Vhs activity leads to enrichment of circRNAs in HSV-1 infection

CircRNA detection from RNA-seq reads is based on identifying junction reads connecting a donor splice site of a downstream exon to the acceptor splice site of an upstream exon. Several algorithms for circRNA de novo detection are available, but due to high false positive rates combination of at least two algorithms is recommended [43]. We thus employed a pipeline combining circRNA_finder [44] and CIRI2 [45] for circRNA de novo detection (see methods and Fig 2A) and analyzed only circRNAs that were independently discovered by both algorithms. We applied this pipeline on our total RNA-seq and 4sU-seq time-courses for the first 8 h of lytic wild-type (WT) and Δvhs HSV-1 infection (Fig 1B). This identified a total of 16,463 circRNAs in the WT time-course and 7,306 in the Δvhs infection time-course, with 1,647 to 6,887 circRNAs detected per sample in total RNA (S1A Fig). In contrast, only a small number of circRNAs were detected in 4sU-RNA (119 and 194 in at least one sample of WT and Δvhs 4sU-seq, respectively, S1B Fig). This is partly consistent with the study by Zhang et al. who also obtained smaller numbers of circRNAs (255, 820, 876 circRNAs in PA1, hESC H9 and H9 differentiated FB Neurons cells, respectively) with 1 h of 4sU-tagging using 4sUDRB-seq (DRB treatment followed by 4sU-seq after DRB removal) compared to total RNA-seq (6,740, 4,523, and 11,185 circRNAs, respectively) [46]. While this highlights a high variability between experiments and/or cell types in the number of circRNAs detected, the higher number of circRNAs detected by Zhang et al. can also partly explained by the use of only one circRNA detection algorithm (CIRCexplorer [47]). This is less restrictive than requiring independent identification of circRNAs by two algorithms as done in our study, but likely has a higher false positive rate. Moreover, circRNA_finder and CIRI2 require presence of canonical GT-AG splicing signals, while CIRCexplorer also identifies potential circRNAs with other less common splicing signals. Most circRNAs in our study (>97%) were only detected in total RNA and almost all of the identified circRNAs (82.6% and 92.7% for the WT and Δvhs infection time-courses, respectively) were already annotated in circBase, a database of circRNAs identified in large-scale RNA-seq studies [48].

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 2. vhs-dependent enrichment of circRNAs in HSV-1 infection.

(a) Outline of our pipeline for de novo circRNA detection (see also methods). It combines two different algorithms, circRNA finder [44] and CIRI2 [45], and retains only circRNAs identified independently by both algorithms. circRNA finder is based on analysis of chimeric reads determined with the RNA-seq mapper STAR [49]. Chimeric reads are reads that align to two distinct regions of the genome in manner not consistent with “normal” linear transcripts. CIRI2 is based on identifying pairs of clipped read alignments (i.e., local alignments for substrings of the read, determined with BWA [50]) for the same read where a downstream part of the read aligns upstream of an upstream part of the read (= back-spliced junction (BSJ) reads). Both methods use further filtering steps to remove false positive results. (b) Number of circRNAs identified for each sample of total RNA for WT and Δvhs infection after adjusting the required number of reads for detecting a circRNA to the number of reads mapping to the host genome in each sample. Thus, a threshold of 2 reads was used for the sample with the lowest number of reads and higher thresholds were used for samples with higher number of host reads. Numbers of identified circRNAs with a threshold of 2 reads for each sample (as outlined in (a)) are shown in S1A Fig for total RNA and in S1B Fig for 4sU-RNA. (c, d) Scatterplots comparing normalized circRNA counts (normalized to the number of linear junction reads mapped to the host genome) between mock and 8 h post infection (p.i.) WT infection (c) and mock and 8 h p.i. Δvhs infection (d). Linear regression analysis across all circRNAs (red line) was used to estimate the enrichment of circRNAs relative to linear mRNAs in HSV-1 infection compared to mock infection. The regression estimate for the enrichment is shown on the bottom right. The gray line indicates the diagonal, i.e., equal values on the x- and y-axis. The five most highly expressed circRNAs are marked by name. Corresponding scatterplots for 2, 4 and 6 h p.i. are shown in S2 Fig. (e) Percentage of expressed circRNAs (= normalized circRNA count >0 in uninfected cells) at each time-point of WT HSV-1 infection for which at least one exon within the genomic region of the circRNA shows differential exon usage for the corresponding gene (determined with DEXSeq, multiple testing adjusted p-value ≤0.005). (f) Boxplots showing the distribution of log2 fold-changes for exons located within circRNAs. For each circRNA, only the exon with the maximum absolute log2 fold-change is shown.

https://doi.org/10.1371/journal.pone.0276467.g002

In total RNA, a substantially larger number of circRNAs was identified for WT infection than for mock and Δvhs infection (S1A Fig). Since the number of reads mapping to the host genome in total RNA was comparable or higher in Δvhs infection than in WT infection (S1C Fig), adjusting the read count threshold for a circRNA to be detected to sequencing depth further increased the difference in detected circRNAs between WT and Δvhs infection (Fig 2A). Thus, the lower number of circRNAs identified in Δvhs infection is not due to lower sequencing depth. Moreover, normalized to sequencing depth, a smaller number of circRNAs was detected in uninfected cells of the Δvhs infection time-course compared to the WT infection time-course. In part, this could be due to shorter read length (76 nt in the Δvhs infection time-course vs. 101 nt in the WT infection time-course, both paired-end sequencing), the effect of which will be further investigated below. However, other experimental differences between the two time-courses, which were performed at different times, likely contribute. To account for these experimental differences, we always compared samples of WT infection against mock infection of the WT infection time-course and samples of Δvhs infection against mock infection of the Δvhs infection time-course.

To evaluate enrichment of individual circRNAs relative to linear transcripts in HSV-1 infection, circRNA read counts were normalized to the total number of reads mapping to linear exon-exon junctions of all human protein-coding genes (not only the “parent” genes from which circRNAs originate) and then averaged between replicates. The reason we did not normalize only against reads from “parent” genes is that reads that map linearly to splicing junctions within circRNA regions could originate from either linear or circular transcripts. Excluding those reads would result in low numbers of reads for normalization, leading to very noisy estimates of enrichment, in particular for HSV-1 infection where linear transcripts are expected to be lost. Including those reads, however, would lead to under-estimation of circRNA enrichment. Since only a relatively small fraction of genes harbour circRNAs, use of junction reads from all genes circumvents this issue. Moreover, there are circRNAs without corresponding linear transcripts, such as CDR1as, one of the few circRNAs with a well-described function [16]. This circRNA would otherwise have to be excluded from our analysis.

