The role of liquid-liquid phase separation in hepatocellular carcinoma: single-cell analysis and identification of prognostic biomarkers

Introduction

The incidence and mortality rates of hepatocellular carcinoma (HCC) are on the rise, and it is the third most common cause of cancer-related deaths globally, posing a major challenge to global health (1,2). The development of immune therapeutics for HCC may be the key to overcoming the barriers in treating this disease (3), with several strategies being devised, including immune checkpoint inhibitors, vaccine therapy, and adoptive cell transfer (4-6). The goal of these therapies is to utilize the human immune system for more effective targeting and destruction of cancer cells. Despite significant advancements in HCC research, the prognosis of patients remains poor. Specifically, the median overall survival (OS) for patients with advanced HCC following sequential systemic therapy is 37.4 months, and the survival rates at 12, 24, and 36 months are 71.4%, 55.6%, and 55.6%, respectively (7). This suggests a critical need for innovative research to identify novel therapeutic targets and prognostic biomarkers.

Liquid-liquid phase separation (LLPS) is a process in which, under specific conditions, distinct components in a mixed liquid system undergo interactions that lead to the formation of two or more phase-separated liquid phases. These phases are typically composed of different components, such as proteins or RNA, which is in contrast to more conventional solid–liquid or gas–liquid phase separations. Recent research has confirmed that LLPS is involved in various critical biological processes (BP) by performing specific functions, including gene regulation, chromatin architecture, X-chromosome inactivation, DNA repair mechanisms, cancer development, and autophagy (8-11). Moreover, clarifying how LLPS is related to cancer can provide novel insights into cancer biology and identify pathways linked to effective therapy. For instance, the dysregulation of LLPS contributes to oncogenic activity, which may constitute a novel anticancer strategy (12). However, research into HCC regulation mediated by an LLPS-related signature (LLPSRS) remains in its early stages, and molecular mechanisms remain to be clarified in depth.

By integrating comprehensive bioinformatics analyses with external validation, we sought to construct an LLPSRS that significantly correlated with HCC pathogenesis, progression, and patient prognosis. Our investigation not only deepens the understanding of LLPS in the context of HCC but also provides a foundation for identifying novel biomarkers and therapeutic targets within the complex biological landscape of HCC. We present this article in accordance with the TRIPOD reporting checklist (available at https://tcr.amegroups.com/article/view/10.21037/tcr-2025-726/rc).

Methods

Data collection

The RNA-sequencing (RNA-seq) expression profiles and clinical data for patients with liver HCC (LIHC) were downloaded from The Cancer Genome Atlas (TCGA) (http://cancergenome.nih.gov/) and the International Cancer Genome Consortium Data Portal (ICGC) (https://dcc.icgc.org). Patients with more than 50% missing data in the original dataset were excluded from the LIHC cohort. Genes related to LLPS were sourced from GeneCards (https://www.genecards.org/). This study was conducted in accordance with the Declaration of Helsinki and its subsequent amendments.

Single‑cell RNA‑seq analysis data collection and processing

From the Gene Expression Omnibus (GEO) dataset, we obtained single-cell RNA-seq (scRNA-seq) data from two patients with HCC in the GSE166635 dataset. The “Seurat” R package (13) (The Foundation for Statistical Computing) was used for data analysis, which began with rigorous quality control and the filtering out of cells with mitochondrial gene expression over 15%. Furthermore, the “SingleR” R package was used for cell annotation. Subsequent analysis focused on the highly variable genes to characterize the cellular diversity. The “Harmony” R package was employed to mitigate sample-specific batch effects and thus ensure the comparability of the two samples. The “FindClusters” and “FindNeighbors” R packages were jointly used to facilitate the recognition of cell clusters, and the t-distributed stochastic neighbor embedding (t-SNE) plot provided a visual representation of our findings. Cell types were annotated through marker gene analysis, and the “AddModuleScore” R package was applied to quantify the specific gene set activities within each cell. Furthermore, the “FindMarkers” R package was used to identify the differentially expressed genes (DEGs) between the different groups, with significance [adjusted P value (P.adj) <0.05] determined by the Wilcoxon test. RNA-seq data from TCGA-LIHC were used to construct coexpression networks. The genes of the LLPSRS identified from single-cell analysis were extracted, and low-variance genes were removed. LLPS enrichment scores were calculated via single-sample gene set enrichment analysis (ssGSEA) and were used as phenotypic traits. LLPS-related DEGs were subjected to batch transcriptome analysis via weighted gene coexpression network analysis (WGCNA), supplemented by the “CellChat” R package (14) to gain a deeper understanding of the cell interaction dynamics.

LLPS score calculation and pathway analysis in TCGA dataset

We used ssGSEA through the “GSVA” R package (15) to compute LLPS scores for each TCGA-LIHC sample, with the upregulation or downregulation of gene sets being determined. The “limma” R package (16) was applied for the calculation of the hallmark pathway gene set variation analysis (GSVA) score to identify pathways with significant differential expression between high-risk and low-risk groups. Moreover, we conducted GSEA for Gene Ontology (GO) gene sets (c5.go.v7.5.1.symbols.gmt) using the “clusterProfiler” R package, with the criterion of P<0.05, to further investigate the BP, cellular components (CC), and molecular functions (MF) that distinguished the risk subgroups.

LLPSRS analysis in TCGA-LIHC

We used the “limma” R package to perform differential expression analysis between normal and tumor samples using TCGA LIHC data [log |fold change (FC)| >1.5 and P.adj <0.05]. Subsequently, these DEGs were matched with those in the LLPS-related module identified via WGCNA, with matching DEGs being considered for the LLPSRS. We first used univariate Cox regression and then least absolute shrinkage and selection operator (LASSO) regression to minimize overfitting through 10-fold cross-validation of gene selection. The efficacy of the result was visualized via LASSO coefficient curves. In order to clarify the association between the genes in the LLPSRS, we employed the “circlize” R package (version 0.4.1) and used the Wilcoxon-rank sum test to statistically compare LLPSRS expression between normal and tumor tissues.

