GEO expression profiling dataset was used to analyze the mRNA expression of NSUN2 in cervical, ovarian, and endometrial cancer. GEPIA (http://gepia.cancer-pku.cn/) was used to analyze the mRNA levels of NOL1/NOP2/SUN domain (NSUN) family proteins in the major gynecologic cancers. UALCAN website (http://ualcan.path.uab.edu/) was used to analyze the mRNA and protein expression of NSUN2 in cervical, ovarian, and endometrial cancer. cBioPortal website (https://www.cbioportal.org/) was used to explore the RNA alteration of NSUN2 in cervical, ovarian, and endometrial cancer. The overall survival plots related to NSUN2 and NSUN6 expression in cervical, ovarian, and endometrial cancer were obtained by Kaplan–Meier plotter (http://kmplot.com/analysis/).

Twenty cervical carcinoma tissues and 20 normal cervical epitheliums, 20 ovarian cancer tissues, and 11 normal ovarian tissues were collected to analyze the mRNA expression of NSUN2 in cervical cancer and ovarian cancer. Five pairs of cervical tumor and adjacent tissue samples and 6 pairs of ovarian tumor and adjacent tissue samples were collected to analyze the protein expression of NSUN2 in cervical and ovarian cancer. We collected 30 paraffin-embedded cervical cancer wax and 29 paraffin-embedded normal cervical tissue, and made a tissue chip to explore the expression of NSUN2 in cervical cancer. All the samples were collected from December 2019 to December 2020 in Women’s hospital, School of Medicine, Zhejiang University, China. Patients provided informed consent to obtain samples, and the study was subjected to approval by the Hospital Ethical Committee. Tissue samples were stored in RNA store solution at 4°C overnight and stored at −80°C until use.

pHAGE-CMV-NSUN2-FLAG-HA, which expresses FLAG-HA-tagged NSUN2, was constructed by inserting NSUN2 cDNA fragment into pHAGE-CMV-FLAG-HA. shNSUN2 resistant NSUN2 wild-type (NSUN2-WT) and mutant (with point mutation at cysteine 271 and 321) plasmids (NSUN2-DM) were constructed by introducing point mutations through the use of KOD One™ PCR Master Mix -Blue- (KMM-201, TOYOBO) as previously reported (9). Plasmids pHAGE-CMV-KRT13-FLAG-HA, pHAGE-CMV-YBX1-FLAG-HA, and pHAGE-CMV-ALYREF-FLAG-HA, which express FLAG-HA-tagged KRT13, YBX1, and ALYREF, respectively, were constructed by inserting corresponding cDNA fragment after CMV promoter of pHAGE-CMV-FLAG-HA. The KRT13 promoter and GAPDH promoter were cloned from genomic DNA. Plasmid pGL3-KRT13-luciferase was constructed by inserting KRT13 promoter into pGL3-luciferase. Plasmid pGL3-GAPDH-mcherry was constructed by inserting GAPDH promoter into pGL3-mcherry. Plasmids pLKO.1-shNSUN2-1 and pLKO.1-shNSUN2-2, which were designed to suppress endogenous NSUN2, were constructed by inserting shRNA-targeting sequence after U6 promoter of pLKO.1 using KOD One™ PCR Master Mix -Blue-. The sense sequences are listed below:

shNSUN2-1: 5’-GAGCGATGCCTTAGGATATTACTCGAGTAATATCCTAAGGCATCGCTC-3’,

shNSUN2-2: 5’-CAGTGGAAGGTAATGACGAAACTCGAGTTTCGTCATTACCTTCCACTG-3’.

The generation and infection of lentivirus were conducted according to the protocol from Addgene.

