Data-driven peptide drug discovery and design

Ribosomally synthesized and post-translationally modified peptides (RiPPs) represent a diverse class of natural products that have obtained significant attention in recent years due to their broad spectrum of biological activities. These peptides undergo extensive post-translational modifications, which impart unique structures not achievable through natural ribosomal peptides. Such modifications often restrict conformational flexibility, enhancing target recognition, metabolic and chemical stability, and functional diversity. RiPPs are particularly attractive for engineering applications, as their precursor peptides are genetically encoded, and many biosynthetic enzymes exhibit substrate tolerance. Despite their potential, the discovery and design of novel RiPPs for therapeutic applications remain challenging due to a limited understanding of the substrate tolerance mechanisms in RiPP synthetases and the folding processes involved in their biosynthesis. In this dissertation, I have employed computational approaches, including deep learning and molecular dynamics (MD) simulations, to advance peptide drug discovery and design. My research primarily focuses on two natural peptide families: lasso peptides and lanthipeptides. For lasso peptides, I developed a specialized language model, LassoESM, leveraging advancements in pre-trained protein language models to enhance lasso peptide-related properties prediction. Furthermore, I conducted extensive atomic-level MD simulations to investigate the universal folding mechanisms of 20 structurally characterized lasso peptides. Using a deep learning architecture based on Variational Autoencoder, I successfully identified distinct folding pathways for these peptides. For lanthipeptides, I utilized MD simulations to explore the site-selectivity of lanthipeptide cyclization, uncovering critical insights into the molecular mechanisms underlying their biosynthesis. Additionally, I addressed the challenges posed by limited data in peptide discovery and protein-peptide interaction prediction by implementing a deep learning framework based on one-shot learning architecture. Overall, I anticipate that LassoESM and its future iterations will serve as valuable tools for the rational design of lasso peptides with tailored properties. Insights into the folding mechanisms of lasso peptides and the site-selectivity of lanthipeptide cyclization will deepen our understanding of their biosynthetic processes. Moreover, one-shot learning provides a promising framework for accurately predicting peptide-protein interactions, particularly in low-data scenarios.

Name thesis:degree_name

Ph.D.

Level thesis:degree_level

Dissertation

Discipline thesis:degree_discipline

Biophysics & Quant Biology

Grantor

University of Illinois Urbana-Champaign

Year dc:date

2025

Author dc:creator

• Mi, Xuenan

Contributors dc:contributor

• Shukla, Diwakar
• Mitchell, Douglas Alan
• Pogorelov, Taras
• Jackson, Nicholas

• Ribosomally synthesized and post-translationally modified peptides
• Deep Learning
• Molecular Dynamics

Statement dc:rights

• Copyright 2025 Xuenan Mi

Language dc:language

en, eng

Handle dc:identifier

https://hdl.handle.net/2142/129159

OAI record oai:identifier

https://www.ideals.illinois.edu/oai-pmh?verb=GetRecord&metadataPrefix=etdms&identifier=oai%3Awww.ideals.illinois.edu%3A2142%2F129159