A Simple Bias Reduction Algorithm for RNA Sequencing Datasets

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 in the clinical diagnostic space for cancer patients. While the information gained is intended to impact treatment decisions, numerous technical and quality issues remain. This includes inaccuracies in the dissemination of gene-gene relationships. For such reasons, clinical decisions are still mostly driven by DNA biomarkers, such as gene mutations or fusions. In this study, we aimed to correct for systemic bias based on RNA-sequencing platforms in order to improve our understanding of the gene-gene relationships. To do so, we examined standard pre-processed RNA-seq datasets obtained from three studies conducted by two consortium efforts including The Cancer Genome Atlas (TCGA) and Stand Up 2 Cancer (SU2C). We particularly examined the TCGA Bladder Cancer (n = 408) and Prostate Cancer (n = 498) studies as well as the SU2C Prostate Cancer study (n = 208). Using various statistical tests, we detected expression-level dependent, per-sample biases in all datasets. Using simulations, we show that these biases corrupt the results of t-tests designed to identify expression level differences between subpopulations. Importantly, these biases introduce large errors into estimates of gene-gene correlations. To mitigate these biases, we introduce Local Leveling as a novel mathematical approach that transforms count level data and corrects these observed biases. Local Leveling specifically corrects for the bias due to the inherent differential detection of transcripts that is driven by differential expression levels. Based on standard forms of count data (Raw counts, transcripts per million, fragments per kilobase of exon per million), we demonstrate that local leveling effectively removes the observed per-sample biases, and improves the accuracy in simulated statistical tests. Importantly, this led to systemic changes of gene-gene relationships when examining the correlation of key oncogenes, such as the Androgen Receptor, with all other detectable genes. Altogether, Local Leveling improves our capacity towards understanding gene-gene relationships, which may lead to novel ways to utilize the information derived from clinical tests.