Stemness-based gastric cancer classification by machine learning for precision diagnosis and treatment of gastric cancer.

Background: Stemness has been shown to play an important role in immunotherapy and chemotherapy response. Machine learning has been used to predict stemness-based cancer subtypes. We aimed to conduct a stemness-based classification of gastric cancer (GC) for the early identification of patients at risk for GC and guide treatment. Methods: Stemness indices [mRNA stemness index (mRNAsi)] of 389 patients with GC from The Cancer Genome Atlas (TCGA) database were generated using one-class logistic regression (OCLR) algorithm. Consensus clustering was performed to divide the patients with GC into two subtypes based on their stemness indices. Finally, four machine learning algorithms were used to construct a logistic regression model containing 12 critical genes. An external cohort was used as the validation cohort. Results: Stemness subtype cluster1 had higher mRNAsi scores and a significantly better prognosis, while stemness subtype cluster2 had higher immunocompetence. In terms of the prediction of therapeutic efficacy, patients in cluster2 may have a better response to anti-cytotoxic T lymphocyte antigen 4 (anti-CTLA4) therapy, whereas no significant response to anti-programmed cell death 1 (anti-PD1) therapy was observed in either subtype. The two subtypes showed significant differences in tolerance to chemotherapy. A total of 1,863 differentially expressed genes (DEGs) were identified based on the stemness signature of GC, of which 12 critical genes were selected to predict the stemness subtype. The consistency of the results in the validation cohort indicated a promising application of this stemness-based classification and predictive model. Conclusions: Our machine learning approach performed an overall analysis of the relationship between the stemness of GC and the therapeutic effect, identified a promising stemness-based classification of GC to predict prognosis and treatment efficacy, and developed a predictive model to make the stemness-based classification accessible for clinical practice.