Development and validation of novel prognostic models based on RNA-binding proteins in breast cancer

Abstract

Objectives

We aimed to construct novel prognostic models based on RNA-binding proteins (RBPs) in breast cancer (BRCA) and explore their roles in this disease and their effects on tumor-infiltrating immune cells (TIICs).

Methods

Datasets were downloaded from the Gene Expression Omnibus (GEO) database. Functions and prognostic values of RBPs were systematically investigated using a series of bioinformatics analysis methods. TIICs were assessed using CIBERSORT.

Results

Overall, 138 differentially expressed RBPs were identified, of which 86 were upregulated and 52 were downregulated. Of these, 13 RBPs were identified as prognosis-related and adopted to construct an overall survival (OS) model, while 12 RBPs were used for the relapse-free survival (RFS) model. High-risk patients had poorer OS and RFS rates than low-risk patients. The results indicate that the OS and RFS models are good prognostic models with reliable predictive abilities. In addition, the proportions of CD8, CD4 naïve, and CD4 memory resting T cells, as well as resting dendritic cells, were significantly different between the low-risk and high-risk groups in the OS model.

Conclusions

OS and RFS signatures can be used as reliable BRCA prognostic biomarkers. This work will help understand the prognostic roles and functions of RBPs in BRCA.

Introduction

Currently, there is no satisfactory treatment for patients with aggressive or relapsed breast cancer (BRCA).^1,2

Studies have demonstrated that abnormal expression of RNA-binding proteins (RBPs) is closely associated with human cancer progression.^3–6 Wang et al. ⁷ explored the prognostic value and function of RBPs in BRCA, but prognostic models were not developed and validated, nor were the effects on immune-infiltrating cells explored. Thus, in this study, we developed and validated novel prognostic models based on RBPs and explored the relationship between risk models and tumor-infiltrating immune cells (TIICs).

Materials and methods

Data processing

RNA sequencing and clinical datasets for BRCA were downloaded from the Gene Expression Omnibus (GEO) database (https://www.ncbi.nlm.nih.gov/geo/). The GSE42568 dataset included gene expression data from 17 normal breast tissues and 107 tumor tissues obtained using the Affymetrix Human Genome U133 Plus 2.0 Array platform. In addition, the names of 1542 RBPs were extracted from a previous study. ⁸ The differentially expressed genes (DEGs) between tumor tissues and normal breast tissues were identified using the limma package in R software. ⁹ DEGs were defined when P < 0.05 and log2|fold change| values were >1. The validation BRCA dataset, including the transcription profile based on the GPL15048 platform (Rosetta/Merck Human RSTA Custom Affymetrix 2.0) and clinical information, was obtained from the GSE86166 GEO dataset.

Construction and validation of prognostic models

Univariate Cox regression analysis was performed on all differentially expressed RBPs in the GSE42568 dataset using the survival R package. The significance of the candidate genes was verified. Subsequently, a least absolute shrinkage and selection operator (LASSO) regression model was constructed based on the above significant candidate genes, and the risk score (RS) was calculated to assess patient prognosis outcomes. The RS formula for each sample was as follows: RS = β1 × Expgene1 + β2 × Expgene2 +βi × Expgenei, where β represents the coefficient value and Exp represents the gene expression level. According to median RS survival analysis, BRCA patients were divided into low- and high-risk groups. The overall survival (OS) and relapse-free survival (RFS) of the two subgroups were compared using log-rank tests. The predictive capability of the prognostic model was evaluated using the survival R package. ¹⁰ The GSE86166 dataset was used as a validation cohort to confirm the predictive capability of the prognostic models. Finally, a nomogram was constructed using the R package (www.r-project.org) to forecast the likelihood of OS and RFS.

Gene ontology (GO) enrichment analysis and biology network

DEGs were classified using the clusterProfiler package ¹¹ in R software with hypergeometric distributions that were significant when P < 0.05. Functional enrichment analyses were performed for GO enrichment. A protein–protein interaction (PPI) network was established via Cytoscape3.7.2 and the STRING 11.5 database (https://string-db.org/). Candidate hub genes were identified using the degree topology method. ¹² Transcription factor (TF)–gene interaction and microRNA (miRNA)–gene interaction networks of prognosis-related RBPs were established via Cytoscape3.7.2 based on the data from NetworkAnalyst 3.0 (https://www.networkanalyst.ca/).

