Why weight? Modelling sample and observational level variability improves power in RNA-seq analyses.
Liu R, Holik AZ, Su S, Jansz N, Chen K, Leong HS, Blewitt ME, Asselin-Labat ML, Smyth GK, Ritchie ME. · 2015
Read Why weight? Modelling sample and observational level variability improves power in RNA-seq analyses. by Liu R, Holik AZ, Su S, Jansz N, Chen K, Leong HS, Blewitt ME, Asselin-Labat ML, Smyth GK, Ritchie ME. (2015) 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-pmc4551905.txt.
Introduction
Ruijie Liu leads a ten-author team improving statistics for gene expression studies, in Nucleic Acids Research in 2015. RNA sequencing compares gene activity between groups, but small studies often include samples of uneven quality. Dropping noisy samples wastes data and power; keeping them flat can bury real changes in static.
Their approach keeps every sample but weights each observation by its trustworthiness. A variance model estimates how noisy each sample runs across all genes, and those factors merge with per-measurement weights from the count data itself. Trials on simulated and real data showed more power to detect true differences than dropping or flatly keeping poor samples. The method slots into standard expression analysis pipelines.
Catalog record
| CREATOR | Liu R, Holik AZ, Su S, Jansz N, Chen K, Leong HS, Blewitt ME, Asselin-Labat ML, Smyth GK, Ritchie ME. |
|---|---|
| YEAR | 2015 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 64 KB |
| LICENSE | CC BY 4.0 |
| RIGHTS | CC BY |
| SOURCE | Europe PMC |
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
Why weight? Modelling sample and observational level variability improves power in RNA-seq analyses By Ruijie Liu, Aliaksei Z. Holik, Shian Su, Natasha Jansz, Kelan Chen, Huei San Leong, Marnie E. Blewitt, Marie-Liesse Asselin-Labat, Gordon K. Smyth, Matthew E. Ritchie Published in: Nucleic Acids Research DOI: 10.1093/nar/gkv412 ABSTRACT Variations in sample quality are frequently encountered in small RNA-sequencing experiments, and pose a major challenge in a differential expression analysis. Removal of high variation samples reduces noise, but at a cost of reducing power, thus limiting our ability to detect biologically meaningful changes. Similarly, retaining these samples in the analysis may not reveal any statistically significant changes due to the higher noise level. A compromise is to use all available data, but to down-weight the observations from more variable samples. We describe a statistical approach that facilitates this by modelling heterogeneity at both the sample and observational levels as part of the differential expression analysis. At the sample level this is achieved by fitting a log-linear variance model that includes common sample-specific or group-specific parameters that are shared between genes. The estimated sample variance factors are then converted to weights and combined with observational level weights obtained from the mean–variance relationship of the log-counts-per-million using ‘voom’. A comprehensive analysis involving both simulations and experimental RNA-sequencing data demonstrates that this strategy leads to a universally more powerful analysis and fewer false discoveries when compared to conventional approaches. This methodology has wide application and is implemented in the open-source ‘limma’ package. INTRODUCTION Second-generation sequencing technology provides researchers with a high resolution and cost-effective tool for surveying the complexity of the transcriptome in both health and disease. RNA-sequencing (RNA-seq) is now a routine tool for studying differential expression, alternative splicing, allele-specific expression and for discovering novel transcripts (1). In an RNA-seq ‘differential expression’ analysis, the goal is to determine which genes, transcripts or exons, show evidence for changes in expression between experimental groups relative to a realistic assessment of both technical and biological variation. In such analyses, the application of statistical modelling to remove systematic biases and reduce variability via normalization (2,3), batch correction (4,5) and subsequently prioritize changes in gene expression between treatment groups (6,7) have each been shown to play a key role in extracting meaningful insights from RNA-seq data. Variations in sample quality is another source of noise that makes the identification of differentially expressed (DE) genes more difficult. Sample-specific variation is often first observed by inspecting plots from a principal components analysis or multi-dimensional scaling (MDS). Figure 1 shows MDS plots from two RNA-seq experiments where a particular gene (Smchd1) has been mutated. In both experiments, one or more samples exhibit increased variation that is likely to be related to RNA quality (e.g. degraded RNA may be unavoidable