Multi-omics analysis of the mechanism by ATP13A2 regulates the tumor microenvironment and prognosis in hepatocellular carcinoma

Abstract

Background and aims

Hepatocellular carcinoma (HCC) is a prevalent global cancer. Most patients with HCC are diagnosed at an advanced stage. Therefore, new biomarkers and treatments are urgently needed.

Methods

We employed eQTL and intersected the results with aging-related genes, ultimately identifying ATP13A2 as the target gene for our study. The expression and intercellular communication of ATP13A2 in HCC were analyzed using single-cell RNA sequencing (scRNA-seq) and spatial transcriptomics sequencing (stRNA-seq). Subsequently, we utilized the Deep Learning Survival Neural Network (DeepSurv) model to construct a prognostic model. Additionally, we performed RNA sequencing (RNA-seq) analysis. In vitro, CCK8, cell wound healing, and flow cytometry assays were used to identify the potential functions of ATP13A2 in HCC cells.

Results

ATP13A2 is positively associated with HCC risk. scRNA-seq analysis demonstrated that ATP13A2 + malignant cells exhibited stronger interactions with tumor microenvironment (TME) cells. stRNA-seq analysis revealed that ATP13A2 + malignant cells were significantly spatially correlated with TME cells. The DeepSurv prognostic model indicated that HCC patients with high risk scores had a significantly lower survival rate than those with low risk scores. In vitro, knockdown of ATP13A2 affected the proliferation, apoptosis, and migration of HCC cells.

Conclusion

The ATP13A2 gene is closely related to TME, and its high expression is indicative of poor prognosis. ATP13A2 has the potential to serve as a biomarker for prognosis and efficacy assessment of HCC and may offer a new therapeutic target for its treatment.

Graphical abstract

Supplementary Information

The online version contains supplementary material available at 10.1186/s12935-026-04315-1.

Introduction

Hepatocellular carcinoma (HCC) ranks among the most prevalent cancers globally, characterized by a high incidence and poor prognosis. As the primary form of liver cancer, HCC accounts for 75%–85% of all cases [1]. Despite significant advancements in HCC treatment, it remains the third leading cause of cancer-related mortality worldwide. The prognosis for HCC is dismal, with a relative 5-year survival rate of approximately 18% [2]. Current diagnostic methods, including histology and radiology, are available for the diagnosis and evaluation of HCC; however, there is an urgent need for new biomarkers to enhance the diagnosis and treatment of this disease [3]. In recent decades, the rapid advancement of omics technologies has enabled us to uncover novel pathogenic genes associated with HCC [4, 5]. Consequently, a series of omics data-based signatures has been developed to predict clinical outcomes in HCC [6]. Therefore, the integration of multi-omics approaches holds great promise for discovering new biomarkers for HCC.

In recent years, many studies have shown that the tumor microenvironment (TME) is closely related to the occurrence and development of tumors. Tumor-associated macrophages (TAM) usually exhibit multiple functional phenotypes; M1-like TAM have pro-inflammatory and anti-tumor effects, while M2-like TAM have anti-inflammatory and pro-tumor effects [7]. HCC is highly vascular, and angiogenesis significantly promotes tumor growth and metastasis [8]. Rapid tumor growth leads to hypoxia, which stimulates HCC cells to secrete pro-angiogenic factors (such as VEGFA, FGF2, and PDGFA), promote angiogenesis, and further accelerate the development of HCC [9]. Cancer-associated fibroblasts (CAF) can promote chemotherapy resistance and recurrence of HCC by secreting a variety of cytokines, growth factors, or extracellular vesicles [10]. As a hallmark of cancer, aging can regulate many cancer-related biological processes by regulating cellular senescence [11]. Aging induces a decrease in protective enzymes, a reduced response to growth factors, and an increased risk of intestinal pathogen uptake, thereby increasing the risk of liver disease and leading to cellular senescence and low-grade inflammation in the liver [12]. Aging can promote tumor growth and development [13]. Compared with non-alcoholic fatty liver disease, hepatocyte senescence is more pronounced in HCC [14]. Aging induces pathological changes in the liver and enhances the occurrence and progression of non-alcoholic fatty liver disease, HCC, and other age-related liver diseases [15, 16]. Bcl-XL/BCl-2 Protac can reduce metabolic dysfunction-associated steatotic liver disease-driven hepatocellular carcinoma in mice by clearing senescent cells in the liver [17]. However, the relationship between aging-related genes and HCC remains unclear.

Multi-omics integration is an emerging methodology in the post-GWAS era, with Summary-data-based Mendelian Randomization (SMR) being a notable advancement. SMR expands upon traditional Mendelian randomization by combining GWAS data with gene expression data derived from expression quantitative trait loci (eQTL) studies to investigate causal relationships between potential pathogenic genes and diseases [18]. Spatial transcriptomics is a novel sequencing technique that enables visualization of gene distribution within tissue and the direction of intercellular signaling, providing new biological insights into diseases [19]. Additionally, artificial intelligence has made significant inroads into the medical field, with deep learning algorithms increasingly applied in tumor diagnosis, subtype classification, grading, staging, and prognosis prediction.

In this study, we utilized SMR to identify the pathogenic gene ATP13A2 associated with aging. Through a multi-omics integrated approach, we explored the functions of ATP13A2 in HCC using single-cell RNA sequencing (scRNA-seq), spatial transcriptomics sequencing (stRNA-seq), and RNA sequencing (RNA-seq), ultimately constructing a prognostic model using the Deep Learning Survival Neural Network (DeepSurv). Finally, we verified the function of ATP13A2 in HCC through in vitro experiments.

