RNA-sequencing (RNA-seq) counts and tumor data for 368 patients in The Cancer Genome Atlas liver hepatocellular carcinoma (TCGA-LIHC) cohort were downloaded using UCSCXenaTools (19). Counts were normalized using the transcript per million (TPM) method, and genes with expression levels below the lower end of the interquartile range (IQR) of expression across all genes in the sample were excluded. Samples with low-frequency counts (the number of occasions of “count-per-million values > 0.5” in the sample less than 2) were excluded.

We also downloaded somatic mutation and DNA methylation profiling data for 364 and 430 patients, respectively, in the TCGA-LIHC cohort using UCSCXenaTools.

We constructed the network of MALAT1 with interacting proteins using text mining in the STRING database. A first shell was created with no more than 10 interactors, and a second shell with no more than 20 interactors. Enrichment of terms from the Kyoto Encyclopedia of Genes and Genomes (KEGG) in the network were visualized using the GOplot package in R. Genes encoding first-shell proteins were analyzed for potential assocations with HCC patient survival using the GEPIA2 database (http://gepia2.cancer-pku.cn/).

StarBase is an online database that can build ceRNA networks based on thousands of interactions between miRNAs and their target genes in a CLIP-seq (Crosslinking or RNA immunoprecipitation followed by sequencing) database (https://starbase.sysu.edu.cn/). We obtained potential MALAT1 target miRNAs and genes that may compete with MALAT1 for binding miRNAs through StarBase (screening by miRNA number ≥ 30 and p-value ≤ 1.0e-5). In this way, a network of ceRNAs was constructed and visualized by Cytoscape. Genes regulated by MALAT1 and miRNAs were analyzed for their potential associations with survival of HCC patients using the GEPIA2 database.

Somatic mutations were analyzed using the maftools package in R (4.1.0). We selected samples tested for both expression and mutation, then selected 84 samples whose MALAT1 expression was below the lower end of the IQR and 85 samples whose MALAT1 expression exceeded the upper end of the IQR. We summarized the overall mutation profiles of samples expressing low or high MALAT1, then we compared the prevalences of differentially mutated genes between the two groups using Fisher’s exact test and enrichment analysis. Differences associated with p < 0.05 were considered significant. Samples were categorized as wild-type or mutant depending on the presence or absence of somatic mutations. Survival of patients in the wild-type or mutant groups was compared using the Kaplan-Meier method and log-rank test.

After being loaded, methylation data were filtered, quality-controlled, and normalized using the ChAMP package in R. Beta values were the quantification values of methylation levels. The singular value decomposition method (SVD) calculated the correlation between biological factors (MALAT1 and other clinical features) and the variation of beta values. Differentially methylated positions (DMPs) were identified by comparing samples expressing high or low MALAT1. We considered that the following DMPs could upregulate target genes of MALAT1: hypermethylated DMPs in the coding sequence of the gene, or hypomethylated DMPs in the promoter or transcriptional start site (TSS). Conversely, we considered that the following DMPs could downregulate target genes of MALAT1: hypomethylated DMPs in the coding sequence of the gene, or hypermethylated DMPs in the promoter or TSS. We then calculated correlation coefficients between MALAT1 and target genes to identify genes regulated by MALAT1-associated methylation. Results associated with p < 0.05 were considered significant. Cox regression was performed on the beta values of each DMP. DMPs that emerged as significant (p < 0.05) were shown as forest plots. Enrichment of Molecular Signatures Database’s terms or pathways in DMPs was assessed using gene set enrichment analysis (GSEA) using the “champ.GSEA()” function.

We used the Immuno-Oncology-Biological-Research (IOBR) package in R integrating six commonly used algorithms (MCPcounter, TIMER, xCell, CIBERSORT, EPIC, and quanTiseq), to calculate scores for tumor-infiltrating lymphocytes (TILs) and tumor-associated signatures. Scores that reached statistical significance were determined for 92 samples whose MALAT1 expression was below the lower end of the IQR and 92 samples whose MALAT1 expression exceeded the upper end of the IQR. Kaplan-Meier analysis was used to compare survival of patients whose MALAT1 expression was above or below the median, based on data for the TCGA-LIHC cohort. Differences between the two groups were compared using the log-rank test.

We analyzed samples from 179 HCC patients who underwent radical resection in the Affiliated Tumor Hospital of Guangxi Medical University between April 2004 and December 2012. None of the patients received TACE, radiotherapy, chemotherapy, or other anti-tumor treatments before surgery. Tumor tissues from all patients, as well as paired adjacent non-tumor tissues (PANTs) from 19 patients, were frozen in liquid nitrogen immediately after surgery and stored at −80°C until RNA extraction. Two pathologists participated in the diagnosis of histopathology of resected tumor samples. One of these two pathologists is responsible for routine pathological diagnosis in the clinical practice, while the other is responsible for routine quality control in the clinical study. This work was approved by the Institutional Ethics Committee of the Affiliated Tumor Hospital of Guangxi Medical University (Nanning, China, Approval Number: LW2022145), which waived the requirement for informed consent because patients, at the time of treatment, consented for their anonymized data to be analyzed and published for research purposes.