Validation of LLPSRS model accuracy in predicting patient prognosis

Initially, individual patients were classified into high- or low-risk categories according to the median risk score of the LLPSRRS. A heatmap was used to display the expression differences of nine LLPSRS genes between the high- and low-risk groups. Subsequently, scatter plots were employed to exhibit the association between the risk scores of patients with HCC and their survival times, with data from TCGA and ICGC being used. Meanwhile, Kaplan-Meier survival curves were generated for both TCGA and ICGC cohorts to evaluate the prognostic value of the LLPSRS. Finally, receiver operating characteristic (ROC) analysis was conducted to assess the predictive accuracy of the LLPSRS model.

Independent prognostic evaluation and nomogram design

Univariate and multivariate Cox regression analyses were used to determine the independence of risk scores and clinical features as prognostic factors for HCC. Additionally, a predictive nomogram was created with the previously identified independent prognostic indicators via “rms” R package to forecast the OS of patients with HCC. Finally, we assessed the prognostic accuracy of the nomogram using calibration curves and concordance index (C-index) analysis, demonstrating its effectiveness in predicting survival.

Clarifying the underlying molecular mechanisms of the LLPSRS via bulk transcriptome analysis

To clarify the molecular mechanisms underlying the link between LLPSRS and HCC prognosis, we applied functional enrichment analyses through GSEA, targeting GO gene sets in high- and low-risk groups. Additionally, we investigated the association between LLPSRS and biomarker pathway scores.

Analysis of genomic variation between the LLPSRS risk subgroups

Tumor mutation burden (TMB) serves as an important biomarker in cancer research, and a deeper understanding of cancer mutations can be gained via the quantification of mutations in the tumor genome (17). We computed the TMB score of patients with HCC by calculating the mutation frequency from sequencing data and then conducted survival analysis based on their TMB scores. To identify the somatic mutations related to LLPSRS, the “maftools” R package was used to create a waterfall plot showing the mutation status of patients with HCC in both the high-risk and low-risk groups. Furthermore, we conducted copy number variation (CNV) analysis on the gene with the largest difference between the high- and low-risk groups.

Comprehensive analysis and validation of LLPSRS’s association with immune cell infiltration in HCC

To determine the association between LLPSRS and immune cell infiltration in the HCC tumor microenvironment (TME), we quantitatively analyzed 22 different immune cell types using the cell-type identification by estimating the relative subsets of RNA transcripts via the CIBERSORT algorithm. Furthermore, we validated the CIBERSORT results by applying the ESTIMATE, ssGSEA, and xCell methods to ensure the accuracy and reliability of our findings.

Prediction of the correlation between the LLPSRS and drug sensitivity

The R software package “pRRophetic” (18) was used to predict the chemotherapy sensitivity of patients with HCC based on their LLPSRS risk score in order to facilitate personalized treatment. By comparing patient tissue gene expression profiles with those of tumor cell lines, we calculated the half maximal inhibitory concentration (IC₅₀). The Wilcoxon test was used to detect differences in drug IC₅₀ between the high- and low-risk groups, with P<0.05 indicating statistical significance.

Statistical analysis

All statistical analyses were performed with R software (version 4.2.1). Continuous variables were compared with the Wilcoxon rank-sum test, while categorical variables were analyzed with the chi-squared test. Survival outcomes were evaluated via Kaplan-Meier analysis and log-rank tests. Univariate and multivariate Cox proportional hazards regression models were applied to identify independent prognostic factors. The prognostic performance of the LLPSRS model was assessed via time-dependent ROC curves and the C-index. LASSO regression with ten-fold cross-validation was employed to reduce overfitting and construct the prognostic signature. Statistical significance was defined as a two-tailed P<0.05.

Results

LLPS characteristic in the single‑cell transcriptome

In this study, 17,854 scRNA-seq datasets from two HCC samples were collected. To eliminate batch effects, we successfully fused the two samples using the “Harmony” R package. Next, to reduce dimensionality, we conducted principal component analysis (PCA) and t-SNE on the 2,000 genes exhibiting the highest expression differences. The obtained cells were annotated into 11 principal categories (Figure 1A), with 1 being used as the resolution for the specific cell type of marker genes. The key genes of each cell type were visualized in a heatmap (Figure 1B). In addition, we assessed the greatest expression variation of the 970 LLPS-related genes in all cells (Figure 1C) and found significant differences in the LLPS activity between hepatocytes and CD4⁺ T cells (Figure 1D). Finally, based on the differences in LLPS activity, we identified 3,338 DEGs for further analysis.

Figure 1 LLPS characteristics in the liver cancer cohort. (A) t-SNE plot showing the cell types identified by the marker genes. (B) Heatmap showing the top 4 marker genes in each cell cluster. (C) The activity score of LLPS in each cell. (D) The distribution of the LLPS score in different cell types. LLPS, liquid-liquid phase separation; t-SNE, t-distributed stochastic neighbor embedding.

