Precision evolution of transaminase under machine learning guidance and its application

Abstract

This invention discloses a machine learning-guided directed evolution method for ω-transaminases and its application in the synthesis of key chiral amine intermediates in retinoic acid. Using wild-type transaminase CA113 as the starting enzyme, consensus sequences were generated through multiple sequence alignment (MSA). Combined with 5-ns alanine scanning molecular dynamics (MD) simulations and CAVER channel analysis, 18 single-point mutations were screened to construct a mini-library (library capacity 48). The steady-state kinetic parameters (kcat / Km) and melting temperature (Tm) of the mutants were uploaded to a small-sample transfer learning regression model based on a pre-trained algorithm for training. Through two rounds of iterative cycles of "wet experiment-model training-virtual screening," the optimal double mutant M5 (R68F / D125V) was obtained. This mutant, under the conditions of a 1 L reactor, a final substrate concentration of 25 mg·mL⁻¹, an enzyme loading of 1 g·L⁻¹, 25 °C, and pH 9.0, achieved a space-time yield of 92 g·L⁻¹·d⁻¹, a separation yield of 81%, an ee > 99%, and RSD < 3% for three consecutive batches prepared at the 100g scale, meeting API impurity profile requirements. This invention provides a reproducible new paradigm for green, efficient, and low-cost pharmaceutical enzyme engineering.

Description

Technical Field

[0001] This invention belongs to the fields of enzyme engineering and directed protein evolution technology, specifically relating to a machine learning-guided method for directed evolution of ω-transaminase, and the industrial application of highly active ω-transaminase mutants obtained by this method in the synthesis of key chiral amine intermediates of retinoic acid. Background Technology

[0002] In recent years, the rapid development of machine learning technology has brought new opportunities to protein engineering. Traditional directed evolution methods suffer from problems such as large screening workload and blind mutation direction. They typically require the construction of large-scale libraries containing thousands to tens of thousands of mutants, and high-throughput screening to obtain beneficial mutations. The entire process is time-consuming, labor-intensive, and has a low success rate. In contrast, machine learning-guided directed evolution adopts a "dry-wet" strategy, using computational methods to quickly screen for beneficial mutations and establishing a feedback mechanism to correct deviations in a timely manner. This can significantly shorten the research and development cycle and improve the success rate.

[0003] ω-transaminases are enzymes that use pyridoxal-5'-phosphate (PLP) as a coenzyme to catalyze the transfer of amino groups from donor molecules to acceptor keto acids, thus synthesizing chiral amines. Compared with chemical catalysts, ω-transaminase catalysis offers significant advantages such as high catalytic efficiency, mild reaction conditions, low energy consumption, and less chemical pollution. Furthermore, their broad substrate range, high stereoselectivity, and strong reactivity make them promising candidates for the synthesis of drugs containing chiral amines. However, industrial catalysis using wild-type ω-transaminases often fails to achieve the desired catalytic effect, necessitating structural modification through protein engineering and other methods.

[0004] Rimegepant is a novel calcitonin gene-related peptide (CGRP) receptor antagonist used for the acute treatment of adult migraines. Its key chiral amine intermediate is a -(R)-3-amino-2-oxopiperidine derivative. Currently, the synthesis of this intermediate mainly relies on a chemical reductive amination route, which suffers from harsh reaction conditions, difficulty in controlling stereoselectivity, and severe environmental pollution. Therefore, developing a green and efficient enzyme-catalyzed process has significant industrial value. Summary of the Invention

[0005] To address the shortcomings of existing technologies, this invention provides a machine learning-guided method for directed evolution of ω-transaminases, as well as a highly active ω-transaminase mutant obtained by this method and its application in the synthesis of key chiral amine intermediates in retinoic acid.

[0006] The technical problems to be solved by this invention include: (1) Traditional directed evolution methods involve a large amount of screening work, blind mutation direction, and low success rate; (2) Wild-type ω-transaminase has insufficient catalytic activity for key chiral amine intermediates of Remegpam; (3) Existing enzyme catalysis processes are difficult to meet the requirements of industrial production.

