Identification and characterization of a 25-lncRNA prognostic signature for early recurrence in hepatocellular carcinoma

Gene expression profile of HCC and corresponding clinical information were downloaded from TCGA-LIHC (http://​cancergenome.​nih.​gov/​). Total 314 out of all 374 HCC samples with complete follow-up information (overall survival (OS) time, OS status, disease free survival (DFS) time and status) were retained. Among these 314 patients, some patients’ follow-up time was less than 1 month, and their OS and DFS status were labeled as “alive” and “recurrence free”. Therefore, these patients were not suitable for early recurrence analysis and they were excluded. Thus, we used 299 patients for signature construction in this study. The 299 HCC patients were then randomly divided into a training cohort (N = 150) and a validation cohort (N = 149). Based on the information of annotated lncRNAs in GENCODE V30, 14,847 human lncRNAs with Ensembl gene ID were obtained and their corresponding expression profile was extracted from the TCGA-LIHC.

Most bioinformatics analyses were conducted using R software. DEG analysis was performed between the 150 HCC samples in the training cohort and 50 normal tissue samples from TCGA-LIHC project by using R package “edgeR” [28, 29]. Univariate Cox regression analysis was performed to select early recurrence related lncRNAs by using R package “survival” [30]. FunRich (version 3) was used to draw Venn diagram between differentially expressed lncRNAs and early recurrence related lncRNAs to obtain candidate lncRNAs for signature construction [31]. Candidate lncRNAs were then further analyzed in LASSO regression analysis by running R package “glmnet” for 1000 times [32], and the most powerful prognostic lncRNAs were selected through 10-fold cross-validation with lambda.min as the optimized cut-off [33]. Risk score of each patient was calculated in a linear combination of lncRNAs weighted by their corresponding regression coefficients and expression levels in indicated HCC patients by formula (risk score =  ∑ coefficient × expression(gene)). Receiver operating characteristic curve (ROC) analysis was conducted by using R package “pROC” [34], and the predictive performance was assessed by calculating the area under curve (AUC). Finally, a combination of 25 lncRNAs was chosen for establishing risk signature because this 25-lcnRNA risk signature gave the largest AUC in ROC analysis. The 150 HCC patients were divided into the low-risk group (N = 75) and the high-risk group (N = 75) by using the median risk score as cut-off. A correlation analysis was performed between the risk score and early recurrence. Kaplan-Meier analysis, cumulative hazard and cumulative events analyses were conducted by using R package “survival” in the training cohort, the validation cohort and the total 299 HCC patients to investigate the early recurrence survival between low risk patients and high risk patients. Univariate and multivariate Cox analysis were done in the total 299 HCC patients with R package “survival” to evaluate whether the risk score could serve as an independent factor for early recurrence prediction in HCCs. Nomogram was constructed by using the 25-lncRNA signature, AFP, vascular invasion, TNM stages and their corresponding multivariate Cox regression coefficients, and calibration plots were generated with R package “regplot” [35]. C-index was used to evaluate the model performance for predicting early recurrence.

GSEA was conducted by using GSEA JAVA program (version 4.0.3) downloaded from official website (http://​software.​broadinstitute.​org/​gsea/​index.​jsp) to find out enriched gene sets. MsigDB h.all.v7.1.symbols.gmt gene set collection was chosen for identifying hallmarks of HCC early recurrence. The random sample permutations were set to be 1000 with the significance set as |NES| > 1, FDR q < 0.25 and nominal P < 0.05.

Immune infiltration analysis was performed with single sample Gene Set Enrichment Analysis (ssGSEA) by using “GSVA” package in R [36]. A group of 28 cellmarker sets were used for calculating normalized enrichment score (NES) for each cell type in every 299 HCC samples [37]. Correlation analysis between risk scores and NES of immune cells was performed by function “cor.test” in R. TIDE (Tumor Immune Dysfunction and Exclusion) algorithm and SubMap modules from GenePattern were used to predict the response to immune checkpoint blockade for all 299 HCC samples [38‐40].

Chemotherapeutic response prediction for every 299 HCC samples was conducted in R by using “pRRophetic” package based on the Genomics of Drug Sensitivity in Cancer (GDSC) pharmacogenomics database. The half maximal inhibitory concentration (IC₅₀) was estimated by ridge regression and the prediction accuracy was evaluated by 10-fold cross-validation [41].

