Enzyme turnover rate prediction method based on adaptive data enhancement and dual information mining
By employing adaptive data augmentation and dual information mining methods, the problems of insufficient dataset size and multimodal information utilization in enzyme turnover prediction are solved, achieving high accuracy and wide applicability of enzyme turnover prediction.
Patent Information
- Application Number
- CN202511244403.5
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-09-02
- Publication Date
- 2026-01-13
AI Technical Summary
Existing methods for predicting enzyme turnover rate suffer from problems such as limited dataset size, inconsistent standards, insufficient utilization of multimodal information, and unbalanced data distribution, resulting in poor prediction accuracy for extreme value intervals.
We employ an adaptive data augmentation and dual information mining approach. By constructing a standardized enzyme turnover dataset, we implement adaptive data augmentation, extract multimodal embedding features, construct a dual information mining module, and design an adaptive prediction network. Combined with a density-weighted loss optimization model, we achieve feature fusion and interaction between enzyme amino acid sequences and the linear input canonical sequences of simplified substrate molecules.
It significantly improves the accuracy and generalization ability of enzyme turnover prediction, especially in the extreme value range, and enhances the overall prediction accuracy and model robustness.
Smart Images
Figure CN121331237A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of enzyme kinetic parameter prediction technology, specifically to a multimodal deep learning method for predicting enzyme turnover rate based on enzyme amino acid sequences and simplified linear input canonical sequences of substrate molecules, integrating adaptive enhancement mechanisms and multimodal feature mining. Background Technology
[0002] Enzymes, as indispensable biocatalysts in life activities, play a vital role in many fields such as biosynthesis, drug development, and bioenergy. Enzyme kinetic parameters reflect various properties of enzymes in catalytic reactions, such as enzyme turnover (k). cat The Michaelis constant (Km) represents the maximum number of substrate molecules that a single enzyme molecule can convert per unit time, reflecting the catalytic efficiency of the enzyme. m The inhibition constant (K) represents the affinity of an enzyme for its substrate, defined as the substrate concentration required for the enzyme-catalyzed reaction to reach half of its maximum reaction rate. i () represents the dissociation constant of the enzyme-inhibitor complex formed by the binding of the inhibitor and the enzyme. Accurate determination of these enzyme kinetic parameters is crucial for processes such as drug design, optimized biosynthesis, and directed enzyme evolution. Among these kinetic parameters, k... cat It is a core indicator reflecting enzyme catalytic efficiency, and therefore extremely important. However, traditional methods for obtaining k... cat Values often require time-consuming and labor-intensive biological experiments to determine, making it difficult to cover a vast number of unknown or artificially designed enzyme proteins.
[0003] In recent years, with the development of deep learning technology, various deep learning models have made considerable progress in the field of enzyme data mining due to their powerful feature extraction and nonlinear modeling capabilities. Especially in the multimodal domain, various information mining and feature fusion methods have emerged. The excellent performance of these methods provides a basis for efficiently and accurately predicting k using deep learning models. cat The value reveals a very promising new direction.
[0004] The paper titled "Deep learning-based kcat prediction enables improved enzyme-constrained model reconstruction" innovatively proposes the DLKcat model. This model uses a convolutional neural network to extract enzyme protein sequence features and a graph neural network to extract substrate molecule features, enabling k... catValue prediction no longer relies on manually extracted features. The paper titled "UniKP: a unified framework for the prediction of enzyme kinetic parameters" proposes the UniKP framework. This framework extracts embedded features using a pre-trained protein language model and the SMILES Transformer model, and conducts extensive experiments on the concatenated features to select the optimal extreme random tree model for k-value prediction. cat Value prediction. The paper titled "Robust enzyme discovery and engineering with deep learning using CataPro" additionally introduces substrate molecular fingerprints as substrate features, constructing a robust CataPro framework to predict k-values. cat or K m Despite the above representative k cat Prediction methods have demonstrated some effectiveness, but there is still significant room for improvement. First, existing studies generally use datasets that are limited in size and inconsistent in standards, lacking a unified, high-quality dataset. Second, current methods primarily focus on the consistency information between enzyme and substrate modalities, neglecting the specific information within each modality. Finally, the experimentally measured k... cat Values naturally exhibit an unbalanced distribution, but existing methods do not pay enough attention to this, resulting in poor prediction accuracy for extreme value intervals.
[0005] How to solve the above-mentioned technical problems is the challenge facing this invention. Summary of the Invention
[0006] The purpose of this invention is to provide an enzyme turnover rate prediction method based on adaptive data augmentation and dual information mining, which solves the technical problems of unbalanced distribution of enzyme turnover rate data and insufficient utilization of multimodal information. It aims to improve the accuracy and generalization ability of multimodal deep learning models in enzyme turnover rate prediction tasks, and has broad application prospects in many fields such as biosynthesis, drug development, and bioenergy.
