NASNet-DTI: accurate drug-target interaction prediction using heterogeneous graphs and node adaptation

Area of Science:

• Computational biology
• Drug discovery
• Bioinformatics

Background:

• Drug-target interactions (DTIs) are crucial for drug development, but experimental prediction is slow and resource-intensive.
• Existing deep learning methods for DTI prediction often neglect relational features and suffer from over-smoothing in graph neural networks (GNNs).

Purpose of the Study:

• To develop an advanced framework, NASNet-DTI, for accurate and efficient prediction of drug-target interactions.
• To address the limitations of existing DTI prediction methods, specifically the neglect of relational features and the over-smoothing problem in GNNs.

Main Methods:

• NASNet-DTI utilizes graph convolutional networks (GCNs) to extract features from drugs and targets separately.
• It constructs heterogeneous networks representing drugs and targets with multiple relationship types (drug-drug, target-target, drug-target).
• A node adaptive learning strategy dynamically determines optimal aggregation depth for each node, mitigating over-smoothing.

Main Results:

• NASNet-DTI demonstrated significantly superior performance compared to existing methods across multiple datasets.
• The node adaptive learning strategy effectively alleviated the over-smoothing problem inherent in GNNs.
• The framework successfully integrated intrinsic and relational features for improved DTI prediction.

Conclusions:

• NASNet-DTI offers a powerful and effective approach for predicting drug-target interactions, advancing drug discovery and development.
• The proposed node adaptive learning strategy provides a robust solution to the over-smoothing challenge in GNN-based DTI prediction.
• This framework holds significant potential for accelerating the identification of novel therapeutic agents.