Identification of the key modules and genes associated with LLPS in bulk RNA‑seq

We used the ssGSEA algorithm to assess the variations in BP and pathway activities within TCGA-LIHC samples, which helped to generate LLPS activity scores for subsequent WGCNA after removal of the outlier samples (Figure 2A). Using WGCNA, we selected modules with significant LLPS score associations. A network was constructed with a soft threshold power of 9 (R² =0.85), and six modules were generated by setting the minimal gene counts to 50 and the reassignment threshold to 0 (Figure 2B). To evaluate their biological relevance, we correlated module eigengenes with LLPS enrichment scores. The MEturquoise module showed a strong relationship with LLPS scores in bulk RNA-seq (correlation coefficient =0.7) (Figure 2C). Additionally, a strong positive correlation between gene significance and module membership was observed in the turquoise module, indicating that the hub genes in this module are also highly correlated with LLPS-related traits (correlation coefficient =0.92; P<1e−200) (Figure 2D). Volcano plots (Figure 2E) were used to identify the DEGs between tumor and normal tissues in TCGA-LIHC (|log FC|>1.5 and P.adj <0.05). It was found that there were 175 genes at the intersection of the turquoise module and DEGs (Figure 2F), which were included in the LLPSRS and implicated in LLPS at the bulk level. Finally, GO analysis (Figure 2G) of the LLPSRS genes revealed significant enrichment in BP, including cadherin binding, adenosine triphosphate (ATP) hydrolysis activity, and actin filament binding; in CC, including cell-substrate junction and focal adhesion; and in MF, including messenger RNA processing and RNA splicing.

Figure 2 Identification of the LLPS-related genes. (A) Dendrogram showing the hierarchical clustering of TCGA-LIHC samples. The bottom heatmap represents each sample’s LLPS score, as calculated by the ssGSEA algorithm. (B) Cluster dendrogram of the WGCNA. (C) Module-trait heatmap showing that the MEturquoise module was closely related to the LLPS trait. (D) Scatter plot showing the relationship between GS and MM in the turquoise module. (E) Volcano plot showing the results of the differential analysis of TCGA-LIHC tumor samples and normal samples, with the top 5 upregulated or downregulated genes being marked. (F) Venn plot showing the LLPS-related genes between the MEturquoise module and DEGs according to bulk RNA-seq. (G) GO enrichment of the LLPS genes. BP, biological processes; CC, cellular components; cor, correlation coefficient; DEG, differentially expressed gene; GS, gene significance; LLPS, liquid-liquid phase separation; MF, molecular function; MM, module membership; RNA-seq, RNA sequencing; ssGSEA, single-sample gene set enrichment analysis; TCGA-LIHC, The Cancer Genome Atlas-liver hepatocellular carcinoma; WGCNA, weighted gene coexpression network analysis.

Selection of prognostic markers and their differential expression in LIHC

To construct the novel LLPSRS, we performed univariate Cox regression analysis to analyze all 175 LLPSRS genes obtained from the intersection of genes in the turquoise module and DEGs from the bulk RNA-seq, identifying 33 genes with significant prognostic values (P<0.001). The results of the univariate Cox regression analysis are presented in Figure 3A. After a LASSO analysis of the 10-fold cross-validation framework, we identified nine genes (Figure 3B,3C) to form the basis of the LLPSRS. As shown in Figure 3D, there exist robust interactions between the genes in the LLPSRS, which highlight their coordinated effect in LIHC pathology. Furthermore, the Wilcoxon rank-sum test indicated that the expression of LLPSRS genes was drastically different between LIHC and normal tissues (Figure 3E), marking them as potential biomarkers for LIHC diagnosis and treatment.

Figure 3 Identification of candidate LLPSRS genes. (A) Univariate Cox regression analysis was used to assess the prognostic value of the LLPS-related genes. (B) Adjusted parameter selection in the LASSO model with 10-fold cross-validation. (C) LASSO coefficient curves. (D) Correlation between the LLPS-related genes. (E) The Wilcoxon rank-sum test was used to analyze the differential expression of 9 LLPS-related genes in LIHC tissues and normal tissues. ***, P<0.001. CI, confidence interval; LASSO, least absolute shrinkage and selection operator; LIHC, liver hepatocellular carcinoma; LLPS, liquid-liquid phase separation; LLPSRS, liquid-liquid phase separation-related signature; TPM, transcripts per kilobase of exon model per million mapped reads.

Validation and evaluation of the LLPSRS model

Nine genes were selected based on the results of Cox LASSO regression analysis. A heatmap was used to illustrate the upregulated expression of nine LLPSRS in patients classified as high risk (Figure 4A). Through a risk score formula, we identified key genes and obtained their corresponding coefficients using a multivariate Cox regression model as follows: risk score =(0.0649136516579006 × RAB10 expression level) + (0.00994540720909403 × RAP2A expression level) + (0.158139987801821 × ARL8B expression level) + (0.0520014415255968 × RSU1 expression level) + (0.0179103306404882 × ZMPSTE24 expression level) + (0.101833491608971 × NRAS expression level) + (0.0710264399891766 × UBAP2L expression level) + (0.0819650378765034 × AGFG1 expression level) + (0.0224195630787008 × GOLT1B expression level). Individual patients were classified into high- or low-risk categories based on the median risk score. This model effectively differentiated patients based on risk score (Figure 4B,4C). In both the training set and the ICGC dataset, high-risk patients exhibited significantly poorer OS compared to their low-risk counterparts, with higher risk scores being correlated with higher mortality rates (P<0.05; Figure 4D,4E). ROC curve analyses yielded area under the curve (AUC) values of 0.762, 0.649, and 0.641 for 1-, 2-, and 3-year survival predictions in the training set and 0.686, 0.669, and 0.672 in the ICGC dataset, respectively (Figure 4F,4G), underscoring the LLPSRS’s effective discriminatory power and predictive ability.

