Analysis of m7G-Related signatures in the tumour immune microenvironment and identification of clinical prognostic regulators in breast cancer

Gene expression profiling data of breast cancer patients were obtained from two independent patient cohorts, The Cancer Genome Atlas (TCGA) dataset TCGA-BRCA and the Gene Expression Omnibus (GEO) dataset GSE1456. The TCGA-BRCA cohort included 113 normal samples and 1113 breast cancer samples. We used the ‘limma’ package to normalise gene expression data and remove genes with a mean expression level of 0. Somatic data for breast cancer were also obtained from TCGA. Somatic data were preprocessed through Perl-related code. TCGA-BRCA data were annotated with gene names using GENCODE22 annotation files, and TCGA-BRCA patient survival data and clinical data were obtained from UCSC Xena, including survival time, survival status, age, tumour stage and sex. CNV data for breast cancer were also obtained from UCSC Xena. In our study, patient samples with clinical information and a survival time greater than 30 days were retained. Finally, 1023 patient samples were included from TCGA-BRCA and 159 from GSE1456; TCGA-BRCA was used for model construction, whereas GSE1456 was used for model validation.

m7GRGs (DCP2, IFIT5, EIF3D, EIF4G3, NSUN2, GEMIN5, AGO2, NUDT10, EIF4E, EIF4E2, NCBP2, NUDT11, NUDT3, NCBP1, METTL1, LARP1, NUDT4, EIF4E3, SNUPN, WDR4, LSM1, NUDT16, DCPS and CYFIP1) were obtained from the existing literature [22] and related gene sets: GOMF_m7G_5_PPPN_DIPHOSPHATASE_ACTIVITY, GOMF_RNA_CAP_BINDING and OMF_RNA_7_METHYLGUANOSINE_CAP_BINDING. We extracted the expression of 24 m7GRGs from the TCGA-BRCA queue and then used the ‘limma’ algorithm to analyse the difference between normal samples and breast cancer samples. m7GRGs with a p-value less than 0.05 were considered differentially expressed. Next, we explored the CNV incidence of the 24 m7GRGs and mapped their altered locations on 23 chromosomes using the ‘RCircos’ package. To elucidate the correlation of the 24 m7GRGs with breast cancer patients’ prognosis, we used the ‘igraph’, ‘psych’, ‘reshape2’ and ‘RColorBrewer’ packages to plot the prognosis-related network of the 24 m7GRGs.

Based on the expression profile data of these 24 genes, the K-means clustering algorithm of the ‘ConsensusClusterPlus’ package was used to perform consensus clustering on TCGA-BRCA patients to obtain breast cancer subtypes. The clustering was repeated 1000 times to ensure the accuracy and stability of the results. The optimal number (K-value) of breast cancer subtypes was calculated using the cumulative distribution function (CDF) and a consensus heatmap.

To explore the differences between different breast cancer subtypes, we used the ‘limma’ package to identify DEGs across the breast cancer subtypes, with thresholds set at p < 0.05 and |logFold Change|> 1. To identify the enriched pathways among the different subtypes, we conducted Gene Ontology (GO) and Kyoto Encyclopaedia of Genes and Genomes (KEGG) enrichment analyses [23, 24] for DEGs between subtypes and visualised the top 5 pathways with the most significant enrichment results. Using the gene set enrichment analysis (c2.cp.kegg.v7.5.1.symbols.gmt of the MSigDB reference gene set), the top 10 pathways were visualised. The above enrichment analysis was conducted using the ‘clusterProfiler’ package.

The ‘maftools’ package was used to draw waterfall plots related to TMB. Kaplan–Meier analysis was applied to evaluate the effect of TMB on the prognosis of breast cancer patients. Moreover, we used t-tests to determine differences in TMB levels between the high- and low-risk groups. Spearman's rank correlation was applied to calculate the correlation between TMB and risk score. Then, we estimated each patient's sensitivity to chemotherapeutic drugs using the Genomics of Cancer Drug Sensitivity (GDSC) database. The half-maximal inhibitory concentration (IC₅₀) values for high- and low-risk groups were quantified with the ‘pRRophetic’ package. We conducted the Wilcoxon test to calculate the differences in the IC₅₀ of the drugs among the breast cancer risk subgroups. The CellMiner database was used to mine sensitive drugs to build a predictive model. In addition, Pearson analysis was used to calculate drugs related to risk scores. The overall flowchart of this study is shown in Fig. 1.

