Microarray data of GSE76427 were downloaded from the Gene Expression Omnibus (https://www.ncbi.nlm.nih.gov/geo/) database. GSE764267 was conducted by GPL10558 (Illumina HumanHT-12 V4.0 expression beadchip). It contained 115 HCC tissue samples and 52 adjacent non-tumor tissue samples. Out of 115 tumor samples, 108 with complete follow-up information (recurrence status and recurrence-free survival [RFS]) were included in the discovery dataset. All 108 participants underwent curative resection for HCC. Among the 108 participants, 22 were female and the other 86 were male. The average age of all the participants in the discovery dataset was 63.4 years old. With respect to HCC stage, 86 of 108 were in stage I or II, 21 in stage III/IV, and only 1 patient had no record of cancer stage. The median follow-up was 1.17 years. HCC recurrence was diagnosed according to the established criteria reached by International Working Party[16]. Each array from GPL10558 consisted of more than 47,000 probes corresponding to more than 31,000 annotated genes including coding and non-coding genes, microRNAs, rRNAs, and other short RNAs. We extracted all the long non-coding RNAs from GPL10558 for the preliminary screening of prognostic lncRNAs. Three hundred and thirty-seven HCC patients with recurrence information from The Cancer Genome Atlas (TCGA, http://cancergenome.nih.gov/) constituted the validation dataset. The lncRNA expression profiles, clinical characteristics, follow-up data, and genetic mutation information of the TCGA cohort were downloaded. Besides, the RNA-Seq data of the 49 paired non-tumor tissue samples from the TCGA database were also obtained. The median RFS of GSE76427 and TCGA series was 8.4 and 13.0 months, respectively.

Univariate COX regression analysis was performed to screen the prognostic lncRNAs. Then, LASSO was applied to the construction of an HCC prognostic signature with the screened lncRNAs[15]. LASSO statistical algorithm was conducted using “glmnet” package in the R software (version 3.4.0, https://www.r-project.org/)[17]. Based on the expression levels of each sample, LASSO identified the eligible lncRNAs for the risk system and generated the corresponding coefficients for each of them.

The risk scores of each sample from the discovery and validation groups were calculated according to the risk model. The respective medians of two groups were used as the cut-off value to divide patients into high-risk and low-risk groups. Kaplan-Meier curves were plotted to compare the RFS of high-risk and low-risk patients. Meanwhile, P-values and hazard ratio (HR) with 95% confidence interval (CI) were generated by log-rank tests. Stratified survival analysis was carried out to identify the best candidates for the prognostic signature. The overall group was divided into subgroups by their clinical characteristics. Kaplan-Meier analysis was performed in the subgroups using the same cut-off value as the overall group. Kaplan-Meier curves were plotted using GraphPad Prism software (version 7.0).

Expression levels of the identified lncRNAs in tumor and non-tumor tissues were compared using TCGA RNA-seq data. The receiver operating characteristic (ROC) curve was plotted with GraphPad Prism (version 7.0). The area under the ROC curve (AUC) for evaluating discriminatory ability was calculated as well. Besides, the distribution of high-risk or low-risk patients in early- and late-stage subgroups were also compared via Chi-square test.

Pearson correlation analyses were conducted between the identified lncRNAs and the protein-coding genes in TCGA dataset based on their expression levels. The correlation coefficient > 0.4 and P < 0.001 were considered significantly correlated. The significantly co-expressed mRNAs were thrown into a publicly available web tool, Enrichr, for BioCarta (http://cgap.nci.nih.gov/Pathways/BioCarta_Pathways) pathway enrichment[18]. The enriched BioCarta terms were sorted by rank based ranking, an algorithm assessing the deviation from the expected rank to the mean rank[19].

Univariate and multivariate COX regression analyses were carried out in TCGA dataset to identify the independent risk factors for the RFS of HCC patients. A composite nomogram predicting the RFS of HCC was established based on the independent factors using the “rms” package of R statistical software. The concordance index (C-index) was calculated to evaluate the discriminatory ability of the nomogram. And calibration curves were plotted to compare the predicted and actual probabilities of RFS. Each component of the nomogram gives points and the sum of them represents the total points a patient receives. All the participants were divided into different risk groups according to their total points. Kaplan-Meier analysis was utilized to compare the RFS of different risk groups. Statistical analyses were performed with SPSS 23.0 (SPSS, Chicago, IL), unless otherwise indicated. A P-value < 0.05 was considered statistically significant.

