Extended Data Figure 3: Clinicopathological predictors of ctDNA detection.

a, 96 patients in the preoperative cohort stratified by pathological tumour, node and metastasis (TNM) stage. b, LUSCs and ctDNA-positive LUADs are significantly more necrotic that ctDNA-negative LUADs. Significant differences in necrosis between groups: LUSCs (median necrosis, 40%; n = 31), ctDNA-positive LUADs (median necrosis, 15%; n = 11) and ctDNA-negative LUADs (median necrosis, 2%; n = 47), Kruskal–Wallis test, P < 0.001, two-sided pairwise comparisons were performed using Dunn’s procedure with Bonferroni correction. c, Univariate (left) and multivariate analyses (right) were performed, by logistic regression to determine significant predictors of ctDNA detection in early-stage NSCLC. ctDNA detection was defined as detection of two or more SNVs in preoperative plasma samples. Details regarding multivariable analysis methodology are in the Methods. d, Receiver operating characteristic (ROC) curve analysis of preoperative PET scan FDG avidity (normalized to tumour background ratio (TBR), see Methods), as a predictor of ctDNA detection (92 out of 96 PET scans were available for central review), determined by the area under the curve (AUC). Median PET TBR of detected tumours = 9.01, n = 45. Median PET TBR of undetected tumours = 3.64, n = 47. P value based on Wilcoxon rank-sum test. e, LUAD subtype analyses based on ctDNA detection and the presence of an EGFR, KRAS or TP53 driver mutation.