Machine learning-based approaches for ubiquitination site prediction in human proteins

Area of Science:

• Biochemistry and Molecular Biology
• Computational Biology
• Bioinformatics

Background:

• Protein ubiquitination is a crucial post-translational modification (PTM) regulating cellular processes.
• Identifying ubiquitination sites (Ubi-sites) is vital for understanding protein function and regulation.
• Traditional Ubi-site detection methods are costly and time-consuming.

Purpose of the Study:

• To evaluate and compare various artificial intelligence (AI) approaches for computer-aided Ubi-site prediction.
• To identify the most effective computational methods for predicting Ubi-sites in human proteins.
• To develop a benchmark dataset for Ubi-site prediction research.

Main Methods:

• Collected experimentally verified Ubi-sites from the dbPTM database for human proteins.
• Compared ten machine learning (ML) methods: conventional ML, deep learning (DL) sequence-based, and hybrid DL models.
• Evaluated models using standard metrics and a validation strategy, analyzing the impact of amino acid fragment length.

Main Results:

• Deep learning (DL) models significantly outperformed classical ML methods.
• The best DL model achieved an F1-score of 0.902, accuracy of 0.8198, precision of 0.8786, and recall of 0.9147.
• DL model performance positively correlated with amino acid fragment length, indicating whole-sequence analysis is beneficial.

Conclusions:

• AI, particularly DL techniques, shows high potential for accurate Ubi-site prediction.
• The developed benchmark dataset facilitates standardized comparison of prediction methods.
• This research advances the understanding of protein ubiquitination and cellular regulation.