Method

SMR and aging

For SMR analysis, cis-eQTLs for liver tissue in the exposure data were obtained from the Genotype Tissue Expression (GTEx) project v8 (https://yanglab.westlake.edu.cn/software/smr/#eQTLsummarydata). GWAS data for HCC outcome data were obtained from Hepatocellular carcinoma (controls excluding all cancers) in FINNGEN [20]. SMR (v1.3.1) was used to analyze these two independent datasets, and a p value < 0.05 was considered to be causal between the exposure and outcome. The 1000 Genomes Project European individuals (1000G EUR) were used as reference genotype data. The HEIDI test was used as a test for heterogeneity of results, and a p value < 0.05 was considered to be heterogeneous. In order to identify functional, translatable causal genes, protein-coding genes were retained. Genes related to aging were searched in GeneCards (https://www.genecards.org/), and genes with a relevance score greater than 10 were selected [21]. The SMR results were intersected with aging-related genes to obtain genes with a causal relationship between HCC and aging. We searched for the prognostic relationship between these genes and HCC in GEPIA (http://gepia.cancer-pku.cn/index.html), and finally found that ATP13A2 was positively correlated with the poor prognosis of HCC with p(HR) < 0.05 and p(HR) > 0.05 for the other three genes (Fig. 1B-D). Therefore, we determined ATP13A2 as the gene to be studied next. We analyzed the expression of ATP13A2 in different cancers in TIMER2.0 (http://timer.cistrome.org/) [22]. Finally, we plotted SMRLocusPlot and SMREffectPlot of ATP13A2 and HCC in R (version 4.3.1).

Fig. 1.

Results of SMR analysis for aging-related genes and HCC. (A) Venn diagram of Mendelian randomization results and aging-related genes. (B—D) Survival curve of ATP13A2, CHRNA4 and NOD2 in hepatocellular carcinoma. High expression of ATP13A2 is associated with poor prognosis. (E) Pan-cancer analysis of ATP13A2 expression (TIMER2.0 database) showing significantly increased ATP13A2 expression in multiple human cancers including HCC compared with normal tissues. (F – G) SMRLocusPlot and SMREffectPlot of ATP13A2 and HCC, demonstrating the positive causal correlation between ATP13A2 and HCC risk. SMR: Summary-data-based Mendelian Randomization; HCC: hepatocellular carcinoma

scRNA-seq analysis

The scRNA-seq data of HCC were obtained from the Gene Expression Omnibus (GEO) accession number GSE125449 of the National Center for Biotechnology Information [23]. scRNA-seq analysis was performed using the “Seurat” package (version 4.3.0). The data was normalized using the NormalizeData function, with scale.factor set to 10,000. Principal component analysis was performed on HVFs (RunPCA(), npcs = 30), and the optimal number of principal components was determined using the elbow plot. A shared nearest neighbor graph was constructed, and cells were clustered using the Louvain algorithm. Non-linear dimensionality reduction was performed for visualization: t-SNE (RunTSNE(), perplexity = 30) and UMAP (RunUMAP(), n.neighbors = 30, min.dist = 0.3). The original file had annotated and filtered cells, and the results were visualized (Fig. 2A). A density map of ATP13A2 expression was plotted (Fig. 2B). Malignant cells were divided into ATP13A2 + Malignant cells (ATP13A2 + Mali) and ATP13A2- Malignant cells (ATP13A2- Mali) based on whether ATP13A2 was expressed, and intercellular communication was performed using the CellChat package. A CellChat object was created with createCellChat using log-normalized expression data. The human CellChatDB database was filtered to retain biologically relevant L-R pairs: "Secreted Signaling" and "Cell–Cell Contact", reducing noise from non-canonical interactions. Communication probability was calculated with computeCommunProb(). Pathway-level communication strength was computed with computeCommunProbPathway (thresh = 1), retaining only pathways with total communication probability ≥ 1. Low-confidence communications were filtered out, and significant interactions were subset with subsetCommunication (cellchat, thresh = 1). Concurrently, we conducted a sensitivity analysis. We stratified the malignant cells expressing ATP13A2 into two groups—ATP13A2-high Malignant and ATP13A2-low Malignant—based on the median expression level, and subsequently performed an intercellular communication analysis. Then, Malignant cells were extracted, and pseudotime analysis was performed using the Vector algorithm.

Fig. 2.

scRNA-seq analysis reveals ATP13A2 + Mali exhibits enhanced intercellular communication with TME cells. (A) UMAP dimensionality reduction plot of annotated cell subpopulations in HCC scRNA-seq data (GSE125449) analyzed by Seurat package. (B) Density map of ATP13A2 expression, normalized by NormalizeData function (scale.factor = 10,000). (C – F) Intercellular communication analysis results by CellChat package. ATP13A2 + Mali has stronger interactions with CAF, TAM, and TEC than ATP13A2- Mali. (G – H) UMAP of pseudotime analysis, showing the developmental and differentiation trajectories of ATP13A2 in HCC malignant cells. scRNA-seq: Single-cell RNA sequencing. TME: tumor microenvironment. ATP13A2 + Mali: ATP13A2 + Malignant cell; ATP13A2- Mali: ATP13A2- Malignant cell; CAF: cancer-associated fibroblasts; TAM: tumor-associated macrophages; TEC: tumor-associated endothelial cells

stRNA-seq analysis

stRNA-seq processing

stRNA-seq for HCC were obtained from the study by Ma et al. [24]. The raw data were loaded using the Load10X_Spatial function of the Seurat package. Tissue coordinate information was extracted from the raw data using the GetTissueCoordinates function. The subset function was employed to select cells located within the tissue. Quality control was conducted to filter out cells expressing mitochondrial and ribosomal genes. Following preprocessing, the SCTransform function was applied to adjust for technical differences, and the NormalizeData and ScaleData functions were used for normalization and standardization of all features, respectively. Principal component analysis and UMAP dimensionality reduction were performed thereafter.

