DE Visualization - Usage Guide

Overview

This skill covers creating publication-quality visualizations for differential expression results, including MA plots, volcano plots, PCA plots, and heatmaps. Works with both DESeq2 and edgeR output.

Prerequisites

install.packages(c('ggplot2', 'pheatmap', 'RColorBrewer', 'ggrepel'))
BiocManager::install('EnhancedVolcano')  # Optional

Quick Start

Tell your AI agent what you want to do: - "Create a volcano plot from my DESeq2 results" - "Make a heatmap of the top 50 differentially expressed genes" - "Generate a PCA plot colored by treatment group"

Example Prompts

MA and Volcano Plots

"Create an MA plot highlighting significant genes"

"Make a volcano plot with gene labels for top hits"

"Generate an EnhancedVolcano plot for my results"

PCA and Sample Clustering

"Show PCA of my samples colored by condition and shaped by batch"

"Create a sample distance heatmap"

"Plot MDS for my edgeR data"

Heatmaps

"Make a heatmap of significant genes clustered by expression"

"Create a heatmap for my genes of interest"

"Show expression patterns across samples"

Individual Genes

"Plot counts for gene X across conditions"

"Show expression of my candidate genes"

What the Agent Will Do

Extract and format results from DESeq2/edgeR
Apply appropriate transformations (vst, log)
Create publication-quality figures
Add annotations and labels
Save in requested format (PDF, PNG)

Common Plot Types

Plot	Shows	Use For
MA plot	LFC vs expression	QC, global view
Volcano	LFC vs significance	Identifying top genes
PCA	Sample relationships	Batch effects, outliers
Heatmap	Expression patterns	Gene clusters, validation

Tips

Always use variance-stabilized counts (vst or rlog) for PCA and heatmaps; raw counts make PC1 = library size
vst() function default is blind=TRUE; the current DESeq2 vignette recommends blind=FALSE for downstream visualization after the model is fit -- pass blind= explicitly
Scale heatmap rows (z-score) for comparable gene patterns; for QC heatmaps showing batch shifts, use scale='none' instead
Check p-value histogram for analysis quality (uniform + spike near 0 is correct; U-shape means batch effects)
Use shrunken LFCs for volcano plot x-axis and un-shrunken p-values for y-axis; shrinkage compresses the cloud but does NOT recompute p-values, so the same genes remain significant
Always set max.overlaps = Inf (or options(ggrepel.max.overlaps = Inf)) when labeling more than ~10 genes -- ggrepel's default silently drops labels
For outlier-robust top-variable-gene selection, use matrixStats::rowMads(assay(vsd)) instead of rowVars -- one outlier sample can inflate rowVars massively
Use colorblind-friendly palettes for publications
Save vector formats (PDF) for publications, raster (PNG) for presentations
MA plot cloud should be symmetric around LFC=0; asymmetry suggests normalization failure (TMM/RLE assumption violated)

deseq2-basics - Generate DESeq2 results for visualization
edger-basics - Generate edgeR results for visualization
de-results - Filter genes before visualization
data-visualization/volcano-and-ma-plots - Custom ggplot2/matplotlib volcano + MA with LFC shrinkage
data-visualization/dimensionality-reduction-plots - PCA / UMAP / t-SNE / PHATE
data-visualization/heatmaps-clustering - Advanced heatmap customization