Abstract
RNA sequencing (RNA-seq) is the conventional genome-scale approach used to capture the expression levels of all detectable genes in a biological sample. This is now regularly used for population-based studies designed to identify genetic determinants of various diseases. Naturally, the accuracy of these tests should be verified and improved if possible. In this study, we aimed to detect and correct for expression level-dependent errors which are not corrected by conventional normalization techniques. We examined several RNA-seq datasets from the Cancer Genome Atlas (TCGA), Stand Up 2 Cancer (SU2C), and GTEx databases with various types of preprocessing. By applying local averaging, we found expression-level dependent biases that differ from sample to sample in all datasets studied. Using simulations, we show that these biases corrupt gene-gene correlation estimations and t tests between subpopulations. To mitigate these biases, we introduce two different nonlinear transforms based on statistical considerations that correct these observed biases. We demonstrate that these transforms effectively remove the observed per-sample biases, reduce sample-to-sample variance, and improve the characteristics of gene-gene correlation distributions. Using a novel simulation methodology that creates controlled differences between subpopulations, we show that these transforms reduce variability and increase sensitivity of two population tests. The improvements in sensitivity and specificity were of the order of 3-5% in most instances after the data was corrected for bias. Altogether, these results improve our capacity to understand gene-gene relationships, and may lead to novel ways to utilize the information derived from clinical tests.