Multi-omics and machine learning refine HCC molecular subtypes and prognosis based on liquid–liquid phase separation related genes

Abstract

Accumulating evidence has demonstrated that biological processes associated with liquid–liquid phase separation (LLPS) play a critical role in cancer development. However, the effect of LLPS on hepatocellular carcinoma (HCC) remains largely unknown. In this study, we integrated consensus clustering with an ensemble machine learning framework to establish robust LLPS-related molecular subtypes and a consensus machine learning–driven LLPS-related signature (CMLLS) for HCC. The consensus clustering robustly identified three fundamental LLPS-driven subtypes (LS1–LS3), and the subsequent machine learning integration, which encompassed 101 algorithm combinations, objectively identified the most generalizable prognostic signature from multiple candidate genes. Our analysis revealed that LS3 exhibits the worst prognosis, significant upregulation of cell cycle and epithelial–mesenchymal transition (EMT)–related pathways, and enhanced immune resistance. Conversely, LS2 displays the best prognosis, enrichment in metabolism-related pathways, and increased sensitivity to immunotherapy. The CMLLS demonstrated robust predictive performance in prognostic stratification and effectively distinguished patients who would benefit from immunotherapy. This study provides novel insights into HCC heterogeneity at the LLPS level and offers a powerful tool for individualized treatment decision-making.

Supplementary Information

The online version contains supplementary material available at 10.1038/s41598-025-28551-z.

Introduction

Liver cancer is a malignant tumor that poses a serious threat to human health and is the third leading cause of cancer-related deaths globally¹. Primary hepatocellular carcinoma (HCC) is the main histologic type of liver cancer. The incidence of HCC involves a variety of complex biological processes and the interaction of many different causative factors, including genetic alterations, viral infections, alcohol, and environmental exposures (e.g., aflatoxin)².

Tumor heterogeneity in HCC is a huge obstacle to precision medicine and individualized treatment for HCC patients³. Previous views considered that genomic alterations contribute to the hallmark characteristics of specific cancers. A rising number of studies have demonstrated that intrinsically disordered regions (IDRs) of biological macromolecules have a non-negligible effect on the generation of malignant phenotypes in cancer. However, the biological process of liquid-liquid phase separation (LLPS) is an important mechanism for regulating the function of proteins with IDRs^4–6.

LLPS is a key process in the dynamic binding of biomolecules and the formation of membrane-less organelles, which plays a crucial role in regulating biomolecular interactions and organelle functions, involving scaffolds, regulators, and clients. It forms liquid condensates of biomolecules without membrane encapsulation through weak multivalent interactions and is a key mechanism for the formation of membrane-less intracellular organelles, including nucleolus, stress granules (SGs), promyelocytic leukemia (PML)-body, processing bodies (P-bodies), Cajal bodies, etc⁷. Notably, LLPS participates in the formation of multiple features of cancer, including leading to sustained cancer cell proliferation, increased cancer cell death resistance, telomere maintenance, and DNA damage repair⁸. Previous studies have indicated that LLPS-related gene signatures can predict the prognosis of a wide range of tumors and provide a useful basis for individualized treatment of tumors^9–12. In hepatocellular carcinoma, GAS2L3 accelerates cytokinesis and enhances tumorigenesis by forming CHMP4B condensates at the midbody through a phase separation scaffold¹³. RIOK1 mediates liquid-liquid phase separation by incorporating IGF2BP1 and G3BP1 into stress granules, thereby sequestering PTEN messenger RNA to reduce its translation and consequently promoting tyrosine kinase inhibitors (TKI) resistance in HCC¹⁴. EEF1E1 can undergo LLPS condensation, which enhances the expression of the DNA damage marker γH2AX by activating the PTEN/AKT pathway, thereby augmenting the stemness of HCC cells¹⁵. Overall, the in-depth progress of LLPS research in tumors, will help to further understand the pathogenesis of tumors, accurately stratify patients, and develop diversified anticancer therapies.

In this study, a comprehensive investigation was conducted on LLPS-related genes and their association with HCC. Three molecular subtypes of HCC related to LLPS were ultimately identified, demonstrating significant prognostic differences. Correlations with published subtypes, clinical characteristics, genomic alterations, tumor immune microenvironment, and potential therapeutic approaches were examined. The robustness of LLPS-related genes driving molecular subtypes in HCC was confirmed through validation across three independent multicenter cohorts. Furthermore, consensus LLPS-related prognostic genes (LRGs) were determined based on distinct cohorts, and a consensus machine learning-driven LLPS-related signature (CMLLS) was constructed by integrating ten machine learning algorithms. CMLLS maintained strong prognostic value in both training and validation cohorts and enabled accurate stratification of patients based on their response to immunotherapy. Our research contributes to understanding HCC heterogeneity and optimizing the therapeutic decision strategy for HCC patients.

Materials and methods

Data acquisition and processing

Overview of the workflow of this study (Fig. 1).

Fig. 1.

Overview of the analysis workflow for this study. (Created with BioRender.com).

Human HCC cohorts

The four independent cohorts of HCC, including TCGA-LIHC, GSE14520, GSE54236 and ICGC-LIRI were systematically collected from the publicly available databases TCGA (http://cancergenome.nih.gov/), the Gene Expression Omnibus (GEO, https://www.ncbi.nlm.nih.gov/) and the International Cancer Genome Consortium (ICGC, https://dcc.icgc.org/) databases. Gene expression data and matching phenotype information were gathered for each cohort in parallel. Please refer to Supplementary materials and methods Doc. S1 for the data processing details. In the end, 894 samples were used in the analyses that followed. The four cohorts’ baseline data are shown in Table S1.

