Enzyme catalysis constant kcat prediction method and system based on multi-modal feature fusion

Abstract

The invention discloses an enzyme catalysis constant kcat prediction method and system based on multi-modal feature fusion, and the method comprises the steps: S1, obtaining an enzyme-substrate pair which comprises an amino acid sequence of a target enzyme and a chemical structure expression of a corresponding substrate; s2, inputting the amino acid sequence into a pre-trained protein language model, and carrying out average pooling treatment on hidden layer state vectors of amino acid residues in the sequence by the protein language model to obtain enzyme feature vectors for representing functional features of enzyme; s3, on the basis of the chemical structure representation of the substrate, coding the chemical structure through the set radius parameter and bit length parameter by adopting an extended connectivity chemical fingerprint algorithm to obtain a substrate feature vector for representing the structural features of the substrate; s4, splicing the enzyme feature vector and the substrate feature vector to obtain a fused feature vector; and S5, inputting the fusion feature vector into a random forest regression model, and outputting a predicted value of the catalytic constant kcat of the enzyme-substrate pair.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of bioinformatics and artificial intelligence, and particularly relates to a method and system for predicting enzyme catalytic constant k cat based on multi-modal feature fusion. BACKGROUND

[0002] Enzymes, as efficient biological catalysts, have been widely used in the fields of medicine, chemical industry, food, etc. The catalytic constant k cat of enzyme, as a core kinetic parameter to measure the catalytic efficiency of enzyme, has important significance for understanding the cell metabolic mechanism, guiding the directed evolution of enzyme and optimizing the biological manufacturing process. However, the large-scale determination of k cat values by traditional biochemical experimental methods faces technical problems such as high cost, long cycle, sensitivity to experimental conditions, etc., which limits the development of enzyme research and application.

[0003] To solve the above problems, various k cat prediction technologies based on computational methods have been developed in the field. Early methods mainly rely on traditional machine learning models to predict by extracting physical and chemical properties of proteins and other hand-designed features, but such methods have the problem of insufficient feature representation ability, and it is difficult to capture long-range sequence dependencies and complex biological context information that determine enzyme function, resulting in limited prediction accuracy and generalization ability. In recent years, end-to-end deep learning methods have improved prediction performance to some extent, but their complex network structure is prone to overfitting on limited experimental data, and the prediction performance significantly decreases when facing new enzymes or mutants not appearing in the training set. In addition, the decision-making process of deep learning models lacks explainability and is difficult to provide valuable biological information on catalytic mechanisms, which limits its practical value in downstream applications such as enzyme engineering.

[0004] Therefore, there is an urgent need for a method and system for predicting enzyme catalytic constant k cat based on multi-modal feature fusion. SUMMARY

[0005] The present application provides a method and system for predicting enzyme catalytic constant k cat based on multi-modal feature fusion to solve the above problems existing in the prior art.

[0006] In order to achieve the above purpose, the present application provides the following technical scheme:

[0007] A method for predicting enzyme catalytic constant k cat based on multi-modal feature fusion, comprising:

[0008] S1: obtaining an enzyme-substrate pair, the enzyme-substrate pair comprising an amino acid sequence of a target enzyme and a chemical structure representation of a corresponding substrate;

[0009] S2: input the amino acid sequence into the pre-trained protein language model, and the protein language model performs average pooling processing on the hidden layer state vector of each amino acid residue in the sequence to obtain an enzyme feature vector for representing the functional characteristics of the enzyme;

[0010] S3: based on the chemical structure representation of the substrate, an extended connectivity chemical fingerprint algorithm is used to encode the chemical structure by setting a radius parameter and a bit length parameter to obtain a substrate feature vector for representing the structural characteristics of the substrate;

[0011] S4: splicing the enzyme feature vector and the substrate feature vector to obtain a fusion feature vector;

[0012] S5: inputting the fusion feature vector into a random forest regression model to output the predicted value of the catalytic constant k of the enzyme-substrate pair. cat

[0013] Further, S2 comprises:

[0014] S21: inputting the amino acid sequence into the pre-trained protein language model to obtain a hidden layer state vector corresponding to each amino acid residue in the sequence;

[0015] S22: performing an average pooling operation on the hidden layer state vector of each amino acid residue to compress and convert the context dependence and long-range interaction information encoded in the sequence into a fixed-dimensional enzyme feature vector.

[0016] Further, the pre-trained protein language model is a ProtTrans-T5 model; and the dimension of the enzyme feature vector is 1024.