MDA-MB-231, MDA-MB-468, SKBR3 and MCF-10A cells were purchased from Procell Life Science & Technology Co., Ltd. (Wuhan, China). MDA-MB-231 cells were incubated in DMEM culture medium (Gibco, 11,965,092) with 10% foetal bovine serum (FBS). MDA-MB-468 cells were incubated in RPMI 1640 culture medium (Gibco, 11,875,093) with 10% FBS. SKBR3 cells were cultured in special culture medium (Procell, CM-0211) containing McCoy's 5A, 10% FBS and 1% penicillin/streptomycin. MCF-10A was cultured in special culture medium (Procell, CM-0525) containing DMEM/F12, 5% Horse serum, 20 ng/mL EGF, 0.5 μg/mL Hydrocortisone, 10 μg/mL Insulin, 1% NEAA and 1% P/S. Total RNA was isolated with the TRIzol reagent (Invitrogen). Complementary DNA was synthesized from 1 μg total RNA using MightyScript Plus First Strand cDNA Synthesis Master Mix (Sangon Biotech). Quantitative real-time polymerase chain reaction (qRT-PCR) was performed with a PowerUp SYBR Green Master Mix (Thermo Fisher, A25742) after RNA extraction and reverse transcription from all four cell lines. Relative mRNA levels were calculated using the comparative Ct method (ΔCt). The primer sequences are listed in Supplementary Material S1.

First, we explored the expression of 24 m7GRGs in breast cancer tissues and normal tissues in the TCGA-BRCA cohort. Overall, 19 m7GRGs were differentially expressed between breast cancer tissues and normal tissues. In particular, the expressions of NSUN2, EIF4E, EIF4E2, NCBP2, NUDT3, NCBP1, LARP1, WDR4 and LSM1 were upregulated (Fig. 2A, Supplementary Material S2). Subsequently, we identified the CNV incidence of the 24 m7GRGs. In the TCGA-BRCA cohort, CNV alterations were found in all 24 m7GRGs. In addition, more than half of the m7GRGs had copy number amplification, while CYFIP1, NUDT4, DCPS, NCBP1, EIF3D, DCP2, EIF4E3, IFIT5 and EIF4G3 had CNV deletions (Fig. 2B). Figure 2C shows the locations of the CNV alterations of the 24 m7GRGs on chromosomes. Finally, to explore the association of the 24 m7GRGs with breast cancer prognosis, we mapped the network of interactions among the 24 m7GRGs and the effect of their expression on breast cancer prognosis. The results showed that 17 m7GRGs were risk factors and significantly impacted breast cancer prognosis, whereas NUDT10, NUDT3, EIF4E3, SNUPN, NUDT16, IFIT5 and EIF3D were favourable prognostic factors (Fig. 2D). In addition, based on the MethSurv website (https://​biit.​cs.​ut.​ee/​methsurv/​), we explored the changes in DNA methylation in the 24 m7GRGs (Supplementary Material S3).

Correlation analysis of the 24 m7GRGs in the TCGA-BRCA dataset showed a close correlation among these genes (Fig. 3A) (p < 0.05). Based on the expression of the 24 m7GRGs, we used consensus clustering analysis to classify 1023 breast cancer samples into four subtypes: C1, C2, C3 and C4 (Fig. 3B-D, Supplementary Material S4). Principal component analysis (PCA) revealed that the four subtypes were clearly separated (Fig. 3E), indicating there were significant differences in expression profiles among the different subtypes. At the same time, the expression levels of the 24 genes were significantly different among the four subtypes (Fig. 3F), showing that our typing results had good stability and accuracy.

