Machine Learning and Molecular Scaffolds Supercharge Production of a Licorice-Derived Drug Molecule

Chemists and biologists have long coveted liquiritigenin, a flavonoid pulled from the rhizomes of licorice plants that shows cardioprotective, antidiabetic, antioxidant, anti-inflammatory, and antitumor activity in studies. Yet getting meaningful quantities of the compound has never been easy. Extracting it from plants is inefficient and subject to the vagaries of agriculture, while synthesizing it chemically demands convoluted reaction sequences and laborious purification. Microbial fermentation has offered only modest help: engineered yeast and bacteria have produced liquiritigenin at titers ranging from a few milligrams to tens of milligrams per liter, far below what industrial applications would demand. Now a team at Sun Yat-Sen University in Guangzhou, China, reports a dramatically more productive route that abandons living cells altogether, combining cell-free enzyme chemistry with machine learning and self-assembling protein scaffolds to reach a titer of 439.42 milligrams per liter, roughly a hundredfold improvement over the starting point.

The study, published in the journal Advanced Biotechnology, describes a five-enzyme cascade that converts the cheap amino acid L-tyrosine into liquiritigenin through a series of intermediates including p-coumaric acid and isoliquiritigenin. The enzymes themselves come from an eclectic cast of organisms: phenylalanine ammonia-lyase and chalcone isomerase from maize, 4-coumarate-CoA ligase from thale cress, chalcone synthase from soybean, and chalcone reductase from alfalfa. In nature, the pathway is buried inside plant cells, where enzymes cluster into multienzyme complexes amid the crowded cytoplasm. The researchers set out to recreate that spatial intimacy in a test tube, where they could control every variable without worrying about keeping cells alive or diverting resources into biomass.

Cell-free biosynthesis comes in two flavors, and the team exploited both. In cell-free metabolic engineering, or CFME, the pathway enzymes are pre-expressed in bacteria, and crude lysates enriched with the target proteins are mixed directly into the reaction. In cell-free protein synthesis-driven metabolic engineering, or CFPS-ME, the enzymes are manufactured on the spot by a cell-free expression system, adding DNA plasmids encoding each enzyme to a bacterial extract loaded with amino acids, nucleotides, and an energy source. When the group compared the two approaches, CFPS-ME consistently outperformed CFME, producing 4.55 milligrams per liter of liquiritigenin from tyrosine, whereas the lysate-only system lagged behind. The difference likely reflects the higher effective enzyme concentrations achievable when protein synthesis proceeds inside the reaction itself.

The next challenge was choosing the right versions of each enzyme and calibrating their proportions. Because a cascade is only as good as its slowest and least compatible links, the researchers screened homologs for every step, tapping the NCBI database and published catalytic data to shortlist candidates. Chalcone reductase deserved particular scrutiny: it works in concert with chalcone synthase to divert flux toward isoliquiritigenin, and without enough malonyl-CoA, the synthase stalls and the pathway shunts toward naringenin instead of the desired product. The screening revealed an optimal combination of the maize, thale cress, soybean, alfalfa, and maize enzymes, with plasmid concentrations in the cell-free expression system fine-tuned to 35, 15, 20, 20, and 20 nanograms per microliter respectively. Excess chalcone reductase actually proved inhibitory, peaking at 25 milligrams per milliliter of enzyme and declining thereafter, a reminder that more catalyst is not always better chemistry.

With the enzyme roster settled, the team moved to reaction conditions, and here the workflow grew ambitious. They began with one-factor-at-a-time tests, identifying 37 degrees Celsius, pH 8.0, a 36-hour reaction time, and a 75-microliter reaction volume as favorable. A Plackett-Burman design, a statistical screening method that tests each variable at two levels across a small set of experiments, then winnowed thirteen parameters down to the five that mattered most: the concentrations of ZmPAL, GmCHS, and MsCHR, plus pH and reaction volume. Steepest-ascent experiments pushed the yield to 104.42 milligrams per liter. But the interplay among enzymes, cofactors like ATP, CoA, and NADPH, and environmental variables creates a rugged optimization landscape that classical statistical designs traverse only slowly.

