We developed an analytic approach to identify genes whose expression in CBMCs are associated with newborns with higher or lower risk for asthma. We conducted a meta-analysis to increase power and generalizability (Figure 1A,B). In our approach, we queried NCBI’s Gene Expression Omnibus for CBMC microarray datasets that included metadata regarding the demographics and birth characteristics that are risk factors for pediatric asthma. We identified 354 datasets from our original search, of which 17 datasets contained relevant metadata. Of the 17 studies, the most common maternal and neonatal characteristics reported were newborn sex, gestational age at birth, birthweight, and maternal pre-pregnancy BMI (PP BMI) at 69.96%, 51.59%, 30.27%, and 19.32% of samples, respectively. Metadata regarding maternal smoking and mode of delivery (i.e., vaginal vs. cesarean section) was reported for three datasets; however, two of these datasets originated from the same laboratory group and contributed the majority of samples reporting these factors. None of the identified datasets reported metadata about maternal atopy or child’s ethnicity. Therefore, we chose to focus the analysis on gene expression associated with newborn sex, gestational age at birth, newborn weight, and maternal PP BMI. Six datasets were excluded due to homogenous metadata (e.g., all female). Datasets and the corresponding analysis are reported; a total of 605 unique transcriptomes were included (Table 1; Bukowski et al., 2017; Edlow et al., 2016; Kallionpää et al., 2014; Mason et al., 2010; Rager et al., 2014; Smith et al., 2014; Stünkel et al., 2012; Turan et al., 2012; Votavova et al., 2011; Votavova et al., 2012; Winckelmans et al., 2017).

To validate the findings of the meta-analysis, we assessed three independent cohorts to confirm which genes are associated with asthma risk (Figure 1C). Further details regarding validation and outcomes are discussed in the Results section. In brief, the goal of validation was to assess gene expression in an independent cohort (UIH Cohort [Koenig et al., 2020]) where all subjects had complete metadata regarding newborn sex, gestational age, birthweight, and PP BMI. We sought to further understand whether transcriptomic differences in CBMCs corresponded to differences at the protein level (all three cohorts) or cell population level (Olin et al., 2018). Finally, we assessed two identified proteins in another independent cohort (Project Viva Cohort [Oken et al., 2015]) to test for association with pediatric asthma and pulmonary function outcomes at two follow-up time points.

Univariate random-effects models were generated to assess transcriptional changes in CBMCs with regards to newborn sex, gestational age, birthweight, and maternal PP BMI (Figure 2A). Differential expression (false discovery rate [FDR] < 1%) was observed in 122, 34, 4, and 12 genes when comparing fetal sex, gestational age, birthweight, and maternal pre-pregnancy BMI, respectively (Supplementary file 1–4). When evaluating sex and gestational age, several expected genes were identified to have large transcriptional changes. With regards to sex-associated transcriptional changes, although there was no X or Y chromosome-wide gene enrichment, several genes located on X (KDM5C, SMC1A, TXLNG, and KDM6A) and Y (KDM5D and EIF1AY) chromosomes exhibited the largest effect sizes and most significant differences. In addition to expected sex-associated transcriptional changes, HBE1, a hemoglobin subunit associated with fetal erythropoiesis, was significantly associated with preterm gestational ages. To further validate our gestational age findings, we compared our univariate analysis with previously published results of differentially methylated regions associated with gestational age at birth estimated by last menstrual period (Bohlin et al., 2016; Figure 2—figure supplement 1). A significant association was observed between differentially methylated genes and the effect size in our gestational age analysis, such that genes whose methylation increased with gestational age as reported by Bohlin et al., 2016 showed on average decreased expression with gestational age in our meta-analysis.

