A method for predicting protein / enzyme substrate binding sites based on convolutional neural network

Abstract

The present invention discloses a method for predicting protein / enzyme substrate binding sites based on convolutional neural networks, and belongs to the field of bioinformatics. The method establishes a deep convolutional neural network model DUnet based on Densenet module and Unet module. In the model, because of the presence of Densenet module, each layer of the network is densely connected to the remaining layers, so as to provide the possibility of realizing enhanced feature extraction. At the same time, the present application rotates the protein-ligand in the training set to realize enhanced feature extraction, further improves the prediction accuracy of protein / enzyme substrate binding sites, and uses public data sets COACH420 and BU48 to verify the recognition substrate binding position and range of the prediction method of the present application, proving that its prediction result is more accurate. It is of great significance to structure-based protein modification, shortening the evolutionary cycle of protein / enzyme, and reducing experimental costs.

Description

Technical Field

[0001] The invention relates to a method for predicting protein / enzyme substrate binding sites based on convolutional neural networks, and belongs to the technical field of bioinformatics. Background Art

[0002] The modification of enzyme activity and specificity has been a research hotspot in recent years. Research in this area can broaden the application field of enzymes and improve the application value of enzymes. In the process of protein drug design, improving the affinity between the target protein and the target protein will help improve the targeting of the drug and thus improve the efficacy. Structure-based protein modification is the research focus at this stage. The overall modification is more rational, which can shorten the evolution cycle of proteins / enzymes and reduce experimental costs. Obtaining the substrate binding site of proteins / enzymes is particularly important for their activity modification. In conventional algorithms, the substrate is docked with the substrate binding pocket of the target protein / enzyme, and the key binding site of the protein / enzyme and the substrate is obtained based on the docking conformation. On this basis, the affinity of the protein / enzyme and the substrate is improved by modifying the binding site.

[0003] In previous studies, the method of obtaining substrate binding sites was basically based on classical mechanics. The method is to dock a virtual substrate with a protein / enzyme and score the cavity on the surface of the protein / enzyme by evaluating the affinity energy. In addition, some proteins / enzymes can also be used as templates to infer the substrate binding site of the target protein by using the experimentally verified homologous protein and substrate binding site as a template. However, both of the above methods have defects: first, many proteins / enzymes and related homologous proteins have not been studied enough, and it is impossible to achieve accurate substrate pocket positioning through homology retrieval; second, virtual substrates cannot effectively interpret the diversity of substrates, and some proteins / enzymes bind to substrates with high specificity, which does not meet the principle of minimum affinity energy.

[0004] Therefore, Jeevan Kandel et al. proposed a method for predicting protein / enzyme substrate binding sites based on deep learning in 2021 (see Kandel, J., Tayara, H., & Chong, KT. (2021). Puresnet: prediction of protein-ligand binding sites using deep residual neural network. Journal of Cheminformatics). This method integrates ResNet and Unet to construct a framework PUResNet prediction model, which contains a coding region and a decoding region, which can respectively realize protein feature extraction and prediction functions. However, since ResNet itself is to solve the problem of network gradient explosion, it belongs to residual learning, which is equivalent to adding an identity mapping to the internal structure of the deep network to retain the initial input features. This makes it impossible to obtain more information when extracting protein features, resulting in its prediction accuracy to be further improved. Moreover, the prediction model proposed by Jeevan Kandel et al. uses a small amount of training samples, which also leads to its low prediction accuracy. Summary of the invention

[0005] In order to further improve the prediction accuracy of protein / enzyme substrate binding sites, the present invention provides a method for predicting protein / enzyme substrate binding sites based on a convolutional neural network, the method comprising:

[0006] Step S1: Based on the Densenet module and the Unet module, a deep convolutional neural network model DUnet is constructed;

[0007] Step S2: Use the public data set to train the deep convolutional neural network model DUnet constructed in step S1, and rotate the proteins and ligands in the public data set by 15° so as to obtain the rotated structure and realize enhanced feature extraction;

[0008] Step S3: Use the trained deep convolutional neural network model Dune to predict protein / enzyme substrate binding sites.

[0009] Optionally, the deep convolutional neural network model DUnet constructed in step S1 is divided into two parts: an encoding area and a decoding area; the encoding area is used to extract features of the training set data and realize high-order feature extraction by increasing the dimension; the decoding area is used to reduce the dimension of the data and enlarge the image size, and realize back propagation adjustment of model parameters by comparing the prediction results and labels;

[0010] In the encoding region, the input vector is initially 36×36×36×18, and after multiple rounds of convolution, activation layers, and batch normalization layers, the size becomes 3×3×3×640;

[0011] In the decoding area, the vector is restored from its initial size of 3×3×3×640 to 36×36×36×1 through multiple rounds of upsampling, convolution, batch normalization, activation layers, and connections, where feature 1 only shows 0 and 1, 1 represents that there is an atom at that position, and 0 represents that there is no atom at that position.

