Skip to main content

Table 3 Ranking performance statistics

From: VarSight: prioritizing clinically reported variants with binary classification algorithms

Ranking System Case Rank - Median (Mean)
  All (n=189) VUS (n=111) LP (n=42) Path. (n=36)
CADD Scaled 57.0 (99.13) 69.0 (107.78) 39.5 (91.24) 28.0 (81.67)
HPO-cosine 22.0 (53.96) 22.0 (56.05) 26.0 (56.38) 19.5 (44.69)
Exomiser(hiPhive) 79.0 (105.34) 85.0 (116.33) 93.5 (101.10) 34.0 (76.42)
Exomiser(hiPhive, human only) 35.0 (53.60) 37.0 (63.84) 34.0 (45.60) 24.5 (31.36)
Phen-Gen 55.0 (48.66) 65.0 (52.91) 47.0 (47.48) 24.0 (36.92)
DeepPVP 15.0 (76.95) 23.0 (79.68) 19.5 (84.95) 6.0 (59.19)
RandomForest(sklearn) 10.0 (29.64) 15.0 (39.27) 8.0 (20.07) 4.0 (11.11)
LogisticRegression(sklearn) 6.0 (29.24) 14.0 (39.87) 3.0 (22.05) 1.0 (4.83)
BalancedRandomForest(imblearn) 8.0 (28.24) 14.0 (38.64) 5.0 (17.67) 3.0 (8.50)
EasyEnsembleClassifier(imblearn) 7.0 (28.72) 15.0 (40.15) 6.0 (18.40) 2.0 (5.50)
  1. This table shows the ranking performance statistics for all methods evaluated on our test set. CADD Scaled and HPO-cosine are single value measures that were used as inputs to the classifiers we tested. The middle four rows (two Exomiser runs, Phen-Gen, and DeepPVP) represent external tools that ranked the same set of variants as the classifier algorithms. Phen-Gen was the only external tool that did not rank every variant in the set, so we conservatively assumed unranked variants were at the next best position despite being unranked. The bottom four rows are the tuned, binary classification methods tested in this paper. Each method was used to rank (prioritize) the Codicem-filtered variants from each proband in the test set, and the position of reported variants was recorded such that lower values indicate better performance with “1” indicating the first variant in the list. The “Case Rank” columns show the median and mean ranks for all reported variants along with the variants split into their reported pathogenicity (variant of uncertain significance (VUS), likely pathogenic (LP), or pathogenic (Path.)) derived from the ACMG guidelines. All values in this table were generated using only the Codicem-filtered variants from testing set