Reverse compositional transcriptomics deconvolution (RCTD) and mistyR

RCTD is a method for inferring cellular composition from stRNA-seq. This method analyzes the comprehensive gene expression profile of the tissue or sample to reverse-calculate the spatial distribution and relative abundance of each cell type [25]. The spatial distributions of ATP13A2 + Mali, ATP13A2- Mali, CAF, TAM, and tumor-associated endothelial cells (TEC) were visualized. The spatial expression pattern of ATP13A2 in HCC tissues was explored using the mistyR package, and a network diagram of these cells was generated using the SPOTlight package.

Cell degree and commot

We further utilized cell degree analysis to explore interactions between different cell types. The spatial coordinate data were obtained and transformed into image coordinates using a scaling factor. The scaled row and column coordinates were represented by imagerow_scaled and imagecol_scaled, respectively. The DBSCAN algorithm was employed, with nNeighbours set to 6 and maxdist to 200, to construct a spatial adjacency network and generate an adjacency matrix. Effective spatial connections were screened with minK set to 4, and the threshold for cell interaction was set to 0.1. ATP13A2 + Mali was represented in yellow, while CAF, TAM, and TEC were depicted in purple in the spatial distribution map of cell interactions. Subsequently, in Python, we utilized the CellChat database within the Commot library to illustrate the communication signals of ligand-receptor pairs related to the research samples. Ligand-receptor pairs were selected based on the intercellular communication results from the previous scRNA-seq analysis (Fig. 2D-F). For ATP13A2 + Mali and TEC, we selected the VEGFA-VEGFR2 ligand-receptor pair. For ATP13A2 + Mali and CAF, we chose the MDK-SDC2 pair. For ATP13A2 + Mali and TAM, we selected the MDK-NCL pair. The spatial distribution of ligand-receptor pairs signals was visualized.

Deep learning survival neural network

From the above analysis, we can conclude that ATP13A2 + Mali play an important role in HCC. Therefore, we used the Deepsurv method to construct a prognostic model based on ATP13A2 + Mali. We extracted differentially expressed genes from ATP13A2 + Mali, setting logfc.threshold to 0.1 and removing ribosome-related genes. GDC TCGA Liver Cancer (LIHC) was obtained from Xena (https://xenabrowser.net/datapages/) as the training set, and the HCC dataset from ICGC (https://dcc.icgc.org/) served as the validation set. Correct the batch effect of training and test sets. We performed univariate Cox regression filtering (pFilter = 0.05) on the differentially expressed genes, followed by lasso regression analysis with tenfold cross-validation repeated 100 iterations on the filtered genes to identify key genes for deep learning. The lambda.min value, which represents the penalty parameter corresponding to the minimum cross-validation mean squared error, was selected as the optimal hyperparameter to screen the key prognostic genes with non-zero coefficients. DeepSurv was employed in Python for in-depth analysis of the training and validation sets. DeepSurv is a novel deep learning method that integrates Cox proportional hazards for survival analysis, demonstrating performance comparable to or exceeding existing survival models [26, 27]. We set 1000 training cycles and employed the Nesterov momentum optimization algorithm (lasagne.updates.nesterov_momentum) to update model parameters, ensuring efficiency during training. L2 regularization penalty of 10.0, batch normalization set to true, a dropout rate of 0.4 to prevent overfitting, and a network architecture consisting of two fully connected hidden layers, each with 25 nodes. The trained model was subsequently used to predict the risk for the training and validation sets, yielding risk scores for each sample. Samples were categorized into high-risk and low-risk groups based on the median risk score. Survival analysis was performed using the survival and survminer packages in R, with survival curves for the high-risk and low-risk groups represented in red and blue, respectively. Additionally, we calculated survival ROC curves at different time points (1 year, 2 years, and 3 years) using the timeROC package.

RNA-seq analysis

Gene correlation and immune checkpoint analysis

We selected the GDC TCGA Liver Cancer (LIHC) dataset and clinical information of participants from Xena for RNA-seq analysis. Gene correlation analysis was performed on the HCC gene expression data to explore the relationship between ATP13A2 and other genes. We set the correlation threshold (corFilter = 0.6) and significance threshold (pFilter = 0.001). A correlation coefficient greater than 0 indicates a positive regulatory relationship with ATP13A2, while a coefficient less than 0 indicates a negative regulatory relationship. We selected the top 6 genes with the highest correlation coefficients and the bottom 5 genes with the lowest correlation coefficients, presenting the results in a circular graph, with red indicating positive regulation and green indicating negative regulation. Immune checkpoint analysis was then performed, with a filtering threshold (pFilter = 0.001) applied to ATP13A2. The Pearson correlation coefficients between ATP13A2 and immune checkpoint genes were calculated, with coefficients greater than 0 shown in red and those less than 0 shown in green.

