Figure 4
The effect of swapping regulatory data between cancer types on model performance. The z-score is computed by comparing the total AUC difference in a swap across all SV types to the mean of performance differences from the original run to all other swaps, divided by the standard deviation of these differences. Higher z-scores thus mean that the performance is better with data from that tissue type relative to all other tested tissue types in the swap. For example, out of all swaps made, nervous system relatively performs best with data from nervous system, ovary and prostate, while the performance is worst with data from skin, urinary tract and uterus. The asterisks indicate cancer types with some missing data for which GM12878 was used.