[0007] To solve the above-mentioned technical problems, the present invention adopts the following technical solution:

[0008] This invention provides an ω-transaminase mutant, which, based on the amino acid sequence of the wild-type ω-transaminase CA113 shown in SEQ ID NO: 1, contains at least one of the following mutation sites: R68F, D125V, I224V, I262A, H318A, M56F, H191F, V435I, T322C, T322N, T322A, D416R, D315I, E168I, or E168L. Preferably, the mutant is a double mutant M5 (R68F / D125V), whose amino acid sequence is shown in SEQ ID NO: 3.

[0009] Preferably, the nucleotide sequence corresponding to amino acid SEQ ID NO: 1 is SEQ ID NO: 2, and the nucleotide sequence corresponding to amino acid SEQ ID NO: 3 is SEQ ID NO: 4.

[0010] This invention also provides a machine learning-guided directed evolution method for ω-transaminase, comprising the following steps: (a) using wild-type ω-transaminase as the starting enzyme, generating consensus sequences through multiple sequence alignment, and screening potential hotspot sites by combining 5-ns alanine scanning molecular dynamics simulation and CAVER channel analysis; (b) constructing a miniature mutant library (library capacity 40-60) containing the screened mutant sites; (c) measuring the enzymatic parameters of each mutant in the mutant library, using kcat / Km as the fitness index, and uploading it to a small-sample regression prediction module based on a pre-trained model for training; (d) performing virtual screening based on the trained model to predict mutants with an activity increase of ≥30%; (e) verifying the activity of the predicted mutants in wet experiments, and appending the verification data to the training set for iterative optimization; (f) repeating steps (d)-(e) for 2-3 rounds until convergence (R²pred-obs≥0.80) to obtain the optimal mutant.

[0011] Preferably, in step (a), bottleneck residues with a radius of <1.2 Å at the narrowest point of the substrate channel are identified.

[0012] Preferably, the training in step (c) uses ESM-1v or a similar ensemble learning algorithm for parameter fine-tuning and employs 5-fold cross-validation. When q²≥0.70, the virtual screening stage is entered.

[0013] The present invention also provides a nucleic acid molecule encoding the ω-transaminase mutant, a recombinant expression vector containing the nucleic acid molecule, and a recombinant host cell containing the recombinant expression vector.

[0014] The present invention also provides the use of the ω-transaminase mutant in the preparation of the key chiral amine intermediate (R)-3-amino-2-oxopiperidine derivative of retinoic acid.

[0015] Preferably, the transamination reaction is carried out in a 1 L reactor, with a final substrate concentration of 25 mg·mL⁻¹, an enzyme loading of 1 g·L⁻¹, a reaction temperature of 45 °C, and a pH of 9.0.

[0016] Compared with the prior art, the present invention has the following beneficial effects:

[0017] 1. Data-driven small-sample directed evolution: With only 96 mutants and two rounds of experiments, the conversion rate of wild-type CA113 to high-concentration retinoic acid intermediates was increased from about 8% to 81%, proving that the small-sample regression prediction module based on the pre-trained model can converge quickly in the "gram-level substrate load + small library" scenario, significantly reducing screening costs.

[0018] 2. Simultaneous optimization of multiple constraints: The AI ​​model simultaneously considers activity, thermal stability and substrate channel patency, which increases the Tm of the double mutant M5 by 4.3℃, extends the half-life at 50℃ by 6 times, and still has residual activity after reacting at 45℃ for several days, laying a thermally stable foundation for subsequent continuous flow reactions.

[0019] 3. Feasibility of green process: The reaction takes place in the aqueous phase at 45°C. o The process is carried out at C and normal pressure, with an enzyme loading as low as 1 g·L⁻¹, and the E-factor is reduced by 65% ​​compared to the chemical reduction amination route, which meets the requirements of the guidelines for green manufacturing of active pharmaceutical ingredients.

