Low RNA stability signifies increased post-transcriptional regulation of cell identity genes

Abstract

Cell identity genes are distinct from other genes with respect to the epigenetic mechanisms to activate their transcription, e.g. by super-enhancers and broad H3K4me3 domains. However, it remains unclear whether their post-transcriptional regulation is also unique. We performed a systematic analysis of transcriptome-wide RNA stability in nine cell types and found that unstable transcripts were enriched in cell identity-related pathways while stable transcripts were enriched in housekeeping pathways. Joint analyses of RNA stability and chromatin state revealed significant enrichment of super-enhancers and broad H3K4me3 domains at the gene loci of unstable transcripts. Intriguingly, the RNA m⁶A methyltransferase, METTL3, preferentially binds to chromatin at super-enhancers, broad H3K4me3 domains and their associated genes. METTL3 binding intensity is positively correlated with RNA m⁶A methylation and negatively correlated with RNA stability of cell identity genes, probably due to co-transcriptional m⁶A modifications promoting RNA decay. Nanopore direct RNA-sequencing showed that METTL3 knockdown has a stronger effect on RNA m⁶A and mRNA stability for cell identity genes. Our data suggest a run-and-brake model, where cell identity genes undergo both frequent transcription and fast RNA decay to achieve precise regulation of RNA expression.

Graphical Abstract

Graphical Abstract.

Cell identity genes are regulated by unique epigenetic mechanisms, such as broad H3K4me3 and super-enhancers. These genes also undergo both frequent transcription and frequent co-transcriptional m6A modifications on their RNAs, which results in rapid RNA decay, like a "run-and-brake" model. The rapid mRNA turnover of cell identity genes may serve as a robust system to precisely regulate the homeostasis of transcripts and achieve optimal gene expression.

INTRODUCTION

Gene expression is known to undergo regulation at multiple levels, including transcriptional and post-transcriptional regulation. An increasing number of publications indicate that post-transcriptional regulation of RNA stability, which can be measured by the RNA decay rate, is fundamentally important for cell differentiation (1–9), development and the progression of multiple diseases (10–16). In particular, it is becoming clear that RNA modifications play a crucial role in the regulation of RNA stability during various biological processes such as stem cell differentiation (1), T-cell homeostasis (17), development (4,12,18) and cancer progression (3,16). For example, N⁶-methyladenosine (m⁶A) modifications decrease the RNA stability of SOX2, NANOG, MYC and SMAD2, which play key roles in the self-renewal and differentiation of embryonic stem cells (ESCs) (1). In addition, our recent work demonstrates that EZH2 enhances 2′-O-methylation of rRNAs to promote prostate cancer progression (19). The effect of RNA modification on RNA stability is often mediated by RNA-binding proteins, which recognize and interact with modified RNAs (20,21). RNA sequencing at multiple time points following inhibition of transcription or metabolic labeling of newly synthesized RNAs has been used in analyzing RNA decay rates at the whole-transcriptome level in several cell types (22–28). Despite this, little is known about the specificity or conservation of RNA stability, e.g. in various types of cells.

It is apparent that different functional gene categories might be transcriptionally regulated by different epigenetic mechanisms. For instance, the transcription of cell identity genes tends to be regulated by super-enhancers, each consisting of a large cluster of enhancers (29). A super-enhancer can be defined by multiple transcription factor-binding sites that are close to each other or by a broad enrichment of the histone modification H3K27ac (30). Similarly, a broad enrichment pattern of H3K4me3 marks promoter regions of both cell identity genes and tumor suppressor genes, while a sharp enrichment pattern of H3K4me3 marks promoters of other genes such as housekeeping genes (31–33). Intriguingly, the broad H3K4me3 domain at cell identity genes tends to be cell type specific, while the broad H3K4me3 domain at tumor suppressor genes tends to be conserved among cell types (31,32). Furthermore, both super-enhancers and the broad H3K4me3 domain are found to be associated with increased transcription of the associated genes. In contrast, oncogenes tend to have a broad genic repression domain (BGRD), defined by a broad enrichment pattern of the histone modification H3K27me3 on the gene body (34). Interestingly, compared with other silent genes, the cell identity genes are more likely to be repressed by H3K27me3 (35,36). Knockout of H3K27me3-rich regions causes changes of the phenotype associated with cell identity (37). Other unique epigenetic modification patterns on the chromatin, e.g. bivalent domains (38), DNA methylation canyons (39) or DNA valleys (40), large organized chromatin K9 modifications (LOCKs) (41) and broad H3K27me3 (42,43), have also been reported to denote unique transcriptional regulatory mechanisms. Transcriptional regulation of cell identity genes, e.g. by super-enhancers and broad H3K4me3 domains, has been a subject of extensive investigation in different cell and tissue types (29,32,44–50). However, the post-transcriptional regulation patterns and mechanisms for cell identity genes remain largely unexplored.

Two key components of the m⁶A methyltransferase complex, METTL3 and METTL14, were reported to bind chromatin and co-transcriptionally catalyze RNA m⁶A modification (51–53). An additional report showed that METTL3 could bind to promoters and co-localize with H3K4me3 around transcription start sites (TSSs) (54). In addition to the transcriptional regulation of target genes, super-enhancers were reported to recruit the microprocessor proteins DGCR8 and Drosha to bind chromatin for post-transcriptional processing of primary microRNAs (55). Based on these findings, we speculate that super-enhancers and broad H3K4me3 domains denote unique mechanisms for post-transcriptional regulation of cell identity genes.

In this work, we investigated the landscape of RNA decay at the transcriptome scale to systematically determine the biological implications of RNA stability. We observed that unstable transcripts tend to be highly cell type specific and are enriched in cell identity-related pathways. In contrast, stable transcripts are enriched in housekeeping pathways and are conserved among cell types. Transcription tends to be more active at the gene loci of unstable transcripts than at those of stable transcripts. Lastly, we revealed a connection between the broad H3K4me3 domains or super-enhancers on chromatin and the co-transcriptional m⁶A modification on transcripts of cell identity genes.

MATERIALS AND METHODS

The sources of public RNA stability, ChIP-seq and m⁶A datasets are described in Supplementary Table S4. The software used in this project was installed and configured by BioGrids (56).

T-cell isolation and RNA-seq library preparation

Human primary T lymphocytes were isolated from buffy coat samples obtained from Gulf Coast Regional Blood Center using the Ficoll density gradient method. In brief, 20 ml of buffy coat samples were taken in a sterile 50 ml Falcon tubes containing 20 ml of Lymphoprep™ (Stem Cell Technologies). The tubes were centrifuged at 800 g for 30 min at room temperature. The mononucleated cells were separated from erythrocytes, which settled at the bottom of the tube. The upper layer containing plasma was carefully removed and discarded. The opaque interface in the middle containing lymphocytes was carefully collected, washed with phosphate-buffered saline (PBS) with 0.5% fetal bovine serum (FBS) and cultured. Post-isolation, the cells were activated using anti-CD3/CD28 microbeads (Thermo Fisher) for 24 h. T cells were further passaged in OpTmizer T Cell SFM (ThermoFisher) with interleukin-2 (IL-2; 200 IU/ml). Upon reaching confluence, cells were incubated with ActD (actinomycin D) (10 μg/ml) for four different time points (0, 0.25, 1 h and 2 h) with three biological replicates. The cells were washed with PBS, and RNA was isolated using an RNeasy kit (Qiagen) post-treatment. The integrity of the isolated RNA was analyzed using a Bioanalyzer prior to sequencing.

