Determining sufficient sequencing depth in RNA-Seq differential expression studies

Abstract

RNA-Seq studies require a sufficient read depth to detect biologically important genes. Sequencing below this threshold will reduce statistical power while sequencing above will provide only marginal improvements in power and incur unnecessary sequencing costs. Although existing methodologies can help assess whether there is sufficient read depth, they are unable to guide how many additional reads should be sequenced to reach this threshold. We provide a new method called superSeq that models the relationship between statistical power and read depth. We apply the superSeq framework to 393 RNA-Seq experiments (1,021 total contrasts) in the Expression Atlas and find the model accurately predicts the increase in statistical power gained by increasing the read depth. Based on our analysis, we find that most published studies (> 70%) are undersequenced, i.e., their statistical power can be improved by increasing the sequencing read depth. In addition, the extent of saturation is highly dependent on statistical methodology: only 9.5%, 29.5%, and 26.6% of contrasts are saturated when using DESeq2, edgeR, and limma, respectively. Finally, we also find that there is no clear minimum per-transcript read depth to guarantee saturation for an entire technology. Therefore, our framework not only delineates key differences among methods and their impact on determining saturation, but will also be needed even as technology improves and the read depth of experiments increases. Researchers can thus use superSeq to calculate the read depth to achieve required statistical power while avoiding unnecessary sequencing costs.