Enhanced stereodivergent evolution of carboxylesterase for efficient kinetic resolution of near-symmetric esters through machine learning

doi:10.21203/rs.3.rs-3897762/v1

Download PDF

Article

Enhanced stereodivergent evolution of carboxylesterase for efficient kinetic resolution of near-symmetric esters through machine learning

https://doi.org/10.21203/rs.3.rs-3897762/v1

This work is licensed under a CC BY 4.0 License

Journal Publication

published 20 Oct, 2024

Read the published version in Nature Communications →

Version 1

posted

You are reading this latest preprint version

Carboxylesterases serve as potent biocatalysts in the enantioselective synthesis of chiral carboxylic acids and esters. However, naturally occurring carboxylesterases exhibit limited enantioselectivity, particularly towards ethyl 3-cyclohexene-1-carboxylate (CHCE), due to its nearly symmetric structure. While machine learning has proven effective in expediting directed evolution, the lack of models for prediction of enantioselectivity for carboxylesterases has hindered progress, primarily due to challenges obtaining high-quality training datasets. In this study, we devised a high-throughput method by coupling alcohol dehydrogenase to determine the apparent enantioselectivity of the carboxylesterase AcEst1 from Acinetobacter sp. JNU9335, thereby generating a high-quality dataset. Leveraging seven features derived from biochemical considerations, we quantitively described the steric, hydrophobic, hydrophilic, electrostatic, hydrogen bonding, and π-π interaction effects of residues within AcEst1. A robust gradient boosting regression tree model was trained to facilitate stereodivergent evolution, resulting in the enhanced enantioselectivity of AcEst1 towards CHCE. Through this approach, we successfully obtained two stereocomplementary variants, DR3 and DS6, demonstrating significantly increased and reversed enantioselectivity. Notably, DR3 and DS6 exhibited utility in the enantioselective hydrolysis of various symmetric esters. Comprehensive kinetic parameter analysis, molecular dynamics simulations, and QM/MM calculations provided insights into the kinetic and thermodynamic aspects underlying the manipulated enantioselectivity of DR3 and DS6.

Biological sciences/Biochemistry/Biocatalysis

Biological sciences/Computational biology and bioinformatics/Machine learning

Carboxylesterase

Machine learning

Nearly symmetric substrate

3-Cyclohexene-1-carboxylic acid

Stereodivergent evolution

Enzymes have garnered considerable attention in the realms of synthetic biology and biocatalysis, being widely hailed as the preferred choice for the biosynthesis of optically active chemicals^1–4. The intricate stereochemical structures of enzymes often manifest in distinctive spatial, hydrophobic, and electrostatic properties, particularly within the active center, forming the basis for high stereoselectivity^5–7. However, enzymes face challenges maintaining high stereoselectivity, especially when confronted with substrates boasting nearly symmetric structures^8–10. These substrates are commonly deemed “hard-to-be-discriminated,” not only by chemical catalysts but also by biocatalysts^11,12.

Chiral cyclohex-3-ene-1-carboxylic acid (CHCA) features a nearly symmetric hexatomic ring, serving as a crucial building block for synthesizing a diverse array of pharmaceuticals, agrochemicals, and natural products (Fig. 1A). Compared to cyclohexane, cyclohexadiene, and phenyl groups, CHCAs often exhibit distinctive biological activities due to their unique cyclohexene structure¹³. Enantiomeric pairs of CHCAs showcase applicability, with both enantiomers serving as versatile building blocks. For example, (S)-CHCA is utilized in the synthesis of Edoxaban, an effective oral medication for treating venous thrombosis and surgical bleeding through the inhibition of coagulation factor Xa^14,15. Other significant applications include the synthesis of immunosuppressant FK-506¹⁶, aglycone of antitumor drug (+)-phyllanthocin¹⁷, toxin pumiliotoxin C¹⁸, and repellent SS220¹⁹, etc. Conversely, (R)-CHCA forms the pivotal building block of Oseltamivir, a widely-used oral antiviral drug for treating influenza A by inhibiting neurosine glycosidase²⁰. It has also been applied to the natural products leustroducsin B²¹, phyllanthocin²², potential antibacterilas²³, and E-selectin antagonists²⁴. Because of their structural similarities with only minor differences in the position of C = C that is distant from the chiral center, distinguishing between (S)- and (R)-CHCA proves challenging for both chemical and biological catalysts.

Current chemical approaches for synthesizing chiral CHCA involve cumbersome resolution steps and necessitate large amounts of chiral reagents and hazardous organic solvents^25,26, resulting in low atomic economy and environmental concerns (Fig. 1B). Biocatalytic synthesis of chiral CHCA presents its own set of challenges. Non-symmetric esters with varying substituent sizes are relatively easily discriminated by the substrate binding pocket, which excludes the incompatible configuration due to steric hindrance or unfavorable repulsive force (Fig. 1C)²⁷. However, discriminating nearly symmetric CHCA esters proves challenging for the substrate binding pocket, as the differences between (R)- and (S)-esters are minor (Fig. 1D)²⁸. Commercial enzymes such as Novozym 435 and pig liver esterase (PLE) exhibit no selectivity in discriminating (S)- and (R)-CHCA esters²⁸. While carboxylesterase BioH is reported to possess activity toward CHCA esters, its low E value of merely 2.1 necessitated the iterative construction of a triple mutant, resulting in an increased E value of 7.1 at a 40 mM substrate concentration²⁹. However, further concentration increases led to a significant decrease in stereoselectivity. Consequently, the enantioselective synthesis of chiral CHCA and esters remains a formidable challenge, posing difficulties for both chemical and biocatalysts alike.

Directed evolution plays a pivotal role in expediting the development of stereoselective enzymes, employing diverse strategies categorized as quantity-intensive and quality-intensive approaches^4,30. Quantity-intensive strategies involve random mutagenesis across the entire sequence space³¹. This approach necessitates a reliable high throughput screening (HTS) method with an ideal throughput exceeding 1×10⁷ mutants/day³². However, the use of chromogenic or electronic substrate analogs in HTS may deviate from research objectives, especially in evolving catalytic activity, enantioselectivity, and substrate specificity. Quality-intensive approaches focus on rational or semi-rational mutagenesis using restricted genetic codons to construct a “smart library”^33,34. Identifying potential hotspots is crucial for these approaches and is usually determined through empirical, experimental, or computational analysis. While quality-intensive approaches are effective, they may face challenges when dealing with “hard-to-be-discriminated” substrates.

Machine learning (ML) emerges as an alternative, offering a more direct shortcut for boosting directed evolution based on extensive high-quality data and statistical models^35,36. ML, being data-driven, can identify catalytic patterns, predict promising mutants, and excel in predicting new substitution combinations for directed evolution³⁷. ML has found application in manipulating the activity, thermostability, and stereoselectivity of various enzymes, including epoxide hydrolases³⁸, imine reductase³⁹, P450 monooxygenase⁴⁰, and transaminases⁴¹. The success of ML predictors depends on the quality of the training dataset⁴¹, with recent efforts by Ran et al. showing promise in predicting enantiomeric excess ratios of hydrolases toward non-symmetric substrates, with positive prediction of 18 out of 28 reactions.⁴² The positive prediction ratio was 64%, might be attributed to the low-quality dataset of hydrolase-catalyzed kinetic resolution reactions, lower in the case of nearly symmetric substrates. Hence, generation of high-quality dataset and selection of suitable descriptors are the key to the success of ML predictors.

In this study, we present the establishment of a high-throughput method to generate a high-quality dataset for carboxylesterase AcEst1⁴³ from Acinetobacter sp. JNU9335 using the “real substrate” ethyl cyclohex-3-ene-1-carboxylate (CHCE), characterized by its nearly symmetric structure. Subsequently, we constructed an ML predictor utilizing features extracted from biochemical considerations of AcEst1. The trained ML predictor was then applied to design combinatorial mutants. Finally, we successfully achieved stereodivergent evolution of AcEst1, resulting in the generation of two stereocomplementary mutants. These mutants were further employed in the synthesis of both enantiomers of chiral CHCE.

