Multi-Objective Engineering of Therapeutic Antibodies

Despite the success of antibody therapeutics, there are several outstanding challenges related to antibody drug development that I have sought to address in this dissertation. First, it is challenging to generate large and high-quality datasets of antibody biophysical properties. Second, even if this data is available, it is also difficult to develop predictive models sufficient for co-optimization of antibody properties. To address these challenges, I have developed ultra-dilute (antibody concentrations <0.02 mg/mL) and highly reproducible experimental techniques for evaluating two forms of non-affinity interactions (self-association and non-specific binding) that significantly impede therapeutic development. For the measurement of self-association, I have contributed to the development of the charge-stabilized self-interaction nanoparticle spectroscopy (CS-SINS) assay, which distinguishes between antibodies with optimal high-concentration solution behaviors and those with suboptimal high-concentration solution behaviors (high viscosity and high opalescence) with high accuracy. I have also developed the polyspecificity particle assay (PSP assay) for the measurement of antibody non-specific binding, which distinguishes between optimal (low non-specific binding) and suboptimal (high non-specific binding) antibodies with high accuracy. The application of these and related experimental techniques has facilitated the acquisition of relatively large and high-quality datasets for the application of machine learning. I demonstrated that modest datasets of self-association and non-specific binding measurements of a diverse panel of clinical-stage antibodies can be analyzed via interpretable machine learning models. Gaussian Naïve Bayes machine learning models accurately classify antibody properties and improve the development process by obviating the need for time- and resource intensive experimentation. Moreover, the interpretation of these models facilitates re-engineering of suboptimal antibodies by identifying mutations that co-optimize the suboptimal biophysical properties of three clinical stage antibodies – panitumumab, gantenerumab, and cinpanemab – while maintaining their high affinity. I also developed machine learning models applied to high-throughput screening data, which enable prediction of continuous biophysical property values from binary experimental labels. Linear discriminant analysis (LDA) facilitates the co-optimization of a therapeutic antibody candidate (emibetuzumab) by identifying variants of this antibody that display both increased antigen binding and reduced non-specific binding despite these properties exhibiting strong tradeoffs. Overall, this work serves to improve the therapeutic antibody development pipeline by increasing the accuracy, throughput, and reliability with which we can experimentally measure and computationally predict important properties that impact the clinical potential of therapeutic antibodies.