Public microarray datasets were extracted from the GEO and TCGA database. The checklist and pipeline for proper organization of the integrated analysis were determined following the reporting guidelines of microarray meta-analysis recommended by Ramasamy et al. (2008). Only original experimental studies to screen miRNAs that were differentially expressed (DE) between CRC and ANT in at least 40 human samples were included. Selection criteria, probe annotation, and data normalization were the same as described in previous reports (Sun et al., 2017a, b; Lin et al., 2019).

Differentially expressed miRNAs between CRC and ANT were determined by MetaOmics software¹ in the MetaDE package (Wang et al., 2012). The filter thresholds of the mean and SD were set to 30% in integrated analysis. Fisher’s method was performed for statistically significant analysis to counterpoise the different stringencies of the methods. A permutation method with a modified t-test was used for the removal of the batch effect and estimation of the P-values (Tseng et al., 2012). P-value < 0.001 was used as the cutoff for statistically significant DE miRNAs.

The candidate miRNAs were determined by a LASSO algorithm with penalty parameter tuning with 10-fold cross-validation and an SVM-Recursive Feature Elimination (RFE) algorithm (Qiu et al., 2017). In total, nine common candidate miRNAs were further selected for the training cohort.

The diagnosis model was developed for evaluation of the differential capacity of CRC and ANT in GSE49246 using logistic regression of the nine candidate miRNAs. Furthermore, candidate miRNAs were used to validate the results of the ROC curve analysis in an independent validation set (GSE115513) by forecasting the grades of CRC and ANT, as described previously (Sun et al., 2017b).

A nine-miRNA prognostic signature model was developed by multivariate Cox hazard model analysis of GSE29622 data to estimate prognostic risk score (PRS) using the formula: PRS = Σ(C × EXP_miRNA), where EXP was the FPKM value of the miRNA, and C was the regression coefficient for the corresponding miRNA in multivariate Cox hazard model analysis. The median PRS of the testing dataset GSE29622 was used to distinguish the high-risk cluster from the low-risk cluster. The prognostic performance of the miRNA signature model was evaluated by comparing the AUC of ROC curves. The association of miRNA signature with patient survival was analyzed, and the miRNA signature model was finally validated in an independent dataset, TCGA-COAD.

The speculated targets of integrated-signature miRNAs, particularly for the tumor suppressor miRNA hsa-miR-200a, were predicted by four different target-prediction algorithms: TargetScan v7.1², miRanda³, DIANA-TarBase v7.0⁴, and PicTar⁵ via the “miRNAtap” package (Dhawan et al., 2018).

Human colon cancer cell lines DLD1 and SW480 and normal intestinal epithelium cell line NCM460 were provided by the Cell Bank of Wuhan University (Wuhan, China). hsa-miR-200a-3p mimic, inhibitor, mimic NC, and inhibitor-negative control were from RiboBio Co., Ltd. (Guangzhou, China). Cell culture and transient transfection were performed as described previously (Sun et al., 2017a; Lin et al., 2019).

Colorectal cancer tissues and ANT (distance to cancer >5 cm) were sampled from sixty patients with primary CRC that was diagnosed by pathological assessment of tissues at Taihe Hospital of Hubei University of Medicine from January 2017 to December 2018. The study was authorized by the Research Ethics Committee of Hubei University of Medicine (Shiyan, Hubei, China). Informed consent was provided by all patients who participated in the present study.

Total RNA was isolated using Trizol reagent (Invitrogen, United States). The expression of mature miR-200a-3p was determined using qRT-PCR with the Bulge-Loop^TM miRNA qRT-PCR Primer Set and Control Primer Set (RiboBio, Guangzhou, China); 2 μg of RNA was used to synthesize cDNA using the RevertAid First Strand cDNA Synthesis Kit (Thermo, United States). qRT-PCR was performed using the FastStart Universal SYBR Green Master (Rox) (Roche, United States) in the ABI PRISM^® 7300 real-time PCR system (Applied Biosystems, Foster City, CA, United States). GADPH and U6 were added as endogenous controls, and melting curves were used to check non-specific amplifications. Relative expression level was calculated by the 2^{–Δ ΔCt} method. The primer sequences are shown in Supplementary Data Sheet S1.

