Development and validation of a prognostic prediction model for gastric cancer based on lipophagy-related genes

Introduction

Gastric cancer (GC) is a highly pathologically heterogeneous disease revealing unique pathological manifestations and molecular features, and is the fifth most common malignant tumor globally and the fourth leading cause of cancer-related deaths (1,2). The incidence of GC varies globally, with the highest rates in East Asia (Japan and Mongolia) and Eastern Europe, and generally lower rates in Northern Europe and North America, comparable to those in Africa (3). In recent years, the incidence of GC in young people (aged <50 years) has been increasing in both low and high-risk countries. In addition to H. pylori infection, GC is associated with genetic risk factors and lifestyle factors such as alcohol and tobacco use (4,5). Although surgery is the mainstay of treatment, it has limited effect on advanced or metastatic GC (3). Chemotherapy and radiotherapy are usually used as adjuvant treatments, but they have limited therapeutic effects on certain types of GC (6). Unfortunately, due to the lack of clear clinical indications, most patients are diagnosed at an advanced stage and have a poor prognosis (7). Intratumoural and intertumoral heterogeneity is a distinguishing feature of GC; however, histological classification only is insufficient to effectively individualise treatment and improve clinical outcomes in patients with GC (8). Current prognostic biomarkers for GC include carcinoembryonic antigen (CEA), carbohydrate antigen 19-9 (CA19-9), and HER2 amplification, but their low specificity limits predictive accuracy. While clinical factors such as Tumor-Node-Metastasis (TNM) staging and lymph node metastasis are widely used, they fail to reflect molecular heterogeneity. Previous prognostic models primarily relying on clinical factors or single genes have demonstrated poor performance in distinguishing between early- and late-stage disease (9). Hence, discovering fresh prognostic markers linked to GC holds substantial importance for the identification, management, and predictive judgment of GC sufferers.

Lipophagy refers to the autophagic degradation process of lipid droplets (LDs) within cells. Serving as a primary alternative energy source during nutrient deprivation, it regulates lipid storage, free fatty acid concentrations, and energy homeostasis across organisms, cell types, and pathological conditions (10,11). Lipophagy-related genes (LRGs) regulate lipid metabolism, a critical driver of GC progression (12), which is closely associated with metabolic functions (13). In tumors, lipophagy promotes disease progression by providing both energy (via fatty acid β-oxidation) and lipid building blocks required for cancer cell growth and proliferation. Its overactivation is associated with poor prognosis in cancers such as hepatocellular carcinoma (HCC) (14,15). This biological mechanism also promotes HCC metastasis by facilitating fluid shear stress-induced epithelial-mesenchymal transition (EMT) (16). Similar lipid signaling-associated EMT regulatory mechanisms have also been observed in breast cancer (17). Unlike conventional biomarkers, LRGs reflect metabolic reprogramming, which is closely associated with treatment resistance and metastasis (18). Integrating LRGs with clinical factors can address the limitations of existing models by capturing both pathological and molecular characteristics. Therefore, this study aims to screen prognosis-associated LRGs in GC and construct a risk model incorporating these genes with clinical factors to improve prognostic accuracy.

In this study, based on the transcriptome-related data on GC from The Cancer Genome Atlas (TCGA) and Gene expression Omnibus (GEO) databases, prognostic genes causally related to lipophagy treatment for GC were identified by differential expression analysis, the weighted gene co-expression network analysis (WGCNA) analysis, functional enrichment analysis, Mendelian randomization (MR) analysis, protein-protein interaction (PPI) network construction, univariate and multivariate Cox regression analyses, and a risk model was constructed. In addition, functional enrichment analysis, immune infiltration analysis, regulatory network construction, and prognostic gene expression were performed. We present this article in accordance with the TRIPOD reporting checklist (available at https://tcr.amegroups.com/article/view/10.21037/tcr-2025-725/rc).

Methods

Data source

GC-related transcriptome data, clinical data and survival data were acquired from TCGA database (https://portal.gdc.cancer.gov/), including 373 GC samples (358 with information on survival) and 32 normal samples, and were treated as training set [TCGA-stomach adenocarcinoma (STAD)]. The GSE183904 dataset (platform: GPL24676) was obtained from the GEO database (https://www.ncbi.nlm.nih.gov/geo/), comprising single-cell RNA sequencing data derived from tumor tissues of 26 GC patients and 10 normal samples. A total of 192 tumor tissue samples in the GSE15459 dataset (platform: GPL570) from the GEO database were used as validation set. The gene set REACTOME_LIPOPHAGY was mined from the Molecular Signatures Database (MSigDB, https://www.gsea-msigdb.org/gsea/msigdb) with Lipophagy as the keyword, and nine LRGs were obtained, namely PLIN3, PRKAG2, HSPA8, PRKAB1, PRKAG3, PRKAB2, PLIN2, PRKAA2 and PRKAG1. Due to the current limitation of MSigDB in LRG annotation, no additional LRGs were identified in other sub-databases. We acknowledge that this may restrict the comprehensiveness of the candidate pool, and future studies will integrate other databases (e.g., Autophagy Database) to expand the gene set. In addition, this study used the Integrative Epidemiology Unit (IEU) Open genome-wide association study (GWAS) database (https://gwas.mrcieu.ac.uk/) was adopted to acquire GWAS datasets for candidate genes and ebi-a-GCST90018849 related to GC. The ebi-a-GCST90018849 consisted of 24,188,662 single nucleotide polymorphisms (SNPs) of 476,116 European samples (n_case =1,029, n_control =475,087).

This study was mainly divided into the following four steps: (I) data collection: GC datasets (TCGA-STAD, GSE183904, GSE15459) and LRGs were retrieved from public databases. (II) Gene screening: differential expression analysis, WGCNA, and MR analysis were used to identify LRGs causally related to GC. (III) Model development: prognostic genes were screened via Cox regression, and a risk model was constructed; clinical factors were integrated to build a nomogram. (IV) Validation: the model was validated in an independent cohort, with accuracy evaluated by area under the curve (AUC).

Differential expression analysis

The DESeq2 package (v 1.36.0) (19) was utilized to compare the differentially expressed genes (DEGs) in the TCGA-STAD dataset between the GC group and normal group, and the filtering criteria applied was as follows: |log₂ fold change (FC)| >1 and P.adj <0.05. The ggplot2 (v 3.3.6) (20) and pheatmap (v 1.1.9) (21) packages were separately adopted to draw volcano map and heat map.

