Supplementary Figure 11: Comparing the accuracy of the PrimateAI network and other classifiers at separating pathogenic and benign variants in 605 disease-associated genes.

a, Scatterplot showing the performance of each of the classifiers on DDD cases versus controls (y axis) and benign prediction accuracy on the withheld primate dataset (x axis). b, Comparison of different classifiers at separating de novo missense variants in cases versus controls within the 605 genes, shown on a receiver operator characteristic (ROC) curve, with area under the curve (AUC) indicated for each classifier. c, Classification accuracy and AUC for the PrimateAI network and the 20 classifiers listed in Supplementary Fig. 9. The classification accuracy shown is the average of the true positive and true negative rates, using the threshold where the classifier would predict the same number of pathogenic and benign variants as expected based on the enrichment in Fig. 4a. The maximum achievable AUC for a perfect classifier is indicated with a dotted line, assuming that de novo missense variants in DDD cases are 67% pathogenic variants and 33% benign, and de novo missense variants in controls are 100% benign.