Fig. 3: ctDNA fraction prediction based on routine clinical variables.

A Predicted probability of ctDNA≥2% based on our 17-feature XGBoost model applied to 463 first-line mCRPC samples (see Supplementary Data 6 for complete list of clinical variables used for model training plus model performance metrics). True observed ctDNA ≥ 2% status is indicated with color. In-set confusion matrix for classification of ctDNA ≥ 2%. B Receiver operating characteristic curves for four separately trained and optimized XGBoost models evaluating different sets of clinical input features. C Average contribution of individual clinical input features to model predictions, quantified using Shapley (SHAP) values (evaluated on the 17-feature XGBoost model). Stars indicate clinical features selected for the parsimonious 8-variable ctDNA% prediction model. D SHAP scores for cfDNA concentration and PSA as continuous variables (evaluated on the 17-feature XGBoost model). E Uniform model prediction error across sequential lines of mCRPC treatment (pairwise comparisons use the Mann-Whitney U test and are not corrected for multiple hypothesis testing). F Scatterplot showing observed minus predicted ctDNA% in the earlier (x-axis) versus later (y-axis) timepoint for 288 same-patient sample pairs across different lines of treatment (p-value is two-sided). Positive correlation between axes suggests the existence of patient and/or tumor-specific multipliers on ctDNA% (i.e., clinical or biological variables not accounted for in our model). G Predicted probability of ctDNA ≥ 2% based on our 8-feature XGBoost model applied to 463 first-line mCRPC samples. H Validation of our 8-feature XGBoost model in two external clinical trial cohorts. var variables, conc. concentration.