RNA modification gene WDR4 facilitates tumor progression and immunotherapy resistance in breast cancer

Graphical abstract

Highlights

• •The strategy derived from RNA modifications effectively predicts patient survival and therapy response in breast cancer.
• •The RMscore has great performance with remarkable improvement in comparison with previously reported models in breast cancer.
• •We have provided the first evidence about the tumor-promoting function of m7G-related WDR4 in breast cancer.
• •The tight relationship between RNA modifications and tumor immune microenvironment is confirmed by multi-omics investigation.

Abstract

Introduction

Growing interest toward RNA modification in cancer has inspired the exploration of gene sets related to multiple RNA modifications. However, a comprehensive elucidation of the clinical value of various RNA modifications in breast cancer is still lacking.

Objectives

This study aimed to provide a strategy based on RNA modification-related genes for predicting therapy response and survival outcomes in breast cancer patients.

Methods

Genes related to thirteen RNA modification patterns were integrated for establishing a nine-gene-containing signature-RMscore. Alterations of tumor immune microenvironment and therapy response featured by different RMscore levels were assessed by bulk transcriptome, single-cell transcriptome and genomics analyses. The biological function of key RMscore-related molecules was investigated by cellular experiments in vitro and in vivo, using flow cytometry, immunohistochemistry and immunofluorescence staining.

Results

This study has raised an effective therapy strategy for breast cancer patients after a well-rounded investigation of RNA modification-related genes. With a great performance of predicting patient prognosis, high levels of the RMscore proposed in this study represented suppressive immune microenvironment and therapy resistance, including adjuvant chemotherapy and PD-L1 blockade treatment. As the key contributor of the RMscore, inhibition of WDR4 impaired breast cancer progression significantly in vitro and in vivo, as well as participated in regulating cell cycle and mTORC1 signaling pathway via m7G modification.

Conclusion

Briefly, this study has developed promising and effective tactics to achieve the prediction of survival probabilities and treatment response in breast cancer patients.

Introduction

As a result of high morbidity and mortality rates, breast cancer (BC) has evolved as a dreadful threat to female health globally [1], [2]. The combination of traditional surgical treatment, chemotherapy, radiotherapy, targeted therapy and immunotherapy has made great contributions to improve the clinical outcomes of BC patients. Notwithstanding, a considerable portion of BC patients still derive no benefit from these advanced treatments, leaving many with a significant unmet need [3], [4]. Therefore, we are keen to delve deeper into the underlying biological mechanisms leading to poor outcomes and therapy resistance of BC patients [5].

The identification of RNA modification acting as post-transcriptional regulation has shed new light on BC patients with unfavorable survival probabilities [6]. Several types of RNA modifications have been identified, with modifications of N⁶-methyladenosine (m6A) being the most prevalent. m6A methyltransferases catalyze the modification, m6A binding proteins act as “readers” and recognize the modified peptides, and demethylase “erasers” can remove the m6A mark after catalyzing the modification [7]. For N⁶-2′-O-dimethyladenosine (m6Am), writers include PCIF1, METTL3, and METTL4; eraser involves FTO; and no readers have been discovered [8]. 5-Methylcytosine (m5C) modification is found in transfer RNAs and ribosomal RNAs [9]. Presently, it has been reported that m5C RNA modification is regulated by methyltransferases (NSUN1-7, DNMT1, DNMT2, DNMT3A, and DNMT3B), demethylases (TET2), and readers (ALYREF and YBX1) [10], [11], [12]. As with m5C, N¹-methyladenosine (m1A) modification has been identified frequently in transfer RNAs (tRNAs) [13]. The only known methyltransferase of m1A is the TRM6–TRM61 complex, while m1A can be erased by ALKBH1 and ALKBH3 [14], [15]. Moreover, the 7-methylguanosine (m7G) modification, which is distributed at the 5′ cap region of mRNA and internally within tRNA and rRNA, occurs at the initial stages of transcription [16]. For the newly identified 3-methylcytosine (m3C), only the METTL8 writer has been detected, while erasers and readers are unknown [17]. RNA polymerase II-transcribed RNAs are 5′ capped and then methylated in a process called RNA cap methylation [15]. Pseudouridylation is the most prevalent modification found in total RNA of human cells [18], [19]. mRNA uridylation targets transcripts for degradation mediated by TUT4 or TUT7 [20]. The most prevalent form of RNA editing, adenosine-to-inosine (A-to-I) editing, involves the adenosine deaminases acting on RNA (ADARs) [21]. 5-Methoxycarbonylmethyl-2-thiouridine (mcm⁵s²U) is one wobble modification form of tRNAs, which is correlated with translational fidelity and proteostasis [22], [23]. Alternative polyadenylation (APA) is a widespread mechanism involving in regulation of the diversity of mRNAs and protein diversification [24]. It has long been known that RNA ribose methylation 2′-O-methylation) affects RNA stability through the methylation of the ribose 2′-OH moiety that occurs in all nucleotides [25], [26].

In-depth research of RNA modifications indicates that m6A regulatory proteins are promising therapeutic targets [27]. FTO inhibitors have shown marked anti-tumor effects in vitro and in vivo [28], and the ALKBH5 inhibitor (ALK-04) has been reported to decrease tumor growth by enhancing the efficacy of anti-PD-1 therapy [29]. Moreover, the whole-genome circulating tumor DNA methylation landscape reveals that DNA methylation can serve as a prognostic marker for breast cancer [30]. However, these strategies have failed to achieve the expected efficacy, and targeting RNA modifications remains a long step ahead.

Different modifiers are believed to be responsible for cancer progression and therapy resistance [31]. Studies have shown that dysregulation of RNA modifications facilitates breast cancer development. Most regulatory proteins of RNA modification can promote the proliferation, invasion and metastasis of breast cancer. And these RNA modification regulators have an effect on the breast cancer stem-like cells, epithelial-to-mesenchymal transition, carcinogenesis, glycolysis and immune escape [32]. Hence, the key proteins involved in RNA modification are expected to be promising candidates as therapy targets for breast cancer.

Despite this, a summary understanding of the relationship between RNA modifications and breast cancer is still lacking, and the detailed function of key regulators in RNA modifications requires more investigation. Thus, our study proposed a promising strategy by establishing an available biomarker, RMscore, as a practical means to evaluate patient survival and therapy response with great performance. As a pivotal member of the RMscore, WDR4 promoted breast cancer progression through its involvement in the cell cycle and the regulation of the mTORC1 signaling pathway.

Material and methods

Data collection

An extensive collection of RNA modification-related genes was gathered from review articles and manual collation. The definitive gene set was a tandem collection of key regulatory genes from 13 RNA modification patterns. A total of 13 genes derived from m1A, 4 genes derived from m3C, 15 genes derived from m5C, 24 genes derived from m6A, 1 gene derived from m6Am, 29 genes derived from m7G, 7 genes derived from RNA cap methylations, 8 genes derived from uridylation, 19 genes derived from pseudouracil, 7 genes derived from adenosine-to-inosine RNA editing, 10 genes derived from mcm5s2U, 20 genes derived from APA and 12 genes derived from RNA ribose methylation. A final 114 genes related to RNA modifications were included in the study (Table S1).

