DESeq2 Differential Expression Analysis (R/Bioconductor)

Overview

DESeq2 is the Bioconductor R package for differential gene expression analysis from bulk RNA-seq count data. It fits a negative binomial generalized linear model per gene, estimates dispersion parameters using empirical Bayes shrinkage across genes, and tests differential expression using Wald tests (two-group) or likelihood ratio tests (complex designs). DESeq2 is the R gold standard for RNA-seq DE analysis, with native Bioconductor integration for seamless import from Salmon (tximeta/tximport), featureCounts, or HTSeq.

When to Use

• Identifying differentially expressed genes between two experimental conditions (treated vs. control, disease vs. healthy) from bulk RNA-seq count data
• Analyzing multi-factor designs that account for batch effects or covariates (e.g., ~ batch + condition)
• Testing complex hypotheses with interaction terms (e.g., time × treatment) or reduced models using likelihood ratio tests (LRT)
• Importing Salmon pseudoalignment output via tximeta or tximport for transcript-level uncertainty propagation
• Performing LFC shrinkage with apeglm for ranked gene lists, volcano plots, and downstream pathway analysis
• Conducting time-series experiments or any design with more than two levels requiring model comparison
• Working in an R/Bioconductor ecosystem where integration with SummarizedExperiment, clusterProfiler, or EnhancedVolcano is needed
• Use pydeseq2-differential-expression instead for Python-based pipelines with the same statistical model
• Use edgeR for negative binomial DE with TMM normalization, quasi-likelihood F-tests, or TREAT testing
• Use omics-plotting SKILL after DE and for publication-quality plots of DESeq2 results
• Use gseapy-gene-enrichment SKILL after DE to interpret results at the pathway level

Prerequisites

• R packages: DESeq2 (Bioconductor), tximeta or tximport (Salmon import), apeglm (LFC shrinkage), pheatmap, ggplot2, EnhancedVolcano
• Data requirements: Raw (unnormalized) integer count matrix — gene rows × sample columns — plus a sample metadata data frame with matching column names. If using Salmon: per-sample quant.sf files and a transcript-to-gene mapping
• Environment: R ≥ 4.2, Bioconductor ≥ 3.16

# Install Bioconductor packages
if (!require("BiocManager", quietly = TRUE))
    install.packages("BiocManager")

BiocManager::install(c("DESeq2", "tximeta", "tximport", "apeglm",
                        "EnhancedVolcano"))
install.packages(c("pheatmap", "ggplot2", "dplyr"))

Pre-flight Interview

Settle these with the user before writing any analysis code.

decisions:
  - id: D1
    param: design
    kind: required
    source: data
    ask: "Which column of the sample sheet separates the groups you want to compare?"
    default: null

  - id: D2
    param: contrast
    kind: required
    source: user
    depends_on: [D1]
    ask: "Which two groups should be compared, and which group should have positive log2 fold changes?"
    default: null

  - id: D3
    param: design
    kind: required
    source: data
    depends_on: [D1]
    ask: "Which available variables should be included to adjust for nuisance variation (batch, donor, sex, sequencing run)?"
    default: "none"
    skip_if: "sample sheet carries no column besides the grouping variable"

  - id: D4
    param: test
    kind: required
    source: user
    depends_on: [D1, D2]
    ask: "Is the primary hypothesis a named comparison, or an overall effect of this factor across its levels?"
    default: "named comparison (Wald test)"

  - id: D5
    param: alpha
    kind: optional
    source: user
    ask: "How strong must the evidence be before a gene counts as changed (false-discovery rate)?"
    default: 0.05

  - id: D6
    param: lfcThreshold
    kind: optional
    source: user
    ask: "Test against a zero change, or against a minimum fold change worth caring about?"
    default: 0

  - id: D7
    param: lfcShrink.type
    kind: derived
    source: upstream
    ask: "Shrink fold-change estimates so low-count genes stop dominating the rankings and plots?"
    default: "apeglm, applied for ranking and visualization"

  - id: D8
    param: blind (vst/rlog)
    kind: derived
    source: upstream
    ask: "Should the variance-stabilizing transform ignore the design?"
    default: "TRUE for QC/clustering, FALSE for design-aware downstream uses"

  - id: D9
    param: independentFilter
    kind: optional
    source: user
    ask: "Should DESeq2 use its adaptive mean-count filter to maximize discoveries at the chosen FDR threshold?"
    default: true

design appears twice because the formula is assembled from both D1 and D3; D2 defines the direction of the reported effect. The pipeline explicitly passes alpha: 0.05 to results() so independent filtering is optimized for the reported FDR threshold rather than DESeq2's default of 0.1.

