Low Preservation of Rank Order between mRNA and Protein Levels across All Individuals
We first compared the correlations between mRNA expression levels within individuals measured by three expression platforms: Affymetrix exon arrays,25 Illumina expression arrays,24 and RNA-sequencing (RNA-seq).26 We observed a strong correlation between these independent measures of transcript abundance for all genes (Figure 2A; RNA-seq versus Illumina expression array, median ρ = 0.67; RNA-seq versus Affymetrix exon array, median ρ = 0.82; Illumina expression array versus Affymetrix exon array, median ρ = 0.62). These observations support previous reports that have demonstrated similarly highly correlated mRNA expression measurements between RNA-seq and expression array technologies when measuring mRNA expression from a single individual (ρ = 0.60–0.77).
However, for expression QTL analysis, the more relevant comparison is how well expression levels correlate for each gene measured across individuals. We examined the preservation of rank order between 173 overlapping gene-level protein and mRNA measurements across the 52 individuals that were examined in each study (Figure 2). Notably, the correlation of interindividual mRNA expression measurements was low across mRNA expression platforms and laboratories (median ρ = 0.09–0.17), consistent with results from a previous eQTL analysis on this cohort using all genes (median ρ = 0.12).26 The correlation between mRNA levels was significantly higher between microarray platforms (median ρ = 0.17) than between microarrays and RNA-seq (median ρ = 0.10, p = 2.80 × 10−4, Wilcoxon rank sum test), indicating that either biological or platform variance contributed substantial variability to previous mRNA expression studies. Similarly to previous observations in yeast and mice,10,11,15 little correlation was observed between transcript and protein levels within genes, across individuals (exon array median ρ = 0.03, expression array median ρ = 0.01, RNA-seq median ρ = 0.02) (Figure 2B). Although global mRNA and protein levels were not strongly correlated across individuals, they were enriched to be correlated with 12% of genes displaying significant preservation of interindividual rank order between mRNA and protein levels, even in the presence of biological variation associated with the propagation of cells across different laboratories.
Several cellular characteristics including intrinsic growth rate, ATP levels, and EBV copy number have previously been shown to associate with gene expression levels and cellular phenotypes measured in LCLs.37,41,42 mRNA expression data sets have previously been adjusted for these cellular covariates to increase the ability to observe relationships between genotypes and mRNA expression levels.8,26,37 We surmised that these cellular variables might also be related to protein levels. We therefore tested for association between interindividual variation in these variables and with protein levels. We identified 197 protein variants that were nominally associated with at least one of six variables (Table S7). At an FDR of 5%, we found that 36 proteins were associated with intrinsic growth rate,35 28 proteins were associated with baseline ATP levels, and 21 proteins were associated with EBV copy number. Levels of phospho-S6 ribosomal protein (RPS6) and structural maintenance of chromosomes protein (SMC1A) were negatively correlated with cell growth (R = −0.33, p = 1.86 × 10−6 and R = −0.19, p = 2.62 × 10−4, respectively). Notably, we found that β-actin and α-tubulin protein levels were positively correlated with intrinsic growth rate (R = 0.12, p = 1.36 × 10−5 and R = 0.02, p = 0.04, respectively), suggesting that their use in total protein load normalization as housekeeping proteins, rather than the median sample load normalization we performed, would have resulted in an erroneous adjustment for intrinsic growth rate differences between cell lines.
Nominally associating with 21% of protein-level and 25% of RNA-seq-derived mRNA-level measurements, intrinsic growth rate was correlated with the highest number of mRNA and protein levels (Table S8). Indeed, of the 18 significant surrogate variables (SV) identified in the protein data set, the first SV was significantly associated with intrinsic growth rate (p = 0.03) and EBV copy number (p = 0.05), and the third SV was associated with intrinsic growth rate (p = 7.0 × 10−4), underscoring the high degree of association between global protein levels, intrinsic growth rate, and EBV copy number in LCLs.