Exercise 10 — PCA, QC and Sample Structure in RNA-seq

In real RNA-seq analysis, we never start with differential expression.

We first ask:

Do the samples behave as expected?

This is the role of:

  • PCA (Principal Component Analysis)
  • sample distance heatmaps
  • clustering

These steps help detect:

  • outliers
  • batch effects
  • mislabeled samples
  • biological structure

Dataset

We reuse:

:contentReference[oaicite:0]{index=0}

from Exercise 8.

Part 1 — Variance Stabilization

Raw counts are not suitable for PCA.

Task 1.1

Create a transformed dataset:

vsd <- vst(dds)

Task 1.2 — Questions

  1. Why can’t we use raw counts for PCA?
  2. What does variance stabilization do conceptually?

Part 2 — Sample Distance Heatmap

Task 2.1

Compute distances:

sampleDists <- dist(t(assay(vsd)))

Task 2.2

Convert to matrix:

sampleDistMatrix <- as.matrix(sampleDists)

Task 2.3

Plot heatmap:

pheatmap(sampleDistMatrix)

Task 2.4 — Interpretation

Answer:

  1. Which samples cluster together?
  2. Do replicates cluster properly?
  3. Any signs of outliers?

Part 3 — PCA (Built-in DESeq2 method)

Task 3.1

Run PCA:

plotPCA(vsd, intgroup = "dex")

Task 3.2 — Questions

  1. What does PC1 represent?
  2. What does PC2 represent?
  3. Do treated vs untreated samples separate?

Task 3.3 — Variance Interpretation

If PCA shows:

  • PC1 = 70%
  • PC2 = 15%

What does this mean?

Part 4 — Custom PCA Plot (ggplot2)

Task 4.1

Extract PCA data:

pcaData <- plotPCA(vsd, intgroup = "dex", returnData = TRUE)

Task 4.2

Plot manually:

library(ggplot2)

ggplot(pcaData, aes(PC1, PC2, color = dex)) +
  geom_point(size = 3)

Task 4.3 — Improve plot

Add:

  • title
  • axis labels
  • theme

Task 4.4 — Interpretation

Answer:

  1. Are samples clearly separated?
  2. Which axis drives separation?
  3. Does biology or noise dominate structure?

Part 5 — Hierarchical Clustering

Task 5.1

Compute clustering:

hc <- hclust(sampleDists)

Task 5.2

Plot dendrogram:

plot(hc)

Task 5.3 — Questions

  1. Which samples cluster together?
  2. Are clusters consistent with treatment?
  3. Any unexpected grouping?

Part 6 — Outlier Detection

Scenario

Imagine one sample is far away in PCA space.

Task 6.1 — Questions

  1. Should you remove it immediately?
  2. What checks should you perform first?
  3. Possible biological vs technical reasons?

Part 7 — Biological Interpretation

Answer:

# Does treatment explain most variation?

# Are replicates consistent?

# Is there evidence of batch effects?

# Would you proceed to DESeq2 confidently?

# Why is QC before DE analysis critical?

Part 8 — Multi-panel Figure (Optional)

Create a combined figure:

  • PCA plot
  • sample distance heatmap

Hint:

library(patchwork)

Bonus

B1 — Colored PCA by sample type

Add shape or label for clarity.

B2 — Identify strongest genes driving PC1

Hint:

attr(plotPCA(vsd), "percentVar")

B3 — Biological reflection

Write:

  • what drives variation in this dataset?
  • is it treatment or something else?
  • would you trust DESeq2 results if PCA looked random?

Key Takeaways

  • PCA is a QC tool, not just visualization
  • sample structure must be understood before DE analysis
  • clustering reveals hidden batch effects
  • variance stabilization is essential for meaningful PCA
  • good RNA-seq analysis always starts with QC
⬅️ Previous Exercise 🏠 Home ➡️ Next Exercise