Modification methods to improve the catalytic activity of UDP glycosyltransferase and the resulting mutants.

Abstract

This invention relates to a method for improving the catalytic activity of UDP glycosyltransferase and the resulting mutant, belonging to the field of enzyme engineering technology. The invention establishes a homology spatial structure model of UDP glycosyltransferase, cuts the rate-limiting region of the starting enzyme, and randomly intercalates the cut regions to obtain a chimeric enzyme sequence space. A greedy algorithm is used to determine the initial sequence with the highest information content in the sequence space. Through multiple iterations and in vitro verification, the iterative sequence with the highest enzyme activity is obtained as the optimal enzyme nucleotide sequence. The optimal UDP glycosyltransferase mutant obtained by this method has a catalytic activity 2.26 times that of the starting enzyme and can be applied to the production of rebaudine D using rebaudine A as a substrate. While promoting the industrial production of rebaudine D, it provides a novel method and approach for enzyme modification and high-throughput screening, demonstrating significant value and broad application prospects in the field of related compound production.

Description

Technical Field

[0001] This invention relates to the field of enzyme engineering technology, and in particular to a method for improving the catalytic activity of UDP glycosyltransferase and the resulting mutant. Background Technology

[0002] Excessive sugar intake is a global public health problem. Obesity, tooth decay, type II diabetes, and cardiovascular disease caused by high-sugar diets are common health issues in human society, with diabetes, in particular, becoming one of the most common metabolic diseases worldwide. To meet the ever-increasing demand for sugar while reducing the intake of high-calorie sugars, the development of sucrose substitutes with high sweetness and low calories has become a focus. Currently, various sweeteners have been developed as sucrose substitutes. Among these natural sweeteners, setviol glycosides (also known as stevia) are highly favored by consumers due to their high sweetness (300-450 times that of sucrose), low calorie value (1 / 300th that of sucrose), safety, stability, and inability to be directly absorbed and metabolized by the human body. As a natural sucrose substitute, it is widely used in the food and beverage industries and is hailed as the "world's third glycogen." Since its discovery in the last century, stevia has obtained national safety certifications in China, the United States, Europe, and other countries and has been approved for use as a food additive.

[0003] Stevia is a product derived from the purification of steviol glycosides extracted from stevia, a plant native to South America. Most steviol glycosides extracted from stevia by manufacturers are of the Reb A type. This type of steviol glycoside is the easiest to extract because of its high content, but it is the source of bitterness. In 2016, a stevia manufacturer, PureCircle, announced that they had extracted new, non-bitter steviol glycosides—Reb D and Reb M. Rebaudioside M (RebD) is a sweet component of stevia that has a pleasant taste and no aftertaste of bitterness, but its content in stevia is extremely low—Reb A accounts for one-third of the dried leaves of stevia, while Reb D and Reb M account for less than 1%. To address this issue, PureCircle is developing new stevia varieties rich in Reb D and Reb M, but the advanced technology and equipment required are heavily funded, making them unaffordable for most manufacturers.

[0004] Previous studies have successfully resolved the crystal structure of the rice-derived UDP-glucosyltransferase eugt11. While its ability to glycosylate Reb A to produce Reb D has been demonstrated, low catalytic efficiency, poor substrate specificity, and the lack of suitable high-throughput screening methods still hinder its potential industrial application. Directed evolutionary enzyme engineering still relies on an iterative two-step process. Random mutations first generate molecular diversity and in vitro recombination, followed by high-throughput screening to identify target proteins from a large library. This approach requires significant work, materials, and financial resources to select the desired phenotype, and completing a comprehensive search of the average protein sequence space remains a considerable challenge. Summary of the Invention

[0005] To address the aforementioned technical problems, this invention uses rice-derived UDP-glycosyltransferase as the starting enzyme. By establishing a UDP-glycosyltransferase domain model, the rate-limiting region of the starting enzyme is cut, and the cut regions are randomly intercalated to obtain a sequence space. The initial sequence with the largest information content in the sequence space is determined by Gaussian mutual information calculation and a greedy algorithm. Through multiple rounds of iteration combined with in vitro verification, the iterative sequence with the highest enzyme activity is obtained as the optimal enzyme nucleotide sequence.

[0006] The first objective of this invention is to provide a method for improving the catalytic activity of UDP glycosyltransferase, comprising the following steps:

[0007] Step S1: Using the UDP glycosyltransferase to be modified as the starting sequence, obtain the homologous structure of the UDP glycosyltransferase to be modified from the database through sequence comparison, and model the starting sequence and homologous structure to obtain the homologous spatial structure model of the UDP glycosyltransferase.

[0008] Step S2: Determine the cleavage site based on the homology spatial structure model of UDP glycosyltransferase, divide the starting sequence into multiple regions, and randomly intercalate the corresponding residues or peptides from the starting sequence or the corresponding position of the homology structure to obtain the sequence space of the chimeric enzyme.

