Deep learning framework with interpretable feature selection for accurate SUMOylation site prediction.

Area of Science:

• Biochemistry and Molecular Biology
• Computational Biology
• Bioinformatics

Background:

• Small ubiquitin-like modifiers (SUMOs) are critical regulators of diverse biological processes.
• SUMOylation, a key post-translational modification (PTM), significantly impacts cellular regulation.
• Accurate prediction of SUMOylation sites is essential for understanding protein function.

Purpose of the Study:

• To develop a deep learning model for accurate prediction of SUMOylation sites.
• To integrate protein structural and sequence features for enhanced prediction accuracy.
• To establish a robust computational tool for SUMOylation site analysis.

Main Methods:

• Developed Hybrid-Sumo, a deep learning model integrating Half-Sphere Exposure (HSE), PSSM-DWT, and BERT features.
• Utilized SHapley Additive exPlanations (SHAP) for optimal feature selection.
• Employed a Deep Neural Network (DNN) for classification and validated using 10-fold cross-validation.

Main Results:

• Hybrid-Sumo achieved 99.74% accuracy on benchmark datasets.
• Achieved 96.15% accuracy on balanced and 95.83% on imbalanced independent datasets.
• Demonstrated superior performance over existing models, with significant improvements in training and testing accuracy.

Conclusions:

• Hybrid-Sumo is a highly effective computational tool for predicting SUMOylation sites.
• The model's integration of diverse features enhances prediction accuracy.
• This tool can accelerate research in protein function and post-translational modification analysis.