MODEL-BASED FEATURE SELECTION AND CLUSTERING OF RNA-SEQ DATA FOR UNSUPERVISED SUBTYPE DISCOVERY.

Area of Science:

• Bioinformatics
• Computational Biology
• Genomics

Background:

• Clustering algorithms identify groups in data, useful for discovering cancer subtypes from gene expression.
• Existing methods struggle with selecting informative genes, determining optimal cluster numbers, and handling normalization factors or confounders in RNA-seq data.

Purpose of the Study:

• To introduce Feature Selection and Clustering of RNA-seq (FSCseq), an advanced algorithm for unsupervised clustering of RNA-seq samples.
• To address limitations in existing methods by incorporating gene selection, normalization factor adjustment, and confounder control.

Main Methods:

• FSCseq employs a model-based approach using a finite mixture of regression (FMR) model.
• It utilizes a penalized Classification EM algorithm with a Smoothly-Clipped Absolute Deviation (SCAD) penalty for feature selection and clustering.
• The framework integrates normalization factors and potential confounding variables.

Main Results:

• FSCseq demonstrates superior performance in simulations and real-world data analysis compared to existing methods.
• The algorithm successfully identifies novel cancer subtypes and allows for subtype prediction in new patients.
• It effectively handles batch effects and confounding variables during the clustering process.

Conclusions:

• FSCseq offers a robust and comprehensive solution for unsupervised clustering of RNA-seq data.
• The method enhances the discovery of cancer subtypes by integrating feature selection and confounder adjustment.
• FSCseq provides a valuable tool for biomedical research, improving the analysis of gene expression data.