Genomic Benchmarking Reveals Reduced and Context-Dependent Transferability of AI-Engineered Bxb1 Recombinases for Large-Fragment Genome Writing.

Programmable large-fragment genome integration using prime-editing-coupled serine integrases, such as PASTE and PASSIGE, remains constrained by the limited activity of wild-type Bxb1 (WT Bxb1) in mammalian cells. Recently, the AI-guided protein engineering framework EVOLVEpro enabled efficient identification of functional protein variants from limited experimental sampling and nominated epBxb1(T166R) as a highly active Bxb1 variant in episomal plasmid-to-plasmid recombination screens. Here, we systematically benchmarked epBxb1 (T166R) against WT Bxb1 and the previously validated high-activity eeBxb1 (V74A/E229K/V375I) variant in genome-integrated reporter and endogenous-locus PASSIGE assays at three well-characterized benchmarking loci (AAVS1, CCR5, and ACTB). epBxb1 showed reduced and variable transferability, with no detectable advantage over WT Bxb1 in the genome-integrated reporter system or at the AAVS1 safe-harbor locus, but produced modest, locus-dependent improvements at CCR5 and ACTB, with 1.89-fold and 1.58-fold increases, respectively. By contrast, eeBxb1 consistently showed superior activity, achieving up to ∼12-fold improvement over WT Bxb1. Four additional rounds of EVOLVEpro-guided optimization on the eeBxb1-BPNLS scaffold identified no reproducibly improved variant among 40 tested single-amino-acid substitutions. These results indicate reduced and genomic context-dependent transferability of EVOLVEpro-nominated Bxb1 variants and highlight the importance of application-matched genomic benchmarking when implementing AI-guided protein optimization for therapeutic genome writing.