table 1.

Demographics of ILI and HC Cohorts

Patient Population	n	Age (y, Mean)	Asthma Pre-study	Asthma Currently	Concern for Bacterial Infection
HC	18	6.4	ND	ND	NA
CHOP ILI cohort (below are subgroups)	29	4.8	7/29	9/29	9/29
Influenza A only (14/15 H1N1; 1/15 untyped)	15	7.6	5/15	5/15	5/15
Influenza A and rhinovirus	6	3.4	2/6	1/6	1/6
Rhinovirus only	4	1.2	0/4	1/4	1/4
RSV only	1	0.03	0/1	0/1	0/1
RSV-rhinovirus	2	1	0/2	2/2	2/2
RSV-adenovirus	1	1	0/1	0/1	1/1

Within ILI cohort, 17 males, 12 females. NA, not applicable; ND, not determined.

ILI and Gene Circuitry

To dissect the transcriptional signatures of the CD8 T cells associated with ARTIs, we interrogated gene expression changes caused by IFV, HRV, and/or RSV infection. Compared with CD8 T cells from HCs, ILI patients showed focused upregulation of interferon-stimulated genes (ISGs) as expected, but also broad downregulation of many genes (Figure 1A). These downregulated genes included integrins (e.g., ITGA4, ITGAL, ITGB2, and ITGAM), signaling molecules (e.g., STIM1, MAP3K12, MAP4K1, FOSB, and JUN), cell surface receptors (e.g., KLRG1 and, CD226) and the pro-survival gene BCL2 (full gene lists in Tables S1 and S2). To understand the contribution of individual viruses and the effect of co-infection, we next compared the effect of the three most common infection scenarios in this cohort (IFV, HRV, and IFV-HRV co-infection) with HCs (Figure 1B). This analysis again revealed prominent downregulation of genes in the infected groups, but also pathogen-specific upregulation of ISGs broadly (IFV), LRP6 (HRV), or IFI44L (IFV-HRV) (Figure 1B). A partial least-squares discriminant analysis (PLS-DA) plot compared the effects of IFV and HRV (and co-infection) on gene expression (Figure 1C) and revealed two groups of IFV patients bridged by a group of IFV-HRV co-infection, with the HRV group more proximal to HCs, suggesting a more prominent transcriptional impact of IFV on CD8 T cells than HRV. Circos plots (Figures 1D and 1E) showed that signatures from IFV infection appeared to dominate in patients with IFV-HRV co-infection. Analysis of Gene Ontology (GO) Biological Processes (BP) gene sets demonstrated that transcriptional changes were often shared between IFV-only and IFV-HRV groups, but not between IFV-only and HRV-only groups (Figure S2 and S3). Predicted upstream regulators were assessed in the IFV-only group using ingenuity pathway analysis (IPA; Figure 1F). This set of predicted regulators was then evaluated in the IFV-only, HRV-only, or co-infected groups versus HCs. Upregulation of potential transcriptional regulators that are also ISGs in IFV (and IFV-HRV co-infection) was noted (IRF3, IRF4, IRF5, IRF7, IRF9, STAT2, STAT1, and STAT6) as expected. In contrast, HRV-only patients displayed higher expression of NOTCH, FOX, SOX2, and KLF2. Finally, the top three GO BP gene sets for IFV-alone and HRV-alone patient data were interferon (IFN)- α response, IFN- γ response, and inflammatory response for IFV-only and IFN- α response, MYC targets, and cholesterol homeostasis for HRV-alone. Although there was overlap in some signatures between IFV and HRV, it was also clear that IFV-only and IFV-HRV versus HRV only had distinct transcriptional imprints on circulating CD8 T cells (Figure 1G). Thus, these data support the idea that different viral ARTIs cause distinct transcriptional imprints on cells in the peripheral blood. Moreover, the focus on CD8 T cells revealed transcriptional changes of particular relevance for the induction of this key adaptive immune cell.

Figure 1. Effect of Acute Pediatric Upper Respiratory Viral Infection on CD8⁺ T Cell Gene Expression.

(A) Heatmap of 100 genes most significantly different between CD8 T cells from HCs (HC) and patients with influenza-like illness (influenza virus [IFV], rhinovirus [HRV] or co-infection [IFV-HRV]). Row normalized z scores are presented.

(B) Heatmap of 20 class neighbors from the CD8 T cells from each of the four groups of patients (HC, IFV, IFV-HRV co-infection, and HRV). Row normalized z scores are presented.

(C) Three-dimensional partial least-squares discriminant analysis (PLS-DA) plot illustrating the differences in gene expression in CD8 T cells from IFV (red), IFV+HRV (blue), HC (light blue), and HRV (green).