Quick Start

Complete two-group comparison from a count matrix in under 20 lines.

library(DESeq2)

# Load count matrix (genes x samples) and metadata
counts <- as.matrix(read.csv("counts.csv", row.names = 1))
coldata <- read.csv("metadata.csv", row.names = 1)
coldata$condition <- factor(coldata$condition)

# Build DESeqDataSet, run full pipeline
dds <- DESeqDataSetFromMatrix(countData = counts,
                               colData   = coldata,
                               design    = ~ condition)
dds <- dds[rowSums(counts(dds)) >= 10, ]   # pre-filter
dds <- DESeq(dds)

# Extract results (treated vs. control)
res <- results(dds, contrast = c("condition", "treated", "control"),
               alpha = 0.05)
summary(res)
# LFC shrinkage for visualization
res_shrunk <- lfcShrink(dds, contrast = c("condition", "treated", "control"),
                         type = "apeglm", coef = 2)

Workflow

Step 1: Prepare Count Matrix and Sample Metadata

Build a DESeqDataSet from a gene × sample count matrix and a colData data frame. Column names of the count matrix must match row names of colData.

library(DESeq2)
library(dplyr)

# Load raw count matrix (genes as rows, samples as columns)
counts <- as.matrix(read.csv("featureCounts_matrix.csv", row.names = 1))
# Or from featureCounts tab-delimited output — skip first comment line
# fc <- read.table("featurecounts_output.txt", header = TRUE, skip = 1, row.names = 1)
# counts <- as.matrix(fc[, 7:ncol(fc)])  # columns 7+ are sample counts

# Load sample metadata
coldata <- data.frame(
    condition = factor(c("control", "control", "control",
                          "treated", "treated", "treated")),
    batch     = factor(c("A", "A", "B", "A", "B", "B")),
    row.names = colnames(counts)
)

# Build DESeqDataSet
dds <- DESeqDataSetFromMatrix(countData = counts,
                               colData   = coldata,
                               design    = ~ condition)

cat("Samples:", ncol(dds), "\n")
cat("Genes:", nrow(dds), "\n")
cat("Condition levels:", levels(coldata$condition), "\n")

Step 2: Import from Salmon via tximeta

When reads were quantified with Salmon, use tximeta to import transcript-level estimates with proper offset correction that accounts for transcript length and GC bias.

library(tximeta)

# Build a coldata with a "files" column pointing to quant.sf files
quant_dirs <- file.path("salmon_output", coldata$sample_id, "quant.sf")
coldata$files <- quant_dirs
coldata$names <- coldata$sample_id

# Import with tximeta — automatically fetches transcript metadata from Ensembl
se <- tximeta(coldata)           # SummarizedExperiment with transcript-level data

# Summarize to gene level (requires Ensembl or custom txdb)
gse <- summarizeToGene(se)       # gene-level SummarizedExperiment

# Build DESeqDataSet from the SummarizedExperiment
dds <- DESeqDataSet(gse, design = ~ condition)

cat("Gene-level SE dimensions:", dim(gse), "\n")
cat("Assay names:", assayNames(gse), "\n")
# Assay "counts" = estimated counts; "abundance" = TPM; "length" = eff. lengths

Step 3: Pre-Filtering and Quality Control

Remove genes with very low counts to improve statistical power and reduce the multiple testing burden. Explore sample quality with PCA on variance-stabilized counts.

library(ggplot2)

# Pre-filter: keep genes with at least 10 reads total
keep <- rowSums(counts(dds)) >= 10
dds <- dds[keep, ]
cat("Genes after pre-filtering:", nrow(dds), "\n")

# Variance-stabilizing transformation for QC visualization
vsd <- vst(dds, blind = TRUE)    # blind = TRUE for exploratory QC

