Enzyme kinetic parameter and domain tag prediction method and device, equipment and medium

Abstract

The invention discloses an enzyme kinetic parameter and domain label prediction method and device, equipment and a medium, and relates to the technical field of enzymes, and the method comprises the steps: training an initial enzyme kinetic parameter prediction model comprising a feature extraction sub-model and a domain adversarial neural network through employing an intra-domain data set and an extra-domain test set, obtaining a target enzyme kinetic parameter prediction model; inputting the target reaction substrate and the target enzyme sequence into a target enzyme kinetic parameter prediction model so that the target enzyme kinetic parameter prediction model performs enzyme kinetic parameter prediction to map to a corresponding enzyme kinetic parameter prediction value, and performing domain tag mapping to obtain a corresponding domain prediction tag, outputting an enzyme kinetic parameter prediction value and a domain prediction tag corresponding to the target reaction substrate and the target enzyme sequence. Through a special training process, the characteristics of a substrate and an enzyme sequence can be extracted more accurately, so that the prediction accuracy of enzyme kinetic parameters and domain tags is improved.

Description

Technical Field

[0001] This invention relates to the field of enzyme technology, and in particular to a method, apparatus, device, and medium for predicting enzyme kinetic parameters and domain labels. Background Technology

[0002] Enzyme kinetics is an important research area in biochemistry, involving key parameters such as enzyme activity and reaction rates in living organisms. Predicting enzyme kinetic parameters is crucial for understanding chemical reaction processes, drug metabolism, and biocatalysis in vivo, and is of great significance for drug development, bioengineering, and medical diagnostics. However, the prediction of enzyme kinetic parameters faces many challenges, including insufficient data, incomplete characterization, and insufficient model generalization ability.

[0003] Data scarcity is one of the primary challenges in predicting enzyme kinetic parameters. The data required for this prediction is typically acquired under laboratory conditions, and the amount of data is often limited. Because experiments such as enzyme activity assays consume significant time and resources, the acquired data is often insufficient to cover all possible scenarios, leading to inadequate training of the prediction model. Furthermore, data obtained under different experimental conditions may exhibit variations, which also challenges the model's generalization ability, resulting in inconsistent prediction performance on new datasets.

[0004] Another challenge is the complexity of biological systems. Predicting enzyme kinetic parameters involves modeling complex biological systems, which often exhibit variable properties and intricate interactions. Traditional feature representation methods often struggle to capture the underlying structural features in the data, preventing models from fully utilizing the information within the data. Furthermore, traditional machine learning models often rely heavily on manual feature engineering, which requires expertise and experience and is susceptible to subjective influences, limiting the model's expressive and generalization capabilities.

[0005] To address the above challenges, existing methods for predicting enzyme kinetic parameters have proposed several solutions. One approach utilizes the epigenetic catalytic rate obtained from proteomics and metabolomics analysis. This method obtains enzyme kinetic parameter values ​​by dividing the measured reaction flux by a quantitative protein abundance. Although this method has proven successful in various microbial resource allocation models, the high cost of metabolomics and proteomics measurements makes it difficult to implement. Another approach utilizes deep learning models of compound-protein interactions. These models have been developed to predict biological parameters such as binding affinity, Michaelis-Menten constant, and enzyme conversion rate. The input to this method is typically a simplified molecular input line entry system (SMILES) string of the compound and a protein sequence. Then, deep learning models such as graph neural networks, recurrent neural networks, or convolutional neural networks are used to extract features of the compound and protein, and these features are concatenated for regression of the target value, i.e., predicting the enzyme kinetic parameter value. This method achieves the prediction of enzyme kinetic parameters by learning the interaction patterns between compounds and proteins.

[0006] While existing methods can predict enzyme kinetic parameters to some extent, they still have limitations and challenges. For example, methods using epigenetic catalytic rates derived from proteomics and metabolomics analyses are costly, limiting their practical application. Deep learning models may be limited in practice by data quality and model generalization ability. The application scenarios for enzyme kinetic parameter prediction are becoming increasingly diverse and complex, requiring models to achieve stable and accurate prediction results under different data domains and experimental conditions. Traditional model training methods often fail to fully consider data distribution differences and domain bias, resulting in insufficient generalization ability of models in new data domains and failing to achieve the expected results. Summary of the Invention

[0007] In view of this, the purpose of this invention is to provide a method, apparatus, device, and medium for predicting enzyme kinetic parameters and domain labels, capable of accurately predicting enzyme kinetic parameters in different data domains. The specific solution is as follows:

[0008] In a first aspect, this application discloses a method for predicting enzyme kinetic parameters and domain tags, including:

[0009] The initial enzyme kinetic parameter prediction model, which includes a feature extraction sub-model and a domain adversarial neural network, is trained using an in-domain dataset and an out-of-domain test set to obtain the target enzyme kinetic parameter prediction model.

[0010] The target reaction substrate and the target enzyme sequence are input into the target enzyme kinetic parameter prediction model, so that the target feature extraction sub-model of the target enzyme kinetic parameter prediction model can extract the substrate feature information of the target reaction substrate and the enzyme sequence feature information of the target enzyme sequence, respectively. The target feature information, which is the result of splicing the substrate feature information and the enzyme sequence feature information, is input into the target domain adversarial neural network of the target enzyme kinetic parameter prediction model.

[0011] The target feature information is transformed by the target domain adversarial neural network to obtain a target feature vector of a preset dimension. The target feature vector is then used to predict enzyme kinetic parameters to map to the corresponding predicted enzyme kinetic parameter values. The target feature vector is also used to map domain labels to obtain corresponding domain prediction labels. Finally, the predicted enzyme kinetic parameters and domain prediction labels corresponding to the target reaction substrate and the target enzyme sequence are output.

[0012] Optionally, training the initial enzyme kinetic parameter prediction model, which includes a feature extraction sub-model and a domain adversarial neural network, using an in-domain dataset and an out-of-domain test set to obtain the target enzyme kinetic parameter prediction model includes:

[0013] The enzyme kinetic parameter samples are split into an in-domain dataset and an out-of-domain test set using hierarchical clustering; wherein the enzyme kinetic parameter samples include substrate information, enzyme sequence and corresponding enzyme kinetic parameter values;

[0014] The feature extraction sub-model of the initial enzyme kinetic parameter prediction model is trained and validated using the substrate information and enzyme sequence in the in-domain dataset and the out-of-domain test set. Then, the pre-trained feature extraction sub-model containing the target pre-training weights is fine-tuned using low-rank adaptation technology to obtain the molecular language model and the protein language model, respectively.

[0015] The domain adversarial neural network of the initial enzyme kinetic parameter prediction model is trained using the in-domain dataset and the out-of-domain test set to obtain a target domain adversarial neural network that includes a target feature converter, a target enzyme kinetic parameter predictor, and a target domain discriminator.

[0016] Based on the protein language model, the molecular language model, and the target domain adversarial neural network, a target enzyme kinetic parameter prediction model is obtained.

[0017] Optionally, the step of splitting the enzyme kinetic parameter samples into an in-domain dataset and an out-of-domain test set through hierarchical clustering includes:

[0018] The substrate information and enzyme sequence in the enzyme kinetic parameter sample are each divided into a first preset number of clusters to obtain a first preset number of substrate information clusters and a first preset number of enzyme sequence clusters;

[0019] A second preset number of substrate information clusters and enzyme sequence clusters are randomly selected as target substrate information clusters and target enzyme sequence clusters. The paired data of substrate information and enzyme sequence in the target substrate information clusters and target enzyme sequence clusters are used as an in-domain dataset. The substrate data in the target substrate information clusters and the enzyme sequence data in the target enzyme sequence clusters, excluding the in-domain datasets, are paired as an out-of-domain test set. The second preset number is less than the first preset number.

[0020] The domain dataset is split according to a preset splitting ratio to obtain a training set, a domain validation set, and a domain test set.

[0021] Accordingly, the feature extraction sub-model of the initial enzyme kinetic parameter prediction model is trained and validated using substrate information from the in-domain dataset and the out-of-domain test set, and then fine-tuned using low-rank adaptation techniques to obtain the molecular language model and the protein language model, respectively.

[0022] The feature extraction sub-model of the initial enzyme kinetic parameter prediction model is trained and validated using substrate information and enzyme sequences from the in-domain dataset and the out-of-domain test set. Then, the pre-trained feature extraction sub-model containing the target pre-training weights is fine-tuned using low-rank adaptation technology to obtain a molecular language model containing a first low-rank adaptation matrix and a first pre-training weight matrix, and a protein language model containing a second low-rank adaptation matrix and a second pre-training weight matrix, respectively.