TIIC analysis based on CIBERSORT

The TIICs used in the GSE42568 dataset were assessed using the CIBERSORT analytical tool (https://cibersort.stanford.edu/). ¹³ The abundance ratio matrix for the 22 immune cell types was determined at P < 0.05.

Statistical analysis

R software was used for most bioinformatics and statistical analyses in this study, including RNA-seq data normalization and transformation, CIBERSORT, DEG analysis, survival analyses, receiver operating characteristic (ROC) curve analysis, and Spearman rank correlation analysis. Univariate and multivariate Cox regression analyses were performed using the “coxph” command of the survival package. Statistical significance was set at P < 0.05.

Results

Differentially expressed RBPs in BRCA

Differentially expressed RBPs in BRCA were identified using the limma package for R software. Among the 1542 RBPs (Table S1), 138 genes were differentially expressed, with 86 upregulated and 52 downregulated genes in the GSE42568 dataset (Figure 1, Table S2).

Figure 1.

Volcano plot of differentially expressed RNA-binding proteins (RBPs) in the GSE42568 dataset.

Development and validation of prognostic models

To analyze the effects of RBPs on the prognosis of BRCA patients, we assessed the relationship between the differentially expressed RBPs and OS using univariate Cox regression analysis and the Kaplan–Meier method in the training cohort (GSE42568 dataset). The results showed a significant association between 34 candidate RBPs and OS (Table S3). Similarly, 19 candidate RBPs were significantly associated with RFS (Table S4).

LASSO regression analysis was performed to select reliable RBPs for predicting OS in the GSE42568 dataset (Figure 2a and 2b). Subsequently, we selected the composition of the final gene signature and used it to generate the RS.

$$\begin{array}{l}\text{RS}=\left(0.\text{479}*\text{ExpDDX39A}\right)+\left(-0.0\text{25}*\text{ExpPAPD4}\right)\\ +\left(-0.\text{1}0\text{4}*\text{ExpSRSF6}\right)+\left(0.00\text{5}*\text{ExpNOP2}\right)\\ +\left(-0.\text{138}*\text{ExpZCCHC24}\right)\\ +\left(-0.0\text{32}*\text{ExpSIDT1}\right)+\left(-0.\text{118}*\text{ExpEIF4E3}\right)\\ +\left(-0.\text{239}*\text{ExpDIS3L}\right)\\ +\left(-0.\text{732}*\text{ExpSFPQ}\right)+\left(-0.\text{356}*\text{ExpSETD1B}\right)\\ +\left(0.0\text{12}*\text{ExpCTIF}\right)\\ +\left(0.\text{158}*\text{ExpUBAP2L}\right)+\left(0.\text{419}*\text{ExpCSTF3}\right)\end{array}$$

Figure 2.

Least absolute shrinkage and selection operator (LASSO) regression analysis based on RNA-binding proteins (RBPs) and overall survival (OS) prediction in the GSE42568 dataset. (a) Confidence interval of each lambda. (b) Changing trajectory of each independent variable. (c) Survival curves for the low- and high-risk subgroups. (d) Receiver operating characteristic (ROC) curves for forecasting OS based on risk score and (e–g) Expression heat map, risk score, and survival status.

According to the median RS, 104 patients from the GSE42568 dataset were assigned to the low- and high-risk groups. Survival analysis demonstrated that patients in the high-risk group had significantly poorer OS than those in the low-risk group (P < 0.001, Figure 2c). Time-dependent ROC analysis was performed to further evaluate the prognostic capability of the 13 identified RBPs. The area under the ROC curve (AUC) was 0.904, 0.85, and 0.828 at 1-, 3-, and 5-years, respectively (Figure 2d), indicating a moderate diagnostic performance for this model. The survival status of patients, RS, and expression heat map of the signature consisting of the 13 RBPs in the low- and high-risk subgroups are shown in Figure 2e–g.

