Identification of Hub Genes Associated with Tumor-Infiltrating Immune Cells and ECM Dynamics as the Potential Therapeutic Targets in Gastric Cancer through an Integrated Bioinformatic Analysis and Machine Learning Methods

Abstract

Background: Stomach cancer, also known as gastric adenocarcinoma, remains the most common and deadly cancer worldwide. Its early diagnosis and prevention are effective to improve the 5-year survival rate of the patients. Therefore, it is important to discover specific biomarkers for early diagnosis and drug treatment. This study investigates the potential key genes and signaling pathways involved in gastric cancer. Methods: The gene expression profiles, GSE63089, GSE33335, and GSE79973, were retrieved for the identification of Differentially Expressed Genes (DEGs) within a total of 80 gastric cancer samples and 80 normal samples. A total of 1423 uP- and 1155 downregulated genes were screened for overlapping DEGs visualized via Venn diagrams along with 58 upregulated and 43 downregulated genes. These overlapping DEGs were evaluated with Gene Ontology (GO) enrichment, Kyoto Encyclopedia of Genes and Genomes (KEGG) enrichment, and Protein-Protein Interaction (PPI) network analysis. Using DAVID software, we identified several genes enriched in both GO and KEGG analyses. PPI analysis was performed with STRING software, and 3 submodules were obtained with Cytoscape software. Then, we used Cytohubba with 12 classification methods to select candidate hub genes. The group 1 genes enriched in GO and KEGG pathway intersected with group 2 genes, which were approved by nine algorithms, and group 3 genes clustered in three submodules. 9 hub genes were intersected from group 1/2/3 genes and the prognostic values were estimated through GEPIA. We found that the LUM and COL1A1 expression levels and survival outcomes displayed a favorable prognostic value (P-value = 0.013 for LUM and P-value =0.042 for COL1A1). Results: Finally, 5 machine learning methods were employed for the validation of two hub genes (COL1A1, LUM) to distinguish between the cancer samples and non-cancer samples. The accuracy of XGBoost was estimated to be 0.9375, and the precision and specificity as 1.000. The highest recalls of LR and MLP were 1.0000, and the AUC was 1.0000. In the test set GSE65801, the accuracy of all models was greater than 80%, and the XGBoost model obtained the highest prediction accuracy of 0.8906. The precision of 0.9301 and the specificity of 0.9375 were obtained. The highest recall of MLP was 0.8750 and AUC was 0.9082. The correlation of prognostic indicators with the tumor-infiltrating immune cell levels was analyzed using TIMER. Conclusion: The identified hub genes explored in this study would enhance the understanding of the molecular mechanism of gastric cancer and may be regarded as a potential therapeutic target as assessed by integrating bioinformatics and machine learning methods.