(D and E) Circos plots and summary tables of GO biological process terms enrichment, demonstrating shared upregulation (D) and downregulation (E) of genes between CD8 T cells from groups of virally infected patients (e.g. IFV, IFV+HRV, and HRV) compared with HC (i.e. IFV versus HC compared with HRV versus HC). The p values are for the shared genes in the viral infections (denoted by the colored squares) versus HC. The summary table has the top 10 BP gene signatures by p value.

(F) Predicted upstream regulatory transcription factors for each group. Each group was compared with HC and the list of significant genes was calculated. IPA was used to determine the upstream genes most likely to be causative. The z scores are presented.

(G) Radar plots of the top three GSEA hallmark signatures for IFV, IFV-HRV, or HRV. The normalized enrichment scores from GSEA of hallmark gene sets are shown.

See also Figures S1–S3 and Tables S1 and S2.

Acute Influenza and Alterations in CD8⁺ T Cell Gene Circuitry

Because of the high incidence in this cohort and transcriptional dominance of IFV, we focused on IFV. We compared global gene expression changes between CD8 T cells in pediatric HC versus pediatric acute IFV (Figure 2A; Table S3), revealing upregulated genes and numerous downregulated genes. PLS-DA showed a clear distinction between IFV and HC groups, as well as two subgroups of IFV (Figure 2B). Co-infected patients were excluded from this analysis, and these two groups of IFV patients did not segregate based on any clinical parameter tested (including age, sex, return to primary care physician [PCP] within one month, asthma status at study visit, or weight; data not shown). To understand the transcriptional circuitry underlying the major differences between HCs and IFV patients, we assessed the predicted upstream regulator network for HC versus IFV-alone infected patients (Figure 2C). These analyses revealed predicted upregulation of STATs and IRFs and downregulation of NFκB and effector molecules, including IFN- γ and granzyme A in IFV-infected patients. These signatures are unlikely to reflect major contributions of virus-specific CD8 T cells since at the time points examined here (likely day 2–6 post-infection; see below), because IFV-specific CD8 T cells are rare to undetectable (data not shown and Wilkinson et al., 2012). Moreover, because IFV does not cause viremia and peripheral CD8 T cells are not directly infected by IFV, the alterations in CD8 T cell transcriptional networks are more likely due to conditioning of the global CD8 T cell population by early innate antiviral events. The prominent signature of IFN signaling in circulating CD8 T cells in IFV infection is expected (Brinza et al., 2016) and reveals both the systemic impact of IFN circuits in respiratory IFV infection and the ability of circulating CD8 T cells to sense and respond to this key signal of viral infection. In contrast, the downregulated genes and pathways including an NFκB-focused hub was unexpected. When top GO BP signatures were examined in the genes upregulated, the expected upregulation of ISGs was again prominent (Figure 2D, upper). Gene sets from the Molecular Signatures Database (mSigDB) C7 that focus on immunological signatures also identified prominent features of activation, infection, or vaccination in the transcriptional profiles of CD8 T cells (Figure 2E). In addition to the upregulated genes and pathways, the GO BP analysis also revealed downregulated metabolic pathways, targets of MYC, and pathways of regulation of protein synthesis (E2F targets) (Figure 2D, lower). A leading-edge analysis of the manually curated signatures in Figure 2F demonstrated a number of complement factors (SERPING1, C2, and C1QL1), ISGs (MX2), chemokines and receptors (CCL2 and CCR1), and interleukin-1 (IL-1) receptor antagonist and IL-1 receptor subunit (IL1RN and IL1R2; Figure 2F). The preponderance of downregulated genes in CD8 T cells in acute IFV was surprising, and we next examined the links between transcriptional changes and patient clinical status.

Figure 2. Effect of Pediatric Acute IFV Infection on CD8⁺ T Cell Gene Expression.

(A) Heatmap of 100 genes most significantly different between CD8 T cells from HC and IFV patients.

(B) Three-dimensional PLS-DA plot displaying the differences between gene expression in CD8 T cells from HC (blue) and patients with IFV (red).

(C) Upstream regulator network for transcriptional changes in CD8 T cells from IFV patients. IPA was used to determine the upstream genes most likely to be causative.

(D) Top three GSEA hallmark signatures upregulated (top) and downregulated (lower) in CD8 T cells from patients with IFV versus HC.

(E) Selected GSEA MSigDB C7 signatures in CD8 T cells from patients with IFV versus HCs. GSE29615 (human PBMC control versus day 7 post-live-attenuated influenza vaccine, upregulated with vaccine); GSE29614 (human PBMC control versus day 7 post-trivalent-inactivated influenza vaccine, upregulated with vaccine); GSE13485 (human PBMC control versus yellow fever vaccine at day 7, upregulated with vaccine); GSE36476 (CD4 young versus elderly CD4 T memory cells, higher in young); GSE19825 (upregulated in mouse effector T cells at day 3.5 after LCMV infection versus naive cells); GSE24634 (upregulated in CD25⁺ T effector versus CD25⁻ T effector cells after in vitro activation with interleukin-4 [IL-4]).

