Benchmarking Sparse Variable Selection Methods for Genomic Data Analyses.

Area of Science:

• Genomics
• Statistical Genetics
• Computational Biology

Background:

• Genomic studies involve numerous features, necessitating accurate variable selection.
• Bayesian inference has advanced for variable selection, but practical implementation details and performance comparisons are lacking.

Purpose of the Study:

• To conduct a comparative analysis of Bayesian variable selection approaches for genomic data.
• To evaluate the performance of shrinkage, global-local, mixture priors, SUSIE, and a proposed RFSFS method.

Main Methods:

• Comparative analysis of Bayesian variable selection methods.
• Evaluation using metrics like False Discovery Rate (FDR), False Negative Rate (FNR), F-score, and mean squared prediction error.
• Simulation studies under various scenarios, including uncorrelated and correlated features.

Main Results:

• No single method uniformly outperforms others across all scenarios and metrics.
• LASSO, spike-and-slab prior with normal slab (SN), and RFSFS are competitive for FDR and F-score with uncorrelated features.
• SN, SuSIE, and RFSFS are competitive for FDR with correlated features; LASSO excels in F-score over SuSIE.

Conclusions:

• Method performance varies depending on feature correlation and evaluation metrics.
• The proposed RFSFS method demonstrates competitive performance alongside established techniques like LASSO and SN.
• Findings offer methodological direction for variable selection in genomic analyses, including The Cancer Genome Atlas (TCGA) data.