The human cervical cancer cell lines SiHa and CaSki, and the human ovarian cancer cell lines A2780 and SKOV3 were purchased from the American Type Culture Collection (ATCC, USA). The SiHa cell line was cultured in DMEM (BI, Israel) containing 10% FBS and maintained at 37°C in 5% CO₂; CaSki and A2780 cell lines were cultured in RPMI-1640 (BI, Israel) containing 10% FBS. The SKOV3 cell line was cultured in MCCOY’S 5A (M9420, Solarbio) containing 10% FBS. We used PEI MAX^®–Transfection Grade Linear Polyethylenimine Hydrochloride (49553-93-7, Polysciences) to transfect plasmid and Lipofectamine™ RNAiMAX Transfection Reagent (13778150, Invitrogen) to transfect siRNA according to the manufacturer’s instructions to regulate gene expression. The siRNA sequences are listed below:

SiYBX1-1: 5’-GGACGGCAATGAAGAAGAT-3’

SiYBX1-2: 5’-CCACGCAATTACCAGCAAA-3’

SiALYREF-1: 5’-GCACGATCTTTTCGACAGT-3’

SiALYREF-2: 5’-GAGGTGGCATGACTAGAAA-3’

SiKRT13-1: 5’-CUACUACAAGACCAUUGAATT-3’

SiKRT13-2: 5’-GCCAGAACCAAGAGUACAATT-3’

The total RNA of cells was extracted with RNAiso Plus (9109, Takara) under the manufacturer’s guidance. The cDNA was synthesized by PrimeScript™ RT Reagent Kit with gDNA Eraser (RRO47A, Takara) according to the manufacturer’s instructions. Light Cycler 480 (Roche) was used to detect the expression levels of target genes using 2X Universal SYBR Green Fast qPCR Mix (RK21203, ABclonal). The examined genes were normalized to GAPDH expression. The specific primer sequences of targeted genes are displayed in Supplementary Table S1.

The total RNA of cells was extracted with RNAiso Plus (9109, Takara). RNA-seq was performed by Cloudseq Biotech Inc. (Shanghai, China) following the published procedure (25). The detailed information is listed below. Total RNA was used for removing the rRNAs with NEBNext rRNA Depletion Kit (New England Biolabs, Inc., Massachusetts, USA) following the manufacturer’s instructions. RNA libraries were constructed by using NEBNext^® Ultra™ II Directional RNA Library Prep Kit (New England Biolabs, Inc., Massachusetts, USA) according to the manufacturer’s instructions. Libraries were controlled for quality and quantified using the BioAnalyzer 2100 system (Agilent Technologies, Inc., USA). Library sequencing was performed on an illumina Hiseq instrument with 150-bp paired-end reads.

The total RNA of cells was extracted with RNAiso Plus (9109, Takara). RNA bisulfite sequencing and the following bioinformatic analysis were performed by Cloudseq Biotech Inc. (Shanghai, China). Briefly, 2 μg of total RNA for each sample was rRNA depleted using NEBNext^® rRNA Depletion Kit (New England Biolabs, Inc., Massachusetts, USA). rRNA-depleted RNA was bisulfite converted and purified using the EZ RNA Methylation Kit (Zymo Research). RNA libraries were then constructed with TruSeq Stranded Total RNA Library Prep Kit (Illumina, San Diego, CA, USA) according to the manufacturer’s instructions. The library quality was evaluated with the BioAnalyzer 2100 system (Agilent Technologies, Inc., USA). Library sequencing was performed on an Illumina Hiseq instrument with 150-bp paired-end reads.

Paired-end reads were harvested from Illumina HiSeq 4000 sequencer and were quality controlled by Q30. After 3’ adaptor-trimming and low-quality reads removing by cutadapt software (v1.9.3), the high-quality trimmed reads (clean reads) were aligned to the human reference genome (UCSC HG19) using meRanGs (one component of meRankTK) software with default parameters. The methylation status of each C within the genome was extracted by meRanCall (one component of meRankTK) software. meRanCompare (one component of meRankTK) software was used to identify differentially methylated sites (DMSs) on mRNA. The DMSs in mRNA were annotated with the Ensembl database to connect DMCs with the gene/transcript annotation. Then, GO and pathway analysis were conducted on those DMC-related genes.

