Capturing Heterogeneity in Gene Expression Studies by Surrogate Variable Analysis

Abstract

It has unambiguously been shown that genetic, environmental, demographic, and technical factors may have substantial effects on gene expression levels. In addition to the measured variable(s) of interest, there will tend to be sources of signal due to factors that are unknown, unmeasured, or too complicated to capture through simple models. We show that failing to incorporate these sources of heterogeneity into an analysis can have widespread and detrimental effects on the study. Not only can this reduce power or induce unwanted dependence across genes, but it can also introduce sources of spurious signal to many genes. This phenomenon is true even for well-designed, randomized studies. We introduce “surrogate variable analysis” (SVA) to overcome the problems caused by heterogeneity in expression studies. SVA can be applied in conjunction with standard analysis techniques to accurately capture the relationship between expression and any modeled variables of interest. We apply SVA to disease class, time course, and genetics of gene expression studies. We show that SVA increases the biological accuracy and reproducibility of analyses in genome-wide expression studies.

Genetic effects on gene expression across human tissues

“…In our primary analysis, we applied aggressive correction of potential confounders, controlling for 15–35 probabilistic estimation of expression residuals (PEER) factors 12 capturing 59–78% of total variance in gene expression levels (Supplementary Information 5). However, PEER and related approaches 30 may also remove variance in gene expression levels arising from regulatory pathways and broad trans effects. Indeed, several loci with numerous associations were found in uncorrected data, but disappeared after controlling for PEER factors (Supplementary Fig.…”

Sufficient dimension reduction and prediction in regression

“…Model (4.1), which leads to PCs as the MLE of the sufficient reduction, has a very general mean function, but its error structure is restrictive. Nevertheless, this error structure has recently been used in studies of gene expression (X) that are complicated by stratification and heterogeneity (Leek & Storey 2007). On the other hand, the usual linear model has a restrictive mean function, and under that model alone there seems to be no clear rationale for reduction by PCs (Cox 1968).…”

On Multiple Hypothesis Testing Maximizing the Average Power

“…For the purpose of applying this method to data sets that arise for example from gene expression studies via microarrays, there is much evidence that the data for different genes cannot be considered to be independent of each other, and correction methods such as surrogate variable analysis (SVA) (Leek & Storey, 2007) or its Partial Least Squares variant (SVA-PLS) (Chakraborty, Datta, Somnath, & Datta, Susmita, 2012) need to be applied to correct for this to allow much of this dependence to be accounted for in extra surrogate variables to be fitted. The p-values from the corrected tests for significance of the genes are almost (theoretically exactly) independent (Leek, & Storey, 2008).…”

“…Probability models for these patterns underlie the statistical analysis used in the search for such genes. Much recent theoretical work involves correction for effects not explicitly modelled that cause correlation among the data for individual tests (Leek & Storey, 2007;Leek & Storey, 2008;Lunceford et al, 2011;Chakraborty, Datta, Somnath, & Datta, Susmita, 2012), while the simpler problem of handling independent tests (Storey, 2007;Hwang & Liu 2010), does not seem to have been fully explored in its practical implementation when one or both hypotheses have unknown (hyper)parameters (Nixon, 2012). This may be because the perception of the need to deal with dependence makes such a study almost irrelevant.…”