Notably, the fraction of linearly spliced host reads decreased considerably during WT infection (S1D Fig). This is largely a consequence of HSV-1-induced read-through transcription. Read-through transcripts are rarely spliced downstream of the poly(A) site and are retained in the nucleus [21, 22], thus escaping vhs-mediated decay. In contrast, in vhs infection, the fraction of spliced reads did not decrease (S1E Fig) due to reduced read-through transcription and absence of vhs-mediated decay of cytosolic mRNAs. Thus, normalization by linear splice junction read counts better quantifies the impact of vhs-mediated decay on linear transcripts than normalization by the total number of reads mapped to the host genome as in WT infection a large fraction of host genome reads originate from read-through transcripts that escape vhs-mediated decay. Normalized circRNA counts increased substantially with increasing duration of WT infection, indicating increasing enrichment of circRNAs relative to linear transcripts (Fig 2C, S2A–S2C Fig, red line = linear regression estimate over all circRNAs). By 8 h p.i. WT infection, circRNAs were enriched ~6.5-fold in total RNA compared to uninfected cells. In contrast, Δvhs infection exhibited only a very modest ~1.3-fold enrichment in circRNAs by 8 h p.i. (Fig 2D, S2D–S2F Fig). This can likely be attributed to the global loss of host transcriptional activity in HSV-1 infection [51], which affects circRNAs less than linear transcripts as circRNAs are much more stable. Notably, if we instead normalize by the total numbers of reads mapped to the host (as done in the study by Shi et al. [19]), we also observe enrichment of circRNAs in WT infection, albeit to a lower degree (S3 Fig).

As noted above, one factor that might negatively impact circRNA detection in the Δvhs infection time-course is the shorter read length. CIRI2 and circRNA_finder require that at least one read overlaps by at least 19 and 20 nt, respectively, on both sides of identified circular junctions. Thus, longer reads could potentially lead to identification of more circRNAs. To address this issue, we pursued two alternative approaches. First, we trimmed reads from the WT infection time-course to 76 nt and applied our de novo circRNA detection method to these trimmed reads. Second, we implemented an alternative circRNA detection method based on aligning reads against circRNA junction sequences constructed from annotated circRNAs, which allowed to reduce or increase the required overlap on either side of the junction (see methods and S4 Fig). In this case, false positive rates are reduced by the restriction to annotated circRNAs. Both with trimmed reads and the alignment-based pipeline with a required overlap of only 10 nt, we observed approximately the same circRNA enrichment in WT infection as with the de novo detection pipeline on full reads (S5 and S6 Figs). Moreover, since circRNA enrichment is consistent between highly and lowly expressed circRNAs as shown by the low deviation from the linear regression line, exclusion of lowly expressed circRNAs by adapting circRNA detection thresholds to sequencing depth would not alter conclusions. Unless explicitly noted otherwise, all results reported in the following use the alignment-based pipeline requiring an overlap of at least 10 nt on either side of the junction.

We also analyzed data from a separate RNA-seq experiment in which we obtained total RNA for mock, WT, Δvhs and ΔICP27 infection in the same experiment as well as subcellular fractions (chromatin-associated, nucleoplasmic and cytosolic RNA) (all 8 h p.i., n = 2) [18, 22, 52]. This showed enrichment of circRNAs in total RNA of WT and ΔICP27 infection, but not Δvhs infection (S7 Fig), Notably, although linear regression analysis over all circRNAs (red line) estimated only a 1.5- to 2-fold enrichment in WT infection, some of the most highly expressed circRNAs were more strongly enriched. To confirm that enrichment of circRNAs in HSV-1 infection was not due to increased abundance of circRNAs, we also analyzed recently published total RNA-seq data for mock and HSV-1 strain KOS (WT KOS) infection (n = 1 replicate), where ERCC spike-ins were added to allow absolute quantification of expression levels [53]. This confirmed the strong enrichment of circRNAs relative to linear RNAs in a second HSV-1 strain (~10-fold, S8A Fig). However, normalization to ERCC spike-in read counts showed that absolute levels of most circRNAs in total RNA remained largely unchanged during HSV-1 infection (S8B and S8C Fig). We conclude that vhs-mediated degradation of linear mRNAs that spares circRNAs leads to an enrichment of circRNAs relative to linear mRNAs during HSV-1 infection.

As circRNA reads that do not cross circular junctions can be aligned by standard RNA-seq mapping programs, they are counted towards circRNA “parent” genes and exons in standard differential gene expression (DGE) and differential exon usage (DEU) analyses. Consequently, genes encoding for circRNAs tend to have significantly higher fold-changes in DGE analysis of total RNA in WT infection from 6 h p.i. (one-sided Wilcoxon rank sum test, p-value <0.0001, S9A Fig). This was not observed in Δvhs infection. Similarly, DEU analysis with DEXSeq showed differential exon usage in WT infection for many exons contained in circRNAs (multiple testing adjusted p-value ≤0.005). By 8 h p.i. WT infection, ~17% of circRNAs showed differential exon usage for at least one exon within the genomic range of the circRNA (Fig 2E), with increased exon usage for almost all of these circRNAs (Fig 2F). This effect was most pronounced for highly expressed circRNAs (S9B Fig). Consistent with the small degree of circRNA enrichment late in Δvhs infection, 3.6% of circRNAs also showed differential exon usage by 8 h p.i. Δvhs infection (S9C and S9D Fig). In summary, our results demonstrate that enrichment of circRNAs in WT infection biases not only differential circRNA analyses, but also DGE and DEU analyses of total RNA and thus needs to be taken into account for the latter.

