PepMNet: a hybrid deep learning model for predicting peptide properties using hierarchical graph representations

Peptides are a powerful class of molecules that can be applied to a range of problems including biomaterials development and drug design. Currently, machine learning-based property prediction models for peptides primarily rely on amino acid sequence, resulting in two key limitations: first, they are not compatible with non-natural peptide features like modified sidechains or staples, and second, they use human-crafted features to describe the relationships between different amino acids, which reduces the model's flexibility and generalizability. To address these challenges, we have developed PepMNet, a deep learning model that integrates atom-level and amino acid-level information through a hierarchical graph approach. The model first learns from an atom-level graph and then generates amino acid representations based on the atomic information captured in the first stage. These amino acid representations are then combined using graph convolutions on an amino acid-level graph to produce a molecular-level representation, which is then passed to a fully connected neural network for property prediction. We evaluated this architecture by predicting two peptide properties: chromatographic retention time (RT) as a regression task and antimicrobial peptide (AMP) activity as a classification task. For the regression task, PepMNet achieved an average R2 of 0.980 across eight datasets, which spanned different dataset sizes and three liquid chromatography (LC) methods. For the classification task, we developed an ensemble of five models to reduce overfitting and ensure robust classification performance, achieving an area under the receiver operating curve (AUC-ROC) of 0.978 and an average precision of 0.981. Overall, our model illustrates the potential for hierarchical deep learning models to learn peptide properties without relying on human engineering amino acid features.