Uncertainty-Driven Experiment Design in Cell-Free Protein Synthesis with Bayesian Optimization

Cell-free protein synthesis (CFPS) offers rapid protein production, yet yield variability from volume constraints and batch-to-batch differences remains a challenge. Bayesian optimization (BO) identifies optimal points under fixed constraints but does not address constraint variability. Here we use Gaussian process (GP) predictive uncertainty in two complementary directions, lower and upper confidence bounds for exploitation and exploration, to design condition spaces under variable constraints. Closed-loop optimization of two CFPS systems improved green fluorescent protein (GFP) yield beyond the standard condition within a few rounds. For exploitation, we defined a yield assurance space (YAS), a low-uncertainty set of conditions selected to exceed a conservative yield threshold. For exploration, we proposed Pareto front expansion under tightened volume constraints via uncertainty-driven sampling. This approach illustrates how GP predictive uncertainty can support both assured condition design and targeted exploration under altered constraints.