Figure 4 Validation of the LLPSRS model accuracy in predicting patient prognosis. (A) A heatmap showing the differential expression of nine LLPS-related differentially expressed genes in groups at high and low risk. Scatter plots of LIHC patients’ risk ratings and prognostic LLPS-related differentially expressed genes and their association with survival time for the (B) TCGA and (C) ICGC cohorts. (D,E) Survival curves for the TCGA cohorts and ICGC cohorts, respectively. (F,G) Time-dependent ROC curves for the TCGA and ICGC cohorts, respectively. AUC, area under the curve; ICGC, International Cancer Genome Consortium; LIHC, liver hepatocellular carcinoma; LLPS, liquid-liquid phase separation; LLPSRS, liquid-liquid phase separation-related signature; ROC, receiver operating characteristic; TCGA, The Cancer Genome Atlas.

Establishment and validation of a nomogram combining clinical characteristics and LLPSRS

To further evaluate whether risk score and clinical factors can serve as independent prognostic factors for HCC, we employed univariate and multivariate Cox regression analyses focusing on OS within the TCGA-LIHC cohort (Figure 5A,5B). It was found that stage (P<0.001) and risk score (P<0.001) maintained their status as independent prognostic factors for OS both in the univariate analysis and multivariate Cox analysis, highlighting the risk score’s substantial prognostic value for patients with HCC. In order to broaden the clinical applicability of our risk assessment models for HCC, we devised a nomogram that included clinical characteristics and risk score to predict 1-, 2-, and 3-year survival (Figure 5C). The C-index indicated the nomogram’s strong predictive performance, as it surpassed other clinical indicators in predicting OS for 1–10 years (Figure 5D). Traditional clinicopathological characteristics fell short in precision compared to the nomogram (AUC =0.771) for predicting the prognosis of HCC (Figure 5E). Additionally, the calibration curves indicated close agreement between the nomogram’s predictions and the actual clinical results (Figure 5F).

Figure 5 Independent prognostic analysis of LLPS-related differentially expressed genes risk scores and clinical parameters. (A) Univariate and (B) multivariate Cox regression analyses of characteristics and different clinical features. (C) Nomogram for predicting the 1-, 2-, and 3-year OS in patients with LIHC. (D) C-index curves for the nomogram, risk scores, and clinical parameters. (E) Multi-indicator ROC analysis of the TCGA cohort. (F) The calibration plots showing the comparison between predicted and actual OS for 1-, 2-, and 3-year survival probabilities. **, P<0.01; ***, P<0.001. AUC, area under the curve; CI, confidence interval; HR, hazard ratio; LIHC, liver hepatocellular carcinoma; LLPS, liquid-liquid phase separation; OS, overall survival; ROC, receiver operating characteristic; TCGA, The Cancer Genome Atlas.

The molecular mechanisms of the LLPSRS at the transcriptome level

To clarify the molecular basis of LLPSRS’s relationship and HCC prognosis, we conducted functional enrichment analysis with the TCGA dataset. Using GSEA, we found significant enrichment in processes such as epithelial-mesenchymal transition, MYC targets V1, and protein secretion in high-risk cohort (Figure 6A,6B). In addition, GSVA revealed that the high-risk group exhibited increased pathway activity associated with mitotic spindle, PI3K/AKT/mTOR signaling, and G2/M checkpoint, while the low-risk group was more active in the oxidative phosphorylation, fatty-acid metabolism, and xenobiotic metabolism pathways (Figure 6C). Finally, correlation analyses between LLPSRS and hallmark pathway scores solidified these findings (Figure 6D).

Figure 6 The transcriptomic features of HCC patients with various LLPS-related differentially expressed genes. (A) Ridge plot showing the GO terms enriched in the low-risk group. (B) GO terms enriched in the high-risk group as determined by GSEA. (C) Differences in hallmark pathway activities between the high- and low-risk groups as determined by GSVA. (D) Correlation between the risk score and hallmark pathway activities as determined by GSVA. GO, Gene Ontology; GSEA, gene set enrichment analysis; GSVA, gene set variation analysis; HCC, hepatocellular carcinoma; LLPS, liquid-liquid phase separation.

Landscape of genomic variation and TMB across the LLPSRS subgroups

To examine somatic mutations, we classified patients with HCC into high- and low-risk groups. Through analysis of somatic mutation data in the TCGA dataset, we found that the TMB was not significantly different between the low-risk group and the high-risk group (P=0.93; Figure 7A). Subsequently, we analyzed the prognostic potential of TMB in HCC. Dividing patients by median TMB values revealed that those in the low-TMB group had better survival outcomes than did those in the high-TMB group (P=0.03; Figure 7B). Specifically, low-risk patients with low TMB exhibited a superior prognosis, whereas high-risk patients with a high TMB had a worse prognosis (P<0.001). Our findings suggest that TMB, in conjunction with risk score, can predict outcomes in patients with HCC (Figure 7C). Further analysis of the genomic mutation landscape revealed distinct mutation profiles between the groups, with TP53, TTN, CTNNB1, and MUC being the most commonly mutated genes among high-risk patients, contrasting with the mutation patterns in low-risk patients (Figure 7D,7E). We compared the associations of co-occurring and mutually exclusive mutations among the top 20 most frequently mutated genes between the high- and low-risk groups. We also noted a higher frequency of co-occurring mutations in the high-risk group (Figure 7F) and investigated the CNVs in the top 15 most frequently mutated genes, discovering significant differences between the LLPSRS risk subgroups (Figure 7G).