To identify the biological processes that are enriched by genes with transcriptional changes associated with higher or lower risk of asthma, the z-scores from each univariate meta-analysis were averaged, such that negative z-statistics were associated with lower risk (female, older gestational ages, higher birthweights, and lower maternal PP BMI) and positive z-statistics were associated with increased risk. Thus, the pooled z-score indicates the average probability that a gene’s expression is associated with either increased or decreased risk of asthma based on an individual’s demographics and birth characteristics. The averaged z-statistic was used as a pre-ranked list for gene set enrichment analysis. GO terms were assessed for enrichment; 18 and 19 GO terms were significantly enriched (FDR < 1%) with regards to low- and high-risk profiles, respectively (Figure 2B). Genes associated with lower risk exhibited increased representation in GO terms involving innate immune signaling and defense, whereas high-risk genes were enriched in pathways involving translation and RNA metabolic processes.

Gene expression studies from pooled cellular populations (e.g. CBMCs, peripheral blood mononuclear cells, and tissues) can be influenced by the cellular composition. To determine whether specific cell population enrichment was associated with the pooled z-score, the Human Protein Atlas (HPA; Uhlen et al., 2010, Monaco et al., 2019) was utilized to assess the abundance of transcripts in peripheral blood leukocyte RNA transcriptomes in relationship to the pooled z-score. We observed a generalized increase in expression of low-risk genes in myeloid cells and high-risk genes in lymphoid cells. This pattern of expression was most pronounced in neutrophils (Figure 2C). These results suggest that lower risk individuals have increased populations of myeloid cells in their CBMCs.

A potential confounder in pooling results is the potential over-representation of any one analysis. To assess bias in the pooled z-score, two analyses were performed (Figure 2D). In the first assessment, z-scores from each individual analysis correlated with the pooled z-score. The highest correlations with the pooled z-score were the z-scores from the meta-analyses assessing gestational age at birth, newborn sex, and birthweight. Individual dataset z-scores for each dataset demonstrated a similar trend. In the second assessment, z-scores from each individual analysis were correlated with the combination of all other z-scores. Again, the pooled z-score had the highest average correlation followed by the meta-analyses. Together, this demonstrates that the pooled z-score does, indeed, amalgamate information across all of the analyses, with the most influence arising from gestational age at birth, newborn sex, and birthweight.

To confirm gene expression changes associated with asthma risk stratification, pooled z-scores from the meta-analysis were compared with the UIH cohort, a cohort of individuals in which newborn sex, maternal PP BMI, gestational age at birth, and birthweight were known (Supplementary file 5). UIH cohort z-scores were calculated from mRNAseq of CBMCs, where gene expression was modeled as a function of number of risk factors (Supplementary file 6). We developed a method to validate the congruence between the UIH cohort and the pooled meta-analysis, which we termed the replication score (RS). This RS is the product of the pooled z-score from the meta-analysis and the UIH z-score (see Materials and methods). We assessed the relationship between RS cutoff, p-values, and number of genes (Figure 3—figure supplement 1). Genes with a RS greater than three were identified as being sufficiently congruent. Fifty-one genes, 0.4% of all genes, tested had a RS greater than 3 (Figure 3A). These identified genes corresponded well with the results of pooled z-score for gestational age at birth, newborn sex and birthweight, but showed limited correlation to maternal PP BMI. They had median p-values of 0.02, 0.01, 0.11, and 0.58 in the gestational age at birth, newborn sex, birthweight, and maternal PP BMI meta-analyses, and median p-value of 0.02 in the UIH cohort.

Replicating genes were enriched for processes involved in vesicle biology (Figure 3—figure supplement 2). Specifically, replicating genes associated with low risk were enriched for genes that are components of granulocyte specific granules (MS4A3, CEACAM8, OLR1, CAMP, LTF, CHI3L1, SLPI, PGLYRP1). With regards to genes associated with higher risk, genes involved in vesicle sorting/production (VPS28, VTI1B, FIS1) as well as several genes involved in vesicle membrane biology (AGPAT3, ELOVL6, TM7SF2) were identified.

