Genome-wide identification and characterization of circular RNA m⁶A modification in pancreatic cancer

Abstract

Background

N⁶-methyladenosine (m⁶A) is the most abundant modification of RNA in eukaryotic cells and play critical roles in cancer. While most related studies focus on m⁶A modifications in linear RNA, transcriptome-wide profiling and exploration of m⁶A modification in circular RNAs in cancer is still lacking.

Methods

For the detection of m⁶A modification in circRNAs, we developed a new bioinformatics tools called Circm6A and applied it to the m⁶A-seq data of 77 tissue samples from 58 individuals with pancreatic ductal adenocarcinoma (PDAC).

Results

Circm6A performs better than the existing circRNA identification tools, which achieved highest F1 score among these tools in the detection of circRNAs with m⁶A modifications. By using Circm6A, we identified a total of 8807 m⁶A-circRNAs from our m⁶A-seq data. The m⁶A-circRNAs tend to be hypermethylated in PDAC tumor tissues compared with normal tissues. The hypermethylated m⁶A-circRNAs were associated with a significant gain of circRNA-mRNA coexpression network, leading to the dysregulation of many important cancer-related pathways. Moreover, we found the cues that hypermethylated m⁶A-circRNAs may promote the circularization and translation of circRNAs.

Conclusions

These comprehensive findings further bridged the knowledge gaps between m⁶A modification and circRNAs fields by depicting the m⁶A-circRNAs genomic landscape of PDAC patients and revealed the emerging roles played by m⁶A-circRNAs in pancreatic cancer. Circm6A is available at https://github.com/canceromics/circm6a.

Supplementary Information

The online version contains supplementary material available at 10.1186/s13073-021-01002-w.

Background

m⁶A is the most abundant posttranscriptional modification in eukaryotic RNAs and play critical roles in various normal bioprocesses [1]. The m⁶A modification is deposited by the m⁶A methyltransferase complex (writer) composed of METTL3, METTL14, and WTAP and can be removed by m⁶A demethylases (erasers) such as FTO and ALKBH5. Distinct proteins (readers) can recognize m⁶A-modified mRNAs and decide their fates by affecting the stability, splicing, nuclear export, and translation of target RNAs [2, 3]. High-throughput m⁶A RNA immunoprecipitation sequencing (MeRIP-seq) studies revealed that m⁶A modification mainly occurs in messenger RNAs (mRNAs) with the consensus motif “RRm⁶ACH” (R = G or A; H = A, C or U) and is enriched near stop codons and in the 3′UTRs of mRNAs [4, 5].

Circular RNAs (circRNAs) are a special class of RNAs that are produced by a covalent linkage between the 5′ and 3′ ends of an RNA molecule [6]. In recent years, circRNAs have been intensively studied in a wide range of cells and tissues, revealing their critical roles in many important biological processes and diseases [7–9]. Compared to linear RNAs, circRNAs are highly stable and are frequently detected in exosomes and cell-free body fluids; therefore, circRNAs may be potential biomarkers or therapeutic targets in precise medicine [10, 11]. While most m⁶A studies focus on linear RNAs, a few studies show that m⁶A modifications also occur in circular RNAs [12], and m⁶A modifications affect the translation efficiency of circRNAs [13]. Moreover, there is evidence showing that m⁶A modifications in circRNAs play important roles in cancer progression [14]. However, to date, few studies on the transcriptome-wide mapping of m⁶A modifications in circRNAs in human cancer tissues have been reported.

Pancreatic ductal adenocarcinoma (PDAC) is an aggressive tumor that is typically diagnosed at an advanced stage and will become the second leading cause of cancer mortality within this decade [15, 16]. To date, there is still no efficient method for the treatment of PDAC, posing an urgent need for the development of new therapeutic strategies, which relies on a clearer understanding of cancer biology. Recent genome-wide association studies (GWAS) and whole-genome/exome/transcriptome sequencing studies on PDAC have revealed a complex network consisting of genetic and genomic/transcriptomic alterations that is closely associated with PDAC occurrence and development [17, 18]. There is widespread dysregulation of circRNAs, which has critical roles in cancer progression [9, 19]. However, little is known about the transcriptome mapping and functions of m⁶A modifications in circRNAs in PDAC.

In this study, we presented a de novo algorithm called “Circm6A” to detect m⁶A modification in circular RNAs from MeRIP-seq data, facilizing further researches on m⁶A modification in circRNAs. We then performed m⁶A sequencing (m⁶A-seq) on total RNAs of 53 tumor and 24 adjacent normal tissue samples from 58 individuals with PDAC and applied Circm6A to the m⁶A-seq data to explore the landscape of m⁶A modification in circRNAs. We found that approximately 23.1% of circRNAs in PDAC tissues harbored m⁶A modifications, and they tended to be hypermethylated in PDAC tumor tissues compared to adjacent normal tissues. Intriguingly, we found that hypermethylated m⁶As in circRNAs caused a gain of circRNA-mRNA coexpression in many cancer-related pathways. Moreover, our study also indicated that m⁶A modifications in circRNAs might promote the circularization and translation of circRNAs. Taken together, we provided a state-of-art tool to detect m⁶A modification in circRNAs, which will facilitate the following study of m⁶A modification in circRNA. Our findings shed new light on the function and mechanism of circRNA dysregulation in PDAC at RNA epigenetics level.

Methods

Construction of Circm6A

Circm6A (https://github.com/canceromics/circm6a) is a de novo algorithm for the detection of circRNAs and their m⁶A modifications from MeRIP-seq data. Circm6A will scan the sequencing alignment files from both MeRIP-seq IP (the IP library represents the RNA fragments captured by m⁶A-antibody pull-down) and INPUT (the paired INPUT library is derived from initial fragmented RNAs before immunoprecipitation) samples [4]. For the detection of circRNAs, Circm6A first detects junction reads with PCC signals (two segments of one read are aligned to the reference genome in a chiastic order) that reflect a circRNA candidate. Paired-end mapping signals (PEM signal, the paired mate of the BSJ read is properly aligned within the candidate circRNA region) are utilized for additional filtering when paired-end sequencing was used, as previously described [4], and this step will be automatically skipped in single-end sequencing data. Furthermore, Circm6A checks whether AG and GT dinucleotides and exon boundaries flank the back-splicing junction (BSJ) sites of candidate circRNAs. The identified circRNAs are annotated according to the GTF file downloaded from the GENCODE database (https://www.gencodegenes.org/). If only RNA-seq data is provided (parameter: -input input_sample.bam), Circm6A will only detect circRNAs from files.

For the detection of circRNAs with m⁶A modification (m⁶A-circRNAs), Circm6A will examine whether the circRNAs are significantly enriched in IP samples compared to INPUT samples in the MeRIP-seq data. Firstly, the whole genome is split into 25-bp bins. The read coverages in the genome bins around the BSJ sites of all identified circRNAs are calculated in the IP samples and INPUT samples, respectively. Then, the read coverage difference for the genome bins around BSJ sites between IP and INPUT samples are calculated using Fisher’s exact test. The P values from Fisher’s exact test are adjusted using the Benjamini-Hochberg method. A false discovery rate (FDR) of less than 0.05 is considered significant. The genome bins with significant enrichment in IP samples over INPUT samples are concatenated. If the concatenated bins are across the BSJ site, the circRNA with that BSJ site is considered as a candidate m⁶A-circRNA. Next, the loose and strict filters are applied to further filter the candidate m⁶A-circRNAs. If there is at least one BSJ read in IP samples (N_{BSJ, IP} ≥ 1), the circRNA was defined as low confidence m⁶A-circRNA. If the fraction of BSJ reads should be higher in IP samples than in INPUT samples ($\frac{N_{\mathit{BSJ},\mathit{IP}}}{N_{\mathit{\text{total}},\mathit{IP}}}:\frac{N_{\mathit{BSJ},\mathit{\text{non}}-\mathit{IP}}}{N_{\mathit{\text{total}},\mathit{\text{non}}-\mathit{IP}}}\ge 1.0)$ and there is at least one BSJ read in IP samples (N_{BSJ, IP} ≥ 2), the circRNA was defined as high-confidence m⁶A-circRNA.

m⁶A peaks in linear RNAs are also detected according to previously published methods [4].

