Method for predicting enzyme turnover number based on deep learning model and protein structure information
By constructing an enzyme sequence-enzyme structure-substrate combination dataset and combining it with deep learning models, especially graph convolutional neural networks and Transformer models, the problem of low accuracy in enzyme turnover prediction was solved, achieving higher prediction accuracy and resistance to sequence similarity.
Patent Information
- Application Number
- CN202311512721.6
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2023-11-14
- Publication Date
- 2025-12-05
- Estimated Expiration
- 2043-11-14
AI Technical Summary
Existing deep learning models have low accuracy in predicting enzyme turnover and fail to effectively utilize the protein structure information of enzymes, resulting in a significant impact from sequence similarity.
We constructed a dataset mapping enzyme sequence-enzyme structure-substrate combination and enzyme turnover, combined it with deep learning models, especially graph convolutional neural networks and Transformer models, and incorporated a multi-head attention mechanism to make predictions using protein structure information.
It significantly improves the accuracy of enzyme turnover prediction, reduces the negative impact of sequence similarity, and can predict the impact of point mutations on enzyme turnover.
Smart Images
Figure CN120015107B_ABST
Abstract
Description
TECHNICAL FIELD
[0001] The present application relates to the technical field of bioengineering, and particularly relates to an enzyme turnover number prediction method based on a deep learning model and protein structure information. BACKGROUND
[0002] Accurate prediction of enzyme turnover number is crucial for protein rational modification and cell metabolism modeling. Although a large number of enzyme turnover numbers are provided in databases, the number of enzymes with experimentally measured enzyme turnover numbers is much less than the number of sequenced proteins. Using artificial intelligence technology can significantly reduce the experimental cost of obtaining large-scale enzyme turnover numbers. However, the existing deep learning model has low prediction accuracy and basically does not exploit the deep relationship between the enzyme turnover number and the protein structure of the enzyme. SUMMARY
[0003] The present application proposes an enzyme turnover number prediction method based on a deep learning model and protein structure information to solve the problems in the prior art. By adding protein structure information to enzyme turnover number prediction training and using deep learning methods to mine the potential relationship between the two, the model generalization effect can be effectively enhanced, and the problems of low prediction model accuracy and limitation by sequence similarity can be solved.
[0004] The present application is implemented by the following technical solutions:
[0005] The present application relates to an enzyme turnover number prediction method based on a deep learning model and protein structure information. A dataset with unique enzyme sequence-enzyme structure-substrate combination and enzyme turnover number mapping is constructed in the offline stage to train the enzyme turnover number prediction model, and the trained enzyme turnover number prediction model is used to predict the enzyme turnover number of the protein structure to be processed in real time in the online stage.
[0006] The enzyme sequence-enzyme structure-substrate combination and enzyme turnover number mapping dataset is constructed by the following method:
[0007] Step 1: Obtain enzyme catalytic reaction data from the Brenda database and the SABIO-RK database, including: EC number of the enzyme, species source, sequence, name of the substrate molecule, SMILES of the substrate molecule, and turnover number value of the enzyme catalytic reaction;
[0008] Step 2: To solve the problem that the same enzyme sequence in the existing dataset has high similarity, which may lead to false prediction, the dataset is filtered and reconstructed. The enzymes with the same substrate molecule and a sequence similarity higher than 90% in the dataset are removed, and only the enzyme-substrate pair with the longest sequence length is retained to construct the enzyme sequence-substrate combination and enzyme turnover number mapping dataset.
