The distribution of m¹A modifications in the transcriptome is uncharacterized in the nervous system, especially in important component neurons. We performed m¹A RNA immunoprecipitation sequencing (MeRIP-seq) to clarify the m¹A transcriptomic landscape in primary neurons and OGD/R-treated neurons (Figure 1A). A total of 48,260, 13,588, and 16,397 m¹A peaks were identified in mRNAs, lncRNAs, and circRNAs, respectively (Figure 1B). For example, regarding the m¹A peaks in mRNAs, 29,066, 37,893, and 41,334 m¹A peaks were identified in the Control, OGD/R 1.5 hr, and OGD/R 3 hr samples, respectively (Figure 1—figure supplement 1A). We examined the genomic distribution of those m¹A peaks and found that in the Control group, 54.2% were located in coding sequences (CDSs) and 25.1% were located in the 5' untranslated region (5’UTRs) in mRNAs, with 88.4% located in CDSs and 8.9% in 5’UTRs in circRNAs (Figure 1C). A similar proportion was found in the OGD/R 1.5 hr and 3 hr groups (Figure 1—figure supplement 1B). The genomic distribution pattern was slightly different from previous findings (Dominissini et al., 2016). However, more m¹A peaks were in the CDS in circRNAs than in mRNAs (88.4% vs 54.2% in the Control group), and a clear decreasing trend in near-5’UTRs was observed (Figure 1D and Figure 1—figure supplement 1C). The m¹A genomic distribution in mRNAs and circRNAs differed significantly between normal and OGD/R-treated neurons (chi-squared test, p<0.05, Figure 1—figure supplement 1B). The above results indicated the presence of abundant m¹A modifications in neurons and showed that different OGD/R treatments may affect these modifications.

Some studies have shown that m¹A motifs, such as GUUCNANNC motifs (Safra et al., 2017), GUUCRA motifs (Han et al., 2017), and GA-rich consensus sequences (Li et al., 2016), exist in different cells. However, there are no accepted highly conserved regions for m¹A modification. We used unbiased motif detection to reveal the potential motifs for m¹A peaks with MEME Suite (Bailey et al., 2015) (v5.4.1). Various motifs that were also GA-rich were identified in mRNAs and circRNAs (Figure 1E and Figure 1—figure supplement 1D). Indeed, more than 90% (4510/4646 in the Control group) of motifs in mRNA were GA-rich, and approximately 10% (579/4391 in the Control group) of motifs in circRNA were GA-rich. This pattern indicated that GA-rich m¹A motifs are more prevalent in mRNA and that various mechanisms may control m¹A modification in circRNA. By comparing these motifs with the known motifs in the JASPAR database (Castro-Mondragon et al., 2022) (v 2022), we found that the GA-rich motifs are highly similar to those in MA0528.1 from ZNF263 (Figure 1E and Figure 1—figure supplement 1D), and the functions of these motifs were mainly focused on RNA polymerase II promoter regulation. Interestingly, although the motifs in circRNAs in the OGD/R 3 hr group were also GA-rich, their main function was to regulate axon guidance and the G-protein coupled receptor protein signalling pathway. These results suggest that GA-rich m¹A motifs are present in neurons, although they vary across RNAs with different functions.

The dynamic regulation and function of m¹A mainly depend on m¹A demethylase (‘erasers’), m¹A methyltransferase (‘writers’), and m¹A read-binding protein (‘readers’). We wondered whether these m¹A regulators were differentially expressed after OGD/R-treatment. The results showed that the expression of Alkbh3, Trmt10c, Trmt61a, Ythdf2, and Ythdf3 was statistically different among the three groups (Figure 1F). However, the trends of these m¹A regulators were different under different OGD/R treatments. This result implies that the m¹A modification process in neurons may be a very complex process, and the specific mechanisms involved in this process still need to be further explored.

m⁶A modifications can functionally alter the expression of mRNAs, pre-mRNAs, miRNAs, and noncoding RNAs, such as rRNAs and tRNAs (Chua et al., 2020). Although m¹A is another abundant RNA modification, the characteristics of this modification on mRNAs and noncoding RNAs in neurons remain unclear. We identified m¹A modifications on different kinds of RNAs (mRNAs, lncRNAs, and circRNAs) in normal primary neurons and OGD/R-treated neurons to determine whether there are characteristic modifications of different RNAs after OGD/R treatment. First, we determined the numbers of m¹A mRNAs identified by four different methods: the conventional, treatment, mismatch, and trough methods (see the Materials and methods section for details). The most m¹A mRNAs were identified by the conventional method, while the mismatch method identified the least (Figure 2—figure supplement 1A). We also identified the unique and common m¹A mRNAs in different OGD/R-treated neurons by four different methods (Figure 2—figure supplement 1B). The percentages of common (5.82–76.44%) and unique mRNAs (8.48–70.1% in the Control group, 11.56–74.17% in the OGD/R 1.5 hr group, and 11.59–76.03% in the OGD/R 3 hr group) identified by the different methods varied greatly. To ensure the accuracy of subsequent analyses as much as possible, we selected the m¹A mRNAs identified by at least two of the abovementioned methods for subsequent analyses (Figure 2—figure supplement 1C), resulting in 4047, 5756, and 4948 (Figure 2—source data 1) m¹A mRNAs for the Control, OGD/R 1.5 hr, and OGD/R 3 hr groups, respectively (Figure 2A).

