Apparatus and method for the prediction of protein binding pockets

Abstract

The application discloses a kind of prediction device and method of protein binding pocket, and the present application is by extracting features to protein and binding pocket data and the transformation of voxel grid, it is applied to the research of predicting protein binding pocket by deep learning.The present application is based on position information features and atomic information features simultaneously for prediction, which greatly improves the accuracy and precision of predicting protein binding pocket, and provides reliable molecular target binding information for drug design.

Description

Technical Field

[0001] This invention belongs to the field of artificial intelligence drug development, and specifically relates to a device and method for predicting protein binding pockets. Background Technology

[0002] Proteins are essential components of all cells and tissues in the human body. All vital bodily functions require protein participation, and proteins play crucial roles in the body, such as participating in the transport of various substances, immune regulation, catalysis, and other regulatory functions. Predicting protein binding sites is of great guiding significance for drug design, protein function research, and molecular docking. The first step in structure-based drug design (SBDD) is to identify feasible drug binding sites on target proteins and the amino acid information constituting these sites—that is, to find surface protein cavities that may bind to small molecule compounds.

[0003] Due to their folding and spatial conformation, proteins often have complex, uneven surfaces with pockets or cavities. These pockets or cavities are often where drugs bind to proteins, and are called protein-ligand binding sites. Protein binding pockets refer to cavities on or inside a protein surface suitable for binding to ligands. The amino acid residues surrounding the binding pocket determine its shape, location, physicochemical properties, and function.

[0004] The method of using artificial intelligence to find protein-small molecule binding sites first analyzes the three-dimensional structure, physicochemical, and biological characteristics of known protein-small molecule complexes to obtain characteristic conditions that can describe protein-small molecule binding sites. Then, the three-dimensional structure of the protein to be predicted is analyzed, and the binding sites are defined on the protein using the previously obtained characteristics. In recent years, algorithms for predicting protein binding sites have been divided into algorithms based on geometric features and algorithms based on biochemical and physical features. Algorithms based on geometric features use computational geometry methods to search and define the protein surface, and then use a scoring function to score and rank the search results to determine the binding sites on the protein. Algorithms based on geometric features mainly include methods based on three-dimensional grids, spheres, and α-shapes. Algorithms based on biochemical and physical features include energy-based and sequence-based methods. Traditional algorithms for predicting protein binding sites have the following limitations:

[0005] (1) Geometry-based algorithms lack information about the interaction between proteins and ligands, and the scoring functions of classic geometric methods usually rely on custom features based on binding site attributes, which have limited scoring capabilities.

[0006] (2) The three-dimensional structure of a protein is closely related to the interaction between protein molecules. Algorithms based on biochemical and physical characteristics lack the positional information of proteins and molecules. Relying solely on the biochemical and physical characteristics of protein molecules cannot accurately predict the binding sites of proteins. Summary of the Invention

[0007] The purpose of this invention is to address the problems of low accuracy and poor precision in existing protein binding site prediction methods by proposing a device and method for predicting protein binding pockets. This invention can significantly improve the accuracy and precision of prediction.

[0008] The objective of this invention is achieved through the following technical solution:

[0009] A first aspect of the present invention provides a device for predicting protein binding pockets, comprising a storage unit and a control unit, characterized in that the storage unit comprises:

[0010] Protein structure information storage unit, used to store protein structure information related to the protein's structure; and

[0011] A protein binding pocket structure information storage unit is used to store protein binding pocket structure information related to the protein binding pocket structure.

[0012] The control unit includes:

[0013] A protein structure information extraction unit is used to obtain candidate protein structure information as a candidate protein from the protein structure information storage unit, wherein the candidate protein is a candidate protein predicted as a protein binding pocket.

[0014] A protein binding pocket information extraction unit is used to obtain the protein binding pocket structure information from the protein binding pocket information storage unit.

[0015] The protein feature extraction unit obtains the candidate protein based on the protein structure information extraction unit, and extracts the structural information features and atomic information features of the candidate protein to determine the three-dimensional voxel representation of the protein to be finally submitted into the protein binding pocket prediction model.

