Enzyme kinetic parameter prediction method and application thereof

Abstract

The invention relates to an enzyme kinetic parameter prediction method and application thereof. A brand-new enzyme kinetic parameter prediction framework based on artificial intelligence is designed, the framework is used for modeling enzyme-substrate information based on a protein language model, a compound molecule language model and molecular fingerprints, and then a machine learning algorithm is applied to fit enzyme kinetic parameters including the enzyme conversion number kcat, the Michaelis constant Km and the catalytic efficiency kcat / Km. And training is carried out on a clustered 10-fold cross validation data set, the prediction accuracy and the model generalization ability are improved, a new enzyme mining scheme is further developed, and the success rate of enzyme mining can be remarkably improved by combining a traditional sequence alignment method.

Description

Technical Field

[0001] The invention belongs to the field of biotechnology and relates to an enzyme kinetic parameter prediction method and application thereof. Background Art

[0002] As a biological catalyst, enzymes can significantly increase the speed of chemical reactions, and their catalytic efficiency is many times higher than that of inorganic catalysts. New enzyme discoveries provide new candidate enzymes for the development of industrial enzyme preparations, improving the efficiency and sustainability of enzyme-catalyzed processes in industrial production. For example, in biofuel production, food processing, and pharmaceutical synthesis, new enzymes can improve reaction efficiency, reduce costs, and reduce the production of by-products. Traditional wet laboratory screening usually requires high time and economic costs, and does not meet the needs of industrial production. Therefore, the prediction of enzyme kinetic parameters has gradually become one of the hot topics in the field of enzyme research.

[0003] Enzyme kinetic parameter prediction is an important research field, especially in biotechnology and industrial applications, where accurate predictions can help design and optimize enzyme performance. For example, a study reported a framework called UniKP, which is an enzyme kinetic parameter prediction framework based on a pre-trained large language model and a machine learning model. The key components of the UniKP framework include a representation module and a machine learning module. The representation module is responsible for converting complex enzyme and substrate information into vector representations, while the machine learning module uses these vectors for parameter prediction. In the representation module, the enzyme sequence representation module uses the ProtT5-XL-UniRef50 model to encode the enzyme information, and the substrate structure representation module uses the SMILESTransformer model to encode the substrate information. It can predict a variety of different enzyme kinetic parameters, including enzyme turnover number (kcat), Michaelis constant (Km), and catalytic efficiency (kcat / Km) given the amino acid sequence of the enzyme and the structural information of the substrate.

[0004] However, the current enzyme kinetic parameter prediction methods still have some limitations, such as the predicted kcat value is not accurate enough and the applicability is low. Therefore, developing new enzyme kinetic parameter prediction tools to improve accuracy and applicability is still one of the urgent problems to be solved in the enzyme research field. Summary of the invention

[0005] In view of the deficiencies of the prior art and actual needs, the present invention provides an enzyme kinetic parameter prediction method and its application, in order to expand the enzyme kinetic parameter prediction tool, improve the accuracy and applicability, and promote the discovery of new enzymes.

[0006] To achieve this object, the present invention adopts the following technical solutions:

[0007] In a first aspect, the present invention provides a method for predicting enzyme kinetic parameters, characterized in that the method for predicting enzyme kinetic parameters comprises:

[0008] The kinetic parameters and substrate information of known enzymes are collected as a data set; the kinetic parameters include the turnover number of the enzyme (kcat), the Michaelis constant of the enzyme (Km), and the ratio between the turnover number of the enzyme and the Michaelis constant of the enzyme (kcat / Km);

[0009] The enzyme representation was obtained using the ProtT5 language model; the substrate representation was obtained using the MolT5 language model and the MACC fingerprint generated by the RDKit software package;

[0010] The obtained enzyme representation and substrate representation are spliced ​​together to obtain an enzyme-substrate representation, which is used as a training set. A machine learning algorithm is used to train and establish the relationship between the enzyme-substrate representation and the enzyme turnover number and the enzyme's Michaelis constant, respectively, to obtain an enzyme turnover number prediction model and an enzyme Michaelis constant prediction model, and the enzyme turnover number and the enzyme Michaelis constant are predicted based on the prediction model;

[0011] The enzyme turnover number prediction model and the enzyme Michaelis constant prediction model are used to predict and calculate the ratio between the enzyme turnover number and the enzyme Michaelis constant as the initial value, and then the machine learning algorithm is used to train to establish the relationship between the enzyme-substrate representation and the ratio between the enzyme turnover number and the enzyme Michaelis constant, and the ratio between the enzyme turnover number and the enzyme Michaelis constant is predicted as the correction value. Finally, the ratio between the enzyme turnover number and the enzyme Michaelis constant is the initial value plus the correction value in proportion, and the sum of the weights of the two is 1.

