A comprehensive analysis of cell-type specific nuclear RNA from neurons and glia of the brain

Abstract

Background

Studies in psychiatric genetics have identified over 100 loci associated with disease risk, yet many of these loci are distant from protein coding genes. Recent characterization of the transcriptional landscape of cell lines and whole tissues has suggested widespread transcription in both coding and non-coding regions of the genome, including differential expression from loci that produce regulatory non-coding RNAs which function within the nucleus; however, the nuclear transcriptome of specific cell types in the brain has not been previously investigated.

Methods

Here we have defined the nuclear transcriptional landscape of the three major cellular divisions of the nervous system using flow sorting of genetically labeled nuclei from bacTRAP mouse lines. This was followed by characterization of the unique expression of coding, non-coding and intergenic RNAs in the mature mouse brain with RNAseq and validation with independent methods.

Results

Our findings reveal diverse expression across the cell-types of all classes of RNAs, including long non-coding RNAs – several of which were confirmed as highly enriched in the nuclei of specific cell-types using anatomical methods. Finally, we also discovered several examples of cell-type specific expression of tandem gene fusions, and report the first cell-type specific expression of circular RNAs, notably a neuron-specific and nuclear-enriched RNA arising from the gene Hnrnpu.

Conclusion

These data will provide an important resource for studies evaluating the function of a variety of ncRNAs in the brain, including those that may play a role in psychiatric disease.

Introduction

The CNS has remarkable cellular diversity with hundreds of distinct cell-types, each with unique morphology, connectivity, and function. However, all of this diversity must arise from essentially identical copies of the genome in every cell, and recent work has highlighted the substantial range of epigenetic and transcriptional variation across cultured cells and whole tissues in the body(1). Likewise, gene expression plays a key role in defining the cellular diversity of the brain, and previous work demonstrated significant differences in transcript abundance for protein-coding genes(2–4).

It is now recognized that transcription also occurs beyond protein-coding genes(5). Furthermore, the nucleus has a clearly distinct RNA profile relative to the cytoplasm, suggesting there may be a number of additional RNA molecules with roles particularly within this subcellular compartment(5, 6). However, previous work focused largely on cell lines, and the CNS remained relatively unstudied.

In the last decade there has also been a substantial expansion of the known roles for non-coding RNAs(ncRNAs) and many new species have been defined, including the long intergenic non-coding RNAs(lincRNAs)(7–10). Work in embryonic stem(ES) cells suggests that many lincRNAs localize to the nucleus and regulate cell fate decisions: 38% bind known chromatin remodeling complexes(11), and experimental overexpression or suppression of ~100 lincRNAs caused changes in gene expression and cell fate in vitro(12). Some lincRNAs have been identified as important for neural stem cell differentiation in primary cultures(13). In addition, recent studies have highlighted the prevalence of a new class of circular non-coding RNAs(ciRNAs)(14, 15), at least one of which has a role in microRNA regulation(16, 17).

The major cell-types of the nervous system – neurons, astrocytes, and oligodendrocytes – have distinct mRNA profiles in vivo(2, 3). These profiles have subserved a wide range of informative secondary analyses(e.g.(18–20)). Yet, little is known about the nuclear transcriptome of these cells. Also, given a nuclear role for some lincRNAs, and generally low expression levels, analysis of nuclear transcriptomes might provide better assessment of ncRNAs. The current study describes a powerful method to study cell-type specific nuclear RNA in vivo. We demonstrate a diversity of nuclear transcripts, replicating previous analyses of protein-coding genes and greatly extending them to include a thorough description of nuclear non-coding RNA, ciRNA, and other RNA species. We provide an overview of the data, describe differences in protein-coding and lincRNA transcription across these three cell-types, and identify novel tandem gene fusions and circular RNAs detectable in each.

Materials and Methods

Mice

All mouse procedures were approved by the Washington University ASC. Lines included B6.FVB-Tg(Snap25-EGfp/Rpl10a)^JD362Htz/J, B6.FVB-Tg(Aldh1L1-EGfp/Rpl10a)^JD130Htz/J, and B6.FVB-Tg(Cnp-EGfp/Rpl10a)^JD368Htz/J.

Nuclear Capture and RNA purification

Nuclear sorting was conducted as described(21). We processed four animals from each line(6–10 weeks, 2 of each sex) as independent replicates. Gates were set for GFP using nuclei from wildtype mice. Nuclei were sorted with Beckman Coulter MoFlo directly into Trizol LS for purification, followed by RNeasy Minelute columns. Quality and quantity were confirmed with Bioanalyzer and qPCR.

