MVFAN-Kcr: A Multi-View Feature Fusion and Attention-Based Network for Lysine Crotonylation Site Identification

Area of Science:

• Biochemistry and Molecular Biology
• Computational Biology and Bioinformatics
• Epigenetics and Post-Translational Modifications

Background:

• Lysine crotonylation (Kcr) is a critical post-translational modification influencing gene expression and chromatin dynamics.
• Existing Kcr site detection methods suffer from high costs, complexity, false positives, and poor model generalization due to limited data and class imbalance.
• There is a need for accurate, robust, and efficient computational tools for Kcr site identification.

Purpose of the Study:

• To develop an innovative intelligent recognition framework, MVFAN-Kcr, for accurate and robust prediction of lysine crotonylation (Kcr) sites.
• To overcome the limitations of existing Kcr detection techniques and computational models.
• To provide an interpretable and effective tool for Kcr site prediction in proteins.

Main Methods:

• Integration of multi-view feature fusion and attention mechanisms within the MVFAN-Kcr framework.
• Combination of physicochemical properties with global sequence semantics from the ESM-2 protein language model for feature representation.
• Application of analysis of variance for feature selection and a chi-square-based stratified undersampling strategy for data balancing.
• Utilizing a convolutional neural network with attention for efficient pattern extraction and feature enhancement.

Main Results:

• MVFAN-Kcr achieved excellent predictive performance, outperforming baseline approaches with an accuracy (ACC) of 79.18% and an ROC curve area of 0.8618.
• Rigorous evaluation using five-fold cross-validation and an independent test set confirmed the model's robustness and accuracy.
• SHAP and gradient-based analyses highlighted the model's reliance on pKa, charge properties, and specific residue characteristics, demonstrating interpretability.

Conclusions:

• MVFAN-Kcr effectively combines data balancing, multi-view feature fusion, and attention mechanisms to provide high accuracy, robustness, and interpretability in Kcr site prediction.
• The developed framework offers a significant advancement over existing methods for identifying lysine crotonylation sites.
• The study provides accessible data, code, and a web platform for broader application and research in protein Kcr site prediction.