Background and Summary

The advent of next-generation sequencing has spurred the discovery of a growing list of RNA biotypes, many of which are detectable across species, detected in numerous biofluids, and have biological function. While many studies have focused on microRNAs (miRNA), several other small RNA species (e.g. piwi-interacting RNAs (piRNA), tRNA fragments, and Y RNA fragments have been detected across a range of biofluids and are being developed as clinical biomarkers1,2,3,4. In addition to these linear RNAs, the discovery and detection of circular RNAs (circRNA), those with a covalently closed loop structure, have gained attention.

CircRNAs were initially discovered by electron microscopy, in the 1970s, as viroid molecules5. Nearly two decades later, circRNA were identified for a handful of mammalian genes6,7,8. Though initially thought to be rare splicing events, circRNAs have recently been identified as an abundant, endogenous RNA species in a number of organisms from Archaea to yeast, plants, worms, flies, fish, and mammals9,10,11. Additionally, circRNAs are abundantly expressed in a number of human tissues and cell types, and circRNA expression changes during development, and as a response to extrinsic factors such as stress, immune response, and hormonal stimuli12,13,14,15,16,17. These endogenous RNAs are characterized by their circular structures, which are formed by a back-splicing event that covalently links the 3′ “tail” splice donor with the upstream 5′ splice acceptor “head” of the transcript, forming a back-spliced, or “head-to-tail” junction. While circRNA function is still being elucidated, there are examples of circRNA inhibiting microRNA, regulating alternative splicing, and modulating the expression of parental genes18,19,20,21,22.

In comparison to their linear counterparts, circRNA transcripts can be more abundant and have greater stability as they are resistant to linear decay mechanisms and do not contain 5′-3′ polarity nor polyadenylated tails14,21,23,24, suggesting feasibility as stable biomarkers. CircRNA stability and detection in biofluids, saliva25, blood24,26,27,28,29,30, and urine31,32,33, comes, in part, from their being protected in extracellular vesicles28,34,35,36. Changes in circRNA expression is altered in multiple diseases, including preeclampsia, glioblastoma and colorectal cancer30,37,38. More recently, circRNAs in tumor tissues, as determined by next-generation sequencing, correlated with disease progression39,40. Urine circRNAs correlated with kidney rejection post-transplant31, while differentially expressed circRNAs have been determined in plasma exosomes of lung cancer patients versus controls41. Most studies of circRNA have small sample sizes or are based on targeted microarray data, rather than discovery-based methods. This dataset includes more than 100 samples from 54 volunteers from two easily accessible biofluids (plasma and urine). In some cases, multiple samples were collected from the same participant longitudinally, allowing us to assess the reliability of circRNA detection in biofluids.

The stability and abundance of circRNAs led us to investigate detection in two easily accessed biofluids: plasma and urine. As the volunteers were part of a larger study elucidating concussion biomarkers in male, college athletes, the samples are derived from young (18–25), healthy, male volunteers as depicted in Table 1. The longitudinal sample collections of plasma and urine are depicted in Online-only Table 1, including the number of circRNAs identified in each biofluid and those circRNAs observed concurrently in the biofluids. We identified circRNA in plasma (n = 134) and urine (n = 114), using RNAseq data followed by one of six different bioinformatic tools (Fig. 1). The intersection of the 6 bioinformatic tools provides a catalog for circRNA in plasma (Fig. 2a) and urine (Fig. 2c).

Table 1 Healthy Participant Characteristics.
Full size table
Fig. 1
figure 1

Study Workflow.

Full size image
Fig. 2
figure 2

CircRNAs were predicted from 134 plasma (a,b) and 114 urine (c,d) samples using 6 different bioinformatic tools. 965 circRNA were identified by all 6 tools in plasma (a; red bar), and 72 circRNA were identified by all 6 tools in urine (c; red bar). Genomic features located within predicted back-spliced junctions in plasma (b) and urine (d), respectively.

