OmicVerse Microbiome — Differential-Abundance Method Comparison

Goal

Take a preprocessed microbiome AnnData (output of the 16S amplicon skill — samples × ASVs with 7-rank SINTAX taxonomy in var) and run all three DA methods exposed by ov.micro.DA on the same two-group contrast — Wilcoxon (rank), pyDESeq2 (NB-GLM), ANCOM-BC (compositional). Compare their hit sets at a common FDR cutoff, report consensus / method-specific genera, and surface the biology around the three methods' different statistical assumptions so the user can pick (or report) the right one.

Quick Workflow

1. Load the AnnData produced by omicverse-microbiome-16s-amplicon-dada2. Drop control / non-relevant groups; keep exactly two for the contrast.
2. Collapse to a chosen taxonomic rank (typically genus): ov.micro.collapse_taxa(adata, rank='genus'). DA at species/ASV level is noisy on small cohorts; genus is the canonical reporting rank for 16S.
3. Wilcoxon: ov.micro.DA(adata_genus).wilcoxon(group_key, group_a, group_b, min_prevalence=0.1). Non-parametric; fastest; tests ranks of relative abundance.
4. pyDESeq2: ov.micro.DA(adata_genus).deseq2(group_key, group_a, group_b, min_prevalence=0.1). NB-GLM on raw counts; uses RNA-seq-style shrinkage of small-count fold-changes.
5. ANCOM-BC: ov.micro.DA(adata_genus).ancombc(group_key, min_prevalence=0.1, pseudocount=1.0). Compositional bias-corrected ANCOM; closest to the compositional ground truth.
6. Build sets of significant features at a common FDR cutoff (typically 0.05). Watch for column-name differences across methods (fdr_bh for Wilcoxon / DESeq2; q_value or fdr_bh for ANCOM-BC).
7. Tabulate the 3-way Venn (Wilcoxon-only / DESeq2-only / ANCOM-BC-only / pairwise overlaps / all-three) and render with matplotlib_venn.venn3 if installed.
8. Report two numbers in the writeup: (a) the consensus hit set (intersection across all three) for the strongest claim, and (b) the Wilcoxon ∪ ANCOM-BC set as a slightly more permissive convention if compositional correctness matters.

Interface Summary