WGCNA

Using the GSVA package (v 1.46.0) (22), we computed the LRG scores for GC samples in the TCGA-STAD dataset. Subsequently, the GC samples were divided into two distinct groups based on high and low LRG scores, using the median LRG score value as the cutoff point, and the difference in survival rates between these groups was evaluated. WGCNA (v 1.17) (23) was utilized to establish a scale-free network in the TCGA-STAD dataset’s GC samples, utilizing the LRG scores as a trait, with the aim of identifying gene modules most relevant to lipophagy. Specifically, the initial clustering of the samples aimed to identify any potential outliers that could be excluded to enhance the precision of subsequent analyses. The soft threshold was established by examining the interactions among genes that adhered to a scale-free network distribution. Following this, with the optimal soft threshold identified, a minimum of 100 genes per module was mandated, and a merge height of 0.2 was applied for the segmentation and amalgamation of modules, adhering to the criteria of the hybrid dynamic tree cutting algorithm. Ultimately, the correlation between modules and the trait of interest was determined, and the module most relevant to the trait was selected to obtain module genes.

Functional enrichment analysis

The candidate genes were identified through the overlap of DEGs and module genes. The clusterProfiler package (v 4.4.4) (24) was utilized to perform Gene Ontology (GO) and Kyoto Encyclopedia of Genes and Genomes (KEGG) enrichment analyses on the candidate genes, aiming to identify their common functions and associated pathways (P<0.05).

MR analysis to identify candidate key genes

MR analysis was performed adopting the TwoSampleMR package (v 0.5.6) (25) with candidate genes as exposure factors and GC as outcome. In MR analysis, it is imperative to ensure the fulfillment of the following three conditions: (I) establishing a significant correlation between instrumental variables (IVs) and exposure factors, (II) ensuring that IVs remain unaffected by confounding factors, and (III) confirming that any impact of IVs on the outcome solely occurs through exposure factors (26).

The extract_instruments function was utilized to read the exposure factors and screen the IVs (P<5×10^-8) in order to identify IVs significantly associated with the exposure factors. IVs exhibiting linkage disequilibrium (LD) were eliminated (clump = TRUE, r²=0.001 and kb =10,000). IVs with an F-value less than 10 were removed. Subsequently, based on the outcome GWAS data and the previously screened IVs, IVs significantly related to the outcome were excluded, followed by conducting expose-IV-outcome matching. The harmonise_data function was utilized to synchronize effect alleles and their magnitude of impact. Subsequently, MR analysis was executed with the mr function, employing a quintet of methodologies, specifically in simple mode (27), weighted median (28), inverse variance weighted (IVW) (29), simple median (30), and weighted mode (31). The primary findings were based on the IVW algorithm (P<0.05). Candidate genes causally related to GC were selected as candidate key genes.

PPI network

The STRING database (https://string-db.org/) was adopted to predict interactions of candidate key genes at protein level to establish a PPI network (interaction score =0.15).

Screening and MR analysis of prognostic genes

The candidate key genes were analyzed by univariate Cox regression analysis (P<0.2) and proportional hazards (PH) assumption test (P˃0.05) in TCGA-STAD dataset. The screened gene expression and survival information were subjected to a stepwise and multivariate Cox regression analysis to further identify prognostic genes (P<0.2).

The MR analysis was conducted using the same method as before, with prognostic genes considered as exposure factors and GC as the outcome. Sensitivity analyses were performed to test the accuracy of the MR analysis, including tests for heterogeneity and horizontal pleiotropy, as well as Leave-One-Out (LOO) analysis. The P value of the heterogeneity and horizontal pleiotropy test was greater than 0.05 suggests an absence of sample heterogeneity and indicates that the analysis is devoid of confounding factors, respectively. The LOO analysis examined the impact of a single SNP on the overall outcome, avoiding SNPs that exhibited high sensitivity to the outcome. Finally, a Steiger direction test was performed on the data processed with the harmonise_data function to verify the causal relationship between prognostic genes and GC.

Localization analyses and network construction

Localization analysis enables us to determine the specific cellular and subcellular sites where genes exert their functions. The mRNA expression data of prognostic genes in various organs and tissues were acquired from the BioGPS database (http://biogps.org), and the network depicting the relationship between prognostic genes, tissues, and organs was visualized using Cytoscape software. The subcellular localization information of prognostic genes was retrieved from GeneCard database (https://www.genecards.org/), and those of confidence ≥3 were selected to construct a network. The RCircos package (v 1.2.2) (32) was utilized to demonstrate the chromosomal localization of prognostic genes. The regulatory network encompasses a diverse array of RNA molecules such as mRNAs, pseudogenes of coding genes, lncRNAs, and miRNAs, offering researchers a novel perspective for transcriptome studies and facilitates a more comprehensive and in-depth understanding of certain biological phenomena. In the Starbase database (http://starbase.sysu.edu.cn/), miRNAs and lncRNAs associated with prognostic genes were predicted. Based on the intersection miRNAs from miRanda (http://www.microrna.org), PITA (http://genie.weizmann.ac.il/), and miRmap (http://mirnamap.mbc.nctu.edu.tw/) databases in Starbase database, we identified key miRNAs. The criterion for selecting predicted lncRNAs associated with key miRNAs was clipExpNum ≥20, and a mRNA-miRNA-lncRNA network was constructed.

Development and validation of risk models

A risk model was developed using expressions (expr) and coefficients (codf) derived from the multivariate Cox analysis of prognostic genes: riskScore=∑i=1ncoef(genei)∗expr(genei). Coef represents the coefficient of the gene, and expr represents the expression of the gene. Risk scores for GC patients in the TCGA-STAD dataset was concurrently calculated to plot a risk curve, and patients were subsequently categorized into high and low-risk parts based on the median risk score to perform survival analysis. Based on the patient’s risk score, the survival ROC package (v 1.0.3) (https://CRAN.R-project.org/package=survivalROC) was applied for drawing the receiver operating characteristic (ROC) curve, and the AUC was computed to characterize the accuracy of risk model. According to previous studies (33), a model with AUC ≥0.7 was considered to have good prognostic value, and AUC ≥0.8 was regarded as excellent. Again, the above analysis was performed on the GSE15459 dataset to verify the generalizability of the risk model.

Independent prognostic analysis

The univariate Cox analysis incorporated the clinical characteristics of TCGA-STAD samples and risk score, encompassing factors such as age, gender, pathologic_M, N, T, and tumor_stage (P<0.05), followed by PH assumption test (P˃0.05). Based on above results, multivariate Cox analysis was performed to explore the independent prognosis factors, resulting in the construction of a nomogram for GC (P<0.2). In addition, calibration curves were drawn to evaluate the nomogram. Meanwhile, ROC curves and decision curves were drawn to evaluate the prediction accuracy of the nomogram at 1-, 3-, and 5-year.

Gene set enrichment analysis (GSEA)

The DESeq2 package (v 1.36.0) was applied for analyzing the distinctions between high and low risk parts, with the log₂FC value from the differential analysis serving as the ranking criterion. Subsequently, GSEA based on KEGG background gene sets (c2.cp.kegg.v2023.1.Hs.symbols) in MSigDB (https://www.gsea-msigdb.org/gsea/msigdb) was performed using clusterProfiler package (v 4.4.4).