[0012] In the present invention, an artificial intelligence-based enzyme kinetic parameter prediction framework is designed. The framework models enzyme-substrate information based on a protein language model, a compound molecular language model, and a molecular fingerprint, and then applies a machine learning algorithm to fit the enzyme's kinetic parameters, including the enzyme's turnover number kcat, the Michaelis constant Km, and the catalytic efficiency kcat / Km, to improve prediction accuracy and model generalization ability.

[0013] Preferably, the sources of the known enzyme kinetic parameters and substrate information include the BEANDA database and the SABIO-RK database.

[0014] Preferably, the weight ratio of the initial value to the correction value is (0.3-0.7):(0.3-0.7), for example, it can be 0.6:0.4 or 0.5:0.5.

[0015] Preferably, the enzyme kinetic parameter prediction method also includes the step of clustering and grouping the data set, specifically including: clustering according to the enzyme sequence similarity of 0.4, and dividing into 10 groups, the enzyme sequence similarity between each group is less than 0.4, and obtaining a 10-fold unbiased cross-validation data set.

[0016] In the present invention, for each sample in the data set, it is clustered by protein sequence similarity of 0.4 and divided into 10 groups. For each model (kcat, or KM, or kcat / Km model), it is trained on the clustered 10-fold cross-validation data set, and finally 10 groups of model weights are saved to further improve the generalization ability of the model.

[0017] Preferably, the machine learning algorithm comprises a deep learning algorithm.

[0018] Preferably, the deep learning algorithm includes a neural network algorithm or an attention mechanism algorithm.

[0019] Preferably, the enzyme kinetic parameter prediction method comprises the following steps:

[0020] (1) Based on the BEANDA database and the SABIO-RK database, the kinetic parameters and substrate information of known enzymes are collected as a data set; the kinetic parameters include the turnover number of the enzyme, the Michaelis constant of the enzyme, and the ratio between the turnover number of the enzyme and the Michaelis constant of the enzyme; clustering is performed according to the enzyme sequence similarity of 0.4, and the enzyme sequences of each group are clustered into 10 groups, and the enzyme sequence similarity between each group is less than 0.4, and a 10-fold unbiased cross-validation data set is obtained;

[0021] (2) Based on the obtained 10-fold unbiased cross-validation dataset, the ProtT5 language model was used to obtain the enzyme representation; the MolT5 language model and the MACC fingerprint generated by the RDKit software package were used to obtain the substrate representation;

[0022] The obtained enzyme representation and substrate representation are spliced ​​together to obtain an enzyme-substrate representation, which is used as a training set. A neural network algorithm is used to train and establish the relationship between the enzyme-substrate representation and the enzyme turnover number and the enzyme's Michaelis constant, to obtain an enzyme turnover number prediction model and an enzyme Michaelis constant prediction model, and the enzyme turnover number and the enzyme Michaelis constant are predicted based on the prediction model;

[0023] (3) The enzyme turnover number prediction model and the enzyme Michaelis-Menten constant prediction model are used to predict and calculate the ratio of the enzyme turnover number to the enzyme Michaelis-Menten constant as an initial value, and then the attention mechanism algorithm is used to train to establish the relationship between the enzyme-substrate representation and the ratio of the enzyme turnover number to the enzyme Michaelis-Menten constant, and the ratio of the enzyme turnover number to the enzyme Michaelis-Menten constant is predicted as a correction value. Finally, the ratio of the enzyme turnover number to the enzyme Michaelis-Menten constant is the sum of the initial value and the correction value in a certain ratio.

