PreMLS: The undersampling technique based on ClusterCentroids to predict multiple lysine sites

Area of Science:

• Biochemistry
• Computational Biology
• Proteomics

Background:

• Post-translational modifications (PTMs) at lysine residues are crucial for protein function and physiological processes.
• Existing research often focuses on single lysine PTMs, neglecting concurrent modifications and leading to data imbalance issues.

Purpose of the Study:

• To develop a classification system for predicting concurrent multiple modifications at a single lysine residue.
• To address the challenge of class imbalance in predicting multiple lysine PTMs.

Main Methods:

• Utilized a multi-label position-specific triad amino acid propensity algorithm for feature encoding.
• Introduced PreMLS, a novel undersampling algorithm (ClusterCentroids based on MiniBatchKmeans) to handle class imbalance.
• Constructed a convolutional neural network for biological sequence analysis to predict multiple lysine modification sites.

Main Results:

• The developed convolutional neural network model significantly outperformed existing methods (iMul-kSite, predML-Site) in predicting multiple lysine modification sites.
• The model demonstrated high accuracy through five-fold cross-validation and independent testing.
• An open-access predictive script was created for enhanced accessibility.

Conclusions:

• The study provides a valuable tool for prioritizing potential lysine modification sites for further biological assays.
• The findings advance the understanding of complex PTMs and support drug development efforts.
• Accurate prediction of concurrent K-PTMs is essential for comprehensive biological research.