Then, we used density distribution (Figure 2B, left) and cumulative distribution function curves (Figure 2B, right) to compare the m¹A level among the groups. After OGD/R treatment, the m¹A level in neurons was higher than that in the Control group (OGD/R 1.5 hr vs Control: p<1.1e-15, OGD/R 3 hr vs Control: p<2.2e-16, Kolmogorov‒Smirnov test). Since the modification density differed by treatment, we sought to determine whether the amount of m¹A modification on each chromosome differs before and after treatment. The numbers of m¹A mRNAs on each chromosome were increased after OGD/R treatment (Figure 2C).

Finally, differential methylation analysis was conducted between OGD/R-treated neurons and normal neurons (Figure 2—source data 2). During the different OGD/R treatments, the mRNAs with large changes in methylation levels were not totally consistent between the two groups, possibly implying that distinct m¹A modification patterns exist in the different OGD/R-treated groups (Figure 2D). To further clarify the functions of the differentially m¹A methylated mRNAs, GO and KEGG enrichment analyses were performed. The m¹A mRNAs with decreased methylation in the OGD/R 1.5 hr group regulated synapses, the release of neurotransmitters, the cAMP signalling pathway, and axon guidance (Figure 2E). In contrast, the m¹A mRNAs with decreased methylation in the OGD/R 3 hr group regulated synaptic plasticity, the membrane potential, glutamatergic synapses, and the calcium signalling pathway (Figure 2F). The m¹A mRNAs with increased methylation in the OGD/R 1.5 hr group influenced morphogenesis, cell fate commitment, chemical carcinogenesis, and apoptosis, while the m¹A mRNAs with increased methylation in the OGD/R 3 hr group affected transmembrane receptors, cell substrate adhesion, focal adhesion, and regulation of the actin cytoskeleton (Figure 2—figure supplement 1D–E). The above results indicate that under different OGD/R treatments, the increases and decreases in m¹A in neurons affect different biological functions. With prolonged OGD/R time, differential m¹A modification regulates the fate of neurons.

m⁶A modification has been reported to be present on lncRNAs and to affect their biological functions (Patil et al., 2016; Yang et al., 2018; Wen et al., 2020). However, the m¹A modification of lncRNAs and its possible role in neurons remain uncharacterized. We identified lncRNAs with m¹A modification in normal neurons and OGD/R-treated neurons to explore the potential functions of m¹A lncRNAs. First, we applied the four abovementioned methods to identify m¹A lncRNAs in normal and OGD/R-treated neurons. The conventional method identified the most m¹A lncRNAs (Figure 3—figure supplement 1A). We then identified the unique and common m¹A lncRNAs in different OGD/R-treated neurons by the four different methods (Figure 3—figure supplement 1B). Fewer m¹A lncRNAs than m¹A mRNAs were identified by the four methods (Figure 3—figure supplement 1C). The m¹A lncRNAs identified by at least two methods were selected for further analysis (Control: 1078, OGD/R 1.5 hr: 1661, OGD/R 3 hr: 1294) (Figure 3A, Figure 3—figure supplement 1D and Figure 3—source data 1).

We next analysed the genomic sources of m¹A lncRNAs in the Control and OGD/R-treated groups. The two main genomic sources of m¹A lncRNAs were exon sense-overlapping regions and intergenic regions (Figure 3B). This distribution did not differ significantly between the Control and OGD/R groups (chi-squared test, p=0.303). Similar to the mRNA modification characteristics (Figure 2B), the differences in the lncRNA m¹A modification level between the Control and OGD/R-treated groups were statistically significant (Figure 3C, OGD/R 1.5 hr vs Control: p<0.001, OGD/R 3 hr vs Control: P<2.508e-5; Kolmogorov‒Smirnov test). We also counted the m¹A lncRNAs on each chromosome in normal and OGD/R-treated neurons and found that the number of m¹A lncRNAs on each chromosome was increased after OGD/R treatment (Figure 3D).