The transfected CaSki cells were washed twice with ice-cold PBS and lysed in 1 ml of RIP Lysis buffer [150 mM KCl, 25 mM Tris, 5 mM EDTA, 0.5% Triton X-100, 0.5 mM DTT, Protease inhibitor (1:100), and RNase inhibitor (1:1,000)] on ice for 30 min. Cell lysates were centrifuged at 14,000 g at 4°C for 10 min to obtain clear lysates. Ten percent supernatant was collected as input, and the remaining supernatant was incubated with primary antibody and protein A/G Magnetic Beads in 900 μl of RIP Lysis buffer on a rotary homogenizer at 4°C for 4 h. Bounded RNAs were immunoprecipitated with beads and the beads were washed four times with RIP buffer (150 mM KCl, 25 mM Tris, 5 mM EDTA, and 0.5% Triton X-100). Then, beads were resuspended with 10 μl of 10% SDS, 10 μl of Proteinase K, and 130 μl of RIP buffer and digested at 55°C for 30 min. RNA in immunoprecipitation or input group was recovered with RNAiso Plus according to the manufacturer’s instruction. The relative enrichment of interest RNA was determined by RT-qPCR. Data were normalized to input.

The total RNA of cells was extracted with RNAiso Plus (9109, Takara). The extracted total RNA was randomly fragmented by RNA fragmentation reagents (AM8740, Thermo Scientific). Ten percent volume RNA was used as input, and the remaining RNA was incubated with anti-m⁵C antibody (A18870, ABclonal) and protein A/G Magnetic Beads in 900 μl of RIP Lysis buffer on a rotary homogenizer at 4°C for 4 h. Bounded RNAs were immunoprecipitated with beads and beads were washed four times with RIP buffer (150 mM KCl, 25 mM Tris, 5 mM EDTA, and 0.5% Triton X-100). Then, beads were resuspended with 10 μl of 10% SDS, 10 μl of Proteinase K, and 130 μl of RIP buffer and digested at 55°C for 30 min. RNA in immunoprecipitation or input group was recovered with RNAiso Plus according to the manufacturer’s instruction. The levels of enriched fragmented RNA and input were measured by one-step qPCR using TransScript^® II One-Step RT-PCR SuperMix (AH411-02, TransScript). IP enrichment ratio was calculated as the ratio of its amount in IP to that in the input. The specific primer sequences are displayed in Supplementary Table S1.

Transfected CaSki cells were treated with 10 μg/ml actinomycin D after transfection (48 h). Cells were collected at the indicated time points. The total RNA was extracted and analyzed by RT-qPCR.

Before the experiment, the KRT13 promoter and GAPDH promoter were cloned from genomic DNA. Plasmid pGL3-KRT13-luciferase was constructed by inserting KRT13 promoter into pGL3-luciferase. Plasmid pGL3-GAPDH-mcherry was constructed by inserting GAPDH promoter into pGL3-mcherry. Then, YBX1 knockdown and control CaSki cells were co-transfected with pGL3-KRT13-luciferase and pGL3-GAPDH-mcherry. After transfection, the cells were harvested and the total RNA of cells was extracted with RNAiso Plus (9109, Takara). The relative luciferase level was determined by RT-qPCR. Data were normalized to mcherry level.

Cells were lysed on ice in ice-cold lysis buffer (50 mM Tris-HCl, 150 mM NaCl, 5 mM EDTA, 0.1% SDS, 1% Triton X-100, and 5% Glycerol) containing protease inhibitors (B14001, Bimake). Equal amounts (20 μg) of total extract protein was separated on a 4%–20% polyacrylamide gel (M00656, GenScript) and was electrically transferred onto NC membrane. The antibodies used are listed as follows: NSUN2 (20854-1-AP, proteintech), KRT13 (A7697, ABclonal), GAPDH (60004-1-lg, proteintech), HA-tag (3724S, Cell Signaling Technology), YBX1 (A7704, ABclonal), ALYREF (A6010, ABclonal), and β-actin (A3854, Sigma-Aldrich).

