An Efficient Algorithm for Sensitively Detecting Circular RNA from RNA-seq Data

Abstract

Circular RNA (circRNA) is an important member of non-coding RNA family. Numerous computational methods for detecting circRNAs from RNA-seq data have been developed in the past few years, but there are dramatic differences among the algorithms regarding the balancing of the sensitivity and precision of the detection and filtering strategies. To further improve the sensitivity, while maintaining an acceptable precision of circRNA detection, a novel and efficient de novo detection algorithm, CIRCPlus, is proposed in this paper. CIRCPlus accurately locates circRNA candidates by identifying a set of back-spliced junction reads by comparing the local similar sequence of each pair of spanning junction reads. This strategy, thus, utilizes the important information provided by unbalanced spanning reads, which facilitates the detection especially when the expression levels of circRNA are unapparent. The performance of CIRCPlus was tested and compared to the existing de novo methods on the real datasets as well as a series of simulation datasets with different configurations. The experiment results demonstrated that the sensitivities of CIRCPlus were able to reach 90% in common simulation settings, while CIRCPlus held balanced sensitivity and reliability on the real datasets according to an objective assessment criteria based on RNase R-treated samples. The software tool is available for academic uses only.

Keywords: RNA-seq; de novo detection; high sensitivity; local similar sequence.

Figure 1

(a) Sensitivity analyses under different read depths of circRNAs (The read length was fixed to 100 bp). (b) Different parameter configurations of BWA–MEM affected the performance of detection on true positives (TP). (c) Sensitivity analyses under different read lengths. (d) Sensitivity analyses under different read depths of linear transcripts (The read length was fixed to 100 bp). (e) Sensitivity Analyses under different read depths of linear transcripts (The read length was fixed to 50 bp).

Figure 2

Comparisons of CIRCPlus and three popular circRNA detection methods. (a) Sensitivity and precision comparison among CIRCPlus and other methods, when the average read depth of linear transcripts was set to 10-fold (read length was set to 100 bp). (b) Sensitivity and precision comparison among CIRCPlus and other methods, when the average read depth of linear transcripts was set to 50-fold (read length was set to 100 bp)

Figure 3

(a) Overlap of circRNAs predicted by CIRCPlus and CIRI2 on the HEK293 datasets. (b) RNase R resistance of each chromosome predictions only detected by CIRCPlus.

Figure 4

CIRCplus workflow. (A) Extracting and classifying the unmapped reads from the alignments of each pair of reads. (B) Filtering the unmapped reads based on PEM signal to get the candidate BSJ reads. (C) Identifying the putative BSJ reads based on local similar sequences. (*) denotes the circular junction. (D) Clustering all putative BSJ reads from the same circRNA. In each box, the longest line on the bottom denotes the reference genome, where the colored regions represent different exons, and the black lines between exons simply represent introns. The short lines denote reads. For each read, the red triangle represents its direction, and a pair of dotted lines indicate a possible alignment of it mapping to the reference genome.

Figure 5

(a) The typical two-segment junction reads align to the reference genome separately in reverse orientation. (b) If one exon flanking the junction is shorter than the read length, the rest of the segment can be aligned to the nearby exon(s) contained in the circRNA. (c) If the length of a circRNA is shorter than the read length, it may map to the reference genome in another three-segment style. (d) The pair read of a junction read should align within the inferred circRNA area. (e) Any two of BSJ reads within a circRNA should have a similar sequence. The dotted lines suggest the possible alignment of each part of BSJ reads.

Figure 6

(a) A read from left-junction cluster should have a similar sequence with a read from right-junction cluster. (b) Two reads from two clusters are accepted during the detection. A red cross denotes a mismatch base-pair in the similar region. (c) Clustering the BSJ reads within a circRNA.