UC Irvine

Antibodies are a central class of biotherapeutics and reagents. Recent advances in computational protein design have enabled the de novo generation of antibody fragments against user-defined epitopes, transforming antibody discovery from a screening problem into a design problem. However, computationally designed antibodies typically exhibit modest binding affinities that limit their practical utility, creating a designto-optimal affinity gap that remains a major challenge in antibody engineering. This dissertation addresses that gap by integrating computational antibody design with OrthoRep-based continuous evolution and by using the resulting evolutionary trajectories to investigate the molecular basis of affinity maturation.First, to substantially improve the scalability of antibody evolution and to make evolution from diverse starting points practical, I developed a TP901-mediated integration strategy to enable highly efficient loading of large gene libraries onto OrthoRep. This advance was incorporated together with high-mutation-rate polymerases and a rapid inducible surface-display circuit into a set of standardized AHEAD (autonomous hypermutation yeast surface display) strains for continuous antibody evolution. Second, using this platform, I demonstrated that continuous evolution could close the design-tooptimal affinity gap. Five de novo VHHs generated by RFdiffusion against distinct epitopes were affinity matured through AHEAD, producing order-of-magnitude affinity gains and sub-nanomolar binders in multiple campaigns. Structural and computational analyses confirmed that maturation preserved the designed epitope and binding orientation while substantially improving functional performance. Finally, I investigated the origins of the design-to-affinity gap by reconstructing the complete combinatorial fitness landscape connecting a computationally designed nanobody to its affinity-matured descendant. Integrating quantitative affinity measurements with molecular dynamics simulations, ProteinMPNN analysis, and probabilistic evolutionary pathway modeling revealed that this gap arises from multiple interacting limitations, including cryptic conformational states, extensive epistatic interactions, and biases in current computational sequence-design models. These findings explain how continuous evolution overcomes limitations that remain difficult to predict computationally and identify opportunities for improving future computational antibody design.Together, these results establish a unified framework in which computational design generates epitope-specific starting binders, continuous evolution optimizes their functional properties, and evolutionary analysis reveals the sequence–structure–function relationships needed to improve the next generation of computational design methods.