Full size image

As there are few datasets with circular RNAs cataloged in clinically-relevant biofluids, we expect this data to contribute to the characterization of circRNAs in young, healthy males. While this might be a direct comparator for concussions, or other diseases more prevalent in young men, we also expect this dataset to help begin to fill out a broader assessment of circRNAs present in healthy populations.

Methods

Sample collection and participants

Samples were collected from healthy, male volunteers, ages 18–25, with consent and approval from the Western Institutional Review Board (WIRB) study ID #1307009395. All participants provided written consent prior to enrollment. We obtained plasma (n = 134) and urine (n = 114) samples from 54 healthy male volunteers. In 71.4% of participants, both biofluid types were collected from the same individual. Blood samples were collected in EDTA tubes, and urine was collected in sterile cups. After collection, samples were placed in a cooler with ice packs and transported from Arizona State University to the Translational Genomics Research Institute, within 2–3 hours of collection. Blood samples were spun down at 1320 x G for 10 minutes at 4 °C, and 1 mL aliquots of plasma were collected in RNase/DNase free microcentrifuge tubes (VWR) and stored at −80 C. Urine samples were spun at 1900 x G for 10 minutes at 4 °C and 15 mL aliquots were collected in 50 mL conical tubes for storage at 80 °C.

RNA isolation, library preparation, and sequencing

For plasma samples, total RNA was isolated from 1 mL plasma using the mirVana PARIS RNA and Native Protein Purification Kit (Thermo Fisher, Cat. No.: AM1556) as in Burgos et al.42, treated with the DNA-free DNA Removal Kit (Thermo Fisher, Cat. No.: AM1906), and purified and concentrated with RNA Clean & Concentrator – 5 columns (Zymo Research, Cat. No.: R1016) by following Appendix C in the kit’s protocol. For urine samples, total RNA was isolated from 15 mL urine using Norgen’s Urine Total RNA Purification Maxi Kit (Slurry Format) (Norgen, Cat. No.: 29600), treated with the RNase-Free DNase Set (Qiagen, Cat. No.: 79254), and concentrated with the speed vacuum. The isolated RNA was quantitated with Quant-iT Ribogreen RNA Assay (Thermo Fisher, Cat. No.: R11490). Samples were not ribo-depleted, double-stranded cDNA was synthesized from 10 ng total RNA with the SMARTer Universal Low Input RNA Kit for Sequencing (Clontech, Cat. No.: 634940) using thirteen PCR cycles. The double-stranded cDNA was quantitated with the Qubit dsDNA HS Assay Kit (Thermo Fisher, Cat. No.: Q32854). For each healthy control sample, Illumina-compatible libraries were synthesized from 2 ng double-stranded cDNA with Clontech’s Low Input Library Prep Kit (Clontech, Cat. No.: 634947) using four mandatory PCR cycles plus ten additional cycles. Each library was measured for size via Agilent’s High Sensitivity D1000 Screen Tape and reagents (Agilent, Cat. No.: 5067–5602 & 5067–5585) and measured for concentration via the KAPA SYBR FAST Universal qPCR Kit (Kapa Biosystems, Cat. No.: KK4824). Libraries were then combined into equimolar pools, and each pool was measured for size and concentration. Pools were clustered onto a paired-end flowcell (Illumina, Cat. No.: PE-401–3001) with a 20% v/v PhiX v3 spike-in (Illumina, Cat. No.: FC-110-3001) and sequenced on Illumina’s HiSeq. 2500 with TruSeq v3 chemistry (Illumina, Cat. No.: FC-401-3002). The first and second reads were each 83 bases.