Immune analysis

The abundance scores for 28 immune cells in GC patients of TCGA-STAD were calculated using single sample GSEA (ssGSEA). The difference of immune cell infiltration between the two risk parts was compared by the Wilcoxon test (P<0.05). Subsequently, we conducted an analysis to determine the correlation between the prognostic genes and various types of immune cells. Additionally, Tumor Immune Dysfunction and Exclusion (TIDE) scores were determined for the two risk groups among patients to evaluate their potential responsiveness to immunotherapy (P<0.05).

Single-cell RNA sequencing analysis

First, the Seurat package (v 3.0) (34) was utilized for cell quality control in the GSE183904 dataset: (I) genes that were common to at least three cells and cells exhibiting a gene count between 500 and 6,000 were selected for retention; (II) cells exhibiting mitochondrial RNA percentages exceeding 20 were excluded from the analysis. The vst method was employed to identify genes exhibiting a high coefficient of variation across cells after logarithmic normalization, followed by data standardization and centralization. The next step involved employing principal component analysis (PCA) to perform dimensionality reduction on these genes, followed by the selection of PCs for subsequent analysis. The cells were subjected to unsupervised clustering analysis adopting the FindNeighbors and FindClusters functions in Seurat package (v 4.3.0). Subsequently, the uniform manifold approximation and projection (UMAP) method was employed for cell clustering, and the cell clusters were annotated based on literature annotation information (35). The Wilcoxon test was applied to evaluate the variability in the expression levels of prognostic genes within identified cell types, as well as to discern the disparities in cell counts between GC and normal control groups. Cells demonstrating a statistically significant difference between these groups were considered as pivotal for further analysis. The CellChat package (v 1.6.1) (36) was utilized for the analysis of intercellular communication and interaction. Furthermore, the monocle package (v 2.26.0) (37) was employed to simulate the pseudotime analysis of cells. Then, the expression of prognostic genes in the pseudotime cells was analysed. Finally, using single-cell analysis tools, plot the heatmap of differentiation trajectories of prognostic genes in key cells with the plot_pseudotime_heatmap function in the monocle package (version 2.30.1).

Reverse transcription quantitative polymerase chain reaction (RT-qPCR)

To validate the findings from the analysis of public databases, we obtained 10 pairs of GC and adjacent non-cancerous tissue samples from clinical sources. Cancer tissue samples and paracancerous tissue samples were taken from 10 GC patients in the Tumour Hospital of Xinjiang Medical University in the experimental and control groups, respectively. All samples were subjected to RT-qPCR. This study was conducted in accordance with the Declaration of Helsinki and its subsequent amendments. The study was approved by the Scientific Research Ethics Review Committee of Xinjiang Medical University (No. S-2024004, 30th April 2024) and informed consent was obtained from all participants. RNA isolation and RT-qPCR were performed on these samples. The TRIzol (Ambion, Austin, USA) was used to extract total RNA according to the manufacturer’s instructions. The First-strand-cDNA-synthesis-kit (Servicebio, Wuhan, China) was employed for reverse transcription of total RNA into cDNA as per the manufacturer’s guidelines. Subsequently, RT-qPCR was conducted using the 2xUniversal Blue SYBR Green qPCR Master Mix (Servicebio, Wuhan, China) following the provided instructions. The sequences of primers used for RT-qPCR are listed in Table S1. Gene expression levels were normalized to GAPDH as an internal reference and calculated using the 2^−ΔΔCt method (38).

Statistical analysis

R software (v 4.2.1) was adopted to process data. A significance level of P<0.05 was deemed as statistically significant.

Results

Acquisition of candidate key genes

The differential expression analysis yielded 5,235 DEGs in TCGA-STAD, comprising 2,683 up-regulated DEGs and 2,552 down-regulated DEGs (Figure 1A,1B). The survival rate was notably different between the high and low LRG score groups, suggesting that the prognosis of GC was associated with the lipophagy (Figure 1C). First of all, in WGCNA, the results of sample clustering did not find any outlier samples (Figure S1A). The optimal threshold of 5 was determined based on an R² value of 0.85 for dividing and merging modules (Figure S1B-S1D). Then, 16 modules were obtained to calculate the correlation with the LRG score. The paleturquoise module (R=0.59, P<0.05) exhibited the highest statistically significant positive correlation with LRGs and emerged as the key gene module comprising a total of 7,365 module genes (Figure 1D). Finally, 5,235 DEGs in TCGA-STAD were overlapped with 7,365 module genes to obtain 1,815 candidate genes enriched in muscle system processes, contractile fibers, structural constituent of muscle, and other GO functions, as well as calcium signaling pathways, adrenergic signaling in cardiomyocytes, and cAMP signaling pathways, etc. KEGG pathways (Figure 1E-1G). Employing candidate genes as exposure factors and GC as outcome, MR analysis identified 35 genes that were causally related to GC which were used as candidate key genes for subsequent analysis (P<0.05 in IVW) (available online: https://cdn.amegroups.cn/static/public/tcr-2025-725-1.xlsx). After excluding discrete proteins, the PPI network consisted of 29 nodes and 47 edges (Figure 1H). Among them, MAPK3 and LRRK2 exhibited intricate interactions with other genes.

Figure 1 Identification and enrichment analysis of candidate key genes. (A) Volcano plot of differential gene analysis. Green dots represent upregulated genes, yellow dots represent downregulated genes, and grey dots represent genes with no significant difference or small fold changes. (B) Heatmap of differential gene analysis. (C) Survival in the high LRGs score and low LRGs score groups. (D) Heatmap of the correlation between module genes and LRGs. The values in the heatmap indicate the correlation, and the values in parentheses denote the P values, with red indicating a positive correlation and blue indicating a negative correlation. (E) Venn diagram of DEGs genes and key module genes. (F) GO enrichment results of candidate genes. (G) KEGG enrichment results of candidate genes. (H) The PPI network of candidate key genes. DEGs, differentially expressed genes; GC, gastric cancer; GO, Gene Ontology; KEGG, Kyoto Encyclopedia of Genes and Genomes; LRGs, lipophagy related genes; PPI, protein-protein interaction.