Simulate MeRIP-seq data for performance evaluation

We developed an in-house simulator called MeRIP-simulator (https://github.com/canceromics/MeRIP-simulator) to simulate the MeRIP-seq data. GRCh38 genome FASTA and GTF annotation files (GENCODE version 25 [20], downloaded from https://www.gencodegenes.org/) were used as references for simulation.

For the simulation of MeRIP-seq Input sample, the linear genes were simulated based on the linear transcript information extracted from the GENCODE GTF annotation file and the circRNA genes were simulated based on the circRNA transcript information extracted from the five circRNA database (CircAtlas [21], MiOncoCirc [22], CSCD [23], TSCD [24], CIRCpedia [25]). A total of 18,689 circRNAs were simulated. The read counts of a transcript (C) were simulated using three coefficients: C = F · d · l, where F is the expression factor of the transcript, d is sequencing depth and l is the length of exons in the transcript. The expression factor of a transcript is defined as F = e^λ and λ is based on a Poisson distribution P(λ = 1.0). The sequencing depth is given by user. The read length R follow a binominal distribution $R~B\left(⌈\frac{L_r}{0.95}⌉,0.95\right)$, where L_r means expected read length given by Tuser or 150 bp by default. We let the insert size S follow another binominal distribution S~B(L_s · 2, 0.5), where L_s means expected insert size given by user or 300 bp by default. The simulated reads will be outputted as FASTQ format.

For the simulation of MeRIP-seq IP sample, the m⁶A sites were simulated based on m⁶A peaks obtained from public MeRIP-seq dataset (GSE120229 [26]: SRR7881528 and SRR7881532). The m⁶A peak was called using STAR [27] and MACS2 [28]. The parameters of STAR were “--twopassMode Basic --chimSegmentMin 20 --outFilterIntronMotifs RemoveNoncanonical --outFilterMultimapNmax 20 --alignIntronMin 20 --algigIntronMax 1000000 --alignMatesGapMax 1000000”. The cutoff of the P value for significant peak for MACS2 was set at 1.0e−6. Two library construction strategies (fragmentation of RNA before MeRIP and fragmentation of RNA after MeRIP) were simulated. For the strategy of fragmentation of RNA after MeRIP, the read coverages of the linear and circular transcripts with m⁶A modifications were simulated 20-fold higher than the read coverages of those linear and circular transcripts without m⁶A modifications. For the strategy of fragmentation of RNA before MeRIP, the read coverages of the m⁶A regions (200 bp) in linear and circular transcripts were simulated 20-fold higher than the read coverages of non-m⁶A regions. We also simulated background reads to mimic the nonspecific IP of m⁶A antibody in IP sample and their abundance was 0.01-fold of that in INPUT sample.

Collection of PDAC samples

A total of 77 (53 cancer and 24 normal) tissues were obtained from pancreatic cancer patients who underwent pancreatectomy at the Sun Yat-sen University Cancer Center and Sun Yat-sen Memorial Hospital between 2010 and 2018 (detailed in Additional file 1: Table S1). The PDAC tumor and non-tumor tissue (≥ 5 cm away from tumor) samples were collected at surgery from each patient and immediately placed in liquid nitrogen. All patients received no chemotherapy or radiotherapy before surgery. The diagnosis of PDAC was histopathologically confirmed and tumor stage was classified according to the 7th edition of AJCC Cancer Staging System [29]. Tumor samples and distant normal tissues were embedded in optimal cutting temperature medium, and histological sections stained with hematoxylin and eosin were reviewed by at least two pathologists for quality assurance that tumor specimens contained at least 60% tumor cell nuclei, whereas normal specimens contained no tumor cells. Clinical data for PDAC patients were collected, which included age at diagnosis, differentiation, lymph node metastasis, vascular invasion, neural invasion, tumor stage, and treatment. The survival time of individuals with PDAC was measured from the date of diagnosis to the date of last follow-up or death. Whether and when a subject had died were obtained from inpatient and outpatient records, subjects’ families, or through follow-up telephone calls.

Public RNA-seq data

We included two separate independent circRNA-related studies [7, 30], which included control (only rRNA-depleted) samples and matched samples additionally treated with RNase R. The first dataset contains 4 runs of RiboMinus RNA-Seq libraries of the HeLa cell line downloaded from the NCBI Sequence Reads Archive (accession numbers: SRR1636985, SRR1636986, SRR1637089, and SRR1637090) [31]. The second public dataset comprises 2 runs of rRNA-depleted RNA-seq data derived from the Hs68 cell line (accession numbers: SRR445016 and SRR444975) [7].

Public CLIP-seq data

All reported binding sites for “readers” (YTHDF1-3, YTHDC1-2, IGF2BP1-3, HNRNPC, and HNRNPA2B1) and EIFs (EIF3A, EIF3B, EIF3D, EIF3G, EIF3H, EIF4A3, EIF4G2) were obtained from POSTAR2 [32]. MeRIP-seq data of hESC cell line were retrieved from Gene Expression Omnibus (GEO accession number: GSE55572 [33] and GSE85324 [12]). ADAR binding sites of hESC cell line were obtained from a public iCLIP-seq study (GEO accession number: GSE63709 [34]).

Detection of circRNAs in public and simulated data using eight known tools

For each public and simulated RNA-seq sample, circRNA detection was also performed using other eight known tools (ACFS [35], AutoCirc [12], circRNA_finder [36], CIRI2 [31, 37], CIRCexplorer2 [38], DCC [39], Find Circ [8], and MapSplice [40]) according to the suggested parameters of each tool. RNase R-resistant (enriched) circRNAs were defined as circRNAs with at least a 5-fold enrichment of abundance in RNase R-treated samples, RNase R-sensitive (depleted) circRNAs were defined as those circRNAs with an abundance reduction in RNase R-treated samples, and RNase R-unaffected circRNAs were defined as those with a 1~5-fold change in RNase R-treated samples, as previously described [41].

Sample processing

Total RNA from tissue samples was extracted with TRIzol reagent (Invitrogen). RNA samples were quantified by measuring the absorption value at 260 nm with a UV spectrophotometer and then analyzed via the RNA6000 Nano assay (Agilent) for determination of an RNA integrity number (RIN), and only analytes with an RIN ≥ 7.0 were included in this study.

m⁶A-specific methylated RNA immunoprecipitation and high-throughput sequencing (MeRIP-seq)

Total RNA from tissue samples was isolated using TRIzol (Life Technologies), digested by DNase I (NEB), and then subjected to two rounds of RiboMinus treatment to reduce rRNA content (Illumina). For m⁶A immunoprecipitation, the Magna MeRIP m⁶A Kit (Millipore) was used according to the manufacturer’s instructions. Briefly, 20 μg of Ribo-off-treated RNA was sheared to approximately 100–200 nt in length by metal-ion-induced fragmentation, purified, and incubated with 10 μg of anti-m⁶A antibody (Synaptic Systems, 202003)-conjugated beads in 500 μL 1× immunoprecipitation buffer supplemented with RNase inhibitors at 4 °C overnight. The m⁶A-modified RNA was recovered by treatment with proteinase K, acidic phenol/chloroform extraction, and ethanol precipitation. Sequencing libraries were prepared using the Illumina protocol, and sequencing was performed on an Illumina HiSeq2500.

m⁶A peak-specific qRT-PCR

Total RNA from tissue samples was isolated according to the procedure shown above. For RNase R treatment, 20 μg of total RNA was incubated for 15 min at 37 °C with or without 5 U/μg RNase R (Epicenter Technologies). For MeRIP-qPCR, the RNA samples were fragmented and immunoprecipitated by anti-m⁶A antibody as described above. The purified m⁶A-containing RNAs were reverse transcribed using a RevertAid First Strand cDNA Synthesis Kit (Thermo) with random hexamers. The enrichment of m⁶A was quantified via quantitative real-time PCR in triplicate on a Roche LightCycler 480 using the SYBR Green method. Gene-specific primers are shown in Additional file 1: Table S2.

