DRExplainer: Quantifiable interpretability in drug response prediction with directed graph convolutional network.

Area of Science:

• Computational Biology
• Bioinformatics
• Machine Learning in Medicine

Background:

• Accurate prediction of cancer cell line response to therapeutic drugs is crucial for advancing personalized medicine.
• Existing deep learning models face challenges in integrating diverse biological data and predicting directional drug responses.
• Interpretability of predictive models is essential for clinical decision-making.

Purpose of the Study:

• To propose DRExplainer, a novel interpretable predictive model for cancer drug response prediction.
• To enhance prediction accuracy by integrating multi-omics profiles, drug chemical structures, and known responses within a directed bipartite network.
• To provide a quantifiable method for model interpretability and identify key biological features driving predictions.

Main Methods:

• Development of DRExplainer, a directed graph convolutional network model.
• Construction of a directed bipartite network integrating cell line multi-omics data, drug chemical structures, and drug response information.
• Implementation of a mask-learning mechanism to identify relevant subgraphs for prediction interpretability.
• Creation of a ground truth benchmark dataset for quantifiable model interpretability.

Main Results:

• DRExplainer demonstrated superior performance compared to state-of-the-art predictive methods and existing graph-based explanation methods.
• The model successfully identified relevant subgraphs, enhancing the interpretability of drug response predictions.
• Case studies validated the model's effectiveness in predicting novel drug responses and its interpretability.

Conclusions:

• DRExplainer offers an effective and interpretable approach for predicting cancer drug response.
• The model's ability to integrate diverse data and provide interpretable insights holds significant potential for personalized medicine.
• The developed interpretability method offers a quantifiable approach to understanding model predictions based on biological features.