The slides were put into the dewaxing solution I/II, 100% ethanol, 90% ethanol, 85% ethanol, and 75% ethanol for 5 min, respectively. Antigen retrieval was performed in Tris-EDTA (pH 9.0) using heat-induced protocol. The slides were incubated at 4°C with NSUN2 antibody (20854-1-AP, proteintech,1:400) overnight.

The total RNA of cells was extracted with RNAiso Plus (9109, Takara), followed by using the Magnetic mRNA Isolation Kit (s1550s, New England Biolabs) to purify polyadenylated mRNA twice according to the manufacturer. After 5-min denaturation at 95°C, the same amounts of serially diluted mRNA were loaded onto an Amersham Hybond N + membrane (RPN303B, GE Healthcare). After UV crosslinking twice for 2 min each time, membrane was stained with Methylene blue (A610622-0025, Sangon Biotech) according to the protocols of the manufacturer and was washed by 1× PBST. After the scanning to indicate the loading RNA, membrane was blocked with 5% non-fat milk in 1×PBST for 1 h at room temperature and then was incubated with anti‐m⁵C antibody (A19841, ABclonal) overnight (4°C). Membrane was washed by 1×PBST and was incubated by HRP-conjugated Affinipure goat anti-rabbit IgG (A0277, Beyotime) diluted 1:5,000 for 1 h at room temperature.

In vitro cell viability was detected by Cell Counting Kit-8 (40203ES80, YEASEN) and real-time cell analysis (RTCA). For CCK-8 detection, 5×10⁴ cells were seeded into a 96-well plate per well and pre-incubated overnight at 37°C, 5% CO₂. After that, 10 μl of CCK-8 solution was added to each well, and the absorbance at 450 nM was measured after 2-h incubation using SpectraMax^® Absorbance reader CMax plus. RTCA was performed by the real-time cell analyzer (xCELLigence RTCA SP, Roche) according to the operation manual. In brief, the same amounts of cells (about 5×10⁴) in 150 μl of medium containing 10% FBS were seeded in E-Plate 96. Relative cell counts were recorded and analyzed by xCELLigence RTCA SP for 48 h.

Cell migration and invasion capabilities were evaluated via transwell assays. The same amounts of cells were resuspended with serum-free Opti-MEM medium and then added to the upper chamber coated with or without Matrigel (BD Biosciences) after transfection. DMEM, RPMI1640, or MCCOY’S 5A with 20% FBS was added to the lower chambers as an attractant. After incubation for 8 h for migration or 24 h for invasion of Siha cells (24 h for migration or 48 h for invasion of CaSki cells and SKOV3 cells) at 37°C, non-migrating or non-invading cells were gently erased. The bottoms of the upper chambers were fixed and stained with 0.1% Crystal violet solution for 30 min. Then, three random fields were taken under a 100-fold microscope. Cell migration was also evaluated by wound healing assay. In brief, the same amounts of cells were seeded in 6-well plates. Three scratches were made using 200-μl pipette tips. After washing the plates three times with phosphate buffered saline (PBS), the cell was cultured in medium containing 2% FBS. The degrees of cell migration were recorded by microscopy at 0 and 24 h and analyzed by Image J.

Cell cycle was detected by using Cell cycle staining buffer (70-CCS01, MultiSciences) according to the protocols of the manufacturer. Briefly, the treated cells were harvested by trypsinization and washed with PBS, and then the cells were stained by 1 ml of DNA Staining solution and 10 μl of Permeabilization solution for 30 min. After staining, cell cycle analysis was performed by using CytoFLEX LX (Beckman Coulter Life Sciences).

