CPI-IGAE

CPI-IGAE predicts compound-protein interactions (CPIs) to identify potential molecular interactions relevant to drug discovery.

Key Features:

• Inductive Graph AggrEgator (IGAE): learns low-dimensional representations of compounds and proteins from a homogeneous graph in an end-to-end manner.
• Heterogeneous-to-homogeneous graph transformation: converts complex heterogeneous graphs into homogeneous graphs for downstream graph-based learning.
• Ligand-based protein representations: integrates ligand-based protein representations into the graph representation.
• Overall similarity associations: incorporates overall similarity associations between entities into the homogeneous graph.
• Graph representation learning: employs inductive graph neural network methods for representation learning of molecular entities.
• Embedding visualization: provides visualization of learned embeddings to inspect feature separation and structure.
• Ablation studies: uses ablation analyses to evaluate contributions of model components.
• Empirical performance: reports superior performance relative to several state-of-the-art CPI prediction methods and identifies top-ranked CPIs corroborated by literature.

Scientific Applications:

• CPI prediction for drug development: prioritizes compound–protein pairs for target identification and lead selection.
• Representation learning for molecular entities: generates compact embeddings of compounds and proteins for downstream analysis.
• Prediction validation and prioritization: produces ranked CPI predictions that can be cross-referenced with literature for validation.

Methodology:

Transforms heterogeneous graphs into homogeneous graphs by integrating ligand-based protein representations and overall similarity associations, and applies the Inductive Graph AggrEgator (IGAE) inductive graph neural network to learn low-dimensional compound and protein representations end-to-end, with embedding visualization and ablation studies for evaluation.