To test whether replicating genes were associated with the number of risk factors, these 51 genes were assessed using principal component analysis. Using the first Eigenvector (explaining 41.3% of variance in replicating genes), a significant association (R [95% confidence interval (CI)=−0.51 [–0.73, –0.18], p-value<0.01) was observed between UIH cohort Eigenvalues and number of epidemiologic risk factors (Figure 3B). Positive eigenvalues represent increased expression of low-risk genes and negative eigenvalues represent increased expression of high-risk genes. Although the associations with individual risk factors were in the expected directions, they were not significant (Figure 3C), suggesting that the additive effect is greater than the individual.

To further determine whether these changes are due to differences in cellular populations, we analyzed mass cytometry data published by Olin et al. for abundance of 21 different cell types in relationship to number of epidemiologic risk factors for pediatric asthma (Votavova et al., 2011). Cord blood neutrophil abundance was inversely associated (R [95% CI]=−0.57 [–0.73, –0.34], Bonferroni p-adj<0.001) with the number of risk factors (Figure 4A). Other myeloid cell types, CD14+ monocytes, and myeloid-derived dendritic cells also had negative correlations but were weaker and not significant after multiple testing correction (data not shown).

Extending these findings from gene expression to protein abundance, reported umbilical cord plasma-protein abundance data was correlated with number of risk factors in a secondary analysis (Olin et al., 2018). Serum proteins from genes that replicated with low risk had on average negative correlations with number of risk factors (Figure 4B). No proteins from the high-risk genes were tested in plasma, due to their intracellular localization. Notably, proteins (CEACAM8, PGLYRP-1, CHIT1, sIL6Rα, MMP-9, and OSM) predicted to be enriched in neutrophils by the HPA (Uhlen et al., 2010) had strong correlations with both neutrophil abundance and number of risk factors (Figure 4B,C).

We hypothesized that the serum concentration of proteins identified as low risk in our transcriptomic analysis would correlate with mRNA abundance in CBMCs, whereas those not associated with risk would not correlate with mRNA in CBMCs. To test this hypothesis, we used the UIH cohort to correlate mRNA abundance with serum protein concentration of PGLYRP-1 (low risk) and sIL6Rα (no risk). We observed a significant (R [95% CI]=0.39 [0.03, 0.66], p<0.05) association between PGLYRP-1 protein concentration and mRNA (Figure 4D). Consistent with transcriptomic results, PGLYRP-1 cord blood serum concentration was inversely associated with number of risk factors (R [95% CI]=−0.51 [–0.74, –0.17], p<0.01). sIL6Rα was neither associated with its mRNA in CBMCs (R [95% CI]=0.37 [–0.22, 0.77], p=0.21) nor associated with the number of risk factors (R [95% CI]=0.20 [–0.39, 0.67], p=0.50).

We tested the association between PGLYRP-1 and sIL6Rα, individual risk factors, and demographics in the UIH and Project Viva cohorts. Cord blood serum was available in a subset of individuals (n = 358) from Project Viva (Supplementary file 7). There was no significant difference (p>0.05, Wilcoxon rank sum test) in PGLYRP-1 or sIL6Rα between UIH and Project Viva cohorts (Figure 4—figure supplement 1). Consistent with our previous observations, PGLYRP-1 and sIL6Rα were positively correlated in both UIH (R [95% CI]=0.21 [−0.16, 0.54]) and Project Viva (R [95% CI]=0.19 [0.09, 0.29]) cohorts. Similar to the observation in the UIH cohort, there was a negative association between PGLYRP-1 and number of risk factors in Project Viva (β [95% CI]=−0.22 [−0.36, –0.07], p-value=0.003) (Table 2). This association was driven by the relationship between PGLYRP-1, gestational age, and sex. There was no association with PP BMI or birthweight when taking into account gestational age and sex. Furthermore, there was no association of sIL6Rα with the number or risk factors or any individual risk factor. Interestingly, there was an inverse relationship observed in both the UIH and Project Viva cohorts between sIL6Rα and self-reported maternal race as Black/African-American. Collectively, these results suggest that increased abundance of mRNA from genes localizing to neutrophil-specific granules are associated with the number of risk factors for pediatric asthma. These changes in mRNA are reflected in the abundance of these specific granule proteins in serum and plasma.