Induction of NEAT1 linear and circular splicing in HSV-1 infection

In contrast to other circRNAs, one circRNA (circBase ID: hsa_circ_0003812) was actually induced in HSV-1 infection even in absence of vhs. hsa_circ_0003812 is encoded in the unique part of the long NEAT1_2 transcript (Fig 1C), was essentially absent in mock-infected cells (0–2 reads in 4sU-RNA, total RNA and all subcellular fractions, ≤0.22 normalized circRNA count) and highly abundant in HSV-1 infection (Fig 3A, S10 Fig). ERCC spike-in normalization in mock and WT KOS infection confirmed that abundance of this NEAT1_2 circRNA indeed increased in absolute levels (S8B and S8C Fig), indicating high levels of de novo synthesis of this circRNA in HSV-1 infection. Consistent with this, induction of the NEAT1_2 circRNA during HSV-1 infection was also confirmed in samples of very recently transcribed RNA, i.e., 4sU-, nucleoplasmic and chromatin-associated RNA. This is particularly noteworthy considering the small number of circRNAs we recovered in these samples (S1B, S1C, S11 Figs). From 6 h p.i., the NEAT1_2 circRNA was among the most highly expressed circRNAs in total, 4sU-, chromatin-associated and nucleoplasmic RNA in WT, ΔICP27 and Δvhs infection (S10 Fig). Consistent with the nuclear localization of NEAT1, only few circular NEAT1_2 reads were found in cytosolic RNA. Other NEAT1_2 circRNAs were also detected, but only with much fewer reads (Fig 3A). In summary, our data indicate that biogenesis of the hsa_circ_0003812 NEAT1_2 circRNA is induced during HSV-1 infection.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 3. Induction of a NEAT1_2 circular and linear splicing in WT, Δvhs and ΔICP27 infection.

(a) Heatmap of normalized circRNA counts (normalized to the number of linear junction reads mapped to the host genome) for all identified NEAT1_2 circRNAs in total and 4sU-RNA time-courses of WT and Δvhs infection and nucleoplasmic, chromatin-associated, cytoplasmic and total RNA for 8 h p.i. WT, Δvhs and ΔICP27 infection. Columns represent individual circRNAs, which are ordered according to their genomic coordinates from the most 5’ to the most 3’. The hsa_circ_0003812 NEAT1_2 circRNA is marked by a blue rectangle. (b, c) Sashimi plots showing NEAT1 read coverage (overlay of both replicates) and circular (red) and linear (same color as read coverage) splice junctions as arcs connecting acceptor and donor splice site in (b) selected samples of the total RNA time-courses of WT and Δvhs infection and (c) chromatin-associated RNA in mock, WT, Δvhs and ΔICP27 infection. Number of junction reads are annotated to arcs separately for the two replicates. Junctions are only shown if at least 4 reads align by at least 10 nt on both sides of the junction. Genomic coordinates of NEAT1 transcripts and the HSV-1-induced circRNA are shown at the bottom. Sashimi plots of the full total RNA and 4sU-RNA time-courses and nucleoplasmic and total RNA from the subcellular fractions experiment are shown in S13 and S15 Figs. (d) Heatmap of linear splicing rates (= number of reads for a linear junction / number of exon-intron reads crossing the corresponding acceptor and donor splice sites) for the WT and Δvhs infection time-courses and the subcellular fraction experiment (dark red = high, white = low, see color bar on the right). Columns represent splice junctions, which are ordered according to their genome positions, first by the donor and then by the acceptor splice site. Counts include only junction and exon-intron reads overlapping the junction or exon-intron boundary, respectively, by at least 10 nt on either side. A pseudocount of 0.01 was used to avoid division by zero. Rectangles on top indicate whether the splice junction is within the genomic region or downstream of the hsa_circ_0003812 circRNA. Colors on the left indicate the virus and time-point of infection (light colors early in infection, darkest colors 8 h p.i, cyan = WT, blue = Δvhs, gold = ΔICP27). Numbers in brackets at the bottom refer to the four most frequent splice junctions outlined in (e). (e) Schematic overview on the circular and the four most frequent linear NEAT1_2 splice junctions induced in HSV-1 infection and sequences around splice sites (“|” indicates exon-intron boundaries in sequences). The height of boxes indicates the relative abundance of each exon (= never or rarely spliced out) and intron (= frequently spliced out) by 8 h p.i. WT infection. Genomic coordinates on chromosome 11 are indicated for the start and end of exons. Numbers in brackets on top of introns indicate the ranking of corresponding splice junctions according to their splicing rate shown in (d). This ranking is also indicated in (d) at the bottom.

https://doi.org/10.1371/journal.pone.0276467.g003

The hsa_circ_0003812 NEAT1_2 circRNA was previously found to be enriched in RNA-seq data of human fibroblasts after treatment with the 3’->5’ RNA exonuclease RNase R [54], confirming that it is a circRNA. Notably, however, NEAT1_2, is partially resistant to RNase R due to the stabilizing triple helical structure at its 3’end, similar to MALAT1 [20]. Thus, the linear NEAT1_2 transcript is not strongly depleted by RNase R treatment even with an improved protocol using A-tailing prior to RNase R treatment and a Li+ buffer (S12A Fig) [20]. Nevertheless, we identified a study of circRNAs in Akata cells, in which the linear NEAT1_2 transcript was fully depleted by RNase R treatment in total and cytoplasmic RNA, yet partially resistant in nucleoplasmic RNA, while the hsa_circ_0003812 circRNA was always retained (S12B Fig) [55]. Interestingly, we observed significant enrichment of the circRNA region in our total RNA-seq time-course of WT infection from 6 h p.i. compared to both up- and downstream regions (Fig 3B, S13A Fig, 1.5- to 1.76-fold increase in the DEXSeq analysis, multiple testing adjusted p<0.005). This effect is likely not mediated by vhs since NEAT1_2 is located in the nucleus while vhs is active in the cytosol. Furthermore, some enrichment of the circRNA region was also observed in Δvhs infection by 8 h p.i., consistent with an attenuated progression of Δvhs infection (S13B Fig). Both NEAT1_1 and NEAT1_2 were found to be highly unstable in mouse, with half-lives of ~30 and ~60 min, respectively [56]. Thus, enrichment of the NEAT1_2 circRNA in HSV-1 infection is likely due to its high stability compared to the linear transcript combined with high de novo synthesis of this circRNA.