[0023] Optionally, the step of using the substrate information and enzyme sequence from the in-domain dataset and the out-of-domain test set to train and validate the feature extraction sub-model of the initial enzyme kinetic parameter prediction model, and then using low-rank adaptation technology to fine-tune the pre-trained feature extraction sub-model containing the target pre-training weights, to obtain a molecular language model containing a first low-rank adaptation matrix and a first pre-training weight matrix and a protein language model containing a second low-rank adaptation matrix and a second pre-training weight matrix, respectively, includes:

[0024] The molecular language extraction sub-model of the initial enzyme kinetic parameter prediction model is trained using the molecular structure data and chemical properties of the substrate information in the training set of the molecular language model. This allows the molecular language extraction sub-model to learn the relationship between the molecular structure data and chemical properties of the substrate information, thus obtaining a pre-trained molecular language model.

[0025] The protein language extraction sub-model of the initial enzyme kinetic parameter prediction model is trained using the amino acid sequence and structural features of the enzyme sequence in the training set of the protein language model, so that the protein language extraction sub-model learns the relationship between the amino acid sequence and the structural features of the enzyme sequence to obtain a pre-trained protein language model.

[0026] The initial enzyme kinetic parameter prediction model, which includes the pre-trained molecular language model and the pre-trained protein language model, is validated and evaluated using the intra-domain validation set and the intra-domain test set to obtain a first validation evaluation result.

[0027] The initial enzyme kinetic parameter prediction model, which includes the pre-trained molecular language model and the pre-trained protein language model, is validated and evaluated using the out-of-domain test set to obtain a second validation evaluation result.

[0028] When the difference between the verification accuracy of the second verification evaluation result and the verification accuracy of the first verification evaluation result is less than a preset threshold, the pre-trained molecular language model and the pre-trained protein language model are used as pre-trained feature extraction sub-models.

[0029] The target pre-training weights of the pre-trained molecular language model and the pre-trained protein language model are decomposed into the product of two low-rank adaptation matrices to obtain a molecular language model containing a first low-rank adaptation matrix, a second low-rank adaptation matrix, and a first pre-training weight matrix, and a protein language model containing a third low-rank adaptation matrix, a fourth low-rank adaptation matrix, and a second pre-training weight matrix.

[0030] Optionally, training the initial enzyme kinetic parameter prediction model, which includes a feature extraction sub-model and a domain adversarial neural network, using an in-domain dataset and an out-of-domain test set to obtain the target enzyme kinetic parameter prediction model includes:

[0031] The intradomain substrate feature information, extradomain substrate feature information, intradomain enzyme sequence feature information, and extradomain enzyme sequence feature information output by the pre-trained protein language model and the pre-trained molecular language model are obtained. The intradomain substrate feature information and the intradomain enzyme sequence feature information are concatenated to obtain intradomain data feature information. The extradomain substrate feature information and the extradomain enzyme sequence feature information are concatenated to obtain extradomain data feature information.

[0032] Construct a domain adversarial neural network that includes a feature converter, an enzyme kinetic parameter predictor, and a domain discriminator;

[0033] The data feature information within the domain and the data feature information outside the domain are input into the domain adversarial neural network so that the feature converter of the domain adversarial neural network performs multi-layer feature mapping processing on the data feature information within the domain and the data feature information outside the domain, thereby updating the converter parameter information of the feature converter and obtaining the target feature converter.

[0034] The feature vector output by the target feature converter is passed to the enzyme kinetic parameter predictor so that the enzyme kinetic parameter predictor learns the relationship between the feature vector and the corresponding true value of the enzyme kinetic parameter, outputs the enzyme kinetic parameter prediction value of the feature vector, and updates the parameters of the enzyme kinetic parameter predictor according to the information difference between the predicted value of the enzyme kinetic parameter and the true value of the enzyme kinetic parameter to obtain the target enzyme kinetic parameter predictor.

[0035] The feature vector output by the target feature converter is passed to the domain discriminator so that the domain discriminator can learn the relationship between the feature vector and the corresponding real domain label, output the predicted domain label of the feature vector, and update the parameters of the domain discriminator with the information difference between the predicted domain label and the real domain label to obtain the target domain discriminator.

[0036] A target domain adversarial neural network is constructed based on the target feature converter, the target enzyme kinetic parameter predictor, and the target domain discriminator.

[0037] Optionally, the enzyme kinetic parameters and domain tag prediction method further includes:

[0038] A first loss function is constructed based on the difference between the predicted value of the enzyme kinetic parameter output by the enzyme kinetic parameter predictor during the training process of the domain adversarial neural network and the true value of the enzyme kinetic parameter; a second loss function is constructed based on the difference between the predicted domain label output by the domain discriminator during the training process of the domain adversarial neural network and the true domain label.

[0039] Construct a target loss function with the first objective of minimizing the first loss function and the second objective of maximizing the second loss function;

[0040] The domain adversarial neural network is iteratively trained until a preset training termination condition is met. The current domain adversarial neural network is then output to obtain a target domain adversarial neural network that includes the target feature converter, the target enzyme kinetic parameter predictor, and the target domain discriminator.

[0041] Optionally, the step of performing feature transformation on the target feature information through the target domain adversarial neural network to obtain a target feature vector of a preset dimension, predicting enzyme kinetic parameters on the target feature vector to map it to the corresponding predicted enzyme kinetic parameter values, and mapping the target feature vector to domain labels to obtain corresponding predicted domain labels, and outputting the predicted enzyme kinetic parameters and domain predicted labels corresponding to the target reaction substrate and the target enzyme sequence, includes:

[0042] The target feature information is transformed by the target feature converter of the target domain adversarial neural network to obtain a target feature vector of a preset dimension.

[0043] The target feature vector is used to predict enzyme kinetic parameters by the target enzyme kinetic parameter predictor of the target domain adversarial neural network, so as to map it to the corresponding enzyme kinetic parameter prediction value;

[0044] The target feature vector is mapped to a domain label by the target domain discriminator of the target domain adversarial neural network to obtain the corresponding domain predicted label.

[0045] Output the predicted values ​​of the enzyme kinetic parameters and the domain prediction labels corresponding to the target reaction substrate and the target enzyme sequence.

[0046] Secondly, this application discloses an enzyme kinetic parameter and domain tag prediction device, comprising:

[0047] The model training module is used to train the initial enzyme kinetic parameter prediction model, which includes a feature extraction sub-model and a domain adversarial neural network, using the in-domain dataset and the out-of-domain test set, so as to obtain the target enzyme kinetic parameter prediction model.

[0048] The information extraction module is used to input the target reaction substrate and the target enzyme sequence into the target enzyme kinetic parameter prediction model, so that the target feature extraction sub-model of the target enzyme kinetic parameter prediction model can extract the substrate feature information of the target reaction substrate and the enzyme sequence feature information of the target enzyme sequence, respectively, and input the target feature information after splicing the substrate feature information and the enzyme sequence feature information into the target domain adversarial neural network of the target enzyme kinetic parameter prediction model;

[0049] The parameter prediction module is used to perform feature transformation on the target feature information through the target domain adversarial neural network to obtain a target feature vector of a preset dimension, perform enzyme kinetic parameter prediction on the target feature vector to map it to the corresponding enzyme kinetic parameter prediction value, and perform domain label mapping on the target feature vector to obtain the corresponding domain prediction label, and output the enzyme kinetic parameter prediction value and domain prediction label corresponding to the target reaction substrate and the target enzyme sequence.

[0050] Thirdly, this application discloses an electronic device, including:

[0051] Memory, used to store computer programs;

[0052] A processor is configured to execute the computer program to implement the steps of the aforementioned disclosed enzyme kinetic parameters and domain tag prediction method.

[0053] Fourthly, this application discloses a computer-readable storage medium for storing a computer program; wherein, when the computer program is executed by a processor, it implements the steps of the aforementioned disclosed enzyme kinetic parameters and domain label prediction method.

[0054] As can be seen, this application discloses training an initial enzyme kinetic parameter prediction model, including a feature extraction sub-model and a domain adversarial neural network, using an in-domain dataset and an out-of-domain test set to obtain a target enzyme kinetic parameter prediction model; inputting the target reaction substrate and the target enzyme sequence into the target enzyme kinetic parameter prediction model, so that the target feature extraction sub-model of the target enzyme kinetic parameter prediction model extracts the substrate feature information of the target reaction substrate and the enzyme sequence feature information of the target enzyme sequence, respectively, and inputting the target feature information, which is the concatenation of the substrate feature information and the enzyme sequence feature information, into the target domain adversarial neural network of the target enzyme kinetic parameter prediction model; performing feature transformation on the target feature information through the target domain adversarial neural network to obtain a target feature vector of a preset dimension, performing enzyme kinetic parameter prediction on the target feature vector to map it to the corresponding enzyme kinetic parameter prediction value, and performing domain label mapping on the target feature vector to obtain the corresponding domain prediction label, and outputting the enzyme kinetic parameter prediction value and the domain prediction label corresponding to the target reaction substrate and the target enzyme sequence. Therefore, by training an initial enzyme kinetic parameter prediction model incorporating a domain adversarial neural network (DAN), a target enzyme kinetic parameter prediction model is obtained. The DAN can map and transform features across different data domains, thereby improving the adaptability and generalization ability of the target enzyme kinetic parameter prediction model across different data domains. It can simultaneously acquire substrate and enzyme sequence features, effectively splicing and processing them, comprehensively considering the influence of multiple factors on enzyme kinetic parameters. Through a specialized training process, it can more accurately extract features from the substrate and enzyme sequence, thus improving the accuracy of enzyme kinetic parameter prediction. Furthermore, the output predicted domain labels provide more comprehensive analytical results, helping to better understand the source and nature of the data, making the process of predicting enzyme kinetic parameters and domain labels more efficient, accurate, and stable. Attached Figure Description

[0055] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort.

