Path-Based Graph Neural Network for Drug Synergy Prediction and Interpretation

Area of Science:

• Computational biology
• Pharmacology
• Artificial intelligence in medicine

Background:

• Combination therapy offers improved efficacy and reduced toxicity for complex diseases.
• The growing number of drug combinations presents challenges in drug screening and synergy prediction.
• Existing prediction methods often lack interpretability regarding mechanisms of action.

Purpose of the Study:

• To develop an interpretable model for predicting synergistic drug combinations.
• To elucidate the mechanisms underlying drug synergy through model interpretation.
• To address the limitations of current drug synergy prediction methods.

Main Methods:

• Proposed SDCInterpreter, a path-based interpretable graph neural network.
• Constructed a heterogeneous graph integrating drugs, genes, pathways, and cell line entities.
• Employed relational graph convolutional networks, mask learning, and Dijkstra's algorithm for prediction and interpretation.

Main Results:

• SDCInterpreter demonstrated strong performance in predicting drug synergy.
• The model successfully generated interpretable insights into the mechanisms of synergistic drug combinations.
• Experimental results validated the model's predictive and interpretive capabilities.

Conclusions:

• SDCInterpreter provides an effective approach for predicting and interpreting drug synergy.
• The model enhances understanding of drug combination mechanisms in cell lines.
• This interpretable AI method can aid in clinical drug discovery and development.