Library Preparation

We generated double stranded cDNA using Nugen Ovation RNAseq system V2, starting from 5 ng of RNA. Standard Illumina sequencing libraries were generated from 1–2 ug of cDNA, sheared to ~200 bp.

Sequencing analysis

Reads were trimmed with Trimmomatic(22). rRNA reads were removed by mapping to rRNA sequences using Bowtie2(23). Remaining reads were mapped to Ensembl version 75 using STAR(24), counted with BEDTools(25), and tested by edgeR(26).

Circular RNA

A custom genome was created by extracting every gene +/− 1kb from repeat masked Ensembl 75 mouse genome. Each gene sequence was duplicated in tandem and treated as a separate chromosome. Cleaned reads were mapped using STAR with additional parameters: mapping 90% of read length, canonical splice junctions only, and no maximum intron size. An additional filtering removed reads that mapped twice to a ‘chromosome’(unspliced), lacking a splice junction spanning the duplicate gene copies, or where splicing was linear.

Gene Ontologies

The mRNAs with FDR<.05 and fold change > 10 were ranked by CPM and the 25 most highly expressed from each cell-type were analyzed with BINGO(v2.44) in Cytoscape(v2.8) for Biological Processes over-represented at p<1e-4 with B-H correction. Results using longer lists were substantially similar.

qPCR validation

Reverse transcription was performed using qScript cDNA Supermix. cDNA was amplified with a ViiA 7 Real-Time PCR system using comparative CT protocol and Qantas 2X mastermix. For circular RNA, DNase1-treated total RNA was incubated with or without RNase R, and purified by phenol-chloroform extraction. Reverse transcription was performed using either random hexamers or oligodT, and Superscript III.

In Situ Hybridization

Probe templates were amplified by PCR from cDNA. cRNA probes were generated using Digoxigenin RNA Labeling Mix. Brains were cryosectioned after 4% paraformaldehyde perfusion, post-fixed in 4% paraformaldehyde, and prehybridized for 1 hour. Probes were hybridized overnight at 65°C, and washed for stringency, then detected with sheep anti-Digoxigenin and developed using Cy3 Tyramide Signal Amplification kit, followed by anti-Gfp(Aves, GFP-1020) and mouse anti-CNP1(Millipore, MAB326), detected with appropriate Alexa-dye labeled secondaries. Slides were imaged using an UltraView Vox spinning-disk confocal microscope.

Results

Isolation of nuclear RNA from neurons, astrocytes, and oligodendrocytes

We previously generated and validated bacTRAP lines expressing Gfp-ribosomal fusion proteins to capture mRNA from neurons(Snap25-eGfp-Rpl10a), astrocytes(Aldh1L1-eGfp-Rpl10a), and oligodendrocytes(Cnp-eGfp-Rpl10a) in vivo(2, 27), as outlined in Figure 1A. Because ribosomes are assembled in the nucleolus, bacTRAP lines have a non-diffusible Gfp signal within the nucleus(Figure 1B,C), facilitating its purification with FACS.(21) Here, we adapted this to study nuclear-localized coding(mRNA) and non-coding(ncRNA) transcripts, in vivo, in a cell-type specific manner.

Figure 1. Purification of nuclear RNA from targeted cell-types in the brain.

A) Method for identifying cell-specific nuclear RNA. Brains from adult bacTRAP mice are gently homogenized and a nuclear fraction is prepared by density fractionation. Individual Gfp positive nuclei are purified by FACS, and RNA is extracted and sequenced. B) Neurons from neuronal bacTRAP mice contain Gfp+ nucleoli(red arrow) within the nucleus(white, nuclear label). Scale bars, 5 uM. C) Gfp+ nucleolus also permits purification of astrocyte specific nuclei by FACS(blue, nuclear label). D) qPCR from an astrocyte mouse nuclear RNA fraction for two known nuclear non-coding RNAs(U2 and U6) demonstrate enrichment in nuclei fractions(−ΔΔCt method [28], normalized to Actb mRNA and homogenate, means +/− SD of technical replicates). E) Gfp+ Nuclei flow sorted from astrocyte bacTRAP mouse are enriched in mRNA for two known protein-coding markers of astrocytes(−ΔΔCt method, normalized to Actb mRNA and presorted nuclei, means +/− SD of technical replicates). F) Graphical representation of significant biological processes over-represented in astrocytes. Detected by the BINGO software at p<1e-4, with Benjamini-Hochberg multiple testing correction. Only significant nodes or parents of significant nodes are shown.