[0012] Optionally, the public data set used in step S2 is the scPDB database, which contains all atomic coordinates of 5020 protein-ligand complexes, functional explanations of proteins, chemical descriptions of ligand complexes, and direct intermolecular interactions of protein-ligand complexes.

[0013] Optionally, before the step S2 uses a public data set to train the deep convolutional neural network model DUnet constructed in the step S1, the method further includes:

[0014] Atomic features of protein and ligand structures in the training set are described separately, including: atom type, atomic hybridization, type of bonds between heavy atoms, number of bonds between each atom and surrounding atoms, whether the atom participates in hydrophobic groups, benzene rings, non-bonded force receptors or donors, and cyclic groups; whether it is charged; whether it is a protein or a ligand;

[0015] The proteins and ligands in the public dataset scPDB were rotated 15° to obtain the rotated structures for enhanced feature extraction;

[0016] According to the above atomic feature description, proteins and ligands are treated as matrices and vector data: proteins are regarded as 3D boxes with a precision of The protein is cut into pieces so that each piece contains only one atom; when treating the ligand, only 0 and 1 are used to represent the atomic characteristics.

[0017] Optionally, when step S2 uses a public data set to train the deep convolutional neural network model DUnet constructed in step S1, the accuracy of the model is optimized by adjusting the learning rate and batch size, and after a specific training generation, the change in the loss value is used to evaluate the convergence interval of the model.

[0018] Optionally, before the step S3 uses the trained deep convolutional neural network model DUnet to predict the protein / enzyme substrate binding site, it also includes performing an accuracy evaluation on the trained deep convolutional neural network model DUnet. When the accuracy meets the predetermined requirements, the deep convolutional neural network model DUnet is considered to have been trained.

[0019] Optionally, the accuracy evaluation of the trained deep convolutional neural network model DUnet includes:

[0020] The trained deep convolutional neural network model DUnet is used to predict the substrate binding sites of proteins in the public data sets COACH420 and BU48, and the corresponding prediction results are obtained; the prediction results include the range data of the protein active center;

[0021] The predicted protein active center was compared with the center of the real ligand in the public data set COACH420 and BU48, and the prediction success rate was calculated. The formula is as follows:

[0022]

[0023] Among them, DCC represents the distance between the predicted three-dimensional center of the substrate and the actual three-dimensional center of the substrate;

[0024] The predicted protein active center space range is compared with the intersection of the actual ligand space range, and the intersection ratio DVO of the three-dimensional space points contained in the predicted substrate and the three-dimensional space points of the actual substrate is calculated. The formula is as follows:

[0025]

[0026] Among them, Pbs is the predicted binding site, and Abs is the actual binding site;

[0027] The prediction accuracy of the trained deep convolutional neural network model DUnet is evaluated based on the prediction success rate and DVO.

[0028] The beneficial effects of the present invention are:

[0029] The present invention is to achieve more accurate prediction for protein / enzyme substrate binding sites, establish a deep convolutional neural network model DUnet based on Densenet module and Unet module, in this model, because of the existence of Densenet module, each layer of the network is densely connected with the remaining layers, so as to provide the possibility of realizing enhanced feature extraction, at the same time, the present application rotates the protein-ligand in the training set, realizes enhanced feature extraction, further improves the prediction accuracy of protein / enzyme substrate binding sites, and uses public data sets COACH420 and BU48 to verify the recognition substrate binding position and range of the prediction method of the present application, proving that its prediction result is more accurate. It is of great significance to structure-based protein modification, shortening the evolutionary cycle of protein / enzyme, and reducing experimental costs. BRIEF DESCRIPTION OF THE DRAWINGS

[0030] In order to more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings required for use in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings in the following description are only some embodiments of the present invention. For ordinary technicians in this field, other accompanying drawings can be obtained based on these accompanying drawings without paying creative work.

[0031] Figure 1 It is a simulation comparison diagram of the prediction results of 11 enzyme substrate pockets by a method for predicting protein / enzyme substrate binding sites based on a convolutional neural network provided in one embodiment of the present invention and a traditional prediction method. DETAILED DESCRIPTION

[0032] In order to make the objectives, technical solutions and advantages of the present invention more clear, the embodiments of the present invention will be further described in detail below with reference to the accompanying drawings.

