Construction method, medium and device of general prediction model for classifying mutant enzyme-substrate

Abstract

The invention discloses a method, medium and device for constructing a general prediction model for classifying mutant enzyme-substrates, and belongs to the technical field of bioinformatics, protein engineering and machine learning, the method comprises the following steps: step 1, expanding data to construct a data set, specifically comprising positive data screening and negative data generation, 2, building a multi-task coupling framework; and 3, optimizing the performance of the prediction model. The invention further provides a medium and a device carrying the model, and the accuracy of interaction prediction is improved through joint optimization of protein and small molecule characterization. Mutation sites capable of remarkably improving the catalytic efficiency are screened out, the activity of the enzyme is optimized, the catalytic ability of the mutant enzyme to different substrates is predicted, and the design of the enzyme with specific substrate selectivity is facilitated.

Description

Technical Field

[0001] The present invention relates to the fields of bioinformatics, protein engineering and machine learning technology, and in particular to a method, medium and device for constructing a universal prediction model for classifying mutant enzymes and substrates. Background Art

[0002] In the field of protein engineering, enzymes are widely used as biocatalysts in scientific research and industrial production. The high efficiency and specificity of enzymes make them of great value in drug development, biofuel production, environmental remediation and other fields. However, traditional enzyme engineering methods mainly rely on in-depth research on a single enzyme family, optimizing the catalytic efficiency of the enzyme or changing the substrate specificity through site-directed mutagenesis or gene editing. This method has limitations when facing complex biological reactions, especially in understanding the three-dimensional structural variations of different enzyme families and their relationship with enzyme activity.

[0003] Currently, there are many methods for predicting enzyme-substrate reactions. For example, SCANEER relies on sequence co-evolution to analyze the effect of single-point mutations on enzyme activity, but its algorithm can only evaluate about 47% of variants and has a limited scope of application. ECNet and eUniRep are prediction methods based on deep learning. Although they can provide relatively accurate predictions, they require experimental data of the target protein as a training set, which limits their application in the absence of activity measurement data. MutCompute learns amino acid evolution information in protein structure through 3D convolutional neural networks (CNNs), but it relies on known structural data. Methods such as UniKP, DLTKcat and CmpdEnzymPred use deep learning methods to predict enzyme activity, but are limited to small molecule substrate structures as input, which limits their effectiveness in wider applications.

[0004] Although these methods have made some progress in specific situations, they generally face the challenge of being unable to be effectively extended to new substrates or variants. This limitation is manifested in practical applications as the inability to efficiently predict the activity of new substrates, difficulty in dealing with complex biological reactions, and reliance on a large amount of experimental data, which increases R&D costs and time. Summary of the invention

[0005] This paper proposes a method for predicting mutant enzyme-substrate pairs based on a bidirectional data expansion strategy and a MEP protein characterization model. This method significantly improves the ability of machine learning models in simulating and predicting the diversity and complexity of enzyme activity by mining the association information of single-point and multi-point parallel mutations, and mining the interpretability of the effects of multidimensional mutations on enzyme activity.

[0006] The present invention is achieved through the following technical solutions: A method for constructing a universal prediction model for classifying mutant enzymes and substrates. Step 1: Build a data set through data expansion, including positive data screening and negative data generation; The positive data screening: screening the enzyme-substrate pair data and phylogenetic inference positive impact data from the existing database; The negative data generation described above: Based on the existing mutant enzyme database, a large number of mutant enzyme sequences are generated through single-point and multi-point mutations, and their activity changes (enhanced, weakened, unchanged, disappeared) are marked. According to the 1:1 pairing strategy of positive and negative samples, negative samples matching the positive data are selected; Step 2: Build a multi-task coupling framework In terms of protein sequence characterization, the MEP (Multi-scale Enzyme Protein) model is used to characterize the mutant enzyme sequence. The MEP model adopts the Transformer+CNN architecture: the ESM-2 Transformer is used to obtain the global dependency representation of the amino acid sequence, and then the convolutional neural network (CNN) is used to focus on the local mutation site features. Finally, the global and local information are fused through residual connections to obtain the multi-scale feature vector of the mutant enzyme. Two complementary characterization methods are used for substrate small molecules, supporting both Extended Connectivity Fingerprints (ECFP) and GNN; After completing the feature representation of the enzyme and substrate, the mutant enzyme representation vector (dimension dp) generated by MEP is concatenated with the substrate small molecule representation vector (dimension dm) generated by ECFP / GNN to form a joint feature vector v∈R dp+dm ; By integrating models including gradient boosting tree (GBDT), support vector machine (SVM), extreme gradient boosting (XGBoost), random forest (RF) and logistic regression (Logistic Regression), and adopting a dynamic weight fusion strategy, the final prediction type is constructed to classify the interaction between mutant enzymes and substrates. The aforementioned spliced ​​feature vectors are used as input to train the prediction models respectively; Step 3: Adjust hyperparameters (such as learning rate, tree depth, regularization parameters, etc.) through grid search, and use 5-fold cross validation to evaluate the model prediction performance to avoid overfitting.