AKAP12, BST1, DCBLD1, PDK4, and SPART were prognostic genes

Based on 35 candidate key genes, univariate Cox regression analysis and PH assumption test selected 16 genes related to prognostic (P<0.2) (Figure 2A, Table S2). Finally, five prognostic genes (AKAP12, BST1, PDK4, DCBLD1, and SPART) were obtained by multivariate Cox regression analysis (Figure 2B). The F-statistics for the SNPs of five prognostic genes were shown in available online: https://cdn.amegroups.cn/static/public/tcr-2025-725-2.xlsx. MR analysis was performed to further confirm the causal relationship among these five prognostic genes and GC. The odd ratio (OR) of AKAP12 [OR =0.7633, 95% confidence interval (CI): 0.6205–0.939], BST1 (OR =0.9228, 95% CI: 0.867–0.9822), DCBLD1 (OR =0.9341, 95% CI: 0.8752–0.9969), PDK4 (OR =0.8067, 95% CI: 0.7068–0.9207), and SPART (OR =0.9438, 95% CI: 0.9101–0.9787) in IVW were all less than one (P<0.05), indicating they were protective factors for GC (Table 1). The scatter plot demonstrated a consistent negative slope for all five prognostic genes, while the forest plot revealed effect sizes that were consistently less than zero, aligning with previous findings (Figure 3A-3J). In the funnel plot, the SNPs showed relatively uniform distribution, which conformed to Mendelian’s second law (Figure 3K-3O). The P value of five prognostic genes in heterogeneity and horizontal pleiotropy tests were greater than 0.05, suggesting the reliability of MR analysis (Tables S3,S4). The LOO results indicated that the removal of any SNPs did not exert a disproportional impact on the outcome. There were no SNPs exhibiting remarkably high sensitivity (Figure S2). The results of Steiger’s direction test were TRUE, indicating the establishment of one-way causality in this study (Table S5). The aforementioned findings suggested that these five prognostic genes were causally related to GC and were associated with the prognosis of GC.

Figure 2 Identification of prognostic genes. (A) Results of univariate Cox regression analysis. (B) Results of multivariate Cox regression analysis.

Figure 3 MR analysis of five prognostic genes. (A-E) Scatterplot of MR analysis of 5 prognostic genes. (F-J) The forest plot of MR analysis of 5 prognostic genes. (K-O) The funnel plot of MR analysis of 5 prognostic genes. MR, Mendelian randomisation.

Localization and regulatory network of prognostic genes

AKAP12, BST1, DCBLD1, PDK4, and SPART were mainly expressed in liver, adipocytes, dentritic cells, and heart (Figure 4A). Subcellular localization analysis further found that BST1, DCBLD1, and SPART were expressed in plasma membrane, and AKAP12, SPART, and PDK4 were expressed in nucleus (Figure 4B). Moreover, BST1 was located on chromosome 4, while DCBLD1 and AKAP12 were situated on chromosome 6. Additionally, PDK4 was found on chromosome 7, and SPART was positioned on chromosome 13 (Figure 4C). In addition, the intersection of predicted miRNAs in the miRanda, PITA, and miRmap databases yielded 51 key miRNAs, and 72 lncRNAs were predicted based on these key miRNAs. A mRNA-miRNA-lncRNA network consisting of 127 nodes and 394 edges was established (Figure 4D). There used to be complex regulatory relationship between mRNAs miRNAs and lncRNAs. For example, three lncRNAs (AL162431.1, RMRP, and NEAT1) regulated AKAP12 through hsa-miR-1-3p.

Figure 4 Localization and regulatory network of prognostic genes. (A) Organ localisation network of prognostic genes. Prognostic genes in red; tissue organs in blue. (B) Subcellular localization analysis of prognostic genes. (C) Chromosomal localisation of prognostic genes. (D) The mRNA-miRNA-lncRNA network of prognostic genes.

The risk model exhibited predictive accuracy and universality

A risk model was developed using the findings from multivariate Cox regression analysis, which was riskScore = 0.1682 × AKAP12 + 0.2697 × BST1+ 0.1258 × PDK4 + 0.3581× DCBLD1 + (−0.2234) × SPART. In TCGA-STAD, GC patients were categorized into two parts, namely high- (n=174) and low-risk (n=174), based on median risk score (1.1449) (Figure 5A). The survival outcome of the two risk components showed a marked disparity, with high-risk patients demonstrating notably inferior prognosis (P=0.0077) (Figure 5B). The AUC values for 1- (0.61), 3- (0.64), and 5-year (0.71) in TCGA-STAD were all above 0.6 (Figure 5C), suggesting the risk model has a basic ability to distinguish between high- and low-risk groups. In the external validation set GSE15459, GC patients were divided into high-risk (n=96) and low-risk (n = 96) subgroups based on the median risk score of 6.3312 (Figure 5D). The AUC values for 1- (0.70), 3- (0.69), and 5-year (0.74) all exceeding 0.6 and the net benefit of nomogram greater than each independent prognosis factor suggested the nomogram demonstrated a certain degree of predictive accuracy (Figure 5E,5F). However, the limited number of clinical factors may restrict the model’s accuracy, as more detailed clinical information (e.g., treatment history, pathological subtypes) that potentially affects GC prognosis was not available in the public datasets used.

Figure 5 Construction and validation of risk models. (A) Construction of risk models. (B) KM survival curves for high and low risk groups. (C) ROC curves for risk models. (D) Construction of risk models in GSE15459. (E) KM survival curves for high and low risk groups in GSE15459. The ordinate in the figure represents the survival rate, and the abscissa represents the overall survival time. The green curve indicates the high-risk group, and the orange curve indicates the low-risk group; the significance of difference between the high-risk and low-risk groups is P=0.003, showing a significant difference. (F) ROC curves for risk models in GSE15459. AUC, area under the curve; KM, Kaplan-Meier; ROC, receiver operating characteristic.

The nomogram demonstrated good predictive accuracy

Age, gender, pathologic_M, N, T, tumor_stage, and risk score were included in both univariate and multivariate Cox regression analyses. Subsequently, age, gender, tumor_stage, and risk score were selected as independent prognosis factors to build a nomogram (Figure 6A,6B, Table S6). The corresponding scores were assigned to each factor, and the scores of all factors were aggregated to obtain the total scores. Based on the total scores, predictions were made for the 1-, 3-, and 5-year survival rates, with a higher score indicating a lower survival rate (Figure 6C). The slope of the calibration curve was close to one, suggesting a high level of accuracy in the nomogram prediction (Figure 6D).

Figure 6 Construction and validation of nomogram. (A,B) Construction of the nomogram. (C) The results of 1-, 3-, and 5-year survival rates. (D) The calibration curve of the nomogram.

AKAP12, BST1, DCBLD1, and PDK4 were positively correlated with most of immune cells

In GSEA, focal adhesion, olfactory conduction, neuroactive ligand-receptor interaction, and calcium signaling pathways were significantly enriched in the high-risk part, whereas ECM-receptor interaction was reversed (Figure 7A). The high-risk part had a significantly higher TIDE score than the low one, indicating a higher possibility of immune escape (Figure 7B). The infiltration ratio of 24 immune cells was significantly different between the two risk parts (P<0.05), such as mast cells, monocytes, and macrophage, and AKAP12, BST1, DCBLD1, and PDK4 were positively correlated with most of differential immune cells (Figure 7C-7E).