[0020] 4. Excellent industrial performance: The optimal mutant M5 achieved a space-time yield of 92 g·L⁻¹·d⁻¹ under 1 L reactor conditions, with a separation yield of 81%, ee>99%, and RSD<3% for three consecutive batches prepared in 100g scale, meeting the API impurity profile requirements. Attached Figure Description

[0021] Figure 1 The transamination reaction of the key chiral amine intermediate of Rimegepant, (R)-3-amino-2-oxopiperidine derivative; Figure 2 This is a schematic diagram illustrating the prediction of the active pocket of CA113. Figure 3 This is a schematic diagram of the CA113 consensus sequence; Figure 4This is a schematic diagram of molecular docking and alanine scanning results; Figure 5 A schematic diagram of the molecular dynamics simulation of CA113 and its flexible region; Figure 6 Schematic diagram of CA113 coenzyme binding pocket and channel analysis; Figure 7 For the construction of micro-libraries and virtual screening libraries; Figure 8 For the HPLC analysis of related substances in the catalytic system; Figure 9 The product after catalysis was analyzed by HPLC to determine its chirality. Detailed Implementation

[0022] The present invention will be further described in detail below with reference to specific embodiments. These embodiments are for illustrative purposes only and are not intended to limit the scope of the invention. Experimental methods in the following embodiments, unless otherwise specified, are generally performed under conventional conditions or as recommended by the manufacturer.

[0023] Example 1: Candidate site screening and micro-library construction

[0024] Using wild-type transaminase CA113 as the research object, this enzyme catalyzes the transamination reaction of the key chiral amine intermediate -(R)-3-amino-2-oxopiperidine derivative of retrimazole.

[0025] (1) Multiple sequence alignment and consensus sequence analysis: Consensus sequences are generated through multiple sequence alignment (MSA) to identify 10 potential hotspot sites, including A31, C47, F71, D125, R133, S164, F166, E168, S301, H320, T322 and other positions.

[0026] (2) Alanine scanning and molecular dynamics simulation: 5-ns alanine scanning molecular dynamics (MD) simulation was used to evaluate the effect of each mutation on the substrate Cα-N distance and the Asp-Arg catalytic dimmer angle. By analyzing the RMSF (root mean square fluctuation) results, key residues in the flexible region were identified, including G39, W60, R68, G120, Y153, K288, F86, and S223.

[0027] (3) Substrate channel analysis: CAVER software was used to analyze the substrate channels and identify bottleneck residues with a minimum radius of <1.2 Å, including D315, I262, H318, M56, H191, V435, T322, D416, E168, etc.

[0028] Based on the above analysis, a total of 18 sites were selected for constructing a micro-library, namely A31, G39, C47, W60, R68, F86, G120, D125, Y153, S164, E168, H191, S301, D315, H318, H320, T322, and D416, with a library size of 48.

[0029] Example 2: Expression and preparation of recombinant transaminases and their mutants

[0030] This embodiment refers to general molecular biology methods, combined with site-directed mutagenesis PCR technology, to complete the PCR mutation construction, recombinant expression, and enzyme solution preparation of the ω-transaminase mutant. The specific steps are as follows, ensuring that each step is repeatable and the parameters are clearly defined:

[0031] I. Construction of mutants by PCR site-directed mutagenesis

[0032] 1. Primer Design: In this embodiment, the amino acid sequence of wild-type ω-transaminase CA113 is shown in SEQ ID NO: 1, and its corresponding cDNA sequence is shown in SEQ ID NO: 2. Using the cDNA sequence of wild-type ω-transaminase CA113 shown in SEQ ID NO: 2 as a template, site-directed mutagenesis primers were designed targeting the mutation site. The primer design followed these principles: the mutation site was located in the middle of the primer, the primer length was 25-30 bp, the GC content was controlled at 45%-55%, the annealing temperature difference between the upstream and downstream primers did not exceed 2℃, and the primers had no hairpin structures, self-dimers, or cross-dimers at both ends.