GO and KEGG enrichment analysis

HCC samples were divided into two groups based on median ATP13A2 expression: high expression and low expression of ATP13A2. Differential genes were extracted from these groups, with logFCfilter set to greater than 1 and FDR set to less than 0.05. Differential genes identified were subjected to GO and KEGG enrichment analysis, with pvalueFilter and qvalueFilter both set to less than 0.05. Bubble charts and circular charts for GO enrichment analysis, along with bubble charts for KEGG enrichment analysis, were generated.

Immune cell infiltration, TME difference, and drug sensitivity analysis

The deconvo_mcpcounter function from the IOBR package was employed to conduct immune cell infiltration analysis for the high and low ATP13A2 expression groups within HCC. To analyze differences in TME, we utilized the deconvo_estimate function in IOBR. Drug sensitivity analysis files were sourced from the Genomics of Drug Sensitivity in Cancer database (https://www.cancerrxgene.org/), and the oncoPredict package was utilized to perform the drug sensitivity analysis.

Clinical characteristics analysis

We investigated the relationship between various clinical characteristics—namely age, gender, TNM classification, grade, and stage of HCC—and the expression levels of ATP13A2. Univariate Cox regression analysis and multivariate Cox regression analysis were then conducted to assess the relationships between these clinical characteristics, ATP13A2 expression, and survival time. Finally, we examined whether there were significant differences in clinical characteristics between the high and low ATP13A2 expression groups.

Immunohistochemical staining

To evaluate whether ATP13A2 is differentially expressed between normal liver tissue and HCC tissue, the ATP13A2 immunohistochemical staining patterns of normal liver tissue and HCC tissue were compared using the Human Protein Atlas database (https://www.proteinatlas.org/).

Cell culture and transfection

Human hepatoma cell lines HuH-7 (Chinese Academy of Sciences Cell Bank, Shanghai, China) and LM3 (Chinese Academy of Sciences Cell Bank, Shanghai, China) were cultured in DMEM medium (Gibco, Suzhou, China) containing 10% fetal bovine serum (Absin, Shanghai, China) and 1% penicillin–streptomycin (Beyotim, Shanghai, China) at 37 °C, 5% CO2, and saturated humidity. The sequences of ATP13A2 SI-RNA were siRNA 1#: forward: 5’- CUCUAAGGGACAUGGUCAATT-3’ reverse: 5’- UUGACCAUGUCCCUUAGAGTT-3’, 2#: forward: 5’- CGUUCAGCGUCUUCAAGUATT-3’ reverse: 5’- UACUUGAAGACGCUGAACGTT-3’. The control (SI-Control) sequences were forward: 5’-UUCUCCGAACGUGUCACGUTT-3’ and reverse: 5’-ACGUGACACGUUCGGAGAATT-3’. The siRNAs were synthesized by Genepharma Co., Ltd. (Shanghai, China). Cells were transfected according to the manufacturer’s instructions.

Quantitative real-time PCR (qRT-PCR)

All cellular RNA was extracted using RNA extraction reagent (Takara, Dalian, China). Reverse transcription was performed using PrimeScript RT reagent kit (Takara, Dalian, China) according to the manufacturer’s instructions. qRT-PCR was performed using TB Green PreMix Ex TaqTM Kit II (Takara, Dalian, China). GAPDH was selected as a reference gene, and the relative expression of the target gene was calculated using the 2^−ΔΔCt method. The primer sequences of the mentioned genes can be found in Supplementary Table 1.

Western blotting

HuH-7 cells and LM3 cells were lysed in RIPA lysis buffer (Beyotime, Shanghai, China) and PMSF (Beyotime, Shanghai, China). Protein lysates were separated by SDS-PAGE, then transferred to PVDF membranes for blocking, incubated with antibodies, and finally developed using the BeyoECL moon kit (Beyotime, Shanghai, China). Antibodies used in the experiment included rabbit anti-human ATP13A2 (1:1000, Cell Signaling Technology, Denver, USA), rabbit anti-human GAPDH (1:10,000, Affinity, Jiangsu, China), rabbit anti-human MCL-1 (1:1000, Cell Signaling Technology, Denver, USA), rabbit anti-human BCL-XL (1:1000, Cell Signaling Technology, Denver, USA), rabbit anti-human BCL-2 (1:1000, Cell Signaling Technology, Denver, USA), rabbit anti-human N-Cadherin (1:1000, Cell Signaling Technology, Denver, USA), and rabbit anti-human Vimentin (1:1000, Abmart, Shanghai, China).

Cell viability assay

Cell growth rate was assessed using the Cell Counting Kit-8 (CCK-8) (MedChemExpress, USA). An appropriate number of cells were selected for seeding on a 96-well plate. CCK-8 reagent was added to each well at a predetermined time. After incubation for about 2 h, the absorbance at 450 nm (OD450) was recorded.

Flow cytometry

The FITC-coupled Annexin-V apoptosis detection kit (BD biosciences, American) was used to detect cell apoptosis. The cells were collected into a centrifuge tube and the reagents were incubated with the cells according to the product instructions. The prepared cells were identified using a BD FACSCelesta flow cytometer. All results were analyzed using FlowJo software version 10.8.1.

