Amyloid proteins are characterized by β-sheet structures, and their excessive aggregation has been identified as a hallmark of numerous neurological disorders. Therefore, efficient prediction of amyloid proteins is crucial for the early warning and potential treatment of neurodegenerative diseases such as Alzheimer's. In this study, we propose the FA-Amy algorithm, designed to accurately predict amyloid proteins. To capture the global structure and complex patterns of peptide sequences, we leverage large-scale pretrained models for feature encoding. To effectively integrate the rich sequence features extracted by these models, we develop a hybrid attention mechanism that combines both local and global information, thereby enhancing prediction accuracy. FA-Amy outperforms existing amyloid protein prediction algorithms across multiple evaluation metrics, demonstrating substantial improvements in predictive performance. Moreover, the model’s effectiveness has been validated in real-world applications, exhibiting superior efficiency and accuracy compared to conventional amyloid protein identification methods.