[0056] Figure 1 This is a flowchart of an enzyme kinetic parameter and domain tag prediction method disclosed in this application;

[0057] Figure 2 This is a schematic diagram of the structure of a target enzyme kinetic parameter prediction model disclosed in this application;

[0058] Figure 3 This is a flowchart of a target domain adversarial neural network training method disclosed in this application;

[0059] Figure 4 This is a schematic diagram of an enzyme kinetic parameter and domain tag prediction device disclosed in this application;

[0060] Figure 5 This is a structural diagram of an electronic device disclosed in this application. Detailed Implementation

[0061] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0062] Enzyme kinetics is an important research area in biochemistry, involving key parameters such as enzyme activity and reaction rates in living organisms. Predicting enzyme kinetic parameters is crucial for understanding chemical reaction processes, drug metabolism, and biocatalysis in vivo, and is of great significance for drug development, bioengineering, and medical diagnostics. However, the prediction of enzyme kinetic parameters faces many challenges, including insufficient data, incomplete characterization, and insufficient model generalization ability.

[0063] Data scarcity is one of the primary challenges in predicting enzyme kinetic parameters. The data required for this prediction is typically acquired under laboratory conditions, and the amount of data is often limited. Because experiments such as enzyme activity assays consume significant time and resources, the acquired data is often insufficient to cover all possible scenarios, leading to inadequate training of the prediction model. Furthermore, data obtained under different experimental conditions may exhibit variations, which also challenges the model's generalization ability, resulting in inconsistent prediction performance on new datasets.

[0064] Another challenge is the complexity of biological systems. Predicting enzyme kinetic parameters involves modeling complex biological systems, which often exhibit variable properties and intricate interactions. Traditional feature representation methods often struggle to capture the underlying structural features in the data, preventing models from fully utilizing the information within the data. Furthermore, traditional machine learning models often rely heavily on manual feature engineering, which requires expertise and experience and is susceptible to subjective influences, limiting the model's expressive and generalization capabilities.

[0065] To address the above challenges, existing methods for predicting enzyme kinetic parameters have proposed several solutions. One approach utilizes the epigenetic catalytic rate obtained from proteomics and metabolomics analysis. This method obtains the enzyme kinetic parameter value by dividing the measured reaction flux by the quantitative protein abundance. Although this method has proven successful in various microbial resource allocation models, the high measurement costs of metabolomics and proteomics make it difficult to implement. Another approach utilizes deep learning models of compound-protein interactions. These models have been developed to predict biological parameters such as binding affinity, Michaelis-Menten constant, and enzyme conversion rate. The input to this method is typically a simplified molecular input line of the compound and the sequence of the protein. Then, deep learning models such as graph neural networks, recurrent neural networks, or convolutional neural networks are used to extract features of the compound and protein, and these features are concatenated for regression of the target value, i.e., predicting the enzyme kinetic parameter value. This method achieves the prediction of enzyme kinetic parameters by learning the interaction patterns between compounds and proteins.

[0066] While existing methods can predict enzyme kinetic parameters to some extent, they still have limitations and challenges. For example, methods using epigenetic catalytic rates derived from proteomics and metabolomics analyses are costly, limiting their practical application. Deep learning models may be limited in practice by data quality and model generalization ability. The application scenarios for enzyme kinetic parameter prediction are becoming increasingly diverse and complex, requiring models to achieve stable and accurate prediction results under different data domains and experimental conditions. Traditional model training methods often fail to fully consider data distribution differences and domain bias, resulting in insufficient generalization ability of models in new data domains and failing to achieve the expected results.

[0067] To this end, the present invention provides an enzyme kinetic parameter and domain label prediction scheme that can accurately predict enzyme kinetic parameters in different data domains.

[0068] Reference Figure 1 As shown in the figure, an embodiment of the present invention discloses a method for predicting enzyme kinetic parameters and domain labels, including:

[0069] Step S11: Train the initial enzyme kinetic parameter prediction model, which includes the feature extraction sub-model and the domain adversarial neural network, using the in-domain dataset and the out-of-domain test set to obtain the target enzyme kinetic parameter prediction model.

[0070] In this embodiment, the enzyme kinetic parameter samples are split into an in-domain dataset and an out-of-domain test set through hierarchical clustering. The enzyme kinetic parameter samples include substrate information, enzyme sequences, and corresponding enzyme kinetic parameter values. It is understood that, in order to train the target enzyme kinetic parameter prediction model, the in-domain dataset and the out-of-domain dataset are first obtained. The steps for obtaining the in-domain dataset and the out-of-domain dataset are as follows:

[0071] The substrate information and enzyme sequence in the enzyme kinetic parameter sample are each split into a first preset number of clusters to obtain a first preset number of substrate information clusters and a first preset number of enzyme sequence clusters. A second preset number of the substrate information clusters and enzyme sequence clusters are randomly selected as target substrate information clusters and target enzyme sequence clusters. The substrate information-enzyme sequence pairing data in the target substrate information clusters and target enzyme sequence clusters are used as the in-domain dataset. The substrate data in the target substrate information clusters and the enzyme sequence data in the target enzyme sequence clusters, excluding the in-domain datasets, are paired as the out-of-domain test set. The second preset number is less than the first preset number. The in-domain dataset is split according to a preset splitting ratio to obtain a training set, an in-domain validation set, and an in-domain test set.

[0072] Specifically, the enzyme kinetic parameter samples are experimental sample sets generated from existing biological enzyme kinetic parameter experiments. This sample set contains 16,838 enzyme kinetic parameter samples (including enzyme sequences, substrates, and enzyme kinetic parameter values), of which 7,822 are independent enzyme sequences and 2,672 are independent substrates. Hierarchical clustering was used to split the data into an in-domain dataset containing a training set, an in-domain validation set, an in-domain test set, and an out-of-domain test set. The specific steps for splitting the dataset using hierarchical clustering are as follows:

[0073] First, the substrate information and enzyme sequence in the enzyme kinetic parameter sample are split into ten clusters to obtain corresponding substrate information clusters and enzyme sequence clusters. Then, six target clusters are randomly selected from each of the substrate information cluster and enzyme information cluster. The substrate information-enzyme sequence pairing data of the selected target clusters are used as the in-domain dataset. Within the in-domain dataset, the training set, in-domain validation set, and in-domain test set are split in an 8:1:1 ratio. The substrate information-enzyme sequence pairing data in the remaining clusters (excluding the target clusters) are used as out-of-domain data. Hierarchical clustering is used to split the enzyme kinetic parameter sample into in-domain and out-of-domain datasets to obtain the out-of-domain dataset to verify the generalization ability of the enzyme kinetic parameter prediction model.

[0074] In this embodiment, after splitting the enzyme kinetic parameter samples to obtain an intra-domain dataset containing a training set, an intra-domain validation set, an intra-domain test set, and an extra-domain test set, the initial enzyme kinetic parameter prediction model is further trained using these datasets. Specifically, the feature extraction sub-model of the initial enzyme kinetic parameter prediction model is trained and validated using the substrate information and enzyme sequence from the training sets of the molecular language model and the protein language model. Then, the pre-trained feature extraction sub-model containing target pre-training weights is fine-tuned using low-rank adaptation techniques to obtain the molecular language model and the protein language model, respectively. The domain adversarial neural network of the initial enzyme kinetic parameter prediction model is trained using the intra-domain dataset and the extra-domain test set to obtain a target domain adversarial neural network containing a target feature converter, a target enzyme kinetic parameter predictor, and a target domain discriminator. Based on the protein language model, the molecular language model, and the target domain adversarial neural network, the target enzyme kinetic parameter prediction model is obtained. Specifically, the feature extraction sub-model is trained using the training sets of the molecular language model and the protein language model. The initial enzyme kinetic parameter prediction model is validated and evaluated using substrate information and enzyme sequences from the in-domain dataset and the out-of-domain test set. Then, the pre-trained feature extraction sub-model containing the target pre-training weights is fine-tuned using low-rank adaptation technology to obtain the molecular language model containing the first low-rank adaptation matrix, the second low-rank adaptation matrix, and the first pre-training weight matrix, and the protein language model containing the third low-rank adaptation matrix, the fourth low-rank adaptation matrix, and the second pre-training weight matrix, respectively.

