Ubigo-X: Protein ubiquitination site prediction using ensemble learning with image-based feature representation and...

Area of Science:

• Biochemistry and Molecular Biology
• Bioinformatics and Computational Biology
• Proteomics

Background:

• Accurate identification of ubiquitination sites is essential for understanding biological functions.
• Existing prediction tools may lack sufficient accuracy or generalizability.

Purpose of the Study:

• To develop a novel and accurate protein ubiquitination prediction tool named Ubigo-X.
• To integrate diverse feature types, including sequence, structure, and function, for improved prediction performance.
• To provide a species-neutral tool for ubiquitination site prediction.

Main Methods:

• Developed three sub-models: Single-Type sequence-based features (SBF), k-mer sequence-based features (Co-Type SBF), and structure/function-based features (S-FBF).
• Utilized amino acid composition, AAindex, one-hot encoding, k-mer encoding, secondary structure, solvent accessibility, and signal peptide cleavage sites.
• Employed XGBoost for S-FBF and Resnet34 for image-based SBF and Co-Type SBF, combining models via weighted voting.

Main Results:

• Ubigo-X achieved high performance on independent datasets: AUC of 0.85, ACC of 0.79, and MCC of 0.58 on balanced PhosphoSitePlus data.
• Demonstrated robust performance on imbalanced data (AUC 0.94, ACC 0.85, MCC 0.55) and GPS-Uber data (AUC 0.81, ACC 0.59, MCC 0.27).
• Outperformed existing tools in Matthews correlation coefficient (MCC) and accuracy (ACC) for balanced data.

Conclusions:

• Ubigo-X effectively integrates diverse features using image-based representation and weighted voting for superior ubiquitination prediction.
• The tool shows significant potential as a species-neutral predictor for ubiquitination sites.
• Ubigo-X is accessible online, providing a valuable resource for researchers in the field.