(F) Leading edge analysis of the manually selected C7 signatures in Figure 2E, with a heatmap including genes present in the LE of at least two gene signatures (black = included, gray = not included).

See also Table S3.

WGCNA Reveals Gene Modules with Clinical Characteristics

Clinical data were available both at the time of the ILI (i.e., 2009) and during the subsequent seven years (i.e., up to 2016). Thus, weighted gene co-expression network analysis (WGCNA) (Horvath et al., 2006) was used to interrogate predicted underlying mechanisms in CD8 T cells from patients with acute IFV and link these modules to clinical features (Figure 3). WGCNA identified modules of transcriptional changes in CD8 T cells and their relationship to key clinical features, such as asthma status, sex, PCP visits within one month of the study visit, previous and current weight, and age (Figures 3A and 3B; Table S4). In particular, four modules revealed prominent associations with distinct patient features (Figure 3B). Module D2 inversely correlated with current asthma status (Figures 3C and 3G). GO BP analysis of genes in this module revealed mainly RNA processing and metabolic pathways (Figure 3C), perhaps reflecting an imprint of systemic metabolic changes linked to asthma status (Adamko et al., 2016; Xu et al., 2016). Module E2 was positively correlated with age and weight (not BMI) at the time of the ARTI and study visit (Figures 3D and 3I). This module enriched for genes involved in response to oxidative stress and mRNA regulation, with modest enrichment for cytokine and/or toll-like receptor (TLR) signaling (Figure 3D). The relationship of age to the host response to IFV was further analyzed in an expanded WCGNA with additional parameters, five modules correlated with age (two inversely and three positively) were combined, and shared pathways, genes, and networks were interrogated (Figure S4). These analyses indicated that RNA processing is downregulated (Figure S4) and highlighted altered regulation of splicing, transcription, protein translation, leukocyte differentiation, IFN production, and apoptosis (Figure S4). Module B3 was positively correlated with male sex, and whereas enrichment of biological pathways was less statistically robust than for other modules, chromosomal organization and microtubule-based processes were enriched (Figures 3E and 3I). Module A2 was positively correlated with return to the PCP within one month of the study visit (Figures 3F and 3J), a feature that is likely similar to the well-studied correlation of return to the ED within 72 h with clinical worsening (Alessandrini et al., 2004; Goldman et al., 2011). The most enriched biological pathway in module A2 was RNA processing, with other pathways also involved in RNA metabolism and mitosis (Figure 3C, lower left). To assess potential links to the severity of infection, we combined three modules inversely correlated with the presence of fever, return to ED within one month of study visit, and whether the participant/patient was admitted from the ED visit at which they were enrolled into the study (Figure S5). This analysis revealed shared BP gene sets, including downregulation of RNA processing (Figure S5) and cell cycle, protein translation, and protein transport (Figures S5B and S5C). It was remarkable that this network analysis consistently revealed associations of clinical features with changes related to CD8 T cell metabolism, RNA processing, and gene regulation, although signatures of inflammatory responses were also revealed (module E2). These observations may link CD8 T cell proliferation or activation status to key metabolic processes (Bengsch et al., 2016), as well as asthma status, age, weight, and likelihood to experience more severe or prolonged consequences of IFV infection.

Figure 3. WGCNA of Transcriptional Modules Correlated to Clinical Characteristics in Acute IFV.

(A) Hierarchical clustering dendrogram of the 15 module eigengenes by WGCNA (top panel) and gene hierarchical clustering dendrogram (bottom panel).

(B) Gene expression module to clinical characteristic Pearson correlation coefficients (heatmap with color coding). Modules are shown on the bottom, and the upper table is a color scale for module-trait correlation. For each, GO BP signatures with an absolute correlation >0.5 and a correlation p value <0.05 were included.

(C–J) Four selected gene modules.

(C) D2, anti-correlated with current (2016) asthma status.

(D) E2, correlated with weight (kg) and age (years) and anti-correlated with concern for bacterial infection on the initial study visit.

(E) B3, correlated with male sex.

(F) A2, correlated with whether the patient returned to the PCP within one month.

(G) Network for module D2.

(H) Network for module E3.

(I) Network for module B3.

(J) Network for module A2.

For each, the list of GO BP signatures is listed to the left, and the network is presented to the right (G–J), respectively; smaller nodes have a p value >0.2 and larger nodes have a p value <0.2, color code as below networks). BP, biological process.

See also Table S4 and Figures S4 and S5.