[0024] In a second aspect, the present invention provides an enzyme kinetic parameter prediction device, the system comprising:

[0025] A data set construction module is used to collect kinetic parameters of known enzymes and substrate information as a data set; the kinetic parameters include the turnover number of the enzyme, the Michaelis constant of the enzyme, and the ratio between the turnover number of the enzyme and the Michaelis constant of the enzyme; the enzyme representation is obtained using the ProtT5 language model; the substrate representation is obtained using the MACC fingerprint generated by the MolT5 language model and the RDKit software package;

[0026] The enzyme turnover number and enzyme Michaelis-Menten constant prediction module is used to splice the obtained enzyme representation and substrate representation together to obtain the enzyme-substrate representation as a training set, and use the machine learning algorithm to train and establish the relationship between the enzyme-substrate representation and the enzyme turnover number and the enzyme Michaelis-Menten constant, obtain the enzyme turnover number prediction model and the enzyme Michaelis-Menten constant prediction model, and predict the enzyme turnover number and the enzyme Michaelis-Menten constant based on the prediction model;

[0027] The ratio prediction module between the enzyme turnover number and the enzyme's Michaelis constant uses the enzyme turnover number prediction model and the enzyme's Michaelis constant prediction model to predict and calculate the ratio between the enzyme turnover number and the enzyme's Michaelis constant as an initial value, and then uses a machine learning algorithm to train and establish the relationship between the enzyme-substrate representation and the ratio between the enzyme turnover number and the enzyme's Michaelis constant, and predicts the ratio between the enzyme turnover number and the enzyme's Michaelis constant as a correction value. The final ratio between the enzyme turnover number and the enzyme's Michaelis constant is the sum of the initial value and the correction value in a certain ratio.

[0028] Preferably, the sources of the known enzyme kinetic parameters and substrate information include BEANDA and SABIO-RK databases.

[0029] Preferably, the enzyme kinetic parameter prediction method also includes the step of clustering and grouping the data set, specifically including: clustering according to the enzyme sequence similarity of 0.4, and dividing into 10 groups, the enzyme sequence similarity between each group is less than 0.4, and obtaining a 10-fold unbiased cross-validation data set.

[0030] Preferably, the machine learning algorithm comprises a deep learning algorithm.

[0031] Preferably, the deep learning algorithm includes a neural network algorithm or an attention mechanism algorithm.

[0032] In a third aspect, the present invention provides an electronic device, comprising a processor and a memory, wherein the processor is used to implement the enzyme kinetic parameter prediction method as described in the first aspect when executing a computer program stored in the memory.

[0033] In a fourth aspect, the present invention provides a computer-readable storage medium having a computer program stored thereon, and when the computer program is executed by a processor, the enzyme kinetic parameter prediction method as described in the first aspect is implemented.

[0034] In a fifth aspect, the present invention provides a new enzyme mining method, the new enzyme mining method comprising:

[0035] The natural enzyme that catalyzes the reaction is used as the initial enzyme, and a multiple sequence alignment algorithm is used to obtain a new sequence similar to the initial enzyme sequence, and the new sequences are sorted from high to low according to the sequence similarity, filtered based on a certain sequence similarity threshold, or the sequences with the highest preset number of sequence similarity rankings are retained;

[0036] Calculate the difference in sequence length between the remaining sequences and the initial enzyme, retain the sequences with a length difference within a certain range, use the enzyme kinetic parameter prediction method described in the first aspect to predict the kinetic parameters of the enzyme corresponding to the retained sequence, sort them from high to low according to the enzyme kinetic parameters, retain the top preset number of sequences, use the protein structure prediction tool to predict the protein structure of the retained sequence, calculate the structural similarity with the initial enzyme, and screen new enzymes based on structural similarity.

[0037] The present invention further develops a new enzyme mining scheme based on the designed enzyme kinetics prediction method, and can significantly improve the success rate of enzyme mining by combining it with the traditional sequence alignment method.