In context of our previous results, we hypothesized that specific granule protein abundance in serum is associated with risk of pediatric asthma and this process is independent of neutrophil abundance. To evaluate this hypothesis, we measured PGLYRP-1 (present in neutrophil-specific granule and correlates with its mRNA in CBMCs) and sIL6Rα (derived from neutrophils but not present in specific granules and does not correlate with its mRNA in CBMCs) in umbilical cord blood serum. At two follow-up time points, asthma outcomes and expiratory flow volumes were modeled as a function of PGLYRP-1 and sIL6Rα in a subset of individuals in Project Viva (Figure 5—figure supplement 1). The demographics of the subset of individuals from Project Viva from which umbilical cord serum was available had a similar demographic profile as the full cohort. One notable difference was a sizeable decreased response rate for asthma outcomes at the early-teenage follow-up compared to the full cohort (32% subset vs. 47% full cohort).

PGLYRP-1 and sIL6Rα were modeled as predictors for current asthma at mid-childhood (median age ~7.7 years old) and early-teen (median age ~12.3 years old) follow ups (Table 3). Four regression models were used to estimate the association between asthma outcomes: univariate, adjustment for child’s birth characteristics and demographics, adjustment for mother’s demographics, and adjustment for birth characteristics and all demographics (reported in manuscript). The abundance of PGLYRP-1 was significantly associated with current asthma at mid-childhood (adjusted odds ratio [OR] [95% CI]: 0.50 [0.31, 0.77] per 1 SD increase, p-value=0.003) (Figure 5A). There were no significant associations between current asthma and PGLYRP-1 at the early-teen follow up; however, the CI at this time point was much wider, likely secondary to the smaller sample size. There were no significant associations between sIL6Rα with any asthma outcome at either time point.

We also performed analyses to estimate the relationship of cord blood PGLYRP-1 and sIL6Rα with FEV₁/FVCx100 ratio and bronchodilator response (BDR) at mid-childhood and early-teen follow-up time points (Table 4). There was no significant association between sIL6Rα and FEV₁/FVCx100 ratio at either time point. PGLYRP-1 and sIL6Rα were not associated with BDR at either time point. However, there was trend toward an association between PGLYRP-1 concentration and FEV₁/FVCx100 ratio at mid-childhood (adjusted β [95% CI]: 1.18 [–0.18, 2.56] per 1 SD increase, p-value=0.09) and significant association at the early-teen follow up (adjusted β [95% CI]: 1.15 [0.20, 2.10] per 1 SD increase, p-value=0.02) (Figure 5B).

To further our understanding of the relationship between PGLYRP-1 and outcomes, we performed two secondary analyses of models adjusted for all demographics and birth characteristics. First, we assessed the total variance explained by the regression model adjusting for both birth characteristics and demographics, and the variance explained by each of the individual predictors in the model. Assessing current asthma risk at mid-childhood, the regression model explained approximately 18% of the variance, and PGLYRP-1 was the most important predictor. Assessing current FEV₁/FVCx100 at early-teen time point, the model explained approximately 26% of variance, and PGLYRP-1 was the second most important predictor (Figure 5—figure supplement 2). Second, to identify covariates that modify the effect of PGLYRP-1, we performed subset analyses (Figure 5—figure supplement 3). Small for gestational age and children identified by their mothers as ‘Other’ race/ethnicity displayed significantly different associations between PGLYRP-1 and mid-childhood asthma. Small for gestational age and children with obese mothers displayed significantly different associations between PGLYRP-1 and FEV₁/FVC.

