Biocatalytic reductive amination from discovery to commercial manufacturing applied to abrocitinib JAK1 inhibitor

Enzymatic reductive amination, being a direct, selective and green methodology, has attracted significant interest in a short period of time and is emerging as a powerful tool for the synthesis of chiral alkylated amines. The discovery of an increasing number of imine reductases with reductive aminase (RedAm) activity has enabled mechanistic and substrate profiling studies. However, their potential for commercial applications has not been realized. Here, we report the discovery of RedAm activity in an imine reductase enzyme for the direct reductive amination of a cyclic ketone with methylamine. We also investigate engineering the enzyme to access a cis-cyclobutyl-N-methylamine for the manufacturing of a late-stage drug candidate, Janus kinase 1 (JAK1) inhibitor abrocitinib. The engineered enzyme, SpRedAm-R3-V6, showed >200-fold improvement in performance over the wild-type enzyme and was successfully used to develop a commercial manufacturing process with 73% isolated yield at 99.5% purity and high selectivity (>99:1 cis:trans). This process has been successfully used to manufacture multi-metric tons of the amine, demonstrating the potential of RedAm technology for commercial manufacturing. Reductive aminases show strong potential for the sustainable synthesis of chiral amines, but their application in industrial scale processes is lacking. Now, such an enzyme is screened and engineered allowing its use in commercial manufacturing of abrocitinib JAK1 inhibitor in multi-metric tons.

E fficient and sustainable synthesis of chiral amines, a promi nent synthetic motif in drug molecules 1 , has spurred the recent advances in the development of innovative and sustain able synthetic methods, including both traditional chemical [1][2][3][4] and enzymatic methods [5][6][7][8][9] . Reductive amination, being highly versatile, is one of the most frequently used transformations for the synthesis of a variety of amines 2 . Chemical methods for reductive amination commonly use stoichiometric amounts of reducing agents at low temperature, or transition metal catalysis, making them unsafe at scale and environmentally unsustainable 1,2 . Enzymatic methods are more desirable 7,10-13 for organic synthesis and manufacturing due to their potential for high selectivity, coupled with safety and environ mental benefits. Enzymatic synthesis of primary amines has been well established 5,6 using transaminases 14,15 or amino acid and amine dehydrogenases 8,9 . Recently, considerable progress has been made to evolve amino acid and amine dehydrogenases to perform reduc tive amination of ketones with ammonia and in some cases with methylamine 16 for preparative scale applications. Although signifi cant advances have been made for the enzymatic synthesis of primary amines, reductive amination of ketones with alkylamines remains a challenge. One of the recent innovations in enzyme catalysis is the discovery of imine reductases (IREDs) [17][18][19][20][21][22] for catalysing the reduction of C=N bonds to give amines. The ability of a subfamily of IREDs classified as reductive aminases (RedAms) to perform reductive amination of a ketone or aldehyde with an alkylamine to give alkylated amines is transformative, as this requires concurrent binding of both substrates in the active site with proper orienta tion [23][24][25][26] . Enzymatic reductive amination has attracted significant interest in a short period of time, with several studies reporting on the potential reaction mechanism for in situ imine formation fol lowed by reduction and identification of new enzymes with broad tolerance for carbonyl substrates and amine nucleophiles 17,25,27 . More recently, an IRED enzyme was successfully engineered and was applied to greater than 100 g scale for the reductive amination of an aldehyde to resolve an amine 28 , further underscoring the poten tial value of RedAm technology. A versatile biocatalytic reductive amination could revolutionize the synthesis of chiral alkylamines, by providing a direct, selective, safe, green and sustainable alterna tive to traditional methods.
Abrocitinib belongs to a group of Janus kinase (JAK) inhibi tors and is in late stage development as a JAK1 inhibitor for the treatment of atopic dermatitis 29 . Abrocitinib is structurally related to tofacitinib (US Food and Drug Administration approved for the treatment of rheumatoid arthritis) and contains a unique Nmethylaminesubstituted ciscyclobutane headpiece 1 (Fig. 1). In early synthetic approaches to abrocitinib, the amine was syn thesized by a chemical reductive amination at low temperatures, providing an approximately 80:20 mixture of cis:trans isomers. The desired cis isomer 1 could be obtained after multiple crystal lizations to purge the undesired diastereomer. While effective for synthesis of smaller amounts of abrocitinib to support clinical stud ies, a more efficient synthetic route, including access to the key ciscyclobutylNmethylamine 1 is required for commercial scale manufacturing. We envisioned the application of RedAm technol ogy to address the structural complexity of the N-methylamine head piece 1 to support the commercial manufacturing of abrocitinib.