[0009] Step S3: Determine the 20 sequences with the most information content in the sequence space as the initial sequences, and obtain the enzyme catalytic activity of the initial sequences;

[0010] Step S4: Based on the enzyme catalytic activity results of the initial sequence, train a Gaussian process regression model and a Gaussian Naive Bayes classifier, use the trained model to predict the enzyme catalytic activity of the sequence, and calculate the upper confidence bound of each sequence to obtain the 10 iterative sequences with the highest upper confidence bound and the enzyme catalytic activity of the iterative sequences.

[0011] Step S5: Repeat step S4 until the preset number of iterations is reached or the enzyme catalytic activity of the iteration sequence no longer increases. Output the iteration sequence with the highest enzyme catalytic activity as the optimal enzyme nucleotide sequence.

[0012] Further, in step S2, the distance between any two atoms in two residues is defined as less than... For each contact, residue pairs with more than 50 contacts or present in more than 50% of the homologous spatial structure models of UDP glycosyltransferases are counted, the relative abundance of the residue pairs is calculated, the degree of disruption of the chimeric protein structure by the residue pairs is assessed, and residue pairs with less disruption are selected as cleavage sites.

[0013] Furthermore, the degree of damage is calculated using the following formula:

[0014]

[0015] Wherein, DD(Chimera) represents the degree of damage; I(i,j) represents 1 if residues i and j interact or contact in the structure, and 0 otherwise; δ(g(i)≠g(j)) represents 1 if the pair of interacting residues come from different parent sequences, and 0 otherwise; N represents the number of residues in the UDP glycosyltransferase.

[0016] Furthermore, in step S3, the selection of the initial sequence is based on a greedy algorithm and the calculation of Gaussian entropy. Starting from the starting sequence and the homologous structure, a selected set is constructed, all unselected sequences in the sequence space are scanned, and the entropy increment brought by each sequence after being added to the selected set is calculated to obtain the 20 initial sequences with the most information.

[0017] Furthermore, the formula for calculating the Gaussian entropy is as follows:

[0018]

[0019] Where H(S) represents the Gaussian entropy of the sequence subset S; det(Σ(S)) represents the determinant of the covariance matrix Σ(S); the covariance matrix Σ(S) reflects the relationship between residues or features at different positions in the sequence subset S; n is the size of the sequence subset; ln is the logarithm with the natural constant e as the base.

[0020] The formula for calculating the entropy increment is as follows:

[0021] Δ r (x)=H(S∪{x})-H(S)=lndet(Σ(S∪{x}))-lndet(Σ(S))

[0022] Where, Δ rH(x) represents the entropy increment brought about by adding the candidate sequence x to the selected sequence subset S; H(S∪{x}) represents the Gaussian entropy of the new set after adding the candidate sequence x to the sequence subset S, and H(S) represents the Gaussian entropy of the original sequence set S.

[0023] Further, in step S4, the hyperparameters of the Gaussian process regression model are optimized using the following formula, adjusting the parameters to minimize NLML(θ):

[0024]

[0025] Where NLML(θ) represents an indicator used to evaluate the goodness of a Gaussian process regression model, θ represents the hyperparameters; y represents the vector of true experimental values ​​of the measured sequence; y T It is the transpose of y; K(θ) represents the kernel matrix constructed from the hyperparameter θ and the measured sequence. -1 It is its inverse matrix; ln|K(θ)| represents taking the natural logarithm of the determinant of the kernel matrix; n1 represents the number of sequences measured.

[0026] Further, in step S4, a Gaussian Naive Bayes classifier is used to distinguish between active and inactive sequences. The specific steps include:

[0027] Step S41: Collect known sequences and their corresponding enzyme activity data, and divide these data into a training set and a test set. The training set contains known sequences and their enzyme activities, and the test set contains sequences whose activities are to be predicted.

[0028] Step S42: Train the Gaussian Naive Bayes model using the training set data. The model calculates the prior probability and conditional probability of the features at different activity levels based on the features of the input sequence and the corresponding activity labels.

[0029] Step S43: Input the sequences in the test set into the trained Gaussian Naive Bayes model, calculate the posterior probability of each sequence belonging to a different activity category, and determine whether the sequence is an active or inactive sequence based on the magnitude of the posterior probability.

[0030] A second objective of this invention is to provide the application of the above-described modification method in improving the catalytic activity of UDP glycosyltransferase.

[0031] A third objective of this invention is to provide a UDP glycosyltransferase mutant obtained using the above-described modification method, wherein the nucleotide sequence of the UDP glycosyltransferase mutant is shown in SEQ ID NO.4 or SEQ ID NO.5.

[0032] A fourth objective of this invention is to provide an expression vector comprising the above-described UDP glycosyltransferase mutant.

[0033] A fifth objective of this invention is to provide a strain comprising the above-described UDP glycosyltransferase mutant or the above-described expression vector.

[0034] A sixth object of the present invention is to provide a microbial preparation comprising the above-described strains.