[0007] Furthermore, the specific steps of generating negative data in step 1 are as follows: (1) Screen out mutant enzyme data entries with reduced or absent activity through in silico simulation; (2) Based on the co-evolutionary information of the mutation sites, the mutation sites in the negative samples are systematically replaced to ensure that the negative samples are comparable to the positive data in terms of mutation sites, that is, to avoid introducing model training bias due to significant differences in the selection of mutation sites between negative and positive samples (such as evolutionary conservation, functional importance, or structural relevance). Through systematic adjustments, this model can focus more on learning the impact of the mutation itself on the enzyme-substrate interaction, rather than being disturbed by irrelevant site differences, thereby improving the accuracy and generalization of predictions; (3) Finally, negative samples that match the number of positive data are generated, and a 1:1 pairing strategy of positive and negative samples is adopted to ensure the balance of the data set and avoid bias problems in model training.

[0008] Further, the step 2 specifically includes protein characterization and small molecule characterization; The protein characterization described above: the mutant enzyme sequence is characterized by using the MEP (Multi-scale Enzyme Protein) model; the MEP includes an evolutionary language model (ESM-2 Transformer) and a convolutional network, and the ESM-2 Transformer layer is fine-tuned on the basis of pre-training to extract the evolutionary conservative features and function-related clues of the mutant enzyme sequence; the multi-scale mutant enzyme feature extraction is achieved through the following technologies: (1) Global dependency modeling: The mutant enzyme sequence is encoded using the Transformer layer of ESM-2; the Transformer self-attention mechanism calculates the global dependency of amino acids in the sequence using the following formula: ; Among them, Q, K, and V are query, key, and value matrices respectively, and dk is the dimension of the key vector; this mechanism captures the remote evolutionary conservation and functional relevance of mutant enzyme sequences.

[0009] (2) Local feature focusing: A two-layer 3-core convolutional neural network (CNN) is used to extract features from the local region of the mutant enzyme sequence; the convolution operation is defined as: ; Among them, i represents the output position of the current convolution operation (i.e., a node or position on the feature map); k is the index of the convolution kernel, which is used to distinguish different convolution kernels, thereby realizing multi-core parallel feature extraction; j represents the neighbor position in the input sequence fragment covered by the convolution kernel (i.e., the position in the local sliding window relative to i); W is the 3-core convolution weight matrix, which is used to perform weighted summation on the local area of ​​the input fragment X; X is a local fragment of the input sequence (such as a window of an amino acid sequence); b is a bias term. Through the sliding window strategy, CNN strengthens the ability to characterize active sites and adjacent mutation regions;

[0010] (3) Feature fusion and robustness enhancement: Introducing residual connection to achieve layer-by-layer superposition of global and local features: H(x)=F(x)+x; Among them, H(x) is the final output of the residual block, F(x) is the transformation result of the current layer (such as the convolutional layer or the Transformer layer) on the input x, and x is the original input (the input of the skip connection); this design alleviates the gradient vanishing problem and improves the model's adaptability to local structural changes.

