A microRNA-based prediction model for lymph node metastasis in hepatocellular carcinoma

We developed an efficient microRNA (miRNA) model that could predict the risk of lymph node metastasis (LNM) in hepatocellular carcinoma (HCC). We first evaluated a training cohort of 192 HCC patients after hepatectomy and found five LNM associated predictive factors: vascular invasion, Barcelona Clinic Liver Cancer stage, miR-145, miR-31, and miR-92a. The five statistically independent factors were used to develop a predictive model. The predictive value of the miRNA-based model was confirmed in a validation cohort of 209 consecutive HCC patients. The prediction model was scored for LNM risk from 0 to 8. The cutoff value 4 was used to distinguish high-risk and low-risk groups. The model sensitivity and specificity was 69.6 and 80.2 %, respectively, during 5 years in the validation cohort. And the area under the curve (AUC) for the miRNA-based prognostic model was 0.860. The 5-year positive and negative predictive values of the model in the validation cohort were 30.3 and 95.5 %, respectively. Cox regression analysis revealed that the LNM hazard ratio of the high-risk versus low-risk groups was 11.751 (95 % CI, 5.110–27.021; P < 0.001) in the validation cohort. In conclusion, the miRNA-based model is reliable and accurate for the early prediction of LNM in patients with HCC.


INTRODUCTION
Hepatocellular carcinoma (HCC) is the most frequent histological form of primary liver cancer [1] and the sixth most common cancer in the world [2]. HCC mortality ranks third among all cancers worldwide and causes 600,000 deaths per year [2]. Metastasis is the major risk factor for long-term survival of patients with post-hepatectomy HCC, and it contributes to the high recurrence rate of HCC [3,4]. Lymph node metastasis (LNM) occurs in approximately 10.0 % of HCC patients during the follow-up period after hepatectomy [5], whereas the LNM rate is estimated as 33.8 % in HCC patients with extrahepatic metastasis [6]. Although LNM of HCC is reported to be sensitive to external beam radiotherapy [7], the prognosis of LNM patients is worse than that of patients without LNM [8]. Therefore, it is important to establish a predictive model and identify molecular markers for LNM to accurately assess the LNM risk in HCC patients. The factors involved in LNM of HCC are poorly documented, especially microRNAs (miRNAs).
The miRNAs responsible for LNM in HCC are largely unknown. If miRNA expression profiling could be used to identify HCC patients at high risk for LNM, the high-risk population could receive prophylactic radiotherapy for regional lymph nodes. So, it might reduce LNM incidence and prolong overall survival time for patients. In this study, we established tissue microarrays (TMAs) and performed in situ hybridization (ISH) to discover novel and annotated miRNAs, which are associated with LNM in HCC patients. We identified significant miRNA signatures and developed a novel miRNA-based prediction model to identify HCC patients at high risk for LNM. Figure 1 shows that miR-31 and miR-92a expression was identified in the cytoplasm and in the nucleus, whereas miR-145 and miR-10b expression was localized primarily to the cytoplasm of tumor cells. In the training cohort of 192 patients, positive miR-145 expression was detected in 47 patients (24.5 %), positive miR-31 expression was detected in 35 patients (18.2 %), positive miR-92a expression was detected in 55 patients (28.6 %), and high miR-10b expression was detected in 59 patients (30.7 %). In the validation cohort of 209 patients, positive miR-145 expression was detected in 54 patients (25.8 %), positive miR-31 expression was detected in 39 patients (18.7 %), positive miR-92a expression was detected in 66 patients (31.6 %), and high miR-10b expression was detected in 65 patients (31.1 %).

