AFP-MCDF: Multi and cross-dimensional feature fusion methods for antifreeze protein prediction

Abstract

Antifreeze proteins can effectively inhibit the formation of ice crystals and enhance cell survival in low-temperature environments. They protect the texture prolong the shelf life of food and maintain cell and tissue integrity in medical treatments, thereby improving the success rate of surgery and transplantation. Accurate prediction of Antifreeze proteins is important to advance these fields. Traditional wet-experiment methods, while providing reliable validation results, are usually time-consuming and costly. And existing computational methods still have room for improvement in predicting performance. In this study, a novel antifreeze protein prediction method, AFP-MCDF, is proposed. The AFP-MCDF method first extracts one- and two-dimensional feature representations of Antifreeze protein sequences using the pre-trained protein language models ProtBERT and ESM-2. Subsequently, these features are fused multidimensionally via BiLSTM and TextCNN to capture long-term dependencies and local features. Finally, the method predicts the frost resistance of Antifreeze protein sequences by cross-dimensional fusion and linear mapping from N to 2 dimensions. Experimental results show that AFP-MCDF performs well in the antifreeze protein prediction task, outperforming traditional computational methods and reaching the current state-of-the-art.