[0033] 2. PCR Reaction System Preparation: In a sterile PCR tube, prepare a 25 μL reaction system according to the following proportions, with precise control of the amount of each component: 2.5 μL of 10×PCR Buffer (containing Mg²⁺), 2 μL of dNTP Mixture (each 2.5 mmol / L), 1 μL of forward mutation primer (10 μmol / L), 1 μL of reverse complementary primer (10 μmol / L), 1 μL of template DNA (wild-type CA113 cDNA, concentration 50 ng / μL), and 0.2 μL of high-fidelity DNA polymerase (5 U / μL). Add sterile deionized water to a final volume of 25 μL. Keep the entire preparation process on ice to avoid enzyme inactivation and non-specific primer binding.

[0034] 3. PCR reaction program settings: Place the PCR tube in the PCR instrument and set the following program for amplification to ensure efficient introduction of mutations and absence of nonspecific bands: ① Pre-denaturation: 95℃, 5min, to completely denature the template DNA; ② Cyclic amplification: 95℃ denaturation for 30s, annealing temperature (adjusted according to primer Tm value, usually 55-60℃) for 30s, extension at 72℃ (extension time adjusted according to the length of the target gene; the full length of the CA113 gene is approximately 1200bp, so the extension time is set to 1.5min), for a total of 30 cycles; ③ Final extension: 72℃, 10min, to ensure complete extension of the PCR product; ④ Incubation: 4℃, ∞, to prevent product degradation.

[0035] 4. PCR Product Validation and Purification: Take 5 μL of PCR product and validate it using 1.2% agarose gel electrophoresis (120V, 30 min). Use the DNA Marker (1000bp-5000bp) as a reference and observe whether a single, clear target band (approximately 1200bp) appears. The absence of any extraneous bands indicates successful amplification. For successfully amplified PCR products, purify them using a PCR product purification kit (centrifuge column type), strictly following the kit instructions: Add 5 volumes of binding buffer, mix well, load onto a centrifuge column, centrifuge at 12000 rpm for 1 min, and discard the waste liquid; add washing buffer, centrifuge at 12000 rpm for 1 min, discard the waste liquid, and repeat the washing once; finally, add 50 μL of sterile deionized water, incubate at room temperature for 2 min, centrifuge at 12000 rpm for 1 min, collect the purified PCR mutant product, and store at -20℃ for later use.

[0036] II. Construction and Screening of Recombinant Genetically Engineered Bacteria

[0037] 1. Transformation of competent cells: Take 10 μL of recombinant mutant plasmid and add it to 100 μL of E. coli BL21(DE3) competent cells, and incubate on ice for 30 min; then heat shock in a 42℃ water bath for 90 s, and quickly incubate on ice for 2 min to complete the transformation; add 900 μL of antibiotic-free LB liquid medium to the transformation system, and culture at 37℃ and 200 rpm for 1 h to restore the activity of competent cells and express the resistance gene.

[0038] 2. Positive clone screening: Spread 100 μL of the transformed solution evenly onto LB solid medium containing 50 μg / mL kanamycin in a dish. Incubate at 37°C with the medium inverted for 12-16 h until single colonies appear. Pick 3-5 single colonies and inoculate them into LB liquid medium containing 50 μg / mL kanamycin. Incubate at 37°C with shaking at 200 rpm for 8 h. Extract the recombinant plasmid from each colony. Verify the results using PCR (primers are universal primers for pET-28a(+) vector) and sequencing. Colonies whose sequencing results match the target mutant sequence are considered positive recombinant genetically engineered bacteria. Inoculate these colonies into LB medium and store at 4°C for later use.

[0039] III. Large-scale culture and induction expression of recombinant transaminase mutants

[0040] 1. Seed culture: Select a single colony of positive recombinant genetically engineered bacteria from LB medium and inoculate it into 50 mL of LB liquid medium containing 50 μg / mL kanamycin. Incubate at 37℃ and 200 rpm for 12 h to obtain seed culture.

[0041] 2. Expanding the culture: Inoculate the seed culture solution into 500 mL of TB liquid medium containing 50 μg / mL kanamycin at an inoculation rate of 1% (v / v). (TB medium components: 12 g tryptone, 24 g yeast extract, 4 mL glycerol, 2.31 g KH2PO4, 16.42 g K2HPO4·3H2O, add water to a final volume of 1000 mL, autoclave at 121℃ for 20 min). Incubate at 37℃ with shaking at 200 rpm, and measure the OD600 value every 2 h until the OD600 reaches 0.6-0.8.