The current study was approved by the University of Illinois at Chicago IRB (#2016–0326) and the IRB of Harvard Pilgrim Health Care. Volunteers were recruited from women attending their first prenatal visit at one of eight practices of Atrius Harvard Vanguard Medical Associates. The exclusion criteria were multiple gestation, inability to answer questions in English, gestational age ≥ 22 weeks at recruitment, and plans to move away before delivery. The cohort profile was previously described by Oken and colleagues (Oken et al., 2015).

Current mid-childhood asthma was defined as a ‘yes’ response to ‘Has a health professional ever told you that your child has asthma?’ and ‘yes’ to either ‘In the past 12 months, has child taken or been prescribed Albuterol, Cromolyn, Nedocromil, Montelukast, inhaled corticosteroids, or Prednisone’ or ‘In the past 12 months, has your child ever had wheezing (or whistling in the chest)?’. We used as a comparison group those with no asthma diagnosis and no asthma medication use or wheezing in the past 12 months. We used the same definition for current asthma in early adolescence, except the time reference for asthma medication was ‘in the past month’. Ever asthma was defined as a ‘yes’ response to ‘Has a health professional ever told you that your child has asthma?’ within either the mid-childhood or early-teen follow-up periods. We used as a comparison group those with no asthma diagnosis. Individuals with missing data were not used in regression models assessing current or ever asthma outcomes. Individuals with missing data are reported in demographics tables and displayed in figure as ‘Missing Data’.

Methods for obtaining spirometric measurements and BDR have been described previously (Tse et al., 2013). In brief, spirometry was performed with the EasyOne Spirometer (NDD Medical Technologies, Andover, MA). Post-bronchodilator spirometric measures were obtained at least 15 min after administration of two puffs (90 μg per puff) of albuterol. Spirometric performance was required to meet American Thoracic Society criteria for acceptability and reproducibility, with each subject producing at least three acceptable spirograms, two of which must have been reproducible (Oken et al., 2015; Miller, 2005).

The current study was approved by the University of Illinois at Chicago IRB (#2015–0353). Women seeking prenatal care at the University of Illinois at Chicago (UIC) Center for Women’s Health were recruited as volunteers in their third trimester (29–33 weeks of gestation) between 2014 and 2017. The cohort profile has been previously described by Koenig and colleagues (Koenig et al., 2020). Inclusion criteria were as follows: singleton pregnancy; naturally conceived pregnancy; 17–45 years of age; pre-pregnancy BMI ≥ 18.5; <34 weeks of gestation; sufficient fluency in English to provide consent and complete the study; and ability to independently provide consent. Exclusion criteria were as follows: live birth or another pregnancy (including ectopic and molar pregnancies) in the previous 12 months; pre-eclampsia; gestational diabetes mellitus or previously diagnosed type 1 or type 2 diabetes; autoimmune disorder; current or previous premature rupture of membranes or chorioamnionitis; previous spontaneous premature birth; current bacterial or viral infection; current steroid or anti-inflammatory treatment; history of bariatric surgery; malabsorptive condition (e.g., celiac disease); current hyperemesis; hematologic disorder (e.g., sickle cell anemia or trait, hemochromatosis); current tobacco use; alcohol consumption or illicit drug use; and current use of medications that decrease nutrient absorption (e.g., proton pump inhibitors). All women provided written informed consent.