Culture of CD4+ T cells

Human peripheral blood CD4+ T cells were purchased from STEMCELL. Upon activation by ImmunoCult human CD3/CD28/CD2 T Cell activator (STEMCELL), T cells were expanded in ImmunoCult-XF T Cell expansion medium (STEMCELL) with a supplement of human recombinant IL-2 (STEMCELL) at 37°C and 5% CO₂ in a humidified atmosphere.

Lentivirus production and infection

Pre-designed lentiviral short hairpin RNA (shRNA) vectors targeting METTL3 were purchased from Sigma (shMETTL3-1, TRCN0000289742; and shMETTL3-2, TRCN0000289743). For lentivirus production, the shRNA vector and helper plasmids of pVSVG and psPAX2 were co-transfected into HEK293T cells by Lipofectamine 3000 (Invitrogen) following the manufacturer's protocol. The medium was renewed at 24 h post-transfection. The supernatants containing viruses were collected at 48 h post-transfection and directly used for spin infection into CD4+ T cells.

Nanopore direct RNA-seq

Total RNA was extracted from both control and shMETTL3-1 CD4+ T cells using the RNeasy Plus Mini Kit (Qiagen) and mRNA was purified using the PolyATtract mRNA Isolation System (Promega). The mRNA concentration was measured by Qubit RNA HS assay kit (ThermoFisher) and the quality was confirmed using an Agilent Bioanalyzer RNA Pico assay chip. Then, the mRNA was subjected to library construction using the Oxford Nanopore direct RNA sequencing kit (SQK-RNA002) according to the manufacturer's manual. A total of 500 ng of mRNA was first annealed and ligated to a reverse transcriptase adapter and then reverse-transcribed to form DNA–RNA hybrid products. Next, an RNA adapter was ligated to the DNA–RNA hybrid products. The resulting library was sequenced on an R9.4.1 flow cell (FLO-MIN106D) using a MinION sequencer, and FAST5 raw sequencing data were obtained in real time. Following base calling with Guppy (version 3.6.1 + 249406c) and mapping to human genome hg19 from the GENCODE database using Minimap2 (57) (version 2.22-r1101), we used a machine learning-based software, nanom6A (version 2021_10_22) to detect m⁶A from Nanopore direct RNA-seq with default parameters (58). The RNA methylation ratios and gene expression levels from Nanopore direct RNA-seq data are summarized in Supplementary Tables S5 and S6.

Measurement of mRNA stability via Click-iT chemistry

The half-lives of selected mRNAs were determined using the Click-iT Nascent RNA Capture Kit (Thermo) as previously described with minor revision (59). Briefly, both control and METTL3-deficient T cells were labeled with 0.2 mM ethynyl uridine (EU) through incubation at 37°C overnight. Cells were then recovered in EU-free fresh medium for 0, 0.25, 1 and 2 h, respectively. Total RNA was extracted from samples collected at each time point using an RNeasy Plus Mini Kit (Qiagen), and 10 μg of total RNA was subjected to a Click-iT reaction for biotinylation. Subsequently, the biotinylated RNA was precipitated and bound to Dynabead Streptavidin T1 magnetic beads. The captured RNA was in-bead converted to cDNA as per the manufacturer's instructions using the SuperScript VILO cDNA synthesis kit (Thermo). Finally, relative levels of mRNAs at each time point were measured by real-time quantitative polymerase chain reaction (RT-qPCR) analyses using Universal SYBR Green Supermix (Bio-Rad) in the QuantStudio 6 Flex Real-time PCR System (GE Healthcare). The relative mRNA level was calculated using the 2^−ΔΔCt method with the Ct values normalized using glyceraldehyde phosphate dehydrogenase (GAPDH) as an internal control. All qPCR primers used are summarized in Supplementary Table S7.

Half-life analysis for mRNA

The human reference genome sequence version hg19 and UCSC Known Genes were downloaded from the iGenomes database (https://support.illumina.com/sequencing/sequencing_software/igenome.html). RNA-seq raw reads (Supplementary Table S4) were mapped to the human genome version hg19 using TopHat version 2.1.1 with default parameter values. mRNA half-lives of the T cells were calculated as described before (60). In short, we first normalized the mRNA levels to RPKM (reads per kilobase million). To reduce the gene expression noise due to genes that showed higher expression levels after inhibition of transcription, we then scaled the RPKM values based on the expression levels of the top 10 genes which were most stable between time points, so that these stable genes would have no change of expression between the time points. The RNA degradation rate k_decay was calculated as the average value of the ratio from RPKM at time 0 over the RPKM value at each other time point (0.25, 1 and 2 h). Thereafter, the pseudo-mRNA half-life t_1/2 was calculated as ln2/k_decay. RNA half-life values in other cell types downloaded from public datasets are listed in Supplementary Table S4 (22,61).

ChIP-seq analysis

The chromatin immunoprecipitation (ChIP)-seq dataset of RNA polymerase II (Pol II), H3K4me3 and H3K27ac from HEK293 cells, the ChIP-seq dataset of Pol II, H3K4me3, H3K27ac, METTL3, DGCR8 and Drosha from mouse ESCs (mESCs), and ChIP-seq of H3K4me3, H3K27ac and METTL3 from human MOLM cells were downloaded from the Gene Expression Omnibus (GEO) database, with accession numbers listed in Supplementary Table S4.

For ChIP-seq data, the read mapping and processing were performed as described previously (62). In brief, the reads were first mapped to the human genome version hg19 by bowtie2 (version 2.4.4). A Wig file was generated using DANPOS 2.2.3 (63) with the command line: python danpos.py dpeak sample –b input –smooth_width 0 -c 25000000 –frsz 200 –extend 200 –o output_dir. Bigwig file was generated using the tool WigToBigWig with the following command line: wigToBigWig -clip sample.bgsub.Fnor.wig hg19.sizes.xls sample.bw.

