A complete graph-based approach with multi-task learning for predicting synergistic drug combinations.

Area of Science:

• Computational biology
• Drug discovery
• Artificial intelligence in medicine

Background:

• Drug combination therapy offers advantages over monotherapy for cancer treatment.
• Experimental screening of drug combinations is challenging due to the vast combinatorial space.
• Current deep learning methods for drug synergy prediction suffer from instability and limited generalization.

Purpose of the Study:

• To address the instability and generalization limitations of existing deep learning models for drug combination prediction.
• To develop a computational method for identifying novel synergistic drug combinations.
• To improve the reliability and applicability of AI in drug discovery.

Main Methods:

• Proposed CGMS, a model that represents drug combinations and cell lines as heterogeneous graphs.
• Utilized heterogeneous graph attention networks to generate whole-graph embeddings for interaction characterization.
• Employed multi-task learning, training simultaneously on drug synergy and drug sensitivity prediction tasks to enhance generalization.

Main Results:

• CGMS provides stable, order-independent predictions of drug synergy.
• Demonstrated superior generalization ability compared to six state-of-the-art methods across various cross-validation scenarios (leave-drug combination-out, leave-cell line-out, leave-drug-out).
• Validated the effectiveness of whole-graph embeddings, attention mechanisms, and multi-task learning in improving prediction accuracy and stability.

Conclusions:

• CGMS offers a robust and generalizable deep learning framework for predicting synergistic drug combinations.
• The model's stability and order-independent nature make it a reliable tool for pre-screening potential cancer therapies.
• CGMS advances the application of AI in accelerating drug discovery and optimizing combination therapy strategies.