That is where machine learning took over. The researchers fed seventy-five data points from their preliminary designs into an iterative loop of data input, model training, Bayesian optimization, and experimental validation. Seven algorithms competed for the job, spanning classics like Random Forest, Support Vector Machine, and Multilayer Perceptron, ensemble methods including XGBoost, LightGBM, and CatBoost, and Gaussian Process Regression, prized for its performance on small datasets. Nested cross-validation, with three-fold tuning inside a nine-fold generalization assessment, guarded against overfitting, and when several models performed comparably the team fused them using Stacking, simple averaging, and non-negative least squares weighting. The Stacking ensemble, with RidgeCV as its meta-learner, ultimately achieved a test R-squared of 0.903, meaning it explained more than ninety percent of the variance in experimental outcomes.

Across three rounds of machine-guided experimentation, the average yield of the top twelve predicted conditions climbed from 106.55 to 125.46 and then 138.11 milligrams per liter, while variability shrank, a signature of genuine convergence rather than lucky guesses. Explainable AI techniques sharpened the picture. SHAP analysis, which decomposes each prediction into the contributions of individual features, identified ZmPAL, reaction volume, ATP, and NADPH as the strongest positive drivers, with ZmPAL alone contributing an estimated 35.8 percent uplift in one prediction round. CoA, tyrosine, and GmCHS showed negative correlations, so the model prescribed less of them. The final recipe, including 4 milligrams per milliliter of ZmPAL, 2.2 millimolar ATP, 2.4 millimolar NADPH, pH 7.3, and 39 degrees Celsius, delivered 155.32 milligrams per liter, a 48.7 percent jump over the statistical-only optimization and more than a doubling of conversion efficiency.

The final act of the study tackled spatial organization, borrowing a trick from biology’s own playbook. The team fused covalent self-assembling peptide tags, including the celebrated SpyTag/SpyCatcher pair along with SnoopTag/SnoopCatcher, CC-Di, and RIAD/RIDD systems, onto the pathway enzymes so that specific pairs could lock together like molecular Velcro. Assembling chalcone synthase and chalcone reductase with these peptide pairs already lifted production in the CFME system, with the best combination reaching 22.69 milligrams per liter. Then came the scaffold. The researchers tested eight candidate scaffold proteins, ranging from bacterial microcompartment shells to viral capsid proteins, and the winner was gamma-prefoldin, a filamentous chaperone from hyperthermophilic archaea, fused to SpyCatcher. Attaching all five enzymes to this scaffold nearly doubled yield relative to unscaffolded reactions, and an N-C-C-N-C binding configuration combined with 5 milligrams per milliliter of scaffold protein pushed the titer to 439.42 milligrams per liter, an overall 2.83-fold enhancement attributable to spatial assembly.

Structural modeling helps explain why the scaffold works so well. Predictions generated with AlphaFold 3 depict the complex as a compact, prism-like architecture in which SpyTag placement exposes every catalytic pocket, verified through active-site analysis with the ProteinsPlus platform. The gamma-prefoldin backbone holds the enzymes close enough that intermediates pass directly from one active site to the next rather than diffusing into the bulk solution, where they would be lost to dilution or side reactions. At 2 millimolar tyrosine, the optimized system converts 85.8 percent of the substrate into product, an efficiency few multi-step biosyntheses achieve in vitro. The results suggest a template for future work: pair automated, model-driven tuning of reaction conditions with rational physical organization of the catalytic machinery, and even pathways that have stalled in living cells can be made to sing in a test tube. For liquiritigenin, and potentially for the wider family of medicinal flavonoids it anchors, the bottleneck may finally be beginning to break.

Subject of Research: Machine learning-guided optimization and scaffold-assisted spatial assembly of a cell-free multi-enzyme system for liquiritigenin biosynthesis

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

Article References: Liu, F., Zhao, S.-B., Liu, Y.-H., Li, J.-F., Lin, N.-Q., Mutailifu, M., Xu, P., & Liu, J.-Z. (2026). Machine learning-driving optimization and spatial assembly of a cell-free system for high-yield liquiritigenin production. Advanced Biotechnology, 4(2), Article 12. https://doi.org/10.1007/s44307-026-00103-0

Image Credits: AI Generated

DOI: 10.1007/s44307-026-00103-0

Keywords: liquiritigenin, cell-free biosynthesis, machine learning, Bayesian optimization, enzyme scaffolds, SpyTag/SpyCatcher, gamma-prefoldin, flavonoid biosynthesis, metabolic engineering, CFPS-ME, spatial enzyme assembly, biocatalysis