DSGAT

DSGAT predicts drug side effect frequencies using Graph Attention Networks applied to drug molecular graphs to support computational drug safety evaluation.

Key Features:

• Graph Attention Networks (GATs): DSGAT employs Graph Attention Networks to model drug molecular structures instead of conventional drug–side effect interaction graphs, producing molecular-level representations for side effect frequency prediction.
• Cold start drugs: DSGAT generates embeddings for drugs absent from training data by leveraging drug molecular graphs, enabling side effect frequency prediction for new compounds.
• Weighted ε-insensitive loss function: DSGAT incorporates a weighted ε-insensitive loss function to mitigate sparsity in interaction-data-based approaches and improve model robustness.
• Benchmark performance: Experimental evaluations report improved predictions for cold start drugs and better performance than existing methods for warm start drugs on benchmark datasets.

Scientific Applications:

• Pharmacovigilance: DSGAT provides predicted side-effect frequency estimates to inform post-marketing safety monitoring and signal assessment.
• Drug development and regulatory assessment: DSGAT supports risk–benefit evaluation during clinical trial planning and regulatory review by estimating side-effect frequencies.
• Novel drug discovery and safety evaluation: DSGAT enables safety profiling of novel compounds lacking prior interaction data.

Methodology:

DSGAT constructs graph representations of drug molecular structures, processes them with Graph Attention Network attention mechanisms, and trains using a weighted ε-insensitive loss function.