Additional file 2 of PRIME: an interpretable artificial intelligence model based on liquid biopsy improves prediction of progression risk in non-small cell lung cancer

Additional file 2. Fig. S1 Survival outcomes. Fig. S2 RNA-sequencing (RNA-seq) analysis of KEAP1, STK11, and CDKN2A mutations in the Cancer Genome Atlas (TCGA) database. Fig. S3 Prognostic effects of KEAP1/STK11 mutations in the overall patient cohort. Fig. S4 Immune infiltration analysis of KEAP1/STK11/CDKN2A mutations. Fig. S5 Gene set enrichment analysis (GSEA) ridge plots showing the normalized enrichment score (NES), adjusted P-value, and false discovery rate (FDR) for significantly enriched gene sets. Fig. S6 Receiver operating characteristic (ROC) curve for the combined model based on logistic regression. Fig. S7 Comparison of model performance with composite vs. separate genomic features. Fig. S8 SHapley Additive exPlanations (SHAP) for neural network (NN)-based PRIME model. Fig. S9 Progression-free survival stratified by NN-PRIME in all patients (a) and in the training set (b). Fig. S10 Sensitivity analyses of NN-PRIME model robustness. Fig. S11 Progression-free survival stratified by NN-PRIME in patients receiving personalized-panel (tumor-informed) ctDNA testing (a) and in patients receiving fixed-panel (tumor-naïve) ctDNA testing (b). Fig. S12 Subgroup analyses of NN-PRIME model performance in different ctDNA sequencing platforms, detection techniques, and patient cohorts. Fig. S13 Kaplan-Meier curves of survival outcomes in patients with different treatment regimens. Fig. S14 Proportion of disease progression in low-risk vs. high-risk patients stratified by NN-PRIME in resectable or unresectable stage II–III NSCLC patients. Fig. S15 Forest plot of logistic regression indicating high-risk patients identified by NN-PRIME correlated with poorer outcomes across various clinical subgroups.