Feature selection and dimension reduction for single-cell RNA-Seq based on a multinomial model.

Townes FW, Hicks SC, Aryee MJ, Irizarry RA. · 2019

Read Feature selection and dimension reduction for single-cell RNA-Seq based on a multinomial model. by Townes FW, Hicks SC, Aryee MJ, Irizarry RA. (2019) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/europmc-pmc6927135.txt.

Introduction

Townes and co-authors rethink how single-cell RNA data should be cleaned and reduced. Their study appeared in Genome Biology. Such data record gene activity cell by cell. Many data sets now carry unique molecular identifiers, or UMIs, which tag each original molecule. Standard cleaning steps can invent false variation.

Using negative controls, the team show that UMI counts follow multinomial sampling, with no extra zero inflation. That finding supports simpler, better matched methods. They propose a generalized form of principal component analysis for count data. They also rank genes by deviance to pick informative features. On data sets with known answers, their approach beat common practice at grouping cells correctly.

Catalog record

CREATORTownes FW, Hicks SC, Aryee MJ, Irizarry RA.
YEAR2019
COLLECTIONPAPERS
FORMATTXT
FILE SIZE71 KB
LICENSECC BY 4.0
RIGHTSCC BY
SOURCEEurope PMC

Source: https://europepmc.org/articles/PMC6927135

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

Feature selection and dimension reduction for single-cell RNA-Seq based on a multinomial model By F William Townes, Stephanie C Hicks, Martin J Aryee, Rafael A Irizarry Published in: Genome Biology DOI: 10.1186/s13059-019-1861-6 ABSTRACT Abstract Single-cell RNA-Seq (scRNA-Seq) profiles gene expression of individual cells. Recent scRNA-Seq datasets have incorporated unique molecular identifiers (UMIs). Using negative controls, we show UMI counts follow multinomial sampling with no zero inflation. Current normalization procedures such as log of counts per million and feature selection by highly variable genes produce false variability in dimension reduction. We propose simple multinomial methods, including generalized principal component analysis (GLM-PCA) for non-normal distributions, and feature selection using deviance. These methods outperform the current practice in a downstream clustering assessment using ground truth datasets. Keywords: Gene expression, Single cell, RNA-Seq, Dimension reduction, Variable genes, Principal component analysis, GLM-PCA BACKGROUND Single-cell RNA-Seq (scRNA-Seq) is a powerful tool for profiling gene expression patterns in individual cells, facilitating a variety of analyses such as identification of novel cell types [1, 2]. In a typical protocol, single cells are isolated in liquid droplets, and messenger RNA (mRNA) is captured from each cell, converted to cDNA by reverse transcriptase (RT), then amplified using polymerase chain reaction (PCR) [3–5]. Finally, fragments are sequenced, and expression of a gene in a cell is quantified by the number of sequencing reads that mapped to that gene [6]. A crucial difference between scRNA-Seq and traditional bulk RNA-Seq is the low quantity of mRNA isolated from individual cells, which requires a larger number of PCR cycles to produce enough material for sequencing (bulk RNA-Seq comingles thousands of cells per sample). For example, the popular 10x Genomics protocol uses 14 cycles [5]. Thus, many of the reads counted in scRNA-Seq are duplicates of a single mRNA molecule in the original cell [7]. Full-length protocols such as SMART-Seq2 [8] analyze these read counts directly, and several methods have been developed to facilitate this [9]. However, in many experiments, it is desirable to analyze larger numbers of cells than possible with full-length protocols, and isoform-level inference may be unnecessary. Under such conditions, it is advantagous to include unique molecular identifiers (UMIs) which enable computational removal of PCR duplicates [10, 11], producing UMI counts. Although a zero UMI count is equivalent to a zero read count, nonzero read counts are larger than their corresponding UMI counts. In general, all scRNA-Seq data contain large numbers of zero counts (often >90% of the data). Here, we focus on the analysis of scRNA-Seq data with UMI counts. Starting from raw counts, a scRNA-Seq data analysis typically includes normalization, feature selection, and dimension reduction steps. Normalization seeks to adjust for differences in experimental conditions between samples (individual cells), so that these do not confound true biological differences. For example, the efficiency of mRNA capture and RT is variable between samples (technical variation), causing different cells to have different total UMI counts, even if the number of molecules in the original cells