[0009] Step 3, using the rapid protein structure prediction software: ColabFold, obtaining the protein structure predicted based on the enzyme sequence data in step 2, constructing the enzyme sequence-enzyme structure-substrate combination and enzyme turnover number mapping dataset without data feature transformation;
[0010] Step 4, transforming protein sequence data, protein structure data and substrate molecular SMILES data into initial features as model input, specifically including:
[0011] Step 4.1, using the RDKit software package to convert the substrate molecular SMILES string into a molecular fingerprint and a contact matrix, wherein: the molecular atoms are represented as nodes of the contact matrix, and the intermolecular chemical bonds are represented as edges of the contact matrix; obtaining the initial features of the substrate molecule;
[0012] Step 4.2, encoding the enzyme sequence, taking every 4 adjacent amino acids as a group, and encoding from 0, the same amino acid combination is the same encoding, obtaining the initial features representing the enzyme sequence. There are 137026 different amino acid combinations in the constructed dataset;
[0013] Step 4.3, based on the predicted enzyme protein three-dimensional structure in step 3, obtaining the three-dimensional coordinate information of the protein amino acid;
[0014] Step 4.4, taking the amino acid node as the graph structure node, calculating the Euclidean distance between each pair of amino acids based on the three-dimensional coordinates of the amino acids If the distance is less than a certain threshold, it is considered that there is an edge between the two amino acids, where (x1, y1, z1) and (x2, y2, z2) are the three-dimensional coordinates of the corresponding amino acids, obtaining the initial features of the protein structure;
[0015] The enzyme turnover number prediction model comprises: a Transformer model and a graph convolutional neural network with a multi-head attention mechanism, wherein: the initial features are further extracted by the graph convolutional neural network with a multi-head attention mechanism, specifically: and the attention mechanism wherein: σ is an activation function, is an adjacency matrix, is a degree matrix, H (l) is a node feature, W (l) is a graph convolutional neural network model parameter; and the initial features of the enzyme sequence are processed by the Transformer model, specifically: are used to obtain the attention weight feature matrix of the enzyme sequence, wherein: softmax is an activation function, Q, K, and V are task representation vectors, d kis the number of dimensions; finally, the initial features of the enzyme structure are processed by using a graph convolutional neural network, and the initial features are obtained by: further feature extraction is performed, wherein: sigma is an activation function, is an adjacency matrix, is a degree matrix, H (l) is a node feature, W (l) is a graph convolutional neural network model parameter.
[0016] The encoding module in the Transformer model obtains the initial features of the enzyme sequence, and the decoding module uses a full connection layer to comprehensively decode the enzyme catalytic reaction matrix.
[0017] A plurality of self-attention modules combined into a multi-head attention module are arranged between each layer of the graph convolutional neural network, so as to mine the different influences of different amino acid points of the enzyme on the enzyme turnover number, and enhance the accuracy and generalization ability of the model.
[0018] The training is performed by taking R 2 and RMSE as the accuracy standards of the enzyme turnover number prediction model, and updating and optimizing the model parameters based on a preset learning rate.
[0019] The real-time prediction refers to: combining the substrate feature matrix, the enzyme sequence matrix and the enzyme structure matrix of each group of enzyme-substrate pairs output by the enzyme turnover number prediction model to obtain an enzyme catalytic reaction matrix, using a full connection layer to operate the enzyme catalytic reaction matrix, and finally obtaining the final turnover number value by: y = Wx.
[0020] The present application relates to a system for implementing the above method, comprising: a data collection unit, a data preprocessing unit, a model building unit, a model training unit and a result prediction unit, wherein: the data collection unit obtains enzyme catalytic reaction data from the Brenda database and the SABIO-RK database; the data preprocessing unit predicts protein structure by using ColabFold, obtains initial features of protein structure according to structure information, obtains initial features of substrate molecules by using RDKit, and obtains initial features of protein sequence by continuing to encode amino acids; the model building unit combines graph convolutional neural networks with different numbers of layers and Transformers to establish a deep learning model framework; the model training unit trains the model according to the initial features obtained in the data preprocessing part to obtain a trained deep learning model; and the result prediction unit predicts the test set according to the enzyme and substrate molecule information of the test set and obtains an evaluation result.
[0021] Technical effects
[0022] The application constructs a new enzyme sequence-enzyme structure-substrate molecule combination and enzyme turnover number one-to-one mapping dataset capable of reducing the influence of sequence similarity; adds three-dimensional structure data of the enzyme, and adds a multi-head attention module based on a graph convolutional neural network. Compared with the prior art, the application significantly improves the prediction effect of enzyme turnover number while significantly reducing the negative impact of different sequence similarities on enzyme turnover number prediction; to a certain extent, it can predict the impact of point mutations on enzyme turnover number. BRIEF DESCRIPTION OF DRAWINGS
[0023] Figure 1 is a flowchart of the application;
[0024] Figure 2 is a schematic diagram of the overall structure of the system;
[0025] Figure 3 is a prediction effect diagram of the embodiment;
[0026] Figure 4 is a prediction effect diagram of the embodiment under different enzyme sequence similarities;
[0027] Figure 5 is a prediction effect diagram of the embodiment under different point mutations. DETAILED DESCRIPTION
[0028] After specific actual experiments, the embodiment uses Python to build a model in a Windows environment and uses GPU for training, starting with randomly initialized parameters to train the model. The model prediction effect of the test set is as shown in Figure 3 : the prediction accuracy is 0.6, and the specific prediction results are shown in Table 1, which illustrates the predicted enzyme EC number, substrate name, and predicted enzyme turnover number of 5 enzyme catalytic reactions. The model prediction effect under different enzyme sequence similarities in the test set is as shown in Figure 4 .