[0038] Compared with the prior art, the present invention has at least the following beneficial effects:

[0039] The present invention designs a brand-new enzyme kinetic parameter prediction framework based on artificial intelligence. The framework models enzyme-substrate information based on protein language model, compound molecular language model and molecular fingerprint, and then applies machine learning algorithm to fit the enzyme kinetic parameters, including the enzyme turnover number kcat, Michaelis constant Km, and catalytic efficiency kcat / Km. The framework is trained on a clustered 10-fold cross-validation data set to improve prediction accuracy and model generalization ability, and further develop new enzyme mining schemes. Combining with traditional sequence alignment methods can significantly improve the success rate of enzyme mining. BRIEF DESCRIPTION OF THE DRAWINGS

[0040] Figure 1aSchematic diagram of clustering based on protein sequence similarity;

[0041] Figure 1b Schematic diagram of the kcat and Km model output through a neural network;

[0042] Figure 1c Schematic diagram of the kcat / Km model output through the attention mechanism;

[0043] Figure 2 Graphs showing the comparison results of the prediction method of the present invention and other prediction schemes on a 10-fold cross validation data set, wherein graph a is the Pearson correlation coefficient PCC result predicted by kcat, graph b is the Spearman rank correlation coefficient result predicted by kcat, graph c is the root mean square error RMSE result predicted by kcat, graph d is the Pearson correlation coefficient PCC result predicted by Km, graph e is the Spearman rank correlation coefficient result predicted by Km, and graph f is the root mean square error RMSE result predicted by Km;

[0044] Figure 3 Schematic diagram of the new enzyme mining process. DETAILED DESCRIPTION

[0045] The technical solution of the present invention is further described below in conjunction with the accompanying drawings and through specific implementation methods. However, the following examples are only simplified examples of the present invention and do not represent or limit the scope of protection of the present invention. The scope of protection of the present invention shall be subject to the claims.

[0046] If no specific techniques or conditions are specified in the examples, the techniques or conditions described in the literature in the field or the product instructions are used. If no manufacturer is specified for the reagents or instruments used, they are all conventional products that can be purchased through regular channels.

[0047] The present invention proposes an enzyme kinetic parameter prediction framework named CataPro. First, enzyme kcat, Km, kcat / Km data are clustered according to the enzyme sequence similarity of 0.4 and divided into 10 components respectively, and the enzyme sequence similarity between each component is lower than 0.4. All trainings are carried out on corresponding 10-fold unbiased crossover data sets. For kcat and Km prediction models, the expression of enzyme is represented by the feature vector generated by ProtT5 model, and the substrate is represented by the MACC fingerprint generated by MolT5 feature vector and RDKit. The obtained enzyme vector and substrate vector are spliced ​​together to obtain the final enzyme-substrate representation, and then kcat and Km are output by neural network (such as perceptron, multi-layer perceptron, convolutional neural network and recurrent neural network, etc.). For the kcat / Km model, training is divided into two steps. First, use kcat and Km to pre-train the kcat and Km models, and use these two models to pre-generate initial kcat / Km values; then use attention mechanisms (such as soft attention, response attention, self-attention, cross attention, and multi-head attention, etc.) to learn a correction term based on the enzyme-substrate feature vector as a supplement to the initial kcat / Km value, and then add the initial kcat / Km value to the correction term in a certain proportion to obtain the final kcat / Km value. When training the correction term, the weights of the pre-trained kcat and Km models are no longer updated.

[0048] RDKit is an open source cheminformatics software package that provides a variety of functions, including molecular representation, geometry optimization, substructure query, and machine learning integration. In RDKit, you can calculate multiple types of molecular fingerprints, including MACC fingerprints. MACC (Molecular Access System) fingerprints are a binary code based on molecular structure that describes molecules through a series of predefined chemical structure features.

[0049] ProstT5 is a novel protein language model (pLM) that can process both one-dimensional (1D) sequence and three-dimensional (3D) structural information of proteins. This model is based on the existing ProtT5 model and fine-tuned to achieve translation between protein sequence and structure.