[0007] To achieve the aforementioned objectives, the present invention employs the following technical solution: a method for predicting enzyme turnover rate based on adaptive data augmentation and dual information mining, comprising the following steps:
[0008] S1: Construct a standardized enzyme turnover rate dataset, using enzyme protein sequences S e and substrate simplified molecule linear input canonical sequence S s As input data, the corresponding y-value is the labeled logarithmic enzyme turnover rate. 10 (k catValue; divide the dataset proportionally into training and training sets. and test set
[0009] S2: Implement adaptive data augmentation through kernel density estimation. Calculate the sample weights w(y) k Weighted sampling of enzyme protein sequences generates an enhancement set. And by performing similar amino acid substitutions, we obtained Screening non-standard simplified molecular linear input canonical sequences to generate substrate enhancement sets The final training set is formed by merging the augmentation set and the training set.
[0010] S3: Extract multimodal embedding features and use evolutionary scale modeling - 2 pre-trained model to process enzyme protein sequences S e Obtain the embedding matrix E e And the initial feature x is obtained through average pooling. e Using a molecular bidirectional encoding representation model to pre-train a model to process substrate sequence S s Obtain the embedding matrix E s And the initial feature x is obtained through average pooling. s Finally, the two types of features are grouped and their dimensions are transformed.
[0011] S4: Construct a dual-information mining module to extract preliminary enzyme protein features x using a specific encoder. e Corresponding specific feature c e and preliminary characteristics of the substrate x s Corresponding specific feature c s And apply domain classification loss to specific features. DCo Meanwhile, preliminary features of the enzyme protein were extracted using a shared encoder. e The corresponding shared feature r e and preliminary characteristics of the substrate x s The corresponding shared feature r s And apply a distribution alignment loss to the shared features. DAO Subsequently, cross-attention is used to fuse shared features and specific features and concatenate them to generate a comprehensive feature h. fuse ;
[0012] S5: Design an adaptive prediction network that integrates features h fuse Input adaptive density weight network to predict target value Combined density-weighted loss DenseWeight The overall objective function Loss is constructed using the two additional losses from dual information mining. Total Perform model optimization.
[0013] Further, step S2 includes the following steps:
[0014] Step S2.1: For enzyme proteins, fit the data distribution using the kernel density function:
[0015]
[0016] Where N is the total number of samples in the dataset, h is the bandwidth hyperparameter, and φ is the kernel function. That is, the kernel density estimate at point x;
[0017] Step S2.2: For any sample in the dataset Take its target value y k The complement of the kernel density is used as the enhanced sampling weight for this sample:
[0018]
[0019] Where w(y) k The target value y is... k The weights for augmentation are assigned to the corresponding samples. Based on the normalized sample weights, weighted sampling with replacement is performed to extract N′ = γ × N samples as the augmentation database. Where γ is a hyperparameter;
[0020] Step S2.3: For the database enzyme protein sequence Each element Determine whether the given amino acid substitution dictionary v contains the following: [[A,V],[S,T],[F,Y],[K,R],[C,M],[D,E],
[0021] [N,Q],[V,I],[T,S],[Y,F],[R,K],[M,C],[E,D],[Q,N],[I,V]]; if the element In dictionary v, the hyperparameter probability p = 0.05 will be used to... Replace with the corresponding amino acid in v to obtain the element. Conversely, if elements If it's not in dictionary v, it remains unchanged; for the database... After replacing all elements of each sequence as described above, the enzyme protein enhancement dataset is obtained.
[0022] Step S2.4: For the substrate, collect the linear input canonical sequence of the simplified molecule of the non-standard substrate and the corresponding enzyme protein sequence, k... cat Value samples build an enhanced database From enhancing the database Samples with a similarity higher than 99% to any enzyme protein sequence in the training set were selected and used as substrate sequence enhancement samples corresponding to these enzyme protein sequences. The resulting substrate enhancement dataset is as follows:
[0023]
[0024] in N represents a sample composed of the k-th enzyme protein sequence obtained through screening, the linear input canonical sequence of the substrate simplified molecule, and the target value. ″ The total number of enhanced samples selected;
[0025] Step S2.5: Take the union of the two augmented datasets and the initial training set as the final training set:
[0026]
[0027] Further, step S3 includes the following steps:
[0028] Step S3.1: For the enzyme protein sequence, extract its embedding representation using the Evolutionary Scale Modeling-2 (ESM-2) pre-trained model, given the input enzyme protein sequence S. e Let the ESM-2 model be F. esm If its embedding dimension is H, then:
[0029] E e =F esm (S e (26)of which The enzyme protein sequence S e The corresponding embedding feature matrix;
[0030] Step S3.2: For the substrate sequence, the embedding representation is extracted using a pre-trained model of the Mole-BERT bidirectional encoding representation model, given the input substrate sequence S. s Let the Mole-BERT model be F. mol If its embedding dimension is H′, then:
[0031] E s =F mol (S s (27)
[0032] in To simplify the linear input canonical sequence S of the substrate molecule s The corresponding embedding feature matrix;
