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

Jiadi Zhu,Mingtao Zhao,Wenhui Liao,Yan Zhou,Ziyang Yuan,Lianjie Shu

doi:10.3389/fgene.2021.642227

Abstract

Next-generation sequencing has emerged as an essential technology for the quantitative analysis of gene expression. In medical research, RNA sequencing (RNA-seq) data are commonly used to identify which type of disease a patient has. Because of the discrete nature of RNA-seq data, the existing statistical methods that have been developed for microarray data cannot be directly applied to RNA-seq data. Existing statistical methods usually model RNA-seq data by a discrete distribution, such as the Poisson, the negative binomial, or the mixture distribution with a point mass at zero and a Poisson distribution to further allow for data with an excess of zeros. Consequently, analytic tools corresponding to the above three discrete distributions have been developed: Poisson linear discriminant analysis (PLDA), negative binomial linear discriminant analysis (NBLDA), and zero-inflated Poisson logistic discriminant analysis (ZIPLDA). However, it is unclear what the real distributions would be for these classifications when applied to a new and real dataset. Considering that count datasets are frequently characterized by excess zeros and overdispersion, this paper extends the existing distribution to a mixture distribution with a point mass at zero and a negative binomial distribution and proposes a zero-inflated negative binomial logistic discriminant analysis (ZINBLDA) for classification. More importantly, we compare the above four classification methods from the perspective of model parameters, as an understanding of parameters is necessary for selecting the optimal method for RNA-seq data. Furthermore, we determine that the above four methods could transform into each other in some cases. Using simulation studies, we compare and evaluate the performance of these classification methods in a wide range of settings, and we also present a decision tree model created to help us select the optimal classifier for a new RNA-seq dataset. The results of the two real datasets coincide with the theory and simulation analysis results. The methods used in this work are implemented in the open-scource R scripts, with a source code freely available at https://github.com/FocusPaka/ZINBLDA.

Highlights

RNA sequencing (RNA-seq), which involves directly sequencing complementary DNAs and aligning the sequences to the reference genome or transcriptome, has emerged as a powerful technology for measuring gene expression (Mardis, 2008; Morozova et al, 2009; Wang et al, 2009)
Poisson linear discriminant analysis (PLDA) and zero-inflated Poisson logistic discriminant analysis (ZIPLDA) showed similar performance, and both were slightly worse than negative binomial linear discriminant analysis (NBLDA) and zero-inflated negative binomial logistic discriminant analysis (ZINBLDA) in different dispersion settings
From the expressions of the negative binomial and Poisson distributions, the former reduced to the latter when the dispersion parameter was reduced to zero, which indicates that NBLDA and ZINBLDA are more suitable for classifying overdispersion data

Summary

Introduction

RNA sequencing (RNA-seq), which involves directly sequencing complementary DNAs and aligning the sequences to the reference genome or transcriptome, has emerged as a powerful technology for measuring gene expression (Mardis, 2008; Morozova et al, 2009; Wang et al, 2009). Existing popular methods include edgeR (Robinson and Smyth, 2008; Robinson et al, 2010), DESeq (Love et al, 2014), and LFCseq (Lin et al, 2014). Another important application is the diagnosis of diseases. Because the expression matrix entries are non-negative integers, classification methods that follow a Gaussian distribution may not perform well for RNA-seq data

Methods

Results

Conclusion