PTM-Related Signatures Predict Lymph Node Metastasis and Shape Microenvironment in Hepatocellular Carcinoma: A Single-Cell and Bulk Integrated Analysis Authors: Zhu Ai, Yuying Liang, Yuzhi Cai, Jian Gao, Lin Ren, Qiong Wang, Yutong Chen, Yongye Huang, Zhiming Xiang Published: 2026 ABSTRACT AbstractBackgroundHepatocellular carcinoma (HCC) has a poor prognosis with lymph node metastasis (LNM), severely impacting survival. Protein post-translational modifications (PTMs) regulate tumor microenvironment (TME) dynamics, but their role in HCC LNM is unknown.MethodsThis study systematically analyzed single-cell RNA sequencing (scRNA-seq) and bulk transcriptomic data from TCGA-LIHC and validation cohorts. Key cellular subpopulations and differentially expressed genes (DEGs) were identified. PTM-related candidate genes were identified through intersection analysis. Prognostic genes were determined via Cox regression and integrated into a random survival forest (RSF) model. Patients were stratified to assess associations with TME remodeling, genomic alterations, immune checkpoints, and drug sensitivity. Pseudotime trajectory and transcription factor activity analyses elucidated molecular dynamics.ResultsHepatocyte populations were significantly expanded in both LNM and non-LNM HCC groups. The intersection of scRNA-seq DEGs, bulk DEGs, and PTM-related genes identified five candidate genes, among which PJA1 and RFPL4B identified as potential prognostic risk factors and were incorporated into a robust RSF model. High-risk patients exhibited poor survival outcomes, increased tumor mutational burden, and pronounced tumor microenvironment alterations. PJA1 expression was negatively correlated with hepatocyte infiltration. Pseudotime analysis revealed that LNM-associated hepatocytes were arrested in early differentiation states, coinciding with dysregulated transcription factor activity.ConclusionThis study computationally identified PJA1 and RFPL4B as post-translational modification–related candidate prognostic genes potentially associated with HCC LNM. The proposed risk model stratified patients into subgroups characterized by distinct tumor microenvironment features and genomic instability. Computational analysis suggested that arrest of hepatocyte differentiation and dysregulation of transcription factor networks may be implicated in metastatic progression, thereby offering novel therapeutic insights.Keywords: hepatocellular carcinoma, lymph node metastasis, post-translational modifications, tumor microenvironment Introduction Characterized by aggressive invasiveness and extensive metastatic potential, Hepatocellular carcinoma (HCC) ranks among the foremost contributors to global cancer mortality, presenting formidable challenges in clinical management.1 Lymphangiogenesis and lymph node metastasis (LNM) are key hallmarks of progression in hepatobiliary malignancies such as HCC, and are closely associated with a significantly worsened prognosis for patients.2 The liver, as the primary source of lymphatic fluid within the body (accounting for 25% to 50%), possesses a dense lymphatic network and plays a central role in maintaining fluid balance and regulating inflammatory and immune responses.3 It is particularly noteworthy that extrahepatic metastases in HCC (such as lymph node metastases) exhibit organ-specific pathological characteristics, and once lymph node metastasis occurs, patient survival rates decline sharply.4 Although localised treatment may improve tumour control in patients with LNM,5 effective methods for the early detection and risk stratification of LNM remain lacking.6 While multiple HCC prognostic models exist (immune, metabolic, ferroptosis/cuproptosis-based), most focus on bulk tumor biology and lack cell-type resolution, limiting insights into lymph node metastasis (LNM). Importantly, no study has integrated single-cell data with PTM-related gene sets to build an LNM-focused prognostic model in HCC. Consequently, comprehensive elucidation of the molecular pathways driving LNM in HCC, coupled with the discovery of reliable prognostic indicators, represents a critical unmet clinical necessity. Recent spatial multi-omics studies show that DLL4–NOTCH3 signaling and POSTN+ CAF-derived migrasomes drive fibrotic remodeling, angiogenesis, and immune exclusion in HCC, promoting immune checkpoint blockade resistance and poor prognosis,7,8 highlighting stromal-vascular-immune crosstalk as a key feature of aggressive tumor microenvironment. Post-translational modifications (PTMs) represent a pivotal mechanism regulating immune activity within the tumour microenvironment (TME). Protein-translocation molecules directly regulate immune cell function by altering protein stability, localisation, and interaction networks, thereby influencing tumour progression.9,10 Accumulating evidence has established that protein ubiquitination, along with the opposing deubiquitination process, constitutes a crucial regulatory axis in the metastatic progression of malignancies. The deubiquitinating enzyme USP2a significantly enhances tumour cell proliferation, migration and invasive capacity in hepatocellular carcinoma by stabilising the RAB1A protein, with its overexpression directly correlating with poor patient prognosis.11 The neddylation modification of sterol regulatory element-binding protein SREBP-1 antagonises ubiquitin-mediated degradation, co-expresses with UBC12 in lymph node metastatic breast cancer, and influences survival outcomes.12 The aforementioned evidence indicates that PTMs exert a universal regulatory role in tumour metastasis. PTMs are promising biomarkers for HCC LNM as they rapidly and reversibly regulate metastasis-related signaling (eg, MAPK, p53), PTM enzymes (E3 ligases/DUBs) are frequently dysregulated in HCC and affect invasion and immune evasion, and PTM-related signatures integrate oncogenic signals into stable transcriptomic readouts suitable for validation. However, their role in HCC LNM remains insufficiently characterized. The breakthrough in single-cell RNA sequencing (scRNA-seq) technology has provided a novel perspective for deciphering tumour heterogeneity. Its single-cell resolution enables precise identification of functional cell subpopulations and gene expression dynamics associated with TMEs and tumour-tissue microenvironment modifications (PTMs), thereby overcoming the limitations of traditional bulk RNA sequencing.13,14 Through integrative bioinformatics approaches, the present investigation employed bulk RNA sequencing (bulk-RNA) and scRNA-seq to explore PTMs-related prognostic genes in HCC patients with LNM. Subsequently, a risk stratification model was developed and rigorously validated, with particular emphasis on correlations between risk categories and variables including tumor immune landscape and therapeutic agent responsiveness. Furthermore, the distribution and expression of prognostic genes were analysed within single-cell datasets. This integrative framework provides hypothesis-generating insights into PTM-mediated regulation of HCC LNM and supports identification of candidate biomarkers requiring experimental validation. To our knowledge, this is the first study integrating single-cell transcriptomics with PTM-specific gene sets to construct an LNM-focused prognostic model in HCC. Materials and Methods Data Acquisition Gene expression data were obtained from multiple public databases. First, the scRNA-seq dataset GSE149614 was downloaded from the Gene Expression Omnibus (GEO) database (https://www.ncbi.nlm.nih.gov/geo/). This dataset, generated on the GPL24676 platform, included transcriptomic profiles from 19 samples: one metastatic lymph node tissue sample from a HCC patient (designated as LNM group), eight adjacent non-tumor liver tissue samples (designated as NLNM group), and ten primary tumor tissue samples (designated as control group). All samples were sequenced using high-throughput methods. Second, transcriptomic data and clinical information for TCGA-LIHC cohort were retrieved from the UCSC-Xena database (https://xenabrowser.net/datapages/). This dataset, downloaded on July 10, 2025, comprised 374 HCC tumor tissue samples (with survival information available for 368 samples) and 50 adjacent non-tumor tissue samples, all generated via high-throughput sequencing. Clinical variables included survival data, gender, age, tumor grade, and pathological stage, as well as mutation data. Third, the microarray dataset GSE76427 was acquired from the GEO database, utilizing the GPL10558 platform; it contained gene expression profiles from 115 HCC tumor tissue samples. To minimize