To validate the 25-lncRNA signature in clinical samples, 3 lncRNAs from the signature were selected and their relative expressions in HCC samples were detected by RT-qPCR. Total RNA from 36 paired HCC tumor and adjacent tissues provided by Xinhua Hospital were extracted by using TRIzol (Invitrogen, 15596026) according to the manufacturer’s instructions. cDNA was synthesized by using ReverTra Ace® qPCR RT Master Mix with gDNA Remover (TOYOBO, FSQ-301) in a SimpliAmp Thermal Cycler (Applied Biosystems). The 20 μL PCR reaction system consist of 2 μL cDNA, 0.8 μL forward primer, 0.8 μL reverse primer, 10 μL CYBR Premix Ex TaqII, 0.4 μL ROX Reference Dye II and 6 μL deionized water (Takara CYBR Premix Ex TaqII, RR820A). RT-PCR was performed in ABI Biosystems™ 7500 Real-Time qPCR System (Applied Biosystems). 18s was used as a housekeeping gene for normalization and the relative expression of selected genes was calculated by using 2^−ΔΔCT method. Primers used were synthesized by GENEWIZ and the sequences of primers were ENSG00000231918 (GTGGCTCTGCCTTGGGTAAT, TTCCAGAACAACCTTGTCAGA), ENSG00000248596 (GCCAGAATTGGCGGTTTCTC, ATCGCTGAGTGTGTCGAGTG), and ENSG00000223392 (ATCCTTACCCTGCATTGCCC, ATGATCCAACCATCTGCAGGG).

DeLong’s test was used to compare the sensitivity and specificity of two ROC curves. Chi-square test was used to evaluate the impact of risk score group distribution on recurrence cases between 1 year and 2 years. The correlation of risk scores with disease free survival (DFS), NES of tumor-infiltrating lymphocytes and levels of immune checkpoints was analyzed by nonparametric Spearman’s rank correlation analysis. The log-rank test was used for Kaplan-Meier survival analyses, cumulative hazard and cumulative events analyses. The Cox proportional hazards regression model was used for univariate and multivariate analyses. Wilcoxon test was used for comparing NES of immune cells and IC₅₀ of drugs between the low-risk and high-risk group. The difference was considered statistically significant when P < 0.05 in all statistical analysis.

HCC RNA-seq data and corresponding clinical information were downloaded from the TCGA-LIHC (Liver Hepatocellular Carcinoma) project. After removal of samples without complete survival information, total 299 out of all 374 HCC samples were enrolled in this study for further analysis. Table S1 shows the clinical characteristics of the 299 HCC samples, in which more than 50% HCC patients had recurrence. Because there are no suitable GEO datasets which are comparable to the TCGA-LIHC project containing comprehensive data on both lncRNA expression profile and patients’ clinical characteristics, we then randomly divided the 299 HCC patients into a training cohort (n = 150) and a validation cohort (n = 149) by using “split” function in R software instead of setting an external validation cohort. Bioinformatics analyses were first performed in the training cohort and further validated in the validation cohort (Fig. 1A).

To establish a lncRNA-based risk signature, differentially expressed gene (DEG) analysis on lncRNAs was performed between the training cohort and 50 normal controls from the TCGA-LIHC project. A total of 1495 DEG lncRNAs were found significantly dysregulated in HCC samples (1159 up-regulated and 336 down-regulated, log2|FC| > 1, FDR < 0.05) (Fig. 1B and C). Meanwhile, a univariate Cox regression analysis revealed that total 1973 lncRNAs were associated with HCC early recurrence (ER lncRNAs) (P < 0.05) (Fig. 1D). Finally, a Venn diagram between the 1495 DEG lncRNAs and the 1973 ER lncRNAs identified 358 lncRNA candidates which may have potential prognostic value for HCC early recurrence (Fig. 1D).

