Fig. 3: Comparison of 28 classifiers and the performance of BELIVE in predicting the primary sites of metastatic cancer (n = 215).

a The radar plot illustrates the area under the curve (AUC) values of 28 classifiers. b BELIVE performance in detecting the primary sites of ten common metastatic cancers. Sample size and recall are plotted at the top of the confusion matrix, while precision is plotted on the right. The rows in the matrix show the primary cancer sites predicted by BELIVE and the columns show the authentic primary cancer sites. The colored squares along the diagonal represent the percentage of primary cancers correctly identified by BELIVE. c The bar chart (top) shows the proportion of samples whose primary sites were correctly identified with different confidence levels; the area charts (bottom) show the proportion of samples (y-axis) whose primary sites of metastatic cancers were correctly classified with greater than or equal to a confidence level (x-axis). d ROC curves for the classification of primary metastatic sites. e Top-k accuracies for predicting primary sites of metastatic cancers. f Sensitivities of BELIVE based on top-k predictions. The red line shows the median sensitivity of BELIVE for predicting primary sites across the ten cancers, while the blue and green lines correspond to the sensitivities for the best and worst performing cancers. g Prediction accuracies of BELIVE for different bins of tumor cell content. h Prediction accuracies were calculated over different inputs of sequencing data. The top 70 FFPE-RRBS libraries with more than 24 million paired-end reads were subjected to a downsampling analysis. After randomly dropping a fraction of the sequencing reads, the remaining data was used to test BELIVE’s prediction accuracy. Of the 70 libraries, 53 with 32 million or more sequencing reads were also evaluated. Source data are available in a supplementary file.