[0042] 3. Induction of expression: Add IPTG (isopropyl-β-D-thiogalactoside) to the expanded bacterial culture to a final concentration of 0.2 mmol / L, adjust the culture temperature to 25℃, and induce culture at 200 rpm for 14-24 h.

[0043] IV. Extraction of crude enzyme solution from recombinant transaminase mutant (ultrasonic disruption method)

[0044] 1. Cell collection: Transfer the induced bacterial culture to a centrifuge tube, centrifuge at 8000 rpm for 10 min at 4 °C, discard the supernatant, and collect the cell pellet; wash the cell pellet twice with 100 mmol / L phosphate buffer (pH 8.0, containing 1 mmol / L EDTA and 0.5 mmol / L LTT, autoclaved at 121 °C for 20 min, and cooled to room temperature for later use). After each wash, centrifuge at 8000 rpm for 10 min at 4 °C, discard the supernatant, and remove culture medium residue and impurities.

[0045] 2. Bacterial resuspension: Add an appropriate amount of 100 mmol / L phosphate buffer (pH 8.0) to the washed bacterial pellet, and resuspend the bacterial pellet at a ratio of 1g of bacterial pellet to 10mL of buffer. Place the pellet in an ice bath and gently agitate it to prevent the bacterial pellet from clumping.

[0046] 3. Ultrasonic disruption: Place the resuspended bacterial solution in an ice bath and disrupt it using an ultrasonic cell disruptor. Set the parameters as follows: power 150~200W, working time 2s, interval time 3s, total number of cycles 150. Maintain the ice bath throughout the disruption process to control the bacterial solution temperature below 4℃ to prevent denaturation of the target protein. Samples can be taken during disruption to observe the state of the bacterial solution until it changes from turbid to clear, indicating complete bacterial disruption.

[0047] 4. Collection of crude enzyme solution: Transfer the broken bacterial culture to a centrifuge tube, centrifuge at 12000 rpm for 20 min at 4℃, and collect the supernatant, which is the crude enzyme solution of the recombinant transaminase mutant; the precipitate consists of bacterial fragments and unbroken bacterial cells, which should be discarded. Use the BCA protein quantification kit to detect the protein content in the crude enzyme solution, adjust the protein concentration to 1-2 mg / mL, aliquot and store at -80℃ for later use, or use directly for subsequent enzymatic property determination.

[0048] Example 3: Machine Learning-Guided Iterative Optimization of Dry and Wet Experiments and Obtaining the Double Mutant M5

[0049] I. First Round of Wet Experiments: Enzymatic Characterization of Single-Point Mutant Libraries

[0050] Eighteen single-point mutation sites (including redundant substitutions such as C / N / A at the T322 site) obtained through screening in Example 1 were used to construct a mini mutant library with a capacity of 48, combined with the wild-type control. Following the method described in Example 2, the 48 single mutants were subjected to site-directed mutagenesis PCR construction, BL21(DE3) transformation expression, and ultrasonic disruption to prepare crude enzyme solution. Steady-state kinetic parameters and melting temperatures were measured for both the 48 mutants and the wild-type CA113.

[0051] 1. Steady-state kinetic parameters (kcat / Km): Using the key chiral amine intermediate precursor ketone of retrimazole as the acceptor substrate and isopropylamine as the amino donor, the amount of acetone or product generated was continuously monitored by spectrophotometry under the conditions of 100 mmol / L phosphate buffer (pH 8.0, containing 0.1 mmol / L PLP) and 30℃. Five to eight different substrate concentration gradients were prepared, and kcat and Km were obtained by nonlinear fitting of the Michaelis-Menten equation, and kcat / Km was calculated.

[0052] 2. Melting temperature (Tm): Differential scanning fluorescence (DSF) was used with SYPRO Orange as the fluorescent probe. The melting curves of each mutant were determined at 25-95℃ and a heating rate of 1℃ / min. The temperature corresponding to the lowest point of the first derivative was taken as Tm.