We first tested this approach with astrocytes. RNA Bioanalyzer confirmed that RNA from the sorted nuclei was intact, and qPCR confirmed the presence of both ncRNA and mRNA(Figure 1D,E). As expected, ncRNAs with known roles in the nucleus were clearly enriched in nuclei compared with total brain homogenate(1D), and mRNAs known to be translated in astrocytes were already enriched in their sorted nuclei(1E).

We then repeated this procedure with four animals from each line, confirmed sort quality with fluorescence microscopy(not shown), RNA integrity with RNA Bioanalyzer, and specificity via qPCR for mRNAs known to be enriched in each cell-type(Figure S1 in Supplement 1). Finally, we generated RNAseq libraries from Gfp+ nuclei, presorted nuclei, and cytoplasmic controls.

Examining the transcriptional profiles with multidimensional scaling confirms reproducibility across replicates(Figure S2A in Supplement 1). We also conducted pairwise comparisons for each cell-type and examined mRNAs previously identified with enriched ribosome-binding in each cell-type in the cortex(27)(Figure S2B in Supplement 1). Most of the previously identified mRNAs were already enriched in the nuclei, thus many of the previously seen differences in translation are likely mediated by transcription. The 82% of genes previously identified as oligodendrocyte-specific by TRAP were enriched >2-fold in oligodendrocyte nuclei compared to astrocyte nuclei. The small fraction(1.19%) that are substantially discordant(>2 fold enriched in astrocytes) could reflect either technical differences between the studies(e.g. whole brain vs. cortex, RNAseq vs. microarray), or potentially interesting biology: a subset of these might be substantially regulated post-transcriptionally. Nonetheless, this analysis confirmed reproducible and accurate measurement the nuclear transcriptome of these cell-types.

Compared with total brain cytoplasmic RNA, all nuclear RNA samples demonstrated enrichment of intronic reads, indicating a substantial proportion of pre-spliced primary transcripts(Figure S3A in Supplement 1), as expected. Excluding rRNA, each of the cell-types devoted >95% of their transcriptional resources to the generation of mRNA(Figure S3B in Supplement 1). Of the remaining reads, >75% of both cytoplasmic and nuclear RNA are devoted to lincRNAs, with minor contributions from small nuclear and nucleolar RNAs(Figure S3C,D in Supplement 1). This overview establishes that our analyses will be able to identify cell-type specific differences in nuclear expression of both mRNA and ncRNA.

The protein-coding nuclear transcriptome of CNS cell-types

The protein-coding transcriptome of these three cell-types has been evaluated in adulthood(27) and during development(3, 29). Of annotated protein-coding genes, 11,097 were transcribed at >5 counts per million(CPM) in at least one of the three cell-types. Of these, 7,878 were differentially expressed and 1,693 had a ‘fold-change’ magnitude of >10 in one or more pairwise comparisons(Table S1 in Supplement 2). Our mRNA results are largely consistent with our previous microarray studies; however, RNAseq provides better absolute quantification of expression across genes and a greater dynamic range(30), which permits us to investigate the most highly expressed transcripts in each cell-type.

Amongst the 1,693, the most highly-expressed protein-coding gene in astrocytes, accounting for 0.7% of all transcripts in the nucleus, is the glutamate transporter Glt1(Slc1a2). A close second is the Alzheimer’s gene ApoE at 0.4%. Consistent with the role of astrocytes as master regulators of the extracellular environment, nine of the twenty-five most highly-expressed genes(each >0.1%) are transporters for the neurotransmitters glutamate(Slc1a2, Slc1a3) and GABA(Slc6a1, Slc6a11), or pumps/channels for ions(Atp1a2, Atp1b2, Slc4a4, Ttyh1). Two others are regulators of the extracellular matrix(Clu, Cst3), and several are receptors(Ptprz1, Ntrk2, Gpr37l1, Fgfr3). The categorical enrichment of these biological processes is highly significant via Gene Ontologies(GO) analysis(Figure 1F; Table S2 in Supplement 1).