Figure 7 Genetic alterations associated with LLPSRS between the low- and high-risk groups. (A) Violin plot showing the difference in TMB score between the high- and low-risk groups. (B) Kaplan-Meier curve showing the difference in OS between the high- and low-TMB score groups. (C) Kaplan-Meier curve analysis for OS according to the combination of TMB score and LLPSRS risk score. (D,E) The waterfall plot of the somatic mutation landscape in (D) the high-risk patients and (E) low-risk patients in the TCGA-LIHC cohort. (F) Heatmaps showing the association of co-occurrence and exclusive mutation among the top 20 mutated genes in the high- and low-risk groups. (G) Distribution of CNV frequency among DEGs between the high- and low-risk groups. CNV, copy number variation; DEG, differentially expressed gene; LLPSRS, liquid-liquid phase separation-related signature; OS, overall survival; TCGA-LIHC, The Cancer Genome Atlas-liver hepatocellular carcinoma; TMB, tumor mutation burden.

Association of the immune landscape and the LLPSRS

To evaluate immune infiltration in HCC samples, we applied the ssGSEA algorithm to the TCGA-LIHC transcriptomic dataset to compute activity scores for immune-related pathways. Our findings indicated significantly elevated activity in the nucleotide-binding oligomerization domain (NOD)-like receptor signaling pathway, leukocyte transendothelial migration, and B-cell receptor signaling pathway within the high-risk group (Figure 8A). Differentiation between high- and low-risk groups was further refined in specific immune cell populations through the CIBERSORT algorithm (see the violin plots in Figure 8B). Notably, the high-risk group exhibited higher levels of resting natural killer (NK) cells, M0 macrophages, resting dendritic cells, and neutrophils, while the low-risk group had a greater abundance of CD8 T cells, activated CD4 memory T cells, gamma delta T cells, and activated NK cells. Validation via the ssGSEA and xCell algorithms confirmed these patterns (see the boxplots in Figure 8C). Additionally, analysis revealed that the nine LLPSRS genes correlated strongly with the abundance of tumor-infiltrating immune cells; for instance, there were generally positive associations with M0 macrophages, and the genes ZMPSTE24 and AGFG1 were associated with neutrophils (Figure 8D).

Figure 8 The immune landscape associated with LLPSRS in HCC. (A) Immune-related pathways’ activity showing a significant difference between the high- and low-risk groups. The abundance of each TME-infiltrated cell type between the high- and low-risk groups as determined by (B) the CIBERSORT algorithm and (C) the ssGSEA algorithm. (D) The association between TME-infiltrated cells and genes in the LLPSRS. ns, P>0.05; *, P<0.05; **, P<0.01; ***, P<0.001. HCC, hepatocellular carcinoma; LLPSRS, liquid-liquid phase separation-related signature; ns, not significant; ssGSEA, single-sample gene set enrichment analysis; TME, tumor mutation burden.

Analysis of the correlation between the LLPSRS and drug sensitivity

Immunotherapy stands as the cornerstone of first-line therapy for patients with advanced HCC. However, the efficacy of immunotherapy is often hindered by drug resistance, which is attributed to the tumor’s dynamic and heterogeneous environment. To address this, we evaluated the sensitivity of the LLPSRS risk subgroups to various tyrosine kinase inhibitors (TKIs)—axitinib, sunitinib, and midostaurin—as well as the epidermal growth factor receptor (EGFR)-TKI inhibitor gefitinib. Our results demonstrated that the IC₅₀ for axitinib, sunitinib, and gefitinib were markedly lower in the low-risk group, suggesting higher sensitivity (Figure 9A-9C). Additionally, a direct correlation was observed between lower risk scores and reduced IC₅₀ values for these drugs (Figure 9D-9F). In contrast, midostaurin displayed a lower IC₅₀ for the high-risk group, showing increased drug sensitivity (Figure 9G), with risk score inversely correlated with IC₅₀ values (Figure 9H). These findings suggest that low-risk patients may derive greater benefit from axitinib, sunitinib, and gefitinib, whereas midostaurin could be more effective in high-risk patients.

Figure 9 Association between the ICDRS and drug sensitivity and the validation of the genes. (A-C,G) Comparison of the sensitivity to TKIs and EGFR-TKI inhibitors, including axitinib, sunitinib, midostaurin, and gefitinib, between the high- and low-risk groups. (D-F,H) The correlation between the risk score and the IC₅₀ of small-molecule drugs, including axitinib, sunitinib, gefitinib, and midostaurin, in HCC. ****, P<0.0001. EGFR, epidermal growth factor receptor; IC₅₀, half maximal inhibitory concentration; ICDRS, ICD-related gene signature; TKIs, tyrosine kinase inhibitors.

Discussion

HCC represents a major health challenge, characterized by a high mortality rate and complex pathogenesis (2,19,20). HCC’s notable heterogeneity and capacity for rapid evolution under treatment pressure have led to an increased resistance to the therapies, which is a critical research bottleneck (21). Recently, the application of bioinformatics in parsing genetic, transcriptomic, and proteomic data has shed light on possible solutions, uncovering potential biomarkers and therapeutic targets for HCC (22,23). Despite the extensive identification of novel biomarkers, the construction of reliable predictive models remains in the early stages. Addressing these challenges is essential for the advancement of personalized medicine and for improving the outcomes of patients with HCC.

LLPS functions critically in regulating gene expression, protein synthesis, and signal transduction through controlling the organization of biomolecules within membrane-less organelles (10,24-26). A growing body of research has confirmed that LLPS is closely associated with cancer development, involving alterations in cellular metabolism, growth signal pathways, and the TME, crucial elements for cancer initiation and progression (9,27-29). Furthermore, proteins and RNA molecules capable of phase separation have been explored as biomarkers for early diagnosis, disease monitoring, and prognosis (30,31). Evidence indicates that both LLPS-associated long noncoding RNAs and circular RNAs are linked with HCC progression (32,33). Although the relationship between LLPS and HCC has provided new perspectives for further research into HCC, reliable LLPS-based biomarkers for targeting the phase separation process are lacking. Our study applied scRNA-seq analysis to construct a novel signature and clarify the related molecular mechanisms, providing a molecular basis for understanding the correlation between LLPSRS, prognosis, and treatment response, while facilitating precision medicine among patients with HCC.