batch effects, bulk RNA-seq data were normalized using DESeq2 size factors; no ComBat correction was applied due to cohort homogeneity (TCGA-LIHC). Single-cell data (GSE149614) were processed within one dataset using Seurat normalization and scaling, so Harmony/ComBat integration was not required. Additionally, a set of 822 PTM-related genes was compiled from supplementary materials in a prior literature report.15 These genes were merged and deduplicated to form a candidate gene list, as detailed in Supplementary Table 1. Cell Type Annotation and Functional Enrichment Analysis Initial data processing was conducted utilizing the R software package Seurat (v 5.3.0).16 The raw gene expression matrix was first constructed using CreateSeuratObject with parameters min.cells = 3 and min.features = 200. Quality control measures were implemented to filter cellular populations according to specified parameters: cells displaying gene expression profiles below 200 or above 6,000 were eliminated; those exhibiting unique molecular identifier (UMI) counts surpassing 50,000 were excluded; additionally, cells demonstrating mitochondrial gene proportions greater than 10% were removed from subsequent analysis. The mitochondrial gene percentage (percent.mt) was calculated based on genes starting with “^MT-”. Normalization of gene expression matrices was accomplished through LogNormalize method (normalization.method = “LogNormalize”, scale.factor = 10,000). For computational efficiency optimization, genes exhibiting high variability (HVGs) were detected via the variance stabilizing transformation (vst) algorithm within Seurat package (v 5.3.0), and these identified HVGs served as the foundation for subsequent analytical procedures. Expression data underwent normalization via the ScaleData function available in the R package Seurat (v5.3.0). Dimensionality reduction was subsequently achieved through principal component analysis (PCA) implementation (RunPCA, npcs = 50). The optimal number of principal components (PCs) was determined. PCs significantly enriched for low p values were retained (p < 0.05). Based on a sharp decline in p values, some PCs were selected for further analysis. Cellular cluster identification was executed through application of the FindNeighbors (dims = 1:30) and FindClusters (resolution parameter set at 0.1, Louvain algorithm) after which uniform manifold approximation and projection (UMAP) algorithm was employed for dimensional visualization purposes (RunUMAP, dims = 1:30). Annotation of distinct cellular clusters was accomplished by leveraging marker gene signatures derived from previously published literature,17 in conjunction with the CellMarker database (https://ngdc.cncb.ac.cn/databasecommons/; accessed July 22, 2025), combined with SingleR automatic annotation for validation. UMAP projections were generated separately for control, LNM, and NLNM groups. Cell type-specific marker expression was visualized. Cell proportions were calculated as the fraction of each annotated cell type per clinical group. Compositional differences between groups were evaluated using Fisher’s exact test via 2×2 contingency tables. P-values were corrected using the Benjamini–Hochberg method, with significance defined as adjusted p < 0.05. Pathway enrichment analysis was performed using the R package “ReactomeGSA” (v 3.10) based on a GSVA/ssGSEA framework.18 Pathway activity scores were calculated using Reactome pathway gene sets (latest version). Differential pathway activity was defined using FDR-adjusted P < 0.05. The top 10 pathways exhibiting the largest differences in activity scores across cell types were visualize. Key Cell Types Screening and Differential Gene Expression Analysis Cell types across sample groups were compared to identify key cell types. Cell types exhibited significant differences in control, LNM and NLNM groups. Integrating observations and supporting literature, key cell types were identified. Analysis of differential gene expression was performed through implementation of the FindMarkers function available within the R software package Seurat (v 5.3.0). Comparisons were performed between HCC tumors vs non-tumor controls within key cell types, HCC LNM vs HCC tumors within key cell types. Genes demonstrating absolute log2-transformed fold change (|log2FC|) values exceeding 0.5 combined with p-values less than 0.05 were designated as differentially expressed genes (DEGs). Identification of shared DEGs across both comparative cohorts was accomplished utilizing the R package ggvenn (v 0.1.10).19 The common DEGs (scRNA-seq DEGs) were identified for further analysis. Identification of Candidate Genes Related to PTMs and LNM For the purpose of detecting DEGs distinguishing malignant tissues from adjacent non-neoplastic hepatic tissues, the R package DESeq2 (v 1.42.0)20 was implemented to conduct differential expression profiling. To capture subtle but coordinated transcriptional changes for downstream pathway analyses, a relatively relaxed threshold of |log2FC| > 0.25 and a p value < 0.05 was applied to define bulk DEGs. Visualization of expression profiles for detected DEGs was executed through multiple graphical representations. Volcano plot construction was accomplished via the R package ggplot2 (v 3.5.2),21 wherein the most prominently upregulated and downregulated genes (top 10 for each category) were emphasized according to their log2FC magnitude rankings. Furthermore, hierarchical clustering heatmap generation was performed utilizing the R package ComplexHeatmap (v 2.16.0)22 to display expression dynamics. To pinpoint candidate genes potentially relevant to both LNM in HCC and PTMs, an intersection analysis was conducted using the R package “ggvenn” (v 0.1.10). Genes common to all three sets were designated as candidate genes. Gene Set Enrichment Analysis (GSEA) of Candidate Genes For elucidating functional implications and pathway associations of screened candidate genes, GSEA was implemented. Initially, Spearman’s rank correlation coefficient computation was executed to assess relationships between individual candidate genes and the complete transcriptome across all sample specimens, employing the R package psych (v 2.5.3).23 Following this, comprehensive gene ranking was established according to correlation coefficient (cor) magnitudes in relation to respective candidate genes. The resulting ordered gene catalog subsequently served as input for GSEA execution through the R package clusterProfiler (v 4.10.1).21 The reference gene collection utilized for this analysis was “c2.cp.kegg.v7.4.symbols.gmt” obtained from the Molecular Signatures Database (MSigDB) repository (https://www.gsea-msigdb.org/gsea/msigdb). Biological pathways demonstrating absolute normalized enrichment score (|NES|) values greater than 1 in conjunction with nominal p-values below 0.05 were designated as statistically significant enrichments. Graphical representation of the ten most significantly enriched pathways, prioritized by enrichment statistical significance, was achieved through application of the R package enrichplot (v 1.22.0).24 Prognostic Gene Screening and Risk Model Construction For the purpose of screening prognostically relevant genes from the candidate gene pool, univariate Cox proportional hazards regression modeling was implemented within the TCGA-LIHC dataset, exclusively incorporating specimens possessing complete overall survival (OS) data. Execution of this statistical analysis was accomplished via the R package survival (v 3.8.3).25 Survival endpoints were defined as overall survival time (OS.time) and survival status (OS), and the Cox model was constructed as coxph(Surv(OS.time, OS) ~ gene_expression). Gene expression levels were treated as continuous variables and assessed for their association with patient survival time. Genes exhibiting hazard ratio (HR) values deviating from unity coupled with p-values inferior to 0.05 were classified as bearing statistical significance. HR and 95% confidence intervals were estimated using the Wald test. Visualization through forest plot construction was performed utilizing the R package forestplot (v 3.1.6).26 Validation of the proportional hazards (PH) assumption pertaining to these genes was subsequently verified using Schoenfeld residuals test (cox.zph function) with p-values exceeding 0.05 indicating no violation. Genes violating the PH assumption (global test p < 0.05) were excluded from subsequent analysis. Integration of prognostic signatures from these identified genes was thereafter achieved through development of a random survival forest (RSF) algorithm employing