To explore differences in immune infiltration levels among the four breast cancer subtypes, we performed immune infiltration analysis on the four subtypes. The CIBERSORT results showed that the infiltration degree of the 22 types of immune cells differed among the different samples (Fig. 4A), and a significant correlation was present between T cells and natural killer cells (Fig. 4B), indicating a synergistic effect between these cells in BRCA patients. Moreover, the expression of most immune cells differed significantly among different subtypes. Specifically, Plasma_cells, T_cells_CD8, T_cells_follicular_helper, T_cells_regulatory_(Tregs), NK_cells_resting, NK_cells_activated, Macrophages_M0, Macrophages_M1, Dendritic_cells_resting, Dendritic_cells_activated, Mast_cells_resting, Mast_cells_activated and Neutrophils exhibited significant differences in immune infiltration abundance among the four breast cancer subtypes (Fig. 4C). The ESTIMATE results showed that the stromal score significantly differed among the subtypes (p < 0.05) (Fig. 5A-D), indicating a significant difference in stromal cell composition among subtypes.

To identify the biological functions and pathways involved in the DEGs among the four breast cancer subtypes in breast cancer, we performed enrichment analysis of DEGs between the different breast cancer subtypes (C1 vs C2, C1 vs C3, C1 vs C4, C2 vs C3, C2 vs C4, C3 vs C4). GO enrichment analysis (Fig. 6A-F) revealed significant enrichment in ribosome biogenesis, RNA splicing and macrophage pathways. KEGG enrichment analysis results (Fig. 7A-F) showed that the DEGs were mainly enriched in Ribosome, TNF signalling pathway and Salmonella infection. The GSEA results (Fig. 8A-F) showed significant enrichment in KEGG_RIBOSOME, KEGG_VIBRIO_CHOLERAE_INFECTION, KEGG_PROTEASOME and other pathways. These results suggest that m7GRGs may be involved in numerous biological processes and pathways related to immunity.

To screen for m7GRGs associated with breast cancer patient survival time, we performed SVM analysis on the 24 m7GRGs. We obtained 21 m7GRGs associated with prognosis using the SVM (Fig. 9A and B) and subjected them to multivariate Cox regression analysis. Four genes (AGO2, EIF4E3, DCPS and EIF4E) were identified and used to construct a prediction model (Fig. 9C) that could provide the risk score of each sample in the TCGA-BRCA. Finally, the samples were divided into high- and low-risk groups according to the median risk score. The survival analysis showed a significant difference between the high- and low-risk groups (Fig. 9E). The difference in survival was also significant between high- and low-risk group samples in the validation set GSE1456 (Fig. 9G). The high-risk group was correlated with a lower survival rate (Fig. 9D, F). The expression levels of the four genes used to construct the predictive model were also significantly different between the high- and low-risk groups (Fig. 9H). Single-gene survival analysis was performed in the GSE1456 (Fig. 10A, C, E and G) and TCGA-BRCA (Fig. 10B, D, F and H) datasets based on the expression levels of the four genes. The results showed significant differences in survival between the high- and low-expression groups, further indicating that these four genes were significantly correlated with the prognosis of BRCA. To more deeply evaluate the clinical impact of AGO2, EIF4E3 and EIF4E in breast cancer, we used the Kaplan–Meier plotter database to plot the DMFS (distant metastasis-free survival)-related Kaplan–Meier curves of AGO2, EIF4E3 and EIF4E. The results showed that the DMFS rates of AGO2 and EIF4E3 were significantly different, which may be related to the metastasis of breast cancer (Supplementary Material S5).

