Machine-learning-assisted universal protein activation in living mice

Abstract

A universal strategy to precisely control protein activation in living animals is crucial for gain-of-function study of proteins under in vivo settings. We herein report CAGE-Prox^vivo, a computer-aided proximal decaging strategy for on-demand protein activation as well as protein-protein interaction modulations in living mice. Through machine-learning-assisted evolution of desired aminoacyl-tRNA synthetases (aaRSs), we successfully incorporated chemically caged amino acids into rationally designed "decaging sites" to transiently block target proteins' function, which can be restored in situ via a small-molecule-triggered bioorthogonal cleavage reaction. This method demonstrates broad applicability ranging from activating proteins of interest to cell-type-specific modulation of distinct phenotypes in living systems. Beyond the active-pocket decaging, CAGE-Prox^vivo also enables precise control of protein-protein interactions, as exemplified by a "gated" anti-CD3 antibody that permits chemically regulated T cell recruitment and activation at tumor sites. Overall, CAGE-Prox^vivo offers a universal platform for time-resolved biological studies and on-demand therapeutic interventions under living conditions.

Keywords: Anti-tumor immunotherapy; Bioorthogonal decaging; Enzyme evolution; Gain-of-function protein studies; Genetic code expansion; Machine learning; On-demand protein activation; Protein-protein interaction modulation; T cell engagement; Tumor-specific pyroptosis.