[0050] The present invention also proposes an enzyme mining process based on AI combined with traditional sequence alignment methods (such as Figure 2As shown), first, the natural enzyme that catalyzes the reaction is used as the initial enzyme, and a multiple sequence alignment algorithm is used to obtain a new sequence similar to the initial enzyme sequence, and the sequence is sorted from high to low according to the sequence similarity, and filtered based on a certain sequence similarity threshold (such as 0.4), or only the sequences with the highest sequence similarity ranking (such as the top 1000) are retained. Then, the difference in sequence length between the remaining sequences and the initial enzyme is calculated, and the sequences with length differences within a certain range are retained (such as the length difference within 10 amino acids), and then the model for predicting enzyme kinetic parameters is applied to predict the kcat and Km of these sequences, and these sequences are sorted by kcat and Km, and only the top-ranked sequences (such as the top 10%) are retained. Finally, the structures corresponding to these sequences are generated by protein structure prediction tools (such as AlphaFold2), and the structural similarity of these enzymes with the initial enzyme is calculated and sorted, and the proteins with the highest structural similarity ranking are retained. Finally, candidate enzymes are selected based on clustering methods, such as PCA, t-SNE and other algorithms.

[0051] Example 1

[0052] This embodiment provides a method for predicting enzyme kinetic parameters.

[0053] Construct a data set, the process diagram is as follows Figure 1a As shown in the figure, all kcat and Km data were collected from the BEANDA and SABIO-RK databases and cleaned. The enzyme-substrate data containing both kcat and Km data constituted the kcat / Km data set. For each kcat, Km and kcat / Km data set, CD-HIT was used to cluster the enzyme sequences according to the sequence similarity of 0.4, and then the clustered clusters were divided into 10 components, thereby obtaining a 10-fold unbiased cross-validation data set.

[0054] Model training was performed on the corresponding 10-fold unbiased crossover dataset. For the kcat and Km prediction models, the enzyme was represented by the feature vector generated by the ProtT5 model, and the substrate was represented by the feature vector of MolT5 and the MACC fingerprint generated by RDKit. The enzyme vector and substrate vector were concatenated to obtain the final enzyme-substrate representation, and then kcat and Km were output by the perceptron. Figure 1b ); For the kcat / Km model, training is divided into two steps ( Figure 1c), first, use kcat and Km to pre-train the kcat and Km models, and use these two models to pre-generate the initial kcat / Km value; then use the self-attention mechanism to learn a correction term based on the enzyme-substrate feature vector as a supplement to the initial kcat / Km value, and then add the initial kcat / Km value to the correction term in a certain ratio (for example, the ratio of the initial kcat / Km to the correction term weight is 0.6:0.4) to obtain the final kcat / Km value, wherein when training the correction term, the weights of the pre-trained kcat and Km models are no longer updated.

[0055] The performance of UniKP, DLKcat and other feature combination schemes were tested on the 10-fold cross validation dataset of kcat and Km. The results are as follows Figure 2 As shown, the protein features we tested include embeddings from ESM2, ProtT5, and SaProt protein language models, and small molecule features include embeddings from MolT5 and Mole-BERT language models as well as MACC, morgan, and RDKitFP fingerprints. Figure 2 The performance of different feature combinations is listed in . CataPro+SaProt represents a model in which the embedding of SaProt is added to the CataPro enzyme feature, ProtT5-SaProt_MolT5 represents the enzyme using the embeddings of ProtT5 and SaProt, and the embedding of MolT5 represents the substrate information; and so on. The Pearson correlation coefficient PCC ( Figure 2 Figure a), Spearman rank correlation coefficient SCC ( Figure 2 (middle b) and root mean square error RMSE ( Figure 2 (c in the figure) are significantly better than UniKP and DLKcat. Similarly, in the Km prediction task, CataPro also achieves the most beneficial performance ( Figure 2 Middle D Figure - Figure 2 (Fig. f).

[0056] Example 2

[0057] This example takes the discovery of a new enzyme that replaces carotenoid cleavage oxygenase (CSO2) to catalyze 4-vinylguaiacol to produce vanillin as an example to verify the new enzyme discovery method of the present invention.

