深度学习驱动的肽段从头测序模型演进研究OA

We systematically review recent advances in deep learning techniques for de novo peptide sequencing,with a specific focus on the model evolution and performance optimization of three mainstream architectures:convolutional neural networks(CNN),graph convo-lutional networks(GCN),and Transformers.Through longitudinal comparisons of representative model architectures across different de-velopmental stages and transversal analysis of technical characteristics among different models,we systematically elucidate the pivotal role of deep learning in enhancing the accuracy and efficiency of de novo peptide sequencing.Experimental validation demonstrates that CNN models demonstrate significant advantages in processing speed owing to their efficient local feature extraction capability;GCN models enhance cross-species generalization ability by modeling topological relationships among spectral peaks;while the Transformer architecture excels in sequencing long peptides and handling non-canonical fragmentation patterns,owing to its global dependency modeling and parallel decoding mechanism.This study provides effective theoretical support and valuable technical references for database-independent peptide identification in proteomics and identifies preliminary directions for future model optimization.