Sample size calculations and statistical analyses

Experimental design and study procedures

This was a prospective cohort study involving a convenience sample of previously healthy children <2 years of age with acute RSV infection, and a cohort of healthy, asymptomatic age-matched controls. The study was conducted at Nationwide Children’s Hospital (NCH; Columbus, OH) from 2011 to 2015.  Children were enrolled at NCH urgent care clinics, the emergency department (ED), or in the inpatient hospital units (ward or PICU) within 24 (17-38) hours of hospitalization, and blood and nasopharyngeal samples obtained. RSV was diagnosed per standard of care using a rapid antigen assay or PCR test. Healthy controls were enrolled during well-child visits or minor elective surgical procedures not involving the respiratory tract (13). Exclusion criteria included: documented bacterial co-infections, premature birth(≤36 weeks of gestation), congenital or chronic medical conditions, and immunodeficiency. For healthy controls, additional exclusion criteria included presence of fever or symptoms of respiratory tract infection within two weeks of enrollment. 

Blood samples were analyzed for: 1) transcriptome; 2) cell immunophenotype; and 3) white blood cell count (WBC) with differential. Samples were drawn in this order but due to the small blood volumes obtained in some children, we were unable to obtain all samples in all participants. Nasopharyngeal (NP) swabs were collected in RSV inpatients and outpatients at enrollmentby trained study personnel using an approach that has been validated in multicenter trials(33, 34). To confirm and quantify RSV loads, qRT-PCR analyses targeting the N gene were performed in NP samples of all children included in the study (14). Other respiratory viruses were detected using PCR assays (either Luminex xTag Respiratory Panel, Austin, TX; or the Filmarray respiratory viral panel, BioFire, BioMerieux, Durham, NC) (23).

Demographic and clinical parameters were collected using a standardized questionnaire and by reviewing electronic healthcare records. Disease severity was assessed primarily by the need for hospitalization (inpatients [=severe disease] vs. outpatients [=mild disease]) and using a standardized clinical disease severity score (CDSS) (14, 23). In hospitalized patients, additional parameters of severity were collected including administration of supplemental oxygen, PICU admission, and total duration of hospitalization. In addition, families were contacted 2 and 4 weeks after enrollment to confirm the lack of subsequent readmissions in the inpatient group or hospitalization in the outpatient cohort. 

Study objectives

We sought to define the clinical, viral loads and immune profile differences between children with mild RSV (outpatients) vs. severe RSV infection (hospitalized) across different ages. We hypothesized that RSV viral loads and host immune responses would differ according to disease severity and age at the time of the infection. Specifically, we hypothesized that children with mild infection would mount a more effective innate immune response that will be associated with protection from developing severe disease.

Transcriptome analyses

Sample collection, processing and RNA hybridization

Blood was collected in Tempus tubes (Applied Biosystems, Foster City, CA) and stored within 2-4 hours at -20°C until further analyses. RNA was extracted, processed and hybridized into Illumina Human HT-12 v4 microarray chips (47,323 probes; Illumina, San Diego, CA) as described (13). After hybridization data was scanned on Illumina Beadstation 500. Illumina GenomeStudio software was used for background subtraction and to scale average signal intensities.

Data pre-processing

We first selected transcripts that were ‘present’ (signal precision <0.01) in ≥10% of samples (PAL10%; 18,213 transcripts) as described (13, 35). Next, raw expression values <10 were set to 10 and the data log₂-transformed. Using PCA and principal variance component analysis (PVCA) in R (36)we identified a technical batch effect within the discovery cohort that was associated with globin reduction (Fig S4A). The batch effect was corrected using the ComBat function of the SVA package in R (37, 38)(Fig S4B).  

Differential gene expression analysis (transcriptional signatures)

Two datasets were used for transcriptome analyses. We first derived the transcriptional signature for mild (outpatients) and severe (inpatients) RSV infection by comparing each RSV cohort to the same healthy controls in the discovery cohort. Patients for this cohort were matched for age and gender. We used the limma package in R (39)and applied stringent statistical filtering (FDR p<0.01, Benjamini-Hochberg multiple test correction and 1.25-fold change). The transcriptional signatures identified were validated with PCA in a not age-matched validation cohortthat included RSV outpatients, inpatients and healthy controls (Figure 2B&D). The two cohorts were independent except for four healthy controls whose samples were hybridized twice and used in the discovery and validation cohorts (Fig 1). The samples included in the study were processed in two different batches (Fig S5) which largely overlapped in terms of enrollment (batch#1: 2012-2014, and batch#2: 2012-2015). To avoid any potential batch effect, samples included in the discovery cohorts were hybridized in batch #1, while those used for validation were hybridized in batch #2.