Development of a high throughput method to obtain a high-quality enantioselectivity dataset

Enantiomeric excess (e.e.) is a commonly used parameter to gauge the enantioselectivity of given enzymes. However, when dealing with hydrolases that exhibit low enantioselectivity, particularly towards nearly symmetric substrates, e.e. proves to be an ineffective parameter as its values fluctuate during reaction⁴⁴. To address this limitation and devise a high-throughput method for obtaining a high-quality dataset of enantioselectivity, we adopted the ratio of initial reaction rates between (R)- and (S)-CHCE, referred to as the apparent enantioselectivity (E_app). This approach relies on the “real substrate” and accurately mirrors the actual reaction dynamics (Figure 2A). The initial reaction rates towards (R)- and (S)-CHCE were determined by coupling the hydrolytic reaction with a chromogenic reaction catalyzed by alcohol dehydrogenase (ADH). An ideal ADH for this method should possess the following characteristics: 1) NADP⁺-dependent instead of NAD⁺-dependent to eliminate background interference; 2) high catalytic efficiency (k_cat/K_M), even at lower ethanol concentrations. Consequently, genome mining and rational mutagenesis were employed to identify an NADP⁺-dependent ADH with high k_cat/K_M.

At high ethanol concentration (800 mM), NADP⁺-dependent ADH6 proved the most efficient, displaying a specific activity as high as 421 U·g^–1(Figure 2B), while at lower ethanol concentration (2 mM), NADP⁺-dependent ADH10 from Kluveromyces polysporus exhibited the highest specific activity of 9.34 U·g^–¹. However, the activity of ADH10 was low and required high loading, which was deemed unsatisfactory for developing a high-throughput method to determine E_app. To address this limitation, ethanol was docked into the active center of ADH10 (PDB: 5Z2X), and residues surrounding ethanol were identified (Figure 2C). In response to the insufficient activity, rational mutagenesis was undertaken to manipulate the hydrophobic interaction and steric hindrance between the methyl group of ethanol and ADH10. Mutations were introduced, turning them into Cys, Phe, Leu, Ile, and Val. Five single mutants, including V84L, V84I, F197L, F197V, and F197I, were obtained, displaying higher activity towards ethanol (Figure 2D). Among these, F197V proved to be the most efficient, with a specific activity of 17.1 U·g^–¹. Consequently, double mutants ADH10_V84L/F197V and ADH10_V84I/F197V were constructed, with ADH10_V84L/F197V displaying a synergistic effect, exhibiting a specific activity of 70.4 U·g^–¹, approximately 7.6-fold higher than ADH10. Kinetic parameters analysis revealed that ADH10_V84L/F197V not only demonstrated increased ethanol binding affinity, with a K_M value of 2.34 mM, much lower than 55.9 mM of WT, but also enhanced catalytic activity, with a k_cat of 5.68 min^–¹. The k_cat/K_M of ADH10_V84L/F197V was significantly increased to 2.43 min^–¹·mM^–¹ from 0.07 min^–¹·mM^–¹ of WT. The impressive performance of ADH10_V84L/F197V encouraged further optimization of conditions for high-throughput determination of E_app of AcEst1.

Various factors including ADH dosage, NADP⁺ concentration, pH, and CHCE concentrations, were systematically investigated. In Figure 2E, the changes in A₃₄₀ were similar at ADH dosages higher than 7.5 U·mL^–¹. With an increase in NADP⁺ concentrations from 2 to 5 mM, the slopes of A₃₄₀ increased accordingly until exceeding 4.0 mM. The effect of CHCE concentrations was also explored, and the slopes of A₃₄₀ increased steadily until reaching 5.0 mM CHCE. Although a further increase in CHCE concentration would result in a higher A₃₄₀ slope, the low solubility of CHCE might interfere with the coupled reaction. Consequently, an ADH dosage of 7.5 U·mL^–¹, NADP⁺ concentration of 4 mM, and CHCE concentration of 5.0 mM were adopted as the optimum conditions for the high-throughput method. Subsequently, the feasibility of this method was evaluated using the entire plate of AcEst1. The average E_app of WT was 4.1 ± 0.3, with a CV of less than 8%. This high-throughput method, based on a “real substrate,” was developed to obtain a high-quality dataset on the enantioselectivity of AcEst1. Importantly, this method can be extended to characterize other hydrolases with ethanol as the product.

Development of ML predictors for enantioselectivity of AcEst1 toward CHCE

To gather a diverse and high-quality dataset on the enantioselectivity of AcEst1 mutants for constructing an ML predictor, 20 unconservative residues located in the first and second layers surrounding the catalytic S201 were identified for saturation mutagenesis (Figure 3A). Degenerative codons of 22c-trick⁴⁵ including NDT (A/T/C/G, A/T/G, T), VHG (A/C/G, A/T/C, G), and TGG were utilized at molar ratio of 12:9:1 to develop the unbiased library, and a 3-fold excess of mutants was evaluated at each position to ensure the quality. A total of 1920 mutants were subjected to determining E_app values using the high-throughput method. Mutants from F66, F78, I82, V133, Y228, V230, D253, V254, V257, and L297 exhibited significantly enhanced E_app values, while mutants from T248, L249, V254, T258, L297, and N329 displayed decreased E_app values (Figure 3B). Among all the single mutants, V257M and L297F had E_app values of 9.84 and 0.77, respectively, ranking as the highest and lowest records. After removing redundancy, a high-quality dataset containing 760 out of 1920 mutants was obtained for training ML models.

While ML is recognized for its potential in advancing directed evolution, its effectiveness is often hindered by incompatible descriptors or features⁴⁶. In this study, biochemical features, including volume (feature #1), hydrophobicity (feature #2), hydrophilicity (feature #3), electrostatic (feature #4), hydrogen bond (feature #5), π-π interaction (feature #6), and distance to catalytic residue (feature #7), based on 1-D sequence from biochemical considerations, were introduced in this study (Figure 3C). The use of atom numbers or SMILES format of amino acids as one feature is common. However, this feature is inaccurate for describing the properties of residues, as there is no linear relationship between atom number and spatial volume. For instance, Phe and Lys have the same spatial volume but different molecular weights. Therefore, spatial volume was selected as a feature. The hydrophobicity of the active center is crucial for enzyme performance. Both the hydrophobicity index⁴⁷ denoting hydrophobicity (feature #2) and the Kyte-Doolittle hydropathy scale⁴⁸ denoting hydrophilicity (feature #3) were introduced as features to describe the hydrophobic interactions. The isoelectric point of an amino acid (feature #4) was selected to describe electrostatic properties. Hydrogen bonding was described by the number of electronegative atoms, such as O and N (feature #5). Stacking interactions, including π-π or alkyl-π interactions, were represented by the number of double bonds (feature #6). Moreover, the distance to nucleophilic S201 was also introduced as feature #7. These features from experimental and biochemical considerations were first introduced in the ML model, providing a robust foundation for training the ML predictor.

Various ML algorithms have been employed to address biocatalytic-related challenges, including Kernel Ridge Regression (KRR), Gaussian Process Regression (GPR), Gradient Boosting Regression Tree (GBRT), Random Forest Regression (RFR), Support Vector Regression (SVR), and Bayesian Ridge Regression (BRR)^35,49. However, no single algorithm is universally optimal for all tasks. Consequently, we assessed the performance of six regression models in correlating the enantioselectivity of AcEst1 with seven features, namely GBRT, GPR, KRR, RFR, SVR, and BRR. The dataset was randomly divided into a training set (90%) and a testing set (10%), and hyperparameters were tuned for each model. Subsequently, the learning process was executed using the high-quality dataset and the aforementioned models. Based on the regression results, GBRT outperformed GPR, KRR, RFR, SVR, and BRR. The Pearson correlation (R²) between predicted and experimental E_app values of the GBRT model reached a high value of 0.93, with a Root Mean Square Error (RMSE) of 0.119 (Figure 3D). Landscape analysis revealed a smooth distribution of data, indicating the excellent performance of GBRT. KRR, SVR, and BRR were ineffective in predicting the E_app of AcEst1, exhibiting lower R² values (below 0.55), and almost all the E_app values were underestimated. RFR showed better performance than GPR, with higher R²and lower RMSE. However, RFR’s ability to predict mutants with elevated E_app was less robust than GBRT and GPR. The performance of the GBRT model aligned with data-driven protein engineering for the activity and stereoselectivity of transaminase, achieving an R² of 0.803⁴¹. The higher R² in our model can be attributed to the high-quality dataset and incorporation of biochemical features. Attempts to reduce the number of features and retrain the GBRT model resulted in decreased correlation.

