Large miRNA survival analysis reveals a prognostic four-biomarker signature for triple negative breast cancer

Abstract

Triple negative breast cancer (TNBC) is currently the only major breast tumor subtype without effective targeted therapy and, as a consequence, usually presents a poor outcome. Due to its more aggressive phenotype, there is an urgent clinical need to identify novel biomarkers that discriminate individuals with poor prognosis. We hypothesize that miRNAs can be used to this end because they are involved in the initiation and progression of tumors by altering the expression of their target genes. To identify a prognostic biomarker in TNBC, we analyzed the miRNA expression of a cohort composed of 185 patients diagnosed with TNBC using penalized Cox regression models. We identified a four-biomarker signature based on miR-221, miR-1305, miR-4708, and RMDN2 expression levels that allowed for the subdivision of TNBC into high- or low-risk groups (Hazard Ratio - HR = 0.32; 95% Confidence Interval - CI = 0.11-0.91; p = 0.03) and are also statistically associated with survival outcome in subgroups of postmenopausal status (HR = 0.19; 95% CI = 0.04-0.90; p= 0.016), node negative status (HR = 0.12; 95% CI = 0.01-1.04; p = 0.026), and tumors larger than 2cm (HR = 0.21; 95% CI = 0.05-0.81; p = 0.021). This four-biomarker signature was significantly associated with TNBC as an independent prognostic factor for survival.

Figure 1. Scheme for the data analysis. We selected 185 individuals with an absent expression of ER, PR, and HER2 from a breast cancer data set composed of 1,290 individuals. This subset of 185 TNBCs was split into training (120 individuals) and validation (65 individuals) sets. To select biomarkers associated with overall TNBC survival, we used the Lasso regularized Cox regression model on the training set. Four genes were selected by the method. Then, to better estimate the parameters of the model, we used the standard Cox regression model. Finally, we confirmed the results obtained in the training set by applying the four-biomarkers in the validation set.

Figure 2. Kaplan-Meier analysis showing that the four-biomarker signature is associated with survival in triple negative breast cancer. (A) Kaplan-Meier curve of the four-biomarker signature in the training set. (B) Kaplan-Meier curve of the four-biomarker signature in the validation set. CI = confidence interval; HR = hazard ratio. HR and 95% CI were estimated by multivariate Cox regression with age at diagnosis, grade, the presence of nodes, and menopausal status included as covariates. The p-value was obtained by the log-rank test of the Kaplan-Meier curve.

Figure 3. Expression of miR-221, miR-1305, miR-4708, and RMDN2 between good and poor prognosis in the (A) training and (B) validation sets. Horizontal bars represent the median. All genes are statistically overexpressed in patients with a good prognosis at a p-value threshold of 0.05.

Figure 4. Kaplan-Meier analysis of overall survival in subgroups of triple-negative breast cancer patients in the validation set. (A) Premenopausal patients; (B) postmenopausal patients; (C) patients without nodes identified; (D) patients with nodes; (E) patients with a tumor size of 2 cm or less; (F) patients with tumor size greater than 2 cm. Hazard ratios and confidence intervals were obtained by multivariate Cox regression with age at diagnosis, tumor grade, tumor size, and presence of nodes status as covariates. P-values were obtained by log-rank tests for the Kaplan-Meier curves. HR = hazard ratio; CI = confidence interval.

Figure 5. Kaplan-Meier analysis showing that the five target genes are associated with survival in triple negative breast cancer. The p-value is obtained by the log-rank test of the Kaplan-Meier curve.

Figure 6. Gene expression heatmaps. (A) For the four prognostic biomarkers and (B) the five genes targets of the three miRNAs. In panel (A), the four biomarkers are not correlated, while in panel (B) there are two sets of genes: one composed of genes SMAD1, SMAD4, and p27, and the second set composed of RUNX2 and c-kit. Genes within the sets are positively highly correlated, while between sets, they are negatively highly correlated.

Figure 7. Time lengths for the last follow-ups. There is no statistical difference in terms of time to death between individuals that died in the EGpA (average of 39.44 months) and TCGA (average of 45.69 months) data sets (p=0.973, Wilcoxon rank test). However, when we analyzed the time length for the last follow-up of people who did not die, we noticed that EGpA has a much longer follow-up (average of 118.55 months) than TCGA (average of 21.12 months) (p-value < 0.001, Wilcoxon rank test).