Clinical significance of genetic variation in hypertrophic cardiomyopathy: comparison of computational tools to prioritize missense variants

Figure 1. Distribution of HCM-associated variants (Pathogenic/Likely pathogenic) with a review status of > 1 star in ClinVar (N = 768). (A) Number and proportion of overall variants per gene. (B) Number and proportion of overall variants per category. (C) Category of variants located in the MYH7 gene. (D) Category of variants located in the MYBPC3 gene.

Figure 2. Workflow of the study. Number of variants on each dataset are presented.

Figure 3. Performance of prediction tools in classifying HCM missense variants using fixed thresholds for ClinVar (A), SHaRe (B) and Walsh_2017 (C) datasets. For each dataset, the numbers of pathogenic/likely pathogenic (N pos) and benign/likely benign (N neg) variants are indicated. Tools were ranked according to the weighted normalized MCC (weighted_norm_mcc).

Figure 4. Performance of prediction tools in classifying HCM missense variants using ROC curve analysis for ClinVar (A), SHaRe (B) and Walsh_2017 (C) datasets. For each dataset, the numbers of pathogenic/likely pathogenic (N pos) and benign/likely benign (N neg) variants are indicated. Tools were ranked according to the area under the ROC curve (auROC). The number (n) of variants predicted by each tool is indicated. Tools with more than 50% of missing predictions were not included. (D) Differences in the metrics when evaluating with auROC and weighted normalized MCC. For comparison, auROC values were weighted by the fraction of variants predicted by each tool.

Figure 5. Performance of prediction tools using adjusted thresholds on each dataset (ClinVar, SHaRe, and Walsh_2017). Optimized thresholds at Beta = 0.5 minimize the false positives (variants predicted as pathogenic that are benign). Optimized thresholds at Beta = 1 give the same importance to false positives and false negatives. Optimized thresholds at Beta = 1.5 minimize the false negatives (variants predicted as benign that are pathogenic). Tools highlighted in blue were selected as the best by averaging the ranks between the three datasets.

Figure 6. Performance of prediction tools after addressing circularity issues. Tools were ranked using the weighted normalized MCC on two new test datasets (A–C). (A) Variants identified as present in the training sets of the tools highlighted in blue were removed from the merged ClinVar, SHaRe, and Walsh_2017 datasets. (B) HCM ClinVar variants submitted after the tools highlighted in blue were developed. (C) Variants in the whole ClinVar irrespective of disease context. The tools selected as best-performers for HCM are highlighted in red (bold).

Figure 7. High-confidence prioritization of HCM-associated VUS based on predictions of the 5 top-performant tools (ClinPred, MISTIC, FATHMM, MPC and MetaLR). On the left, 63 variants for which 100% of the tools predict pathogenicity. On the right, variants predicted to be benign by more than 50% of the tools.