HemoNet: Predicting hemolytic activity of peptides with integrated feature learning

Area of Science:

• Computational chemistry
• Bioinformatics
• Drug discovery

Background:

• Quantifying peptide hemolytic activity is vital for discovering therapeutic peptides.
• Computational methods aid in screening peptides by hemolytic activity.
• Existing methods struggle to model N/C-terminal modifications and amino acid types.

Purpose of the Study:

• To develop a novel neural network-based approach, HemoNet, for predicting peptide hemolytic activity.
• To improve the accuracy of computational prediction of hemolytic activity.
• To facilitate the discovery of novel therapeutic peptides.

Main Methods:

• Developed HemoNet, a neural network model for peptide hemolytic activity prediction.
• Incorporated specialized feature embedding and SMILES-based fingerprint representation for N/C-terminal modifications.
• Utilized stratified cross-validation, non-redundant cross-validation, and external validation on clinical antimicrobial peptides.

Main Results:

• HemoNet achieved a significantly higher predictive performance (AUC-ROC of 88%) compared to previous methods (HemoPI and HemoPred, AUC-ROC of 73%).
• The model accurately captures the contextual importance of amino acids and N/C-terminal modifications.
• Validation on external datasets confirmed the robustness and generalizability of HemoNet.

Conclusions:

• HemoNet represents a significant advancement in the computational prediction of peptide hemolytic activity.
• The tool can effectively guide the discovery and screening of novel therapeutic peptides.
• HemoNet's Python implementation is publicly available for research use.