[0016] The protein binding pocket feature extraction unit obtains the protein binding pocket based on the protein binding pocket structure information extraction unit, extracts the protein binding pocket information features from the protein binding pocket, and determines the three-dimensional voxel representation of the protein binding pocket to be finally submitted into the protein binding pocket prediction model.

[0017] The protein binding pocket prediction unit predicts protein binding pockets using a deep learning method based on the protein's three-dimensional voxel representation. It then calculates the similarity between the predicted protein binding pocket and the protein binding pocket predicted by the protein binding pocket prediction model, and comprehensively evaluates the accuracy of the prediction model based on this similarity to determine the final predicted protein binding pocket.

[0018] The protein binding pocket prediction processing output unit performs screening processing on the predicted protein binding pockets and converts them into a specified format to obtain the final predicted protein binding pockets.

[0019] Furthermore, the protein binding pocket structural information includes the protein binding site, the protein binding cavity, and the protein-binding ligand.

[0020] Furthermore, the protein structure information extraction unit preprocesses the protein data, removing water molecules and impurities, and retaining 20 standard amino acids in the protein to obtain the candidate protein structure information, which is a three-dimensional structure file of the protein in pdb format.

[0021] Furthermore, the protein binding pocket structural information is a three-dimensional structural file of the binding pocket in pdb format.

[0022] Furthermore, the protein structure information extraction unit converts the 3D structure file of the protein in PDB format into a mol2 format file of the protein structure; the protein binding pocket information extraction unit converts the 3D structure file of the binding pocket in PDB format into a mol2 format file of the protein binding pocket.

[0023] Furthermore, the protein feature extraction unit performs feature extraction processing on the mol2 format file of the candidate protein, extracting structural information features and atomic information features of the protein respectively. The atomic information features include atom type, PyBel atom features, and SMARTS features.

[0024] Furthermore, the protein feature extraction unit converts the protein's mol2 format file into a three-dimensional voxel representation of the protein; the protein binding pocket feature extraction unit converts the protein binding pocket's mol2 format file into a three-dimensional voxel representation of the protein binding pocket.

[0025] Furthermore, a protein binding pocket prediction model is built using a deep learning algorithm. The processed training data is input, and the protein binding pocket prediction model is trained iteratively using mini-batch gradient descent. Based on the loss function, a set similarity metric function is used to calculate the similarity between the predicted value and the true value of the protein binding pocket prediction model.

[0026] Furthermore, the protein binding pocket prediction model includes an encoder part and a decoder part. The encoder part includes 16 3D convolution operations, 16 PReLU activation operations, and 5 residual connection operations. The decoder part includes 4 3D transposed convolution operations, 19 3D convolution operations, 21 PReLU activation operations, 9 regularization operations, and 1 sigmoid activation operation.

[0027] A second aspect of this invention provides a method for predicting protein binding pockets, the method using the aforementioned protein binding pocket prediction device, characterized in that the prediction method includes the following steps:

[0028] (1) Extracting protein structure information: The protein structure information extraction unit obtains the candidate protein structure information, i.e., the three-dimensional structure file of the protein in pdb format, from the protein structure information storage unit. The protein data is preprocessed to remove water molecules and impurities from the protein molecules, and the amino acid residues in the protein are tested and 20 standard amino acids are retained.

[0029] (2) Extract protein binding pocket information. The protein binding pocket information extraction unit obtains the protein binding pocket structure information, i.e., a three-dimensional structure file of the protein binding pocket in pdb format, from the protein binding pocket information storage unit.

[0030] (3) The protein structure information extraction unit extracts the atomic information of the protein according to the three-dimensional structure file of the protein in pdb format and converts it into a mol2 format file of the protein structure; the protein binding pocket information extraction unit extracts the atomic information of the binding pocket according to the three-dimensional structure file of the binding pocket in pdb format and converts it into a mol2 format file of the protein binding pocket.