The following patient selection criteria was established: (a) the diagnoses were confirmed by histological examination; (b) the overall survival (OS) data were available and follow-up time was greater than 3 months; (c) technical replications should be removed if necessary [33]. We captured the raw data of bulk transcriptome counts from 1113 BC patients and 113 normal samples in The Cancer Genome Atlas (TCGA) database. Masked somatic mutation profiles, clinical characteristics, and CNV information were downloaded from the General Genomic Data Center (GDC, https://portal.gdc.cancer.gov/repository, accessed on July 7, 2022). The clinical information and the microarray gene expression data (Log2 transformed intensity values) of 1904 BC patients were collected from the Molecular Taxonomy of Breast Cancer International Consortium (METABRIC, http://www.cbioportal.org/, accessed on July 4, 2022) dataset. The normalized and log2 converted RNA-sequencing (RNA-seq) profiles and clinical characteristics of 4474 BC patients were downloaded from the Gene Expression Omnibus (GEO, https://www.ncbi.nlm.nih.gov/geo/, accessed on July 26, 2023) databases (GSE96058, GSE20685, GSE24450, GSE42568, GSE41994 and GSE86166). The “AnnoProbe” R package was used to map the probes for gene expression microarray data. In case multiple probes were involved, average values were calculated using the R package “limma” (version 3.52.2). Data from GSE176078 was collected and analyzed with the R package “Seurat” (version 4.3.0.1) to identify single-cell RNA transcriptomes of BC patients [34], [35], [36]. Transcriptome data and treatment response information of breast cancer patients in GSE21974, GSE18728, GSE106977, and GSE33658 datasets were also downloaded from the GEO database.

Differential expression, somatic mutation, and copy number variation analyses

Differentially expressed genes (DEGs) within the TCGA-BRCA cohort among breast cancers and normal tissues were filtered using the “limma” R package. Genes with |log2 (fold change)| > 1 and false discovery rate (FDR)-adjusted P-value < 0.05 were considered to be differentially expressed at the mRNA level. Two types of DNA alterations were recognized: mutations (truncating and missense) and copy number changes (amplification and deep deletion). Mutation Annotation Format (MAF) analysis of TCGA-BRCA was performed with the “maftools” R package. In the TCGA-BRCA dataset, CNV values > 0.2 and <  − 0.2 were considered to be “gain” and “loss”, respectively.

Correlation analysis of RNA modification-related genes and clinical characteristics

The diverse attributes and relationships with other RNA modification-related genes were displayed in a circular diagram with the assistance of the R package called “RCircos”. More precise correlations between RNA modification-related genes and relationships between gene expression matrices and survival status matrices, survival time matrices, stage matrices, and pathology type matrices were demonstrated with the R package “linkET” [37].

Establishment of RNA modification-related gene signature

To identify RNA modification-related genes significantly associated with OS, the “survival” R package was utilized for performing univariate Cox regressions. To avoid missing important genes, we set the cut-off p value at 0.1. The “glmnet” package was applied to achieve a least absolute shrinkage and selection operator (LASSO)-Cox regression model analysis. Ultimately, RNA modification-related risk score (RMscore) was calculated using weighted LASSO-Cox coefficients based on individual expression levels of each Σ_i=1ⁿsignature gene: RMscore = (β_i ∗ exp_i), where β_i indicates the corresponding coefficient of a particular RNA modification-related gene which was selected previously, and exp_i represents its expression level. The RMscore was adjusted using a linear transformation to make plots more intuitive according to the following formula:$\mathit{Adjusted}RMscore=\frac{\mathit{calculated}RMscore-RMcor{e}_{\min }}{\mathit{RMscor}{e}_{\max }-RMscor{e}_{\min }}$ [38]. Subsequently, patients were divided into low-RMscore and high-RMscore groups according to their median value. Principal component analysis (PCA) was carried out using the “stats” package to examine the clustering effect by RMscore. Survival curves of patients in the low-RMscore group and high-RMscore group were estimated by Kaplan-Meier method using the R package “survival” and “survminer” comparison by log-rank test.

Functional enrichment analysis

Based on DEGs, we identified potential biological functions between the high- and low-RMscore groups with the “h.all.v7.5.1.entrez.gmt” [HALLMARK] database described in the Molecular Signatures Database (MSigDB) (R package “clusterProfiler”). In addition, the implementation of Gene Set Variation Analysis (GSVA) using the “GSVA” R software package was employed to investigate the correlation between different RMscore groups and relevant biological processes.

Construction and verification of a prognostic nomogram

After collecting clinical data of BC patients from TCGA-BRCA cohort for analyzing together with the RMscore, we performed univariate and multivariable Cox regression analyses. Constructing the nomogram based on the RMscore and prognostic clinical factors was achieved by using the “rms” and “regplot” R packages. The predictable efficacy was evaluated by calibration plots and Receiver Operating Characteristic (ROC) analyses by means of the R package “ggDCA” and “timeROC”.

We retrospectively collated three RNA modification-related signatures for comparison, including a N6-methyladenosine regulators-related gene signature (m6A1) from Wang et al. [39], a m6A-related gene signature (m6A2) from Li et al. [40], and a RNA adenosine modification-related gene signature (RMW) from Ni et al. [41]. We also collected two published signatures with high predictive performance in breast cancer for comparison, including a diverse cell-death patterns-related gene signature (CDI) from Zou et al. [38], and an immune cell infiltration-based signature (immscore) from Sui et al. [42]. Each patient's score was calculated according to the formula given in the corresponding publication and modeled by incorporating the clinical characteristics required in their article. Subsequently, we calculated the AUC values and the Harrell's concordance index (C-index) of each model in the training and validation sets, and obtained the net reclassification index (NRI) of our model compared to the other models through the R package “nricens”. In more detail, we speculated on the benefit of our model by calculating the hazard ratio (HR) value of each gene signature and the integrated discrimination improvement (IDI) value of RMscore compared to other risk scores (R package “survIDINRI”).

Evaluation of RMscore on tumor immune infiltration

We chose two different dimensions to speculate on tumor immune infiltration levels and analyzed their relationship with RMscore. CIBERSORT algorithm quantified tumor-infiltrating immune cells by deconvolution of cell mixtures based on expression signatures [43], and xCell algorithm [44] was a marker gene-based method for quantifying tumor-infiltrating immune cells. In GSE176078, the immune cells affiliation of each patient and their relative abundances of both major type and minor type of immune cells were depicted by Krona charts constructed using the ktImportText command of the KronaTool v2.7 (https://github.com/marbl/Krona) [45]. The CytoTRACE algorithm was used to compare cellular differentiation states among malignant cells (R package "CytoTRACE" v0.3.3) [46]. Based on CytoTRACE scores, higher scores indicated increased tumor stemness (less differentiation) and vice versa [47]. Communication differences of immune cells between high and low RMscore groups were also computed by R package “CellChat” [48].

Prediction value of the RMscore on therapy benefits

We characterized the connection between the RMscore and the expression of known immunomodulators. R package “oncoPredict” was applied for drug sensitivity prediction [49]. We collected the expression profile, follow-up survival information, and therapy records from the R package "Imvigor210CoreBiologies", GSE21974, GSE18728, GSE106977, and GSE33658 [50]. From the normalized raw count data, we calculated the RMscore and investigated its impact on prognosis and effectiveness of therapy.

