Machine learning guided engineering of conditional split inteins for regulated protein splicing in mammalian cells

Inteins are proteins that excise themselves from precursor proteins and connect the flanking polypeptides with a peptide bond. Split inteins consist of two independently translated fragments that must associate to become splice-competent. They can be used for diverse post-translational protein modifications. Using the ML Int&in algorithm, we predicted unnatural split sites in two of the fastest and most efficient split inteins, gp41-1 and NrdJ-1, to generate functional variants with fragments of reduced mutual affinity. We harness this feature to create conditional versions of these inteins by controlling the physical proximity of the intein fragments with a light-inducible heterodimerization system. The resulting light-activatable gp41-1 and NrdJ-1 inteins enabled blue light-dependent control of Cre recombinase activity in mammalian cells, which we exploited to spatially control apoptosis via localized expression of truncated BID (tBID) and caspase-8. This work highlights the versatility of Int&in for designing conditional inteins for precise spatiotemporal protein regulation.