in studies involving patient tissue) or sample purity. In our experience, sample quality variation is relatively common in RNA-seq data sets, and once identified, the researcher may choose to re-run the suspect samples or, in extreme cases, collect fresh samples. These options may not always be feasible due to resource constraints, therefore developing analytical methods that can address this variability and make the most of the data available would be advantageous. [Figure: RNA-seq data sets where variations in sample quality are evident. Each panel shows a multi-dimensional scaling (MDS) plot, with samples colour-coded by experimental group. One or more samples that exhibit higher variability than average are present in each case ((A): sample 6; (B): samples 1 and 7). In these experiments, cells carrying a mutant allele of the gene Smchd1 were compared against wild-type cells.] There are currently no methods that can model sample-specific variation in small-scale RNA-seq experiments in a systematic way. One option is to remove samples with higher variability from the analysis. This has the benefit of reducing variation, but comes at the expense of reducing power, which can hamper our ability to detect DE genes. This approach assumes that samples with higher than average variation can be readily identified, which is not always the case. The decision of whether to keep or discard a particular sample is often based on ad-hoc cut-offs and a trial and error approach. At the other extreme, one could retain all samples in the analysis. In this situation, our ability to distinguish genuine differences between experimental conditions from the noise will be limited by the increased variation present. An intermediate approach would be to analyse the complete experiment but down-weight the observations from more variable samples, thus retaining the maximum degrees of freedom whilst discounting noisy observations. Experience with microarray data has shown that this approach can be beneficial, with the use of quantitative quality weights at the sample-level offering improved results (8–10). Allowing for variations in quality between RNA samples is particularly important when samples are difficult to obtain such as when collected from human subjects or very rare cell types (11). RNA sample quality is typically related to the availability of tissue samples or cells and the ease of RNA isolation, a consideration that is relevant irrespective of whether microarray or RNA-seq is used to profile gene expression. The approach of Ritchie et al. (2006) (9) has allowed researchers to make discoveries that would otherwise have been missed, for example in phase I trials of cancer drugs (12). We adapt this method to enable the use of quantitative quality weights in RNA-seq data sets to improve the results obtained in the presence of more variable samples. We model this variation using a combined approach that takes into account both global intensity-dependent trends in the variability using ‘voom’ (13) as well as sample-specific variability using a log-linear model that shares parameters between genes (9). Our new method is compared against a number of alternatives on simulated data, a control experiment where samples were mistreated to simulate increased variability and in a gene mutation study. Across a range of scenarios, we show that our combined observational and sample-specific weighting approach improves power for detecting known changes in gene expression and generates fewer false positives than other methods. This approach is widely applicable, suitable for use in most designed RNA-seq analyses and is available in the popular R-based ‘limma’ package (14). MATERIALS AND METHODS Linear models for RNA-seq differential expression The entry point for our analysis is a matrix of counts that have been appropriately normalized and transformed into log2 counts per million (CPM) so that they are approximately normally distributed. The work of Law et al. (2014) (13) has shown that although the distributional assumption of normality is imperfect, the log2(CPM) transformation when combined with weights that take into account the unequal variabilities on this scale can lead to a more powerful analysis compared to methods that assume RNA-seq counts follow a negative-binomial distribution. Assuming normality allows us to model the variances, which we will exploit later. Variance modelling is not currently available for RNA-seq analysis using count distributions. We begin by assuming a linear model where the systematic expression effects for each gene (or exon or other genomic feature of interest) can be described as(1)\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} }{}\begin{equation*} E({\bf y}_g)=X {\boldsymbol \beta}_g, \end{equation*}\end{document}where yg = (yg1, …, ygJ)T is the vector of log2(CPM) values for