iGRLDTI: an improved graph representation learning method for predicting drug-target interactions over heterogeneous...

Area of Science:

• Bioinformatics
• Computational Drug Discovery
• Machine Learning

Background:

• Drug-target interactions (DTIs) are crucial for novel drug discovery.
• Computational methods offer efficient and cost-effective DTI prediction.
• Graph neural networks (GNNs) show promise but suffer from over-smoothing in heterogeneous biological information networks (HBINs).

Purpose of the Study:

• To propose an improved graph representation learning method, iGRLDTI, for enhanced DTI prediction.
• To address the over-smoothing issue in GNN-based DTI prediction.
• To capture more discriminative representations of drugs and targets.

Main Methods:

• Constructing a heterogeneous biological information network (HBIN) integrating drug and target information.
• Employing a node-dependent local smoothing strategy to mitigate over-smoothing.
• Utilizing a Gradient Boosting Decision Tree classifier for DTI prediction.

Main Results:

• iGRLDTI demonstrates superior performance compared to state-of-the-art methods on benchmark datasets.
• The method successfully alleviates over-smoothing, enhancing feature representation discriminability.
• Case studies confirm iGRLDTI's ability to identify novel DTIs with distinguishable features.

Conclusions:

• iGRLDTI offers an effective solution for accurate DTI prediction by addressing GNN limitations.
• The proposed method enhances the discriminative power of drug and target representations.
• iGRLDTI facilitates efficient and accurate drug discovery through improved DTI prediction.