Approximately 0.2 g CRC and paired normal tissue samples were ground in liquid nitrogen and lysed with RIPA lysis buffer (CelLytic, Sigma-Aldrich, United States) and proteinase inhibitor cocktail (Merck Millipore, United States). Total protein was extracted and immunoblotted as described previously (Sun et al., 2017a; Lin et al., 2019). The antibody used in the present study is shown in Supplementary Data Sheet S1. The dual-luciferase reporter assays, colony formation assay, wound healing assay in vitro, transwell migration/invasion assay, transfection reagents, primers, and western blot antibody are described in Supplementary Data Sheet S1.

All statistical analyses were performed using R/BioConductor (version 3.5.1) with two-tailed P-values. In the training phase, the dependent variable of the LASSO-logistic regression and SVM-RFE algorithm is dichotomized into two groups (tumor vs. normal), whereas the independent variable is the continuous expression values of 68 miRNAs.

In the testing phase, logistic regression is used for the diagnostic model of the binary variables, tumor and normal, and the independent variable is the continuous expression values of nine miRNAs. ROC was analyzed with R (“pROC,” “MASS,” “nnet,” “Daim,” and “epicalc” packages) to evaluate the diagnostic efficiency for the diagnostic model on the CRC outcome for the expression level of the nine candidate miRNAs. For Cox regression, the dependent variable is survival data showing the state of survival and the time of follow-up, whereas the independent variable is the continuous expression values of the nine miRNAs. Kaplan-Meier plots and ROC curve for the prognosis model were constructed using R (“survival,” “survminer,” “ggplot2,” and “survivalROC” packages). The median value was used as the cutoff value between low and high expression levels of each miRNA. The Cox proportional hazards regression model was applied to determine OS, and log-rank test was used to detect differences in OS.

For the validation phase, the nomogram prognostic model contains clinical features such as the total TNM Stage, T stage, N stage, M stage, and risk score. The clinical features (TNM Stage, T stage, N stage, M stage) are factor variables (categorical variables), and risk score is a continuous variable. The performance of the nomogram for the testing and validation data was evaluated by the concordance index and was distributed graphically with calibration plots using the “rms” package in R.

In the verification phase, ANOVA or Student’s t-test was used to detect significant differences between the groups with P < 0.05 as the threshold. The data from at least three independent experiments were averaged and presented as mean ± SD.

Three CRC miRNA datasets consisting of 170 CRC samples and 172 ANT samples (Supplementary Table S1) were selected by integrated analysis (Figure 1), from which 135 miRNAs with steady DE patterns were identified by a moderated t-test with addition of a fudging parameter, and Fisher’s method with summarization of -log (p-value) across studies to run 300 selections (Wang et al., 2019). The effect size was merged, and a total of 68 DE miRNAs with profiles of DE miRNAs similar to those from Fisher’s method of combining P < 0.001 (Supplementary Table S2) were determined (Figure 2A). As expected, the CRC was differentiated from ANT samples by hierarchical clustering of the 68 DE miRNAs (Figure 2B).

The miRNAs that were selected by the LASSO (Figure 3A) and SVM-RFE (Figure 3B) algorithms were merged, among which nine miRNAs that were overlapped in both algorithms (Figure 3C, Supplementary Table S3 and Supplementary Figure S1) and constituted a minimum candidate list for model prediction of CRC samples (Figures 3A–C).