[0033] Step S3.3: Apply average pooling to the embedding feature matrices of the two modalities to embed the variable-length enzyme protein sequence into the feature matrix E. e And the substrate simplified molecule linear input canonical sequence embedding feature matrix Es Transform into a feature vector of fixed length and For a dataset containing N samples, after embedding extraction and average pooling, the two-dimensional feature matrix of the enzyme protein sequence of the entire dataset is obtained. And substrate simplified molecule linear input canonical sequence two-dimensional feature matrix
[0034] Step S3.4: Convert the two-dimensional feature matrix X of the enzyme protein sequence e The two-dimensional feature matrix X of the simplified linear input canonical sequence of the substrate molecule. s The sequence is divided into groups G and G′ based on its feature dimension, thereby transforming it into a three-dimensional tensor of the enzyme protein sequence. And substrate simplified molecule linear input canonical sequence three-dimensional tensor After grouping, the feature vectors of the two modes described in step S3.3 are respectively transformed into as well as
[0035] Further, step S4 includes the following steps:
[0036] Step S4.1: Analyze the enzyme protein sequence feature matrix x extracted in step S3. e and substrate simplified molecule linear input canonical sequence feature matrix x s Construct two self-attention modules respectively. and Feature extraction as a specific encoder:
[0037]
[0038] in These are the learnable projection parameter matrices for Query, Key, and Value in the self-attention module, respectively. head is a hyperparameter, representing the projection dimension. From this, we can obtain the specific feature c of the enzyme protein modality. e With substrate mode-specific features c s ;
[0039] Step S4.2: Specific characteristics of the enzyme protein c e With substrate-specific features c s An independent binary classifier F is constructed using a fully connected layer and a sigmoid function. DCO , for F DCO A binary cross-entropy loss function is constructed as the domain classification loss function to enhance the specificity of features for each modality:
[0040]
[0041]
[0042] Step S4.3: Analyze the enzyme protein modal feature matrix x extracted in step S3. e and substrate modal characteristic matrix x s Then construct two self-attention modules respectively. and Perform deep information extraction:
[0043]
[0044] in Intermediate features calculated for enzyme protein modalities. Intermediate features calculated for the substrate modes;
[0045] Step S4.4: Using intermediate features of enzyme proteins As the Query in the attention mechanism, the intermediate features of the substrate As the key and value in the attention mechanism, an enzyme protein-substrate attention (ESA) module is constructed to extract shared interaction information:
[0046]
[0047] Where r e The extracted enzyme proteins share common characteristics;
[0048] Step S4.5: Using intermediate features of the substrate As a query in the attention mechanism, intermediate features of enzyme proteins As the Key and Value in the attention mechanism, a substrate-enzyme protein attention (SEA) module is constructed to extract shared interaction information:
[0049]
[0050] Where r s For the extracted substrate shared features;
[0051] Step S4.6: For enzyme-protein shared characteristics r e Shared features with substrate r s Constructing a distribution alignment loss to enhance common information mining:
[0052]
[0053] Step S4.7: Specific characteristics of the enzyme protein c e With substrate-specific features c s and enzyme protein shared characteristics r e Shared features with substrate r sCross-attention modules are used to achieve the fusion of shared and characteristic information:
[0054]
[0055] in This represents the high-level semantic features of enzyme proteins after projection fusion. This represents the high-level semantic features of the substrate after projection fusion.
[0056] Step S4.8: Concatenate the two high-level semantic features obtained in step S4.7 and introduce a residual connection operation to obtain the final output feature:
[0057]
[0058] in This represents a vector concatenation operation.
[0059] Further, step S5 includes the following steps:
[0060] Step S5.1: Combine the comprehensive feature h obtained in step S4 fuse Input to feedforward neural network F ffn To predict the final target value log 10 (K cat ):
[0061]
[0062] Step S5.2: Use density-weighted loss to balance the loss weights of different samples during training:
[0063]
[0064] Where f w (ω,y k ) is the weight calculation function based on kernel density value, used to balance the loss value of mean square error, and ω is a hyperparameter used to control the unbalanced weighting intensity;
[0065] Step S5.3: Combining the density weight loss, the distribution alignment loss, and the domain classification loss from step S4, construct the total loss function:
[0066] Loss Total =Loss DenseWeight +α×Loss DAO +β×Loss DCO (42)
[0067] Here, α and β are hyperparameters used to balance the contributions of the two additional objective loss functions.
[0068] Compared with the prior art, the beneficial effects of the present invention are as follows:
[0069] 1. Effective Mining of Multimodal Information: An enzyme turnover prediction method based on adaptive data augmentation and dual-information mining can simultaneously process two modalities of data: enzyme amino acid sequences and simplified linear input canonical sequences of substrate molecules. Its core dual-information mining module effectively decouples and utilizes common and characteristic information from enzyme-substrate multimodal data, achieving more comprehensive and deeper multimodal information fusion and interaction, significantly improving feature representation capabilities.
