Moving Protein PEGylation from an Art to a Data Science

PEGylation is a well-established and clinically proven half-life extension strategy for protein delivery. Protein modification with amine-reactive poly(ethylene glycol) (PEG) generates heterogeneous and complex bioconjugate mixtures, often composed of several PEG positional isomers with varied therapeutic efficacy. Laborious and costly experiments for reaction optimization and purification are needed to generate a therapeutically useful PEG conjugate. Kinetic models which accurately predict the outcome of so-called "random" PEGylation reactions provide an opportunity to bypass extensive wet lab experimentation and streamline the bioconjugation process. In this study, we propose a protein tertiary structure-dependent reactivity model that describes the rate of protein-amine PEGylation and introduces "PEG chain coverage" as a tangible metric to assess the shielding effect of PEG chains. This structure-dependent reactivity model was implemented into three models (linear, structure-based, and machine-learned) to gain insight into how protein-specific molecular descriptors (exposed surface areas, pK_a, and surface charge) impacted amine reactivity at each site. Linear and machine-learned models demonstrated over 75% prediction accuracy with butylcholinesterase. Model validation with Somavert, PEGASYS, and phenylalanine ammonia lyase showed good correlation between predicted and experimentally determined degrees of modification. Our structure-dependent reactivity model was also able to simulate PEGylation progress curves and estimate "PEGmer" distribution with accurate predictions across different proteins, PEG linker chemistry, and PEG molecular weights. Moreover, in-depth analysis of these simulated reaction curves highlighted possible PEG conformational transitions (from dumbbell to brush) on the surface of lysozyme, as a function of PEG molecular weight.