To verify the validity of the OS predictive model, the GSE86166 dataset was analyzed. The results showed that patients with a high-risk score had significantly worse OS than those with a low-risk score (P < 0.001, Figure 3a). The AUCs of the GSE86166 dataset were 0.552, 0.683, and 0.673 at 1-, 3-, and 5-years, respectively (Figure 3b), which suggests good sensitivity and specificity of the predictive model. The survival status of patients, RS, and the expression heat map of the signature consisting of the 13 RBPs in the GSE86166 dataset are displayed in Figure 3c–e.

Figure 3.

Validation of the overall survival (OS) model using the GSE86166 dataset. (a) Survival curves for the low- and high-risk subgroups. (b) Receiver operating characteristic (ROC) curves for forecasting OS based on risk score and (c–e) Expression heat map, risk score, and survival status.

Using the same method, 12 gene signatures were selected, and a predictive RFS model was generated for the GSE86166 dataset (Figure 4a and 4b).

$$\begin{array}{l}\text{RS}=\left(0.\text{498}*\text{ExpDDX39A}\right)+\left(-0.0\text{34}*\text{ExpPAPD4}\right)\\ +\left(-0.0\text{79}*\text{ExpSRSF6}\right)+\left(0.0\text{83}*\text{ExpESRP1}\right)\\ +\left(-0.0\text{24}*\text{ExpSUGP2}\right)+\left(-0.0\text{33}*\text{ExpSIDT1}\right)\\ +\left(-0.0\text{59}*\text{ExpEIF4E3}\right)+\left(-0.\text{185}*\text{ExpDIS3L}\right)\\ +\left(-0.0\text{65}*\text{ExpNOVA1}\right)+\left(-0.\text{529}*\text{ExpSETD1B}\right)\\ +\left(0.0\text{81}*\text{ExpHINT3}\right)+\left(0.0\text{6}0*\text{ExpCSTF3}\right)\end{array}$$

Figure 4.

Least absolute shrinkage and selection operator (LASSO) regression analysis based on RNA-binding proteins (RBPs) and relapse-free survival (RFS) prediction from the GSE42568 dataset. (a) Confidence interval of each lambda. (b) Changing trajectory of each independent variable. (c) Survival curves for the low- and high-risk subgroups. (d) Receiver operating characteristic (ROC) curves for forecasting overall survival (OS) based on risk score. (e–g) Expression heat map, risk score, and survival status.

Survival analysis for the GSE42568 dataset demonstrated that patients in the high-risk group had significantly poorer RFS than those in the low-risk group (Figure 4c). The AUCs for the dataset were 0.817, 0.812, and 0.794 at 1-, 3-, and 5-years, respectively (Figure 4d). The relapse status of patients, RS, and expression heat map of the 12 RBPs in the GSE42568 dataset are displayed in Figure 4e–g.

We also verified the validity of the RFS predictive model by analyzing the GSE86166 dataset. Patients with high-risk scores had significantly worse RFS than those with low-risk scores (P < 0.001, Figure 5a). The AUCs of the GSE86166 dataset were 0.735, 0.714, and 0.689 at 1-, 3-, and 5-years, respectively (Figure 5b). The relapse status of patients, RS, and expression heat map of the 12 RBPs in the GSE86166 dataset are displayed in Figure 5c–e.

Figure 5.

Validation of the relapse-free survival (RFS) model using the GSE86166 dataset. (a) Survival curves for the low- and high-risk subgroups. (b) Receiver operating characteristic (ROC) curves for forecasting overall survival (OS) based on risk score and (c–e) Expression heat map, risk score, and survival status.

Additionally, the prognostic significance of different clinicopathological variables was assessed among patients in the GSE42568 dataset using Cox regression analysis. The results demonstrated that RS could independently predict both OS (Table 1) and RFS (Table 2). Additionally, there were similar results in the GSE86166 dataset (Table S5 and S6). Taken together, these data show that the RBP-based OS and RFS models are reliable in predicting the outcomes of BRCA patients.

