Identification of metastasis-related microRNAs in hepatocellular carcinoma^†

Clinical Specimens.

Hepatic tissues were obtained with informed consent from patients who underwent radical resection between 2002 and 2003 at the Liver Cancer Institute and Zhongshan Hospital (Fudan University, Shanghai, China). The study was approved by the Institutional Review Board of the Liver Cancer Institute and the US National Institutes of Health. Gene expression profiles were conducted in primary HCC and corresponding noncancerous hepatic fresh frozen tissues from 241 Chinese patients with HCC. Among them, 93% had underlying cirrhosis and 68% had a serum alpha-fetoprotein level > 20 ng/mL (Table 1). The sample inclusion criteria are included in the supplemental text. The general strategy for partitioning cases and testing the miRNA signature is outlined in Fig. 1. A total of 131 metastasis or nonmetastasis cases were used to build a miRNA signature of metastasis (Fig. 1, Step 1). These cases were associated with clear outcomes (that is, those with accompanying metastasis at surgery and those without metastasis and recurrence at surgery and at 3-year follow-up). Among them, 29 had primary HCC lesions (19 solitary and 10 multinodular) accompanied by tumor emboli found in the major branches of the portal vein (n = 23), inferior vena cava (n = 2), or common bile duct (n = 4; one also with tumor thrombi in the inferior vena cava) and 102 had solitary HCC with no metastasis/recurrence found at follow-up (3 years). The median overall survival of cases with metastasis was 349 days whereas that of the nonmetastasis cases was 664 days. In the validation analysis, we used the remaining 110 independent cases (Fig. 1, Step 3). This group appeared to consist of cases whose prognosis could not be accurately determined at the time of resection by several HCC staging systems. The 110 cases included 43 multinodular and 67 solitary HCC. Of the 43 multinodular HCC cases, 18 developed intrahepatic recurrence and one developed extrahepatic metastasis in addition to an intrahepatic recurrence. Of the 67 solitary HCC cases, 4 patients had a solitary tumor with an appearance of aggregated nodules, 10 developed intrahepatic and/or extrahepatic metastases whereas 49 developed intrahepatic recurrence confirmed at follow-up (3 years). In patients with multinodular tumors, the largest tumor was used for miRNA profiling.

RNA Isolation,miRNA Arrays, and qRT-PCR.

The RNA isolation and miRNA array methodology were essentially as described.9, 14 In the analysis of the 241 HCC cases, RNA was isolated in a pairwise fashion from tumor or nontumor tissue and samples were selected in random order for miRNA analysis to avoid grouping bias. The tumor and nontumor tissues were run separately on a total of 482 single-channel microarrays (see Supplementary data). The microarray platform and data have been submitted to the Gene Expression Omnibus (GEO) public database at the National Center for Biotechnology Information, following MIAME guidelines. The accession numbers are GPL4700 (platform) and GSE6857 (samples; release date January 2008). In addition, REMARK guidelines have been followed to report the miRNA-based metastasis markers in this study.38 For quantitative reverse transcription polymerase chain reaction (qRT-PCR), 70 M or NM cases were randomly chosen from the 131-case training set and RNA was isolated and converted to cDNA using the High Capacity cDNA Archive Kit (Applied Biosystems, Foster City, CA). Reactions were performed using ABI-purchased primer probe sets with the ABI-Prism 7700 Sequence Detector System (Applied Biosystems) (see Supplementary Methods). U6 RNA was used as a control. The statistical P value, generated by the student t-test, and the Spearman correlation constant were calculated using GraphPad Prism, version 5.

Statistical Analyses.

Unsupervised hierarchical clustering analysis was performed by the GENESIS software version 1.5 developed by Alexander Sturn (IBMT-TUG, Graz, Austria). The BRB ArrayTools software version 3.3 was used for supervised analysis to search for differentially expressed genes, as described9, 13 (see also Supplementary Methods). In the qRT-PCR-based profiling for class prediction of 70 M or NM cases utilizing the 4-miRNA signature, we used PAM (Prediction Analysis of Microarrays) developed by Tibshirani et al.39 PAM uses nearest shrunken centroids to predict unknown samples with cross-validated training and testing, which appears to be superior in performance to identify smaller sets of genes when compared to other class prediction algorithms. PAM permutation was performed using R version 2.5. The Kaplan-Meier survival analysis was used to compare patient survival based on prediction results, using Excel-based WinSTAT software. The statistical P value was generated by the Cox-Mantel log-rank test. Survival Risk Prediction analysis was performed using BRB ArrayTools software version 3.3. Cox proportional hazards regression was used to analyze the effect of fourteen clinical variables on patient survival using STATA 9.2 (College Station, TX) (see Supplementary Methods). The statistical significance was defined as p<0.05. TargetScan analysis was based on a website tool developed by Ben Lewis (http://genes.mit.edu/targetscan/index.html) (see Supplementary Methods).40 A search for experimentally proved targets of the 20 miRNAs was performed using the Tarbase database (www.diana.pcbi.upenn.edu/tarbase).