[0058] application Figure 3The enzyme mining process in , first, the natural enzyme of the catalytic reaction is used as the initial enzyme, and the multiple sequence alignment algorithm is used to obtain new sequences similar to the initial enzyme sequence, and the sequences are sorted from high to low according to the sequence similarity, filtered based on a certain sequence similarity threshold (0.4), or only the sequences with the highest sequence similarity ranking (the top 1000) are retained, then the difference in sequence length between the remaining sequences and the initial enzyme is calculated, and the sequences with length differences within a certain range are retained (the length difference is within 10 amino acids), and then the model for predicting enzyme kinetic parameters is applied to predict the kcat and Km of these sequences, and these sequences are sorted by kcat and Km, and only the top-ranked sequences (the top 10%) are retained, and finally the structures corresponding to these sequences are generated by AlphaFold2, and the structural similarity of these enzymes with the initial enzyme is calculated and sorted, and the proteins with the highest structural similarity ranking are retained, and finally the candidate enzymes are selected based on the clustering method, and finally 5 candidate enzymes are mined, and their amino acid sequences and DNA sequences are shown in Tables 1 and 2, respectively. The relative activity of the enzymes is tested by wet experiments, and the specific experimental steps include:

[0059] (1) Gene cloning and expression

[0060] The synthesized enzyme sequence was cloned into the NdeI and XhoI sites of the pET28a(+) vector and the expression of the enzyme was confirmed by Hieff The recombinant expression vector was cloned according to the instructions of PlusOne Step Cloning Kit (Cat#10911), and then transferred into E. coli BL21 (DE3). Then, colony PCR was performed to screen positive bacteria. The selected single clone was inoculated into 4 mL of liquid LB medium containing 50 μg / mL kanamycin, and cultured at 37°C and 220 r / min for 6 hours. The bacterial liquid was transferred to 50 mL of liquid 2YT medium containing 50 μg / mL kanamycin at a 2% inoculation rate, and the OD was continued at 37°C and 220 r / min. 600 To about 1.0, add inducer IPTG to a final concentration of 0.2mM, FeCl2 to a final concentration of 1mM, 25℃, 200r / min to induce expression for 16h. Collect the bacteria by refrigerated centrifugation at 4℃, 10000r / min for 5min. Resuspend the collected bacteria with (50mM pH 7.0PB, 150mM NaCl, 10% glycerol) buffer, ultrasonically disrupt the bacteria (50% power, on 2s, off 3s, total ultrasonic 15min), centrifuge at 4℃, 10000r / min for 20min, collect the supernatant as the crude enzyme solution, and analyze the soluble expression amount by SDS-PAGE.

[0061] (2) Enzyme activity assay

[0062] The enzyme activity assay reaction system 200μL consists of 60mM 4VG, 10μL 10-fold diluted crude enzyme solution, and 50mM pH 9.5 Na2CO3-NaHCO3 buffer. The reaction solution was placed in a 25℃ incubator for 60min, 20μL 2.5M HCl was added to terminate the reaction, and centrifuged at 4000rpm for 10min. 50μL 2,4-dinitrobenzene (8g / L, dissolved in 2.5M HCl) was added to a 96-well ELISA plate, 10μL supernatant reaction solution was added, and the mixture was shaken at room temperature for 3min, 80μL 4M NaOH solution was added, and the mixture was shaken at room temperature for 2min, and the absorbance at 480nm was measured.

[0063] The results of enzyme activity measured by wet experiment are shown in Table 3. It was found that the activities of two enzymes, PpCSO and SsCSO, were higher than that of the original CSO2, among which the activity of SsCSO was 19.53 times that of CSO2 (Table 3), which shows the practicality of the enzyme mining method proposed in the present invention.

[0064] Table 1 Amino acid sequences of five enzymes

[0065]

[0066]

[0067] Table 2 DNA sequences of enzymes

[0068]

[0069]

[0070]

[0071] Table 3 Enzyme activities measured by wet test

[0072]

[0073]

[0074] In summary, the present invention designs a new artificial intelligence-based enzyme kinetic parameter prediction framework, which models enzyme-substrate information based on protein language model, compound molecular language model, and molecular fingerprint, and then applies a machine learning algorithm to fit the enzyme's kinetic parameters, including the enzyme's turnover number kcat, Michaelis constant Km, and catalytic efficiency kcat / Km. It is trained on a clustered 10-fold cross-validation data set to improve prediction accuracy and model generalization ability, and further develop new enzyme mining solutions. Combined with traditional sequence alignment methods, the success rate of enzyme mining can be significantly improved.