Clinicopathological data on all patients were collected, including age, sex, Child-Pugh classification, tumor size, tumor numbers, clinical TNM stage, the presence or absence of vascular invasion, capsular infiltration, and cirrhosis. Laboratory results were also collected, such as serum levels of alpha-fetoprotein (AFP) and hepatitis B surface antigen (HBsAg).

Follow-up information of patients was collected for survival analysis. Patients were followed up every 3 months during the first 2 years, and thereafter every 6 months, for 2–10 years. Tumor recurrence was detected based on radiology and serum AFP examination. Relationships of MALAT1 expression with PFS and overall survival (OS) were explored.

Total RNA was extracted from HCC tumors and PANTs using TRIzol (Invitrogen, Carlsbad, CA, United States) and stored at −80°C. RNA (1 μg) was reverse-transcribed into complementary DNA (cDNA) using the PrimeScriptTM RT Reagent Kit (Takara Bio, Kusatsu, Japan) in reactions with a total volume of 20 μl. Real-time PCR (RT-PCR) was performed in 10-μl reactions on a LightCycler^® 480 II Real-time PCR system (Roche, Basel, Switzerland) using GAPDH as an internal reference. Reactions were performed in a microassay plate (Roche) at 95°C for 10 min, followed by 40 cycles at 95°C for 10 s and 60°C for 30 s. Primer sequences were: MALAT1 forward, 5′-CAG​GCG​TTG​TGC​GTA​GAG​GA-3′; MALAT1 reverse, 5′-TGC​CGA​CCT​CAC​GGA​TTT​T-3′; GAPDH forward, 5′-GTC​AGC​CGC​ATC​TTC​TTT-3′; and GAPDH reverse, 5′-CGC​CCA​ATA​CGA​CCA​AAT-3′. All samples were tested in triplicate. Expression levels of MALAT1 were normalized to those of GAPDH and were calculated using the 2^−ΔΔCT method (20).

All statistical analyses were performed using SPSS 22.0 (IBM, Chicago, IL, United States) and R version 4.1.0. Differences in continuous variables were compared using Student’s t-test for normally distributed data or Mann-Whitney U-test for skewed data. Patients’ survival was analyzed by the Kaplan-Meier method, and differences were assessed for significance using the log-rank test. Univariate analysis was performed to identify clinicopathological features associated with survival, and results were assessed for significance using the Gehan-Breslow-Wilcoxon test.

Covariates that were associated with p < 0.05 in univariate analysis were included in multivariate Cox analysis, which was used to select features for constructing Cox proportional hazard models to predict survival. Models for predicting PFS and OS were assessed using the area under time-dependent receiver operating characteristic curves (AUC), as calculated with the timeROC package in R. Net benefits of models was assessed using decision curves, plotted with the ggDCA package in R. The ability of MALAT1 expression to diagnose clinicopathological features of HCC patients was assessed using AUCs. All hypothesis testing was two-sided, with p < 0.05 considered statistically significant.

The protein interaction network related to MALAT1 was obtained using the text mining algorithm in STRING (Figure 1A). The text mining identified 10 proteins interacting with MALAT1 as EZH2, SRSF1, SMARCA4, HNRNPC, SUZ12, ELAVL1, SFPQ, SP1, AGO2, HDAC9. The proteins in the network were most enriched in KEGG terms associated with cancer (Figure 1B). Among these terms, HCC had the highest z-score, indicating that the proteins in this network were most closely related to HCC (Figure 1C). Five protein-coding genes in the first shell of the network were significantly associated with worse survival when highly expressed: AGO2, HNRNPC, EZH2, SFPQ, and SRSF1 (Figure 2; log-rank p < 0.05).

From the ceRNA network for MALAT1, we obtained 10 target genes: ZBTB6, LCOR, PDE7A, SORL1, RPL37, RPL7L1, SREK1IP1, DIS3, PPP1R3C, CAMK2G. All these genes may competitively bind to the same miRNAs as MALAT1 (Figure 3). Higher expression of LCOR, PDE7A, RPL7L1, RPL37, and SREK1IP1 was associated with worse survival of HCC patients (log-rank p < 0.05; Figure 4).

