AN-IHP: Incompatible Herb Pairs Prediction by Attention Networks.

Area of Science:

• Pharmacology
• Computational Chemistry
• Traditional Chinese Medicine

Background:

• Adverse drug-drug interactions (DDIs) are a significant safety concern in drug development.
• Traditional Chinese Medicine (TCM) combinations, while therapeutically potent, carry risks of incompatible herb pairs (IHPs).
• Existing methods for IHP inference primarily analyze known interactions, leaving undiscovered IHPs unaddressed.

Purpose of the Study:

• To develop a novel deep attention network (AN-IHP) for predicting IHPs.
• To effectively utilize diverse data types for accurate IHP prediction.
• To provide interpretability for IHP analysis at the ingredient level.

Main Methods:

• AN-IHP employs an attention-aggregation block to learn ingredient-level herb features.
• Similarity profiles are used to represent herb efficacy and properties.
• Commonality and specificity constraints enhance feature representations.
• A gated attention unit (GAU) dynamically fuses representations across herb pairs.
• A deep neural network (DNN) predicts IHPs.

Main Results:

• AN-IHP demonstrated superior performance compared to existing methods on the IHPTCM dataset.
• The model provides ingredient-level interpretability for analyzing IHPs.
• The approach is beneficial for guiding wet-lab experiments and predicting DDIs.

Conclusions:

• AN-IHP is an effective deep learning model for predicting incompatible herb pairs in TCM.
• The method enhances the safety of TCM by identifying potential adverse drug-drug interactions.
• AN-IHP offers valuable insights for both computational and experimental research in TCM safety.