PT  - JOURNAL ARTICLE
AU  - Tanbin Rahman
AU  - Hsin-En Huang
AU  - An-Shun Tai
AU  - Wen-Ping Hsieh
AU  - George Tseng
TI  - A sparse negative binomial classifier with covariate adjustment for RNA-seq data
AID  - 10.1101/636340
DP  - 2019 Jan 01
TA  - bioRxiv
PG  - 636340
4099  - http://biorxiv.org/content/early/2019/05/15/636340.short
4100  - http://biorxiv.org/content/early/2019/05/15/636340.full
AB  - Supervised machine learning methods have been increasingly used in biomedical research and in clinical practice. In transcriptomic applications, RNA-seq data have become dominating and have gradually replaced traditional microarray due to its reduced background noise and increased digital precision. Most existing machine learning methods are, however, designed for continuous intensities of microarray and are not suitable for RNA-seq count data. In this paper, we develop a negative binomial model via generalized linear model framework with double regularization for gene and covariate sparsity to accommodate three key elements: adequate modeling of count data with overdispersion, gene selection and adjustment for covariate effect. The proposed method is evaluated in simulations and two real applications using cervical tumor miRNA-seq data and schizophrenia post-mortem brain tissue RNA-seq data to demonstrate its superior performance in prediction accuracy and feature selection.In the past two decades, microarray and RNA sequencing (RNA-seq) are routine procedures to study transcriptome of organisms in modern biomedical studies. In recent years, RNA-seq [5, 20] has become a popular experimental approach for generating a comprehensive catalog of protein-coding genes and non-coding RNAs [13], and it largely replaces the microarray technology due to its low background noise and increased precision. The most important difference between RNA-seq and microarray technology is that RNA-seq outputs millions of sequencing reads rather than the continuous fluorescent intensities in microarray data. Unlike microarray, RNA-seq can detect novel transcripts, gene fusions, single nucleotide variants, and indels (insertion/deletion). It can also detect a higher percentage of differentially expressed genes than microarray, especially for genes with low expression [24].In machine learning, classification methods are used to construct a prediction model based on a training dataset with known class label so future independent samples can be classified with high accuracy. For example, labels in clinical research can be case/control, disease subtypes, drug response or prognostic outcome. Many popular machine learning methods have been widely applied to microarray studies, such as linear discriminate analysis [9], support vector machines [3] and random forest [7]. However, for discrete data nature in RNA-seq, many powerful tools for microarray assuming continuous data input or Gaussian assumption may be inappropriate. A common practice to solve this problem is to transform RNA-seq data into continuous values such as FPKM or TPM [6] and possibly taking additional log-transformation. However, such data transformation can lead to loss of information from the original data [14, 18], producing less accurate inference. Particularly, the transformation often produces greater loss of information for genes with lower counts [15]. To accommodate discrete data in RNA-Seq, Poisson distribution and negative binomial distribution are two common distributions expected to better fit the data generation process and data characteristics. Witten [22] proposed a sparse Poisson linear discriminant analysis (sPLDA) based on Poisson assumption for the count data. However, Poisson distribution assumes equal mean and variance, which is often not true. In real RNA-seq data, the variance is often larger than the mean, leading to the need of an overdispersion parameter. Witten [22] reconciled this problem by proposing a power transformation to the data for eliminating overdispersion. However, as we will see later, the power transformation can perform well when the overdispersion is small but performs poorly when overdispersion becomes large. Hence, direct modeling by negative binomial assumption rather than a Poisson distribution is more appropriate. To this end, Dong et al. [8] proposed negative binomial linear discriminant analysis (denoted as NBLDAPE) by adding a dispersion parameter. They, however, borrowed the point estimation from sPLDA in [22] and did not pursue a principled inference such as maximum likelihood, consequently producing worse performance than the method we will propose later.Since the number of genes is often much larger than the number of samples in transcriptomic studies (a standard “small-n-large-p” problem), feature selection is critical to achieve better prediction accuracy and model interpretation. Witten [22] proposed a somewhat ad hoc soft-thresholding operator, similar to univariate Lasso estimator in regression, for gene selection in sPLDA but the method is not applicable to the NBLDAPE model due to the addition of dispersion parameter. In the NBLDAPE model proposed by [8], feature selection issue was not discussed, except that they used “edgeR” package to reduce the number of genes in the input data. Such a two-step filtering method is well-known to have inferior performance than methods with embedded feature selection. In fact, Zararsiz et al. [23] have compared sPLDA and NBLDAPE, and showed that the power transformed sPLDA generally performed better than NBLDAPE in their simulations and the worse performance in NBLDAPE mainly came from the lack of feature selection. Finally, another critical factor to consider in transcriptomic modeling is the adjustment of covariates such as gender, race and age since it is well-known that many genes are systematically impacted by these factors. For example, Peters et al. [16] have identified 1,497 genes that are differentially expressed with age in a whole-blood gene expression meta-analysis of 14,983 individuals. A classification model allowing for covariate adjustment is expected to provide better accuracy and deeper biological insight.To account for all aforementioned factors, we propose a sparse negative binomial model (snbClass) for classification analysis with covariate selection and adjustment. The method is based on generalized linear model (GLM) with a first regularization for feature sparsity. The GLM framework also allows straightforward covariate adjustment and a second regularization term on covariates, facilitating further covariate selection. Such covariate adjustment is not possible through existing sPLDA or NBLDAPE methods. The paper is structured as following. In Section 1.1, we will briefly describe the two existing methods sPLDA [22] and NBLDAPE [8], and then followed by our proposed methods sNBLDAGLM and sNBLDAGLM.sC in Section 1.2. Section 1.3 and 1.4 will discuss parameter estimation and model selection of the proposed method. Benchmarks for evaluation are described in Section 1.5. Section 2 presents simulation studies and Section 3 shows two real applications of cervical tumor miRNA data and schizophrenia RNA-seq data. Conclusion and discussion are included in Section 4.