Enzymatic reductive amination reactions are reported to be highly stereoselective in creating new chiral centres 30 and in some cases are able to control the stereoselectivity at adjacent chiral centres 31 . However, this study entails introduction of cisselectivity dur ing enzymatic reductive amination, which requires controlling the configuration of remotely lying substituents during reduction of an imine intermediate. Also, fourmembered cyclic ketones are reported to have much lower activity compared to fivemembered and sixmembered cyclic ketones 32 . Therefore this study was a sig nificant undertaking, and in the absence of any prior reports on the scalability of RedAm activity at the outset of this work, was an even larger challenge under an accelerated development timeline driven by the 'Breakthrough Therapy Designation' from the US Food and Drug Administration for abrocitinib.
In this Article, we report on the successful transition of RedAm technology from the initial identification of enzymatic activity to kilogram scale production and eventually to com mercial manufacturing for the synthesis of key intermediate 1 for the new drug, abrocitinib. Our production of the intermediate ciscyclobutylNmethylamine (1) is a successful demonstration of RedAm technology for the commercial scale synthesis of substituted amines by the reductive amination of a ketone with methylamine.

Results
Route selection and identification of RedAm enzyme. Reductive amination of ketones with alkylamines remains a challenge. Several drawbacks include incomplete reaction, overreaction to give bisalkylated byproducts and low selectivity leading to challenges in isolation and loss of yield. Transaminases have proven to efficiently transfer an ammonia equivalent, but require the sub sequent alkylation of the primary amine, which often results in overalkylation and the use of potentially genotoxic alkylating reagents 33,34 . A direct reductive amination was identified as the most efficient method to introduce a methylamine moiety to synthesize the Nmethylaminesubstituted ciscyclobutane headpiece 1 (Fig. 1).
A screen of the Pfizer inhouse enzyme panel, consisting of over 80 wildtype IRED enzymes from various sources, was performed to identify an enzyme capable of performing reductive amination of ketone 2 with methylamine to give the desired Nmethylamine 1. Several enzymes were identified with reductive amination activ ity that resulted in the formation of the desired cis isomer 1. A few enzymes also showed activity for the undesired trans isomer (Supplementary Table 1). The three best enzyme hits from initial screening were retested to confirm their performance, resulting in the selection of SpRedAm from Streptomyces purpureus 18 as the best candidate with both reductive aminase activity and high selectivity (diastereometric ratio (d.r.) > 99:1) for the desired cis isomer 1. The reaction with SpRedAm was scaled to 7.5 g using a substrate load ing of 20 g l −1 of 2 and 8 g l −1 of enzyme as a lysate, to give 27% iso lated yield of amine 1 with high selectivity for the desired cis isomer (d.r. > 99:1), confirming the performance seen in initial screening. Further testing of SpRedAm wildtype enzyme under more reason able process conditions (100 g l −1 ketone 2 and 1.5 g l −1 enzyme load ing) showed only 0.75% conversion (24 h) to product 1, but retained high selectivity. The significant drop in percentage conversion with increase in substrate ketone 2 loading from 20 to 100 g l −1 , suggested  a low substrate tolerance for wildtype SpRedAm enzyme and it was not suitable for targeted 100 g l −1 substrate loading for a commer cial process. In addition, the wildtype enzyme SpRedAm showed a narrow range of pH (7)(8) and temperature (25-30 °C) for opti mal activity, which could be challenging to manage in commercial manufacturing facilities.
Enzyme engineering. After initial evaluation, over 200aggregate fold (~127× in activity and 2× methylamine tolerance, see Table 1) improvement in enzyme performance over the parent wildtype (WT) SpRedAm was needed to enable a commercial manufactur ing process. In addition, to ensure robust performance at scale a broader window for operational stability (pH and temperature) was highly desirable. Therefore, we pursued enzyme engineering to improve the performance of the SpRedAm wildtype enzyme, targeting increased substrate tolerance and activity while retaining high selectivity (see the estimated process targets for commercial manufacturing, Table 1).
