Quantitative prediction of hemolytic activity of peptides

• Published in: Computational toxicology Vol. 32; p. 100335
• Main Authors: Karasev, Dmitry A., Malakhov, Georgii S., Sobolev, Boris N.
• Format: Journal Article
• Language: English
• Published: Elsevier B.V 01.12.2024
• Subjects: Membrane toxicity; Peptide safety; Predictive model; Membrane toxicity; R2; ANN; RF; Predictive model; Peptide safety; kNN; RMSE
• ISSN: 2468-1113; 2468-1113
• DOI: 10.1016/j.comtox.2024.100335

Peptides are currently considered promising therapeutic agents, ranging from antimicrobial to anticancer drugs. Damage to the cell membrane is the most studied mechanism of action of antibacterial peptides. The membrane toxicity of peptides towards human cells is assessed using hemolysis estimation. Several in silico methods have been developed to predict the hemolytic activity of potential antibacterial drugs. Most of the programs use classification models whose results are difficult to interpret. Usually, a researcher does not have the opportunity to understand under what conditions the prediction results can be realized. Furthermore, the authors often use the same external data as training ones not considering the principles of dividing the active and non-active subjects despite that underlying results were obtained under differed conditions. To overcome the gap between the prognosis and real study, we developed the regression models involving the details of differed experimental protocols. We reviewed the literature and supplemented the training data for 951 peptides with quantitative descriptors of the experimental conditions. The resulting regression models predicted the peptide concentration that would cause a certain level of hemolysis at a certain incubation time. Under different validation schemes, our models achieved acceptable performance estimates of 0.69 for R2 and 58 µM for RMSE. Having evaluated the impact of descriptors on model performance, we confirmed the importance of accounting for the experimental conditions for reliable prediction of the peptide membrane toxicity.