To assess the prognostic value of the risk score, we performed a prognostic analysis of risk scores and other clinical characteristics in the TCGA-BRCA cohort. Univariate Cox regression analysis (Fig. 11A) revealed that age, tumour stage and risk score were significantly associated with prognosis (p < 0.05). Further multivariate Cox regression analysis confirmed (Fig. 11B) that age, tumour stage and risk score were independent predictors of prognosis (p < 0.05). Therefore, we constructed a nomogram (Fig. 11D) based on these three clinical characteristics to predict the 1-, 3- and 5-year survival rates of patients. Calibration and ROC curves were used to validate the accuracy of the nomogram in predicting survival time in breast cancer patients. The calibration curve results showed that the predicted values at 1, 3 and 5 years slightly deviated from the diagonal line (Fig. 11C), indicating that the nomogram can satisfactorily predict the prognosis of breast cancer patients compared to the ideal model. We conducted ROC curve analysis on the nomogram for 1 year, 3 years, and 5 years (Fig. 11E-G). The ROC curve showed that the AUC values at 1 year (AUC = 0.778, Fig. 11F) and 3 years (AUC = 0.678, Fig. 11G) were higher. Furthermore, we calculated the C-index of the column chart model, and the results further confirmed its high predictive accuracy (Supplementary Material S6). Overall, both methods demonstrated that the nomogram has better accuracy for predicting patient prognosis.

Because genetic mutations are an essential cause of the development of BRCA-related breast cancer, we explored differences in the distribution of somatic mutations between high- and low-risk populations. The 20 most frequently mutated genes for these two groups are shown in Fig. 12A and B, respectively. There were no significant differences in the mutation frequency of the top 20 genes between the high- and low-risk groups. The expression levels of risk scores were statistically different between the low- and high-TMB groups (Fig. 12C). The correlation between the mutational burden of BRCA and the population risk score was weak but statistically significant (Fig. 12D). The Kaplan–Meier curve of OS indicated that the OS of the patients in the high-TMB group was significantly lower than that of the patients in the low-TMB group (p = 0.03) (Fig. 12E). In addition, the OS of patients with high/low mutational burden in the high-risk group was statistically different from that in the low-risk group with high/low mutational burden (p = 0.03) (Fig. 12F). Therefore, high TMB may be an important factor leading to poor OS in breast cancer patients.

To evaluate the drug prediction ability of the 4-m7GRG prognostic risk model, we used the ‘pRRophetic’ package to compare the differences in the estimated IC₅₀ levels of six chemotherapeutic agents, namely, erlotinib (Fig. 13A), gemcitabine (Fig. 13B), cytarabine (Fig. 13C), gefitinib (Fig. 13D), the Akt1/2/3 inhibitor MK-2206 (Fig. 13E) and the PPM1D (WIP1) inhibitor CCT007093 (Fig. 13F). Our data showed that the high-risk score group was more sensitive to gemcitabine and cytarabine than the low-risk group. The above results indicated that the 4-m7GRG prognostic model has good drug sensitivity. Furthermore, to screen possible therapeutic drugs for breast cancer, based on the CellMiner database, we identified the drugs related to the risk score of the prognostic model. The results indicated that the low-risk score group was more sensitive to gefitinib and CCT007093. Drug sensitivity analysis of the four-gene signature showed that the DCPS, EIF4E, EIF4E3 and AGO2 genes were significantly associated with dasatinib, chelerythrine, E7820 and imexon, respectively (Fig. 14A–P).

Breast cancer cell lines (MDA-MB-231, MDA-MB-468 and SKBR3) and a normal breast cell line (MCF-10A) were used to validate the expression levels of the signature m7GRGs (AGO2, EIF4E3, DCPS and EIF4E). The results showed that the expression levels of AGO2 (Fig. 15A) and EIF4E3 (Fig. 15B) were significantly lower in breast cancer cell lines than in the normal breast cell line, whereas the expression levels of DCPS (Fig. 15C) and EIF4E (Fig. 15D) were significantly higher. The above results were consistent with the results of the bioinformatics analysis. Therefore, further exploration of the exact mechanism of these four m7Gs in breast cancer would be of great significance for improving the prognosis and treatment of breast cancer patients. In addition, we also explored the expression of AGO2, EIF4E3, DCPS and EIF4E in pan-cancer (Supplementary Material S7).