Machine learning has delivered a green-chemistry win for the pharmaceutical industry, as researchers in China report the discovery of a bacterial enzyme that manufactures a key drug-building block with near-perfect precision on an industrial scale. Writing in the journal Catalysis Letters, a team from Hangzhou Weiyuan Biotechnology Co., Ltd. describes how a computational screening strategy pinpointed a transaminase enzyme from Arthrobacter sp., designated TA-6, that converts a simple ketone precursor into (R)-(+)-1-Boc-3-aminopyrrolidine with 99.5% enantiomeric excess in a 20-liter reaction vessel. The achievement is notable not only for the purity of the product but for the efficiency of the process: the whole-cell catalyst was used at a loading of just 10 grams per liter of dry cell weight while processing an extraordinary 100 grams per liter of ketone substrate, a substrate-to-biocatalyst ratio that few enzymatic processes of this kind can match.
The target molecule is far from an obscure laboratory curiosity. (R)-1-Boc-3-aminopyrrolidine serves as an indispensable chiral intermediate in the synthesis of central nervous system drugs and Janus kinase inhibitors, a class of anti-inflammatory therapeutics that includes blockbuster treatments for rheumatoid arthritis and other autoimmune conditions. Because the pyrrolidine ring bearing the Boc-protected amine appears in so many active pharmaceutical ingredients, drug manufacturers have long sought reliable ways to produce the single enantiomer they need. The trouble with traditional chemical synthesis is that it demands harsh conditions, generates substantial pollution, and often struggles to control the handedness of the product. When a molecule exists in two mirror-image forms, or enantiomers, biological systems typically respond very differently to each, so a drug intermediate contaminated with the wrong enantiomer can compromise efficacy or safety.
Transaminases offer an elegant way out of this predicament. These enzymes, which rely on the vitamin B6-derived cofactor pyridoxal-5′-phosphate (PLP), catalyze the transfer of an amino group from a donor molecule such as isopropylamine to a carbonyl acceptor, in this case the prochiral ketone N-Boc-3-pyrrolidinone. Because the enzyme’s active site holds the substrate in a defined orientation during the amination step, the amino group is delivered to one face of the molecule with exquisite selectivity, yielding almost exclusively the desired R-enantiomer. In principle, the reaction runs in water at ambient temperature and pressure, avoiding the heavy-metal catalysts, cryogenic conditions and hazardous reagents that plague conventional asymmetric amination chemistry.
The practical difficulty has always been finding the right enzyme. Nature offers thousands of transaminase variants scattered across the tree of life, and their performance on any given substrate is notoriously hard to predict. Enzymes that work beautifully on simple aliphatic ketones often fail completely on bulky, ring-containing substrates like N-Boc-3-pyrrolidinone, and even closely related enzymes can differ wildly in activity and stereoselectivity. The traditional approach—cloning, expressing and testing candidates one at a time in the laboratory—is slow, expensive and frequently fruitless. It is precisely this bottleneck that the Hangzhou team set out to break by putting machine learning at the front end of the discovery pipeline.
The researchers assembled a panel of seven phylogenetically diverse transaminases, spanning a broad swath of sequence space in the hope of capturing the full range of catalytic behaviors these enzymes can exhibit. Rather than subjecting all seven to exhaustive experimental characterization, the team used machine learning models to evaluate each candidate computationally, integrating structural and sequence-derived features to predict which enzyme was most likely to deliver high activity and R-selectivity on the pyrrolidinone substrate. The computational workflow drew on tools that have become standard in modern enzyme informatics: AlphaFold-type structure prediction to generate reliable three-dimensional models of each enzyme, molecular docking with software such as AutoDock Vina to position the substrate within the PLP-dependent active site, and molecular dynamics simulations to assess how stably the productive binding pose is maintained in solution. Visual analysis tools including VMD and PyMOL allowed the team to inspect the binding pockets and rationalize the predictions at the level of individual amino acid residues.
When the predictions were put to the test in the laboratory, one candidate stood out decisively. TA-6, an enzyme sourced from an Arthrobacter species—soil bacteria famous for their metabolic versatility—proved to be an outstanding R-selective biocatalyst for the asymmetric amination of N-Boc-3-pyrrolidinone. Kinetic analysis using HPLC-based assays confirmed that the enzyme combined high catalytic efficiency with the strict stereoselectivity the project demanded. The machine learning-guided screening had effectively compressed what would normally be months of trial-and-error enzyme testing into a rapid, data-driven selection process, identifying a single winner from a phylogenetically diverse field on the strength of computational evidence alone.
Perhaps the most striking aspect of the work is that it did not stop at a milliliter-scale curiosity in a shaking flask. The team carried their lead enzyme all the way to preparative scale, a stage where many promising biocatalysts fall apart. At 20 liters, the reaction presents challenges that small-scale experiments never reveal: the ketone substrate and its amine product can inhibit or even denature the enzyme, oxygen transfer and mixing become limiting, and the economics of catalyst loading come under intense scrutiny. Yet TA-6, deployed as a whole-cell catalyst—meaning the engineered bacteria expressing the enzyme are used directly, with all their cofactor regeneration machinery intact—handled 100 grams per liter of ketone substrate while making up only 10 grams per liter of dry cell weight. The whole-cell format is a significant practical advantage on its own, because the host organism regenerates the PLP and pyridoxamine phosphate cofactors internally, sparing process chemists from the expense of adding stoichiometric cofactor and the complexity of external recycling systems.