[0031] (4) Extracting protein features: The protein feature extraction unit further processes the mol2 format file of the candidate protein obtained after processing by the protein structure information extraction unit in step (3), and extracts the structural information features and atomic information features of the candidate protein respectively, so as to determine the three-dimensional voxel representation of the protein to be finally submitted into the protein binding pocket prediction model.

[0032] (5) Extract protein binding pocket features. The protein binding pocket feature extraction unit further processes the mol2 format file of the protein binding pocket obtained after processing by the protein binding pocket structure information extraction unit in step (3) to extract the protein binding pocket structure information features, so as to determine the three-dimensional voxel representation of the protein binding pocket that is finally submitted into the protein binding pocket prediction model.

[0033] (6) Construct a training dataset and save the processed proteins and their binding pocket data in a folder in a one-to-one correspondence manner; convert the protein data into a 36×36×36×18 feature vector and its binding pocket into a 36×36×36×1 feature vector.

[0034] (7) Use deep learning algorithms to build a protein binding pocket prediction model, input the processed training data, and iteratively train the model using mini-batch gradient descent to determine the optimal protein binding pocket prediction model.

[0035] (8) Determine the predicted protein binding pocket. The predicted protein binding pocket determination unit predicts the protein binding pocket based on the three-dimensional voxel representation of the protein. Based on the three-dimensional voxel representation of the protein binding pocket, it calculates the similarity between the protein binding pocket and the protein binding pocket predicted by the protein binding pocket prediction model. Based on the similarity, it comprehensively evaluates the accuracy of the protein binding pocket prediction model to determine the finally predicted protein binding pocket.

[0036] (9) Process the predicted protein binding pockets. The predicted protein binding pocket processing output unit performs screening processing based on the predicted protein binding pockets in step (8) and converts them into a specified format to obtain the final predicted protein binding pockets.

[0037] The beneficial effects of this invention are that it establishes a workflow that can process protein three-dimensional structure data and build a protein pocket prediction model using deep learning algorithms to predict possible drug binding pockets of proteins. By applying deep learning methods to the study of protein three-dimensional structure binding pocket prediction, the protein binding pocket prediction device can predict its binding pocket based on the protein's positional and atomic features in three-dimensional space, greatly improving the accuracy and precision of protein binding pocket prediction and providing reliable molecular target binding information for drug design. Attached Figure Description

[0038] Figure 1 This is a schematic diagram illustrating the basic principle of the present invention;

[0039] Figure 2This is a block diagram illustrating an example of the structure of the protein binding pocket prediction device of the present invention.

[0040] Figure 3 This is a flowchart illustrating an example of the processing of the protein binding pocket prediction device of the present invention;

[0041] Figure 4 This is a schematic diagram of the data processing flow of the present invention;

[0042] Figure 5 This is a structural diagram of the protein pocket prediction model of the present invention;

[0043] Figure 6 This is a visualization of the predicted binding pocket of protein 6BQH;

[0044] Figure 7 This is a scatter plot showing the characteristic number of non-hydrogen atoms in protein 6BQH;

[0045] Figure 8 This is a scatter plot showing the characteristic number of heteroatoms in protein 6BQH;

[0046] Figure 9 This is a scatter plot showing the characteristic partial charge of protein 6BQH;

[0047] Figure 10 This is a scatter plot visualizing the predicted protein 6BQH pocket features. Detailed Implementation

[0048] The following is based on Figures 1-10 This invention will be described in detail.

[0049] It should be noted that the following embodiments can be combined where there is no conflict.

[0050] The protein binding pocket prediction device in this embodiment of the invention includes a storage unit and a control unit. The storage unit includes a protein structure information storage unit and a protein binding pocket structure information storage unit.

[0051] This embodiment only conceptually illustrates the parts of the prediction device relevant to the present invention. For example... Figure 2 As shown, the external system can be interconnected with the protein binding pocket prediction device via a network, and has the function of providing a database related to protein structure information and protein binding pocket structure information, and / or a website that executes a user interface, etc.

[0052] Furthermore, external systems can function as web (Internet) servers or ASP (Active Server Page) servers, etc. Moreover, the various functions of an external system can be implemented by its hardware structure, including the CPU, disk drive, storage device, input device, output device, communication control device, etc., and the programs that control them.

