Characterization of Analytic and Experimental Uncertainty of RNA-seq Co-expression Network Determination: Application to SCA2
Bioinformatics sequencing pipelines produce results that contain a degree of uncertainty which stem from a variety of sources. For example, uncertainty can arise from sampling bias during sample preparation, sequencing platform bias, alignment errors and/or mathematical uncertainty from algorithmic assumptions. While these sources are well-known, few studies exist on the overall quantitative determination of this uncertainty. Here, we report on a formal approach to characterize the effect of uncertainty on a published Spinocerebellar Ataxia Type2 (SCA2) co-expression network analysis and show how the method can be extended to other bioinformatics pipelines.To quantify the effect of uncertainty, we used a RNA sequencing dataset from a previously published study examining important hub genes and pathways in Spinocerebellar Ataxia Type2 (SCA2). We systematically built 108 network analysis pipelines using three annotation databases, six alignment tools and six well-established normalization methods. We show low concordance in hub gene analysis, with few genes being identified in over 90% of all pipelines and no functional enrichment terms in all pipelines. Assessing the pipelines in aggregate, we find a new set of frequently occurring hub genes that were not identified in the original analysis. Finally, we quantify the effect of raw experimental uncertainty using estimated variations of gene expression and Monte Carlo simulations. We show that this uncertainty also can significantly affect subnetwork clustering which impacts hub gene and enrichment analysis.Our results suggest that best-practice guidelines and methodologies for uncertainty quantification are needed to create more reliable and reproducible analytical conclusions when determining co-expression.