The flow chart of the study procedure is presented in Figure 1. All the lncRNAs in the discovery dataset (GES76427) were subjected to univariate COX analysis and those significantly associated with RFS (P < 0.05) were considered as prognostic ones for LASSO modelling. The risk score formula for RFS was calculated as follows: risk score = 0.021355462 × (expression value of MSC-AS1) + 0.018051929 × (expression value of POLR2J4) + 0.016385849 × (expression value of EIF3J-AS1) + 0.01340867 × (expression value of SERHL) + 0.012263937 × (expression value of RMST) + 0.007303891 × (expression value of PVT1). From the formula, it is seen that these lncRNAs were all risk factors for HCC recurrence (coefficient > 0). And the value of their respective coefficients represented how much impact they had on the RFS prediction. It is obvious that MSC-AS1 had the most while PVT1 had the least impact. The risk model generates a risk score for each participant. Using the median of the risk scores of the whole discovery group, 9.100, as the cut-off value, 108 patients were classified as high-risk or low-risk ones (Figure 2A). The recurrence status, RFS period, and six lncRNAs’ expression value of each patient are presented in Figure 2A as well. Kaplan-Meier curves showed that the low-risk group had significantly longer RFS than the high-risk group with a P-value from log-rank test of 0.024 (HR = 1.842, 95%CI: 1.026-3.309) (Figure 2B).

Open in New Tab   Full Size Figure   Download Figure

Figure 1 Overall design of the present study. HCC: Hepatocellular carcinoma; RFS: Recurrence-free survival.

Open in New Tab   Full Size Figure   Download Figure

Figure 2 Construction and validation of a prognostic lncRNA signature for hepatocellular carcinoma. A: LncRNA risk score distribution (Upper), the recurrence status and recurrence-free survival (RFS) period (Middle), and expression profiles of the six lncRNAs (Lower) of the 108 patients in the discovery dataset. B: The Kaplan-Meier curve of the RFS between the high-risk and low-risk groups stratified by the median risk score in GSE76427 series. C: The Kaplan-Meier curve of the RFS between the high-risk and low-risk groups stratified by the median risk score in The Cancer Genome Atlas cohort.

To confirm the predictive ability of the lncRNA-signature, validation analysis was carried out in a group of 337 HCC patients from the TCGA project. The whole validation group was divided into high-risk and low-risk groups accordingly in the discovery dataset. The median score (0.0357) of the whole validation group was adopted as the cut-off value. Survival analysis showed great performance of the risk model in stratifying high-risk and low-risk patients with a log-rank P-value being less than 0.001 (HR = 1.807, 95%CI: 1.329-2.457) (Figure 2C).

Stratified survival analysis in the validation series was conducted to further investigate the suitable patient group of the six-lncRNA signature. The cut-off value in the subgroups was consistent with that in the overall group (0.0357). The result of stratification analysis was shown in Table 1. Our risk score system was more applicable to the patients possessing the following characteristics: TNM stage I/II, male gender, younger than 60 years, Asian, with family history, hepatitis B virus (HBV) infection, alcohol consumption, ECOG = 0, and higher levels of preoperative serum albumin (ALB, > 3.5g/dL) and alpha-fetoprotein (AFP, > 20 ng/mL).

To investigate the expression profile of the identified lncRNAs, we compared the expression values of the six lncRNAs between tumor and non-tumor samples in TCGA dataset. The results showed that all six lncRNAs were significantly upregulated in HCC presenting with remarkably significant P-values (P < 1.0 × 10^-10) (Figure 3A-F). In addition, the ROC curve showed the great utility of the combined six-lncRNA signature in discriminating tumor from non-tumor tissues. The AUC was 0.932 (95%CI: 0.898-0.966) (Figure 3G). Furthermore, the distribution of high-risk and low-risk patients in different stages was also examined. Patients in late-stage (TNM stage III/IV) had a higher likelihood of being high-risk patients than those in early-stage (TNM stage I/II) (P < 0.05), implying that higher risk score was associated with relatively late HCC stage (Figure 4H).

Open in New Tab   Full Size Figure   Download Figure

Figure 3 Expression patterns of the six lncRNAs in tumor and non-tumor tissues. Differential expression of RMST (A), EIF3J-AS1 (B), SERHL (C), PVT1 (D), MSC-AS1 (E), and POLR2J4 (F) between hepatocellular carcinoma and non-cancerous samples. G: The ROC curve of tumor tissue vs non-tumor tissue discriminated by the six-lncRNA signature. (H) Comparisons of the distribution of high-risk and low-risk patients in early stage (TNM I/II) and late stage (TNM III/IV) by the chi-square test.

Open in New Tab   Full Size Figure   Download Figure

Figure 4 Functional annotation of the six lncRNAs. Top 10 significantly enriched BioCarta pathways using the co-expressed mRNAs of the six lncRNAs in The Cancer Genome Atlas database.