The Search for Metastasis-Related miRNAs in HCC Tissues.

HCC-related mortality and recurrence could be attributed to at least 2 problems whose underlying biological activities can differ significantly, that is, metastasis and/or development of de novo HCC due to cirrhosis. Due to the limited number of HCC metastasis cases available for analysis, in a set of 241 archived HCC cases, we first searched for metastasis-related miRNAs based on the comparison between primary HCC or noncancerous tissues from all available cases with venous metastases (M; n = 29) at the time of surgery and 102 nonmetastasis cases (NM) with no evidence of metastasis at the time of surgery and after 3 year follow-up by a supervised class comparison approach (that is, the worst-prognosis and best-prognosis cases, respectively; see supplementary methods) (Fig. 1, Step 1; Table 1). We identified 20 miRNAs that could significantly discriminate the tumor tissues of M from NM cases (Fig. 2A; Table 2). It appeared that mir-219-1, mir-207, and mir-338 were most highly up-regulated, whereas mir-34a, mir-30c-1, and mir-148a were most highly down-regulated in metastasis cases. When the noncancerous tissue miRNA expression data were separately analyzed, we could not identify any miRNA capable of distinguishing M from NM at the same statistical significance. This may be a reflection of genetic make-up or epigenetic factors accounting for gene expression changes in the microenvironment versus gain or loss of a gene in the tumor or perhaps differences in the sensitivity between the mRNA and miRNA array platforms used to assay the liver microenvironment. Thus, unlike the mRNA profile associated with the HCC microenvironment,13 it appears that there are more measurable changes in miRNA expression in tumor cells compared to that of the surrounding tissue, and suggests that analysis of miRNA expression in tumor tissues may be better suited for differentiating groups of HCC patients. We note that significant miRNAs could not be identified when a comparison of these tissues was made with other clinical variables including multinodular status, microvascular invasion, and 4 clinical staging systems.

To validate the microarray data, we performed qRT-PCR of the top 2 up-regulated (mir-219 and mir-338) and the top 2 down-regulated miRNAs (mir-30c and mir-148a) in the M/NM comparison in a set of 70 randomly selected HCC cases. We found that each of these miRNAs could significantly discriminate M versus NM cases based on their RT-PCR based expression changes and significantly correlated with the miRNA array data (Fig. 2B; Supplementary Fig. 1). Next, we performed PAM analysis with 10-fold cross-validation and 1000 permutations of the class labels to assess the predictive capacity of these 4 miRNAs based on qRT-PCR readings. We found that the expression changes of these miRNAs could correctly classify M and NM samples with 100% accuracy (62% mean permutation accuracy, P < 0.0001) (Fig. 2C). Therefore, it appeared that the expression of these miRNAs were significantly associated with metastasis.

The Association of Metastasis-Related miRNAs with Survival.

To establish a classifier, we performed multivariate nearest neighbor class prediction with 10-fold cross-validation and 1000 permutations of the class label in the 131 M and NM cases (Fig. 1A, Step 2). In this analysis, 90% of the samples were randomly chosen to build a classifier which was then used to predict the remaining 10% of the cases. The accuracy of the prediction was calculated after 1000 repetitions of this random partitioning process (See Supplementary Methods). This analysis yielded a 20-miRNA signature that could significantly predict M and NM status with an overall accuracy of 76% (multivariate P < 0.001). Kaplan-Meier survival analysis based on the 20-miRNA prediction results revealed that the predicted metastasis group had a significantly shorter survival period when compared to the nonmetastasis group (P < 0.042) (Fig. 2D). Thus, this signature is associated with patient survival.

Next, we tested whether the predicted propensity for metastasis based on this signature was related to survival in 110 additional cases. Due to the limited number of patients with metastasis at surgery, we used these 110 cases who developed relapse within 3 years after surgery as patients with potential HCC metastasis to test our miRNA signature developed in “extreme” cases of the training set. We found that the miRNA predictor had a 72% overall prediction accuracy in these independent cases. The predicted M group had a significantly worse survival rate than the NM group (P = 0.009) (Fig. 3A). The 500-day cumulative survival rate based on the 20-miRNA predictor was 41% (95% confidence interval [CI] = 19% to 63%) for patients classified as M compared to 75% (95% CI = 64%–83%) for patients classified as NM (Fig. 3A). Interestingly, available HCC prognostic staging systems (that is, TNM, OKUDA, CLIP, or BCLC) were incapable of predicting patient survival in this independent case set (Supplementary Fig. 2).41-44