The least absolute shrinkage and selection operator (LASSO) Logistic Regression is a selection and shrinkage technique designed for regression model initially applied to Ordinary Logistic Regression [42]. LASSO can better identify those risk factors strongly linked to the outcome and is widely employed in signature construction [43]. To identify key lncRNAs suitable for establishing a risk signature for predicting HCC early recurrence, those 358 candidate lncRNAs (Fig. 1D) were further analyzed in LASSO regression. A total of 1000 LASSO regression iterations were performed by using the R package “glmnet”. Lambda.min was chosen as the optimized cut-off to select key lncRNAs for risk model (Fig. S1) [44]. Consequently, 7 lncRNA combinations were obtained after LASSO analysis (Table S2 and Fig. S1). Thus, 7 lncRNA risk signatures were individually constructed based on these combinations. The risk score of each HCC patient was calculated in a linear formula risk score =  ∑ coefficient × expression(gene) (expression: lncRNA expression in individual HCC patients; coefficient: regression coefficients of indicated lncRNAs). To determine which lncRNA risk signature gives the best predictive performance on early recurrence, receiver operating characteristics (ROC) analysis was conducted between the 7 lncRNA risk signatures. As shown in Fig. 1E, all the 7 lncRNA risk signatures gave high area under the ROC curve (AUC, AUC > 80%), suggesting the reliability of our LASSO analysis. Among them, the 25-lncRNA risk signature gave the highest AUC (AUC = 86.70%) (Fig. 1F), suggesting the 25-lncRNA risk signature has the best predictive performance for HCC early recurrence.

Since the 25-lncRNA risk signature gave the best predictive performance, we then selected this signature to establish a risk model for HCC early recurrence. The detailed information of the 25 lncRNAs, including Ensembl gene ID, gene symbol, hazard ratio and coefficients, was summarized in Table 1. Among them, 19 lncRNAs (ENSG00000253417, ENSG00000272205, ENSG00000269894, ENSG00000275437, ENSG00000223392, ENSG00000248596, ENSG00000268201, ENSG00000247675, ENSG00000231918, ENSG00000234129, ENSG00000269974, ENSG00000236366, ENSG00000275223, ENSG00000253406, ENSG00000232079, ENSG00000255980, ENSG00000267905, ENSG00000176912, ENSG00000254333) had positive coefficients and were negatively associated with disease free survival (DFS), and the remainder 6 lncRNAs (ENSG00000259834, ENSG00000254887, ENSG00000259974, ENSG00000273837, ENSG00000231246, ENSG00000234283) had negative coefficients and were positively associated with DFS (Table 1). Here, we named those lncRNAs with positive coefficients as risk lncRNAs and those with negative coefficients as protective lncRNAs. The risk score could be calculated according to the coefficients of individual lncRNAs and their expression in corresponding HCC patients.

To determine whether the 25-lncRNA risk signature could predict HCC early recurrence, we first calculated the risk scores of the 150 HCC patients in the training cohort and then distributed them according to their risk scores from low to high (Fig. 2A). The median risk score was set as the cut-off to separate those patients into low-risk group (n = 75, patients’ risk scores < the median risk score) and high-risk group (n = 75, patients’ risk scores > the median risk score) (Fig. 2A). As shown in Fig. 2B, the 19 risk lncRNAs were mostly enriched in the high-risk group whereas the 6 protective lncRNAs were mainly enriched in the low-risk group. Moreover, 81.25% of recurrence cases in 1-year and 76.06% in 2-year came from the high-risk group, while the percentages were respectively 18.75 and 23.94% in the low-risk group (Fig. 2C). These results indicate that the 25-lncRNA risk signature have satisfying predictive potential for HCC early recurrence in the training cohort.