[0017] Further, S3 comprises:

[0018] S31: parsing the chemical structure representation of the substrate into molecular structure data;

[0019] S32: setting the radius parameter and the bit length parameter of the extended connectivity chemical fingerprint algorithm;

[0020] S33: inputting the molecular structure data and the radius parameter and the bit length parameter into the extended connectivity chemical fingerprint algorithm to perform encoding operation to obtain the substrate feature vector.

[0021] Further, the radius parameter is 3 and the bit length parameter is 2048; and the substrate feature vector is a 2048-bit binary fingerprint vector.

[0022] Further, the random forest regression model in S5 is a pre-trained model, and the training process of the model comprises:

[0023] ​S51: Based on the development dataset containing known catalytic constant logarithm values, use K-fold cross-validation method to evaluate multiple sets of hyperparameter combinations, and obtain the performance evaluation results of each combination in cross-validation;

[0024] S52: According to the performance evaluation results, determine the target hyperparameter combination from the multiple sets of hyperparameter combinations;

[0025] S53: Train the random forest regression model on the development dataset using the target hyperparameter combination to obtain the trained prediction model.

[0026] Further, K is 5;

[0027] The target hyperparameter combination includes: the number of decision trees n_estimators is 1500, and the maximum depth of the tree max_depth is 25.

[0028] Further, the catalytic constant logarithm values in the development dataset are base 10 logarithm values, i.e. 10 (k cat );

[0029] The training target of the random forest regression model is the logarithm value of the catalytic constant.

[0030] Further, after S5, it further includes a verification step:

[0031] S6: Divide the test set containing known catalytic constant values into seen and unseen subsets according to whether they appear in the training set;

[0032] S7: Input the enzyme-substrate pairs in the seen subset into the random forest regression model to obtain the first prediction result, and calculate the first performance indicator based on the first prediction result and the known catalytic constant value in the seen subset;

[0033] S8: Input the enzyme-substrate pairs in the unseen subset into the random forest regression model to obtain the second prediction result, and calculate the second performance indicator based on the second prediction result and the known catalytic constant value in the unseen subset.

[0034] Further, the system applied to the enzyme catalytic constant k cat prediction method based on multi-modal feature fusion, comprising:

[0035] A data acquisition module is configured to acquire and temporarily store enzyme-substrate pair data containing enzyme amino acid sequences and substrate chemical structure representations;

[0036] An enzyme feature extraction module is configured with a pre-trained ProtTrans-T5 protein language model, configured to receive enzyme amino acid sequences, perform hidden layer state vector acquisition and average pooling operations, and output enzyme feature vectors with a dimension of 1024.

[0037] The substrate feature extraction module is equipped with an extended connectivity chemical fingerprint algorithm with a radius parameter of 3 and a bit length parameter of 2048. It is used to receive the chemical structure representation of the substrate and output a substrate feature vector in 2048-bit binary fingerprint format.

[0038] The feature fusion module receives enzyme feature vectors and substrate feature vectors, performs vector concatenation operations, and outputs a fused feature vector with a dimension of 3072.

[0039] The prediction module is configured with a pre-trained random forest regression model, which receives the fused feature vector, performs regression calculations, and outputs the catalytic constant k. cat The predicted value.

[0040] Compared with the prior art, the present invention has the following advantages:

[0041] This invention achieves industry-leading prediction performance by synergistically integrating the deep sequence representation of ProtTrans-T5 with the precise structural representation of ECFP. On an independent test set containing 1192 samples, the proposed method achieved a coefficient of determination (R²) of 0.875 and a root mean square error (RMSE) as low as 0.539, comprehensively surpassing existing technologies. This method even outperforms wild-type on more challenging mutant datasets (R² = 0.905 vs 0.822). More importantly, when faced with a "rigorous test subset" consisting of enzymes or substrates never seen in the training set, the R² shows almost no decay, reaching an extremely high level of 0.871, demonstrating its powerful inference ability and great application potential in exploring novel chemical spaces. Employing a random forest model allows for the deconstruction of the model's decision-making basis through built-in feature importance analysis. Analysis shows that enzyme sequence features contribute over 88% of the decision weights. Furthermore, the prediction results of this method are highly consistent with known biological principles and can significantly distinguish between high-activity and low-activity mutants in specific cases.

[0042] Other features and advantages of the invention will be set forth in the description which follows, and will be apparent in part from the description, or may be learned by practicing the invention.