[0011] The small molecule characterization described above: two complementary characterization methods are used for substrate small molecules - fingerprint and GNN are supported at the same time to fully capture the chemical environment and topological structure information of small molecules, and then two layers of CNN with a window size of 3 are used to convolve the sequence, focusing on the local pattern of the mutation site and the neighborhood, and the two features are fused through residual connection to obtain a fixed-dimensional mutant enzyme representation; the specific method is as follows: (1) Extended Connectivity Cycle Fingerprint (ECFP): Generates a fixed-length molecular fingerprint, encodes the chemical environment of atoms and their neighborhoods in the molecule through a hash function, and captures functional groups and multi-level topological structure information; (2) Pre-trained Graph Neural Network (GNN): Encode the molecular graph structure based on the Graph Attention Network (GAT); the node feature update formula is: ; Among them, h' is the updated feature vector of node i, h j is the feature vector of neighbor node j, α ij is the attention weight of node i to neighbor j, indicating the importance of j to i, W is a learnable parameter matrix used to linearly transform node features, and σ is the activation function. Finally, a 100-dimensional task-specific vector is extracted to characterize the functional groups and spatial conformations of small molecules.

[0012] Furthermore, in step 3, the performance of the prediction model is optimized by the following steps: Hyperparameter grid search: Combined search of parameters including learning rate (0.01–0.2), tree depth (3–10), and subsample ratio (0.6–1.0) to screen the optimal parameter set; Five-fold cross validation: Divide the training set into five subsets, and iteratively evaluate the prediction model on the validation set, including accuracy, ROC-AUC, and F1 score, to ensure generalization ability; Test set validation: The final prediction model was validated by an independent test set, showing high accuracy in high confidence regions (prediction scores close to 0 or 1) to screen for efficient enzyme variants.

[0013] The present invention also provides a computer-readable storage medium for universal prediction of classified mutant enzyme-substrate pairs, wherein the computer-readable storage medium stores a computer program, and the computer program is suitable for being loaded by a processor and executing the universal prediction method for classified mutant enzyme-substrate pairs.

[0014] The present invention also provides a device for universal prediction of classified mutant enzyme-substrate pairs, wherein the device is equipped with a medium for running the universal prediction model of classified mutant enzyme-substrate pairs.

[0015] The beneficial effects of the present invention compared with the prior art are as follows: 1. The data expansion strategy in this step effectively solves the following problems by introducing negative samples and phylogenetic inference data: (1) Insufficient samples: Through silicon simulation and phylogenetic inference, the data scale is expanded to provide richer training samples; (2) Category imbalance: A 1:1 pairing strategy of positive and negative samples is adopted to ensure the balance of the data set and avoid bias problems in model training; (3) Data diversity: Combining experimental data and inferred data enhances the diversity and relevance of the data set, providing more comprehensive information for the model.

[0016] 2. The multi-task coupling framework described in the present invention is designed to capture information about long-range dependencies (such as structural / evolutionary conserved sites) and local changes (neighborhood of mutation sites) at the same time. The MEP model is based on this multi-scale design, reflecting the consideration of both long-range and short-range sequence features: Transformer provides a comprehensive understanding of sequence context, while CNN focuses on the neighborhood pattern of mutation sites. Combining the two can learn long- and short-range dependencies, and has shown competitiveness in fields such as mutation effect prediction. The present invention applies this hybrid architecture to the context of enzyme-substrate prediction, paying special attention to mutation sites, which may better distinguish the functional changes brought about by key mutations compared to models that only use Transformer.

[0017] 3. The constructed model of the present invention has a high degree of accuracy in the high confidence region (the region where the prediction score is close to 0 and 1), can effectively screen out potential high-efficiency enzyme variants, reduce the cost and time of experimental verification, and provide a powerful tool for enzyme design and optimization. Multi-model comparison and selection and dynamic weight fusion are used to improve performance and interpretability. This integrated optimization idea can be distinguished from a single deep learning framework to a certain extent, so that the model maintains both high accuracy and result interpretation capabilities.

[0018] 4. The present invention integrates the global modeling of Transformer and the local feature extraction of CNN to enhance the ability to capture mutation sites and functional changes, and combines experimental verification with phylogenetic inference data to significantly expand the size and diversity of the training set. In addition, the accuracy of interaction prediction is improved by jointly optimizing protein and small molecule characterization. This model helps to screen out mutation sites that can significantly improve catalytic efficiency and optimize enzyme activity by predicting the interaction between mutant enzymes and substrates, proving that the distal synergistic sites of mutant enzymes play an important role in enzymatic reactions in industrial production; it can also predict the catalytic ability of mutant enzymes for different substrates, helping to design enzymes with specific substrate selectivity. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] Figure 1 This is the prediction result diagram. DETAILED DESCRIPTION

[0020] The technical solution of the present invention is further explained below by way of embodiments, but the protection scope of the present invention is not limited in any form by the embodiments.

