Identification of influential rare variants in aggregate testing using random forest importance measures

Abstract

Aggregate tests of rare variants are often employed to identify associated regions compared to sequentially testing each individual variant. When an aggregate test is significant, it is of interest to identify which rare variants are “driving” the association. We recently developed the rare variant influential filtering tool (RIFT) to identify influential rare variants and showed RIFT had higher true positive rates compared to other published methods. Here we use importance measures from the standard random forest (RF) and variable importance weighted RF (vi-RF) to identify influential variants. For very rare variants (minor allele frequency [MAF] < 0.001), the vi-RF:Accuracy method had the highest median true positive rate (TPR = 0.24; interquartile range [IQR]: 0.13, 0.42) followed by the RF:Accuracy method (TPR = 0.16; IQR: 0.07, 0.33) and both were superior to RIFT (TPR = 0.05; IQR: 0.02, 0.15). Among uncommon variants (0.001 < MAF < 0.03), the RF methods had higher true positive rates than RIFT while observing comparable false positive rates. Finally, we applied the RF methods to a targeted resequencing study in idiopathic pulmonary fibrosis (IPF), in which the vi-RF approach identified eight and seven variants in TERT and FAM13A, respectively. In summary, the vi-RF provides an improved, objective approach to identifying influential variants following a significant aggregate test. We have expanded our previously developed R package RIFT to include the random forest methods.