To further investigate the potential biological roles of the identified six lncRNAs, BioCarta pathways were enriched using the co-expressed protein-coding genes of these lncRNAs. A gene significantly correlating with at least one of the six lncRNAs (Pearson coefficient > 0.4 and P < 0.001) was considered eligible for pathway enrichment. Top ten highly enriched pathways are shown in Figure 4. These co-expressed genes of the lncRNAs clustered most significantly in TGF-β signaling pathway, internal ribosome entry pathway, granzyme A mediated apoptosis, FAS signaling pathway, calcium signaling by HBx, p38/MAPK signaling pathway, etc. Most of them are classical and vital pathways involved in HCC initiation and progression.

To develop a composite predictor for the RFS of HCC patients, we combined the six-lncRNA signature, clinicopathological characteristics, and TP53 mutation status together for the screening of the independent factors for RFS. The results from univariate and multivariate COX regression analyses showed that the identified independent risk factors for RFS included the six-lncRNA score, TNM stage, and ECOG [Eastern Cooperative Oncology Group] score (P < 0.05) (Table 2). The nomogram for RFS prediction was comprised of the three factors (Figure 5A). The C-index of the nomogram was 0.684 (95%CI: 0.635-0.733). The calibration curves for the probability of recurrence at 1 year and 3 years showed good agreement between the prediction from the nomogram and the actual observations (Figure 5B). Each patient got the total points according to the scoring of the nomogram. The tertiles of all the total points were used as the cut-off value (6.800 and 3.200) to divide the patients into high-, intermediate- and low-risk groups. The Kaplan-Meier analysis of the three risk subgroups indicated the great utility of the composite nomogram in discriminating HCC patients with good, intermediate, and poor prognosis (Figure 5C).

Open in New Tab   Full Size Figure   Download Figure

Figure 5 Construction of a nomogram for recurrence-free survival prediction in hepatocellular carcinoma. A: The composite nomogram consists of the six-lncRNA score, TNM stage, and ECOG score. Each component generates their respective points according to the “Points” line drawn above. Add the points from three variables together and find the location of the total points on “Total Points” line. Then draw a vertical line from “Total Points” axis to the two lower lines which corresponds to the predicted 1-year and 3-yr recurrence-free survival (RFS) rates by the nomogram. B: Calibration curves of the nomogram for the estimation of RFS rates at 1-year (Left) and 3-years (Right). The predicted and actual 1-yr and 3-yr RFS-probabilities were drawn on the x and y axis, respectively. C: The Kaplan-Meier curve of three risk subgroups stratified by the tertiles of total points derived from the nomogram.

Hepatocellular carcinoma (HCC) is the most common type of liver cancer which remains a severe health issue worldwide. In recent years, genetic markers and predictive models have been put forward for improving the management of HCC. Meanwhile, many statistical techniques have been used for data mining in a series of large public databases involving the high-throughput genetic data of cancers. With the help of the most advanced clinic-practical methods, more accurate and robust prognostic models can be constructed for HCC.

Researchers have tried to constitute a prognostic model based on molecular biomarkers for HCC over these years. Long non-coding RNAs (lncRNAs) are novel predictive indicators. Although a few attempts have been made to construct lncRNA-based models for HCC, more are needed for further really significant findings.

By analyzing data from two databases, Gene Expression Omnibus (GEO) and The Cancer Genome Atlas (TCGA), we wanted to identify a prognostic signature for HCC which is comprised of the potential functional lncRNAs.

The latest statistical algorithm, the least absolute shrinkage and selection operator (LASSO), was utilized to constitute our predictive model. This method was performed based on the significant lncRNAs screened based on the lncRNA expression profiles from the GEO database. The expression values of the candidate lncRNAs were also examined in the HCC and normal liver tissues. The robustness of this model was validated using TCGA dataset. The suitable patients and other clinical applicability of the lncRNA-signature were explored as well.

The risk score system for predicting the recurrence of HCC was constructed based on the six lncRNAs (MSC-AS1, POLR2J4, EIF3J-AS1, SERHL, RMST, and PVT1) using LASSO. All six lncRNAs were aberrantly expressed in HCC and non-tumor tissue and they were significantly enriched in TGF-β signaling pathway and cellular apoptosis-related pathways. The best candidates we identified were younger early-staged male patients with HBV infection and family history in better physical condition but with higher preoperative AFP. To broaden the application scope of the model, a nomogram involving the lncRNA signature and other clinicopathological characteristics was formulated.

The six-lncRNA signature showed great predictive ability in prognostic evaluation of HCC patients. This tool may help perform risk stratification and provide more individualized clinical advice for each patient.

Our study offered extra evidence that lncRNAs are potential functional regulators in HCC progression. Finding effective molecular biomarkers and predictive signatures of HCC prognosis are future direction calling urgently for groundbreaking attempts.