It is important to note that all genes included in our LLPSRS model are closely associated with tumor development, progression, and immunity. For instance, the expression of RSU1 is higher in more aggressive HCC cells, correlating with reduced cell proliferation, but may also exert tumor-suppressing effects (34). Lee et al. reported a significant association of ZMPSTE24 mutation and expression alterations with the microsatellite instability-high phenotype in gastric and colorectal cancers, suggesting ZMPSTE24’s critical role in tumorigenesis, potentially through pathways involving genomic instability and impaired DNA damage response (35). Moreover, NRAS has been shown to enhance the proliferation, migration, and invasion of HCC cells via the miR-145-5p/NRAS signaling pathway (36). Meanwhile, UBAP2L plays a carcinogenic role in HCC, and the suppression of UBAP2L activity has been shown to substantially curb the expansion and spread of HCC cells, a process likely linked with the modulation of the PI3K/AKT and P53 signaling networks (37). In esophageal squamous cell carcinoma (ESCC) tissues and cell lines, circ-AGFG1 is upregulated, promoting cell proliferation, migration, invasion, and glutamine catabolism while inhibiting apoptosis by targeting the miR-497-5p/SLC1A5 signaling pathway (38). In one study, a higher expression of GOLT1B in patients with HCC correlated with poorer OS (39). In another study, knockdown of RAB10 mediated by short hairpin RNA significantly reduced HCC cell proliferation and increased apoptosis, demonstrating that RAB10 overexpression promotes tumor growth and is associated with poor prognosis in HCC (40). In HCC tissues, RAP2A exhibits notably elevated expression levels, which contribute to enhanced tumor cell proliferation and apoptosis resistance through mTOR pathway activation (41). In cancer cells that survive radiation treatment, the activation of ARL8B enhances lysosomal exocytosis, thereby increasing the invasiveness of the cells. In cancer cells that survive radiation treatment, ARL8B activation enhances lysosomal exocytosis and increases cell invasiveness through strengthened binding to its effector, SKIP, which is regulated by BORC subunits after radiation (42).

In this study, we demonstrated that the LLPSRS had high predictive accuracy and substantial clinical translational potential using an integrated approach involving scRNA-seq and machine learning frameworks, broadening the scope of its application in HCC. The application of the LLPSRS not only enables the prediction of 3-year survival rate for patients with HCC but also serves as a foundation for forecasting adverse outcomes. Within the clinical setting, the nomogram consisting of the LLPSRS and clinical features offers a practical tool for estimating patient survival. Our findings support the use of LLPS in forecasting prognosis and treatment responses in patients with HCC, as it may improve clinical outcomes and the efficacy of therapeutic interventions. Noninvasive methods are generally favored for patient monitoring and diagnosis. Notably, patient sequencing data can be derived from both tumor tissues and blood samples (43). Within this context, LLPS holds considerable promise as an innovative noninvasive diagnostic technique, as its expression levels in blood and circulating tumor cells may significantly refine the prognostic prediction.

Notable progress has been achieved in the treatment of HCC, especially with the use of immune checkpoint inhibitors and TKIs (44,45). There is a correlation between longer median OS and treatment with immune checkpoint inhibitor therapy among patients with HCC, particularly when followed by TKI administration (46). Through a comprehensive examination involving single-cell analysis and the identification of prognostic biomarkers, we have gained a deeper understanding of how LLPS influences HCC progression. The LLPSRS can be used for patient stratification and ascertaining responsiveness to certain medications, including previously mentioned axitinib, sunitinib, midostaurin, and gefitinib. This strategy not only facilitates a more individualized administration of pharmaceuticals in a clinical setting but also represents a proactive approach to HCC treatment via the early identification of high-risk individuals and subsequent intervention. Moreover, a detailed investigation into the relationship between LLPSRS and drug responsiveness contributes to a more accurate prediction of treatment outcomes across different risk groups, aiding in the crafting of personalized treatment regimens. High-risk individuals, for example, may benefit more from certain TKIs, and thus a broader spectrum of treatment options may be provided for those at lower risk. This strategy of segmentation based on LLPSRS not only ensures more targeted treatment choices for patients with HCC but also lays the groundwork for proactive cancer prevention and early intervention, aligning with the aims of personalized and precision medicine in HCC treatment.

Despite providing valuable insights into the role of LLPS in HCC, this study involved limitations that should be acknowledged. Most notably, the sample size was small and derived from only two patients, which may not adequately capture the heterogeneity of HCC. Moreover, the reliance on scRNA-seq data, while innovative, necessitates further validation through larger, more diverse datasets to ensure the findings’ generalizability. Additionally, this study depended heavily on computational analyses, and experimental validation to confirm the mechanism by which the LLPSRS is connected to HCC progression was lacking.

Conclusions

In this study, we demonstrated that the LLPSRS is associated with HCC outcomes and established a robust framework for utilizing LLPS-related biomarkers in clinical prognostication and treatment strategy optimization. These findings suggest avenues for further investigation into targeting LLPS mechanisms for treating HCC, offering hope for improved patient outcomes.

Acknowledgments

None.

Footnote

Reporting Checklist: The authors have completed the TRIPOD reporting checklist. Available at https://tcr.amegroups.com/article/view/10.21037/tcr-2025-726/rc

Peer Review File: Available at https://tcr.amegroups.com/article/view/10.21037/tcr-2025-726/prf

Funding: This work was supported by Natural Science Foundation of Sichuan Province (No. 2022NSFSC1378 to X.T.).