Significant predictors of lymph node metastasis
Detailed characteristics of all patients analyzed in this study are presented in Table 1. The following 19 clinicopathological factors were analyzed in the training cohort: age, gender, HBsAg, hepatitis C virus antibody (HCV-Ab), alpha-fetoprotein (AFP), alanine aminotransferase (ALT), gamma-glutamyl transferase (γ-GT), liver cirrhosis, Child-Pugh score, tumor differentiation, tumor size, tumor number, tumor encapsulation, vascular invasion, Barcelona Clinic Liver Cancer (BCLC) stage, miR-145, miR-31, miR-92a, and miR-10b. Table 2 summarizes the association of these clinicopathological factors with the LNM status of HCC patients in the training cohort, which were determined by univariate analysis. These significant variables were adopted for multivariate analysis, which indicated that the following five independent variables could significantly predict the LNM risk in HCC: vascular invasion (P = 0.001), BCLC stage (P = 0.018), miR-145 (P = 0.014), miR-31 (P = 0.006), and miR-92a (P = 0.034) ( Table 3). We also identified miR-145, miR-31 and miR-92a were correlated with vascular invasion and BCLC stage (Supplementary  Tables S1 and S2).

Constructing the miRNA-based prediction model of lymph node metastasis in hepatocellular carcinoma
The miRNA-based prediction model was constructed as follows using multivariate analysis. Each chi-square (χ 2 ) value was divided by the minimum multivariate analysis χ 2 value of 4.496 to obtain a simple risk score for each significant variable according to its relative contribution to the multivariate analysis model ( Table 4). The total risk score of every patient was the aggregate of all simple risk scores, ranging from 0 to 8 in both the training and the validation cohorts. The cutoff point of 4 for both cohorts was the best determinant to discriminate patient categories for low-risk and high-risk LNM groups using the χ 2 test for linear trend.
Using the score of 4 as a cutoff point in the training cohort of 192 patients, 176 (91.7 %) and 16 (8.3 %) patients were in the low-risk and high-risk groups, respectively. In the low-risk and high-risk groups of the training cohort, 10/176 (5.7 %) and 13/16 (81.3 %) patients developed LNM (P < 0.05). The three-miRNA-based prognostic model sensitivity was 73.9 % and specificity was 79.4 % over 5 years and the area under the receiver operating characteristic (ROC) curve (AUC) was 0.906 (95 % CI, 0.842-0.969; P < 0.001), which indicates good reliability and validity (Supplementary Figure S1A). The 1-and 2-year cumulative LNM rates in the high-risk group were 78.1 and 85.4 %, respectively, whereas these values were 3.7 and 4.8 %, respectively, in the low-risk group. Patients in the high-risk group were found to have higher LNM rates (Supplementary Figure S1B).The 5-year positive and negative predictive values of the LNM prediction model were 32.8 and 95.7 %, respectively. Cox regression analysis determined that the LNM hazard ratio for high-risk versus low-risk groups was 32.071 (95 % CI, 12.599-81.636; P < 0.001).

Validating the prognostic value of the miRNAbased model
The prognostic value of the miRNA-based model was validated using an independent validation cohort of 209 patients and the cutoff point of 4. The results indicated that 190/209 (90.9 %) and 19/209 (9.1 %) patients were categorized as low risk and high risk, respectively, in the validation cohort. In the low-risk and high-risk groups of the validation cohort, 11/190 (5.8 %) and 12/19 (63.2 %) patients developed LNM, respectively. Analysis of the validation cohort by ROC curve proved that the miRNA prediction model could predict LNM in HCC patients, with very high AUC of 0.860 (95 % CI, 0.773-0.948; P < 0.001) (Supplementary Figure S1C). The 5-year sensitivity and specificity of the prediction model were 69.6 and 80.2 %, respectively. The positive and negative values of the cutoff score for predicting LNM in HCC were 30.3 and 95.5 %, respectively, over 5 years.
Kaplan-Meier and log-rank tests were applied for analysis of time-to-LNM. The 1-and 2-year cumulative LNM rates in the high-risk category were 38.6 and 71.2 %,  respectively, whereas these values were 4.5 and 11.7 %, respectively, in the low-risk category. Patients in the highrisk group had a higher risk for developing LNM than patients in the low-risk group (Supplementary Figure  S1D). The Cox regression analysis hazard ratio for developing LNM of the high-risk versus low-risk groups was 11.751 (95 % CI, 5.110-27.021; P < 0.001).

