Batch effects correction for microbiome data with Dirichlet-multinomial regression

Abstract

Motivation Metagenomic sequencing techniques enable quantitative analyses of the microbiome. However, combining the microbial data from these experiments is challenging due to the variations between experiments. The existing methods for correcting batch effects do not consider the interactions between variables—microbial taxa in microbial studies—and the overdispersion of the microbiome data. Therefore, they are not applicable to microbiome data. Results We develop a new method, Bayesian Dirichlet-multinomial regression meta-analysis (BDMMA), to simultaneously model the batch effects and detect the microbial taxa associated with phenotypes. BDMMA automatically models the dependence among microbial taxa and is robust to the high dimensionality of the microbiome and their association sparsity. Simulation studies and real data analysis show that BDMMA can successfully adjust batch effects and substantially reduce false discoveries in microbial meta-analyses. Availability and implementation An R package” BDMMA” for Windows and Linux is available at https://github.com/DAIZHENWEI/BDMMA/BDMMA, and a version for MacOS is provided at https://github.com/DAIZHENWEI/BDMMA/BDMMA_MacOS. Supplementary information Supplementary data are available at Bioinformatics online.

Population Structure Discovery in Meta-Analyzed Microbial Communities and Inflammatory Bowel Disease

“…ComBat correction 15 , suited for gene expression data, was capable of reducing batch effects to a lesser degree, but also tended to reduce desirable “biological” variation in the process, likely due to noise introduced by it changing many zero counts to non-zero values. Previously proposed techniques for microbial community data, namely quantile normalization 18 and BDMMA 19 , are only appropriate for differential abundance analysis and do not provide batch-normalized profiles, thus precluding PERMANOVA batch effect quantification; their per-feature testing performance is evaluated together with MMUPHin in the following section. MMUPHin thus provides batch-corrected microbial community profiles that retain biologically meaningful variation more than (or not even possible using) existing methods.…”

“…For differential abundance testing, MMUPHin successfully corrected for false associations when batch/cohort effect s were confounded with variables of interest, which is a common concern for ‘omics meta-analysis 38 , while quantile normalization 18 and BDMMA 19 had either inflated or overly conservative false positive rates ( Fig. 2c-d, Supplemental Fig.…”

“…2c, Supplemental Fig. 6 ): a) naive MaAsLin2 model on the study effect confounded null data (without explicitly modelling the batches), b) the quantile normalization procedure, paired with two-tailed Wilcoxon tests, as proposed in 18 , c) BDMMA as proposed in 19 , with the default 1,0000 total MCMC sampling and 5,000 burn-in, d) the complete MMUPHin meta-analysis model for the batch corrected data as described above. Note that due to its computational cost we were only able to evaluate the Dirichlet-multinomial regression model on a subset of parameter combinations, namely number of samples per batch = 100, number of features = 200, and percent of associated microbes = 5%.…”

“…Consequently, methods to correct for cohort and batch effects from other ‘omics settings 14-17 are not directly appropriate. Two recent studies have suggested quantile normalization 18 and Bayesian Dirichlet-multinomial regression (BDMMA) 19 for microbial profiles, which are applicable to a limited subset of differential abundance tests and do not provide batch-corrected profiles. To date, there are no methods permitting the joint analysis of batch-corrected microbial profiles for most study designs.…”

A multivariate method to correct for batch effects in microbiome data

“…Correcting for batch effect offers the flexibility to apply any type of downstream analysis, including dimension reduction, visualisation and clustering. Methods accounting for batch effects are often restricted to differential abundance analysis with models that hold strong assumptions about data distribution, they include zero-inflated Gaussian model (Paulson et al, 2013) and Bayesian Dirichlet multinomial regression (Dai et al, 2018)). In terms of evaluating the effectiveness of batch effect handling methods, batch effect removal methods are more straightforward and transparent than methods that account for batch effects.…”