To enhance the accuracy of the trained GBRT predictor, we systematically combined advantageous single mutants to generate double mutants. All single mutants with increased or decreased E_app values were paired, resulting in various double mutants. Notably, V257M/Y228M stood out with the highest E_app value of 13.8, while L297F/L249A displayed the lowest E_app value of 0.36. The incorporation of these double mutants further enriched the dataset used for retraining the GBRT model. The GBRT predictor was then retrained using both the single and double mutants (Figure 4A). The retrained GBRT predictor demonstrated exceptional performance, achieving a Pearson correlation of 0.97 with an RMSE of 0.105, indicative of a robust regression. The remarkable performance of the trained GBRT predictor instilled confidence in its application to guide combinatorial mutagenesis for the stereodivergent evolution of AcEst1.

ML guided stereodivergent evolution of AcEst1

Considering that V257M and L297F were the single mutants with the highest and lowest E_app values, they were selected as starting points for the stereodivergent evolution of (R)-selective and (S)-selective AcEst1 mutants. Consequently, V257M and L297F were designated as DR1 and DS1, respectively. For the (R)-selective evolution, we predicted the E_app values of all double mutants starting with DR1 and incorporating other mutations (I82M, V133S, Y228M, V230W, D253H, V254Q, V257M, and L297I) with increased E_app (Figure 4B). DR1/Y228M (DR2) emerged as the most (R)-selective, consistent with experimental results. Subsequently, we predicted the E_app values of triple mutants starting from DR2, leading to DR2/I82M (DR3) with an impressive E_app of 18. We further forecasted the E_app values of quadruple mutants based on DR3, including DR3/V133S, DR3/D253H, DR3/V254Q, and DR3/L297I. V230W was excluded due to the lack of a synergistic effect. As depicted in Figure 4B, the predicted E_app values of quadruple mutants were all higher than the corresponding triple, double, and single mutants, suggesting a synergistic effect among them. However, none of the quadruple mutants exhibited predicted E_app values surpassing DR3, which can be attributed to the limited contributions of V133S, D253H, V254Q, and L297I. To validate the accuracy of the GBRT-predicted results, we experimentally constructed DR1, DR2, and DR3, determining their enantioselectivity values (E value) through resolution reactions. The E values progressively increased from 7.3 for WT to 40.1 for DR1, 59.1 for DR2, and 103 for DR3 (Figure 4D). Quadruple mutants, such as DR3/V254Q and DR3/L297I, were also experimentally constructed; however, their E values were lower than DR3, aligning with the prediction result. The newly developed DR3 thus represents a promising biocatalyst with excellent enantioselectivity in the resolution of near-symmetric CHCE.

The successful guidance of (R)-selective evolution further motivated us to assess the applicability of the trained GBRT predictor in (S)-selective evolution. To depict the evolution pathway more clearly, the reciprocal of the predicted E_app ((S)-CHCE/(R)-CHCE) was adopted for (S)-selective evolution (Figure 4C). Among all the single mutants, only L297F exhibited an E_app value lower than 1.0, indicating reversed (S)-selectivity. Therefore, L297 was designated as DS1 and served as the starting point for (S)-selective evolution. All the single mutants with E_app values lower than WT, including L247F, T248W, L249A/I, V254K, T258A, and N329F, were added to DS1 to form double mutants. Among all the double mutants, the reciprocal of the predicted E_app value for DS1/L249A was 2.8, ranking the highest. Consequently, DS1/L249A was designated as DS2a and subjected to triple mutation. Other mutations were iteratively added to DS2a to predict their E_app values. DS2a/T248W displayed a reciprocal predicted E_app of 3.0, slightly higher than DS2a, and was regarded as DS3 for further combination. The quadruple mutant DS3/T258A (DS4) exhibited increased enantioselectivity and was combined with other mutations to form quintuple mutants. DS4/V254K showed the highest (S)-selectivity among all the quintuable mutants and was regarded as DS5. Furthermore, L247F and N329F were introduced in DS5, resulting in DS5/N329F (DS6) with a reciprocal of predicted E_app value as high as 5.9 toward (S)-CHCE, even higher than the 4.1 of WT towards (R)-CHCE, implying the concrete reversal of the enantioselectivity of AcEst1 from (R)-selective to (S)-selective.

Although the single mutant L247F exhibited decreased enantioselectivity compared to WT, its contribution to (S)-selectivity was limited. The reciprocal predicted E_app values of different combinatorial mutants containing L247F were overall increased while locally decreased (Figure 4C), suggesting the existence of an antagonistic effect. The introduction of L247F in DS6 resulted in decreased enantioselectivity. DS1-DS6 predicted by the GBRT model were experimentally constructed and evaluated in the resolution of rac-CHCE. The E value of DS1 was –1.6, reversed from the 7.3 of WT. Since DS2a (DS1/L249A) and DS2b (DS1/L249I) had similar predicted E_app values, both were investigated, and DS2a displayed a lower E-value of –4.2. From DS3 to DS6, the E value further decreased from –5.7 to –11, significantly lower than WT AcEst1. Stereodivergent evolution of AcEst1 toward nearly symmetric esters has been achieved using this newly trained GBRT predictor.

Characterization of stereocomplementary AcEst1 mutants

To gain a deeper understanding of enantioselectivity manipulation, kinetic parameters of AcEst1 mutants toward (R)- and (S)-CHCE were meticulously characterized (Table 1). The WT AcEst1 exhibited k_cat/K_M values toward (R)- and (S)-CHCE of 156 and 13 s^–1∙mM^–1, respectively, with a resulting calculated ratio of 12.0. Notably, the (R)-selective DR3 mutant displayed k_cat/K_M values of 9.82 and 0.08 s^–1∙mM^–1toward (R)- and (S)-CHCE, respectively, leading to a drastically increased ratio of 123, approximately 10.3-fold higher than WT, aligning with the observed E value. DR3 exhibited k_cat values toward (R)- and (S)-CHCE of 60.6 and 0.66 s^–1, respectively, with a ratio of 92, significantly surpassing the 6.2 of WT. In terms of kinetic dynamics, the substantial decrease in the k_cat value of DR3 toward (S)-CHCE is pivotal for its heightened enantioselectivity. Deconvolution analysis indicated elevated k_cat/K_M ratios for Y257M and Y228M (59.0 and 32.9, respectively) compared to WT, suggesting their significant contributions to enhanced enantioselectivity. I82M exhibited a modest increase in the k_cat/K_M ratio (13.0).

Kinetic parameter and deconvolution analysis of the (S)-selective DS6 were also conducted. The k_cat/K_M value of DS6 toward (R)-CHCE significantly decreased to 0.74 s^–1∙mM^–1 from the 156 s^–1∙mM^–1 of WT, indicating a substantial reduction in catalytic efficiency toward (R)-CHCE. Meanwhile, the k_cat/K_M value of DS6 toward (S)-CHCE was 4.00 s^–1∙mM^–1, with a k_cat/K_M ratio of (S)-CHBE to (R)-CHCE of 5.41, around 67.6-fold higher than that of WT, in agreement with the reversed E value of DS6. Deconvolution analysis revealed that L297F exhibited the highest k_cat/K_M ratio of 1.74. While L297F maintained the same k_cat/K_M value of 13.3 s^–1∙mM^–1 toward (S)-CHCE as WT, it drastically decreased the k_cat/K_M value to 7.65 s^–1∙mM^–1 toward (R)-CHCE. Additionally, L249A also displayed a higher k_cat/K_M ratio of 1.08, about 13.5-fold higher than WT. In contrast to L297F and L249A, single mutants T248W, T258A, and V254K exhibited increased catalytic efficiency toward (S)-CHCE, surpassing the increment toward (R)-CHCE. T248W, T258A, and V254K contributed not only to the increased (S)-selectivity but also to the catalytic efficiency of DS6. T248W and N329F were beneficial for the enhanced binding affinity of DS6 toward (S)-CHCE. From the perspective of kinetic dynamics, the significantly decreased catalytic efficiency toward (R)-CHCE and enhanced binding affinity toward (S)-CHCE are both crucial for the (S)-selectivity of DS6.