Supplementary Figure S1 showed the mutation profiles of the samples. Samples expressing high or low MALAT1 were compared using Fisher’s exact test to identify 10 differentially mutated genes (Figures 5A, B). Eight mutated genes were enriched in samples expressing high MALAT1 and 13 mutated genes were enriched in samples expressing low MALAT1 (Figure 5C). Patients with higher MALAT1 expression showed more mutations in IRX1 or TP53 and mutations in IRX1 and TP53 indicated worse survival than those with the corresponding wild-type alleles (log-rank p < 0.05; Figure 6). Patients with lower MALAT1 expression were associated with more LRP1B mutations and LRP1B mutations indicated worse survival (log-rank p < 0.05; Figure 6). Mutations in the 10 differentially mutated genes altered the protein sequences (Supplementary Figure S2).

After normalizing the methylation signal value matrix, samples expressing high or low MALAT1 were analyzed by multidimensional scaling (Supplementary Figure S3A). The beta distributions for each sample showed that most probes were completely unmethylated (beta value ≤ 0.2), and a smaller number of probes were partially methylated (0.4 < beta value < 0.6) or completely methylated (beta value ≥ 0.8) (Supplementary Figure S3B). Sample clustering showed no outliers (Supplementary Figure S3C). The SVD plot showed that several principal components that best explained the methylation variation significantly correlated with MALAT1 (Figure 7). More principal components significantly correlated with MALAT1 than the clinical features of tumor stage and tumor grade, indicating that MALAT1 correlated with methylation variants more strongly than stage or grading did (Figure 7). In samples expressing high MALAT1, hypomethylation occurred in 88 positions and hypermethylation in 250 positions (Figure 8A).

Beta values of DMPs were visualized using a clustering heatmap (Figure 8B). In the case of high MALAT1 expression, 157 DMPs were hypermethylated and located in coding regions of genes, while 28 DMPs were hypomethylated and located in promoters and TSS. These DMPs may positively correlated with gene expression. Conversely, 60 DMPs were hypomethylated and located in coding regions of genes, while 93 DMPs were hypermethylated and located in promoters and TSS. These DMPs therefore may negatively correlated with gene expression (Supplementary Figure S4).

We further analyzed the correlation of these genes with MALAT1 expression at the transcriptome level. We assumed that 87 genes were upregulated by MALAT1 through methylation, and confirmed that their expression positively correlated with that of MALAT1. Conversely, three genes were assumed down-regulated by MALAT1 through methylation, and their expression negatively were confirmed correlated with that of MALAT1 (Supplementary Figure S5).

Univariate Cox regression was performed for 87 up-regulated genes and 3 down-regulated genes, of which 38 were significantly upregulated; all genes emerged as risk factors for poor survival except ANKRD24, which was a protective factor (Figure 8C). Eight terms with AUCs > 0.75 in GSEA were associated with tumors. One of the terms “IIZUKA_LIVER_CANCER_PROGRESSION_G2_G3_DN” was directly related to liver cancer progression (Figure 8D). Therefore, the MALAT1-related methylation probes were associated with HCC progression and survival of patients.

In tumor signatures and TIL analysis, we found significant enrichment in some tumor-related signatures scores in samples expressing high MALAT1. These signatures included genes involved in the cell cycle, DNA damage repair (DDR), mismatch repair, homologous recombination, molecular cancer m6A, exosome, and positive regulation of exosomal secretion (Figure 9). In contrast, ferroptosis scores were significantly lower in samples expressing high MALAT1.

We used six algorithms to estimate TILs in HCC. Several algorithms showed consistent results for the analysis of macrophages and fibroblasts. Macrophage scores were significantly higher in samples expressing low MALAT1, whereas fibroblasts scores were significantly higher in samples expressing high MALAT1 (p < 0.05; Figure 9). In analyses of other cell types, there were few significant results or different algorithms were inconsistent in predicting the trends (Figure 9).

Among HCC patients in the TCGA-LIHC cohort, there was no significant difference in OS between those with high or low MALAT1 expression (Figure 10A). PFS was significantly shorter in patients with high MALAT1 expression than in those with low expression (Figure 10B; p = 0.014). In a multivariate Cox regression model for MALAT1 expression and baseline clinical information (age, gender, tumor stage, and tumor grade), MALAT1 independently predicted shorter PFS (Figure 10C; p = 0.023).

The median age of the 179 HCC patients was 51 years (19–77 years). There were 160 males and 19 females. The median diameter of tumors was 5.5 cm (range, 2–19 cm). The proportion of patients positive for HBsAg was 86.6%, and the proportion with serum AFP was 65.4%. The baseline features of the cases are shown in Table 1.

Median relative expression of MALAT1 was significantly higher in cancer tissues than in PANTs (5.42 ± 3.84 vs. 0.048 ± 0.032, p < 0.001), confirming that MALAT1 was markedly upregulated in HCC (Figure 11A).

