DRExplainer: Quantifiable Interpretability in Drug Response Prediction with Directed Graph Convolutional Network

2024-08-22

Artificial Intelligence in Medicine

AI Summary

Poor model interpretability and difficulty in directional biological modeling hinder clinical adoption of drug response prediction in cancer cell lines. Method: We propose the first quantifiable causal attribution framework for drug response prediction, defining causal attribution scores based on directed graphs to enable consistent cross-cell-line and cross-drug interpretability assessment. The framework integrates a Directed Graph Convolutional Network (DGCN), pathway-aware attention, multi-scale gene graph construction, and a graph-adapted Grad-CAM variant. Contributions/Results: Our method achieves AUC > 0.91 on GDSC and CCLE benchmarks. Causal attributions are significantly enriched in KEGG pathways (FDR < 0.001). In blinded clinician evaluations, interpretability scores improved by 37%, substantially enhancing clinical trustworthiness and biological traceability.

Technology Category

• Machine Learning: Causal Learning
• Reasoning under Uncertainty: Causality
• Computer Vision: Interpretability, Explainability, and Transparency

Application Category

• Graph Algorithms and Modeling for the Web: Foundation models and LLMs for Web-related graphs
• Search and Retrieval-Augmented AI: Web learning to rank, online learning, and counterfactual learning for ranking
• User Modeling, Personalization and Recommendation: Explainable and interpretable methods for personalization

Problem

• Predicting cancer cell line drug response accurately
• Integrating multi-omics data for directed drug response prediction
• Providing quantifiable interpretability in drug response models

Innovation

• Uses directed graph convolutional network
• Integrates multi-omics and chemical data
• Quantifiable interpretability with benchmark dataset