The following evidence further confirms that hsa_circ_0003812 is a circRNA. First, it was identified by three distinct algorithms for circRNA detection that employ different approaches: CIRI2, circRNA_finder and our aligment-based approach. Combination of two or more different circRNA detection methods has been reported as highly successful in removing false positive predictions [43]. Second, since false positives in circRNA detection largely originate from repetitive sequences, we performed a BLAST search in the human and HSV-1 genomes for the 20 nt on either side of the circular junction. This confirmed that both sequences are unique. Third, we ran our alignment-based circRNA detection approach with a required overlap of at least 40 nt on either side of circular junctions on the WT 4sU-seq and total RNA-seq time-courses. This still identified up to 170 reads for the NEAT1_2 circRNA in total RNA and up to 44 in 4sU-RNA (17–36% of reads identified with a ≥10 nt overlap) and confirmed the strong induction of this circRNA (S14A and S14B Fig). Finally, we implemented a method to determine so-called “confirming read pairs”, i.e., pairs of reads where both reads could be aligned linearly to the genome within a circRNA region but only in crosswise direction that is only consistent with a circRNA but not a linear transcript (S14C Fig). We selected those confirming read pairs for which their genomic distance would imply a fragment size ≥500 nt if they originated from a linear transcript and the fragment size would be smaller if they originated from the circRNA. In this way, we identified up to 45 confirming read pairs per replicate in WT infection for hsa_circ_0003812 (S14B Fig).

NEAT1_2 circular splicing was accompanied by linear splicing both within and downstream of the circRNA region but not within the NEAT1_1 region (Fig 3B and 3C, S13 and S15 Figs). None of these linear splicing events were annotated, but some were already observed at very low rates in uninfected cells. As read counts for the most frequent linear splice sites were comparable or even higher than for the induced hsa_circ_0003812 circular junction, NEAT1_2 linear splicing was also increased in absolute levels. Calculation of splicing rates (= number of junction reads / number of exon-intron reads crossing the corresponding acceptor and donor splice sites) showed that linear splicing was indeed induced and not simply observed more frequently due to the previously reported up-regulation of NEAT1 transcription in HSV-1 infection [29, 31] (Fig 3D). As linear splicing was also induced downstream of identified circRNAs and several of the confirming read pairs for the hsa_circ_0003812 circRNA included linear splice junctions, both linear and circular NEAT1_2 transcripts were further spliced in HSV-1 infection. Canonical GT-AG splice signals were used both for the circular and the most frequent linear splice junctions (see Fig 3E for a diagram) and they tended to be conserved at least among primates (S16 Fig). Since circRNA biogenesis by back-splicing requires the spliceosome [11], our data suggest that circular and linear splicing of NEAT1_2 are linked, with the circRNA originating from back-splicing during NEAT1_2 linear splicing (see also Fig 1A). Splicing rates and their increases compared to mock differed substantially between the different linear splice junctions within the circRNA region (Fig 3D), indicating presence of alternative splicing isoforms of the circRNA. Induction of linear splicing was also observed in 4sU-, chromatin-associated and nucleoplasmic RNA and in ΔICP27 and Δvhs infection (Fig 3C and 3D, S15 Fig), although it was less pronounced in ΔICP27 infection. In summary, both linear and circular splicing of NEAT1_2 is induced during HSV-1 infection while the RNA is still associated with the chromatin. Although ICP27 is not required for NEAT1_2 splicing, it may contribute.

Recently, we also performed total RNA-seq of infection with mock, HSV-1 strain F (WT-F) and its ΔICP22 mutant at 8 and 12 h p.i. (n = 2) [57]. We furthermore generated telomerase-immortalized human foreskin fibroblasts (T-HFs) that express either ICP22 in isolation (T-HFs-ICP22 cells) or in combination with ICP27 (T-HFs-ICP22/ICP27 cells) upon doxycyclin (Dox) exposure. Total RNA-seq of T-HFs-ICP22 and T-HFs-ICP22/ICP27 cells was performed both with and without Dox exposure (n = 2). Ectopic co-expression of ICP22 and ICP27 lead to induction of both circular and linear splicing of NEAT1_2 (Fig 4A). However, ICP22 knockout did not abolish HSV-1-induced NEAT1_2 circular and linear splicing (Fig 4A, S17 Fig), thus ICP22 expression is not required. This also confirmed induction of linear and circular NEAT1_2 splicing in a third HSV-1 strain. Although circRNA reads were recovered after Dox-induced expression of ICP22, this was limited to 2 reads in one replicate (Fig 4B).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 4. Ectopic co-expression of ICP22 and ICP27 is sufficient for induction of NEAT1_2 splicing.

Sashimi plots as in Fig 3 showing read coverage and circular and linear splice junctions in (a) T-HFs-ICP22/ICP27 cells without and with Dox-induced co-expression of ICP22 and ICP27 as well as mock, WT-F and ΔICP22 infection at 12 h p.i., (b) T-HFs-ICP22 cells without and with Dox-induced ICP22 expression and (c) mock and WT KOS infection and ICP27 overexpression. A minimum read count of 4 was again required for (a), while in (b) and (c) all junction reads are shown covering at least 10nt on either side of the junction. For an explanation of sashimi plots, see caption to Fig 3. Circular splice junctions are marked in red, linear splice junctions in the same color as read coverage.

https://doi.org/10.1371/journal.pone.0276467.g004

To investigate whether ICP27 expression alone could induce NEAT1_2 splicing, we analyzed our recently published 4sU-seq data of mock and WT KOS infection and of cells transfected with an ICP27-expressing plasmid [34]. This again confirmed induction of both linear and circular NEAT1_2 splicing in WT KOS infection but recovered only one circular read upon ICP27 overexpression in one replicate (Fig 4C). Notably, NEAT1 was upregulated upon ectopic co-expression of ICP22 and ICP27 (2.2-fold, p<10⁻¹¹) and in ΔICP22 (2.1-fold, p<10⁻⁴) and ΔICP27 (4.9-fold, p<10⁻¹²) infection, but not upon expression of either ICP22 or ICP27 alone. In summary, our results show that neither ICP27 nor ICP22 are required for induction of linear and circular NEAT1_2 splicing but co-expression of both proteins is sufficient. Considering the small number of circular reads recovered when expressing either one of these proteins alone, we can neither confidently confirm nor exclude that either ICP22 or ICP27 alone may be sufficient to (weakly) induce NEAT1_2 circular splicing.

