LinDA: linear models for differential abundance analysis of microbiome compositional data.

Zhou H, He K, Chen J, Zhang X. · 2022

Read LinDA: linear models for differential abundance analysis of microbiome compositional data. by Zhou H, He K, Chen J, Zhang X. (2022) 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-pmc9012043.txt.

Introduction

Zhou and co-authors present LinDA, a method for finding which microbes differ in abundance between sample groups. Microbiome counts are compositional, so standard tests produce too many false positives. LinDA fits plain linear models to transformed data and then corrects the bias that composition creates. The method keeps the false discovery rate under control and extends to mixed models for linked samples.

In tests on simulated and real data, LinDA stayed accurate where rival methods slipped. Because it rests on ordinary regression, it scales to large studies and is simple to run. The authors offer it as a flexible default for a core task in microbiome analysis.

Catalog record

CREATORZhou H, He K, Chen J, Zhang X.
YEAR2022
COLLECTIONPAPERS
FORMATTXT
FILE SIZE102 KB
LICENSECC BY 4.0
RIGHTSCC BY
SOURCEEurope PMC

Source: https://europepmc.org/articles/PMC9012043

Browse the full library · About the Tower of Babel

Full text

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

LinDA: linear models for differential abundance analysis of microbiome compositional data By Huijuan Zhou, Kejun He, Jun Chen, Xianyang Zhang Published in: Genome Biology DOI: 10.1186/s13059-022-02655-5 ABSTRACT Differential abundance analysis is at the core of statistical analysis of microbiome data. The compositional nature of microbiome sequencing data makes false positive control challenging. Here, we show that the compositional effects can be addressed by a simple, yet highly flexible and scalable, approach. The proposed method, LinDA, only requires fitting linear regression models on the centered log-ratio transformed data, and correcting the bias due to compositional effects. We show that LinDA enjoys asymptotic FDR control and can be extended to mixed-effect models for correlated microbiome data. Using simulations and real examples, we demonstrate the effectiveness of LinDA. Supplementary Information The online version contains supplementary material available at (10.1186/s13059-022-02655-5). BACKGROUND The role of the human microbiome in health and disease has been intensively studied over the past few years, see, e.g., [1, 2], for several reviews. Potentially pathogenic or probiotic microorganisms can be identified by analyzing their abundances in a microbial ecosystem (e.g., the human gut) with respect to some covariate of interest such as disease status. Current prevailing technologies for studying the human microbiome use metagenomic sequencing, where either the DNA of a taxonomically informative gene (e.g., 16S rRNA) or all the genomic DNA in the microbial genome is sequenced. After obtaining the raw sequencing reads, the reads can be clustered into operational taxonomic units (OTUs), denoised into amplicon sequence variants (ASVs), or mapped to a microbial reference database (taxa) using existing bioinformatics pipelines such as UPARSE, DADA2, and MetaPhlAn [3–5]. For simplicity, we use the term taxon (pl. taxa) to represent any taxonomic unit (OTU/ASV/taxon) from a bioinformatics pipeline. Therefore, after bioinformatics processing, we have an abundance table recording the frequencies of detected taxa in the samples, together with a meta data table capturing the sample-level information. Differential abundance analysis is then carried out based on the abundance and meta data table. Ideally, we want to measure the absolute abundance of the microorganisms, i.e., the number of microorganisms per unit area/volume at the microbial ecosystem, and differential abundance analysis is performed on the absolute abundance data. However, the data from a sequencing experiment only captures the relative abundance (compositional) information since the total sequence read count, also known as sequencing depth or library size, does not reflect the total microbial load in the specimen due to the complex chemistry involved in sequencing [6, 7]. Although there are several experimental techniques such as qPCR, spike-in and flow cytometry that can be used to achieve absolute abundance measurement, the severe limitations of these techniques prevent their wide adoption [8]. Therefore, the prevailing sequencing protocol is still only able to measure the relative abundances. Drawing inferences about the changes on the unknown absolute abundance based on the measured relative abundance data is challenging due to missing the total microbial load information. The increase or decrease in the abundance of some taxa with respect to a covariate of interest automatically results in changes in the relative abundances of all other taxa, a statistical phenomenon known as compositional effects. Therefore, using the standard statistical techniques such as two-sample t-test, Wilcoxon rank sum test, and linear regression analysis ignoring the compositional nature of the data could lead to a large number of false discoveries. For the differential abundance problem to be well defined, one has to make assumptions. One major assumption is that the differential signal is sparse, i.e., only a small proportion of taxa are associated with the covariate of interest. Although many studies have supported the sparse signal assumption, there are also studies support dense signal hypotheses, where a large number of taxa are differential with small effect sizes [9, 10]. Therefore, the validity of a method and the definition of true or false positive depends on the specific assumption one is willing to accept. Here our goal is to provide a statistical tool that could be potentially useful for pinpointing top candidate taxa for further biological validation. To address compositional effects in differential analysis, one popular approach is robust normalization. It involves calculating a normalizing factor (scale factor), which is robust to a small number of differential taxa and could well capture the sequencing effort for the non-differential part. Therefore, dividing by such a normalizing factor will bring the abundance of the non-differential taxa to the same scale while retaining the differences for those differential ones. Assuming the number of differential taxa is small, different strategies have been used to calculate a robust normalizing factor including TMM, RLE, CSS, and GMPR [11–14]. We list these methods in Additional file 1: Table S1. In contrast, the naive total sum scaling (TSS) normalization, which divides the counts by the library size, is not a robust normalization method [14]. These normalization techniques can be combined with different statistical procedures in differential abundance analysis. For example, we can divide the counts by the normalizing factor from the normalization techniques in Additional file 1: Table S1 and then apply standard statistical tools based on the normalized data. The normalizing factor could also be included as an offset in regression models such as edgeR [15], DESeq2 [16], MicrobiomeDDA [17], and metagenomeSeq [13], where the TMM, RLE, GMPR, and CSS normalization are the accompanying normalization methods. The recently developed MaAsLin2 [18] uses log linear models on the normalized abundance data. Different normalization approaches including TSS, TMM, CSS, and CLR are options in MaAsLin2. A variant to the robust normalization approach is to find a reference taxon or a set of reference taxa, which are assumed to be non-differential with respect to the covariate of interest. The data are then normalized by the count of the reference taxon (or the sum of the counts of the reference taxa). This strategy was used in RAIDA [19] and DACOMP [20]. Another line of methods to tackle the compositional effect uses (log) ratio approach since only ratios are well defined for compositional data [21]. The ALDEx2 method by [22] uses the centered log-ratio (CLR) transformation, where the counts of a sample are divided by their geometric mean before taking logarithms. Differential abundance analysis is then performed using Wilcoxon rank sum test or t-test based on the CLR transformed data. In the CLR approach, the geometric mean can also be regarded as a robust normalizing factor. The ANCOM proposed by [23] computes the pairwise ratios of the relative abundances and identifies the taxa with the most differential ratios. This is based on the observation that the abundance ratios for those differential taxa to other taxa are all differential assuming distinct effect sizes while the ratios for those non-differential taxa are mostly non-differential. Therefore, by analyzing the pattern of the pairwise ratios, one could distinguish the differential taxa from a background of non-differential taxa with high accuracy. Recently, a bias-corrected version of ANCOM (called ANCOM-BC) has been proposed [24], which uses a linear regression framework based on log-transformed taxa count and estimates the unknown bias term due to the compositional effect through an EM algorithm. In the work of [25, 26], the authors evaluated several popular methods in differential abundance analysis (ANCOM-BC/MaAsLin2 not included) and showed that

Read the complete text