[0053] The (site, kcat / Km, Tm) triplet data of 48 mutants were organized into a structured training set. The sequence / structural features of the mutation site were used as input features, and kcat / Km (after normalization) and Tm were used as regression targets.

[0054] II. First Round of Model Training and Virtual Screening

[0055] A pre-trained protein language model (ESM-1v or a similarly functional pre-trained model) was used as the feature extraction backbone. A few-shot transfer learning regression head was then integrated into its embedding representation. Only the regression head and top-level parameters were fine-tuned to avoid overfitting under small-sample conditions. A 5-fold cross-validation was used to evaluate the model's performance. After the first round of training, the cross-validation determination coefficient q² = 0.73, meeting the inclusion criterion of q² ≥ 0.70 as described in claim 10, and the model entered the virtual screening stage. Using the trained model, sites with high synergistic potential (from consensus sequence analysis, alanine scanning MD, and channel analysis, respectively) from the 18 single-point mutations in the first round were paired to construct a virtual double-mutant library. Each pair was scored and predicted, and candidate double mutants with a predicted activity (kcat / Km) ≥ 30% higher than wild-type and a predicted Tm not lower than wild-type were selected. These were ranked by prediction score, and approximately 48 of the top-ranked combinations were selected for the second round of wet experimental validation, totaling 96 mutants along with the 48 single mutants from the first round.

[0056] III. Second Round of Wet Experiment Verification

[0057] Following the method described in Example 2, the double mutants obtained from the first round of virtual screening were constructed, expressed, and purified by PCR site-directed mutagenesis, and the measured values ​​of kcat / Km and Tm were determined as described above. The predicted values ​​were compared with the measured values.

[0058] The results showed that the measured kcat / Km of the double mutant R68F / D125V (M5) was 9.8 times higher than that of the wild type, and the Tm was increased by 4.3℃. All indicators ranked first among all candidate double mutants in the second round, and the model's prediction of the activity enhancement of this mutant had the smallest deviation from the measured values. Therefore, the double mutant M5 (R68F / D125V) was determined to be the optimal mutant for this invention.

[0059] IV. Summary of the Iteration Process

[0060] This invention employs a two-round "dry-wet combination" iterative process: "first round of single-point mutation wet experiment determination → model training → virtual screening of double-mutant combinations → second round of wet experiment verification → data supplementation and model convergence." Using only 96 mutants in wet experiments, the optimal double mutant M5 was obtained from a mini-library of 48 single-point mutations. This verifies the rapid convergence capability of the method in the "small sample + gram-level substrate loading" scenario, providing direct mutant source evidence for the comprehensive characterization of enzymatic properties in Example 3 and the hundred-gram-level scale-up preparation in Example 5.

[0061] Example 4: Characterization of the enzymatic properties of mutants

[0062] Key enzymatic parameters of wild-type CA113 and mutant M5 were compared and determined, and the results are shown in Table 1:

[0063] Table 1 Comparison of enzyme parameters between wild-type and mutant M5

[0064] kcat / Km (mM⁻¹s⁻¹) 12.5 122.5 9.8× Tm (°C) 52.3 56.6 +4.3℃ <![CDATA[t1 / 2 at 50℃ (h)]]> 2.5 15.0 6× Conversion rate (%) ~8 81 10× ee (%) >99 >99 - Spacetime productivity (g·L⁻¹·d⁻¹) 8.5 92 10.8×

[0065] The results showed that the mutant M5 had a 9.8-fold increase in kcat / Km, a 4.3-fold increase in Tm, a 6-fold increase in half-life at 50℃, an increase in conversion efficiency from about 8% to 81%, and a space-time yield of 92 g·L⁻¹·d⁻¹.

