Fig. 1

Modeling simple somatic mutations. a We divided ICGC prostate cancer donors into two classes, Low Burden (LB) or High Burden (HB), based on the number of somatic mutations in their tumors and labeled their mutations accordingly. b After modeling with a linear Support Vector Classifier (SVC), we generated a ROC curve of LB classification. Accuracy was 76% +/− 12%. c We visualized classification probabilities for test mutations. The model predicts fewer LB mutations and classifies both LB and HB with high confidence. d We show model feature weights for both classes when features were used as lone predictors. Repressed regions of the genome were more predictive of HB mutations whereas regulatory, transcribed regions of the genome or ‘deleterious’ mutations were more predictive of LB mutations