DEML: Drug Synergy and Interaction Prediction Using Ensemble-Based Multi-Task Learning.

Area of Science:

• Computational biology
• Pharmacology
• Artificial intelligence

Background:

• Synergistic drug combinations show promise in cancer treatment.
• Deep learning accelerates drug discovery but often overlooks adverse drug-drug interactions (DDIs).
• Existing methods lack the ability to predict both synergy and potential DDIs simultaneously.

Purpose of the Study:

• To develop a novel deep learning model, DEML, for simultaneous prediction of drug synergy and DDIs.
• To improve the accuracy and safety assessment of drug combinations in cancer therapy.
• To leverage chemical and transcriptomics data for enhanced predictive capabilities.

Main Methods:

• Developed DEML, an ensemble-based multi-task neural network.
• Utilized chemical and transcriptomics data as input features.
• Implemented a hybrid ensemble layer and task-specific fusion with gating for representation learning.
• Optimized five synergy regression tasks, synergy classification, and DDI classification concurrently.

Main Results:

• DEML outperformed state-of-the-art methods in Loewe synergy prediction, improving RMSE by 7.8% and R2 by 13.2%.
• The model effectively addressed the multi-task learning 'seesaw effect' through soft parameter sharing and ensemble learning.
• DEML demonstrated superior ability in predicting high-confidence drug pairs with minimal adverse DDIs.
• No performance loss was observed on other prediction tasks.

Conclusions:

• DEML offers a robust computational approach for identifying effective and safe cancer drug combinations.
• The model's simultaneous prediction of synergy and DDIs provides a significant advancement in drug discovery.
• DEML paves the way for guiding novel combination therapy strategies in cancer treatment.