MULGA

MULGA predicts drug–protein interactions and supports drug discovery and repositioning by integrating multi-view learning, a graph autoencoder, and guilty-by-association negative sampling to infer missing DPIs and learn affinity matrices for drugs and target proteins.

Key Features:

• Multi-View Learning Technique: Employs multi-view learning to capture and integrate diverse data representations for accurate estimation of affinity matrices of drugs and target proteins.
• Graph Autoencoder Architecture: Uses a graph autoencoder to encode biological data into a lower-dimensional space and reconstruct it to infer missing DPI interactions and predict potential drug-target associations.
• Guilty-by-Association-Based Negative Sampling: Applies a guilty-by-association strategy to select highly reliable non-DPIs as negative samples, improving training robustness and prediction reliability.

Scientific Applications:

• Drug discovery and repurposing: Identifies candidate drug-target associations to support drug discovery and repositioning efforts.
• Benchmarking and validation: Demonstrated superior performance over state-of-the-art methods in DPI prediction through benchmark experiments and validated component contributions via ablation studies.
• SAR-CoV-2 spike glycoprotein targeting: Applied to identify potential drugs targeting the spike glycoprotein of severe acute respiratory syndrome coronavirus 2 (SAR-CoV-2).

Methodology:

Computational methods explicitly include multi-view learning for affinity estimation, a graph autoencoder for encoding and reconstruction to infer missing DPIs, guilty-by-association negative sampling for selecting non-DPIs, extendable and unbiased similarity calculation, and heterogeneous information integration within a unified framework.