Functional characterization of differentially expressed transcripts

To define the biological function of the “mild” and “severe” RSV signatures, we applied modular transcriptional analyses. Briefly, this is a systems scale that aims to reduce the abundance of transcriptional data into functional pathways or modules. Transcriptional modules are formed by genes coordinately expressed, thus allowing functional interpretation of the microarray data into biologically useful information. Modular over- and under-expression was defined by the percentage of transcripts within each module that were differentially expressed in RSV outpatients and inpatients vs. healthy controls. A detailed description of this mining analysis strategy has been reported elsewhere (19).

Modular maps for the outpatient and inpatient discovery cohorts were derived by comparison with the same age-matched healthy controls (FDR p<0.05, Benjamini-Hochberg multiple test correction). Results were confirmed by a) computing correlations of modular expression between the discovery and validation cohorts (Spearman’s correlation coefficient (Fig. 3B); b) by DAVID bioinformatics tool (version 6.7, available at https://david-d.ncifcrf.gov) (40), and c) Ingenuity Pathway Analysis tool (IPA; QIAGEN, Redwood City, CA, USA; Table S7). 

Age analyses using modular expression were performed in children <6 months vs. 6-24 months of age according to disease severity (outpatients vs. inpatients) using chi-square test and adjusted for multiple comparisons by Benjamini-Hochberg multiple test (Table S5).

Blood immune cell populations 

Blood samples (1-2 ml) were obtained in ACD tubes (BD vacutainer ACD Solution B; BD, Franklin Lakes, NJ) and processed within 2-4 hours of collection. Five blood aliquots of ~200 μL each, were stained with different antibody panels for characterization of innate (neutrophils, monocytes--including HLA-DR low monocytes as a functional marker of activation--, NK cells, DCs), and adaptive immune cell populations (T cells, peripheral T follicular-like helper cells (Tfh), and B cells; Tables S11&12). Samples were incubated with antibody panels for 15 minutes, and red blood cells lysed with BD FACS lysing solution (BD Biosciences, San Jose, CA). Stained cells were washed twice with phosphate-buffered saline (PBS) before storage at 4°C until cell acquisition, which was performed in 1-3 days on a LSRII flow cytometry instrument (BD Biosciences, San Jose, CA). Data were analyzed using FlowJo software v9.8.2 (Tree Star, Ashland, OR). Data are presented as absolute numbers (Table 2; Fig 5) in children with total WBC data available (89%; 93/104 children), and percentages (Table S10). Individial patient data is provided in Table S1. We used immune cell counts for all analyses except to compute correlations between immune cell populations and modular gene expression, in which we used percentages because of the lack of WBC count data in 5 of 37 patients with paired data. 

Sample size calculations and Statistical analyses

For sample size calculation, best practices in the transcriptome field dictate utilization of at least two independent sets of samples for the purpose of validating candidate signatures (or profiles). In previous studies in individuals with acute infections, others and we have obtained robust profiles using groups of 15-20 subjects per group (13, 35, 38, 41, 42). We analyzed demographic and clinical data using SPSS v22.0 (IBM Corp, Armonk, NY) and Graphpad Prism v7.0b (La Jolla, CA) software packages, that was presented as medians with 25%-75% interquartile ranges (IQRs). We compared continuous variables using either Mann-Whitney or Kruskal-Wallis test with Dunn’s or Benjamini-Hochberg post hoc tests for multiple testing, and categorical variables using either chi-square or Fisher’s exact test. 

For transcriptome and cellular immunophenotype data we used R (R Foundation for Statistical Computing, Vienna, Austria). The statistical tools used for microarray analyses are included in each of the above sections. For correlations between transcriptomic and cellular, clinical or virology data we usedSpearman’s correlation coefficient. 

Last, we analyzed whether the main immune variables found to be relevant in the study (IFN expression & HLA-DR low monocyte numbers) were associated with hospitalization. To this end we conducted multivariable analyses using logistic regression with Firth’s correction for small sample size, after accounting for other factors.  Due to the limited sample size, additional covariates were added in separate models, so that a maximum of two risk factors were included within any given model. Analyses were conducted using SAS 9.4 (SAS Institute, Cary, NC) with two-sided p-values <0.05 considered statistically significant.

Category