Feature-Optimized LightGBM Ensemble Framework for Interpretable and High-Accuracy Antifreeze Protein Prediction

Antifreeze proteins (AFPs) are known to occur universally in different species of plants, insects, and fish. These antifreeze proteins are capable of inhibiting the growth of ice crystals, helping the species survive under severely low-temperature conditions. Because of the unique properties of these proteins, their applications are diversifying into metabolic engineering, food processing like yogurt production, and cryopreservation. Various researchers have developed different computer-aided approaches to identify antifreeze proteins. However, the current available predictors are limited by low levels of accuracy and lack of strong predictive potential. This necessitates the development of more accurate predictive methods to identify antifreeze proteins. In this paper, a comprehensive study of the different methods available for predicting antifreeze proteins is conducted. A comparative study of different machine models for the identification of antifreeze proteins is presented. The study reveals that the proposed LightGBM classifier with the help of pentamer feature representations surpasses all the other machine learning models. The proposed model reflected a sensitivity of 97.50%, a recall of 98.30%, and an F1-score of 99.10%.