Benchmark study of feature selection strategies for multi-omics data.

Area of Science:

• Bioinformatics
• Computational Biology
• Genomics

Background:

• Multi-omics data, integrating diverse molecular variables from the same samples, is increasingly prevalent.
• Existing feature selection studies primarily focus on single-omics data, leaving optimal strategies for multi-omics data unclear.
• This study addresses the gap by comparing feature selection methods specifically for cancer multi-omics datasets.

Purpose of the Study:

• To compare the performance of various feature selection methods on cancer multi-omics data.
• To evaluate how different feature selection strategies impact predictive accuracy for binary outcomes.
• To identify the most effective feature selection methods for multi-omics data analysis.

Main Methods:

• Evaluated 8 feature selection methods (4 filter, 2 embedded, 2 wrapper) across 15 cancer multi-omics datasets.
• Utilized support vector machines and random forests as classifiers.
• Employed repeated fivefold cross-validation with accuracy, AUC, and Brier score as performance metrics.

Main Results:

• The number of selected features significantly impacts predictive performance for many methods.
• Selecting features by data type versus concurrently showed minimal impact on performance, but concurrent selection was slower.
• mRMR, random forest permutation importance, and Lasso generally outperformed other methods.
• mRMR and random forest permutation importance achieved strong performance with few features.
• Wrapper methods were computationally more expensive than filter and embedded methods.

Conclusions:

• Recommends random forest permutation importance and mRMR for feature selection in multi-omics data.
• Highlights mRMR's computational cost as a consideration despite its strong performance.
• Suggests that feature selection method choice is critical for effective multi-omics data analysis.