Batch effects removal for microbiome data via conditional quantile regression

Abstract

Batch effects in microbiome data arise from differential processing of specimens and can lead to spurious findings and obscure true signals. Strategies designed for genomic data to mitigate batch effects usually fail to address the zero-inflated and over-dispersed microbiome data. Most strategies tailored for microbiome data are restricted to association testing or specialized study designs, failing to allow other analytic goals or general designs. Here, we develop the Conditional Quantile Regression (ConQuR) approach to remove microbiome batch effects using a two-part quantile regression model. ConQuR is a comprehensive method that accommodates the complex distributions of microbial read counts by non-parametric modeling, and it generates batchremoved zero-inflated read counts that can be used in and benefit usual subsequent analyses. We apply ConQuR to simulated and real microbiome datasets and demonstrate its advantages in removing batch effects while preserving the signals of interest.Advances in 16S rRNA 1 and full metagenome 2 sequencing technologies have enabled large-scale human microbiome profiling studies involving hundreds to thousands of individuals. The large sample sizes of these studies and the rich availability of metadata promise a comprehensive understanding of the role of microorganisms in health and disease. These studies have already revealed associations between bacterial taxa and both diseases and exposures, such as obesity 3 , type 2 diabetes 4 , bacterial vaginosis 5 , antibiotics 6 , and environmental pollutants 7 . However, although large-scale studies facilitate more robust and powerful analyses, they are often subject to serious batch effects-systematic variation in the data originating from differential handling and processing of specimens 8 . Many large studies include samples collected across times or locations and processed in different runs. In a more extreme situation, several studies may be pooled together for integrative analysis, with inter-study heterogeneity introducing even more severe variation. These batch effects pose serious challenges to analysis and can lead to excessive false positive discoveries, obscure true associations between microbes and clinical variables, and hinder prediction modeling and biomarker development. Unfortunately, despite the importance of batch effects,

Critical review of 16S rRNA gene sequencing workflow in microbiome studies: From primer selection to advanced data analysis

“…As an alternative, non‐parametric multivariate approaches can be used, or more recent methods explicitly developed for considering the specificities of microbiome data (Goh et al., 2017; Wang & Lê Cao, 2020). The latter include, from oldest to most recent percentile normalisation (Gibbons et al., 2018), the Bayesian Dirichlet–multinomial regression meta‐analysis (BDMMA) (Dai et al., 2019), the conditional quantile regression (ConQuR) (Ling et al., 2022), the ‘adjust_batch’ tool from the MMUPHin package (Ma, 2022) and three approaches based on the PLS‐DA from the PLSDAbatch package (Wang & Lê Cao, 2023). The latter three, named PLSDA‐batch, sparse PLSDA‐batch (sPLSDA‐batch), and weighted PLSDA‐batch (wPLSDA‐batch), require prior abundance filtering and CLR transformation.…”

“…Moreover, both approaches are only appropriate for a limited subset of differential abundance tests and do not provide batch‐normalised profiles (Ma et al, 2022). ConQuR requires comprehensive metadata to accurately estimate conditional distributions of read counts, which can lead to over‐optimism in association analysis and cannot work if the batch completely confounds the critical variable (Ling et al., 2022). In addition, MMUPHin (Ma, 2022) assumes the data to be zero‐inflated Gaussian, which is only suitable for certain transformations of relative abundance data (Ling et al., 2022), and the PLS‐DA‐based methods require pre‐defined batch group information, so, if unknown, it should be identified with PCA or any clustering approach (Wang & Lê Cao, 2023).…”

“…ConQuR requires comprehensive metadata to accurately estimate conditional distributions of read counts, which can lead to over‐optimism in association analysis and cannot work if the batch completely confounds the critical variable (Ling et al., 2022). In addition, MMUPHin (Ma, 2022) assumes the data to be zero‐inflated Gaussian, which is only suitable for certain transformations of relative abundance data (Ling et al., 2022), and the PLS‐DA‐based methods require pre‐defined batch group information, so, if unknown, it should be identified with PCA or any clustering approach (Wang & Lê Cao, 2023). Lastly, all the methods suffer from the presence of too many small batches and low numbers of sequences/library sizes, and none work very well for low‐frequency taxa (Ling et al., 2022).…”

Compositional variable selection in quantile regression for microbiome data with false discovery rate control

“…Besides aforementioned considerations in statistical power and model interpretability, another potential issue is that most existing compositional variable selection methods are based on mean regression models [10,12,15,16], which often fail to accommodate data heterogeneity, probably due to either heteroscedastic variance or other forms of non-location-scale covariate effects [17]. A common solution to this issue would be the quantile regression [18], which has been widely used in biomedical data analysis [19][20][21][22]. In the context of compositional data analysis, Ma and Zhang [23] proposed an adaptive Lasso-based method for compositional quantile regression, whose finite-sample selection accuracy has not yet been investigated.…”

The salivary microbiome as a diagnostic biomarker of health and periodontitis: a large-scale meta-omics analysis before and after the removal of batch effects

“…BEs include any sources of unwanted biological, technical or computational variations that are unrelated to, but obscure, the biological element of interest (18). Although microbiomespeci c methods have been developed to remove BEs (19)(20)(21)(22)(23), their use is not yet widespread and, to the best of our knowledge, they have not been employed in any 16S rRNA gene sequencing research on salivary microbiota to date.…”