[0021] A general prediction method for classifying mutant enzyme-substrate pairs. Step 1: Build a data set through data expansion, including positive data screening and negative data generation; 1.1 Forward data screening: The integrated enzyme-substrate pair dataset was used for correlation screening, including 18,351 experimentally verified data and 274,030 phylogenetically inferred data.

[0022] Screen enzyme-substrate pair data and phylogenetic inference positive impact data from existing databases; as a specific implementation method, accurately screen entries matching the Alex data set from the D3 database, and obtain a total of 18,351 experimentally verified data and 274,030 phylogenetic inference data. Use the integrated enzyme-substrate pair data set for correlation screening to obtain 103 experimentally verified enzyme-substrate pair data and 250 phylogenetic inference positive impact data, and screen out instances of reduced or disappeared activity as positive samples through computational simulation. These data have been rigorously experimentally verified and phylogenetically analyzed to ensure the reliability and representativeness of the data.

[0023] 1.2 Negative data generation: Based on the D3 mutant enzyme database, more than 70,000 data were obtained, and a large number of mutant enzyme sequences were generated through single-point and multi-point mutations. Their activity changes (enhanced, weakened, unchanged, and disappeared) were marked. The 70,000 data were subdivided into four categories: activity enhancement, activity weakening, activity unchanged, and activity disappearance. According to the 1:1 pairing strategy of positive and negative samples, negative samples that matched the positive data were selected. The specific steps are as follows: 1.2.1 Screen out mutant enzyme data entries with reduced or absent activity through in silico simulation; 1.2.2 Based on the co-evolution information of the mutation sites, the mutation sites in the negative samples are systematically replaced to ensure that the negative samples and positive data are comparable in terms of mutation sites; Finally, negative samples matching the number of positive data are generated, and a 1:1 pairing strategy of positive and negative samples is adopted to ensure the balance of the data set and avoid bias problems in model training.

[0024] 1.3 Dataset Construction Through the above steps, we finally constructed a dataset containing 206 experimentally verified enzyme-substrate pairs and 500 phylogenetically inferred enzyme-substrate pairs, achieving two-way expansion of the data. This two-way expansion refers to the simultaneous increase of positive and negative examples: both using phylogenetic inference to expand positive examples to cover a wider range of enzyme-substrate pairs, and introducing simulated negative examples to expand non-interacting pairs, enriching the data from two directions. By combining experimental data and inferred data, the diversity and relevance of the dataset are enhanced, providing more comprehensive information for the model. The specific dataset structure is as follows: Class 1 dataset: contains 206 experimentally verified enzyme-substrate pairs, used for training and verification of high-precision models; Type 2 dataset: contains 500 phylogenetically inferred enzyme-substrate pairs, which is used to enhance the generalization ability of the model.

[0025] Step 2: Build a multi-task coupling framework 2.1 In terms of protein sequence representation, the Transformer+CNN architecture is used: the ESM-2 Transformer is used to obtain the global dependency representation of the amino acid sequence, and then the convolutional neural network (CNN) is used to focus on the local mutation site features. Finally, the global and local information are fused through residual connections to obtain the multi-scale feature vector of the mutant enzyme. This architecture is designed to capture information about long-range dependencies (such as structural / evolutionary conserved sites) and local changes (neighborhood of mutation sites) at the same time. The MEP model is based on this multi-scale design, reflecting the consideration of both long-range and short-range sequence features: Transformer provides a comprehensive understanding of sequence context, while CNN focuses on the neighborhood pattern of mutation sites. Combining the two can learn long- and short-range dependencies, and has shown competitiveness in areas such as mutation effect prediction. This hybrid architecture is applied to enzyme-substrate prediction, with special attention to mutation sites, so that it may better distinguish the functional changes caused by key mutations compared to models that only use Transformer. Specifically, the following steps are included: The MEP (Multi-scale Enzyme Protein) model is used to characterize mutant enzyme sequences. MEP essentially integrates an evolutionary language model (ESM-2 Transformer) and a convolutional network to output a protein embedding vector that covers global and local information. Specifically, the ESM-2 Transformer layer is fine-tuned based on pre-training to extract evolutionary conservation features and function-related clues of the sequence. The innovation lies in the embedding of multi-scale proteins, that is, compared with simple Transformer embedding, the additional CNN layer enables the mutation site to be prominently represented, which is different from most existing enzyme-substrate models. This method achieves multi-scale feature extraction through the following techniques: 2.1.1 Global Dependency Modeling: The mutant enzyme sequence is encoded using the Transformer layer of ESM-2. The Transformer self-attention mechanism calculates the global dependency of amino acids in the sequence using the following formula: ; Among them, Q, K, and V are query, key, and value matrices respectively, and dk is the dimension of the key vector. This mechanism captures the remote evolutionary conservation and functional relevance of sequences.