Wound healing

An appropriate number of cells were seeded in a 6-well plate. After the cells were fully fused on the plate, a straight line was drawn on the monolayer of cells with a 10 µL pipette tip. The cells were cultured in DMEM medium containing 1% fetal bovine serum. The wound healing was observed and recorded under a microscope at 0 h, 24 h, and 48 h after the scratch. Image J was used to analyze the data.

Statistical analysis

All statistical charts and statistical results were analyzed using GraphPad Prism 9.5 software. Unpaired t-test and analysis of variance (ANOVA) were used to compare the experimental data of two groups and multiple groups. All data are expressed as mean ± SD, and p < 0.05 was considered statistically significant.

Result

ATP13A2 is positively correlated with HCC risk and associated with poor prognosis

We screened genes with p_SMR values less than 0.05 and that encode proteins, yielding a total of 70 genes (Supplementary Table 2). By intersecting these with aging-related genes (Fig. 1A), we identified 3 genes: ATP13A2, CHRNA4, and NOD2 (Fig. 1A, Table 1). The p_HEIDI values for these 3 genes were all greater than 0.05, indicating no heterogeneity among them. The prognostic relationships of these genes with HCC were examined in GEPIA, revealing that increased expression of ATP13A2 correlates with poor prognosis in HCC, while the other 2 genes were not associated with prognosis (Fig. 1B-D). Pan-cancer analysis demonstrated that ATP13A2 is expressed at higher levels in multiple cancer types (Fig. 1E). The SMRLocusPlot and SMREffectPlot for ATP13A2 in relation to HCC are depicted in Fig. 1F-G. Collectively, these analyses suggest that ATP13A2 is an aging-related gene positively correlated with HCC risk and associated with poor prognosis.

Table 1.

SMR results of 3 aging-related genes in hepatocellular carcinoma

Gene	b_GWAS	se_GWAS	p_GWAS	b_eQTL	se_eQTL	p_eQTL	b_SMR	se_SMR	p_SMR	p_HEIDI
ATP13A2	0.25	0.11	0.019	0.76	0.12	8.96E-11	0.32	0.15	0.028	0.40
CHRNA4	0.19	0.064	0.0026	−0.79	0.070	1.65E-29	−0.24	0.083	0.0037	0.99
NOD2	0.15	0.070	0.029	−0.61	0.062	7.19E-23	−0.25	0.12	0.033	0.62

SMR: Summary-data-based Mendelian Randomization

ATP13A2 is closely associated with the TME

The annotated cell subpopulations were visualized (Fig. 2A). The ATP13A2 expression density was plotted (Fig. 2B). Intercellular communication analysis indicated that ATP13A2 + Mali exhibited stronger interactions with three types of TME cells: CAF, TAM, and TEC compared to ATP13A2- Mali (Fig. 2C-F). ATP13A2 + Mali primarily interacted with TEC through the VEGF and semaphorin families, and with CAF and TAM mainly via MDK and SPP1. This suggests that ATP13A2 is closely linked to TME and may correlate with HCC by TME. Pseudotime analysis provided insights into the developmental and differentiation trajectories of ATP13A2 in HCC cells (Fig. 2G-H). The results of this sensitivity analysis revealed that, compared to ATP13A2-low Mali cells, ATP13A2-high Mali cells exhibit closer cellular communication relationships with TME cells (Supplementary Fig. 1); this finding is consistent with the results presented above.

ATP13A2 + Mali spatially interact with the TME

To further explore the spatial locations of ATP13A2 + Mali, ATP13A2- Mali, and TME cells and to visually observe the locations of cell-to-cell interactions, we performed stRNA-seq analysis. The spatial distributions of ATP13A2 + Mali and ATP13A2- Mali are significantly different, and the spatial positions of ATP13A2 + Mali and CAF, TAM, and TEC are closer than those of ATP13A2- Mali (Fig. 3A-E). Notably, ATP13A2 + Mali demonstrated a stronger correlation with CAF, TAM, and TEC compared to ATP13A2- Mali, corroborating the intercellular communication results (Fig. 3F). In terms of spatial distribution, it can also be observed that ATP13A2 + Mali has many interactions with CAF, TAM, and TEC (Fig. 3G-I). Based on the previous results of the intercellular communication, we selected some ligand-receptor pair signals for spatial visualization, and it can be observed that these signals are mainly concentrated in the area containing ATP13A2 + Mali (Fig. 3J-L). These stRNA-seq and scRNA-seq results revealed the close connection between ATP13A2 and TEC.

Fig. 3.

stRNA-seq analysis reveals spatial clustering of ATP13A2 + Mali drives TME crosstalk in HCC. (A – E) Visualization of spatial distribution of ATP13A2 + Mali, ATP13A2- Mali, CAF, TAM, and TEC. (F) Network diagram of cell interactions. ATP13A2 + Mali has a stronger correlation with CAF, TAM, and TEC than ATP13A2- Mali. (G) Spatial distribution of ATP13A2 + Mali interaction with TEC. (H) Spatial distribution of ATP13A2 + Mali interaction with TAM. (I) Spatial distribution of ATP13A2 + Mali interaction with CAF. (J) Visualization of VEGFA-VEGFR2 ligand-signaling. (K) Visualization of MDK-SDC2 ligand-signaling. (L) Visualization of MDK-NCL ligand-signaling. stRNA-seq: Spatial transcriptomics sequencing. TME: tumor microenvironment. ATP13A2 + Mali: ATP13A2 + Malignant cell; ATP13A2- Mali: ATP13A2- Malignant cell; CAF: cancer-associated fibroblasts; TAM: tumor-associated macrophages; TEC: tumor-associated endothelial cells