Design and Validation of an Influenza Pediatric Signature

We designed a pediatric influenza gene expression signature to better understand the host circuitry altered in CD8 T cells in acute IFV. To design an influenza pediatric signature (IPS), we started with the IFV-only CHOP ILI cohort patients and three published pediatric IFV datasets (GSE6269, GSE34205, and GSE38900; Table S5) as a training set (Figure 4A). We restricted our gene list to genes common to all four datasets and ordered each gene list by significance comparing HC with IFV patients. Next, we used a leave-one-out strategy and a strict threshold (false discovery rate [FDR] < 0.0001) to generate a final list of the 16 genes present in all five lists. Notably, due to prominent downregulation of genes in CD8 T cells revealed by the analyses above, we included both upregulated and downregulated genes (Figures 4B and 4C). The expression pattern of this set of genes identified clear differences between the groups of HC, IFV-only, HRV-only, and IFV-HRV co-infected patients (Figures 4C and 4D).

Figure 4. Developing and Validating an IPS for Gene Expression Data.

(A) Strategy for designing IPS.

(B–C) IPS genes and their expression in IFV and HC. IPS score calculated as geometric mean of upregulated genes minus geometric mean of downregulated genes.

(B) Fold change of gene expression of each IPS member in all 4 training cohorts.

(C) Heatmap of IPS gene expression in HC and IFV-only patients from CHOP cohort.

(D) IPS score for HC (purple), HRV (green), IFV-HRV (blue), and IFV (red) within our CHOP dataset. IPS score calculated by sum of geometric mean of upregulated genes minus sum of geometric mean of downregulated genes.

(E) Assessment of IPS sensitivity. Left IPS score in a separate pediatric IFV cohort and right IPS score in a published adult IFV cohort (Parnell et al., 2011).

(F) Assessment of the sensitivity of the Influenza Meta-Signature (IMS) score (Andres-Terre et al., 2015). Left is the IMS score in our CHOP ILI cohort and right is the IMS score in a published adult IFV cohort (Parnell et al., 2011).

(G) Assessment of IPS score for HIV progressors versus non-progressors (Quigley et al., 2010) and chronic versus resolver HCV patients (Gupta et al., 2015) (left and right, respectively).

(H) Assessment of IMS score for (left) HIV progressors versus non-progressors and (right) chronic versus resolver HCV patients.

(I) Comparison of IPS with published IFV or bacterial/viral classification gene lists. The CHOP ILI dataset was used as a basis for calculating p values for two pan-age gene lists, three pediatric IFV gene lists, four pediatric ARTI gene lists, seven adult IFV gene lists, and two adult ARTI gene lists (including HC versus IFV, HC versus HRV, and HC versus IFV co-infected with HRV). The p values were calculated using Mann-Whitney; and top, CD8 T cells from IFV versus HC; middle, CD8 T cells from IFV+HRV versus HC. Lower, CD8 T cells from HRV versus HC. See also Figure S3 and Table S5 for dataset details and Table S6 for gene lists.

(J) IPS calculated for published dataset (Zhai et al., 2015) that tracks gene expression during acute ARTI (baseline, day 0, day 2, day 4, day 6, day 21, and the following spring) with IFV A, IFV B, IFV A+HRV, IFV B+HRV, and HRV alone.

See also Figure S6 and Table S5 and S6.

A score based on this signature (an “IPS score”) was significantly elevated in IFV-alone and IFV-HRV co-infected patients in our ILI cohort (Figure 4D). The IPS score also distinguished IFV infection from HCs in the three previously published datasets that were part of the training dataset (data not shown). The sensitivity of the IPS was demonstrated in separate pediatric influenza datasets (Figure 4E, left; Table S5) and adult influenza datasets (Figure 4E, right). Other IFV signatures exist, including a recently published pan-age Influenza Meta-Signature (IMS) (Andres-Terre et al., 2015). As published, this IMS accurately identified adult IFV infection (Figure 4F, right; Table S5) and pediatric IFV infection (Figure 4F, left), but failed to correctly identify the IFV signature in pediatric IFV-HRV co-infection (Figure 4F, left). In addition, the IMS misclassified signatures from HIV progressors versus HIV controllers (Quigley et al., 2010), while the IPS developed here displayed no off-target signal in HIV or HCV infection (Gupta et al., 2015) (Figures 4G and 4H; Table S5).

We next identified 18 previously defined gene lists involving discrimination of IFV (or simply viral/bacterial infections) from HCs. Using the infections in our ILI cohort, the IPS developed here had the most significant p value for correctly distinguishing IFV from HCs among all 19 gene lists tested, including several from other pediatric studies (Figure 4I, top; Tables S5 and S6). Many of the gene lists were sufficiently sensitive to allow detection of an IFV signature in IFV-HRV co-infection (Figure 4I, middle), although several of these 19 gene sets lost sensitivity to detect IFV in the co-infected dataset. Finally, only one signature identified HRV alone (Figure 4I, lower). It is important to note that some of these gene lists were published originally as part of classifiers or complex predictors.

