PoGB-pred: Prediction of Antifreeze Proteins Sequences Using Amino Acid Composition with Feature Selection Followed by a Sequential-based Ensemble Approach

Abstract

Background: Proteins contribute significantly in every task of cellular life. Their functions encompass the building and repairing of tissues in human bodies and other organisms. Hence they are the building blocks of bones, muscles, cartilage, skin, and blood. Similarly, antifreeze proteins are of prime significance for organisms that live in very cold areas. With the help of these proteins, the cold water organisms can survive below zero temperature and resist the water crystallization process, which may cause the rupture in the internal cells and tissues. AFP’s have also attracted attention and interest in food industries and cryopreservation. Objective: With the increase in the availability of genomic sequence data of protein, an automated and sophisticated tool for AFP recognition and identification is in dire need. The sequence and structures of AFP are highly distinct, therefore, most of the proposed methods fail to show promising results on different structures. A consolidated method is proposed to produce the competitive performance on highly distinct AFP structure. Methods: In this study, machine learning-based algorithms including Principal Component Analysis (PCA) followed by Gradient Boosting (GB) were proposed to be used for anti-freeze protein identification. To analyze the performance and validation of the proposed model, various combinations of two segments' composition of amino acid and dipeptides are used. PCA, in particular, is proposed for dimension reduction and high variance retaining of data, which is followed by an ensemble method named gradient boosting for modeling and classification. Results: The proposed method obtained the superfluous performance on PDB, Pfam and Uniprot dataset as compared with the RAFP-Pred method. In experiment-3, by utilizing only 150 PCA components a high accuracy of 89.63 was achieved which is superior to the 87.41 utilizing 300 significant features reported for the RAFP-Pred method. Experiment-2 is conducted using two different dataset such that non-AFP from the PISCES server and AFPs from Protein data bank. In this experiment-2, our proposed method attained high sensitivity of 79.16 which is 12.50 better than state-of-the-art the RAFP-pred method. Conclusion: AFPs have a common function with distinct structure. Therefore, the development of a single model for different sequences often fails to AFPs. A robust results have been shown by our proposed model on the diversity of training and testing dataset. The results of the proposed model outperformed compared to the previous AFPs prediction method such as RAFP-Pred. Our model consists of PCA for dimension reduction followed by gradient boosting for classification. Due to simplicity, scalability properties and high performance result our model can be easily extended for analyzing the proteomic and genomic dataset.