Circular RNAs are the predominant transcript isoform from hundreds of human genes in diverse cell types

Abstract

Most human pre-mRNAs are spliced into linear molecules that retain the exon order defined by the genomic sequence. By deep sequencing of RNA from a variety of normal and malignant human cells, we found RNA transcripts from many human genes in which the exons were arranged in a non-canonical order. Statistical estimates and biochemical assays provided strong evidence that a substantial fraction of the spliced transcripts from hundreds of genes are circular RNAs. Our results suggest that a non-canonical mode of RNA splicing, resulting in a circular RNA isoform, is a general feature of the gene expression program in human cells.

Figure 1. Models to explain exon scrambling.

The canonical linear reference transcript is depicted with exons as colored boxes with four exons 1, 2, 3, and 4. Two simple models of RNA structure that could explain scrambled transcripts are depicted at left and right. At left, model 1 depicts how a scrambled exon 3-exon 2 junction could arise from a tandem duplication of exons 3 and 2, positioning the first copy of exon 3 upstream of exon 2. At the RNA level, this event could arise from post-transcriptional exon rearrangement, or a genomic duplication of exons 2 and 3. Under the model of tandem duplication, when one side of a paired-end read maps to the junction between exon 3 and 2, the other may map to any of exons 1, 2, 3 or 4 with probabilities determined by the library's insert length distribution and the exon lengths. Our data supports paired-end mapping between a junction and exons 2 or 3, but not exons 1 and 4. We note that in principle, the scrambled exon 3 - exon 2 junction could arise from other splicing events and does not necessarily entail tandem duplication. At right, model 2 depicts how a scrambled exon 3 - exon 2 junction could arise from splicing of exons 2 and 3 into a circular RNA molecule, again positioning exon 3 upstream of exon 2. In this model, when one side of a paired-end read maps to the junction between exon 3 and 2, the other will map to exon 2 or exon 3.

Figure 2. Expression levels of scrambled exons.

Analysis of paired-end RNA-Seq data from random primed libraries reveals evidence that scrambled exons are present at high stoichiometries compared to the canonical linear transcript transcribed from a large number of human genes. This phenomenon persists across cell types and is illustrated by the expression patterns of 3 leukocyte cell types: CD19 (B cells), CD34 (stem cells) and neutrophils. The fraction of each scrambled transcript as a fraction of total gene expression is computed. The bar plot depicts the number of circular isoforms with estimated abundance relative to all transcripts of the gene in the following ranges: between 0–25%, 25–50%, 50–75% and 75+%. Hundreds of isoforms in each cell type are estimated to represent more than half of all transcripts from each gene.

Figure 3. RNaseR assay confirms scrambled exons arise from circular RNA.

Panel A: Total RNA from HeLa cells was digested with RNaseR at varying enzyme concentrations (0, 3, 10, and 100 units) after the RNA was depleted of ribosomal RNA. Primers capable of amplifying the canonical linear transcript and the predicted circular transcript (by outward facing primers within a single exon predicted in the scramble) were used in a RT-PCR experiment for each of the digestion conditions. Canonical transcripts were consistently degraded by RNaseR, only detectable by PCR at 0 units of RNaseR, whereas predicted circular transcripts consistently resisted the RNaseR challenge, providing strong evidence of circularity. FBXW4 and MAN1A2 respectively show 2 and 4 circular isoforms, both of which were predicted by the sequencing data. The predicted lengths of circular isoforms are respectively a 3-2 junction of CAMSAP1 (predicted to produce a 435 bp circle), a 4-2 and 5-2 junction of FBXW4 (predicted to produce 415 and 510 bp circles), a 4-2, 5-2 and 6-2 junction of MAN1A2 (predicted to produce 471, 553, and 648 bp circles), a 3-3 junction in REXO4 (predicted to produce a 338 bp circle), a 2-2 junction of RNF220 (predicted to produce a 742 bp circle) and a 3-2 junction of ZKSCAN1 (predicted to produce a 667 bp circle). Panel B: A northern blot on total and cytoplasmic lysate from HeLa cells shows hybridization of a 481 bp probe complementary to the MAN1A2 5-2 exon scramble. 3.7 and 6.2 ug of total and cytoplasmic RNA were loaded onto a 1% agarose gel and 10 pM of probe was hybridized for 24–48 hours. Detection was performed using the BrightStar BioDetect Kit (Ambion, Austin, TX). The specific band at 553 bp corresponds to the predicted size of a circular RNA containing exons 2,3,4 and 5 of MAN1A2.