For neurons, there were fewer transcripts above the 0.1% level, perhaps reflecting the greater heterogeneity of this cell-type. In the top 25, there was a significant over-representation of transcripts previously identified as highly abundant in cerebellar granule neurons(p<.005 using our cell-specific expression analysis tool(19)). This likely reflects the abundance of this cell-type as there are more cerebellar granule neurons in the brain than all other types of neurons combined(31). Nonetheless, markers of other neuron types are represented in the data too(e.g. GABAergic marker Gad1, 32 CPM), indicating our results represent an averaging of expression across all neuronal nuclei. GO analysis(Figure S4A and Table S3 in Supplement 1) reveals a significant over-representation of synaptic and membrane localized proteins, particularly ion channels and receptors(Grin1, Scn1a, Scn2a1), and cell-cell interaction molecules(e.g. Cbln1, Cbln3, Syt1, Fat2). For oligodendrocytes, the most highly and differentially expressed transcripts were significantly over-represented in categories related to generating fatty acids(Scd2, Ptgds, Enpp2) and myelination(Plp1, Mal, Mbp, Qk, Mobp)(Figure S4B and Table S4 in Supplement 1), with the latter accounting for ~2% of transcription.

The non-coding nuclear transcriptome of CNS cell types

Current annotation indicates that only ~2% of the genome contributes to mRNA(32), yet a larger fraction is certainly transcribed. In the CNS, ncRNAs are suspected to play diverse roles in synaptic function, brain development, neural plasticity and repair(33), and some ncRNA producing loci have been identified in human genetic studies, though they are less frequently annotated than protein-coding genes.(34) Indeed, the majority of lincRNA transcripts have little or no functional annotation. For example, only 17 of the highest 50 expressed in our dataset have functional studies associated with them.

LincRNAs were expressed at a lower level than mRNAs, and were more likely to be enriched in the nucleus(Table S5 in Supplement 2), consistent with putative roles in regulation of gene expression. Among these transcripts, ~300 were detected at a level of >5 CPM in one of the three cell-types and most were transcribed at a level of >1 CPM(i.e. at least 7–10 counts per replicate). The most highly expressed overall was Malat1(>1% of nuclear transcription), the deletion of which is known to cause subtle deficits in neural functioning(35). Several other highly expressed transcripts were also found in all three cell-types. For example, Xist and its regulators Ftx(36) and Jpx(37) were robustly detectable, as was the non-coding oncogene Pvt1(38). By contrast, 169 lincRNAs had a >10-fold difference in one of the pairwise comparisons(FDR of 0.05, Table S3 in Supplement 1). For example, Mirg and the other adjacent maternally expressed ncRNAs(Meg3, Rian) from the Dlk1-Dio3 imprinted cluster are highly enriched in neuronal nuclei. Meg3 and Mirg both overlap with microRNAs, and Meg3 in particular is thought to be regulated by Dlk1 and may serve as a tumor suppressor in various tissues(39, 40). All three have previously been detected in the embryonic brain under the control of proneural genes(41), and here we have confirmed their continued expression into adulthood.

To validate and extend our analysis of these three genes, we examined their distribution in adult brain using in situ hybridization(ISH). Overall, Meg3 showed robust expression in gray matter where neurons are found(Figure 2A), and depletion from white matter areas, where only glia are found. Meg3 showed selective neuronal expression, as demonstrated by co-labeling with neurons(anti-Gfp in Snap25-eGfp-Rpl10a brains, Figure 2B). Additionally, we co-labeled these sections for oligodendrocyte precursors(anti-NG2, Figure 2B), astrocytes(anti-Gfp in Aldh1L1-eGfp-Rpl10a brains, Figure 2C), oligodendrocytes(anti-CNP, Figure 2D). Meg3 did not show apparent expression in these cell-types. Furthermore, unlike typical mRNA ISH where robust signal is found in cytoplasm, Meg3 RNA was almost entirely within the nucleus(Figure 3A,B). Likewise, Rian showed a similar ISH pattern(not shown), suggesting a nuclear role for both. Mirg1 showed even more precise localization within bright subnuclear puncta in neurons(Figure 3C,D). This localization suggests that Mirg1 may have a specific role within a unique subnuclear compartment or perhaps is bound to a particular region of the genome.

Figure 2. Identification of nuclear non-coding RNA gene expression in neurons.