Conflicts of Interest: All authors have completed the ICMJE uniform disclosure form (available at https://tcr.amegroups.com/article/view/10.21037/tcr-2025-726/coif). X.T. reports that this work was supported by Natural Science Foundation of Sichuan Province (No. 2022NSFSC1378). The other authors have no conflicts of interest to declare.

Ethical Statement: The authors are accountable for all aspects of the work in ensuring that questions related to the accuracy or integrity of any part of the work are appropriately investigated and resolved. The study was conducted in accordance with the Declaration of Helsinki and its subsequent amendments.

Open Access Statement: This is an Open Access article distributed in accordance with the Creative Commons Attribution-NonCommercial-NoDerivs 4.0 International License (CC BY-NC-ND 4.0), which permits the non-commercial replication and distribution of the article with the strict proviso that no changes or edits are made and the original work is properly cited (including links to both the formal publication through the relevant DOI and the license). See: https://creativecommons.org/licenses/by-nc-nd/4.0/.

References

1. McGlynn KA, Petrick JL, El-Serag HB. Epidemiology of Hepatocellular Carcinoma. Hepatology 2021;73:4-13. [Crossref] [PubMed]
2. Sung H, Ferlay J, Siegel RL, et al. Global Cancer Statistics 2020: GLOBOCAN Estimates of Incidence and Mortality Worldwide for 36 Cancers in 185 Countries. CA Cancer J Clin 2021;71:209-49. [Crossref] [PubMed]
3. Sprinzl MF, Galle PR. Current progress in immunotherapy of hepatocellular carcinoma. J Hepatol 2017;66:482-4. [Crossref] [PubMed]
4. Xu F, Jin T, Zhu Y, et al. Immune checkpoint therapy in liver cancer. J Exp Clin Cancer Res 2018;37:110. [Crossref] [PubMed]
5. Xu G, Feng D, Yao Y, et al. Listeria-based hepatocellular carcinoma vaccine facilitates anti-PD-1 therapy by regulating macrophage polarization. Oncogene 2020;39:1429-44. [Crossref] [PubMed]
6. Shin SK, Mishima Y, Lee Y, et al. Current Landscape of Adoptive Cell Therapy and Challenge to Develop "Off-The-Shelf" Therapy for Hepatocellular Carcinoma. J Gastroenterol Hepatol 2025;40:791-807. [Crossref] [PubMed]
7. von Felden J, Karkmann K, Ittrich H, et al. Sequential Systemic Treatment in Advanced Hepatocellular Carcinoma Is Able to Prolong Median Survival to More than 3 Years in a Selected Real-World Cohort. Visc Med 2021;37:87-93. [Crossref] [PubMed]
8. Pu X, Zhang C, Jin J, et al. Phase separation of EEF1E1 promotes tumor stemness via PTEN/AKT-mediated DNA repair in hepatocellular carcinoma. Cancer Lett 2025;613:217508. [Crossref] [PubMed]
9. Zhang H, Ji X, Li P, et al. Liquid-liquid phase separation in biology: mechanisms, physiological functions and human diseases. Sci China Life Sci 2020;63:953-85. [Crossref] [PubMed]
10. Nozawa RS, Yamamoto T, Takahashi M, et al. Nuclear microenvironment in cancer: Control through liquid-liquid phase separation. Cancer Sci 2020;111:3155-63. [Crossref] [PubMed]
11. Peng PH, Hsu KW, Wu KJ. Liquid-liquid phase separation (LLPS) in cellular physiology and tumor biology. Am J Cancer Res 2021;11:3766-76.
12. Mehta S, Zhang J. Liquid-liquid phase separation drives cellular function and dysfunction in cancer. Nat Rev Cancer 2022;22:239-52. [Crossref] [PubMed]
13. Stuart T, Butler A, Hoffman P, et al. Comprehensive Integration of Single-Cell Data. Cell 2019;177:1888-1902.e21. [Crossref] [PubMed]
14. Jin S, Guerrero-Juarez CF, Zhang L, et al. Inference and analysis of cell-cell communication using CellChat. Nat Commun 2021;12:1088. [Crossref] [PubMed]
15. Hänzelmann S, Castelo R, Guinney J. GSVA: gene set variation analysis for microarray and RNA-seq data. BMC Bioinformatics 2013;14:7. [Crossref] [PubMed]
16. Ritchie ME, Phipson B, Wu D, et al. limma powers differential expression analyses for RNA-sequencing and microarray studies. Nucleic Acids Res 2015;43:e47. [Crossref] [PubMed]
17. Falk M, Tiemann M, Ficker JH, et al. To be or low TMB (tumor mutational burden). Transl Cancer Res 2025;14:667-9. [Crossref] [PubMed]
18. Geeleher P, Cox N, Huang RS. pRRophetic: an R package for prediction of clinical chemotherapeutic response from tumor gene expression levels. PLoS One 2014;9:e107468. [Crossref] [PubMed]
19. Vogel A, Meyer T, Sapisochin G, et al. Hepatocellular carcinoma. Lancet 2022;400:1345-62. [Crossref] [PubMed]
20. Chidambaranathan-Reghupaty S, Fisher PB, Sarkar D. Hepatocellular carcinoma (HCC): Epidemiology, etiology and molecular classification. Adv Cancer Res 2021;149:1-61. [Crossref] [PubMed]
21. Salazar J, Le A. The Heterogeneity of Liver Cancer Metabolism. Adv Exp Med Biol 2021;1311:127-36. [Crossref] [PubMed]
