Optimal use of statistical methods to validate reference gene stability in longitudinal studies

Multiple statistical approaches have been proposed to validate reference genes in qPCR assays. However, conflicting results from these statistical methods pose a major hurdle in the choice of the best reference genes. Recent studies have proposed the use of at least three different methods but there is no consensus on how to interpret conflicting results. Researchers resort to averaging the stability ranks assessed by different approaches or attributing a weighted rank to candidate genes. However, we report here that the suitability of these validation methods can be influenced by the experimental setting. Therefore, averaging the ranks can lead to suboptimal assessment of stable reference genes if the method used is not suitable for analysis. As the respective approaches of these statistical methods are different, a clear understanding of the fundamental assumptions and the parameters that influence the calculation of reference gene stability is necessary. In this study, the stability of 10 candidate reference genes (Actb, Gapdh, Tbp, Sdha, Pgk1, Ppia, Rp113a, Hsp60, Mrp110, Rps26) was assessed using four common statistical approaches (GeNorm, Norm Finder, Coefficient of Variation or CV analysis and Pairwise Delta Ct method) in a longitudinal experimental setting. We used the development of the cerebellum and the spinal cord of mice as a model to assess the suitability of these statistical methods for reference gene validation. GeNorm and the Pairwise Delta Ct were found to be ill suited due to a fundamental assumption in their stability calculations. Highly correlated genes were given better stability ranks despite significant overall variation. NormFinder fares better but the presence of highly variable genes influences the ranking of all genes because of the algorithm's construct. CV analysis estimates overall variation, but it fails to consider variation across groups. We thus highlight the assumptions and potential pitfalls of each method using our longitudinal data. Based on our results, we have devised a workflow combining NormFinder, CV analysis along with visual representation of mRNA fold changes and one-way ANOVA for validating reference genes in longitudinal studies. This workflow proves to be more robust than any of these methods used individually.