[0035] A seventh object of the present invention is to provide the use of the above-mentioned UDP glycosyltransferase mutant, the above-mentioned expression vector, the above-mentioned strain or the above-mentioned microbial preparation in the production of one or more of rebaudioside D, rebaudioside M and stevia.

[0036] An eighth object of the present invention is to provide a method for producing rebaudioside D, the method comprising the step of adding the above-described strain or the above-described microbial preparation to a fermentation system.

[0037] A ninth object of the present invention is to provide a method for producing betaine, the method comprising the step of adding the above-mentioned strain or the above-mentioned microbial preparation to a fermentation system.

[0038] The beneficial effects of this invention are:

[0039] This invention provides a highly efficient enzyme redesign method based on machine learning, which can systematically modify UDP glycosyltransferases to improve their catalytic activity. This method combines techniques such as sequence comparison, domain modeling, greedy algorithms, and Gaussian process regression models to achieve precise modification of enzyme structure and activity prediction. Through iterative optimization, the iterative sequence with the highest enzyme activity is output as the optimal enzyme nucleotide sequence, thereby significantly improving the catalytic activity of UDP glycosyltransferase. The optimal UDP glycosyltransferase mutant obtained after modification has a catalytic activity 2.26 times that of the starting enzyme, which has important value in the production of compounds such as rebaudioside D, can significantly improve production efficiency, and has broad application prospects. Attached Figure Description

[0040] To make the content of this invention easier to understand, the invention will be further described in detail below with reference to specific embodiments and accompanying drawings, wherein...

[0041] Figure 1 This is a schematic diagram of the sequence space construction steps in an embodiment of the present invention;

[0042] Figure 2 This is a schematic diagram illustrating the principle of multi-round rapid iteration in an embodiment of the present invention;

[0043] Figure 3 This is a graph showing the relative catalytic activity detection results of multiple iteration sequences in an embodiment of the present invention. Detailed Implementation

[0044] The present invention will be further described below with reference to the accompanying drawings and specific embodiments, so that those skilled in the art can better understand and implement the present invention. However, the embodiments described are not intended to limit the present invention.

[0045] The detection methods involved in the embodiments are as follows:

[0046] High performance liquid chromatography-mass spectrometry (LC-MS) detection method: Waters (MICROMASS QUATTRO MICRO), BEH-C18 column, mobile phase acetonitrile and 0.1% formic acid aqueous solution, column temperature 45℃, injection volume 1μL.

[0047] Enzyme activity is defined as the ability of 1 U of enzyme to catalyze the conversion of 1 μmol of substrate or the generation of product within 1 minute under specified conditions (temperature, pH, etc.).

[0048] The formula for calculating enzyme activity is:

[0049] (1) Total enzyme activity (U) = [amount of substrate produced or consumed (μmol)] / [reaction time (min)];

[0050] (2) Specific enzyme activity (U / mL) = Total enzyme activity ÷ Enzyme solution volume used (mL) or (3) Specific enzyme activity (U / mg) = Total enzyme activity ÷ Enzyme protein content (mg).

[0051] The strains involved in the examples are as follows:

[0052] Plasmid storage and construction: Escherichia coli DH5α;

[0053] Target protein expression: Escherichia coli BL21(DE3).

[0054] The culture media involved in the examples are as follows:

[0055] LB medium: yeast extract 5 g / L, peptone 10 g / L, NaCl 10 g / L.

[0056] TB medium: yeast extract 24 g / L, peptone 12 g / L, glycerol 5.04 g / L, potassium dihydrogen phosphate 2.31 g / L, dipotassium hydrogen phosphate 12.54 g / L.

[0057] Example 1: Construction of original and iterative sequence plasmids

[0058] The rice-derived UDP-glucosyltransferase eugt11 (NCBI ID: XP_015629141.1) was modified. Studies have shown that eugt11 can add glucose to both ends of steviol glycoside substrates. eugt11 consists of two Rossmann-like domains at the N-terminus and C-terminus, a typical GT-B folding pattern and a characteristic of glycosyltransferases. Sequence comparison identified two other UDP-glucosyltransferases highly homologous to eugt11: the tomato-derived UDP-glucosyltransferase ugtsl2 (NCBI ID: XP_004250485.1) and the stevia-derived UDP-glucosyltransferase ugt91d2 (NCBI ID: B3VI56.1). Following the codon preference of *E. coli*, the nucleotide sequences of these three UDP-glucosyltransferases were codon optimized, and the entire genome was synthesized. The codon-optimized nucleotide sequence of eugt11 is shown in SEQ ID NO.1.

[0059] The nucleotide sequence of ugt91d2 is shown in SEQ ID NO.2, and the nucleotide sequence of ugtsl2 is shown in SEQ ID NO.3.

[0060] I. Plasmid Construction: Using pET-22b plasmid as a vector, and the fully synthesized eugt11, ugt91d2, and ugtsl2 plasmids as templates, primers were designed in E. coli DH5α to construct the original plasmids pET22b-eugt11, pET22b-ugt91d2, and pET22b-ugtsl2 according to the overlapping extension PCR primer design method. Subsequent iteration sequence fragments were constructed using the three original plasmids as templates.