[0043] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description

[0044] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:

[0045] Figure 1This invention provides an example of an enzyme catalytic constant k based on multimodal feature fusion. cat Flowchart of the prediction method;

[0046] Figure 2 This is a schematic diagram of the model framework provided in the embodiments of the present invention;

[0047] Figure 3 To demonstrate the overall performance of R on independent test sets in this embodiment of the invention, 2 A density scatter plot of predicted and experimental values ​​with a value of 0.875;

[0048] Figure 4 (a) is a performance comparison analysis of wild-type (a) and mutant (b) subsets in the embodiments of the present invention, and (c) is a quantitative comparison of the core performance indicators of the two.

[0049] Figure 5 This is a cross-species generalization performance analysis diagram in an embodiment of the present invention;

[0050] Figure 6 This is a strict generalization ability analysis diagram in an embodiment of the present invention, where (a,b,c) compares R2, RMSE and MAE on the "seen", "unseen" and full test sets, respectively;

[0051] Figure 7 The biological consistency analysis diagram in the embodiments of the present invention includes (a) metabolic system analysis and (b) enzyme family case analysis;

[0052] Figure 8 This is a high-fidelity prediction scatter plot for a specific case in an embodiment of the present invention, in which 94.6% of the predicted points fall within the 95% prediction interval;

[0053] Figure 9 The model deconstruction analysis diagram in the embodiments of the present invention includes (a) feature importance analysis and (b) interval RMSE analysis;

[0054] Figure 10 This is an error distribution depth analysis diagram in an embodiment of the present invention, wherein (a) shows the distribution of the absolute prediction error in different kcat value ranges, and (b) shows the distribution of the absolute prediction error in different enzyme functional categories;

[0055] Figure 11 The diagram shows the hyperparameter optimization process and results in this embodiment of the invention. Detailed Implementation

[0056] The preferred embodiments of the present invention will be described below with reference to the accompanying drawings. It should be understood that the preferred embodiments described herein are for illustration and explanation only and are not intended to limit the present invention.

[0057] Example 1:

[0058] The embodiments of the present invention provide, as follows Figure 1 As shown, an enzyme catalytic constant k is derived based on the fusion of multimodal features. cat Prediction methods include:

[0059] S1: Obtain enzyme-substrate pairs, which include the amino acid sequence of the target enzyme and the chemical structure representation of the corresponding substrate;

[0060] S2: Input the amino acid sequence into the pre-trained protein language model. The protein language model performs average pooling on the hidden layer state vector of each amino acid residue in the sequence to obtain the enzyme feature vector used to characterize the functional features of the enzyme.

[0061] S3: Based on the chemical structure representation of the substrate, the extended connectivity chemical fingerprint algorithm is used to encode the chemical structure through the set radius parameter and bit length parameter to obtain the substrate feature vector used to characterize the structural features of the substrate;

[0062] S4: Concatenate the enzyme feature vector with the substrate feature vector to obtain the fused feature vector;

[0063] S5: Input the fused feature vector into the random forest regression model, and output the catalytic constant k of the enzyme-substrate pair. cat The predicted value.

[0064] The following is a detailed description with reference to specific embodiments.

[0065] I. Dataset Construction and Preprocessing

[0066] The data in this embodiment comes from the publicly available enzyme databases BRENDA and SABIO-RK. After integration and cleaning, a dataset containing the enzyme amino acid sequence, substrate SMILES string, and k... cat A comprehensive dataset containing experimental values, enzyme type (wild type / mutant), EC number, and species information. The raw data was processed as follows: k-values ​​were removed. cat Entries with values ​​that are not positive or whose SMILES structure is invalid; for k cat Perform a base-10 logarithmic transformation on the value (log 10 (k cat Each entry was labeled with its corresponding metabolic subsystem (e.g., core metabolism, major metabolism, etc.) based on its EC number; abnormal entries with an amino acid sequence length exceeding 2000 were filtered out. The processed complete dataset was randomly divided into a training set (80%), a validation set (10%), and a test set (10%).

[0067] S1: Obtain enzyme-substrate pairs, which include the amino acid sequence of the target enzyme and the chemical structure representation of the corresponding substrate;

[0068] The enzyme-substrate pairs to be predicted are obtained from the dataset. Each enzyme-substrate pair contains the amino acid sequence of the target enzyme and the chemical structure representation of the corresponding substrate. In this embodiment, the chemical structure representation of the substrate is in SMILES string format.

[0069] II. Feature Extraction

[0070] The feature extraction process in this embodiment is as follows: Figure 2 As shown in the "Featurization" module.

[0071] S2: Input the amino acid sequence into the pre-trained protein language model. The protein language model performs average pooling on the hidden layer state vector of each amino acid residue in the sequence to obtain the enzyme feature vector used to characterize the functional features of the enzyme.