Immunotherapy cohorts

The study enrolled four immunotherapy cohorts from the GEO database including GSE35604, GSE93157, GSE100797 and GSE126044 were used to evaluate immune response trends. Responders included complete response (CR) or partial response (PR) and non-responders included stable disease (SD) or progressive disease (PD), the Response Evaluation Criteria In Solid Tumors (RECIST, v1.1)¹⁶.

LLPS-related genes

The data resource of liquid–liquid phase separation (DrLLPS, http://llps.biocuckoo.cn/) is an integrated database for the proteins involved in the LLPS process¹⁷. The database contains 9,285 experimentally confirmed LLPS-related proteins. Following Homo sapiens screening parameters, 3633 LLPS-related genes were retrieved (Table S2). Of these, 3321 genes with available gene expression data in the TCGA cohort were selected for inclusion in subsequent analyses.

Identification and validation of LLPS-related subtypes (LS)

Univariate Cox regression analysis based on 3321 LLPS-related genes identified 670 prognostic LLPS-related genes (PLRGs) in the TCGA-LIHC cohort (Table S3). A consensus clustering algorithm based on PLRGs identified three LLPS-related subtypes (LS). We validated the robustness of the subtyping by NTP algorithm and submap analysis in 3 external cohorts. The genes were sorted in descending order according to log2 fold change (logFC), and the top 500 genes were utilized as subtype feature genes to prevent instability of LS caused by overfitting and underfitting of the model (Table S4).

Human HCC tissue samples

Tumor tissues and paired normal liver tissues were taken from 19 patients at the First Affiliated Hospital of Zhengzhou University during 2019–2021. All of these patients were confirmed as HCC patients by pathologic investigations and clinical diagnosis. The study was conducted in accordance with the Declaration of Helsinki and was approved by the Ethics Committee of the First Affiliated Hospital of Zhengzhou University, with all patients signing a written informed consent (2019-KY-21). All elements related to this study were conducted under the supervision and guidance of the ethics committee.

Statistical analysis

The statistical analysis of this research was accomplished based on R software (version 4.1.1) and GraphPad Prism (version 9.0). For multiple population comparisons, the chi-square test was applied to assess the correlation between categorical variables, while the Kruskal-Wallis rank-sum test was performed to evaluate the variations between continuous variables. The Wilcoxon rank sum test or Student’s t-test was used to assess the correlation between the two continuous variables. The optimal cutoff value for CMLLS scores was determined via the “surv-cutpoint” function of the survminer package. Kaplan–Meier (K–M) analysis with the log-rank test and Cox regression in the survival and survminer packages were used to distinguish survival differences. For all unadjusted comparisons, a two-tailed P < 0.05 was regarded as statistically significant.

Results

Identification and comprehensive analysis of prognostic LLPS-related genes

A total of 3633 LLPS-related genes were collected in the DrLLPS database, and 3321 genes with expression data in the TCGA-LIHC cohort were selected for inclusion in the univariate Cox regression analysis (P < 0.001). The expression patterns of LLPS-related genes in tumor and normal tissues were found to be significantly different in the TCGA-LIHC cohort (Fig. S1A). We obtained 670 PLRGs that were significantly associated with survival in HCC patients. Additionally, functional enrichment analysis revealed that PLRGs were mainly enriched in processes related to cell cycle, proliferation and metabolism (Fig. S1B).

Discovery of liquid–liquid phase separation-related molecular subtypes

The samples of TCGA cohort were categorized by consensus clustering via ConsensusClusterPlus package. According to the consensus clustering matrix heatmap (K ranked 2 to 9), CDF curve, delta area plot for K, and referring to the result of the PAC analysis, the optimal number of clusters was determined (K = 3) (Fig. 2A,B and S1C,D). Ultimately, three LLPS-related subtypes (LS) were identified within the TCGA cohort, LS1 (n = 141), LS2 (n = 126) and LS3 (n = 74). Principal component analysis (PCA) revealed significant differences among LS groups distinguished on the basis of 670 prognostic LLPS-related genes (Fig. 2C). We found that LS3 displayed a higher preference for the expression of PLRGs (Fig. 2D). Furthermore, K–M survival analysis indicated a significant difference in overall survival (OS) among the 3 subtypes, as evidenced by the worst prognosis for LS3 and the best prognosis for LS2 (P < 0.001, Fig. 2E).

Fig. 2.

Identification and validation of three subtypes of HCC based on 670 prognosis LLPS-related genes. (A) Sample clustering heatmap of the TCGA cohort. (B) The curves of proportion of ambiguous clustering for clustering numbers (K = 2–9). (C) PCA plot to distinguish LLPS clusters. (D) Heatmap shows the expression patterns of 670 prognostic LLPS-related genes in HCC subtypes identified by consensus clustering. (E) Kaplan–Meier (K–M) survival curves for overall survival (OS) of HCC subtypes in the TCGA cohort. (F-H) K–M survival curves for OS of HCC subtypes in validation cohorts (GSE14520, GSE54236, ICGC-LIRI), respectively. (I) SubMap analysis reveals a significant correlation of gene expression profiles between HCC subtypes of TCGA and validation datasets.

Comprehensive validation of LS classification

A comprehensive validation of the LS classification was performed in three independent cohorts to confirm the robustness and accuracy of the classification. Based on the characteristic genes of each subtype, the NTP algorithm was used to predict the categorization of HCC samples in three separate cohorts. Consistent with the expected results, each dataset could be identified into three clusters and the signature genes had similar expression patterns in the corresponding subtypes of the TCGA and the validation cohorts (Figure S2A–C). Similarly, K–M survival analysis revealed that the prognostic characteristics of the three clusters of the validation cohorts were consistent with the TCGA cohort (P < 0.001, Fig. 2F–H). SubMap analysis provided additional evidence that the validation cohort clusters’ gene expression profiles were strikingly similar to those of the TCGA cohort. (Fig. 2I).