the R package randomForestSRC (v 3.3.3).27 Model parameters were set as follows: number of trees (ntree) = 3 and number of variables tried at each split (mtry) = 4. The splitting rule was based on log-rank splitting, and node size was set to default. A grid search strategy was not applied; parameters were empirically selected. The relative importance of the prognostic genes within the RSF model was visualized. To assess the robustness of the RSF model, we additionally reconstructed the model using more conventional parameter settings (ntree = 500, mtry = 2, minprop = 0.1). Risk Model Evaluation and Validation For validation of the prognostically-oriented risk prediction model, individual risk scores were calculated for all HCC patients within the TCGA-LIHC training cohort through application of the predict function derived from the RSF algorithm. Patient stratification into elevated-risk and reduced-risk categories was accomplished by establishing an optimal cutoff value using the surv_cutpoint function in the survminer package (minprop = 0.1). A risk curve plot was generated, ordering patients by ascending risk score, to visualize the distribution of survival time and death events across the risk spectrum. Hierarchical heatmap construction was implemented to illustrate prognostic gene expression profiles across distinct risk stratification categories within HCC specimens. Comparative survival outcome assessment between elevated-risk and reduced-risk cohorts was executed via Kaplan-Meier (K-M) survival curve analysis employing the R package survminer (v 0.5.0).25 Temporal predictive performance of the risk stratification model was evaluated through time-dependent receiver operating characteristic (ROC) curve methodology. Implementation of this analytical approach was achieved utilizing the R package timeROC (v 0.4),28 with computation of area under the curve (AUC) metrics for 1-year, 3-year, and 5-year survival prediction endpoints. The predictive performance of the established prognostic risk model was further validated using a validation cohort (GSE76427). The risk score calculation and subsequent risk group stratification procedures applied to the TCGA-LIHC cohort were identically applied to the samples within the validation cohort (GSE76427). Association of Risk Scores with Clinical Features For examination of clinical and molecular pathological characteristic distributions distinguishing elevated-risk from reduced-risk categories, K-M survival curves were constructed with stratification according to diverse clinical parameters. This analysis utilized the R package “survminer” (v 0.5.0). Clinical features assessed included age, race, T stage, N stage, M stage, and overall pathological stage. Log rank tests were performed within each clinical feature subgroup to compare survival differences between the high-risk and low-risk groups, with a significance threshold of p < 0.05. Subsequently, the correlation between clinical features and the calculated risk score was analyzed using the R package “reshape2” (v 1.4.4).29 Features exhibiting an |cor| > 0.3 and p < 0.05 were considered significantly correlated, with higher |cor| values indicating stronger correlation strength. Assessment of Transcription Factor Activity To evaluate the activity of transcription factors within the key cell type, the R package “dorothea” (v 1.14.1)30 was employed. This package utilized the DoRothEA database, which contained curated TF-target gene interactions with confidence levels (A, B, or C). Transcription factor activity scores (viper scores) were computed for each cell. The computed TF activity scores were then visualized using the R package “pheatmap” (v 1.0.12).31 Cell-Cell Communication Analysis For inferring intercellular communication networks through investigation of receptor-ligand expression patterns and pairing dynamics among pivotal cellular populations, comprehensive cell-cell interaction analysis was executed across all annotated cell types present in the scRNA-seq dataset (GSE149614). Execution of this analytical framework was accomplished via the R package CellChat (v 2.2.0).32 Furthermore, to elucidate specific ligand-receptor interactions mediating communication between distinct cell types, bubble plots were generated. Pseudotime Trajectory Analysis To understand the dynamics of prognostic gene expression during state transitions and infer the cellular evolution within HCC, pseudotime trajectory analysis was conducted specifically on key cell types. Initially, pivotal cellular populations extracted from the GSE149614 dataset were subjected to dimensionality reduction procedures followed by sub-clustering operations utilizing the R package Seurat (v 5.3.0). Following this, pseudotemporal trajectory inference was implemented on these critical cell type subsets to characterize the chronological evolution of transcriptional alterations, with particular emphasis on the screened prognostic gene candidates. The trajectory was constructed using the “DDRTree” algorithm implemented in the DDRTree package (v 0.1.5).33 GSEA Based on Risk Groups For exploration of potential functional divergences distinguishing elevated-risk from reduced-risk categories within the TCGA-LIHC cohort, GSEA was implemented. Comparative differential expression profiling contrasting elevated-risk versus reduced-risk cohorts was initially executed employing the R package DESeq2 (v 1.42.0). Genes were ranked based on their log2FC values from highest to lowest. The resulting ordered gene catalog was subsequently applied for GSEA execution via the R package clusterProfiler (v 4.10.1). The benchmark gene collection utilized was “c2.cp.kegg.v7.4.symbols.gmt” retrieved from the MSigDB. Biological pathways satisfying statistical criteria including p-values below 0.05, false discovery rate (FDR) inferior to 0.25, and absolute normalized enrichment score (|NES|) exceeding 1 were designated as significantly enriched. Graphical representation of the five most statistically significant pathways, prioritized by ascending p-value rankings, was accomplished through the R package enrichplot (v 1.22.0). Analysis of Immune Cell Infiltration For comparative assessment of immune cell infiltration patterns distinguishing elevated-risk from reduced-risk categories within the TCGA-LIHC cohort, the xCell computational algorithm was implemented. This analytical tool computed enrichment scores spanning 64 discrete immune and stromal cellular populations for individual disease specimens harboring comprehensive survival data. Subsequently, Wilcoxon rank-sum tests were employed to detect immune cellular populations exhibiting statistically significant enrichment score variations between elevated-risk and reduced-risk stratifications (p < 0.05). Visualization of enrichment score distributions for these differentially represented immune cell types across both risk categories was performed utilizing the R package ggplot2 (v 3.5.2). To explore potential interactions among the differentially infiltrated immune cells, pairwise Spearman correlation analysis was performed utilizing the cor.test function from the R package “stats” (v 4.3.3)34 (|cor| > 0.3, p < 0.05). Analysis of Somatic Mutation and Tumor Mutational Burden Somatic mutation data for HCC samples were analyzed. To visualize the mutation landscape, waterfall plots depicting the top 10 most frequently mutated genes in the entire cohort and stratified by high-risk and low-risk groups were generated using the R package “maftools” (v3.13).35 Tumor mutational burden (TMB) quantification was performed for individual specimens. Statistical evaluation of TMB score disparities distinguishing elevated-risk from reduced-risk cohorts was conducted via Wilcoxon rank-sum testing with significance threshold established at p < 0.05. Analysis of Immune Checkpoint Expression Within the TCGA-LIHC dataset, comparative analyses were conducted to assess immune checkpoint inhibitor (ICI) expression profiles between patients stratified into high-risk versus low-risk categories. The comprehensive ICI inventory was retrieved from the TISIDB repository (http://cis.hku.hk/TISIDB/index.php). Statistical evaluation of ICI expression disparities across both risk stratifications was accomplished through Wilcoxon rank-sum testing, employing p < 0.05 as the criterion for statistical significance. ICIs showing significant differential expression between two risk groups were defined as differential ICIs. Utilizing the “psych” package (v 2.5.3) in R environment, Spearman’s rank correlation coefficients were computed to examine associations between derived risk