Figure 7 Enrichment analysis of prognostic genes and correlation with immune cells. (A) GSEA enrichment analysis of prognostic genes. The figure demonstrated the enrichment pathways of the top 5. (B) TIDE score for high and low risk groups. (C) The ssGSEA score for immune cells in high and low risk groups. (D) Difference in 24 immune cell infiltration rates in high and low risk groups. (E) Correlation of prognostic genes with immune cells. Yellow represented negative correlation, green represented positive correlation. ns, no significance; *, P<0.05; **, P<0.01; ***, P<0.001; ****, P<0.0001. GSEA, gene set enrichment analysis; KEGG, Kyoto Encyclopedia of Genes and Genomes; ssGSEA, single sample GSEA; TIDE, Tumor Immune Dysfunction and Exclusion.

All five prognostic genes were notably differentially expressed in fibroblast cells

After performing quality control and logarithmic normalization in the GSE183904 dataset, top 2,000 genes with large variation coefficients were selected for PCA, and the top 15 genes were named (Figure 8A, Figure S3). PCA was carried out to reduce the 2,000 genes to 50 dimensions. Then, elbowplot was used to view the effect of dimensionality reduction, and the top 35 PCs were selected for downstream analysis (Figure 8B). The clustering and cell annotation process resulted in the identification of 14 distinct cell types, that is, T cells, plasma cells, basal gland mucous cells, macrophages, pit mucous cells, endothelial cells, fibroblast cells, natural killer (NK) cells, B cells, chief cells, mast cells, proliferative cells, enteroendocrine cells, and enterocyte (Figure 8C,8D). The expression of marker genes in individual cells was shown in Figure 8D. All five prognostic genes were notably differentially expressed in fibroblast cells between GC and normal patients (Figure 8E-8I). Basal gland mucous cells, macrophages, enteroendocrine cells and proliferative cells showed significant differences between the two groups, which were recorded as the key cells (Figure S4). Through pseudotime analysis combined with differentiation trajectory heatmaps, we revealed the dynamic expression patterns of five prognostic genes (AKAP12, BST1, PDK4, DCBLD1, and SPART) in key cell types, including basal glandular mucous cells, macrophages, enteroendocrine cells, and proliferating cells (Figure S5). In basal glandular mucous cells (Figure S5A), BST1, PDK4, DCBLD1, and SPART were initially lowly expressed, and their expression increased along the differentiation trajectory over time; while AKAP12 was initially lowly expressed, showed moderate expression in the middle stage, and high expression in the late stage. In enteroendocrine cells (Figure S5B), the expression of AKAP12, SPART, and DCBLD1 started from low levels and gradually transitioned to moderate levels; the expression of PDK4 showed a gradual trend from moderate to low levels, with a significant low-expression characteristic in the late stage; whereas the expression of BST1 fluctuated obviously, with a red (high expression) peak in the middle region. In proliferating cells (Figure S5C), BST1, AKAP12, SPART, and DCBLD1 transitioned from low to high expression, showing a gradual upregulation trend, while the expression of PDK4 gradually changed from high to low, with high expression in the early stage and downregulation in the late stage. In proliferating cells (Figure S5D), BST1, PDK4, and SPART were highly expressed in the late stage; AKAP12 was lowly expressed in the early and late stages, with relatively higher expression in the middle stage; while DCBLD1 showed a gradual change from high to low expression, with high expression in the early stage and downregulation in the late stage. These results indicate that the five prognostic genes exhibit cell type-specific dynamic expression patterns during the pseudotime course, providing a new perspective for understanding their stage-dependent regulatory roles in GC development. The interactions between endothelial cells and intestinal epithelial cells were significantly enhanced in the GC group compared to the normal group, whereas the interaction between endothelial cells and fibrocytes was pronounced in the normal group (Figure 9A-9D). In basal gland mucous cells, DCBLD1 was highly expressed at stage one and three in pseudotime, while SPAPT was highly expressed in stage three (Figure 9E). In macrophages and enteroendocrine cells, BST1, DCBLD1, and SPAPT were highly expressed at stage four (Figure 9F,9G). In proliferative cells, AKAP12 was highly expressed at stage three, and DCBLD1 was highly expressed at stage two (Figure 9H). Furthermore, RT-qPCR results showed that AKAP12, BST1, and DCBLD1 were significantly up-regulated in the GC samples (Figure 10A-10E).

Figure 8 Differential expression of prognostic genes in immune cells. (A) Data standardisation to screen for highly variable genes. (B) Results of PCA downscaling. (C) UMAP plots for single-cell annotations. (D) Bubble diagrams for single-cell annotations. (E) Differential intergroup expression of AKAP12 in various cell clusters. (F) Differential intergroup expression of BST1 in various cell clusters. (G) Differential intergroup expression of DCBLD1 in various cell clusters. (H) Differential intergroup expression of PDK4 in various cell clusters. (I) Differential intergroup expression of SPART in various cell clusters. GC, gastric cancer; PCA, principal component analysis; UMAP, uniform manifold approximation and projection.

Figure 9 Changes in the expression of prognostic genes during the proposed chronological process in key cells. (A) The cell communication network of endothelial cells in the normal group. (B) The cell communication network of intestinal epithelial cells in the normal group. (C) The cell communication network of endothelial cells in the GC group. (D) The cell communication network of fibroblasts in the GC group. (E) Changes in expression of prognostic genes during cytomimetic chronotropy in basal gland mucous cells. (F) Changes in expression of prognostic genes during cytomimetic chronotropy in macrophages. (G) Changes in expression of prognostic genes during cytomimetic chronotropy in enteroendocrine cells. (H) Changes in expression of prognostic genes during cytomimetic chronotropy in proliferative cells. GC, gastric cancer.

Figure 10 RT-qPCR results of prognostic genes. (A) RT-qPCR results of AKAP12. (B) RT-qPCR results of BST1. (C) RT-qPCR results of DCBLD1. (D) RT-qPCR results of PDK4. (E) RT-qPCR results of SPART. ns, no significance; **, P<0.01. GC, gastric cancer; RT-qPCR, reverse transcription quantitative polymerase chain reaction.

Discussion

Reprogramming of cancer metabolism is a hallmark of cancer (39). Tumor metabolism encompasses a range of biochemical activities, including glucose metabolism, amino acid metabolism, and lipid metabolism. These metabolic processes are interconnected with the tumor microenvironment and contribute to the advancement of the tumor (40). Solid tumors are often in a malnourished microenvironment due to their rapid growth and the relative lack of neovascularisation. Serving as a vital energy source and a key component of cellular membranes, fatty acids are essential for fulfilling the metabolic requirements of cancer cells, especially as they contend with the constrained availability of lipids from serum in the tumor microenvironment (41). A comprehensive analysis of multiple histological data revealed that disorders of lipid metabolism induced by lipophagy may confer pre-metastatic features and directly influence the metastatic process (42). Our investigation centered on the influence of genes associated with lipophagy, known as LRGs, on the survival rates of individuals with GC. A prognostic risk assessment model was crafted that incorporated five LRGs specifically linked to GC, establishing its effectiveness as a robust and standalone tool for predicting patient outcomes. The risk model developed in this study demonstrated moderate prognostic discrimination in both the training set (TCGA-STAD) and external validation set (GSE15459), as evidenced by significant survival curve differences between high-risk and low-risk groups. However, the relatively low AUC values (ranging from 0.61 to 0.71) indicate limited predictive accuracy that currently falls short of clinical application requirements. Future research will focus on three key aspects: (I) expanding the pool of LRGs by integrating additional databases (e.g., autophagy databases) to enrich candidate genes; (II) optimizing the model by incorporating more clinical factors (e.g., treatment history, pathological subtypes) and multi-omics data (e.g., methylation, mutation profiles); (III) conducting large-scale clinical validation in multicenter cohorts to enhance its translational value. Subsequently, we conducted RT-qPCR experiments to validate the expression patterns of these five genes. The alignment of the experimental outcomes with the bioinformatics analyses suggests that this study yields valuable implications for future research pursuits.

