Circulating microRNAs in the early prediction of disease recurrence in primary breast cancer

A total of 209 consecutive patients with early breast cancer who underwent surgery followed by adjuvant chemotherapy administered at the Department of Medical Oncology of the University Hospital of Heraklion (Crete, Greece) between years 2003 and 2010 and had available plasma samples, were included in the present study. Plasma samples were obtained after the surgical resection of the primary tumor and before the initiation of adjuvant chemotherapy. Plasma samples were also collected from 23 normal blood donors to serve as controls. All patients and normal donors had provided signed informed consent to participate in the study, which was approved by the Ethics and Scientific Committee of Department of Medical Oncology of the University Hospital of Heraklion (ID 13998/8–10-2104; Crete, Greece). Clinical characteristics and follow-up information for each patient were prospectively collected. Peripheral blood from healthy donors and patients was drawn early in the morning and was collected in EDTA tubes. Plasma was subsequently isolated within 2 h by centrifugation at 2500 rpm for 15 min at 4 °C, followed by a second centrifugation at 2000 g for 15 min at 4 °C to remove cellular debris. Samples were kept in aliquots at ˗ 80 °C until further use. Plasma samples presenting a change in color to pink, suggesting the presence of hemolysis, were not processed for further analysis (Fig. 1).

Reverse transcription and RT-qPCR was performed according to manufacturer’s instructions and as previously described [13]. Total RNA input of 1.67 μl was reverse transcribed using the TaqMan miRNA Reverse Transcription kit and miRNA specific stem-loop primers (Applied Biosystmes, Foster City, CA, USA) in a 5-μl reaction comprising 1 mM dNTPs, 1 × PCR Reverse Transcription Buffer, 0.787 μl H₂O, 3.3 units Multiscribe Reverse Transcriptase, 0.252 units RNase inhibitors and 0.2 × RT-specific stem-loop primers. The reaction was performed in a Peltier Thermal Cycler PTC-200 at 16 °C for 30 min, 42 °C for 30 min and 85 °C for 5 min. Complementary DNA (cDNA) was diluted at 30 μl and each miRNA was assessed by RT-qPCR in a 5-μl reaction comprising 1 × of TaqMan 2× Universal PCR Mater Mix, No AmpErase UNG, 0.25 μl of TaqMan miRNA Assay and 2.25 μl of diluted cDNA. The quantitative real- time PCR reaction was carried out at 95 °C for 10 min, followed by 40 cycles of 95 °C for 15 min and 60 °C for 1 min on a ViiA 7 Real-Time PCR System (Applied Biosystems, Foster City, CA, USA). All the assays were performed in triplicates. Appropriate negative controls were used in both cDNA synthesis and RT-qPCR reactions where RNA input was replaced by H₂O and no template control was used, respectively. The average expression levels for each miRNA were calculated by the 2^-ΔCt method relative to the average of miR-23a.

Due to the lack of consensus concerning the normalization of circulating miRNAs we used miR-23a as a reference gene that was stably and reproducibly expressed among patients’ groups (Mann-Whitney test, p = 0.458) and among patients and normal donors (Mann-Whitney test, p = 0.12) [22]. Finally, the fold change in target miRNAs relative to miRNAs expressed in normal controls was calculated by the 2^-ΔΔCt method [23]. Samples with mean cycle threshold (Ct) > 35 for target miRNAs (n = 17) were excluded from the analysis. In addition, samples with mean Ct > 22 or Ct < 20 of cel-miR-39 (n = 5), suggesting RNA extraction was not efficient, were also excluded. Moreover, plasma samples were tested for contamination with red blood cells by measuring miR-451 and miR-23a expression levels [24]. Samples contaminated with red blood cells were not processed for further analysis.

