<scp>ConsAMPHemo</scp>: A computational framework for predicting hemolysis of antimicrobial peptides based on machine learning approaches

Many antimicrobial peptides (AMPs) function by disrupting the cell membranes of microbes. While this ability is crucial for their efficacy, it also raises questions about their safety. Specifically, the membrane-disrupting ability could lead to hemolysis. Traditionally, the hemolytic activity of AMPs is evaluated through experiments. To reduce the cost of assessing the safety of an AMP as a drug, we introduce ConsAMPHemo, a two-stage framework based on deep learning. ConsAMPHemo performs conventional binary classification of the hemolytic activities of AMPs and predicts their hemolysis concentrations through regression. Our model demonstrates excellent classification performance, achieving an accuracy of 99.54%, 82.57%, and 88.04% on three distinct datasets, respectively. Regarding regression prediction, the model achieves a Pearson correlation coefficient of 0.809. Additionally, we identify the correlation between features and hemolytic activity. The insights gained from this work shed light on the underlying physics of the hemolytic nature of an AMP. Therefore, our study contributes to the development of safer AMPs through cost-effective hemolytic activity prediction and by revealing the design principles for AMPs with low hemolytic toxicity. The codes and datasets of ConsAMPHemo are available at https://github.com/Cpillar/ConsAMPHemo.