To examine the temporal sensitivity of the IPS, a time course transcriptional dataset was used (Zhai et al., 2015) (Figures 4J and S6). In this dataset, transcriptional data are available from blood prior to infection and at multiple time points within the first week of infection. Unlike many other studies, this study includes downregulated genes, like our IPS, and the overlapping genes include upregulated (IFI27, RSAD2, OAS3, and OASL) and downregulated (EIF3L and EIF4B) genes. When we used the IPS to analyze their data, the scoring peaked during acute viral respiratory infection at 48 h (noted as “day 0,” but patients could present up to 48 h after symptom onset; subsequent times are all relative to this timepoint) in patients infected with IFV-A or IFV-B or co-infected with either IFV or HRV. The IPS score remained significantly above baseline until day 6–8 of illness. Although HRV infection led to a significant increase in IPS above baseline at day 0 (i.e., 0–48 h post-infection), at all time points through day 4–6 of symptoms, the IPS score for any IFV infection group was significantly higher than HRV alone. This analysis correlates with the clinical assumption that pediatric patients present to the ED between day 3 and 5 after acute respiratory infection. Moreover, these results highlight that although there are shared components of the IPS in HRV infection (i.e., HRV infection leads to elevated IPS), the IPS can easily distinguish between IFV and HRV. This dataset also allowed dissection of four different timelines for changes in transcription that included five clear gene modules within the IPS, one with all the downregulated genes (EIF4B, EIF3L, FBL, QARS, EEF2, UXT, NYT5E, and ESYT1) and three with upregulated genes that peak over day 0 (HEX2 and USP18), day 2 (IFI27) and day 4 (FAM64) (Figures S6A–S6D). Across the dataset, networks of altered gene expression could be defined on each day, highlighting the role of the IPS genes in defining distinct temporal phases of transcriptional changes in CD8 T cells during ARTI (Figures S6E–S6G).

The Effect of Age on Transcriptional Circuits in CD8 T Cells in Acute Influenza Infection

Given that the IPS was most effective in the pediatric patients, we focused on the role of age in altered transcriptional circuitry with acute IFV. Children <7 years of age (7 years) with IFV had more significant elevation in their IPS versus HCs when compared with children >7 years (Figure 5A). Circos plots revealed a prominent gene expression downregulation in younger versus older pediatric patients, where transcriptional changes were dominated by upregulation (Figures 5B and 5C). The significant genes from the GO BP gene sets were used to construct networks for both the younger and older pediatric IFV patients (Figure 5D). A heatmap of a pan-ISG gene set (created by merging all GO BP gene sets that contain the term “IFN”; Table S7) demonstrated that ISGs were upregulated in both groups, although more robustly in older pediatric patients (Figure 5E). Of note, neither ISGs nor STATs were upregulated at baseline in older versus younger non-IFV pediatric patients (Figures S7A and S7B), likely ruling out a pre-infection bias in the older pediatric group. In addition, it is unlikely that CMV status played a role. To address this question, we used a published study that analyzed CMV⁺ and CMV⁻ HC individuals (Baitsch, 2011) and assessed the IPS in both groups. Other than IFI27 and PLSCR1, all of the IPS genes were downregulated in CMV-infected HCs (Figure S7C). These data indicate that CMV status is unlikely to account for differences in IPS enrichment in our cohort.

Figure 5. Effect of Age on IPS.

(A) IPS scores in our CHOP IFV dataset divided by age (<7 years versus >7 years) and comparing CD8 T cells from HC and IFV patients.

(B) Circos plots and tables of genes downregulated (upper) and upregulated (lower) between CD8 T cells from IFV patients <7 years and >7 years.

(C) Assessing the contribution of upregulated and downregulated genes to the IPS in each age group. Left, The difference in expression in upregulated genes comparing the four groups (young HC = <7 years, older HC = >7 years, young IFV = <7 years IFV, older IFV = >7 years IFV); middle, the difference in absolute expression of downregulated genes in CD8 T cells; right, the upregulated CD8⁺ T cell gene expression minus the downregulated CD8⁺ T cell gene expression (IPS score calculation).

(D) Networks of predicted upstream regulators in CD8 T cells from IFV patients <7 years of age (upper panel) and >7 years of age (lower panel) using IPA.

(E) Comprehensive ISG list based on the union of BP IFN-related lists was used to generate a heatmap of gene expression for CD8 T cells from HC and IFV patients by age.

See also Figure S7 and Table S7.