[0072] S21: Input the amino acid sequence into the pre-trained protein language model to obtain the hidden layer state vector corresponding to each amino acid residue in the sequence;

[0073] For each enzyme amino acid sequence obtained in S1, a pre-trained ProtTrans-T5 model (specifically the ProtTrans-T5-XL-U50 model) is loaded using the bio_embeddings library. The amino acid sequence is input into this model to obtain the hidden layer state vector corresponding to each amino acid residue. In this embodiment, the hidden layer state vector for each amino acid residue has a dimension of 1024.

[0074] S22: Perform average pooling on the hidden layer state vector of each amino acid residue to compress the context dependency and long-range interaction information encoded in the sequence into a fixed-dimensional enzyme feature vector;

[0075] Mean pooling is performed on the hidden layer state vectors of all amino acid residues to compress and convert the contextual dependencies and long-range interactions encoded in the variable-length sequence into a fixed-dimensional enzyme feature vector that represents the entire enzyme function. The enzyme feature vector has a dimension of 1024 so that it can be concatenated with the subsequently generated substrate feature vector.

[0076] S3: Based on the chemical structure representation of the substrate, the extended connectivity chemical fingerprint algorithm is used to encode the chemical structure through the set radius parameter and bit length parameter to obtain the substrate feature vector used to characterize the structural features of the substrate;

[0077] S31: Resolve the chemical structure representation of the substrate into molecular structure data;

[0078] For each SMILES string (i.e. chemical structure representation) of a substrate obtained in S1, the RDKit library is used to parse it into molecular structure data for subsequent chemical fingerprinting.

[0079] S32: Set the radius and bit length parameters for the extended connectivity chemical fingerprint algorithm;

[0080] The radius parameter (radius) of the Extended Connectivity Chemical Fingerprint (ECFP) algorithm is set to 3, and the bit length parameter (nBits) is set to 2048. This parameter combination aims to ensure that local chemical substructure information crucial to enzyme catalytic activity is captured while avoiding excessive sparsity of the feature vector due to excessive bit length, thereby providing a substrate structure representation with optimal information density for subsequent machine learning regression models.

[0081] S33: Input the molecular structure data, radius parameters, and position length parameters into the extended connectivity chemical fingerprint algorithm for encoding operations to obtain the substrate feature vector;

[0082] Using the AllChem.GetMorganFingerprintAsBitVect function from the RDKit library, the molecular structure data parsed in S31, along with the radius and bit length parameters set in S32, are input into the extended connectivity chemical fingerprint algorithm for encoding operations. This process calculates the extended connectivity chemical fingerprint, ultimately yielding a substrate feature vector that characterizes the substrate's structural features. The substrate feature vector is a 2048-bit binary structural feature fingerprint vector.

[0083] III. Feature Fusion and Model Prediction

[0084] S4: Concatenate the enzyme feature vector with the substrate feature vector to obtain the fused feature vector;

[0085] The 1024-dimensional enzyme feature vector obtained in S2 and the 2048-dimensional substrate feature vector obtained in S3 are concatenated to form a 3072-dimensional fused feature vector. The fused feature vector is a multimodal feature vector formed by concatenating a continuous high-dimensional semantic vector (i.e., enzyme feature vector) for characterizing enzyme function and a sparse binary fingerprint vector (i.e., substrate feature vector) for characterizing substrate structure.

[0086] S5: Input the fused feature vector into the random forest regression model and output the predicted value of the catalytic constant kcat of the enzyme-substrate pair;

[0087] The fused feature vector obtained in S4 is input into a pre-trained random forest regression model to perform regression calculations, thereby outputting the predicted value of the catalytic constant kcat of the enzyme-substrate pair, achieving high-precision prediction of the kcat value.

[0088] The random forest regression model in S5 is a pre-trained model. The training process of this model includes:

[0089] Model selection: In this embodiment, RandomForestRegressor from the scikit-learn library is selected as the machine learning regression model.

[0090] S51: Based on the development dataset containing known logarithmic values ​​of catalytic constants, the K-fold cross-validation method is used to evaluate multiple hyperparameter combinations and obtain the performance evaluation results of each combination in cross-validation.

[0091] The logarithmic constants in the development dataset are logarithmic values ​​to the base 10, i.e., logarithmic. 10 (kcat). The training target of the random forest regression model is the logarithm of the catalytic constant, that is, during training, its prediction target is the logarithm of the enzyme catalytic constant kcat, specifically the logarithm with base 10.

[0092] To achieve optimal model performance, the training and validation sets were merged into a development set. On this development set, K-fold (K = 5) cross-validation was used to optimize the key hyperparameters of the random forest model. Multiple combinations of hyperparameters, including the number of decision trees (n_estimators) and the maximum tree depth (max_depth), were tested and evaluated to obtain the performance assessment results of each combination in cross-validation.