# PCA plot
pca_data <- plotPCA(vsd, intgroup = c("condition", "batch"), returnData = TRUE)
percent_var <- round(100 * attr(pca_data, "percentVar"))

ggplot(pca_data, aes(PC1, PC2, color = condition, shape = batch)) +
    geom_point(size = 3) +
    xlab(paste0("PC1: ", percent_var[1], "% variance")) +
    ylab(paste0("PC2: ", percent_var[2], "% variance")) +
    ggtitle("PCA — Variance-Stabilized Counts") +
    theme_bw()

ggsave("pca_plot.pdf", width = 6, height = 5)
cat("Saved pca_plot.pdf\n")

Step 4: Run DESeq() — Normalization, Dispersion, and Model Fitting

DESeq() runs three sequential steps: (1) median-of-ratios size factor estimation, (2) gene-wise and shrunken dispersion estimation, (3) negative binomial GLM fitting and Wald statistics.

# Run full DESeq2 pipeline
dds <- DESeq(dds)

# Inspect size factors (library normalization)
cat("Size factors:\n")
print(sizeFactors(dds))
# Expected range: 0.3–3.0; outliers indicate library quality issues

# Dispersion plot — should show decreasing dispersion trend with mean
plotDispEsts(dds, main = "Dispersion Estimates")

# Inspect fitted model coefficients
resultsNames(dds)
# [1] "Intercept" "condition_treated_vs_control"

Step 5: Extract Results and Apply FDR Correction

Use results() to extract Wald test statistics. Specify the contrast explicitly for clarity. Independent filtering maximizes the number of detectable genes.

# Extract results for the contrast of interest
res <- results(dds,
               contrast          = c("condition", "treated", "control"),
               alpha             = 0.05,   # FDR threshold for independent filtering
               lfcThreshold      = 0,      # Test H0: |LFC| = 0 (Wald test)
               independentFilter = TRUE)   # Maximize power via adaptive filtering

# Summary: how many genes are up/down/NA
summary(res)
# out of N with nonzero total read count:
# adjusted p-value < 0.05: ...
# LFC > 0 (up): ...
# LFC < 0 (down): ...
# outliers [Cook's distance]: ...
# low counts [below independent filtering threshold]: ...

# Convert to data frame and sort by adjusted p-value
res_df <- as.data.frame(res)
res_df <- res_df[order(res_df$padj, na.last = TRUE), ]

cat("Significant genes (padj < 0.05):", sum(res_df$padj < 0.05, na.rm = TRUE), "\n")
cat("Upregulated:", sum(res_df$padj < 0.05 & res_df$log2FoldChange > 0, na.rm = TRUE), "\n")
cat("Downregulated:", sum(res_df$padj < 0.05 & res_df$log2FoldChange < 0, na.rm = TRUE), "\n")

write.csv(res_df, "deseq2_all_results.csv")

Step 6: LFC Shrinkage with apeglm

Shrink log2 fold changes using the adaptive t prior (apeglm). Use shrunk LFCs for MA plots, volcano plots, and gene ranking — NOT for significance calling (padj comes from the unshrunk Wald test).

library(apeglm)

# apeglm shrinkage — specify coefficient name from resultsNames(dds)
res_shrunk <- lfcShrink(dds,
                         coef = "condition_treated_vs_control",
                         type = "apeglm",
                         res  = res)    # pass original results for speed

# Compare shrunk vs. unshrunk LFC
cat("Max |LFC| unshrunk:", max(abs(res$log2FoldChange), na.rm = TRUE), "\n")
cat("Max |LFC| shrunk:  ", max(abs(res_shrunk$log2FoldChange), na.rm = TRUE), "\n")

# Export shrunk results
res_shrunk_df <- as.data.frame(res_shrunk)
res_shrunk_df <- res_shrunk_df[order(res_shrunk_df$padj, na.last = TRUE), ]
write.csv(res_shrunk_df, "deseq2_shrunk_results.csv")

# Filter significant with effect size threshold
sig_genes <- res_shrunk_df[
    !is.na(res_shrunk_df$padj) &
    res_shrunk_df$padj < 0.05 &
    abs(res_shrunk_df$log2FoldChange) > 1, ]
cat("Significant (padj<0.05, |LFC|>1):", nrow(sig_genes), "\n")
write.csv(sig_genes, "deseq2_significant.csv")