[0053] A protein binding pocket prediction device generally includes a control unit, a communication control interface unit, an input / output control interface unit, and a storage unit. Here, the control unit is a CPU or similar device that comprehensively controls the protein binding pocket prediction device; the communication control interface unit is an interface to a communication device such as a router connected to a communication line; the input / output control interface unit is an interface to a display unit and an input unit; and the storage unit is a device for storing various databases or tables. These components of the protein binding pocket prediction device are connected communicatively via any communication path. Furthermore, the protein binding pocket prediction device is communicatively connected to a network via a communication device such as a router and a wired or wireless communication line such as a dedicated line.

[0054] The storage unit contains various databases or tables (protein structure information database, protein binding pocket structure information database), which are storage units for fixed disk devices, etc. The storage unit stores various programs, tables, files, databases, and web pages used in various processes. Among the components of this storage unit, the protein structure information database is a storage unit that stores protein structure information related to the protein's structure. This protein structure information can be pre-stored in the protein structure information database. The control unit of the protein binding pocket prediction device can periodically, or according to the processing of the control unit, download the latest data from an external system via a network to update the protein structure information stored in the protein structure information database. Furthermore, the protein binding pocket structure information database is a storage unit that stores the structural information of binding sites, binding pockets, and ligands that can bind to proteins. This protein binding pocket structure information can be pre-stored in the protein binding pocket structure information database. The control unit of the protein binding pocket prediction device can periodically, or according to the processing of the control unit, download the latest data from an external system via a network to update the protein binding pocket structure information stored in the protein binding pocket structure information database.

[0055] In addition, see Figure 2 The input / output control interface section controls the display section and the input section. The display section can be a display unit that displays the screen of applications, etc., and the input section can be a key input section, control panel, mouse, keyboard, etc.

[0056] The control unit has internal memory for storing control programs such as the OS (Operating System), programs that define various processing steps, and necessary data. The control unit performs information processing according to these programs to execute various processes. Functionally, the control unit includes a protein structure information extraction unit, a protein binding pocket information extraction unit, a protein feature extraction unit, a protein binding pocket feature extraction unit, a predicted protein binding pocket determination unit, and a predicted protein binding pocket processing output unit.

[0057] In this embodiment, the protein structure information storage unit is used to store protein structure information related to the protein's structure; the protein binding pocket structure information storage unit is used to store protein binding pocket structure information related to the protein binding pocket structure. The protein binding pocket structure information includes the protein binding site, the protein binding cavity, and the protein-binding ligand.

[0058] In this embodiment, the protein structure information extraction unit is used to obtain the candidate protein structure information as a candidate protein from the protein structure information storage unit, wherein the candidate protein is a candidate protein predicted as a protein binding pocket.

[0059] Specifically, the protein structure information extraction unit obtains the structure information of candidate proteins from the protein structure information database, acquires the three-dimensional structure file of the protein, preprocesses the protein data to remove water molecules and impurities, examines the amino acid residues in the protein and retains 20 standard amino acids, and can obtain the structure information of candidate proteins, which is a three-dimensional structure file of the protein in pdb format.

[0060] It should be understood that the 20 standard amino acids include glycine, alanine, valine, leucine, isoleucine, methionine, proline, tryptophan, serine, tyrosine, cysteine, phenylalanine, asparagine, glutamine, threonine, aspartic acid, glutamic acid, lysine, arginine, and histidine.

[0061] In this embodiment, the protein binding pocket information extraction unit is used to obtain protein binding pocket structural information from the protein binding pocket information storage unit. The protein binding pocket structural information is a three-dimensional structural file of the binding pocket in PDB format.

[0062] Furthermore, the protein structure information extraction unit converts the 3D structure file of the protein in PDB format into a mol2 format file of the protein structure; the protein binding pocket information extraction unit converts the 3D structure file of the binding pocket in PDB format into a mol2 format file of the protein binding pocket.