Quantitative real-time PCR (qRT-PCR) analysis

Total RNA from tissue samples was isolated and subjected to RNase R treatment according to the procedure shown above. Reverse transcription for mRNA and circRNAs was performed with a RevertAid First Strand cDNA Synthesis Kit (Thermo) using random hexamers. Relative RNA levels were determined by qRT-PCR in triplicate on a Roche LightCycler 480 using the SYBR Green method. β-Actin was employed as an internal control for quantification of the level of each gene. The primer sequences for the analyzed genes are summarized in Additional file 1: Table S2. The relative expression levels were calculated according to the 2^−ΔCT method.

Cell lines and cell culture

Human PDAC cell line PANC-1 was purchased from the Cell Bank of Type Culture Collection of the Chinese Academy of Sciences Shanghai Institute of Biochemistry and Cell Biology. PANC-1 was maintained in DMEM medium supplemented with 10% fetal bovine serum in an atmosphere of 5% CO₂ and 99% relative humidity at 37 °C. Cells passaged for < 6 months were authenticated by DNA fingerprinting analysis using short-tandem repeat (STR) markers.

RNA interference

PDAC cells were pre-seeded in 6 cm dish with 50–60% confluency before night. Small interfering RNA (siRNA) targeting METTL3 and scramble control ((Additional file 1: Table S2) were purchased from GenePharma. Transfection of siRNA (75 nM) or scramble control was performed using lipofectamine 3000 (Life Technologies) according to the manufacturer’s instructions. The transfected cells were harvested 48 or 72 h after transfection and followed by downstream application. Efficiency of RNA interfering was validated by qRT-PCR as described above.

Metabolic labeling of nascent RNAs with 4sU and nascent RNA purification

Metabolic labeling of newly transcribed RNAs was performed as described previously [42, 43]. PANC-1 cells were incubated with 100 μM 5, 6-dichloro-1-β-D-ribofuranosylbenzimidazole (DRB) for 3 h to block Pol II transcription. Transcription recovered after DRB release and newly transcribed RNAs were labeled with 200 μM 4sU. After terminating transcription with TRIzol treatment, 200 μg total RNAs were used for biotinylation and purification of 4sU-labeled nascent RNA. After incubating with 0.2 mg/mL EZ-link biotin-HPDP (Pierce, 21341, dissolved in dimethylformamide (DMF, Sigma, D4551) at a concentration of 1 mg/mL) in biotinylation buffer (10 mM Tris PH 7.4, 1 mM EDTA) at room temperature (RT) for 1.5 h with rotation, bound biotin-HPDP were purified with chloroform and precipitated using equal volume of isopropanol and 1:10 volume of 5 M NaCl. Biotinylated 4sU-labeled RNAs were isolated using 150 μL streptavidin-coated magnetic beads (Invitrogen) for 30 min at RT. After substantially washed on beads, nascent RNAs were eluted twice with 100 μL 0.1 M dithiotheitol (DTT) and precipitated with 40 μL of 4 M LiCl, 2 μL glycogen, and 600 μL ice-cold ethanol.

RNA immunoprecipitation assays

RNA immunoprecipitation (RIP) assays were performed using the Magna RIP RNA-Binding Protein Immunoprecipitation kit (Millipore). 2 × 10⁶ Cells were rinsed twice with ice-cold phosphate buffered saline (PBS), harvested and then centrifuged at 1000 rpm for 5 min at 4 °C. Cell lysates were centrifuged at 12,000 rpm for 15 min at 4 °C and the supernatants were precleared with 15 μL Dynabeads Protein G (Invitrogen) to get rid of nonspecific binding. Then, the precleared lysates were used for IP by incubating with anti-EIF3A (ab86146) or anti-EIF3B (ab133601) antibodies from Abcam or anti-EIF3H (#PA5-87290) antibody from Invitrogen at 4 °C overnight with rotation. Protein-RNA mixture were substantially washed on beads, followed by extraction with phenol/chloroform/isoamyl alcohol and subjected to qRT-PCR using the primers listed in Additional file 1: Table S2.

Polysome profile analysis

PANC-1 cells (1 × 10⁷) were treated with 100 mg/mL cycloheximide (Sigma-Aldrich) at 37 °C for 5 min, washed twice with ice-cold PBS containing 100 mg/mL cycloheximide, and lysed with 500 μL polysome lysis buffer (15 mM Tris-HCl, 5 mM MgCl₂, 100 mM KCl, 2 mM DTT, 1% Triton X-100, 100 μg/mL cycloheximide, 1 mg/mL heparin sodium). Cell lysates were incubated on ice for 10 min and centrifuged at 16,000×g at 4 °C for 7 min. The supernatant was loaded onto top of a 5–50% sucrose gradient followed by centrifuging at 222,228×g at 4 °C for 120 min (Beckman). The gradient was divided and collected into 15 fractions by monitoring RNA absorbance at 254 nm with an ISCO fractionator (Brandel). RNA from each fraction was extracted using TRIzol reagent (Invitrogen) and analyzed by qRT-PCR. RNA distributions across the polysome profile are presented as percentages.

Identification and quantification of circRNAs in PDAC and normal tissue samples

Reads were first mapped to the human reference genome (GRCh38) by BWA (v0.7.15) [44] and the annotation file obtained from the GENCODE database (https://www.gencodegenes.org/). Circm6A was applied to detect the circRNAs and their m⁶A modifications from all our PDAC tissue and normal tissue samples. Only circRNAs that had read counts ≥ 2 and were detected in at least 2 samples were retained. The circRNA expression level was calculated by the spliced reads per billion mapped reads (SRPBM) method, as previously mentioned [7]. Differential circRNA abundance analysis was carried out with edgeR [45] based on the quantification results from Circm6A. The thresholds for differential expression were a P ≤ 0.05 and a fold change ≥ 1.5 between PDAC and normal tissue samples.

Construction of a random forest model for prediction of m⁶A modifications in circRNA

m⁶A-circRNAs identified by Circm6A were defined as the positive set. Non-m⁶A-circRNAs identified by Circm6A were defined as the negative set. We included some significant differential features of circRNA characteristics, the relative position between a circRNA and the m⁶A peak region of colinear transcripts, and related features of host genes as previously reported [46]. Finally, the top 14 features were selected for the construction of the random forest model (detailed in Additional file 1: Table S3).

Relative m⁶A level quantification and differential methylation analysis

The relative m⁶A level for each peak of m⁶A-circRNA was quantified according to the approach described by Schwartz et al. [33]. Briefly, the read coverage for each peak in IP and INPUT (control) samples was calculated during the peak-calling module of Circm6A. The spliced reads per billion mapped reads (SRPBM) method was then used to normalize the read coverage. The relative m⁶A level was obtained by calculating the ratio between the IP SRPBM value and the INPUT SRPBM value for each m⁶A-circRNA. To obtain the dysregulated m⁶A modifications of circRNAs between PDAC and normal tissue samples (DM circRNAs), we utilized the Wilcoxon rank sum test on common peaks between PDAC and normal tissue samples (FDR ≤ 0.05; absolute fold change ≥ 1.5). CircRNAs that were specifically m⁶A modified in PDAC samples were also included in the hypermethylated-m⁶A-circRNAs; in the same way, those specifically m⁶A modified in normal samples were included in the hypo-m⁶A-circRNAs.

Survival analysis with the Cox proportional hazards model

To independently investigate the prognostic role of each hyper-/hypomethylated m⁶A-circRNA, the relative m⁶A level of the dysregulated m⁶A-circRNAs, survival data in terms of OS/PFS and other clinical features, such as sex, age, were subjected to multivariate analysis using the Cox proportional hazards model. Hyper-/hypomethylated m⁶A-circRNAs with an FDR ≤ 0.25 were identified as significantly correlated with PFS/OS.

mRNA quantification and differential expression analysis

mRNA levels were quantified using RSEM [47] (parameters: “-paired-end, -star”) on RiboMinus-treated (control/input) RNA-seq data. mRNA expression was measured in fragments per kilobase per million (FPKM). The “DESeq2” R package [48] was applied for differentially expressed mRNAs (DE mRNAs). mRNAs with an expression absolute fold change ≥ 1.5 and an FDR ≤ 0.05 were identified as DE mRNAs.

