Evaluation of variable selection methods for random forests and omics data sets.

Area of Science:

• Bioinformatics
• Computational Biology
• Genomics

Background:

• Machine learning, particularly random forests, shows promise for analyzing high-dimensional omics data.
• Variable importance measures are crucial for ranking predictors in omics studies.
• Selecting relevant variables is key for identifying active biological networks and pathways.

Purpose of the Study:

• To evaluate and compare the performance of various variable selection procedures for high-dimensional omics data.
• To identify the most effective methods for variable selection in both simulated and experimental datasets.
• To provide recommendations for choosing appropriate variable selection techniques based on study objectives and data characteristics.

Main Methods:

• Comparison of Boruta algorithm, Vita method, recurrent relative variable importance, permutation approach (Altmann), and recursive feature elimination (RFE).
• Evaluation using simulated datasets to assess power and stability.
• Analysis of publicly available experimental methylation and gene expression data.

Main Results:

• Boruta demonstrated the highest power in simulation studies, closely followed by Vita.
• Both Boruta and Vita showed similar stability in variable selection.
• Vita proved most robust under a null model and was more computationally efficient on experimental data.
• Vita offered slightly better stability and was faster than Boruta for experimental datasets.

Conclusions:

• Boruta and Vita are recommended for high-dimensional data analysis.
• Vita is preferable for large datasets due to its speed and efficiency.
• Boruta is suitable for low-dimensional settings and offers high power.
• The choice between Boruta and Vita depends on dataset size and specific analytical goals.