Fig. 3: Comparing pathogenicity predictor performances in splicing.

A A heatmap of each tool’s Spearman correlation to all other predictors; tools such as CADD exhibit high correlation to several tools – especially Oncosplice – while others such as S-CAP and Oncosplice are less correlated indicating they may capture orthogonal information. B Ratio of pathogenic, benign, and ambiguous variants found in ClinVar for subsets of top 5% of predicted deleterious mutations using eight pathogenicity predictors’ scores. C The positive predictive values for incremental percentiles for each of 8 tools’ scores indicate that Oncosplice functional divergence can generate meaningful and narrowed search spaces on parr to TraP and MMSplice. D ROC analysis of all the tools indicates that on the task of pathogenicity performance, the scoring algorithm employed by Oncosplice is effective, with CADD being the only model to outperform in terms of AUC. E Confusion matrices for showing the predictions using binary thresholds based on Oncosplice scores for all mutations and missplicing mutations, as well as the prediction quality for considering missplicing mutations as pathogenic alone.