[0075] The applicant declares that the above is only a specific implementation mode of the present invention, but the protection scope of the present invention is not limited thereto. Those skilled in the art should understand that any changes or substitutions that can be easily thought of by those skilled in the art within the technical scope disclosed by the present invention shall fall within the protection scope and disclosure scope of the present invention.

Claims

1. A method for predicting enzyme kinetic parameters, characterized in that: The enzyme kinetic parameter prediction method comprises: The kinetic parameters and substrate information of known enzymes are collected as a data set; the kinetic parameters include the turnover number of the enzyme, the Michaelis constant of the enzyme, and the ratio between the turnover number of the enzyme and the Michaelis constant of the enzyme; The enzyme representation was obtained using the ProtT5 language model; the substrate representation was obtained using the MolT5 language model and the MACC fingerprint generated by the RDKit software package; The obtained enzyme representation and substrate representation are spliced ​​together to obtain an enzyme-substrate representation, which is used as a training set. A machine learning algorithm is used to train and establish the relationship between the enzyme-substrate representation and the enzyme turnover number and the enzyme's Michaelis constant, respectively, to obtain an enzyme turnover number prediction model and an enzyme Michaelis constant prediction model, and the enzyme turnover number and the enzyme Michaelis constant are predicted based on the prediction model; The enzyme turnover number prediction model and the enzyme Michaelis constant prediction model are used to predict and calculate the ratio between the enzyme turnover number and the enzyme Michaelis constant as the initial value, and then the machine learning algorithm is used to train to establish the relationship between the enzyme-substrate representation and the ratio between the enzyme turnover number and the enzyme Michaelis constant, and the ratio between the enzyme turnover number and the enzyme Michaelis constant is predicted as the correction value. Finally, the ratio between the enzyme turnover number and the enzyme Michaelis constant is the initial value plus the correction value in proportion, and the sum of the weights of the two is 1.

2. The method for predicting enzyme kinetic parameters according to claim 1, characterized in that: The sources of the known enzyme kinetic parameters and substrate information include BEANDA database and SABIO-RK database; Preferably, the weight ratio of the initial value to the correction value is (0.3-0.7):(0.3-0.7).

3. The method for predicting enzyme kinetic parameters according to claim 1 or 2, characterized in that: The enzyme kinetic parameter prediction method further comprises the step of clustering the data set, specifically comprising: Clustering was performed based on the enzyme sequence similarity of 0.4 and divided into 10 groups. The enzyme sequence similarity between each group was less than 0.4, and a 10-fold unbiased cross-validation data set was obtained.

4. The method for predicting enzyme kinetic parameters according to any one of claims 1 to 3, characterized in that: The machine learning algorithm includes a deep learning algorithm; Preferably, the deep learning algorithm includes a neural network algorithm or an attention mechanism algorithm.

5. The method for predicting enzyme kinetic parameters according to any one of claims 1 to 4, characterized in that: The enzyme kinetic parameter prediction method comprises the following steps: (1) Based on the BEANDA database and the SABIO-RK database, the kinetic parameters and substrate information of known enzymes are collected as a data set; the kinetic parameters include the turnover number of the enzyme, the Michaelis constant of the enzyme, and the ratio between the turnover number of the enzyme and the Michaelis constant of the enzyme; clustering is performed according to the enzyme sequence similarity of 0.4, and the enzyme sequences of each group are clustered into 10 groups, and the enzyme sequence similarity between each group is less than 0.4, and a 10-fold unbiased cross-validation data set is obtained; (2) Based on the obtained 10-fold unbiased cross-validation dataset, the ProtT5 language model was used to obtain the enzyme representation; the MolT5 language model and the MACC fingerprint generated by the RDKit software package were used to obtain the substrate representation; The obtained enzyme representation and substrate representation are spliced ​​together to obtain an enzyme-substrate representation, which is used as a training set. A neural network algorithm is used to train and establish the relationship between the enzyme-substrate representation and the enzyme turnover number and the enzyme's Michaelis constant, to obtain an enzyme turnover number prediction model and an enzyme Michaelis constant prediction model, and the enzyme turnover number and the enzyme Michaelis constant are predicted based on the prediction model; (3) The enzyme turnover number prediction model and the enzyme Michaelis-Menten constant prediction model are used to predict and calculate the ratio of the enzyme turnover number to the enzyme Michaelis-Menten constant as an initial value, and then the attention mechanism algorithm is used to train to establish the relationship between the enzyme-substrate representation and the ratio of the enzyme turnover number to the enzyme Michaelis-Menten constant, and the ratio of the enzyme turnover number to the enzyme Michaelis-Menten constant is predicted as a correction value. Finally, the ratio of the enzyme turnover number to the enzyme Michaelis-Menten constant is the sum of the initial value and the correction value in a certain ratio.