[0033] Embodiment 1:

[0034] This embodiment provides a method for predicting protein / enzyme substrate binding sites based on a convolutional neural network, the method comprising:

[0035] Step S1: Based on the Densenet module and the Unet module, a deep convolutional neural network model DUnet is constructed;

[0036] The deep convolutional neural network model DUnet is divided into two parts: the encoding area and the decoding area; the encoding area is used to extract the features of the training set data and realize high-order feature extraction by increasing the dimension; the decoding area is used to reduce the dimension of the data and enlarge the image size, and realize back propagation to adjust the model parameters by comparing the prediction results and labels;

[0037] In the encoding region, the input vector is initially 36×36×36×18, and after multiple rounds of convolution, activation layers, and batch normalization layers, the size becomes 3×3×3×640;

[0038] In the decoding area, the vector is restored from its initial size of 3×3×3×640 to 36×36×36×1 through multiple rounds of upsampling, convolution, batch normalization, activation layers, and connections, where feature 1 only shows 0 and 1, 1 represents that there is an atom at that position, and 0 represents that there is no atom at that position.

[0039] Step S2: Use the public data set to train the deep convolutional neural network model DUnet constructed in step S1, and rotate the proteins and ligands in the public data set by 15° so as to obtain the rotated structure and realize enhanced feature extraction;

[0040] The public data set used in this example is the scPDB database, which is a data archive of drug binding sites extracted from the protein database. After processing, the database contains all atomic coordinates of 5020 protein-ligand complexes. scPDB provides functional annotations for proteins, chemical descriptions for ligands, and detailed molecular interactions for complexes. scPDB now includes a hierarchical classification of all binding sites in a functional class.

[0041] Atomic features of protein and ligand structures in the training set are described separately, including: atom type, atomic hybridization, type of bonds between heavy atoms, number of bonds between each atom and surrounding atoms, whether the atom participates in hydrophobic groups, benzene rings, non-bonded force receptors or donors, and cyclic groups; whether it is charged; whether it is a protein or a ligand;

[0042] The proteins and ligands in the public dataset scPDB were rotated 15° to obtain the rotated structures for enhanced feature extraction;

[0043] According to the above atomic feature description, proteins and ligands are treated as matrices and vector data: proteins are regarded as 3D boxes with a precision of The protein is cut into pieces so that each piece contains only one atom; when treating the ligand, only 0 and 1 are used to represent the atomic characteristics.

[0044] The accuracy of the model is optimized by adjusting the learning rate and batch size. After a specific number of training generations, the change in the loss value is used to evaluate the convergence interval of the model.

[0045] The loss value is used to define the error between a single training sample and the true value. The adjustment of the learning rate and batch size can effectively capture the point where the loss value gradient decreases with the true value at the minimum. At the same time, after the loss value change converges, the gradient descent will disappear. Based on the loss value change, it can be inferred that after how many generations of training, the model converges and the training is stopped.

[0046] The accuracy of the trained deep convolutional neural network model DUnet is evaluated. When the accuracy meets the predetermined requirements, the training of the deep convolutional neural network model DUnet is considered complete. Specifically,

[0047] The trained deep convolutional neural network model DUnet is used to predict the substrate binding sites of proteins in the public data sets COACH420 and BU48, and the corresponding prediction results are obtained; the prediction results include the range data of the protein active center;

[0048] Among them, COACH420 contains 298 protein-ligand data, and BU48 contains 62 protein-ligand data.

[0049] The predicted protein active center was compared with the center of the real ligand in the public data set COACH420 and BU48, and the prediction success rate was calculated. The formula is as follows:

[0050]

[0051] The predicted protein active center space range is compared with the intersection of the actual ligand space range, and the intersection ratio DVO of the three-dimensional space points contained in the predicted substrate and the three-dimensional space points of the actual substrate is calculated. The formula is as follows:

[0052]

[0053] Among them, Pbs is the predicted binding site, and Abs is the actual binding site;

[0054] The prediction accuracy of the trained deep convolutional neural network model DUnet is evaluated based on the prediction success rate and DVO.

[0055] Step S3: Use the trained deep convolutional neural network model Dune to predict protein / enzyme substrate binding sites.