In this study, we ultimately identified five prognostic genes (AKAP12, BST1, PDK4, DCBLD1, and SPART) using MR analysis through the TCGA database. A-Kinase anchoring protein 12 (AKAP12) has been reported identified in literature as playing a role in the modulation of cellular division and has been recognized to act as a suppressor of tumorigenesis in GC cells. It achieves its inhibitory effect on GC cells through apoptotic processes (43,44). Conversely, findings from a different study indicate that colorectal cancer stem cells exhibit an upregulation of AKAP12, which is instrumental in preserving their stem-like attributes via the AKAP12/PKC/STAT3 signaling pathway. Within the realm of cancer stem cell research, AKAP12 could emerge as a significant target for therapeutic strategies aimed at curbing the advancement of colorectal cancer (45). Cancer cells have a high rate of glycolysis and pyruvate dehydrogenase kinase (PDK) plays an important role in this phenomenon. The expression of PDK4, which has an important role in lipid regulation, is partly regulated by lipid availability, therefore an increase in PDK4 may be followed by an increase in lipid accumulation (46). Studies have reported that up-regulation of PDK4 expression enhances the malignant cellular phenotype of GC cells, and topoisomerase can regulate the migration and glycolytic processes of GC cells by increasing the expression of PDK4 (47,48). Cervical cancer chemoresistant cells exhibit active glycolytic metabolism characterised by up-regulation of PDK4 (49). Discoidin, CUB and LCCL domain containing type I (DCBLD1) is widely distributed and sequence conserved in mammals (50). Overexpression of DCBLD1, an emulsifying substrate, inhibited autophagic degradation of glucose-6-phosphate dehydrogenase and activated pentose phosphate pathway to promote cervical cancer progression (51). SPART, a regulatory target of STAT3, is also localised to the outer mitochondrial membrane and has been implicated in metabolic processes such as cellular fatty acid and oxidative metabolism (52). There is evidence that SPART is down-regulated in tumor tissues compared to neighbouring normal samples, a finding that is consistent with results from our predictive model (52). SPART interacts with the ubiquitin ligase AIP4 and functions in lipophagy. Binding to perilipin 3 (PLIN3), it recruits and activates AIP4. AIP4 ultimately triggers macrolipophagy by ubiquitinating proteins associated with LDs (53). To encapsulate, the modulation of the aforementioned genes’ expression significantly contributes to the facilitation of cancer advancement. Our experimental outcomes are largely in alignment with the current body of research, suggesting that these prognostic genes potentially exert a substantial influence on the progression of GC.

The fluctuations in the immune system’s functionality are crucial in the progression of cancer. The research findings revealed notable disparities in the presence of immune cells—such as mast cells, monocytes, and macrophages—between the groups categorized as high risk and those as low risk. Lv et al. found that IL-33 expression is elevated in tumor of GC patients, and mast cells receive stimulation from tumor-derived IL33 to prolong survival through its mediated inhibition of apoptosis. In addition, this stimulatory effect caused mast cells to produce IL-2 thereby inducing Treg expansion, thus showing immunosuppressive activity (54). The lipophagy-associated gene RAB7A serves as a potential tumor prognostic and immune infiltration-associated biomarker for predicting immunotherapeutic efficacy in certain cancers, particularly HCC, in addition to being involved in a multi-pathway target for malignant progression of HCC (55). Based on our analytical findings, LRGs potentially alter the immune milieu of GC patients, thereby impacting their prognostic outcomes. In single-cell analysis, we observed that all five of these prognostic genes exhibited varying expression patterns within fibroblast cells. Cancer associated fibroblast cells (CAFs) contribute to the progression and metastasis of GC. In a CAF-associated GC prognostic model, GC patients in the high CAFS score group had worse OS, more active cancer-associated malignant pathways, and low CAFS scores were associated with high microsatellite instability (MSI-H), mutational load, and increased immune activation (56). CAFs are an important member of the cellular compartment of the tumor microenvironment which are responsible for the assembly and remodelling of the extracellular matrix. Originating from fibroblasts that have been stimulated and from endothelial or epithelial cells that are experiencing EMT, these cells produce a range of proteins and soluble factors. These secretions can intensify the process of tumorigenesis by stimulating the activity of receptors that span the cell membrane (57). In addition, CAFs can form a protective shield against immune surveillance and the spread of anticancer drugs by remodelling the extracellular matrix (58).

Given that GC is a highly heterogeneous cancer, single-cell analyses were performed in order to reveal the specific molecular mechanisms by which five prognostic genes regulate GC-related processes. The results showed that most of the prognostic genes were expressed at higher levels in the normal group in cells that differed between groups. The research identified and assessed five prognostic genes associated with lipophagy in GC, and developed a risk model based on these findings. This model is anticipated to enhance the existing diagnostic and prognostic frameworks for GC, offering more refined tools for patient care. Furthermore, these genes hold the potential to be novel targets for both the diagnosis and therapeutic intervention of GC. Their discovery may pave the way for innovative strategies in the clinical management of GC patients.

There are still limitations in our study. Fewer LRGs, only nine, were searched from the MSigDB database. Secondly, the expression levels of the five prognostic genes in the pseudotime analysis did not significantly change during the cellular pseudotime process. In addition to this, the development of the nomogram was constrained by limited clinical variables (only age, sex, and tumor stage) due to insufficient detailed information in the public datasets (TCGA-STAD and GSE15459), particularly regarding treatment regimens, pathological subtypes, and comorbidities—all known prognostic factors in GC. This data limitation may have contributed to the model’s modest predictive accuracy. We plan to address this issue in future studies by incorporating comprehensive clinical data from multicenter cohorts.

Conclusions

We screened and obtained five prognostic genes associated with GC and established prognostic models, which provided some new insights and references for the diagnosis and treatment of GC.

Acknowledgments

None.