Figure 4. Scrambled exons are enriched in poly-A depleted samples.

Single-end 76-bp RNA-Seq was performed on matched experiments on HeLa, and H9 Human embryonic stem cell lysates were polyA selected and polyA depleted (data from Yang et al [22]). The numbers of scrambled exons detected in each sample which appeared in our curated database of scrambled junctions from the leukocyte data are depicted as colored bars. Roughly equal numbers of sequencing reads were available from each of 4 samples. Left panels of bar plot: both H9 and HeLa cells show markedly more exon scrambles in polyA depleted fractions compared to polyA enriched fractions, consistent with scrambles arising from circular transcripts which lack polyA tails. Right panels of bar plot: conversely, in the much smaller subset of scrambled exon pairs where we have evidence of internal tandem duplication (i.e. evidence against circularity), we find the opposite enrichment: more exon scrambles in polyA enriched fractions compared to polyA depleted fractions, consistent with this small subset of scrambles arising from linear, polyA transcripts.

Figure 5. qPCR shows scrambled exons are enriched in the cytoplasm.

HeLa whole cells lysates were fractioned into cytoplasmic and nuclear. The nuclear localized noncoding RNA XIST served as a control for fractionation:, and as expected, was enriched in the nuclear fraction. In addition, precursor ribosomal RNA bands were present in the nucleus but not the cytoplasm (see Figure S4). Using probes specific to each canonical and circular isoform (corresponding to those examples depicted in Figure 3), we compared Ct values calculated from qPCR on cDNA from the cytoplasmic fraction to the Ct value from qPCR on cDNA from the nuclear fraction. Bar heights show this average Ct value difference across 2 biological replicates. Error bars represent 2.5 standard deviations computed from biological variation of the qPCR assay. These results show that most circular isoforms are more enriched in the cytoplasm compared to the canonical linear isoforms.

Figure 6. Models for generation of circular RNA.

At left: a schematic diagram of the canonical splicing process splicing out the first intron of the a pre-mRNA of a 4 exon gene, and subsequent removal of introns 2 and 3. Canonical splicing of exon 1 to exon 2 occurs when the splicing machinery catalyzes the formation of the intron lariat and the attack of the free 3′ OH of exon 1 on the 3′ splice site upstream of exon 2. This produces a lariat containing intron 1 and a pre-mRNA with exons 1 and 2 spliced together. At right: a model for the production of circular transcripts. If there is a canonical transcriptional start, and if intron excision does not proceed sequentially in time from the 5′ to 3′ direction of the pre-mRNA, non-canonical pairing of 3′ and 5′ splice sites could be generated. Since the sequences of each 5′ splice site of the pre-mRNA contain the same splicing signals, it is possible that the 3′ splice site upstream of exon 2 is paired with the 5′ splice site downstream of exon 3 and splicing proceeds as if this 5′ splice site were paired with the 3′ splice site upstream of exon 4. In this case, exon 3 would be spliced upstream of exon 2, creating a pre-mRNA intermediate comprised of these two exons and intron 2. Canonical splicing would be predicted to excise this intron, leaving a circular RNA composed of exons 2 and 3. Non-canonical transcription start, as suggested in , could produce an orphan 3′ splice site corresponding to the first transcribed exon. This splice site could be paired with a downstream 5′ splice site, generating a circular RNA. In both models, the excised intron would be linear and branched, and expected to be quickly degraded.