Influenza A virus infection also induces NEAT1_2 splicing

Since NEAT1_2 is also up-regulated in IAV infection [29], we investigated presence of circRNAs in previously published total RNA-seq time-courses of influenza A/California/04/09 (H1N1), A/Wyoming/03/03 (H3N2), and A/Vietnam/1203/04 (H5N1) HALo virus infection (3, 6, 12 and 18 h p.i. plus time-matched controls) of human tracheobronchial epithelial cells (HTBE, multiplicity of infection (MOI) = 5) and monocyte-derived macrophages (MDM, MOI = 2) [41]. The H5N1 HALo mutant virus is an attenuated virus generated from WT influenza A/Vietnam/1203/04. H1N1 infection of HTBE cells included 24 h p.i. and two H5N1 infection time-courses of MDM cells were performed. This revealed a strong enrichment of NEAT1_2 circular splicing in H5N1 infection of both cell types and H3N2 infection of MDM cells as well as a weak enrichment in H3N2 infection of HTBE cells and H1N1 infection of MDM cells (Fig 5A, S18 Fig). The most highly expressed NEAT1_2 circRNA was again hsa_circ_0003812 circRNA as for HSV-1 infection (marked by a blue rectangle in Fig 5A) and its enrichment increased with the duration of infection. Similar trends were observed for linear splicing events, both within and downstream of the circRNA region (Fig 5B). Notably, both the hsa_circ_0003812 circRNA and linear splicing events were also detected in some of the time-matched controls, however normalized circRNA counts and linear splicing rates were significantly lower than in infected cells (with the exception of H1N1 infection, S18A Fig).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 5. Induction of NEAT1_2 circular and linear splicing in IAV infection.

(a) Heatmap of normalized circRNA counts (normalized to number of linear splice junctions mapped to the host genome) for the three most highly expressed NEAT1_2 circRNAs identified in H1N1, H3N2 and H5N1 infection of HTBE and MDM cells or time-matched controls. Columns represent individual circRNAs, which are ordered according to their genomic coordinates from the most 5’ to the most 3’. The NEAT1_2 hsa_circ_0003812 also induced in HSV-1 infection is marked by a blue rectangle. IAV strain and time-points of infection are color-coded on the left (white = time-matched control, cyan = H1N1, green = H3N2, purple = H5N1, darker colors indicate later time-points). (b) Heatmap of linear splicing rates (= number of reads for a linear junction / number of exon-intron reads crossing the corresponding acceptor and donor splice sites) in IAV infection and time-matched controls (dark red = high, white = low, see color bar on the right). Columns represent individual splice junctions, which are ordered according to their genome positions. Counts include only junction and exon-intron reads overlapping the junction or exon-intron boundary, respectively, by at least 10 nt on either side. A pseudocount of 0.01 was used to avoid division by zero. Colors on the left indicate time-points of infection as in (a). Colored rectangles on top indicate whether linear splice junctions were also observed in HSV-1 infection (cyan) or are within (blue) or downstream (gray) of the hsa_circ_0003812 circRNA region. Numbers in brackets at the bottom refer to the introns outlined in the schematic in Fig 3E.

https://doi.org/10.1371/journal.pone.0276467.g005

IAV infection also induces the degradation of host RNAs via two mechanisms. Cap-snatching by the viral RNA-dependent RNA polymerase leads to cleavage of host RNAs downstream of the cap and is required for synthesis of viral mRNAs [58]. The PA-X protein produced by all IAV strains [59] selectively degrades RNAs transcribed by Pol II, including long non-coding RNAs, and is localized and acts predominantly in the nucleus [60]. To exclude that enrichment of the NEAT1_2 circRNA is simply due to degradation of linear transcripts without increased biogenesis of the circRNA, we also investigated overall enrichment of circRNAs in IAV infections using linear regression analysis as for HSV-1 infection. This indeed showed an enrichment of circRNAs by up to 1.4-fold in H1N1 infection, 6.8-fold in H3N2 infection and 4.6-fold for H5N1 infection (S19A–S19G Fig). It also revealed a much higher enrichment for several other circRNAs, such as a circRNA of the ZC3HAV1 (zinc finger CCCH-type containing, antiviral 1) gene. ZC3HAV1 is induced during influenza A/WSN/33 (WSN/H1N1) infection and ectopic expression of ZC3HAV1 inhibits WSN replication [61]. Thus, the ZC3HAV1 circRNA is likely increasingly produced during IAV infection due to up-regulation of ZC3HAV1 transcription and associated increased circRNA biosynthesis and is not only enriched due to linear host RNA degradation. We compared fold-changes in normalized circRNA counts between infection and time-matched controls for 7077 well-expressed circRNAs (normalized circRNA count >1 in at least one condition), which showed that the hsa_circ_0003812 NEAT1_2 circRNA was among the 1% most enriched circRNAs in H3N2 and H5N1 infection of MDM cells (S19H Fig). Occasionally, it was even more enriched than the ZC3HAV1 circRNA. Moreover, splicing rates of linear NEAT1_2 transcripts, which should not escape host mRNA degradation in IAV infection, also increased during infection (Fig 5B). Thus, our results indicate that–at least in H3N2 and H5N1 infection–enrichment of NEAT1_2 circular splicing results from increased biogenesis of the circRNA and not only from IAV-induced degradation of host mRNAs.

Since both HSV-1 and IAV infection disrupt transcription termination, we also evaluated read-through for IAV infection to correlate this to the extent of induction of NEAT1_2 splicing. Paralleling NEAT1_2 induction, read-through was higher in MDM cells than in HTBE cells and higher in H5N1 infection than in H3N2 infection (S20 Fig), while H1N1 infection led to the lowest extent of read-through. Disruption of transcription termination in IAV infection is mediated by the viral NS1 protein via inhibition of the CPSF30 subunit of the cleavage and polyadenylation specificity factor (CPSF) [62]. Our results are consistent with previous reports that the NS1 protein of A/California/04/09 H1N1 virus cannot bind and inhibit CPSF30 [63], while NS1 of the H3N2 and H5N1 strains can [64, 65]. However, a recent report indicated that the NS1-CPSF30 interaction is not necessary for disruption of transcription termination in IAV infection [40]. This could explain why read-through is still observed in H1N1 infection. A similar observation can be made in HSV-1 infection. Although HSV-1-induced disruption of transcription termination is mediated by ICP27 via interaction with CPSF subunits [34], reduced but significant read-through is also observed in ΔICP27 infection likely as a stress response [22, 39]. Due to this correlation between induction of NEAT1_2 splicing and read-through and a recent report that heat stress also up-regulates NEAT1_2 expression [38], we also analyzed our previously published 4sU-seq data of 1 and 2 h salt and heat stress regarding NEAT1_2 splicing. However, neither circular nor linear NEAT1_2 splicing was induced in either of the two stress conditions.