Application potential of stereocomplementary DR3 and DS6 in resolving near-symmetric esters

The potential application of stereocomplementary DR3 and DS6 in the synthesis of chiral CHCA was thoroughly investigated. To qualify as promising biocatalysts with industrial relevance, substrate loading should exceed 100 g·L^–¹, and the e.e. value should surpass 99% within 24 hours. After optimization, a substantial 1.0 M rac-CHCE (154 g·L^–¹) could be enantioselectively hydrolyzed into (S)-CHCE by DR3 (Figure 5A). By the 8.0-hour mark, the e.e._s reached >99% (S) at a conversion ratio of 53% and e.e._p of 88%. (S)-CHCE was isolated from the reaction system with a yield of 45.7%. Similarly, DS6 was applied in the resolution of 1.0 M rac-CHCE (154 g·L^–¹) for the synthesis of (R)-CHCE (Figure 5B). At 18 hours, e.e._s value reached >–99% (R), with a conversion ratio of 74% and e.e._p of –41%. (R)-CHCE was isolated with a yield of 23.3%. Notably, the enantioselective resolution of nearly symmetric CHCE for the synthesis of both enantiomers of chiral CHCE was achieved at 1.0 M for the first time, employing stereocomplementary DR3 and DS6.

The application potential of stereocomplementary DR3 and DS6 was further examined in the enantioselective resolution of esters with a nearly symmetric structure (Table 2). Initially, methyl and isopropyl cyclohex-3-ene-1-carboxylate (S2 and S3) could be enantioselectively hydrolyzed by DR3 and DS6. The E values of DR3 towards S2 and S3 were 92 and 110, respectively, dramatically higher than 5.6 and 8.3, consistent with CHCE. Furthermore, with the increase of alcohol groups from methyl to ethyl and isopropyl, the E values of DR3 increased accordingly, implying that a larger volume of alcohol groups is favorable for high enantioselectivity. DS6 exhibited reversed (S)-selectivity towards S2 and S3 with E values of –7.0 and 12, respectively. DR3 and DS6 demonstrated opposite priorities in the enantioselective accommodation of CHCA esters.

Moreover, other esters with a nearly symmetric structure, specifically oxyheterocyclic esters, were also evaluated by DR3 and DS6 (Table 2). WT AcEst1 displayed a relatively higher activity towards S4 and S6 with odd-member rings than S5 and S7 with even-member rings. The highest E value was 12 towards S7. DR3 could enantioselectively hydrolyze (R)-esters with higher E values than WT. The highest E value of DR3 was also observed with S7. Conversely, DS6 could enantioselective hydrolyze (S)-esters with reversed E values. The lowest E value was observed with S5 by DS6. All of the above results demonstrate that stereocomplementary DR3 and DS6 can be applied in the enantioselective hydrolysis of other esters containing a nearly symmetric structure.

Molecular mechanism of stereoselectivity control according to MD and QM/MM calculation

In-depth insights into the molecular mechanism of carboxylesterase in the enantioselectivity resolution of nearly symmetric esters were sought by attempting to resolve the crystal structure of sterecomplementary DR3 and DS6. Despite optimization of crystallization conditions and different truncation attempts, crystals with the desired quality for X-ray crystallography were not obtained. Consequently, structural models of DR3 and DS6 were constructed using AlphaFold2, a widely accepted software for building reliable protein structures based on artificial intelligence. The structural models of WT, DR3, and DS6 were minimized and subjected to multiple 100-ns MD simulations. The RMSD achieved stability at about 5 ns, ranging from 1.5 to 2.5 Å.

Near-attack conformation (NAC) analyses were conducted to gain insights into the transition state of nucleophilic attack⁵⁰. According to the transition state, the angle among O of S201, carbonyl C, and carbonyl O of the substrate (θ1: ∠OG-C7-O2) should be within 90° ± 15°, and the nucleophilic attack distance between O of S201 and carbonyl C of the substrate (d1: dist_OG-C7) should be less than 3.4 Å. As illustrated in Figure 6, the percentages of conformations satisfying NAC parameters of WT toward (R)- and (S)-CHCE were 38.7% and 14.9%, respectively (Figure 6A&6B), consistent with the (R)-preference of WT AcEst1. However, the NAC percentage of 14.9% in WT and (S)-CHCE hinted that (S)-CHCE could also be accommodated by WT.

For DR3, the NAC percentage of (R)-CHCE was 39.9%, significantly higher than the 5.6% of (S)-CHCE (Figure 6C&6D), indicating that hydrolysis of (R)-CHCE is more favorable than (S)-CHCE in DR3. Regarding DS6, the NAC percentage of (S)-CHCE was 29.4%, much higher than the 12.9% of (R)-CHCE, proving that (S)-CHCE is more preferable than (R)-CHCE in DS6 (Figure 6E&6F). Binding free energy was also analyzed employing the MMPB/GBSA method⁵¹. The free energy difference between (R)- and (S)-CHCE of DR3 was –3.7 kcal·mol^–¹, higher than the –3.0 kcal·mol^–¹ of WT. For DS6, the free energy difference between (R)- and (S)-CHCE was 3.3 kcal·mol^–¹, suggesting that (S)-CHCE was preferable.

Representative conformations with the highest distribution ratio were extracted from MD simulations and are presented in Figure 7. In the case of WT AcEst1, the distance between the O atom of catalytic S201 and the carbonyl C (d₁) of (R)-CHCE measured 3.1 Å, while d₁ of (S)-CHCE was 3.9 Å (Figure 7A&7B). Turning to DR3, the d₁ of (R)-CHCE decreased to 2.5 Å, and the d₁ of (S)-CHCE increased to 4.0 Å (Figure 7C&7D). The mutation of V257 into M257 introduced an alkyl-π interaction with the double bond of (R)-CHCE. For DS6, the d₁ of (S)-CHCE decreased to 2.6 Å, while the d₁ of (R)-CHCE remained the same as in WT (Figure 7E&7F). A favorable π-π interaction between F297 and (S)-CHCE was also observed, providing evidence for the increased catalytic efficiency of DS6 toward (S)-CHCE.

The free energy barriers (DG^‡) for (R)- and (S)-CHCE of WT, DR3, and DS6 were calculated using QM/MM⁵². The free energy difference between (R)- and (S)-CHCE (DDG^‡) of WT was calculated to be –1.1 kcal·mol^–¹, favoring the hydrolysis of (R)-CHCE. In contrast, the DDG^‡ values of DR3 and DS6 were –2.7 and 0.9 kcal·mol^–¹, respectively. The significant decrease in DDG^‡ of DR3 and the increase for DS6 were consistent with their performance in the enantioselectivity resolution of CHCE. These findings provide molecular insights into the manipulation of enantioselectivity of stereocomplementary DR3 and DS6 in the accommodation and resolution of CHCE with a nearly symmetric structure.