[0093] S52: Based on the performance evaluation results, determine the target hyperparameter combination from multiple hyperparameter combinations;

[0094] like Figure 11 As shown, the cross-validation results indicate that the performance of multiple parameter sets is robust and similar. Considering both the robustness of cross-validation and the final performance on independent test sets, an optimal hyperparameter combination (i.e., the target hyperparameter combination) is determined from the multiple hyperparameter combinations based on the performance evaluation results. In this embodiment, the target hyperparameter combination includes: a decision tree number n_estimators of 1500, and a maximum tree depth max_depth of 25 (corresponding to...). Figure 10 (run5 in the middle).

[0095] S53: Train the random forest regression model on the development dataset using the target hyperparameter combination to obtain the prediction model after training;

[0096] Using the above combination of target hyperparameters, the random forest regression model is trained on the complete development dataset (i.e., the original training set + validation set) to obtain the final prediction model after training, and saved as a .pkl file for subsequent prediction use.

[0097] IV. Performance Evaluation and Verification

[0098] The final model, trained using the aforementioned optimal parameters (n_estimators = 1500, max_depth = 25), was comprehensively evaluated on an independent test set, and the results are as follows. Figures 3 to 10 As shown.

[0099] Overall performance: such as Figure 3 As shown, the overall performance of this embodiment of the invention on an independent test set containing 1192 samples demonstrates R... 2 A scatter plot of the density of predicted and experimental values ​​with a value of 0.875. The model predictions are highly correlated with the experimental values, R0. 2 The performance is 0.875, and the root mean square error (RMSE) is 0.539, which comprehensively surpasses the existing technology.

[0100] Generalization ability: such as Figure 4 As shown, embodiments of the present invention are in the wild type ( Figure 4 (a) and mutants Figure 4 (b) Performance comparison analysis on subsets Figure 4 (c) A quantitative comparison of the core performance indicators of the two models. The model in the wild-type (R...) 2 =0.822) and more challenging mutants (R 2 The method performed excellently on all datasets with a mean value of 0.905. Furthermore, the method of this invention outperformed the wild-type dataset (R0.905) on even more challenging mutant datasets. 2 =0.905 vs 0.822).

[0101] like Figure 5 As shown in the cross-species generalization performance analysis diagram of this invention embodiment, the model maintains high-precision prediction across a variety of different species.

[0102] Following S5, a verification step is also included:

[0103] S6: Divide the test set containing known catalytic constant values ​​into seen subsets and unseen subsets according to whether they have appeared in the training set;

[0104] The test set is divided into "seen subsets" and "unseen subsets" based on whether the enzymes or substrates in them have appeared in the training set used to train the random forest regression model. The enzymes or substrates in the "unseen subsets" have never appeared in the training set used to train the machine learning regression model.

[0105] S7: Input the enzyme-substrate pairs in the seen subset into the random forest regression model, obtain the first prediction result, and calculate the first performance index based on the first prediction result and the known catalytic constant values ​​in the seen subset.

[0106] The enzyme-substrate pairs from the "seen subset" are input into a random forest regression model, regression calculations are performed, and the first prediction result is obtained. Based on the first prediction result and the known catalytic constant values ​​in the "seen subset," the first performance index (such as R) is calculated. 2 (e.g., RMSE) to evaluate the model's performance on existing data.

[0107] S8: Input the enzyme-substrate pairs in the unseen subset into the random forest regression model to obtain the second prediction result, and calculate the second performance index based on the second prediction result and the known catalytic constant values ​​in the unseen subset.

[0108] The enzyme-substrate pairs from the "unseen subset" are input into a random forest regression model, regression calculations are performed, and a second prediction result is obtained. Based on the second prediction result and the known catalytic constant values ​​from the "unseen subset," a second performance index (such as R0) is calculated. 2 RMSE and other metrics are used to verify the generalization ability of the model.

[0109] like Figure 6 As shown, this is a strict generalization ability analysis diagram of an embodiment of the present invention, wherein... Figure 6 (a, b, c) respectively compare R on the "seen", "unseen", and complete test sets. 2 RMSE and MAE. In the most stringent rigorous generalization tests, the model achieves R-values ​​on the "unseen subset". 2 With a value as high as 0.871 and almost no performance degradation, it demonstrates its powerful reasoning ability and great application potential in exploring new chemical spaces.