A) Low(10×) magnification confocal fluorescent images showing overall Meg3 in situ hybridization(ISH) expression pattern is consistent with expression in neurons – there is robust expression of Meg3 in both cortex and striatum and lack of Meg3 positive cells in the corpus callosum(dashed lines), a Cnp positive white matter track that contains glial but not neuronal nuclei. B) High magnification(63×) Meg3 ISH and anti-Gfp labeling in an neuronal(Snap25-eGfp-Rpl10a) bacTRAP mouse cortex shows Meg3 is present in the nuclei of neurons(green arrows). Anti-NG2 staining(white) shows absence of Meg3 in oligodendrocyte-precursor cells(white arrows). C) Similarly, Meg3 ISH and anti-Gfp labeling in an astrocyte(Aldh1L1-eGfp-Rpl10a) bacTRAP mouse cortex shows Meg3 is absent from the nuclei of astrocytes(green arrows). D) Likewise, anti-CNP, which labels both oligodendrocyte-myelinated axons(fibers) and the cytoplasm of oligodendrocytes themselves(yellow arrows), demonstrates Meg3 is absent from the nuclei of these cells. E) Mirg RNA ISH is strongly localized to discrete sub-nuclear puncta within most presumptive neuronal nuclei, and is largely absent from glial nuclei(green arrow, astrocyte). Neuronal-like labeling by all three ncRNAs was confirmed in multiple brain regions beyond the cortex, including hippocampus, striatum, midbrain & brainstem.

Figure 3. Mirg1 is localized specifically in a subnuclear compartment of neurons.

A) ISH for the mRNA Gad, a marker of interneurons, shows robust signal in the cytoplasm of sparse neurons in cortex, surrounding the DAPI stained nucleus. B) ISH for Meg3 is found largely within the nucleus, as is Mirg1(C); however, it is typically localized to very discrete puncta. Quantifying these across 35 confocally reconstructed neuronal nuclei revealed an average 1.7 puncta per nucleus(range 1–3). This number is consistent with the number of nucleoli in neurons. However,(D) Mirg1 ISH in the Snap25-eGfp-Rpl10a confirms these puncta are found in neurons but also demonstrates that Mirg1RNA does not localize to the eGfp-Rpl10a-labeled nucleoli.

In astrocytes, our sequencing results similarly indicate a variety of transcripts, though most do not have known functions. The notable exception is the ncRNA Rmst, recently reported to be a key co-regulator of neurogenesis with the Sox2 transcription factor(42). Consistent with this, we see robust expression of this transcript in the nuclei of astrocytes, a subset of which serve as neural stem cells in the CNS(43, 44)(Table S5 in Supplement 2). In oligodendrocytes, there were at least 16 ncRNAs with CPM of greater than 5; among those annotated, both dLeu2 and Neat1 were highly expressed(Table S5 in Supplement 2). Dleu2 has been described in the context of B-cell chronic lymphocytic leukemia as a potential tumor suppressor(45), and our data suggest it may serve an anti-proliferative or pro-differentiation function in oligodendrocytes. Neat1 shows highest expression in oligodendrocytes but is also present in astrocytes. Thus we have identified a variety of differentially expressed ncRNAs in the nuclei of CNS cell-types.

Detection of tandem gene fusions

Some reads in our analysis were not consistent with the expression of known transcripts. Therefore, we tested for the presence of two additional categories of mRNA and ncRNA species: tandem gene fusions and ciRNAs. Overall, we identified 94,435 splice junctions, of which 0.4% were novel splicing events between known exons within a gene. We also saw clear evidence for 9,591 novel splicing events, many involving previously unannotated exons detected only in specific cell-types.

Interestingly, we also detected splicing from an annotated exon of one gene into an exon of a nearby gene, suggesting a novel fusion of two adjacent genes into a single transcript(Figure 4), a type of event previously described as a tandem gene fusion(46, 47). This includes a glial-specific fusion of the Metrn and Fam173a genes(Figure 4A), and a fusion of the N-terminal ~20 amino acids of Cx3cl1 to the body of Ccl17 in neurons(Figure 4B). We manually curated those with >5 reads(Figure 4C). One was an artifact mediated by repetitive protein sequence, and two appeared to be novel unannotated 5’ exons for known genes that fall in the introns of adjacent genes. However, the remaining 17 events appear to represent transcriptional fusions of adjacent genes. Of these, at least seven could maintain coding frame for one exon, and five could continue to the end of the second transcript and thus are likely to produce novel fusion proteins or modify the N-terminal motifs of particular proteins in a cell specific manner. Thus, for at least these events, adjacent genes could be merged into a single gene(consistent with the definition of a gene by their ability to produce a single protein product(48)), and indeed three of our 20 candidate tandem fusions of Ensembl genes are annotated as a single gene in alternative databases. However, if the other fusion events continue to be seen only in a specific cell-type, it would suggest that the definition of a gene model may be unique to the cell-type it is expressed in.

Figure 4. The discovery of tandem gene fusions in specific CNS cell-types.

