UCLA

Amyloid proteins call into question our long-standing notions of the relationship between a proteinsequence and its three-dimensional fold. These proteins exhibit distinct conformations associated with soluble and fibrillar states and potentially various fibril polymorphs. This conformational plasticity contributes to the diverse roles of amyloids, which fall into two distinct categories within biological systems. Pathogenic amyloids are observed in patients suffering from debilitating neurodegenerative and systemic disorders such as Alzheimer’s Disease and Parkinson’s Disease and can be infectious as is the case with prion-related diseases. In contrast to pathogenic amyloids, functional amyloid assemblies have a defined function in living systems. Such functional amyloids are prevalent across both eukaryotic and prokaryotic domains of life, implying the fold may be an early and conserved mechanism with value to cellular function. This dichotomy raises an important question: What dictates the fold adopted by amyloid- forming sequences and in turn, what determines the function of a particular amyloid fold? This thesis attempts to address this puzzle with computational tools and experimental validation, as a stepping stone toward a broader understanding of amyloids across the tree of life.

Advances in structural biology have provided atomic-level views of amyloids, from short 'steric zipper’segments and to full amyloid fibril structures, offering researchers detailed insights into the features of amyloid proteins. Concurrently, the advent of machine learning techniques has allowed for the exploration of data that previously had been unfeasible. Importantly, despite their differences, pathogenic and functional amyloids have been found to share a common architecture: long unbranched fibrils with polypeptide chains folded nearly two-dimensionally within a layer, and with adjacent layers stacked parallel in register. The work presented in Chapter 2 of this dissertation explores the application of machine learning strategies to expand the capabilities of steric zipper prediction and investigates their utility in mining organism proteomes for high zipper propensity proteins that may harbor amyloid-forming domains. The approach offers a fast way to survey proteomes for putative amyloid-forming proteins or domains, providing insight into the prevalence and distribution of steric zippers across proteomes. Chapter 3 presents an alternative application of the methods developed in Chapter 2. Here, we employ computational approaches to design novel peptide sequences indented to bind a known fibril surface. Subsequently, we investigate the effects of these designed peptides on amyloid fibril growth and their resulting structures. By combining structural biology, machine learning, and targeted peptide design, the work presented in this thesis aims to deepen our understanding of amyloid proteins across diverse biological contexts, potentially paving the way for novel therapeutic strategies and insights into both pathogenic and functional amyloid formations.