Fig. 4

Multiclass predictive modeling for liquid biopsy (LB)-based histological classification. a One-vs-all receiver operating characteristic (ROC) analysis was executed in combination with leave-one-out cross-validation (LOOV) using the public training set for classifier selection [16]. Evaluated classifiers include random forest (RF); support vector machine (SVM); and logistic regression (LR) with ridge, elastic net (enet), and lasso regularization. Lines represent average ROC curves. Performance is quantified by the mean area under the curve (mAUC). b Solid (SBs) and LBs were evaluated with the best model (LR with ridge penalty) from a, using one-vs-all ROC analysis (dotted lines). Abnormal (abn) LBs, defined by copy number profile abnormality (CPA > 0.623), are shown separately in addition (solid line). c LBs evaluated using default ROC analysis. d β coefficients from the best model (LR with ridge penalty) from a. For perceptibility, the most prominent regions are colored (absolute value > 1), where the six most important loci to distinguish non-small cell lung cancer from small cell lung cancer (SCLC), according to the model, are emphasized by arrows. Coefficients were multiplied by 100. e Scatter plot of the relation between the CPA score and the prediction probability for LBs. The gray box (left) shows the 1% false discovery rate (FDR) cutoff. Colors indicate histology according to pathologists. The dotted line represents an ordinary least squares fit with the corresponding Pearson correlation (r). f Custom performance plot, where numbers represent patient IDs. Paired LBs (top) and SBs (bottom) are connected. Colors represent predicted type, grid position type according to pathologists (adenocarcinomas (LUADs), left; squamous cell carcinomas (LUSCs), central; SCLCs, right). The prediction probability linearly sets character size. Position within grid squares is random