Fig. 5

Testing of NCBoost against a fully independent set of recently reported pathogenic variants. The figure shows the area under AUROC (a) and the AUPRC (b) obtained for NCBoost (configuration of features ABCD) together with the eight reference methods (CADD, DeepSEA, Eigen, Eigen-PC, FunSeq2, ReMM, GWAVA, and ncER; “Methods” section) tested on a fully independent set of 70 positive and 700 negative variants matched per genomic region (“Methods” section). Only the GWAVA Region version is depicted for the sake of visualization. GWAVA Unmatched and TSS versions led to AUROC of 0.59 and 0.58 and to AUPRC of 0.02 and 0.02, respectively. Notice that GWAVA and ncEM methods were considered here assuming no overlap of their training sets with the recently reported pathogenic non-coding variants evaluated in these figures. There was no re-training of NCBoost, but the same NCBoost ABCD bundle used in Figs. 3 and 4 was applied