Predicting Drug-Target Interactions via Within-Score and Between-Score.

Area of Science:

• Computational Biology
• Bioinformatics
• Drug Discovery

Background:

• Network inference and local classification models aid in predicting drug-target interactions (DTIs).
• Existing methods face challenges with isolated subnetworks, biased classifiers due to limited positive samples, and the need for multiple local classifiers.
• There is a need for improved approaches to predict novel DTIs and understand relationships between known and unknown drug-target pairs (DTPs).

Purpose of the Study:

• To develop a more effective approach for predicting potential drug-target interactions (DTIs).
• To address limitations of existing methods in DTI prediction, including issues with isolated subnetworks and biased classifiers.
• To establish a visualized relationship between known DTIs and unapproved DTPs.

Main Methods:

• Improved drug and target similarity measures were developed.
• Each drug-target pair (DTP) was characterized by a feature vector using within-scores and between-scores.
• A single global classifier was trained on these feature vectors.

Main Results:

• The proposed method generates a uniform vector representation for all DTPs.
• A single global classifier demonstrated reduced bias, benefiting from sufficient positive samples.
• The approach successfully visualized the relationship between known DTIs and unapproved DTPs.
• The method's effectiveness was validated through cross-validation and prediction in existing databases, outperforming popular methods.

Conclusions:

• The developed method offers a superior approach to DTI prediction compared to existing techniques.
• This approach enhances drug discovery and repositioning by providing more accurate predictions and clearer insights into DTI relationships.
• The characterized feature vectors and global classifier offer a robust and less biased framework for DTI prediction.