Identification of a five-gene prognosticsignature for gastric cancer based on hypoxia and stemnessscores and utilizing a decision-tree approach

Background: Gastric cancer (GC) is among the deadliest cancersworldwide. Many patients are diagnosed in the late stagesof the disease, so a poor overall survival is generally observed.One of the hallmarks of solid tumors is hypoxia, which acceleratesthe formation of stemness characteristics in cancer cellsleading to a more severe malignancy. In the present study, ouraim was to conduct a decision-tree approach based on hypoxiaand stemness scores to predict the prognosis of GC patients.Materials and Methods: We first downloaded the transcriptomeand clinical data of tumoral samples of GC from TheCancer Genome Atlas database via the TCGAbiolinks package.Next, we normalized the transcriptome data using the edgeRpackage. To calculate the hypoxia and stemness scores, we usedGSVA package and the mRNAsi method, respectively. Afterwards,we carried out weighted gene co-expression networkanalysis (WGCNA) and correlated the clusters with hypoxiaand stemness scores. We selected two clusters with the highestcorrelation with these two traits, and selected ۸۱ genes withgene significance (GS) ≥ ۰.۵۵ and module membership (MM)≥ ۰.۷۵. We then ranked the genes based on their importance usingimportance function from the caret package, and conducteda decision-tree approach with selected genes by using the rpartpackage to predict the prognosis of GC patients.Results: We retrieved the transcriptome and clinical data of۳۷۵ GC samples from TCGA database. After calculation of hypoxiaand stemness scores, we scaled them between ۰ and ۱.WGCNA was carried out on the GC transcriptome data, andwe selected blue and dark turquoise modules as they showedthe highest correlation with hypoxia and stemness scores. By using GS ≥ ۰.۵۵ and MM ≥۰.۷۵, we selected ۸۱ shared genesbetween the hypoxia and stemness scores. For feature selection,we ranked these ۸۱ genes based on their importance andselected LUM, MSC-AS۱, COL۳A۱, LRRC۳۲ and COL۸A۱ toconduct a decision-tree approach. We parted the data by ۷۰/۳۰ratio to create the training and the test sets. Finally, we proposeda decision tree with the accuracy of ۶۴% to predict the prognosisof GC patients.Conclusion: The results of our study indicated that consideringhypoxia and stemness scores, as two crucial factors in cancerpathogenesis, can be used to construct machine learning modelsto predict GC patients’ prognosis