[0075] Understandably, by choosing a suitable neural network architecture, such as the Transformer architecture, which can handle long-term dependencies in sequences, an initial protein language model can be constructed. Then, the amino acid sequences from the training set of the protein language model, encompassing various enzyme sequences, are encoded. The training set uses the UniRef dataset. Common encoding methods include one-hot encoding or more complex embedding encoding. The encoded amino acid sequences are then input into the initial protein language model, allowing it to predict the next amino acid or perform other related tasks. A loss function is calculated to measure the difference between the prediction and the actual result. Model parameters are updated using backpropagation, and appropriate regularization methods are employed to avoid overfitting. Suitable optimization algorithms, such as stochastic gradient descent and its variants, are used to adjust the parameters. This process is repeated until the model converges or reaches a certain performance metric, resulting in a pre-trained feature extraction sub-model (pre-trained protein language model). Simultaneously, a suitable model is selected based on the molecular characteristics of the substrate information; for example, a graph neural network-based model can be chosen as the initial molecular language model. Alternatively, other models suitable for handling structural information can also be selected as the initial molecular language model; no specific limitations are imposed. Then, the molecular structure data of the substrate information in the training set of the molecular language model are preprocessed appropriately, such as cleaning and normalization, to obtain the target molecular structure data. This target molecular structure data is then transformed into feature representations that can be processed by the initial molecular language model, including atom types and bond information. The training set for the molecular language model uses the ChEMBL dataset. Molecular feature data is input into the model, and the initial molecular language model learns the relationship between molecular structure and chemical properties, calculates the loss, and updates the parameters. Different combinations of hyperparameters, such as learning rate and number of network layers, are experimented with to optimize the performance of the initial molecular language model. The current molecular language model that meets the training stopping condition is used as the trained molecular language model, i.e., the pre-trained feature extraction sub-model (pre-trained molecular language model).

[0076] Then, the pre-trained protein language model and the pre-trained molecular language model are fine-tuned in low rank using the training set within the domain dataset to update their parameters. Furthermore, the initial enzyme kinetic parameter prediction model, incorporating the pre-trained protein language model and the pre-trained molecular language model, is evaluated and tested using the domain validation set and the domain test set within the domain dataset. The domain validation set is primarily used during training to evaluate the performance of the initial enzyme kinetic parameter prediction model and to adjust hyperparameters. It helps determine if the model is overfitting or underfitting, allowing for the selection of appropriate model architecture, learning rate, and other parameters. The domain validation set does not directly update model parameters; it only provides an intermediate evaluation. The domain test set is used after model training to objectively verify and evaluate the final performance of the initial enzyme kinetic parameter prediction model. It represents the model's performance on new, unseen data. The domain test set should be independent of the training set and the domain validation set to ensure the reliability of the evaluation results.

[0077] Subsequently, the performance of the initial enzyme kinetic parameter prediction model, which includes pre-trained protein language models and pre-trained molecular language models, was evaluated using an out-of-domain test set to examine differences in data distribution. When the initial enzyme kinetic parameter prediction model performed well on the in-domain test set but poorly on the out-of-domain test set, it indicated a data distribution discrepancy, requiring fine-tuning of both the pre-trained protein language model and the pre-trained molecular language model. This also indicated that the split out-of-domain data met the splitting requirements.

[0078] Low-rank adaptation techniques are employed to fine-tune the pre-trained feature extraction sub-models (pre-trained molecular language model and pre-trained protein language model). Specifically, this is achieved by combining the pre-trained weight matrices of the pre-trained feature extraction sub-models with two low-rank adaptation matrices. In particular, two low-rank adaptation matrices are added to the pre-trained weight matrices of the pre-trained protein language model and the pre-trained molecular language model, forming new weight matrix forms. That is, by introducing low-rank adaptation matrices, the large target pre-trained weight matrix is ​​decomposed into the product of two smaller matrices (the product of two low-rank matrices), reducing model complexity and the number of parameters, improving generalization ability, and reducing the risk of overfitting. Specifically, the target pre-trained weights of the pre-trained molecular language model are decomposed into the product of the first and second low-rank adaptation matrices; the target pre-trained weights of the pre-trained protein language model are decomposed into the product of the third and fourth low-rank adaptation matrices. In this way, during fine-tuning, only the parameters of the low-rank adaptation matrices are updated, while the pre-trained weight matrices remain unchanged. This approach fully leverages the knowledge gained from pre-trained models, allowing for parameter adjustments tailored to specific tasks, thereby improving prediction accuracy and stability. This technique will accelerate model convergence and enhance its effectiveness and reliability in practical applications.

[0079] Specifically, such as Figure 2 As shown, the steps for obtaining the fine-tuned protein language model and molecular language model using an in-domain dataset, an out-of-domain training set, and low-rank adaptation techniques are as follows:

[0080] The molecular language extraction sub-model of the initial enzyme kinetic parameter prediction model is trained using the molecular structure data and chemical properties of the substrate information in the training set of the molecular language model. This allows the molecular language extraction sub-model to learn the relationship between the molecular structure data and chemical properties of the substrate information, thus obtaining a pre-trained molecular language model.

[0081] The protein language extraction sub-model of the initial enzyme kinetic parameter prediction model is trained using the amino acid sequence and structural features of the enzyme sequence in the training set of the protein language model, so that the protein language extraction sub-model learns the relationship between the amino acid sequence and the structural features of the enzyme sequence to obtain a pre-trained protein language model.

[0082] The initial enzyme kinetic parameter prediction model, which includes the pre-trained molecular language model and the pre-trained protein language model, is validated and evaluated using the domain validation set and the domain test set to obtain a first validation evaluation result. Specifically, the domain validation set is input into the initial enzyme kinetic parameter prediction model so that the pre-trained molecular language model extracts the corresponding molecular features and the pre-trained protein model extracts the corresponding protein features. The two features are then concatenated to obtain concatenated features. The enzyme kinetic parameter prediction model then predicts enzyme kinetic parameters based on the concatenated features to obtain the corresponding predicted enzyme kinetic parameters. The predicted enzyme kinetic parameters are compared with the corresponding true values ​​of enzyme kinetic parameters in the domain validation set to verify the performance of the enzyme kinetic parameter prediction model. The validated enzyme kinetic parameter prediction model is then further tested on the domain test set, with the testing process being the same as the validation process described above. The corresponding test results are then obtained as the first validation evaluation result.

[0083] The initial enzyme kinetic parameter prediction model, which includes the pre-trained molecular language model and the pre-trained protein language model, is validated and evaluated using the out-of-domain test set with in-domain data to obtain a second validation evaluation result. Specifically, after the in-domain test set validation evaluation, the initial enzyme kinetic parameter prediction model is validated and evaluated using the out-of-domain test set. The validation evaluation process is the same as the testing process for the in-domain validation set and the in-domain test set, as described above, to obtain the corresponding second validation evaluation result.

[0084] When the difference between the verification accuracy of the second verification evaluation result and the verification accuracy of the first verification evaluation result is less than a preset threshold, the pre-trained molecular language model and the pre-trained protein language model are used as pre-trained feature extraction sub-models.

[0085] The target pre-training weights of the pre-trained molecular language model and the pre-trained protein language model are decomposed into the product of two low-rank fitness matrices to obtain a molecular language model containing a first low-rank fitness matrix, a second low-rank fitness matrix, and a first pre-training weight matrix, and a protein language model containing a third low-rank fitness matrix, a fourth low-rank fitness matrix, and a second pre-training weight matrix. The pre-training weights in the molecular language model are the first pre-training weight matrix, and the pre-training weights in the protein language model are the second pre-training weight matrix. Therefore, decomposing the pre-training matrices into the product of two low-rank matrices reduces the number of parameters that need to be adjusted during fine-tuning (only these two low-rank matrices need to be fine-tuned). This operation transforms the parameters of the original pre-trained molecular language model or the pre-trained protein language model (only the pre-training weight matrix) into an original pre-training weight matrix (parameters frozen, no fine-tuning) plus two low-rank matrices (fine-tuning).

[0086] The improved protein language model and molecular language model will be trained on the same training data, and their performance will be tested on both in-domain and out-of-domain test sets. If the model performs well on both test sets, it indicates that the model's generalization ability has been improved. Therefore, the trained protein language model and molecular language model possess good generalization ability.

[0087] Therefore, the fine-tuned protein and molecular language models were used to extract features from enzyme sequences and substrate information. The protein language model captures key information from the enzyme sequence, including amino acid sequences and structural features; the molecular language model extracts the molecular structure and chemical properties of the substrate. These feature extraction methods transform biological information into numerical representations, providing input for subsequent machine learning models. Fine-tuning the pre-trained model using low-rank adaptation techniques further improves the model's generalization ability and prediction accuracy.