scores and expression magnitudes of differentially expressed ICIs. Correlation relationships demonstrating absolute coefficient values exceeding 0.3 alongside p-values below 0.05 were designated as statistically meaningful. Comparative evaluation of immune checkpoint (ICP) gene expression patterns was undertaken across high-risk and low-risk patient subgroups within the TCGA-LIHC cohort, with ICP gene identities sourced from supplementary data accompanying a previous study.36 Differences in ICP gene expression were assessed using the Wilcoxon test (p < 0.05). The expression distribution of significantly different ICP genes across the risk groups were visualized. Additionally, correlation analyses employing Spearman’s methodology were executed to investigate relationships among prognostic gene expression profiles, computed risk scores, and ICP gene expression quantities via the R-based “psych” package (v 2.5.3), with significance criteria defined as absolute correlation coefficients surpassing 0.3 and p-values remaining under 0.05. To further explore the potential immunomodulatory role of RFPL4B, a protein–protein interaction (PPI) network was constructed. Functionally associated genes of RFPL4B were retrieved from the GeneMANIA database (https://genemania.org/), selecting the top 20 genes based on co-expression, co-localization, pathway co-membership, and physical interactions. The resulting 21-gene set (RFPL4B plus 20 genes) was integrated with a curated immune-related gene panel covering immune checkpoints, immune cell markers, cytokines, chemokines, and antigen presentation pathways. PPI network visualization and analysis were performed using Cytoscape. Chemotherapeutic Drug Sensitivity Profiling To investigate the therapeutic efficacy of chemotherapeutic drugs in HCC patients, a comprehensive list of common chemotherapeutic agents was obtained from the Genomics of Drug Sensitivity in Cancer (GDSC) database (https://www.cancerrxgene.org/) on July 23, 2025. Within the TCGA-LIHC cohort, the half-maximal inhibitory concentration (IC50) values for 196 chemotherapeutic drugs were estimated using the R package “oncoPredict” (v 1.2).37 Drug sensitivity variations between antecedently established high-risk and low-risk cohorts, quantified through IC50 metrics, underwent statistical evaluation via Wilcoxon rank-sum tests with p < 0.05 denoting significance. Graphical representation of IC50 value distributions for respective therapeutic agents spanning both risk stratifications was generated utilizing the “ggplot2” package (v 3.5.2) within the R framework. Genomic Localization, Subcellular Localization Prediction, and Protein Expression of Prognostic Genes The chromosomal locations of the identified prognostic genes were determined based on their genomic annotation information. Visualization of the chromosomal distribution of these prognostic genes was performed using the R package “RCircos” (v 1.2.2).38 Data concerning anticipated subcellular compartmentalization of protein entities encoded by prognostic gene candidates were obtained from the GeneCards repository (https://www.genecards.org/) as of July 23, 2025, with subsequent graphical presentation accomplished through the R-based “ggplot2” package (v 3.5.2). For investigating protein expression disparities of prognostic gene products between HCC specimens and control tissues, immunohistochemical (IHC) staining imagery was acquired from the Human Protein Atlas (HPA) repository (https://www.proteinatlas.org/) on July 29, 2025. Reverse Transcription Quantitative Polymerase Chain Reaction (RT-qPCR) Validation Clinical HCC specimens were obtained from surgical resection samples (primary tumors and adjacent non-tumor tissues) of three patients at Panyu Central Hospital Affiliated to Guangzhou Medical University. The study was approved by the Ethics Committee (Approval No. PYRC-2022-007), conducted in accordance with the Declaration of Helsinki, and written informed consent was obtained from all participants. Total RNA extraction from three paired specimen sets comprising hepatocellular carcinoma tissues alongside their corresponding adjacent non-malignant counterparts was executed utilizing the RNA Isolation Total RNA Extraction Reagent manufactured by Vazyme. cDNA synthesis was accomplished from isolated total RNA employing the NovoScript® One-Step cDNA Synthesis Kit produced by Novoprotein. GAPDH functioned as the endogenous reference standard. Real-time quantitative PCR (RT-qPCR) analysis was performed with SYBR Green Master Mix (TaKaRa) using different primer sequences. Gene-specific primers used for RT-qPCR are listed in Supplementary Table 2. Statistical Analysis Computational bioinformatics procedures were implemented through R statistical environment (v 4.3.3). Inter-group data comparisons were conducted via Wilcoxon rank-sum testing methodology, with statistical significance established at p < 0.05. Results Single-Cell Transcriptomic Profiling PCA revealed that the first 30 PCs exhibited significant variation, suggesting their utility in capturing biologically relevant variance (Supplementary Figure 1A–C). ScRNA-seq profiling resolved the hepatic microenvironmental landscape into 22 discrete clusters (Supplementary Figure 1D), which were annotated into six major cell lineages using canonical marker genes: T lymphocytes, myeloid lineage cells, endothelial cells, hepatocytes, B lymphocytes, and fibroblast populations (Figure 1A). Marker gene expression demonstrated high specificity within each annotated cell type (Figure 1B). While the global clustering architecture remained stable across groups, the relative abundance of these cellular subsets shifted markedly among the control, non-lymph node metastasis (NLNM), and lymph node metastasis (LNM) groups (Figure 1C). Functionally, GSVA/GSEA revealed that hepatocytes possessed highly distinct metabolic and survival profiles compared to other lineages, characterized by robust enrichment in MGMT-dependent DNA damage repair, cardiolipin (CL) synthesis, and serotonin metabolism (Figure 1D). Cell clustering and annotations (A) Cell UMAP annotation results, with different colors indicating different cell types; (B) Heatmap of marker gene expression for cell clusters. (C) Results of cell annotation by different subgroups. (D) Functional enrichment of key cells. Identification of Candidate Genes and Functional Characteristics Sectional analysis of cellular proportions revealed a marked expansion of hepatocytes in both the LNM and NLNM groups compared to healthy controls (Figure 2A). This expansion of hepatocyte populations in disease states was aligned with prior reports implicating hepatocytes in HCC pathogenesis and metastasis.39 Differential expression analysis of this hepatocyte population identified 3,997 DEGs (3,089 upregulated, 908 downregulated) in Control vs NLNM, and 5,180 DEGs (2,621 upregulated, 2,559 downregulated) in LNM vs NLNM, yielding 1,265 overlapping scRNA-seq DEGs (Figure 2B). To isolate post-translational modification (PTM)-related drivers of metastasis, these single-cell DEGs were intersected with 10,785 bulk-level DEGs from the TCGA-LIHC cohort (Figure 2C) and a curated list of 822 PTM-related genes. This cross-database intersection yielded five high-confidence candidate genes: PJA1, RFPL4B, TRIM11, TRIM31, and ZNRF2 (Figure 2D). GSEA was performed to elucidate the biological pathways associated with the five candidate genes. Results indicated significant enrichment in 116 distinct pathways, such as MAPK signaling pathway, p53 signaling pathway, and glycolysis gluconeogenesis (Figure 2E–I and Supplementary Table 3). Identification of candidate genes and functional characteristics (A) Different colours denote distinct cell types: the left panel represents the control group, the middle panel the hepatocellular carcinoma metastasis group, and the right panel the hepatocellular carcinoma non-metastasis group. (B) Venn diagram. Yellow indicates genes showing differences between non-metastatic liver cancer and the control group, while blue indicates genes showing differences between metastatic and non-metastatic liver cancer groups. (C) Differentially expressed genes (volcano plot and heatmap). (i) log-transformed difference multiplier; top 10 up/down genes shown before sorting. Red: upregulation; blue: downregulation; gray: no change. (ii) Heatmap: genes (rows) vs samples (columns); red: high, blue: low expression. (D) Venn diagram of candidate genes. Note: Yellow indicates directional genes, blue indicates differentially expressed genes in TCGA-LIHC, and green indicates key cell-specific differentially expressed