Interaction with RBPs is an important regulatory mechanism of lncRNAs (Yang et al., 2020a; Søndergaard et al., 2022). We thus sought to determine whether lncRNAs with m¹A modifications play vital roles in biological processes, such as RNA processing. Considering the differentially m¹A-methylated lncRNAs in the OGD/R groups (Figure 3—figure supplement 1E) (Figure 3—source data 2), RBPs binding to these lncRNAs with high reliability were predicted in the ENCORI database, and lncRNA–RBP interaction networks were then constructed (Figure 3E). In the OGD/R 1.5 hr and OGD/R 3 hr lncRNA–RBP networks, Meg3 and Neat1, which have been reported to play an important role in nervous system-related diseases, were predicted to bind to various RBPs (Zhong et al., 2017; Sanli et al., 2018; Cui et al., 2019; Liang et al., 2020). Functional enrichment analysis showed that these RBPs participated in the processes of mRNA metabolism, mRNA processing, and mRNA splicing (Figure 3F). These results suggested that the ability of nerve-related lncRNAs to regulate RNA metabolism and then affect the pathophysiological processes in neurological diseases might depend on their m¹A modification.

As RNA molecules regulate diverse pathophysiological processes, circRNAs are specifically enriched in the nervous system (Gokool et al., 2020; Mehta et al., 2020). Some studies have demonstrated that m⁶A modification can affect circRNA biogenesis, immunogenicity, translation, etc. (Chen et al., 2019b; Tang et al., 2020; Xu et al., 2020). However, the features of m¹A modification of circRNAs in the nervous system remain unknown. We investigated m¹A modifications in neuronal circRNAs using various approaches. The four abovementioned methods identified different numbers of m¹A circRNAs in each group, with the most m¹A circRNAs identified by the conventional method (Figure 4—figure supplement 1A). We then examined the common and unique m¹A circRNAs in the normal and OGD/R-treated groups identified by the four methods. The fewest associated m¹A circRNAs were identified by the mismatch method (Figure 4—figure supplement 1A–B). Fewer m¹A circRNAs than m1A mRNAs were identified by all four methods (Figure 4—figure supplement 1C). However, m¹A circRNAs were more abundant than m¹A lncRNAs. To further investigate the role of m¹A circRNAs in each group, we selected the m¹A circRNAs identified by at least two methods for downstream analysis (Figure 4A, Figure 4—figure supplement 1D and Figure 4—source data 1).

The genomic sources of those m¹A circRNAs were counted in each group, and exonic regions and sense-overlapping regions were found to be the main sources (chi-squared test, p=0.9379) (Figure 4B). Next, we generated density distribution and cumulative distribution function curves to explore the m¹A modification degree in the three groups. The m¹A modification level was increased after OGD/R treatment, and the difference between the OGD/R 3 hr and Control groups was statistically significant (OGD/R 3 hr vs Control p<0.0009, Kolmogorov‒Smirnov test). However, the m¹A level did not differ significantly between the OGD/R 1.5 hr and Control groups (Figure 4C). m¹A circRNAs on each chromosome were also analysed, and the numbers of m¹A circRNAs were increased after the different OGD/R treatments (Figure 4D).

To clarify the biological functions of the differentially m¹A-methylated circRNAs in the OGD/R 1.5 hr and OGD/R 3 hr groups (Figure 4—source data 2), we obtained the source genes of these differentially methylated circRNAs and performed GO and KEGG functional enrichment analyses. In the OGD/R 1.5 hr group, the differentially methylated circRNAs with decreased m¹A levels were enriched in synapse organization, regulation of membrane potential, the cAMP signalling pathway, and the neurodegeneration pathway, while the circRNAs with increased m¹A levels were enriched in morphogenesis, GTPase activity, axon guidance, and the MAPK signalling pathway (Figure 4E and Figure 4—figure supplement 1E). In the OGD/R 3 hr group, the differentially methylated circRNAs with decreased m¹A levels were enriched in synaptic plasticity, neurotransmitter transport, glutamatergic synapse, and dopaminergic synapse, while the circRNAs with increased m¹A levels were enriched in supramolecular fibre organization, cellular component disassembly, adherens junction, and endocytosis (Figure 4F and Figure 4—figure supplement 1F). These results indicate that the differentially modified circRNAs regulate different biological functions in different OGD/R processes. The m¹A circRNAs with decreased m¹A levels were related mainly to synapses and the release of neurotransmitters, while the m¹A circRNAs with increased m¹A levels played regulatory roles in many biological processes.

