The normalized expression data were adjusted for age, gender and smoking status in a robust linear model to accommodate potential outliers in expression level (Figure 1). In parallel, we performed genotype QC to exclude SNPs of low call rate (<0.9) and deviating from Hardy-Weinberg equilibrium (P<1×10−6). Genotype imputations were based on the cleaned sets. In the end, assayed and imputed genotypes were used to identify cis and trans acting expression quantitative trait loci (eQTLs) following a method similar to that previously described in Schadt et al. [19]. The strategy for assembling of the asthma candidate gene list (Table S5) and for constructing Bayesian networks, and co-expression modules are given in Text S1.