[0061] II. Homology Modeling: Homology modeling was performed using MODELLER software, with the sequences of eugt11, ugt91d2, and ugtsl2 as templates (the structures of ugt91d2 and ugtsl2 were predicted by AlphaFold). 100 homology models were created for each of the three template sequences. The aim was to comprehensively cover all possible conformations of the protein structure and increase the diversity of the models.

[0062] III. Contact Counting: A contact counting matrix is ​​created for each template. The residue information from the PDB database is aligned with the template sequence, creating a zero-value matrix with both sides being the aligned length of the template. The corresponding data are then counted in the counting matrix. Except for hydrogen and oxygen atoms, the distance between any two atoms in two residues is less than [missing information]. A contact is considered a high-abundance contact if it occurs more than 50 times or in more than 50% of the models. The relative abundance of this type of contact is then calculated for subsequent analysis.

[0063] IV. Determining Breakpoints: Based on the relative abundance calculated above, breakpoints suitable for gene recombination are identified. The locations of these breakpoints are selected by balancing structural disruption with library diversity to ensure that structural integrity is preserved as much as possible during recombination.

[0064] Suppose a protein has N residues. A contact matrix I(i,j) represents whether there is a significant contact (or contact weight) between residue i and residue j. During the construction of a chimeric protein, partial sequences are selected from different parent sequences and spliced ​​together to form a new chimeric protein. To track the origin of each residue in the new chimeric protein, a function g(i) is introduced to record which parent sequence residue i comes from.

[0065] The “destructive quantity” defined by SCHEMA is as follows:

[0066]

[0067] Where DD(Chimera) represents the degree of damage; I(i,j) means that if residues i and j interact (or contact) in the structure, it is 1 (or the corresponding weight), and if not, it is 0. It takes into account the interaction between residues i and j, and reflects the importance or strength of the interaction between them.

[0068] δ(g(i)≠g(j)) represents the interaction of the two residues (residue i and residue j) from different parent sequences, i.e. from different parents during the splicing process to form a chimeric protein. In this case, δ(g(i)≠g(j))=1, indicating that the two interacting residues have undergone parent switching, which may disrupt the interaction between residues in the original structure. When δ(g(i)≠g(j))=0, it means that no such disruption has occurred.

[0069] The amount of damage can be used as an indicator to assess the degree of structural damage to chimeric proteins. The greater the amount of damage, the more severe the damage to the original structure may be during the splicing of chimeric proteins from different parent sequences, ultimately affecting the structural stability and function of the chimeric protein.

[0070] The SCHEMA-RASPP tool was used to analyze breakpoint compatibility to ensure effective binding between different sequence blocks during recombination. Breakpoints were determined as follows: for eugt11, base pairs 219, 315, 432, 495, 558, 600, and 1344; for ugt91d2, base pairs 207, 294, 411, 474, 537, 603, and 1353; and for ugtsl2, base pairs 192, 279, 396, 459, 495, 537, and 1272. All breakpoints were based on codon-optimized sequences. Each of the eight sequence blocks could be genetically defined from one of three natural enzymes. 8 The combined sequence space of the sequences, therefore the total diversity of the library is 3. 8 = 6561 sequences.

[0071] Example 2: Iterative modification of eugt11 using semi-rational modification of machine learning

[0072] I. Generation of the initial sequence

[0073] To select the 20 most informative sequences from a total of 6561 chimeric enzyme sequences, a diversity initial sampling method based on a greedy algorithm was employed. The key to this method is that the selected sequences must broadly cover the diversity of the sequence space and reflect the variability of different structural and functional regions. The specific process is as follows: starting with three parental sequences, a selected set S is constructed iteratively. In each step, all unselected enzyme sequences are scanned, and the entropy increment resulting from adding each sequence to the current selected set S is calculated.

[0074] For a sequence subset S of size n, its corresponding covariance matrix is ​​Σ(S), which reflects the relationship between residues or features at different positions in the sequence subset.

[0075] The Gaussian entropy H(S) (ignoring the constant term) is defined as H(S)∝ln·det(Σ(S)), specifically:

[0076]

[0077] Where H(S) represents the Gaussian entropy of the sequence subset S; det(Σ(S)) represents the determinant of the covariance matrix Σ(S); the covariance matrix Σ(S) reflects the relationship between residues or features at different positions in the sequence subset S; n is the size of the sequence subset; ln is the logarithm with the natural constant e as the base.

[0078] At each step of the greedy algorithm, it is necessary to compare the information increment resulting from adding the candidate sequence x to the selected set S. The specific information increment is:

[0079] Δ r (x)=H(S∪{x})-H(S)=lndet(Σ(S∪{x}))-ln det(Σ(S))

[0080] Where, Δ r H(x) represents the entropy increment brought about by adding candidate sequence x to the selected sequence subset S; H(S∪{x}) represents the Gaussian entropy of the new set after adding candidate sequence x to the sequence subset S, and H(S) represents the Gaussian entropy of the original sequence set S. The information increment brought about by candidate sequence x can be obtained by calculating the difference between the two.