Clinicopathologic features of LS classification

To understand the clinicopathologic features of the distinct subtypes, the clinical features of the subtypes were compared in each dataset separately based on pre-existing clinical information. We observed that LS3 was associated with more advanced stages (III or IV) and higher histologic grades (G3 or G4) compared with LS1 and LS2 (P < 0.05), which was consistent with LS3 having the most pessimistic prognostic feature. Five widely accepted molecular subtypes of HCC, the classifications of Lee, Hoshida, Boyault, Chiang and TCGA iCluster^18–22 were collected to assess the relevance of these classifications to LS (Fig. 3A,B and Figure S3A). In the correlation analysis, LS3 was observed to be associated with the SURVIVAL_DN subclass in Lee’s classification, the S1 and S2 subclasses in Hoshida’s classification, the G3 subclass in Boyault’s classification, and the PROLIFERATION subclass in Chiang’s classification (P < 0.001). The LS2 displayed a correlation with SURVIVAL_UP in Lee’s classification, S3 in Hoshida’s classification, G56 in Boyault’s classification, and CTNNB1 in Chiang’s classification, while LS1 showed a wide distribution of these subclasses (P < 0.001). In the TCGA cohort, LS1/2/3 correspond to iC3/2/1 of the TCGA iCluster subtype, respectively (P < 0.001). However, by observing the distribution of HCC patient samples, it was noted that the LS classification could synthesize the characteristics of several subtypes more clearly than the TCGA iCluster classification. Furthermore, LS3 was found to be an independent prognosis risk factor by multivariate Cox regression analysis, while LS1 and LS2 were found to be prognostic protective factors. This also yielded consistent results across the three external datasets (Figure S3B,C).

Fig. 3.

Comparison of clinical and cancer hallmarks characterization of HCC subtypes. (A,B) Correlations of HCC subtypes with clinical features and previous HCC classifications in TCGA and GSE14520 cohorts, respectively. (C) Heatmap illustrates the ssGSEA Z-scores of 50 hallmarks among HCC subtypes. (D) The enrichment of 19 oncogenic pathways in distinct HCC subtypes. Statistical p values were calculated by Kruskal–Wallis rank sum test for multiple comparison. ‘ns’ represents no significance, ∗P < 0.05, ∗∗P < 0.01, ∗∗∗P < 0.001, ∗∗∗∗P < 0.0001.

Comprehensive analysis of the biological peculiarities within LS clusters

To decipher the pertinent biological characteristics of various LSs, we first carried out functional enrichment analysis on the signature genes of various subtypes. (Figure S2D–F). Specifically, the pathways linked to metabolism, such as those involving lipid and amino acid metabolism, were primarily enriched in LS2. LS3 was mainly enriched in cell cycle-related processes and epithelial-mesenchymal transition (EMT) pathways. Notably, LS1 combined some of the characteristics of both. Subsequently, the levels of 50 hallmarkers in different clusters were quantified by ssGSEA (Fig. 3C). LS2 was enriched in a variety of metabolism-related pathways, such as fatty acid, glycolysis, and bile acid metabolism. LS3 was significantly enriched in wntβ-catenin, Notch, TGF-β, as well as cell cycle-related pathways (E2F, G2M, PI3K-AKT-mTOR), which indicated that stemness, immunosuppression, and high proliferative features were abundant in LS3. However, LS1 showed moderate levels of enrichment in multiple pathways.

Furthermore, the enrichment levels of LS in oncogenic pathways, stromal activation, mismatch repair and immune response-related (SA-MR-IR) processes, and multiple metabolic processes were investigated to fully characterize the distinct subtypes. Analysis of the enrichment levels of 19 oncogenic pathways revealed that LS3 was significantly enriched in the activated cell cycle, Hippo, Notch, TP53 and Wnt pathways, which confirmed the enhanced proliferative and invasive properties of LS3 (Fig. 3D). In SA-MR-IR analysis (Fig. S4A), LS2 was found to be upregulated in immune response-related processes (CD8 + T effector and antigen processing machinery) compared to LS3. The LS3 was significantly enriched in processes such as cell cycle, mismatch repair, homologous recombination, DNA and nucleotide repair, whereas LS2 was significantly downregulated in these processes. Interestingly, LS1 remains intermediate in these processes, indicating that LS1 may be a mixed subtype. Additionally, in metabolic characterization, it was observed that LS2 showed the most active generalized enrichment of metabolic processes, followed by LS1, while LS3 appeared to be a metabolically “cold” subtype (Fig. S4B).

Genomic features of distinct subtypes

We mapped the mutational landscape of LS to decode the differences among the three subtypes at the genomic level. Specifically, the mutation frequency of TP53 was significantly upregulated in LS3; CTNNB1 and ALB were mutated most frequently in LS2; AXIN1 and ARID1A exhibited the most frequent mutations in LS1, while LRP1B showed more frequent mutations in both LS1 and LS3. Of the top 20 segments with the most frequent CNV variants, only the deletions of 17p13.3 and 17p11.2 displayed significant differences among the three subtypes and were most prominent in LS3 (P < 0.05, Fig. 4A). We further explored the effect of CNV alterations on subtypes through FGA, FGG and FGL. Compared to LS3 and LS1, the distribution of FGA, FGG and FGL was lowest in LS2, indicating that LS2 was minimally affected by CNV and possessed a more stable genome (P < 0.001, Fig. 4B). In the mutational signature analysis, it was found that all subtypes exhibited strong correlations with signature 16 and signature 1B, which was consistent with a previous study²³, in which signature 16 was a unique mutational signature of hepatocellular carcinoma, demonstrating that hepatocellular carcinoma tissues exhibit a strong transcriptional strand bias of unknown origin, which may provide new insights into the pathogenic mechanisms of HCC.

Fig. 4.