[0088] In this embodiment, after the feature extraction sub-model is trained, a protein language model and a molecular language model are obtained. The enzyme sequence feature information and substrate feature information of the enzyme kinetic parameter samples output by the protein language model and the molecular language model are concatenated to form the input vector for training the domain adversarial neural network. The enzyme kinetic parameter prediction ability and domain label prediction ability of the domain adversarial neural network are further trained using the intra-domain and extra-domain label information of the enzyme kinetic parameter samples. The output of the current domain adversarial neural network that meets the training stopping condition is used as the target domain adversarial network. A target enzyme kinetic parameter prediction model is constructed based on the protein language model, the molecular language model, and the target domain adversarial network. The target enzyme kinetic parameter prediction model can be a common machine learning algorithm such as a linear model, decision tree, or support vector machine, or it can be a deep learning model; no specific limitation is made.

[0089] Step S12: Input the target reaction substrate and the target enzyme sequence into the target enzyme kinetic parameter prediction model, so that the target feature extraction sub-model of the target enzyme kinetic parameter prediction model can extract the substrate feature information of the target reaction substrate and the enzyme sequence feature information of the target enzyme sequence, respectively, and input the target feature information after splicing the substrate feature information and the enzyme sequence feature information into the target domain adversarial neural network of the target enzyme kinetic parameter prediction model.

[0090] In this embodiment, the trained target enzyme kinetic parameter prediction model is used to extract and concatenate features from the target reaction substrate and target enzyme sequence to be predicted. This concatenated target feature information is then further input into the target domain adversarial neural network via the target enzyme kinetic parameter prediction model. It is understood that simultaneously acquiring substrate feature information and enzyme sequence feature information, and effectively concatenating and processing them, comprehensively considers the influence of multiple factors on enzyme kinetic parameters.

[0091] Step S13: The target feature information is transformed by the target domain adversarial neural network to obtain a target feature vector of a preset dimension. The target feature vector is then used to predict enzyme kinetic parameters to map to the corresponding predicted enzyme kinetic parameters. The target feature vector is then used to map domain labels to obtain the corresponding domain prediction labels. The predicted enzyme kinetic parameters and domain prediction labels corresponding to the target reaction substrate and the target enzyme sequence are then output.

[0092] In this embodiment, the target feature information is transformed by the target feature converter of the target domain adversarial neural network to obtain a target feature vector of a preset dimension; the target enzyme kinetic parameter predictor of the target feature vector is used to predict enzyme kinetic parameters to map to the corresponding predicted enzyme kinetic parameter values; the target domain discriminator of the target domain adversarial neural network is used to map domain labels to the target feature vector to obtain the corresponding predicted domain labels; and the predicted enzyme kinetic parameters and the predicted domain labels corresponding to the target reaction substrate and the target enzyme sequence are output. Understandably, the target feature vector in the target domain adversarial neural network transforms the target feature information to obtain a target feature vector of a preset dimension, namely, a D-dimensional target feature vector. This target feature vector is then passed to the target enzyme kinetic parameter predictor and the target domain discriminator, respectively. In this way, the target enzyme kinetic parameter predictor and the target domain discriminator each output their corresponding prediction information. That is, the target enzyme kinetic parameter predictor predicts the corresponding enzyme kinetic parameter value based on the D-dimensional target feature vector, while the target domain discriminator predicts the corresponding domain label based on the D-dimensional target feature vector, i.e., predicting whether the target reaction substrate and target enzyme sequence belong to in-domain or out-of-domain data. Here, the D-dimensional target feature vector refers to a vector composed of D values. Each value represents an abstract representation of the input data in a specific dimension or feature. It can be understood as a set of descriptive values ​​obtained after highly condensing and summarizing the input information.

[0093] As can be seen, this application discloses training an initial enzyme kinetic parameter prediction model, including a feature extraction sub-model and a domain adversarial neural network, using an in-domain dataset and an out-of-domain test set to obtain a target enzyme kinetic parameter prediction model; inputting the target reaction substrate and the target enzyme sequence into the target enzyme kinetic parameter prediction model, so that the target feature extraction sub-model of the target enzyme kinetic parameter prediction model extracts the substrate feature information of the target reaction substrate and the enzyme sequence feature information of the target enzyme sequence, respectively, and inputting the target feature information, which is the concatenation of the substrate feature information and the enzyme sequence feature information, into the target domain adversarial neural network of the target enzyme kinetic parameter prediction model; performing feature transformation on the target feature information through the target domain adversarial neural network to obtain a target feature vector of a preset dimension, performing enzyme kinetic parameter prediction on the target feature vector to map it to the corresponding enzyme kinetic parameter prediction value, and performing domain label mapping on the target feature vector to obtain the corresponding domain prediction label, and outputting the enzyme kinetic parameter prediction value and the domain prediction label corresponding to the target reaction substrate and the target enzyme sequence. Therefore, by training an initial enzyme kinetic parameter prediction model incorporating a domain adversarial neural network (DAN), a target enzyme kinetic parameter prediction model is obtained. The DAN can map and transform features across different data domains, thereby improving the adaptability and generalization ability of the target enzyme kinetic parameter prediction model across different data domains. It can simultaneously acquire substrate and enzyme sequence features, effectively splicing and processing them, comprehensively considering the influence of multiple factors on enzyme kinetic parameters. Through a specialized training process, it can more accurately extract features from the substrate and enzyme sequence, thus improving the accuracy of enzyme kinetic parameter prediction. Furthermore, the output predicted domain labels provide more comprehensive analytical results, helping to better understand the source and nature of the data, making the process of predicting enzyme kinetic parameters and domain labels more efficient, accurate, and stable.

[0094] Reference Figure 3 As shown, this embodiment of the invention discloses a specific method for constructing a target domain adversarial neural network. Compared with the previous embodiment, this embodiment further explains and optimizes the technical solution. Specifically:

[0095] Step S21: Obtain the intradomain substrate feature information, extradomain substrate feature information, intradomain enzyme sequence feature information, and extradomain enzyme sequence feature information output by the pre-trained protein language model and the pre-trained molecular language model. Concatenate the intradomain substrate feature information and the intradomain enzyme sequence feature information to obtain intradomain data feature information. Concatenate the extradomain substrate feature information and the extradomain enzyme sequence feature information to obtain extradomain data feature information.

[0096] In this embodiment, in order to train the domain adversarial neural network to predict domain labels and enzyme kinetic parameter values, during the processing of training data, the protein language model and the molecular language model extract enzyme sequence feature information and substrate feature information from the training data. Since there are corresponding intra-domain or extra-domain labels in the intra-domain dataset and extra-domain data, the extracted enzyme sequence feature information and substrate feature information carry not only the enzyme kinetic parameter values ​​(true values ​​of enzyme kinetic parameters) in the enzyme kinetic parameter samples, but also the corresponding domain label information. Therefore, using the feature information carrying domain label information and true values ​​of enzyme kinetic parameters, the corresponding feature information is spliced ​​according to the substrate-sequence pairing data to obtain the corresponding intra-domain / extra-domain data feature information respectively.

[0097] Step S22: Construct a domain adversarial neural network that includes a feature converter, an enzyme kinetic parameter predictor, and a domain discriminator.

[0098] In this embodiment, a domain adversarial neural network is constructed based on a feature converter, an enzyme kinetic parameter predictor, and a domain discriminator. The feature converter is used to convert feature information into feature vectors of a preset dimension; the enzyme kinetic parameter predictor is used to predict the enzyme kinetic parameter values ​​corresponding to the input data; and the domain discriminator is used to distinguish whether the input data belongs to the domain or not.

[0099] Step S23: Input the intra-domain data feature information and the extra-domain data feature information into the domain adversarial neural network, so that the feature converter of the domain adversarial neural network performs multi-layer feature mapping processing on the intra-domain data feature information and the extra-domain data feature information, so as to update the converter parameter information of the feature converter and obtain the target feature converter.

[0100] In this embodiment, firstly, in each iteration, the input vector (intra-domain data feature information and out-of-domain data feature information) is fed into the feature converter. The feature converter performs feature mapping based on the current parameters, obtaining a D-dimensional feature vector. This feature vector is then passed to the enzyme kinetic parameter predictor to calculate the loss for enzyme kinetic parameter prediction. This loss is fed back to the feature converter, prompting it to adjust its parameters so that the output feature vector is more conducive to accurate prediction of enzyme kinetic parameters. Simultaneously, the feature vector is also passed to the domain discriminator. The domain discriminator predicts the domain label based on this and generates a domain classification loss. This loss is also backpropagated to the feature converter, guiding it to learn to extract features that can confuse the domain discriminator, i.e., features with domain invariance. During training, optimization algorithms such as stochastic gradient descent or its variants can be used to update the parameters of the feature converter. By continuously receiving feedback from the enzyme kinetic parameter predictor and the domain discriminator, the feature converter gradually learns how to extract common features from data in different domains that are helpful for predicting enzyme kinetic parameters. At the same time, regularization techniques can be introduced to avoid overfitting or instability in the feature converter. For example, L2 regularization can be used to constrain the size of parameters.

[0101] Furthermore, the feature converter includes multiple feedforward layers for progressively performing feature extraction and transformation. That is, the feature converter is built on a multi-layered neural network, with each layer further processing and transforming the input features to progressively extract higher-level features. Advanced techniques such as attention mechanisms are used in the multi-layered neural network to better process the input vector and perform feature mapping, achieving more efficient feature extraction and transformation to meet the requirement of converting the input vector into a specific D-dimensional feature vector.