Cell culture and small inhibitory RNA (siRNA) transfection

Cell lines purchased from American Type Culture Collection (ATCC) included MCF-10A, MCF-7, SKBR3, MDA-MB-231, BT-549, and SUM159PT, and all were maintained in culture under standard 5 % CO₂ and 37 °C conditions. Cancer cell lines were grown in DMEM with 10 % FBS and 1 % penicillin–streptomycin, and the MCF-10A cells were cultured in DMEM/F12 with necessary supplements. For siRNA transfection, 1 × 10⁶ cells were plated in six-well plates and cultured overnight, then siRNA transfection was conducted using Lipofectamine™ 3000 reagent (Invitrogen, USA) following the product instructions (Ribio, China). The siRNA sequences targeting WDR4 were as follows: siRNA#1-AAGAGGCTGTCATCATCACTG; siRNA#2-TCTAACAGCATAGACAGGTGC. Total RNA was extracted from the indicated cells using a total RNA Quick Extraction Kit (Qiagen, China) after incubating for 24 h, and whole-cell proteins were isolated for western blot analysis in RIPA (Beyotime, China) from indicated cells 72 h after transfection.

Lentivirus infection for construction of stable cell lines

Lentivirus containing small hairpin RNA targeting human WDR4 (5′-AAGAGGCTGTCATCATCACTG-3′, Obio Technology) or negative control (NC) was separately transfected into human breast cancer cell lines (MCF-7 and SUM159PT) for 24 h. Afterwards, cells were screened with puromycin (1 µg/ml) for 10 to 14 days to obtain stable WDR4 knockdown cell lines. The GFP-positive cells were pooled as shWDR4 and shNC and used for subsequent assays [51].

Quantitative real-time PCR (RT-qPCR) and western blots

After RNA extraction, cDNA synthesis (Takara, RR036A) and real-time qPCR (Takara, RR420A) were performed using the BioRad CFX96 RT-qPCR instrument. Normalizing to β-actin, all target gene expression levels were calculated with the 2^-ΔΔCt method after three repetitions of each reaction. Primer sequences used are listed in Table S1. The Western blots were performed according to standard protocols, and proteins electrophoresed on SDS-PAGE, transferred to PVDF membranes, blocked in 5 % milk/TBST buffer for 1 h, incubated with secondary antibodies for 1 h at room temperature followed by incubation with primary antibodies overnight at 4 °C and detection with an ECL luminescence kit (Yeasen, China). Antibodies used for Western blots are also listed in Table S1.

RNA m7G dot blot

RNA was extracted and concentrations were determined. The RNA was then diluted to 500 ng/uL, 1000 ng/uL. After treating with the decapping enzyme (NEB, M0608S), the RNA was denatured by heating at 95 °C for 5 min and 1 uL of each of these concentrations was then dropped onto a nylon membrane, positively charged (Roche, USA). Subsequently, the membranes were cross-linked twice using a UV crosslinker, blocked with 5 % milk/TBST buffer for 1 h, incubated with m7G antibody overnight at 4 °C, and incubated with HRP-conjugated secondary antibody, and detected with an ECL luminescence kit. The membrane stained with methylene blue (MB) was used as a control.

m7G RNA immunoprecipitation assay (MeRIP)

m7G RNA immunoprecipitation assay experimental methods and detailed procedures were carried out as described previously [52]. It should be noted that RNA in this assay were treated with decapping enzyme for further experiments. The experiment was performed by using BersinBio™ RNA Immunoprecipitation (RIP) kit (Catalog Bes5101). TRIzol (Invitrogen, CA, USA) was used to extract coprecipitated RNA and RT-qPCR was performed to analyze it.

In vitro cell experiments

All cell experiments in vitro were repeated at least for three times, including the cell counting kit 8 (CCK8) proliferation, EdU proliferation, plate clone formation, transwell migration, and scratch wound healing assays. The cells were re-plated as required for each experiment 48 h after siRNA transfection, and subsequent experimental procedures followed those described in our previous study [53].

Cell cycle experiment

For the cell cycle experiment, cells were transfected with siRNAs targeting WDR4 for 48 h, then collected for flow cytometry, which was performed according to instructions for the Cell Cycle Kit (KeyGen, China). The Modifit LT software was used to conduct the cell cycle analysis.

Immunohistochemistry (IHC) staining

To carry out the following procedures, paraffin-embedded tissue blocks were serially sectioned into 3 µm sections. The sections were dewaxed and dehydrated, followed by antigen retrieval and endogenous peroxidase blocking, then treated with primary antibodies overnight at 4 °C, followed by secondary antibodies at room temperature for 30 min. The sections were stained using DAB and hematoxylin, dehydrated, and covered for imaging. The IHC images were assessed by a semiquantitative scoring system (H-score), which evaluated staining intensity as well as percent positivity [54].

In vivo tumor xenograft model

Four-week-old female BALB/c nude mice were purchased from Vital River Laboratories Animal Technology (Guangzhou, China). Sixteen nude mice were randomly divided into four groups, and four mice of each group were injected with indicated breast cancer cells (MCF-7-shNC, MCF-7-shWDR4, SUM 159PT-shNC, SUM 159PT-shWDR4) at 1 × 10⁶ were suspended in 100 µl PBS and injected into the unilateral mammary fat pad of each nude mouse. Tumors were measured every seven days; the mice were sacrificed four weeks after inoculation and the tumors were collected. Tumor volumes were calculated using the following formula: $V=(length\times width$²$)/2$. In vivo animal experiments were performed in accordance with the principles for the use of laboratory animals of the Animal Care and Use Committee of Sun Yat-sen University.

Statistical analyses

All statistical analyses were conducted by applying R software (Version 4.3.1) and GraphPad Prism (Version 8.0) software. Comparisons between two groups were analyzed using t-tests, while multiple groups were compared using one-way ANOVA. The correlation coefficient was evaluated by Spearman’s test. Statistical significance was set as p < 0.05 and denoted by asterisks (*p value < 0.05, ** p value < 0.01, *** p value < 0.001, and **** p value < 0.0001, ns > 0.05). All experiments in the study were repeated three times.

Results

A comprehensive picture of RNA modification-related genes in breast cancer

The far-reaching role of various RNA modifications in cancer is coming to light, including m6A, m5C, m7G, m1A, m3C, m6Am, RNA cap methylations, uridylation, pseudouracil (ψ), A to I RNA editing, mcm5s2U, APA, and RNA ribose methylation. For a more outright understanding of RNA modifications in breast cancer, we obtained 114 RNA modification-related genes co-expressing in three datasets (TCGA-BRCA, METABRIC, and GSE96058) after intersection (Fig. S1A). Ten genes were found to be significantly differentially expressed between cancer and paracancerous tissues in TCGA-BRCA (adjusted, p value < 0.05 and |log2FC| > 1) (Fig. S1B). The chromosomal localizations, expression, and relationships with each other were visualized in Fig. S1C. We observed 42.39 % BC patients with gene mutation after assessing the variation of RNA modification genes in TCGA-BRCA. The top 20 mutated genes displayed mutation frequency ranging from 2 % to 5 % (Fig. 1A). CNV analysis revealed that RNA ribose methylation-related TARBP1 had the greatest copy number amplification while m6A-related RBMY1A1 had the most notable CNV deletion (Fig. 1B). Respectively, we identified 33 and 55 OS-associated genes in TCGA-BRCA and METABRIC cohorts by univariate COX regression analysis with log-rank p value < 0.1. There were 18 genes involved in the intersection of the two outputs, including NSUN2, EIF4G3, GEMIN5, METTL2A, DNMT3B, FUS, MTPAP, CPSF6, RNGTT, DKC1, YTHDF1, LARP1, WDR4, EIF4E3, TET2, TRMT61B, YTHDF3, and NUDT3. Correlation and mantel test analyses in all three cohorts revealed that uridylation-related MTPAP expression was prominently correlated with survival outcome and PAM50 subtype, while expressions of pseudouracil-related DKC1 and m7G-related WDR4 were markedly associated with survival time and PAM50 subtype (Fig. 1C).