[0026] Local feature focus: A two-layer 3-core convolutional neural network (CNN) is used to extract features from the local area of ​​the sequence. The convolution operation is defined as: ; Among them, i represents the output position of the current convolution operation (i.e., a node or position on the feature map); k is the index of the convolution kernel, which is used to distinguish different convolution kernels, thereby realizing multi-core parallel feature extraction; j represents the neighbor position in the input sequence fragment covered by the convolution kernel (i.e., the position in the local sliding window relative to i); W is the 3-core convolution weight matrix, which is used to perform weighted summation on the local area of ​​the input fragment X; X is a local fragment of the input sequence (such as a window of an amino acid sequence); b is a bias term. Through the sliding window strategy, CNN strengthens the ability to characterize active sites and adjacent mutation regions; 2.1.2 Feature fusion and robustness enhancement: Introducing residual connection to achieve layer-by-layer superposition of global and local features: H(x)=F(x)+x; Among them, H(x) is the final output of the residual block, F(x) is the transformation result of the current layer (such as the convolutional layer or the Transformer layer) on the input x, and x is the original input (the input of the skip connection); this design alleviates the gradient vanishing problem and improves the model's adaptability to local structural changes.

[0027] 2.2 Two complementary characterization methods are used for substrate small molecules - supporting both fingerprints and GNN. This model can fully capture the chemical environment and topological structure information of small molecules. Then, a two-layer CNN with a window size of 3 is used to convolve the sequence, focusing on the local pattern of the mutation site and the neighborhood. The two features are fused through residual connection to obtain a fixed-dimensional mutant enzyme representation. This method ensures that the model can "see" the global sequence relationship without missing the local effects of key mutations. It is achieved in the following ways: 2.2.1 Extended Connected Cyclic Fingerprint (ECFP): Generates a fixed-length molecular fingerprint, encodes the chemical environment of atoms and their neighborhoods in the molecule through a hash function, and captures functional group and multi-level topological structure information.

[0028] 2.2.2 Pre-trained Graph Neural Network (GNN): Encode the molecular graph structure based on the Graph Attention Network (GAT). The node feature update formula is: ; Among them, h' is the updated feature vector of node i, h j is the feature vector of neighbor node j, α ijis the attention weight of node i to neighbor j, indicating the importance of j to i, W is a learnable parameter matrix used to linearly transform node features, and σ is the activation function. Finally, a 100-dimensional task-specific vector is extracted to characterize the functional groups and spatial conformations of small molecules.

[0029] 2.3 After completing the feature representation of the enzyme and substrate, the mutant enzyme representation vector (dimension dp) generated by MEP is concatenated with the small molecule representation vector (dimension dm) generated by ECFP / GNN to form a joint feature vector v∈R dp+dm By integrating models including gradient boosting tree (GBDT), support vector machine (SVM), extreme gradient boosting (XGBoost), random forest (RF) and logistic regression (Logistic Regression), and adopting a dynamic weight fusion strategy, the final prediction model is constructed to classify the interaction between mutant enzymes and substrates; the aforementioned spliced ​​feature vectors are used as input to train the prediction models separately, and the performance is optimized mainly through the following steps: 2.3.1 Hyperparameter grid search: Combined search of learning rate (0.01–0.2), tree depth (3–10), subsample ratio (0.6–1.0) and other parameters to screen the optimal parameter set; 2.3.2 Five-fold cross validation: Divide the training set into 5 subsets, iteratively evaluate the accuracy, ROC-AUC, F1 score and other indicators of the prediction model on the validation set to ensure generalization ability; 2.3.3 Test set validation: The final prediction model was validated by an independent test set, showing high accuracy in the high confidence region (prediction scores close to 0 or 1), and screening for efficient enzyme variants.