DISCUSSION
The incidence of LNM in extrahepatic metastases of HCC is 33.8 % [6], which contributes to the high mortality and poor prognosis of patients with HCC. However, an accurate model that can predict the LNM risk for patients with HCC has not been available. The LNM status depends primarily on LN biopsy, which may increase distant metastatic rates and reduce survival [21]. PET and CT analyses have low sensitivity for detecting tumorpositive small LNs <1 cm [22]. Therefore, the discovery of molecular markers to accurately assess the risk of LN involvement is of great importance. Aberrant forms of miRNAs have been reported in many cancer types, and considerable attention is focused on understanding the role of miRNAs in cancer development [23][24][25]. MiRNAs appear to have potential as molecular markers that can be developed to assess the LNM risk in many cancers [26][27][28][29]. In the current study, we sought to identify novel miRNAs that predict LNM in HCC.
We constructed and then validated a novel tool based on three miRNAs to predict LNM development for patients with HCC after hepatectomy. Our miRNA prediction model incorporates the following five factors: vascular invasion, BCLC stage, miR-145, miR-31, and miR-92a. Vascular invasion and BCLC stage are known prognostic factors for HCC [30,31]. MiR-145 is known to promote LNM in colorectal cancer (CRC) [32]. The increased expression of heat shock protein 27 (Hsp-27), an up-regulated target of miR-145, is responsible for LNM in   [32]. MiR-31 promotes LNM in lung ADC through the activation of ERK1/2 signaling [10]. MiR-92a targeted suppressed E-Cadherin (CDH1) expression to promote LNM in esophageal squamous cell carcinoma [33]. And we are going to investigate the molecular mechanism of abnormal miRNAs expression associated with LNM in HCC by a large prospective research. It may have a little limitation for that some potential miRNAs may be not included in the study. We are going to screen significant miRNAs on a large scale through miRNA microarray in future research to enrich and perfect the microRNA-based prediction model for LNM in HCC.
To our knowledge, the significance of such a prediction model has not been unveiled in HCC, especially with respect to LNM. The miRNA-based prediction model effectively and accurately classifies post-hepatectomy HCC patients into groups with low risk and high risk of developing LNM. This model could be implemented for routine clinical use, and serve as a valuable tool for determining optimal treatment strategies for HCC patients. Traditional approaches include palliative radiotherapy after LNM. However, our model can identify high-risk HCC patients in advance, just after hepatectomy, which may help to guide the therapeutic strategy. These highrisk patients can be treated with prophylactic radiotherapy for regional lymph nodes during the earliest stages before LNM develops. This strategy could replace the current passive palliative treatment with active preventative treatment to suppress LNM development. Thus, the LNM incidence in HCC can be reduced, and the life quality and survival of patients can be improved. This predictive model also may provide new therapeutic targets and facilitate decisions regarding individualized clinical therapies.
In summary, the results confirm that our miRNAbased model is a good prognostic system to predict LNM in

Ethics statement
Investigation had been approved by the Ethical Review Board of Zhongshan Hospital, Fudan University. Informed consent had been obtained from all participants included in this study.

Patients and tissue specimens
All patients with HCC included in this study were diagnosed based on pathology. Patients with a history of other solid tumors were excluded from the study. None of the patients had distant metastasis before surgery, nor had they received anticancer therapy before surgery. All patients received chest radiography and abdominal ultrasonography examinations before surgery, and patients with extrahepatic metastasis were excluded from the study. Bone scanning was performed if bone pain was reported. If extrahepatic metastasis was suspected, computed tomography (CT) and/or magnetic resonance imaging (MRI) was performed to verify the occurrence of extrahepatic metastasis. Suitable FFPE tissue samples, complete clinicopathological examinations, and followup data were available for all patients. The tumor stage was defined in accordance with the BCLC staging system (2010 version). The histological grade of tumor differentiation was determined by the Edmondson grading system. Liver function was classified by the Child-Pugh scoring system. Tumor size depended on the maximum diameter of the tumor specimen. The extent of vascular invasion was identified by microscopic examination of the resected specimen.
The retrospective study was performed in two independent cohorts. The training cohort consisted of 192 consecutive HCC patients who underwent hepatectomy (all performed by the same surgical team) at the Liver Cancer Institute, Fudan University, from October 1999 to January 2006. The training cohort was used to develop a predictive model for LNM in HCC. The validation cohort consisted of 209 patients with HCC, who were recruited from August 2000 to May 2006. All patients in the validation cohort underwent hepatectomy by a different surgical team at the same institution. Variables were adopted for their prognostic significance on univariate analysis. BCLC stage Barcelona Clinic Liver Cancer stage, CI confidence interval, * significance values.

