Development of a predictive miRNA signature for breast cancer risk among high-risk women

Abstract

Significant limitations exist in our ability to predict breast cancer risk at the individual level. Circulating microRNAs (C-miRNAs) have emerged as measurable biomarkers (liquid biopsies) for cancer detection. We evaluated the ability of C-miRNAs to identify women most likely to develop breast cancer by profiling miRNA from serum obtained long before diagnosis. 24 breast cancer cases and controls (matched for risk and age) were identified from women enrolled in the High-Risk Breast Program at the UVM Cancer Center. Isolated RNA from serum was profiled for over 2500 human miRNAs. The miRNA expression data were input into a stepwise linear regression model to discover a multivariable miRNA signature that predicts long-term risk of breast cancer. 25 candidate miRNAs were identified that individually classified cases and controls based on statistical methodologies. A refined 6-miRNA risk-signature was discovered following regression modeling that distinguishes cases and controls (AUC0.896, CI 0.804-0.988) in this cohort. A functional relationship between miRNAs that cluster together when cases are contrasted against controls was suggested and confirmed by pathway analyses. The discovered 6 miRNA risk-signature can discriminate high-risk women who ultimately develop breast cancer from those who remain cancer-free, improving current risk assessment models. Future studies will focus on functional analysis of the miRNAs in this signature and testing in larger cohorts. We propose that the combined signature is highly significant for predicting cancer risk, and worthy of further screening in larger, independent clinical cohorts.

Keywords: benign breast disease; high risk breast cancer; liquid biopsy; microRNA; risk signature.

Figure 1. Schematic of study design and serum collection timeframe

Forty-eight patients were selected from a database of 605 women at increased risk for developing breast cancer based on clinical factors: 24 women developed cancer at least six months after blood was drawn (cases) and 24 age and risk matched women who remain free of breast cancer (controls).

Figure 2. The expression of 25 candidate C-miRNAs separate women who develop cancer from those who remain cancer free

Expression of all 2578 mature miRNAs within miRBase v20 were assayed in circulation. A: A standardized workflow was developed and optimized to reduce sample-to-sample variability that includes precise miRNA isolation, profiling, and normalization (see methods). We identified 2 sets of candidate miRNAs totaling 25 in combination. AUC: 20 miRNAs with the highest individual AUC (range 0.632-0.766). ANOVA: 19 miRNAs significantly different between cases and controls with an ANOVA p < 0.05. B, C: Principal component analysis was performed using normalized expression levels of those miRNAs within each set; top 20 AUC miRNAs or 19 miRNAs with and ANOVA p < 0.05 between cases and controls. In both instances, women with an eventual breast cancer diagnosis cluster together in the top left quadrant of the PCA graphs. Black circles represent cancer-free controls. Red circles represent eventual breast cancer cases. D, E: Heat maps show log₂ expression of candidate miRNAs in cancer-free controls and future breast cancer cases. MicroRNAs are ordered based on hierarchical clustering (UPGMA, Euclidian distance). Each column within the stacked heat maps are matched case/control. Blue is background (0) and red is high miRNA log₂ expression.

Figure 3. Identification of miRNAs predictive of future breast cancer diagnosis

In order to develop a miRNA risk signature and score we generated a multivariable proportional hazards model based on candidate miRNA expression. Both AUC and ANOVA sets were examined. A: Schematic of iterative model generation using randomly selected training sets of 32 samples and corresponding validation sets of 16 samples. B: A total of 9 combined miRNAs were identified as signature miRNA (AUC: 6 of 20 miRNAs and ANOVA: 6 of 19 miRNAs) being present in greater than 50% of the models and did not change when the validation set AUC was required to be greater than 0.8. Cox proportional hazards (CoxPH) models were built using the expression of each 6-miRNA set across all 48 samples. C, D: The predictive ability of each 6-miRNA set in B was assessed by ROC curve and AUC based on calculated risk score. Hsa-miR-7855-5p was computationally excluded from the final ANOVA miRNA set model. 95% confidence intervals (CI) are indicated by gray area around each curve. Sensitivity is the true positive rate and 1-specificity is the false positive rate. E-H: Graphical representation of risk scores for each of the 48 samples based on a CoxPH model generated from the 6-miRNA AUC set (E, G) or 5-miRNA ANOVA set (F, H). Dotted line is the threshold used to distinguish cases from controls. Squares represent eventual breast cancer cases and circles represent cancer-free controls (G, H). p < 0.0001 between cases and controls.

Figure 4. Biological implications of risk signature miRNAs

A-I: Box and whisker plots of normalized log₂ microarray expression for 9 candidate miRNAs identified through an iterative multivariable CoxPH model. Each open circle represents an individual patient. A-F: 6 AUC set miRNAs ordered from high to low individual AUC. B, C, G-I: 5 ANOVA set miRNAs ordered from low to high p-value. G-I: The 3 miRNAs unique to the ANOVA set are only detected in a small percentage of samples and were not included in downstream pathway analysis. J, K: Targets of the 6 AUC set miRNAs were identified in Ingenuity Pathway Analysis (IPA, www.ingenuity.com) and core analyses run. No genes were present in the interaction network for hsa-miR-7855-5p. The top 5 enriched IPA canonical pathways (J) and diseases and biological functions (K) were identified through comparison analysis of these miRNA interaction networks. Colors in the heat maps are based on the -log(p-value) of enriched pathway over genomic background where a value > 1.3 corresponds to p < 0.05. n = 24 controls and n = 24 cases. * p < 0.05 ** p < 0.01 for both paired (matched case/control) and unpaired t-tests.