Step 7: Visualize — MA Plot, Volcano Plot, and Heatmap

Compute/viz separation (required): Do not combine DESeq2 computation and plotting in the same code block. Step 6 (or Step 5) saves the results CSV; this step loads that CSV and produces plots only. This way plots can be regenerated or restyled without re-running the expensive computation. Plotting: for the volcano, read skills/data-visualization/omics-plotting/SKILL.md and follow its "Volcano" recipe on the exported DEG CSV (→ figures/volcano_plot.pdf). The heatmap below stays native R (pheatmap).

# Volcano: use the omics-plotting SKILL (`skills/data-visualization/omics-plotting/SKILL.md`) "Volcano" recipe on the exported DEG CSV -> figures/volcano_plot.pdf

library(pheatmap)
# ---Heatmap of top 50 DE genes ---
top50 <- head(rownames(sig_genes[order(sig_genes$padj), ]), 50)
mat   <- assay(vsd)[top50, ]            # variance-stabilized expression
mat   <- mat - rowMeans(mat)            # center by gene mean

annotation_col <- data.frame(
    Condition = coldata$condition,
    Batch     = coldata$batch,
    row.names = colnames(mat)
)
# Use the code in the omics-plotting SKILL (`skills/data-visualization/omics-plotting/SKILL.md`) for publication-quality heatmap plots or use below code to generate a heatmap
pheatmap(mat,
         annotation_col  = annotation_col,
         cluster_rows    = TRUE,
         cluster_cols    = TRUE,
         show_rownames   = TRUE,
         fontsize_row    = 6,
         filename        = "figures/heatmap_top50.pdf",
         width           = 8,
         height          = 10)

Step 8: Multi-Factor Design and Interaction Terms

For designs with batch correction, paired samples, or interaction effects between two variables.

# --- Batch-corrected two-group comparison ---
dds_batch <- DESeqDataSetFromMatrix(countData = counts,
                                     colData   = coldata,
                                     design    = ~ batch + condition)
dds_batch <- dds_batch[rowSums(counts(dds_batch)) >= 10, ]
dds_batch <- DESeq(dds_batch)

res_batch <- results(dds_batch,
                      contrast = c("condition", "treated", "control"),
                      alpha    = 0.05)
cat("Significant (batch-corrected):", sum(res_batch$padj < 0.05, na.rm = TRUE), "\n")

# --- Interaction design: test whether treatment effect differs between genotypes ---
# H0: the effect of treatment is the same in WT and KO
coldata$genotype <- factor(coldata$genotype)  # "WT" or "KO"

dds_int <- DESeqDataSetFromMatrix(
    countData = counts, colData = coldata,
    design = ~ genotype + condition + genotype:condition
)
dds_int <- DESeq(dds_int)
resultsNames(dds_int)
# Interaction coefficient: "genotypeKO.conditiontreated"

# Extract interaction term — genes where treatment effect differs by genotype
res_int <- results(dds_int, name = "genotypeKO.conditiontreated", alpha = 0.05)
cat("Genes with significant interaction:", sum(res_int$padj < 0.05, na.rm = TRUE), "\n")

# --- Likelihood Ratio Test for complex designs (e.g., time series) ---
# LRT compares full model vs. reduced model
dds_lrt <- DESeq(dds_int, test = "LRT", reduced = ~ genotype + condition)
res_lrt  <- results(dds_lrt)
cat("Genes with genotype:condition interaction (LRT):",
    sum(res_lrt$padj < 0.05, na.rm = TRUE), "\n")

