Two precision-boosting features for prevalence modeling are a data-augmentation step to reduce linear separability and comparing prevalence coefficients to abundance coefficients to flag abundance-driven prevalence effects.

The method handles presence/absence with logistic regression and nonzero abundance with log-linear models, using a median-based comparison to address compositionality, effectively functioning as a hurdle model.

MaAsLin 3 resources—including software, documentation, tutorials, and datasets—are freely available at huttenhower.sph.harvard.edu/maaslin3/.

Spike-in experiments show MaAsLin 3 can further boost precision when incorporating experimental absolute-abundance information, though median-based testing remains strong with large samples.

Across simulations and real data, MaAsLin 3 shows robustness to assumption violations (extreme sparsity, structural zeros) with high precision and competitive recall, and accurate effect-size correlations with true values.

MaAsLin 3 refines differential abundance testing by separating prevalence from abundance and by addressing compositionality and high dimensionality in meta-omic data.

The workflow normalizes to relative abundances, builds a parallel prevalence profile, log-transforms nonzero abundances, runs logistic regression on prevalence and linear regression on abundance, and combines results into an overall feature–covariate effect.

Separating prevalence from abundance reveals distinct signals, such as a feature more likely present with a covariate but at lower abundance when present, illustrated by Eubacterium rectale in the IBDMDB cohort.

Analyses of real datasets with absolute abundance quantification show that relative-abundance coefficients can diverge from absolute-abundance coefficients, especially under high sparsity, underscoring MaAsLin 3’s options for absolute-abundance or median-based testing.

MaAsLin 3 analyzes data on both relative and absolute abundance scales, including a method to infer absolute-scale associations by testing whether a feature’s coefficient differs from the median of all features’ coefficients.

In longitudinal and repeated-measures settings, MaAsLin 3 preserves precision and recall, and applying post-hoc coefficient thresholds further stabilizes precision.

MaAsLin 3 outperforms several differential abundance methods (ALDEx2, ANCOM-BC2, MaAsLin 2) in simulations, especially as sample sizes exceed 50–100, with robust performance across varying zero-inflation.