genes. (E–I) Analysis of the GSEA for candidate genes in the training set represent the GSEA plots for PJA1 (E), RFPL4B (F), TRIM11 (G), TRIM31 (H), and ZNRF2 (I). Construction and Validation of a Prognostic Risk Model Among the five candidate genes screened, PJA1 and RFPL4B exhibited significant associations with overall survival (p < 0.001). Both genes functioned as risk factors, with HR of 1.427 (95% CI: 1.169–1.743) for PJA1 and 1.598 (95% CI: 1.246–2.048) for RFPL4B (Figure 3A). PH assumption testing confirmed the validity of the Cox model for both genes (Schoenfeld test: p= 0.0898 for PJA1; p= 0.7371 for RFPL4B) (Figure 3B). A Random Survival Forest (RSF) model was established using both genes, with RFPL4B demonstrating higher variable importance (Figure 3C). In the TCGA-LIHC dataset (n = 368) an optimal risk cutoff of 72.80 stratified patients into high-risk (n = 130) and low-risk (n = 238) (Figure 3D), with both risk genes significantly overexpressed in the high-risk subset (Figure 3E). K-M survival curve assessment validated significantly inferior prognostic outcomes for the high-risk cohort compared to the low-risk stratum (p < 0.0001) (Figure 3F). Time-dependent ROC curve evaluation revealed the risk stratification model’s considerable discriminatory capacity, yielding AUC values of 0.75, 0.74, and 0.76 corresponding to 1-year, 3-year, and 5-year survival intervals, respectively (Figure 3G). Construction and validation of a prognostic risk model. (A) Forest plot of prognostic-related genes. (B) PH hypothesis test scatter plot. (C) The significance of prognostic genes. (D) Risk score distribution (i) and survival status of patients (ii) in the TCGA-LIHC training cohort. (E) Training set TCGA-LIHC high- and low-risk group prognostic gene expression heatmap. (F) Training set TCGA-LIHC KM survival curve. (G) Training set TCGA-LIHC ROC curve. (H) Risk curve (i) and survival distribution (ii) of the validation set. (I) Heatmap of prognostic gene expression levels for high- and low-risk subgroups in the validation set. (J) KM survival curve for the validation set. (K) ROC curve for the validation set. These findings were robustly validated in the independent GSE76427 cohort (n = 115). Utilizing a cutoff of 53.57, patients were classified into high-risk, (n = 42) and low-risk (n = 73) (Figure 3H). Heightened expression levels of both PJA1 and RFPL4B were recurrently detected within the high-risk patient subset (Figure 3I). The high-risk group again demonstrated significantly inferior survival (p = 0.016) (Figure 3J). The model maintained predictive power in the validation cohort, with AUC values exceeding 0.6 for 1-year, 3-year, and 5-year survival predictions (Figure 3K). Reconstructed RSF analysis (ntree=500) further supported the stability of this PTM-related prognostic signature (Supplementary Figure 2). Association of Risk Stratification with Clinical Subgroups This detrimental prognostic effect of high-risk status was consistently observed in subgroups defined by age, gender, T1 stage, T3 stage, N0 stage, M0 stage, stage I, and stage III (p < 0.05 for all comparisons) (Figure 4A). Robust positive association with high statistical significance was established between T staging classification and comprehensive pathological stage designation (cor = 0.9657, p < 0.001) (Figure 4B). Association of risk stratification with clinical subgroups. (A) Risk scoring analysis of K-M survival curves for different clinical characteristics. (B) Correlation significance between clinical features and risk scores was determined by Spearman’s rank correlation; *p < 0.05, **p < 0.01, ***p < 0.001. NS, not significant. Altered Intercellular Communication in the HCC Tumor Microenvironment Concomitantly, both the occurrence rate and magnitude of intercellular communications between hepatocyte and fibroblast populations exhibited marked elevation within HCC tissue specimens (Figure 5A and B). Conversely, a reduction in interactions between hepatocytes and myeloid cells was observed in LNM group compared to NLNM group (Figure 5A and B). Further characterization of specific receptor-ligand pairs identified PPIA-BSG as a predominant mediator of hepatocyte communication with other cell types in HCC samples relative to controls. In hepatocytes of LNM group, additional interactions were implicated, including MDK-NCL for communication with various cell types and PPIA-BSG for specific communication with B cells (Figure 5C). Altered intercellular communication in the HCC tumor microenvironment (A) Network diagram showing the number of connections between different cell clusters. (i) The number of interactions between hepatocyte cells and other cells in normal tissue; (ii) The number of interactions between hepatocyte cells and other cells in the liver cancer metastasis group; (iii) The number of interactions between hepatocytes and other cells in non-metastatic liver cancer tissue. (B) Network diagram showing the weight of connections between different cell clusters. (i) Total interaction intensity between hepatocyte cells and other cells in normal tissue; (ii) Interaction intensity between hepatocyte cells and others in the liver cancer metastasis group; (iii) Total interaction intensity between hepatocytes and other cells in non-metastatic liver cancer tissue. (C) Differential ligand–receptor interactions between hepatocytes and other cell types. (i) The levels of receptor and ligand expression between hepatocyte cells and other distinct cell types in normal tissues; (ii) The levels of receptor and ligand expression between liver cancer metastatic tissue and other distinct cell types; (iii) The levels of receptor and ligand expression between hepatocyte cells in non-metastatic liver cancer tissue and other distinct cell types. Dynamic Expression of Prognostic Genes During Hepatocyte Differentiation Sub-clustering partitioned hepatocytes into 12 transcriptionally distinct subpopulations (Figure 6A). Pseudotime trajectory analysis mapped these cells along a five-stage differentiation continuum. While normal liver and NLNM hepatocytes spanned the entire evolutionary spectrum, LNM-derived cells were predominantly arrested at early, poorly differentiated stages (Figure 6B). Mechanistically, PJA1 expression decreased upon differentiation initiation and stabilized at lower levels, whereas RFPL4B displayed a biphasic pattern, peaking during late-stage maturation (Figure 6C). Single-cell regulon analysis uncovered profound transcriptional reprogramming accompanying this differentiation arrest (Figure 6D). In LNM hepatocytes, master regulators of mature hepatic identity (RFX5, FOXO4, HNF4A, CEBPA, and FOXA2) were significantly deactivated. Conversely, stemness- and proliferation-associated transcription factors (POU5F1, MAFK, NR2F1, TCF4, E2F7, TFAP2C, and ARNT) were selectively hyperactivated, while NLNM hepatocytes maintained near-baseline activities across these factors (Figure 6D). Dynamic expression of prognostic genes during hepatocyte differentiation. (A) Clustering results for different subpopulations of hepatocytes. (B) Pseudotime analysis of liver cell trajectories. (C) Pseudotime analysis of liver cell subtypes. (D) Analysis of gene regulation in key cells. Tumor Microenvironment and Genomic Alterations Stratified by Risk Groups GSEA identified 51 significantly enriched pathways between the risk groups (p < 0.05, FDR < 0.25, |NES| > 1). High-risk tumors exhibited severe metabolic and biosynthetic perturbations, characterized by marked suppression of fatty acid metabolism, peroxisome, cytochrome P450 drug metabolism, complement/coagulation cascades, and steroid hormone biosynthesis (Figure 7A and Supplementary Table 4). Such observations suggested substantial perturbation of metabolic homeostasis and biosynthetic pathway coordination within the high-risk patient stratum, conceivably facilitating neoplastic advancement and unfavorable clinical endpoints. Tumor microenvironment and genomic alterations stratified by risk groups. (A) GSEA analysis of the TCGA-LIHC training set’s high- and low-risk groups. (B) Heatmap of immune cell or stromal cell scores. (C) Inter-group differences in immune cells. (D) Analysis of correlations between different immune cells. (E) Analysis of the correlation between different immune cells and prognostic genes. Wilcoxon rank-sum test (C) and Spearman correlation (D) were used; *p < 0.05, **p < 0.01, ***p < 0.001; NS, not significant. Microenvironmental profiling revealed 29 cell types with significantly disparate infiltration levels (p < 0.05; Figure 7B and C). Low-risk tumors are enriched with stromal and