Predicting protein functions using positive-unlabeled ranking with ontology-based priors.

Area of Science:

• Bioinformatics
• Computational Biology
• Genomics

Background:

• Automated protein function prediction is vital in bioinformatics.
• Current methods face challenges with multilabel classification and a high number of unlabeled annotations, leading to false negatives.
• Existing approaches often incorrectly classify unlabeled protein functions as negative.

Purpose of the Study:

• To introduce a novel approach, PU-GO, for protein function prediction.
• To address the limitations of existing methods by framing the problem as a positive-unlabeled ranking task.
• To improve the robustness and accuracy of automated protein function prediction.

Main Methods:

• Developed PU-GO, a novel positive-unlabeled (PU) learning approach for protein function prediction.
• Applied empirical risk minimization to minimize classification risk.
• Utilized Gene Ontology (GO) hierarchical structure to obtain class priors.

Main Results:

• PU-GO effectively addresses the false negative issue inherent in protein function prediction.
• The proposed method demonstrates superior robustness compared to state-of-the-art techniques.
• Performance was validated on both similarity-based and time-based benchmark datasets.

Conclusions:

• PU-GO offers a more robust and accurate solution for automated protein function prediction.
• The positive-unlabeled ranking framework is effective for handling imbalanced and noisy biological data.
• This work advances computational methods for understanding protein functions and their biological roles.