edgeR Basics - Usage Guide

Overview

This skill covers differential expression analysis using edgeR, a powerful Bioconductor package that uses empirical Bayes methods to moderate gene-wise dispersion estimates. The quasi-likelihood F-test framework provides robust results even with small sample sizes.

Prerequisites

if (!require('BiocManager', quietly = TRUE))
    install.packages('BiocManager')
BiocManager::install('edgeR')

Quick Start

Tell your AI agent what you want to do: - "Run edgeR analysis on my count matrix with treated vs control comparison" - "Perform differential expression with edgeR controlling for batch" - "Get top differentially expressed genes with FDR < 0.05"

Example Prompts

Basic Analysis

"Create a DGEList from my count matrix"

"Run the standard edgeR quasi-likelihood workflow"

"Normalize my RNA-seq data with TMM"

Design and Contrasts

"Set up edgeR with multiple comparisons"

"Create contrasts to compare all treatments against control"

"Test for interaction effects between genotype and treatment"

Results

"Get genes with FDR < 0.05 and |logFC| > 1"

"Export edgeR results to CSV"

"Compare edgeR and DESeq2 results"

What the Agent Will Do

Create DGEList object from counts
Filter low-expression genes with filterByExpr()
Normalize using TMM (calcNormFactors)
Set up appropriate design matrix
Estimate dispersions (estimateDisp)
Fit quasi-likelihood model (glmQLFit)
Perform statistical testing (glmQLFTest)
Extract and filter results

Input Requirements

Input	Format	Description
Count matrix	Integer matrix	Genes (rows) x Samples (columns)
Group factor	Factor	Sample group assignments
Design matrix	Matrix	From model.matrix()

Tips

Use filterByExpr() for automatic low-count filtering -- skipping it inflates the multiple-testing burden; edgeR has no automatic independent filtering like DESeq2
The quasi-likelihood framework (glmQLFit) is recommended over exact test; QL F-test is the modern default with proper finite-sample FPR control
Set BOTH estimateDisp(robust=TRUE) AND glmQLFit(robust=TRUE) -- they propagate Phipson 2016 robust hyperparameter estimation through separate EB steps
For complex designs, use design matrices without intercept (~ 0 + group)
glmTreat() tests for log fold changes above a threshold (H0: |LFC| <= threshold), preferred over post-hoc LFC filtering
edgeR v4 introduced bias-corrected APL with glmQLFit(legacy=FALSE) as the default -- pre-2025 results are reproducible with legacy=TRUE, but the v4 defaults are improved
normLibSizes() is the v4 canonical name (replaces calcNormFactors()); method='TMM' remains the documented default, with method='TMMwsp' as the preferred alternative for sparse / single-cell-pseudobulk data
For transcript-level DE, use catchSalmon() / catchKallisto() (Baldoni 2024) instead of standard tximport
For prokaryotic data, TMM normalization assumes most genes are not DE -- verify this holds

deseq2-basics - Alternative DE analysis with DESeq2
de-visualization - MA plots, volcano plots, heatmaps
de-results - Extract and export significant genes
pathway-analysis/go-enrichment - Functional analysis of DE gene lists