vascular cell lineages (adipocytes, hepatocytes, hepatic stellate cells, lymphatic endothelial cells, microvascular endothelial cells, pericytes, preadipocytes), all of which decrease in high-risk tumors. In contrast, high-risk tumors have increased infiltration of common lymphoid progenitors, epithelial cells, keratinocytes, γδ T cells and Th2 cells. Strong positive correlations were observed between endothelial cells and lymphatic endothelial cells (cor = 0.97, p< 0.001) and between endothelial cells and microvascular endothelial cells (cor = 0.95, p< 0.001). Hepatocyte populations additionally demonstrated statistically meaningful positive association with adipocyte compartments (cor = 0.77, p < 0.001). Conversely, the strongest inverse relationship was identified linking pericyte and Th2 cell subsets (cor = −0.55, p < 0.001) (Figure 7D). Furthermore, PJA1 expression was significantly negatively correlated with the infiltration levels of hepatocytes (cor = −0.39, p < 0.05) and lymphatic endothelial cells (cor = −0.33, p < 0.05). No significant correlations were found between RFPL4B expression and any of the differentially infiltrated cell types (Figure 7E). Genomically, CTNNB1, TTN, and TP53 were the top three altered loci across the entire cohort (Figures 8A and B). The high-risk group harbored a significantly elevated tumor mutational burden (p < 0.001) (Figure 8C and D), indicating heightened genomic instability. Genomic alterations associated with risk stratification. (A-B) Top 10 most frequently mutated genes in high- (A) and low-risk (B) groups. (C) Tumor mutation burden (TMB) differences between groups. (D) Correlation between TMB and risk score. Wilcoxon rank-sum test (C) and Spearman correlation (D). Alterations in the Immune Checkpoint Landscape High-risk stratification captured extensive remodeling of the immune checkpoint landscape. Among 15 differentially expressed immune checkpoint inhibitors (ICIs), KDR was significantly downregulated in the high-risk group and inversely correlated with risk scores (cor = −0.12, p < 0.05) whereas TGFB1, TGFBR1, and CSF1R were significantly upregulated (Figure 9A). Additionally, CSF1R was positively correlated with HAVCR2 (cor = 0.85, p < 0.05) (Figure 9B). Furthermore, 68 immune checkpoint (ICP) genes exhibited significant divergence between risk groups, characterized by alterations in SIRPA, HLA-DMA, TNFRSF14, and CD47 (p < 0.01) (Figure 9C). Risk scores positively correlated with BTN2A1 (cor = 0.33, p < 0.05) and BTN2A2 (cor = 0.34, p < 0.05). Expression of PJA1 was significantly positively correlated with several ICPs, such as BTN2A1 (cor = 0.48, p < 0.05), BTN2A2 (cor = 0.46, p < 0.05), and CD47 (cor = 0.49, p < 0.05) (Figure 9D). Chemosensitivity profiling indicated that high-risk tumors had significantly lower IC50 values for BPD.00008900, dinaciclib, and fulvestrant (p < 0.0001) (Figure 9E). Genomic and expression characterization of prognostic genes. (A) Differences in the expression of ICIs between high- and low-risk groups. (B) Analysis of the correlation between differences in ICIs and risk scores. (C) Differences in the expression of immune checkpoint genes between high- and low-risk groups. (D) Analysis of the correlation between prognostic genes, risk scores, and immune checkpoint markers. (E) Analysis of drug sensitivity between high-risk and low-risk groups. (F) Chromosome localization analysis. (G) Subcellular localization. (H) Expression of prognostic gene PJA1 in normal tissues (i) and liver cancer tissues (ii). (I) Expression of the prognostic gene RFPL4B in normal tissues (i) and hepatocellular carcinoma tissues (ii). Wilcoxon rank-sum test was used for group comparisons (A–E). *p < 0.05, **p < 0.01, ***p < 0.001, ****p < 0.0001. Abbreviations: NS, not significant. To explore the immunomodulatory potential of RFPL4B, a PPI network was constructed using its top 20 GeneMANIA-associated neighbors and immune regulators. Despite its peripheral location, RFPL4B indirectly linked to core checkpoint and T-cell genes via a BTN2A1/2 subcluster, which further integrated with an immunosuppressive hub consisting of SIGLEC15, ADORA2A, ENTPD1, NT5E, ARG1, and TAB2 (Supplementary Figure 3). Genomic and Expression Characterization of Prognostic Genes The genomic locations of PJA1 and RFPL4B were determined. PJA1was localized to the X chromosome, while RFPL4B resided on chromosome 6 (Figure 9F). Prediction of subcellular localization indicated that the PJA1 protein was most likely localized to the nucleus, whereas the RFPL4B protein was predominantly predicted to reside in the cytosol (Figure 9G). Comparative analysis of protein expression between tumor tissues and normal liver control tissues revealed distinct patterns. Significantly elevated nuclear expression of PJA1 was observed in tumor tissues compared to control liver tissues (Figure 9H). This finding aligned with its predicted nuclear localization and suggested a potential role in nuclear processes during hepatocarcinogenesis. Consistent with its predicted cytosolic localization, RFPL4B protein expression was markedly increased within the cytoplasm of tumor cells relative to control tissues (Figure 9I). Independent Validation RT-qPCR was employed to validate gene expression differences in three patients from Panyu Central Hospital, Guangzhou Medical University. RT-qPCR results from three patient samples showed variable expression patterns of PJA1. In one case, PJA1 expression was higher in the tumor tissue, while in the other two cases, its expression was higher in the adjacent non-tumor tissue (Figure 10). This preliminary validation suggests heterogeneity in PJA1 expression among HCC patients, warranting further investigation in a larger cohort (Figure 10). RT-qPCR analysis of RFPL4B in the same cohort showed no detectable expression in either tumor or adjacent tissues, likely due to low transcript abundance, subpopulation-specific expression, and limited sample size and heterogeneity. Although mRNA-level validation was not achieved, its prognostic significance in TCGA-LIHC and GSE76427 supports biological relevance and underscores the need for protein-level and single-cell validation in future studies. qRT-PCR results for tumour tissue and adjacent non-tumour tissue in hepatocellular carcinoma patients. Relative gene expression levels of PJA1. Discussion HCC is among the leading causes of cancer-related mortality worldwide, with LNM significantly worsening clinical outcomes.2,40 PTMs of proteins play a crucial regulatory role in the dynamics of the TME,11,41 yet their involvement in HCC lymph node metastasis remains unclear. To our knowledge, this study represents one of the first integrating single-cell transcriptomics with bulk RNA sequencing data to evaluate the prognostic value of PTM-related genes in HCC with LNM. The predictive signature developed based on two pivotal genes, PJA1 and RFPL4B demonstrated promising prognostic stratification performance across both the training dataset and independent validation cohort. Individuals stratified into the high-risk category displayed markedly increased tumour mutational burden (TMB) alongside distinct immunological microenvironmental profiles. Single-cell trajectory analysis further revealed that hepatocyte differentiation in LNM samples was arrested at an early stage, accompanied by reprogramming of transcription factor activity, providing a novel cellular evolutionary perspective on metastatic HCC. In HCC progression, hepatocytes undergo dynamic changes in proportion and pathway enrichment, indicating their crucial role in remodeling the tumor microenvironment. Enrichment of the MGMT-mediated DNA repair pathway in hepatocytes suggests enhanced genomic maintenance under stress, which may also contribute to chemotherapy resistance.42 Additionally, hepatocytes show marked enrichment in cardiolipin (CL) synthesis, essential for mitochondrial integrity and function.43 This may represent an adaptive response to support the high metabolic demands of tumors. Notably, mitochondrial dysfunction—often driven by dysregulated PTMs in mitochondrial proteins—can disrupt energy balance and promote apoptosis,44 suggesting a potential link between CL metabolism and PTM-regulated mitochondrial stability. These intrinsic metabolic adaptations align with the emerging role of hepatocyte-derived signaling in tumor progression. For instance, hepatocyte-secreted molecules such as FGL1 can suppress CD8⁺ T and NK cell activity via LAG-3, facilitating immune escape and metastasis.45 Similarly, extracellular vesicles from hepatocytes carry molecules like ACSL4, which induce