Accurate Antifreeze Protein Prediction Using a LightGBM-Based Model

Abstract

Antifreeze proteins (AFPs) play a role in a large number of organisms such as plants, insects, and fish, in which they prevent ice crystal formation and allow them to survive in very cold temperatures. The AFPs can be used in the metabolic engineering, food engineering, yogurt and cryopreservation. Despite the fact that a number of computational tools have been designed to detect AFPs, most of them are ineffective in prediction with the need to have a more precise prediction technique. During this research, we have reviewed the existing predictors of AFP and we have compared the various machine learning models. We find that LightGBM performs better with pentamer-based feature vectors, which is better than other models on benchmark and independent data, with a sensitivity of 97.50, recall of 98.30, and an F1-score of 99.10. We also emphasize effective classifiers that may be used to enhance the generation of new generation predictors to identify fast and reliable AFPs.