Broad H3K4me3 and super-enhancer gene lists were obtained from two published datasets (31,64) and the SEdb database (65). For the cell types which were not included in the published dataset, H3K27ac and H3K4me3 ChIP-seq data were obtained from the GEO database (Supplementary Table S4). The raw data were downloaded from the GEO database and mapped to the human genome version hg19 or mouse genome version mm10 using bowtie2. To define super-enhancers by the histone modification H3K27ac, we followed the methods described in Whyte et al. (9). Briefly, the reads were mapped to the genome using bowtie2, and then the peaks of the ChIP-seq signal over input were called by the MACS2. Next, the ROSE algorithm (9) was applied to identify super-enhancers and typical enhancers by the command line: python ROSE_main.py -g HG19 -i *peaks.bed –c *_input.sort.bam -r *_ChIP.sort.bam -o * -s 12500. The super-enhancers and typical enhancers were annotated by the command line: python ROSE_geneMapper.py -g HG19 -I *_AllEnhancers.table.txt -f. For the broad H3K4me3 genes, the methods were the same as described before (31). Briefly, after processing by the DANPOS dregion function, the widths of the peaks around the TSS were computed by the Selector function in DANPOS. At the same time, typical enhancer genes and sharp H3K4me3 genes were used as controls to compare with super-enhancers and broad H3K4me3 genes, respectively. The RNA polymerase density ratio (R/P ratio) was calculated using Pol II ChIP-seq and mRNA expression (RNA-seq) (66).

Functional enrichment analysis

Gene Ontology (GO) and Kyoto Encyclopedia of Genes and Genomes (KEGG) functional enrichment analyses were performed using the R package clusterProfiler (version 4.0.5)(67). Gene set enrichment analysis was performed using GSEA Java software (version 4.0.3) (68). We calculated the Z-score for mRNA half-life in each cell type and used the GSEAPreranked method to analyze enrichment of cell identity genes. Only the expressed genes were included in the GSEA analysis. The GSEA based on differential gene expression was performed by ranking genes based on the fold change of RNA expression between the knockdown and control samples. The spectrum from blue to red in GSEA figures indicates the rank of mRNA half-life or fold change of gene expression. Each vertical bar underneath the spectrum indicates one gene in the respective gene set. Overlap analysis between the super-enhancer genes or broad H3K4me3 genes and the unstable genes was performed using the R package GeneOverlap (69). The P-values were calculated by Fisher's exact test. The radar plot was draw by the R function radar chart.

Analysis of m⁶A modification

The m⁶A modification data for mESCs, induced pluripotent stem cells (iPSCs) and fibroblasts were download from the GEO database with the accession number GSE52662. The sequencing reads were mapped to the mm10 (mouse) or hg19 (human) genome by Tophat (v2.1.1) (70). The m⁶A peaks were called by the R package exomePeak (v 1.9.1) (71), which requires both the m⁶A reads as the signal and RNA reads as the background for normalization. To calculate the m⁶A signal in each peak, the m⁶A IP reads number was divided by the input RNA reads number. The m⁶A peak numbers in the super-enhancer or typical enhancer genes and broad or sharp H3K4me3 genes were plotted by the Guitar R/Bioconductor package (72). Lists of m⁶A methylated genes in HEK293, HeLa, HepG2 and B cells were obtained from the results in public datasets (73,74).

Comparison between cell identity genes and housekeeping genes with similar expression levels

Human and mouse housekeeping genes were downloaded from published datasets (75,76). To find the housekeeping genes that have expression levels close to those of the cell identity genes, we merged the cell identity genes with housekeeping genes, ranked the genes by expression levels and divided the rank into 20 bins. Then we randomly selected the same number of housekeeping genes and cell identity genes from each bin to perform the comparison.

AU-rich elements, A-to-I RNA editing and APA analyses of mRNA

We downloaded the list of genes containing AU-rich elements from a public database (77). We used the REDItoolDenovo.py script from REDItools (v.1.0.3) to detect the A-to-I RNA editing from RNA-seq data (78,79), with known single nucleotide polymorphisms (SNPs) from dbSNP (version 151) removed as potential false positives. For RNA editing and alternative polyadenylation (APA) analysis, we merged the biological replicates for deeper sequencing depth. The A-to-I editing sites were defined by requiring the editing ratio to be >0.1 and the reads coverage to be >20. APA analysis based on RNA-seq data was performed using the software Dapars2 with default parameters (80,81).

RESULTS

Transcripts of cell identity genes tend to have a low RNA stability

To investigate transcriptome-wide RNA stability, we treated T cells with ActD to inhibit transcription and performed RNA sequencing (RNA-seq) before and after ActD treatment (Figure 1A). The decrease of transcripts for a gene following the inhibition of transcription reflects the decay rate of RNA and was used to calculate the half-life of RNA for each gene. Our results indicated that the RNA half-lives of different genes range from <1 h to >12 h in T cells (Figure 1B; Supplementary Table S1). We observed a similar result by applying this method to HEK293 cells (Supplementary Figure S1A). To investigate the functional implications of RNA stability, we retrieved the top 1000 genes of the most unstable transcripts in T cells and performed a GO pathway analysis. We observed significant enrichment of pathways associated with T-cell identity, e.g. T-cell differentiation (GO:0030217) and alpha–beta T-cell activation (GO:0046632) (Figure 1C; Supplementary Table S2). Meanwhile, the top 1000 genes of the most unstable transcripts in HEK293 cells, which were derived from human embryonic kidney cells, were enriched in kidney development (GO:0001822), embryonic organ development (GO:0048568), regionalization (GO:0003002) and urogenital system development pathways (GO:0001655) (Supplementary Figure S1B; Supplementary Table S2). In contrast, genes of stable transcripts were enriched in housekeeping pathways related to basic molecular activities, such as a protein targeting to membrane (GO:0006614) or to endoplasmic reticulum (GO:0045047), mRNA catabolic process (GO:0000184), precursor metabolites and energy process (GO:0006091) and nucleoside monophosphate metabolic process (GO:0009123) (75) (Figure 1D; Supplementary Figure S1C; Supplementary Table S2). We further consolidated this observation by analyzing RNA stability based on sequencing of newly synthesized RNAs that were metabolically labeled using 4sU (4-thiouridine) in HEK293 cells (Supplementary Figure S2). These results indicate that cell identity genes tend to be unstable, and this observation is not limited to a single cell type and not dependent on a specific method for RNA stability analysis.

Figure 1.

T-cell unstable transcripts were enriched in T-cell identity-related pathways. (A) A flowchart of RNA-seq experiments to profile mRNA stability at multiple time points during transcription inhibition by treating the cells with ActD. (B) The cumulative fraction of genes plotted against mRNA half-life in T cells. Example genes were labeled to show genes of unstable (red) or stable (blue) transcripts. (C and D) GO enrichment analysis performed for genes (gene number =1000) of unstable (C) and stable (D) transcripts in T cells. (E) Enrichment of each functional gene set in individual gene groups defined and ranked by RNA half-life from 1 to 24 h. To define gene groups by RNA half-life (x-axis), 10 165 genes were ranked by RNA half-life from short to long and divided into 10 groups, with groups 1 and 10 having the shortest and longest half-life, respectively. Odds ratios were determined by one-tailed Fisher's exact tests. (F) Genome browser tracks to show the RNA-seq signal of individual genes in individual samples. RNA-seq analyses were based on three independent biological replicates.