Construction of the circRNA-mRNA coexpression network

Similarities between circRNA and mRNA expression patterns were determined by computing a Pearson correlation coefficient matrix for each circRNA-mRNA pair, and pairs with a correlation (r) ≥ 0.5 were defined as coexpression pairs. To avoid false positives in the coexpression network analysis, the nodes on the network were restricted to mRNAs or circRNAs expressed in at least 3 tissues, as previously described [21]. TargetScan miRNA site predictions were used to find RNAs (circRNAs and mRNAs) with MREs [49], and miRTarBase [50] was used as a further experimentally validated source. Coexpression pairs sharing the same MERs were ceRNAs [21].

Pathway enrichment analysis

All pathway enrichment analyses of KEGG and GO terms were performed with clusterProfiler [51]. KEGG pathways or GO terms with an FDR ≤ 0.05 were considered significantly enriched.

Permutation test

The Bioconductor package regioneR [52] was employed to test the significance of the overlap between two sets of regions. Peak sets were shuffled 10,000 times for each of the permutation tests. The widths of the peaks and total number of peaks from the tested sets were used for the widths of the random peaks. The following options were used for all three permutation tests: permTest (ntimes = 10,000, randomize.function = randomizeRegions, evaluate.function = numOverlaps, count.once = TRUE).

Results

De novo algorithm to identify m⁶A modification in circRNAs from m⁶A-seq data

We developed a de novo algorithm called Circm6A to detect circRNAs and their m⁶A modifications from m⁶A-seq data. The workflow of Circm6A is illustrated in Fig. 1. First, Circm6A identifies circRNAs by searching back-splicing junction (BSJ) reads with paired chiastic clipping signals (PCC signals) from both MeRIP-seq IP and INPUT read alignments. If paired-end data are dealt with, paired-end mapping signals (PEM signals) will be automatically utilized to check that mapping of mates properly matches within the relevant circRNAs, as previously described [31]. To further remove potential false positives, the candidate circRNAs are filtered with additional features such as splicing signals (such as “GT-AG”) and exon boundaries. Next, we identify the circRNAs with m⁶A modification (m⁶A-circRNAs) by examining whether the circRNAs are significantly enriched in IP samples compared to INPUT samples. To do this, we count the reads coverage in the up- and downstream 100-bp regions around the BSJ sites of all circRNAs in the IP samples and INPUT samples, respectively. We split the 200-bp BSJ region into 25-bp bins and concatenate the bins with significant enrichment in IP samples over INPUT samples (Fisher’s exact test). If the concatenated bins are across the BSJ site and there is at least one BSJ read in IP samples, the circRNA with that BSJ site is considered as a m⁶A-circRNA. To obtain high-confidence m⁶A-circRNAs, we apply additional filters: first, there should be more than one BSJ reads in IP samples; second, the fraction of BSJ reads should be higher in IP samples than in INPUT samples. In addition, m⁶A peaks in linear RNAs are also detected according to previously published methods [4]. The full method and implementation details are given in the “Methods” section.

Fig. 1.

The framework of Circm6A. The pipeline of Circm6A consists of two steps: first, identification of circRNAs from MeRIP-seq IP and INPUT samples according to PCC and PEM signals; second, identification of circRNAs and linear transcripts with m⁶A modifications by examining whether the circRNAs or regions in linear transcripts are significantly enriched in IP samples compared to INPUT samples in the MeRIP-seq data. PCC, paired chiastic clipping signals; PEM, paired-end mapping signals

To evaluate the performance of Circm6A in the identification of circRNAs, we compared Circm6A with eight known circRNA detection tools including ACFS [35], AutoCirc [12], circRNA_finder [36], CIRI2 [31, 37], CIRCexplorer2 [38], DCC [39], Find Circ [8], and MapSplice [40], which apply different strategies to detect circRNAs. The fraction of RNase R-depleted circRNAs is usually used to represent the false positives [41]. By utilizing two public datasets (HeLa and Hs68 cell lines) that consist of RNase R-treated or untreated samples, we calculated the fraction of RNase R-depleted circRNAs by comparing the circRNA level in the RNase R-treated samples with that in untreated samples. As a result, Circm6A detected 32836 circRNAs in the two public datasets and 15.4% of circRNAs were RNase R-depleted, which was comparable to the performance of the other eight tools (Fig. 2A). Moreover, we evaluated the performance with simulated RNA-seq datasets generated with our in-house simulator. Among the simulated datasets, Circm6A achieved the highest F1 score (0.99) compared to other eight tools (average F1 score: 0.91, F1 score: 0.62 ~ 0.94) (Fig. 2B and Additional file 1: Table S4).

Fig. 2.

Performance evaluation of Circm6A. A Bar plot showing the performance of Circm6A and other eight known circRNA detection tools in the identification of circRNAs using public RNA-seq dataset consist of RNase R-treated or untreated samples. RNase R-depleted represents circRNAs that were depleted after RNase R treatment. Unaffected represents circRNAs that were enriched 1–5-fold after RNase R treatment. RNase R resistant represents circRNAs that were enriched ≥ 5-fold after RNase R treatment. B Bar plot showing the performance of Circm6A and other eight known circRNA detection tools in the identification of circRNAs using simulated datasets. C Bar plot showing the performance of Circm6A and other eight known circRNA detection tools in the identification of m⁶A-circRNAs using simulated MeRIP-seq datasets with fragmentation of RNA after MeRIP. D Bar plot showing the performance of Circm6A and other eight known circRNA detection tools in the identification of m⁶A-circRNAs using simulated MeRIP-seq datasets with fragmentation of RNA before MeRIP. BSJ-proximal, distance of m⁶A in circRNAs to BSJ within fragment length; BSJ-distal, m⁶As of circRNAs distal to BSJ (distance to BSJ over fragment length). E Saturation curve to access the impact of sequencing depth on the identification of circRNAs and m⁶A-circRNAs, respectively. M, million

To further evaluate the performance of Circm6A in the identification of m⁶A-circRNAs, we developed an in-house simulator (MeRIP-simulator, see “Methods” section) to simulate MeRIP-seq data. There are two MeRIP-seq library construction strategies, one which conducts fragmentation of RNA before m⁶A antibody IP (MeRIP) and the other conducts fragmentation of RNA after MeRIP, which theoretically have great impact on the identification of m⁶A-circRNAs (Additional file 1: Fig. S1A). Therefore, we evaluated the performance of Circm6A by simulating both of the two MeRIP-seq library construction strategies. For the strategy of RNA fragmentation after MeRIP, we simulated three MeRIP-seq datasets with sequencing depth range from 60 to 80 million reads. Previous study on genome-wide detection of m⁶A-circRNAs treated those circRNAs detected in IP sample as m⁶A-circRNAs [12]. Due to the nonspecific binding of m⁶A antibody, we hypothesized that some circRNAs without m⁶A modification will be randomly pulled down in IP samples. Therefore, taking all the circRNAs detected in IP samples as m⁶A-circRNAs might cause false discoveries. To avoid false discoveries, Circm6A utilizes Fisher’s exact test to examine whether the circRNAs are enriched in IP sample compared to INPUT sample (Additional file 1: Fig. S1B). We compared Circm6A with eight known circRNA detection tools using the simulated MeRIP-seq datasets. As expected, Circm6A has highest F1 score (0.99) compared to the approach of detecting circRNAs in IP sample using the eight known circRNA detection tools (average F1 score 0.76, F1 score 0.62 ~ 0.94) (Fig. 2C and Additional file 1: Table S5). Even if low confidence m⁶A-circRNAs from Circm6A were included, Circm6A still achieve high F1 score (0.89) on detection of m⁶A-circRNAs (Fig. 2C and Additional file 1: Table S5).

