Abstract
This thesis develops a computationally assisted framework for peptide discovery that combines structural modelling, machine learning, and experimental validation across antiviral and molecular-targeting applications. The work is motivated by the large size of peptide sequence space and the practical difficulty of screening candidate libraries exhaustively, particularly when target specificity and translational relevance are required. Against this background, the thesis first reviews the evolution of AI-assisted antiviral peptide discovery, from early sequence-based predictors to target-aware, structure-guided, and generative design strategies, with respiratory viruses used as a central test case.
The original research contributions focus on accelerating peptide screening and applying the resulting workflow to biologically relevant targets. First, a machine-learning-guided strategy was developed to complement peptide-protein docking and make large peptide spaces computationally tractable. This framework was then extended through the generation of a large peptide conformational repository to support future virtual-screening applications. The platform was subsequently applied to the design of peptides against receptor-binding regions of SARS-CoV-2 and influenza A virus, yielding candidates that reduced viral burden in tissue-engineered human lung models and were also associated with improved cell viability, reduced apoptosis, and altered inflammatory responses. Finally, the same general strategy was applied to endothelial targeting through E-selectin, where direct clamp-peptide targeting emerged as the more promising approach. Together, these studies establish a generalizable peptide-discovery framework and demonstrate its potential across distinct translational settings.