To further investigate the association between low RNA stability and genes in cell identity pathways, we retrieved the genes from the T-cell differentiation pathway (GO:0030217) and T-cell activation pathway (GO:0042110). We assessed the enrichment of these pathways in individual gene groups associated with different levels of RNA stability. The genes from these pathways were only enriched in the gene group associated with the most unstable transcripts (Figure 1E). The enrichment level diminished quickly along with the rank of gene groups by RNA stability from unstable to stable. A similar enrichment pattern was observed in HEK293 cells for genes from the embryonic organ development pathway (GO:0048568) and kidney development pathway (GO:0001822) (Supplementary Figures S1D and S2D). As a control, we also analyzed housekeeping genes collected from the literature (75), in which the housekeeping genes were defined based on RNA expression patterns in 16 tissue types from humans. In contrast to the genes from cell identity-related pathways, the housekeeping genes were enriched in genes of stable transcripts, with an increase in enrichment level along the rank of gene groups by RNA stability from unstable to stable (Figure 1E; Supplementary Figures S1D and S2D). These results confirmed the low RNA stability of genes from cell identity-related pathways and, in contrast, the high RNA stability of housekeeping genes.

We also observed low RNA stability for cell identity genes by visual inspection of genome browser tracks at individual loci. For instance, the genes XBP1 (82), SMAD7 (83) and RHOH (84) play important roles in T-cell differentiation. The transcripts of these genes in T cells decayed 2- to 4-fold faster than the transcripts of housekeeping genes such as RPL24, PHPT1 and BABAM1 (Figure 1B, F). The RNA half-lives of XBP1, SMAD7 and RHOH were <2 h, while RPL24, PHPT1 and BABAM1 showed RNA half-lives of >8 h. We also found that mRNA stability is low for the genes HES1, GATA3 and HOXA5-A7 in the human embryonic kidney cell line HEK293. Of these genes, HES1 regulates the differentiation of ESCs by suppressing Notch signaling (85); GATA3 is a transcription factor gene critical for embryonic development (86); and the homeobox (HOX) family genes HOXA5, HOXA6 and HOX7 are spatially and temporally regulated during embryonic development (87). In contrast, we did not observe a significant change of RNA expression levels for the housekeeping genes RPL24, PHPT1 and BABAM1 in the HEK293 cells upon inhibition of transcription or under 4sU-labeled method, indicating strong RNA stability (Supplementary Figures S1A, E and S2A, E).

Unstable transcripts tend to be cell type specific

To assess the specificity or conservation of RNA stability across cell types, we expanded our investigation of RNA stability to six different cell types based on a combination of an in-house RNA-seq dataset from T cells as well as public datasets from five additional cell types, namely HEK293 cells, B cells, ESCs, HeLa cells and HepG2 cells (22–28). Our data indicated that the genes of unstable transcripts in each cell type were enriched in their respective cell type-specific pathways (Figure 2). For example, GSEA confirmed that the genes of unstable transcripts in T cells and HEK293 cells were enriched in T-cell differentiation pathway (GO:0030217) and the embryonic organ development pathway (GO:0048568), respectively (Figure 2A). Furthermore, the genes of unstable transcripts in B cells, human ESCs, HeLa cells and HepG2 cells were enriched in the lymphocyte differentiation pathway (GO:0030098), embryonic epithelial tube formation pathway (GO:0001838), utero embryonic development pathway (GO:0001701) and Wnt signaling pathway (GO:0016055) (Figure 2A). The enrichment of these pathways in genes of unstable transcripts appeared to be highly cell type specific (Figure 2B). For instance, the enrichment of the T-cell differentiation pathway in genes of unstable transcripts was only observed for T cells but not in any of the other five cell types. In contrast, the enrichment of housekeeping pathways in genes of stable transcripts appeared to be conserved in these different cell types. For instance, the oxidative phosphorylation (GO:0006119), mitochondrial translation (GO:0032543), mitochondrial gene expression (GO:0140053), generation of precursor metabolites and energy (GO:0006091) and ribonucleoprotein complex biogenesis (GO:0022613) pathways showed a similar degree of enrichment in genes with stable transcripts from each of the six cell types (Figure 2C; Supplementary Table S3). These results indicated that the low stability of transcripts from cell identity genes tends to be cell type specific, while the high stability of transcripts from housekeeping genes tends to be conserved between cell types.

Figure 2.

Genes of unstable transcripts in individual cell types were enriched with cell identity-related pathways. (A) GSEA of individual cell identity-related pathways to show their association with RNA stability. The spectrum from blue to red indicates the mRNA stability from the most stable to the most unstable, and each vertical bar underneath the spectrum indicates one gene in the respective gene sets. (B and C) GO analysis for genes (gene number = 1000) of unstable transcripts in individual cell types to show the specificity of their enrichment in cell identity-related pathways (B) and the conservation of their enrichment in housekeeping pathways (C).

Gene loci of unstable transcripts are enriched with epigenetic signatures of cell identity genes

We and others recently identified an epigenetic signature called broad H3K4me3, which is enriched at cell identity genes and associated with increased transcription (31,64). Meanwhile, super-enhancers, which could be defined by broad enrichment of the enhancer marker H3K27ac, were also found to activate cell identity genes (29,88). Intrigued by the enrichment of cell identity-related pathways in genes of unstable transcripts, we investigated the relationship between genes of unstable transcripts and genes associated with broad H3K4me3 or super-enhancers. Our results indicated that the genes associated with broad H3K4me3 or super-enhancers were significantly enriched with the genes of unstable transcripts in each of the six cell types described above (Figure 3A). Compared with genes of stable transcripts, genes of unstable transcripts have significantly broader H3K4me3 and H3K27ac modifications, e.g. in both T cells (Figure 3B–E) and HEK293 cells (Supplementary Figure S3A–D). The broad H3K4me3 of cell identity genes denotes increased transcription, which can be manifested as a high binding density of Pol II (31). Consistently, compared with genes of stable transcripts, the genes of unstable transcripts showed significantly higher binding intensity of Pol II (Figure 3F; Supplementary Figure S3E). In addition, we observed that the ratio of mRNA-seq to Pol II ChIP-seq read density was significantly lower for genes of unstable transcripts than for genes of stable transcripts (Figure 3G; Supplementary Figure S3F). These results indicated that like cell identity genes, the genes of unstable transcripts tended to be associated with broad H3K4me3, super-enhancers and active transcription.

Figure 3.

Genes of unstable transcripts are associated with epigenetic signatures of cell identity genes. (A) GSEA of genes marked by super-enhancers or broad H3K4me3 to show their association with RNA stability. (B) H3K4me3 signal density plotted around TSSs (n = 1000). (C) Cumulative fraction of genes plotted against width of H3K4me3 at the gene loci. (D) H3K27ac signal density plotted around TSSs. (E) Cumulative fraction of genes plotted against width of H3K27ac at the gene loci. (F) Pol II binding density plotted around TSSs. (G) Cumulative fraction of genes plotted against RNA-seq to ChIP-seq ratio of read numbers for individual genes (B–G, gene number = 1000, biological replicate number = 2). P-values were determined by two-tailed unpaired Wilcoxon's tests (C, E, G).

