Effectively Identifying Compound-Protein Interactions by Learning from Positive and Unlabeled Examples.

Area of Science:

• Bioinformatics
• Computational Chemistry
• Machine Learning

Background:

• Accurate prediction of compound-protein interactions (CPIs) is vital for drug discovery.
• Existing machine learning methods struggle due to the lack of validated negative CPI datasets.
• Randomly selected unknown interactions as negative examples limit prediction performance.

Purpose of the Study:

• To propose a novel method, PUCPI, for CPI prediction using only positive and unlabeled data.
• To address the limitations of existing methods by leveraging PU learning.
• To develop a more reliable approach for identifying potential drug candidates.

Main Methods:

• PUCPI utilizes PU learning (learning from positive and unlabeled data) with biased-Support Vector Machines (SVM).
• Protein domains and compound substructures are extracted as features.
• A tensor product combines compound and protein features for CPI representation.

Main Results:

• PUCPI demonstrates superior performance compared to six typical classifiers and three existing CPI prediction models.
• The method effectively predicts CPIs even without validated negative examples.
• Experimental results validate the efficacy of the PU learning approach for CPI identification.

Conclusions:

• PUCPI offers a robust and effective solution for predicting compound-protein interactions.
• This work pioneers the use of positive and unlabeled data in CPI prediction.
• The PUCPI method holds significant potential for advancing drug design and discovery.