Fig 1.
Theoretical limits of ctDNA CNV detection.
A) Density plot for healthy donor cfDNA sequencing reads mapped to 10kb genomic bins. A negative binomial (red) and Poisson (blue) distribution was fit to the resultant data. B) The ctDNA CNV size limit of detection (in megabases) is plotted as a function of sequencing depth for single copy change at various ctDNA fractions. C) Same as panel B but for four copies gained.
Fig 2.
Percentage of cancer samples with ctDNA detectable CNV events.
The fraction of samples with at least one detectable CNV event (top panels) or two and more detectable CNV events (bottom panels) at 5 Mb (left panels) and 100 Mb (right panels) resolution are plotted per cancer type. All CNV events were considered (black) as well as only those deemed important for cancer type discrimination by our random forest model (grey).
Fig 3.
Unsupervised cancer sample clustering with ctDNA detectable CNV events.
Heat maps representing the results of unsupervised clustering of cancer samples using 100 Mb resolution (top panel) and 5 Mb resolution (bottom panel) of ctDNA CNV events. Deletions are depicted in blue and amplifications are depicted in red.
Fig 4.
Performance of the classification models.
The predictive performance of the KNN model (left panel) and the random forest model (right panel) is plotted. In general, random forest models outperform the KNN models. Positive predictive value (light gray) (PPV) remains stable across models and CNV size resolution. Accuracy (black) and true positive rate (dark gray) (TPR) remain stable at 5Mb and 100Mb resolutions for the KNN model but increase at 5Mb resolution for the random forest model.
Table 1.
Performance of the KNN and random forest classification models for determination of cancer type.
Fig 5.
Tumor classification performance.
ROC curves at 5Mb (top panels) and 100Mb ctDNA CNV resolution (bottom panels) showing performance of cancer detection (left panels) and cancer type determination (right panels) for 11 major types of solid tumors—breast adenocarcinoma (BRCA), lung adenocarcinoma (LUAD), lung squamous cell carcinoma (LUSC), uterine corpus endometrial carcinoma (UCEC), glioblastoma multiforme (GBM), head and neck squamous cell carcinoma (HNSC), colon and rectal carcinoma (COAD, READ), bladder urothelial carcinoma (BLCA), kidney renal clear cell carcinoma (KIRC), ovarian serous carcinoma (OV). A small overall increase in the AUC values when going from a 100 Mb resolution to 5 Mb resolution can be observed for both detection of cancer and determination of cancer type.
Fig 6.
Cancer type misclassification heatmap.
The frequency of cross misclassifications is depicted in a heatmap for 5 Mb (panel A) and 100 Mb (panel B) ctDNA CNV resolution. Columns correspond to the known cancer type and rows correspond to the predicted cancer type. Misclassification frequency is depicted by the darkness of each cell, with darker color reflected a higher misclassification frequency. Correct classifications are set to 0. White = 0% misclassification. Dark blue = 100% misclassification.