Run all three differential-abundance methods (Wilcoxon, pyDESeq2, ANCOM-BC) on the same microbiome AnnData, compare their hit sets via 3-way Venn / overlap counts, and decide which to trust on a given cohort. Use when the user wants to benchmark DA methods on a 16S study, when picking between methods on a small or zero-inflated cohort, or when reporting consensus features that survive multiple tests.
Instrucciones de origen · Vista previa de solo lectura
name
omicverse-microbiome-da-comparison
description
Run all three differential-abundance methods (Wilcoxon, pyDESeq2, ANCOM-BC) on the same microbiome AnnData, compare their hit sets via 3-way Venn / overlap counts, and decide which to trust on a given cohort. Use when the user wants to benchmark DA methods on a 16S study, when picking between methods on a small or zero-inflated cohort, or when reporting consensus features that survive multiple tests.
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
Load the AnnData produced by omicverse-microbiome-16s-amplicon-dada2. Drop control / non-relevant groups; keep exactly two for the contrast.
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.
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.
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.
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).
Tabulate the 3-way Venn (Wilcoxon-only / DESeq2-only / ANCOM-BC-only / pairwise overlaps / all-three) and render with matplotlib_venn.venn3 if installed.
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
Same ov.micro.DA class, three different methods, all returning a pd.DataFrame indexed by the feature column with method-specific columns:
Running all three DA methods on the same AnnData and the same two-group contrast.
Set-arithmetic comparison of hit sets at a common FDR.
Method-choice guidance based on cohort size / zero inflation / compositional concerns.
Documenting consensus and method-specific features.
Outside scope:
Building / preprocessing the input AnnData — see omicverse-microbiome-16s-amplicon-dada2.
Cross-cohort meta-analysis (combining DA results from multiple studies) — see omicverse-microbiome-meta-analysis.
Three-or-more-group DA — ov.micro.DA is two-group; use Kruskal-Wallis externally.
Repeated-measures / paired DA — out of scope for ov.micro (use mixed models in ov.metabol or external nlme etc.).
Phylogenetically-aware DA (e.g. PERMANOVA on UniFrac distances) — see phylogeny skill.
Branch Selection
Wilcoxon (default for moderate cohorts)
Pros: non-parametric (no NB / log-normal assumption); fast (sub-second on a 22-sample × 250-genus cohort); robust to outliers.
Cons: weak power on small cohorts (n<10/group); doesn't model compositional bias; relative-abundance scaling makes interpretation of log2fc less rigorous than DESeq2.
Use when: n>=10/group, low-to-moderate sparsity, no strong compositional concern, you need fast / portable.
pyDESeq2 (RNA-seq-style NB-GLM)
Pros: well-calibrated p-values on raw counts; LFC shrinkage stabilises small-count fold-changes; mature implementation with extensive vignettes.
Cons: NB assumption fails on highly zero-inflated data (microbiome can be 80%+ zeros); doesn't correct for compositional bias; slower than Wilcoxon.
Use when: n is moderate, sparsity is moderate, you want NB-style shrinkage, reviewers expect DESeq2-style methodology.
ANCOM-BC (compositional bias-corrected)
Pros: explicit compositional correction (the only method here that doesn't ignore the simplex); recovers true effect direction on highly biased data; bias-corrected per-feature log-ratio model is closest to the ground-truth.
Cons: slowest of the three; requires skbio>=0.7.1; can be conservative on small cohorts; the bias-correction adds a per-feature constant which shifts log2fc relative to DESeq2.
Use when: compositional correctness matters (always for true microbiome relative-abundance reporting), n is moderate-to-large, downstream analysis uses log-ratio interpretation.
Consensus reporting strategy
For the strongest claim: intersection across all three (rare but unambiguous; survives method assumptions).
For a defensible claim: Wilcoxon ∩ ANCOM-BC (combines a model-free test with the compositional-aware test; bypasses DESeq2's NB assumption).
For exploratory hits: union; flag method-specific hits explicitly so the reader knows the assumption that drove them.
Always report cohort size + sparsity + chosen FDR — DA results without those numbers are uninterpretable.
Column-name pitfalls
Wilcoxon / pyDESeq2 use fdr_bh; ANCOM-BC may use q_value or fdr_bh depending on skbio version. Pattern in the tutorial: sig_col = 'q_value' if 'q_value' in ab.columns else 'fdr_bh'.
