Evaluating the bias of circRNA predictions from total RNA-Seq data

CircRNAs are a group of endogenous noncoding RNAs. The quickly developing high throughput RNA sequencing technologies along with novel bioinformatics approaches have enabled researchers to systematically identify circRNAs and their biological functions in cells. Deep sequencing of rRNA-depleted RNAs treated with RNase R, which digests linear RNAs and leaves circRNAs enriched, is an efficient way to identify circRNAs. However, very few of RNase R treated data are at hand but a large amount of total RNA-Seq data with no sequencing costs is available, for circRNA predictions. In this study, we systematically investigated the prediction bias from total RNA-Seq data as well as the influence of sequencing depth, sequencing quality and single-end or paired-end sequencing strategy on the predictions. We also identified circRNA properties that may contribute to the improved prediction performance. Our analysis shows that circRNA predictions from total RNA-Seq data gain similar to 50% true positive. Sequencing error dramatically worsens the predictions, rather than single-end sequencing strategy or low sequencing depth. However, false positive can be carefully controlled by using data with good quality and narrowing down circRNAs guided by their properties.