TransPhos: A Deep-Learning Model for General Phosphorylation Site Prediction Based on Transformer-Encoder...

Area of Science:

• Biochemistry and Molecular Biology
• Computational Biology
• Bioinformatics

Background:

• Protein phosphorylation is a critical post-translational modification in eukaryotes, regulating numerous biological processes.
• Accurate prediction of phosphorylation sites is essential for understanding protein function and signaling pathways.
• Existing computational predictors often rely on domain knowledge or feature selection, limiting their generalizability.

Purpose of the Study:

• To develop a novel deep learning-based predictor, TransPhos, for accurate phosphorylation site prediction.
• To evaluate the performance of TransPhos against established deep learning models and state-of-the-art prediction tools.
• To assess the model's efficacy across different phosphorylation sites (Serine, Threonine, Tyrosine).

Main Methods:

• TransPhos utilizes a transformer encoder combined with densely connected convolutional neural network (CNN) blocks.
• Model performance was evaluated on the PPA (version 3.0) and Phospho.ELM datasets.
• 10-fold cross-validation was employed to assess prediction accuracy.

Main Results:

• TransPhos demonstrated superior performance compared to various deep learning models (CNN, LSTM, RNN, FCNN) and existing tools (GPS2.1, NetPhos, DeepPhos, etc.).
• The model achieved high AUC values: 0.8579 for Serine (S), 0.8335 for Threonine (T), and 0.6953 for Tyrosine (Y).
• TransPhos significantly outperformed competing predictors in general phosphorylation site prediction.

Conclusions:

• TransPhos represents a significant advancement in computational phosphorylation site prediction.
• The deep learning architecture effectively captures complex patterns for improved prediction accuracy.
• This tool offers a robust and accurate solution for researchers studying protein phosphorylation.