[0063] Specifically, based on the 3D structure files of proteins and binding pockets in PDB format, atomic information starting with ATOM is extracted, yielding information such as atom number, atom type, base sequence or amino acid sequence, chain number, nucleotide number, XYZ 3D Cartesian coordinates (in Å), and atom name. This information is then converted into a mol2 format file starting with @ <tripos>The information starting with ATOM includes the atom number, atom name, XYZ three-dimensional rectangular coordinates (in angstroms), atom type, residue number, residue name, and charge. Then, the connection information between atoms starting with CONECT is extracted from the PDB file and converted to the corresponding @ symbol in the mol2 file. <tripos>The binding information starting with BOND includes the binding sequence number, starting atom number, target atom number, and binding type. The binding type can be a single bond, double bond, triple bond, amide bond, or aromatic bond.

[0064] In this embodiment, the protein feature extraction unit can obtain candidate proteins based on the protein structure information extraction unit, extract the structural information features and atomic information features of the candidate proteins, and thus determine the three-dimensional voxel representation of the protein to be finally submitted into the protein binding pocket prediction model.

[0065] Furthermore, the protein feature extraction unit performs feature extraction processing on the mol2 format files of candidate proteins, extracting structural information features and atomic information features of the proteins respectively. The structural information features correspond to the positional features of the proteins, i.e., the atomic coordinates of the proteins. The positional features are represented by a two-dimensional array N*3, where N is the number of atoms in the molecule and 3 is the three-dimensional coordinates (x, y, z) of the atoms. The atomic information features include atom type, PyBel atom features, and SMARTS features.

[0066] It should be noted that the atom types include: 'B (boron)', 'C (carbon)', 'N (nitrogen)', 'O (oxygen)', 'P (phosphorus)', 'S (sulfur)', 'Se (selenium)', 'halogen (halogen elements: fluorine, chlorine, bromine, iodine)', and 'metal (metal elements: lithium, beryllium, sodium, magnesium, etc.)'. Atom types are encoded as binary vectors based on their atomic numbers. Atoms within the defined atom types are marked as 1, and those not within the defined atom types are marked as 0. Pybel atomic characteristics include: 'hyb' (hybridization: sp, sp2, sp3, etc.), 'heavyvalence' (number of non-hydrogen atoms), 'heterovalence' (number of heteroatoms), and 'partial charge'. SMARTS features include 'hydrophobic', 'aromatic', 'acceptor', 'donor', and 'ring'. SMARTS features are encoded as a two-dimensional array N*M, where N is the number of atoms in the molecule and M is the number of SMARTS features. The generated feature is represented as [i,j], which is the j-th attribute of the i-th atom. The atomic features in the molecule are matched with the defined SMARTS features. An atom in the molecule belonging to a defined SMARTS feature is marked as 1, and an atom not belonging to a defined SMARTS feature is marked as 0.

[0067] Optionally, the protein feature extraction unit converts the protein's mol2 format file into a three-dimensional voxel representation of the protein.

[0068] In this embodiment, the protein binding pocket feature extraction unit can obtain the protein binding pocket according to the protein binding pocket structure information extraction unit, extract the protein binding pocket information features, and thus determine the three-dimensional voxel representation of the protein binding pocket to be finally submitted into the protein binding pocket prediction model.

[0069] Optionally, the protein binding pocket feature extraction unit converts the mol2 format file of the protein binding pocket into a three-dimensional voxel representation of the protein binding pocket.

[0070] Furthermore, based on the protein structure information extraction unit, protein feature extraction unit, protein binding pocket information extraction unit, and protein binding pocket feature extraction unit, protein structure information features, atomic information features, and protein pocket structure information features are obtained, and then a training dataset is constructed.