Genomic alteration landscape of HCC subtypes in TCGA cohort. (A) Samples were sorted in descending order according to the cumulative contribution of the HCC-associated mutational signature 16. The TMB, weights for different mutation signatures, somatic mutational landscape of the top 20 frequently mutated genes and the CNVs landscape of the top 20 AMP and Homdel chromosome fragments for HCC subtypes are shown from the top to the bottom panels. The proportion of these variants in the HCC subtypes is denoted on the right. (B) Distribution of fraction genome altered (FGA) and fraction genome gain/loss (FGG/FGL). Bar charts are presented as the mean ± standard error of the mean. Statistical p values were calculated by Kruskal–Wallis rank sum test for multiple comparison. ‘ns’ represents no significance, ∗P < 0.05, ∗∗P < 0.01, ∗∗∗P < 0.001.

Assessment of the immune features of LLPS-related subtypes

Cancer immunity plays an indispensable role in the development and progression of cancer. To understand the potential relevance of LS classification with the tumor immune microenvironment (TIME) and immunotherapeutic responses, an in-depth investigation of the immune landscapes of the three subtypes was conducted. First, the infiltration abundance of 28 immune cells across the three subtypes was assessed using the GSVA algorithm. Specifically, patients classified as LS2 exhibited the greatest total immune cell infiltration abundance, especially of activated CD8 + T cells, in contrast to patients in the LS3 category, who had the lowest infiltration abundance (Fig. 5A). Furthermore, in the ICP analysis, it was revealed that LS2 displayed a significantly downregulated most of the expression levels of the ICPs, while ICP expression levels were significantly upregulated in LS3 (Fig. 5B). Among the co-stimulatory molecules, members of the tumor necrosis factor receptor (TNFR) superfamily exhibited significantly up-regulated levels, and these molecules may provide new options for immunotherapy of LS3. (P < 0.05, Fig. 5C). Some molecules that are recurrently used in clinical immunotherapy practice, such as PDCD1 and CTLA-4, also showed high levels of distribution in LS3 (P < 0.05, Fig. 5D). Therefore, it is possible that targeting these specific molecules could benefit patients in the LS3 category.

Fig. 5.

Differential immune profiles and prediction of potential capacity for immune evasion of distinct HCC subtypes. (A) Heatmap displays immunograms in the TCGA cohort, with the top panel displaying the immune enrichment score, stromal enrichment score, and the enrichment levels of the 28 microenvironmental cell types are annotated at the bottom panel. (B) Heatmap of expression profiles for immune checkpoint-related genes of HCC subtypes. (C,D) Boxplot of co-stimulatory and co-inhibitory immune checkpoint profiles for HCC subtypes, respectively. (E–H) The distribution of T-cell inflammatory signature score (TIS) (E), antigen presentation score (APS) (F), pan-fibroblast transforming growth factor b (TGF-β) response signature score (G), cancer/testis-antigens (CTA) score (H) in three HCC subtypes. Statistical p-values were calculated by Wilcox test for two-sample comparisons and Kruskal–Wallis rank sum test for multiple comparisons. ‘ns’ represents no significance, ∗P < 0.05, ∗∗P < 0.01, ∗∗∗P < 0.001, ∗∗∗∗P < 0.0001.

Immune checkpoint blocking therapy as an auspicious form of tumor therapy has significantly advanced anti-tumor therapy. Therefore, we predicted the possible benefit of different subtypes in response to ICI treatment by several immune-related indicators. Consistently, TIS and APS showed the highest distribution in LS2 subtypes and the lowest in LS3 subtypes (Fig. 5E,F). The higher TIS and APS indicated that patients in the LS2 category shared strong T-cell activity and antigen-presenting capacity. Notably, pan-TGF-βsignature response and CTA score were particularly prominent in LS3, whereas they were lowest in LS2, the former indicating a greater likelihood of TGF pathway-mediated immunotherapy resistance and immune evasion in LS3, and the latter emphasizing that patients in the LS3 category showed a stronger capacity for tumor progression (Fig. 5G,H). In addition, the LS2 subtype exhibited a significant up-regulation of MSI, which is considered an indicator of the effectiveness of immunotherapy, while LS3 displayed the highest TIDE scores, suggesting a stronger immune evasion capacity (Fig. 6A,B). Likewise, TIDE analysis showed that LS2 had a significantly higher rate of immune response than the other subtypes (P < 0.05, Fig. 6C). These features indicated that LS2, which may benefit more substantially from immunotherapy, was an immune-activating subtype, whereas LS3, characterized by a low infiltration of immune cells and a low responsiveness of the ICB, we defined as an immune-excluding subtype. The LS1 subtype, which exhibited an intermediate state in all of these features, supported the opinion that LS1 was a mixed subtype.

Fig. 6.

Correlation of HCC subtypes with potential therapeutic response and candidate therapeutic agents for LS3 Patients. (A,B) Comparison of MSI and TIDE scores across subtypes. (C) Distribution of immunotherapy response outcomes across the three subtypes of TCGA cohort predicted by the TIDE algorithm. (D,E) Response of our classification in the TACE cohort (D) and sorafenib cohort (E). (F) Schematic outlining strategies for identifying agents with higher drug sensitivity in patients with LS3. (G,H) Consistency and robustness of two prediction models (CTRP and PRISM), the low expression of STAT3 correlates with sensitivity to sorafenib. (I,J) Potential therapeutic agents for patients with LS3. Statistical p-values were calculated by Wilcox test for two-sample comparisons and Kruskal-Wallis rank sum test for multiple comparisons. The error bars indicate 95% confidence interval (CI). ‘ns’ represents no significance, ∗P < 0.05, ∗∗P < 0.01, ∗∗∗P < 0.001, ∗∗∗∗P < 0.0001.

