Fig. 4

Comparison of prognostic accuracy of NAT- and tumor-based models. A The distributions of P-values from estimating the NAT-DEGs and P-values from estimating the tumor-DEGs are compared. The dot in each bar graph shows the median of P-values generated from each tumor-DEG against the P-values estimated from NAT-DEGs in the indicated ranges. B Schematic of the procedure to test model accuracy of two types of prognostic models, i.e., NAT- and tumor-based model from SMC-CRC (left most panel), applying logistic regression to tumor-derived transcriptomes of TCGA-COAD (middle panel). Note that only 186 tumor-derived transcriptomes harboring information on the patients’ prognosis were used for the logistic regression analysis out of the total 450 TCGA-COAD samples with count information. The beta coefficient and 95% CI of each model are depicted as different point shapes and segments depending on the number of genes used in the model, on the rightmost panel. C, D Kaplan–Meier plot of survivals of the TCGA-COAD patient groups: the patients with high risk scores (turquoise) and the patients with low risk scores (crimson red) were classified based on the risk scores estimated from (C) NAT-based models or D tumor-based models