[0102] Step S24: The feature vector output by the target feature converter is passed to the enzyme kinetic parameter predictor so that the enzyme kinetic parameter predictor learns the relationship between the feature vector and the corresponding true value of the enzyme kinetic parameter, outputs the enzyme kinetic parameter prediction value of the feature vector, and updates the parameters of the enzyme kinetic parameter predictor according to the information difference between the enzyme kinetic parameter prediction value and the true value of the enzyme kinetic parameter, so as to obtain the target enzyme kinetic parameter predictor.

[0103] In this embodiment, based on the feature vector output by the feature converter, the enzyme kinetic parameter predictor learns the relationship between the feature vector and the corresponding true values ​​of the enzyme kinetic parameters, and outputs the corresponding predicted values ​​of the enzyme kinetic parameters. Then, it calculates the error between the predicted values ​​and the true values ​​of the enzyme kinetic parameters. The parameters of the enzyme kinetic parameter predictor are updated according to the backpropagation of the error to reduce the prediction error and improve the prediction accuracy.

[0104] Step S25: The feature vector output by the target feature converter is passed to the domain discriminator so that the domain discriminator learns the relationship between the feature vector and the corresponding real domain label, outputs the predicted domain label of the feature vector, and updates the parameters of the domain discriminator with the information difference between the predicted domain label and the real domain label to obtain the target domain discriminator.

[0105] In this embodiment, a domain discriminator learns the relationship between feature vectors and true domain labels (data labels 0 within the domain and data labels 1 outside the domain). This allows the domain discriminator to predict the predicted domain labels of the feature vectors and update its parameters based on the information difference between the predicted and true domain labels. Specifically, it receives feature vectors and attempts to distinguish between domain labels 0 and 1 (data within the domain is represented as 0, and data outside the domain is represented as 1). A loss is calculated based on the difference between the predicted and true domain labels. The parameters of the domain discriminator are then updated through backpropagation of the loss, enabling it to better distinguish between different domains.

[0106] It is important to note that these three parts are interconnected and work collaboratively throughout the training process of the Domain Adversarial Neural Network (DANN): the feature converter must provide effective features for the enzyme kinetic parameter predictor and also address the "challenge" of the domain discriminator, continuously adjusting to extract more domain-invariant features. The enzyme kinetic parameter predictor, while optimizing its own prediction accuracy, also influences the feature extraction direction of the feature converter. The training of the domain discriminator drives the feature converter towards domain invariance, and it itself faces new "challenges" as the feature converter improves. This achieves the learning and extraction of domain invariance in features, thereby improving the model's adaptability and generalization ability across different data domains. By continuously iterating through this training process and adjusting the parameters of each part, the goal is ultimately to achieve both accurate prediction of enzyme kinetic parameters and good domain invariance. That is, by calculating the losses of the enzyme kinetic parameter classifier and the domain classifier, iteratively updating the model parameters of the DANN, and maximizing the loss of the domain classifier, the feature converter is encouraged to extract domain-invariant characteristics during the optimization process as input for the subsequent enzyme kinetic parameter predictor, thus enhancing the model's generalization ability.

[0107] In this embodiment, a first loss function is constructed based on the difference between the predicted value of the enzyme kinetic parameter output by the enzyme kinetic parameter predictor and the actual value of the enzyme kinetic parameter during the training process of the domain adversarial neural network (DANN). A second loss function is constructed based on the difference between the predicted domain label output by the domain discriminator and the actual domain label during the training process of the DANN. A target loss function is constructed with the first objective of minimizing the first loss function and the second objective of maximizing the second loss function. The DANN is iteratively trained until a preset training termination condition is met, and the current DANN is output to obtain a target DANN containing the target feature converter, the target enzyme kinetic parameter predictor, and the target domain discriminator. It can be understood that the optimization of the parameters of the domain adversarial neural network (DANN) is achieved by calculating the losses of the enzyme kinetic parameter predictor and the domain classifier, and iteratively updating the model parameters of the DANN.

[0108] First, define the objective function of the domain adversarial network. The goal of the domain adversarial network is to encourage the feature converter to extract domain-invariant features during the optimization process by maximizing the loss of the domain classifier (maximizing the second loss function), making the feature distributions of data from different domains as similar as possible. The domain discriminator is used to distinguish whether the input features belong to in-domain or out-of-domain data based on the transformed feature values. By maximizing the loss of the domain discriminator, the feature converter is encouraged to extract domain-invariant features (because if the feature converter successfully extracts domain-invariant features, the domain discriminator cannot distinguish whether the input features belong to in-domain or out-of-domain data, thus maximizing the prediction error of the domain discriminator). Simultaneously, the model also needs to minimize the loss of the enzyme kinetic parameter predictor (minimizing the first loss function) to ensure the accuracy of the predictor's predictions of enzyme kinetic parameters. Therefore, the objective loss function of the domain adversarial network can be expressed as maximizing the loss of the domain classifier while minimizing the loss of the enzyme kinetic parameter classifier, specifically:

[0109] ;

[0110] Where N represents the number of samples, L enzyme L represents the first loss function of the enzyme kinetic parameter predictor. domain Let λ represent the second loss function of the domain classifier, and let λ represent the hyperparameter that balances the loss of the domain classifier and the loss of the enzyme kinetic parameter classifier.

[0111] Step S26: Construct a target domain adversarial neural network based on the target feature converter, the target enzyme kinetic parameter predictor, and the target domain discriminator.

[0112] In this embodiment, a target domain adversarial neural network is constructed based on the target feature converter, target enzyme kinetic parameter predictor, and target domain discriminator obtained through the aforementioned training, such as... Figure 2 As shown on the right, Figure 2 The domain adversarial neural network in the text is a trained target domain adversarial neural network, which will... Figure 2 After training, the enzyme sequence feature information and substrate feature information output by the protein language model and molecular language model on the left are concatenated to obtain the target feature information, which is then input into the target domain adversarial neural network. The target feature information is transformed into a feature vector with a preset dimension through the target feature converter, target enzyme kinetic parameter predictor and target domain discriminator of the target domain adversarial neural network. The transformed target feature vector is then input into the target enzyme kinetic parameter predictor and the target domain discriminator respectively, so as to predict the corresponding enzyme kinetic parameter value and domain label respectively.

[0113] It is evident that by processing and splicing the feature information of data both within and outside the domain, the characteristics of substrates and enzyme sequences can be captured more comprehensively, improving the quality and richness of feature extraction. The construction and training of the domain adversarial neural network enables the model to better cope with data from different domains, enhancing its adaptability and robustness in various scenarios. The continuous learning and optimization of the enzyme kinetic parameter predictor helps to more accurately predict enzyme kinetic parameters, providing reliable data support for related research and applications. The training of the domain discriminator can more accurately identify the domain to which the data belongs, thereby better distinguishing data from different sources or with different properties, providing an important basis for subsequent analysis. The updating and optimization of the feature converter ensures the effective conversion and mapping of feature information between different domains, enabling the model to process and utilize this information more effectively. The final target domain adversarial neural network integrates multiple key components, achieving functional synergy and complementarity, improving the overall technical effect and performance. It can fully utilize data both within and outside the domain for training and optimization, enhancing the utilization rate and value of data.

[0114] Reference Figure 4 As shown, the present invention also discloses an enzyme kinetic parameter and domain tag prediction device, comprising:

[0115] Model training module 11 is used to train the initial enzyme kinetic parameter prediction model, which includes a feature extraction sub-model and a domain adversarial neural network, using the in-domain dataset and the out-of-domain test set, so as to obtain the target enzyme kinetic parameter prediction model.

[0116] Information extraction module 12 is used to input the target reaction substrate and the target enzyme sequence into the target enzyme kinetic parameter prediction model, so that the target feature extraction sub-model of the target enzyme kinetic parameter prediction model can extract the substrate feature information of the target reaction substrate and the enzyme sequence feature information of the target enzyme sequence, respectively, and input the target feature information after splicing the substrate feature information and the enzyme sequence feature information into the target domain adversarial neural network of the target enzyme kinetic parameter prediction model;

[0117] The parameter prediction module 13 is used to perform feature transformation on the target feature information through the target domain adversarial neural network to obtain a target feature vector of a preset dimension, perform enzyme kinetic parameter prediction on the target feature vector to map it to the corresponding enzyme kinetic parameter prediction value, and perform domain label mapping on the target feature vector to obtain the corresponding domain prediction label, and output the enzyme kinetic parameter prediction value and domain prediction label corresponding to the target reaction substrate and the target enzyme sequence.