Correlation of three subtypes with classical treatment modalities for HCC

In addition, we incorporated two public cohorts, the transcatheter arterial chemoembolization (TACE)-treated cohort (GSE104580) and the sorafenib-treated cohort (GSE109211)²⁴, which were utilized to correlate our molecular subtypes with these two classical treatment modalities. Patients in both cohorts were classified as responsive and nonresponsive based on whether they responded or not after receiving treatment. Patients from the two cohorts were categorized into three LS subtypes by the NTP algorithm based on 1500 subtype signature genes. We observed the largest number of patients responding to TACE therapy in the LS2 subtype, while the majority of patients responding to sorafenib were classified as LS3 (P < 0.001, Fig. 6D,E).

Potential therapeutic agents for the LS3

Since patients in the LS3 category show the worst prognosis and have a high possibility of immune tolerance. Through differential drug response analysis, drugs with the lowest AUC compared to the other two subtypes were screened for prospective therapeutic agents targeting the LS3 subtype to provide theoretical support for the clinical development of new therapeutic agents. To ensure the accuracy of the obtained drug sensitivity information, sorafenib, a drug widely used in the treatment of HCC was investigated to check whether the predicted drug sensitivity data were consistent with the clinical efficacy. Previous studies revealed that low STAT3 expression or dysfunction predicted higher sensitivity to sorafenib in HCC patients²⁵. Patients in the TCGA cohort were categorized into STAT3 high-expression and STAT3 low-expression subgroups by median STAT3 expression level. Consistent with the published findings, both in the CTRP database and the PRISM database, the AUC of sorafenib was lower in the STAT3 low-expression subgroup, indicating a higher accuracy of the predicted drug sensitivity data (P < 0.001, Fig. 6G,H). The patients with LS1 and LS2 in the TCGA cohort were combined and candidate drugs were obtained based on two screening strategies (Fig. 6F). Finally, we screened 5 CTRPderived agents (paclitaxel, SB-743921, daporinad, gemcitabine and BI-2536; Fig. 6I) and 5 PRISM-derived agents (vindesine, LY2606368, YM-155, verubulin and dolastatin-10; Fig. 6J).

Construction of CMLLS

The consensus prognostic genes obtained from univariate Cox analysis were intersected with 670 PLRGs, which finally generated 81 consensus PLRGs. Subsequently, 81 consensus PLRGs were included in the ensemble framework to perform CMLLS. Among the 101 models, the lasso and superpc approach produces the best signature, maintaining the highest average C-index (0.702) in the external validation cohorts (Fig. 7A). The model consisted of 10 hub genes (CDCA8, DCAF13, FTCD, GTPBP4, MRTO4, NOL10, RRP12, SAC3D1, TRIB3 and TTK), and risk scores were further calculated for each patient based on gene expression and weighted regression coefficients (Fig. S5A,B). Patients from the four cohorts were categorized into high and low CMLLS subgroups according to the optimal cutoff value for CMLLS. K–M survival analysis revealed worse OS in the high CMLLS subgroup of patients (Fig. 7B and Fig. S5C,G,K). The high accuracy of CMLLS in assessing OS at 1,2,3 years was confirmed by the ROC curves (Fig. 7C and Fig. S5D,H,L). Sankey plots showing altered attributes of HCC patients in LS classification and CMLLS subgroups (Fig. 7D and Fig. S5E,I,M). Interestingly, we observed that the vast majority of patients in the LS3 category were categorized into the high CMLLS subgroup, suggesting that these 10 hub genes appear to have the potential to differentiate patients in the LS3 category. Consistently, patients in the LS3 category had the highest CMLLS scores (Fig. 7E and Fig. S5F,J,N). Furthermore, the high CMLLS group was revealed by multivariate Cox regression analysis to be an independent prognostic risk factor (Fig. S6A–D).

Fig. 7.

Construction of consensus machine learning-driven LLPS-related signature (CMLLS) in HCC patients. (A) Through a comprehensive computational framework, a combination of 101 machine learning algorithms was generated. The C-index of each model was calculated through TCGA, GSE14520, GSE54236 and ICGC-LIRI cohorts and ranked according to the average C-index of the three validation datasets. (B) Survival analysis of HCC patients with high and low CMLLS groups in TCGA cohort. (C) Receiver operating characteristic (ROC) curves demonstrate the accuracy of CMLLS groups in predicting 1, 3, and 5-year OS. (D) Sankey plots showing property changes in LLPS subtypes and CMLLS groups. (E) Distribution of CMLLS groups among LS classification. (F) Distribution of immunotherapy response outcomes in CMLLS groups predicted by the TIDE algorithm. (G) Comparison of TIDE scores across CMLLS groups. (H) SubMap analysis reveals the similarity of GEPs between CMLLS groups and 4 GEO immunotherapy cohorts (P < 0.05 were marked). Statistical p-values were calculated by Wilcox test for two-sample comparisons. ∗P < 0.05, ∗∗P < 0.01, ∗∗∗P < 0.001, ∗∗∗∗P < 0.0001.

Low CMLLS scores predicted better immune responses

Firstly, we assessed differences in the immune landscape of different CMLLS subgroups. Specifically, the high CMLLS group had a lower abundance of immune cell infiltration and higher levels of immune checkpoint expression (Fig. S7A,B). In addition, based on TIDE and SubMap, we further validated the predictive performance of CMLLS in immune responses. We systematically collected four immunotherapy cohorts respectively GSE35640 (anti-MAGE-A3), GSE93157 (anti-PD-1), GSE100797 (CAR-T) and GSE126044 (anti-PD-1). We observed that patients in the low CMLLS group had more efficient immune responses, lower TIDE scores (Fig. 7F,G), and better efficacy of immunotherapy (Fig. 7H), suggesting a strong stability of CMLLS in predicting HCC immune responses. Conversely, poor immunotherapy benefits were predicted for the high CMLLS subgroup.