Fig. 1.

The alteration landscape and relationship network of RNA modification (RM) related genes in breast cancer (BC). (A) Oncoplot of RM-related genes in the TCGA-BRCA cohort. (B) Copy number variations of RM-related genes in the TCGA cohort. (C) The correlation analysis and mantel test analysis of RM-related genes which were associated with patient survival.

Development and validation of an effective signature derived from RNA modification-related genes

The hazard ratio (HR) of OS-related genes in TCGA-BRCA was shown in a circle graph (Fig. 2A). A total of 18 OS-related genes were encompassed with LASSO Cox regression analysis (Fig. 2B and C). Then, a nine-gene signature was established after integrating five genes derived from m7G (EIF4G3, EIF4E3, WDR4, LARP1 and NUDT3), one gene from m3C (METTL2A), one gene from APA (FUS), one gene from m5C (TET2), and one gene from m6A (YTHDF3). The Kaplan-Meier survival analyses of nine model genes for OS in TCGA-BRCA were shown in Fig. S1D. We assigned each BC sample an RNA modification-related score (RMscore) according to the following formula: RMscore = (0.132620 * EIF4G3 exp.) +(−0.282585 * EIF4E3 exp.) +(0.280154 * WDR4 exp.) +(0.260614 * LARP1 exp.) + (−0.125143 * NUDT3 exp.) + (0.035712 * METTL2A exp.) + (−0.476939 * FUS exp.) + (−0.021100 * TET2 exp.) + (0.090359 * YTHDF3 exp.).

Fig. 2.

Construction and evaluation of the RMscore in predicting prognosis of BC patients. (A) Hazard ratio (HR) of survival-associated RM-related genes in TCGA-BRCA cohort. (B) In the 10-fold cross validation, the LASSO regression model revealed partial likelihood deviance. (C) Cross-validation of the LASSO coefficients of selected RM-related genes. The optimal values were marked with vertical dotted lines. (D) Box plots of the relationship between RMscore and vital status, N stage, and PAM50 subtype in the TCGA-BRCA cohort. The color of * means that the group with the corresponding color is the comparison. (E) Kaplan–Meier plots of overall survival in the high-RMscore and low-RMscore patient groups in TCGA-BRCA, METABRIC, GSE96058, GSE20685, GSE24450, GSE42568, GSE41994 and GSE86166 cohorts. (F) Ridge plot of ten biological functions showing the top enrichment score between high-RMscore and low-RMcsore groups in TCGA-BRCA, METABRIC, and GSE96058 cohorts.

Consequently, BC patients in the TCGA-BRCA dataset were classified into high-RMscore and low-RMscore groups according to the median value of the RMscore. The correlations between each signature gene, the RMscore, and clinical features in TCGA-BRCA, METABRIC, and GSE96058 cohorts were visualized by heatmaps (Fig. S2A–C). The RMscore was tightly coupled with patient survival status, N stage, and PAM50 subtypes in the TCGA-BRCA (Fig. 2D), as well as the METABRIC (Fig. S3A) and GSE96058 cohorts (Fig. S3B). We next surveyed the predictive performance of the RMscore in patient prognosis through a series of analyses. BC patients in each set, including the TCGA-BRCA training set and other three validation sets (METABRIC, GSE96058, and GSE20685), were categorized into high and low RMscore subgroups, respectively (Fig. S3C). A noticeable discrepancy in OS time between the two groups revealed that patients with high RMscores tended to have higher death rates (Fig. S3D). Principal component analysis (PCA) suggested that RMscore-based segmentation was reliable (Fig. S3E). We then performed Kaplan–Meier survival analyses in the eight cohorts (TCGA-BRCA, p value < 0.0001; METABRIC, p value = 0.04; GSE96058, p value < 0.0001; GSE20685, p value < 0.0001; GSE24450, p value = 0.0071; GSE42568, p value = 0.0011; GSE41994, p value = 3e − 04; GSE86166, p value = 0.038), which also demonstrated that high RMscores could be an available marker for unfavorable OS of BC patients (Fig. 2E). Moreover, further Kaplan–Meier survival analyses in TCGA-BRCA and GSE20685 indicated that high RMscores always resulted in shorter time of progression-free survival (p value = 0.00025), disease-free survival (p value = 0.0085), disease-specific survival (p value = 0.0011), and distant metastasis-free survival (p value = 0.00095) (Fig. S3F). Thus, the RMscore signature derived from RNA modifications has great value for predicting patient survival outcomes. Further, we applied GSVA to exploit the biological distinction within the two RMscore subgroups. From the results, mTORC1 signaling and G2M checkpoint hallmarks were identified as the most important biological functions affected by the RMscore levels (Fig. 2F).

Construction and evaluation of an effective nomogram based on RMscore

As part of our efforts to determine if the RMscore could be a reliable prognostic indicator, we conducted the following univariate Cox regression analysis in TCGA-BRCA and identified the RMscore as a disadvantage for OS (HR = 3.629, 95 % CI: 2.425–5.430, and p value < 0.001, Fig. S4A). Moreover, the multivariate Cox regression analysis also regarded the RMscore as an independent characteristic of BC (HR = 3.646, 95 % CI: 2.409–5.518, and p value < 0.001, Fig. S4B). After integrating age, N stage, and the RMscore by multivariable Cox regression analyses in the TCGA-BRCA, a nomogram model was established to assess the 2-, 3-, and 5-year OS (Fig. S4C). The calibration curve of this model was accurate in predicting the 2-, 3-, and 5-year OS survival rates (Fig. S4D). The DCA curve for 5-year OS proved a superior prognostic utility of this model compared to other individual parameters (Fig. S4E). Naturally, we conducted ROC analyses in the four independent datasets and the area under curve (AUC) values disclosed promising predictive accuracy of the nomogram in predicting 2-, 3-, and 5-year survival probabilities of BC patients (Fig. S4F). The 2-year, 3-year, and 5-year AUC values were above 0.7 for the TCGA-BRCA, GSE96058, and GSE20685 cohorts. Only for the METABRIC cohort, the AUC values were around 0.680.