A multipronged approach for enzyme engineering was designed and applied in view of an accelerated development timeline. This included a computationalbased and a bioinformaticsbased approach for initial site selection and library design, coupled with a datadriven approach to identify hot spots and their synergistic recombination to achieve the desired performance.
Initial library design and site selection were done using struc tural homology models as there was limited mechanistic and structural information available for this emerging class of enzyme. A total of 93 sites (out of 296) were selected for single site satura tion mutagenesis (SSM), covering both binding site and second ary shell residues. Overall, 34 binding site residues, 55 secondary shell residues and 4 additional sites identified by bioinformatics, were selected for SSM library synthesis. SSM libraries from the first round of enzyme engineering were initially screened at a low sub strate concentration (20 g l −1 of ketone 2) and were progressively increased in subsequent rounds of screening. This resulted in the rapid identification of over 20 improved variants with single amino acid substitution, representing 12 amino acid residues as hot spots for improved activity (Fig. 2). The most active variants exhibited up to fivefold improvement over the parent (FIOP) and included the mutations Q13R, N131H, A170C/M, M176A, F180M and F214I/N. Screening of the remainder of the SSM libraries, comprised of sec ondary shell residues, identified an additional 12 variants, from 10 additional amino acid residues, with up to 3fold increased activity. The most notable variants included D220R, G242I and D250H (see Fig. 2 for screening data). The top hits from round 1 (Fig. 3a) were further challenged by retesting at higher substrate loading of 50 g l −1 (Fig. 3b). Interestingly, many top hits (Q13R, N131H, A170C/M, F180M and F214N) retained their improved activity (fourfold to fivefold over WT), suggesting increased tolerance for higher sub strate and methylamine concentrations. Variants A170C/M, F180M and F214N showed the highest increase in activity and substrate tolerance, resulting in an aggregated improvement of approximately 20fold.
The top five variants from round 1 were tested at multigram scale at 50 g l −1 loading of 2 with 8 g l −1 enzyme lysate and gave >75% conversion over 96 h with >99:1 cis:trans isomers. This represented a significant improvement in enzyme performance for substrate tolerance, given the final substrate loading target of 100 g l −1 . In parallel, we also started to evaluate various process parameters to provide insights into the design of the next set of screening conditions. Computational analysis of the top hot spots from round 1 of enzyme engineering showed that most were associated with either the enzyme active site or cofactor nicotin amide adenine dinucleotide phosphate reduced (NADPH) binding pocket (Fig. 2b,c).
In round 2 of enzyme engineering, recombination of beneficial mutations was performed and they were screened under more strin gent conditions at 75 g l −1 (480 mM) of substrate loading using 2 equiv. of methylamine, close to the desired process targets for a scalable process (Fig. 3b). Initially, targeted recombination was per formed to produce double and triple mutation variants. Several double mutants (Q13R/A170M, Q13R/N131H and A170C/F180M) showed up to threefold improvement in percentage conversion over their single mutation parents with retention of high selectivity (Fig. 3a). Double mutant Q13R/F214I showed a threefold improvement in percentage conversion. However, it resulted in a drop in selectivity and gave a cis:trans isomer ratio of 98.5:1.5. Addition of F214I/N to A170C or F180M, individually, showed a twofold improvement. However, their combination as a triple mutant was nonsynergistic and resulted in only a modest increase in activity (Fig. 3a).
Computational analysis showed amino acid residue 214 is part of the active site (chain B), which is formed at the interface between two chains and is adjacent to other binding site residues 180 and 176 (from chain A, Fig. 4). This provided a useful insight for future recombination of hot spots and residue 214 was evaluated by both ISM (ref. 35 ) and CASTing (ref. 36 ), using multiple parental tem plates. Double variants N131H/A170C and A170C/F180M showed the highest performance and were selected to progress further by targeted mutagenesis. Based on the crystal structure of SpRedAm, both the 170 and 180 amino acid residues are present on the same αhelix and residue 180 is part of the binding site. On the contrary, amino acid residue 170 is a secondary shell residue which is 8 Å away from the binding site and is oriented parallel or slightly away from the site. As the number of positive mutations increased, a multisite random recombination approach was also introduced. Several of the most active double and triple mutation variants were used as parental templates to create diverse variant libraries, in reac tions including up to 12 additional mutations 37 . Screening of several hundred variants from the resultant libraries identified improved variants with four to six mutations. Screening of all the libraries from round 2 was performed at 75 g l −1 and any identified hits were retested under multiple screening conditions (increasing substrate and cosubstrate loadings).