senescence and promote HCC progression.46 Together, these findings highlight hepatocytes as active contributors to HCC lymph node metastasis through both cell-autonomous pathway alterations and microenvironmental crosstalk. PJA1 (Praja ring finger ubiquitin ligase 1) functions as an E3 ubiquitin ligase. Emerging evidence indicates that PJA1 facilitates proteasomal degradation of the mitochondrial protein PGAM5 via augmentation of K48-linked polyubiquitin chain formation specifically at the lysine 88 residue. This subsequently leads to increased phosphorylation of DRP1 at position S637, reducing mitochondrial reactive oxygen species (ROS) production. Consequently, GSDME-mediated pyroptosis is inhibited, thereby weakening the anti-tumour immune response. Inhibition of PJA1 (eg, using the small-molecule inhibitor RTA402) enhances chemotherapy sensitivity.47 Moreover, PJA1 can influence the immune microenvironment by mediating the degradation of glycogen synthase 1 (GYS1) through ubiquitination. Flavin monoxygenase 2 (FMO2) competitively binds GYS1, inhibiting its degradation by PJA1 and thereby promoting nuclear factorκB/p65-mediated CCL19 expression. CCL19 has been demonstrated to enhance the efficacy of anti-PD-1 therapy, with elevated serum levels correlating with favourable responses to immunotherapy.48 In HCC, the deubiquitinating enzyme OTUB2 stabilises PJA1 through deubiquitination, thereby promoting malignant proliferation and migration in HCC cells.49 As an E3 ubiquitin ligase, the overexpression of PJA1 may potentially disrupt the ubiquitin balance of key proteins—such as those involved in transfer—by antagonising the action of other deubiquitinating enzymes (eg, the known pro-transfer enzyme USP2a).11 Such disruption could theoretically affect their stability, ultimately and may contribute to tumour progression. These findings support the potential oncogenic role of PJA1 and are consistent with our bioinformatics-based identification of PJA1 as a prognostic candidate, warranting further experimental validation in HCC. RFPL4B (Ret finger protein-like 4B) belongs to the “Ret finger protein-like” family. Research indicates that RFPL4B may play a significant role in HCC subtypes with poorer prognosis (highly proliferative, highly heterogeneous, and prone to metastasis). Furthermore, in vitro experiments confirm its capacity to promote progression in Huh7 cells.50 Research on its homologue RFPL3S has provided important clues. Low expression of RFPL3S correlates with poor prognosis in testicular germ cell tumours (TGCT), including increased metastatic potential. Its expression levels positively correlate with infiltration by immunologically activated cells (B cells, CD8+ T cells, NK cells, etc.) and negatively correlate with infiltration by immunosuppressive cells (Th17, Th2). High RFPL3S expression correlates with immunotherapy benefit.51 We acknowledge that RFPL4B was not detected at the transcript level in all three clinical samples, representing a clear discrepancy between computational prediction and experimental validation. This may reflect its restricted expression in specific hepatocyte subpopulations identified by scRNA-seq, leading to signal dilution in bulk tissue. In addition, RFPL4B has been associated with highly proliferative and heterogeneous HCC subtypes, which may not be represented in this limited cohort. Moreover, as a RING finger family member, RFPL4B may exert key functions at the post-translational level rather than through mRNA abundance. Importantly, the prognostic value of RFPL4B was consistently observed in independent large-scale cohorts (TCGA-LIHC and GSE76427), supporting its robustness at the population level. While our bioinformatics analysis identified PJA1 and RFPL4B as key genes significantly associated with HCC lymph node metastasis and demonstrated their robust predictive power in our risk model, further validation in clinical samples was crucial. We conducted gene expression analysis on tumor and paracancerous tissues from three HCC patients. The results showed consistent expression of PJA1 across all samples, whereas RFPL4B was not detected. This discrepancy highlights that while bioinformatics analysis can identify potential key genes from large datasets, empirical validation with diverse methods is essential to confirm their presence and expression patterns in specific disease contexts. The non-detection of RFPL4B could be attributed to extremely low expression levels, its absence in the specific tissue types analyzed, or limitations in the sensitivity of our detection methods. This finding directs future research towards optimizing detection strategies for RFPL4B or exploring alternative, non-expression-level mechanisms through which RFPL4B might influence HCC lymph node metastasis. Importantly, our model demonstrated favorable external validation performance in the independent GSE76427 cohort, with AUCs of 0.69, 0.66, and 0.63 at 1-, 3-, and 5-year survival, respectively. Notably, previous HCC prognostic models reported AUC values in this dataset typically ranging from 0.55 to 0.75, including prominent published signatures.52,53 Regarding PJA1, we observed an intriguing and varied expression pattern: in two patients, PJA1 expression was higher in paracancerous tissues than in tumor tissues, while in one patient, the opposite was true, with higher PJA1 expression in tumor tissues compared to paracancerous tissues. This heterogeneous expression profile, particularly between tumor and adjacent non-tumor tissues, suggests that PJA1’s role in HCC lymph node metastasis may be complex and multifaceted, rather than a simple up- or downregulation. The scenario where PJA1 is highly expressed in paracancerous tissue and lowly expressed in tumor tissue might indicate a protective role for PJA1 in maintaining normal hepatocyte function or inhibiting early-stage carcinogenesis, with its expression being suppressed during tumor progression, thereby promoting tumor development. Conversely, higher PJA1 expression in tumor tissue compared to paracancerous tissue could suggest that in certain tumor subtypes or disease stages, its overexpression actively promotes malignant progression, for instance, by influencing the immunosuppressive tumor microenvironment or enhancing tumor cell invasiveness. As previously indicated by our bioinformatics analysis, PJA1 expression exhibited significant positive correlations with multiple immune checkpoint molecules, particularly BTN2A1, BTN2A2, and CD47, suggesting that PJA1 may contribute to immune evasion by regulating the expression or stability of these immunosuppressive molecules. Therefore, this observed heterogeneity in PJA1 expression patterns might reflect the intricate molecular mechanisms underpinning HCC lymph node metastasis. It underscores the necessity for future studies to delve deeper into the functional differences associated with varying PJA1 expression levels and their specific impacts on tumor progression. This could potentially reveal differential responses to PJA1-targeted therapies among patients, thereby offering preliminary clinical evidence for PJA1’s complex biological role in HCC LNM and providing new avenues for developing precision medicine strategies. GSEA revealed candidate genes enriched in pathways including MAPK, p53, and glycolysis. These pathways are closely associated with tumour metastasis. The MAPK pathway serves as a central regulator of cell proliferation, invasion, and metastasis. Studies have demonstrated that bacterial lipopolysaccharide (LPS) can upregulate pro-metastatic factors by activating the TLR4/MAPK signalling pathway, thereby promoting lymphangiogenesis and metastasis.54 p53 is a crucial tumour suppressor gene, and dysregulation of its signalling pathway is closely associated with tumour initiation and progression. Research has revealed that the integrin-associated factor FAM27E3 promotes lymph node metastasis in thyroid carcinoma via the p53 signalling pathway.55 Malignant cells characteristically demonstrate upregulated aerobic glycolytic metabolism, a phenomenon termed the Warburg effect, which supplies both energetic substrates and biosynthetic intermediates essential for accelerated cellular expansion. Activation of the C-kinase receptor 1 (RACK1) has been reported to promote lymph node metastasis in cervical carcinoma via glycolysis-dependent AKT/mTOR signalling.56 PJA1 and RFPL4B may regulate the metastatic potential of hepatocellular carcinoma by influencing post-translational modification processes within these key pathways, such as acting upon critical enzymes or substrates within the ubiquitination