De novo peptide sequencing is important for proteomics to predict unknown peptide sequences based on mass spectrometry data. Traditional methods for peptide sequence prediction have certain limitations because they often rely on handcrafted feature extraction and statistical models. To enhance prediction accuracy and rationality, a novel de novo peptide prediction model based on Transformer and Graph Convolutional Networks (GCN) is designed. In this model BERT is firstly used to extract peptide embedding features. These extracted features are integrated with physicochemical characteristics such as precursor mass, mass-to-charge ratio (m/z), and intensity to improve global feature representation. A model of the graph structural relationships between peptides using GCN can capture complex interactions to further enhance the capability of representation. Finally, the autoregressive decoder of a GPT-2 generation module is used to generate peptide sequences. And some strategies such as Top-k sampling, Top-p sampling, and temperature control are incorporated to ensure the rationality and diversity of the generated sequences. Experimental results demonstrate that the proposed model achieves an average improvement of 5.29% in amino acid recall and 4.18% in accuracy across multiple datasets compared to DeepNovo. The model effectively enhances the accuracy of de novo peptide sequencing.