CircRNA prediction

Samples were demultiplexed and raw fastqs generated using CASAVA (v1.8.2, Illumina). Raw fastqs were trimmed using cutadapt (v1.9) with a quality score cutoff of 30 and a minimum length of 30 bp43. For each sample, 6 different algorithms (Table 2) were used to predict circRNA: KNIFE v1.444, find_circ21, MapSplice245, CIRCexplorer46, CIRI247, and DCC48. Indices of the GRCh37/hg19 genome were created using bwa and STAR v2.4.0j using default parameters49,50; bowtie and bowtie2 genome indices were downloaded with the KNIFE package51,52. Reads were mapped to the genome with the recommended aligner and alignment parameters for each program: STAR v2.4.0j for DCC and CIRCexplorer, bowtie2 v2.2.1 for find_circ and KNIFE, bowtie v0.12.9 for MapSplice2, and bwa v0.7.13 for CIRI2. CircRNA prediction was then completed with the suggested parameters for each program, with the exception of incorporating a minimum 18nt overlap on either side of the junction. CircRNAs were kept for downstream analysis if they 1) had 2 or more junction counts and 2) were identified in at least 5 samples for each respective program.

Table 2 CircRNA program characteristics.
Full size table

Analysis of predicted circRNA

The version of CIRCexplorer used here does not support paired-end data; therefore, circRNA prediction was performed on each pair separately and then combined for analysis. For each program, BED files containing count expression data were created from the output data. CIRCexplorer, KNIFE, and find_circ output files all produce output files with 0-based coordinates while CIRI2, MapSplice, and DCC output files have 1-based coordinates; therefore, all coordinates were converted to a 0-based system for comparison. BED12 GRCh37 RefSeq gene annotation files were obtained from UCSC (http://genome.ucsc.edu/cgi-bin/hgTables), and bedtools v2.26.0 was used to infer genes from reported backsplice junction genome locations53. Data were analyzed using the R v3.3.2 statistical package (https://cran.r-project.org). UpSet plots were generated using the UpSetR v1.3.3 package54.

Quantification of circRNA expression

CircRNA count expression data was obtained from each respective bioinformatic program. Junction reads per million (JRPM) were calculated according to the total number of junction reads found in each sample as identified by STAR (both canonical and chimeric); therefore, JRPM = (circRNA count/junction reads) * 1,000,000. The circular-to-linear ratio (CLR) for each circRNA was calculated as described previously13,27, by counting the linear spliced reads identified by STAR on the 5′ and 3′ flanks of each circRNA junction, and dividing the back-spliced read count by the flank with the highest count; therefore, CLR = circRNA count/max (5′ linear junction count, 3′ linear junction count). In order to avoid division by zero, if no linearly spliced reads were detected, a pseudo count of 1 was added to the denominator. The number of reads assigned to the transcriptome was calculated using featureCounts (subread v1.5.1) with the Ensembl75 gene annotation55. Differential expression analysis was performed using DESeq. 2 v1.14.156, after filtering to select samples which had detected at least 300 circRNA/sample as well as exclusion of circRNA that were expressed in less than 50% of samples.

DNA isolation and qRT-PCR

After centrifugation of blood samples, DNA was isolated from the buffy coat using the DNeasy Kit (Qiagen, Cat. No.: 69504). Previously isolated RNA from samples matching those used for library prep were selected for cDNA synthesis. cDNA was synthesized with random hexamers using the SuperScript III First-Strand Synthesis System for RT-PCR following manufacturer’s protocols (Invitrogen, Cat. No.: 18080-051) with three nanograms of total RNA as input, and stored at −20 °C. Inward-facing (crossing the back-splice junction) custom primers were designed with Primer3 and LabReady primers (100 µM in IDTE pH 8.0) were ordered from Integrative DNA Technologies with Standard Desalting Purification57,58. Real-time qRT-PCR was performed with SYBR Select Master Mix (Thermo Fisher, Cat. No.: 4472919) on the QuantStudio 7 (Applied Biosystems), with 0.2 µM of primer and 0.2 µL of cDNA template or 2 ng of gDNA template per 10 µL reaction. U6 was used as a positive control and no template controls (NTCs) were used as a negative control. All results are expressed as the mean of three independent reactions, with a standard deviation less than 0.5. The ReadqPCR v1.20.0 and NormqPCR v1.20.0 Bioconductor v3.4 packages were used for qRT-PCR data analysis59.