[0056] In order to demonstrate the predictive performance of the method for predicting protein / enzyme substrate binding sites based on convolutional neural networks proposed in this application, this application conducted a simulation comparison on the above-mentioned public data sets COACH420 and BU48 with the method for predicting protein / enzyme substrate binding sites based on deep learning proposed by Jeevan Kandel et al. in 2021 (because it uses PUResNet, PUResNet is used to represent the existing method in the subsequent introduction and simulation drawings; the method proposed in this application is represented by DUnet). The simulation results are shown in Table 1 below. In Table 1, DVO Refers to the prediction results of the two models, DCC is less than or equal to DVO (DCC top 150): refers to the specific value of DVO of the top 150 prediction results with the lowest DCC values ​​in the prediction results of the two models; DVO (DCC top 31): refers to the specific value of DVO of the top 31 prediction results with the lowest DCC values ​​in the prediction results of the two models.

[0057] Note: When the prediction accuracy is low, that is, when the DCC value is greater than 5, the predicted points and the actual ligand 3D points generally have no intersection, so it is meaningless to calculate DVO. Therefore, under most conditions, only the predicted DCC is less than or equal to The case is characterized, and DCC is less than or equal to The sample size is less than 50% of the overall sample size. In order to demonstrate the high accuracy of the constructed model, this application evaluates the data that accounts for half of the sample library. Therefore, in the COACH420 dataset, the DVO evaluation of the top 150 DCC rankings and the DVO evaluation of the top 31 DCC rankings in the BU48 dataset are added;

[0058] Table 1 Performance of PUResNet and DUnet on the validation set

[0059]

[0060] It can be seen from Table 1 above that on the public data sets COACH420 and BU48, the DVO obtained by the prediction method proposed in this application is 0.028, while the DVO of the existing deep learning-based prediction method is 0.02. It can be seen that the prediction ability of the method of this application for the intersection of the coordinate space of the predicted substrate and the real substrate is much higher than the existing method.

[0061] In practical applications, relatively accurate DVO prediction can maximize the simulation of the relative position of the substrate and infer the binding site between the substrate and the enzyme. Therefore, DUnet has achieved a significant improvement over PUResNet.

[0062] In addition, in order to illustrate that DUnet has a significant improvement in accuracy compared to PUResNet, this application also collected 11 different enzyme-substrate structures, as shown in Table 2, and used PUResNet and DUnet to predict substrate pockets. The prediction results are as follows: Figure 1 As shown. In order to illustrate the advantages of DUnet, the prediction results of FTSite (refer to the introduction in doi: https: / / doi.org / 10.1038 / nprot.2015.043) were compared. FTSite uses virtual molecular probes to analyze the binding energy of protein surface cavities and sorts the cavities based on the obtained binding energy to determine whether they are substrate binding regions. The results show that the prediction method DUnet proposed in this application can more accurately predict the enzyme-substrate binding position, such as Figure 1 shown.

[0063] Table 2: 11 different enzyme-substrate structures

[0064]

[0065]

[0066] Figure 1In (A), the abscissa PDB ID is the crystal conformation data of the protein, which comes from www.rcsb.org; the ordinate represents the corresponding DCC prediction value. It can be seen that the distance DCC between the three-dimensional center of the substrate obtained by the method of the present application and the existing protein / enzyme substrate binding site prediction method based on deep learning and the actual three-dimensional center of the substrate is smaller, as shown in Table 3 below:

[0067] Table 3: DCC obtained by PUResNet and DUnet respectively

[0068] PDB ID This application method DUnet Existing method PUResNet 1oa7 5.5207760840000010 9.3179935650000000 1pig 2.1154143760000000 3.1989701650000000 1r50 1.7763744040000000 0.9088345040000000 1scn 4.6383586890000010 No prediction results 1q6c 2.4366502600000000 1.3883605760000000 1v2w 2.8667493010000000 4.2151694060000000 2bnj 5.1867075390000010 4.1497794210000000 3esb 3.3254862990000000 3.4996163660000000 4pse 0.7093608500000001 2.9154317480000000 5i79 9.9778902429999990 9.5882805080000000 6gmg 7.6770147540000000 9.0221674830000000

[0069] As can be seen from Table 3, most of the DCCs between the three-dimensional center of the substrate predicted by the method of the present application and the actual three-dimensional center of the substrate are smaller than the prediction results of the existing methods, indicating that the prediction results of the method of the present application are more accurate.

[0070] Figure 1 (B) The spatial position of the substrate pocket predicted by PUResNet, FTSite and DUnet, where yellow is the prediction by PUResNet, pink is the most recommended prediction result by FTSite, green and blue are also the prediction results of FTSite, and light gray is the prediction by DUnet; this shows that among the three models, the points predicted by DUnet have a higher degree of overlap with the real ligand, so the model has higher accuracy; Figure 1 (C) is the position of the substrate pocket of glutamine transaminase, where Acyl donor represents the acyl donor binding region and Acyl acceptor represents the acyl acceptor binding region. The yellow and gray colors are the prediction results of PUResNet and DUnet for the substrate pocket, respectively, indicating that in this case, DUnet can simultaneously predict the binding positions of the donor and ligand in the substrate, indicating that it has a good substrate binding position prediction effect for multi-substrate enzymes.