[0081] The goal of the greedy algorithm is to select candidate sequences that maximize information increment at each step. This means that the selected sequence x can maximize the uncertainty or information content of set S, thereby enabling subsequent models to better learn and capture features in the sequence space. By continuously selecting such sequences, the greedy algorithm gradually builds a set containing the most informative sequences to achieve effective coverage and exploration of the entire sequence space. In this process, since Gaussian entropy is a subset function, the greedy algorithm, while not guaranteeing to find the global optimum, can efficiently approximate the optimum, providing a relatively good selection strategy under limited computational resources and time, until the selected set S contains 20 sequences. Because entropy is a subset function, the greedy algorithm can efficiently approximate the optimum. The selection of these sequences is based on the principle of maximizing Gaussian mutual information, that is, selecting those sequences that can reflect the features of the entire sequence space to the greatest extent. In this study, Gaussian mutual information is simplified to Gaussian entropy, that is, selecting those sequences that maximize the entropy value. Through this method, 20 most informative sequences were gradually selected. These sequences not only have broad coverage and diversity but also provide comprehensive initial data for subsequent machine learning models.

[0082] II. Verification of the Initial Sequence

[0083] A plasmid containing 20 initial sequences was constructed. This first-round iteration sequence plasmid was transformed into *E. coli* BL21(DE3) competent cells. Transformants with correct colony PCR results were selected for sequencing verification. The catalytic efficiency was verified through in vitro enzyme catalysis experiments to obtain the relative catalytic activity corresponding to the first-round iteration sequences. The experimental steps for verifying the catalytic efficiency in vitro are as follows:

[0084] (1) Inoculate a loopful of a single colony into LB medium containing 50 μg / mL ampicillin and pre-culture overnight at 37°C and 220 rpm. Inoculate the seed culture at an initial inoculum volume of 4% into 50 mL of TB fermentation medium containing 50 μg / mL ampicillin and incubate at 37°C and 220 rpm for 1-2 h until OD. 600The concentration was adjusted to 0.8-1.0. Then, 1 mM isopropyl-β-D-thiogalactopyranoside (IPTG) was added to induce protein expression. After induction at 16°C for 20 hours, cells were harvested by centrifugation at 7,000 rpm for 3 minutes at 4°C. Cells were washed twice with Tris-HCl buffer (pH 8.0) and disrupted by sonication. The supernatant was collected by centrifugation at 10,000 rpm for 30 minutes at 4°C to obtain the crude enzyme.

[0085] (2) 0.3 mM steviol glycoside and Reb A were added to 200 μL of a reaction mixture consisting of 20 mM Tris-HCl buffer (pH = 8.0) and crude enzyme, respectively. The mixture was incubated at 30 °C for 20 h. After the reaction was completed, 800 μL of acetonitrile was added to terminate the reaction and the hydrophobic steviol glycoside substrate and product (stevioside and Reb D, respectively) were extracted. After centrifugation at 12,000 rpm for 10 min, the supernatant was filtered through a 0.22 μm aqueous filter membrane and analyzed by LC-MS.

[0086] III. Generation and Optimization of Iteration Sequences

[0087] Using experimental data from the first round of stevioside-based synthesis of stevioside (stevioside yield), a Gaussian Process Regression (GP) model and a Gaussian Naive Bayes (GNB) classifier were trained. The GP model was primarily used to predict sequence titers, while the GNB classifier was used to distinguish between active and inactive sequences. These two models played a crucial guiding role in subsequent sequence design and optimization.

[0088] The hyperparameters of the Gaussian process model need to be optimized, specifically:

[0089]

[0090] Where NLML(θ) represents an indicator used to evaluate the goodness of a Gaussian process regression model, θ represents all hyperparameters such as the kernel function and noise; y represents the vector of true experimental values ​​of the measured sequence (such as enzyme activity, product titer); y T K(θ) is the transpose of y; K(θ) represents the (covariance) kernel matrix constructed from the hyperparameter θ and the measured sequence. -1 It is its inverse matrix; ln|K(θ)| represents the natural logarithm of the determinant of the kernel matrix; n1 represents the number of training data (measured sequences).

[0091] The hyperparameter θ represents all parameters including the kernel function and noise. The kernel function determines how the similarity between data points in the Gaussian process model is measured; different kernel functions lead to different model behaviors and fitting abilities. The noise parameter considers the unavoidable random errors during experimental measurements. The goal of model optimization is to find a suitable value that minimizes NLML(θ). This is because minimization implies balancing the difference between the model's predictions and the true values ​​(via y). T K(θ) -1 While controlling for y (the parameter) and model complexity (controlled by ln|K(θ)|), an optimal model configuration is found. Through continuous adjustment using gradient descent or other optimization algorithms, the model's performance on the training data can be optimized, meaning it can fit the data well, thereby improving the model's predictive ability on unseen data. This allows the model to provide more accurate predictions when forecasting the catalytic activity of different UDP glycosyltransferase sequences.