is identical. Feature selection refers to excluding uninformative genes such as those which exhibit no meaningful biological variation across samples. Since scRNA-Seq experiments usually examine cells within a single tissue, only a small fraction of genes are expected to be informative since many genes are biologically variable only across different tissues. Dimension reduction aims to embed each cell’s high-dimensional expression profile into a low-dimensional representation to facilitate visualization and clustering. While a plethora of methods [5, 12–15] have been developed for each of these steps, here, we describe what is considered to be the standard pipeline [15]. First, raw counts are normalized by scaling of sample-specific size factors, followed by log transformation, which attempts to reduce skewness. Next, feature selection involves identifying the top 500–2000 genes by computing either their coefficient of variation (highly variable genes [16, 17]) or average expression level (highly expressed genes) across all cells [15]. Alternatively, highly dropout genes may be retained [18]. Principal component analysis (PCA) [19] is the most popular dimension reduction method (see for example tutorials for Seurat [17] and Cell Ranger [5]). PCA compresses each cell’s 2000-dimensional expression profile into, say, a 10-dimensional vector of principal component coordinates or latent factors. Prior to PCA, data are usually centered and scaled so that each gene has mean 0 and standard deviation 1 (z-score transformation). Finally, a clustering algorithm can be applied to group cells with similar representations in the low-dimensional PCA space. Despite the appealing simplicity of this standard pipeline, the characteristics of scRNA-Seq UMI counts present difficulties at each stage. Many normalization schemes derived from bulk RNA-Seq cannot compute size factors stably in the presence of large numbers of zeros [20]. A numerically stable and popular method is to set the size factor for each cell as the total counts divided by 106 (counts per million, CPM). Note that CPM does not alter zeros, which dominate scRNA-Seq data. Log transformation is not possible for exact zeros, so it is common practice to add a small pseudocount such as 1 to all normalized counts prior to taking the log. The choice of pseudocount is arbitrary and can introduce subtle biases in the transformed data [21]. For a statistical interpretation of the pseudocount, see the “Methods” section. Similarly, the use of highly variable genes for feature selection is somewhat arbitrary since the observed variability will depend on the pseudocount: pseudocounts close to zero arbitrarily increase the variance of genes with zero counts. Finally, PCA implicitly relies on Euclidean geometry, which may not be appropriate for highly sparse, discrete, and skewed data, even after normalizations and transformations [22]. Widely used methods for the analysis of scRNA-Seq lack statistically rigorous justification based on a plausible data generating a mechanism for UMI counts. Instead, it appears many of the techniques have been borrowed from the data analysis pipelines developed for read counts, especially those based on bulk RNA-Seq [23]. For example, models based on the lognormal distribution cannot account for exact zeros, motivating the development of zero-inflated lognormal models for scRNA-Seq read counts [24–27]. Alternatively, ZINB-WAVE uses a zero-inflated negative binomial model for dimension reduction of read counts [28]. However, as shown below, the sampling distribution of UMI counts is not zero inflated [29] and differs markedly from read counts, so application of read count models to UMI counts needs either theoretical or empirical justification. We present a unifying statistical foundation for scRNA-Seq with UMI counts based on the multinomial distribution. The multinomial model adequately describes negative control data, and there is no need to model zero inflation. We show the mechanism by which PCA on log-normalized UMI counts can lead to distorted low-dimensional factors and false discoveries. We identify the source of the frequently observed and undesirable fact that the fraction of zeros reported in each cell drives the first principal component in most experiments [30]. To remove these distortions, we propose the use of GLM-PCA, a generalization of PCA to exponential family likelihoods [31]. GLM-PCA operates on raw counts, avoiding the pitfalls of normalization. We also demonstrate that applying PCA to deviance or Pearson residuals provides a useful and fast approximation to GLM-PCA. We provide a closed-form deviance statistic as a feature selection method. We systematically compare the performance of all combinations of

Read the complete text