[0070] 2. Enhancing Prediction Performance in Extreme Value Ranges: The enzyme turnover rate prediction method based on adaptive data augmentation and dual information mining integrates an adaptive data augmentation module and an adaptive density weight loss to address the challenge of the inherently highly imbalanced distribution of enzyme turnover rate data. Adaptive data augmentation specifically expands the sample distribution in low-density regions, mitigating the bias caused by data imbalance. The adaptive density weight loss assigns higher learning weights to extreme value samples, guiding the model to pay more attention to the prediction accuracy of these regions during optimization. The synergistic effect of both significantly improves the model's prediction accuracy and overall robustness in extreme value ranges of catalytic efficiency.
[0071] 3. Improved Overall Prediction Accuracy: The enzyme turnover rate prediction method based on adaptive data augmentation and dual information mining significantly improves overall prediction accuracy by combining the powerful feature extraction capabilities of a pre-trained language model, the deep utilization of common and specific information by the dual information mining module, and the effective handling of data imbalance by adaptive data augmentation and density weight loss. Experiments have verified that its prediction performance outperforms existing advanced methods.
[0072] 4. Enhanced Model Generalization Ability: The enzyme turnover rate prediction method based on adaptive data augmentation and dual information mining incorporates domain classification loss and distribution alignment loss in dual information mining, which helps strengthen the encoder's feature extraction process. Combined with effective preservation of modal characteristics and adaptive mechanisms to correct for data distribution imbalance, this invention exhibits excellent generalization ability. Not only does it perform superiorly on general test sets, but its generalization ability is significantly better than existing methods when dealing with sparsely distributed and highly variable extreme catalytic efficiency samples, thus providing a reliable tool for covering a wider range of enzyme-catalyzed reaction scenarios in practical applications. Attached Figure Description
[0073] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used to explain the invention but do not constitute a limitation thereof.
[0074] Figure 1 This is an overall flowchart of the enzyme turnover rate prediction method based on adaptive data augmentation and dual information mining of the present invention.
[0075] Figure 2 This is a complete framework diagram of the enzyme turnover rate prediction method based on adaptive data augmentation and dual information mining of the present invention;
[0076] Figure 3 This is a comparison chart of the evaluation metrics of this invention with other models. Detailed Implementation
[0077] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. Of course, the specific embodiments described herein are merely illustrative and not intended to limit the invention.
[0078] Example 1
[0079] See Figures 1 to 2 This embodiment provides a technical solution: an enzyme turnover rate prediction method based on adaptive data augmentation and dual information mining, comprising the following steps:
[0080] S1: Data was collected from multiple public sources to construct a standardized enzyme turnover dataset. After cleaning, standardization, and rigorous deduplication, a final dataset containing 32,350 entries was obtained. This dataset covers 2,581 different enzyme EC numbers and 4,843 different substrate sequences. Enzyme protein sequences were used as the basis for the dataset. e and substrate simplified molecule linear input canonical sequence S s As input data, the corresponding y-value is the labeled logarithmic enzyme turnover rate. 10 (k cat Value; divide the dataset into training and training sets in a 4:1 ratio. and test set We obtained 25,880 training samples and 6,470 test samples.
[0081] S2: Implement adaptive data augmentation through kernel density estimation. Calculate the sample weights w(y) k Weighted sampling of enzyme protein sequences generates an enhancement set. And by performing similar amino acid substitutions, we obtained Ultimately, 7764 enhancement entries were obtained; simultaneously, non-standard simplified molecular linear input canonical sequences were screened to generate substrate enhancement sets. A total of 16,155 augmentation entries were obtained; the augmentation set and the training set were merged to form the final training set D′. train It contains a total of 49,799 entries;
[0082] S3: Extract multimodal embedding features and use evolutionary scale modeling - 2 pre-trained model to process enzyme protein sequences S e Obtain the embedding matrix Ee And the initial feature x is obtained through average pooling. e Simultaneously, a pre-trained model using a molecular bidirectional encoding representation model is employed to process the substrate sequence S. s Obtain the embedding matrix E s And the initial feature x is obtained through average pooling. s Finally, grouping and dimensional transformations are performed on the two types of features respectively;
[0083] S4:
[0084] A dual-information mining module is constructed to extract preliminary features x of enzyme proteins through a specific encoder. e Corresponding specific feature c e and preliminary characteristics of the substrate x s Corresponding specific feature c s And apply domain classification loss to specific features. DCO Meanwhile, preliminary features of the enzyme protein were extracted using a shared encoder. e The corresponding shared feature r e and preliminary characteristics of the substrate x s The corresponding shared feature r s And apply a distribution alignment loss to the shared features. DAO Subsequently, cross-attention is used to fuse shared features and specific features and concatenate them to generate a comprehensive feature h. fuse ;
[0085] S5: Design an adaptive prediction network that integrates features h fuse Input adaptive density weight network to predict target value Combined density-weighted loss DenseWeight The overall objective function Loss is constructed using the two additional losses from dual information mining. Total Perform model optimization.
