Genetic algorithm-guided sample design enables efficient machine learning-driven optimization of tauroursodeoxycholic acid production in engineered Escherichia coli.

Data-driven machine learning (ML) approaches offer powerful tools for modeling and optimizing complex fermentation processes, yet their application is often limited by the need for large experimental datasets, which is difficult to meet in low-throughput microbial fermentation systems. Here, we present a practical genetic algorithm (GA)-assisted adaptive sampling strategy to support sample-efficient ML-assisted optimization of microbial biotransformation processes. In this strategy, GA is employed upstream to iteratively select informative experimental conditions and generate a compact dataset for artificial neural network (ANN) modeling, rather than relying on uniformly distributed or predefined experimental designs. Using tauroursodeoxycholic acid (TUDCA) biotransformation in engineered Escherichia coli as a case study, the proposed GA-assisted adaptive sampling strategy coupled with ANN (GA-ANN) workflow identified favorable fermentation medium compositions under a limited experimental budget. Experimental validation yielded a maximum TUDCA/TCDCA ratio of 2.18 ± 0.06 and a TUDCA titer of 7.11 ± 0.12 g/L using 20 g/L chicken bile powder as substrate. ANN-based response surface analysis further revealed a biphasic effect of corn dextrin concentration on TUDCA conversion, indicating a nonlinear behavior that was not fully captured by quadratic regression models. This study provides a feasible and sample-efficient strategy for low-throughput fermentation optimization under constrained experimental resources.