Carboxylesterases represent highly promising biocatalysts for synthesizing optically active carboxylic acids and esters⁵³. Their appeal lies in their independence from cofactors, high catalytic efficiency, and ease of operation. Recent surveys indicate that approximately 25% of commercialized pharmaceuticals and 40% of crop protection compounds contain at least one carboxylic group^54,55. This characteristic not only enhances solubility but also improves pharmaceutical efficacy, allowing easy derivatization into versatile functional groups like amines or hydroxyl groups⁵⁶. While biocatalysts are generally known for their high enantioselectivity, they often fall short of demonstrating the desired enantioselectivity, particularly for industrially relevant chiral compounds. This challenge becomes more pronounced when dealing with substrates possessing nearly symmetric structures, considered “hard-to-discriminate.” Chiral cyclohex-3-ene-1-carboxylic acids (CHCAs), characterized by a nearly symmetric hexatomic ring, are essential building blocks. Examples include (S)-CHCA for Edoxaban¹⁵ and (R)-CHCA for Oseltamivir²⁰. However, commercial and naturally-evolved enzymes typically exhibit low enantioselectivity toward CHCA esters²⁸. As a result, manipulating carboxylesterase enantioselectivity for the stereocomplementary synthesis of chiral CHCAs has become an area of particular interest. The enzyme enantioselectivity can be engineered through either quantity-intensive directed evolution or quality-intensive (semi-)rational design. To enhance screening throughput, substrate analogs such as p-nitrophenol ester, featuring an absorption peak at 405 nm, or 2,3-dibromopropanol ester (toxic to bacterial growth), which replaces glycerol (beneficial to bacterial growth), are commonly employed⁵⁷. However, these substrate analogs often differ, either wholly or partially, from the “real substrate.” Consequently, screening outcomes may deviate from the intended objectives. Machine learning (ML) has emerged as a powerful data-driven strategy for expediting directed evolution. The accuracy of a trained ML predictor relies heavily on a high-quality dataset and suitable features. Obtaining accurate enantioselectivity data for carboxylesterases at a high volume can be challenging. A high-throughput method, involving an alcohol dehydrogenase coupling, was developed to generate a high-quality dataset using the ratio of initial activity between (R)-CHCE and (S)-CHCE. Additionally, using atom numbers or a binary model of 0 and 1 as features to describe the volume and hydrophobicity of residues and substrates may yield implausible results. Considering parameters from biochemical perspectives, such as spatial volume for residue or substrate size, hydrophobic index for hydrophobicity, Kyte-Doolittle hydropathy scale for hydrophilicity, and isoelectric point for electrostatic properties, is of interest. To address these challenges, we envisioned to introduce biochemical features from experimentalists’ considerations in training ML predictors to expedite the stereodivergent evolution of carboxylesterase towards nearly symmetric esters. Seven features – spatial volume, hydrophobic, hydrophilicity, electrostatic, hydrogen bonding, π-π interaction, and distance to catalytic residue – were introduced as descriptors for training ML algorithms. Six ML models were trained, and a GBRT model was derived, exhibiting a high Pearson correlation of 0.93 and a low RMSE of 0.119. With the introduction of a dataset containing double mutants, the correlation was further enhanced to 0.97. The GBRT model demonstrated effectiveness in predicting the enantioselectivity of carboxylesterase AcEst1 towards CHCE. Moreover, the potential application of the GBRT model using the aforementioned features is under evaluation in learning and predicting the enantioselectivity of alcohol dehydrogenase from Kluyveromyces polysporus⁵⁸ and Baeyer–Villiger monooxygenase from Rhodococcus aetherivorans⁵⁹.

The application of GBRT accelerated the stereodivergent evolution of AcEst1, creating stereocomplementary mutants DR3 and DS6. Kinetic parameter analysis revealed that DR3 displayed a significantly increased k_cat/K_M ratio of (R)-CHCE to (S)-CHCE, while DS6 exhibited an increased k_cat/K_M ratio of (S)-CHCE to (R)-CHCE. DR3 and DS6 were successfully employed in the enantioselective resolution of other nearly symmetric esters. Molecular dynamic simulations and QM/MM calculations provided kinetic and thermodynamic evidence for manipulating the enantioselectivity of AcEst1 toward CHCE with a nearly symmetric structure. This study presents an effective ML predictor for expediting the directed evolution of carboxylesterase enantioselectivity. It introduces two stereocomplementary carboxylesterase mutants for the preparation of enantiomerically enriched 3-cyclohexene-1-carboxylic acids for the synthesis Edoxaban and Oseltamivir, and many other significant compounds.

Screening and protein engineering of alcohol dehydrogenases

Alcohol dehydrogenases (ADHs) stored in our lab were submitted for determination the oxidative activity toward ethanol. A 200 µL reaction system containing 0.5 mM NAD⁺ or NADP⁺, 800 mM or 2 mM ethanol, 10 µL ADH solutions and 170 µL Tris-HCl (pH 8.0, 100 mM) was performed at 30°C. Absorbance changes at 340 nm referring to the changes of NAD(P)H were monitored. One unit (U) of activity was defined as the amount of ADH required for the production of 1 µmol NAD(P)H per minute. All assays were performed in triplicate.

Substrate ethanol was docked into the crystal structure of ADH10 in complex with NADP⁺ (PDB ID. 5Z2X). Residues at the methyl terminal of ethanol including V84, F86M Y127, A128, F161, P195, S196, F197, V198, and T215 were selected for mutagenesis employing whole-plasmid PCR as previously reported and primers listed in Table S1. The resultant plasmids were digested with DpnI to remove the template plasmids and were chemically transformed into E. coli BL21(DE3). ADH10 mutants were induced by IPTG and obtained by ultrasonication as previously reported. Then, the specific activity toward 2 mM ethanol of ADH10 mutants were determined as above mentioned. Mutations of V84L and F197V were combined to obtain ADH10_V84L/F197V. WT and double mutant were purified employing nickel-affinity chromatography. The kinetic parameters of purified enzymes were determined by varying the concentrations of ethanol (1–800 mM). The K_M, V_max, k_cat values were calculated by non-linear fitting to Michaelis-Menten equation. All determinations were conducted in triplicate. Supernatant of ADH10_V84L/F197V was obtained by centrifuge and lyophilized under vacuum to obtain crude enzyme power, which was stored at 4°C for further use.

Establishment of high throughput method for determination of E_app

A high throughput method was established for determination of apparent enantioselectivity (E_app) of AcEst1 by coupling ADH10 _V84L/F197V based on released ethanol from (R)- and (S)-CHCE. A 200 µL reaction system consisting of (R)- or (S)-CHCE, NADP⁺, ADH10_V84L/F197V and AcEst1 was conducted at 30°C to spectrometrically monitor the increase of NADPH at 340 nm. First, background interference was exclude using the ADH-free, AcEst1-free, and CHCE-free systems. Then, different dosages of ADH10 _V84L/F197V including 5.0, 7.5, 15, and 30 U·mL^–1, NADP⁺ concentrations of 2.0, 3.0, 4.0, and 5.0 mM, CHCE concentrations of 0.75, 1.0, 2.0, and 5.0 mM, were systematically optimized. The final reaction system included 10 µL ADH10_V84L/F197V (7.5 U·mL^–1), 10 µL NADP⁺ (4 mM), 10 µL (R)- or (S)-CHCE (5.0 mM), 10 µL AcEst1 solution in 160 µL Tris-HCl (pH 8.0, 100 mM). E_app value was obtained by calculating the ratio of (R)-CHBE to (S)-CHBE or (S)-CHBE to (R)-CHBE (Eq. 1). A total of 96 replicates of WT AcEst1 was cultivated for determining their E_app and calculating the average E_app value and standard deviation of this method.