CDK7 inhibition induces NEAT1_2 splicing and disrupts transcription termination in cancer cell lines

To shed additional light on the molecular mechanism(s) governing NEAT1_2 splicing, we searched the recount3 database [66] using Snaptron [67] for presence of the novel linear NEAT1_2 splicing events. Recount3 provides read coverage data for >700,000 publicly available human and mouse RNA-seq samples from the Sequence Read Archive (SRA). Snaptron allows rapidly searching samples in recount3 for (linear) splice junctions within specific genomic regions. Our Snaptron query recovered 6,391 human samples in which at least one of the four most frequent NEAT1_2 linear splice junctions (see Fig 3E) induced in HSV-1 infection was covered by ≥10 reads. This included our HSV-1 RNA-seq experiments, several other studies on HSV-1 infection, the IAV infection time-courses, 62 studies focusing on circRNAs and 21 studies using RNase R (S21A Fig). The latter provide further support for the association of linear and circular splicing of NEAT1_2. Particularly high numbers of NEAT1_2 splice junctions were found in samples of blood cells, such as platelets, erythrocytes, peripheral blood mononuclear cells (PBMCs), erythroleukemia (K562) cells, and more (S21A Fig). Consistently, one of these studies reported that circRNAs are strongly enriched in platelets and erythrocytes [68], which both lack a cell nucleus and thus de novo transcription. In absence of de novo transcription, the high stability of circRNAs leads to their enrichment relative to linear RNAs.

Application of our circRNA alignment pipeline to two RNA-seq experiments for platelets [69] and erythrocytes [70] identified multiple NEAT1_2 circRNAs including the hsa_circ_0003812 circRNA induced by HSV-1 infection (S21B Fig). Interestingly, almost no expression was detected in the ~4 kb between the 3’end of the NEAT1_1 transcript and the 5’end of hsa_circ_0003812, however, expression extended beyond hsa_circ_0003812. This either implied novel linear transcripts beginning near the 5’end of hsa_circ_0003812 or a novel circRNA not included in circBase. Indeed, our de novo circRNA detection pipeline identified a novel circular junction connecting the 5’ end of hsa_circ_0003812 with the 3’end of the downstream expressed region. This confirms that only circular, but not linear, NEAT1_2 transcripts were present in these cells and that these circRNAs are also linearly spliced. It also shows that NEAT1_2 circRNAs naturally occur in uninfected cells, however likely at such low levels that they only become detectable upon degradation of linear transcripts. It should be repeated that NEAT1_2 has been shown to be up-regulated in HSV-1 and IAV infection by RT-qPCR [29], thus enrichment of NEAT1_2 circular splicing in HSV-1 and IAV infection cannot be explained by loss of de novo transcription followed by degradation of linear transcripts. However, this partly explains the presence of NEAT1_2 circular and linear junctions in uninfected cells from the IAV infection time-courses as MDM cells were obtained from blood.

Interestingly, the Snaptron search also identified NEAT1_2 linear splicing after inhibition of CDK7 by THZ1 in multiple cancer cell lines, which was not observed in untreated cells. CDK7 is an essential part of the TFIIH transcription factor complex, phosphorylates Pol II CTD at Ser5 residues and plays key roles in maintenance of promoter-proximal Poll II pausing and the regulation of transcription elongation [71]. Overexpression of CDK7 has been associated with a number of cancers and correlated to poor prognosis [71]. THZ1 is a highly selective covalent inhibitor of CDK7 with anti-tumor activity against multiple different types of cancer [72], but also inhibits CDK12 and CDK13 at higher concentrations [73]. Experiments identified with the Snaptron search cover prostate cancer (VCaP, LNCaP and DU145 cells) [42], bladder cancer (HCV-29 cells) [74], esophageal squamous cell carcinoma (TE7 and KYSE510 cells) [75], nasopharyngeal carcinoma (C666-1, HK1 and HNE1 cells) [76], pancreatic ductal adenocarcinoma (BxPC3, MiaPaCa-2 and PANC1 cells) [77] and chordoma (UM-Chor1 cells) [78]. Further literature search identified a second study of chordoma cells (UM-Chor1 and CH22) [79] with up to 24 h THZ1 treatment and a study on THZ1 treatment of B-cell acute lymphocytic leukemia cells (Nalm6) [80]. Realignment of corresponding samples confirmed linear splicing upon THZ1 treatment in all cell lines apart from DU145 and showed that splicing rates increased with the duration and dosage of THZ1 treatment (Fig 6A and 6B and S22A Fig).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 6. NEAT1_2 splicing and read-through upon CDK7 inhibition.

(a) Heatmap of linear splicing rates (= number of reads for a linear junction / number of exon-intron reads crossing the corresponding acceptor and donor splice sites) upon treatment by DMSO/control, THZ1 or Act-D or knockdown (KD) of CDK7 or MED1 (dark red = high, white = low, see color bar on the right). Columns represent individual splice junctions, which are ordered according to their genome positions. Counts include only junction and exon-intron reads overlapping the junction or exon-intron boundary, respectively, by at least 10 nt on either side. A pseudocount of 0.01 was used to avoid division by zero. Colored rectangles on top indicate whether linear splice junctions were also observed in HSV-1 (cyan) or IAV (purple) infection or are within (blue) or downstream (gray) of the hsa_circ_0003812 circRNA region. Numbers in brackets at the bottom refer to the introns outlined in the schematic in Fig 3E. (b, c) Sashimi plots as in Fig 3 showing read coverage and linear splice junctions in (b) UM-Chor1 cells after 4, 8, 12 and 24 h THZ1 or control treatment (3 replicates each) and (c) CDK7 and MED1 knockdown in VCaP and LNCaP cells (2 replicates each). All junctions are shown that are covered by at least 4 reads aligning to at least 10 nt on either side of the junction. (d) Boxplots showing distribution of %read-through (for calculation see methods) after THZ1 treatment in example studies or CDK7 or MED1 knockdown in VCaP and LNCaP cells.