[0066] Example 5: Hundred-gram scale preparation and batch stability

[0067] Based on the excellent enzymatic properties of the mutant CA113-M5, a scale-up preparation experiment of the key chiral amine intermediate of 100 g grade remedypa was conducted. Simultaneously, standardized methods for detecting related substances and chirality were established. The specific procedures are as follows:

[0068] I. Preparation of Tris-isopropylamine buffer (0.1M, initial pH 9.5)

[0069] Prepare Tris-isopropylamine buffer solution at a substrate concentration of 20 equivalents: Take an appropriate volume beaker, add an appropriate amount of ice water, add Tris reagent to prepare a 0.1M concentration, then add 354 g of isopropylamine, and slowly add about 400 mL of concentrated hydrochloric acid dropwise while stirring. Maintain the system at a low temperature (0-4℃) throughout the process, and adjust the initial pH of the system to about 9.5.

[0070] II. Construction and reaction process of 100 g-scale substrate reaction system

[0071] Catalytic synthesis was carried out in a 10 L reaction flask with a total volume controlled at 4 L. 0.1 M Tris-isopropylamine buffer (initial pH 9.5) was added to the flask, followed by 2.4 L of the crude enzyme solution of the recombinant transaminase mutant M5 prepared in Example 1, and 1 g of coenzyme PLP. The mixture was stirred to dissolve. Stirring was started (200 rpm), and the pH of the system was finely adjusted to 9.0 ± 0.2 with isopropylamine, maintaining the reaction temperature at 45 °C. 100 g of substrate was weighed, dissolved in 200 mL of DMSO, transferred to a dropping funnel, and slowly added to the reaction system. Samples were taken periodically during the reaction to monitor the progress. After the reaction, separation and purification were performed.

[0072] III. Detection Methods for Related Substances and Chirality in Reaction Systems

[0073] (a) Detection of related substances

[0074] Chromatographic conditions: XBridge®Shield RP18 column (4.6*250 mm, 5μm); mobile phase A was 0.1% phosphoric acid aqueous solution, and mobile phase B was methanol; flow rate was 1 mL / min; detection time was 16 min; column temperature was 40℃; detection wavelength was 254 nm; injection volume was 5 μL.

[0075] Gradient elution procedure:

[0076] A% 95.0 95.0 0.0 0.0 95.0 95.0 B% 5.0 5.0 100.0 100.0 5.0 5.0

[0077] Detection method: Take an appropriate amount of reaction solution, dilute it with methanol, filter it through a 0.22μm filter membrane, and inject it for detection under the above chromatographic conditions. Record the chromatogram and calculate the content of related substances.

[0078] (ii) Product chirality detection

[0079] Chromatographic conditions: Daicel IC-3 column (250×4.6 mm, 3 μm); detection wavelength 260 nm; flow rate 1.0 mL / min; column temperature 20 ℃; injection volume 10 μL; mobile phase A was hexane-isopropanol (100:10), and mobile phase B was isopropanol.

[0080] Gradient elution procedure:

[0081] A% 80.0 70.0 70.0 80.0 80.0 B% 20.0 30.0 30.0 20.0 20.0

[0082] Detection method: Take an appropriate amount of the purified product, dissolve it in a suitable solvent, filter it through a 0.22μm filter membrane, and inject it for detection under the above chromatographic conditions. Record the chromatogram and calculate the ee value of the product.

[0083] IV. Scale-up Preparation Results and Batch Stability

[0084] After the reaction was completed at a scale-up of 100 g of substrate, the product was separated and purified, with a separation yield of 81% and ee>99%, which met the requirements of the API impurity profile.

[0085] To verify the process stability, three consecutive batches of 100 g grade Remegapan intermediates were synthesized using enzyme catalysis. The results are shown in Table 2.

[0086] Table 2 Batch stability of 100g-level preparation

[0087] Batch 1 100 81.2 >99 Batch 2 100 80.5 >99 Batch 3 100 81.8 >99 average value - 81.2 >99 RSD (%) - 2.1 -

[0088] The yields of the three batches were RSD < 3% and ee > 99%, meeting the API impurity profile requirements, demonstrating that the process has good reproducibility and industrial feasibility.

