Interpretable graph deep learning framework for drug synergy prediction by integrating functional and clinical...

Area of Science:

• Computational biology
• Drug discovery
• Artificial intelligence in medicine

Background:

• Single-agent cancer therapies are limited by tumor heterogeneity and drug resistance.
• Combination treatments are essential but traditional screening methods are costly and inefficient.
• Predicting drug synergy computationally can accelerate the discovery of effective combination therapies.

Purpose of the Study:

• To develop and validate DSimSynergy, a graph deep learning framework for predicting drug synergy.
• To reduce the cost and improve the efficiency of identifying synergistic drug combinations.
• To elucidate the underlying biological mechanisms of drug synergy for personalized treatment strategies.

Main Methods:

• Constructed drug similarity networks based on biological processes and clinical applications.
• Employed graph convolution and graph attention mechanisms to learn drug and cell line representations.
• Integrated drug molecular fingerprints and cell line gene expression data for synergy prediction.

Main Results:

• DSimSynergy consistently outperformed state-of-the-art methods across multiple independent datasets.
• Identified key genes and pathways contributing to drug synergy through model interpretability.
• Validated clinical translational potential and discovered 'pathway complementary networks' driving synergy.

Conclusions:

• DSimSynergy offers an efficient computational approach for identifying synergistic drug combinations.
• The framework aids in overcoming drug resistance and guiding personalized cancer treatment.
• Elucidating molecular mechanisms of synergy facilitates the development of novel therapeutic strategies.