\({\text{E}}_{\text{app}}\text{=}\frac{\left(\text{R}\right)\text{-CHBE}}{\left(\text{S}\right)\text{-CHBE}}\) (Eq. 1)

Site-directed saturation mutagenesis and combinatorial mutagenesis of AcEst1

Site-directed saturation mutagenesis of AcEst1 was conducted with plasmid pET28-AcEst1 as template. Degenerative codons of 22c-trick including NDT, VHG, TGG were employed to introduce saturation mutagenesis (Table S2). Primers were mixed at ratio of 12:9:1 to encode all the twenty amino acids with only two redundant. Whole plasmid PCR was performed using KOD polymerase, and the resultant product was digested by DpnI as above mentioned and transformed into E. coli BL21(DE3). After evaluated by colony PCR and sequencing, positive colonies were inoculated into 96-well plate supplemented with LB medium and kanamycin. Then, the plate was cultivated at 37°C and sub-transferred to another 96-well plate for induction expression at 25°C for 12 h. The cells were collected by centrifuge and lysed by lysozyme at 37°C for 2 h. After centrifugation at 8000 × g and 4°C for 20 min, the supernatants were subjected for E_app analysis using above-described high throughput method. All measurements were performed in triplicate. The mutants with distinctly different E_app (higher or lower) compared to WT were recultivated for further evaluation. A high-quality dataset was established for subsequent machine learning.

Beneficial single mutants were randomly combined to construct double mutants employing the primers listed in Table S3. PCR reaction and transformation were conducted as above mentioned. Double mutants were verified by sequencing, and positive clones were inoculated in LB medium and cultivated overnight. After inductive expression by IPTG, double mutants were obtained and subjected for determination of E_app. Triple, quadruple, quintuable and sextuple mutants guided by machine learning were also constructed using primers listed in Table S3. After verification, their E_app and enantioselectivity value (E value) were determined.

Machine learning

The high-quality dataset of E_app was built based on the features including volume, hydrophobicity, hydropathy scale, isoelectric point, hydrogen bond, π-interaction, and distance to nucleophilic S201 listed in Table S4. Machine learning was performed on Jupyter Notebook installed with Scikit-Learn package. Six regression algorithms including GBRT, KRR, GPR, RF, SVR, and BRR were trained using the high-quality dataset. These hyperparameters were tuned with a ten-fold cross-validated grid-search on the training set (Table S5). Subsequently, these models were retrained on the training set with the above optimized hyperparameters and evaluated on the test set. The quality of ML predictors was evaluated by the Pearson coefficient (R²) and root mean squared error (RMSE) (Eq. 2–4). Based on the initial dataset with 760 mutants, the GBRT was proved to be the best algorithm to build the ML model, using hyperparameters of n_estimators = 250, learning_rate = 0.04, max_depth = 10, min_samples_split = 2, min_samples_leaf = 4.

\({\text{R}}^{\text{2}}\left(\text{y,}\widehat{\text{y}}\right)\text{=1-}\frac{{{\sum }_{\text{i}\text{=1}}^{\text{n}}\text{(}{\text{y}}_{\text{i}}-{\widehat{\text{y}}}_{\text{i}}\text{)}}^{\text{2}}}{{{\sum }_{\text{i}\text{=1}}^{\text{n}}\text{(}{\text{y}}_{\text{i}}-\stackrel{\text{-}}{\text{y}}\text{)}}^{\text{2}}}\) (Eq. 2)

\(\stackrel{\text{-}}{\text{y}}\text{=}\frac{\text{1}}{\text{n}}{\sum }_{\text{i}\text{=1}}^{\text{n}}{\text{y}}_{\text{i}}\) (Eq. 3)

\(\text{RMSE}\left(\text{y,}\widehat{\text{y}}\right)\text{=}\sqrt{\frac{\text{1}}{\text{n}}{\sum }_{\text{i}\text{=1}}^{\text{n}}{\text{(}{\text{y}}_{\text{i}}-{\widehat{\text{y}}}_{\text{i}}\text{)}}^{\text{2}}}\) (Eq. 4)

Where y_i and are \({\widehat{\text{y}}}_{\text{i}}\) the true and predicted values of the i^th mutant, respectively.

E _app data of (R)- and (S)-selective double mutants was added into the dataset, which was subsequently retrained the GBRT predictor. The refined GBRT predictor was evaluated by R² and RMSE. Furthermore, combinatorial mutants were virtually constructed and their E_app values were predicted by GBRT algorithm. Mutants with the highest E_app values at each round of combination were further experimentally verified using the above-mentioned method. The activity and enantioselectivity of combinatorial mutants were measured to achieve the stereodivergent evolution of (R)- and (S)-selective mutants.

Characterization of AcEst1 mutants

General protocol for determination the activity and enantioselectivity of AcEst1 and mutants was performed in a 10-mL reaction mixture containing 50 mM rac-CHCE and appropriate amount of enzyme solution in PBS buffer (pH 8.0, 100 mM) at 30°C and 180 rpm. At different time intervals, samples were withdrawn from the reaction mixture and the reaction was terminated by addition of 1.0 M HCl, followed by extraction with equal volume of ethyl acetate supplemented with 1 mM dodecane as the internal standard. The upper organic phase was isolated and dried over anhydrous Na₂SO₄, analyzed by chiral GC or HPLC (Table S6). The activity (U) was defined as the amount of enzyme required for the decrease of 1 µmol CHCE at above conditions. The enantiomeric excess (ee) and enantioselectivity value (E value) were calculated according to Eq. 5–7.

\({\text{ee}}_{\text{s}}\text{=}\frac{\left[\text{(}\text{S}\text{)-CHCE}\right]\text{ }-\text{ [(}\text{R}\text{)-CHCE]}}{\left[\text{(}\text{S}\text{)-CHCE}\right]\text{ + [(}\text{R}\text{)-CHCE]}}\text{×100\%}\) (Eq. 5)

\({\text{ee}}_{\text{p}}\text{=}\frac{\left[\text{(}\text{R}\text{)-CHCA}\right]\text{ }-\text{ [(}\text{S}\text{)-CHCA]}}{\left[\text{(}\text{R}\text{)-CHCA}\right]\text{ + [(}\text{S}\text{)-CHCA]}}\text{×100\%}\) (Eq. 6)

\(\text{E}\text{ value=}\frac{\text{ln}\left[\left(\text{1}-\text{conversion ratio}\right)\text{×}\left(\text{1}-{\text{ee}}_{\text{s}}\right)\right]}{\text{ln}\left[\left(\text{1}-\text{conversion ratio}\right)\text{×}\left(\text{1+}{\text{ee}}_{\text{s}}\right)\right]}\) (Eq. 7)

Kinetic parameters of AcEst1 and mutants were determined using standard protocol except for different substrate concentrations ranging from 1 mM to 20 mM. The initial activities were obtained. Subsequently, the K_M, V_max and k_cat values were calculated by non-linear fitting to Michaelis-Menten equation employing Origin package. All assays were performed for at least three times.

Application potential of AcEst1, DR3, and DS6 in the resolution of other nearly symmetric esters were performed employing the standard activity assay conditions. Different substrates were added in the reaction system and the specific activity was calculated according to the initial reaction rates. E values toward different substrates were determined at conversion ratios of 50% based on Eq. 7. All determinations were conducted in triplicate.

Enantioselective resolution of rac-CHCE by DR3 and DS6 at gram scale

Recombinant E. coli whole cells expressing (R)-selective DR3 and (S)-selective DS6 were prepared as above mentioned. The cells were lyophilized under vacuum to form the dry cells, which were stored at 4°C for further use. To achieve stereoselective synthesis of chiral CHCA at gram scale, 100 mmol (1.0 M) rac-CHCE, appropriate of dry cells of DR3 and DS6 were dissolved in 100 mL PBS buffer (pH 8.0, 100 mM) in a tri-neck flask. Preparative reaction was magnetically stirred at 180 rpm and 30°C. 1.0 M Na₂CO₃ was spontaneously titrated to maintain the pH of the reaction mixture at 8.0. The reaction was terminated until the ee_s achieving > 99% by adjusting the pH to basic condition and adding equal volume of ethyl acetate for three times. All the organic phases were combined and dried over anhydrous Na₂SO₄ overnight. Products were obtained by evaporation under vacuum and verified by ¹H-NMR and ¹³C-NMR.