[0030] Step 3: Adjust hyperparameters (such as learning rate, tree depth, regularization parameters, etc.) through grid search, and use 5-fold cross validation to evaluate the performance of the prediction model to avoid overfitting. The model has high accuracy in high confidence areas (areas where the prediction scores are close to 0 and 1), and can effectively screen out potential efficient enzyme variants, reducing the cost and time of experimental verification, and providing a powerful tool for enzyme design and optimization. Use multi-model comparison and selection and dynamic weight fusion to improve performance and interpretability. This integrated optimization idea can be distinguished from a single deep learning framework to a certain extent, so that the model maintains both high accuracy and the ability to interpret results.

[0031] According to the experimental test results (as shown in Table 1-2), this technical solution significantly improves the prediction accuracy and generalization ability by adjusting the hyperparameters through grid search and optimizing the model performance in combination with 5-fold cross-validation: In cross-validation, GBDT showed excellent parameter tuning efficiency with an accuracy of 0.9512 and an F1 score of 0.9394, SVM verified its comprehensive advantages in nonlinear classification tasks with an ROC-AUC of 0.9845, and XGBoost achieved a 23% increase in F1 score (0.7872→0.8611) through optimization, highlighting the efficiency of the gradient boosting framework; on the test set, LogisticRegression showed extremely strong reliability in the high confidence area (prediction scores close to 0 or 1) with an accuracy of 0.88 and an ROC-AUC of 0.9541, and its true prediction accounted for more than 90% (such as Figure 1 As shown in the figure, the error rate is less than 5%, which greatly reduces the risk of false positives / false negatives in experimental verification and improves the efficiency of enzyme variant screening by more than 60%. Through multi-model dynamic weight fusion (integrating GBDT's high F1, SVM's high AUC, etc.), the model maintains interpretability while outperforming a single deep learning framework in comprehensive performance (e.g., all models ROC-AUC>0.93), fully demonstrating the technical advantages of this scheme in enzyme design and optimization, which has high precision, strong robustness, and practicality.

[0032] Table 1 is the cross-validation model evaluation index table ; Table 2 shows the test set model evaluation results. .

Claims

1. A method for constructing a universal prediction model for classifying mutant enzymes-substrates, characterized in that: The steps of the method are as follows: Step 1: Build a dataset through data expansion, including positive data screening and negative data generation The positive data screening: screening the enzyme-substrate pair data and phylogenetic inference positive impact data from the existing database; The negative data generation: Based on the existing mutant enzyme database, a large number of mutant enzyme sequences are generated through single-point and multi-point mutations, and their activity changes are marked. According to the 1:1 pairing strategy of positive and negative samples, negative samples matching the positive data are selected; Step 2: Build a multi-task coupling framework In terms of protein sequence characterization, the MEP model is used to characterize the mutant enzyme sequence. The MEP model adopts the Transformer+CNN architecture: the ESM-2 Transformer is used to obtain the global dependency representation of the amino acid sequence, and then the convolutional neural network is used to focus on the local mutation site characteristics. Finally, the residual connection is used to fuse the global and local information to obtain the multi-scale feature vector of the mutant enzyme. Two complementary characterization methods are used for substrate small molecules, supporting both extended connection cycle fingerprints and pre-trained graph neural networks; After completing the feature representation of the enzyme and substrate, the mutant enzyme representation vector generated by MEP is concatenated with the substrate small molecule representation vector generated by ECFP / GNN to form a joint feature vector v∈R dp+dm ; By integrating models including gradient boosting trees, support vector machines, extreme gradient boosting, random forests and logistic regression, and adopting a dynamic weight fusion strategy, the final prediction type is constructed to classify the interaction between mutant enzymes and substrates, and the aforementioned spliced ​​feature vectors are used as input to train the prediction models separately; Step 3: Adjust hyperparameters through grid search and evaluate model prediction performance using 5-fold cross validation to avoid overfitting.

