Prediction of antimicrobial Peptides with Advanced Neural Networks

نوع: Type: Thesis

مقطع: Segment: masters

عنوان: Title: Prediction of antimicrobial Peptides with Advanced Neural Networks

ارائه دهنده: Provider: Erfan Safari Barzideh

اساتید راهنما: Supervisors: Dr Hassan Khotanlou

اساتید مشاور: Advisory Professors:

اساتید ممتحن یا داور: Examining professors or referees: Dr Muharram Mansourizadeh, Dr Mahlagha Afrasiabi

زمان و تاریخ ارائه: Time and date of presentation: 2025

مکان ارائه: Place of presentation: آمفی تئاتر

چکیده: Abstract: Abstract: Peptides are short chains of amino acids that can exhibit very different biological functions due to their diversity in sequence length, amino acid composition, net charge, hydrophobicity, amphipathy, and structural features. Antimicrobial peptides, as one of the most important groups of therapeutic peptides, often have short lengths, positive charges, and specific patterns of distribution of hydrophilic/hydrophobic regions that enable effective interaction with the membrane of microorganisms; however, this high diversity and sequence similarities among peptides with different functions have made their accurate identification a challenging problem. Despite significant advances in the application of deep learning to predict antimicrobial peptides, this field still faces several fundamental challenges. The most important of these challenges include the limited and heterogeneous peptide datasets, the severe imbalance between active and inactive classes, and the high structural and functional diversity of peptides, which make the learning process of models difficult. In addition, the short peptide sequences and sequence similarity between peptides with different biological functions reduce the discriminatory power of the models. The issue of overfitting in deep models, limited interpretability of outputs, and the inability of some architectures to simultaneously extract local, global dependencies, and structural relationships are also considered important obstacles. These challenges highlight the need to design advanced, multi-modal, attention-based and transfer learning-based deep learning frameworks to identify and predict effective antimicrobial peptides with higher accuracy and stability. In this study, in order to overcome the aforementioned challenges, a multimodal and fusion-oriented deep learning architecture is proposed that simultaneously models the dual nature of peptides as linguistic sequences and structured biological entities. The framework consists of three parallel processing branches that respectively use a transformer-based language model (PeptideBERT) to extract semantic representations of sequences, a transformer-based graph encoder to model the ordinal and topological relationships of amino acids without the need for an explicit 3D structure, and a biochemical and evolutionary feature branch based on One Hat, BLOSUM62, and Z-scale. The resulting representations from these branches are combined through adaptive fusion strategies including concatenation, gating fusion, and mutual attention, and the final fusion vector is sent to a neural classifier for peptide class prediction. The results show that the simultaneous integration of sequence, graph, and explicit biological features leads to a richer representation and significantly improves the accuracy and stability of predicting the properties of therapeutic peptides compared to single-source approaches.