DeepMPF

DeepMPF predicts drug-target interactions by integrating multi-modal representations and meta-path semantic analysis on heterogeneous biological networks to prioritize candidate protein–drug pairs for drug discovery.

Key Features:

• Heterogeneous network construction: Constructs a protein–drug–disease heterogeneous network that represents interactions among proteins, drugs, and diseases.
• Multi-modal feature extraction: Extracts features under three views—sequence modality, heterogeneous structure modality, and similarity modality.
• Meta-path semantic analysis: Employs six representative meta-path schemas to capture high-order nonlinear structural information in the heterogeneous network.
• Comprehensive feature descriptors and joint learning: Uses joint learning to generate comprehensive feature descriptors and to compute interaction probabilities for DTI prediction.

Scientific Applications:

• Benchmark evaluation: Demonstrated competitive performance across four gold-standard datasets.
• Drug repositioning: Applied to drug repositioning experiments for diseases including COVID-19 and HIV.
• Molecular docking validation: Supported predicted drug candidates with molecular docking experiments.

Methodology:

Construct a protein–drug–disease heterogeneous network, extract features from sequence, heterogeneous structure, and similarity modalities, apply meta-path semantic analysis using six meta-path schemas, and perform joint learning to generate comprehensive feature descriptors and predict drug-target interaction probabilities.