For the strategy of RNA fragmentation before MeRIP, we also simulated three MeRIP-seq datasets with sequencing depth range from 60 to 80 million reads. This strategy is a frequently used approach in the m⁶A profiling studies. Compared to the strategy of RNA fragmentation after MeRIP, fragmentation of RNA before MeRIP will obtain higher resolution m⁶A sites. However, this strategy is not able to identify m⁶A-circRNAs with m⁶As distal to BSJ sites, since the sequenced fragments of these BSJ-distal m⁶A-circRNAs pulled down by m⁶A antibody do not contain BSJ site, which is the critical signal for detecting m⁶A-circRNAs needed by both Circm6A and other circRNA detection tools (Additional file 1: Fig. S1A). By applying Circm6A and other known tools to our simulated MeRIP-seq dataset, we indeed found that Circm6A and other known tools performed poorly in identifying m⁶As distal to BSJ sites (called “BSJ-distal m⁶A-circRNAs”) (F1 score < 0.01). However, Circm6A had much better performance in detecting m⁶As near BSJ (called “BSJ-proximal m⁶A-circRNAs”) than other tools (Circm6A: F1 score > 0.9; other tools: F1 score 0.28 ~ 0.74) (Fig. 2D).

To evaluate the depth of sequencing required for the identification of m⁶A-circRNAs, we simulated the MeRIP-seq data with sequencing depth from 1 M to 100 M and applied Circm6A to these simulated MeRIP-seq datasets. We found at sequencing depth of 40–50 M, Circm6A achieved satisfactory performance in the identification of both circRNA and m⁶A-circRNA (Fig. 2E). In addition, the time consumption of Circm6A was low (Additional file 1: Fig. S1C), and the Memory usage of Circm6A was approximately 32 GB (which can be set freely) (Additional file 1: Fig. S1D). Circm6A was implemented in JAVA and user-friendly.

Taken together, these results showed that Circm6A could accurately detect both BSJ-proximal and BSJ-distal m⁶A-circRNAs from MeRIP-seq data using the strategy of fragmentation of RNA after MeRIP, while could only accurately detect BSJ-proximal m⁶A-circRNAs from MeRIP-seq data using the strategy of fragmentation of RNA before MeRIP.

The identification, validation, and characterization of m⁶A-circRNAs in PDAC tissue samples

We performed MeRIP-seq using fragmentation before MeRIP on the total RNAs from 77 samples from 58 PDAC patients, including 53 tumor and 24 normal tissue samples (Additional file 1: Table S1), and then applied Circm6A to these MeRIP-seq data (Fig. 3A). We detected a total of 38,164 circRNAs in all these samples (Additional file 2: Table S6), and among these circRNAs, 13,935 (36.5%), 12,634 (33.1%), and 5564 (14.6%) were recorded in the MiOncoCirc (PAAD cohort) [22], circBase [53], and CircRic [54] databases, respectively (Additional file 1: Fig. S2A). Moreover, the distribution of host gene features was similar to that of the circRNAs in the two public databases (Additional file 1: Fig. S2B, C), and the majority of circRNAs were flanked by the canonical splicing motif AG-GT, as were those from the two public databases (Additional file 1: Fig. S2D). We selected 9 circRNAs for further experimental validation, and 100% of circRNAs were validated by qRT-PCR in the PDAC samples (Additional file 1: Fig. S3A and Additional File 4). The back spliced exon-exon junctions of six circRNA candidates were also validated by Sanger sequencing (Additional file 1: Fig. S3B). These results suggested the reliability of Circm6A in the identification of circRNAs.

Fig. 3.

Identification and characterization of m⁶A-circRNAs in PDAC. A Workflow for the identification and characterization of m⁶A-circRNAs from MeRIP-seq of 77 tissue samples from 53 PDAC patients using Circm6A. B Validate the m⁶A modification of 9 m⁶A-circRNAs using MeRIP-qPCR after RNase R treatment. C Differential analysis results for 14 features for predicting m⁶A-circRNAs. Bars show FDR-corrected P values (q value). P values for quantitative features were calculated with the Wilcoxon rank sum test, and P values for qualitative variables were calculated with Pearson’s chi-squared test. D The power of prediction of m⁶A-circRNAs for the 14 features. E Receiver operating characteristic (ROC) curve for the random forest model in ten-fold cross-validation. F Mean decrease accuracy (MDA) for each feature in the random forest model. G The predicted result for the m⁶A status of “ambiguous circRNAs” by the random forest model. m⁶A, “ambiguous circRNAs” predicted as m⁶A-circRNAs; non-m⁶A, “ambiguous circRNAs” predicted as non-m⁶A circRNAs

Among all these circRNAs identified from Circm6A in our PDAC MeRIP-seq data, Circm6A detected 6369 m⁶A-circRNAs. We selected 9 m⁶A-circRNAs and 1 non-m⁶A-circRNA for experimental validation using MeRIP-qPCR of the PDAC samples after RNase R treatment. All 9 m⁶A-circRNAs were found m⁶A-modified, and the non-m⁶A-circRNA did not harbor m⁶A modification (Fig. 3B and Additional File 4), further suggesting high accuracy of Circm6A. However, the 6369 m⁶A-cirRNAs detected from Circm6A were mainly BSJ-proximal m⁶A-circRNAs, due to the limitation of Circm6A in detecting m⁶A-circRNAs using the classic MeRIP-seq technology. During the library building of classic MeRIP-seq, the m⁶A modifications and their BSJ sites are likely to be separated during RNA fragmentation (Additional file 1: Fig. S1A), which will result in the miss of the BSJ-distal m⁶A-circRNAs. To retrieve those missed BSJ-distal m⁶A-circRNAs, we next built a machine learning model to predict them. “Potential BSJ-distal m⁶A-circRNA” are defined as non-m⁶A-circRNA whose genomic region exists a m⁶A modification in the region distal to its BSJ sites. Accordingly, we found 3356 “potential BSJ-distal m⁶A-circRNAs” that have linear m⁶A peaks located in their BSJ-distal region in our PDAC MeRIP-seq data. The machine learning model was then constructed using the well-defined m⁶A-circRNAs and non-m⁶A-circRNAs as the training set. We found that 14 features, such as exon count, exon length, circRNA level, circRNA length, m⁶A motif, and GC content, were significantly different between the m⁶A-circRNAs and non-m⁶A-circRNAs (Fig. 3C) and these 14 features showed ability of distinguishing m⁶A-circRNAs from non-m⁶A-circRNAs (Fig. 3D). Based on these 14 significant features, we constructed a model using random forest algorithm to predict the m⁶A modification of circRNA. Ten-fold cross-validation of the prediction model achieved a high AUC (0.87) (Fig. 3E). Among the 14 features, circRNA level and GC content were the two most important factors contributing to accuracy (Fig. 3F). We then applied our random forest model to predict the m⁶A modification of “potential BSJ-distal m⁶A-circRNAs”. Among 3356 “potential BSJ-distal m⁶A-circRNAs,” the majority (2438, 72.6%) were predicted as m⁶A-circRNAs (Fig. 3G). We considered the 2438 m⁶A-circRNAs predicted by random forest model as BSJ-distal m⁶A-circRNAs. We next integrated the 8807 BSJ-distal and BSJ-proximal m⁶A-circRNAs for the further analysis.

Dysregulation of m⁶As in circRNAs and their functional significance in PDAC

We found that among the 8807 m⁶A-circRNAs, 1142 were hypermethylated while only 181 m⁶A-circRNAs were hypomethylated in tumor tissues compared with normal tissues (Fig. 4A and Additional file 2: Table S7). Eight out of 9 selected hyper/hypomethylated m⁶A-circRNAs could be validated by MeRIP qRT-PCR, further suggesting the reliability of the results (Fig. 4B). We found some hyper/hypomethylated m⁶A-circRNAs have clinical significance. After multiple hypothesis correction, 8 dysregulated m⁶A-circRNAs showed a negative correlation (FDR ≤ 0.25) with progression-free survival (PFS) (Fig. 4C, D) and 22 dysregulated m⁶A-circRNAs showed a negative correlation (FDR ≤ 0.25) with overall survival (OS) (Additional file 1: Fig. S4A, B).