All data in the manuscript represent three independent experiments giving similar results. Data are presented as mean ± standard deviation (SD). We performed the statistical analysis using GraphPad Prism 6 software. Kaplan–Meier curves were based on the log-rank test. The HR was performed using the Cox model. The significance between groups was determined using Student’s unpaired t-test. Statistical significance is reported as ***p < 0.001.

We primarily evaluated the role of NSUN2 in gynecological cancers through collecting the gene expression profile from multiple databases and websites including the GEO database, UALCAN website, oncomine, cBioPortal, and Kaplan–Meier Plotter. The results from GEO datasets showed no statistically significant difference of NSUN2 expression in uterine corpus endometrial cancer (UCEC) (Figure 1A), while NSUN2 is significantly upregulated in ovarian cancer (OV) and cervical squamous cell carcinoma (CESC) (Figures 1B, C). Similar results were found on NSUN2 expression from TCGA database assessed by UALCAN website (Figure 1D). Interestingly, cBioPortal database showed lower proportion of NSUN2 alteration in UCEC (7%) and OV (8%) as compared with that of CESC (29%) (Figure 1E). Additionally, the Kaplan–Meier plot of 5-year overall survival curves stratified by NSUN2 expression revealed a significant relationship between NSUN2 expression and 5-year overall survival in CESC but not UCEC and OV (Figures 1F–H). Taken together, these findings suggest that NSUN2 likely contributes to the carcinogenesis of ovarian cancer and cervical cancer but not endometrial cancer. Thus, we conducted further study on ovarian cancer and cervical cancer.

Consistent with upregulation of gene expression, the protein level of NSUN2 was also increased in ovarian cancer (Supplementary Figure S1A). In addition, NSUN2 protein level was positively associated to the individual cancer stages (Supplementary Figure S1B). Then, we examined the expression levels of NSUN2 in a set of clinical samples and found that NSUN2 mRNA was highly expressed in ovarian cancer tissues (n = 20) compared to normal ovary tissues (n = 11) (Supplementary Figure S2A). Meanwhile, 4 out of 6 patient samples exhibited higher expression of NSUN2 protein in tumor tissues than their paired adjacent normal tissues as examined by Western blot (Supplementary Figure S2B). These results demonstrated that both transcript and protein levels of NSUN2 were upregulated in ovarian cancer.

To further explore the role of upregulated NSUN2 in ovarian cancer, we established two NSUN2-knockdown stable clones in ovarian cancer cell lines A2780 and SKOV3 (Supplementary Figures S2C, D). NSUN2 knockdown has marginal effects on cell proliferation (Supplementary Figures S2E, F) and cell cycle distribution (Supplementary Figures S2G, H). The wound-healing assay also showed no obvious change of migration in the NSUN2 knockdown clones (Supplementary Figure S2I). Taken together, these results suggest that the overexpression of NSUN2 in ovarian cancer may be an incidental event.

As mentioned above, bioinformatics analysis showed that NSUN2 was upregulated and negatively related to overall survival in cervical cancer (Figures 1C, D, H). In line with this, we found that NSUN2 mRNA level was significantly upregulated in cervical cancer tissues (n = 20) compared to normal cervix samples (n = 20) (Figure 2A), and NSUN2 protein levels were also higher in randomly selected tumor tissues (n = 5) than their paired adjacent normal tissues (Figure 2B). We next collected 29 normal cervix and 30 cervical cancer paraffin sections to prepare tissue chip, and the immunohistochemistry analysis showed higher levels of NSUN2 in cervical cancer than in normal tissues (Figures 2C, D and Supplementary Figure S3A). These results indicate that both mRNA and protein levels of NSUN2 were upregulated in cervical cancer.