Data Records

Raw FASTQ files for the RNAseq libraries were deposited into dbGap (accession # phs001258.v2.p1) (https://identifiers.org/dbgap:phs001258.v2.p1)60. Data (circRNAs identified across all informatic tools and raw cirRNA expression) are also provided in figshare: https://doi.org/10.6084/m9.figshare.c.542083261.

Technical Validation

CircRNA set size and genomic alignment

The set size (all circRNA in any sample by one tool) ranges from 1,835 to 7,462 and 163 to 1,349 in plasma and urine, respectively (Table 3). 965 and 72 circRNA were detected across all six tools in plasma and urine, respectively (Fig. 2a,c, red bars; Table 4; full list in figshare File 1 and 261). KNIFE predicted the most circRNA per sample in plasma and urine, while MapSplice predicted the fewest (Table 3). Table 5 displays the correlations between all of the tools, CIRCexplorer and DCC had the highest correlation. 85% (61 of the 72) of the circRNAs found in urine were also detected in plasma (Table 4). Figure 2b(plasma) and 2d (urine) display the number of detected circRNAs and the number that span introns, exons, and UTRs for both plasma and urine. The majority of circRNA identified in plasma and urine contain at least two exons and span an intron; 671 in plasma and 52 in urine; green bars (Fig. 2b, plasma and 2d, urine). A small number of circRNA are transcribed from a single exon (15 in plasma and 2 in urine).

Table 3 CircRNA totals detected across six informatic tools in plasma and urine.
Full size table
Table 4 Number of circRNA detected in plasma and urine by all 6 bioinformatic tools.
Full size table
Table 5 Pearson’s correlation of circRNA expression (JRPM) between informatic tools.
Full size table

Highly expressed, back-spliced junctions were validated by qRT-PCR

In order to validate predicted back-spliced junctions by qRT-PCR, we designed inward-facing primers for the 15 most highly expressed circRNA in each biofluid and tested each primer pair in samples from 10 different individuals, using the same source RNA for cDNA synthesis that was used for RNAseq (Fig. 3a,b). Figure 3a shows that the 15 circRNAs are detected in most of the 10 plasma samples. The numbers of samples are described in Table 6, and compared with the RNASeq detection for those circRNAs in the same samples. 13 primer pairs were validated in urine. Detection in urine samples was sparse, with fewer samples positive for each circRNA than for plasma (Fig. 3b and Table 6). For the two back-spliced junctions detected in RNASeq data, but not validated by qRT-PCR in urine (circMYO5B and circPHC3), it is possible that the circRNA primers did not work, or there were qPCR inhibitors in the sample, or the circRNA was not present. Two of the samples did not have enough assigned reads via RNASeq to be included, so the total number of samples was 8. In order to rule out chimeric junctions that might be present in DNA or resemble artifacts introduced during library preparation, we also used genomic DNA (gDNA) from each individual as a negative control. All 15 primer pairs used in the plasma and urine samples were not detected in gDNA (data not shown). Table 7 describes the rank from highest to lowest expression for each of the circRNA validated by qRT-PCR, and compares it with the expression detected with sequencing. Their ranks do not correlate well between the two platforms.

Fig. 3
figure 3

(a,b) Highly-expressed, predicted back-spliced junctions were validated by qRT-PCR. qRT-PCR validation of the 15 most highly expressed circRNA found in plasma (a) and urine (b), respectively. Each circRNA was examined in 10 cDNA samples from the same source RNA as sequenced samples. (c,d) Circular-linear ratios are higher in plasma than urine. Linear splice junction expression plotted against circular splice junction expression in plasma (c) and urine (d). Points representing circRNA between 1-fold and 5-fold higher than their linear counterparts are blue; 5x or higher are red.