[0110] Biological consistency and case analysis: such as Figure 7 As shown, the biological consistency analysis diagram of this invention includes... Figure 6 (a) Metabolic system analysis and Figure 6 (b) Case Analysis of Enzyme Families. The model's predictions align with the higher biological principles that core metabolic enzymes exhibit higher activity, and it can accurately distinguish mutants with different activities within enzyme families. The prediction results of the method in this invention are highly consistent with known biological principles and can significantly differentiate between high-activity and low-activity mutants in specific cases.

[0111] like Figure 8 As shown, the high-fidelity prediction scatter plot of specific cases in this embodiment of the invention shows that, in the analysis of multiple specific cases, as many as 94.6% of the predicted points fell within the 95% statistical prediction interval, proving the high reliability of the model.

[0112] Model deconstruction and interpretability: such as Figure 9 As shown, the model deconstruction analysis diagram of this embodiment of the invention includes... Figure 9 (a) Feature importance analysis andFigure 9 (b) Interval RMSE analysis. Feature importance analysis showed that enzyme sequence features were the core driver of prediction, contributing over 88% of the decision weight. Error analysis revealed that the model was most accurate in predicting the intermediate activity range.

[0113] like Figure 10 As shown, this is a depth analysis diagram of the error distribution in an embodiment of the present invention, wherein... Figure 10 (a) shows the absolute prediction error at different k cat Distribution of value intervals Figure 10 (b) illustrates the distribution of absolute prediction error across different enzyme functional classes. The model exhibits the lowest and most stable prediction error in the medium activity range, while the error increases when predicting enzymes with extremely high activity. The model demonstrates universally applicable low-error prediction capability for the mainstream enzyme functional classes (EC1-3), but the uncertainty increases slightly when dealing with enzymes with more complex mechanisms (such as EC4-6). This analysis clearly reveals the advantageous range and potential optimization directions of the method of this invention.

[0114] Example 2:

[0115] This invention provides an embodiment of the enzyme catalytic constant k applied to multimodal feature fusion. cat Systems of prediction methods include:

[0116] The data acquisition module is used to acquire and temporarily store enzyme-substrate pair data containing the enzyme's amino acid sequence and the substrate's chemical structure representation;

[0117] The enzyme feature extraction module is equipped with a pre-trained ProtTrans-T5 protein language model, which is used to receive the enzyme amino acid sequence, perform hidden layer state vector acquisition and average pooling operation, and output an enzyme feature vector with a dimension of 1024.

[0118] The substrate feature extraction module is equipped with an extended connectivity chemical fingerprint algorithm with a radius parameter of 3 and a bit length parameter of 2048. It is used to receive the chemical structure representation of the substrate and output a substrate feature vector in 2048-bit binary fingerprint format.

[0119] The feature fusion module receives enzyme feature vectors and substrate feature vectors, performs vector concatenation operations, and outputs a fused feature vector with a dimension of 3072.

[0120] The prediction module is configured with a pre-trained random forest regression model, which receives the fused feature vector, performs regression calculations, and outputs the catalytic constant k. cat The predicted value.

[0121] The following is a detailed description with reference to specific embodiments.

[0122] This embodiment provides an application to the enzyme catalytic constant k based on multimodal feature fusion as described in Example 1. cat A system of prediction methods, comprising:

[0123] The data acquisition module is used to acquire and temporarily store enzyme-substrate pair data containing the enzyme's amino acid sequence and the substrate's chemical structure representation;

[0124] The data acquisition module is configured to retrieve enzyme-substrate pair data, containing the amino acid sequence of the target enzyme and the chemical structure representation of the corresponding substrate, from user input, a database, or a file, and temporarily store it in system memory for subsequent module processing. In this embodiment, the chemical structure representation of the substrate uses the SMILES string format.

[0125] The enzyme feature extraction module is equipped with a pre-trained ProtTrans-T5 protein language model, which is used to receive the enzyme amino acid sequence, perform hidden layer state vector acquisition and average pooling operation, and output an enzyme feature vector with a dimension of 1024.

[0126] The enzyme feature extraction module is equipped with a pre-trained ProtTrans-T5 protein language model (specifically, the ProtTrans-T5-XL-U50 model) to receive the enzyme amino acid sequence from the data acquisition module. This module first inputs the amino acid sequence into the ProtTrans-T5 model, performing the acquisition of the hidden layer state vector for each amino acid residue in the sequence, resulting in a 1024-dimensional hidden layer state vector for each amino acid residue. Then, it performs mean pooling on the hidden layer state vectors of all amino acid residues, compressing the contextual dependencies and long-range interactions encoded in the variable-length sequence into a fixed-dimensional enzyme feature vector. Finally, it outputs a 1024-dimensional enzyme feature vector that represents the entire enzyme function.

