Selecting Classification Methods for Small Samples of Next-Generation Sequencing Data.

Area of Science:

• Bioinformatics
• Statistical Genetics
• Computational Biology

Background:

• RNA sequencing (RNA-seq) is crucial for gene expression analysis and disease identification.
• Discrete RNA-seq data require specialized statistical methods, unlike microarray data.
• Existing methods like PLDA, NBLDA, and ZIPLDA have limitations in modeling RNA-seq data characteristics.

Purpose of the Study:

• To propose a novel classification method, zero-inflated negative binomial logistic discriminant analysis (ZINBLDA), for RNA-seq data.
• To compare ZINBLDA with existing methods (PLDA, NBLDA, ZIPLDA) based on model parameters and performance.
• To provide guidance for selecting optimal classifiers for RNA-seq datasets.

Main Methods:

• Development of ZINBLDA, extending existing distributions to handle excess zeros and overdispersion in count data.
• Comparative analysis of four classification methods (PLDA, NBLDA, ZIPLDA, ZINBLDA) using parameter perspectives.
• Simulation studies and analysis of two real RNA-seq datasets to evaluate classifier performance.
• Creation of a decision tree model to aid in optimal classifier selection.

Main Results:

• ZINBLDA effectively models RNA-seq data, addressing excess zeros and overdispersion.
• The study reveals interrelationships between the four classification methods under specific conditions.
• Simulation and real data analyses demonstrate ZINBLDA's superior or comparable performance.
• A decision tree model was developed to guide the selection of appropriate classifiers.

Conclusions:

• ZINBLDA offers a robust approach for RNA-seq data classification, enhancing disease identification accuracy.
• Understanding model parameters is vital for selecting the best method for specific RNA-seq datasets.
• The developed decision tree provides a practical tool for researchers to choose optimal classification strategies.