[0029] As shown in Figure 1 , the specific steps of the embodiment include:
[0030] Step 1, collect enzyme catalytic reaction data, preprocess the collected enzyme catalytic reaction data, remove enzymes with the same substrate molecule and sequence similarity higher than 90% in the dataset, and only keep the longest enzyme-substrate pair, construct an enzyme catalytic reaction dataset capable of reducing the influence of enzyme sequence similarity, use the ColabFold protein structure prediction method to predict the structure of the enzyme protein sequence, and construct an enzyme sequence-enzyme structure-substrate molecule combination and enzyme turnover number one-to-one mapping dataset based on each enzyme catalytic reaction.
[0031] Step 2: Based on the enzyme sequence data, structural data, and substrate molecule data, the RDKit software package is used to convert the substrate molecule SMILES string into a molecular fingerprint and a contact matrix to obtain the initial features characterizing the substrate molecule. The enzyme sequence is encoded, with every four adjacent amino acids grouped together and encoded starting from 0. Identical amino acid combinations are assigned the same code, thus obtaining the initial features characterizing the enzyme sequence. Using amino acid nodes as graph structure nodes, the Euclidean distance between each pair of amino acids is calculated from their three-dimensional coordinates. When the two amino acids are considered to have an edge, (x1,y1,z1) and (x2,y2,z2) are the three-dimensional coordinates of the corresponding amino acids, respectively, the initial structural features of the protein are obtained; the initial features of the enzyme sequence, enzyme structure and substrate molecule are used as input, and the input data are randomly divided into 80% training set, 10% validation set and 10% test set.
[0032] Step 3: Utilizing a deep learning model: A graph neural network with multi-head attention mechanism, a Transformer encoding module, and a fully connected layer as a decoding module are incorporated to build an enzyme turnover prediction model based on protein structure information. See [link to model structure] for details. Figure 2 .
[0033] Step 4: Based on the dataset built in Step 1 and the initial features extracted in Step 2, train the prediction model in Step 3 to obtain the model prediction accuracy for the prediction task, and repeat the process multiple times to find the model parameters with the highest prediction accuracy.
[0034] Step 5: Train the weights using the model with the best prediction accuracy to predict enzyme turnover, and obtain the final enzyme turnover based on protein structure information.
[0035] Table 1 Experimental Data
[0036]
[0037] Compared with existing technologies, this method significantly improves the prediction accuracy while significantly reducing the negative impact of different sequence similarities on enzyme turnover prediction; it can predict the impact of point mutations on enzyme turnover and identify key amino acid residue sites to a certain extent.
[0038] The above-described specific implementations can be partially adjusted by those skilled in the art in different ways without departing from the principles and purpose of the present invention. The scope of protection of the present invention is defined by the claims and is not limited to the above-described specific implementations. All implementation schemes within the scope of the claims are bound by the present invention.