[0086] Specifically, step S2 includes the following steps:
[0087] Step S2.1: For enzyme proteins, fit the data distribution using the kernel density function:
[0088]
[0089] Where N = 25880 is the total number of samples in the dataset, h = 0.5 is the bandwidth hyperparameter, and φ is selected using the Gaussian kernel function. This is the estimated kernel density at point x.
[0090] Step S2.2: For any sample in the dataset Take its target value y k The complement of the kernel density is used as the enhanced sampling weight for this sample:
[0091]
[0092] Where w(y) k The target value y is... k Weights are assigned to the corresponding samples for augmentation. Based on the normalized sample weights, weighted sampling with replacement is performed to extract N′=γ×N=7764 samples as the augmentation database. Where γ = 0.3 is a hyperparameter.
[0093] Step S2.3: For the database enzyme protein sequence Each element Determine whether the given amino acid substitution dictionary V contains the following: [[A,V],[S,T],[F,Y],[K,R],[C,M],[D,E],
[0094] [N,Q],[V,I],[T,S],[Y,F],[R,K],[M,C],[E,D],[Q,N],[I,V]]. If element In dictionary v, the hyperparameter probability p = 0.05 will be used to... Replace with the corresponding amino acid from dictionary v to obtain the element. Conversely, if elements If it's not in dictionary v, it remains unchanged. (Regarding the database...) After replacing all elements of each sequence as described above, the enzyme protein enhancement dataset is obtained. It contains a total of 7,764 entries.
[0095] Step S2.4: For the substrate, collect the linear input canonical sequence of the simplified molecule of the non-standard substrate and the corresponding enzyme protein sequence, k... cat Value samples build an enhanced database It contains a total of 20,256 data entries. (From the enhanced database) Samples with a similarity higher than 99% to any enzyme protein sequence in the training set were selected and used as substrate enhancement samples for these enzyme protein sequences. The resulting substrate enhancement dataset is as follows:
[0096]
[0097] in N represents a sample composed of the k-th enzyme protein sequence obtained through screening, the linear input canonical sequence of the substrate simplified molecule, and the target value. ″ =16155 represents the total number of enhanced samples selected.
[0098] Step S2.5: Take the union of the two augmented datasets and the initial training set as the final training set:
[0099]
[0100] in It contains a total of 49,799 entries.
[0101] Specifically, step S3 includes the following steps:
[0102] Step S3.1: For the enzyme protein sequence, extract its embedding representation using an evolutionary scale modeling-2 (ESM-2) pre-trained model. Given the input enzyme protein sequence S... e Let the ESM-2 model be F. esm If its embedding dimension is H = 1280, then:
[0103] E e =F esm (S e (47)
[0104] in The enzyme protein sequence S e The corresponding embedding feature matrix.
[0105] Step S3.2: For the substrate sequence, extract its embedding representation using a pre-trained Mole-BERT (Mole-BERT) model. Given the input substrate sequence S... s Let the Mole-BERT model be F. mol If its embedding dimension is H′=300, then:
[0106] E s =F mol (S s (48)
[0107] in To simplify the linear input canonical sequence S of the substrate molecule s The corresponding embedding feature matrix.
[0108] Step S3.3: Apply average pooling to the embedding feature matrices of the two modalities to embed the variable-length enzyme protein sequence into the feature matrix E. e And the substrate simplified molecule linear input canonical sequence embedding feature matrix E s This is transformed into a fixed-length enzyme protein feature vector. and substrate feature vector For a dataset containing 49,799 samples, after embedding extraction and average pooling, the two-dimensional feature matrix of the enzyme protein sequence, the substrate simplified molecule linear input canonical sequence, and the two-dimensional feature matrix of the entire dataset can be obtained. and
[0109] Step S3.4: Convert the two-dimensional feature matrix X of the enzyme protein sequence e The two-dimensional feature matrix X of the simplified linear input canonical sequence of the substrate molecule. s The sequence was divided into 32 groups along the feature dimension, thus transforming it into a three-dimensional tensor of the enzyme protein sequence. And substrate simplified molecule linear input canonical sequence three-dimensional tensor This facilitates subsequent attention calculations. After grouping, the feature vectors of the two modalities described in step S3.3 are respectively transformed into... as well as
[0110] Specifically, step S4 includes the following steps:
[0111] Step S4.1: Analyze the enzyme protein sequence feature matrix x extracted in step S3. e and substrate simplified molecule linear input canonical sequence feature matrix x s Construct two self-attention modules respectively. and Feature extraction as a specific encoder:
[0112]
[0113] in These are the learnable projection parameter matrices for Query, Key, and Value in the self-attention module, respectively. head is a hyperparameter representing the projection dimension, automatically determined by PyTorch. From this, the specific features c of the enzyme protein modality can be obtained. e With substrate mode-specific features c s ;
[0114] Step S4.2: Specific characteristics of the enzyme protein c e With substrate-specific features c s An independent binary classifier c is constructed using a fully connected layer and a sigmoid function. DCO For F DCO A binary cross-entropy loss function is constructed as the domain classification loss function to enhance the specificity of features for each modality:
[0115]
[0116] Step S4.3: Analyze the enzyme protein modal feature matrix x extracted in step S3. e and substrate modal characteristic matrix x s Then construct two self-attention modules respectively. and Perform deep information extraction:
[0117]
[0118] in Intermediate features calculated for enzyme protein modalities. Intermediate features calculated for substrate modes; Step S4.4: Using intermediate features of enzyme proteins As the Query in the attention mechanism, the intermediate features of the substrate As the key and value in the attention mechanism, an enzyme protein-substrate attention (ESA) module is constructed to extract shared interaction information:
[0119]
[0120] Where r e The extracted enzyme proteins share common characteristics.
