(a) Evolutionary statistical energies ΔE computed using the independent model separate human disease-associated variants from frequent alleles in the population, but not as strongly as the epistatic model (Fig. 3b). The separation increases with the minimum allele frequency (AF) of the variants assumed to be neutral (area under the ROC curve (AUC)=0.88 for AF≥0.1, AUC=0.90 for AF≥0.25, AUC=0.92 for AF≥0.5). (b) The epistatic model outperforms all other tested methods on the HumVar dataset without any training on disease variants, as measured by the area under the ROC curve (colored lines for individual methods; grey line: expectation for random classifier; inset: AUC across the full range of specificities (left) and up to a false positive rate of 20% (right); AUCs of SIFT are < 0.5). Since PolyPhen-2 was trained on HumVar, the results here may overestimate its performance (see Online Methods for explanation). (c) On the subset of "difficult" variants that are predicted differently by SIFT and PolyPhen-2, the epistatic model is more accurate than all other methods but overall AUCs are lower than on the full dataset (figure elements as in b).