A method for predicting protein lactation modification sites based on small sample learning

Abstract

This invention discloses a method for predicting protein lactation modification sites based on few-shot learning. The method includes: collecting independent lactation modification site data; integrating lactation modification sites encoded with multiple features; training a model using a few-shot learning strategy; and designing a multi-feature hybrid system to collaboratively predict lactation modification sites. This invention constructs positive and negative datasets of lactation modification sites, and uses various types of sequence and structural features for grouped feature encoding. It employs both the SMOTE algorithm and the Random Undersampling (RUS) algorithm to enhance positive data and weaken negative data, respectively. Deep neural networks are used to construct prediction models based on the features (numerical vectors) of the positive and negative data, resulting in multiple prediction models. New features are integrated using the prediction results of each model, and penalized logistic regression is used to construct the final model. This invention can largely overcome the extreme imbalance of lactation modification site data and the overfitting of the trained model, enabling rapid and large-scale identification of lactation modification sites.

Images