To further characterize the effectiveness of RMscore, we collected several valid models from previous researches, including m6A1, m6A2 and RMW derived from RNA modification, cell death patterns score (CDI), and immune cell infiltration scores (immscore) (Fig. 3A). ROC curves performed in the training and validation sets indicated that the RMscore significantly outperformed the other features in the TCGA-BRCA cohort, despite an only slightly improved advantage in the other validation set. Given these results, we applied NRI (a more sensitive method than AUC) to evaluate the improvement of the RMscore-based nomogram (Fig. 3B). In the TCGA-BRCA, GSE96058, and GSE20685 cohorts, the RMscore-based model showed a greater degree of improvement in predicting probability of 5-year survival of BC patients compared to all other models. From the calculated C-index of each model, we concluded that the prognostic nomogram derived from the RMscore performed better than the other models, which was consistent with the above findings (Fig. 3C). However, the enhancement of predictive performance for our model was limited compared to the m6A2-based nomogram, which might be a result of the m6A2-based model combining clinical indicators with p value > 0.05 in the multiple regression Cox analysis. In view of this, we calculated HR values from one-way Cox regression analysis for each risk score (Fig. 3D, E) and assessed the predictive benefit of the RMscore over other scores by calculating integrated discrimination improvement (IDI) (Fig. 3F). As expected, the RMscore was more effective in predicting BC patient’s survival (all p value < 0.05) with higher HR values than other risk scores in both the training and validation sets. The IDI values further validated this conclusion. Thus, our analyses disclosed the non-negligible importance of RNA modifications in breast cancer through the generation of RMscore.

Fig. 3.

Comparison of the RMscore-based nomogram to other published nomograms. (A) Receiver operator characteristic (ROC) analyses of nomograms for 5-year OS in TCGA-BRCA, METABRIC, GSE96058, and GSE20685 cohorts. (B) Net reclassification index (NRI) values of the RMscore-based nomogram compared to other nomograms for 5-year OS. Dots represent NRI values. The upper and lower bounds of the bars indicate 95% confidence interval. (C) The Harrell's concordance index (C-index) values of the RMscore-based nomogram and the other nomograms. Dots represent C-index values. The upper and lower bounds of the bars indicate 95% confidence interval. (D) Hazard ratio (HR) values of the RMscore and other risk models. Dots represent −log10(p value). The upper and lower bounds of the bars indicate 95% confidence interval. (E) Radar chart of HR values of the RMscore and other risk models. (F) The integrated discrimination improvement (IDI) values of the RMscore compared to other risk models.

High RMscores denoted immunotherapy resistance in breast cancer

Immunotherapy has shown great benefits in improving BC patient prognoses, such as the monoclonal antibody blocking the PD-L1 immune checkpoint [55]. The RMscore and signature-contained genes were significantly correlated with widely recognized immunomodulators, suggesting a potential effect of the RMscore in predicting clinical response of immunotherapy treatment (Fig. 4A). Specifically, BC tumors with higher RMscores seem to express lower levels of immunomodulators, including most of major histocompatibility complex (MHC) molecules, which primarily acts as an antigen presenter. We speculated on the immune cell infiltration of tumors with different RMscores by two algorithms based on different dimensions. By applying the CIBERSORT and xcell algorithms in TCGA-BRCA, we found that macrophage M0 and M2 cells infiltrated more in tumors with high RMscores (Fig. S5A and B). Infiltration levels of mast cells, memory B cells, NK cells, and Tregs similarly diverged between the high and low RMscore groups. The above findings suggested that patients with high RMscores would have a more severely immunosuppressive microenvironment.

Fig. 4.

Predictive role of the RMscore in the tumor immune microenvironment. (A) Dot plot of the correlation between immunomodulators and the RMscore in BC patients. (B) Violin plot of the cytolytic activity, immune signature score and TLS signature score across high and low RMscore groups in BC patients. (C and D) Distribution of patients with different RMscore levels regarding to anti-PD-L1 therapy in the IMvigor210 cohort. (E) Kaplan-Meier curve of OS for patients with different RMscore levels in the IMvigor210 cohort. (F) Heatmap of RMscore, clinicopathologic features, immunotherapy responses, gene mutation status, and core biological pathways.

There are other reliable immune biomarker scores that can predict response to immune checkpoint inhibitors, such as the cytolytic activity [56], the immune signature score [57], and the tertiary lymphoid structure (TLS) signature score [58]. We performed statistical calculations in TCGA-BRCA cohort based on formulas previously reported in the literature. We found that the high RMscore group had lower cytolytic activity, immune signature score and TLS signature score, indicating that BC patients with high RMscores may have immune microenvironment suppression and therapy resistance to immune checkpoint inhibitors (Fig. 4B).

To delve into the potential predictive value of RMscore in immunotherapy responses, BC patients of the IMvigor210 cohort (anti-PD-L1 treatment cohort, n = 348) were divided into high and low RMscore groups. Among BC patients receiving anti-PD-L1 therapy, fewer patients in the high RMscore group achieved partial or complete response, which led to eventual disease progression (Fig. 4C, D). BC patients with low RMscores gained significant clinical benefits by prolonging OS time (Fig. 4E). Next, we examined the effect of RMscore on PD-L1 levels in immune cells (IC) and tumor cells (TC), tumor mutation burden (TMB), mutation frequency of common mutation genes in breast cancer, and activation of core biological pathways. The heatmap indicated that tumors with high RMscores, accompanied by lower PD-L1 expressions in both IC and TC, had higher TMB and greater mutation frequency of common mutated genes (BRCA, FGFR3, CDKN2A and ATM). The inhibition of the CD8 + T effector, APM, immune checkpoint pathway activities in the high RMscore group versus the low RMscore group were also shown in the heatmap (Fig. 4F). In summary, these results verified that the RMscore could be a definite biomarker for immunotherapy response in BC patients.

Dissection of tumor immune microenvironment based on RMscores

Exact distribution of the RMscore in BC tumor immune microenvironment was investigated using transcriptome data from single cells in GSE176078. The major cell types were annotated and the RMscore of each cell was calculated and presented in t-SNE plots (Fig. 5A). To better demonstrate this, a Krona pie chart visualized the proportion of the respective major and minor cell subpopulations of nine patients with the highest cell counts detection (Fig. 5B). Next, we illustrated the distribution correlation between the RMscore and model gene expressions across cell types (Fig. 5C, D). Consistently, we found that the RMscore was relatively higher in cancer epithelial cells and myeloid cells (mainly composed of macrophages) compared to other cell types. Further, we surveyed the profiles of the RMscore and model genes in cancer epithelial cell subtypes (Fig. 5E, F and Fig. S5C). Shown in the UMAP plots, a high level of RMscore was observed in both cancer basal cells and cancer high cycling cells. Among the model genes, writer and reader genes (YTHDF3, METTL2A, WDR4) were co-expressed in cells with high RMscores, while the eraser gene TET2 was mainly expressed in cells with low RMscores. Growing evidence strengthens a significant association between tumor stemness and drug resistance resulting from immune evasion. To minimize inter-sample variation for better prediction, we downloaded the bulk RNA-Seq data of these patients and filtered out triple-negative breast cancer patients who were grouped by RMscore levels. Surprisingly, the high RMscore group had a higher prevalence of cancer cell stemness (p value = 3.49E−11, Fig. 5G and Fig. S5D), which implied that the RMscore might be indicative of breast cancer stemness.

Fig. 5.

Landscape of tumor microenvironment based on the RMscore signature. (A) t-SNE plot visualization of all cell subtypes and RMscore of each cell from 26 BC patients. (B) Krona pie chart of major and minor cell subpopulations of 9 patients with the highest cell count detection in GSE176078. (C) Violin plot of RMscore values across all cell types. (D) The bubble plot of the average expression and percent expression of model genes in different cell types. (E and F) UMAP plot visualization of the distribution of RMscore value and expression of model genes in the different cancer epithelial subtypes. (G) Left tSNE plot depicts the distribution of CytoTRACE scores among TNBC patients’ cancer epithelial cells. Dark-blue indicates lower CytoTRACE scores (low stemness) while dark-red indicates higher CytoTRACE scores (high stemness). Right tSNE plot labeled the TNBC patients’ cancer epithelial cells by the RMscore groups. (H) Cell chat analysis between all cell subtypes comparing the number and strength of interaction between high-RMscore group and low-RMscore group. Red indicates more numbers and stronger strength in the high-RMscore group than in low-RMscore group and vice versa. (I) Projecting cell types onto a two-dimensional manifold according to their incoming and outgoing interaction strenth. (For interpretation of the references to color in this figure legend, the reader is referred to the web version of this article.)