In round 3, hits obtained in round 1 were recombined with mul tiple approaches to generate various combinations. In this round, an additional 20 plates containing various combinations of mutations were screened and analysed. This resulted in identification of mul tiple variants with an additional fourfold to fivefold improvement in enzyme performance. The key variants from each round were further characterized to calculate kinetic parameters K cat (catalytic constant), K m (Michaelis constant) and turnover numbers (TONs) (Fig. 5). The final variant R3V6 from round 3 with substitutions N131H, A170C, F180M and G217D was selected for reaction and process loading, glucose loading, cofactor NADP+ loading and methylamine concentration were screened to identify optimum reaction condi tions to achieve high conversion and selectivity. The engineered enzyme showed high selectivity and specificity for the desired amine 1, controlling the formation of the only observed byproduct, the trans isomer of 1, to <0.5%. An increase in temperature and pH resulted in higher conversion. However, reaction temperature above 30 °C and maintaining pH > 8 resulted in an increased rate of hydro lysis of the isopropyl ester of both ketone 2 and product amine 1.
Hydrolysis was controlled by maintaining the reaction temperature between 20 and 30 °C and the pH between 6 and 8. Enzyme loading studies showed an increase in percentage conversion to the prod uct with higher enzyme loadings and the loading was optimized to 1.5 wt% to enable efficient enzyme removal downstream for prod uct isolation. The optimized laboratory process was further refined and engineered to fit the scale and manufacturing equipment train and was successfully scaled from the laboratory (gram scale) to kilo laboratory (1-10 kg of 2 per batch), pilot plant (50-100 kg of 2 per batch) and finally commercial manufacturing plant (>200 kg of 2 per batch). Consistent reaction performance of >91% conversion in 48 h was observed irrespective of scale (Fig. 6a).
Under optimized process conditions, the engineered enzyme SpRedAmR3V6 showed on average ~77% conversion (versus 0.75% with WT) after 24 h (see Fig. 6a), resulting in a 103× improve ment in enzyme performance. In addition, it showed 2× improve ment in tolerance for methylamine since the final process was run with >1 M of methylamine (versus 0.5 M for wildtype SpRedAm), leading to an overall improvement of 206fold (103 × 2) over wild type. The product amine 1 was isolated by removal of the enzyme and extraction at pH > 11.5 using methyl tertbutyl ether (MTBE) and then crystallized as the succinate salt. A total of >3.5 MT of amine 1 as the succinate salt was manufactured in >99% purity and >99:1 cis:trans selectivity. representing the successful implemen tation of RedAm technology on a commercial manufacturing scale (Fig. 6b).

Conclusions
RedAm technology was successfully applied for the commercial scale manufacturing of a secondary amine via direct reductive ami nation of a ketone with methylamine. This was accomplished by discovering an IRED enzyme with the desired RedAm activity cou pled with stateoftheart enzyme engineering and highthroughput   screening to generate an engineered enzyme SpRedAmR3V6 to provide an overall performance improvement of >200fold over the wild type. The engineered enzyme SpRedAmR3V6 was success fully implemented in the commercial process to give a space-time yield of 60 g per litre per day with high purity (>99.5%) and selec tivity (>99:1 cis:trans) to access the amine 1 required to synthesize abrocitinib. The successful transition of initial laboratory activity to commercial industrial scale manufacturing was demonstrated under significantly accelerated timelines, highlighting the potential of rapid development of enzyme catalysis as a competitive green alternative to traditional chemical methods. This work contributes to the rapidly developing field of enzyme catalysis that is emerging as a critical strategy for pharmaceutical and fine chemical manufac turing as a sustainable alternative to existing methods.