MD simulations and QM/MM calculation

Structure models of WT AcEst1, DR3 and DS6 were built employing AlphaFold2 and verified by SAVES. (R)- and (S)-CHCE were docked into the active center of AcEst1, DR3 and DS6 using AutoDock Vina. Docking poses with relative higher scores were regarded as the initial structure from molecular dynamic (MD) simulations. Subsequently, AcEst1, DR3 and DS6 were protonated using H⁺⁺. All the MD simulations were performed using the Amber20 packages with GPU acceleration on the hypercomputer at the School of Biotechnology. The force fields for protein, CHCE and water molecules were ff14SB, gaff and TIP3P respectively. Sodium and chloride were added to neutralize simulation system, and a TIP3P water box with a clearance distance of 15 Å around the protein was added. There were 27208 atoms in the final simulation systems First, 10000-step of energy minimization with the steepest descent method was employed, followed by 1 ns heating at NVP from 0 K to 300 K, and 2 ns equilibrating at NPT and 300 K. Production was performed in NPT ensembles for 100 ns with dt of 0.002 ps. All MD simulations were conducted in at five replicates. Binding free energy was analyzed using MM-PB/GBSA, distance and angle were statistically analyzed employing Chimera 1.6, interactions between enzymes and substrates were analyzed with Discover studio 4.5 and visualized with Pymol.

Quantum mechanics/molecular mechanics (QM/MM) calculations were performed employing Turbomole 7.21⁶⁰ and DL-POLY2 modules within the Chemshell 3.7.1 software package3⁵². The QM region includes the catalytic triad Asp200-Ser201-Hid325 and the substrate molecule, totaling 65 atoms. The atoms in the QM region are treated with DFT methods, while the atoms in the MM region are subjected to energy calculations using the AMBER ff14SB force field4. For the polarization effects in the QM region, the static electric field embedding method is employed, and the QM/MM boundary is treated using the hydrogen-linking method of the charge transfer model. Considering the medium- and long-range dispersion interactions of the system, the dispersion correction method Gimme-D3 is applied in all QM/MM calculations. The structural optimizations are completed using M06-2X-D3/def-SVP/AMBER ff14SB, and single-point energy calculations are performed using M06-2X-D3/def2-TZVP.

Data availability

All data needed to evaluate the conclusions in the paper are available in the main text or the supplementary information. Source data are provided with this paper. There is no restriction on data availability.

Acknowledgements

We are grateful to the National Key R&D Program (2019YFA0906401), National Natural Science Foundation of China (22078127, 22378169), the Fundamental Research Funds for the Central Universities (111-2-06) for the financial support of this research.

Author information

Authors and Affiliations

Key laboratory of industrial Biotechnology, Ministry of Education, School of Biotechnology, Jiangnan University, Wuxi 214122, Jiangsu, China

Zhe Dou, Xuanzao Chen, Xiangyu Zheng, Xiaoyu Chen, Jiayu Xue, Guochao Xu, Ye Ni

Collaborative Innovation Center of Yangtze River Delta Region Green Pharmaceuticals, College of Pharmacy, Zhejiang University of Technology, Hangzhou 310014, Zhejiang, China

Zhe Dou

Graduate School of Engineering, Muroran Institute of Technology, 27-1, Mizumoto-cho, Muroran, Hokkaido 050-8585, Japan

Satomi Niwayama

Contributions

Y. N., G. X., and Z. D. designed the study. Z. D. and X. C. performed the experiments. X. Z., X. C., and J. X. analyzed the data. L. Z. performed QM/MM calculations. Z. D., S. N., G. X., and Y. N. wrote the manuscript. G. X. and Y. N. supervised the project. All authors contributed to the manuscript.

Contributions

Correspondence to Guochao Xu, Email: [email protected].

Ethics declarations

All authors declare no competing interests.

