HSTrans: Homogeneous substructures transformer for predicting frequencies of drug-side effects

Area of Science:

• Pharmacovigilance
• Computational Chemistry
• Machine Learning

Background:

• Accurate drug-side effect (SE) frequency identification is vital for risk-benefit assessment but limited by clinical trial constraints.
• Existing computational methods struggle with novel drug predictions and effectively capturing drug-SE associations.
• Current approaches often rely heavily on drug-SE interaction graphs or simple feature concatenation, limiting predictive power.

Purpose of the Study:

• To develop a novel computational approach, HSTrans, for accurately predicting drug-side effect (SE) frequencies.
• To address limitations of existing methods in predicting novel drug-SE associations and capturing complex relationships.
• To enhance the assessment of drug risk-benefit profiles through improved SE frequency prediction.

Main Methods:

• HSTrans treats drugs and side effects (SEs) as sets of substructures.
• A transformer encoder is utilized for unified substructure embedding, capturing complex relationships.
• A specialized algorithm extracts drug substructures, and an indicator identifies effective SE-related substructures. Convolutional Neural Networks (CNNs) are employed for association capture.

Main Results:

• The proposed HSTrans method demonstrated superior performance compared to state-of-the-art approaches.
• Experimental results validated the effectiveness of HSTrans in predicting drug-side effect frequencies.
• The approach successfully addresses challenges in novel drug prediction and feature association.

Conclusions:

• HSTrans offers a significant advancement in predicting drug-side effect frequencies.
• The substructure-based approach and transformer encoder architecture improve prediction accuracy.
• This method holds promise for enhancing drug safety monitoring and risk-benefit evaluations.