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

Area of Science:

• Biotechnology
• Computational Biology
• Pharmacology

Background:

• Peptides are promising drug alternatives but often exhibit hemolytic activity, limiting their therapeutic use.
• Accurate prediction of peptide toxicity against human red blood cells (RBCs) is crucial for drug development.
• Existing methods for assessing hemolytic activity may not be sufficiently accurate or efficient.

Purpose of the Study:

• To develop a deep learning model, BERT-HemoPep60, for quantitative prediction of peptide hemolytic activity.
• To assess the model's performance across various hemolytic measures and species.
• To provide a tool for accurate estimation of peptide hemolytic activity in drug research.

Main Methods:

• Utilized a transformer-based deep learning architecture with domain-adaptive pretraining (DAPT).
• Employed a prefix-prompt approach integrating experimental hemolysis data from six mammalian species.
• Trained and evaluated the model on multiple hemolytic measures (HC$_{5}$, HC$_{10}$, HC$_{50}$) using five-fold cross-validation.

Main Results:

• BERT-HemoPep60 achieved high prediction accuracy, with Pearson Correlation Coefficient (PCC) values of 0.7431 (HC$_{5}$), 0.8088 (HC$_{10}$), and 0.7606 (HC$_{50}$).
• The model demonstrated superior performance compared to traditional machine learning and other deep learning models.
• The model effectively predicts hemolytic activity for peptide sequences up to 60 amino acids.

Conclusions:

• BERT-HemoPep60 accurately predicts peptide hemolytic activity, addressing a critical challenge in peptide drug development.
• This deep learning approach offers a significant advancement over existing methods for toxicity assessment.
• The tool facilitates the identification and development of safer and more effective peptide therapeutics.