A 15-Gene risk stratification model derived from ATP13A2 + Mali delineates survival outcomes in HCC cohorts

Differentially expressed genes from ATP13A2 + Mali were extracted, resulting in 1011 genes after removing ribosome-related genes. Univariate Cox regression analysis of the differentially expressed genes yielded 536 genes. Subsequently, LASSO regression analysis identified 15 key genes for deep learning (Fig. 4A-C). We employed DeepSurv to construct a prognostic model, observing that as training iterations increased, the negative log-likelihood loss decreased, indicating improved model performance (Fig. 4D). The trained model was used to predict risks for the samples, categorizing them into high-risk and low-risk groups based on median risk scores. Survival analysis results for both the training and validation sets are illustrated in Figs. 4E-H, demonstrating a significant difference in survival rates between high-risk and low-risk patients (p < 0.05).

Fig. 4.

Construction and validation of a 15-gene risk model for HCC based on ATP13A2 + malignant cells using Deep Learning Survival Neural Network. (A) LASSO regression coefficient path diagram for differentially expressed genes from ATP13A2 + Mali. (B) LASSO regression cross-validation curve. (C) PPI network of the 15 key genes identified by LASSO regression. (D) Negative log-likelihood loss curve of DeepSurv model during training, showing decreasing loss and improved model performance with increasing training iterations. (E—F) Survival curve and ROC curve of the training set. (G—H) Survival curve and ROC curve of the test set. HCC patients with high risk scores have worse prognosis than those with low risk scores. DeepSurv: Deep Learning Survival Neural Network; HCC: hepatocellular carcinoma; PPI: protein–protein interaction; DeepSurv: Deep Learning Survival Neural Network

ATP13A2 Overexpression links to poor HCC prognosis, immune hotspots, and altered therapeutic vulnerability

Gene correlation analysis indicated that ATP13A2 was positively regulated by NCDN, PLEKHM2, DGKZ, RCC2, LIMK1, and MAP7D1, while negatively regulated by APOC3, SERPINC1, APOC1, SLC27A5, and RBP4 (Fig. 5A). Immune checkpoint analysis revealed that ATP13A2 was positively correlated with most immune checkpoint genes (Fig. 5B). GO enrichment analysis indicated that differentially expressed genes were enriched in cellular transport processes, including synapse organization, transmembrane transporter complex, and monoatomic ion channel activity (Fig. 5C). KEGG enrichment analysis showed that differentially expressed genes were enriched in pathways such as neuroactive ligand-receptor interaction, cytokine-cytokine receptor interaction, and the cAMP signaling pathway (Fig. 5D). Immune cell infiltration analysis indicated that the abundance scores of various immune cells in the ATP13A2 high-expression group were higher than those in the low-expression group (Fig. 5E). Notably, endothelial cells and fibroblasts had higher abundance scores in the ATP13A2 high-expression group, corroborating findings from intercellular communication and stRNA-seq analyses. TME difference analysis revealed that both StromalScore and ImmuneScore were higher in the ATP13A2 high-expression group compared to the low-expression group (Fig. 5F), further supporting the association of ATP13A2 with theTME. Drug sensitivity analysis indicated that the ATP13A2 low-expression group may exhibit greater sensitivity to selumetinib, BI − 2536, AZD2014, Doramapimod, and SB505124, suggesting potential new therapeutic avenues for HCC treatment (Figs. 5G-K). Clinical characteristics analysis (Figs. 6A-G) revealed that female patients had higher ATP13A2 expression than male patients. Additionally, ATP13A2 expression levels were higher in patients with T2 and T3 stages compared to those with T1 stage, and higher in Grade 3 patients compared to Grade 1 patients. ATP13A2 expression was also elevated in Stage II and Stage III patients relative to Stage I patients. Both univariate and multivariate Cox regression analyses indicated that ATP13A2 expression and tumor stage were associated with survival time (Figs. 6H-I). Differences in tumor stage were observed between the high and low ATP13A2 expression groups (Fig. 6J).

Fig. 5.

RNA-seq analysis reveals ATP13A2 overexpression links to poor HCC prognosis, immune hotspots, and altered therapeutic vulnerability. (A) Circle plot of gene correlation analysis (corFilter = 0.6, pFilter = 0.001). (B) Results of immune checkpoint analysis (pFilter = 0.001). (C) Bubble plot of GO enrichment analysis results. (D) Bubble plot of KEGG enrichment analysis results. (E) Results of immune cell infiltration analysis. The abundance scores of various cells in the ATP13A2 high expression group in HCC were higher than those in the ATP13A2 low expression group. (F) Violin plot of TME difference analysis. StromalScore and ImmuneScore were higher in the ATP13A2 high-expression group compared to the low-expression group (p < 0.05). (G—K) Results of drug sensitivity analysis. Selumetinib, BI − 2536, AZD2014, Doramapimod, and SB505124 in hepatocellular carcinoma were more sensitive to drugs in the ATP13A2 low expression group. HCC: hepatocellular carcinoma; TME: tumor microenvironment

Fig. 6.