We next analyzed the mRNA stability of genes marked by the super-enhancer and broad H3K4me3 in eight cell types, i.e. T cells, neural progenitor cells (NPCs), human umbilical vein endothelial cells (HUVECs), ESCs, HepG2 cells, HeLa cells, HEK293 cells and B cells. In each of these cell types, we observed significant overlap between genes of unstable transcripts and genes marked by broad H3K4me3 (Figure 4A) or super-enhancers (Supplementary Figure S4A). In contrast, the overlap was less or not significant when the unstable transcripts and super-enhancers or broad H3K4me3 were defined in different cell types. For example, the genes of unstable transcripts in T cells were enriched with the super-enhancers defined in T cells but not enriched with the super-enhancers defined in NPCs, HUVECs or ESCs. Similarly, the RNA half-life was shorter for genes marked by broad H3K4me3 compared with genes marked by sharp H3K4me3 (Figure 4B), and for genes marked by super-enhancers compared with genes marked by typical enhancers (Supplementary Figure S4B). We next asked if the low stability of cell identity genes was biased by their high expression. Accordingly, we compared two gene sets with similar expression levels in each cell type, one of which contains putative cell identity genes defined by the broad H3K4me3 or super-enhancer and the other containing housekeeping genes (see the Materials and Methods). In this context, the mRNA half-lives of the putative cell identity genes were still shorter than those of the housekeeping genes (Supplementary Figure S5A–D). Collectively, these results suggested that the genes marked by known epigenetic signatures of cell identity genes, including the broad H3K4me3 and super-enhancers, tend to have a low RNA stability.

Figure 4.

Genes marked by broad H3K4me3 tend to have a low RNA stability. (A) Radar plot to show the significance of overlap between genes marked by broad H3K4me3 in T cells (top left), HEK293 cells (top middle), B cells (top right), ESCs (bottom left), HeLa cells (bottom middle) and HepG2 cells (bottom right), and genes of unstable transcripts in individual cell types. P-values were determined by one-tailed Fisher's exact tests. (B) Violin plots to show the half-life of genes (gene number = 500) marked by broad or sharp H3K4me3 in T cells. The three lines indicate quartile positions [Q1, Q2 (median) and Q3], and the black dot indicates the mean value. P-values were determined by two-tailed unpaired Wilcoxon's tests.

The RNA stability regulator METTL3 preferentially binds to chromatin at cell identity gene loci

It was recently reported that the microprocessor proteins DGCR8 and Drosha were co-transcriptionally recruited to super-enhancer-associated gene loci for post-transcriptional processing of pre-mRNA (55,89). This suggested that transcripts from super-enhancer-associated genes may be subjected to unique post-transcriptional regulation. Also, emerging evidence suggests that m⁶A, an RNA modification catalyzed by METTL3 and known to trigger RNA degradation, could be deposited on RNAs in a co-transcriptional manner (51–53). Co-IP assays confirmed that Pol II could recruit METTL3 to add m⁶A on RNA when the transcriptional machinery runs on the chromatin (90). However, little is known about the connection between the m⁶A modification on transcripts and the super-enhancers or broad H3K4me3 modification on chromatin.

Considering that the genes associated with super-enhancers or broad H3K4me3 showed lower RNA stability, we hypothesized that these genes might display stronger binding intensity of METTL3 on DNA and more m⁶A modifications on RNA. To test this hypothesis, we analyzed a set of data from mESCs (89). We first confirmed that unstable transcripts in mESCs were much more enriched with m⁶A modifications when compared with stable transcripts (Supplementary Figure S7A). We further observed that the binding intensities of METTL3, DGCR8 and DROSHA were positively correlated with the intensity of the histone modification H3K4me3, enhancer marker H3K27ac and the binding of Pol II on chromatin (Figure 5A). Compared with typical enhancers, super-enhancers showed a higher binding intensity of METTL3 (Figure 5B). METTL3 also showed higher binding intensity at gene loci associated with super-enhancers than gene loci associated with typical enhancers (Figure 5C). Similar results were observed when we analyzed genes associated with broad H3K4me3 and sharp H3K4me3 (Figure 5D), and when we analyzed independent data from the HEK293 cell line (Supplementary Figure S6A–D) (91) and the MOLM3 cell line (Supplementary Figure S7B–D). Pol II Ser5P is known to be associated with transcription initiation and is enriched in the promoter, while Pol II Ser2P is associated with transcription elongation and enriched at the 3′ end of the gene body (92). Pol II Ser5P and Ser2P both showed enrichment on super-enhancers in addition to their associated genes (Supplementary Figure S8A–D). Also, Pol II Ser5P showed higher ChIP-seq signals at promoters of genes associated with broad H3K4me3 compared with those of genes associated with sharp H3K4me3 (Supplementary Figure S8E). Meanwhile, Ser2P showed substantially stronger ChIP-seq signals at the 3′ ends of genes associated with broad H3K4me3 (Supplementary Figure S8F). Considering that Pol II was known to recruit METTL3, these results suggest that the interaction between METTL3 and Pol II happens with a higher frequency at the loci associated with the epigenetic signatures of cell identity genes.

Figure 5.

Genes marked by the epigenetic signature of cell identity genes are preferentially bound by METTL3 on chromatin and modified with m⁶A on RNA in mESCs. (A) Heat map to show the binding intensity of individual chromatin proteins or the density of individual histone modifications on chromatin flanking individual TSSs. (B–D) Binding density of METTL3 plotted in regions flanking typical or super-enhancers (B), gene bodies associated with typical and super-enhancers (C) and gene bodies associated with broad or sharp H3K4me3 modification patterns (D) (gene number = 500). (E) Number of m⁶A enrichment peaks on RNA plotted in genomic regions flanking gene bodies associated with typical or super-enhancers. (F) Violin plots to show half-lives of transcripts from genes associated with typical or super-enhancers (gene number = 500). (G) Number of m⁶A enrichment peaks on RNA plotted in genomic regions flanking gene bodies associated with broad or sharp H3K4me3 modification patterns (gene number = 500). (H) Violin plots to show the half-life of transcripts from genes associated with broad or sharp H3K4me3 modification patterns (gene number = 500). The three lines indicate quartile positions [Q1, Q2 (median) and Q3], and the black dot indicates the mean value. (I) Genome browser tracks to show the binding intensity of individual chromatin proteins or the density of individual histone modifications on chromatin flanking the gene locus of Nanog. P-values were determined by two-tailed unpaired Wilcoxon's tests (F, H).