Currently, research on the translation ability and products of circRNAs is gradually increasing. We asked whether the m¹A circRNAs could be translated. By using the riboCIRC database (Li et al., 2021) (http://www.ribocirc.com/index.html), we identified some circRNAs (237/811) with translation ability among the identified differentially m¹A-modified circRNAs. In addition, m⁶A modifications annotated in the database were present on some m¹A circRNAs. The polypeptide structures were predicted to show the possible translation products. Then, mmu_circ_0000705 (encoded by App) and mmu_circ_0002207 (encoded by Foxo3) were selected to show the abovementioned features (Figure 4G–H, Figure 4—figure supplement 1G–H). App can regulate the stability of synapses that bind to diverse proteins and regulates the occurrence and development of nervous system diseases (Lee et al., 2020; Eysert et al., 2021). mmu_circ_0000705 is composed of 5 exons. In this circRNA, m⁶A modification occurs in the 85025441–85025456 region, while m¹A modification occurs in the 85043575–85043600 region. Foxo3 also regulates multiple functions of the nervous system (Deng et al., 2018; Du et al., 2021). mmu_circ_0002207 is composed of 1 exon, with m⁶A modification occurring in the 42196615–42196629 region and m¹A modification occurring in the 42196961–42197760 region. These results suggest that m¹A modification may play a regulatory role in the translation of circRNAs into peptides. However, the specific mechanism needs further exploration.

The ceRNA mechanism is a main posttranscriptional regulatory mechanism of circRNAs and lncRNAs. Studies have shown that the ceRNA mechanism plays an important role in nervous system diseases (Huang et al., 2020; Moreno-García et al., 2020). However, the function and physiological effect of the m¹A modifications in these noncoding RNAs remain unclear. We identified the differentially expressed miRNAs in the OGD/R-treated groups; 69 differentially modified miRNAs were found in the OGD/R 1.5 hr group, and 81 differentially modified miRNAs were found in the OGD/R 3 hr group (Figure 5A and Figure 5—figure supplement 1A). Differentially expressed mRNAs in the OGD/R-treated groups were also identified by RNA-seq. A total of 1579 and 3259 differentially expressed mRNAs were identified in the OGD/R 1.5 hr and OGD/R 3 hr groups, respectively (Figure 5B and Figure 5—figure supplement 1B). GO and KEGG enrichment analyses were conducted to explore the biological functions of the differentially expressed mRNAs (Figure 5C). In the OGD/R 1.5 hr group, the differentially expressed mRNAs were enriched in extracellular matrix organization, the PI3K−Akt signalling pathway, neuroactive ligand−receptor interaction, and the calcium and Hippo signalling pathways. In the OGD/R 3 hr group (Figure 5—figure supplement 1C), the differentially expressed mRNAs were enriched in synapse organization, axonogenesis, regulation of membrane potential, regulation of neurogenesis, etc. The functions of the differentially expressed genes were mainly related to the influence on cell structure in the OGD/R 1.5 hr group but to synapse, axon, and other neuron functions in the OGD/R 3 hr group.

Based on the differentially expressed miRNAs (Figure 5—source data 1), we performed a screen to identify the differentially expressed and differentially methylated mRNAs, lncRNAs, and circRNAs that bind to these miRNAs. Then, the identified miRNAs, mRNAs, lncRNAs, and circRNAs were used to construct expression- and methylation-specific ceRNA regulatory networks. In the OGD/R 1.5 hr expression ceRNA network, H19, a well-known lncRNA, was shown to regulate circRNAs by sponging miRNAs (Figure 5D, left and Figure 5—figure supplement 1D, top). In the OGD/R 3 hr expression ceRNA network, Neat1 and Rian were shown to regulate various miRNAs and affect downstream mRNA and circRNA expression (Figure 5D, right and Figure 5—figure supplement 1D, bottom). In the OGD/R 1.5 hr methylation ceRNA network, Neat1, which plays critical regulatory roles in diverse neurological diseases, was shown to sponge many miRNAs and affect other lncRNA/circRNA–miRNA axes (Figure 5E, left and Figure 5—figure supplement 1E). In the OGD/R 3 hr methylation ceRNA network, Neat1 was also shown to play important roles (Figure 5E, right). Interestingly, the miRNAs sponged by Neat1 in the methylation ceRNA network were different from those in the expression ceRNA network. Neat1, present in both the expression and methylation ceRNA networks, sponged different miRNAs in the two networks, suggesting that m¹A modification of lncRNAs impacts the ceRNA mechanism.

