Dynamic Control for Directed Evolution: A New Tool for Robust Bioengineering

Microorganisms on Earth are crucial to providing oxygen, food, and waste remediation. Providing the same functionality in space, as in closed-loop human environmental control and life support systems (ECLSS) or in situ resource utilization (ISRU), are key areas for NASA's technology advancement; however, space environments present a complex battery of challenges to life. Synthetic biology's toolbox for addressing this problem currently relies heavily on rational design. Directed evolution (DE) is a complementary bioengineering technique that empirically alters the powerful, yet often poorly understood, natural mechanisms within a biological population. Although DE shows great promise in creating robust, stable biological populations with a desired form or function, existing implementations have not yet reach life's known limits. We propose combining DE with machine learning techniques via automated state sensing and environmental adjustment: the resulting fast, scalable, dynamically-controlled directed evolution (DCDE) will be a new synthetic biology tool. Its potential applications include rapid generation and screening of gene libraries (synthetic biology), identification of novel environmental tolerance mechanisms (astrobiology, ECLSS, ISRU), and improvement of drug and biomaterial production systems (biomedicine, etc.). We propose to develop an automated DCDE system, including both hardware and control software. We have an existing prototype capable of iterative cycles of microbial population growth, population state sensing, environmental exposure, and population regrowth. The hardware is currently single-input/ single-output (SISO), using optical density and UV-C exposure. To enable testing of the software DCDE approach, we must improve the robustness of the prototype to allow for at least seven continuous exposure/regrowth cycles, add the extra sensors (minimum of dissolved CO2/O2 and pH, with conductivity and OPR as time permits) necessary for the multiple-input/multiple-output (MIMO) case, and replace the current one-path fluidics chamber with one that will allow for parallel operation (necessary for proper experimental control). To perform a direct comparison of DCDE and uncontrolled DE, we will repeat a well-established DE experiment, inducing UV-C tolerance in Escherichia coli, both with and without dynamic control. (We have already performed the 'by hand' experimental cases, eliminating automation as a variable.) Pilot SISO runs will consist of UV-C radiation as input and UV-C survival as measured output. Full experimental runs, with at least three replicates, will consist of inputs UV-C and temperature and outputs UV-C and thermal survival and metabolic rate. We anticipate that DCDE will show a higher final resistance in both cases. Hardware-controlled sensing and regulation of living systems has been demonstrated at the single gene pathway level, but such approaches are not scalable to complete biological systems. Attempts at population-level directed evolution controlled 'by hand' have shown that even limited adaptive control increases the end result substantially; however, this is both a non-reproducible 'feedback' technique and a highly labor-intensive and error-prone laboratory procedure. Efforts at true closed-loop population-level DE systems exist, but have been ad hoc efforts to enable a single experiment. These cases, in addition to lacking generality, have used at best SISO linear control techniques. However, a biological culture in its environment, even a highly controlled experimental environment, consists of many relevant input (temperature, light, presence of nutrients, etc.) and output (survival, growth rate, metabolic products, etc.) variables, is highly non-linear, and cannot be assumed to be time-invariant. Thus, in addition to creating a novel DCDE system, we are expanding for the first time the approach of DCDE to the MIMO case, making it possible to apply powerful techniques from the field of machine learning such as Bayesian modeling, state space systems control, and black-box systems identification. The flexibility, extensibility, and modularity of such an approach will make it possible for this hardware to be used to test a wide variety of control techniques, as well as enable non-bioengineering science. If the work proposed here demonstrates DCDE's potential, a follow-on collaboration with UCSC is lined up to sequence the microbes sampled from each iteration. This ability to track population-level phenotypic and genotypic changes to environmental pressure with a time resolution matching that of a single generation will be unprecedented. More »