BERT-HemoPep60: Transformer-based Deep Learning Method with Domain-Adaptive Pretraining for Quantitative Hemolytic Activity Prediction of Peptides.

Abstract

Peptides have emerged as a promising alternative to traditional small-molecule drugs and therapeutic proteins for treating various diseases. However, their potential therapeutic use is often limited by their inherent hemolytic activity. Consequently, quantitative assessment of peptide toxicity against human red blood cells (RBCs) is crucial for peptide research. Here we present BERT-HemoPep60, a transformer-based deep learning method using domain-adaptive pretraining (DAPT) to quantitatively predict the hemolytic activity of peptides toward human RBCs for sequences up to 60 amino acids. Our prefix-prompt approach integrates experimental hemolysis data from six common mammalian species (human, mouse, rat, horse, sheep, rabbit) across multiple hemolytic measures (HC$_{5}$, HC$_{10}$, HC$_{50}$). In a comprehensive evaluation using five-fold cross-validation, our model achieved PCC values of 0.7431, 0.8088, and 0.7606 for HC$_{5}$, HC$_{10}$, and HC$_{50}$ predictions, respectively. These results outperform traditional machine learning and deep learning models based on various sequence encoding methods. BERT-HemoPep60 enables accurate estimation of peptide hemolytic activity for drug research.