The engineering of living cells able to learn algorithms by themselves, such as playing board games —a classic challenge for artificial intelligence— will allow complex ecosystems and tissues to be chemically reprogrammed to learn complex decisions. However, current engineered gene circuits encoding decision-making algorithms have failed to implement self-programmability and they require supervised tuning. We show a strategy for engineering gene circuits to rewire themselves by reinforcement learning. We created a scalable general-purpose library of Escherichia coli strains encoding elementary adaptive genetic systems capable of persistently adjusting their relative levels of expression according to their previous behavior. Our strains can learn the mastery of 3×3 board games such as tic-tac-toe, even when starting from a completely ignorant state. We provide a general genetic mechanism for the autonomous learning of decisions in changeable environments.One-Sentence SummaryWe propose a scalable strategy to engineer gene circuits capable of autonomously learning decision-making in complex environments.

E. coli cells, which are commonly used to study gene expression, show membraneless organization of gene expression components within the cell. Cellfree expression systems are also widely used to study gene expression, but a major limitation is the lack of a means to spatially organize DNA plasmids and ribosomes to mimic the cellular environment. In this work, we used computer simulations to guide experimental efforts to control the spatial organization of DNA and ribosomes in cell-sized vesicles using macromolecular crowding. Using Brownian dynamics simulations of coarse-grained models of DNA plasmids and crowders, we showed that plasmids remain uniformly distributed at low levels of crowding but become strongly adsorbed at confining surfaces at high levels of crowding. These effects are due to entropic depletion interactions resulting from the presence of crowding molecules. We validated the simulation results experimentally by using fluorescently-labelled DNA plasmids and ribosomes in cell-sized vesicles at different concentrations of the macromolecular crowder Ficoll-70. Large crowder concentrations resulted in preferential localization of plasmids at the walls of the vesicle, while ribosomes remained uniformly distributed throughout the vesicle. We then used kinetic Monte Carlo simulations to study how protein production was affected by crowding-induced changes in spatial organization and diffusion. The localization of DNA to the wall resulted in lower protein abundance and a decrease in translational efficiency with an increase in system size. Experimentally, we tracked the dynamics of transcription and translation in crowded vesicles using a coupled mRNA/protein reporter technique. These results were consistent with results from simulations. We thus used computer simulations to design a cell-free experimental platform capable of spatial organization of gene expression components. This platform can be used to better understand the spatial control of gene expression in cells.