For Project Viva, procedures for obtaining umbilical cord blood serum have been describe previously (Schaub et al., 2005). For the UIH cohort, umbilical cord blood was obtained by venipuncture shortly after time of delivery. Blood (approximately 5 mL per tube) was drawn into Red Top Serum Plus and Green Top Sodium Heparin 95 USP Units Blood Collection Tubes (BD Vacutainer). Red Tops were allowed to stand upright at room temperature for 30 min prior centrifugation at 1500 × g for 10 min at room temperature. Supernatants (serum) were collected, aliquoted, and stored at −80°C until further processing. Heparinized blood obtained in Green Tops was diluted 1:1 in 1× phosphate-buffered saline, pH 7.4 (PBS) and overlaid onto Ficoll-Paque Plus (GE Healthcare) density gradients. Density gradients were centrifuged at 400 × g for 30 min at room temperature without brake. Upper phase (diluted plasma) was drawn off, aliquoted, and stored at −80°C. Buffy coats (CBMCs) were drawn off, washed twice with 10 mL of 1× PBS, aliquoted, and stored in 500 mL of RNAlater (Qiagen). Viability and number of cells isolated were determined by diluting cellular suspensions 1:1 with Trypan Blue Solution 0.4% (wt/vol) in PBS (Corning) and counting live/dead cells > 7 μM using a TC20 Automated Cell Counter (Bio-Rad). The viability of isolated cells was >90% for all samples. Time from delivery to storage was recorded for every sample.

Total RNA was extracted from CBMCs using RNeasy kits (QIAGEN) following manufacturers protocol except for switching 70% ethanol for 100% ethanol. The quality and quantity of all the extracted RNA were analyzed with a RNA 6000 Nano Kit on the 2100 Bioanalyzer Instrument (Agilent) and ssRNA High Sensitivity Kit and Qubit (Invitrogen). RIN for all samples was >8. RNA was constructed into barcoded libraries using the TruSeq Stranded mRNA Library Prep Kit (Illumina). The pooled libraries were sequenced for a paired-end 151 read length. The DNA libraries were sequenced on HiSeq X Ten platform using HiSeq Reagent v2.5 kit (Illumina), following manufacturer's protocol.

PGLYRP-1 and sIL6Rα were assessed using Human PGLYRP1/PGRP-S DuoSet Elisa and Human IL-6 R alpha DuoSet Elisa (R and D Systems). Serum was diluted with 1% bovine serum albumin in PBS (pH 7.2–7.4, 0.2 micron filtered) at 1:100 for PGLYRP-1 and 1:300 for sIL6Rα. ELISAs were performed according to manufacturer’s protocol. All samples were run in duplicate. Optical densities were assessed at 450 and 540 using a Spectra Max M5 (Molecular Devices). The intra- and inter-plater CVs for PGLYRP-1 were 2.4% and 11.0%. The intra- and inter-plater CVs for sIL6Rα were 3.8% and 18.9%.

All statistical analyses were performed in R (https://www.r-project.org/) unless otherwise specified.

To identify dataset studies used in the meta-analysis, NCBI’s Gene Expression Omnibus (https://www.ncbi.nlm.nih.gov/geo) was searched using the search ‘(cord blood) AND ‘Homo sapiens’ [porgn:_txid9606]’ and was limited to study types that included expression profiling by array. This search yielded 352 studies that were further examined for cell types assessed and metadata reported. Seventeen studies met inclusion criteria of reporting metadata regarding at least one perinatal risk factor (e.g. gestational age at birth, newborn sex, birthweight, maternal pre-pregnancy BMI, smoke exposure, mode of delivery), and expression data from either whole cord blood or CBMCs derived from human subjects. Expression, feature, and subject demographic data were extracted using geoquery (Davis and Meltzer, 2007). If expression data was non-normalized, it was quantile normalized and log2 transformed. Six studies were excluded due to no variability in demographic data (i.e. only males) or low data quality, leaving 605 unique cord blood gene expression samples. To assess associations between gene expression and perinatal risk factors, univariate, inverse variance weighted, random-effects models were constructed for genes using the GeneMeta package (GeneMeta, 2020). Newborn sex (Male vs. Female), maternal pre-pregnancy BMI (continuous: 0 = BMI < 18.5, 1 = 18.5 ≤ BMI < 25, 2 = 25 ≤ BMI < 30, 3 = 30 ≤ BMI), gestational age at birth in weeks (continuous), and birthweight in grams (continuous) were assessed as perinatal risk factors. Significant genes were defined as Benjamini–Hochberg correct p-value<0.01. The z-score for each gene was averaged across each univariate test (Equation 1) and termed pooled z-score. It was used to assess how likely a gene is effected by multiple risk factors. If a gene was not assessed in one of the univariate analyses while assessed in others, the missing data was inferred as a z-score of zero. To determine cell enrichment, genes expression in each cell as defined by HPA (Uhlen et al., 2010; Monaco et al., 2019) were modeled as a function of the pooled z-score using general additive model with cubic splines function. To determine biological processes enriched in the low- vs. high-risk individuals, the pooled z-score was used a pre-ranked list for gene set enrichment analysis for GO biologic processes using GSEA (Subramanian et al., 2005). Significantly enriched GO terms were defined as Benjamini-Hochberg correct p-value<0.01. Spearman’s correlations between pooled z-score, univariate z-scores, and individual dataset z-scores were determined in R.(1)Zpooled=((ZMale+ ZPPBMI) −(ZGA+ ZBW))4