The expression levels of selected CMLLS hub genes

Since the majority of patients in the LS3 category were categorized in the high CMLLS group, we sought in turn to further explore the value of the 10 CMLLS hub genes by whether they could identify patients in the LS3 category. Univariate Cox analysis obtained a consistent trend in risk values for each CMLLS hub gene across all cohorts (Fig. S8A). Except for FTCD, the high expression of all the other genes is a prognostic risk factor for HCC patients. Given that LS3 carried the worst prognosis, we calculated the area under the ROC curve for identifying LS3 patients based on gene expression of hub genes in the TCGA cohort (Fig. S8B). Notably, the performance of these genes was excellent, especially CDCA8, FTCD, GTPBP4, MRTO4, NOL10 and TTK, with an AUC value of more than 0.8. We also calculated a feature score for each hub gene in CMLLS (Fig. S8C). Among these genes, the effects of DCAF13 and MRTO4 in hepatocellular carcinoma are not clear; therefore, we first looked at the prognostic relevance of DCAF13 and MRTO4 in patients with HCC. In the four cohorts, patients were categorized into high and low groups based on median DCAF13 or MRTO4 expression, and patients in the high expression group were consistently found to suffer from a poor prognosis (P < 0.05, Figure S8D–K). The expression of both genes was higher in tumor samples from HCC patients (P < 0.05, Fig. S8L–Q).

Biological function of MRTO4 and DCAF13

IHC staining was performed to assess the protein expression levels of MRTO4 and DCAF13, and differential analysis was carried out via IOD/area. The expression of MRTO4 and DCAF13 was significantly upregulated in HCC tissues compared to non-tumor tissues (P < 0.05, Fig. 8A–D). Further, qPCR results of tissue samples revealed that the transcript levels of MRTO4 and DCAF13 were upregulated in HCC tissues compared to non-tumor tissues (P < 0.05, Fig. 8E,F). Since LS3 was significantly enriched with proliferation- and stemness-related pathways (EMT and Notch), as well as MRTO4 and DCAF13 were significantly associated with LS3. We performed several experiments to explore the biological functions of MRTO4 and DCAF13 in HCC cell lines. MHCC97H and Huh7 cell lines were overexpressed using plasmids carrying MRTO4 or DCAF13. SiRNA transfection was used to construct MHCC97H and Huh7 knockdown cell lines. We first identified the transfection efficiency by RT-qPCR (Fig. S9A–D). Wound healing assays demonstrated that MRTO4 and DCAF13 were able to significantly promote the migratory ability of MHCC97H and Huh7 cell lines (Fig. 8G,H and Fig. S10A,B). Transwell assay further indicated that overexpression of MRTO4 or DCAF13 significantly promoted the migration and invasion ability of HCC cell lines, while knockdown of MRTO4 or DCAF13 significantly inhibited the migration and invasion ability of HCC cell lines (Fig. 8I,J and Fig. S10C,D). Colony formation assays revealed that elevated expression of MRTO4 and DCAF13 resulted in the formation of more and larger colonies in MHCC97H and Huh7 cells. In contrast, knockdown of MRTO4 and DCAF13 significantly inhibited the colony forming ability of the cell lines (Fig. 8K,L and Fig. S10E,F). The results of the EdU incorporation assay revealed that overexpression of MRTO4 and DCAF13 significantly increased the percentage of EdU-positive cells in both HCC cell lines, whereas decreased expression of MRTO4 and DCAF13 decreased the positive cell ratio (Fig. 8M,N and Fig. S10G,H). Overall, MRTO4 and DCAF13, as key factors of LS3, could enhance the proliferation, migration and invasion of HCC.

Fig. 8.

In vitro validation of the biological functions of MRTO4 and DCAF13. (A,C) Representative IHC staining images of MRTO4 (A) and DCAF13 (C). (B,D) Differences in mean IOD/Aera levels of MRTO4 and DCAF13 between tumor tissues and adjacent normal tissues of HCC patients. (E,F) Differential transcript levels of MRTO4 and DCAF13 between tumor tissues and adjacent normal tissues of HCC patients. (G,H) The migration activity of MHCC97H cells was examined via wound-healing assay after knocking down or enforcing the expression of MRTO4 (G) and DCAF13 (H). (I,J) Transwell assay in the MHCC97H cell lines. (K,L) The proliferative capacity of MRTO4 (K) and DCAF13 (L) knockdown or overexpression in MHCC97H cell lines was detected by clone formation assay. (M,N) EDU incorporation assay of the MHCC97H cell lines.

Discussion

Increasing evidence confirms the non-negligible role of the LLPS process in cancer development and progression. For example, the LLPS process can induce oncogene activation. Previous research has shown that LLPS in the NUP98-HOXA9 IDRs enhances oncogene activation, which results in mutations and carcinogenesis²⁶. Fusion proteins, which are novel products formed by chromosomal translocations, are a major cause of the development of some malignant tumors, whereas fusion proteins can promote abnormal gene expression patterns through LLPS²⁷. LLPS plays a non-negligible role in the regulatory progression of HCC. the study of He et al. found that phase separation of RNF214 promotes the progression of HCC. the CC structural domain is required for the oncogenic role of RNF214 in HCC, and the CC structural domain-mediated phase separation process dimerizes RNF214, which results in a significant enhancement of HCC invasiveness²⁸. Meng et al. revealed that the transcriptional complex formed by Twist1, YY1 and p300 could form a localized high-concentration phase-separated condensate at the super-enhancer of miR-9, which activated the expression of miR-9 and mediated the up-regulation of the EMT pathway, and thus promoted the migration and invasive ability of hepatocellular carcinoma cells, and facilitated the malignant evolution of HCC²⁹. In addition, the impact of biomolecular condensates on HCC drug resistance has important implications for revolutionizing therapeutic modalities. Meng et al. discovered that RIOK1 activates the pentose phosphate pathway by phase-separately capturing PTEN mRNA and inhibiting its translation, helping cellular stress resolution and resisting the toxicity of tyrosine kinase inhibitors (TKI). The small molecule inhibitor chidamide was able to down-regulate RIOK1 and enhance the efficacy of TKI, providing a new potential means for HCC treatment¹⁴. The study of LLPS provides new avenues for understanding the formation of malignant phenotypes in cancer, contributing to the precise stratification of patients and the development of promising anticancer targets. Here, we found that a specific collection of LLPS-related genes contributes to the classification of HCC and that each subtype possesses distinct clinical and molecular features, which was consistently validated in four multicenter cohorts containing 894 samples. Furthermore, considering individual heterogeneity, we integrated a combination of 10 machine learning algorithms to comprehensively select consensus LRGs affecting clinical outcomes in HCC patients and construct CMLLS. It was proven that CMLLS demonstrated and strong predictive ability and was able to predict the response to ICI therapy in HCC patients. To our knowledge, this is the first study to integrate LLPS with clinical subtypes of HCC, providing an important reference for understanding the development of HCC and developing new therapeutic targets.