All three methods may also expose a raw pvalue column. Don't confuse pvalue and FDR when filtering — pvalue<0.05 is not FDR<0.05.
Input Contract
An AnnData from the 16S amplicon skill; obs[group_key] is a categorical with at least two values; the two-group slice should have n>=5/group for any of these tests to behave reasonably.
adata.X is integer counts (DESeq2 requires raw counts; Wilcoxon and ANCOM-BC handle either, but raw counts are the canonical input).
For ANCOM-BC: pip install skbio>=0.7.1 (function raises ImportError if missing).
For pyDESeq2: pip install pydeseq2.
Minimal Execution Patterns
import omicverse as ov
import anndata as ad
import matplotlib.pyplot as plt
ov.plot_set()
# 1) Load AnnData from the 16S amplicon skill, restrict to two groups.
adata = ad.read_h5ad('mothur_sop_16s.h5ad')
adata = adata[adata.obs['group'].isin(['Early', 'Late'])].copy()
print(adata.obs['group'].value_counts().to_dict())
# 2) Collapse to genus.
adata_genus = ov.micro.collapse_taxa(adata, rank='genus')
print('samples × genera:', adata_genus.shape)
# 3) Run all three DA methods on the same contrast.
wx = ov.micro.DA(adata_genus).wilcoxon(
group_key='group', group_a='Early', group_b='Late', min_prevalence=0.1,
)
ds = ov.micro.DA(adata_genus).deseq2(
group_key='group', group_a='Early', group_b='Late', min_prevalence=0.1,
)
ab = ov.micro.DA(adata_genus).ancombc(
group_key='group', min_prevalence=0.1,
)
print(f' Wilcoxon : {(wx["fdr_bh"] < 0.05).sum():3d} / {len(wx):3d} genera at FDR 0.05')
print(f' DESeq2 : {(ds["fdr_bh"] < 0.05).sum():3d} / {len(ds):3d}')
sig_col_ab = 'q_value'if'q_value'in ab.columns else'fdr_bh'print(f' ANCOM-BC : {(ab[sig_col_ab] < 0.05).sum():3d} / {len(ab):3d}')
# 4) Build sets and tabulate the 3-way Venn.
sig_wx = set(wx.loc[wx['fdr_bh'] < 0.05, 'feature'])
sig_ds = set(ds.loc[ds['fdr_bh'] < 0.05, 'feature'])
sig_ab = set(ab.loc[ab[sig_col_ab] < 0.05, 'feature'])
print('Wilcoxon only :', len(sig_wx - sig_ds - sig_ab))
print('DESeq2 only :', len(sig_ds - sig_wx - sig_ab))
print('ANCOM-BC only :', len(sig_ab - sig_wx - sig_ds))
print('All three :', len(sig_wx & sig_ds & sig_ab))
# 5) Render Venn (optional dependency)try:
from matplotlib_venn import venn3
fig, ax = plt.subplots(figsize=(5, 5))
venn3([sig_wx, sig_ds, sig_ab],
set_labels=('Wilcoxon', 'DESeq2', 'ANCOM-BC'), ax=ax)
ax.set_title('Genera significant at FDR/q < 0.05')
plt.show()
except ImportError:
pass
Validation
For each method: report n_tested (pre-min_prevalence filter), n_significant at FDR/q < 0.05, and the percentage. With a 22-sample mothur SOP demo, expect ~5–25 genus-level hits per method.
Hit-set agreement: at least the top-3 by absolute log2fc per method usually agree; if Wilcoxon and ANCOM-BC have zero overlap, you have a problem (cohort too small, or the contrast doesn't actually exist in the data).
pyDESeq2 hits that don't appear in Wilcoxon are often spurious driven by NB shrinkage on rare features; check the raw counts for those genera and consider tighter min_prevalence.
ANCOM-BC's log2fc is bias-corrected — it may differ in magnitude from DESeq2's by a small constant offset (per the compositional correction), but the sign must agree on consensus hits. Sign disagreement on a consensus genus is a red flag.
Reporting: always disclose cohort size, the FDR cutoff, and which min_prevalence was used. A "significant" feature at min_prevalence=0.0 on n=5/group is uninterpretable.