Footnote

Reporting Checklist: The authors have completed the TRIPOD reporting checklist. Available at https://tcr.amegroups.com/article/view/10.21037/tcr-2025-725/rc

Data Sharing Statement: Available at https://tcr.amegroups.com/article/view/10.21037/tcr-2025-725/dss

Peer Review File: Available at https://tcr.amegroups.com/article/view/10.21037/tcr-2025-725/prf

Funding: This work was supported by Xinjiang Uygur Autonomous Region Graduate Innovation Program and Xinjiang Uygur Autonomous Region Graduate Innovation Program (No. XJ2025G181), and the National Science Foundation of Xin-jiang Uygur Autonomous Region (No. 2022D01D77).

Conflicts of Interest: All authors have completed the ICMJE uniform disclosure form (available at https://tcr.amegroups.com/article/view/10.21037/tcr-2025-725/coif). The authors have no conflicts of interest to declare.

Ethical Statement: The authors are accountable for all aspects of the work in ensuring that questions related to the accuracy or integrity of any part of the work are appropriately investigated and resolved. This study was conducted in accordance with the Declaration of Helsinki and its subsequent amendments. Approval was granted by the Scientific Research Ethics Review Committee of Xinjiang Medical University (S-2024004, 30th April 2024). The privacy rights of human subjects have been observed, and informed consent was obtained from all participants.