Subsequently, we performed cell-chat analyses in patients of both the high and low RMscore groups. Cellular communications between each specific cell type and other cell types were shown in Fig. S6. Cell communications in both number and strength between cancer epithelial cells and the other cell types were obviously reduced in the high RMscore group, except myeloid cells and T-cells (Fig. 5H and Fig. S7A). Furthermore, the number and strength of cellular communications between T-cells and myeloid cells were significantly increased in high RMscore group. For better visualization, we demonstrated the strength of incoming and outgoing cell signaling communications by a two-dimensional spatial plot, which showed that both myeloid cells and T-cells were significantly enhanced in the high RMscore group (Fig. 5I). And we also demonstrated the number of cellular communications between and within cancer epithelial cells, T-cells and myeloid cells in low and high RMscore groups in Fig. S7B. The results were consistent with the previous findings that there was increased signaling inflow from cancer epithelial cells to T-cells as well as cellular communication between T-cells and myeloid cells in the high-RMscore group compared to the low-RMscore group. For a more in-depth analysis, we compared the relative information flows in the high and low RMscore groups, indicating that the low RMscore group was mainly enriched in information flows related to cell adhesion and intrinsic immunity such as SELE, SPP1, ANGPT, EDN, DESMOSOME, TNF and COMPLEMENT, while the high RMscore group was mainly enriched in information flows related to cell growth and cellular immunity such as MHC-I, MHC-II, BAFF, IL10, CD45, LCK, VEGF and EGF(Fig. S7C). Considering the marked enhancement of cellular communication between T cells and myeloid cells in the high RMscore group, we explored the cellular communication of ligand-receptor pairs between T-cells and myeloid cells (Fig. S7D). And we found that HLA-CD4 and HLA-CD8 signaling were significantly enhanced in the high-RMscore group. Finally, we probed the signal strength of ligand receptor pairs between cancer epithelial cells and other cells (Fig. S7E,F). In agreement with earlier findings, high RMscores were found to be correlated with enhanced intensity of MIF and its ligands primarily with T cells and myeloid cells (associated with tumor tumorigenesis and immunosuppression) [59], [60], enhanced intensity of MDK and its ligands primarily with T cells, B cells and myeloid cells (associated with tumor immunosuppression) [61], and diminished intensity of THBS1 and its ligands primarily with tumor epithelial cells themselves and with endothelial cells (inhibiting angiogenesis and thus tumor progression) [62].

The implication of RMscore in drug sensitivity of breast cancer

To determine whether the RMscore was correlated with drug sensitivity in BC, we computed the half maximal inhibitory concentration (IC50) values for several drugs, including Oxaliplatin, Palbociclib, SCH772984 and PD173074. The correlation landscape between drug sensitivities, the RMscore and expression of each model gene were shown in Fig. 6A. We found that the RMscore was obviously correlated with IC50 values of most drugs, which signaled that multidrug-resistance (MDR) is involved with high RMscores in BC patients. We also observed higher IC50 values of Oxaliplatin, Palbociclib, and SCH772984 and a lower IC50 value for PD173074 in the high RMscore group (Fig. 6B). These results suggested that BC patients with high RMscores may be resistant to therapy regimens such as platinum-based chemotherapy drugs, CDK4/6 inhibitors, and ERK1/2 inhibitors, respectively, while sensitive to the FGFR inhibitor. To further validate the potential efficacy of the RMscore in predicting drug sensitivity, we performed pairwise comparisons in GSE21974 and GSE18728 (Fig. 6C). A decreased RMscore was observed in BC patients after adjuvant chemotherapy treatment (p value = 0.00115 for GSE21974; p value = 0. 0135 for GSE18728). The result from GSE106977 dataset indicated more patients in the high RMscore group (74.58 % vs. 48.33 %) suffered resistance to standard chemotherapy (p value = 0.0033; Chi-square test) (Fig. 6D). Moreover, BC patients receiving neoadjuvant endocrine therapy in GSE33658 dataset were divided into four groups: progressive disease (PD), stable disease (SD), partial response (PR), and complete response (CR). Fig. 6E demonstrated that more patients with CR were included in the low RMscore group, while more patients with PD sorted into the high RMscore group (p value = 0.046; Fisher exact test). Thus, higher RMscores of breast tumors indicated greater resistance to neoadjuvant chemotherapy as well as endocrine therapy and targeted therapy.

Fig. 6.

Potency of the RMscore signature in predicting drug sensitivity. (A) Bubble plot of the correlation between IC50 of drugs, RMscore values, and expression of model genes. (B) Raincloud plots comparing IC50 of drugs between high-RMscore and low-RMscore groups, and correlation plots between the IC50 of drugs and RMscore values in TCGA-BRCA cohort. (C) Boxplots comparing the RMscore alteration before and after neoadjuvant chemotherapy in GSE21974 and GSE18728. (D) Bar graph illustrating the treatment response to neoadjuvant chemotherapy in high-RMscore and low-RMscore patients of GSE106977. (E) Bar graph visualizing the therapeutic response [complete response (CR), partial response (PR), stable disease (SD), and progressive disease (PD)] to neoadjuvant endocrine therapy and targeted therapy in high-RMscore and low-RMscore patients of GSE33658.

High RMscore implied hallmarks of cancer and low PD-1/PD-L1 expression

We reaffirmed the association of RNA modification with cancer cell cycle and mTORC1 signaling in the light of the RMscore. Postoperative BC samples were obtained from 11 patients for extraction of total RNA. After detecting the mRNA levels of nine model genes and calculating the median RMscore for each sample, we were able to classify these patients into high and low RMscore groups (Fig. 7A). After choosing the three highest and three lowest scored specimens, we examined the expression of relevant proteins by IHC. As shown in Fig. 7B, both cell cycle-related proteins (Cyclin A1, Cyclin D1, and STAT3) and mTORC1 downstream proteins (p-4EBP1 and p-p70S6K) were highly expressed in tissues with high RMscores. This finding shed favorable light on how RNA modification affects cell cycle management and mTORC1 regulation. Then, we collected images of breast enhanced CT scans from the above six patients at the time of diagnosis (Fig. 7C). Generally, tumors from patients with high RMscore appeared to have more malignant phenotypes, including larger size, more uncertain borders, and more obvious enhancement. Since PD-1/PD-L1 expression levels correlate with response to immune checkpoint therapy, we attempted to detect PD-1/PD-L1 expression in both high and low RMscore tissue sections using immunofluorescence. We found that PD-1/PD-L1 expression was indeed weaker in the high-RMscore tumors (Fig. 7D and Fig. S8A), which reasonably explained poorer therapeutic response to PD-1/PD-L1 in this population.

Fig. 7.