[0127] The substrate feature extraction module is equipped with an extended connectivity chemical fingerprint algorithm with a radius parameter of 3 and a bit length parameter of 2048. It is used to receive the chemical structure representation of the substrate and output a substrate feature vector in 2048-bit binary fingerprint format.

[0128] The substrate feature extraction module is equipped with an Extended Connectivity Chemical Fingerprinting (ECFP) algorithm, with a radius parameter set to 3 and a bit length parameter set to 2048. This module receives the substrate chemical structure representation (such as the string "SMILES") from the data acquisition module. First, the module uses the RDKit library to parse the "SMILES" string into molecular structure data. Then, it uses the `AllChem.GetMorganFingerprintAsBitVect` function to input the molecular structure data, along with the set radius and bit length parameters, into the ECFP algorithm for encoding and outputting a 2048-bit binary fingerprint format substrate feature vector. This parameter combination aims to ensure the capture of local chemical substructure information crucial to enzyme catalytic activity while avoiding excessively sparse feature vectors due to excessively large bit lengths.

[0129] The feature fusion module receives enzyme feature vectors and substrate feature vectors, performs vector concatenation operations, and outputs a fused feature vector with a dimension of 3072.

[0130] The feature fusion module receives the enzyme feature vector output by the enzyme feature extraction module and the substrate feature vector output by the substrate feature extraction module, and performs a vector concatenation operation to concatenate the 1024-dimensional enzyme feature vector and the 2048-dimensional substrate feature vector, outputting a fused feature vector with a dimension of 3072. This fused feature vector is a multimodal feature vector formed by concatenating a continuous high-dimensional semantic vector (enzyme feature vector) for characterizing enzyme function and a sparse binary fingerprint vector (substrate feature vector) for characterizing substrate structure.

[0131] The prediction module is configured with a pre-trained random forest regression model, which receives the fused feature vector, performs regression calculations, and outputs the catalytic constant k. cat The predicted value;

[0132] The prediction module is configured with a pre-trained random forest regression model, which receives the fused feature vector output by the feature fusion module, performs regression calculations, and outputs the catalytic constant k of the enzyme-substrate pair. cat The predicted value.

[0133] The random forest regression model was pre-trained according to the methods described in S51 to S53 of Example 1. Specifically, the model is based on logarithmic values ​​of known catalytic constants (log... 10 (k catUsing the development dataset, multiple hyperparameter combinations were evaluated using 5-fold cross-validation to determine the optimal hyperparameter combination (number of decision trees n_estimators = 1500, maximum tree depth max_depth = 25), and training was completed on the development dataset. The trained model was saved as a .pkl file and loaded into the prediction module.

[0134] Optionally, the system further includes a training module for training or updating the random forest regression model in the prediction module according to the method steps S51 to S53 in Embodiment 1. The training module can receive new training data, such as... Figure 11 As shown, the hyperparameter optimization and model training process is performed, and the trained model parameters are updated in the prediction module to maintain the prediction performance of the system.

[0135] Optionally, the system further includes a verification module for verifying the performance of the prediction module according to the method steps S6 to S8 in Embodiment 1. The verification module can divide the test set into a "seen subset" and a "unseen subset," and evaluate the model's performance metrics (such as R-squared) on the two subsets respectively. 2 Quantitative verification of the model's generalization ability is achieved through methods such as RMSE and MAE.

[0136] The overall workflow of the system in this embodiment is as follows: the data acquisition module acquires enzyme-substrate pair data → the enzyme feature extraction module and the substrate feature extraction module extract features in parallel → the feature fusion module splices the features → the prediction module outputs the kcat prediction value.

[0137] This embodiment system automates and modularizes the method described in Embodiment 1, enabling efficient and accurate determination of the enzyme catalytic constant k. cat The prediction task has good engineering application value.

[0138] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from the spirit and scope of this invention.

Claims

1. An enzyme catalytic constant k based on multimodal feature fusion cat The prediction method is characterized by, include: S1: Obtain enzyme-substrate pairs, which include the amino acid sequence of the target enzyme and the chemical structure representation of the corresponding substrate; S2: Input the amino acid sequence into the pre-trained protein language model. The protein language model performs average pooling on the hidden layer state vector of each amino acid residue in the sequence to obtain the enzyme feature vector used to characterize the functional features of the enzyme. S3: Based on the chemical structure representation of the substrate, the extended connectivity chemical fingerprint algorithm is used to encode the chemical structure through the set radius parameter and bit length parameter to obtain the substrate feature vector used to characterize the structural features of the substrate; S4: Concatenate the enzyme feature vector with the substrate feature vector to obtain the fused feature vector; S5: Input the fused feature vector into the random forest regression model, and output the catalytic constant k of the enzyme-substrate pair. cat The predicted value.