Key Parameters

Parameter	Function	Default	Range / Options	Effect
design	DESeqDataSetFromMatrix()	(required)	~ var or ~ cov + var	Specifies the GLM formula; put batch/covariates before variable of interest
contrast	results()	NULL	c(var, level1, level2) or coefficient name	Defines the comparison; always specify explicitly for clarity
alpha	results()	0.1	0.01–0.10	FDR threshold used for independent filtering; does not change padj values
lfcThreshold	results()	0	0–2	Test H0:
independentFilter	results()	TRUE	TRUE/FALSE	Adaptive mean-count filtering to maximize the number of rejectible hypotheses
type	lfcShrink()	"apeglm"	"apeglm", "ashr", "normal"	Shrinkage prior; apeglm is fastest and best calibrated
coef	lfcShrink()	(required)	coefficient name from resultsNames()	Which model coefficient to shrink; must match resultsNames() output
test	DESeq()	"Wald"	"Wald", "LRT"	Wald: two-group comparisons; LRT: model comparison for complex designs
reduced	DESeq()	NULL	formula without the term to test	Required for LRT; specifies the null model
blind	vst() / rlog()	TRUE	TRUE/FALSE	TRUE for QC/exploration; FALSE for downstream analysis using dispersion info

Key Concepts

Negative Binomial Model and Dispersion

DESeq2 models each gene's count as Negative Binomial(mean = µ, dispersion = α), where mean µ depends on the experimental design through a log-linear model. The key innovation is dispersion shrinkage: gene-wise dispersion estimates are shrunk toward a fitted trend across all genes using an empirical Bayes approach, borrowing information across genes to stabilize estimates from small sample sizes.

Size Factor Normalization

DESeq2 uses the median-of-ratios method: for each sample, compute the ratio of each gene's count to the geometric mean of that gene across all samples, then take the median ratio across all genes as the size factor. This is robust to differentially expressed genes and does not require any assumptions about the fraction of DE genes.

Wald Test vs. Likelihood Ratio Test

• Wald test (default): tests a single coefficient β = 0; fast and appropriate for two-group comparisons and specific contrasts in multi-factor designs
• LRT: compares the full model to a reduced model by the likelihood ratio; appropriate for testing whether any level of a multi-level factor (e.g., 5 time points) has an effect, or for testing interaction terms as a group

Cook's Distance Outlier Detection

DESeq2 computes Cook's distance per gene per sample to flag potential expression outliers. Genes where any sample has a Cook's distance above a threshold are automatically replaced with NA in the results. With refit = TRUE (default), outlier samples are excluded and the model is refit for that gene.

Common Recipes

Recipe: Multiple Contrasts from One Fitted Model

Efficient when comparing multiple treatment groups — fit DESeq() once, extract multiple contrasts.

# Fit once
dds_multi <- DESeqDataSetFromMatrix(counts, coldata, design = ~ condition)
dds_multi <- dds_multi[rowSums(counts(dds_multi)) >= 10, ]
dds_multi <- DESeq(dds_multi)

# Extract contrasts — all vs. "control"
contrasts_list <- list(
    A_vs_ctrl = c("condition", "treatment_A", "control"),
    B_vs_ctrl = c("condition", "treatment_B", "control"),
    C_vs_ctrl = c("condition", "treatment_C", "control")
)

results_list <- lapply(names(contrasts_list), function(nm) {
    res <- results(dds_multi, contrast = contrasts_list[[nm]], alpha = 0.05)
    res_shrunk <- lfcShrink(dds_multi, contrast = contrasts_list[[nm]],
                             type = "apeglm", res = res)
    write.csv(as.data.frame(res_shrunk), paste0("results_", nm, ".csv"))
    cat(nm, "— significant:", sum(res$padj < 0.05, na.rm = TRUE), "\n")
    as.data.frame(res_shrunk)
})
names(results_list) <- names(contrasts_list)

Recipe: Paired Sample Design

Eliminate within-subject variability by including subject as a blocking factor.

# Paired design: each subject has pre- and post-treatment samples
coldata$subject   <- factor(c("S1","S1","S2","S2","S3","S3"))
coldata$timepoint <- factor(c("pre","post","pre","post","pre","post"))

dds_paired <- DESeqDataSetFromMatrix(counts, coldata,
                                      design = ~ subject + timepoint)
dds_paired <- dds_paired[rowSums(counts(dds_paired)) >= 10, ]
dds_paired <- DESeq(dds_paired)

res_paired <- results(dds_paired,
                       contrast = c("timepoint", "post", "pre"),
                       alpha    = 0.05)
cat("Significant (paired):", sum(res_paired$padj < 0.05, na.rm = TRUE), "\n")

Recipe: Rank Genes for GSEA Using Shrunk LFC × -log10(pvalue)

Create a pre-ranked gene list for input to GSEA or gseapy's prerank function.

