BioERP

BioERP predicts relationships among biomedical entities by applying self-supervised representation learning to biomedical heterogeneous networks (BioHNs) to improve the accuracy of entity relationship predictions.

Key Features:

• Self-Supervised Meta-Path Detection: Employs a self-supervised meta-path detection mechanism to learn structural and semantic context within BioHNs.
• Deep Transformer Encoder: Trains a deep Transformer encoder to capture global structural and semantic features from BioHNs.
• Dual-Level Representation Vectors: Integrates representations from different task-specific models to generate dual-level vectors encoding both local and global associations.
• Biomedical Entity Mask Learning Strategy: Implements a biomedical entity mask learning strategy to model local associations of vertices within the network.
• Concatenation of Representations: Concatenates outputs from task-specific models to construct robust representation vectors for relationship prediction.

Scientific Applications:

• Entity Relationship Prediction in BioHNs: Applies to predicting relationships among biomedical entities within heterogeneous networks, including drug–target interactions.
• Drug–Target Interaction Prediction: Demonstrated near-perfect AUC and AUPR (close to 1) on benchmarks and was evaluated across eight datasets, outperforming 30 state-of-the-art methods.

Methodology:

Uses self-supervised meta-path detection, a deep Transformer encoder, a biomedical entity mask learning strategy, and concatenation of task-specific model outputs to form dual-level representation vectors.