AFPDeepPred: A Deep Learning Framework for Accurate Identification of Antifreeze Proteins

Antifreeze proteins (AFPs) are essential for the survival of organisms in subzero environments and have significant potential in biomedical and agricultural applications. However, their high sequence diversity poses a significant challenge for accurate high-throughput identification. To address this, we propose AFPDeepPred, a novel multimodal deep learning framework that explicitly integrates global and local sequence information for AFP prediction. Specifically, AFPDeepPred combines evolutionary scale modeling, which captures long-range evolutionary dependencies, with chaos game representation, which highlights local motif distributions by transforming protein sequences into image-like matrices. These heterogeneous features are then fused using a bilinear attention network, allowing high-order interaction across both local and global features. Experimental results demonstrate that AFPDeepPred achieves state-of-the-art performance on reviewed Swiss-Prot-based data sets, with an accuracy of 93.46%. Moreover, the model consistently outperforms existing methods in all evaluation metrics, offering a robust and generalizable solution for AFP identification.