[0118] As can be seen, this application discloses training an initial enzyme kinetic parameter prediction model, including a feature extraction sub-model and a domain adversarial neural network, using an in-domain dataset and an out-of-domain test set to obtain a target enzyme kinetic parameter prediction model; inputting the target reaction substrate and the target enzyme sequence into the target enzyme kinetic parameter prediction model, so that the target feature extraction sub-model of the target enzyme kinetic parameter prediction model extracts the substrate feature information of the target reaction substrate and the enzyme sequence feature information of the target enzyme sequence, respectively, and inputting the target feature information, which is the concatenation of the substrate feature information and the enzyme sequence feature information, into the target domain adversarial neural network of the target enzyme kinetic parameter prediction model; performing feature transformation on the target feature information through the target domain adversarial neural network to obtain a target feature vector of a preset dimension, performing enzyme kinetic parameter prediction on the target feature vector to map it to the corresponding enzyme kinetic parameter prediction value, and performing domain label mapping on the target feature vector to obtain the corresponding domain prediction label, and outputting the enzyme kinetic parameter prediction value and the domain prediction label corresponding to the target reaction substrate and the target enzyme sequence. Therefore, by training an initial enzyme kinetic parameter prediction model incorporating a domain adversarial neural network (DAN), a target enzyme kinetic parameter prediction model is obtained. The DAN can map and transform features across different data domains, thereby improving the adaptability and generalization ability of the target enzyme kinetic parameter prediction model across different data domains. It can simultaneously acquire substrate and enzyme sequence features, effectively splicing and processing them, comprehensively considering the influence of multiple factors on enzyme kinetic parameters. Through a specialized training process, it can more accurately extract features from the substrate and enzyme sequence, thus improving the accuracy of enzyme kinetic parameter prediction. Furthermore, the output predicted domain labels provide more comprehensive analytical results, helping to better understand the source and nature of the data, making the process of predicting enzyme kinetic parameters and domain labels more efficient, accurate, and stable.

[0119] Furthermore, embodiments of this application also disclose an electronic device, Figure 5 This is a structural diagram of an electronic device 20 according to an exemplary embodiment. The content of the diagram should not be construed as limiting the scope of this application.

[0120] Figure 5 This is a schematic diagram of the structure of an electronic device 20 provided in an embodiment of this application. Specifically, the electronic device 20 may include: at least one processor 21, at least one memory 22, a power supply 23, a communication interface 24, an input / output interface 25, and a communication bus 26. The memory 22 stores a computer program, which is loaded and executed by the processor 21 to implement the relevant steps in the enzyme kinetic parameter and domain tag prediction method disclosed in any of the foregoing embodiments. Alternatively, the electronic device 20 in this embodiment may specifically be an electronic computer.

[0121] In this embodiment, the power supply 23 is used to provide operating voltage for each hardware device on the electronic device 20; the communication interface 24 can create a data transmission channel between the electronic device 20 and external devices, and the communication protocol it follows can be any communication protocol applicable to the technical solution of this application, and is not specifically limited here; the input / output interface 25 is used to acquire external input data or output data to the outside world, and its specific interface type can be selected according to specific application needs, and is not specifically limited here.

[0122] The processor 21 may include one or more processing cores, such as a quad-core processor or an octa-core processor. The processor 21 may be implemented using at least one hardware form selected from DSP (Digital Signal Processing), FPGA (Field-Programmable Gate Array), and PLA (Programmable Logic Array). The processor 21 may also include a main processor and a coprocessor. The main processor, also known as a CPU (Central Processing Unit), is used to process data in the wake-up state; the coprocessor is a low-power processor used to process data in the standby state. In some embodiments, the processor 21 may integrate a GPU (Graphics Processing Unit), which is responsible for rendering and drawing the content to be displayed on the screen. In some embodiments, the processor 21 may also include an AI (Artificial Intelligence) processor, which is used to handle computational operations related to machine learning.

[0123] In addition, the memory 22, as a carrier for resource storage, can be a read-only memory, random access memory, disk or optical disk, etc. The resources stored thereon can include operating system 221, computer program 222, etc., and the storage method can be temporary storage or permanent storage.

[0124] The operating system 221 manages and controls the various hardware devices and computer programs 222 on the electronic device 20 to enable the processor 21 to perform calculations and processing on the massive amounts of data 223 in the memory 22. It can be Windows Server, Netware, Unix, Linux, etc. The computer program 222, in addition to including a computer program capable of performing the enzyme kinetic parameter and domain label prediction method disclosed in any of the foregoing embodiments, may further include computer programs capable of performing other specific tasks. The data 223 may include data received by the electronic device from external devices, as well as data collected by its own input / output interface 25.

[0125] Furthermore, this application also discloses a computer-readable storage medium for storing a computer program; wherein, when the computer program is executed by a processor, it implements the aforementioned enzyme kinetic parameter and domain label prediction method. Specific steps of this method can be found in the corresponding content disclosed in the foregoing embodiments, and will not be repeated here.

[0126] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on its differences from other embodiments. Similar or identical parts between embodiments can be referred to interchangeably. For the apparatus disclosed in the embodiments, since it corresponds to the method disclosed in the embodiments, the description is relatively simple; relevant parts can be referred to in the method section.

[0127] Those skilled in the art will further recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, computer software, or a combination of both. To clearly illustrate the interchangeability of hardware and software, the components and steps of the various examples have been generally described in terms of functionality in the foregoing description. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application. The steps of the methods or algorithms described in conjunction with the embodiments disclosed herein can be implemented directly in hardware, software modules executed by a processor, or a combination of both. The software module may be located in random access memory (RAM), memory, read-only memory (ROM), electrically programmable read-only memory (EPROM), electrically erasable programmable read-only memory (EEPROM), registers, hard disks, removable disks, CD-ROMs (Compact Disc-Read Only Memory), or any other form of storage medium known in the art.

[0128] Finally, it should be noted that in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0129] The solution provided by the present invention has been described in detail above. Specific examples have been used to illustrate the principle and implementation of the present invention. The description of the above embodiments is only for the purpose of helping to understand the method and core idea of ​​the present invention. At the same time, for those skilled in the art, there will be changes in the specific implementation and application scope based on the idea of ​​the present invention. Therefore, the content of this specification should not be construed as a limitation of the present invention.

Claims

1. A method for predicting enzyme kinetic parameters and domain tags, characterized in that, include: The initial enzyme kinetic parameter prediction model, which includes a feature extraction sub-model and a domain adversarial neural network, is trained using an in-domain dataset and an out-of-domain test set to obtain the target enzyme kinetic parameter prediction model. The target reaction substrate and the target enzyme sequence are input into the target enzyme kinetic parameter prediction model, so that the target feature extraction sub-model of the target enzyme kinetic parameter prediction model can extract the substrate feature information of the target reaction substrate and the enzyme sequence feature information of the target enzyme sequence, respectively. The target feature information, which is the result of splicing the substrate feature information and the enzyme sequence feature information, is input into the target domain adversarial neural network of the target enzyme kinetic parameter prediction model. The target feature information is transformed by the target domain adversarial neural network to obtain a target feature vector of a preset dimension. The target feature vector is then used to predict enzyme kinetic parameters to map to the corresponding predicted enzyme kinetic parameter values. The target feature vector is also used to map domain labels to obtain corresponding domain prediction labels. Finally, the predicted enzyme kinetic parameters and domain prediction labels corresponding to the target reaction substrate and the target enzyme sequence are output.

2. The method for predicting enzyme kinetic parameters and domain tags according to claim 1, characterized in that, The process of training an initial enzyme kinetic parameter prediction model, comprising a feature extraction sub-model and a domain adversarial neural network, using an in-domain dataset and an out-of-domain test set to obtain a target enzyme kinetic parameter prediction model includes: The enzyme kinetic parameter samples are split into an in-domain dataset and an out-of-domain test set using hierarchical clustering; wherein the enzyme kinetic parameter samples include substrate information, enzyme sequence and corresponding enzyme kinetic parameter values; The feature extraction sub-model of the initial enzyme kinetic parameter prediction model is trained and validated using the substrate information and enzyme sequence in the in-domain dataset and the out-of-domain test set. Then, the pre-trained feature extraction sub-model containing the target pre-training weights is fine-tuned using low-rank adaptation technology to obtain the molecular language model and the protein language model, respectively. The domain adversarial neural network of the initial enzyme kinetic parameter prediction model is trained using the in-domain dataset and the out-of-domain test set to obtain a target domain adversarial neural network that includes a target feature converter, a target enzyme kinetic parameter predictor, and a target domain discriminator. Based on the protein language model, the molecular language model, and the target domain adversarial neural network, a target enzyme kinetic parameter prediction model is obtained.