Our above analysis indicated that m¹A modification may affect the ceRNA mechanism of different RNAs. miRNA binding sites in mRNAs are located mainly in the 3'UTR. Therefore, we next explored the changes in miRNA‒mRNA pairs under different OGD/R treatments. By analysing the m¹A-modified peak regions and the positions of miRNA binding seed sequences, we found that there were some miRNA‒mRNA pairs that changed dynamically during different OGD/R treatments (Figure 5F). Among these miRNA‒mRNA pairs, some mRNAs have been shown to be associated with neural tissue development (Rbfox3) (Kim et al., 2013), neural progenitor differentiation (Arid1a) (Liu et al., 2021a), NMDA receptor complexes (Grin2d) (XiangWei et al., 2019), and ubiquitination of proteins (Wdtc1) (Groh et al., 2016). Therefore, for these key miRNA‒mRNA pairs, we designed a dual luciferase assay to detect whether m¹A modification of the 3’UTR affects the binding of miRNAs to mRNAs (Figure 5—source data 2). Arid1a encodes a member of the SWI/SNF family that belongs to the neural progenitor-specific chromatin remodelling complex (npBAF complex) and the neuron-specific chromatin remodelling complex (nBAF complex). We cotransfected Arid1a, Mir101b, and Trmt6 (an m¹A methyltransferase) into HEK293T cells. Trmt6 alleviated the inhibitory effect of Mir101b on Arid1a expression (one-way ANOVA, p<0.05). Rbfox3–Mir129 is another miRNA‒mRNA pair affected by Trmt6. After cotransfection of Rbfox3–Mir129 and Trmt6, Trmt6 alleviated the repression of target genes by the miRNA (albeit not significantly, one-way ANOVA) (Figure 5G). Regarding the other two genes (Grin2d and Wdtc1), after cotransfection with their corresponding miRNAs and Trmt6, Trmt6 did not affect the ability of the miRNAs to bind to their target genes (Figure 5—figure supplement 1F). This result suggests that m¹A modification of the 3’UTR in some mRNAs does affect the binding of miRNAs. The underlying mechanisms appear complex and need further exploration.

We identified features of m¹A modifications in mRNAs and noncoding RNAs and explored the potential functions of those RNAs. However, the m¹A patterns remain unknown. We applied the NMF method to identify m¹A modification patterns; k=3 was the value at which the largest change occurred and was thus selected as the optimal k value (Figure 6—figure supplement 1A). Three m¹A modification patterns, termed Cluster 1, Cluster 2, and Cluster 3, were discovered (Figure 6A). The mixture coefficient matrix also showed that the three clusters had good discrimination (Figure 6B). We further explored whether the different m¹A clusters had different biological functions. Cluster 1 was defined as the ‘Metabolism-associated cluster’ (MAC) since its members were enriched in signal transduction and biosynthetic processes (Figure 6C and Figure 6—figure supplement 1B) and metabolism-associated pathways, such as the chemokine and relaxin signalling pathways. In Cluster 2, autophagy-related biological processes and pathways (Figure 6D and Figure 6—figure supplement 1C) were enriched; thus, we defined Cluster 2 as the ‘Autophagy-associated cluster’ (AAC). Finally, Cluster 3 was defined as the ‘Catabolism-associated cluster’ (CAC) because of the enrichment of catabolic processes (Figure 6E and Figure 6—figure supplement 1D). The above results show that m¹A modification plays different regulatory roles in the different clusters, indicating the existence of specific m¹A modification patterns in neurons. To further explore the upstream regulatory mechanisms of these three m¹A patterns, we performed a transcription factor (TF) analysis. The ‘Enrichr’ database (http://amp.pharm.mssm.edu/Enrichr/) was used to enrich the TFs (Figure 6—figure supplement 1E–G). We found that the representative TFs enriched specifically in clusters. Further prediction of the activity of these transcription factorTFs showed that different TF activation and repression patterns exist in different clusters. In MAC and CAC, most of the TFs had increased inferred activity. However, most of the TFs had decreased inferred activity in AAC (Figure 6F–H). Based on these TFs with altered activity, we constructed a TF-target gene regulatory network (Figure 6I–K). The above results suggested that the three m¹A patterns we identified are indeed regulated by different TFs and that the activities of these TFs are also specific. Another question we are interested in is whether these three different functional m¹A patterns are associated with the expression of m¹A regulators. We calculated the scores of representative functions in the three m¹A patterns and then calculated the correlation between these functional scores and the expression of m¹A regulators. Our results suggested that most of m¹A regulators were positively correlated with metabolism-associated functions and autophagy-associated functions, and negatively correlated with catabolism-associated functions (Figure 6—figure supplement 1H).

C57BL/6 mice purchased from the Laboratory Animal Centre, Academy of Military Medical Science (Beijing, China), were used mainly for harvesting primary cerebral neurons. All experiments were performed in adherence to the National Institutes of Health Guidelines for the Care and Use of Laboratory Animals and were approved by the Medical Ethics Committee of Qilu Hospital of Shandong University (IACUC Issue NO. DWLL-20210061).