Full size image
Table 6 circRNA detection in 10 samples by qRT-PCR and RNASeq.
Full size table
Table 7 qRT-PCR and RNA-Seq expression of the 15 most highly expressed genes in plasma and urine.
Full size table

Circular-to-linear RNA ratios

While the overall expression of most circRNAs is low compared to their linear counterparts, there are a number of circular RNA transcripts that have been described as more abundant than their linear host, cellularly as well as extracellularly23,27,62,63. We examined the circular-to-linear ratio (CLR) of circRNA transcripts found in plasma and urine as described previously; by taking the ratio of the circular, back-spliced junction counts compared to the linear count of the nearest 5′ or 3′ splice junction13,24,27. On average, 28.5% of circRNA transcripts in plasma and 21.5% of circRNA transcripts in urine have higher expression than their linear host gene (Fig. 3c, plasma and 3d, urine). Extracellular RNA is often fragmented and may have a 3′ bias64. Before examining the expression of circular RNA in relation to their host genes, we calculated the overall 5′ to 3′ coverage of linear transcripts and did not find a bias in our samples.

Participants sequenced 5 or more times have less inter-sample variation

A notable feature of this dataset is that many participants were sampled longitudinally, allowing for analysis of circRNA stability in individuals versus the entire dataset. Figure 4a,b show longitudinal circRNA expression in the same participants in plasma and urine, respectively. Broadly speaking, the heatmaps demonstrate similar expression patterners in the same participant over time. In order to assess variability within individuals, we calculated the coefficient of variation (CV) of circRNA expression, normalized to junction reads per million (JRPM). Here, we focus on participants sampled on 5 or more occasions over approximately one year. In both plasma and urine, the CV for each individual participant is displayed along with the CV for all participant samples. The data indicate that individuals have a statistically-significant consistency in circRNA expression pattern over time (Fig. 4c, plasma and 4d, urine).

Fig. 4
figure 4

Participants sequenced five or more times have less inter-sample variation. CircRNA populations identified in plasma (a,c) and urine (b,d) from participants sampled five or more times. (a,b) Heatmaps showing the log-normalized JRPM expression of plasma (a) and urine (b) samples taken longitudinally from the same participant. The coefficient of variation (CV) of circRNA expression is significantly lower across individual participant samples when compared to the entire dataset (c, plasma; and d, urine). ****p <  = 0.0001.

Full size image

Usage Notes

As the approach to detecting circRNA from RNA-Seq data differ with available tools, we employed 6 different bioinformatic tools: CIRCexplorer, CIRI2, DCC, KNIFE, find_circ, and MapSplice, in two clinically relevant biofluids, plasma and urine, using 134 and 114 samples, respectively. Most of these circRNA pipelines use an external aligner, such as bowtie, STAR, or bwa, to align reads to the genome and/or transcriptome (Table 2). After alignment, reads that contiguously align to the genome and/or transcriptome are filtered out, and the remaining unmapped reads are further filtered to identify back-spliced junctions. Differences in circRNA identification algorithms include: 1) how paired-end reads and gene annotations are used, if at all, 2) the amount of overlap over the junction that a read must contain, 3) the types of junctions considered, and 4) various filtering steps (Table 1)65. We sought to generate a high confidence set of circRNA expressed in plasma and urine with the following requirements for each circRNA: 1) detection in at least 5 samples for each respective biofluid, 2) a minimum 18 nt overlap on either side of the junction, 3) at least two reads spanning the back-spliced junction, and 4) identification by all 6 tested bioinformatic tools as identification can vary widely between tools66,67,68.

We tested alignment parameters and their influence on the detection rate of circRNA and found that the number of input reads, genome mapped reads, and junction reads did not correlate well with the number of circRNA detected per sample; rather the number of reads assigned to the transcriptome had the greatest correlation with the number of circRNA (R2 = 0.805; data not shown).