[0071] Specifically, the processed proteins and their binding pocket data are saved in a folder, with one protein corresponding to one ligand or its binding cavity and binding site. Then, a 36×36×36 3D box is placed at the center of the protein, with a distance of [missing information] in each direction. If the distance between the atoms in the molecule and the center point exceeds If an atom is outside the defined box, the out-of-bounds atom is removed from the molecule, generating new coordinate information for the protein and its binding pocket. Finally, based on the new coordinate information and feature information, a voxel grid file of the protein is generated, with a shape of 36×36×36×18, essentially converting the protein data into a 36×36×36×18 feature vector. Similarly, the protein binding pocket structure information extraction unit obtains candidate protein binding pocket structure information from the protein binding pocket structure information database. The protein binding pocket feature extraction unit generates a voxel grid file of the binding pocket based on the structure information. Each binding pocket is represented as a 36×36×36×1 tensor, converting the protein binding pocket into a 36×36×36×1 tensor. In space, points representing protein binding pockets are marked as 1, otherwise 0.

[0072] In this embodiment, the protein binding pocket prediction unit can predict protein binding pockets using deep learning methods based on the three-dimensional voxel representation of the protein. Based on the three-dimensional voxel representation of the protein binding pocket, it calculates the similarity between the protein binding pocket and the protein binding pocket predicted by the protein binding pocket prediction model. Based on the similarity, it comprehensively evaluates the accuracy of the protein binding pocket prediction model, thereby determining the finally predicted protein binding pocket.

[0073] In this embodiment, a protein binding pocket prediction model is built using a deep learning algorithm. The processed training data is input, and the protein binding pocket prediction model is trained iteratively using mini-batch gradient descent. Based on the loss function, a set similarity metric function is used to calculate the similarity between the predicted value and the true value of the protein binding pocket prediction model.

[0074] The protein binding pocket prediction model comprises an encoder and a decoder, and borrows the concept of residual connections. This design is essential for building the protein binding pocket prediction model. The encoder consists of 16 3D convolution operations, 16 PReLU activation operations, and 5 residual connection operations; the decoder consists of 4 3D transposed convolution operations, 19 3D convolution operations, 21 PReLU activation operations, 9 regularization operations, and 1 sigmoid activation operation.

[0075] In addition, the loss function can be expressed as:

[0076]

[0077] Where y is the model prediction value, t is the true value, i, j, k represent the three-dimensional coordinates of the atom, and ε is the interference factor.

[0078] In this way, the trained protein binding pocket model can be used to predict the protein binding pocket. Select the protein to be predicted, process it into a 36×36×36×18 feature vector using the same data processing method as described above, and input it into the protein binding pocket model. Then, the protein binding pocket model outputs multiple 36×36×36×1 tensors as the prediction results.

[0079] In this embodiment, the protein binding pocket prediction processing output unit can perform screening processing based on the predicted protein binding pockets and convert them into a specified format to obtain the final predicted protein binding pockets.

[0080] Specifically, the protein binding pocket prediction processing output unit processes and transforms the binding pockets output by the protein binding pocket prediction model, determines the size of the predicted binding pockets, and defines a minimum size for the bound pocket. The combined pocket size is smaller than The process involves deleting the predicted binding pockets, creating new binding pocket molecules, iterating through each value in turn based on the predicted binding pockets, adding coordinate information to the atoms in the molecule according to the index of each value of the binding pocket, and finally outputting the predicted binding pockets according to the specified output format.

[0081] like Figure 1 As shown, in this embodiment, the control unit of the protein binding pocket prediction device obtains the protein structure information desired by the user from the storage unit, or obtains the protein structure information that needs to be predicted for protein binding pockets that is not stored in the storage unit from an external protein database.

[0082] Next, the obtained candidate protein structures are cleaned, the data format is converted to mol2 format, and the positional and atomic features of the protein data are extracted. The protein data is then saved as an HDF file. The protein input to the model is three-dimensional structural information, which needs to be converted into a three-dimensional voxel representation. Then, the protein features are used as the number of channels to form a 36×36×36×18 feature vector.

[0083] The control unit of the protein binding pocket prediction device uses deep learning to predict potential binding pockets of proteins based on the determined three-dimensional voxel information of the protein. In addition, the control unit can also use the protein binding pocket structure information stored in the storage unit to evaluate the prediction results.

[0084] Finally, the shape and size of the predicted protein binding pockets are screened and converted into mol2 file format for output, thus determining the information of the predicted protein binding pockets.