We next investigated the relationship between m⁶A modifications on RNA and super-enhancers or broad H3K4me3 on DNA. We identified RNA m⁶A modification sites in the mESC data based on methylated RNA immunoprecipitation sequencing (MeRIP-seq). We observed more m⁶A modification sites in the region flanking the stop codon location on transcripts of genes associated with super-enhancers than on transcripts of genes associated with typical enhancers (Figure 5E; Supplementary Figure S6E–G). Consistent with results in human cell types (Figure 4B; Supplementary Figures S4B and S5), the RNA half-life in mESCs was shorter for genes associated with super-enhancers than for genes associated with typical enhancers (Figure 5F; Supplementary Figure S9A, B). We also observed a larger number of m⁶A modification sites and shorter RNA half-lives for genes associated with broad H3K4me3 than in genes associated with sharp H3K4me3 (Supplementary Figure S9D, E). The putative cell identity genes marked with super-enhancers and broad H4K4me3 showed higher m⁶A modification levels when compared with housekeeping genes (Supplementary Figure S9C, F–H). These results suggest an association between chromatin state, RNA m⁶A and RNA stability. For instance, METTL3, DGCR8 and DROSHA bound super-enhancer regions enriched with the enhancer marker H3K27ac in the upstream region of the pluripotent gene Nanog in mESCs (Figure 5I). The gene body of Nanog further displayed broad enrichment of H3K4me3, H3K27ac, Pol II and METTL3 (Figure 5I). Through analysis of the mRNA half-life, we also observed that unstable transcripts from mESCs were enriched in stem cell identity-related pathways, while stable transcripts were enriched with housekeeping pathways (Supplementary Figure S7E, F). Interestingly, the super-enhancer genes were up-regulated upon knockout of Mettl3 in mESCs (Supplementary Figure S7G). Well-known mESC identity genes, such as Nanog, Sox2, Klf4 and Smad3, showed increased mRNA stability upon Mettl3 knockout, while the housekeeping genes were unaffected (Supplementary Figure S7H, I). These results indicate that the expression level and RNA stability of cell identity genes were more sensitive to perturbations in Mettl3. In summary, we observed that METTL3 preferentially binds to genes associated with super-enhancers or broad H3K4me3, and the stronger binding intensity of METTL3 on these genes was correlated with a larger number of RNA m⁶A modification sites as well as lower RNA stability.

METTL3 regulates the mRNA half-life of cell identity genes though m⁶A modification

To verify the role of METTL3-catalyzed m⁶A in the regulation of mRNA stability of cell identity genes, we mapped the global m⁶A profiles using Nanopore direct RNA-seq in METTL3 knockdown and control CD4+ T cells. We first analyzed mRNA half-life data from CD4+ T cells (93) and confirmed that unstable transcripts were enriched in pathways related to T-cell identity (Supplementary Figure S10A–C). As expected, we observed a significant decrease in m⁶A modification levels in CD4+ T cells upon METTL3 knockdown (Figure 6A–C; Supplementary Figure S11A, B). The well-known T-cell identity genes, such as XBP1 (82), ITK (94), CCR7 (95) and SOCS1 (17), showed significantly decreased m⁶A modification levels (Figure 6D). RT-qPCR analyses of mRNA half-lives indicated that the transcripts of these T-cell identity genes became more stable upon METTL3 knockdown (Figure 6E). In contrast, housekeeping genes such as the RPL28, RPL34 and RPS15 showed little change of RNA m⁶A or stability upon METTL3 knockdown (Figure 6F, G). GSEA revealed that the genes from pathways related to T-cell identity, such as the T-cell differentiation and T-cell activation pathways, showed an increase in expression and a decrease in m⁶A modifications upon METTL3 knockdown (Figure 6H, I; Supplementary Figure S11C). In addition to these results from human CD4+ T cells, we also analyzed an independent RNA-seq dataset from mouse T cells (17). Consistently, we observed that expression of genes in pathways related to T-cell identity were up-regulated upon METTL3 knockout (Supplementary Figure S11E, F), suggesting that cell identity genes were vulnerable to the deficiency of METTL3. These results indicated that the low mRNA stability of cell identity genes is regulatable through mechanisms such as METTL3-mediated m⁶A modifications.

Figure 6.

METTL3 regulates the RNA stability of cell identity genes in activated CD4+ T cells. (A) m⁶A ratio determined by Nanopore direct RNA-seq for m⁶A sites of individual genes (site number = 41 710). (B) The density of m⁶A sites plotted across the transcript body. (C) Scatter plot to show the m⁶A ratio of individual mRNA sites (site number = 41 710). (D) Paired boxplot of the m⁶A ratio for individual sites on transcripts of reported cell identity genes (site number for XBP1 = 21; ITK = 23; CCR7 = 26; SOCS1 = 9; RHOH = 25). (E) The RNA decay curves of T-cell identity gene transcripts labeled with EU (ethynyl uridine). Time ‘0’ represents the input fully modified EU-mRNA (biological replicate number = 3). (F) Paired boxplot of the m⁶A ratio for individual sites on transcripts of housekeeping genes (site number for RPL28 = 9; RPL34 = 12; RPS15 = 5). (G) The RNA decay curves of housekeeping gene transcripts labeled with EU. Time ‘0’ represents the input fully modified EU-mRNA (biological replicate number = 3). (H) GSEA to show enrichment of the T-cell differentiation pathway in genes up- or down-regulated in response to knockdown of METTL3. The spectrum from blue to red indicates the mRNA fold change in shMETTL3 compared with shCTRL from the most up-regulated to the most down-regulated. Each vertical bar underneath the spectrum indicates one gene in the respective gene sets. (I) Violin plots to show RNA expression (left) and the m⁶A ratio (right) of T-cell differentiation pathway genes (gene number = 238). The three lines indicate quartile positions [Q1, Q2 (median) and Q3], and the black dot indicates the mean value. P-values were determined by two-tailed t-test (E, G) and paired Wilcoxon's tests (D, F, I).

Reprograming of RNA stability for cell identity genes in cell identity transition

Recent studies showed that mutations of m⁶A writers or readers affected cell differentiation and development (2,3,15,96,97). We therefore decided to investigate the dynamic change of RNA m⁶A modifications and their effects on RNA stability of cell identity genes in iPSCs relative to the parental fibroblasts using a published dataset (98). Consistent with the reported association of m⁶A with RNA decay (99,100), we observed higher m⁶A levels on unstable transcripts than on stable transcripts in iPSCs (Supplementary Figure S12A). Genes of unstable transcripts in iPSCs were enriched in pluripotency-related pathways [e.g. the KEGG signaling pathways regulating the pluripotency of stem cells (hsa04550)] (Figure 7A). Transcripts of genes from this pathway showed shorter half-lives in iPSCs than in fibroblasts (Figure 7B), despite higher mRNA expression levels in iPSCs (Supplementary Figure S12B). Meanwhile, the RNA m⁶A levels of these genes were higher in iPSCs than in fibroblasts (Figure 7C). Furthermore, RNA half-lives of fibroblast cell identity genes, defined by their association with fibroblast super-enhancers, were lower in fibroblasts than in iPSCs (Figure 7D); although their mRNA expression levels were higher in fibroblasts (Supplementary Figure S12C). Consistently, RNA m⁶A levels of fibroblast cell identity genes were higher in fibroblasts than in iPSCs (Figure 7E). For instance, transcripts of iPSC identity genes ID4 (101), BMP4 (102) and AXIN1 (103) showed shorter RNA half-lives and higher m⁶A modifications in iPSCs than in fibroblasts (Figure 7F, G). In contrast, the RNA half-life of fibroblast cell identity genes SNAI1 (104), EGFR (105) and GAS6 (106) was shorter in fibroblasts than in iPSCs (Figure 7F). We also observed higher m⁶A levels for the transcripts of these genes in fibroblasts than in iPSCs (Figure 7H). Taken together, these results revealed that RNA m⁶A modification and its effects on RNA stability were reprogrammed for cell identity genes during cell identity transition.