pathway. In LNM, the tumour microenvironment undergoes profound interactive remodelling, whilst hepatocytes themselves exhibit a distinctive state of differentiation arrest. Cell communication analysis indicates that hepatocyte-fibroblast interactions are significantly enhanced in the LNM context. This aligns with the established mechanism whereby fibroblasts promote lymphatic invasion via ITGA11-SELE,57 potentially providing pro-metastatic signals. Concurrently, cellular crosstalk between hepatocytes and myeloid populations demonstrated marked attenuation within the LNM cohort, implying either intensification of the immunosuppressive milieu pervading the tumour microenvironment or phenotypic reprogramming of myeloid lineages toward tumour-promoting functions. This bears similarities to known pro-metastatic immunosuppressive axes such as HOXC10/CXCL5/MDSCs.58,59 At the ligand-receptor level, LNM-associated hepatocytes were observed to preferentially employ PPIA-BSG (a known lymph node metastasis regulatory axis)60 and MDK-NCL (a pathway shaping immunosuppressive microenvironments across cancer types)61,62 for communication, which may represent candidate mechanisms contributing to the construction of a metastasis-promoting, immune-evasion niche. Pseudo-time analysis suggests that hepatocytes in the LNM cohort tend to accumulate at an early (low-differentiation) stage, existing in a more plastic and metastatic “dedifferentiated” state. The observed decline in PJA1 expression during early pseudotime and the late-stage expression peak of RFPL4B are intriguing patterns that may warrant further investigation; however, the functional significance of these dynamic expression profiles-including whether they reflect pluripotency maintenance or pre-metastatic preparation—remains speculative and requires experimental verification. These observations offer a hypothesis-generating perspective on how PTM-related genes may participate in regulating hepatocyte differentiation states during metastatic progression. Analysis indicates that the high-risk group defined by PJA1 and RFPL4B exhibits distinct TME characteristics, with markedly reduced infiltration levels of endothelial cells, microvascular endothelial cells, and lymphatic endothelial cells (LECs). This decrease in vascular/lymphatic components shows a strong positive correlation, suggesting they may occur synergistically.63 Such alterations could potentially impact lymphatic vessel architecture within tumours or local immune surveillance functions. Notably, LEC function can be regulated by the tumour microenvironment; for instance, PAI-1 derived from cancer-associated fibroblasts (CAFs) promotes lymphatic metastasis by inducing lymphatic endothelium-to-mesenchymal transition (EndoMT).64 The reduced LEC infiltration observed in the high-risk group may reflect tumor-associated lymphatic remodeling rather than a stochastic phenomenon. Potential mechanisms include impaired lymphangiogenic signaling, extracellular matrix remodeling, and EndoMT induced by an immunosuppressive microenvironment. Consistent with this possibility, TGFB1, TGFBR1, and CSF1R were markedly elevated in the high-risk group, while previous studies have shown that CAF-derived PAI-1 can promote EndoMT and lymphatic metastasis. However, direct mechanistic evidence remains lacking and warrants further investigation. Furthermore, PJA1 expression exhibited a significant negative correlation with both hepatocyte and LEC infiltration levels. This aligns with findings that PJA1 serves as a risk factor and is highly expressed in hepatocytes, suggesting PJA1 may indirectly influence the abundance and function of hepatocytes (particularly tumour-associated hepatocytes) or LECs within the microenvironment by regulating downstream target proteins via its E3 ubiquitin ligase activity. Similarly, S100A6 is known to drive LEC lymphangiogenesis by activating the RAGE/NF-κB/VEGF-D axis.2 Whether PJA1 influences LEC function via comparable pathways requires subsequent validation. Finally, the high-risk group exhibited a higher TMB, reflecting greater genomic instability in tumours within this cohort. This is typically associated with increased tumour aggressiveness, heightened heterogeneity, and poorer prognosis,65 further consistent with the notion that tumours in the high-risk group possess greater potential for progression. This study, based on a risk model incorporating PJA1 and RFPL4B, reveals that high-risk hepatocellular carcinoma patients exhibit a distinct immune checkpoint expression profile. Apart from reduced expression of the proangiogenic receptor KDR (VEGFR2), most immune checkpoint molecules are significantly upregulated in the high-risk cohort. This includes key members of immunosuppressive pathways: the ligand TGFB1 and its receptor TGFBR1 (TGF-β signalling frequently exhibits potent immunosuppression in advanced tumours),66 alongside CSF1R—primarily expressed on M2 tumour-associated macrophages—whose signalling pathway drives the recruitment and polarisation of immunosuppressive TAMs.67 Additionally, differential expression patterns were observed across 68 immune checkpoint-related genes within the high-risk patient population, substantiating extensive perturbation of immunoregulatory circuitry. Notably, PJA1 expression showed significant positive correlations with multiple immune checkpoint molecules, particularly BTN2A1, BTN2A2, and CD47. BTN2A1/A2, members of the B7 family, may modulate γδ T cell activity,68 while CD47 functions as a “do not eat me” signal molecule that inhibits phagocytosis by binding to SIRPα on macrophages.69 As an E3 ubiquitin ligase, the high expression of PJA1 correlates with the upregulation of these immunosuppressive molecules, suggesting that PJA1 may be computationally implicated in regulating the expression or stability of these molecules through its ubiquitin modification function. However, whether this reflects a direct mechanistic relationship remains to be established experimentally and should not be overinterpreted from correlation-based analyses alone. This finding resonates with previous observations that PJA1 can influence the immune microenvironment by regulating CCL19.48 Furthermore, positive correlative associations were identified between the calculated risk scores and BTN2A1/A2 expression levels, thereby reinforcing the notion that patients categorized as high-risk by this predictive framework manifest heightened immune escape attributes. Recent studies highlight that PTMs, epigenetic regulation, and translational control jointly shape angiogenesis and immune evasion in HCC via interconnected networks, including ASH2L lactylation-VEGFA axis,70 RBM30-mediated STAT1/PD-L1 activation,71 and RBM28-driven STAT3/VEGFA translational upregulation.72 Our PTM-related risk model may therefore capture one dimension of this dynamic microenvironment. The current work conducts a methodical investigation into both the predictive utility and potential molecular roles of PTM-related genes in the setting of lymphatic metastasis in HCC. It identified PJA1 and RFPL4B as candidate prognostic risk genes and developed a risk stratification model that demonstrated consistent prognostic performance, providing computational evidence for risk stratification, prognosis prediction, and hypothesis-generating insights into potential therapeutic strategies. However, this study has certain limitations. A major limitation of this study is its reliance on computational analyses, which precludes definitive causal inference regarding the roles of PJA1 and RFPL4B in HCC progression. Functional mechanisms, including upstream substrates of PJA1 and downstream regulatory targets of RFPL4B, remain to be experimentally validated. External validation is also limited by the absence of independent cohorts with simultaneous PJA1 and RFPL4B expression profiling, which may affect the generalizability of the prognostic model. In addition, RFPL4B was not detected at the mRNA level in the clinical validation cohort, highlighting a discrepancy between computational predictions and experimental observations and underscoring the need for protein-level and larger cohort validation. Furthermore, the single-cell dataset presents a tissue-of-origin confounding issue, as the LNM, NLNM, and Control groups differ in both anatomical source and disease status, and the lack of a non-metastatic lymph node control limits the interpretation of LNM-specific signatures. The heterogeneous expression pattern of PJA1 further suggests biological complexity that warrants validation in spatially resolved or larger clinical datasets.