(394b) A Computational Meta-Analysis for the Identification of Therapeutic Membrane-Active Peptides | AIChE

(394b) A Computational Meta-Analysis for the Identification of Therapeutic Membrane-Active Peptides

Authors 

Song, H., Inha University
DeFreese, W., Auburn University
Kieslich, C., Auburn University
There have been a growing number of studies reported on membrane active peptides in the last 20 years. These peptides exert their biological activity upon interacting with the cell membrane, either to translocate through it to deliver cargos into the cell and reach their target or to disrupt it and lead to cell lysis. There exist a couple prominent classes of membrane active peptides, namely: cell penetrating peptides (CPPs) and membranolytic or membrane disrupting peptides (MDPs). CPPs can pass through cell membranes and transport to the inside of cells while carrying a wide variety of cargo such as nanoparticles, peptides, proteins, antisense oligonucleotides, small-interfering RNA, double-stranded DNA, and liposomes. Therefore, CPPs represent a significant approach to deliver bioactive molecules into cells for various biomedical applications. On the other hand, MDPs disrupt the membrane integrity or inhibit the cellular functions of bacteria, viruses, fungi, and cancer cells. These peptides are referred to using other names and acronyms in the literature, e.g., cationic antimicrobial peptides (CAPs) or antimicrobial peptides (AMPs). MDPs come from structurally diverse families with varied function and therapeutic activities (i.e., anticancer, antiviral, antibacterial), all of which have numerous advantages over conventional pharmaceuticals that are mainly based on small molecules and antibodies. However, a number of these peptides are prevented from entering clinical trials due to their high hemolytic activity. Therefore, we performed a comprehensive computational analysis using machine learning to investigate the structure-function relationships of different categories of MDPs and CPPs. To this end, we developed accurate models for predicting membrane-active peptides function, by training support vector machines (SVMs). This was achieved by collecting various datasets of each of the classes and subclasses of membrane-active peptides mentioned above from publicly available databases. Additionally, feature selection has been performed to further optimize our SVM models. Monte Carlo cross-validation classification has been applied to train and tune the models based on multiple train/test sets of each dataset. Afterwards, different evaluation metrics have been calculated to quantify our models performance and to compare our predictions with the recently developed models. Among them, we modified and curated the imbalanced datasets or the ones with duplications and then reapplied the whole procedure to make the predictions as accurate and reliable as possible with the ultimate goal of designing most effective therapeutic peptides for clinical trials.

Keywords: Membrane-active peptides, Computer-aided drug design, Sequential feature representation, BLOSUM, feature selection.