Refining and extending generalized multivariate linear models for meta-omic discovery with MaAsLin 3

A common and important step in analyzing microbiome data is differential abundance testing, determining how the abundances of taxa change with respect to a community phenotype or environment. Differential abundance testing is complicated by the fact that microbiome data are usually compositional, sparse, right-skewed, and high-dimensional. Existing methods for differential abundance testing fail to model all of these properties of the data, including how both the abundance of a taxon and its prevalence—its probability of being present or absent in a sample—change in response to sample parameters. To bridge this gap, we have developed MaAsLin 3 (Microbiome Multivariable Associations with Linear Models) to simultaneously identify both prevalence and abundance associations in a biologically motivated and statistically principled manner. In addition to detecting prevalence differences, MaAsLin 3 enables more robust inference for abundance associations by accounting for compositionality with reference spike-ins or an iterative renormalization procedure. MaAsLin 3 also expands inferential abilities beyond traditional linear models by allowing users to test for microbial differences associated with ordered monotonic predictors and differences among more than two categorical groups. Across a variety of simulations, MaAsLin 3’s set of methods is more robust to the statistical properties of microbiome data than current state-of-the-art differential abundance methods. Additionally, when applied to a large dataset of stool samples from an inflammatory bowel disease cohort, MaAsLin 3 indicates that the vast majority of previous abundance associations are actually prevalence associations. In summary, MaAsLin 3 enables researchers to identify more specific and more accurate microbiome associations, especially in large and complex datasets.