DeepLGF

DeepLGF predicts drug-drug interactions by integrating a biomedical knowledge graph (BKG) with fused local features—chemical semantics from drug sequences via natural language processing and biological embeddings from a Biomedical Functional Graph Neural Network (BFGNN)—and global features from knowledge graph embedding methods.

Key Features:

• Biomedical Knowledge Graph Integration: Leverages a biomedical knowledge graph (BKG) to provide comprehensive drug- and biology-related relational context.
• Chemical local information extraction (NLP): Extracts chemical semantics from drug sequences using natural language processing techniques.
• Biological local embeddings (BFGNN): Derives biological local information and embeddings via a Biomedical Functional Graph Neural Network (BFGNN) learning from various functional spaces.
• Global feature extraction (knowledge graph embeddings): Obtains global drug interaction context using knowledge graph embedding methods applied to the BKG.
• Local–Global Feature Fusion: Combines local and global features to capture specific chemical/biological signals and overarching relational patterns.
• Aggregation methods: Implements four distinct aggregating methods to fuse local and global feature sets.
• Deep neural network prediction: Uses aggregated features as input to a deep neural network for training and DDI prediction.

Scientific Applications:

• Multi-task DDI prediction: Evaluated across three distinct drug-drug interaction prediction tasks.
• Comparative evaluation: Demonstrated superior capability compared to other computational methods in DDI prediction.
• Case studies in high-priority therapeutics: Applied to case studies involving cancer-related drugs and COVID-19-related drugs to identify critical interactions.

Methodology:

DeepLGF extracts chemical local information from drug sequences using natural language processing, derives biological local embeddings via a Biomedical Functional Graph Neural Network (BFGNN) learning from various functional spaces, extracts global features using knowledge graph embedding methods on the BKG, fuses features via selected aggregation methods, and trains a deep neural network for prediction.