DeepNeoAGNet: Enhancing Cancer Immunotherapy with General Heterogeneous Sequence Learning for Precise Neoantigen Prediction and Immunogenicity Assessment

Multi-Peptide Personalized drugs have emerged as a prominent focus in pharmaceutical research owing to their high specificity, low toxicity, and excellent biocompatibility. They are increasingly being applied across various fields, including peptide-drug conjugates (PDCs), infectious disease vaccines, and anti-aging interventions. Immunotherapy, leveraging the immune system to combat cancer, increasingly relies on identifying neoantigens—tumor-specific proteins arising from mutations. These neoantigens, pivotal for personalized cancer vaccines, present computational challenges in their identification and binding affinity prediction to T-cell receptors. This study introduces DeepNeoAGNet, a novel sequence deep learning framework designed for gene sequence prediction. DeepNeoAGNet utilizes bidirectional long-short term memory (Bi-LSTM) neural networks to extract important features from peptide sequences by incorporating both global and local bioengineer information. Our approach incorporates new analysis techniques, including comprehensive amino acid feature engineering such as global amino acid composition strategy, dipeptide composition, local amino acid composition, and local dipeptide composition. Inter-residue sequence relationship analysis is employed to capture intricate dependencies within the sequences. In both the training and testing phases, our approach achieves higher R-square values, indicating a correlation enhancement of 46.466%, and lower RMSE, achieving an accuracy improvement of 11.147%, compared to HLAIImaster and TripHLApan models. These advancements indicate the possibility of facilitating precise identification of HLA-peptide binding, offering researchers and clinicians with a more effective and general tool for neoantigen prediction. By leveraging both global and local features within a BiLSTM architecture, DeepNeoAGNet offers interpretable prediction. Furthermore, it serves as a valuable domain-specific knowledge resource for large language models (LLMs).