Figure 7.

Reprogramming of m⁶A modification on cell identity genes in the induction of iPSCs from fibroblasts. (A) GSEA of genes from the signaling pathways regulating pluripotency of stem cells to show their association with RNA stability. The spectrum from red to blue indicates the mRNA stability from the most stable to the most unstable, and each vertical bar underneath the spectrum indicates one gene in the respective gene sets. (B and C) Violin plot to show the half-life (B) and m⁶A signal (C) on transcripts of genes (gene number = 51) from the signaling pathways regulating pluripotency of stem cells. (D and E) Violin plots to show the half-life (D) and m⁶A signal (E) on transcripts of genes (gene number = 468) associated with fibroblast super-enhancers. The three lines indicate quartile positions [Q1, Q2 (median) and Q3], and the black dot indicates the mean value. (F) Dot plot to show the half-life difference of individual genes between iPSCs and fibroblasts. (G and H) Genome browser tracks to show m⁶A modification density on transcripts of iPSC identity genes (G) and fibroblast identity genes (H). P-values were determined by single-tailed paired Wilcoxon's test (B, C, D, E).

The low RNA stability of cell identity genes may also be regulated by some other mechanisms

In additional to the RNA m⁶A modification, we further observed that some other mechanisms might also regulate the RNA stability of cell identity genes. For example, adenylate–uridylate-rich elements (AREs) in the 3′-untranslated region (UTR) are well-known regulators of mRNA stability (107–109). Among putative cell identity genes defined by super-enhancer or broad H3K4me3, we found that 31–40% have m⁶A on RNAs and 21–35% have AREs (Supplementary Figure S13A–C). A-to-I editing and APA were also reported to regulate mRNA stability (110,111) and cell fate decisions (112,113) in cancers (114–116) and other diseases (117–119). We observed A-to-I editing of RNAs for ∼10% of putative cell identity genes (Supplementary Figure S13B, D, E). Interestingly, GSEA revealed little enrichment of the putative cell identity genes in genes that demonstrate preferential usage of distal or proximal poly(A) sites (Supplementary Figure S13F). These data suggest that more investigations are needed to fully understand the regulation of the low RNA stability of cell identity genes.

Lastly, we propose a working model in which cell identity genes have both a more active transcription and a faster RNA decay rate when compared with other genes such as housekeeping genes (Supplementary Figure S14). It was reported that the unique epigenetic signature of cell identity genes, e.g. the broad H3K4me3, is associated with increased release of Pol II from the initiation to the elongation steps of transcription (31). Pol II is known to interact with and recruit the METTL3/14 complex for co-transcriptional m⁶A modification of nascent RNA (120). Thus, the transcripts of cell identity genes, due to an as yet unclear mechanism, have a higher level of m⁶A added during transcription and, therefore, a lower RNA stability. The rapid mRNA turnover of cell identity genes may serve as a robust system to precisely regulate the homeostasis of transcripts and achieve optimal gene expression (Supplementary Figure S14).

DISCUSSION

The key discovery of this work is that cell identity genes tend to show a common pattern of RNA stability. We first observed this pattern in T cells (Figure 1). We then showed that the pattern also exists in five other cell types, including ESCs (Figure 2A). Notably, because different cell types have different cell identity genes, a cell identity gene tends to show low RNA stability in its cognate cell type and not in other cell types (Figure 2B). To show the regulatability of cell identity gene RNA stability during cell identity transition, we employed the iPSC induction model (Figure 7), which is one of the most frequently investigated models of cell identity transition. Because the data for some different cell types were generated by different labs using different protocols, we did not directly compare the difference in RNA stability values between these cell types. Instead, we ranked mRNAs by comparing their stability between different mRNAs within the same dataset. Thereafter, we analyzed the enrichment of cell identity genes in the most unstable mRNAs for each cell type. We found that for each of these cell types, the cell identity genes were always enriched in genes of unstable mRNAs.

Regulation of RNA decay is a key step in the regulation of the expression level of a gene. Our data show that transcripts of housekeeping genes tend to be stable, which is consistent with a recently published study (93), and the stability of housekeeping gene transcripts tends to be conserved between different cell types. In contrast, we found that unstable transcripts tend to be cell type specific and derived from cell identity genes. Although cell identity genes tend to have low RNA stability and fast RNA decay, their transcription frequency tends to be high. Consistent with the reports showing that super-enhancers and broad H3K4me3 tend to regulate cell identity genes (31,32,88), we found that genes regulated by super-enhancers and broad H3K4me3 have shorter RNA half-lives when compared with genes regulated by typical enhancers and sharp H3K4me3. The RNA stability of cell identity genes was dynamically reprogrammed, e.g. in the induction of iPSCs from fibroblasts. The transcripts of iPSC identity genes have more m⁶A modifications and shorter half-lives in iPSCs than in fibroblasts, and vice versa for putative fibroblast identity genes (Figure 7). These results suggest that cell identity genes are different from many other genes with respect to their post-transcriptional regulation of RNA expression.

Based on our results and recent publications, we propose a run-and-brake (run–brake) model of cell identity gene expression (Supplementary Figure S14). In this run–brake model, steady expression of cell identity genes is tightly regulated through both highly active transcription and fast RNA decay. This contrasts with housekeeping genes, for which steady expression results from relatively less transcription and slower RNA decay. While the high frequency of transcription at cell identity genes might be epigenetically regulated by super-enhancers and broad H3K4me3, the greater decay rate of transcripts might be epigenetically regulated by RNA modifications (e.g. m⁶A). Intriguingly, the RNA modification and decay rate of transcripts in the cytoplasm may have been pre-determined at the transcriptional stage in the nucleus through the interaction between RNA modification enzymes and Pol II on the gene body, transcription factors binding to super-enhancers or chromatin modifications such as the broad H3K4me3, which covers both the promoter and body of cell identity genes (Figure 5). After RNAs with a high level of m⁶A modifications in the nucleus were exported to the cytoplasm, they could be sorted synchronously to the fast track for decay. When the m⁶A modification or the associated decay factors are depleted, mRNAs transcribed at a high frequency would accumulate to cause abnormally high expression levels. For example, deletion of the m⁶A ‘writer’ protein METTL3 in mouse T cells decreases the mRNA decay rates of the genes Socs1, Socs3 and Cish. Consequently, the Mettl3-depleted naïve T cells were locked in the naïve state and proliferated more slowly than wild-type cells (17). Furthermore, a defect in YTHDF2-mediated m⁶A-dependent mRNA clearance was found to promote hematopoietic stem cell expansion due to abnormal overexpression of stem cell identity genes (96,97).