A total of 80 testing-phase samples in GSE49246 and 1513 validation-phase samples in GSE115513 were differentiated into the CRC and normal groups using a linear logistic regression model of nine miRNAs. A DRS was estimated and weighted with the coefficients by the univariate linear regression model to evaluate the performance of nine miRNAs for prediction of CRC in the GSE49246 cohort. The DRS was estimated by the formula: DRS = (−5.1928 × EXP_{hsa–miR–492}) + (0.1885 × EXP_{hsa–miR– 200a}) + (−3.5434 × EXP_{hsa–miR–338}) + (−9.1146 × EXP_{hsa–miR– 29c}) + (0.8177 × EXP_{hsa–miR–101}) + (−1.3238 × EXP_{hsa–miR– 148a}) + (2.9460 × EXP_{hsa–miR–92a}) + (−1.2699 × EXP_{hsa–miR– 424}) + (2.7264 × EXP_{hsa–miR–210}). ROC curve analysis indicated high categorization accuracy of a panel of nine miRNA signatures in the testing set GSE49246 (AUC = 0.94) (Figure 3D), in the validation set GSE115513 (AUC = 0.89) (Figure 3E), and in our independent validation set (AUC = 0.978) (Figure 3F and Supplementary Table S4).

Significant associations of candidate miRNAs with the survival of CRC patients were analyzed by multivariate Cox regression in 80 testing-phase GSE29622 samples (Supplementary Table S4). PRS was estimated using the nine-miRNA signature model as follows: PRS = (0.1170 × EXP_{hsa–miR–492}) + (0.3496 × EXP_{hsa–miR–92a}) + (0.3640 × EXP_{hsa–miR–424}) + (−0.5316 × EXP_{hsa–miR–29c}) + (0.0573 × EXP_{hsa–miR–101}) + (0.0882 × EXP_{hsa–miR–338}) + (−0.5038 × EXP_{hsa–miR–200a}) + (0.1920 × EXP_{hsa–miR–210}) + (0.2397 × EXP_{hsa–miR–148a}), where EXP was the FPKM value of the miRNA.

The prognostic performance of the nine-miRNA signature model was characterized using the GSE29622 dataset. The median PRS (1.036) was used to classify CRC patients into a low-risk group (n = 33) and a high-risk group (n = 32) (Figure 4A). As shown in Figure 4B, the high-risk group had significantly shorter survival time or lower survival probability compared with the low-risk group (HR, 4.741, 95% CI, 2.13–10.55, log-rank test p = 0.0002). The AUC for predicting the 1- to 10- year survival of CRC patients was between 0.783 and 0.872 (Figure 4C). High expression of hsa-miR-200a and low expression of hsa-miR-492, hsa-miR-92a, hsa-miR-424, hsa-miR-29c, hsa-miR-101, hsa-miR-338, hsa-miR-210, and hsa-miR-148a were associated with longer predicted survival (Figures 4D–L). The results suggested the nine-miRNA signature enabled effective prediction of the prognosis of CRC patients.

The prognostic performance of the nine-miRNA prediction model for CRC was validated using the validation dataset TCGA-COAD. The CRC patients were classified into the high-risk group (n = 203) and the low-risk group (n = 203) using PRS and the median PRS as cut-off criteria (Figure 5A). The high-risk group had significantly shorter survival time or lower survival probability compared with the low-risk group (HR, 3.9, 95% CI, 2.125-7.16, log-rank test p < 0.001) (Figure 5B). The AUC for predicting 1- to 10- year survival of CRC was between 0.796 and 0.911 (Figure 5C). The results were consistent with those of the testing set (Figures 4, 5), supporting the effectiveness of the nine-miRNA signature in predicting the prognosis of CRC.

The effect of each clinicopathological feature on survival was analyzed by Cox regression to evaluate whether the prognostic classifier acted as an independent indicator in CRC patients. As shown in Table 1, after multivariable adjustments for clinicopathological factors, the nine-miRNA-based classifier was demonstrated to be a powerful and independent factor in the testing cohort of 65 cases (HR 1.232, 95% CI 1.115–1.363, P < 0.001) and in the validation set of 406 cases (HR 1.455, 95% CI 1.287–1.645; P < 0.001).

Time-dependent ROC curve analysis revealed that the nine-miRNA signature was a better predictor of survival compared with clinicopathologcial factors in both the testing and validation sets (Figures 5D,E).

