Practical prediction model of the clinical response to programmed death-ligand 1 inhibitors in advanced gastric cancer

The identification of predictive biomarkers or models is necessary for the selection of patients who might benefit the most from immunotherapy. Seven histological features (signet ring cell [SRC], fibrous stroma, myxoid stroma, tumor-infiltrating lymphocytes [TILs], necrosis, tertiary lymphoid follicles, and ulceration) detected in surgically resected tissues (N = 44) were used to train a model. The presence of SRC became an optimal decision parameter for pathology alone (AUC = 0.78). Analysis of differentially expressed genes (DEGs) for the prediction of genomic markers showed that C-X-C motif chemokine ligand 11 (CXCL11) was high in responders (P < 0.001). Immunohistochemistry (IHC) was performed to verify its potential as a biomarker. IHC revealed that the expression of CXCL11 was associated with responsiveness (P = 0.003). The response prediction model was trained by integrating the results of the analysis of pathological factors and RNA sequencing (RNA-seq). When trained with the C5.0 decision tree model, the categorical level of the expression of CXCL11, a single variable, was shown to be the best model (AUC = 0.812). The AUC of the model trained with the random forest was 0.944. Survival analysis revealed that the C5.0-trained model (log-rank P = 0.01 for progression-free survival [PFS]; log-rank P = 0.012 for overall survival [OS]) and the random forest-trained model (log-rank P < 0.001 for PFS; log-rank P = 0.001 for OS) predicted prognosis more accurately than the PD-L1 test (log-rank P = 0.031 for PFS; log-rank P = 0.107 for OS).Subject terms: Predictive markers, Predictive markers, Translational research, Predictive medicine