Follow-up assessments
All patients received follow-up evaluations every 3 months after hepatectomy; each follow-up evaluation performed physical examinations and collected history documentation. Each follow-up examination included ultrasonographic examination of the liver and abdominal LNs; laboratory tests including liver function, AFP, ALT, and γ-GT; and evaluation of hematologic parameters. These examinations were performed by doctors who were blind to the study. Chest radiography was performed at 6-month intervals and a bone scan was performed annually. CT scanning or MRI was performed immediately when LNM was suspected. Metastatic LNs were indicated by one of the following: hypoechoic masses on ultrasonography, a central hypodensity region bordered by a faint hyperdensity rim on contrast CT scanning, or high signal intensity on T2-weighted MRI [5]. When a diagnosis of LNM was confirmed, the metastatic LNs received external beam radiotherapy [7]. Other recurrent foci were treated with radiotherapy, interventional therapy, or surgery.

Tissue microarray construction and analysis
TMAs were constructed as described previously [13]. We constructed TMA slides (in collaboration with Biochip Company, Ltd., Shanghai, China) using HCC samples from 192 consecutive HCC patients in the training cohort and 209 consecutive HCC patients in the validation cohort. Slides stained with hematoxylin and eosin was screened to identify optimal intratumoral tumor tissue for analysis. Two cores with dimensions of 1.0 mm were punched from non-necrotic areas of tumor foci in the patient paraffin blocks. Sections (4-μm thickness) of the resulting TMA blocks were made using standard techniques.

Extraction of miRNAs from FFPE tissue and validation by real time qRT-PCR
FFPE specimens stained with hematoxylin and eosin were identified optimal intratumoral tissue of 192 hepatocellular carcinoma (HCC) patients. Total RNA including miRNA was extracted from FFPE samples using miRNeasy FFPE kit (Qiagen) according to manufacturer's instructions. Briefly, the samples were deparaffinsed and lysed with proteinase K digestion. After a short incubation at a higher temperature, DNase treatment was performed to eliminate genomic DNA, including very small DNA fragments which were often present in FFPE samples after long-term fixation and storage. Followed by the addition of buffer RBC and 100% ethanol, then the sample was applied to RNeasy MinElute spin columns. Finally, samples were washed and eluted in RNase free water. RNA concentrations were measured via absorbance spectrophotometry on a NanoVue Plus instrument (GE, General Electric Company).
The expression levels of the candidate miRNAs were validated by real time qRT-PCR. Primers for U6, miR-145, miR-31, miR-92a, and miR-10b were purchased from GeneCopoeia. Real time qRT-PCR was performed by the SYBR Green PCR method using All-in-One™ miR qRT-PCR detection kit (GeneCopoeia, China) with miRNA specific primers. All samples were run in triplicate, and the 2 −ΔΔCt method was used to analyse the miRNA expression levels. The qRT-PCR was performed on a 7500 Real-Time PCR System (Applied Biosystems). U6 snRNA was used as the endogenous control.

Statistical analysis
Correlations between clinicopathological features and LNM status were analyzed by Pearson's χ 2 test and Fisher's exact test. Then, univariate analysis and multivariate analysis were performed to determine statistically significant variables and devise a simple risk score, which was the corresponding estimated coefficient divided by the minimum χ 2 value [18]. The final score of every patient was the aggregate of all risk scores. The optimal cutoff point was determined by the χ 2 test for linear trend for the best separation of LNM risk [19,20]. Prediction performance of the miRNA-based model was evaluated using AUC. The temporal interval from the date of surgery to the date of LNM incidence was considered as time-to-LNM, which was analyzed by Kaplan-Meier and log-rank tests. Cox regression analysis was used to estimate relative LNM risk of the high-risk versus low-risk groups. Two-sided P values < 0.05 were considered as statistically significant. All statistical analyses were conducted with SPSS version 20.0 (SPSS IBM).