Machine-Guided Dual-Objective Protein Engineering for Deimmunization and Therapeutic Functions

Abstract

Cell and gene therapies often rely on the expression of exogenous proteins derived from nonhuman organisms. An emerging consensus is to reduce the potential immunogenicity of such therapies by instead using human protein domains. However, as we engineer these human-derived proteins, we create nonhuman peptides at the linkers or junctions between domains and at mutated residues within them, which still pose a risk of immunogenicity that has largely been left unaddressed. Here, we present a modular workflow to simultaneously optimize the functions of proteins and minimize their immunogenic risk using existing machine learning models that predict protein function and nonhuman peptide immunogenicity from their sequences. We first applied this workflow to existing transcriptional activation and bio-orthogonal RNA binding domains. Then we generated a set of small molecule-controllable transcription factors with human-derived zinc finger DNA-binding domains for targeting orthogonal non-genomic DNA sequences. Finally, we established a workflow for creating deimmunized zinc finger arrays to target arbitrary genomic DNA sequences and used it to upregulate expression of two therapeutically relevant genes, UTRN and SCN1A. Our future-proof, modular workflow offers a proof of principle for making cell and gene therapies safer and more efficacious through dual-objective protein optimization using state-of-the-art algorithms.