[0085] It is worth mentioning that the embodiments of the present invention also provide a method for predicting protein binding pockets.

[0086] The prediction method uses a predictive device based on the protein binding pocket, and specifically includes the following steps:

[0087] (1) Extracting protein structure information. The protein structure information extraction unit obtains the candidate protein structure information from the protein structure information storage unit, obtains the protein three-dimensional structure file, i.e., the protein three-dimensional structure file in pdb format, preprocesses the protein data, removes water molecules and impurities from the protein molecules, examines the amino acid residues in the protein and retains 20 standard amino acids.

[0088] (2) Extracting protein binding pocket information. The protein binding pocket information extraction unit obtains the protein binding pocket structure information from the protein binding pocket information storage unit and obtains the three-dimensional structure file of the protein binding pocket, that is, the three-dimensional structure file of the protein binding pocket in pdb format.

[0089] (3) The protein structure information extraction unit extracts the atomic information of the protein from the three-dimensional structure file of the protein in pdb format and converts it into a mol2 format file of the protein structure; the protein binding pocket information extraction unit extracts the atomic information of the binding pocket from the three-dimensional structure file of the binding pocket in pdb format and converts it into a mol2 format file of the protein binding pocket.

[0090] Specifically, based on the 3D structure files of proteins and binding pockets in PDB format, atomic information starting with ATOM is extracted, yielding information such as atom number, atom type, base sequence or amino acid sequence, chain number, nucleotide number, XYZ 3D Cartesian coordinates (in Å), and atom name. This information is then converted into a mol2 format file starting with @ <tripos>The information starting with ATOM includes the atom number, atom name, XYZ three-dimensional rectangular coordinates (in angstroms), atom type, residue number, residue name, and charge. Then, the connection information between atoms starting with CONECT is extracted from the PDB file and converted to the corresponding @ symbol in the mol2 file. <tripos>The binding information starting with BOND includes the binding sequence number, the starting atom number, the target atom number, and the binding type.

[0091] (4) Protein feature extraction. The protein feature extraction unit further processes the mol2 files of candidate proteins obtained after processing by the protein structure information extraction unit in step (3), extracting structural information features and atomic information features of the candidate proteins to determine the final three-dimensional voxel representation of the protein to be submitted into the protein binding pocket prediction model. Atomic coordinates are extracted from the mol2 files of the candidate protein molecules as their positional features. The positional features are two-dimensional arrays N*3, where N is the number of atoms in the molecule and 3 is the three-dimensional coordinates (x, y, z) of the atom. The atomic features of the candidate proteins include atomic type features, Pybel atomic features, and SMARTS features.

[0092] (5) Extract protein binding pocket features. The protein binding pocket feature extraction unit further processes the mol2 file of the protein binding pocket obtained after processing by the protein binding pocket structure information extraction unit in step (3), extracts the protein binding pocket structure information features, and determines the three-dimensional voxel representation of the protein binding pocket to be finally submitted into the protein binding pocket prediction model.

[0093] (6) Construct the training dataset. Save the processed protein and binding site data in a folder, with one protein corresponding to one ligand or its binding cavity and binding site. Then place a 36×36×36 3D box at the center of the protein, with a distance of [missing information] in each direction. If the distance between the atoms in the molecule and the center point exceeds If an atom is outside the defined box, the out-of-bounds atom is removed from the molecule, generating new coordinate information for the protein and its binding pocket. Finally, based on the new coordinate information and feature information, a voxel grid file of the protein is generated, with a shape of 36×36×36×18. Similarly, the protein binding pocket structure information extraction unit obtains candidate protein binding pocket structure information from the protein binding pocket structure information database. The protein binding pocket feature extraction unit generates a voxel grid file of the binding pocket based on the structure information. Each binding pocket is represented as a 36×36×36×1 tensor, where points that are protein binding pockets are marked as 1, and others as 0.

[0094] (7) Use deep learning algorithms to build a protein binding pocket prediction model, input the processed training data, and iteratively train the model using mini-batch gradient descent.