[0092] In machine learning-based research on modifying UDP glycosyltransferases, existing experimental data (such as enzyme activity or product titer data of different mutant sequences) are used as y. The sequence-related features of UDP glycosyltransferases (such as sequence composition and structural features) are constructed into a kernel (covariance) matrix K(θ) using a kernel function. Then, by optimizing the hyperparameter θ, the NLML(θ) is minimized, resulting in a Gaussian process model that fits the existing data well. This trained model can be used to predict the enzyme activity or product titer of new UDP glycosyltransferase sequences. By inputting the relevant features of the new sequence, the model can calculate predicted values ​​based on the optimized kernel function and hyperparameters, helping to screen sequences with potentially higher catalytic activity. This guides further experiments and enzyme modification work, improving the efficiency of finding highly efficient UDP glycosyltransferase mutants.

[0093] The formula for the predicted mean of GP is as follows:

[0094] μ(x * )=k(x * ) T [K+σ 2 I] -1 y

[0095] Where, μ(x) * ) is the model for the new sequence x * The predicted mean of the objective function (such as enzyme activity or product yield); k(x * ) is the new sequence x * The kernel vector between all measured sequences; K is the kernel matrix between measured sequences (plus the noise term σ). 2I), where y is the vector of experimental observations of the measured sequence. Based on the information from the measured sequence, the observations of the measured sequence are weighted and averaged according to the similarity between the new sequence and the measured sequence to obtain the new sequence x. * The predicted mean.

[0096] The formula for predicting variance is as follows:

[0097] σ 2 (x * )=k(x * x * )-k(x * ) T [K+σ 2 I]- 1 k(x * )

[0098] Where, σ 2 (x * ) represents the new sequence x * The predicted mean of the objective function, i.e., the model's prediction of the new sequence x. * Average expected values ​​in terms of enzyme activity or product yield; k(x) * x * ) is the new sequence x * Its own core value, through the new sequence x * Subtract the kernel value of the new sequence x calculated based on the measured sequence information from its own kernel value. * The "expected similarity". This difference represents the new sequence x. * The uncertainty of the predicted value. If the prediction variance is large, it indicates that the model is not accurate for the new sequence x. * The prediction is uncertain, possibly because of the new sequence x. * The new sequence x differs significantly from previously measured sequences in the kernel space, or the data from previously measured sequences are insufficient to accurately describe the new sequence x. * The characteristics of the model; conversely, if the prediction variance is small, it indicates that the model is accurate for the new sequence x. * The predictions are relatively more confident, as the behavior of the new sequence in the kernel space is quite similar to that of the previously measured sequences, and the information from the previously measured sequences can be well used to predict the new sequence x. * The properties of.

[0099] GNB usually stands for "Gaussian Naive Bayes" classifier, used to distinguish between active and inactive sequences. Based on the idea of ​​Naive Bayes, it assumes that the features are independent of each other and that each feature follows a Gaussian (normal) distribution.

[0100] We make the feature independence assumption of Gaussian Naive Bayes. Since it is a "naive" assumption, we define the features of x in each dimension x. iTreated as conditionally independent, specifically:

[0101]

[0102] Under the Gaussian assumption, for each class c and each feature x i They are all assumed to follow a certain one-dimensional normal distribution. Therefore:

[0103]

[0104] Where, μ i，c and σ i，c In the training data, the estimated mean and standard deviation of the i-th feature based on the samples under category c are the feature dimensions (for protein sequences, these features may be physicochemical properties, structure / sequence encoding, or other artificially extracted descriptors). Multiplying these distributions together yields P(x|y=c).

[0105] When applied to protein sequences, each sequence is first encoded into several features (such as...). Figure 2 As shown, in the sequence structure, 1-1 means that sequence block 1 originates from parent sequence 1, and 2-2 means that sequence block 2 originates from parent sequence 2. The sequence blocks correspond to the 8 sequence blocks described in Example 1, and are obtained by cutting the 219th, 315th, 432nd, 495th, 558th, 600th, and 1344th base pairs of eugt11 as 7 breakpoints. Then, these features are input into the GNB classifier to obtain a posterior probability: P(Active|x) and P(Inactive|x). If P(Active|x) > P(Inactive|x), the sequence is judged as an "active sequence"; otherwise, it is judged as an "inactive sequence".

[0106] The classification process selects the maximum posterior probability. In actual classification, if only binary classification is performed, the comparison is made between P(y = Active|x) and P(y = Inactive|x), or a logarithmic comparison is performed, specifically as follows:

[0107]

[0108] Choose the larger category based on the results.

[0109] Therefore, the GNB classification process can be summarized as follows:

[0110] (1) Estimate the prior probability P(y=c) = (number of samples in this class) / (total number of samples) using the training data, and the μ of each feature in this class. i,c and σ i,c ;

[0111] (2) For the test sample x, calculate its likelihood in each category c:

[0112]

[0113] (3) Combine the prior P(y=c) with the posterior P(y=c|x) and select the category corresponding to the maximum value.