The outcome of the scaled reaction speaks for itself: the desired (R)-enantiomer was delivered in 99.5% enantiomeric excess, a purity level that meets the exacting demands of pharmaceutical manufacturing without any additional resolution or purification steps to strip out the unwanted mirror image. For a chiral amine intermediate destined for central nervous system drugs and JAK inhibitors, achieving that level of stereocontrol directly in the production vessel, in water, at ambient conditions, represents a genuine process-intensification milestone. The route eliminates the harsh conditions and heavy pollution associated with the traditional chemical synthesis of this intermediate, aligning with the broader push across the fine-chemicals industry toward sustainable, enzymatic manufacturing.
Beyond the immediate application, the study carries a broader message about how enzyme discovery is changing. The authors frame their work as evidence of a paradigm shift from trial-and-error to data-driven precision in asymmetric synthesis. Rather than relying on intuition, sequence homology alone or brute-force screening, the combination of machine learning evaluation, structural modeling and targeted experimental validation allowed a small team to move from a pool of diverse candidate enzymes to a production-ready biocatalyst in a single coherent workflow. The approach echoes a growing body of literature on machine learning-assisted enzyme engineering, in which predictive models trained on sequence, structure and kinetic data are increasingly able to rank variants before any wet-lab work begins, and it complements parallel advances in automated enzyme engineering and structure-guided transaminase redesign reported by other groups in recent years.
The Hangzhou researchers, led by corresponding author Zhibo Luo, suggest that the strategy should be readily transferable to other chiral amines of pharmaceutical interest, particularly bulky N-heterocyclic amines that have historically challenged both chemocatalysts and biocatalysts. As biocatalytic routes to chiral compounds continue to displace legacy chemical processes across the industry, the ability to identify the right enzyme quickly—and to trust that it will perform at scale—may prove to be the decisive competitive advantage. With TA-6, the team has shown that a machine learning-guided search through biodiversity can end not in a lab notebook but in a 20-liter tank producing a drug intermediate at pharmaceutical grade.
Subject of Research: Machine learning-guided discovery of an R-selective transaminase (TA-6 from Arthrobacter sp.) for the scalable biocatalytic synthesis of (R)-(+)-1-Boc-3-aminopyrrolidine from N-Boc-3-pyrrolidinone.
Subject of Research: Chemistry
Article Title: From Computational Screening to Preparative Scale: A Transaminase Identified by Machine Learning for (R)-1-Boc-3-aminopyrrolidine
Article References: Duan, M., Lu, X., Wu, Y., Huang, P., Zheng, W., Zhang, S., Liu, K., Jia, C., Du, F., Cao, M., Ma, M., & Luo, Z. (2026). From Computational Screening to Preparative Scale: A Transaminase Identified by Machine Learning for (R)-1-Boc-3-aminopyrrolidine. Catalysis Letters, 156(9), Article 240. https://doi.org/10.1007/s10562-026-05489-z
Image Credits: AI Generated
DOI: 10.1007/s10562-026-05489-z
Keywords: Data-driven biocatalysis, machine learning, transaminase, (R)-1-Boc-3-aminopyrrolidine, scalable asymmetric synthesis, N-Boc-3-pyrrolidinone, enantiomeric excess, whole-cell catalysis, pyridoxal-5′-phosphate, chiral amine pharmaceuticals, enzyme discovery, preparative-scale biocatalysis
Cite Scienmag News
APA
MLA
Chicago
Bethany Barker. (September 8, 2026). Machine learning finds transaminase for making chiral aminopyrrolidine at scale. Scienmag. https://scienmag.com/machine-learning-finds-transaminase-for-making-chiral-aminopyrrolidine-at-scale/
Bethany Barker. “Machine learning finds transaminase for making chiral aminopyrrolidine at scale.” Scienmag, 8 September 2026, https://scienmag.com/machine-learning-finds-transaminase-for-making-chiral-aminopyrrolidine-at-scale/. Accessed 8 September 2026.
Bethany Barker. “Machine learning finds transaminase for making chiral aminopyrrolidine at scale.” Scienmag. September 8, 2026. https://scienmag.com/machine-learning-finds-transaminase-for-making-chiral-aminopyrrolidine-at-scale/
Copy citation
Download RIS
Tags: aminopyrrolidine synthesisArthrobacter sp. enzyme applicationbacterial transaminase enzyme for chiral amine synthesisbacterial transaminase enzymesbiocatalyst efficiency in drug synthesischiral amine productioncomputational enzyme screeningcomputational screening for biocatalyst identificationenantiomeric purity in drug intermediatesenantiomeric purity in pharmaceutical intermediatesenzyme discoveryenzyme-driven production of chiral intermedigreen chemistry in enzyme catalysisgreen chemistry pharmaceutical manufacturinghigh substrate-to-catalyst ratio in bioprocessesindustrial biocatalysisindustrial-scale biocatalysis in drug manufacturinglarge-scale enzyme catalysismachine learning in drug synthesismachine learning in pharmaceutical enzyme discoverysustainable pharmaceutical processessynthesis of (R)-1-Boc-3-aminopyrrolidine
