Machine learning-driving optimization and spatial assembly of a cell-free system for high-yield liquiritigenin production

Abstract Liquiritigenin is a medicinal flavonoid whose production is constrained by inefficient plant extraction and complex chemical synthesis. To overcome this, we developed a modular cell-free multi-enzyme system for its efficient biosynthesis from tyrosine, integrating spatial enzyme assembly with machine learning-guided optimization. Using a combined cell-free metabolic engineering (CFME) and cell-free protein synthesis-driven metabolic engineering (CFPS-ME) approach, we screened and optimized five key pathway enzymes to establish a one-pot reaction. The optimal enzyme combination (phenylalanine ammonia-lyase from Zea mays , 4-coumarate-coenzyme A ligase 4 from Arabidopsis thaliana , chalcone synthase from Glycine max , chalcone reductase from Medicago sativa , chalcone flavonone isomerase from Zea mays ) was identified through systematic screening and ratio optimization. After Plackett–Burman and steepest-ascent experiments, three rounds of iterative machine learning fine-tuned key parameters, including enzyme ratios and cofactor concentrations, yielding 155.32 ± 14.39 mg/L. Spatial enzyme assembly was further enhanced via covalent peptide tags and scaffold proteins (γPFD-SpyCatcher) under CFME. Combining CFPS-ME with scaffold-assisted co-immobilization significantly boosted production, reaching a final titer of 439.42 ± 19.53 mg/L. This study demonstrates that machine learning-driven optimization and spatial assembly of multienzyme complexes is a powerful approach for cell-free biosynthesis.