The statistical analysis was performed using the SPSS software package, version 22.0 (SPSS Inc. Chicago IL, USA). Cutoff points were set at the median value for expression of each miRNA. Patients with miRNA expression above or equal to the median values were characterized as having high expression, whereas patients with miRNA expression below the median were characterized as having low expression. Spearman’s test was used to test correlation between expression of the different miRNAs. The Mann-Whitney U-test and Kruskal Wallis test were used to estimate associations between miRNA expression and clinicopathological characteristics. Differences in clinicopathological characteristics between relapsed and non-relapsed patients were evaluated by Pearson’s chi-square test. The associations between circulating miRNA expression levels and disease-free survival (DFS) or overall survival (OS) were assessed by the Kaplan-Meier method, log rank test (Mantel-Cox) and Cox proportional hazard regression models. DFS was calculated from the date of surgery until the date of relapse or death from any cause, whereas OS was calculated from the date of surgery until the date of death from any cause or last follow up. The Mann-Whitney test was used to examine the differential expression between the different groups of patients. To evaluate the value of circulating miRNAs in predicting relapse, receiver operating characteristics (ROC) curves were constructed and area under the curve (AUC) calculated. The Youden index (sensitivity + specificity – 1) was used to set the optimal cutoff point. Logistic regression analyses were performed to identify the best discriminating combinations of miRNAs with clinicopathological features. Cross-validation analysis was implemented in R using a generalized linear model for logistic regression, with recurrence/non-relapse as binary target variables (http://​www.​r-project.​org/). Statistical significance was set at p < 0.05 (two-sided test). This report is written according to the reporting recommendations for tumor marker prognostic studies (REMARK criteria) [25].

The flow diagram of the study and patients’ characteristics are summarized in Fig. 1 and Table 1, respectively. Plasma samples from 155 patients with early breast cancer and from 23 healthy women were processed for RNA extraction. There were 22 patients excluded from the analysis as described above. After a median follow-up period of 94.3 months (range 14.33–159.30), 84 out of the 133 patients with breast cancer who were included in the analysis remained disease-free and 49 had relapsed. Demographics and clinical characteristics were similar between patients who remained disease-free and those who developed recurrence, except for the proportions of patients with tumor size of > 5 cm (T3) and four or more infiltrated axillary lymph nodes, which were higher in patients who had recurrence (p = 0.015 and p = 0.003, respectively; Table 1). Patients were divided into three groups according to the clinical outcome: (i) patients who remained disease-free during the whole follow-up period (n = 84), (ii) patients with early relapse, defined as relapse within 3 years post-surgery (< 3 years; n = 23) and (iii) patients with late relapse, defined as relapse presenting at 5 years or more post-surgery (≥ 5 years; n = 20). Consequently, 6 out of 49 relapses were observed in between 3 and 5 years. Patients’ characteristics for the groups (ii) and (iii) are shown in Table 2. The median age was 52, 55 and 53 years in each group, respectively.

No significant associations were observed between miRNA expression (high expression, low expression) and age, menopausal status, tumor size, histological grade, number of infiltrated lymph nodes, estrogen receptor (ER), progesterone receptor (PR) or human epidermal growth factor receptor 2 (HER2) status (chi-square test, p > 0.05). However, miR-21 expression was higher in PR-negative as compared to PR-positive patients (63.4% vs 36.6%; chi-square test, p = 0.038). As expected, there was strong correlation between expression of miR-200b and miR-200c (Spearman’s Rho 0.628; p < 0.001) that belong to the same miR-200 family. Moreover, there was strong correlation between miR-21 and miR-200b (Spearman’s Rho 0.447; p < 0.001) and miR-200c (Spearman’s Rho 0.540; p < 0.001) expression, as well. Weaker but still significant association was observed between the dormancy-related miR-23b and miR-190 (Spearman’s Rho 0.236; p < 0.001) (Table 3).

Median expression levels of miR-21, miR-23b and miR-200c were significantly higher (p < 0.001, p = 0.028 and p < 0.001, respectively) and median miR-190 expression was significantly lower (p = 0.013) in relapsed compared to non-relapsed patients (Fig. 2). No significant difference was observed in the median expression of miR-200b (p = 0.063) between the two groups. Subsequently, we evaluated the DFS (Fig. 3) and OS (Fig. 4) in patients classified into high and low expression groups, according to the median value of each miRNA. We found that patients with high miR-21 expression had significantly shorter DFS compared to patients with low expression (105.03 months versus not reached; p < 0.001) (Fig. 3a). Similarly, patients with high miR-200c expression had significantly shorter DFS compared to those with low miR-200c (105.03 vs not reached; p = 0.005) (Fig. 3e). Finally, patients with high expression of both miR-21 and miR-200c had shorter DFS compared to patients with only one miRNA high or with both low (81.37 vs 132.9 and not reached, respectively; p < 0.001) (Fig. 3f). No significant differences in DFS were found among patients with high or low expression of miR-23b, miR-190 or miR-200b (Fig. 3b-d). Median survival was not reached by patients with either high or low expression of any of the miRNAs evaluated (Fig. 4b-e). Nevertheless, only patients with high miR-21 had significantly shorter OS compared to those with low miR-21 (p = 0.033) (Fig. 4a).

