Background <p>Antimicrobial peptide (AMP) prediction has been extensively studied in recent years. However, many existing models do not fully leverage the intrinsic chemical structures of AMPs, such as atomic composition and sidechain group characteristics. Instead, these models often focus on letter composition, positional encodings, and pre-defined chemical-physical descriptors. The incorporation of non-canonical amino acids, which enhance peptide stability and reduce toxicity, is getting more attention in peptide design. Despite this, they are overlooked in predictive studies, as traditional deciphering methods and single-letter representation systems are inadequate for this task. Even though some efforts have been made to expand current alphabets, these approaches remain insufficient, impeding the development of novel AMPs.</p> Results <p>A novel deep learning model, termed AmpHGT, was developed based on heterogeneous graphs’ representation of peptides. AmpHGT demonstrates competitive performance against current methods on canonical amino acid benchmarks. Notably, AmpHGT is capable of efficiently classifying antimicrobial peptides with non-canonical amino acids, addressing the limitations of traditional feature extraction methods. In addition, this model is adaptable to handling different conformations, sidechains, and backbones (e.g., α, β, γ), demonstrating its potential to enhance the screening and discovery of AMPs containing non-canonical amino acids.</p> Conclusions <p>Our study suggests that AmpHGT is reliable for antimicrobial peptide classification task. It may serve as an efficient primary filter for evaluating thousands of mined peptides and provides a good foundation for future studies aimed at producing peptide antibiotics containing non-canonical amino acids.</p>