[0121] Step S4.5: Using intermediate features of the substrate As a query in the attention mechanism, intermediate features of enzyme proteins As the Key and Value in the attention mechanism, a substrate-enzyme protein attention (SEA) module is constructed to extract shared interaction information:
[0122]
[0123] Where r s The extracted substrate-shared features.
[0124] Step S4.6: For enzyme-protein shared characteristics r e Shared features with substrate r s Constructing a distribution alignment loss to enhance common information mining:
[0125]
[0126] Step S4.7: Specific characteristics of the enzyme protein c e With substrate-specific features c s and enzyme protein shared characteristics r e Shared features with substrate r s Cross-attention modules are used to achieve the fusion of shared and characteristic information:
[0127]
[0128] in This represents the high-level semantic features of enzyme proteins after projection fusion. This represents the high-level semantic features of the substrate after projection fusion.
[0129] Step S4.8: Concatenate the two high-level semantic features obtained in step S4.7 and introduce a residual connection operation to obtain the final output feature:
[0130]
[0131] in This represents a vector concatenation operation.
[0132] Specifically, step S5 includes the following steps:
[0133] Step S5.1: Combine the comprehensive feature h obtained in step S4 fuse Input to feedforward neural network F ffn To predict the final target value log 10 (K cat ):
[0134]
[0135] Step S5.2: To further optimize the imbalanced data problem, density-weighted loss is used to balance the loss weights of different samples during training:
[0136]
[0137] Where f w (ω,y k ) is the weighting function based on kernel density values, used to balance the loss value of mean square error. ω = 0.6 is a hyperparameter used to control the intensity of unbalanced weighting.
[0138] Step S5.3: Combining the density weight loss, the distribution alignment loss, and the domain classification loss from step S4, construct the total loss function:
[0139] Loss Total =Loss DenseWeight +α×Loss DAO +β×Loss DCO (63)
[0140] Here, α = 0.06 and β = 0.06 are hyperparameters used to balance the contributions of the two additional objective loss functions.
[0141] Example 2
[0142] Referring to Example 1, this example will use the parameters calculated in Example 1, and the comparison results with other algorithms on a unified test set are as follows:
[0143] 1. Traditional Model
[0144] Based on relevant research both domestically and internationally, the following comparative model was selected for this embodiment:
[0145] 1) DLKcat: This model uses word segmentation embedding and convolutional neural networks to extract enzyme protein sequence features, and graph neural networks to extract substrate features. Finally, it combines the features of both to predict the k of the enzyme-catalyzed reaction. cat value.
[0146] 2) CataPro: This model utilizes ProtT5 and MolT5 pre-trained language models to extract enzyme protein sequence and substrate sequence features, respectively, and additionally incorporates MACCS molecular fingerprint information. Subsequently, the three modalities are merged for k-factor analysis. cat Value prediction.
[0147] 3) CatPred: This model utilizes rotational position embedding, evolutionary scale modeling-2 pre-trained language model to extract enzyme protein sequence features, and message passing network to extract substrate molecular graph features. After merging the features, it predicts k using a probabilistic regression model. cat value.
[0148] 2. Comparison Indicators
[0149] This embodiment uses three commonly used regression evaluation indicators: the coefficient of determination R. 2 The root mean square error (RMSE) and mean absolute error (MAE) are used to evaluate the model's prediction results. Let y be the true value. The formulas for the three indicators are as follows, representing the predicted values:
[0150]
[0151] Each of the above evaluation indicators has its unique role and limitations. This embodiment uses a comprehensive evaluation method to objectively compare the present invention with other models, aiming to highlight the superiority of the present invention.
[0152] 3. Comparison Results
[0153] Each model was retrained using a unified training set and evaluated on the same test set. The results are shown in Table 1 and... Figure 3 As shown.
[0154] Table 1 Comparison of the present invention with other models
[0155]
[0156] From Table 1 and Figure 3 The data shows that this invention performs excellently across all evaluation metrics, significantly outperforming other models. This invention achieves an R-value of 0.602. 2The coefficient of determination, RMSE, and MAE error indices were 0.987 and 0.668, respectively, representing improvements of 4.6%, 3.1%, and 6.9% compared to the second-best method, CatPred. Furthermore, the indices of this invention are superior to the CataPro method and significantly superior to the DLKcat method. This indicates that the present invention has higher overall prediction accuracy and can more accurately estimate the turnover rate of enzymatic reactions.