[0089] Comparative Example 1: Traditional Directed Evolution Method

[0090] Constructing mutant libraries using traditional random mutagenesis or error-prone PCR methods generally yields mutants with relatively low viability; most mutants exhibit only 0.05-0.15 times the activity of the wild type, and the positive mutation rate is extremely low. The entire process is time-consuming, labor-intensive, and has a low success rate. Furthermore, in traditional directed evolution, most negative mutants are not adequately analyzed, hindering the accumulation of knowledge and the generation of relevant models, thus failing to effectively guide subsequent modification work.

[0091] In contrast, the machine learning-guided directed evolution method of this invention obtains the optimal mutant with an activity increase of about 10 times through only 96 mutants and two rounds of experiments, which significantly shortens the research and development cycle and reduces screening costs.

Claims

1. An ω-transaminase mutant, characterized in that, The mutant, based on the amino acid sequence of the wild-type ω-transaminase CA113 shown in SEQ ID NO: 1, contains at least one of the following mutation sites: R68F, D125V, I224V, I262A, H318A, M56F, H191F, V435I, T322C, T322N, T322A, D315I, D416R, E168I, E168L, wherein the mutant exhibits enhanced activity in catalyzing the transamination of the key chiral amine intermediate (R)-3-amino-2-oxopiperidine derivative of retinoic acid compared to the wild type.

2. The ω-transaminase mutant according to claim 1, characterized in that, The mutant is a double mutant M5, whose amino acid sequence is shown in SEQ ID NO: 3, containing two mutation sites, R68F and D125V.

3. A nucleic acid molecule encoding the ω-transaminase mutant of claim 1 or 2, characterized in that, The nucleotide sequence corresponding to amino acid SEQ ID NO: 1 is SEQ ID NO: 2, and the nucleotide sequence corresponding to amino acid SEQ ID NO: 3 is SEQ ID NO:

4.

4. A recombinant expression vector, characterized in that, It includes the nucleic acid molecule as described in claim 3.

5. A recombinant host cell, characterized in that, It includes the recombinant expression vector as described in claim 4.

6. A machine learning-guided directed evolution method for ω-transaminase, characterized in that, Includes the following steps: (a) Using wild-type ω-transaminase as the starting enzyme, consensus sequences were generated through multiple sequence alignment, and potential hotspot sites were screened using molecular dynamics simulations and channel analysis; (b) A mini-mutant library containing the mutation sites screened in step (a) was constructed; (c) Enzymatic parameters of each mutant in the mutant library were measured, and kcat / Km was used as the fitness index. The results were then uploaded to a small sample learning module based on a pre-trained model for training; (d) Virtual screening was performed based on the trained model to predict mutants with an activity increase of ≥30%. (e) Verify the activity of the predicted mutant using wet experiments, and add the verification data to the training set for iterative optimization; (f) Repeat steps (d)-(e) until convergence is achieved to obtain the optimal mutant.

7. The method according to claim 6, characterized in that, The molecular dynamics simulation described in step (a) is a 5-ns alanine scan molecular dynamics simulation to evaluate the effect of each mutation on the substrate Cα-N distance and the Asp-Arg catalytic dimmer angle.

8. The method according to claim 6, characterized in that, The channel analysis described in step (a) uses CAVER software to identify bottleneck residues with a radius of <1.2 Å at the narrowest point of the substrate channel.

9. The method according to claim 6, characterized in that, The library size of the micromutation library described in step (b) is 40-60.

10. The method according to claim 6, characterized in that, The training described in step (c) uses ESM-1v or a similar ensemble learning algorithm for parameter fine-tuning and employs 5-fold cross-validation. When q²≥0.70, the virtual screening stage is entered.

11. The use of the ω-transaminase mutant according to claim 1 or 2 in the preparation of the key chiral amine intermediate (R)-3-amino-2-oxopiperidine derivative of Remepiquat chloride.

12. The application according to claim 12, characterized in that, The transamination reaction was carried out in a 1 L reactor with a final substrate concentration of 25 mg·mL⁻¹, an enzyme loading of 1 g·L⁻¹, a reaction temperature of 45 °C, and a pH of 9.0.

Images