DephosNet: A Novel Transfer Learning Approach for Dephosphorylation Site Prediction

Abstract

Protein dephosphorylation is the process of removing phosphate groups from protein molecules, which plays a vital role in regulating various cellular processes and intricate protein signaling networks. The identification and prediction of dephosphorylation sites are crucial for this process. Previously, there was a lack of effective deep learning models for predicting these sites, often resulting in suboptimal outcomes. In this study, we introduce a deep learning framework known as “DephosNet”, which leverages transfer learning to enhance dephosphorylation site prediction. DephosNet employs dual-window sequential inputs that are embedded and subsequently processed through a series of network architectures, including ResBlock, Multi-Head Attention, and BiGRU layers. It generates predictions for both dephosphorylation and phosphorylation site probabilities. DephosNet is pre-trained on a phosphorylation dataset and then fine-tuned on the parameters with a dephosphorylation dataset. Notably, transfer learning significantly enhances DephosNet’s performance on the same dataset. Experimental results demonstrate that, when compared with other state-of-the-art models, DephosNet outperforms them on both the independent test sets for phosphorylation and dephosphorylation.