Results of clinical feature analysis. (A-G) Expression of ATP13A2 in different age, gender, TNM, grade and stage of HCC. (H) Forest plot of univariate Cox regression analysis. (I) Forest plot of multivariate Cox regression analysis. Both univariate Cox regression analysis and multivariate Cox regression analysis showed that high expression of ATP13A2 was not conducive to the survival of HCC patients (p < 0.05). (J) Heat map of clinical features in HCC in the two groups with high and low expression of ATP13A2 (median expression as cutoff), showing significant differences in tumor stage. HCC: hepatocellular carcinoma

Immunohistochemical staining and knockdown of ATP13A2 in HCC cells

Normal liver tissue and HCC tissue were generated using the Human Protein Atlas database. Immunohistochemical analysis showed that the expression level of ATP13A2 in HCC tissue was significantly higher than that in normal liver tissue (Fig. 7A-B). HCC cell lines HuH-7 and LM3 were selected to construct ATP13A2 knockdown cells, and the knockdown effect was identified by qRT-PCR and Western blot. The mRNA and protein expression of ATP13A2 in HuH-7 cells and LM3 cells after knockdown were significantly downregulated compared with the control group, indicating that ATP13A2 knockdown cells were successfully established (Fig. 7C-E).

Fig. 7.

Immunohistochemical staining and the effect of ATP13A2 on HuH-7 and LM3 cells’ proliferation, apoptosis, and migration. (A) Immunohistochemical staining of ATP13A2 in normal liver tissue in the Human Protein Atlas. (B) Immunohistochemical staining of ATP13A2 in HCC tissue in the Human Protein Atlas. (C—D) Compared with the control group, the mRNA expression of ATP13A2 in HuH-7 and LM3 cells after knockdown was significantly downregulated. (E) Compared with the control group, the protein expression of ATP13A2, N-cadherin, and vimentin in HuH-7 and LM3 cells after knockdown was significantly downregulated. (F—I) Cell scratch assay showed that the healing rate of HuH-7 and LM3 cells with ATP13A2 knockdown was lower than that of the control group at 48 h. (J—M) Knockdown of ATP13A2 can increase the apoptosis rate of HuH-7 and LM3 cells. (N) Compared with the control group, the protein expression of MCL-1, BCL-XL and BCL-2 in HuH-7 and LM3 cells after knockdown was significantly downregulated. (O—P) CCK-8 cell viability assays showed that, after 48 h of incubation, the absorbance at 450 nm in the si-ATP13A2 group was significantly reduced. HCC: hepatocellular carcinoma. *p < 0.05, **p < 0.01, ****p < 0.0001

ATP13A2 Depletion inhibits migration, elevates apoptosis, and reduces viability in HCC cells

Knocking down ATP13A2 resulted in decreased expression of N-cadherin and vimentin proteins in HuH-7 cells and LM3 cells (Fig. 7E). The cell scratch experiment found that the healing rate of cells with knocked-down ATP13A2 at 48 h of culture was lower than that of the control group (Fig. 7F-I). Flow cytometry was used to detect the apoptosis of HuH-7 cells and LM3 cells with ATP13A2 knockdown. The results showed that knocking down ATP13A2 could increase the apoptosis rate of cells (Figs. 7J-M). Western blot also showed that knocking down ATP13A2 in HuH-7 cells and LM3 cells resulted in reduced expression of apoptosis-related proteins MCL-1, BCL-XL, and BCL-2 (Fig. 7N). The CCK8 assay found that there was a difference in the viability of HuH-7 cells and LM3 cells between the control group and the ATP13A2 knockdown group. After 48 h of culture, the absorbance of the latter was lower than that of the former. This indicates that HCC cell proliferation was inhibited after ATP13A2 knockdown (Figs. 7O-P). These results indicate that knocking down ATP13A2 reduces the proliferation and migration of HCC cells and enhances the apoptosis of HCC cells.

Discussion

Previous studies have shown that ATP13A2 is a gene associated with HCC prognosis, but the cohort size and experimental studies were insufficient [28]. Our study just makes up for these shortcomings. We identified ATP13A2 as an important gene associated with HCC and aging using the SMR method. Our findings indicate that elevated ATP13A2 levels correlate with poor prognosis in HCC patients. scRNA-seq analysis revealed that ATP13A2 + Mali exhibits stronger interactions with TME cells, including CAF, TAM, and TEC. Furthermore, our stRNA-seq analysis allowed us to visualize the spatial distribution of ATP13A2 + Mali, ATP13A2- Mali, CAF, TAM, and TEC. It also revealed the direction and location of ligand-receptor pair signals that mediate interactions between ATP13A2 + Mali and TME cells, predominantly concentrated near ATP13A2 + Mali. This suggests that ATP13A2 may facilitate HCC growth by enhancing the TME. RNA-seq analysis showed that the ATP13A2 high-expression group exhibited higher immune cell abundance scores, StromalScore, and ImmuneScore, reinforcing the notion that ATP13A2 is closely associated with the TME.

