Gradient Boosting Decision Tree-Based Method for Predicting Interactions Between Target Genes and Drugs.

Area of Science:

• Computational drug discovery
• Bioinformatics
• Machine learning in pharmacology

Background:

• Identifying drug-target interactions is crucial for drug discovery but experimentally challenging.
• Computational methods offer a cost-effective alternative, yet often suffer from class imbalance issues.
• Existing methods struggle to fully utilize known and unknown interactions, impacting prediction accuracy.

Purpose of the Study:

• To develop a novel computational method, DTIGBDT, for predicting drug-target interactions.
• To address and mitigate the class imbalance problem inherent in drug-target interaction datasets.
• To improve the accuracy and efficiency of identifying potential drug-target relationships.

Main Methods:

• Constructed a heterogeneous drug-target network incorporating drug chemical structure and target sequence similarities.
• Utilized random walks on the network to capture topological information and update similarities.
• Extracted features from categorized paths between drugs and targets, feeding them into a gradient boosting decision tree model.

Main Results:

• The DTIGBDT method demonstrated superior performance compared to several state-of-the-art drug-target interaction prediction techniques.
• The ensemble learning approach effectively reduced the adverse effects of class imbalance on prediction accuracy.
• Case studies on antipsychotic drugs (Quetiapine, Clozapine, Olanzapine, Aripiprazole, Ziprasidone) validated DTIGBDT's ability to uncover novel interactions.

Conclusions:

• DTIGBDT provides a robust and accurate computational approach for predicting drug-target interactions.
• The method's ability to handle class imbalance makes it a valuable tool for drug discovery pipelines.
• DTIGBDT can significantly aid researchers in identifying novel therapeutic targets and drug candidates.