[0157] In summary, this invention can extract previously overlooked characteristic information from enzyme and substrate modalities and perform deep interactions on the obtained features, thereby improving the model's predictive performance. Furthermore, experimental results show that DLKcat, due to its simple feature extraction process and framework, failed to achieve good regression results. This further demonstrates the importance of constructing effective feature extraction and interaction modules for capturing the interactions between enzyme proteins and substrates in enzymatic reactions. Overall, this invention significantly outperforms other models on several key evaluation metrics, demonstrating its superior performance and generalization ability in regression tasks.
[0158] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.
Claims
1. A method for predicting enzyme turnover rate based on adaptive data augmentation and dual information mining, characterized in that, Includes the following steps: S1: Construct a standardized enzyme turnover rate dataset, using enzyme protein sequences S e and substrate simplified molecule linear input canonical sequence S s As input data, the corresponding y-value is the labeled logarithmic enzyme turnover rate. 10 (k cat Value; divide the dataset proportionally into training and training sets. and test set S2: Implement adaptive data augmentation through kernel density estimation. Calculate the sample weights w(y) k Weighted sampling of enzyme protein sequences generates an enhancement set. And by performing similar amino acid substitutions, we obtained Screening non-standard simplified molecular linear input canonical sequences to generate substrate enhancement sets The final training set is formed by merging the augmentation set and the training set. S3: Extract multimodal embedding features and use evolutionary scale modeling - 2 pre-trained model to process enzyme protein sequences S e Obtain the embedding matrix E e And the initial feature x is obtained through average pooling. e Using a molecular bidirectional encoding representation model to pre-train a model to process substrate sequence S s Obtain the embedding matrix E s And the initial feature x is obtained through average pooling. s Finally, the two types of features are grouped and their dimensions are transformed. S4: Construct a dual-information mining module to extract preliminary enzyme protein features x using a specific encoder. e Corresponding specific feature c e and preliminary characteristics of the substrate x s Corresponding specific feature c s And apply domain classification loss to specific features. DCO Simultaneously, preliminary features of the enzyme protein were extracted using a shared encoder. e The corresponding shared feature r e and preliminary characteristics of the substrate x s The corresponding shared feature r s And apply a distribution alignment loss to the shared features. DAO Subsequently, cross-attention is used to fuse shared features and specific features and concatenate them to generate a comprehensive feature h. fuse ; S5: Design an adaptive prediction network that integrates features h fuse Input adaptive density weight network to predict target value Combined density-weighted loss DenseWeight The overall objective function Loss is constructed using the two additional losses from dual information mining. Total Perform model optimization.
2. The enzyme turnover rate prediction method based on adaptive data augmentation and dual information mining according to claim 1, characterized in that, Step S2 includes the following steps: Step S2.1: For enzyme proteins, fit the data distribution using the kernel density function: Where N is the total number of samples in the dataset, h is the bandwidth hyperparameter, and φ is the kernel function. This is the kernel density estimate at point x; Step S2.2: For any sample in the dataset Take its target value y k The complement of the kernel density is used as the enhanced sampling weight for this sample: Where w(y) k ) represents the target value y k The weights for augmentation are assigned to the corresponding samples. Based on the normalized sample weights, weighted sampling with replacement is performed to extract N′ = γ × N samples as the augmentation database. Where γ is a hyperparameter; Step S2.3: For the database Middle enzyme protein sequence Each element Determine if it is in a given amino acid substitution dictionary In the dictionary For: [[A,V],[S,T],[F,Y],[K,R],[C,M],[D,E], [N,Q],[V,I],[T,S],[Y,F],[R,K],[M,C],[E,D],[Q,N],[I,V]]; if the element In dictionary v, the hyperparameter probability p = 0.05 will be used to... Replace with dictionary The corresponding amino acids are obtained from the elements. Conversely, if elements Not in the dictionary In the middle, it remains unchanged; for the database After replacing all elements of each sequence as described above, the enzyme protein enhancement dataset is obtained. Step S2.4: For the substrate, collect the linear input canonical sequence of the simplified molecule of the non-standard substrate and the corresponding enzyme protein sequence, k... cat Value samples build an enhanced database From enhancing the database Samples with a similarity higher than 99% to any enzyme protein sequence in the training set were selected and used as substrate sequence enhancement samples corresponding to these enzyme protein sequences. The resulting substrate enhancement dataset is as follows: in N″ represents the sample consisting of the kth enzyme protein sequence obtained through screening, the linear input canonical sequence of the substrate simplified molecule, and the target value, where N″ is the total number of enhanced samples selected. Step S2.5: Take the union of the two augmented datasets and the initial training set as the final training set:
