BERT-HemoPep60

A deep learning method based on transformer architecture and Domain Adaptive Pre-Training (DAPT) for quantitative prediction of peptide hemolytic activity against human red blood cells.

Overview

BERT-HemoPep60 is a deep learning method based on transformer architecture and Domain Adaptive Pre-Training (DAPT) for quantitative prediction of peptide hemolytic activity against human red blood cells. The model employs an innovative prefix prompting approach that integrates experimental hemolysis data from six common mammalian species (human, mouse, rat, horse, sheep, and rabbit) and multiple hemolysis indicators (HC5, HC10, HC50) for peptide toxicity prediction.

Figure 1. BERT-HemoPep60 Model Overview

Model Performance

We employed an innovative prefix prompting approach to integrate experimental data from six mammalian species into a single model, significantly improving prediction accuracy. Compared to traditional methods, BERT-HemoPep60 achieves excellent performance across various metrics.

Results Interpretation

BERT-HemoPep60 provides three key hemolytic activity indicators:

  • HC5: Represents the peptide concentration (μM) required to cause 5% hemolysis of red blood cells
  • HC10: Represents the peptide concentration (μM) required to cause 10% hemolysis of red blood cells
  • HC50: Represents the peptide concentration (μM) required to cause 50% hemolysis of red blood cells

Important Note: Lower HC values indicate stronger hemolytic activity (higher toxicity), while higher HC values indicate weaker hemolytic activity (lower toxicity).