Using recursive feature elimination in random forest to account for correlated variables in high dimensional data

Area of Science:

• Genomics
• Biostatistics
• Machine Learning

Background:

• Random Forest (RF) is effective for high-dimensional data but struggles with correlated predictors.
• The Random Forest-Recursive Feature Elimination (RF-RFE) algorithm addresses correlated predictors in smaller datasets.
• RF-RFE's performance in high-dimensional omics data remains untested.

Purpose of the Study:

• To evaluate the performance of RF and RF-RFE in identifying causal associations in high-dimensional omics data.
• To assess the impact of correlated predictors and genotype-methylation interactions on variable detection.
• To determine the scalability of RF-RFE for large-scale omics datasets.

Main Methods:

• Integrated 202,919 genotypes and 153,422 methylation sites from 680 individuals.
• Utilized RF and RF-RFE algorithms to detect simulated causal associations with triglyceride levels.
• Incorporated simulated genotype-methylation interactions to test algorithm robustness.

Main Results:

• RF identified strong causal variables among a few correlated predictors but missed others.
• RF-RFE reduced the importance of correlated variables.
• However, RF-RFE also diminished the importance of causal variables in the presence of numerous correlated predictors, hindering detection.

Conclusions:

• Both RF and RF-RFE faced challenges in detecting causal variables within high-dimensional omics data.
• RF-RFE's effectiveness is compromised by a high number of correlated predictors, impacting causal variable identification.
• RF-RFE may not be suitable for high-dimensional omics datasets due to scalability issues.