Comparison of thresholds
The ROC was used to determine the best threshold for each dataset to
classify the variants as splice altering or non-splice altering. Table 2
shows the comparison of the thresholds identified with the ROC curve
with the predefined threshold for the different tools suggested by the
developers. The best threshold to maximize the number of true positives
and true negatives depended highly on the dataset. For MaxEntScan,
SpliceSiteFinder-like and NNSplice the best thresholds were higher than
the predefined threshold, whereas the best thresholds for SPIDEX and
SpliceAI were lower. DSSP thresholds observed for NCSS variants were
lower than the predefined threshold whilst the best threshold for DI
variants was higher. The thresholds for S-CAP and CADD were difficult to
compare to the predefined thresholds as these utilize a threshold
depending on the location of the variant. For CADD, the threshold forABCA4 DI variants was lower than the suggested threshold. Two
tools, SpliceAI and SpliceSiteFinder-like, showed thresholds close to
the predefined threshold.