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

Background: Gastric cancer (GC), as a heterogeneous disease, lacks clear clinical indications, and existing methods based on histological classification are insufficient for individualized stratified treatment of GC patients. Lipophagy-related genes (LRGs) are involved in the progression of GC, but their role in GC remains unclear. This study aims to develop and validate a lipophagy-related prognostic model to improve the predictive ability for GC prognosis. Methods: GC-related datasets from public databases were used, including The Cancer Genome Atlas-Stomach Adenocarcinoma (TCGA-STAD, training set), Gene Expression Omnibus 183904 (GSE183904, single-cell dataset), and GSE15459 (external validation set). LRGs were retrieved from the Molecular Signatures Database (MSigDB) using "Lipophagy" as the keyword. Differential expression analysis on TCGA-STAD identified differentially expressed genes (DEGs). Weighted gene co-expression network analysis screened lipophagy-related module genes, overlapped with DEGs for candidate genes. Mendelian randomization (MR) analysis identified key genes with causal relationships to GC. Univariate and multivariate Cox regression analyses screened prognostic genes to construct and validate a risk model. Nomograms for 1-, 3-, and 5-year survival rates were built using significant clinical factors and risk scores. Immune and enrichment analyses compared high- and low-risk groups. Pseudotime processes of prognostic genes in cell clusters and inter-cluster communication were analyzed. Reverse transcription quantitative polymerase chain reaction (RT-qPCR) verified prognostic gene expression. Results: Thirty-five genes had causal relationships with GC; AKAP12, BST1, DCBLD1, PDK4, and SPART were selected as prognostic genes, mainly expressed in adipocytes, dendritic cells, and heart. A risk model and nomogram integrating risk scores and clinical features (age, gender, tumor stage) were constructed. Nomogram predicted values were consistent with observed ones. Survival curves in both sets showed significant differences (P<0.05). Area under the curve (AUC) values: TCGA-STAD (0.61 for 1 year, 0.64 for 3 years, 0.71 for 5 years); GSE15459 (0.67 for 1 year, 0.67 for 3 years, 0.69 for 5 years), indicating limited prognostic discrimination. AKAP12, BST1, DCBLD1, and PDK4 correlated positively with most immune cells. All 5 genes showed differential expression in fibroblasts between GC and normal samples and in pseudotime stages. Endothelial-intestinal epithelial cell interaction was enhanced in GC. RT-qPCR showed AKAP12, BST1, and DCBLD1 were up-regulated in GC samples. Conclusions: In this study, a risk model was constructed based on five prognostic genes (AKAP12, BST1, DCBLD1, PDK4, and SPART). A nomogram was established by integrating each patient's risk score and clinical features (age, gender, and tumor stage). This model exhibits preliminary prognostic discrimination ability in both the training set and the validation set; however, its predictive accuracy is limited and requires further optimization.