Fig. 4.

Differential m⁶A methylation of m⁶A-circRNAs between PDAC tumor and normal tissues. A The circos plot shows the genome-wide hypermethylated (red bar) and hypomethylated (blue bar) circRNAs in PDAC tumor tissues. The height of the bar represents the frequency of m⁶A modification in all samples. B Validation of differentially methylated m⁶A-circRNAs using MeRIP-qPCR. *, P < 0.05; **, P < 0.01; ***, P < 0.001; n.s. P > 0.05 of Student’s t test. C, D The correlation between PFS and m⁶A level of differentially methylated m⁶A-circRNAs in PDAC. Volcano plot depicting log₂ hazard ratios (HRs) and − log₁₀ (FDR) values of differentially methylated m⁶A-circRNAs in the Cox proportional hazards model of PFS. The horizontal dashed line in C and vertical dashed line in D correspond to “FDR = 0.25”. All 8 differentially methylated m⁶A-circRNAs significantly associated with PFS are showed in D. E The number of coexpressed pairs for hypermethylated m⁶A-circRNAs and non-differentially methylated m⁶A-circRNAs (non-DM m⁶A-circRNAs) in PDAC tumor and normal tissue samples, respectively (upper panel). The number in parentheses is normalized to the number of coexpressed pairs shared between the tumor and normal samples. The permutation test shows significant contribution of hypermethylated m⁶A-circRNAs to the gain of coexpression network in tumor tissue samples compared with randomly selected non-DM m⁶A-circRNAs (lower panel). Vertical red lines indicate the gain fold of hypermethylated m⁶A-circRNAs, the histogram indicates the fold gain of randomly selected non-DM m⁶A-circRNAs. F Scatter plots showing the coexpression of circEEFSEC-MACC1 pair (upper panel) and circUBXN7-KDR pair (lower panel) in PDAC tumor and normal tissue samples, respectively. r represents Pearson correlation coefficients; P represents P value. G, H Top enriched KEGG pathways for upregulated mRNAs (red, G) and downregulated mRNAs (blue, H) in the coexpression network that hypermethylated m⁶A-circRNAs were involved in

It has been reported that circRNAs could regulate mRNAs by competing RNA-binding proteins or acting as scaffolds for RNA-binding proteins [55, 56]. We therefore suspected that m⁶A-circRNAs will regulate the expression of mRNAs by recruiting the m⁶A writers/readers/erasers through their m⁶A sites. To explore this, we firstly constructed a circRNA/mRNA coexpression network (detail in “Methods” section) by using all circRNAs and differentially expressed mRNAs between PDAC tumor and normal tissues (1206 upregulated and 1117 downregulated mRNAs were detected in PDAC tumor tissues, as shown in Additional file 1, Fig. S4C). Intriguingly, m⁶A-circRNA/mRNA pairs showed a significantly higher correlation than non-m⁶A-circRNA/mRNA pairs after eliminating the interference of microRNA (Wilcoxon rank sum test, P = 3.14e−13), suggesting that m⁶A can affect the coexpression between the circRNAs and the mRNAs in PDAC. Moreover, hypermethylated m⁶A-circRNAs were widely associated with a gain of coexpression network in PDAC tumor tissues, while nondifferentially methylated m⁶A-circRNAs were not (16.2- versus 5.2-fold gain, Pearson’s chi-squared test P < 2.2e−16, permutation test P < 1e−4) (Fig. 4E). In total, 671 hypermethylated m⁶A-circRNAs resulted in a gain of mRNA-circRNA coexpression in PDAC tumor tissues (Additional file 3: Table S8). For example, the hypermethylated m⁶A in circEEFSEC and circUBXN7 resulted in the gain of coexpression of circEEFSEC-MACC1 (r = 0.74, P = 4.2e−5) and circUBXN7-KDR (r = 0.71, P = 0.0095) in tumor samples and while there was no correlation between circEEFSEC and MACC1 (r = 0.3, P = 0.34) and between circUBXN7 and KDR (r = 0.25, P = 0.53) in normal samples. Moreover, the gains of coexpression in PDAC tumor were also observed in circITGB6-MECOM and circPTK2-ETV1 (Fig. 4F and Additional file 1: Fig. S4D). The 1029 upregulated mRNAs that were involved in the aberrantly methylated m⁶A-circRNA/mRNA coexpression network were mainly enriched in cancer-related pathways, such as the ECM-receptor interaction, mucin type O-glycan biosynthesis, and p53 signaling pathway, while the 925 downregulated mRNAs that were involved in this network were mainly enriched in pathways related to PPAR signaling pathway and Th1 and Th2 cell differentiation (Fig. 4G, H). These results suggested that the dysregulation of m⁶As in circRNAs might have critical functional significance in PDAC.

We speculated that m⁶A-dependent crosstalk between circRNAs and mRNAs in the coexpression network was probably mediated by m⁶A readers, as m⁶A functions through recruiting readers. Indeed, we found that all reported m⁶A readers significantly preferred to bind the comethylated pairs (both the circRNA and mRNA were m⁶A methylated in the coexpression network) (Additional file 1: Fig. S4E), and YTHDF1 and YTHDF2 were correlated with a significant proportion of differentially expressed mRNAs involved in the coexpression network (Additional file 1: Fig. S4F), indicating that these two “readers” may play a role in m⁶A-dependent crosstalk.

m⁶A modification is associated with the circularization of circRNAs

We next examined whether there is any effect of m⁶A modification on circRNA abundance. Interestingly, we found 85.5% (767/897) of circRNAs show the same direction of change between tumor and normal tissue samples at m⁶A and circRNA expression level, meaning hypermethylated m⁶A-circRNAs tended to upregulate circRNA expression in PDAC (hypermethylated/upregulated circRNAs, 767), and hypomethylated m⁶A-circRNAs tended to downregulate circRNA expression (hypomethylated/downregulated, 119) (Fig. 5A). We experimentally validated the differential expression of 9 selected candidates (Additional file 1: Fig. S5A). Further integrated analysis revealed positive correlations between the m⁶A level and the circRNA level (Fig. 5B). As previously reported, m⁶A can regulate the stability of transcripts by recruiting “readers,” such as YTHDF and IGF2BP proteins [3, 57, 58]. We first examined the correlation between the expression level of m⁶A readers and circRNAs. However, we found that there were few hypermethylated-upregulated circRNAs significantly correlated with these reported “readers,” indicating that m⁶A modification in circRNAs might not enhance expression level of circRNAs via recruiting “readers” in pancreatic cancer (Additional file 1: Fig. S5B). Since the circRNA level may also be regulated by circularization, we further investigated the relationship between m⁶As and the circular ratio of circRNAs. Interestingly, we found that BSJ-proximal m⁶A-circRNAs displayed a significantly higher circular ratio than BSJ-distal m⁶A-circRNAs and non-m⁶A-circRNAs (Fig. 5C), suggesting m⁶As locating near BSJ sites play critical role in the formation of circRNAs. The circularization process is regulated by multiple factors [6], such as ADAR (double-stranded RNA-specific adenosine deaminase) [59]. Closer inspection revealed that flanking regions of m⁶A peak centers were less occupied by ADAR1 compared to background (Additional file 1: Fig. S5C), further suggesting that m⁶As are related to the formation of circRNAs. To further validate our result, 4sU pulse labeling and nascent RNA collection were then performed after METTL3 knockdown (Additional file 1: Fig. S5D) to measure circularization index of circRNAs (CI, the relative abundance of circRNA [C] versus spliced linear mRNA [L]) of circRNA-generating genes. In agreement with the sequencing results, the decreased m⁶A level of examined circRNAs significantly reduced their circularization efficiency (Fig. 5D, E). These observations suggested that m⁶A modifications contribute to the biogenesis of circRNAs.

