TCGA-DepMap Drug Profiler

Digital patient generation and drug response prediction through TCGA–DepMap integration.

A computational oncology pipeline that bridges the gap between patient transcriptomes (TCGA-LUAD) and cell-line drug screening data (DepMap) using CVAE-generated virtual patients and multi-level similarity scoring.

---

Motivation

Cell-line drug sensitivity data from DepMap/CCLE covers 578+ compounds but lacks direct clinical relevance — cell lines don't represent individual patient biology. This pipeline answers:

Which cell line drug response data can we trust for a given patient subtype?

Key Results

Metric	Value
Virtual patients generated	400 (100 per stage, CVAE)
Feature genes selected	~3,500 (ANOVA + driver + pathway)
Compounds reinterpreted	578
Patient–cell line pairs scored	10,000
High-confidence matches (>0.8)	~1,800

CVAE Validation

Original vs Generated	Stage-Conditioned Latent Space

Left: expression distribution overlap between real and virtual patients (0.7–0.8 intentional — avoids overfitting). Right: PCA of CVAE latent space showing stage-specific clustering.

Similarity Analysis

Patient–Cell Line Similarity Matrix	Stage-Specific Similarity Networks

Left: 10 virtual patients × 25 LUAD cell lines integrated similarity scores. Right: High-confidence (>0.8) patient–cell line pairs, stratified by cancer stage.

Top 20 Representative Cell Lines

Q75 similarity score by cancer stage. Bubble size = number of matched virtual patients. Lines with large bubbles across multiple stages serve as broadly representative experimental models.

Pipeline

  TCGA-LUAD                         DepMap/CCLE
  ─────────                         ───────────
  [01] Data processing              [03] Cell line preparation
       ↓                                  ↓
  [02] Feature selection ───────→ Common gene space
       ↓
  [CVAE] Virtual patient             [04] Multi-level similarity
         generation (Python)    ──→  S = 0.2·Global + 0.3·Pathway + 0.5·Target
                                          ↓
                                    [05] Drug response integration (578 compounds)
                                          ↓
                                    [06] Network analysis & top cell lines

Steps

Step	Script	Description
01	R/01_tcga_data_processing.R	Download TCGA-LUAD, Ensembl→symbol mapping, MAD filter, log2
02	R/02_feature_selection.R	ANOVA 3000 + Stage I vs IV 500 + driver 38 + Hallmark pathways
CVAE	python/cvae_train.py	Conditional VAE: encoder/decoder [1024→512→256→50], 400 virtual patients
03	R/03_depmap_preparation.R	LUAD cell lines, HVG 5000, 3-tier gene selection
04	R/04_similarity_analysis.R	Spearman + GSVA/Pearson + cosine → integrated score [0,1]
05	R/05_drug_response.R	Drug sensitivity cutoffs, per-cell-line reports, weighted response
06	R/06_network_analysis.R	Similarity networks (>0.8), stage-specific analysis, top 20 dotplot

Multi-Level Similarity

$$S_{integrated} = 0.2 \times S_{global} + 0.3 \times S_{pathway} + 0.5 \times S_{target}$$

Level	Method	Weight	Rationale
S_global	Spearman correlation (full genes)	0.2	Overall transcriptomic similarity
S_pathway	GSVA enrichment → Pearson	0.3	Pathway-level biological correspondence
S_target	Cosine similarity (C6 gene sets)	0.5	Drug-actionable target concordance

Project Structure

tcga-depmap-drug-profiler/
├── R/
│   ├── 01_tcga_data_processing.R
│   ├── 02_feature_selection.R
│   ├── 03_depmap_preparation.R
│   ├── 04_similarity_analysis.R
│   ├── 05_drug_response.R
│   └── 06_network_analysis.R
├── python/
│   ├── cvae_train.py
│   ├── run_pipeline.py
│   └── requirements.txt
├── report/
│   ├── project_overview.Rmd
│   └── digipat_styles.css
├── config.yaml
└── figures/              # Key output figures (committed)
    ├── Original_vs_Generated.png
    ├── Stage_PCAplot.png
    ├── Latent_PCAplot.png
    ├── Ten_patient_cellline_similarity_matrix.png
    ├── stage_specific_similarity_networks.png
    ├── top20_celllines_dotplot.png
    └── cellline_stage_similarity_stats.csv

Quick Start

# 1. Install R dependencies
Rscript -e "install.packages(c('tidyverse','yaml','matrixStats','umap','pheatmap','msigdbr','GSVA','igraph','ggraph','kableExtra'))"
Rscript -e "BiocManager::install(c('TCGAbiolinks','SummarizedExperiment','org.Hs.eg.db'))"

# 2. Install Python dependencies
pip install -r python/requirements.txt

# 3. Run full pipeline
python python/run_pipeline.py

# Or run individual steps
Rscript R/01_tcga_data_processing.R
Rscript R/02_feature_selection.R
python python/cvae_train.py --epochs 300 --n-per-stage 100
Rscript R/04_similarity_analysis.R

Configuration

All parameters are centralized in config.yaml:

• TCGA filtering: min sample %, TPM threshold, MAD quantile
• Feature selection: number of ANOVA/extreme genes
• DepMap: HVG count, expression cutoff
• Similarity weights: global/pathway/target ratios
• CVAE: latent dim, epochs, beta, patients per stage
• Network: similarity threshold, max edges

Data Sources

Source	Description	Access
TCGA-LUAD	Lung adenocarcinoma RNA-seq (STAR counts + TPM)	Public (GDC API)
DepMap	Cell line expression + drug sensitivity	Public (download)
MSigDB	Hallmark (H) + oncogenic (C6) gene sets	Public (msigdbr)

Requirements

• R ≥ 4.3 with Bioconductor
• Python ≥ 3.9 with PyTorch ≥ 2.0
• ~16 GB RAM recommended for GSVA computation

License

MIT