# Build ranking metric: signed -log10(pvalue) with LFC direction
res_rank <- as.data.frame(res_shrunk)
res_rank <- res_rank[!is.na(res_rank$pvalue) & !is.na(res_rank$log2FoldChange), ]
res_rank$rank_metric <- sign(res_rank$log2FoldChange) *
                         (-log10(res_rank$pvalue + 1e-300))
res_rank <- res_rank[order(res_rank$rank_metric, decreasing = TRUE), ]

# Export as two-column TSv for fgsea or gseapy prerank
write.table(data.frame(gene  = rownames(res_rank),
                        score = res_rank$rank_metric),
            "gsea_ranked_list.rnk",
            sep = "\t", quote = FALSE, row.names = FALSE, col.names = FALSE)
cat("Ranked gene list saved:", nrow(res_rank), "genes\n")

Recipe: Save and Reload Fitted DESeq2 Object

Avoid re-running the expensive DESeq() step in subsequent sessions.

# Save fitted DDS object
saveRDS(dds, "dds_fitted.rds")
cat("Saved fitted DESeqDataSet to dds_fitted.rds\n")

# Reload in a new session
dds_loaded <- readRDS("dds_fitted.rds")
# Continue with results(), lfcShrink(), etc.
res_reloaded <- results(dds_loaded, contrast = c("condition", "treated", "control"))

Expected Outputs

File	Description
deseq2_all_results.csv	Full results table: baseMean, log2FoldChange, lfcSE, stat, pvalue, padj for all genes
deseq2_shrunk_results.csv	Results with apeglm-shrunk LFC; use for ranking and visualization
deseq2_significant.csv	Filtered results: padj < 0.05 and
pca_plot.pdf	PCA of variance-stabilized counts; used to check batch structure and outliers
dispersion_plot.pdf	Gene-wise vs. fitted dispersions; should show tight cloud around trend
figures/volcano_plot.pdf	Volcano plot with significance and LFC thresholds labeled
figures/heatmap_top50.pdf	Hierarchically clustered heatmap of top 50 DE genes
gsea_ranked_list.rnk	Pre-ranked gene list for pathway enrichment (fgsea/gseapy)
dds_fitted.rds	Serialized DESeqDataSet for checkpoint/resume

Troubleshooting

Problem	Cause	Solution
Error: design contains one or more variables with all samples having the same value	A design factor has only one level	Check table(coldata$condition); ensure all factor levels are present
Model matrix not full rank	Confounded design (e.g., batch perfectly correlated with condition)	Run table(coldata$batch, coldata$condition); drop the confounded variable or collect more samples
All or most padj = NA	Genes flagged by independent filtering (low mean count) or Cook's outliers	Check summary(res) for count of filtered genes; relax alpha or verify data quality
No significant genes despite strong biology	Under-powered experiment or high within-group variability	Verify n ≥ 3 per group; check PCA for outliers; inspect p-value histogram (should show spike near 0)
Error in lfcShrink: coefficient not found	coef name does not match resultsNames(dds)	Run resultsNames(dds) and copy the exact coefficient string into coef =
apeglm fails with convergence warning	Extreme LFC estimates (sparse data or complete separation)	Switch to type = "ashr" which is more robust to these cases
Very large size factors (> 5)	Extremely different library sizes, or normalized counts accidentally used	Check colSums(counts(dds)); ensure input is raw integer counts from aligner/counter
Dispersion plot shows outlier cloud far from trend	Noisy or failed libraries; incorrect metadata grouping	Examine PCA; remove outlier samples; check that design formula matches biological groups

References

• DESeq2 Bioconductor page — Package home, vignette, and reference manual
• DESeq2 vignette — Comprehensive worked example covering all design types
• Love et al. (2014) Genome Biology — Original DESeq2 paper: moderated estimation of fold change and dispersion for RNA-seq data
• Zhu et al. (2019) Bioinformatics — apeglm LFC shrinkage for DESeq2
• tximeta Bioconductor page — Salmon import with transcript-level metadata
• EnhancedVolcano Bioconductor page — Publication-quality volcano plots for DE results