Table 1.

Cox proportional-hazard regression analysis for overall survival in the GSE42568 dataset.

Characteristic	Univariate analysis			Multivariate analysis
	HR	HR (95% CI)	P-value	HR	HR (95% CI)	P-value
Age	0.998	0.969–1.028	0.889	1.009	0.975–1.046	0.581
T	1.783	0.942–3.374	0.076	1.083	0.529–2.217	0.828
N	4.555	1.877–11.054	0.001	3.112	1.152–8.411	0.025
Grade	3.038	1.534–6.018	0.001	1.503	0.711–3.179	0.286
ER	0.532	0.268–1.056	0.071	0.647	0.274–1.528	0.321
Risk score	1.77E+18	1.75E+13–1.80E+23	<0.001	4.83E+14	1.74E+09–1.34E+20	<0.001

HR, hazard ratio; CI, confidence interval; ER, estrogen receptor.

Table 2.

Cox proportional-hazards regression analysis for relapse free survival in GSE42568.

Characteristic	Univariate analysis			Multivariate analysis
	HR	HR (95% CI)	P-value	HR	HR (95% CI)	P-value
Age	0.996	0.970–1.022	0.747	1.016	0.986–1.047	0.303
T	2.467	1.306–4.659	0.005	1.543	0.854–2.789	0.151
N	4.549	2.184–9.479	<0.001	3.689	1.656–8.218	0.001
Grade	2.365	1.375–4.069	0.002	1.109	0.607–2.029	0.736
ER	0.438	0.243–0.792	0.006	0.504	0.246–1.034	0.062
Risk score	17.425	7.289–41.655	<0.001	9.241	3.205–26.643	<0.001

HR, hazard ratio; CI, confidence interval; ER, estrogen receptor.

Building predictive nomograms

Nomograms for OS (Figure 6a) and RFS (Figure 6c) were constructed to generate clinically practical models that would enable physicians to evaluate the prognosis of BRCA patients using the selected RBPs. According to the results of the LASSO regression analysis, each individual variable was assigned a corresponding point based on the scale obtained in the nomograms. The point for each variable was determined by drawing a horizontal line, and then the points of all the variables were added to obtain the patient’s total score. This was used to estimate the survival rate of each patient at 1, 3, and 5 years. Furthermore, calibration plots (Figure 6b and 6d) were generated to evaluate the validity and accuracy of the nomograms.

Figure 6.

Nomograms for predicting 1-, 3-, and 5-year overall survival (OS) and relapse-free survival (RFS) of breast cancer patients in the GSE42568 dataset. (a) Nomogram for predicting OS. (b) Calibration plots of the nomogram for predicting OS at 3- and 5-years in the dataset. (c) Nomogram for predicting RFS and (d) Calibration plots of the nomogram for predicting RFS at 3- and 5-years in the training cohort.

GO enrichment analysis and biology network

To explore the functions of the differentially expressed RBPs and mechanisms through which they can promote BRCA progression, we performed functional analyses of the downregulated and upregulated RBPs via the clusterProfiler package in R software. As shown in Figure 7a and 7b, significant differences were observed in the functional enrichment of downregulated and upregulated RBPs. Upregulated RBPs were enriched in the spliceosomal complex, nuclear speckles, catalytic step 2 spliceosome, and methyltransferase complex, whereas downregulated RBPs were enriched in ribosomal subunits, organellar ribosomes, mitochondrial ribosomes, mRNA cap binding complexes, and RNA polymerase I complexes. Biological process analysis showed that upregulated RBPs were related to RNA splicing and non-coding RNA processing, while downregulated RBPs were related to RNA splicing, translation, cellular amide metabolic processes, and RNA phosphodiester bond hydrolysis. Molecular functional analysis demonstrated that upregulated RBPs participated in catalytic, methyltransferase, and transferase activities, as well as single-stranded RNA binding, whereas downregulated RBPs participated in catalytic, endonuclease, nuclease, and ribonuclease activities.

Figure 7.