A) Novel splice-junction spanning reads were detected from exon 1 to exon 3 within the Fam173a gene(green line) and from exon 3 of the Metrn gene into exons 2 and 3 of the Fam173a gene(red lines) in astrocyte RNA. One of these Metrn-Fam173a splicing events is predicted to produce an in-frame protein fusion. B) Novel splice-junction spanning reads were detected from exons 1 and 2 of Cx3cl1 to exon 2 Ccl17, producing an in-frame fusion in neurons. C) Table of the top 20 candidate tandem-fusions, selected by read depth. Each was manually evaluated to determine if it would produce a fusion of the canonical reading frames of the two proteins(ORF=Y), or utilize a novel reading frame for at least one exon of the second protein(ORF=1 exon). Three candidates derived from Ensembl annotation were already annotated as single genes by UCSC Genome Browser Database(*), and four could not be evaluated because they contained highly repetitive sequences or were subject to immune related recombination(**). D) Sanger sequencing of RT-PCR products from independent mouse samples confirms the existence of tandem fusions illustrated in A & B. Forty nucleotides bracketing the tandem splice junction are shown.

Detection of circular RNAs

Finally, our analysis included additional reads that did not map to a linear genome. Recent studies have highlighted the widespread presence of ciRNAs created by back-splicing of exon splice donors into upstream exon splice acceptors(14, 15). This results in a circular RNA molecule resistant to exonuclease attack. One such transcript, ciCdr1as, has been shown to have a conserved functional role as a sponge for microRNAs(16, 17), though it remains unclear whether other ciRNAs are also functional or merely occasional aberrations of splicing(49). It has been suggested that for some genes over 10% of the transcripts may undergo circularization(49), though as ciRNAs are predicted to be more stable than their linear counterparts this may overestimate their actual production.

While ciRNAs are typically cyotplasmic, we reasoned they must arise from splicing events occurring in the nucleus and thus be detectable in our data. As ciRNAs can be detected from back-spliced reads spanning exon junctions, we screened all unmapped reads for this property(Figure 5). A key distinction of our approach compared with prior screens for ciRNAs(14, 15, 49, 50) is that we did not presume circularization will occur only between annotated exons; thus we have the opportunity to observe circular splicing between known and novel exons, as well as completely novel exon pairs. This method detected 8,328 back-splicing loci across all samples, many mapping to specific cell-types(Figure S5 in Supplement 1). Consistent with prior reports, the vast majority of these were present in very low abundance, and many were more abundant in cytoplasm. Interestingly, many genes appeared to have multiple circular forms, produced by different combinations of exons(Figure S6 in Supplement 1). Some of these were confirmed by PCR amplification followed by Sanger sequencing(Figure 6; and S6B in Supplement 1). There were moderate correlations(r=0.2–0.6) between back-splicing counts and total RNAseq counts for the same genes, suggesting the occurrence of a back-splice depends in part on the expression level of the host gene. Overall abundance of back-spliced reads ranged from 0.006% for some of the nuclear samples to 0.046% for cytoplasmic ones.

Figure 5. The discovery of circular RNA species in the nuclei of specific CNS cell-types.

A) Schematic for discovering putative circular nuclear RNAs. B) A custom ‘circular genome’ generated by creating a genome definition file where each ‘chromosome’ was a repeat-masked tandem copy of the sequence of each Ensembl annotated gene. After mapping using standard aligners, linear reads(green) will map twice, while candidate circular reads will map only once, span both gene copies, and back-splice from a more 3’ spice donor to a relatively 5’ splice acceptor. C) Example of ciRNA reads in Hnrnpu gene splicing from midway through exon 17 of copy 2, into exon 14 of copy 1 in custom genome(note that gene is on second strand). D) All back-spliced read coordinates are converted to the standard genome and provided as a circular RNA track(http://tinyurl.com/CNS-Nuc-RNA); Hnrnpu shown as example. Note that splice junctions are end to end(blue arrow), and the splice donor is not at an annotated exon boundary. E) Illustration of a predicted circular RNA product, detected by the presence of reads spanning from exon D to A. F) Expression of ciCdr1as is detected in cytoplasmic RNA, while expression of ciHnrnpu is found almost exclusively in neuronal nuclear RNA, even though the Hnrnpu transcript is robustly expressed in all three cell-types. Mean CPM(+/− SEM) shown.

Figure 6. Independent confirmation of circularity, nuclear enrichment, and neuronal expression of ciHnrnpu.