Methods
All reagents and solvents used in this study were purchased from commercial suppliers and were used as received, unless specified. NADP+ was purchased from Oriental Yeast Co., GDH (CDX 901) enzyme was purchased from Codexis Inc. Engineered enzyme SpRedAmR3V6 at large scale was custom produced by commercial enzyme producers.
Identification and cloning of enzymes in screening panel. The Pfizer IRED screening panel included various wildtype IRED enzymes from multiple sources 18,19,25,27 . Enzyme identification and cloning of various wildtype enzymes included in the Pfizer IRED screening panel was previously published 18,25 . Frozen cell pellets were thawed on ice and then fully suspended at 120 mg ml −1 in Bug Buster Master Mix (Millipore Sigma 71456). Samples were incubated at 4 °C for 60 min. Lysates were then clarified by centrifugation at 5,000g for 15 min. Finally, 50 μl of clarified lysates were transferred to 96well plates and frozen at −80 °C.

IRED panel screening.
Screening protocol A (substrate loading 10 mg ml −1 ). Screening was performed in 2 ml, 96well deepwell plates using enzyme lysate. Screening panels were prepared and stored at −80 °C. For screening, enzyme panels were removed from the freezer and allowed to warm at room temperature for 30 min before use.
To each well of a screening plate (2 ml deepwell plate) containing enzyme lysate (50 μl) was added 190 μl of reaction stock solution, which was preadjusted to pH 7.25 using 50% hydrochloric acid. The reaction stock solution contained 0.6 mM NADP+, 0.25 mg ml −1 GDH, 70 mM glucose and 114 mM methylamine in aqueous phosphate buffer (pH 7.2, 100 mM). The substrate (64 mM) stock solution (10 μl) was added to each well as a solution in dimethylsulfoxide (DMSO) (250 mg dissolved in 1 ml DMSO). The plate was sealed and incubated on an Eppendorf shaker at 30 °C and 900 rpm. After 20 h of incubation, 1 ml of acetonitrile was added to each well to quench the reaction and the plate was centrifuged at 4,000 rpm for 10 min. After centrifugation, 400 μl of supernatant was transferred to another 1 ml 96well plate and analysed using liquid chromatography-mass spectroscopy (LCMS).

Bioinformatics. Template identification and homology model construction.
At the start of this engineering project no experimentally determined structures for SpRedAm were available. The Chemical Computing Group's Molecular Operating Environment (MOE) suite of tools was used to calculate amino acid conservation rates at each position along the protein backbone using an alignment file containing the closest 250 nonredundant sequences identified through a BLAST search. A template search for homology modelling was performed in MOE as well as a BLAST search of Protein Data Bank proteins. The selected template, 3ZHB from Streptomyces kanamyceticus, resulted in an alignment with 52% identity to SpRedAm and a pairwise percentage positive (BLSM62) of 61%. A homology model was constructed using MOE. For this model, the conserved cofactor binding motif GxGxxG was constrained between the template and SpRedAm sequences, as were the highly conserved active site residues corresponding to Asp169 and Trp177.
Identification of positions for site saturation mutagenesis. Binding site residues in the homology model were identified using the Site Finder function in MOE. Secondary shell residues were identified by selecting residues within 4.5 Ǻ of binding site residues. All residues were sorted and scored by calculated conservation rates, variability of amino acids seen in the alignment mentioned previously (number and chemical characteristics) and distance from the binding site. Ninetythree positions were selected for site saturation mutagenesis library creation, including all 34 of the identified binding site residues. The remaining positions comprised nonconserved secondary shell and protein surface residues. Substrate docking. Once an experimentally determined structure of SpRedAmR3V6 was determined, MOE was used to first identify the binding site using the Site Finder function and then the product amine was docked using the General Docking function with an Induced Fit refinement.
Enzyme engineering (library synthesis, assembly and expression). Single site saturation variant libraries (93 residues), consisting of 34 active site residues, 55 secondary shell residues and 4 additional residues were identified by bioinformatics.
