NphosNet: Predicting Protein N-Phosphorylation Sites via xLSTM and Enhanced PLM Features with a Weighted...

Area of Science:

• Biochemistry and Molecular Biology
• Computational Biology
• Genomics and Proteomics

Background:

• Protein phosphorylation is crucial for cellular signaling and disease.
• N-phosphorylation, a less-studied modification, faces challenges in site stability and detection.
• Existing methods for N-phosphorylation site prediction are limited.

Purpose of the Study:

• To develop a robust deep learning framework, NphosNet, for accurate N-phosphorylation site prediction.
• To address challenges in N-phosphorylation site identification, including data imbalance and feature representation.
• To provide a computational tool for understanding N-phosphorylation in biological processes and diseases.

Main Methods:

• Construction of a novel, class-imbalanced N-phosphorylation dataset (pH-913, pK-2060, pR-1700).
• Hybrid embedding strategy using amino acid tokenization, positional encoding, and pre-trained models (ProtT5, EMBER2).
• A three-branch deep learning architecture integrating Transformers, xLSTM, CNN, and attention-enhanced ResNet, with a cross-attention mechanism for feature fusion.

Main Results:

• NphosNet achieved superior performance in N-phosphorylation site prediction.
• AUC values for NphosNet were 0.9227 (pH), 0.9099 (pK), and 0.9377 (pR).
• NphosNet significantly outperformed existing methods for N-phosphorylation site prediction.

Conclusions:

• NphosNet offers a powerful computational solution for N-phosphorylation site prediction.
• The framework enhances the study of N-phosphorylation's role in cellular mechanisms and disease pathogenesis.
• This advancement facilitates deeper insights into post-translational modifications and their biological implications.