Gene ontology (GO) enrichment analysis and biology network. (a) GO enrichment of downregulated RNA-binding proteins (RBPs). (b) GO enrichment of upregulated RBPs. (c) Protein–protein interaction (PPI) network and hub RBPs and (d) Transcription factor (TF)–microRNA coregulatory network.

To investigate the functions of the differentially expressed RBPs in BRCA, we created PPI and TF-miRNA coregulatory networks and identified the top 20 hub RBPs (Figure 7c). As shown in Figure 7d, the TF-miRNA co-regulatory network suggested that the prognosis-related RBPs are regulated by many TFs and miRNAs.

TIICs based on CIBERSORT

To explore the impact of the risk score on TIICs, the infiltration proportions of 22 immune cell types in the GSE42568 dataset were calculated using the CIBERSOT algorithm. For the BRCA samples, follicular helper M0, M1, and M2 macrophages were the major constituents of TIICs (Figure S1). Compared with normal breast tissues, the infiltration proportions of follicular helper T cells (P < 0.001), M0 macrophages (P < 0.001), and M1 macrophages (P < 0.001) were significantly increased in BRCA tissues, while the infiltration proportions of activated natural killer (NK) cells (P = 0.001), monocytes (P < 0.001), M2 macrophages (P < 0.001), and resting mast cells (P = 0.004) were reduced (Figure 8a). We also analyzed the relationship between TIICs and OS or RFS. Patients with high proportions of follicular helper T cells (P = 0.001, Figure 8b) or activated mast cells (P = 0.015, Figure 8b) had poorer OS than those with low infiltration proportions, whereas patients with high proportions of resting mast cells had longer OS (P = 0.017, Figure 8b). In addition, patients with high infiltration proportions of follicular helper T cells had poorer RFS than those with low infiltration proportions (P = 0.034, Figure 8b).

Figure 8.

Differences in immune infiltration characteristics between risk groups in the GSE42568 dataset. (a) Quantity of tumor-infiltrating immune cells (TIICs) in normal and breast cancer tissues. (b) Prognostic value of different TIICs and (c) Difference in TIICs between the low- and high-risk groups determined using the CIBERSORT R package.

Compared with the low-risk group designated by the OS model (Figure 8c), the proportions of CD4 naïve T cells (P = 0.016), resting CD4 memory T cells (P = 0.029), resting dendritic cells (P = 0.003), and resting mast cells (P = 0.048) were significantly reduced in the high-risk group, while the infiltration proportions of CD8 T cells (P = 0.028), monocytes (P = 0.031), and activated mast cells (P = 0.028) were increased. In addition, the prognosis-related RBPs had different associations with various TIICs (Figure S2). These results suggest that the prognosis-related RBPs interact with some TIICs and influence the progression and prognosis of BRCA.

Discussion

In the present study, we identified 138 differentially expressed RBPs. OS- and RFS-risk models were established based on prognosis-related RBPs that were identified and validated using the GSE42568 and GSE86166 datasets, respectively. Both models proved reliable in predicting OS and RFS. To improve the clinical practicality, we constructed nomograms to evaluate the survival or relapse of patients at 1-, 3-, and 5-years. Additionally, we analyzed the relevant biological functions and networks of the RBPs, and assessed the effects of risk score on TIICs.

To explore the prognostic significance of the differentially expressed RBPs, we performed Cox survival analysis on the GSE42568 dataset, which showed that 13 RBPs (DDX39A, PAPD4, SRSF6, NOP2, ZCCHC24, SIDT1, EIF4E3, DIS3L, SFPQ, SETD1B, CTIF, UBAP2L, and CSTF3) were OS-related, and 12 RBPs (DDX39A, PAPD4, SIDT1, SETD1B, DIS3L, SRSF6, ESRP1, EIF4E3, CSTF3, SUGP2, NOVA1, and HINT3) were RFS-related. Subsequently, we established OS- and RFS-risk models based on the identified RBPs to predict the prognosis of BRCA patients in the GSE42568 dataset. ROC analysis demonstrated that the RBP-based risk models were reliable in selecting BRCA patients with poor OS and RFS, which was validated by the GSE86166 dataset. Additionally, RS was shown to be an independent prognostic factor in BRCA patients. Furthermore, we constructed a nomogram to enable physicians to more accurately predict survival or relapse at 1-, 3-, and 5-years.

