LapGAT: A Semi-Supervised Learning Framework for Drug-Target Interaction Prediction.

Area of Science:

• Computational chemistry and pharmacology
• Bioinformatics and computational biology
• Drug discovery and development

Background:

• Drug-target interaction (DTI) prediction is crucial for identifying new therapeutics but faces challenges like data scarcity and poor generalization.
• Existing DTI prediction methods struggle with reliable negative sample generation and diverse biological contexts.

Purpose of the Study:

• To develop a robust and generalizable semi-supervised framework for accurate drug-target interaction prediction.
• To overcome limitations in data scarcity and negative sample selection for DTI prediction.
• To accelerate the identification of novel therapeutic candidates and drug repositioning.

Main Methods:

• Proposed LapGAT, a framework combining graph-enhanced Laplacian regularized least squares (LapRLS) and a multilayer graph attention network (GAT).
• LapRLS fuses similarity matrices and generates high-confidence positive/negative samples in the upstream stage.
• GAT learns from pseudo-labeled interactions, capturing local structures and nonlinear dependencies in the downstream stage.

Main Results:

• LapGAT demonstrated robust performance across four target categories: enzymes, GPCRs, ion channels, and nuclear receptors.
• Molecular docking confirmed the physical plausibility of predictions, with binding affinities between -4.5 to -7.3 kcal/mol.
• Literature-based validation achieved high accuracies: 81.3% (enzymes), 100% (GPCRs), 71.4% (ion channels), and 88.9% (nuclear receptors).

Conclusions:

• LapGAT provides a scalable and flexible computational tool for accelerating drug discovery.
• The framework effectively addresses data scarcity and improves generalization in DTI prediction.
• LapGAT holds potential for broad applicability across various therapeutic domains, aiding in identifying novel drug candidates.