To investigate the biological function of highly expressed NSUN2 in cervical cancer, we then established NSUN2 knockdown stable clones in two common cervical cancer cell lines CaSki and SiHa (Figures 2E, G). CCK-8 and real-time cell analysis (RTCA) assays showed that knockdown of NSUN2 has little effect on cervical cancer cell proliferation (Supplementary Figures S3B–E), and downregulation of NSUN2 led to minor changes in cell cycle distribution (Supplementary Figures S3F, G). In contrast, depletion of NSUN2 significantly inhibited migration and invasion of CaSki and SiHa cells in transwell assays (Figures 2F, H). These results suggest that NSUN2 plays an important role in migration and invasion but not proliferation in cervical cancer.

Downregulation of NSUN2 significantly decreased m⁵C levels in CaSki and SiHa cells (Supplementary Figures S4A, B). We next asked whether the effect of NSUN2 on migration and invasion is m⁵C dependent. For this end, we constructed wild-type NSUN2 (NSUN2-WT) expression vector and catalytically inactive mutant NSUN2 (NSUN2-DM) expression vector harboring double point mutations in its releasing (cysteine 271) and catalytic (cysteine 321) sites. Knockdown of NSUN2 in CaSki cells led to the inhibition of cell migration and invasion, and compensation of NSUN2-WT but not NSUN2-DM could reverse the inhibition of cell migration and invasion (Figures 3A–D), suggesting that promotion of cell migration and invasion by upregulated NSUN2 is dependent on its m⁵C RNA methyltransferase activity. In other words, NSUN2 promoted cervical cancer migration and invasion in an m⁵C-dependent manner.

We then performed RNA-BisSeq and RNA-Seq in wild-type and NSUN2 depleted CaSki cells to explore the potential m⁵C-dependent pathways involved in NSUN2-mediated cell migration and invasion. The RNA-BisSeq results showed that knockdown of NSUN2 significantly reduces the number of m⁵C sites of the transcripts across chromosomes, especially across chromosomes 1, 11, 12, and 19 (Supplementary Figure S4C). After sequence context comparison, we found that NSUN2 has no preferred motif among CHH, CHG, and CG (H=A, C, U) in CaSki cells, because there was no significant alteration of sequence distribution proportion between control (shNC) and NSUN2 knockdown cells (Supplementary Figures S4D, E). On the other hand, through profiling analysis, we found downregulation of m⁵C levels on 301 genes, and further identified six genes overlapping between 1204 NSUN2-depletion-affecting genes and 301 m⁵C downregulated genes (Figure 4A). Next, we verified expression of these genes in the other cervical cancer cell line, SiHa, and found that only KRT13 levels displayed significant downregulation in NSUN2-depleted SiHa cells as well (Supplementary Figure S4F). Furthermore, besides mRNA levels, the expression of KRT13 protein was also obviously downregulated in NSUN2 knockdown CaSki and SiHa cells (Figures 4B–E), suggesting KRT13 mRNA as the downstream target gene of NSUN2-m⁵C regulatory pathway in cervical cancer cells. Supporting this notion, inhibition of migration and invasion by NSUN2 knockdown was significantly reversed by overexpression of KRT13 in both CaSki and SiHa cells (Figures 4F–I and Supplementary Figures S4G, H). However, overexpressing NSUN2 cannot rescue cell migration in KRT13 knocking down CaSki cells (Supplementary Figures S5A, B), which indicated that NSUN2 regulates cell migration in a KRT13-dependent manner. Collectively, our results demonstrated that NSUN2 promotes cervical cancer cell migration and invasion via a KRT13-dependent manner.