To validate the predictive potential of the 25-lncRNA risk signature, we evaluated it in the validation cohort. According to the median cut-off in the training cohort, the validation cohort (n = 149) was separated into the low-risk group (n = 69) and the high-risk group (n = 80) (Fig. 3A). In line with the finding in the training cohort, the risk lncRNAs were mainly enriched in the high-risk group whereas the protective ones were mainly enriched in the low-risk group (Fig. 3B). Meanwhile, 69.09% of 1-year recurrence cases and 63.01% of 2-year recurrence cases came from the high-risk group (Fig. 3C). Moreover, the predictive potential of the 25-lncRNA risk signature was also evaluated in the total 299 recruited HCC patients. Similarly, 299 HCC patients were separated into low-risk group (n = 144) and high-risk group (n = 155) according to the median cut-off in the training cohort (Fig. 3D). Most of the risk lncRNAs were enriched in the high-risk group and most of the protective lncRNAs were enriched in the low-risk group (Fig. 3E). Consistent with this finding, non-pair Wilcoxon test confirmed enrichment of the risk lncRNAs and the protective lncRNAs in the high-risk and low-risk groups respectively, except for the lncRNAs (ENSG00000255980, ENSG00000253406, ENSG00000232079, and ENSG00000234283) whose expression showed no significant changes between the low- and high-risk groups (Fig. S2). Patients in the high-risk group contributed 74.76% of 1-year recurrence cases and 69.44% of 2-year recurrence cases (Fig. 3F). More importantly, correlation assays showed significantly negative correlation of risk score with 1-year (Fig. 3G) or 2-year DFS (Fig. 3H) in the recurrent HCC patients in the high-risk group. No correlation was observed between risk score and DFS in recurrent HCC patients in the low-risk group (Fig. S3). These findings further validate the correlation of the 25-lncRNA risk signature with HCC early recurrence and indicate the great predictive potential of the risk signature on HCC early recurrence.

The primary purpose for the signature construction study is to accurately discriminate low- and high-risk patients. Therefore, the cut-off selection is critical for the accuracy of the prediction signature. In this study, we adopted the median risk score as cut-off which has been widely employed by many other groups [16, 45‐48]. To investigate whether there are other cut-offs which could distinguish low- and high-risk of early recurrence better than the median cut-off, we adopted the cut-off derived from Youden index [49]. Although the Youden index/cut-off could separate patients into the low- and high-risk groups (Fig. S4A-C), the prediction performance for early recurrence is much poorer than that by using median cut-off (Fig. S4D-F). Therefore, the median risk score used in this study is an appropriate cut-off to accurately distinguish HCC patients with low or high early recurrence risk.

To further investigate the prognostic value of the 25-lncRNA risk signature for early recurrence, we analyzed cumulative hazard and event in HCC patients. Both cumulative hazards and cumulative events were significantly higher in the high-risk group than those in the low-risk group in either the training cohort (Fig. S5A and B), validation cohort (Fig. S5C and D) or total 299 HCC patients (Fig. S5E and F). Meanwhile, Kaplan-Meier analyses, in the training cohort (Fig. 4A), validation cohort (Fig. 4B) and 299 enrolled HCC patients (Fig. 4C), showed that the patients in the high-risk group had lower 2-year DFS than those in the low-risk group. These findings further indicate the prognostic value of the 25-lncRNA risk signature for HCC early recurrence.

To determine whether the 25-lncRNA risk signature is an independent prognostic factor for HCC early recurrence, we performed univariate and multivariate Cox regression analyses in the enrolled 299 HCC patients. The 25-lncRNA risk score and other clinicopathological factors, including gender, age, race, cirrhosis, vascular invasion, serum AFP level and TNM stage, were used as covariates. As shown in Table 2, the vascular invasion, serum AFP and 25-lncRNA risk score were significantly associated with 1-year and 2-year recurrence in HCC patients, while the TNM stage was significantly associated with 2-year recurrence but not with 1-year recurrence. These findings are consistent with previous studies showing that serum AFP [50], TNM stage [51] and vascular invasion [2] are independent risk factors for HCC early recurrence, and indicate that the 25-lncRNA risk signature could serve as an independent prognostic factor for HCC early recurrence.

To investigate which independent risk factor gives the best predictive performance for HCC early recurrence, ROC analyses were performed by using “pROC”. As shown in Fig. 5A and B, the AUC of risk score for 1-year recurrence (73.86%) and 2-year recurrence (71.98%) were better those of AFP (64.58% for 1-year, 61.39% for 2-year recurrence), TNM (64.99% for 1-year, 67.17% for 2-year recurrence) and vascular invasion (VI) (63.47% for 1-year, 60.33% for 2-year recurrence). Moreover, compared to risk score alone, combining the risk score with AFP, TNM and VI further increased the predictive performance for 1-year recurrence (AUC: 78.79% vs. 73.86%) and 2-year recurrence (AUC: 76.82% vs. 71.98%) (Fig. 5C and D). The 95% confidence interval of AUC and C-index of above signatures were summarized in Table S3. An integrated Nomogram was further constructed by combining the 25-lncRNA signature, AFP, VI and TNM with a C-index 0.739 (Fig. 5E), and the calibration curves of the integrated nomogram for 1-year and 2-year DFS were presented in Fig. 5F. Therefore, the combination of the 25-lncRNA risk signature with AFP, TNM and VI could improve the prognosis evaluation for HCC early recurrence.