Functional enrichment analysis demonstrated the significant enrichment of genes involved in the post-transcriptional regulation of RNA, including RNA splicing, regulation of translation, and non-coding RNA processing. Post-transcriptional regulation of RNA plays an important role in many types of cancers.^14,15 For example, serine and arginine rich splicing factor 1 (SRSF1) has been shown to play an important role in BRCA by regulating alternative splicing, ¹⁶ and heterogeneous nuclear ribonucleoprotein A1 (HNRNPA1) affected the progression of triple-negative breast cancer by regulating hypoxia up-regulated 1 (HYOU1) mRNA expression. ¹⁷ In addition, N-acetyltransferase 10 (NAT10) regulates cell cycle checkpoint control and resistance to DNA-damaging chemotherapy and radiotherapy by affecting MORC family CW-type zinc finger 2 (MORC2) acetylation in BRCA. ¹⁸ Furthermore, interleukin enhancer binding factor 3 (ILF3) promoted breast tumorigenicity by regulating sustained urokinase-type plasminogen activator (uPA) expression, ¹⁹ and DEAD-box helicase 17 (DDX17) enhanced the tumorigenic and stem-like features of SRY-box 2 (SOX2) by promoting its binding to its target genes in ER-positive BRCA. ²⁰ Consistent with the results of the functional enrichment analysis, the hub RBPs mainly affected the post-transcriptional regulation of RNA. However, the roles and mechanisms of other hub RBPs, such as SRSF6, PRPF19, POLR2E, SNRPB, and SF1, in BRCA remain unclear and require further study. The TF-miRNA co-regulatory network revealed that these prognosis-related RBPs might play important roles in BRCA. Indeed, studies have shown that CSTF3, ²¹ SIDT1, ⁷ ESRP1, ²² DDX39A, ²³ NOP2, ²⁴ NOVA1, ²⁵ SFPQ, ²⁶ and UBAP2L ²⁷ participate in BRCA development and progression. Although the mechanisms still need to be defined, these studies support our prognostic models. In addition, the roles of other prognosis-related RBPs (CTIF, EIF4E3, SETD1B, ZCCHC24, and SUGP2) in BRCA warrant further exploration.

Finally, we analyzed TIICs in BRCA. The infiltration proportions of follicular helper T cells ranked among the top four TIICs and correlated with OS and RFS. This finding suggests that follicular helper T cells play an important role in BRCA. Furthermore, the proportion of follicular helper T cells was significantly higher in the BRCA tissues than in the normal tissues. A high proportion of follicular helper T cells in BRCA patients was associated with worse OS and RFS when compared with those with a low proportion of follicular helper T cells. This suggests that follicular helper T cells are associated with a poor prognosis and tumor progression in BRCA. However, follicular helper T cells are direct mediators of the immune checkpoint inhibitor response and are associated with B cell activation and antitumor responses in high mutation burden mouse models of BRCA, ²⁸ which have not yet been studied in humans. Nevertheless, these studies suggest that follicular helper T cells play an important role in BRCA.

This study has some limitations. First, the sample size of the GSE42568 dataset is small. Second, detailed clinicopathological parameters were lacking in the GSE42568 and GSE86166 datasets. Third, this study was based on public data and needs more basic experiments to further verify these concepts.

Collectively, we investigated the prognostic value and potential function of differentially expressed RBPs in BRCA. Risk models that reliably predict the prognosis of BRCA patients were constructed based on identified RBPs and validated using an alternative dataset. In addition, the results showed that follicular helper T cells play an important role in BRCA and are positively correlated with some RBPs. Our study provides important evidence for future studies on the role of RBPs in BRCA.

Data availability statement

The data used to support the findings of this study were obtained from the GEO datasets GSE42568 and GSE86166.

Supplemental material

Supplemental material for this article is available online.