Supplementary information

Supplementary Information

Source data

High-quality dataset of AcEst1

Bornscheuer, G. W., Huisman, G. W., et al. Engineering the third wave of biocatalysis. Nature485, 185–194 (2002).
Devine, P. N., et al. Extending the application of biocatalysis to meet the challenges of drug development. Nat. Rev. Chem.2, 409–421 (2018).
Nazor, J., Liu, J., Huisman, G. Enzyme evolution for industrial biocatalytic cascades. Curr. Opin. Biotechnol.69, 182–190 (2021).
Reetz, M. T. Witnessing the birth of directed evolution of stereoselective enzymes as catalysts in organic chemistry. Adv. Synth. Catal.364, 3326–3335 (2022).
Falivene, L., Cao, Z., et al. Towards the online computer-aided design of catalytic pockets. Nat. Chem.11, 872–879 (2019).
Cheng, F., Chen, Y., et a. Controlling stereopreferences of carbonyl reductases for enantioselective synthesis of Atorvastatin precursor. ACS Catal.11, 2572–2582 (2021).
Wu, L. J., Qin, L., et al. Computer-aided understanding and engineering of enzymatic selectivity. Biotechnol. Adv.54, 107793 (2022).
Sousa, H. A., Afonso, C. A. M., Crespo, J. G. Kinetic study of the enantioselective hydrolysis of a meso-diester using pig liver esterase. J. Chem. Technol. Biotechnol.75, 707–710 (2000).
Niwayama, S., Cho, H. Practical large scale synthesis of half-esters of malonic acid. Chem. Pharm. Bull.57, 508–510 (2009).
Dou, Z., Chen, X. Z., et al. Kinetic resolution of nearly symmetric 3-cyclohexene-1-carbloxylate esters suing a bacterial carboxylesterase identified by genome mining. Org. Lett.23, 3043–3047 (2021).
Garcia-Urdiales, E., Alfonso, I., Gotor, V. Enantioselective enzymatic desymmetrizations in organic synthesis. Chem. Rev.105, 313–354 (2005).
Borissov, A., Davies, T. Q., et al. Organocatalytic enantioselective desymmetrization. Chem. Soc. Rev.45, 5474–5540 (2016).
Herdewijn, P., de Clercq, E. The cyclohexene ring as bioisostere of a furanose ring: synthesis and antiviral activity of cyclohexenyl nucleosides. Bioorg. Med. Chem. Lett.11, 1591–1597 (2001).
Nagata, T., Nagamochi, S., et al. Stereoselective synthesis and biological evaluation of 3,4-diaminocyclohexanecarboxylic derivatives as factor Xa inhibitors. Bioorg. Med. Chem. Lett.18, 4587–4592 (2008).
Michida, M., Ishikawa, H., et al. Development of an efficient manufacturing process for a key intermediate in the synthesis of Edoxaban. Org. Process Res. Dev.23, 524–534 (2019).
Marco, C., Paolo, C., et al. Practical, asymmetric synthesis of the cyclohexyl C28-C34 fragment of the immunosuppressant FK-506 via (S)-(-)-3-cyclohexenecarboxylic acid. Tetrahedron48, 539–544 (1992).
Trost, B. M., Kondo, Y. An asymmetric synthesis of (+)-phyllanthocin. Tetrahedron Lett.32, 1613–1616 (1991).
Toyota, M., Asoh, T., Matsuura, M., Fukumoto, K. Stereoselective transformation of enantiopure cyclohexanol into cis-hydrindan. An enantioselective formal total synthesis route to (+)-pumiliotoxin C. J. Org. Chem.61, 8687–8691 (1996).
Klun, J. A., Gupta, R. Optically active arthropod repellents for use against disease vectors. J. Med. Entomol.37, 182–186 (2000).
Raghavan, S., Babu, V. S. Enantioselective synthesis of oseltamivir phosphate. Tetrahedron67, 2044–2050 (2011).
Miyashita, K., Tsunemi, T., et al. Total synthesis of leustroducsin B via a convergent route. Tetrahedron Lett.48, 3829–3833 (2007).
Martin, S. F., Colapret, J. A., et al. Application of nitrile oxide cycloadditions to a convergent, asymmetric synthesis of (+)-phyllanthocin. J. Org. Chem.54, 2209–2216 (1989).
Miles, T. J., Barfoot, C., et al. Novel cyclohexyl-amides as potent antibacterials targeting bacterial type IIA topoisomerases. Bioorg. Med. Chem.21, 7483–7488 (2011).
Schwizer, D., Patton, J. T., et a. Pre-organization of the core structure of E-selectin antagonists. Chem. Eur. J.18, 1342–1351 (2012).
Wang, X., Ma, M. L., Reddy, A. G. K., Hu, W. H. An efficient stereoselective synthesis six stereoisomers of 3,4-diaminocyclohexane carboxamide as key intermediates for the synthesis of factor Xa inhibitors. Tetrahedron73, 1381–1388 (2017).
Riedl, R., Heilmayer, W., Spence, L. Enantiomerically pure amines. WO2011/146953A1, 2011.
Zhao, D. T., Yue, H., et al. Enzymatic resolution of ibuprofen in an organic solvent under ultrasound irradiation. Biotechnol. Appl. Biochem.61, 655–659 (2014).
Dou, Z., Xu, G. C., Ni, Y. Efficient microbial resolution of racemic methyl 3-cyclohexene-1-carboxylate as chiral precursor of Edoxaban by newly identified Acinetobacter sp. JNU9335. Enzym. Microb. Technol.139, 109580 (2020).
Wu, X. F., Yang, S. L., et al. Improved enantioselectivity of E. coli BioH in kinetic resolution of methyl (S)-3-cyclohexene-1-carboxylate by combinatorial modulation of steric and aromatic interactions. Biosci. Biotechnol. Biochem.83, 1263–1269 (2019).
Arnold, F. H. Innovation by evolution: Bringing new chemistry to life. Angew. Chem. Int. Ed.58, 14420–14426 (2019).
Shivange, A. V., Marienhagen, J. Advances in generating functional diversity for directed protein evolution. Curr. Opin. Chem. Biol.13, 19–25 (2009).
Ma, F. Q., Chung, M. T., et al. Efficient molecular evolution to generate enantioselective enzymes using a dual-channel microfluidic droplet screening platform. Nat. Commun.9, 1030 (2018).
Qu, G., Li, A. T., et al. The crucial role of methodology development in directed evolution of selective enzymes. Angew. Chem. Int. Ed.59, 13204–13231 (2020).
Cheng, F., Zhu, L. L., Schwaneberg, U. Directed evolution 2.0: improving and deciphering enzyme properties. Chem. Commun.51, 9760–9772 (2015).
Yang, K. K., Wu, Z., Arnold, F. H. Machine-learning-guided directed evolution for protein engineering. Nat. Methods16, 687–694 (2019).
Li, G. Y., Dong, Y. J., Reetz, M. T. Can machine learning revolutionize directed evolution of selective enzymes? Adv. Synth. Catal.361, 2377–2386 (2019).
Siedhoff, N. E., Schwaneberg, U., Davari, M. D. Machine learning-assisted enzyme engineering. Methods in Enzymology, 643, 281–315 (2020).
Cadet, F., Fontaine, N., et al. A machine learning approach for reliable prediction of amino acid interactions and its application in the directed evolution of enantioselective enzymes. Sci. Rep.8, 16757 (2018).
Ma, E. J., Siirola, E., et al. Machine-directed evolution of an imine reductase for activity and stereoselectivity. ACS Catal.11, 12433–12445 (2021).
Romero, P. A., Krause, A., Arnold, F. H. Navigating the protein fitness landscape with Gaussian processes. Proc. Natl. Acad. Sci. USA110, E193–E201 (2013).
Ao, Y. F., Pei, S. X., et al. Structure- and data-driven protein engineering of transaminases for improving activity and stereoselectivity. Angew. Chem. Int. Ed.62, e202301660 (2023).
Ran, X. C., Jiang, Y. Y. K., Shao, Q. Z., Yang, Z. Y. J. EnzyKR: a chirality-aware deep learning model for predicting the outcomes of the hydrolase catalyzed kinetic resolution. Chem. Sci.14, 12073 (2023).
Dou, Z., Xu, G. C., Ni, Y. A novel carboxylesterase from Acinetobacter sp. JNU9335 for efficient biosynthesis of Edoxaban precursor with high substrate to catalyst ratio. Bioresour. Technol.317, 123984 (2020).
Chen, C. S., Fujimoto, Y., Girdaukas, G., Sih, C. J. Quantitative analyses of biochemical kinetic resolutions of enantiomers. J. Am. Chem. Soc.104, 7294–7299 (1982).
Kille, S., Acevedo-Rocha, C. G., et al. Reducing codon redundancy and screening effort of combinatorial protein libraries created by saturation mutagenesis. ACS Synth. Biol.2, 83–92 (2013)..
Mazurenko, S., Prokop, Z., Damborsky, J. Machine learning in enzyme engineering. ACS Catal.10, 1210–1223 (2020).
Monera, O. D., Sereda, T. J., et al. Relationship of sidechain hydrophobicity and alphacal propensity on the stability of the single-stranded amphipathic alphax. J. Pept. Sci.1, 319–329 (1995).
Kyte, J., Doolittle, R. F. A simple method for displaying the hydropathic character of a protein. J. Mol. Biol.157, 105–132 (1982).
Pedregosa, F., Varoquaus, G., et al. Scikit-learn: Machine learning in python. J. Mach. Learn. Res.12, 2825–2830 (2011).
Hur, S., Bruice, T. C. The near attack conformation approach to the study of the chorismite to prephenate reaction. Proc. Natl. Acad. Sci. USA100, 12015–12020 (2003).
Hou, T., Wang, J., Li, Y., Wang, W. Assessing the performance of the MM/PBSA and MM/GBSA methods: 1. The accuracy of binding free energy calculations based on molecular dynamics simulations. J. Chem. Inf. Model.51, 69–82 (2011).
Metz, S., Kästner, J., et al. ChemShell—a modular software package for QM/MM simulations. WIREs Comput. Mol. Sci.4, 101–110 (2014).
Romano, D., Bonomi, F., et al. Esterases as stereoselective biocatalysts. Biotechnol. Adv.33, 547–565 (2015).
Halama, A., Zapadlo, M. Synthesis, isolation, and analysis of stereoisomers of sacubitril. Org. Process Res. Dev.23, 102–107 (2019).
Lanigan, R. M., Sheppard, T. D. Recent developments in amide synthesis: Direct amidation of carboxylic acids and transamidation reactions. Eur. J. Org. Chem.33, 7453–7465 (2013).
David, S. E., Timmins, P., Conway, B. R. Impact of the counterion on the solubility and physicochemical properties of salts of carboxylic acid drugs. Drug Dev. Ind. Pharm.38, 93–103 (2012).
Fernández-Álvaro, E., Snajdrova, R., et al. A combination of in vivo selection and cell sorting for the identification of enantioselective biocatalysts. Angew. Chem. Int. Ed.50, 8584–8587 (2011).
Xu, G. C., Wang, Y., et al., Hydroclassified combinatorial saturation mutagenesis: reshaping substrate binding pockets of KpADH for enantioselective reduction of bulky-bulky ketones. ACS Catal.8, 8336–8345 (2018).
Wei, S. Y., Xu, G. C., et al., Engineering of a Baeyer-Villiger monooxygenase reveals key residues for the asymmetric oxidation of omeprazole sulfide. Chem. Commun.58, 13246–13249 (2022).
Furche, F., Ahlrichs, R., et al. Turbomole. WIREs Comput. Mol. Sci.4, 91–100 (2014).

Tables 1 to 2 are available in the Supplementary Files section

There is NO Competing Interest.

Download PDF

Journal Publication

published 20 Oct, 2024

Read the published version in Nature Communications →

Version 1

posted

You are reading this latest preprint version

Enhanced stereodivergent evolution of carboxylesterase for efficient kinetic resolution of near-symmetric esters through machine learning

Status:

Journal Publication

Version 1

Abstract

Figures

Introduction

Results

Discussion

Methods

Screening and protein engineering of alcohol dehydrogenases

Machine learning

Characterization of AcEst1 mutants

MD simulations and QM/MM calculation

Declarations

References

Tables

Additional Declarations

Supplementary Files

Status:

Journal Publication

Version 1