DeepSTFSynergy: A multi-scale structural information fusion method for personalized drug combination prediction

Area of Science:

• Computational biology
• Drug discovery
• Bioinformatics

Background:

• Predicting drug synergy is vital for personalized cancer combination therapies.
• Current methods struggle with multi-scale structural information from drugs and cell lines, limiting understanding of synergistic mechanisms.

Purpose of the Study:

• To develop DeepSTFSynergy, a novel framework for personalized drug combination prediction.
• To effectively model multi-scale structural interactions between drugs and cell lines for synergy prediction.

Main Methods:

• Proposed DeepSTFSynergy, a multi-scale structural information fusion framework.
• Utilized three parallel attention-based subnetworks for atomic, sub-structural, and global drug feature extraction.
• Implemented a cell line-specific cross-modal fusion mechanism with gating units for dynamic information integration.

Main Results:

• DeepSTFSynergy demonstrated superior performance in both regression and classification tasks compared to existing methods.
• Predicted novel drug combinations aligned with previous studies.
• Visualization analysis identified synergistic atomic structures and substructures, providing interpretability.

Conclusions:

• DeepSTFSynergy effectively predicts drug synergy by integrating multi-scale structural information.
• The framework offers an interpretable approach to understanding synergistic mechanisms.
• This aids in personalized cancer treatment decision-making.