To evaluate further the prognostic value of the circulating miRNAs, univariate and multivariate analyses were performed that included demographic and clinical variables and the expression levels of the five miRNAs (classified into high or low). Cox univariate analysis revealed that patients with infiltrated axillary lymph nodes and those with negative hormone receptor expression had significantly shorter DFS (p = 0.013 and p = 0.028, respectively) and OS (p = 0.044 and p = 0.01, respectively) (Table 4). High miR-21 and miR-200c expression levels were significantly associated with shorter DFS (p < 0.001 and p = 0.007, respectively) and only miR-21 high expression was associated with shorter OS (p = 0.042) (Table 4). Cox multivariate analysis revealed that the involvement of axillary lymph nodes and hormone receptor negativity were independent prognostic factors for shorter DFS (p = 0.019 and p = 0.012, respectively) and OS (p = 0.029 and p = 0.006, respectively) (Table 4). Furthermore, only high miR-21 and high miR-200c expression emerged as independent prognostic factors associated with shorter DFS (p = 0.003 and p = 0.037) (Table 4).

We examined further whether the five circulating miRNAs are differentially expressed among patients classified into groups according to the timing of recurrence. For this purpose we compared miRNA expression levels in (i) patients who relapsed early compared to those who did not experience early relapse i.e. in patients who had recurrence within 3 years (n = 23) and those who either relapsed 3 or more years post-surgery or remained disease-free for the whole follow-up period (n = 110) and (ii) in patients with late relapse (at ≥ 5 years; n = 20) compared to those who remained disease-free during the whole follow-up period (n = 84). The Mann-Whitney test revealed that miR-190 expression levels were lower in patients with early relapse (p = 0.0032), whereas no differences were recorded for the remaining miRNAs (Fig. 5). Moreover, miR-21 and miR-200c expression was higher in patients with late relapse as compared to non-relapsed patients (p = 0.015 and p < 0.001, respectively; Fig. 6a and e).

Expression levels of various miRNAs were combined with clinicopathological characteristics in relapse-predicting models. We used binary logistic regression incorporating various combinations of miRNAs and used the corresponding ROC curves to determine the sensitivity and the specificity of plasma miRNA expression, to discriminate patients who subsequently had disease recurrence from non-relapsed patients (Fig. 7 and Table 5). When assessing single miRNAs, the ROC curves showed that the expression of miR-21 and miR-200c had the highest performance with an area under the ROC curve (AUC) of 0.685 (sensitivity 71.4%, specificity 63.9% (p < 0.001; 95% CI 0.592–0.777)) and AUC of 0.678 (sensitivity 75.5%, specificity 61% (p < 0.001; 95% CI 0.586–0.769)), respectively (Fig. 7a and e, respectively and Table 5). When assessing combinations of miRNAs, binary logistic regression analysis resulted in a pattern of three miRNAs (miR-21, miR-23b and miR-190) bearing the highest predictive accuracy. The AUC from ROC analysis of this combined model 0.765, with sensitivity 80% and specificity 65.3% (p < 0.001; 95% CI 0.673–0.850) (Fig. 8a and Table 5). Eventually, the combination of the three miRNAs with the currently used clinical prognostic parameters, axillary lymph node infiltration and tumor grade, resulted in superior discriminatory capability (AUC 0.873, sensitivity 89% and specificity 76.2% (p < 0.001; 95% CI 0.802–0.940)) compared to the expression of the three miRNAs alone or to the clinicopathological features alone (Fig. 8b, c and Table 5). Using the same procedure the combination of miR-200c expression, axillary lymph node infiltration, tumor grade and ER status resulted in an increased AUC of 0.890 with a sensitivity 75% and specificity 89% (p < 0.001; 95% CI 0.818–0.972) for the prediction of late disease relapse (Fig. 8d and Table 5). When the same model was fitted to predict early relapse, there were no differences in the discriminatory power when combining miRNAs with clinicopathological parameters.

The robustness of the predictive performance of our models was assessed through a cross-validation strategy. A 10-fold cross-validation with a 70–30 split (70% training data, 30% testing data) was implemented in R and applied on nine different feature combinations of miRNAs and clinicopathological features. Mean AUC values were calculated for each 10-fold cross-validation. The mean AUC was then compared to the AUC calculated from our initial regression analysis. There were no significant differences in the values of AUC in any variable combinations, indicating that the performance of these models is robust and can be generalized to independent datasets (Additional file 1: Table S1).