SpRedAm gene variant libraries were synthesized by Twist Bioscience. Twist Bioscience was supplied with a codonoptimized gene sequence which also included 30 base pairs of 5′ and 3′ flanking DNA sequences to allow cloning of the linear DNA libraries, into an expression plasmid, using Gibson Assembly. The single site saturation variant libraries were delivered/received as lyophilized linear DNA, in a 96well plate (one targeted position per residue per well). The library pools were rehydrated with 80 μl TE buffer (pH 7.0).   plasmid by Gibson Assembly. The PCRamplified variant library DNA was purified using Qiagen PCR purification kit reagents and the DNA was eluted with water. The library pools were cloned into the pET28b vector (EMD Biosciences), which had been linearized with restriction enzymes (NEB) and purified with Qiagen PCR purification kit reagents. Hifi DNA assembly reagents (NEB) were used to perform the variant library cloning step. Large scale analysis. Scaleup reactions from laboratory scale to kilolab scale and to commercial manufacturing were analysed by a direct injection gas chromatograhy (GC) method using an Agilent 6890 with a flame ionization detector. The GC was equipped with a DB5MS, 60 m x 0.25 mm × 1.0 μm (Agilent Part No. 1225563) column and helium as a carrier gas at a constant flow of 1.0 ml min −1 (pressure of ~24 psi). Analysis was performed with an inlet temperature of 250 °C with a split mode using a split ratio of 20:1 and a detector temperature of 300 °C. The oven was programmed with 100 °C as the initial temperature for 0 min, followed by a ramp at 5 °C min −1 to 210 °C and then held for 0 min, followed by another ramp at 20 °C per min to 325 °C with a hold for 6.25 min. The detector flow rate was 40 ml min −1 for hydrogen and 400 ml min −1 for air, with constant makeup using helium or nitrogen at 15 ml min −1 . The total run time was 34 min (Supplementary Fig. 4: RRTSM ketone (2) 17.554 min, product desired cis amine (1) 20.03 min, trans amine (3) 20.19 min).

Multi-kilogram scale reductive amination reaction with engineered SpRedAm-V6.
To a reactor at 25 °C was charged water (1840 l, 8 l kg −1 of limiting reactant, LR) and dipotassium hydrogen phosphate (23 kg, 0.09 equiv.) followed by methylamine hydrochloride (149 kg, 1.5 equiv.) and glucose monohydrate (344 kg, 1.3 equiv.). The reaction mixture was stirred until homogeneous. Ketone 2 (230 kg, LR) was charged to the reactor. NADP+ (0.6 kg, 0.25 wt% relative to 2), GDH (0.6 kg, 0.25 wt% relative to 2) and the custom enzyme SpRedAmR3V6 (3.4 kg, 1.5 wt% relative to 2) were added. The reaction mixture was held at 25 °C for 48 h using a pH dosing unit to maintain a constant pH of 7 by titration with 20% aqueous sodium hydroxide solution. When the reaction was complete, the reactor was cooled to 3 °C and the pH was adjusted to 2.3-3.3 with 37% aqueous hydrochloric acid (80 kg, 0.6 equiv.). After the mixture was stirred for 30 min, carbon Darco G60 powder (46 kg, 20 wt% relative to 2) was charged to the reactor and the suspension was stirred for an additional 30 min. The carbon was filtered through a layer of Celite and rinsed with water (230 l, 1 l kg −1 , LR). MTBE (4,100 l, 20 l kg −1 , LR) was charged to the reactor and the mixture was cooled to 5 °C, after which 20 wt% aqueous NaOH (551 kg, 2.1 equiv.) was added. The phases were split and the organic layer was collected and concentrated under reduced pressure (500 mbar) to a final volume of 1,150 l. Fresh MTBE (1,150 l) was added. In a separate vessel, MTBE (2,300 l, 10 l kg −1 , LR) and succinic acid (139 kg, 0.90 equiv.) were charged to the reactor at 20 °C. Seed crystals of succinate salt of 1 (2.3 kg, 1 wt% relative to 2) were added, followed by the transfer of the MTBE solution to the succinic acid slurry. The resulting slurry was granulated for 1 h at 20 °C. The solids were filtered off, rinsed with MTBE (1,150 l, 5 l kg −1 , LR) and dried in a vacuum oven at 40 °C to afford the desired amine succinate salt 1 (311 kg, 73% yield) as a white solid. 1

Data availability
Additional data supporting the findings reported in this paper are available as Supplementary Information. All other data are available from the authors upon reasonable request. X X