6. An enzyme kinetic parameter prediction device, characterized in that: The system comprises: A data set construction module is used to collect kinetic parameters of known enzymes and substrate information as a data set; the kinetic parameters include the turnover number of the enzyme, the Michaelis constant of the enzyme, and the ratio between the turnover number of the enzyme and the Michaelis constant of the enzyme; the enzyme representation is obtained using the ProtT5 language model; the substrate representation is obtained using the MACC fingerprint generated by the MolT5 language model and the RDKit software package; The enzyme turnover number and enzyme Michaelis-Menten constant prediction module is used to splice the obtained enzyme representation and substrate representation together to obtain the enzyme-substrate representation as a training set, and use the machine learning algorithm to train and establish the relationship between the enzyme-substrate representation and the enzyme turnover number and the enzyme Michaelis-Menten constant, obtain the enzyme turnover number prediction model and the enzyme Michaelis-Menten constant prediction model, and predict the enzyme turnover number and the enzyme Michaelis-Menten constant based on the prediction model; The ratio prediction module between the enzyme turnover number and the enzyme's Michaelis constant uses the enzyme turnover number prediction model and the enzyme's Michaelis constant prediction model to predict and calculate the ratio between the enzyme turnover number and the enzyme's Michaelis constant as an initial value, and then uses a machine learning algorithm to train and establish the relationship between the enzyme-substrate representation and the ratio between the enzyme turnover number and the enzyme's Michaelis constant, and predicts the ratio between the enzyme turnover number and the enzyme's Michaelis constant as a correction value. The final ratio between the enzyme turnover number and the enzyme's Michaelis constant is the sum of the initial value and the correction value in a certain ratio.

7. The enzyme kinetic parameter prediction device according to claim 6, characterized in that: The sources of the known enzyme kinetic parameters and substrate information include BEANDA and SABIO-RK databases; Preferably, the enzyme kinetic parameter prediction method further comprises the step of clustering the data set, specifically comprising: clustering according to the enzyme sequence similarity of 0.4, and dividing into 10 groups, wherein the enzyme sequence similarity between each group is lower than 0.4, and obtaining a 10-fold unbiased cross-validation data set; Preferably, the machine learning algorithm comprises a deep learning algorithm; Preferably, the deep learning algorithm includes a neural network algorithm or an attention mechanism algorithm.

8. An electronic device, characterized in that: The electronic device comprises a processor and a memory, and the processor is used to implement the enzyme kinetic parameter prediction method according to any one of claims 1 to 5 when executing the computer program stored in the memory.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the method for predicting enzyme kinetic parameters as described in any one of claims 1 to 5 is implemented.

10. A new enzyme mining method, characterized in that: The novel enzyme mining method comprises: The natural enzyme that catalyzes the reaction is used as the initial enzyme, and a multiple sequence alignment algorithm is used to obtain a new sequence similar to the initial enzyme sequence, and the new sequences are sorted from high to low according to the sequence similarity, filtered based on a certain sequence similarity threshold, or the sequences with the highest preset number of sequence similarity rankings are retained; Calculate the difference in sequence length between the remaining sequences and the initial enzyme, retain the sequences with length differences within a certain range, predict the kinetic parameters of the enzyme corresponding to the retained sequences using the enzyme kinetic parameter prediction method described in any one of claims 1 to 5, sort the sequences from high to low according to the enzyme kinetic parameters, retain the top preset number of sequences, use a protein structure prediction tool to predict the protein structure of the retained sequences, calculate the structural similarity with the initial enzyme, and screen new enzymes based on structural similarity.

Images