Primary cultures of cerebral neurons were obtained as described previously (Hilgenberg and Smith, 2007). In brief, mouse foetuses (embryonic day 17 [E17]) were removed from the uterus, and the individual foetuses were freed from the embryonic sacs. The brain and cortical tissue were dissected and placed in high-glucose Dulbecco’s modified Eagle’s medium (DMEM-HG) without phenol red (Gibco, Grand Island, NY, USA; Cat. No. 31053028). Papain solution (10 U/mL; Sigma‒Aldrich, St. Louis, MO, USA; Cat. No. LS003126) was added to these cerebral tissues, which were then incubated for 15 min at 37°C in a 5% CO₂ incubator. Dissociated cortical cells were plated on poly-L-lysine (Sigma‒Aldrich; Cat. No. P4832)-coated cell culture dishes and cultured in DMEM-HG (Gibco, Grand Island, NY, USA; Cat. No. 31053028) containing 10% foetal bovine serum (Gibco, Australia; Cat. No. 10099141) and 1% penicillin/streptomycin (P/S; Invitrogen, Carlsbad, CA, USA; Cat. No. 15140148) at a density of 1.0×10⁶ cells/mL. Four hours after seeding, the medium was replaced with neurobasal medium (NM; Gibco, Carlsbad; Cat. No. 21103049) supplemented with B-27 (Gibco, Grand Island, NY, USA; Cat. No. 17504044). Cells were cultured in a humidified incubator at 37°C with 5% CO₂. The medium was changed every 3 days. Cultures were used for in vitro experiments after 7 days.

The OGD/R model was established using a previously described method with slight modifications (Tasca et al., 2015; Ryou and Mallet, 2018). Cultured primary cerebral neurons were washed twice with phosphate-buffered saline (PBS; Sigma‒Aldrich; Cat. No. D8537) supplemented with 1% P/S after 7 days of culture. Glucose-free Dulbecco’s modified Eagle’s medium (Gibco, Grand Island, NY, USA; Cat. No. 31053028) was added to the dishes. Next, the neurons were cultured with a GENbag anaerobic incubation system (bioMérieux SA, France; Cat. No. 45534) at 37°C. The cultures were kept separate under hypoxic conditions for 1.5 hr and 3 hr to achieve OGD. Thereafter, the neurons were allowed to recover by culture in a normal serum-free medium (NM) under normal incubation conditions (37°C, 5% CO₂) for 24 hr. Neurons cultured in a normal serum-free medium under normoxic conditions served as Controls.

Total RNA was extracted from primary cultured neurons using TRIzol reagent (Invitrogen, Carlsbad, CA, USA; Cat. No. 15596018) according to the manufacturer’s protocol. A NanoDrop ND-1000 system (Thermo Fisher Scientific, Waltham, MA, USA) was used to measure the RNA concentration in each sample. The OD260/OD280 ratio was assessed as an index of RNA purity, and samples with OD260/OD280 values ranging from 1.8 to 2.1 met the qualifications for purity. RNA integrity was evaluated using denaturing agarose gel electrophoresis.

MeRIP-seq was performed by Cloudseq Biotech Inc (Shanghai, China) according to a published procedure with slight modifications (Meyer et al., 2012). In brief, three biological replicates were used for the Control, OGD/R 1.5 hr, and OGD/R 3 hr groups. Isolated RNA was chemically degraded into fragments of approximately 100 nucleotides in length using a fragmentation buffer (Illumina, Inc, CA, USA). GenSeq m1A MeRIP Kit (Cloudseq Biotech Inc, Shanghai, China; Cat. No. GS-ET-002) was used to perform immunoprecipitation (IP) of m1A RNA according to the manufacturer’s recommendations. IP buffer, Protein A/G beads, and m1A antibodies were used to prepare the magnetic beads for IP. Next, the MeRIP reaction mixture was prepared according to the manufacturer’s guidelines, and fragmented RNA was included in this mixture. The magnetic beads and MeRIP reaction mixture were combined in tubes, and all tubes were incubated with rotation for 2 hr at 4°C. Subsequently, elution buffer was prepared according to the manufacturer’s instructions and was used to elute bound RNA from the beads using the anti-m¹A antibody in IP buffer. Both input samples without IP and m¹A input samples were used for library construction using a NEBNext Ultra II Directional RNA Library Prep Kit (New England Biolabs, Inc, MA, USA). Eluted RNA fragments were converted to cDNA and sequenced or treated to induce partial m¹A to m⁶A conversion before cDNA synthesis. Library sequencing was performed using an Illumina HiSeq 4000 instrument (Illumina, Inc, CA, USA) with 150 bp paired-end reads.