Although both age groups demonstrated upregulated target ISGs, younger IFV patients lacked changes in STAT1 and STAT2 expression, as well as proximal IFNAR and IFNGR signaling that were clearly apparent in children >7 years (Figure 6A). One possible reason for the presence of ISGs, but differences in STAT1 and STAT2, is the involvement of other pathways that could induce ISGs, such as interleukin-27 (IL-27) or an underlying difference in bacterial co-infection. However, there was no statistical difference in IL-27-driven signatures in younger versus older patients (Figure S7D). Despite a clinical concern for bacterial infection in some patients (Table 1), this did not correlate with the IPS (Figure S7E), likely ruling out bacterial co-infection as a variable. We next hypothesized that these differences in ISG pathways could reflect a primary influenza infection in the younger patients and a secondary (or memory) response in the older patients. This analysis was complicated by the lack of available human or mouse secondary IFV infection transcriptional datasets. Therefore, we used an adult IFV challenge dataset (GSE61754, Figure S7F) as a proxy for secondary IFV (or memory responses) and an infant IFV dataset (GSE34205, Figure S7G) as a proxy for primary IFV infection and derived predictive signatures based on those datasets. Neither showed a clear difference between younger and older acute IFV patients in our dataset (Figures S7F and S7G). To further interrogate the question of exposure to IFV, IFV nucleoprotein (NP) ELISAs were performed on our CHOP ILI cohort to identify patients who had serological evidence of previous exposure to IFV or vaccine (Figures 6B–6E). The anti-NP antibody titers (from both H1 and H3 viruses) correlated with age (data not shown). The IPS had a trend toward correlation with the NP ELISA (Figure 6B) across ages, although there was no significance in the correlation between the NP ELISA and the ISG signature in that group (Figure 6C). When the patients were split by age (<7 years and >7 years), there was still a trend that the IPS correlated with the NP ELISA in the older patients (Figure 6D). Serological evidence of previous exposure in the older patients showed only weak association with STAT2 (and no correlation with STAT1). Similarly, in the group of younger combined IFV and IFV-HRV patients, relationships between serological memory and STAT1 or STAT2 were not apparent (Figure 6E). However, when patients with co-infections were removed, a significant correlation between previous exposure to IFV and higher STAT2 activity became evident (Figure 6E, lower left). At baseline in HCs, STAT2 was equivalent between age groups and STAT1 was actually lower in older children, compared with what was seen in the IFV CHOP participants (Figure S7B). These observations suggest that the differences in ISGs and STAT circuits in young versus older pediatric patients with IFV may reflect elaboration of different early inflammatory pathways based on immune memory versus primary immune responses.

Figure 6. Mechanisms Underlying Immune Response of Younger versus Older Childhood to Acute IFV Infection.

(A) IPA was used to plot the IFN stimulated gene (ISG) pathway coded by gene upregulation (red nodes) and downregulation (green nodes) in IFV patients <7 years (upper panel) of age and >7 years of age (lower panel) when compared with healthy controls.

(B–E) NP ELISAs for H1 and H3 were performed.

(B) CHOP IFV patients, all ages, with correlation between IPS score and NP ELISA (includes both H1 and H3).

(C) CHOP IFV patients, all ages, with correlation between ISG score and NP ELISA (includes both H1 and H3).

(D) Left column, IPS correlated with NP (includes both H1 and H3); and right column, ISG correlated with NP (includes both H1 and H3). Top row, patients >7 years and with IFV. Middle row, patients <7 years and with IFV or IFV-HRV. Bottom row, patients <7 years with IFV only.

(E) Left column, STAT2 expression level and right column, STAT1 expression level. In all subplots, correlation of STATs with NP (includes both H1 and H3). Top row, patients >7 years and with IFV. Middle row, patients <7 years and with IFV or IFV-HRV. Bottom row, patients <7 years with IFV only.

See also Figure S7.

DISCUSSION

Using focused transcriptomics of the CD8 T cells in IFV-infected pediatric patients, we identified predictive signatures in IFV and leveraged these signatures to assess the effect of host clinical condition on anti-viral molecular pathways. These signatures yielded biological insights into the underlying molecular circuits of anti-viral responses to ARTI, including an age-related difference in signaling to ISGs. In addition, we were able to use extensive clinical data to connect gene circuits altered in ARTI to the clinical characteristics of pediatric patients.

IPS as an Analytic Tool

Whole-blood or whole-PBMC acute influenza and acute viral gene signatures and gene lists have been previously reported (Andres-Terre et al., 2015; Ioannidis et al., 2012; Mejias et al., 2013; Ramilo et al., 2007; Zhai et al., 2015). These studies and others have provided considerable insights into the patterns of gene expression on human viral infection, but using these classifiers and lists to generate fundamental biological understanding has remained challenging. The generation of novel datasets is of value, but additional insights have been gained by integrating new data with published datasets, and meta-analyses drawn exclusively from published data have generated novel mechanistic conclusions (Andres-Terre et al., 2015; Sparks et al., 2016).

