Nephrotoxicity Analysis Pipeline

A comprehensive pipeline for analyzing drug-induced nephrotoxicity using single-cell RNA sequencing data and machine learning approaches. Zenodo link to data: https://doi.org/10.5281/zenodo.15724290 Pre-print: https://doi.org/10.1101/2025.06.17.660070

Overview

This pipeline integrates multiple analysis approaches to study nephrotoxicity:

• Differential Expression Analysis: Compare gene expression between toxic and non-toxic drug treatments
• Drug2Cell Analysis: Perform drug-cell interaction analysis using LINCS L1000 data
• Statistical Analysis: Conduct comprehensive statistical comparisons and visualizations
• Machine Learning: Build predictive models for nephrotoxicity classification
• Simulation Analysis: Perform power analysis and effect size simulations

Prerequisites

Before diving into this analysis, please make sure to run the R workflows provided in /kidney_single_cell_data_processing.

There is an additional readme within the directory for more explanations on that pipeline and how to run it.

Python Dependencies

Install the required packages:

pip install -r requirements.txt

Required Data Files

Download the input data folder from Zenodo and place it in the project root directory. Link: https://doi.org/10.5281/zenodo.15724290

The pipeline expects the following files in the input_data/ directory:

Core Dataset Files:

• kidney_data.h5ad - Single-cell annotated RNA-seq data in AnnData format
• V2_merged_drug_dataset.csv - Merged drug information dataset
• kidney_filtered_expanded_toxicity_df.csv - Toxicity labels and metadata

Reference Data Files:

• DIRIL_508.csv - DIRIL reference dataset
• pryan_reference_set_ades.csv - Reference adverse events dataset
• toxic_drugList_ACS.csv - ACS toxic drug list
• refined_drugbank_targets.csv - DrugBank target information
• recheck-kidney_lincs_matched_df.csv - LINCS matched dataset
• dr2c_var_genes.csv - Variable genes for Drug2Cell analysis
• base_sampled_data_1_percent.csv - Sampled data for simulation analysis

Quick Start

Option 1: Run Complete Pipeline

Execute the full analysis pipeline with default settings:

python run_analysis.py --all

Option 2: Custom Configuration

1. Copy and modify the configuration file:

cp config/default_config.yaml config/my_config.yaml
# Edit my_config.yaml with your specific settings

2. Run with custom configuration:

python run_analysis.py --config config/my_config.yaml

Option 3: Run Specific Components

Run individual analysis components:

# Differential expression analysis only
python run_analysis.py --components differential

# Statistical analysis and visualization
python run_analysis.py --components statistical

# Machine learning analysis
python run_analysis.py --components ml

# Multiple specific components
python run_analysis.py --components differential,statistical,ml

Output Structure

All results are saved in the pipeline_results/ directory:

pipeline_results/
├── differential_expression_results/     # DE analysis outputs
│   ├── direct_analysis_*.csv           # Direct comparison results
│   ├── drug_averaged_analysis_*.csv    # Drug-averaged results
│   └── *.png/*.svg                     # Visualization plots
├── statistical_results/                # Statistical analysis outputs
│   ├── correlation_matrix.csv          # Gene correlation data
│   ├── mannwhitney_results.csv         # Statistical test results
│   └── *.png/*.svg                     # Statistical plots
├── drug2cell_results/                  # Drug2Cell analysis outputs
│   ├── combined_drug_matrix.csv        # Drug signature matrix
│   └── analysis_summary.json          # Summary statistics
├── drug_score_ml_results/              # Drug score-based ML outputs
│   ├── best_model_predictions.csv      # Model predictions from drug scores
│   ├── feature_importance_*.csv        # Feature importance scores
│   ├── cross_validation_metrics.csv    # CV performance
│   └── *.png/*.svg                     # ML visualization plots
├── expression_ml_results/              # Expression-based ML outputs
│   ├── best_model_predictions.csv      # Model predictions from gene expression
│   ├── feature_importance_*.csv        # Feature importance scores
│   ├── cross_validation_metrics.csv    # CV performance
│   └── *.png/*.svg                     # ML visualization plots
├── simulation_results/                 # Power analysis outputs
│   ├── power_analysis_*.html/.png      # Interactive/static plots
│   └── power_analysis_results.csv     # Simulation results
└── pipeline_summary_report.txt         # Overall summary

Configuration

Key Configuration Options

Edit config/default_config.yaml to customize:

Input Data Paths:

input_data:
  merged_drug_dataset: "input_data/V2_merged_drug_dataset.csv"
  single_cell_data: "input_data/kidney_data.h5ad"
  # ... other data paths

Output Settings:

output:
  base_dir: "pipeline_results"
  formats: ["png", "svg"]  # Figure formats
  cleanup_intermediate: false

Analysis Components

1. Differential Expression Analysis

• Direct Method: Direct comparison between toxic vs non-toxic samples
• Drug-Averaged Method: Compare drug-averaged expression profiles
• Output: DE gene lists, fold changes, statistical significance, visualizations

2. Statistical Analysis

• Correlation Analysis: Gene-gene and drug-drug correlations
• Statistical Tests: Mann-Whitney U tests, t-tests
• Visualizations: Correlation heatmaps, boxplots, distribution plots

3. Drug2Cell Analysis

• LINCS Integration: Connect drug perturbations to cellular responses
• Signature Analysis: Analyze drug signature patterns
• Visualization: UMAP projections, drug clustering

4. Machine Learning

Two complementary ML approaches:

Drug Score-Based ML (drug_score_ml_results/):

• Input: Aggregated drug scores from Drug2Cell analysis
• Models: Random Forest, XGBoost, SVM, Logistic Regression
• Purpose: Predict nephrotoxicity using drug signature patterns

Expression-Based ML (expression_ml_results/):

• Input: Differential gene expression data
• Models: Random Forest, XGBoost, Logistic Regression
• Purpose: Predict nephrotoxicity using gene expression profiles
• Features: Differentially expressed genes, top variable genes

Common Outputs: Cross-validation metrics, ROC curves, feature importance, model predictions

5. Simulation Analysis

• Power Analysis: Determine optimal sample sizes
• Effect Size Analysis: Simulate various biological effect sizes

Example Workflows

Research Workflow

# 1. Explore data with statistical analysis
python run_analysis.py --components statistical

# 2. Identify differentially expressed genes
python run_analysis.py --components differential

# 3. Build predictive models
python run_analysis.py --components ml

# 4. Validate with simulation
python run_analysis.py --components simulation