XMR: an explainable multimodal neural network for drug response prediction

Area of Science:

• Computational biology
• Bioinformatics
• Machine learning in oncology

Background:

• Large-scale preclinical cancer drug response databases enable drug discovery.
• Deep learning models excel at cancer drug response prediction but lack interpretability.
• Existing interpretable models show limited performance for clinical application.

Purpose of the Study:

• To develop an explainable multimodal neural network (XMR) for accurate and interpretable cancer drug response prediction.
• To integrate genomic and drug structural features for enhanced predictive modeling.
• To provide biological justification for predicted drug responses in cancer.

Main Methods:

• Developed the XMR model, a multimodal neural network with a visible neural network (VNN) for genomic features and a graph neural network (GNN) for drug structures.
• Integrated VNN and GNN via a multimodal fusion layer to model drug response.
• Utilized pathway hierarchies from Reactome Pathway Database as VNN architecture for triple-negative breast cancer drug response prediction.
• Applied a pruning approach for improved model interpretability.

Main Results:

• The XMR model demonstrated superior predictive performance compared to state-of-the-art interpretable deep learning models.
• The model successfully provided biological insights explaining drug responses in triple-negative breast cancer.
• The combination of VNN and GNN effectively captured key genomic and molecular features.

Conclusions:

• The XMR model offers a balance of predictive accuracy and biological interpretability for cancer drug response.
• This approach holds promise for advancing personalized cancer therapy by identifying effective drugs based on patient-specific data.
• The multimodal fusion of genomic and structural data enhances understanding of drug-cancer interactions.