https://doi.org/10.1371/journal.pone.0276467.g006

Surprisingly, considering the previously observed strong association between linear and circular NEAT1_2 splicing, no NEAT1_2 circRNAs were recovered for any of these samples neither with our alignment-based nor our de novo circRNA discovery pipeline. However, in all cases almost no circRNAs (maximum 105 circRNAs, mean 18) were recovered at all, not even the most ubiquitous circRNAs such as HIPK3, CORO1C or ASXL1. As read length was ≥75 nt for all samples, this cannot be explained by short read length. It rather suggests that circRNAs were generally not or only poorly recovered in all these studies. While a literature search identified no other studies for specific CDK7 inhibition without poly(A) selection, the Snaptron search also recovered linear NEAT1_2 splicing in chromatin-associated RNA of HEK293 cells upon 2 h treatment with DRB [81]. DRB inhibits CDK7 [82] but also the CDK9 kinase component of the positive P-TEFb complex [83] and other kinases [84]. The hsa_circ_0003812 NEAT1_2 circRNA induced by HSV-1 and IAV infection was indeed observed upon DRB treatment, but not any of the other NEAT1_2 circRNAs present in platelets and erythrocytes (S22B Fig). This provides some evidence that CDK7 inhibition leads to NEAT1_2 circular splicing.

In any case, identification of linear splicing in absence of circRNAs shows that the linear NEAT1_2 isoform is spliced upon CDK7 inhibition. As no circRNAs were recovered, induction of linear splicing cannot be an artefact of circRNAs being enriched after THZ1-mediated transcription inhibition. Furthermore, spliced linear NEAT1_2 transcripts should not be more or less stable than unspliced NEAT1_2 transcripts as splicing occurs upstream of the last 100 nt of NEAT1_2 that form the stabilizing helical structure. Interestingly, however, CDK7 inhibition led to a relative increase of NEAT1_2 compared to NEAT1_1, suggesting that NEAT1_2 transcription was less strongly reduced by CDK7 inhibition than NEAT1_2 transcription or potentially even increased. To exclude that inhibition of transcription per se leads to NEAT1_2 splicing, we investigated data from one of the UM-Chor1 experiments that included actinomycin D (Act-D) treatment, which inhibits transcription but not via inhibition of cyclin-dependent kinases (CDKs) [85]. Act-D treatment, while leading to a similar relative increase in the abundance of NEAT1_2 compared to NEAT1_1 as THZ1 treatment, did not induce NEAT1_2 splicing (S22A Fig). In addition, siRNA-mediated knockdown of CDK7 in VCaP and LNCaP cells also led to NEAT1_2 splicing and a relative increase in NEAT1_2 compared to NEAT1_1 (Fig 6C), including many more novel splicing events. In contrast to splicing observed in HSV-1 and IAV infection and upon THZ1 treatment, some of these splicing events extended into the NEAT1_1 region. This indicates that indeed inhibition of CDK7, and not of other CDKs, leads to NEAT1_2 splicing. Interestingly, siRNA-mediated knockdown of the MED1 component of the Mediator complex in VCaP and LNCaP cells had the same effect (Fig 6C). CDK7 directly phosphorylates MED1 at Threonine 1457 [42], which together with a phosphorylation at Threonine 1032 promotes MED1 association with the Mediator complex [86]. Mediator is responsible for communicating regulatory signals from gene-specific transcription factors to Pol II and in this way impacts transcription at multiple stages [87]. Our results suggest that the effect of CDK7 inhibition on NEAT1_2 splicing may follow from the loss of MED1 phosphorylation and thus loss of MED1 association with Mediator.

Interestingly, CDK7 inhibition also induced read-through transcription, albeit to a lesser degree than HSV-1 or H5N1 infection (Fig 6D). Read-through transcription after THZ1 treatment was comparable to H1N1 or H3N2 infection of HTBE cells, most pronounced in THZ1-treated Nalm6 cells and increased with the duration and dosage of THZ1 treatment in UM-Chor1 cells. In addition, both CDK7 and MED1 knockdown induced read-through at levels comparable to THZ1-treated Nalm6 cells. Moreover, THZ1 treatment also resulted in extended antisense transcription at promoters. In summary, our results show that CDK7 inhibition impacts RNA transcription and processing at multiple levels.

RNA-seq data

All RNA-seq data analyzed in this study were downloaded from the SRA. SRA project IDs: HSV-1 WT infection time-course: SRP044766; HSV-1 Δvhs infection time-course: SRP192356; RNA-seq of subcellular fractions: SRP110623, SRP189489, SRP191795 (same experiment but data for mutant viruses submitted separately); total RNA-seq for mock and WT KOS infection with ERCC spike-ins: SRP321121 (samples: SRR14632002, SRR14631995); T-HFs-ICP22 and T-HFs-ICP22/ICP27 cells, WT strain F and ΔICP22 infection: SRP340110; WT KOS, ΔICP27 and ICP27 overexpression: SRP189262; IAV infection time-courses: SRP091886, SRP103821; platelets: ERP003815; erythrocytes: SRP050333; CDK7 inhibition/knockdown: VCaP, LNCaP and DU145 cells: SRP179971; HCV-29 cells: SRP217721; TE7 and KYSE510 cells: SRP068450; C666-1, HK1 and HNE1 cells: SRP101458; BxPC3, MiaPaCa-2 and PANC1 cells: SRP165924; UM-Chor1 and CH22 cells: SRP166943, SRP270819; Nalm6 cells: SRP307127; DRB treatment of HEK293 cells: SRP055770. RNase R treatment: SRP197110, SRP152310.

