Machine learning regression models for predicting hemolytic concentration (HC50) of peptides using curated activity data

Abstract

In this thesis, we first compiled hemolytic activity data of peptides in terms of hemolytic concentration (HC50), defined as the concentration required to lyse 50% of red blood cells (RBCs). We then developed regression models using machine learning techniques to predict HC50 values, which serve as a key indicator of hemolytic potential. This activity data has been integrated into Hemolytik2 (http://webs.iiitd.edu.in/raghava/hemolytik2/), an updated and enhanced version of the Hemolytik database. Hemolytik2 is a manually curated and systematically organized resource that compiles experimentally validated hemolytic peptides from literature and public repositories, including the Antimicrobial Peptide Database (APD), UniProt, and the Dragon Antimicrobial Peptide Database (DAMPD). Over 5,000 of the 13,215 validated peptides in the database have known HC 50 values. Additionally, 2,569 peptides with experimentally established HC50 values against mammalian RBCs were used to train the regression models. With a Pearson correlation coefficient (R) of 0.660 and a coefficient of determination (R2) of 0.408, the top-performing model demonstrated a decent capacity for prediction. All things considered, Hemolytik2.0 is a useful platform for investigating the hemolytic characteristics of peptides and aids in the creation of computational tools meant to create safer and more efficient peptide-based drugs.