[0114] To further enhance enzyme activity, an Upper Confidence Bound (UCB) optimization strategy was employed to select the most promising sequences for optimization. The UCB standard considers not only the predicted uncertainty region but also the predicted mean, thus efficiently balancing exploration and development. It aims to rapidly converge to the optimal sequence, providing strong theoretical support for the optimization process. In practice, the GP model was used to predict the titers and corresponding uncertainties (i.e., standard deviations) of all untested chimeric enzyme sequences. Then, the UCB value for each sequence was calculated, and the sequence with the highest UCB value was selected for further study. These sequences not only have the potential for high activity but also cover the uncertainty region of the model. In each iteration, 10 sequences were designed according to the batch model UCB standard, corresponding plasmids were constructed, and they were transformed into *E. coli*. The titers of betaine produced by these *E. coli* were then measured using LC-MS technology.

[0115] The specific formula for the optimization strategy of the upper confidence bound is as follows:

[0116] UCB(x) = μ(x) + κσ(x)

[0117] Where μ(x) represents the mean prediction of the objective function (e.g., enzyme activity or product yield) by the Gaussian process model; σ(x) represents the uncertainty (standard deviation) of the model's prediction at that point; and κ is a hyperparameter that is manually set or updated with iterations. The trained Gaussian process model is used to predict the titers (objective function values) and their corresponding uncertainties (standard deviations) of all untested chimeric enzyme sequences. Then, for each sequence x, its UCB(x) value is calculated. In each iteration, several sequences are designed according to the batch model UCB standard (e.g., 10 sequences per iteration). The sequences with the highest values ​​are selected for further study; these sequences not only have the potential for high activity but also cover the uncertainty region of the model. By continuously iterating this process, the sequence selection is optimized, and the enzyme activity is gradually improved.

[0118] In early iterations, a larger κ value may be set to explore the sequence space more broadly and discover potentially high-activity sequences that might have been overlooked. As iterations progress and more knowledge of the sequence space is gained, the κ value can be appropriately reduced to utilize sequences surrounding already discovered high-activity regions for optimization, allowing the optimization process to converge to the optimal solution more quickly. In this way, the UCB optimization strategy can effectively utilize existing information while ensuring that no potentially high-activity sequences are overlooked (exploration), improving the efficiency of finding efficient UDP glycosyltransferase mutants, reducing unnecessary experiments, and accelerating the enzyme modification process.

[0119] In this embodiment, a total of 7 rounds of UCB sequence optimization were performed, and the results of the 7 rounds of enzyme catalytic activity verification are as follows: Figure 3 As shown, the first round of iteration sequences are named A1-A20, and the second round of iteration sequences are named B1-B10. The sequences that lost catalytic activity were not included in the results. Figure 3 This is reflected in the results. By integrating the results, it can be seen that the enzyme's catalytic activity gradually increases. However, after the seventh round of sequence detection, it was found that the enzyme's catalytic activity no longer increased and instead showed a decreasing trend. Therefore, it was concluded that the optimal sequence was obtained. The most significant increase in catalytic activity was observed in sequences F4 and F6 from the sixth iteration. The relative catalytic activity of F6 was 2.26 times that of wild-type EUGT11, and the nucleotide sequence of F6 is shown in SEQ ID NO.4. The relative catalytic activity of F4 was 1.99 times that of wild-type EUGT11, and the nucleotide sequence of F4 is shown in SEQ ID NO.5.

[0120] Obviously, the above embodiments are merely illustrative examples for clear explanation and are not intended to limit the implementation. Those skilled in the art will recognize that other variations or modifications can be made based on the above description. It is neither necessary nor possible to exhaustively list all possible implementations here. However, obvious variations or modifications derived therefrom are still within the scope of protection of this invention.

Claims

1. A method for improving the catalytic activity of UDP glycosyltransferase, characterized in that, Includes the following steps: Step S1: Using the UDP glycosyltransferase to be modified as the starting sequence, obtain the homologous structure of the UDP glycosyltransferase to be modified from the database through sequence comparison, and model the starting sequence and homologous structure to obtain the homologous spatial structure model of the UDP glycosyltransferase. Step S2: Determine the cleavage site based on the homology spatial structure model of UDP glycosyltransferase, defining the distance between any two atoms of two residues as less than [missing information]. For each contact, residue pairs with more than 50 contacts or more than 50% of the models in the homology spatial structure model of UDP glycosyltransferase are counted, the relative abundance of the residue pairs is calculated, the degree of damage of the residue pairs to the chimeric protein structure is evaluated, residue pairs with less damage are selected as cleavage sites, the starting sequence is divided into multiple regions, and corresponding residues or peptides are obtained from the starting sequence or the corresponding position of the homology structure for random chimerism to obtain the sequence space of the chimeric enzyme. Step S3: Determine the 20 sequences with the most information content in the sequence space as the initial sequences, and obtain the enzyme catalytic activity of the initial sequences; Step S4: Based on the enzyme catalytic activity results of the initial sequence, train a Gaussian process regression model and a Gaussian Naive Bayes classifier, use the trained model to predict the enzyme catalytic activity of the sequence, and calculate the upper confidence bound of each sequence to obtain the 10 iterative sequences with the highest upper confidence bound and the enzyme catalytic activity of the iterative sequences. Step S5: Repeat step S4 until the preset number of iterations is reached or the enzyme catalytic activity of the iteration sequence no longer increases. Output the iteration sequence with the highest enzyme catalytic activity as the optimal enzyme nucleotide sequence.

