A deep learning-based method for predicting the frequency classes of drug side effects based on multi-source...

Area of Science:

• Pharmacovigilance
• Computational Chemistry
• Bioinformatics

Background:

• Drug side effects are adverse reactions unrelated to therapeutic goals.
• Accurately predicting drug side effect frequency is crucial for safe medication use and drug development.
• Existing computational methods for drug side effect prediction often suffer from overfitting and boundary issues.

Purpose of the Study:

• To develop an advanced computational model for predicting the frequency classes of drug side effects.
• To overcome limitations of existing regression-based models in handling frequency classes and overfitting.
• To introduce a novel approach leveraging multi-source similarity fusion and Bayesian variational inference.

Main Methods:

• Developed a multi-source similarity fusion (MSSF) model.
• Incorporated a multi-source feature fusion module and a self-attention mechanism.
• Utilized Bayesian variational inference for precise prediction of drug side effect frequency classes.

Main Results:

• MSSF demonstrated superior performance over existing models in various evaluation settings.
• Consistent results were observed across cross-validation, cold-start experiments, and independent testing.
• Visual analysis and case studies confirmed MSSF's robust feature extraction and predictive capabilities.

Conclusions:

• The MSSF model offers a significant advancement in predicting drug side effect frequencies.
• Its ability to deeply explore drug-side effect relationships enhances prediction accuracy.
• MSSF shows considerable promise for improving drug safety and informing clinical practice.