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

Area of Science:

• Genetics
• Statistical Genetics
• Bioinformatics

Background:

• Aggregate tests are commonly used to identify genetic regions associated with diseases by analyzing multiple rare variants simultaneously.
• When an aggregate test yields significant results, pinpointing the specific rare variants driving this association is crucial for understanding disease mechanisms.
• Previous methods, like the rare variant influential filtering tool (RIFT), have been developed to identify these influential rare variants, but further improvements are sought.

Purpose of the Study:

• To evaluate the performance of standard random forest (RF) and variable importance weighted RF (vi-RF) methods in identifying influential rare variants following significant aggregate tests.
• To compare the true positive rates (TPR) and false positive rates (FPR) of RF-based approaches against the existing RIFT method, particularly for very rare and uncommon variants.
• To apply the developed RF methods to a real-world genetic study, such as idiopathic pulmonary fibrosis (IPF), to identify disease-associated variants.

Main Methods:

• Utilized importance measures derived from standard random forest (RF) and variable importance weighted RF (vi-RF) algorithms.
• Compared the performance of RF:Accuracy and vi-RF:Accuracy methods against the rare variant influential filtering tool (RIFT).
• Evaluated variant identification performance across different minor allele frequency (MAF) strata: very rare (MAF < 0.001) and uncommon (0.001 < MAF < 0.03).
• Applied the vi-RF approach to a targeted resequencing dataset from an idiopathic pulmonary fibrosis (IPF) study.

Main Results:

• For very rare variants (MAF < 0.001), vi-RF:Accuracy demonstrated the highest median TPR (0.24), followed by RF:Accuracy (0.16), both significantly outperforming RIFT (TPR = 0.05).
• Among uncommon variants (0.001 < MAF < 0.03), RF methods achieved higher TPRs than RIFT with comparable FPRs.
• In the IPF study, the vi-RF approach successfully identified eight influential variants in TERT and seven in FAM13A.
• The R package RIFT has been updated to incorporate these new random forest-based methods.

Conclusions:

• Variable importance weighted random forest (vi-RF) offers a superior and objective approach for identifying influential rare variants after significant aggregate tests.
• RF-based methods provide enhanced sensitivity (higher TPR) for detecting causal rare variants compared to RIFT, especially across different MAF categories.
• The successful application in the IPF study highlights the clinical relevance and utility of vi-RF in genetic association studies.
• The integration of RF methods into the RIFT package expands the toolkit available for rare variant analysis in genetic research.