[0071] Some steps in the embodiments of the present invention may be implemented using software, and the corresponding software program may be stored in a readable storage medium, such as a CD or a hard disk.

[0072] 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 principle of the present invention should be included in the protection scope of the present invention.

Claims

1. A method for predicting protein / enzyme substrate binding sites based on convolutional neural networks, characterized in that: The method comprises: Step S1: Based on the Densenet module and the Unet module, a deep convolutional neural network model DUnet is constructed; Step S2: Use the public data set to train the deep convolutional neural network model DUnet constructed in step S1, and rotate the proteins and ligands in the public data set by 15° so as to obtain the rotated structure and realize enhanced feature extraction; Step S3: using the trained deep convolutional neural network model DUnet to predict protein / enzyme substrate binding sites; The deep convolutional neural network model DUnet constructed in step S1 is divided into two parts: an encoding area and a decoding area; the encoding area is used to extract the features of the training set data and realize high-order feature extraction by increasing the dimension; the decoding area is used to reduce the dimension of the data and enlarge the image size, and realize back propagation adjustment of model parameters by comparing the prediction results and labels; In the encoding region, the input vector is initially 36×36×36×18, and after multiple rounds of convolution, activation layers, and batch normalization layers, the size becomes 3×3×3×640; In the decoding area, the vector is restored from the initial size of 3×3×3×640 to 36×36×36×1 through multiple rounds of upsampling, convolution, batch normalization, activation layers, and connections, where feature 1 only shows 0 and 1, 1 represents that there is an atom at this position, and 0 represents that there is no atom at this position; Before the step S3 uses the trained deep convolutional neural network model DUnet to predict the protein / enzyme substrate binding site, it also includes performing an accuracy evaluation on the trained deep convolutional neural network model DUnet. When the accuracy meets the predetermined requirements, the deep convolutional neural network model DUnet is considered to be trained; The accuracy evaluation of the trained deep convolutional neural network model DUnet includes: The trained deep convolutional neural network model DUnet is used to predict the substrate binding sites of proteins in the public data sets COACH420 and BU48, and the corresponding prediction results are obtained; the prediction results include the range data of the protein active center; The predicted protein active center was compared with the center of the real ligand in the public data set COACH420 and BU48, and the prediction success rate was calculated. The formula is as follows: Among them, DCC represents the distance between the predicted three-dimensional center of the substrate and the actual three-dimensional center of the substrate; The predicted protein active center space range is compared with the intersection of the actual ligand space range, and the intersection ratio DVO of the three-dimensional space points contained in the predicted substrate and the three-dimensional space points of the actual substrate is calculated. The formula is as follows: Among them, Pbs is the predicted binding site, and Abs is the actual binding site; The prediction accuracy of the trained deep convolutional neural network model DUnet is evaluated based on the prediction success rate and DVO.

2. The method according to claim 1, characterized in that The public data set used in step S2 is the scPDB database, which contains all atomic coordinates of 5020 protein-ligand complexes, functional explanations of proteins, chemical descriptions of ligand complexes, and direct intermolecular interactions of protein-ligand complexes.

3. The method according to claim 2, characterized in that Before the step S2 uses the public data set to train the deep convolutional neural network model DUnet constructed in the step S1, the method further includes: Atomic features of protein and ligand structures in the training set are described separately, including: atom type, atomic hybridization, type of bonds between heavy atoms, number of bonds between each atom and surrounding atoms, whether the atom participates in hydrophobic groups, benzene rings, non-bonded force receptors or donors, and cyclic groups; whether it is charged; whether it is a protein or a ligand; The proteins and ligands in the public dataset scPDB were rotated 15° to obtain the rotated structures for enhanced feature extraction; According to the above atomic feature description, proteins and ligands are treated as matrices and vector data: proteins are regarded as 3D boxes with a precision of The protein is cut into pieces so that each piece contains only one atom; when treating the ligand, only 0 and 1 are used to represent the atomic characteristics.

4. The method according to claim 3, characterized in that When the step S2 uses a public data set to train the deep convolutional neural network model DUnet constructed in step S1, the accuracy of the model is optimized by adjusting the learning rate and batch size. After a specific training generation, the change in the loss value is used to evaluate the convergence interval of the model.

Images