mdhewei/bioinfor-claw/drug-sensitivity-for-gene

Drug Sensitivity for Gene

Purpose

Identify drugs whose cellular sensitivity correlates with a gene's expression or essentiality score across a panel of cancer cell lines. This enables:

• Discovery of synthetic lethal interactions (high dependency genes)
• Identification of drugs effective in cell lines with high/low gene expression
• Biomarker-driven drug selection
• Mechanistic understanding of drug efficacy across genetic contexts

Use When

• You want to find drugs whose efficacy correlates with a specific gene's expression
• You're screening for synthetic lethal interactions (genes whose knockout increases drug sensitivity)
• You need to identify patient stratification biomarkers for drug response
• You're prioritizing gene targets based on downstream drug opportunities
• You want to understand the mechanistic connection between a gene and drug response

Do Not Use When

• You need to predict drug response in a specific patient (use pharmacogenomics databases instead)
• You're looking for drug-drug interactions
• You want cell-cycle phase-specific sensitivity analysis
• You need epigenetic or post-translational modification effects
• You lack internet access (requires DepMap/PRISM API or local data)

Expected Inputs & Outputs

Inputs

• --gene (required): Gene symbol (e.g., TP53, BRCA1, KRAS)
• --omics-type: Type of gene measurement (choices: expression, dependency, copy_number; default: expression)
• --correlation-method: Statistical correlation method (choices: pearson, spearman; default: spearman)
• --fdr-cutoff: False Discovery Rate threshold (default: 0.05)
• --top-n: Number of top correlated drugs to visualize (default: 20)
• --min-cell-lines: Minimum cell lines with data for both gene and drug (default: 50)
• --depmap-dir: Path to local DepMap data directory (optional)
• --prism-file: Path to PRISM drug sensitivity TSV/HDF5 (optional)
• --expression-file: Path to DepMap expression matrix (optional)
• --outdir: Output directory

Outputs

• drug_correlation.tsv: Results table (drug_name, moa, target, r, pvalue, fdr, n_cell_lines)
• scatter_plot_top4.png: 2x2 grid of scatter plots for top 4 drugs (gene expression vs sensitivity)
• waterfall_plot.png: Ranked bar chart of top N drug correlations (red=positive, blue=negative)
• correlation_summary.txt: Summary statistics and parameters used

Procedure

Step 1: Obtain Data

If local files not provided:

1. Download DepMap expression data (OmicsExpressionProteinCodingGenesExpectedCountsLogTPMPlusOne)
2. Download PRISM repurposing drug sensitivity screen data
3. Cache data locally to avoid repeated downloads

Step 2: Extract Gene Vector

1. Retrieve gene expression or dependency scores across cell lines
2. Filter for cell lines with complete data (no missing values)
3. Normalize expression if needed (already log-transformed in DepMap)

Step 3: Compute Correlations

1. For each drug in PRISM:

• Identify cell lines with data for both gene and drug
• Require minimum of --min-cell-lines with both measurements
• Compute Pearson or Spearman correlation
• Calculate p-value
• Store: r, p-value, n_cell_lines, drug metadata (MOA, target)

Step 4: Multiple Testing Correction

1. Apply Benjamini-Hochberg FDR correction across all drug correlations
2. Filter results by --fdr-cutoff (default 0.05)
3. Sort by absolute correlation value

Step 5: Visualization

1. Scatter plots: Top 4 drugs, gene expression (x) vs drug sensitivity (y)

• Add regression line
• Color by cell line lineage if available
• Include R and p-value

1. Waterfall plot: Ranked bar chart of top N drugs

• Red bars: positive correlation (high expression = high sensitivity)
• Blue bars: negative correlation (high expression = low sensitivity)
• Sorted by effect size

Key Execution Patterns

Basic usage (gene expression vs all drugs):

python scripts/drug_sensitivity_for_gene.py \
  --gene TP53 \
  --outdir results/tp53_drugs

CRISPR dependency correlation:

python scripts/drug_sensitivity_for_gene.py \
  --gene BRCA1 \
  --omics-type dependency \
  --correlation-method spearman \
  --fdr-cutoff 0.01 \
  --top-n 30 \
  --outdir results/brca1_synthetic_lethal

Copy number and local data:

python scripts/drug_sensitivity_for_gene.py \
  --gene MYC \
  --omics-type copy_number \
  --prism-file /data/prism_repurposing.csv \
  --expression-file /data/depmap_expression.csv \
  --min-cell-lines 100 \
  --outdir results/myc_cn_drugs

High-stringency filtering:

python scripts/drug_sensitivity_for_gene.py \
  --gene EGFR \
  --fdr-cutoff 0.001 \
  --min-cell-lines 200 \
  --top-n 10 \
  --outdir results/egfr_high_confidence

Parameter Decision Guide

Failure Modes

Technical Notes

• Correlation is computed only on cell lines with complete data for both gene and drug
• Spearman correlation recommended for non-normal distributions (typical for sensitivity data)
• FDR correction uses Benjamini-Hochberg procedure (controls false positive rate)
• Drug metadata (MOA, target) sourced from PRISM manifest
• Results are robust to cancer type differences (uses cross-cancer cell line panel)
• Positive correlation: high gene activity -> high drug sensitivity
• Negative correlation: high gene activity -> resistance to drug (protective/therapeutic antagonism)
• Minimum 50 cell lines recommended for reliable correlation estimates
• All plots use matplotlib; vector-compatible output (PNG 300 DPI)