2. The method for constructing a universal prediction model for classifying mutant enzymes and substrates according to claim 1, characterized in that: The specific steps of generating negative data in step 1 are as follows: (1) Screen out mutant enzyme data entries with reduced or absent activity through in silico simulation; (2) Based on the co-evolutionary information of the mutation sites, the mutation sites in the negative samples are systematically replaced to ensure that the negative samples and positive data are comparable in terms of mutation sites; (3) Finally, negative samples that match the number of positive data are generated, using a 1:1 pairing strategy for positive and negative samples.

3. The method for constructing a universal prediction model for classifying mutant enzymes and substrates according to claim 1, characterized in that: The step 2 specifically includes protein characterization and small molecule characterization; The protein characterization described above: the mutant enzyme sequence is characterized by using the MEP model; the MEP includes an evolutionary language model and a convolutional network, and the evolutionary language model is the ESM-2 Transformer, which realizes multi-scale mutant enzyme feature extraction by the following method: (1) Global dependency modeling: The mutant enzyme sequence is encoded using the Transformer layer of ESM-2; the Transformer self-attention mechanism calculates the global dependency of amino acids in the sequence using the following formula: ； Among them, Q, K, V are the matrices of query, key, and value respectively, and dk is the dimension of the key vector; (2) Local feature focusing: A two-layer three-core convolutional neural network is used to extract features from the local area of ​​the mutant enzyme sequence; the convolution operation is defined as: ； Where i represents the output position of the current convolution operation; k is the index of the convolution kernel, j represents the neighbor position in the input sequence segment covered by the convolution kernel; W is the 3-kernel convolution weight matrix, which is used to perform weighted summation on the local area of ​​the input segment X; X is the local segment of the input sequence; b is the bias term; (3) Feature fusion and robustness enhancement: Introducing residual connections to achieve layer-by-layer superposition of global and local features: H(x)=F(x)+x Among them, H(x) is the final output of the residual block, F(x) is the output of the current layer, and x is the input; this design alleviates the gradient vanishing problem and improves the model's adaptability to local structural changes; The small molecule characterization described above: two complementary characterization methods are used for substrate small molecules - fingerprint and GNN are supported at the same time to fully capture the chemical environment and topological structure information of small molecules, and then two layers of CNN with a window size of 3 are used to convolve the sequence, focusing on the local pattern of the mutation site and the neighborhood, and the two features are fused through residual connection to obtain a fixed-dimensional mutant enzyme representation; the specific method is as follows: (1) Extended connection loop fingerprint: Generate a fixed-length molecular fingerprint, encode the chemical environment of atoms and their neighborhoods in the molecule through a hash function, and capture functional groups and multi-level topological structure information; (2) Pre-trained graph neural network: Encode the molecular graph structure based on the graph attention mechanism; the node feature update formula is: ； Among them, h' is the updated feature vector of node i, h j is the feature vector of neighbor node j, α ij is the attention weight of node i to neighbor j, indicating the importance of j to i, W is a learnable parameter matrix used to perform linear transformation on node features, and σ is the activation function.

4. The method for constructing a universal prediction model for classifying mutant enzymes and substrates according to claim 1, characterized in that: In step 3, the performance of the prediction model is optimized by the following steps: Hyperparameter grid search: Combined search of parameters including learning rate, tree depth, and subsample ratio to select the optimal parameter set; Five-fold cross validation: Divide the training set into five subsets, and iteratively evaluate the prediction model on the validation set, including accuracy, ROC-AUC, and F1 score, to ensure generalization ability; Test set validation: The final prediction model was validated by an independent test set to screen for efficient enzyme variants.

5. A computer-readable storage medium for general prediction of classified mutant enzyme-substrate pairs, characterized in that: The computer-readable storage medium stores a computer program, and the computer program is suitable for being loaded by a processor and executing the method for constructing a universal prediction model for classifying mutant enzymes and substrates according to any one of claims 1 to 4.

6. A device for general prediction of classified mutant enzyme-substrate pairs, characterized in that: The device is equipped with the computer-readable storage medium according to claim 5.

Images