Fig. 5.

Performance of the best CScape-somatic model with the original CScape, CADD and FunSeq2 on the ICGC test set for non-coding regions (CSS= CScape-somatic and CS= CScape). Top: CScape-somatic yields accuracy from 60.0 up to 64.2% on the ICGC test sets, substantially higher than competitors. The closest competitor changes at each ICGC recurrence level: FunSeq2 for ICGC $r\ge 2$, at 50.9%; CScape for ICGC $r\ge 3$, at 50.5% and CADD for ICGC $r\ge 4$, at 51.4%. Bottom: CScape-somatic yields AUC scores from 0.64 to 0.73. None of the competitors yield scores better than random chance (0.50), and with the exception of the original CScape, perform worse as the driver threshold r increases