Fig. 4: Assessment and comparison of V2P performance on functional characterization of variant effect.
From: Expanding the utility of variant effect predictions with phenotype-specific models
a Comparison of V2P’s pathogenicity predictions with six previously published methods on variants from 66 deep mutational scanning (DMS) assays of 52 proteins, for which each pair of methods provided predictions. (Top) Distribution of average precision scores per assay. (Bottom) Spearman’s rank correlation coefficient (ρ) per assay. b V2P (top) and inverse DMS output (bottom) averaged at each amino acid for single-nucleotide variant (SNV) missense variants in the PRKN protein (PDB 5C1Z). Highlighted variants colored according to ClinVar4 classification: pathogenic (red) and benign (blue). c (Top) P53 families and domains. (Bottom) Inverse of DMS output, V2P pathogenicity scores, and ClinVar labels, respectively, for SNV missense variants in P53. ClinVar label key. 1: Benign, 2: Benign/Likely benign, 3: Likely Benign, 4: Uncertain significance, 5: Likely pathogenic, 6: Pathogenic/Likely pathogenic, 7: Pathogenic. d V2P and DMS scores for pathogenic and benign P53 variants from ClinVar. e Distribution of V2P scores in regions of P53. f Comparison of V2P’s pathogenicity predictions with Enformer, CADD v1.7, and FATHMM on 16 massively parallel reporter assays (MRPA) of distinct regulatory elements. (left) ρ for each MPRA. (right) Median ρ and average ρ weighted by the number of variants per assay across MPRAs. Variant effect predictor versions are detailed in the dbNSFP 4.7a. Boxes represent the quartiles of the data. Whiskers extend to points that lie within 1.5 times the interquartile range of the lower and upper quartiles. Violins extend to and are clipped at data minima and maxima.