Fig. 5.

m⁶A modification is associated with the circularization of circRNAs. A Distribution of m⁶A-circRNAs with a significant change in both the m⁶A level and circRNA expression level. B The distribution of correlation coefficients (Pearson correlation coefficient) between the m⁶A level and circRNA expression level. C Cumulative fraction of the circular ratio for m⁶A-circRNAs and non-m⁶A-circRNAs, respectively. The P value was calculated with the Wilcoxon rank sum test. D MeRIP-qPCR analysis of m⁶A levels of 7 indicated circRNAs in PDAC cell line with METTL3 knockdown and control. E Circulation index of 7 indicated circRNAs in PDAC cell line with METTL3 knockdown and control (bottom panel). Circulation index represent the relative abundance of each circRNA to its pre-mRNA. The abundance of circRNA and its pre-mRNA were determined using qRT-PCR. Top panel showed the scheme of primer design for circRNA and linear RNA to calculate circularization index. “C” and “L” primer sets are used to quantify circRNAs and linear RNAs. Data in D and E were means ± S.E.M. (n = 3). *, P < 0.05; **, P < 0.01; ***, P < 0.001 of Student’s t test

Hyper-m⁶A-circRNAs possess potential for translation

The m⁶A modification is reported to regulate the translation of circRNAs [13]. We next dedicated to explore the relationship between m⁶A modification and circRNA translation in the identified m⁶A-circRNAs in PDAC samples. To this end, we firstly predicted coding potential using Coding potential assessing tool (CPAT) [60] and found that both BSJ-distal and BSJ-proximal m⁶A-circRNAs displayed significantly higher coding probability scores than non-m⁶A-circRNAs (P < 2.2e−16; Wilcox rank sum test). Interestingly, BSJ-distal m⁶A-circRNAs had significant higher coding probability scores than BSJ-proximal m⁶A-circRNAs (Fig. 6A). Moreover, we found BSJ-distal m⁶A-circRNAs had longer open reading frames (ORFs) compared with BSJ-proximal m⁶A-circRNAs and non-m⁶A-circRNAs (Additional file 1: Fig. S6A). To further investigate the effects of m⁶A on the translation of circRNAs, we systematically analyzed the binding sites within circRNAs for translation-related m⁶A “readers” (such as YTHDF1, YTHDF3) and eukaryotic initiation factors (EIFs) that were previously reported to promote translation [13, 61]. We found that the ratio of binding sites for EIFs (EIF3A, EIF3B, EIF3D, EIF3G, EIF3H, EIF4A3, EIF4G2) and YTHDF1-3 were more enriched in BSJ-distal m⁶A-circRNAs than that in BSJ-proximal m⁶A-circRNAs and non-m⁶A-circRNAs (Fig. 6B), which was further confirmed by permutation tests (Additional file 1: Fig. S6B). These results suggested that m⁶A-circRNAs are more likely to have translational capacity than non-m⁶A-circRNAs, and among these m⁶A-circRNAs, BSJ-distal m⁶A-circRNAs showed stronger translational capacity than BSJ-proximal m⁶A-circRNAs.

Fig. 6.

The hypermethylated m6A-circRNAs possess potential for translation. A The cumulative fraction of coding probability scores predicted by CPAT for BSJ-distal m6A-circRNAs, BSJ-near m6A-circRNAs and non-m6A-circRNAs. The P value was calculated with the Wilcoxon rank sum test. B Significant higher ratio of binding sites of EIFs and YTHDFs in BSJ-distal m6A-circRNAs than BSJ-near m6A-circRNAs and non-m6A-circRNAs. For each clustered bar, from left to right were binding sites ratios of BSJ-distal m6A-circRNAs, BSJ-near m6A-circRNAs, and non-m6A-circRNAs. P values were calculated by chi-squared test; *, P < 0.05; **, P < 0.01; ***, P < 0.001. C Venn diagram showing the number of m6A-circRNAs with coding potential (circRNAs with ORFs and EIFs binding sites) for non-DE hypermethylated-m6A-circRNA and all hypermethylated-m6A-circRNA. The outer layer indicated total number of corresponding m6A-circRNAs; the inner layer indicated the number of corresponding m6A-circRNAs with coding potential. Fisher’s exact test was used to derive the P value. D GO molecular function (MF) terms enrichment analysis for host genes of hypermethylated m6A-circRNAs with coding potential. E Relative fractions of unbound (free) RNAs, monosome- and polysome-bound RNAs for circZFHX3, circTSHZ1, and circLMTK2 in PDAC cell lines with METTL3 siRNA and control. The HPRT mRNA was used as control. F RIP-qPCR analysis showed EIF3A-, EIF3B-, and EIF3H-bound RNA abundance for circZFHX3, circTSHZ1, and circLMTK2 in PDAC cells with or without METTL3 knockdown. Data in E and F were means ± S.E.M. (n = 3). *, P < 0.05; **, P < 0.01; ***, and P < 0.001 of Student’s t test comparing with each control

Interestingly, we found that the percentage of circRNAs harboring binding sites for EIFs (EIF3A, EIF3B, EIF3D, EIF3G, EIF3H, EIF4A3, EIF4G2) in nondifferentially expressed (non-DE) hyper-m⁶A-circRNAs was significantly higher compared to all hyper-m⁶A-circRNAs (Fisher’s exact test, P = 0.00015) (Fig. 6C), indicating that m⁶A modifications of some hyper-m⁶A-circRNAs might not be functional through affecting circRNA formation, but through affecting translation. To elucidate the biological modules those hyper-m⁶A-circRNAs potentially affecting translation (hyper-m⁶A-circRNAs with ORFs and EIFs/YTHDFs binding), we carried out GO enrichment analysis on their host gene. We found that hyper-m⁶A-circRNAs enrich in GO molecular function (MF) terms on transcription process such as core promoter binding and transcription core repressor activity (Fig. 6D). We selected several circRNAs involved in these GO term for further experiment validation. After knockdown of METTL3, the m⁶A level and translation activity of circZFHX, circTSHZ1 and circLMTK2 were significantly reduced (Fig. 6E and Additional file 1: Fig. S6C, D). We further observed that knockdown of METTL3 attenuated the binding of EIFs (EIF3A, EIF3B and EIF3H) to these hyper-m⁶A-circRNAs (Fig. 6F). These findings indicated that the hypermethylation of some m⁶A-circRNAs might cause the dysregulation of circRNA translation.

Discussion

In recent years, both circRNAs and m⁶A modifications have attracted increasing attention in cancer studies due to their important roles in cancer progression in various types of cancer [9, 11, 22, 62–64]. However, few studies have brought together the two rapidly expanding fields in cancer studies. In particular, transcriptome-wide mapping of m⁶A modifications in circRNAs in cancer patients is still lacking. In this study, we developed a de novo algorithm for the transcriptome-wide mapping of m⁶A modifications in circRNAs from m⁶A-seq data and applied this algorithm to the m⁶A-seq data from a large sample of individuals with PDAC with the goal of exploring the function and regulation of m⁶A-circRNA in cancer.

Our de novo algorithm has several advantages compared to the existing methods for transcriptome-wide mapping of m⁶A modifications in circRNAs [12, 13]. Zhou et al. considered that all the circRNAs detected from m⁶A IP samples have potential m⁶A modifications [12]. However, antibodies recognizing m⁶A will also capture RNAs without m⁶A modifications in a nonspecific manner [65, 66]; therefore, the detection of circRNA signals only in IP samples without considering the background noise as presented in INPUT samples will probably introduce false positives of m⁶A-circRNAs. Yang et al. employed m⁶A IP for RNA samples treated with exoribonuclease RNase R (circRNA-m⁶A-seq) to specifically identify m⁶A modifications in circRNAs [13]. Similar to the methods of Zhou et al., circRNA-m⁶A-seq also does not consider the nonspecific precipitation of RNAs by antibodies targeting m⁶A. Moreover, circRNA-m⁶A-seq does not detect the linear transcripts and their m⁶A modification; thus, crosstalk between circRNAs and mRNAs cannot be well explored. To address these issues, we developed a de novo algorithm called “Circm6A” to detect m⁶A modification in circular RNAs and linear transcripts from m⁶A-seq data. We performed Fisher’s exact test on the circRNA signal difference between IP and INPUT samples to determine the probability of observing a circRNA signal in IP samples caused by nonspecific capture.