ATP13A2 (ATPase cation transporting 13A2) encodes a member of the ATPase P5 subfamily, responsible for transporting inorganic cations and other substrates. ATP13A2 has been previously reported to be associated with Parkinson’s disease, colorectal cancer, and other diseases [29, 30]. We speculate that ATP13A2 may promote the development of HCC through multiple mechanisms. First, ATP13A2 may affect tumor development by regulating cellular senescence. Studies on the effect of ATP13A2 on HCC through aging have not yet been reported. Study on Parkinson’s disease has revealed that ATP13A2 is associated with a reduced rate of adenosine triphosphate synthesis, an increased incidence of mitochondrial DNA damage, elevated oxygen consumption, and exacerbated fragmentation of the mitochondrial network [31]. Furthermore, aging affects the behavioral and pathological phenotypes of ATP13A2 heterozygous and knockout mice [32]. Therefore, we hypothesize that the link between ATP13A2 and aging may be associated with mitochondrial damage and apoptosis. Second, our analyses indicate that ATP13A2-expressing tumor cells may enhance HCC growth by modulating the TME through intercellular communication, as evidenced by scRNA-seq, stRNA-seq and immune cell infiltration analysis. Third, immune checkpoint analysis suggests that ATP13A2 may interfere with antigen presentation in tumor immunity, allowing tumor cells to evade immune surveillance by activating immune checkpoints such as PD-1/PDCD1, CTLA4, and LAG3. Lastly, previous research has shown that ATP13A2 is involved in cellular iron homeostasis, oxidative stress, mitochondrial maintenance, exosome release, and inflammation [33–37], all of which are critical factors in HCC pathogenesis and progression.

stRNA-seq and scRNA-seq analysis demonstrated that ATP13A2 can facilitate HCC growth through interactions with TAM, TEC, and CAF. The TME is crucial in promoting cancer progression and influencing treatment responses [38]. TAM typically exhibit diverse functional phenotypes; M1-like TAM are pro-inflammatory and anti-tumor, while M2-like TAM have anti-inflammatory and pro-tumor effects [7]. M2-like TAM secrete various cytokines (e.g., CCL22, TNF-α, ICAM-1, IL-6) that enhance tumor invasion [39–41]. Additionally, CCL17 secreted by M2 macrophages is closely linked to tumor stemness through the TGFβ1 and Wnt/β-catenin signaling pathways [42]. HCC is characterized by high vascularity, and angiogenesis significantly contributes to tumor growth and metastasis [43]. Rapid tumor growth leads to hypoxia, which stimulates the secretion of pro-angiogenic factors (e.g., VEGFA, FGF) by HCC cells, promoting angiogenesis and further accelerating HCC development [44, 45]. Anti-angiogenic therapies, such as Sorafenib, Regorafenib, and Ramucirumab, have emerged as promising strategies for HCC treatment [9]. CAF can contribute to HCC chemotherapy resistance and recurrence by secreting various cytokines (e.g., BGN, VCAN, HGF), growth factors, or extracellular vesicles [46].

Our immune checkpoint analysis revealed that ATP13A2 can activate multiple immune checkpoints, suggesting a role in promoting HCC immune escape. Among these, CTLA-4, PDCD1, LAG3, and TIGIT have been extensively studied. CTLA-4 inhibits T cell-antigen presenting cell immune synapses, and mAbs targeting CTLA-4, such as Tremelimumab, have been developed [47]. PDCD1 can induce extensive dephosphorylation of T cell activation kinases, resulting in T cell inactivation [48]. Nivolumab, a potent PDCD1-targeting antibody, is FDA-approved [49]. Relatlimab, the first mAb targeting LAG-3, promotes the proliferation of CD4 + and CD8 + tumor-infiltrating lymphocytes (TILs) and the production of immune-stimulatory cytokines [50]. TIGIT is predominantly expressed in certain T cell subsets (Tregs and memory T cells) and NK cells. The anti-TIGIT drug MK-7684 has entered phase I clinical trials (NCT02964013) [50].

The advantage of this study is that it is the first time that ATP13A2 has been identified as an aging-related gene associated with HCC by integrating comprehensive analysis of SMR, multi-omics, deep learning, and in vitro experiments. The combination of these analytical methods enhances the credibility of our research results. And for the first time, we explored how ATP13A2 correlates with the development of HCC by TME. However, this study also has some shortcomings. First, a comparison between the DeepSurv model and traditional standard baseline models was not conducted in this study. In the future, we intend to explore this within a larger sample. Second, the SMR results were not corrected for multiple testing, which may increase the risk of false positives. Last, the communication signals between ATP13A2 and the three TME cells and the results of drug sensitivity have not been experimentally verified, which is also the focus of our future research.

Conclusion

In summary, our study concludes that ATP13A2 is a pathogenic gene for HCC. ATP13A2 may promote HCC growth and invasion through multiple pathways, including modulation of the TME. ATP13A2 holds promise as a biomarker for prognosis and efficacy assessment of HCC and may provide a new target for therapeutic intervention. Future studies should focus on elucidating the mechanisms and functions of ATP13A2 in HCC and exploring novel treatment strategies.

Author contributions

D.J. and X.C. conceived and designed the study, supervised the project, and served as the corresponding author. Y.W. and C.C. contributed to data acquisition and interpretation. C.C., Y.W., G.Z., F.G., J.S., Y.C. and J.W. assisted in study design, data collection, and critical revision of the manuscript. C.C. performed data analysis and drafted the initial manuscript. All authors reviewed, edited, and approved the final manuscript.

Funding

This study was supported by the Affiliated hospital of Guangdong Medical University High-level Talent Start-up Project (GCC2023018).

Data availability

The data supporting the findings of this case report are available from the corresponding author upon reasonable request.

Declarations

Ethics approval and consent to participate

This study did not require any ethics approval. Only secondary analyses were performed using publicly accessible data.

Consent for publication

Not applicable.

Informed consent

Not applicable.

Institutional review board

Not applicable.

Competing interests

The authors declare no competing interests.