In brief, quality control of the paired-end reads was performed with FastQC (v0.11.9) prior to trimming of 3’ adaptors and removal of low-quality reads using Cutadapt software (v1.9.3). Then, HISAT2 software (v2.0.4) was used to align the clean reads from all libraries to the reference genome (mm10) downloaded from Ensembl. Peaks for which –10×log10(p value)>3 were detected using Model-Based Analysis of ChIP-Seq (MACS) software (v2.2.7.1). Differentially methylated sites with a fold change of ≥2 and false discovery rate (FDR) of ≤0.0001 were identified with the diffReps differential analysis package (v1.55.6). The peaks identified by MACS and diffReps that overlapped with exons of mRNAs, lncRNAs, and circRNAs were identified and selected for further analysis. In addition to conventional peak calling (conventional method), the peaks with A->T mismatches (mismatch method) were analysed. In addition, m¹A sites have a partial chance of terminating reverse transcription, resulting in lower coverage in the middle of the peak (near the m¹A site) than on both sides, forming a depression. These peaks and depressions (trough method) were also analysed. Moreover, when m¹A was converted into m⁶A, the mismatch and termination properties of m⁶A were lost, while reverse transcription was normal. Thus, after sequencing, peaks with the normal shape could be detected at the sites (treatment method).

Total RNA was used for the removal of ribosomal RNA (rRNA) using Ribo-Zero rRNA Removal Kits (Illumina, USA) following the manufacturer’s instructions. RNA libraries were constructed by using rRNA-depleted RNAs with a TruSeq Stranded Total RNA Library Prep Kit (Illumina, USA) according to the manufacturer’s instructions. Libraries were subjected to quality control and quantified using a BioAnalyzer 2100 system (Agilent Technologies, USA). RNA was purified from each sample using Oligo(dT) Dynabeads (Invitrogen) and subjected to first-strand cDNA synthesis and library preparation using a TruSeq Stranded mRNA Library Prep Kit (Illumina). Libraries (10 pM) were denatured as single-stranded DNA molecules, captured on Illumina flow cells, amplified in situ as clusters, and finally sequenced for 150 cycles on an Illumina HiSeq 4000 instrument according to the manufacturer’s instructions.

Quality control of paired-end reads was performed with FastQC (v0.11.9) prior to trimming of 3’ adaptors and removal of low-quality reads using Cutadapt software (v1.9.3). The high-quality reads were aligned to the mouse reference genome (mm10) with HISAT2 software (v2.0.4). Then, guided by the Ensembl (GRCm39.104) GTF gene annotation file, expression was estimated in units of fragments per kilobase of transcript per million mapped reads (FPKMs). The differentially expressed genes were identified as those with a fold change of ≥2 and adjusted p value of ≤0.05.

miRNA-seq was conducted at Cloudseq Biotech Inc (Shanghai, China). In brief, total RNA from each group was prepared and quantified with the BioAnalyzer 2100 system (Agilent Technologies, USA). Polyacrylamide gel electrophoresis (PAGE) was performed, and the gel was cut to select the band corresponding to a length of 18–30 nt to recover small RNAs. Adaptors were then ligated to the 5’ and 3’ ends of the small RNAs. After cDNA synthesis and amplification, the PCR-amplified fragments were purified from the PAGE gel, and the complete cDNA libraries were quantified with a BioAnalyzer 2100. Sequencing was performed on the Illumina HiSeq 4000 instrument, and 50 bp single-end reads were generated. For miRNA-seq data analysis, the adaptor sequences were trimmed, and the trimmed reads (≥15 nt) were retained by Cutadapt software (v1.9.3). Then, the trimmed reads were aligned to the merged mouse pre-miRNA databases (known pre-miRNAs from miRbase [v22.1]) using NovoAlign software (v3.02.12). The number of mature miRNA-mapped tags was defined as the raw expression level of that miRNA. Read counts were normalized by tag counts per million aligned miRNAs (TPM) values. Differentially expressed miRNAs between the two groups were filtered by the following criteria: fold change ≥2 and adjusted p-value ≤0.05.

Gene Ontology (GO) and Kyoto Encyclopedia of Genes and Genomes (KEGG) pathway enrichment analyses based on the differentially expressed genes were performed with the R package ‘clusterProfiler’ (v4.4.0) (Wu et al., 2021). GO covers three categories: cellular component (CC), molecular function (MF), and biological process (BP). The adjusted p value for a GO term denotes the significance of the enrichment of genes in that term. Pathway enrichment analysis is a functional analysis that maps genes to KEGG pathways. The Fisher p value denotes the significance of the pathway correlation to the conditions. GO analysis was performed using enrichGO function in the R package ‘clusterProfiler’ with the following parameters: pvalueCutoff = 0.005, qvalueCutoff = 0.005, minGSSize = 2, and maxGSSize = 500. KEGG analysis was performed using enrichKEGG function in ‘clusterProfiler’ with the following parameters: pvalueCutoff = 0.05 and qvalueCutoff = 0.05. The GO and KEGG pathway terms with adjusted p values of≤0.05 were considered significantly enriched.

