PIGE: Pathway-Informed Graph Explanation Framework for Drug Response Interpretability

Overview

PIGE (Pathway-Informed Graph Explanation) is a deep learning framework that predicts anti-cancer drug response by understanding the underlying biological mechanisms. PIGE allows you to answer the following questions:

• Why is this cell line sensitive or resistant to this drug?
• Which pathways/genes/interactions drive the drug response or lack thereof?
• What happens to drug response if we block a specific pathway/gene/interaction?

Precomputed results can be explored using Interactive PIGE Atlas: https://www.pigeatlas.com

Approach

PIGE integrates multi-omics data (mutations, copy number alterations, gene expression) with drug chemical structures through a biologically informed pathway interaction network. The model architecture processes genomic features through this pathway graph, creating interpretable pathway-level representations. By systematically knocking out pathways and genes and measuring changes in model predictions, PIGE generates importance scores that identify influential biological mechanisms.

Key Results

• Improved Prediction Performance and Generalizability

  • Spearman ρ=0.84 in 5-fold cross-validation on CTRPv2 training data
  • Spearman ρ=0.56 on external GDSC2 validation (excluding all overlapping cell line-drug pairs)
  • Outperforms DrugCell (ρ=0.52), DRPreter (ρ=0.51), and other state-of-the-art methods on external validation

• Enhanced Interpretability

  • 74% pathway hit rate at K=25 vs 50% for CRISPR differential essentiality, 60% for GSEA, and 29% for DrugCell
  • Over 60% gene hit rate at K=100 (representing 4.8% of all genes) vs 41% for CRISPR screens
  • Recovers specific mechanisms of action including non-obvious targets (e.g. CX3CL1 as a resistance hub to chemotherapy in triple-negative breast cancer, only discovered experimentally in 2017)

• Enhanced Clinical Translation Potential

  • Validated on BeatAML ex vivo patient samples
  • Generalizes to unseen drugs and drug classes (e.g. MDM2 inhibitors not present in training data)
  • Identifies patient-specific resistance mechanisms that suggest rational combination therapies

Architecture

PIGE is based on a biologically informed pathway interaction network constructed from Gene Ontology (GO) biological processes and OmniPath protein-protein interactions. The architecture consists of three integrated components:

This design mirrors biology: each pathway's influence on a cell's response is determined by its interactions with upstream pathways, the cell's genomics, and the drug's properties.

Citation

If you use PIGE in your research, please cite:

[Will be added when paper is published]

Web Application

• Interactive PIGE Atlas: https://www.pigeatlas.com - Explore 45,000+ pathway graphs across 70 drugs and 800+ cell lines

---

Project Structure

The project consists of the following folders:

• data/ - Contains all data on which models were built
  • input_data/ - Raw input data (omics, drug response, pathway data)
  • intermediate_data/ - Processed intermediate files (pathway graphs, precomputed features)
  • output_data/ - Model outputs and trained model checkpoints
• src/ - Source code for PIGE
  • data_processing/ - Data preprocessing and feature generation
  • models/ - Model architectures (Structural Causal Model, GAT layers)
  • pathway_interaction_setup/ - Pathway graph construction from GO and OmniPath
  • model_interpretability/ - Virtual knockout analysis and interpretability tools
  • pige_atlas/ - Interactive graph visualization and atlas generation
  • train/ - Training scripts and experiment management
  • train_test/ - External validation and evaluation scripts
• notebooks/ - Jupyter notebook tutorials and examples
  • 01_train_erlotinib.ipynb - Train PIGE on a single drug
  • 02_knockout_analysis.ipynb - Virtual knockout analysis
  • 03_generate_pige_graphs.ipynb - Generate interactive pathway graphs

Quick Start

Virtual Environment Setup

We recommend using a virtual environment to manage dependencies:

# Create virtual environment (using conda)
conda create -n pige python=3.10
conda activate pige

# Or using venv
python -m venv venv
source venv/bin/activate  # On Windows: venv\Scripts\activate

Installation

# Install PyTorch (adjust CUDA version as needed)
# For CUDA 11.0+
pip install torch==1.12.0+cu116 torchvision==0.13.0+cu116 torchaudio==0.12.0 -f https://download.pytorch.org/whl/torch_stable.html

# Install PyTorch Geometric and dependencies
pip install torch-scatter torch-sparse torch-cluster torch-spline-conv -f https://data.pyg.org/whl/torch-1.12.0+cu116.html
pip install torch-geometric

# Install other dependencies
pip install -r requirements.txt

Basic Usage

Start with the tutorial notebooks in notebooks/:

1. Train a model: notebooks/01_train_erlotinib.ipynb (~15 minutes)
2. Run interpretability: notebooks/02_knockout_analysis.ipynb (~5 minutes)
3. Generate graphs: notebooks/03_generate_pige_graphs.ipynb (~5 minutes)

For programmatic usage, use the main pipeline:

from src.pipeline.pipeline_orchestrator import PipelineOrchestrator
from src.main import resolve_config_variables
import yaml

# Load and resolve config
with open('src/configs/config.yaml', 'r') as f:
    config = yaml.safe_load(f)
config = resolve_config_variables(config)

# Run pipeline
orchestrator = PipelineOrchestrator(config)
orchestrator.run()

Notebooks

Interactive tutorials demonstrating PIGE's capabilities:

Tutorial	Description	Runtime	Link
Train PIGE	Train PIGE on a single drug (Erlotinib) using the full 5-stage pipeline	~15 min	notebooks/01_train_erlotinib.ipynb
Virtual Knockout Analysis	Identify key pathways and genes driving drug sensitivity vs resistance	~5 min	notebooks/02_knockout_analysis.ipynb
Generate PIGE Graphs	Create interactive visualizations of pathway crosstalk and mechanisms	~5 min	notebooks/03_generate_pige_graphs.ipynb

Data

Data Sources

PIGE uses the following publicly available datasets:

• CTRPv2 - Cancer Therapeutics Response Portal v2 (drug response data)
• GDSC - Genomics of Drug Sensitivity in Cancer (external validation)
• DepMap - Dependency Map (omics data: mutations, CNAs, expression)
• Gene Ontology - Biological process annotations (pathway definitions)
• OmniPath - Protein-protein interaction network (pathway crosstalk)

Requirements

System Requirements

• Python: 3.10+
• PyTorch: 1.13+ (with CUDA support recommended)
• CUDA: 11.0+ (GPU recommended, 8GB+ VRAM)
  • Training and interpretability will work on CPU but will be significantly slower
• Disk Space: ~30GB for data and intermediate files (if training on all 68 drugs)
• RAM: 16GB+ needed for interpretability analysis

Contact

For questions, issues, or collaboration inquiries:

• Email: cbahl076@uottawa.ca
• Issues: Please open an issue on GitHub for bug reports or feature requests