In contrast to whole-blood or whole-PBMC studies, there has been a dearth of cell-type-focused human transcriptional profiling. Moreover, pediatric ARTIs have important age-dependent differences in pathogenesis and clinical events, but the mechanisms of age-based differences have remained unclear. Here, focusing specifically on CD8 T cells highlighted the importance of downregulated genes in addition to genes induced by infection. Including downregulated genes in our IPS improved specificity and allowed better resolution to distinguish between different infections. Downregulated genes also provided biological insights in the longitudinal analysis, where transcriptional responses to IFV changed over time, perhaps reflecting waves of bystander versus antigen-specific T cell activation.

Different Strategies for STATs and ISGs

Using our systems immunology approach, we noted that although the IPS predicted influenza accurately in all age groups, the IPS for the youngest participants (<7 years) had an expression signature of predominantly downregulated gene expression in contrast to the older children (>7 years), who displayed predominantly transcriptional upregulation. Moreover, ISG expression was distinct in the younger pediatric participants. Specifically, the upregulated ISGs in acute IFV, well appreciated in mouse (Brinza et al., 2016) and human data (Mejias et al., 2013), characterized the dominant component of the older patients, as expected. However, in younger participants, there was only moderate ISG upregulation with less involvement of STAT1 and STAT2. While non-STAT1/2 pathways, including IL-27 (Yoshida and Hunter, 2015) and perhaps even bacterial recognition by TLR or other innate sensing pathways (Baruch et al., 2014), can induce ISGs in some settings, we found no evidence for a role of IL-27 or concomitant bacterial infection in the age-based differences in ISG gene circuits. One obvious potential difference between younger and older pediatric IFV patients is previous exposure to IFV or vaccine. We therefore considered the possibility that immune memory to IFV could underlie the differences in ISG and STAT pathways. Using serological evidence of previous influenza exposure to define patients with immune memory to influenza, the IPS score tended to correlate with the NP antibody signal across all ages, perhaps reflecting a link between the level of immune memory and some features in the IPS. However, previous IFV exposure did not resolve IPS, ISG, or STAT patterns in children >7 years. This was expected since, epidemiologically, essentially 100% of humans have been exposed to IFV through infection or vaccination by this age (Bodewes et al., 2011), greatly reducing the resolving power of this analysis. However, in children <7 years, where we expected some previous IFV exposure but also many IFV-naive patients, immune memory to IFV was linked to transcriptional circuits. STAT2 (although not STAT1) correlated with anti-NP antibodies in younger IFV-only patients. It was notable that HRV co-infection appeared to distort this connection between IFV memory and STAT2, perhaps reflecting an HRV link to STATs and ISGs distinct from IFV. The connection to STAT2, but not STAT1, is also intriguing. Our cohort was relatively small (and cannot be expanded due to the unique nature of the ARTI viruses circulating in 2009). It is possible that with greater power, a connection to STAT1 would emerge. Nevertheless, these observations might reveal a distinction between IFN- α/β (STAT1/2 heterodimers) versus IFN- γ (STAT1 homodimers) in the response to IFV in naive versus immune pediatric ILI patients. Future studies will be necessary to confirm and extend these observations, but these insights could be of interest for determining the role of memory (although not protective) immune responses in human influenza infection.

Differences in Host Response with IFV and Other Pathogens

The pediatric cohorts studied here provided an opportunity to dissect the relative contributions of distinct co-infections to host transcriptional circuitry. Of note, in our institution in the fall of 2009, an unexpectedly large cluster of patients tested positive for HRV, and samples from a subset of patients were sent to the CDC (CDC, 2011). In our dataset, there was an HRV-only transcriptional signature. This HRV signature included genes involved in negative immune regulation (DUSP22, SMAD5, and PTPN6), chemotaxis (CCL4), NFkB (NFKBIA), and CD8 genes themselves (CD8A and CD8B), but was masked during IFV-HRV co-infection, and co-infection in the <7 years patients altered the STAT2 connection to serological memory as discussed above. It is interesting that there was a clear set of ISGs induced in IFV-HRV co-infection (e.g., IFIT1, IFIT3, IFIT5, IFI44, and IFI44L), but also a set of ISGs induced in IFV infection (MX1, IFI6, IFIT2, and OAS1). Among the ISGs, IFI27 had a robust ability to detect IFV infection as well as IFV-HRV co-infection, but not HRV alone. IFI27 as a single gene biomarker can distinguish IFV from bacterial infection (Tang et al., 2017). Mechanistically, IFI27 may be of considerable interest since this gene product is a mitochondrial protein thought to be involved in apoptosis (Gytz et al., 2017; Rosebeck and Leaman, 2008). Thus, IFI27 might mediate some of the type I IFN-induced bystander attrition affects observed in previous studies (Bahl et al., 2006; McNally et al., 2001). These data further highlight the relevance of defining not only the clinically relevant viruses present, but also co-infecting viruses in ARTI and their impact on transcriptional programs.