A decreasing mode in the miR-200a levels was observed in human CRC (Figure 2C). The levels of miR-200a-3p in DLD1 and SW480 cell lines were lower than those in the normal intestinal epithelium cell line NCM460 (Figure 6A). qRT-PCR (normalized with U6) was used to measure mature miR-200a-3p level in the independent validation sample cohort (n = 60). The results indicated that the expression of miR-200a-3p was significantly (P < 0.05) decreased in the ANT-related tumors (Figure 6B). Both qRT-PCR and Western blot analyses demonstrated significant upregulation of FOXA1 in CRC tissues compared with ANT (Figures 6C,D). The miR-200a-3p expression levels were significantly and inversely associated with FOXA1 levels in both CRC tissues and normal samples (R² = 0.21, P = 1.01E-07) (Figure 6E) in our independent validation cohort, but they were not associated with FOXA1 levels in the TCGA colon or rectal cancer cohort (Supplementary Figure S3).

A total of 66 common targets of miR-200a were identified from both prediction algorithms and experimentally supported databases using four publicly available algorithms with high-stringency (Figure 6F and Supplementary Table S5).

Based on the results of the integrated analysis and in vitro experiments, we hypothesized that the reduced expression of miR-200a-3p might accelerate CRC tumorigenesis via targeting FOXA1. We transfected SW480 and DLD1 cells with miR-200a-3p mimic, miR mimic NC, miR-200a-3p inhibitor, miR inhibitor NC to determine the role of miR-200a-3p in clone formation. Clonogenic assay revealed that the clonogenic survivals of SW480 and DLD1 cells were reduced by miR-200a-3p compared with the NC groups, whereas phenotype was reversed by miR-200a-3p inhibitor (Figures 7A,B). The wound healing assay showed that miR-200a-3p overexpression suppressed and miR-200a-3p inhibitor enhanced migration of CRC cells (Figures 7C–F). The transwell assay showed that miR-200a-3p overexpression suppressed and miR-200a-3p inhibitor enhanced invasion of CRC cells (Figures 7G–J). These data indicated that miR-200a suppressed the proliferation, migration, and invasion of CRC cells.

FOXA1 3′-UTR was cloned into a luciferase reporter plasmid (Figure 8A), and expression of the adjacent hRluc coding region was quantified to evaluate miR-200a-3p-regulated FOXA1 expression. Searching different databases revealed that FOXA1 is a predicted target of miR-200a-3p that presents oncogenic properties. FOXA1 harbors two conserved miR-200a-3p cognate sites: nucleotides 310-317 and 271–276 of FOXA1 3′-UTR (Figures 8B,C). The luciferase reporter plasmid FOXA1-3′-UTR or mutant reporter plasmids with point mutations in the putative miR-200a-3p binding sites were co-transfected with miR-200a-3p mimics or miR mimic NC and inhibitors, separately. The results showed that the luciferase activity of reporter plasmid with wild type FOXA1 3′-UTR was significantly inhibited by miR-200a-3p but increased by miR-200a-3p inhibitor (Figure 8D, P < 0.05). However, no significant change was observed in reporter plasmid containing mutated FOXA1 3′-UTR (i.e., MUT-FOXA1-3′-UTR). These results suggested that miR-200a-3p negatively regulated FOXA1 expression by directly binding to the predicted binding site(s) in the FOXA1 3′-UTR.

DLD1 and SW480 cells were transfected with miR-200a-3p mimic, miR mimic NC, miR-200a-3p inhibitor, and inhibitor NC to verify the direct regulation of FOXA1 expression by miR-200a-3p. Both qRT-PCR and Western blotting revealed that the expression level of FOXA1 was suppressed in miR-200a-3p-overexpressed cells but regained in miR-200a-3p inhibitor-treated cells (Figures 8E,F). Therefore, miR-200a-3p negatively regulated FOXA1 levels in CRC cells. In addition, the results showed that miR-200a-3p negatively regulated expression of N-cadherin and YAP1 proteins and positively regulated E-cadherin expression (Figure 8F), suggesting that miR-200a-3p inhibited EMT of CRC cells.