A) Cartoon of ciHnrnpu showing locations for primers used in B(gray arrows) and C-G(black arrows). B) Agarose gel electrophoresis of PCR products from gray primers showing two isoforms of ciHnrnpu are detected. C) Sanger sequencing confirmed back-splicing. D) RT-qPCR of ciHnrnpu in samples from independent Snap25 mice shows it is more efficiently primed by random hexamers than oligo dT(−ΔΔCt method, normalized to the linear mRNA Actb, p<.05, unpaired T-test). E) RT-qPCR of ciHnrnpu in samples treated with RNAse R shows it is relatively resistant to RNAse R(−ΔΔCt method, normalized to Actb, p<.005, unpaired T-test). Shift in threshold cycle (Ct) was driven entirely by change in Actb following RNAse R digestion in this experiment. F) RT-qPCR of ciHnrnpu in parallel nuclear and cytoplasmic RNA fractions from the same mice confirm its relative enrichment in the nucleus(−ΔΔCt method, normalized to Actb, p=.001, unpaired t-test). G) RT-qPCR of ciHnrnpu across four regions of mouse brain show no substantial differences(−ΔΔCT method, normalized to Rna18s and cortex).

Recently, it has been reported that ciRNA formation is promoted by intronic binding of the RNA binding protein QK during the epithelial to mesenchyme transition(51). QK is known to be a key regulator of translation in oligodendrocytes(52), while it is largely unexpressed in neurons; our data confirm this(Figure 7A). Thus, if QK were the sole mediator of RNA circularization, one would predict a higher rate of back-splicing in oligodendrocytes. However, analyzing loci with evidence from at least three back-spliced reads between annotated splice junctions, we do not see substantially more circularization of the same transcript in oligodendrocytes than in the other cell-types(Figure 7B). Likewise, the overall abundance of back-spliced RNA was slightly lower in oligodendrocytes than neurons(Figure 7C). The presence of robust ciRNA in the nuclei of largely QK-negative neurons indicates there are other mechanisms for ciRNA formation.

Figure 7. Amount of circular RNA detected is comparable across all three cell-types.

A) The Qk gene, previously reported as mediating circularization of RNA(50), is most highly expressed in oligodendrocytes by RNAseq(Oligo vs Neuron, FDR=1.68e-28, Astro vs Oligo, FDR<.02). Mean CPM(+/− SEM) shown. B) Percentage of back-splicing at annotated linear splice junctions, and C) total CPM(mean +/− SEM) devoted to back-spliced reads, is not significantly higher in oligodendrocytes than in neurons or astrocytes.

Since most loci had evidence from few reads, we focused on those with the highest abundance(>1 CPM) for manual curation. Among these, <10% showed normal splicing misattributed by the algorithm. Of the remaining, most showed evidence of preferential detection in particular cell-types(Figure S5 in Supplement 1, Table S6 in Supplement 2), and common for either splice-donor or splice-acceptor to be novel – not mapping to an annotated exon boundary(Figure S6 in Supplement 1). Indeed, 22% of the events detected overall, and 100% of those with CPM>1 had a novel splice-site. We also detected ciCdr1as in the brain, and, consistent with earlier reports(17), it was enriched in the cytoplasm(Figure 5F). It showed only trace reads from nuclei, and only in astrocytes. However, it appears that cytoplasmic enrichment is not required of all ciRNAs. Indeed, we robustly detected a ciRNA arising from the 3’UTR of the RNA binding protein Hnrnpu almost exclusively in the nucleus of neurons(Figure 5F).

As ciHnrnpu (circular Hnrnpu) has not been reported in any prior studies of ciRNA, we validated its existence with multiple methods. Sanger sequencing of RT-PCR products spanning the junction confirmed back-splicing(Figure 6). Further, we determined that two versions of the transcript exist, including and excluding exon 15(Figure 6B). We then confirmed the circularity of the RNA in independently-sorted neuronal nuclei using two standard methods(14, 15): First, as ciRNAs are not poly-adenylated, they should be less efficiently primed by Oligo-dT than random hexamers during cDNA syntheses, and second, they should be resistant to the exonuclease RNAse R(Figure 6D,E). Finally, we confirmed that this species of RNA is enriched in nuclei relative to cytoplasm, and is found throughout the brain(Figure 6F,G). This robust detection of ciRNAs in neuronal nuclei suggests a possible nuclear-specific role for ciHnrnpu.

