Differential Abundance Testing - Usage Guide

Overview

Differential abundance testing asks whether cell-type proportions or composition changed between conditions (treatment vs control, disease vs healthy). Because proportions live on a simplex (they sum to 1 and are not independent), naive per-cluster proportion tests are invalid. This skill covers cluster-free neighborhood testing (Milo) and cluster-based compositional models (scCODA, sccomp, propeller), and how to keep compositional shifts from being misread as differential expression.

Prerequisites

# Milo, sccomp, propeller (R)
BiocManager::install(c('miloR', 'sccomp', 'speckle'))

# scCODA (Python)
pip install scanpy sccoda

Quick Start

Tell your AI agent what you want to do: - "Did any cell-type proportions change between my conditions?" - "Test differential abundance with Milo on my integrated data" - "Run scCODA to find populations that expanded with treatment" - "Check whether my DE signal is actually a compositional shift"

Example Prompts

Cluster-free abundance

"Run Milo differential abundance on my kNN graph and report SpatialFDR neighborhoods" "Find transitional states that expanded with treatment without committing to clusters"

Cluster-based composition

"Test cell-type proportion changes with scCODA using a stable reference cell type" "Run sccomp on my count table, robust to outlier samples" "Use propeller to quickly test proportion differences across groups"

Guarding the DE/DA confound

"Pair my pseudobulk DE with a differential-abundance test" "Is the change in this cluster's expression real or just a shift in substate proportions?"

Interpretation

"Which population actually drives the compositional change relative to the reference?" "Do Milo and scCODA agree on which cell types changed?"

What the Agent Will Do

Confirm there are biological replicates per condition (samples, not cells, are the unit)
Choose cluster-free (Milo) and/or cluster-based (scCODA/sccomp/propeller) testing for the question
Build a per-sample cell-type count table or a Milo object from an integrated embedding
For scCODA, select a stable reference cell type (or automatic) and explain the relative interpretation
Fit the model and report FDR/SpatialFDR-controlled effects with direction and uncertainty
Reconcile cluster-free and cluster-based results when both are run
Cross-check against the condition-DE analysis to separate abundance shifts from expression changes

Tips

Samples are the replicate - more donors increase power for an abundance claim; more cells per donor barely do.
Never run a per-cluster proportion t-test - the simplex constraint makes one expansion force spurious depletions elsewhere.
Choose the scCODA reference deliberately - a reference that actually changes biases every other call; use automatic if unsure.
Report SpatialFDR for Milo - it corrects for overlapping neighborhoods; raw p-values over-call.
Milo k and prop trade resolution for power - larger neighborhoods are better powered but blur fine shifts.
Always pair DA with DE - a compositional shift can masquerade as differential expression and vice versa.
Build Milo's graph on an integrated embedding - batch structure in the kNN graph produces false abundance calls.

clustering - Define the clusters whose abundance is tested
cell-annotation - Annotate cell types before testing their proportions
markers-annotation - Pair condition DE with abundance testing to separate the confound
batch-integration - Build the integrated embedding Milo's kNN graph relies on
differential-expression/deseq2-basics - Pseudobulk condition DE that abundance testing complements
pathway-analysis/go-enrichment - Characterize the populations that expanded or contracted