CCLE Colon (Large Intestine) RNA-seq: PCA, Clustering, and DESeq2

Goal. Explore biologically distinct groups within colon cancer cell lines using bulk RNA-seq, and identify gene programs that separate those groups.

---

Dataset

• Counts: CCLE RNA-seq gene counts (*.gct.gz; genes × samples).
• Metadata: CCLE sample annotations (site, pathology, histology, sex, age, TCGA code, etc.).
• Focus subset: Site_Primary == "large_intestine" (58 samples).

---

Methods (summary)

1. Preprocessing

   • Read GCT, keep columns as samples.
   • Filter to protein-coding genes via Ensembl/biomaRt.
   • Collapse duplicate gene symbols by summing counts; set Description as rownames; drop non-numeric cols.
   • Align counts ↔ metadata by sample intersection & order; clean metadata ("missing-info" for NA in categoricals).

2. Exploration

   • Build DESeqDataSet; filter low counts; VST transform.
   • PCA with plotPCA() and prcomp(); elbow plot; extract gene loadings.
   • Hierarchical clustering (top 150 most-variable genes) → 4 clusters observed; for DE kept 2 clusters for balanced comparison.
   • Heatmaps with ComplexHeatmap + sample annotations.

3. Differential Expression

   • DESeq2 using design = ~ Cluster; shrink LFC & visualize with volcano plot.

---

Key findings (high-level)

• PC1 (~40% var) loads on intestinal epithelial/differentiation markers (e.g., GPA33, CEACAM5, CDH17, LGALS4, PHGR1, PPP1R1B).
  → Right side: more differentiated/epithelial-like; left: de-differentiated.
• PC2 (~9% var) highlights proteolysis & metabolism/invasion axes (e.g., KLK6, KLK10, KRT23, SLC2A3, MSLN, ZBED2).
• Unsupervised clustering using DESeq2 split cell lines into two broad states:
  • Cluster 1 (stromal/mesenchymal-leaning signals): up in SGIP1, PDPN, COL3A1, PTN, ANGPTL2, GDF6, KRTAP1-5, PODN, GNAT3, NTM.
  • Cluster 2 (epithelial/differentiated): up in SLC26A3, OLFM4, GPA33, FABP1, TINAG, UGT1A10, SPINK4, MUC17, PIGR, MOGAT2.
• No single metadata field (sex, pathology, TCGA code) fully explained the variance—supporting a transcriptional state separation rather than a labeled covariate.

---

Reproduce

Requirements: R (≥4.2), Quarto, and R/Bioc packages: DESeq2, ComplexHeatmap, biomaRt, matrixStats, ggplot2, EnhancedVolcano, ggrepel, dplyr, tidyr.

# from repo root
quarto render analysis/bulkRNAseq.qmd   # produces analysis/bulkRNAseq.pdf

If you don’t use Quarto, open the .qmd in RStudio and “Render”.

Limitations & future directions

• Batch/confounders: Metadata incompleteness (e.g., age, race) limits covariate modeling; consider surrogate variable analysis (SVA) or RUV.

• Clustering k: Validate k with silhouette, gap statistic, and stability resampling; try consensus clustering.

• Pathway biology: Run GSEA/ORA on DEGs (Hallmarks/Reactome) to formalize epithelial vs mesenchymal/metabolic programs.

• Generalization: Cross-validate with TCGA COAD/READ and other colon cell-line panels.