We next investigated whether and how NSUN2 regulates KRT13 transcripts via m⁵C modification. The RNA-BisSeq results showed that there were multiple m⁵C modification sites on 3’-UTR and CDS regions of KRT13 mRNA (Supplementary Figure 6A). Among these, three clusters (marked green in Figure 5A) were demethylation in NSUN2 knockdown CaSki cells. We designed 3 pairs of primers specifically amplifying these m⁵C clusters on KRT13 mRNA (named KRT13-1, KRT13-2, and KRT13-3) and a pair of primers amplifying the non-m⁵C region (named KTR13-ctrl) (Supplementary Figure 6A). m⁵C-MeRIP-qPCR results showed high enrichment of m⁵C modification on KRT13 mRNA in CaSki cells, especially at the KRT13-2 (CDS region near 3’-UTR) and KRT13-3 (3’-UTR region); NAPRT1, a gene validated to harbor m⁵C modification (7), was used as a positive control (Figure 5A). Knockdown of NSUN2 significantly reduced the interaction between NSUN2 and KRT13 mRNA (Figure 5B) accompanied by a notable reduction of m⁵C levels of KRT13 (Figure 5C), leading to an obvious downregulation of KRT13 mRNA levels (Figure 5D). GAPDH and Actin were used as negative control. Furthermore, overexpression of wild-type NSUN2 (NSUN2-WT) increased the interaction between NSUN2 and KRT13 mRNA (Figure 5E), accompanied by a notable upregulation of KRT13 m⁵C modification (Figure 5F), leading to a significant increase of KRT13 mRNA level (Figure 5G). Remarkably, overexpression of catalytically inactive mutant NSUN2 (NSUN2-DM) neither stimulated m⁵C levels of KRT13 nor increased KRT13 mRNA expression (Figures 5F, G). These results suggest that NSUN2 binds with and catalyzes m⁵C modification of KRT13 to upregulate KRT13 mRNA.

We next investigated the mechanism by which NSUN2 upregulated KRT13 mRNA. The mRNA stability assay showed that depletion of NSUN2 resulted in destabilization of KRT13 mRNA, suggesting that m⁵C regulates the stability of KRT13 mRNA (Figure 6A). YBX1 and ALYREF are two well-characterized m⁵C readers that could regulate m⁵C mRNA stability and transport, respectively (26). We found that siRNA knockdown of YBX1 but not ALYREF decreases KRT13 mRNA and protein levels (Figures 6B, C), suggesting that YBX1 is the reader protein involved in NSUN2 mediated stabilization of KRT13. Supporting this notion, overexpression of YBX1 but not ALYREF enhanced the levels of KRT13 mRNA and its protein (Figures 6D, E); RNA stability assay further demonstrated that suppression of YBX1 indeed reduced the stability of KRT13 mRNA (Supplementary Figures S6B, C). Inhibiting RNA synthesis by using Actinomycin D (Actd), a transcription inhibitor, cannot abolish the upregulation of KRT13 mRNA induced by YBX1 overexpression (Supplementary Figure S6D). Meanwhile, the KRT13 promoter reporter assay showed YBX1 knocking down cannot decrease the activity of KRT13 promoter (Supplementary Figure S6E). Collectively, these results demonstrated that rather than regulating KRT13 transcription, YBX1 modulated KRT13 expression by regulating its RNA stability. Moreover, YBX1-RIP assay confirmed the binding of YBX1 to KRT13 mRNA (Figure 6F), and knocking down NSUN2 can significantly decrease the binding between YBX1 and KRT13 mRNA (Supplementary Figure S6F), further demonstrating that YBX1 regulates KRT13 in a NSUN2-dependent manner. HDGF was used as a positive control (9). Furthermore, m⁵C-MeRIP-qPCR results showed that knockdown of YBX1 decreases while overexpression of YBX1 increases m⁵C levels on KRT13 mRNA (Figures 6G, H). We performed a rescue experiment and found that overexpressing NSUN2 could not rescue the expression levels of KRT13 mRNA in YBX1 depleted CaSki cells (Supplementary Figure S6G), which indicated that NSUN2 regulates KRT13 mRNA in a YBX1-dependent manner. Together, these results suggest that NSUN2 promotes m⁵C modification on KRT13, and m⁵C methylated KRT13 would be recognized and stabilized by the m⁵C reader protein YBX1. Figure 6I shows the working model by which NSUN2 stabilizes KRT13 via YBX1, through which it regulates cervical cancer invasion and migration.