GraFSyn: An Interpretable Deep Learning Framework for Anticancer Drug Synergy via Graphlet Fingerprints

Area of Science:

• Computational chemistry
• Pharmacology
• Bioinformatics

Background:

• Accurate prediction of drug synergy accelerates the development of effective anticancer combination therapies.
• Current computational methods often lack structural traceability due to implicit feature aggregation.
• Understanding precise chemical substructure interactions is key to drug synergy.

Purpose of the Study:

• To develop a deep learning framework (GraFSyn) for predicting anticancer drug synergy.
• To enhance structural traceability in computational drug synergy prediction.
• To improve the accuracy of identifying effective anticancer drug combinations.

Main Methods:

• Utilized graphlet fingerprints to encode drugs, preserving chemical substructures and topology.
• Introduced a Dynamic Multi-Scale Convolution (DMSC) module for learning from graphlet features.
• Incorporated an interaction module to model drug substructure and cell line gene expression interplay.

Main Results:

• GraFSyn achieved high performance on benchmark datasets (Merck: ROC-AUC/PR-AUC 0.972/0.912; AstraZeneca: 0.823/0.906).
• Outperformed existing representative baseline methods in synergy prediction.
• Demonstrated structural traceability by mapping signals to specific pharmacophoric regions.

Conclusions:

• GraFSyn offers an accurate and structurally traceable deep learning approach for anticancer drug combination screening.
• The framework facilitates substructure-level analysis of synergistic interactions.
• GraFSyn advances the discovery of novel anticancer combination therapies.