The positive association of the RMscore, regulation of cell cycle and mTORC1 signaling, and PD-1/PD-L1 infiltration in breast cancer. (A) Examination of nine model gene expressions in 11 breast cancer samples by RT-qPCR. (B) Comparison of regulatory protein expressions involving cell cycle and mTORC1 signaling between breast cancer samples with high RMscore and low RMscore, respectively. (C) The representative images of breast enhanced CT scan from patients with high and low RMscore, respectively. (D) Representative images detecting PD-1/PD-L1 expressions using immunofluorescence in BC samples with high and low RMscore, respectively.

m7G-related WDR4 boosted breast cancer progression

To clarify the involvement of RNA modifications in BC, we took note of m7G-related WDR4 since it has the most significant positive correlation with patient prognosis. Thus, we conjectured that WDR4 might be a key player of RNA modification participating in BC development. As shown in Fig. 8A, WDR4 expression was closely correlated with vital status and PAM50 subtype, and higher WDR4 expression reflected worse survival probabilities in both TCGA-BRCA and SYSUCC cohorts (Fig. 8B and Fig. S8B). We characterized the protein expression of WDR4 in breast cancer specimens by IHC (Fig. 8C and Fig. S8C) and western blots (Fig. 8D). WDR4 had the highest expression concentration in triple-negative breast cancer, though it was upregulated in all subtypes of breast cancer. We found that WDR4 was abnormally upregulated in most BC cell lines except MDA-MB-231, compared to the non-tumorigenic MCF-10A cells (Fig. 8EandF). In both MCF-7 and SUM 159PT cells, we performed effective knockdown of WDR4 using siRNAs (Fig. 8G and H). By using CCK8 (Fig. 8I), colony formation (Fig. 8J), and EdU assays (Fig. 8K), we observed that WDR4 knockdown impaired the proliferative capacity of BC cells. Furthermore, the transwell (Fig. 8L) and scratch wound healing assays (Fig. 8M) demonstrated that WDR4 silencing significantly attenuated the migration of BC cells. Similarly, we overexpressed WDR4 in MCF-7 and SUM 159PT (Fig. S8D) and further confirmed that WDR4 promoted breast cancer aggressiveness in vitro (Fig. S8E–I). Finally, we confirmed that inhibition of WDR4 restrained breast cancer tumor growth in vivo in xenograft models (Fig. S8J). Immunohistochemical staining of the obtained tumor tissue sections also displayed that KI67 expression was also diminished within tumors developed by WDR4 knockdown cells (Fig. S8K).

Fig. 8.

Verification of WDR4 biological function in breast cancer. (A) Distribution of different clinical indicators between different WDR4 expression groups in TCGA-BRCA cohort. (B) Kaplan-Meier curve of OS for patients with different WDR4 expression groups in both TCGA-BRCA and SYSUCC cohorts. (C and D) Detection of WDR4 expression in breast cancer samples by immunohistochemistry (C) and western blots (D). (E and F) Detection of WDR4 expression in breast cancer cell types by RT-qPCR (E) and western blots (F). (G and H) Detection of WDR4 expression by RT-qPCR in MCF-7 and SUM 159PT cells with siRNA knockdown (G) and western blots (H). (I-K) Examination of the effect of WDR4 knockdown on breast cancer cell proliferation using CCK8 proliferation assay (I), plate clonal formation assay (J), and EdU proliferation assay (K). (L and M) Examination of the effect of WDR4 knockdown on breast cancer cell migration using transwell migration assay (L) and scratch wound healing assay (M).

WDR4 engaged in the regulation of cell cycle and mTORC1 signaling

WDR4 has been identified as a cytoplasmic writer of RNA m7G modification, which was reaffirmed in breast cancer. The immunofluorescence assay showed that m7G levels were reduced in WDR4 knockdown cells while correspondingly increased in WDR4 overexpressing cells (Fig. 9A). And the dot blot assay also revealed that WDR4 inhibition led to a lower m7G level while overexpressed WDR4 led to a higher m7G level (Fig. 9B). Since GSEA has revealed enrichments of G2M checkpoint and mTORC1 signaling hallmarks in the high RMscore group, we hypothesized that WDR4, as a critical participant of the RMscore signature, might promote BC progression by regulating G2M checkpoint and mTORC1 signaling. We began with an analysis of GSEA based on the WDR4 expression matrix, which displayed a robust correlation with G2M checkpoint and mTORC1 signaling (Fig. 9C, p value < 0.001), which was supported by Pearson correlation analysis (Fig. 9D). Next, the effect of WDR4 knockdown on cell cycle was determined by flow cytometry assays in MCF-7 and SUM 159PT cells. And we found that G2/M phase proportion decreased markedly in indicated cells treated with WDR4 siRNAs (p value < 0.0001)), suggesting that WDR4 can induce considerable G2/M phase arrest (Fig. 9E, Fig. S8L). Cells with WDR4 knockdown also exhibited reduced expression of several cell cycle regulatory proteins, including Cyclin A, Cyclin B, Cyclin D1, and STAT3 (Fig. 9F). These observations reflected that WDR4 knockdown impaired protein expressions of cell cycle-related regulators, subsequently breast cancer cells did not proceed to the G2/M stage and finally stopped proliferating. Focusing on the mTORC1 signaling pathway, downstream substrates of mTORC1 include p70S6K and 4EBP1, which are frequently regarded as indicators of mTORC1 activity. WDR4 knockdown suppressed the mTORC1 activity by inhibiting phosphorylation of p70S6K and 4EBP1 (Fig. 9G). It was found that m7G methylation modification exists within mRNAs and regulates gene expressions [63]. We hypothesized whether WDR4 could also realize its function through this way. The m7G RIP experiments suggested that m7G modification were present in several key regulatory molecules upstream of the G2M checkpoint and mTORC1 signaling pathway (Fig. 9H). In conclusion, WDR4 facilitated breast cancer malignancy, providing a potential application for targeted therapy in breast cancer.

Fig. 9.

Regulation of the cell cycle and mTORC1 signaling by WDR4. (A) Immunofluorescence images for m7G levels in breast cancer cells with WDR4 inhibition or overexpression. (B) Dot blot assay for m7G levels of RNA in breast cancer cells with WDR4 knockdown or overexpression. The intensity of dot immunoblotting (above) represented the m7G levels while methylene blue staining (below) indicated the amount of loaded RNA. (C) GSEA result based on WDR4 expression showing enrichments of G2M checkpoint and mTORC1 signaling hallmarks in TCGA-BRCA. (D) The correlation plots of WDR4 expression and G2M checkpoint and mTORC1 signaling hallmarks in TCGA-BRCA. (E) The flow cytometry assay determining changes in the cell cycle induced by WDR4 knockdown. (F and G) Western blots detecting alterations of cell cycle regulatory proteins (F) and downstream substrates of mTORC1 (G) induced by WDR4 knockdown. (H) The m7G RIP-qPCR assay to investigate upstream signaling molecules of G2M checkpoint and mTORC1 pathway which were regulated by WDR4 through m7G modification in breast cancer cells. (For interpretation of the references to color in this figure legend, the reader is referred to the web version of this article.)