Previous studies have shown that lncRNAs function as key regulators of critical biological processes including cell differentiation, development, and apoptosis [52]. To investigate the biological processes associated with HCC early recurrence, gene set enrichment analysis (GSEA) was performed with hallmark pathways based on the gene expression profiling data from HCC patients in the high-risk and the low-risk groups. Eight gene sets were significantly enriched in the high-risk group while no significant gene set enrichment was observed in the low-risk group (|NES| > 1, FDR q-val < 0.25, NOM p-val < 0.05) (Table 3). Among them, the enrichment of gene sets of “E2F TARGETS”, “G2M CHECKPOINT”, “MYC TARGETS V1” and “DNA REPAIR” showed higher significance (|NES| > 1.5, FDR q-val < 0.10, NOM p-val < 0.01) (Table 3). The snapshots of enrichment results were displayed in Fig. 6 and the heatmaps for enriched gene sets were displayed in Fig. S6. These findings suggest that the 25 lncRNAs may affect HCC early recurrence through E2F, Myc, G2M and DNA repair pathways.

Tumor infiltrating lymphocytes (TILs) have been recognized as a prognostic factor in various types of cancers, and accumulation of TILs has been established as a positive prognostic factor in a number of solid cancers including melanoma [53], colon cancer [54] and ovarian cancer [55]. Previous studies have demonstrated that HCC patients with prominent TILs showed reduced recurrence and better prognosis compared with those without prominent TILs [56, 57]. Although the TILs are minority in the tumor bulk, the immune checkpoint molecules specifically express on T cells and antigen presenting cells but not tumor cells or other stromal cells in tumor bulk. Thus, it is a commonly accepted approach to evaluate TILs by using the expression levels of immune checkpoint molecules from bulk-tumor data [45, 58, 59]. To investigate whether the 25-lncRNA prognostic signature could reflect the levels of TILs, comparison of TILs was performed between the low- and high-risk groups. As shown in Fig. 7A, 22 out of 28 TILs showed significant enrichment in the low-risk group compared to the high-risk group (P < 0.05). Correlation analysis between risk scores and normalized enrichment scores (NES) of TILs revealed that the intratumor accumulation of 23 TILs was negatively associated with risk scores (P < 0.05, Fig. 7B). Among them, the type 1 T helper cell, effector memory CD8 T cell and activated CD8 T cell, which are well-known antitumor immune cells, ranked as the top 3 TILs negatively associated with the risk scores (|NES| > 0.4, Fig. 7C-E). These findings suggested that the 25-lncRNA prognostic signature may reflect the levels of TILs and predict the post-surgery prognosis in HCCs.

Since more TILs significantly enriched in the low-risk group patients, we attempted to further investigate whether the immunotherapies response are different in the low- and high-risk group. TIDE prediction suggest that there was no significantly difference in immunotherapies response between the low- (48.61%, 70/144) and high-risk (42.58%, 66/155) group (P = 0.297). However, by mapping the expression profile of the low- and high-risk group with a public dataset of 47 melanoma patients responded to immunotherapies in SubMap modules of GenePattern [60], the low-risk group showed prospective response to anti-PD-1 (programmed cell death protein 1) therapy (Bonferroni-corrected P = 0.008, Fig. 8A). Besides immunotherapies, we attempted to identify whether the 25-lncRNA prognostic signature could be applied to chemotherapies prediction. The results showed that the low-risk group had a lower half maximal inhibitory concentration of docetaxel, gefitinib and vinblastine, while the high-risk group had a lower half maximal inhibitory concentration of doxorubicin, mitomycin C and paclitaxel (Fig. 8B, P < 0.05). Thus, the 25-lncRNA prognostic signature could act as a potential predictor for immunotherapies and chemotherapies.