Table 2.
Evaluation of performance on additional test sets.
| Cardiomyopathies | ||||
|---|---|---|---|---|
| Pathogenic test variants (TPR) | Benign/population test variants (TNR) | |||
| SHaRe (N = 129) | ClinVar (N = 15) | HGMD (N = 145) | gnomAD (N = 2003) | |
| CardioBoost | 62.0a | 66.7 | 41.4a | 51.5a |
| M-CAP | 37.2 | 40.0 | 22.1 | 20.3 |
| REVEL | 24.0 | 53.3 | 22.8 | 5.6 |
| Arrhythmias | ||||
|---|---|---|---|---|
| Pathogenic test variants (TPR) | Benign/population test variants (TNR) | Deep mutational scanning (accuracy) | ||
| OMGL (N = 77) | HGMD (N = 138) | gnomAD (N = 1237) | Calmodulin (N = 576) | |
| CardioBoost | 88.3a | 72.5a | 64.3a | 29.0a |
| M-CAP | 59.7 | 39.9 | 9.8 | 0.3 |
| REVEL | 68.8 | 52.9 | 2.8 | 4.2 |
CardioBoost performance was evaluated against additional variant sets. Four resources provided known pathogenic variants (SHaRe cardiomyopathy registry, ClinVar (two-star submissions), a UK regional genetic laboratory (Oxford Medical Genetics Laboratory [OMGL]) and the Human Gene Mutation Database [HGMD]). Variants found in gnomAD population controls were expected to be predominantly benign. Since gnomAD includes variants seen in the previous ExAC data set that was partially used to train M-CAP and REVEL, we tested against the subset of variants in gnomAD that were not in ExAC. The number of single-nucleotide variants in each set is shown in parentheses. The TPR is reported for pathogenic variant test sets (with threshold Pr ≥ 0.9), and the TNR for benign variant test sets (with threshold Pr ≤ 0.1). We also evaluated the classification accuracies on functional mapping of amino acid substitutions in calmodulin genes obtained through a previous deep functional scanning study. For each performance measure the best algorithm is highlighted in bold. Permutation tests were carried out to evaluate whether the performance of CardioBoost was significantly different from the best value obtained by M-CAP or REVEL.
TNR true negative rate, TPR true positive rate.
aP value ≤ 0.001.