همایش ملی بیوانفورماتیک ایران

صفحه اصلی / 4th international edition and 13th Iranian Conference on Bioinformatics

An efficient method based on transformers for antimicrobial peptide prediction

نویسندگان :

Alireza Khorramfard¹ Jamshid Pirgazi² Ali Ghanbari Sorkhi³

1- دانشگاه علم وفناوری مازندران 2- دانشگاه علم وفناوری مازندران 3- دانشگاه علم وفناوری مازندران

کلمات کلیدی :

Antimicrobial peptides،Deep learning،Transformer network،Positional Embedding،Rotary Embedding

چکیده :

Antimicrobial peptides (AMPs) are recognized as a diverse group of membrane-penetrating peptides that play a crucial role in the host's innate defense.[1] Due to their unique properties in combating various microbes, including bacteria, fungi, and viruses, AMPs have garnered significant attention.[2] However, identifying and designing effective AMPs remains challenging due to their structural and functional complexities. Furthermore, laboratory methods for identifying AMPs are often limited by high costs and time-consuming processes. In this context, computational prediction approaches, particularly machine learning-based models, have gained increasing importance. These models enable the identification of AMPs and the prediction of their functions before conducting laboratory experiments. Nevertheless, a major challenge in these methods lies in their low accuracy in identifying and classifying different types of AMPs, especially regarding the complex and nonlinear features of proteins. Existing methods often face issues such as low accuracy in precisely identifying AMPs and categorizing their functions, and they require large and diverse datasets for more accurate training. Moreover, the use of both local and global features in these models has not been fully explored. In this study, we introduce a novel computational approach based on deep learning and transformer networks, called Tra-AMP, for the identification of AMPs and the prediction of their functional types. In this method, amino acid sequences are first tokenized, and an embedding vector is generated for each token. To account for the position of each amino acid, a positional embedding is also added to the token embeddings. A specific type of positional embedding, known as rotary embedding [3], is employed in this paper. The resulting vectors are then input into an enhanced transformer model, where local and global attention mechanisms are used to improve the model's focus on diverse features. Following the feature extraction by the transformer network, classification is performed using fully connected layers. This approach is particularly effective in AMP identification as the model can focus on the complex structural and functional features of proteins. The results of this study demonstrate that the proposed model achieves high performance in identifying AMPs and predicting their functional types, with evaluation metrics such as accuracy (96.74) and F1-score (95.48). These findings indicate that the Tra-AMP method not only accurately identifies AMPs but also excels in classifying their functional types. This method has the potential to significantly aid in the prediction and design of novel AMPs for therapeutic applications.