Linear RNA-seq read mapping

Sequencing reads were downloaded from SRA using the sratoolkit version 2.10.8 and aligned against the human genome (GRCh37/hg19) and human rRNA sequences using ContextMap2 version 2.7.9 [105] using BWA as short read aligner [50] and allowing a maximum indel size of 3 and at most 5 mismatches. For HSV-1 infection RNA-seq data, alignment also included the HSV-1 genome (Human herpesvirus 1 strain 17, GenBank accession code: JN555585). For the two repeat regions in the HSV-1 genome, only one copy was retained each, excluding nucleotides 1–9,213 and 145,590–152,222 from the alignment. SAM output files were converted to BAM files using samtools [106]. Following read mapping, fragment size distribution was determined from BAM files using the Picard CollectInsertSizeMetrics tool [107]. The linear RNA-seq mapping was used for identification of linear splice sites, filtering of identified circRNA reads (see below) and visualization of read coverage on genes.

circRNA de novo detection

For de novo detection of circRNAs, we applied CIRI2 [45] and circRNA_finder [44] in parallel to all reads for each sample (including both reads aligned and unaligned in the linear RNA-seq mapping) (outline in Fig 2A). Combination of two complementary circRNA detection algorithms, such as CIRI2 and circRNA_finder, has been recommended to remove algorithm-specific false positives commonly observed in circRNA detection [43]. Both CIRI2 and circRNA_finder allow de novo detection of circRNAs using splice sites not annotated in the human genome. CIRI2 is based on identifying so-called paired chiastic clipping (PCC) signals in BWA alignments, which are pairs of clipped read alignments (i.e., local alignments for different substrings of the read) where a downstream part of the read aligns upstream of an upstream part of the read. Such reads are denoted as back-spliced junction (BSJ) reads and are then further filtered by CIRI2 to remove false positives (see Fig 2A and the original publication for details). For this purpose, we first aligned all reads using BWA with parameters recommended by the developers of CIRI2 (= default BWA parameters, except for the -T option (= minimum score to output an alignment), which was set to 19). With default BWA options, matches in alignments are given a score of 1, mismatches a penalty of 4, the gap open penalty was 6 and the gap extend penalty was 1. As a consequence, mismatches on read segments flanking each side of the circular junction are only allowed if corresponding read segments are ≥ 24 nt.

circRNA_finder is based on filtering chimeric alignments obtained with STAR [49], in which different parts of a read align to two distinct regions of the genome in a manner not consistent with “normal” linear transcripts. We thus first aligned all reads with STAR using the default parameters used by the developers of circRNA_finder, i.e., chimSegmentMin (minimum total length of the chimeric segment) = 20, chimScoreMin (minimum total score of the chimeric segments) = 1, alignIntronMax (maximum intron length) = 100000, outFilterMismatchNmax (maximum number of mismatches per pair) = 4, alignTranscriptsPerReadNmax (maximum number of different alignments per read to consider) = 100000, outFilterMultimapNmax (maximum number of loci the read is allowed to map to) = 2. circRNA_finder then further filters chimeric alignments to identify circRNAs and remove false positives (see Fig 2A and the original publication for details).

CircRNAs were only further analyzed if they were detected independently by both algorithms. Circular reads identified by either algorithm were then pooled and subsequently all reads were removed that could be mapped in a linear fashion to the genome using ContextMap2. Only circRNAs with at least two supporting reads remaining in the same sample were further analyzed. Fragment lengths were not explicitly considered for circRNA detection, but all used algorithms have different criteria regarding the allowed distances and positions of the two reads in a read pair relative to each other (see original publications for details). The complete pipeline is available as a workflow for the workflow management system Watchdog [108] in the Watchdog workflow repository (https://github.com/watchdog-wms/watchdog-wms-workflows/tree/master/circRNA_Detection).

Alignment-based circRNA detection

Candidate circular junction sequences were generated from putative spliced circRNAs sequences downloaded from circBase [48]. RNA-seq reads were aligned against circular junction sequences using BWA [50] with default parameters (i.e., -T is set to 30). BWA clips a read, i.e., outputs only a local alignment of the read, if the best local alignment score of the read minus a clipping penalty (default = 5) is at least as good as the best global alignment score. Thus, if a read only aligns on one side of the circular junction, only a local alignment to that side of the junction will be generated by BWA with at most a very short overlap to the other side of the junction with few mismatches. Reads were retained if they aligned to at least X nt (default X = 10) on either side of the junction and could not be mapped in a linear fashion anywhere to the genome using ContextMap2. In case of paired-end sequencing, the second read in the pair also had to align consistently within the circRNA region. circRNAs resulting from repetitive regions were discarded. The complete pipeline is outlined in S4 Fig.

Identification of confirming read pairs

ContextMap2 is also capable of aligning read pairs, where each individual read can be aligned linearly to the genome, but the two reads are aligned in the wrong orientation relative to each other (= crosswise, see S14C Fig). Such crosswise alignments of read pairs are inconsistent with linear transcripts and were considered a confirming read pair for a circRNA if (i) both reads were aligned within the circRNA region with a genomic distance implying a fragment size ≥ 500 nt if they originated from a linear transcript and (ii) the fragment size would be smaller if they originated from the circRNA. This fragment size cutoff was based on the observed distribution of fragment sizes, which showed that <3.7% of fragments exhibited a fragment size ≥ 500 nt.

Read-through calculation

Number of read fragments per gene and in 5kb windows downstream of genes were determined from read alignments using featureCounts [109] and gene annotations from Ensembl (version 87 for GRCh37/hg19) [110]. For strand-specific RNA-seq, the stranded mode of featureCounts was used. All fragments (read pairs for paired-end sequencing or reads for single-end sequencing) overlapping exonic regions on the corresponding strand by ≥25 nt were counted for the corresponding gene. Downstream transcription for a gene was calculated as previously described [22] as the FPKM (Fragments Per Kilobase Million) in the 5kb windows downstream of genes divided by the gene FPKM multiplied by 100. %Read-through transcription was quantified as the difference in downstream transcription between infected/treated and uninfected/control cells, with negative values set to zero.

Other bioinformatics analyses

Sashimi plots were generated with ggsashimi [111]. Heatmaps were generated with the heatmap.2 and pheatmap functions in R. Differential exon usage analysis was performed with the R Bioconductor package DEXSeq [112]. Log2 fold-changes in gene expression for WT and Δvhs infection time-courses were obtained from our recent study [18]. In this study, we used DESeq2 [113] to compare gene expression between mock and each time-point of infection for 4,162 genes without read-in transcription originating from disrupted transcription termination for an upstream gene. Genes with read-in transcription were excluded from the analysis as read-in transcription can be mistaken for induction of gene expression.