The mutational landscape of different LSs was viewed to identify somatic mutations and CNVs that may drive LSs. Specifically, the frequency of AXIN1 and ARID1A mutations was significantly elevated in LS1. AXIN1 mutations lead to hepatocellular carcinoma through pathways independent of the Wnt/β-catenin pathway and patients have a poor prognosis³⁰. The inactivation of ARID1A is likewise associated with a poor prognosis in patients with HCC³¹. TP53 displayed high mutation frequency in LS3. According to earlier research, patients with a high TP53 mutation frequency have a worse prognosis and respond less well to immunotherapy³². The frequency of LRP1B mutations is higher in LS1 and LS3 and patients with mutations in LRP1B have a poor prognosis and down-regulation of LRP1B expression inhibits HCC progression^33,34. The frequency of CTNNB1 and ALB mutations was significantly higher in LS2. CTNNB1 mutations enhanced cell proliferation and survival and consistently showed preserved expression of various genes involved in hepatocyte differentiation and function, such as ALB, and such tumors consistently showed good differentiation^35,36. It was consistent with the biological characterization of the different LSs in the functional analysis. Moreover, by counting the FGA, FGG and FGL information of different isoforms, it was found that LS1 and LS3 were enriched with more CNVs and the genomes were more destabilized compared to LS2, suggesting that the occurrence of LS1 and LS3 may be associated with mutation accumulation.

In addition, it has been described that LLPS has a potential impact on regulating the immune microenvironment and immunotherapeutic sensitivity^37,38. Hence, it helps to understand the significance of LLPS in the formation of the TIME in HCC by characterizing the differences in immunelandscape in the three subtypes. LS2 exhibited the highest abundance of immune cell infiltration among the three subtypes, such as activated CD8 + T cells, NK cells and Th1 cells, whereas LS3 not only had a low abundance of immune cell infiltration but also exhibited a high level of expression of multiple immune checkpoint molecules. A higher abundance of immune cell infiltration is thought to have a greater reserve of immune cells and thus more likely to benefit from immunotherapy. However, overexpression of immune checkpoint molecules is commonly thought to have an immunosuppressive microenvironment, and immunotherapy in such patients tends to be less effective³⁹. Consistently, in studies of several immune-related metrics, LS2 was found to have the highest TIS and APS scores and the lowest pan-TGF-β characteristic response and CTA scores, which was opposed in LS3. This indicated that patients in the LS2 category had strong T-cell activity and antigen presentation whereas patients in the LS3 category were more immune-resistant and capable of tumor progression^40–43. In addition, TIDE analysis showed that patients with the LS2 subtype showed the highest sensitivity to immunotherapy and a low LS3 response rate. These results indicated that LS2 is an immune-activated subtype, whereas the immune profile of LS3 suggests that it is an immuno-desert subtype.

The current status of poor prognosis of patients in the LS3 category is a matter of concern, and the precise refinement of patient treatment is facilitated by correlating different treatment modalities with our subtypes. Specifically, LS2 was more likely to benefit from TACE treatment whereas LS3 would show a higher probability of response when sorafenib was applied. Subsequently, an integrated framework system was used to screen potential therapeutic agents for LS3 with the expectation of providing additional therapeutic options to improve the prognosis of LS3 patients. A comprehensive bioinformatics algorithm based on multiple pharmacological databases identified 10 potentially effective compounds, which provide new insights into the treatment of high-risk patients.

Given the ability of PLRGs to distinguish HCC patients into 3 significantly heterogeneous groups, we believe that screening biomarkers by constructing prognostic signatures is feasible for quantifying the risk of HCC patients as well as for distinguishing immunotherapy-sensitive patients. Machine learning algorithms are powerful for developing robust predictive models, and the creation of multiple machine learning algorithms has led to decisions about which algorithms will produce optimal models being largely unknown. In practice, the type of algorithm to apply depends largely on the personal subjectivity of the researcher. To circumvent this problem as much as possible, the signature that performs optimally in multiple cohorts was selected among 101 models by combining 10 machine learning algorithms. Eventually, lasso and superpc-based CMLLS were created, which had an optimal mean C-index (0.702) in four multicenter cohorts. CMLLS had good robustness in prognostically stratifying patients with HCC. Clinically, the CMLLS score effectively stratified HCC patients into high- and low-risk subgroups with distinct overall survival outcomes across multiple cohorts. Importantly, the CMLLS score was independently associated with prognosis in multivariate analyses, highlighting its potential as a robust clinical biomarker. Furthermore, the low CMLLS subgroup exhibited enhanced immune cell infiltration, lower expression of immune checkpoints, and superior responses to immune checkpoint inhibitors, as validated in four independent immunotherapy cohorts. These findings underscore the clinical utility of CMLLS in guiding prognostic evaluation and immunotherapeutic strategies for HCC patients.

