Related Experiment Videos
LNMGAT: a laplacian regularized pseudo-negative mining graph attention network for robust drug-target interaction
Shuai Guo1,2, Weichi Liu3,4, Jie Zou1,2
1College of Computer and Information Engineering, Inner Mongolia Agricultural University, Hohhot, China.
Frontiers in Bioinformatics
|August 15, 2026
Summary
This study introduces LNMGAT, a novel computational framework for drug-target interaction (DTI) prediction. It effectively identifies reliable pseudo-negative interactions, improving prediction accuracy, especially in cold-start scenarios.
Area of Science:
- Computational biology
- Bioinformatics
- Drug discovery
Background:
- Drug-target interaction (DTI) prediction is crucial for drug discovery but faces challenges due to limited negative interaction data.
- Unlabeled drug-target pairs in databases are often treated as negatives, introducing noise and reducing model reliability, especially in cold-start scenarios.
Purpose of the Study:
- To develop a reliable computational framework for DTI prediction that addresses the issue of negative-label uncertainty.
- To improve the accuracy and robustness of DTI prediction models, particularly for novel drugs and targets.
Main Methods:
- Proposed LNMGAT, a framework combining Laplacian regularized least squares (LapRLS) for pseudo-negative mining with dual graph attention networks (GATs) for representation learning.
- Identified reliable pseudo-negatives from low-confidence unlabeled pairs using LapRLS on similarity graphs.
- Learned drug and target representations using GATs on k-nearest neighbor graphs and predicted interactions via MLP.
Main Results:
- LNMGAT achieved superior performance across multiple benchmarks (Yamanishi, Davis, KIBA, BindingDB), excelling in 10 of 16 evaluation settings for Area Under the Precision-Recall curve (AUPR).
- Demonstrated significant improvements in cold-start settings, achieving the best AUPR in 9 out of 12 cases with notable gains over baselines.
- External validation using DrugBank and molecular docking provided computational support for high-ranked predictions.
Conclusions:
- LNMGAT offers a competitive and interpretable DTI prediction framework that effectively handles negative-label uncertainty.
- The method enhances prediction reliability, particularly in challenging cold-start scenarios, paving the way for more efficient drug discovery pipelines.
- Further experimental validation is recommended for the top-ranked predicted interactions to confirm biological relevance.
