STNN-DDI: a Substructure-aware Tensor Neural Network to predict Drug-Drug Interactions

Area of Science:

• Computational chemistry
• Pharmacology
• Artificial intelligence in drug discovery

Background:

• Drug-drug interactions (DDIs) pose significant risks in poly-drug therapies, necessitating accurate prediction methods.
• Existing computational approaches for DDI prediction often lack interpretability and do not identify specific drug substructures responsible for interactions.

Purpose of the Study:

• To develop a novel computational model, STNN-DDI, for predicting multiple types of DDIs.
• To improve the interpretability of DDI prediction by focusing on local chemical substructures.
• To provide a unified framework for both transductive and inductive DDI prediction.

Main Methods:

• Designed a Substructure-aware Tensor Neural Network (STNN-DDI) model.
• Learned a 3-D tensor representing substructure-substructure interaction (SSI) space.
• Mapped drugs into the SSI space using predefined substructures for DDI prediction.

Main Results:

• STNN-DDI demonstrated superior performance compared to state-of-the-art deep learning baselines, significantly improving AUC, AUPR, Accuracy, and Precision.
• Case studies confirmed the model's interpretability by identifying key substructure pairs involved in specific DDIs.
• The model effectively predicts DDIs and explains underlying interaction mechanisms.

Conclusions:

• STNN-DDI offers an effective and interpretable approach for multiple-type drug-drug interaction prediction.
• The model's ability to identify causative substructures enhances understanding of drug interaction mechanisms.
• This work contributes to safer poly-drug treatments by improving DDI prediction and interpretability.