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

References

1. Onoyama T, Ishikawa S, Isomoto H. Gastric cancer and genomics: review of literature. J Gastroenterol 2022;57:505-16. [Crossref] [PubMed]
2. Siegel RL, Miller KD, Fuchs HE, et al. Cancer Statistics, 2021. CA Cancer J Clin 2021;71:7-33. [Crossref] [PubMed]
3. Sung H, Ferlay J, Siegel RL, et al. Global Cancer Statistics 2020: GLOBOCAN Estimates of Incidence and Mortality Worldwide for 36 Cancers in 185 Countries. CA Cancer J Clin 2021;71:209-49. [Crossref] [PubMed]
4. Lu L, Mullins CS, Schafmayer C, et al. A global assessment of recent trends in gastrointestinal cancer and lifestyle-associated risk factors. Cancer Commun (Lond) 2021;41:1137-51. [Crossref] [PubMed]
5. Leydhecker W. Diagnosis of primary glaucoma. Can J Ophthalmol 1966;1:7-13.
6. Smyth EC, Verheij M, Allum W, et al. Gastric cancer: ESMO Clinical Practice Guidelines for diagnosis, treatment and follow-up. Ann Oncol 2016;27:v38-49. [Crossref] [PubMed]
7. Qiu H, Cao S, Xu R. Cancer incidence, mortality, and burden in China: a time-trend analysis and comparison with the United States and United Kingdom based on the global epidemiological data released in 2020. Cancer Commun (Lond) 2021;41:1037-48. [Crossref] [PubMed]
8. Körfer J, Lordick F, Hacker UT. Molecular Targets for Gastric Cancer Treatment and Future Perspectives from a Clinical and Translational Point of View. Cancers (Basel) 2021;13:5216. [Crossref] [PubMed]
9. Song A, Huang Z, Xu J, et al. Development and validation of a preoperative CT-based body composition nomogram for predicting recurrence-free survival after radical surgery in patients with gastric cancer. J Gastrointest Oncol 2025;16:875-89. [Crossref] [PubMed]
10. Singh R, Kaushik S, Wang Y, et al. Autophagy regulates lipid metabolism. Nature 2009;458:1131-5. [Crossref] [PubMed]
11. Lu J, Fan Y, Liu M, et al. 1,3-dichloro-2-propanol induced lipid accumulation by blocking autophagy flux in HepG2 cells. Toxicology 2021;454:152716. [Crossref] [PubMed]
12. Guo X, Xiao K, Gong J, et al. Guanine monophosphate synthase-mediated nuclear and mitochondrial communication in the progression of gastric cancer. Transl Cancer Res 2025;14:3471-89. [Crossref] [PubMed]
13. Shin DW. Lipophagy: Molecular Mechanisms and Implications in Metabolic Disorders. Mol Cells 2020;43:686-93. [Crossref] [PubMed]
14. Carracedo A, Cantley LC, Pandolfi PP. Cancer metabolism: fatty acid oxidation in the limelight. Nat Rev Cancer 2013;13:227-32. [Crossref] [PubMed]
15. Li X, Zhou Y, Yang L, et al. LncRNA NEAT1 promotes autophagy via regulating miR-204/ATG3 and enhanced cell resistance to sorafenib in hepatocellular carcinoma. J Cell Physiol 2020;235:3402-13. [Crossref] [PubMed]
16. Su G, Feng T, Pei T, et al. Autophagy modulates FSS-induced epithelial-mesenchymal transition in hepatocellular carcinoma cells. Mol Carcinog 2021;60:607-19. [Crossref] [PubMed]
17. Luo J, Yao JF, Deng XF, et al. 14, 15-EET induces breast cancer cell EMT and cisplatin resistance by up-regulating integrin αvβ3 and activating FAK/PI3K/AKT signaling. J Exp Clin Cancer Res 2018;37:23. [Crossref] [PubMed]
18. Rao XH, Qiu H, Zhang W, et al. FBXO32 promotes gastric cancer progression by regulating NME1. Transl Cancer Res 2025;14:3520-31. [Crossref] [PubMed]
19. Love MI, Huber W, Anders S. Moderated estimation of fold change and dispersion for RNA-seq data with DESeq2. Genome Biol 2014;15:550. [Crossref] [PubMed]
20. Gustavsson EK, Zhang D, Reynolds RH, et al. ggtranscript: an R package for the visualization and interpretation of transcript isoforms using ggplot2. Bioinformatics 2022;38:3844-6. [Crossref] [PubMed]
21. Gu Z, Hübschmann D. Make Interactive Complex Heatmaps in R. Bioinformatics 2022;38:1460-2. [Crossref] [PubMed]
22. Hänzelmann S, Castelo R, Guinney J. GSVA: gene set variation analysis for microarray and RNA-seq data. BMC Bioinformatics 2013;14:7. [Crossref] [PubMed]
23. Langfelder P, Horvath S. WGCNA: an R package for weighted correlation network analysis. BMC Bioinformatics 2008;9:559. [Crossref] [PubMed]
24. Wu T, Hu E, Xu S, et al. clusterProfiler 4.0: A universal enrichment tool for interpreting omics data. Innovation (Camb) 2021;2:100141. [Crossref] [PubMed]
25. Hemani G, Zheng J, Elsworth B, et al. The MR-Base platform supports systematic causal inference across the human phenome. Elife 2018;7:e34408. [Crossref] [PubMed]
26. Zheng J, Zhang Y, Rasheed H, et al. Trans-ethnic Mendelian-randomization study reveals causal relationships between cardiometabolic factors and chronic kidney disease. Int J Epidemiol 2022;50:1995-2010. [Crossref] [PubMed]
27. Burgess S, Thompson SG. Interpreting findings from Mendelian randomization using the MR-Egger method. Eur J Epidemiol 2017;32:377-89. [Crossref] [PubMed]
28. Bowden J, Davey Smith G, Haycock PC, et al. Consistent Estimation in Mendelian Randomization with Some Invalid Instruments Using a Weighted Median Estimator. Genet Epidemiol 2016;40:304-14. [Crossref] [PubMed]
29. Burgess S, Scott RA, Timpson NJ, et al. Using published data in Mendelian randomization: a blueprint for efficient identification of causal risk factors. Eur J Epidemiol 2015;30:543-52. [Crossref] [PubMed]
30. Chen X, Kong J, Diao X, et al. Depression and prostate cancer risk: A Mendelian randomization study. Cancer Med 2020;9:9160-7. [Crossref] [PubMed]
31. Hu J, Song J, Chen Z, et al. Reverse causal relationship between periodontitis and shortened telomere length: Bidirectional two-sample Mendelian random analysis. Front Immunol 2022;13:1057602. [Crossref] [PubMed]
32. Zhang H, Meltzer P, Davis S. RCircos: an R package for Circos 2D track plots. BMC Bioinformatics 2013;14:244. [Crossref] [PubMed]
33. Liu W, Wang S, Ye Z, et al. Prediction of lung metastases in thyroid cancer using machine learning based on SEER database. Cancer Med 2022;11:2503-15. [Crossref] [PubMed]
34. Hao Y, Hao S, Andersen-Nissen E, et al. Integrated analysis of multimodal single-cell data. Cell 2021;184:3573-3587.e29. [Crossref] [PubMed]
35. Kumar V, Ramnarayanan K, Sundar R, et al. Single-Cell Atlas of Lineage States, Tumor Microenvironment, and Subtype-Specific Expression Programs in Gastric Cancer. Cancer Discov 2022;12:670-91. [Crossref] [PubMed]
36. Jin S, Guerrero-Juarez CF, Zhang L, et al. Inference and analysis of cell-cell communication using CellChat. Nat Commun 2021;12:1088. [Crossref] [PubMed]
37. Qiu X, Mao Q, Tang Y, et al. Reversed graph embedding resolves complex single-cell trajectories. Nat Methods 2017;14:979-82. [Crossref] [PubMed]
38. Livak KJ, Schmittgen TD. Analysis of relative gene expression data using real-time quantitative PCR and the 2(-Delta Delta C(T)) Method. Methods 2001;25:402-8. [Crossref] [PubMed]
39. Hanahan D. Hallmarks of Cancer: New Dimensions. Cancer Discov 2022;12:31-46. [Crossref] [PubMed]
40. Kroemer G, Pouyssegur J. Tumor cell metabolism: cancer's Achilles' heel. Cancer Cell 2008;13:472-82. [Crossref] [PubMed]
41. Röhrig F, Schulze A. The multifaceted roles of fatty acid synthesis in cancer. Nat Rev Cancer 2016;16:732-49. [Crossref] [PubMed]
42. Cerami E, Gao J, Dogrusoz U, et al. The cBio cancer genomics portal: an open platform for exploring multidimensional cancer genomics data. Cancer Discov 2012;2:401-4. [Crossref] [PubMed]
43. Choi MC, Jong HS, Kim TY, et al. AKAP12/Gravin is inactivated by epigenetic mechanism in human gastric carcinoma and shows growth suppressor activity. Oncogene 2004;23:7095-103. [Crossref] [PubMed]
44. Choi MC, Lee YU, Kim SH, et al. A-kinase anchoring protein 12 regulates the completion of cytokinesis. Biochem Biophys Res Commun 2008;373:85-9. [Crossref] [PubMed]
45. Li K, Wu X, Li Y, et al. AKAP12 promotes cancer stem cell-like phenotypes and activates STAT3 in colorectal cancer. Clin Transl Oncol 2023;25:3263-76. [Crossref] [PubMed]
46. Pin F, Novinger LJ, Huot JR, et al. PDK4 drives metabolic alterations and muscle atrophy in cancer cachexia. FASEB J 2019;33:7778-90. [Crossref] [PubMed]
47. Zhang Z, Han S, Ouyang S, et al. PDK4 Constitutes a Novel Prognostic Biomarker and Therapeutic Target in Gastric Cancer. Diagnostics (Basel) 2022;12:1101. [Crossref] [PubMed]
48. Wang H, Sun X, Yang C, et al. Deficiency of TOP1MT enhances glycolysis through the stimulation of PDK4 expression in gastric cancer. Cancer Metab 2024;12:2. [Crossref] [PubMed]
49. Zhao Z, Ji M, Wang Q, et al. miR-16-5p/PDK4-Mediated Metabolic Reprogramming Is Involved in Chemoresistance of Cervical Cancer. Mol Ther Oncolytics 2020;17:509-17. [Crossref] [PubMed]
50. Schmoker AM, Ebert AM, Ballif BA. The DCBLD receptor family: emerging signaling roles in development, homeostasis and disease. Biochem J 2019;476:931-50. [Crossref] [PubMed]
51. Meng Q, Sun H, Zhang Y, et al. Lactylation stabilizes DCBLD1 activating the pentose phosphate pathway to promote cervical cancer progression. J Exp Clin Cancer Res 2024;43:36. [Crossref] [PubMed]
52. Cusenza VY, Bonora E, Amodio N, et al. Spartin: At the crossroad between ubiquitination and metabolism in cancer. Biochim Biophys Acta Rev Cancer 2022;1877:188813. [Crossref] [PubMed]
53. Filali-Mouncef Y, Hunter C, Roccio F, et al. The ménage à trois of autophagy, lipid droplets and liver disease. Autophagy 2022;18:50-72. [Crossref] [PubMed]
54. Lv Y, Tian W, Teng Y, et al. Tumor-infiltrating mast cells stimulate ICOS(+) regulatory T cells through an IL-33 and IL-2 axis to promote gastric cancer progression. J Adv Res 2024;57:149-62. [Crossref] [PubMed]
55. Liu Y, Ma J, Wang X, et al. Lipophagy-related gene RAB7A is involved in immune regulation and malignant progression in hepatocellular carcinoma. Comput Biol Med 2023;158:106862. [Crossref] [PubMed]
56. Mak TK, Li X, Huang H, et al. The cancer-associated fibroblast-related signature predicts prognosis and indicates immune microenvironment infiltration in gastric cancer. Front Immunol 2022;13:951214. [Crossref] [PubMed]
57. Yin Y, Liu Y, Wang Y, et al. DZIP1 expressed in fibroblasts and tumor cells may affect immunosuppression and metastatic potential in gastric cancer. Int Immunopharmacol 2023;117:109886. [Crossref] [PubMed]
58. Liu T, Zhou L, Li D, et al. Cancer-Associated Fibroblasts Build and Secure the Tumor Microenvironment. Front Cell Dev Biol 2019;7:60. [Crossref] [PubMed]