Protein-ligand interaction prediction based on heterogeneity maps and data enhancement

Area of Science:

• Computational chemistry
• Bioinformatics
• Machine learning

Background:

• Protein-ligand interaction prediction is crucial for drug discovery.
• Existing deep learning models struggle with sparse, imbalanced data and underutilize metadata.
• Computational methods are often intensive and limited in capturing structural dynamics.

Purpose of the Study:

• To develop an advanced deep learning model for accurate protein-ligand interaction prediction.
• To enhance model performance on sparse and imbalanced biological datasets.
• To improve the efficiency and effectiveness of drug discovery and repositioning pipelines.

Main Methods:

• Proposed the Heterogeneous Graph Enhanced Fusion Network (HGEF-Net) model.
• Utilized a heterogeneous information learning module for subgraph analysis and metadata integration.
• Implemented a multi-level contrastive learning strategy for data enhancement.
• Developed a heterogeneous attention framework for feature fusion.

Main Results:

• HGEF-Net demonstrated superior performance compared to state-of-the-art models on BindingDB and Davis datasets.
• Achieved an AUC of 0.826 and AUPRC of 0.811 on BindingDB.
• The data enhancement module significantly improved key metrics (AUC, AUPRC, Precision, Recall) on the Davis dataset.

Conclusions:

• HGEF-Net effectively captures complex biological interactions by integrating heterogeneous graph information and advanced data enhancement techniques.
• The model shows significant potential for accelerating drug discovery by providing accurate and efficient protein-ligand interaction predictions.
• The proposed methods address limitations of existing models in handling sparse and imbalanced biological data.