Discussion

As the most predominant neoplasm in women globally, breast cancer has been extensively characterized with overwhelming research and practice in terms of its development and treatment. However, some patients may not respond well to existing treatments, especially immunotherapy. RNA modifications are a widespread phenomenon, which are discovered to expand the functional diversities of RNA transcripts [64]. According to numerous studies, dysregulation of RNA modification-related genes has been implicated in many human cancers, including breast cancer [15], [65]. To our knowledge, our study is the first to integrate multiple RNA modifications for comprehensive analyses, which provides an important reference for the effects of RNA modifications in cancer, especially breast cancer. We constructed a reliable signature based on RNA modification-related genes in TCGA-BRCA, which was further validated in three other independent cohorts (METABRIC, GSE96058, and GSE20685). The gene signature was combined with clinical prognostic indicators to generate a prognostic nomogram model, which accurately predicted the survival probabilities for breast cancer patients. Finally, further investigation was performed to explore the critical effects of this gene signature on tumor immune infiltration and treatment response.

A nine-gene containing signature (RMscore) was identified to develop into a serviceable predictor of patient prognosis and therapy response. Aberrant expression of m7G regulators often influences target gene expression and ultimately manages cancer biological processes. Previous findings regarded eIF4E3 as a tissue-specific tumor suppressor [66], which agreed with better prognosis of highly expressed eIF4E3 in the current analysis. The RNA binding protein LARP1 is a post-transcriptional regulator that facilitates cell division and cell migration [67]. LARP1 was also reported to regulate mTOR post-transcriptionally and contribute to cancer progression [68]. In our survival analysis, LARP1 was also found to exert an oncogenic function in breast cancer. Emerging evidence indicates that the METTL1/WDR4 methyltransferase complex regulates various tumor processes via m7G methylation modification [69]. Our results showed that WDR4 knockdown strongly impaired both breast cancer proliferation and migration, which may be mediated through the cell cycle and mTORC1 signaling pathway. However, the relationship between two other molecules (EIF4G3 and NUDT3) and BC has not been described, which requires further study. The m6A reader YTHDF3 was reported to affect tumor metastasis in melanoma and breast cancer [70], [71], and the Kaplan–Meier survival analysis in this study confirmed the contribution of YTHDF3 to poor outcome of BC patients. Researches have been intensively conducted into METTL2A to develop targeted drug therapy for BC patients [72]. Known as TLS (Translocated in Liposarcoma), FUS functions as an RNA binding protein by binding to the regulatory regions of target genes [73]. Thus, the role of FUS in human cancer will depend on its target genes. Our result suggested that upregulated FUS expression in breast cancer might lead to improvement of patient prognosis. Summarizing from previous studies, TET2 plays a pivotal role in human tumors. TET2 deletion can decrease chemokine expression and tumor-infiltrating lymphocytes, resulting in tumor resistance to PD-L1 immunotherapy [74]. Also, breast cancer cells with repressed TET2 expression have a more favorable epithelial-mesenchymal transition (EMT) phenotype [75]. However, our analysis indicated the opposite conclusion, that breast cancer patients with high TET2 expressions have shorter survival time. This discrepancy should be further examined.

Over the past two decades, a great deal of research has been conducted on the immune mechanisms monitoring cancer development. Consisting of a huge variety of cell types, the tumor microenvironment tends to undergo immune-associated remodeling, which can further affect tumor evasion from immune surveillance and anti-tumor therapy efficacy [76]. From our results, we found that high RMscore might indicate a suppressive TME in breast cancer. After clustering the population that had received PD-L1 immunotherapy by RMscores, we also found that patients with high RMscores developed relatively resistance to immunotherapy. In search of additional evidence, analysis of single-cell RNA-seq data revealed that the RMscore was relatively higher in cancer epithelial cells and myeloid cells compared to other cell types. In cell-chat analyses, we found that HLA-CD4 and HLA-CD8 signaling were significantly enhanced in the high-RMscore group, which is in contrast to our previous correlation analysis where RMscore was significantly negatively correlated with HLA family expression. We hypothesized that this may be due to the presence of a suppressed immune microenvironment in the high scoring group, resulting in the presentation of tumor antigens while cellular immunity is not activated. And we also found that cancer epithelial cells diminished intensity of MDK and its ligands with CAFs. In previous studies, MDK exerted immunosuppressive effects mainly through interactions with receptors on macrophages and endothelial cells, and its role with CAF still requires more experiments in the future to explore whether it has an immunopromotive function [61], [77].

Previous studies have demonstrated that A-to-I editing protein ADAR1 is essential to maintain normal hematopoietic stem cells and self-renewal of leukemic stem cells. Similarly, aberrant m6A and m5C modifications have been reported to play a vital role in modulating the stemness of cancer stem cells. However, the specific role of RNA modification in regulating tumor stemness remains unknown. Thus, the positive correlation between tumor stemness and the RMscore needs to be proven more convincingly. Generally, tumor cells engage in signal communication with surrounding immune cells, and a decrease of intercellular communication often predicts tumor progression and drug resistance. This phenomenon was also observed in BC patients with high RMscores. Our additional analyses revealed that BC patients with high RMscores had higher IC50 values for Oxaliplatin, Palbociclib, and SCH772984 after undergoing drug therapy, which might explain why patients with high RMscore had a worse prognosis due to resistance to chemotherapeutic drugs. These findings indicated that the RMscore is also valuable in predicting therapy response in BC.

Although the gene signature in this study has excellent predictive strengths for prognosis and therapy response in BC patients, some limitations still exist. Firstly, retrospective recruitment of breast cancer patients leads to some bias. Secondly, except for WDR4, the biological functions of other RNA modification-related genes in BC have not yet been investigated. Thirdly, validation of the gene signature related to RNA modification using real-world data is indispensable.

Conclusions

Our study has provided an effective strategy for predicting prognosis, drug sensitivity, and immunotherapeutic response in BC patients. The m7G-related WDR4 was found to participate in BC development through the regulation of cell cycle and mTORC1 activity. Our findings may provide positive tidings for improving survival outcomes and therapy resistance by targeting RNA modifications in BC patients.

Availability of data and materials

All the datasets could be downloaded directly from the indicated websites in the methods section. Datasets and custom scripts are available upon request.

Compliance with Ethics requirements

This study was approved by Institutional Research Ethics Committee of Sun Yat-sen University Cancer Center (No. 2021-358) and conducted under the guidance of the Declaration of Helsinki.

Ethical approval

This study was approved by Institutional Research Ethics Committee of Sun Yat-sen University Cancer Center (No. 2021-358) and conducted under the guidance of the Declaration of Helsinki.

Funding statement

This research was funded by National Natural Science Foundation of China (Nos.82373369, Weidong Wei and Nos. 82203569, Hao Wu).

CRediT authorship contribution statement

Yongzhou Luo: Conceptualization, Visualization, Methodology, Writing – original draft. Wenwen Tian: Conceptualization, Visualization, Methodology, Writing – original draft. Da Kang: Conceptualization, Visualization, Methodology, Writing – original draft. Linyu Wu: Conceptualization, Visualization, Methodology, Writing – original draft. Hailin Tang: Data curation. Sifen Wang: Data curation. Chao Zhang: Data curation. Yi Xie: Formal analysis, Validation. Yue Zhang: Formal analysis, Validation. Jindong Xie: Formal analysis, Validation. Xinpei Deng: Formal analysis, Validation. Hao Zou: Formal analysis, Validation. Hao Wu: Conceptualization, Supervision. Huan Lin: Conceptualization, Supervision. Weidong Wei: Project administration, Funding acquisition.

Declaration of competing interest

The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.

Appendix A. Supplementary material

The following are the Supplementary data to this article: