Fig. 6: Fragment-based analysis of cfDNA enables accurate tumor detection and classification.

Prediction performance of machine learning classifiers trained to distinguish patients with EwS from healthy controls (a) and from patients with other pediatric sarcomas (b), based on the following sets of input features: global fragment-size distribution (blue); fragment coverage at EwS-specific DHSs (orange); read depth in 5 Mb bins (green); and regional fragmentation patterns (red). Results are also shown for a meta-learner combining the predictions of all individual classifiers into a weighted consensus prediction (purple). The performance of each model was evaluated by and averaged over 100 iterations of bootstrapping, separately for the different sequencing coverage levels (median of 12×, 1×, and 0.1×). CI is the 95% confidence interval obtained by bootstrapping. a ROC curves show, for each feature set, the performance for distinguishing between cfDNA samples from patients with clinical evidence for EwS (n_samples = 103) and healthy controls from three independent sets (22 controls sequenced in this study; 22 controls from Cristiano et al.³²; and 24 controls from Ulz et al.³⁴). Machine learning models were trained separately for each of the 3 control sets; the mean results over the 3*100 bootstrap iterations are shown. b ROC curves show the performance of each feature set for distinguishing between cfDNA samples from patients with EwS (n_samples = 98) and from patients with other pediatric sarcomas (n_samples = 18). For both sets of samples, we ensured the presence of tumor-derived cfDNA in the blood based on genetic evidence.