A combinatorially regulated RNA splicing signature predicts breast cancer EMT states and patient survival

Abstract

During breast cancer metastasis, the developmental process epithelial-mesenchymal transition (EMT) is abnormally activated. Transcriptional regulatory networks controlling EMT are well-studied; however, alternative RNA splicing also plays a critical regulatory role during this process. A comprehensive understanding of alternative splicing (AS) and the RNA binding proteins (RBPs) that regulate it during EMT and their impact on breast cancer remains largely unknown. In this study, we annotated AS in the breast cancer TCGA data set and identified an AS signature that is capable of distinguishing epithelial and mesenchymal states of the tumors. This AS signature contains 25 AS events, among which nine showed increased exon inclusion and 16 showed exon skipping during EMT. This AS signature accurately assigns the EMT status of cells in the CCLE data set and robustly predicts patient survival. We further developed an effective computational method using bipartite networks to identify RBP-AS networks during EMT. This network analysis revealed the complexity of RBP regulation and nominated previously unknown RBPs that regulate EMT-associated AS events. This study highlights the importance of global AS regulation during EMT in cancer progression and paves the way for further investigation into RNA regulation in EMT and metastasis.

Keywords: RNA-binding proteins (RBPs); alternative splicing (AS); breast cancer; epithelial–mesenchymal transition (EMT).

FIGURE 1.

Identification of EMT-associated AS events in human breast cancer. (A) Schematic of the analysis pipeline. (B) Heat map of the PSI values of the 25 significantly altered cassette exons in epithelial and mesenchymal samples. The PSI values were transformed into z-score (mean = 0 and standard deviation = 1) and plotted for each event. Red denotes higher inclusion of the cassette exon, while blue stands for higher skipping of that exon. Rows and columns are clustered based on the Pearson correlation coefficient. (C,D) Genome browser tracts of RNA-sequencing data showing alterations in AS between epithelial and mesenchymal samples. (C) Epithelial samples show higher inclusion of MAP3K7 exon 12 than mesenchymal samples. (D) Epithelial samples show lower inclusion of SPAG9 exon 24 than mesenchymal samples.

FIGURE 2.

Detection of the AS signature in the CCLE data set. (A) Heat map of the 25 EMT-associated AS events between epithelial (n = 7) and mesenchymal (n = 7) cell lines from the CCLE database. The heat map displays z-score transformed PSI values. The columns represent samples and the rows represent the 25 EMT-associated alternative splicing events. (B,C) Genome browser tracts of RNA-sequencing data showing alterations in alternative splicing between epithelial and mesenchymal cell lines. (B) Epithelial cell lines show higher inclusion of MAP3K7 exon 12 than mesenchymal cell lines. (C) Epithelial cell lines show lower inclusion of SPAG9 exon 24 than mesenchymal cell lines.

FIGURE 3.

Experimental and computational validation of the AS signature. (A–F) Semiquantitative PCR validation of six out of 25 AS events in epithelial or mesenchymal cell lines from the CCLE database. (G–I) Prediction of epithelial or mesenchymal cell status from the CCLE database based on the EMT AS signature through machine learning methods (SVM, DT, KNN, and NB). (G) Accuracy. (H) Sensitivity. (I) Specificity distribution is shown for all four methods. All machine learning methods using the AS signature exhibit strong predictive power with average accuracies of SVM, DT, KNN, and NB at 87.81%, 87.47%, 99.13%, and 99.66%, respectively.

FIGURE 4.

Community detection of RBP-AS bipartite networks. (A) Subgroup A correlated with exon inclusion. (B) RBPs and associated exon inclusion events that promote EMT. (C) RBPs and associated exon inclusion events that inhibit EMT and promote epithelial state. (D) Subgroup B correlated with exon skipping. (E,F) RBPs and associated exon skipping events that promote EMT. (G) RBPs and associated exon skipping events that inhibit EMT. Node sizes are proportional to the number of related events. (H,I) Semiquantitative PCR validation of MAP3K7 exon 12 alternative splicing affected by RBP knockdown in MCF7 cells or HS_578T cells.

FIGURE 5.

AS levels predict patient survival. (A–D) Kaplan-Meier survival plots of BRCA patients from TCGA stratified by the exon-inclusion level of four of the 25 EMT AS events. High PSI is 50% above the average PSI value while low PSI is 50% below the average PSI value. The “survdiff” function in R is used to compute P-values using the log-rank test. (E–G) Alternative splicing of the genes ATP5C1, KIF13A, CD44, and LRRFIP2 are used to predict the survival of five randomly selected patients from TCGA using the Cox proportional hazard model. The known patient survival times from TCGA are indicated in the inserted boxes as “t=” with the unit as days. The estimated survival probabilities over time by PSI values are plotted, and the color code of each patient survival curve corresponds to the same patient with indicated survival time from TCGA in the inserted box. Patients are randomly selected without specifying a breast cancer subtype (E), from the basal subtype (F), or from the luminal A subtype (G). Regardless of subtype classification, patients with the longer survival time in TCGA showed better predicted survival and vice versa.