McConnell, Adam2025-04-212025-04-212024-01https://hdl.handle.net/11299/271363University of Minnesota Ph.D. dissertation. January 2024. Major: Biomedical Engineering. Advisor: Benjamin Hackel. 1 computer file (PDF); xi, 128 pages.Binding ligands empower molecular therapeutics and diagnostics, yet discovery of protein ligands is hindered by the vast, sparse, and rugged sequence-function fitness landscape. Protein scaffolds, characterized by a conserved framework to support an engineered variable active site, have proven effective in navigating this landscape. While monoclonal antibodies are established natural scaffolds, miniproteins offer advantages over antibodies such as their modularity, tissue penetration, plasma clearance, and cost-effective manufacturing. Yet, engineering miniproteins as scaffolds is complex due to their small size, which poses a challenge in integrating novel functionality while also maintaining biophysical integrity. The challenge is further compounded by the elusive factors driving their developability – properties like stability and expression that determine a protein’s viability as a product – and evolvability – the capacity to accept mutations necessary for novel function. To tackle this challenge, we evaluated the ability of hyperstable synthetic miniproteins, originally designed for stability, for their potential as binding scaffolds. We synthesized 45 libraries with systematic variations in two topologies, each incorporating five initial framework sequences and four or five distinct paratopes, to determine the impact of topology, paratope, and framework sequence on evolvability and developability. We discovered several library designs that were both evolvable and developable, producing numerous binders across four distinct targets, while maintaining high stability, protease resistance, and expression. Next, we developed an active machine learning framework to improve the efficiency of screening potential framework sequences as scaffolds, through a computational-experimental feedback loop. Employing traditional machine learning models, trained using our data from the initial study, we predicted scaffold fitness using features extracted from structures of scaffold variants. This approach guided our selection of an additional 45 frameworks across diverse topologies for experimental testing. Integrating the new data from the diverse range of newly developed scaffolds, we refined our models for a more reliable method for modeling scaffold developability and evolvability. Furthermore, eigen centrality and betweenness centrality measures were identified as key features critical for modeling scaffold fitness. When trained on evolvability data, the models predominantly utilized paratope-related features. In contrast, models trained with developability data emphasized features from the scaffold’s conserved region. This distinction reveals insights on properties that influence both the developability and evolvability of miniprotein scaffolds. Overall, this work validated hyperstable synthetic miniproteins as viable scaffold precursors and identified multiple designs that yielded high-performance libraries containing high-affinity, developable binders. These studies elucidated determinants of the developability and evolvability of synthetic miniproteins as ligand scaffolds by employing a computational-experimental feedback loop.enantibody alternative scaffoldsprotein engineeringscaffold developabilityscaffold evolvabilityyeast surface displayThesis or Dissertation