Prognostic microRNA/mRNA signature from the integrated analysis of patients with invasive breast cancer

Abstract

The optimal management of breast cancer (BC) presents challenges due to the heterogeneous molecular classification of the disease. We performed survival analysis on a cohort of 466 patients with primary invasive ductal carcinoma (IDC), the most frequent type of BC, by integrating mRNA, microRNA (miRNA), and DNA methylation next-generation sequencing data from The Cancer Genome Atlas (TCGA). Expression data from eight other BC cohorts were used for validation. The prognostic value of the resulting miRNA/mRNA signature was compared with that of other prognostic BC signatures. Thirty mRNAs and seven miRNAs were associated with overall survival across different clinical and molecular subclasses of a 466-patient IDC cohort from TCGA. The prognostic RNAs included PIK3CA, one of the two most frequently mutated genes in IDC, and miRNAs such as hsa-miR-328, hsa-miR-484, and hsa-miR-874. The area under the curve of the receiver-operator characteristic for the IDC risk predictor in the TCGA cohort was 0.74 at 60 mo of overall survival (P < 0.001). Most relevant for clinical application, the integrated signature had the highest prognostic value in early stage I and II tumors (receiver-operator characteristic area under the curve = 0.77, P value < 0.001). The genes in the RNA risk predictor had an independent prognostic value compared with the clinical covariates, as shown by multivariate analysis. The integrated RNA signature was successfully validated on eight BC cohorts, comprising a total of 2,399 patients, and it had superior performance for risk stratification with respect to other RNA predictors, including the mRNAs used in MammaPrint and Oncotype DX assays.

Fig. 1.

Strategy used to derive and validate prognostic mRNAs and miRNAs in breast cancer. mRNAs and miRNAs were integrated in a single 7982-RNA profile (TCGA IDC cohort, n = 466). Survival analysis was performed within the various subgroups of the following clinical and molecular classes: disease stage, lymph node involvement (N stage), surgical margin, pre- or postmenopause, intrinsic subtype, somatic mutations (TP53, PIK3CA pathway, TP53/PIK3CA double mutants, GATA3, and the remaining less frequently altered genes). The subclasses within a class represented disjoint patient sets, thus enabling immediate validation of the prognostic RNAs within that class. The HRs and Kaplan–Meier curve were generated for every RNA in all independent subclass. RNAs that had significant both HRs and log-rank tests (P < 0.05) in at least two subclasses (within the same clinical or molecular class) were initially selected. Additional criteria, required for the selection of coding genes, were the association of DNA methylation with OS and the presence of somatic mutations in the COSMIC database (www.sanger.ac.uk/genetics/CGP/cosmic/). The association between DNA methylation and OS was carried out on the whole cohort (not on each subclass) using univariate Cox regression (SI Appendix, Tables S2 and S3). The HR was the ratio of hazards for a twofold change in the DNA methylation level. A majority-rule voting procedure was applied to all significant HRs for the CpG sites in the prognostic genes (false discovery rate < 0.001); e.g., the DNA methylation of a gene with the most significant CpG HRs lower than 1 would be defined as negatively correlated to outcome or vice versa. A further step for gene reassessment was then performed in the BC tumor subtype, as detailed in Results. Eight independent validation cohorts (total n = 2,399) were used to evaluate the prognostic miRNA/mRNA signature generated in the TCGA IDC cohort.

Fig. 2.

mRNAs and miRNAs associated with prognosis in different clinical and molecular subclasses of invasive ductal carcinoma (TCGA cohort). The matrix visualizes the significant HRs for the 30 mRNAs and seven miRNAs in the TCGA IDC cohort (listed in SI Appendix, Table S5). The HRs for expression with significant univariate Cox regression (P < 0.05) are displayed on a log₂ scale. Red squares indicate HRs >1 and blue squares indicate HRs <1.

Fig. 3.

Kaplan–Meier and ROC curves for the integrated miRNA/mRNA signature (TCGA IDC cohort). (A) The cross-validated Kaplan–Meier curves for IDC risk groups obtained from the TCGA cohort (n = 466), using the integrated signature (“RNA model”). The permutation P value of the log-rank test between risk groups (P < 0.001) was based on 1,000 permutations. (B) The ROC curve had an AUC of 0.74 (P < 0.001). The permutation P value was computed for testing the null hypothesis (AUC = 0.5) using 1,000 permutations.

Fig. 4.

Kaplan–Meier and ROC curves for the integrated miRNA/mRNA signature in the UK validation cohort. (A) The cross-validated Kaplan–Meier curves for breast cancer risk groups obtained from the validation cohort (n = 207), using the prognostic integrated signature. The permutation P value of the log-rank test between risk groups (P = 0.007) was based on 1,000 permutations. (B) The ROC curve had an AUC of 0.65 (P = 0.004). The permutation P value was computed for testing the null hypothesis (AUC = 0.5) using 1,000 permutations.