The sequences were quality controlled by filtering out all low-quality reads (<25 on Phred quality score) and short reads (<50 bp). Transcripts were annotated using salmon v0.12.0 and Ensembl Homo sapiens Genome Assembly GRCh38.12 (Patro et al., 2017). Transcript counts were aggregated in gene-level counts using the tximport package in R (Soneson et al., 2015). Genes with median counts across samples < 10 were filtered out, leaving 14,055 genes remaining whose expression normalized using median sum scaling. Normalized gene expression was modeled as function of the number of perinatal risk factors (gestational age < 37 weeks, birthweight < 3000 g, PP BMI >30, male) using DESeq2 (Love et al., 2014). Genes were ranked for replication by their product of their pooled z-score and RNAseq z-score, termed RS (Equation 2). A cutoff of RS > 3 was used to determine candidate genes associated with pediatric asthma risk. Candidate gene biologic process and cellular component enrichment were performed using STRING with default settings (Szklarczyk et al., 2019). Data from RNAseq is available on NCBI SRA database ID PRJNA577955.(2)RS =Zpooled × ZRNAseq

Methods for data acquisition for cell population percentages and protein abundances using mass cytometry and ProSeek are reported by Olin et al., 2018. Cell population percentages were transformed using centered log-ratios. Cell populations and protein abundances, and number of risk factors for each individual (gestational age < 37 weeks, male, birthweight < 3000 g, birth via c-section) were correlated (Pearson’s method). For cell population correlations with number of risk factors, significance was defined as Bonferroni corrected p-value<0.05. Proteins determined to neutrophil associated were determined by those expressed in CBMC (UIH cohort) and enriched in neutrophils HPA (Uhlen et al., 2010).

Outcomes assessed in Project Viva Categorical outcomes were modeled using logistic regression. Continuous outcomes were modeled using linear regression. Four regression models were used for each outcome: univariate/unadjusted model 1 adjusted for child’s demographics (gestational age at birth in weeks, birthweight adjusted for gestational age and sex, mode of delivery, sex, and race/ethnicity), model 2 adjusted for maternal demographics (PP BMI, race/ethnicity, level of education, atopy, antibiotic use during pregnancy, and smoking during pregnancy, 6 months, or 1 year), and model 3 adjusted for both mother and child’s demographics (excluding maternal race/ethnicity). For regression models, PGLYRP-1 and sIL6Rα were log10 transformed and standardized to internal z-score. Subset analysis was performed by splitting the full data set by categorical variables and modeling outcomes as function of PGLYRP-1 in each subset using the univariate model. R (Jaakkola et al., 2006) for each variable in linear regression model 3 was determined using the relaimpo package (Groemping, 2006). Mcfadden’s pseudo-R (Jaakkola et al., 2006) for each variable in model 3 was calculated for logistic regression using full models minus the variable of interest.