Because the ability to predict risk of cancer spread at early stages of HCC could have a significant clinical impact, we also assayed the prognostic capacity of the 20-miRNA signature in patients with early stage HCC. Because long-term survival can be affected by tumor number and underlying liver disease, we performed univariate Cox proportional hazards regression using TNM, BCLC, CLIP or OKUDA staging of the test set to assess the affect of nodular status accounted for by stage on survival outcome in this cohort. We found that nodularity accounted for by staging systems did not affect outcome in this cohort and therefore, we assayed TNM stage I and II as early stage HCC (n = 89) (data not shown). A significantly worse survival was observed for the predicted M patients versus NM by the 20-miRNA signature in the early stage cases (P = 0.022) (Fig. 3B). The 500 day survival rate based on the 20-miRNA predictor was 38% (95% CI = 14%–63%) for patients classified as M compared to 73% (95% CI = 60%–82%) for patients classified as NM (Fig. 3B). In addition, we also tested the capacity of the miRNA signature to predict recurrence in the entire independent set or in early stage cases and found that the predicted M group based on the 20-miRNA signature had a higher recurrence rate (P = 0.056, or P = 0.002, respectively) than the NM group (Fig. 3C,D).

To further test whether metastasis-related miRNAs were those largely associated with survival, we searched for miRNAs whose expression was correlated with this outcome. A survival risk prediction based on two survival risk groups (that is, high versus low; cutoff defined by a 50% prognostic index; see Supplementary Methods) and a 10-fold cross validated permutation test (1000×) at a statistical risk prediction cutoff of P < 0.01 in the entire cohort yielded 46 miRNAs associated with survival (Table 3). We found that 16 (80%) of these survival-related miRNAs overlapped with the 20 metastasis-related miRNAs (Table 2). Kaplan-Meier survival analysis based on the survival risk prediction results revealed that the predicted high-risk group had a significantly shorter survival period when compared to the low-risk group (P = 0.014; 1000× permuted log-rank test) (Supplementary Fig. 3A). Similarly, in the entire cohort, the 20-miRNA-based predicted M group had a significantly worse survival rate than the NM group (P = 0.04) (Supplementary Fig. 3B). The 500 day cumulative survival rate based on the 20-miRNA predictor was 62% (95% CI = 48%–74%) for patients classified as M compared to 81% (95% CI = 74%–86%) for patients classified as NM in the entire cohort (Supplementary Fig. 3B). Therefore, it appeared that the miRNA metastasis signature identified in this study was significantly associated with patient prognosis, including patients with early-stage HCC.

We further addressed the relationship of the 20-miRNA-based metastasis signature with survival outcome by performing a survival risk prediction analysis on the training and testing data sets. This analysis was performed using a geneset restricted to only the 20 metastasis miRNAs at a P value of 0.01 with 10-fold cross-validation and 1000 permutations for the significance of the log-rank test. The results demonstrated that these miRNAs could significantly differentiate patient groups based on survival outcome (log-rank P = 0.005 and permuted P = 0.03) (Fig. 4). This analysis provided further validation that these 20 miRNAs were associated with survival outcome.

Comparison of the miRNA Predictor and Known Clinical Variables.

Next, we performed Cox proportional hazards regression analysis to determine whether the miRNA predictor was confounded by underlying clinical conditions. A univariate analysis revealed that the miRNA signature was a significant predictor of survival (P = 0.006) (Table 4). The multivariate survival model, which controlled for potential confounding covariates demonstrated that the miRNA predictor was associated with a significant 2.1-fold increased risk of death for patients with the M versus the NM expression profile (Table 4). We also assessed the miRNA signature in relation to other available staging information within this cohort. Although CLIP staging and OKUDA staging were significantly associated with survival in univariate and multivariate analysis, the 20-miRNA metastasis signature maintained its significant association with survival in final models including either CLIP (20miRNA HR =2.3 (CI: 1.3–4.1) and P = 0.007) or OKUDA staging systems (20miRNA HR = 2.3 (CI: 1.3–4.0) and P = 0.004) (data not shown). Analysis of BCLC staging was not significant in univariate analysis (Stage A versus O: P = 0.061; Stage B versus O: P = 0.063), perhaps due to the smaller number of cases available for analysis with known BCLC stage and survival data compared to that of TNM, CLIP, and OKUDA in the cohort studied (data not shown). Thus, the miRNA signature is an independent predictor for survival.

A Search for Metastasis-Related miRNA Targets.

Of the 20 miRNAs, 4 were overexpressed in M whereas 16 were overexpressed in NM (Table 2). This suggested that miRNAs are mainly down-regulated in metastasis and hence, an upregulation of their targets would be found in samples bearing this phenotype. In a previous study using cDNA microarray we identified a 153-gene metastasis signature from hepatic tumors.9 Using a Tarbase bioinformatics approach, we searched for experimentally validated targets of the 20 miRNAs. We then generated a putative miRNA target gene list using TargetScan, whose output was restricted to those 153 tumor-associated metastasis genes with an inverse correlated expression level (Table 2). It appeared that the host targets SPTBN2, GTF2H1, PSCD3, VAMP3, and SLC20A2 are affected by multiple miRNAs related to metastasis and may be part of signaling pathways that significantly contribute to this phenotype.