Subgroup analysis showed that patients with a history of liver cirrhosis, patients who were positive for HBsAg, patients who were positive for AFP, patients with vascular invasion and patients with tumor capsule infiltration had significantly higher MALAT1 expression than the corresponding patient subgroups without these characteristics (all p < 0.05). Nevertheless, there was no significant correlation between MALAT1 expression and age, sex, tumor diameter or number, Child-Pugh classification, clinical TNM stage, serum level of transaminase, total bilirubin, or albumin (all p ≥ 0.05; Table 1).

Across all 179 patients, the median follow-up duration was 41 months (range, 23–59 months) for PFS and 70 months (range, 55–85 months) for OS. Patients were divided based on the mean value of MALAT1 relative expression level of 38.76 into a high expression group (n = 63) and low expression group (n = 116).

Patients with high MALAT1 expression were with 19.2 months mPFS and 40.5 months mOS and patients with low MALAT1 expression were with 52.8 months mPFS and 78.3 months mOS. Patients with high MALAT1 expression showed significantly shorter PFS than those with a low MALAT1 expression (p = 0.033, Figure 11B). Similarly, they showed significantly shorter OS (p = 0.023, Figure 11C).

In 59 patients with vascular invasion, the mPFS of patients with MALAT1 overexpression and low expression were 14 months and 49 months, and PFS was significantly shorter among those with high MALAT1 expression than among those with low expression (HR 2.63, 95% CI 1.294–5.346; p = 0.008, Figure 11D). Similarly, the mOS of patients with MALAT1 overexpression and low expression were 58 and 72 months, and OS was significantly shorter among those with high MALAT1 expression (HR 2.416, 95% CI 1.160–5.031; p = 0.018, Figure 11E).

MALAT1 overexpression was associated with shorter PFS and OS in subgroups who were positive for HBsAg or AFP (Table 2). Among 130 patients with liver cirrhosis, high MALAT1 expression was associated with significantly shorter OS than low expression (p = 0.018) and with a tendency toward shorter PFS (p = 0.0566; Table 2). In contrast, the expression level of MALAT1 was not significantly associated with PFS or OS among patients without vascular invasion, without cirrhosis, who were negative for HBsAg or who were negative for AFP (Table 2).

Univariate analysis showed that sex, HBsAg, tumor number, tumor diameter, TNM stage, and MALAT1 expression were significantly associated with OS and PFS (Table 3).

Variables with p < 0.05 in the univariate analysis were included in the multivariate Cox analysis. Furthermore, the results showed that the independent risk factors for OS included sex, HBsAg, and clinical TNM stage (p < 0.05; Figure 12). At the same time, HBsAg and clinical TNM stage were independent risk factors for PFS (p < 0.05; Figure 12). The p-values of tumor size from OS and PFS multivariate Cox analysis was less than 0.1, and the p-value of MALAT1 expression from PFS multivariate Cox analysis was less than 0.1 (Figure 12).

The multivariate Cox analysis showed that there was strong collinearity between the variable “tumor number” and “clinical TNM stage” and the prognostic value of “sex” was not clear. Therefore, we constructed prediction models based on four promising variables: “MALAT1,” “size,” “clinical TNM stage,” and “HBsAg” (Figures 13A, B). The calibration curves showed that the progression prediction model fit the data well at 1, 2 and 3 years (Figure 13C), while the survival prediction model fit the data moderately well at 1, 3, and 5 years (Figure 13D). Using the progression prediction model, we obtained AUCs of 0.683 for 1-year progression rate, 0.692 for 2-year progression rate, and 0.661 for 3-year progression rate (Figure 14A). Using the survival prediction model, we obtained AUCs of 0.731 for 1-year survival rate, 0.703 for 3-year survival rate, and 0.699 for 5-year survival rate (Figure 14B). Time-dependent AUCs showed that the prediction models outperformed the predictions of individual indicators. Predictive models constructed with other indicators excluding MALAT1 were slightly inferior to MALAT1-based predictive models, for both PFS and OS. The performance of MALAT1 alone was not inferior to that of clinical stage and HBsAg in the prediction of progression, or from that of tumor size and HBsAg in the prediction of survival when data were censored at > 40 months (Figures 14C, D). Decision curves were plotted by depicting the net benefit ratio along the vertical axis and the risk threshold along the horizontal axis (Figures 14E, F). The results showed that the smaller the risk threshold value, the greater the net benefit. Over most threshold intervals, the progression and survival prediction models led to greater net benefit than models based on other independent indicators.

Using an appropriate cut-off value, the AUC of MALAT1 expression level used to distinguish patients with or without liver cirrhosis was 0.631 (cut-off value 3.86, p = 0.006; Table 4; Figure 15A), 0.676 for patients with or without vascular invasion (cut-off value 54.52, p < 0.001; Figure 15B), 0.605 for patients with or without tumor capsule infiltration (cut-off value 54.52, p = 0.039; Figure 15C), and 0.593 for patients positive or negative for AFP (cut-off value 4.23, p = 0.037; Figure 15D).