FA-Amy: An amyloid protein prediction model based on protein pre-trained large models and an attention-fusion...

Amyloid proteins are misfolded proteins that aggregate into insoluble fibrils under specific conditions. Their abnormal accumulation is closely linked to neurodegenerative diseases like Alzheimer's. Accurate prediction of amyloidogenic proteins is thus essential for understanding disease mechanisms and guiding therapeutic development. However, current methods for amyloid protein prediction still face several challenges, such as insufficient feature extraction and the reliance on relatively outdated traditional machine learning techniques. In this study, we present FA-Amy, a novel amyloid protein prediction model based on the ESM C protein language model and an attention fusion mechanism. To overcome the limitations of previous methods that relied on handcrafted features and insufficient utilization of sequence information, we employ the ESM C model to generate deep, comprehensive representations of protein sequences. Additionally, we design a fusion mechanism that integrates global and local attention, enhancing the model's ability to identify crucial regions associated with amyloid aggregation. Results from five-fold cross-validation confirm the complementarity and effectiveness of our multi- attention fusion strategy. FA-Amy demonstrates excellent stability and robustness, and is particularly effective in handling highly imbalanced classification tasks. Notably, the attention mechanism successfully focuses on sequence segments likely to contribute to β-sheet stacking. On an independent test set, FA-Amy consistently outperforms the current state-of-the-art model, achieving an 8.1 % improvement in Matthews Correlation Coefficient. To enhance accessibility, we have built a user-friendly online server for the model, which can be accessed at: http://www.bioai-lab.com/FA-Amy.