Claims
1. A method for predicting the number of enzyme turnovers based on a deep learning model and protein structure information, characterized in that, By constructing a dataset with unique enzyme sequence-enzyme structure-substrate combination and enzyme turnover number mapping in the offline stage for training the enzyme turnover number prediction model, the enzyme turnover number of the protein structure to be processed is predicted in real time in the online stage by the trained enzyme turnover number prediction model; The enzyme sequence-enzyme structure-substrate combination and enzyme turnover number mapping dataset is constructed by the following method: Step 1, obtain enzyme catalytic reaction data from the Brenda database and the SABIO-RK database, including: EC number of enzyme, species source, sequence, name of substrate molecule, SMILES of substrate molecule and turnover number value of enzyme catalytic reaction; Step 2, remove the enzymes with the same substrate molecule and the sequence similarity higher than 90% in the dataset, and only keep the enzyme-substrate pair with the longest enzyme sequence to construct the enzyme sequence-substrate combination and enzyme turnover number mapping dataset; Step 3, use the fast protein structure prediction software: ColabFold to obtain the protein structure predicted based on the enzyme sequence data in step 2, and construct the enzyme sequence-enzyme structure-substrate combination and enzyme turnover number mapping dataset without data feature conversion; Step 4, convert the protein sequence data, protein structure data and substrate molecule SMILES data into the initial features of the model input; The enzyme turnover number prediction model comprises a Transformer model and a graph convolutional neural network with a multi-head attention mechanism, wherein the initial features of a substrate molecule are first processed by the graph convolutional neural network with the multi-head attention mechanism, specifically: the initial features are further extracted by: and an attention mechanism , wherein: is an activation function, is an adjacency matrix, is a degree matrix, is a node feature, is a graph convolutional neural network model parameter; then the initial features of an enzyme sequence are processed by the Transformer model, specifically: is used to obtain an attention weight feature matrix of the enzyme sequence, wherein: is an activation function, , , is a task representation vector, is a dimension number; finally, the initial features of an enzyme structure are processed by the graph convolutional neural network, and the initial features are further extracted by: , wherein: is an activation function, is an adjacency matrix, is a degree matrix, is a node feature, is a graph convolutional neural network model parameter. 2.The method of claim 1, wherein the method is characterized by, The step 4 specifically includes: Step 4.1, use the RDKit software package to convert the substrate molecule SMILES string into a molecular fingerprint and a contact matrix, wherein: the molecular fingerprint represents the small molecule atoms as the nodes of the contact matrix, and the chemical bonds between molecules as the edges of the contact matrix; obtain the initial features of the substrate molecule; Step 4.2, encode the enzyme sequence, every 4 adjacent amino acids as a group, and encode from 0, the same amino acid combination is the same code, and obtain the initial features representing the enzyme sequence; there are 137026 different amino acid combinations in the constructed dataset; Step 4.3, based on the enzyme protein three-dimensional structure predicted in step 3, obtain the three-dimensional coordinate information of the protein amino acid; Step 4.
4. Take amino acid nodes as graph structure nodes, and calculate the Euclidean distance between each pair of amino acids based on the three-dimensional coordinates of the amino acids If the distance is less than or equal to the threshold, then it is considered that there is an edge between the two amino acids, wherein are the three-dimensional coordinates of the corresponding amino acids, respectively, to obtain the initial structural features of the protein. 3.The method of claim 1, wherein the method is characterized by, The encoding module in the Transformer model obtains the initial features of the enzyme sequence, and the decoding module uses a fully connected layer to comprehensively decode the enzyme catalytic reaction matrix. 4.The method of claim 1, wherein the method is characterized by, Each layer of the graph convolutional neural network is provided with a plurality of self-attention combination multi-head attention modules to mine the different influences of different amino acid points of the enzyme on the enzyme turnover number, and enhance the accuracy and generalization ability of the model. 5.The method of claim 1, wherein the method is characterized by, The training, with R 2 RMSE as the accuracy standard of the enzyme turnover number prediction model, updates and optimizes the model parameters based on the preset learning rate. 6.The method of claim 1, wherein the method is characterized by, The real-time prediction refers to: combining the substrate feature matrix, the enzyme sequence matrix and the enzyme structure matrix of each group of enzyme-substrate pairs output by the enzyme turnover number prediction model to obtain an enzyme catalytic reaction matrix, using a fully connected layer to operate the enzyme catalytic reaction matrix, and obtaining the final turnover number value from: The final turnover number value is finally obtained.
7. A deep learning-based protein structure information-empowered enzyme turnover number prediction system implementing any of the methods of claims 1-6, characterized in that, It includes: A data collection unit, a data preprocessing unit, a model building unit, a model training unit and a result prediction unit, wherein: the data collection unit obtains enzyme catalytic reaction data from the Brenda database and the SABIO-RK database; the data preprocessing unit uses ColabFold to predict the protein structure, and obtains the initial features of the protein structure according to the structure information, uses RDKit to obtain the initial features of the substrate molecule, and further encodes the amino acid to obtain the initial features of the protein sequence; The model building unit combines the graph convolutional neural network and the Transformer with different layers to establish a deep learning model framework; the model training unit trains the model according to the initial features obtained by the data preprocessing part to obtain a trained deep learning model; and the result prediction unit predicts the test set according to the enzyme and substrate molecular information of the test set and obtains an evaluation result.