Clinical and pathological findings
The clinical features of samples collected from epithelial of patients (30 BC tissue samples; 29 adjacent normal tissue samples) in our study are summarized in Table 1. 30 bladder cancer tissue samples include 1 case of stage Tis (in training group), 25 cases of stage I (15 cases in training group and 10 cases in double blind group), 2 cases of stage II (including one of low-lever urothelial carcinoma in training group), 1 cases of stage III (in training group) and 1 cases of stage IV (in double blind group). 29 adjacent normal tissue samples include 19 cases in training group and 10 cases in double blind test group.
Biomarker selection based on qRT-PCR
The miRNA expression data from real-time PCR was performed by using R and package e1071 with some modifications as the following: First, set Ct value as 32 to any miRNAs if their measured Ct values were greater than 32. Then, we compared mean Ct-value of each miRNA between tumor tissue and control tissue for 365microRNAs (15 miRNAs without expression were removed from analysis), 2 endogenous controls. There are many miRNAs with significant higher Ct-value in cancer tissue were shown in Fig. 1. Afterwards, we did T-test for each miRNA. 70 miRNAs with P-value < 0.001 were displayed in Fig. 1.
In order to prove there are tumor related markers and to find out these markers, we trained a SVM model, and predicted 20 other samples. The model was built as the following: a). we produced 66430 new variables by computing the difference between Ct of each tow miRNAs and got 66796 total variables with the original 366 variables. b). we did T-test on all these variables and took 2 variables with the greatest P-value and some variables with smallest P-value. Until we had got 20 candidate miRNAs those variables included. c). for each miRNA subset with markers less than or equal to 12 of the 20 candidate miRNAs, we trained a SVM model with default parameters by function SVM and evaluated the accuracy by 50 times 10-fold cross validation. d) Then 10 subsets with the best accuracy and AUC were selected and 10 SVMs based on these 10 marker subsets were trained by total 38 samples and the stat of 21 blind samples was predicted. They all showed good performances on blind samples with accuracy greater than 0.95. The best one was selected for further analysis.
miRNA expression profiling in BC and normal controls
To identify BC specific miRNA expression signatures as biomarkers to diagnose BC, we applied the panel to profile miRNA expression of 38 tissue samples in training group, including 19 BC tissue samples and 19 matched adjacent normal tissue samples. The 365 dysregulated microRNAs and two endogenous controls in BC tissue samples and in adjacent normal tissue samples were presented in the supplemental Table 1. There were 121 microRNAs, including miR-182, miR-431, miR-183, miR-429 and miR-425 etc., with a higher expression levels in the BC tissues compared with adjacent normal tissues. In contrast, 245 microRNAs, such as miR-1, miR-133a, miR-133b miR-125b miR-143 and miR-145 etc, had a lower expression level in the BC tissues relative to adjacent normal tissues. The aberrant expression levels of miRNAs were summarized in Table 2.
Developing miRNAs expression signatures in diagnosis of BC
An unpaired T-test (p< 0.05) with a Benjamini Hochberg FDR multiple testing corrections was used to identify significantly dysregulated miRNAs that distinguish BC from normal controls. Accurate classification of BC patients from normal controls is crucial for successful BC treatment; we investigated the diagnostic value of the miRNA-expression profile in BC patients. Among the 70 significant miRNAs (P-value < 0.001) checked (these miRNAs listed in Table 2 with t-value and p-value), several expression signatures of three or four-miRNAs were developed as predictors of BC from normal controls. These signatures were selected based on a machine learning approach of support vector machine (SVM).
The best 10 groups of miRNA signatures were listed in Table 3. Thereby, in the group 1, the expression levels of miR-133a (log2FC = 9.258; P < 0.0001) were significant down regulated in BC patients. However, expression of miR-431 (log2FC = -3.268; P < 0.0001) was significant higher in BC tissues than in adjacent normal tissues. In spite of level of miR-4251 (P > 0.1) was slightly higher in BC tissues than in adjacent normal tissues. The results showed that the use of improved comparative Ct method seems to be an easily applicable method with potential for general clinical use that avoids the need for large-scale, high-throughput profiling analyses and was therefore used to develop clinically useful signatures based on tissues biomarkers (Supplemental Fig. 1).
Prediction of BC and control subjects by risk score analysis
To verify the accuracy and specificity of these three or four miRNA signatures to be used as BC biomarkers, we further assessed the 3 miRNAs in the former set consisting of 38 samples, including of 19 BC tissue samples and 19 matched adjacent normal tissue samples (Fig. 2). The areas under the ROC curve (AUC) were 1 with 100% sensitivity and 100% specificity was respectively.
Double blind test
To verify the accuracy and specificity of identified miRNA signatures to be used as BC biomarkers, another 21 samples (including 11 BC tissue samples and 10 adjacent normal bladder tissue samples) were tested in a double-blind fashion to validate the predictive ability of the miRNA-based signatures for BC diagnosis. The accuracy 95.2% of the signature consisting of hsa-miR-133a, hsa-miR-431 and hsa-miR-4251 with 100% sensitivity and 90% specificity was respectively (Fig. 3). Especially, 10 of stage I BC tissues were confirmed by cytology.