22. Chi H, Zhao S, Yang J, et al. T-cell exhaustion signatures characterize the immune landscape and predict HCC prognosis via integrating single-cell RNA-seq and bulk RNA-sequencing. Front Immunol 2023;14:1137025. [Crossref] [PubMed]
23. Murai H, Kodama T, Maesaka K, et al. Multiomics identifies the link between intratumor steatosis and the exhausted tumor immune microenvironment in hepatocellular carcinoma. Hepatology 2023;77:77-91. [Crossref] [PubMed]
24. Wang B, Zhang L, Dai T, et al. Liquid-liquid phase separation in human health and diseases. Signal Transduct Target Ther 2021;6:290. [Crossref] [PubMed]
25. Zhang X, Yang Z, Fu C, et al. Emerging roles of liquid-liquid phase separation in liver innate immunity. Cell Commun Signal 2024;22:430. [Crossref] [PubMed]
26. Tong X, Tang R, Xu J, et al. Liquid-liquid phase separation in tumor biology. Signal Transduct Target Ther 2022;7:221. [Crossref] [PubMed]
27. Taniue K, Akimitsu N. Aberrant phase separation and cancer. FEBS J 2022;289:17-39. [Crossref] [PubMed]
28. Jiang S, Fagman JB, Chen C, et al. Protein phase separation and its role in tumorigenesis. Elife 2020;9:e60264. [Crossref] [PubMed]
29. Xu J, Liu W, Yao Y, et al. Liquid-liquid phase separation in hepatocellular carcinoma. Front Cell Dev Biol 2024;12:1454587. [Crossref] [PubMed]
30. Peng Q, Tan S, Xia L, et al. Phase separation in Cancer: From the Impacts and Mechanisms to Treatment potentials. Int J Biol Sci 2022;18:5103-22. [Crossref] [PubMed]
31. Peng W, Li Y, Cheng B, et al. Liquid-liquid phase separation-related lncRNA prognostic signature and ZNF32-AS2 as a novel biomarker in hepatocellular carcinoma. Comput Biol Med 2024;169:107975. [Crossref] [PubMed]
32. Cao P, Jin Q, Feng L, et al. Emerging roles and potential clinical applications of noncoding RNAs in hepatocellular carcinoma. Semin Cancer Biol 2021;75:136-52. [Crossref] [PubMed]
33. Chen S, Cao X, Zhang J, et al. circVAMP3 Drives CAPRIN1 Phase Separation and Inhibits Hepatocellular Carcinoma by Suppressing c-Myc Translation. Adv Sci (Weinh) 2022;9:e2103817. [Crossref] [PubMed]
34. Louca M, Stylianopoulos T, Gkretsi V. Ras Suppressor-1 (RSU1) in Cancer Cell Metastasis: A Tale of a Tumor Suppressor. Int J Mol Sci 2020;21:4076. [Crossref] [PubMed]
35. Lee JH, Yoo NJ, Kim MS, et al. Mutational and expressional alterations of ZMPSTE24, DNA damage response-related gene, in gastric and colorectal cancers. Pathol Res Pract 2016;212:1113-8. [Crossref] [PubMed]
36. Ding B, Fan W, Lou W. hsa_circ_0001955 Enhances In Vitro Proliferation, Migration, and Invasion of HCC Cells through miR-145-5p/NRAS Axis. Mol Ther Nucleic Acids 2020;22:445-55. [Crossref] [PubMed]
37. Liu Z, Yuan J, Zeng Q, et al. UBAP2 contributes to radioresistance by enhancing homologous recombination through SLC27A5 ubiquitination in hepatocellular carcinoma. Biochim Biophys Acta Mol Basis Dis 2024;1870:167481. [Crossref] [PubMed]
38. Chen S, Liu Z, Feng Q, et al. Circular RNA circ-AGFG1 contributes to esophageal squamous cell carcinoma progression and glutamine catabolism by targeting microRNA-497-5p/solute carrier family 1 member 5 axis. Anticancer Drugs 2023;34:195-206. [Crossref] [PubMed]
39. Butterworth J, Gregoire D, Peter M, et al. GOLT1B Activation in Hepatitis C Virus-Infected Hepatocytes Links ER Trafficking and Viral Replication. Pathogens 2021;11:46. [Crossref] [PubMed]
40. Wang W, Jia WD, Hu B, et al. RAB10 overexpression promotes tumor growth and indicates poor prognosis of hepatocellular carcinoma. Oncotarget 2017;8:26434-47. [Crossref] [PubMed]
41. Yang JR, Ling XL, Guan QL. RAP2A promotes apoptosis resistance of hepatocellular carcinoma cells via the mTOR pathway. Clin Exp Med 2021;21:545-54. [Crossref] [PubMed]
42. Wu PH, Onodera Y, Giaccia AJ, et al. Lysosomal trafficking mediated by Arl8b and BORC promotes invasion of cancer cells that survive radiation. Commun Biol 2020;3:620. [Crossref] [PubMed]
43. Liu J, Ren L, Li S, et al. The biology, function, and applications of exosomes in cancer. Acta Pharm Sin B 2021;11:2783-97. [Crossref] [PubMed]
44. Donisi C, Puzzoni M, Ziranu P, et al. Immune Checkpoint Inhibitors in the Treatment of HCC. Front Oncol 2020;10:601240. [Crossref] [PubMed]
45. Rezaee-Zavareh MS, Yang JD. Triple combination of hepatic arterial infusion chemotherapy, immune checkpoint inhibitors, and tyrosine kinase inhibitors for treatment of advanced hepatocellular carcinoma: more robust evidence is still needed. Chin Clin Oncol 2024;13:77. [Crossref] [PubMed]
46. Sharma R, Pillai A, Marron TU, et al. Patterns and outcomes of subsequent therapy after immune checkpoint inhibitor discontinuation in HCC. Hepatol Commun 2022;6:1776-85. [Crossref] [PubMed]