3. The method for predicting enzyme kinetic parameters and domain tags according to claim 2, characterized in that, The process of splitting enzyme kinetic parameter samples into an in-domain dataset and an out-of-domain test set through hierarchical clustering includes: The substrate information and enzyme sequence in the enzyme kinetic parameter sample are each divided into a first preset number of clusters to obtain a first preset number of substrate information clusters and a first preset number of enzyme sequence clusters; A second preset number of substrate information clusters and enzyme sequence clusters are randomly selected as target substrate information clusters and target enzyme sequence clusters. The paired data of substrate information and enzyme sequence in the target substrate information clusters and target enzyme sequence clusters are used as an in-domain dataset. The substrate data in the target substrate information clusters and the enzyme sequence data in the target enzyme sequence clusters, excluding the in-domain datasets, are paired as an out-of-domain test set. The second preset number is less than the first preset number. The domain dataset is split according to a preset splitting ratio to obtain a training set, a domain validation set, and a domain test set. Accordingly, the feature extraction sub-model of the initial enzyme kinetic parameter prediction model is trained and validated using substrate information from the in-domain dataset and the out-of-domain test set, and then fine-tuned using low-rank adaptation techniques to obtain the molecular language model and the protein language model, respectively. The feature extraction sub-model of the initial enzyme kinetic parameter prediction model is trained and validated using substrate information and enzyme sequences from the in-domain dataset and the out-of-domain test set. Then, the pre-trained feature extraction sub-model containing the target pre-training weights is fine-tuned using low-rank adaptation technology to obtain a molecular language model containing a first low-rank adaptation matrix and a first pre-training weight matrix, and a protein language model containing a second low-rank adaptation matrix and a second pre-training weight matrix, respectively.

4. The method for predicting enzyme kinetic parameters and domain tags according to claim 3, characterized in that, The process involves using substrate information from the in-domain dataset and the out-of-domain test set, along with the enzyme sequence, to train and validate a feature extraction sub-model for the initial enzyme kinetic parameter prediction model. Then, low-rank adaptation techniques are used to fine-tune the pre-trained feature extraction sub-model containing the target pre-training weights, resulting in a molecular language model containing a first low-rank adaptation matrix and a first pre-training weight matrix, and a protein language model containing a second low-rank adaptation matrix and a second pre-training weight matrix, respectively. The molecular language extraction sub-model of the initial enzyme kinetic parameter prediction model is trained using the molecular structure data and chemical properties of the substrate information in the training set of the molecular language model. This allows the molecular language extraction sub-model to learn the relationship between the molecular structure data and chemical properties of the substrate information, thus obtaining a pre-trained molecular language model. The protein language extraction sub-model of the initial enzyme kinetic parameter prediction model is trained using the amino acid sequence and structural features of the enzyme sequence in the training set of the protein language model, so that the protein language extraction sub-model learns the relationship between the amino acid sequence and the structural features of the enzyme sequence to obtain a pre-trained protein language model. The initial enzyme kinetic parameter prediction model, which includes the pre-trained molecular language model and the pre-trained protein language model, is validated and evaluated using the intra-domain validation set and the intra-domain test set to obtain a first validation evaluation result. The initial enzyme kinetic parameter prediction model, which includes the pre-trained molecular language model and the pre-trained protein language model, is validated and evaluated using the out-of-domain test set to obtain a second validation evaluation result. When the difference between the verification accuracy of the second verification evaluation result and the verification accuracy of the first verification evaluation result is less than a preset threshold, the pre-trained molecular language model and the pre-trained protein language model are used as pre-trained feature extraction sub-models. The target pre-training weights of the pre-trained molecular language model and the pre-trained protein language model are decomposed into the product of two low-rank adaptation matrices to obtain a molecular language model containing a first low-rank adaptation matrix, a second low-rank adaptation matrix, and a first pre-training weight matrix, and a protein language model containing a third low-rank adaptation matrix, a fourth low-rank adaptation matrix, and a second pre-training weight matrix.

5. The method for predicting enzyme kinetic parameters and domain tags according to claim 1, characterized in that, The process of training an initial enzyme kinetic parameter prediction model, comprising a feature extraction sub-model and a domain adversarial neural network, using an in-domain dataset and an out-of-domain test set to obtain a target enzyme kinetic parameter prediction model includes: The intradomain substrate feature information, extradomain substrate feature information, intradomain enzyme sequence feature information, and extradomain enzyme sequence feature information output by the pre-trained protein language model and the pre-trained molecular language model are obtained. The intradomain substrate feature information and the intradomain enzyme sequence feature information are concatenated to obtain intradomain data feature information. The extradomain substrate feature information and the extradomain enzyme sequence feature information are concatenated to obtain extradomain data feature information. Construct a domain adversarial neural network that includes a feature converter, an enzyme kinetic parameter predictor, and a domain discriminator; The data feature information within the domain and the data feature information outside the domain are input into the domain adversarial neural network so that the feature converter of the domain adversarial neural network performs multi-layer feature mapping processing on the data feature information within the domain and the data feature information outside the domain, thereby updating the converter parameter information of the feature converter and obtaining the target feature converter. The feature vector output by the target feature converter is passed to the enzyme kinetic parameter predictor so that the enzyme kinetic parameter predictor learns the relationship between the feature vector and the corresponding true value of the enzyme kinetic parameter, outputs the enzyme kinetic parameter prediction value of the feature vector, and updates the parameters of the enzyme kinetic parameter predictor according to the information difference between the predicted value of the enzyme kinetic parameter and the true value of the enzyme kinetic parameter to obtain the target enzyme kinetic parameter predictor. The feature vector output by the target feature converter is passed to the domain discriminator so that the domain discriminator can learn the relationship between the feature vector and the corresponding real domain label, output the predicted domain label of the feature vector, and update the parameters of the domain discriminator with the information difference between the predicted domain label and the real domain label to obtain the target domain discriminator. A target domain adversarial neural network is constructed based on the target feature converter, the target enzyme kinetic parameter predictor, and the target domain discriminator.

6. The method for predicting enzyme kinetic parameters and domain tags according to claim 5, characterized in that, Also includes: A first loss function is constructed based on the difference between the predicted value of the enzyme kinetic parameter output by the enzyme kinetic parameter predictor during the training process of the domain adversarial neural network and the true value of the enzyme kinetic parameter; a second loss function is constructed based on the difference between the predicted domain label output by the domain discriminator during the training process of the domain adversarial neural network and the true domain label. Construct a target loss function with the first objective of minimizing the first loss function and the second objective of maximizing the second loss function; The domain adversarial neural network is iteratively trained until a preset training termination condition is met. The current domain adversarial neural network is then output to obtain a target domain adversarial neural network that includes the target feature converter, the target enzyme kinetic parameter predictor, and the target domain discriminator.

7. The method for predicting enzyme kinetic parameters and domain tags according to any one of claims 1 to 6, characterized in that, The step of performing feature transformation on the target feature information through the target domain adversarial neural network to obtain a target feature vector of a preset dimension, predicting enzyme kinetic parameters on the target feature vector to map it to the corresponding predicted enzyme kinetic parameter values, and mapping the target feature vector to domain labels to obtain corresponding predicted domain labels, and outputting the predicted enzyme kinetic parameters and domain predicted labels corresponding to the target reaction substrate and the target enzyme sequence, includes: The target feature information is transformed by the target feature converter of the target domain adversarial neural network to obtain a target feature vector of a preset dimension. The target feature vector is used to predict enzyme kinetic parameters by the target enzyme kinetic parameter predictor of the target domain adversarial neural network, so as to map it to the corresponding enzyme kinetic parameter prediction value; The target feature vector is mapped to a domain label by the target domain discriminator of the target domain adversarial neural network to obtain the corresponding domain predicted label. Output the predicted values ​​of the enzyme kinetic parameters and the domain prediction labels corresponding to the target reaction substrate and the target enzyme sequence.

8. A device for predicting enzyme kinetic parameters and domain tags, characterized in that, include: The model training module is used to train the initial enzyme kinetic parameter prediction model, which includes a feature extraction sub-model and a domain adversarial neural network, using the in-domain dataset and the out-of-domain test set, so as to obtain the target enzyme kinetic parameter prediction model. The information extraction module is used to input the target reaction substrate and the target enzyme sequence into the target enzyme kinetic parameter prediction model, so that the target feature extraction sub-model of the target enzyme kinetic parameter prediction model can extract the substrate feature information of the target reaction substrate and the enzyme sequence feature information of the target enzyme sequence, respectively, and input the target feature information after splicing the substrate feature information and the enzyme sequence feature information into the target domain adversarial neural network of the target enzyme kinetic parameter prediction model; The parameter prediction module is used to perform feature transformation on the target feature information through the target domain adversarial neural network to obtain a target feature vector of a preset dimension, perform enzyme kinetic parameter prediction on the target feature vector to map it to the corresponding enzyme kinetic parameter prediction value, and perform domain label mapping on the target feature vector to obtain the corresponding domain prediction label, and output the enzyme kinetic parameter prediction value and domain prediction label corresponding to the target reaction substrate and the target enzyme sequence.

9. An electronic device, characterized in that, include: Memory, used to store computer programs; A processor for executing the computer program to implement the steps of the enzyme kinetic parameters and domain tag prediction method as described in any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, Used to store a computer program; wherein, when the computer program is executed by a processor, it implements the steps of the enzyme kinetic parameters and domain label prediction method as described in any one of claims 1 to 7.

Images