Based on differentially methylated lncRNA profile (Figure 3—source data 1), we predicted RBPs that might interact with those lncRNA by the ENCORI database (https://starbase.sysu.edu.cn/). We predicted the lncRNA–RBP interaction network using the RBP-Target module in ENCORI with the following parameters: Clade = mammal, Genome = mouse, CLIP Data = medium stringency ≥ 2. We then integrated the lncRNA–RBP interacting pairs using R tidyverse package (v 1.3.2). Network visualization was done using Cytoscape vs 3.9.1 (http://www.cytoscape.org/). The shape and colour of the points are customizable.

circRNA structure and polypeptide structure were all predicted by the riboCIRC database (Li et al., 2021) (http://www.ribocirc.com/index.html). For the circRNA structure prediction, after submitting the query, the site provides several pieces of information, including ribosome associated evidence, m⁶A evidence, and ORF evidence. We redrew the RNA structure from the website using Adobe Illustrator CC 2017 and added additional information. For the protein structure prediction, the I-TASSER Suite was used (Yang et al., 2015). After submitting the query, predicted secondary structure, predicted solvent accessibility, top 10 threading templates used by I-TASSER and top 5 final models predicted by I-TASSER were provided by the website tool.

Based on the miRNA expression data (Figure 5—source data 1), we used three databases (TargetScan, miRDB, and miRTarBase) for comprehensive analysis of miRNA binding target genes. We obtained the intersection of the mRNAs predicted by the three databases, and differentially expressed mRNAs and differentially methylated mRNAs were integrated to construct the miRNA‒mRNA interaction network. For lncRNAs and circRNAs, we used the ENCORI database (https://starbase.sysu.edu.cn/) to analyse the binding of lncRNAs and circRNAs to the differential miRNAs. After obtaining the lncRNA‒miRNA and circRNA–miRNA interaction networks, we intersected the predicted lncRNAs and circRNAs with the differentially expressed lncRNAs and circRNAs and the differentially methylated lncRNAs and circRNAs in our data. Finally, we integrated all data to construct the expression and methylated circRNA/lncRNA‒miRNA‒mRNA interaction networks. Network visualization was done using Cytoscape vs 3.9.1 (http://www.cytoscape.org/). The shape and colour of the points are customizable.

The Rbfox3, Arid1a, Grin2d, and Wdtc1 reporter vectors were constructed by inserting the 3’UTRs of the corresponding genes downstream of the luciferase reporter gene into the GV272 plasmid (GeneChem Technologies, Shanghai, China). The Trmt6 overexpression vector was constructed by subcloning the coding sequence (CDS) into the GV712 plasmid (GeneChem Technologies, Shanghai, China). The Mir129, Mir101b, Mir27, and Mir16 mimics were designed and synthesized by GeneChem. We used Lipofectamine 3000 to cotransfect reporter plasmids, the Trmt6 plasmid, and miRNA mimics into HEK293T cells. Luciferase activity was measured 48 hr later according to the manufacturer’s procedures (E292, Promega, USA). All experiments described were replicated independently with similar results at least three times.

NMF is a widely used tool for the analysis of high-dimensional data because it automatically extracts sparse and meaningful features from a set of nonnegative data vectors. NMF clustering was used to determine the m¹A modification patterns in our nine differently treated samples (Gaujoux and Seoighe, 2010). We performed this analysis using all m¹A-modified mRNAs in the three groups. The k value at which the magnitude of the cophenetic correlation coefficient began to decrease was selected as the optimal number of clusters. The heatmap of m¹A regulators, the basic components, and the NMF connectivity matrix for different clusters were estimated with the NMF package (v0.24.0) in R (v4.1.3).

TF analysis was employed to determine the upstream regulation of three m¹A patterns. Submit the genes contained in each of the three patterns to the ‘Enrichr’ database (http://amp.pharm.mssm.edu/Enrichr/) and select the ChEA 2022 database to view the results of the enriched TFs. To predict the activity of TFs in each pattern, we used the R package ‘decoupleR’ (v 2.5.0). The data containing signed tTF - target gene interactions were harvested from the R package ‘decoupleR’ (v 1.10.0). We employed the run_wmean function to calculate the TF activity with the following parameters: mor=‘mor’, times = 1000, and minsize = 5. Network visualization was done using Cytoscape vs 3.9.1. The shape and colour of the points are customizable.

The R package ‘GSVA’ (v 1.46.0) was exploited to quantify the pathway activity in three m¹A patterns by calculating the GSVA score. Genes for each pathway were obtained from the corresponding GO database (http://geneontology.org/). The gsva function was used to calculate the pathway score. Then we used the corr.test function from R package ‘psych’ (v 2.2.9) calculated the correlation between the m¹A_regulators and pathway score.