Screening drug-target interactions with positive-unlabeled learning.

Area of Science:

• Biochemistry
• Computational Biology
• Pharmacology

Background:

• Drug-target interactions (DTIs) are vital for drug repositioning.
• Existing databases primarily contain positive DTIs, creating a scarcity of negative samples for computational prediction.
• Random negative sample selection often leads to false positives, hindering accurate DTI prediction.

Purpose of the Study:

• To develop a robust method for identifying negative drug-target interaction (DTI) examples.
• To design a novel framework (PUDTI) for predicting new drug repositioning candidates.
• To enhance the accuracy of computational DTI prediction by addressing the lack of negative samples.

Main Methods:

• Developed NDTISE, a negative sample extraction method based on positive-unlabeled learning.
• Designed the PUDTI framework integrating NDTISE, class probability estimation, and SVM optimization.
• Evaluated NDTISE against random selection and NCPIS; compared PUDTI with six state-of-the-art methods on diverse DTI datasets.

Main Results:

• NDTISE demonstrated superior performance over random selection and slightly outperformed NCPIS in screening negative DTIs.
• PUDTI achieved the highest Area Under the Curve (AUC) across four DTI datasets (enzymes, ion channels, GPCRs, nuclear receptors) compared to six other methods.
• Validated predicted DTIs through independent drug databases and literature mining, confirming PUDTI's predictive power.

Conclusions:

• NDTISE effectively screens strong negative DTI examples, mitigating issues with random selection.
• PUDTI offers a powerful framework for inferring novel drug repositioning candidates with high accuracy.
• The PUDTI framework serves as an effective pre-filtering tool for new drug design and development.