However, Circm6A still has limitations when being applied to the MeRIP-seq data that was generated using the strategy of fragmentation before m⁶A-IP, which is the more frequently used library construction strategy since it could detect higher resolution m⁶A sites compared to the strategy of fragmentation after m⁶A-IP. For the strategy of fragmentation before m⁶A-IP, if the m⁶As are located in the circRNAs but distal to BSJs, the circRNA fragments pulled down by m⁶A-IP will not have BSJ signal thus the m⁶A enrichment signal will not be connected with the circRNAs. Therefore, the m6A-circRNAs with m6As distal to BSJs will not be detected by Circm6As and other circRNA detection tools. The analysis on our simulated data validated this assumption. To address this limitation, we developed a random forest model for the prediction of m⁶A modification status from the MeRIP-seq data. The random forest model will complement the Circm6A results.

We described a comprehensive landscape of m⁶A modification in circRNA in PDAC by using Circm6A. The m⁶A modifications in many circRNAs were dysregulated in PDAC tumor tissues compared to adjacent normal tissues, suggesting their functional roles in cancer progression. The identification of dysregulated m⁶A-circRNAs expanded our understanding of the biological role of m⁶A modification in cancer. Previous studies have reported that circRNAs function by regulating the metabolism of linear mRNA [8, 21, 67, 68]. We therefore explored the function of circRNAs and their m⁶A modifications in a circRNA-mRNA coexpression network. Intriguingly, we found that m⁶A modifications had a positive effect on circRNA-mRNA coexpression, and hypermethylated circRNAs significantly amplified the circRNA-mRNA coexpression network in PDAC tumor tissues compared to normal tissues. CircRNAs are intensively reported to function as miRNA sponges to regulate linear mRNA expression [8, 69, 70]. A few studies have reported that circRNAs could also modulate linear mRNAs by competing with RNA-binding proteins or acting as scaffolds for RNA-binding proteins [55, 56]. We propose several potential mechanism models for m⁶A-dependent crosstalk. One possible explanation is that m⁶A-circRNAs perhaps serve as sponges to sequester m⁶A readers from linear mRNAs with m⁶As, thereby reducing the effect of m⁶A readers on linear mRNAs. The other possible mechanism might be that m⁶A-circRNAs might function as scaffolds that promote the binding of readers to linear mRNAs. Indeed, Chen et al. have reported that circNSUN2, an m⁶A-circRNA, can enhance the stability of HMGA2 mRNA by forming a circNSUN2/IGF2BP2/HMGA2 RNA-protein ternary complex [14].

m⁶A modifications have critical roles in many aspects of RNA regulation, including RNA stability, splicing, and translation [2, 3, 71–74]. In this study, we found that the hypermethylation of circRNAs tended to elevate circRNA levels in PDAC tumor tissues compared to normal tissues, and m⁶A levels were significantly positively correlated with circRNA abundance, whereas we observed that few circRNA were significantly coexpressed with well-known m⁶A readers that regulates RNA stability, suggesting that m⁶As might not regulate expression of circRNAs by recruiting “reader” to stabilize circRNAs. Apart from stability, the biogenesis of circRNA is another factor influencing circRNA levels. Interestingly, we found that m⁶A-circRNAs have a higher circular ratio than non-m⁶A-circRNAs, suggesting that m⁶As might play a potential role in the biogenesis of circRNAs. ADAR1, an RNA-editing enzyme, has been reported to suppress the biogenesis of circRNAs because A-to-I editing can diminish RNA pairing and thus prevent back-splicing for circRNA biogenesis [75]. Xiang et al. reported that ADAR1 is unfavorably associated with m⁶A-modified transcripts for further A-to-I editing and that m⁶A modification suppresses A-to-I editing on the same transcripts [76]. Consistently, we observed that the flanking region of m⁶A-circRNAs has fewer binding sites for ADAR1 than that of non-m⁶A-circRNAs, indicating a potential role of m⁶A in the regulation of circRNA biogenesis in which m⁶A modification suppresses A-to-I editing of pre-mRNAs and hence results in recovery of RNA pairing, eventually promoting the biogenesis of circRNAs; however, this mechanism requires further investigation. Previous researches have reported that some circRNAs have translational abilities [77–81]. Moreover, m⁶A is reported to promote the translation efficiency of circRNAs [13].

Conclusions

In conclusion, we firstly developed a computational tool, Circm6A, to detect m⁶A modification in circRNAs from m⁶A-seq data specially, which would facilitate following researches on m⁶A-circRNAs. With application of Circm6A to our PDAC m⁶A-seq data, we showed the genomic landscape of m⁶A-circRNAs in PDAC patients for the first time. Further analysis showed that m⁶A-circRNAs tended to be hypermethylated in PDAC tumor tissues compared to adjacent normal tissues. Surprisingly, we found that hypermethylated m⁶As-circRNAs were related with the expansion of circRNA-mRNA coexpression network involved in many cancer-related pathways. Further, our bioinformatic analysis and experiment results also indicated that m⁶A modifications in circRNAs might promote the circularization and translation of circRNAs. These comprehensive findings shed new light on the regulatory perspective of m⁶A modification in circRNAs and the function importance of m⁶A-circRNAs in PDAC.

Abbreviations

m⁶A

N⁶-methyladenosine

circRNA

Circular circRNA

BSJ

Back-splicing junction

PDAC

Pancreatic ductal adenocarcinoma

IP

Immunoprecipitation

MeRIP-seq

m⁶A-specific methylated RNA immunoprecipitation followed by high-throughput sequencing

m⁶A-circRNA

Circular RNA with m⁶A modification

non-m⁶A-circRNA

Circular RNA with m⁶A modification

qRT-PCR

Quantitative real-time PCR

PFS

Progression-free survival

OS

Overall survival

ORF

Open reading frames

F1 score

The harmonic mean of the precision and recall

FDR

False discovery rate

AUC

Area under the curve

GO

Gene ontology

Authors’ contributions

J.Z., Z.Z., and D.L. conceptualized and supervised the research. Y.Y. and W.F. were engaged in biostatistics and bioinformatics analysis. J. Zhang, X.H., and Y.Z. designed and performed most experiments. R.B., L.Z., L.W., M.L., J.D., and R.C. were responsible for patient recruitment and biospecimen and clinical data collection. J.Z., Z.Z., D.L., and Y.Y. drafted the manuscript. All authors reviewed and edited the manuscript. All author(s) read and approved the final manuscript.

Funding

This study was supported by the National Natural Science Foundation of China (81772586 and 82072617 to J.Z.; 81772614 and U1611261 to Z.Z.); the Program for Guangdong Introducing Innovative and Entrepreneurial Teams (2017ZT07S096 to D.L.), the National Young Top-notch Talent Support Program (to J.Z.), and Sun Yat-sen University Intramural Funds (to D.L. and to J.Z.).

Availability of data and materials

The bioinformatics tools Circm6A (https://github.com/canceromics/circm6a [82]) and MeRIP-simulator at (https://github.com/canceromics/MeRIP-simulator [83]) are both available for free online. Data analysis scripts in the study were deposited in github (https://github.com/finallyisnoone/Circm6A.git [84]).

The raw sequencing data of this study have been deposited in the Genome Sequence Archive of Beijing Institute of Genomics, Chinese Academy of Sciences (https://ngdc.cncb.ac.cn), with accession number HRA000102 (https://ngdc.cncb.ac.cn/gsa-human/browse/HRA000102).

Declarations

Ethics approval and consent to participate

For all of the patients who participated in this study, written informed consent was obtained. This study has been approved by the Ethics committee of Sun-Yat-sen University Cancer Center (Approval number: GZR2016-257) according to the Chinese Ethical Regulations and conducted according to the guidelines of the Declaration of Helsinki.

Consent for publication

Not applicable.

Competing interests

The authors declare that they have no competing interests.