2. The enzyme catalytic constant k based on multimodal feature fusion as described in claim 1 cat The prediction method is characterized by, S2 include: S21: Input the amino acid sequence into the pre-trained protein language model to obtain the hidden layer state vector corresponding to each amino acid residue in the sequence; S22: Perform average pooling on the hidden layer state vector of each amino acid residue to compress the context dependency and long-range interaction information encoded in the sequence into a fixed-dimensional enzyme feature vector.

3. The enzyme catalytic constant k based on multimodal feature fusion according to claim 2 cat The prediction method is characterized by, The pre-trained protein language model is the ProtTrans-T5 model; the enzyme feature vector has a dimension of 1024.

4. The enzyme catalytic constant k based on multimodal feature fusion according to claim 1 cat The prediction method is characterized by S3 include: S31: Resolve the chemical structure representation of the substrate into molecular structure data; S32: Set the radius and bit length parameters for the extended connectivity chemical fingerprint algorithm; S33: Input the molecular structure data, radius parameters, and position length parameters into the extended connectivity chemical fingerprint algorithm for encoding operations to obtain the substrate feature vector.

5. The enzyme catalytic constant k based on multimodal feature fusion according to claim 4 cat The prediction method is characterized by, The radius parameter is 3, the bit length parameter is 2048; the substrate feature vector is a 2048-bit binary fingerprint vector.

6. The enzyme catalytic constant k based on multimodal feature fusion according to claim 1 cat The prediction method is characterized by, The random forest regression model in S5 is a pre-trained model. The training process of this model includes: S51: Based on the development dataset containing known logarithmic values ​​of catalytic constants, the K-fold cross-validation method is used to evaluate multiple hyperparameter combinations and obtain the performance evaluation results of each combination in cross-validation. S52: Based on the performance evaluation results, determine the target hyperparameter combination from multiple hyperparameter combinations; S53: Train the random forest regression model on the development dataset using the target hyperparameter combination to obtain the prediction model after training.

7. The enzyme catalytic constant k based on multimodal feature fusion according to claim 6 cat The prediction method is characterized by, K is 5; The target hyperparameter combination includes: the number of decision trees n_estimators is 1500, and the maximum depth of the trees max_depth is 25.

8. The method for predicting the enzyme catalytic constant kcat based on multimodal feature fusion according to claim 6, characterized in that, The logarithmic constants in the development dataset are logarithmic values ​​to the base 10, i.e., logarithmic. 10 (k cat ); The training objective of the random forest regression model is the logarithm of the catalytic constant.

9. The enzyme catalytic constant k based on multimodal feature fusion according to claim 1 cat The prediction method is characterized by, Following S5, a verification step is also included: S6: Divide the test set containing known catalytic constant values ​​into seen subsets and unseen subsets according to whether they have appeared in the training set; S7: Input the enzyme-substrate pairs in the seen subset into the random forest regression model, obtain the first prediction result, and calculate the first performance index based on the first prediction result and the known catalytic constant values ​​in the seen subset. S8: Input the enzyme-substrate pairs in the unseen subset into the random forest regression model to obtain the second prediction result, and calculate the second performance index based on the second prediction result and the known catalytic constant values ​​in the unseen subset.

10. Applied to the enzyme catalytic constant k based on multimodal feature fusion as described in any one of claims 1-9 cat The system of prediction methods is characterized by, include: The data acquisition module is used to acquire and temporarily store enzyme-substrate pair data containing the enzyme's amino acid sequence and the substrate's chemical structure representation; The enzyme feature extraction module is equipped with a pre-trained ProtTrans-T5 protein language model, which is used to receive the enzyme amino acid sequence, perform hidden layer state vector acquisition and average pooling operation, and output an enzyme feature vector with a dimension of 1024. The substrate feature extraction module is equipped with an extended connectivity chemical fingerprint algorithm with a radius parameter of 3 and a bit length parameter of 2048. It is used to receive the chemical structure representation of the substrate and output a substrate feature vector in 2048-bit binary fingerprint format. The feature fusion module receives enzyme feature vectors and substrate feature vectors, performs vector concatenation operations, and outputs a fused feature vector with a dimension of 3072. The prediction module is configured with a pre-trained random forest regression model, which receives the fused feature vector, performs regression calculations, and outputs the catalytic constant k. cat The predicted value.

Images