Importance of Downregulated Genes

One key aspect of these studies was the prominence of downregulated genes. ISGs and several other gene groups were, as expected, upregulated in acute IFV. However, transcriptionally downregulated gene circuits included co-stimulation (CD226), adhesion, and migration (ITGA4, ITGAL, ITGB2, and ITGAM), survival (BCL2), and several signaling genes (MAP3K12, PLCB2, PLCG1, FOS, JUNB, etc.) were prominent in pediatric acute IFV, IFV-HRV, and HRV. The specificity and sensitivity of our IPS was due in part to downregulated genes, and this feature helped further resolve the distinct biology of younger (<7 years) versus older (>7 years) pediatric patients. The underlying mechanisms for this transcriptional downregulation may include transcriptional regulation by type I IFN signaling (Su et al., 2015). Alternatively, because blood was sampled, transcriptional downregulation could reflect the depletion of sub-populations of CD8 T cells from the blood due to homing and/or retention in the inflamed respiratory tract. Thus, in the future, more focused transcriptional profiling examining phenotypically defined CD8 T cell subsets or interrogating cells in the human respiratory tract may further define the nature of this transcriptional downregulation in the blood.

A major advantage of the current dataset was extensive clinical metadata, including the opportunity for long-term followup in some patients. Network analysis using WGCNA revealed several key connections between clinical features and transcriptional signatures. For example, transcriptional modules were linked to sex, age, short-term return to health care, current asthma status, and severity. Short-term return to health care is a proxy for more severe illness or complications arising from ARTI. The module associated with these clinical features was strongly enriched for RNA processing biological activity, an observation that may reflect inflammation-induced alternative splicing or a role for large-scale transcriptional reorganization. It is notable, however, that this module also includes several key genes involved in the regulation of signaling and/or co-stimulation, including STAT4, SH2D1A (encoding SAP), and CD28, possibly reflecting an altered type of immune activation. Current asthma status is an example of the strength of long-term clinical follow-up. At the time of the initial ED visit, there was no gene module correlated with asthma status, but the development of asthma subsequently was strongly anti-correlated with a transcriptional module enriched in RNA metabolic processes at the time of ED visit. Whether this enrichment reflects an underlying metabolic signature associated with predisposition toward (or against) asthma is unclear. However, age and weight at the time of ED visit were strongly associated with a related set of metabolic signatures in a different module. Thus, CD8 T cell-focused transcriptional profiling linked to clinical data features may reveal an underlying connection between early-life metabolic alterations and disease.

Conclusions

These studies have revealed several fundamental features of pediatric ARTI. First, the use of CD8 T cells (rather than whole blood or PBMCs) revealed several key features of pediatric IFV responses. Second, studying pediatric ARTI patients allows features of disease, especially IFV, to be revealed that might not otherwise be apparent in adults due to previous IFV exposure or other comorbidities. Third, detailed clinical information provided linked transcriptional programs to complex clinical phenotypes. Finally, the inclusion of downregulated genes substantially improved the performance of our IPS compared with previous studies. Through these approaches, we were able to define several sets of mechanistic pathways changed during pediatric IFV infection, reveal the transcriptional imprint of co-infection, and identify a potential role for immune memory in skewing the pathways used to induce ISGs and other transcriptional events in young ARTI patients. Future approaches may be able to build on this foundation to rapidly identify primary IFV infection, patients at risk for more several clinical outcomes, or perhaps even children who become predisposed to asthma.

EXPERIMENTAL PROCEDURES

Recruitment

The CHOP institutional review board approved the studies, and written informed consent was obtained from all subjects.

DNA Microarrays

CD8 T cells were sorted from PBMCs and stored in Trizol. The NuGen WT-Ovation Pico + Exon Module was used for amplification. Affymetrix Human Gene 1.0ST microarray analysis was performed at the University of Pennsylvania Microarray Core Facility.

NP ELISAs

ELISAs were performed using recombinant NP from H3 and H1 viruses.

Highlights.

• Patients with influenza have distinct transcriptional changes in blood CD8 T cells

• Influenza pediatric signature developed that outperforms other prediction gene lists

• Gene expression modules correlate with clinical characteristics

• Age-dependent alterations in gene circuits underlying host response to influenza