Using Supervised Learning Methods for Gene Selection in RNA-Seq Case-Control Studies.

Area of Science:

• Genomics
• Bioinformatics
• Computational Biology

Background:

• Whole transcriptome studies generate vast RNA-Seq data, making gene identification challenging.
• Multivariate analysis is crucial for understanding inter-gene relationships in biological questions.
• Differential expression analysis often yields extensive gene lists, complicating downstream research.

Purpose of the Study:

• To explore supervised learning methods for ranking genes based on RNA-Seq expression data.
• To compare the efficacy of random forests and the EPS pipeline against differential expression analysis.

Main Methods:

• Utilized random forests variable importance measures for gene ranking.
• Developed the Extreme Pseudo-Samples (EPS) pipeline using Variational Autoencoders (VAEs) and regressors.
• Applied methods to 12 cancer RNA-Seq datasets with varying sample sizes.

Main Results:

• Random forests-based gene selection outperformed differential expression analysis in 9 out of 12 datasets for identifying survival-associated genes.
• The EPS pipeline outperformed differential expression analysis in 8 out of 12 datasets.
• Both supervised methods demonstrated superior performance in specific cancer datasets.

Conclusions:

• Supervised learning-based gene selection holds significant potential for RNA-Seq studies.
• Multivariate gene selection methods should complement traditional differential expression analysis.
• These advanced techniques enhance the identification of biologically relevant gene subsets.