2. The method according to claim 1, characterized in that, The degree of damage is calculated using the following formula: Wherein, DD(Chimera) represents the degree of damage; I(i,j) represents 1 if residues i and j interact or contact in the structure, and 0 otherwise; δ(g(i)≠g(j)) represents 1 if the pair of interacting residues come from different parent sequences, and 0 otherwise; N represents the number of residues in the UDP glycosyltransferase.

3. The modification method according to claim 1, characterized in that: In step S3, the selection of initial sequences is based on a greedy algorithm and the calculation of Gaussian entropy. Starting from the starting sequence and the homologous structure, a selected set is constructed. All unselected sequences in the sequence space are scanned, and the entropy increment brought by each sequence after being added to the selected set is calculated to obtain the 20 initial sequences with the most information.

4. The modification method according to claim 3, characterized in that, The formula for calculating the Gaussian entropy is as follows: Where H(S) represents the Gaussian entropy of the sequence subset S; det(Σ(S)) represents the determinant of the covariance matrix Σ(S); the covariance matrix Σ(S) reflects the relationship between residues or features at different positions in the sequence subset S; n is the size of the sequence subset; ln is the logarithm with the natural constant e as the base. The formula for calculating the entropy increment is as follows: D r (x)=H(S∪{x})-H(S)=ln det(Σ(S∪{x}))-ln det(Σ(S)) Where, Δ r H(x) represents the entropy increment brought about by adding the candidate sequence x to the selected sequence subset S; H(S∪{x}) represents the Gaussian entropy of the new set after adding the candidate sequence x to the sequence subset S, and H(S) represents the Gaussian entropy of the original sequence set S.

5. The modification method according to claim 1, characterized in that, In step S4, the hyperparameters of the Gaussian process regression model are optimized using the following formula, adjusting the parameters to minimize the model index: Where NLML(θ) represents an indicator used to evaluate the goodness of a Gaussian process regression model, θ represents the hyperparameters; y represents the vector of true experimental values ​​of the measured sequence; y T It is the transpose of y; K(θ) represents the kernel matrix constructed from the hyperparameter θ and the measured sequence. -1 It is its inverse matrix; ln|K(θ)| represents taking the natural logarithm of the determinant of the kernel matrix; n1 represents the number of sequences measured.

6. The modification method according to claim 1, characterized in that, In step S4, a Gaussian Naive Bayes classifier is used to distinguish between active and inactive sequences. The specific steps include: Step S41: Collect known sequences and their corresponding enzyme catalytic activity data, and divide these data into a training set and a test set. The training set contains known sequences and their enzyme catalytic activities, and the test set contains sequences whose enzyme catalytic activities are to be predicted. Step S42: Train the Gaussian Naive Bayes model using the training set data. The model calculates the prior probability and conditional probability of different enzyme activity levels based on the features of the input sequence and the corresponding activity labels. Step S43: Input the sequences in the test set into the trained Gaussian Naive Bayes model, calculate the posterior probability of them belonging to different enzyme activity categories, and determine whether the sequence is an active sequence or an inactive sequence based on the magnitude of the posterior probability.

7. The application of the modification method according to any one of claims 1-6 in improving the catalytic activity of UDP glycosyltransferase.

8. A UDP glycosyltransferase mutant obtained using any one of the modification methods described in claims 1-6, characterized in that: The nucleotide sequence of the UDP glycosyltransferase mutant is shown in SEQ ID NO.4 or SEQ ID NO.

5.

9. An expression vector comprising the nucleotide sequence of the UDP glycosyltransferase mutant of claim 8.

10. A strain comprising the UDP glycosyltransferase mutant of claim 8 or the expression vector of claim 9.

11. A microbial preparation comprising the strain of claim 10.

12. The use of the UDP glycosyltransferase mutant of claim 8, the expression vector of claim 9, the strain of claim 10, or the microbial preparation of claim 11 in the production of one or more of rebaudioside D, rebaudioside M, and stevia.

13. A method for producing rebaudioside D, characterized in that: The method includes the step of adding the strain of claim 11 or the microbial preparation of claim 12 to the fermentation system.

14. A method for producing betaine, characterized in that: The method includes the step of adding the strain of claim 11 or the microbial preparation of claim 12 to a fermentation system for fermentation production.

Images