Further correlate CMLLS with LS3 stratification to discover the significant value of hub genes. CMLLS includes a total of 10 LLPS-related genes, namely CDCA8, DCAF13, FTCD, GTPBP4, MRTO4, NOL10, RRP12, SAC3D1, TRIB3 and TTK. These genes have excellent performance in distinguishing patients in the LS3 subset and can serve as key factors for both CMLLS and LS3 subtypes. It was reported that CDCA8 overexpression was an independent predictor of HCC progression and knockdown of CDCA8 inhibited the malignant phenotype of HCC through the MEK/ERK pathway⁴⁴. The expression of FTCD progressively decreased at different stages of HCC progression, which was confirmed in animal experiments. In addition, FTCD deficiency significantly upregulated the expression of PPARγ and SREBP2 and their transcriptional targets, leading to lipid accumulation and HCC development⁴⁵. GTPBP4 upregulation promotes HCC growth and metastasis in vitro and in vivo and can facilitate HCC glycolytic reprogramming through the GTPBP4-PKM2 regulatory axis⁴⁶. NOL10 and RRP12 knockdown would significantly inhibit HCC cell invasion and migration during in vitro experiments^47,48. Overexpression of SAC3D1 promotes the proliferation of HCC cells and increases the level of β-catenin⁴⁹. Increased TRIB3 expression promotes the malignant progression of HCC⁵⁰. TTK overexpression is associated with poor prognosis in HCC patients and may promote HCC progression by maintaining cell cycle progression⁵¹. In our study, high expression of MRTO4 and DCAF13 was associated with poor prognosis in HCC patients. Similarly, higher transcript and protein expression levels of MRTO4 and DCAF13 in cancer tissues were observed in tissue samples from HCC patients. More importantly, MRTO4 and DCAF13 play important roles in tumor progression, migration and invasion, which was consistently verified in a series of subsequent functional analyses and multiple cellular and molecular experiments. The available studies on LLPS-related genes are almost limited to their involvement in the type of molecular cohesion formation, and further studies on the mechanisms by which various LLPS-related processes affect tumors are still required.

Although our study comprehensively developed an LLPS-based HCC classification platform and quantified the risk of different HCC patients by CMLLS, which is promising in our opinion. However, there are still some shortcomings worth noting; first, our samples are retrospective, and future validation of LS classification needs to be done in prospective fresh samples. Second, the roles of the LLPS-related genes we identified are limited to the macroscopic level, and more studies are needed to reveal the complex crosstalk between LLPS-related modalities and HCC features. In the future, a deeper understanding of the complex biological behavior of HCC is expected by studying the deeper mechanisms by which LLPS-related genes affect HCC. Looking forward, our findings pave the way for several clinical applications: (1) integration of CMLLS into clinical trials for patient stratification, (2) development of LLPS-targeted therapies for high-risk subgroups, and (3) exploration of LLPS-related biomarkers in liquid biopsies. Future work should focus on elucidating the mechanistic basis of LLPS in HCC and validating our classification in prospective, multi-institutional cohorts.

In summary, we identified three LLPS-based molecular subtypes of HCC and developed a machine learning–driven signature, CMLLS, that robustly predicts prognosis and immunotherapy response. The LS subtypes encapsulate distinct genomic, metabolic, and immune features, with LS3 representing an aggressive, immune-resistant phenotype. CMLLS effectively stratifies patients into risk groups and identifies those likely to benefit from immune checkpoint inhibition. The proposed LS classification and CMLLS provide new directions for individualized decision-making in HCC patients.

Supplementary Information

Below is the link to the electronic supplementary material.

Abbreviations

HCC

hepatocellular carcinoma

LLPS

Liquid–liquid phase separation

LRGs

LLPS-related prognostic genes

CMLLS

Consensus machine learning-driven LLPS-related signature

CNVs

Copy number variations

LS

LLPS-related subtypes

PLRGs

Prognostic LLPS-related genes

NTP

Nearest template prediction

Submap

Subclass mapping

FGA

Fraction of genomic alterations

FGG

Fraction of genomes gained

FGL

Fraction of genomes lost

ICPs

Immune checkpoints

ICI

Immune checkpoint inhibitor

TIS

Tumor inflammation signature

APS

Antigen presentation score

CTAs

Cancer/testis-antigens

TIDE

Tumor immune dysfunction and exclusion

AUC

Area under the dose-response curve

ROC

Receiver operating characteristic

OS

Overall survival

Author contributions

M.H.L., S.J.Z and Y.Z. contributed study design. M.H.L., Q.L. and L.L. integrated and analyzed the data. M.H.L. and J.G. contributed manuscript writing. S.J.Z., Y.Z. and R.L.T. contributed project oversight and manuscript revisiting. M.H.L., Z.H.W., X.Y.S. and P.H.W. contributed manuscript revisiting. All authors read and approved the final manuscript.

Funding

This study was supported by grants from the National Natural Science Foundation of China (82170670, 81971881). The authors declare that they have no competing interests.

Data availability

Public data used in this work can be acquired from the TCGA Research Network portal (https://portal.gdc.cancer.gov/), Gene Expression Omnibus (GEO, http://www.ncbi.nlm.nih.gov/geo/), and the International Cancer Genome Consortium (ICGC, https://dcc.icgc.org/).

Declarations

Competing interests

The authors declare no competing interests.

Approval of the research protocol

The studies involving human participants were reviewed and approved by the Ethics Committee of the First Affiliated Hospital of Zhengzhou University, China (2019-KY-21).

Informed consent

The patients/participants provided their written informed consent to participate in this study.