DESeq2 Basics - Usage Guide

Overview

This skill covers differential expression analysis using DESeq2, the most widely-used Bioconductor package for analyzing RNA-seq count data. DESeq2 uses a negative binomial model to test for differential expression between experimental conditions.

Prerequisites

if (!require('BiocManager', quietly = TRUE))
    install.packages('BiocManager')
BiocManager::install(c('DESeq2', 'apeglm'))

Quick Start

Tell your AI agent what you want to do: - "Run DESeq2 on my count matrix with treated vs control comparison" - "Analyze differential expression controlling for batch effects" - "Get significantly differentially expressed genes with padj < 0.05"

Example Prompts

Basic Analysis

"Create a DESeqDataSet from my count matrix and sample metadata"

"Run the standard DESeq2 workflow on this RNA-seq data"

"Apply log fold change shrinkage to my DESeq2 results"

Design Formulas

"Set up DESeq2 with batch correction"

"Create an interaction model for genotype and treatment"

"Compare multiple conditions against a control"

Results

"Extract genes with adjusted p-value < 0.05 and |log2FC| > 1"

"Order results by significance"

"Export normalized counts and DE results"

What the Agent Will Do

Create DESeqDataSet from counts and metadata
Pre-filter low-count genes
Set reference level for comparisons
Run DESeq() pipeline
Apply lfcShrink() for fold change shrinkage
Extract and filter significant results

Input Requirements

Input	Format	Description
Count matrix	Integer matrix	Genes (rows) x Samples (columns)
Sample metadata	Data frame	Rownames matching count column names
Design formula	R formula	Variables from metadata (~condition)

Tips

Always use biological replicates (minimum 3 per condition)
Pre-filter genes with very low counts before analysis (speed only; independent filtering handles statistics at results())
Use lfcShrink() with type='apeglm' for better fold change estimates -- but note apeglm requires coef= and cannot use contrast=; rebuild design as ~ 0 + group for interaction contrasts, or use type='ashr'
Use vst() instead of rlog() for large datasets (>100 samples)
Set the reference level BEFORE calling DESeq() -- alphabetical default silently inverts fold-change direction
Check resultsNames(dds) to see available coefficients and ALWAYS pass name= or contrast= explicitly; results(dds) with no argument returns the last coefficient and depends on design order
With LRT (test='LRT'), use padj only as an omnibus screen -- the reported LFC is for the last coefficient, NOT an omnibus summary
The shrunken LFC and the Wald p-value come from different models -- shrinking does not recompute p-values
For prokaryotic data, verify normalization assumptions (majority-DE experiments violate median-of-ratios)
Use PyDESeq2 for Python-only environments; results closely match R DESeq2

edger-basics - Alternative DE analysis with edgeR
de-visualization - MA plots, volcano plots, heatmaps
de-results - Extract and export significant genes
pathway-analysis/go-enrichment - Functional analysis of DE gene lists
expression-matrix/counts-ingest - Loading count matrices from various formats