Imbalanced Phosphorylation Site Prediction Based on Federated Learning

With the development of biotechnology, the number and variety of proteins in databases are increasing rapidly. Post-translational modification (PTM) is a major way to regulate protein function and finding protein phosphorylation sites from a large number of protein sequences is crucial for disease research and drug development. Recently, many scholars have been devoted to developing computational predictors based on deep learning architectures for phosphorylation site prediction, but these methods are usually trained on independent databases. This paper explores a new method, PhosFLID, which uses federated learning to improve the accuracy of protein phosphorylation site prediction models and evaluates the effectiveness of this method on different imbalanced dataset. PhosFLID outperforms current leading methods and can be extended to multiple databases while ensuring the privacy of medical data. The research results show that the protein phosphorylation site prediction model using federated learning demonstrates better performance, improving prediction accuracy and stability.