Enhancing the identification of malonylation sites using AlphaFold2 and ensemble learning

Area of Science:

• Biochemistry and Bioinformatics
• Computational Biology
• Artificial Intelligence in Medicine

Background:

• Protein malonylation is implicated in diseases such as diabetes and cancer.
• Accurate identification of malonylation sites is crucial for disease mechanism elucidation.
• Traditional experimental methods for site identification are costly and time-consuming.

Purpose of the Study:

• To develop an accurate computational model for predicting protein malonylation sites.
• To leverage complementary protein features and ensemble learning for improved prediction accuracy.
• To provide a user-friendly tool and accessible resources for malonylation site prediction.

Main Methods:

• Utilized AlphaFold2 for protein tertiary structure information.
• Extracted protein sequence and physicochemical features using CTDC, EAAC, and EGAAC.
• Developed a soft voting ensemble model with Bayesian-optimized CatBoost base classifiers.
• Integrated 3D structural features with sequence/physicochemical features.

Main Results:

• The Catsoft_Kmalsite model demonstrated robust performance and generalization ability.
• Achieved high accuracy in cross-validation (e.g., 87.91% ACC) and on an independent test set (e.g., 89.55% ACC).
• Outperformed other state-of-the-art methods across multiple evaluation metrics (AUC, ACC, Sen, Pre, F1, MCC).

Conclusions:

• The Catsoft_Kmalsite model offers a significant advancement in computational prediction of protein malonylation sites.
• The integration of diverse feature types and ensemble strategy enhances prediction accuracy.
• The developed tool and resources facilitate further research into malonylation-related diseases.