It was recently reported that METTL3 could interact with Pol II to bind chromatin, and that METTL3 could bind chromatin regions carrying H3K4me3 modifications (54,90). H3K36me3, a histone modification known to be associated with transcriptional elongation, was also reported to be recognized by METTL3 and to guide the m⁶A modification of RNA (120). Broad H3K4me3 modification was reported to be associated with increased transcriptional elongation of cell identity genes, whereas the sharp H3K4me3 modification was associated with transcription initiation (31). Our results indicated that METTL3 has stronger binding intensity on genes associated with broad H3K4me3 or super-enhancers when compared with genes associated with sharp H3K4me3 or typical enhancers (Figure 5; Supplementary Figures S6 and S7). Furthermore, we observed a larger number of m⁶A modification sites and lower stability of transcripts for genes associated with super-enhancers and broad H3K4me3 than for transcripts of genes marked by typical enhancers and sharp H3K4me3 (Figure 5; Supplementary Figures S6 and S7).

The underlying molecular mechanism by which METTL3 preferentially binds cell identity genes for co-transcriptional m⁶A modification is not yet clear. Previous reports indicated that the miRNA-processing machine DGCR8 complex was recruited to miRNA genes associated with super-enhancers for co-transcriptional processing of pre-miRNAs and pre-mRNAs (55). We observed the preferable binding of DGCR8 at protein-coding genes associated with super-enhancers. DGCR8 is known to interact with METTL3 and bind to transcriptionally active coding and non-coding genes in a METTL3- and transcription-dependent manner (55,89). Therefore, it is possible that DGCR8 might mediate the preferential binding of METTL3 on cell identity genes, which were known to be regulated by super-enhancers. However, the underlying mechanism that connects DGCR8 to mRNA decay will need to be comprehensively investigated. Cell identity genes display broad H3K4me3 that covers both promoters and gene bodies, whereas many other active genes such as housekeeping genes display sharp H3K4me3 that covers only the promoter region (31). METTL3 could bind chromatin regions with H3K4me3 modifications (54,90). Therefore, it is also possible that the broad H3K4me3 on the gene bodies of cell identity genes might mediate the preferential binding of METTL3 on cell identity genes. Both hypotheses require comprehensive molecular and biochemical experiments to confirm the causal relationship between super-enhancers or broad H3K4me3 and the binding of METTL3, the co-transcriptional m⁶A modification and the low RNA stability of cell identity genes.

Many previous studies have compared tissue-specific genes with housekeeping genes (121,122). Cell identity genes are related to cell type-specific genes but are conceptually not the same. Although some cell identity genes can be very specific to one cell type, other cell identity genes might be expressed in multiple or many cell types. For instance, the expression of OCT4 drives the specification of both ESCs and primitive endoderm lineage commitment when it cooperates with SOX2 and SOX17, respectively (123). We recently performed a manual curation of 247 reported cell identity genes for 10 cell types (62). These genes belong to four major categories: (i) master transcription factors, which drive differentiation toward a cell type when their expression is ectopically induced in another cell type; (ii) required transcription factors, whose depletion impairs differentiation toward a specific cell type; (iii) genes required for key functions or phenotypes of a cell type; and (iv) genes that are widely used as markers for a cell type. Our previous analysis indicates that in each of these categories, different cell identity genes show very different degrees of expression specificity (62).

In our investigation, we utilized a comprehensive set of data including mRNA half-life profiling data, ChIP-seq data of histone modifications and METTL3, and m⁶A modification profiling data. There can be multiple layers of technical variations, e.g. variation between RNA stability based on transcriptional inhibition and metabolic labeling of newly synthesized RNAs, variation caused by different time intervals used for stability measurements and variation caused by different experimental conditions between laboratories. Our comparative analysis in this work was divided into three types based on careful consideration of these technical variations. (i) Analyze the difference in each feature value, e.g. the RNA stability of each gene between fibroblasts and iPSCs. This comparison is feasible because each type of feature value was generated by the same lab using the same protocol in the same published paper. The original publication also demonstrated that the data are comparable between the two cell types. (ii) Analyze the enrichment of the top unstable mRNAs in cell identity genes of each cell type. We performed this analysis for multiple cell types. Because the data for some different cell types were generated by different labs using different protocols, we did not directly compare the differences in RNA stability between cell types. Instead, we ranked mRNAs by comparing RNA stability between different mRNAs within the same dataset. Thereafter, we analyzed the enrichment of cell identity genes in the genes of the top unstable mRNAs for each cell type. We found that for each of these cell types, the cell identity genes are always enriched in genes of unstable mRNAs. The observed enrichment is not specific to the lab that generated the data, a protocol used to generate the RNA stability data or a cell type. Therefore, our results indicate that the observations are resistant to technical variation. (iii) We observed that the cell identity genes/pathways tend to be the most enriched in genes of unstable RNAs from the cognate cell type when compared with unstable RNAs from other cell types. A potential concern is that this cell type specificity is due to technical variation of the data between cell types. Therefore, we analyzed RNA stability data generated by different methods and labs for the same cell type (Supplementary Figures S1 and S2) and found that the enrichment of cell identity genes in genes of unstable RNAs is not specific to a method or a lab that generated the data. Taken together, we carefully considered potential technical variations to design appropriate analyses and, when necessary, performed additional analyses to demonstrate that the observation was not dependent on a technical difference.

Taken together, our results reveal low RNA stability as a key post-transcriptional regulatory signature of cell identity genes. We demonstrated an association between low RNA stability and super-enhancers or broad H3K4me3. This association may be mediated by a co-transcriptional m⁶A modification of RNA. While the link between RNA m⁶A modification and low RNA stability is well known, more evidence is needed to understand the association between RNA modification, Pol II activity and chromatin state (including histone modifications and the binding of transcription factors to enhancers). A comprehensive understanding of these associations could be critical to reveal the exact molecular mechanisms for precise regulation of cell identity gene expression.

DATA AVAILABILITY

Newly generated genomic data in this manuscript were deposited to the GEO database with the accession number GSE185990. GEO accession numbers of public datasets analyzed in this project are listed in Supplementary Table S4.

SUPPLEMENTARY DATA

Supplementary Data are available at NAR Online.

FUNDING

The National Institute of Health [R01GM138407, R01GM125632, R01HL148338 and R01HL133254 to K.C., R01CA208257 and R01CA256741 to Q.C. and R01HL155632 to L.Z.]; the U.S. Department of Defense [W81XWH-17-1-0357, W81XWH-19-1-0563 and W81XWH-20-1-0504 to Q.C.]; Prostate SPORE [P50CA180995 Development Research Program to Q.C.]; and the Polsky Urologic Cancer Institute of the Robert H. Lurie Comprehensive Cancer Center of Northwestern University at Northwestern Memorial Hospital [to Q.C.]. The Nanopore direct RNA-seq library preparation and sequencing was done at Northwestern University NUseq facility core with the support of NIH grant 1S10OD025120.

Conflict of interest statement. None declared.