Differential Abundance Analysis

Modern tools treat sequencing data as compositional, meaning counts are relative and constrained by an arbitrary total sum. This breaks the independence assumption of standard tests.

• Centered Log-Ratio (CLR) Transformation: Used by ALDEx2 to map data from the simplex to real space, making it compatible with standard Gaussian models.
• Additive Log-Ratio (ALR): An alternative transformation that selects a single reference taxon, but results are reference-dependent.
• Median-of-Ratios: Used by DESeq2, which estimates size factors from the median ratio of each feature to the geometric mean across samples.

Metagenomic count tables are sparse, with many zeros arising from both biological absence (structural zeros) and technical undersampling (sampling zeros). Ignoring this leads to inflated false discovery rates.

• Bayesian Multivariate Models: Tools like ANCOM-BC use zero-inflated negative binomial regression to model excess zeros explicitly.
• Imputation Strategies: Some pipelines impute zeros using a small pseudo-count or Bayesian priors before transformation, though this can introduce bias.
• Hurdle Models: Separate the modeling of presence/absence (binary) from abundance given presence (continuous), as implemented in metagenomeSeq.

With thousands of taxa tested simultaneously, multiple testing correction is critical. Tools implement rigorous FDR control to limit the expected proportion of false positives.

• Benjamini-Hochberg Procedure: The standard method for controlling the false discovery rate, applied to raw p-values from per-taxon tests.
• FDR via Empirical Bayes: ALDEx2 uses a Dirichlet-multinomial model to simulate the null distribution and estimate posterior probabilities of differential abundance, providing direct FDR estimates.
• Holm-Bonferroni Correction: A more conservative family-wise error rate control used in ANCOM for pairwise comparisons.

A critical property ensuring that conclusions do not depend on which taxa are included in the analysis. ANCOM and ANCOM-BC are designed to be reference frame invariant.

• ANCOM's Logic: Instead of testing absolute abundance changes, ANCOM tests whether the log-ratio of a taxon to every other taxon changes between groups. A taxon is differentially abundant if a majority of these pairwise log-ratio tests are significant.
• ANCOM-BC's Approach: Introduces a sampling fraction correction term in a linear regression framework, estimating bias due to unequal sampling fractions across samples without requiring a reference taxon.
• Contrast with ALDEx2: ALDEx2's CLR transformation is technically reference-dependent, as it uses the geometric mean of all features as the reference, though its effect sizes are robust in practice.

Statistical significance alone is insufficient; tools must quantify the magnitude and direction of abundance changes to identify biologically meaningful shifts.

• Log2 Fold Change: Standard metric in DESeq2, representing the log-ratio of normalized counts between conditions. Shrinkage estimators (apeglm, ashr) are applied to stabilize estimates for low-abundance features.
• CLR-Based Effect Size: ALDEx2 reports the median difference in CLR-transformed values between groups, which approximates the log-ratio of geometric mean abundances.
• ANCOM-BC's Beta Coefficients: The regression coefficients directly estimate the log-fold change in absolute abundance, corrected for sampling fraction bias.

Clinical studies often collect samples from the same subjects over time. Tools now support mixed-effects models to account for within-subject correlation.

• ANCOM-BC2: Extends the original model to handle repeated measures and mixed-effects designs with fixed and random effects, including subject-specific random intercepts.
• MaAsLin2: A multivariate association framework that fits per-feature linear mixed models, supporting continuous and categorical metadata with random effects for repeated sampling.
• Time Series Analysis: Tools like MetaLonDA model temporal trends using spline-based generalized linear mixed models to identify features with differential temporal abundance patterns.

Metagenomic count data is inherently compositional—it carries only relative information because the total read count is an artifact of the sequencing instrument, not the biological sample. A change in one taxon's abundance forces a change in the relative abundance of all others, creating spurious negative correlations. Standard statistical tests like t-tests or Wilcoxon rank-sum fail here because they assume independence and operate in Euclidean space. The solution is to apply a log-ratio transformation—such as the centered log-ratio (CLR) or additive log-ratio (ALR)—which maps the data from the constrained simplex to unconstrained real space where standard methods become valid.

When testing thousands of microbial taxa simultaneously for differential abundance, the probability of false positives explodes. The Benjamini-Hochberg procedure controls the expected proportion of Type I errors among all rejected null hypotheses—the FDR—rather than controlling the family-wise error rate, which is overly conservative. A typical threshold is an adjusted p-value < 0.05, meaning that among all taxa declared significant, fewer than 5% are expected to be false discoveries. This is the standard for high-dimensional genomic hypothesis testing.

Statistical significance alone is insufficient; a taxon can have a tiny, biologically irrelevant change with a very low p-value due to low variance. Effect size quantifies the magnitude of the difference. In compositional analysis, this is often expressed as the median log2 fold change between groups. A log2 fold change of 1 indicates a doubling of the CLR-transformed abundance. Tools like ALDEx2 produce an effect plot that visualizes the relationship between the expected effect size and the statistical significance, allowing researchers to prioritize taxa that are both statistically and biologically meaningful.

Raw read counts cannot be compared directly between samples due to varying library sizes. Common normalization methods include:

• TSS (Total Sum Scaling): Divides each count by the total library size. Simple but fails to address compositionality.
• CSS (Cumulative Sum Scaling): A median-like quantile normalization implemented in metagenomeSeq that is robust to outliers.
• TMM (Trimmed Mean of M-values): Weights taxa based on their log fold changes, excluding those with extreme values.
• Rarefaction: Randomly subsampling reads to a uniform depth. Statistically inefficient and discards valid data; generally discouraged in favor of model-based normalization.