Discussion

We have characterized the nuclear transcriptome of the major cell-types of the CNS in vivo using a new approach. Our findings are largely consistent with patterns of nuclear transcription seen in various cell lines in vitro(1, 5), with a greater proportion of ncRNA detectable and a larger fraction of the genome transcribed. The data also revealed some rare events including tandem gene fusions and ciRNAs generated by back-splicing of both annotated and unannotated exons. Thus, in addition to defining the differential transcription across these RNA classes in the three major cell-types, we also provide a key resource for future analyses of RNA splicing, editing, and transcriptional regulation in the CNS, which may be of interest for analyses comparing nuclear profiling in vivo to prior analysis in vitro(1, 5, 53, 54). These data may also provide some context for human psychiatric genetic studies, where, for example, up to 15% of loci recently associated with schizophrenia were greater than 20 kb from any protein coding genes, include several regions where the association spans hundreds of kb. Our data indicate that the mouse regions syntenic for a subset of these loci are producing RNA in the nucleus – notably almost continuous coverage from a >600 kb long region upstream of Gria1 that was detected primarily in astrocytes(Figure S7 in Supplement 1). Thus, we provide raw data(GEO:GSE73391) and browser tracks of expression levels, novel splice junctions, and circular RNAs to facilitate leveraging of this data(http://tinyurl.com/CNS-Nuc-RNA).

This work also demonstrates that bacTRAP mice, previously utilized for translational profiling, can also provide access to the nuclear transcriptome. The protocol as currently implemented does exhibit some non-specific background similar to previously reported for other FACS-based methods(55), which we coped with here by focusing on a comparative analysis between cell-types. We also acquired RNAseq data from presorted nuclei to allow for filtering of transcripts below empirically derived background thresholds, as previously described(56). Caution should be used when interpreting results for genes with lower expression in a particular cell-type than that found in presorted nuclei. In addition, though Aldh1L1 is currently the best available marker of astrocytes, it is also expressed in choroid plexus and neural stem cells(43). While these likely contribute only a small fraction of the nuclei, we caution that some small fraction of the Aldh1L1 nuclear RNA here might also be attributable to them. Nonetheless, the in vivo nuclear RNA purification approach used here provides important experimental access to neural cell-types, and even finer dissection of cell-types using more specific bacTRAP lines(2, 57, 58) or Cre-responsive bacTRAP reagents(59).

Finally, we have confirmed the existence of ciRNA in the nervous system(17, 60) and identified potentially cell-type specific ciRNA species in the adult CNS. Our results also suggest that most ciRNAs use at least one splice site not found in mRNA and multiple ciRNAs may be transcribed from a single gene. We have also shown that a subset of ciRNAs are enriched in the nucleus. This suggests some active transport of ciRNAs to the cytoplasm, and that the process can be regulated in a sequence-specific way. In the CNS, passive escape from the nucleus during mitosis, previously proposed as a method of export(15), is highly unlikely as the vast majority of cells are postmitotic. Furthermore, while our current findings do not contradict the recent work indicating that QK may mediate the circularization of RNA in other tissues(51), the detection of ciRNA in non-QK expressing cell-types suggests there are multiple mechanisms that promote back-splicing. Indeed, previous work shows that inverted repeat sequences also can promote RNA circularization(61), suggesting both nucleotide sequence and/or RNA-binding proteins that alter RNA conformation might enhance formation of ciRNAs.

The detection of ciHnrnpu specifically in the nuclei of neurons is notable. The 3’ UTR of the host transcript is unusual: it contains exons after the stop codon, and thus should be subject to nonsense-mediated decay. Evidence from human sequences(e.g. CB053217) demonstrates that this spliced UTR is conserved, suggesting some role for this part of the gene. It is also interesting that microdeletions overlapping Hnrnpu cause intellectual disability and seizures in humans, indicating this genomic region is essential for normal CNS function(62). Finally, despite approximately equal levels of Hnrnpu primary transcript in neuron, oligodendrocyte, and astrocyte nuclei, ciHnrnpu is ~30 fold more abundant in neurons. Though expression is only a prerequisite for function, this substantial and cell-type specific enrichment encourages future studies investigating whether ciHnrnpu may have a role specifically in the nuclei of these cells.

More broadly, the methods described here may be applied to mechanistic questions in psychiatric genetics. In addition to comparisons with GWAS data(Figure S7 in Supplement 1), our approach could assess experimental manipulations for targeted investigation of mechanisms in cell-types of clinical interest. For example, we have previously applied TRAP to profile neurons important for the regulation of sleep(57): nuclear RNA capture from these mice could be combined with administration of lithium, a drug known to influence chromatin state and gene expression, to investigate how this compound alters circadian rhythms during the treatment of bipolar disorder. Given the abundance of non-coding transcripts in the nuclei of healthy CNS cells, techniques that allow investigation of both coding and ncRNA may prove critical in advancing our understanding of psychiatric pathology.