3. The enzyme turnover rate prediction method based on adaptive data augmentation and dual information mining according to claim 1, characterized in that, Step S3 includes the following steps: Step S3.1: For the enzyme protein sequence, extract its embedding representation using an evolutionary scale modeling-2 pre-trained model, given the input enzyme protein sequence S. e Let the ESM-2 model be F. esm If its embedding dimension is H, then: E e =F esm (S e ) (5) in The enzyme protein sequence S e The corresponding embedding feature matrix; Step S3.2: For the substrate sequence, the Mole-BERT pre-trained model is used to extract its embedding representation, given the input substrate sequence S. s Let the Mole-BERT model be F mol If its embedding dimension is H′, then: E s =F mol (S s ) (6) in To simplify the linear input canonical sequence S of the substrate molecule s The corresponding embedding feature matrix; Step S3.3: Apply average pooling to the embedding feature matrices of the two modalities to embed the variable-length enzyme protein sequence into the feature matrix E. e And the substrate simplified molecule linear input canonical sequence embedding feature matrix E s This is transformed into a fixed-length enzyme protein feature vector. and substrate feature vector For a dataset containing N samples, after embedding extraction and average pooling, the two-dimensional feature matrix of the enzyme protein sequence of the entire dataset is obtained. And substrate simplified molecule linear input canonical sequence two-dimensional feature matrix Step S3.4: Convert the two-dimensional feature matrix X of the enzyme protein sequence e The two-dimensional feature matrix X of the simplified linear input canonical sequence of the substrate molecule. s The sequence is divided into groups G and G′ based on its feature dimension, thereby transforming it into a three-dimensional tensor of the enzyme protein sequence. And substrate simplified molecule linear input canonical sequence three-dimensional tensor After grouping, the feature vectors of the two modes described in step S3.3 are respectively transformed into as well as 4. The enzyme turnover rate prediction method based on adaptive data augmentation and dual information mining according to claim 1, characterized in that, Step S4 includes the following steps: Step S4.1: Analyze the enzyme protein sequence feature matrix x extracted in step S3. e and substrate simplified molecule linear input canonical sequence feature matrix x s Construct two self-attention modules respectively. and Feature extraction as a specific encoder: in These are the learnable projection parameter matrices for Query, Key, and Value in the self-attention module, respectively. head Here, is a hyperparameter representing the projection dimension, from which the specific feature c of the enzyme protein mode is obtained. e With substrate mode-specific features c s ; Step S4.2: Specific characteristics of the enzyme protein c e With substrate-specific features c s An independent binary classifier F is constructed using a fully connected layer and a sigmoid function. DCO , for F DCO A binary cross-entropy loss function is constructed as the domain classification loss function to enhance the specificity of features for each modality: Step S4.3: Analyze the enzyme protein modal feature matrix x extracted in step S3. e and substrate modal characteristic matrix x s Then construct two self-attention modules respectively. and Perform deep information extraction: in Intermediate features calculated for enzyme protein modalities. Intermediate features calculated for the substrate modes; Step S4.4: Using intermediate features of enzyme proteins As the Query in the attention mechanism, the intermediate features of the substrate As the Key and Value in the attention mechanism, an enzyme-protein-substrate attention ESA module is constructed to extract shared interaction information: Where r e The extracted enzyme proteins share common characteristics; Step S4.5: Using intermediate features of the substrate As a query in the attention mechanism, intermediate features of enzyme proteins As the Key and Value in the attention mechanism, a substrate-enzyme protein attention SEA module is constructed to extract shared interaction information: Where r s For the extracted substrate shared features; Step S4.6: For enzyme-protein shared characteristics r e Shared features with substrate r s Constructing a distribution alignment loss to enhance common information mining: Step S4.7: Specific characteristics of the enzyme protein c e With substrate-specific features c s and enzyme protein shared characteristics r e Shared features with substrate r s Cross-attention modules are used to achieve the fusion of shared and characteristic information: in This represents the high-level semantic features of enzyme proteins after projection fusion. This represents the high-level semantic features of the substrate after projection fusion. Step S4.8: Concatenate the two high-level semantic features obtained in step S4.7 and introduce a residual connection operation to obtain the final output feature: in This represents a vector concatenation operation.
5. The enzyme turnover rate prediction method based on adaptive data augmentation and dual information mining according to claim 1, characterized in that, Step S5 includes the following steps: Step S5.1: Combine the comprehensive feature h obtained in step S4 fuse Input to feedforward neural network F ffn To predict the final target value log 10 (K cat ): Step S5.2: Use density-weighted loss to balance the loss weights of different samples during training: Where f w (ω,y k ) is the weight calculation function based on kernel density value, used to balance the loss value of mean square error, and ω is a hyperparameter used to control the unbalanced weighting intensity; Step S5.3: Combining the density weight loss, the distribution alignment loss, and the domain classification loss from step S4, construct the total loss function: Loss Total =Loss DenseWeight +α×Loss DAO +β×Loss DCO (21) Here, α and β are hyperparameters used to balance the contributions of the two additional objective loss functions.