A prognostic model for gastric cancer constructed by multiple machine learning algorithms

Gastric cancer (GC) is a highly heterogeneous disease that requires highly accurate prognostic models. Machine learning is a powerful tool for identifying predictive biomarkers and developing prognostic models. Here, we aim to integrate bioinformatics and machine learning algorithms to construct a risk model to predict prognosis of GC patients. Transcriptome data and clinical information of GC patients were obtained from the Cancer Genome Atlas (TCGA) database. Microarray data (GSE84437 and GSE26253) were obtained from the Gene Expression Omnibus (GEO) database. Univariate Cox regression analysis was used to screen prognostic genes. The risk genes closely related to prognosis were screened by machine learning algorithms and the risk score was calculated. Kaplan-Meier survival curve, time-dependent receiver operating characteristic (ROC) curve, univariate and multivariate Cox regression analysis were used to verify the validity of the risk model. The protein expression of hub genes in GC tissues was evaluated by immunohistochemical staining. 7 hub genes (CGB5, FEM1A, MATN3, ZNF101, MARCKS, BRI3BP and APOD) were identified and correlated with GC prognosis. A high-precision risk model based on random survival forest (RSF) and generalized boosted regression modelling (GBM) was constructed using these 7 hub genes. The risk model has good predictive ability for GC patients’ prognosis, and the risk score could be used as an independent prognostic factor for GC. In addition, the protein expression levels of CGB5, MATN3, MARCKS and APOD in GC tissues were significantly higher than those in normal tissues, and correlated with the pathological characteristics of GC patients. The risk model composed of 7 hub genes can accurately evaluate the prognosis of GC patients, which may contribute to the precise and personalized treatment of GC patients.