Machine Learning Meta-analysis of Proteolytic Cleavage Specificity

Aug 4, 2026

Proteolytic enzymes such as cathepsins and matrix metalloproteinases (MMPs) govern metabolism, cell signaling, and apoptosis, making cleavage-site identification a complex problem driven by diverse substrate specificities and regulatory context. This thesis applies machine learning models, notably Support Vector Machines (SVMs) and One-Class SVMs (OCSVMs), to predict proteolytic cleavage specificity. A Fourier Transform-based peptide encoding feeds SVM feature selection for a cross-protease meta-analysis using MEROPS-derived positive and synthetic negative datasets. Cross-validation and hyperparameter tuning yield near-perfect AUC-ROC for multiple proteases, improving biomarker and therapeutic insight.

LanguageEnglish

CategoryResearch & Report

FormatPDF

Pages80

Reading Time3 h 22 min

AccessFree to download

Abstract

Acknowledgments

List of Figures

List of Tables

1 Introduction

1.1 Overview of Vaccines

1.2 Advancements in Peptide-based Vaccines

1.3 Mechanism of Immune Response

1.4 Role of Proteases in the Immune System

1.5 Research Challenges

1.6 Objectives of the Study

2 Background Information

2.1 Specific Types of Proteases

2.2 Overview of Peptidase Databases: Focus on MEROPS

2.3 Foundations of Machine Learning in Biomedical Research

2.4 Kernel Methods

Machine Learning Meta-analysis of Proteolytic Cleavage Specificity by Suhyeon Kim A thesis submitted to the Graduate Faculty of Auburn University in partial fulfillment of the requirements for the Degree of Master of Science Auburn, Alabama August 3, 2024 Keywords: Proteolytic Cleavage Specificity, Machine Learning, Feature Selection Copyright 2024 by Suhyeon Kim Approved by Christopher Kieslich, Chair, Assistant Professor of Chemical Engineering Selen Cremaschi, B. Redd & Susan W. Redd Eminent Scholar Chair Professor of Chemical Engineering Peter He, George E. & Dorothy Stafford Uthlaut Endowed Professor of Chemical Engineering Abstract Proteolytic enzymes, such as cathepsins and matrix metalloproteinases (MMPs), play crucial roles in various physiological processes, including metabolism, cell signaling, and apoptosis. Identifying their cleavage sites is a complex challenge due to the diverse substrate specificities and regulatory mechanisms of these enzymes. This thesis investigates the use of machine learning models, particularly Support Vector Machines (SVMs) and One-Class Support Vector Machines (OCSVMs), to predict proteolytic cleavage specificity. The study introduces a novel approach utilizing Fourier Transform-based encoding of peptide sequences to capture essential biochemical properties and structural characteristics, which are used as inputs into SVM algorithms. The research encompasses a comprehensive meta-analysis using SVM-based feature selection techniques to compare and contrast the substrate specificity of different proteases. This analysis aims to uncover distinct patterns in substrate interaction, offering valuable insights for therapeutic strategies and biomarker discovery. The datasets used in this study were sourced from the MEROPS database and included both positive data points (cleaved sequences) and synthetic negative data points (non-cleaved sequences) to ensure robustness and diversity. Through rigorous cross-validation and hyper-parameter optimization, the SVM models demonstrated high predictive accuracy, achieving Area Under the Receiver Operating Characteristic (AUC-ROC) scores close to 1 .00 for several proteases. The study also explores the performance of OCSVM models, both with and without negative class data, revealing that tailored feature selection and weighting strategies significantly enhance model performance. The findings of this research underscore the potential of machine learning techniques in advancing bioinformatics and protease research. The developed models not only improve the precision of proteomic analyses but also support the broader field of precision medicine by providing deeper insights into protease functions in health and disease. Acknowledgments I would like to extend my heartfelt gratitude to several individuals and institutions that made this research possible. First and foremost, I am deeply indebted to my advisor, Dr. Christopher Kieslich, for his unwavering support, insightful guidance, and invaluable feedback throughout the course of this study. His expertise and dedication were instrumental in shaping the direction and success of this research. I am also profoundly grateful to Dr. Selen Cremaschi, the chair of department, whose profound knowledge and rigorous approach have greatly enriched this work. Her encouragement and constructive critiques were essential in refining my research objectives and methodologies. Special thanks to Dr. Peter He for his critical feedback and for being an inspiring mentor. His thorough reviews and insightful comments significantly improved the quality of this thesis. I would like to acknowledge my colleagues and friends in the Department of Chemical Engineering at Auburn University for their camaraderie, support, and intellectual discussions that have greatly contributed to my academic and personal growth. A special note of thanks to the administrative staff for their assistance and for ens

FAQs

What machine learning models are used to predict proteolytic cleavage specificity?

The study uses Support Vector Machines (SVMs) and One-Class Support Vector Machines (OCSVMs) to model proteolytic cleavage specificity from encoded peptide sequences.

How are peptide sequences represented for the classifiers?

The thesis introduces Fourier Transform-based encoding of peptide sequences to capture essential biochemical properties and structural characteristics, which are then used as inputs to SVM algorithms.

What datasets support the meta-analysis and how are negative examples handled?

Datasets are sourced from the MEROPS database and include positive cleaved sequences plus synthetic negative non-cleaved sequences to improve robustness and diversity. The work also evaluates OCSVM performance with and without negative class data.

Aug 4, 2026

Upload by Theodore