Cell and gene therapies often express nonhuman proteins, which carry a risk of anti-therapy immunogenicity. An emerging consensus is to instead use modified human protein domains, but these domains include nonhuman peptides around mutated residues and at interdomain junctions, which may also be immunogenic. We present a modular workflow to optimize protein function and minimize immunogenicity by using existing machine learning models that predict protein function and peptide-major histocompatibility complex (MHC) presentation. We first applied this workflow to existing transcriptional activation and RNA-binding domains by removing potentially immunogenic MHC II epitopes. We then generated small-molecule-controllable transcription factors with human-derived DNA-binding domains targeting non-genomic DNA sequences. Finally, we established a workflow for creating deimmunized zinc-finger arrays to target arbitrary DNA sequences and upregulated two therapeutically relevant genes, utrophin (UTRN) and sodium voltage-gated channel alpha subunit 1 (SCN1A), using it. Our modular workflow offers a way to potentially make cell and gene therapies safer and more efficacious using state-of-the-art algorithms.