[0095] (7.1) The processed training data is used as input (36×36×36×18), and the labels are ti,j,k (36×36×36×1) and input into the model.

[0096] (7.2) The input tensor enters the encoder part. First, it undergoes a 3D convolution operation with a stride of 1, padding = 'same', 16 convolution kernels, and a kernel size of [5×5×5] to transform the input tensor into a shape of (36×36×36×16). Then, it undergoes a PReLU activation operation, and the shape of the tensor remains unchanged.

[0097] (7.3) Concatenate the tensor output in step (7.2) with the input tensor input to obtain a tensor with a shape of (36×36×36×34).

[0098] (7.4) Perform a 3D convolution operation on the spliced ​​tensor with a stride of 2, padding = 'valid', 32 kernels, and kernel size of [2×2×1]. The tensor shape becomes (18×18×18×32). Then perform a PReLU activation operation, and the tensor shape remains unchanged.

[0099] (7.5) Perform a 3D convolution operation on the tensor output in step (7.4) with a stride of 1, padding = 'same', 32 kernels, and a kernel size of [5×5×5], to output a tensor with a shape of (18×18×18×32). Then perform a PReLU activation operation, and the tensor shape remains unchanged. Then repeat the above convolution activation operation, and the tensor shape is still (18×18×18×32).

[0100] (7.6) Concatenate the tensor output in step (7.4) with the tensor output in step (7.5) to obtain a tensor with a shape of (18×18×18×64).

[0101] (7.7) Perform a 3D convolution operation on the spliced ​​tensor with a stride of 2, padding = 'valid', 64 kernels, and a kernel size of [2×2×1]. The tensor shape becomes (9×9×9×64). Then perform a PReLU activation operation. The tensor shape remains unchanged.

[0102] (7.8) Perform a 3D convolution operation on the tensor output in step (7.7) with a stride of 1, padding = 'same', 64 kernels, and a kernel size of [5×5×5], to output a tensor with a shape of (9×9×9×64). Then perform a PReLU activation operation, and the tensor shape remains unchanged. Then repeat the above convolution activation operation twice, and the tensor shape is still (9×9×9×64).

[0103] (7.9) Concatenate the tensor output in step (7.7) with the tensor output in step (7.8) to obtain a tensor with a shape of (9×9×9×128).

[0104] (7.10) Perform a 3D convolution operation on the spliced ​​tensor with a stride of 2, padding = 'valid', 128 kernels, and a kernel size of [2×2×2]. The tensor shape becomes (4×4×4×128). Then perform a PReLU activation operation, and the tensor shape remains unchanged.

[0105] (7.11) Perform a 3D convolution operation on the tensor output in step (7.10) with a stride of 1, padding = 'same', 128 kernels, and a kernel size of [5×5×5], to output a tensor with a shape of (4×4×4×128). Then perform a PReLU activation operation, and the shape of the tensor remains unchanged. Then repeat the above convolution activation operation twice, and the shape of the tensor is still (4×4×4×128).

[0106] (7.12) Concatenate the tensor output in step (7.10) with the tensor output in step (7.11) to obtain a tensor with a shape of (4×4×4×256).

[0107] (7.13) Perform a 3D convolution operation on the spliced ​​tensor with a stride of 2, padding = 'valid', 256 kernels, and a kernel size of [2×2×1]. The tensor shape becomes (2×2×2×256). Then perform a PReLU activation operation. The tensor shape remains unchanged.

[0108] (7.14) Perform a 3D convolution operation on the tensor output in step (7.13) with a stride of 1, padding = 'same', 256 kernels, and a kernel size of [5×5×5], to output a tensor with a shape of (2×2×2×256). Then perform a PReLU activation operation, and the shape of the tensor remains unchanged. Then repeat the above convolution activation operation twice, and the shape of the tensor is still (2×2×2×256).

[0109] (7.15) Concatenate the tensor output in step (7.13) with the tensor output in step (7.14) to obtain a tensor with the shape (2×2×2×512).

[0110] (7.16) The input tensor enters the decoder section. First, the tensor output in step (7.15) is subjected to a 3D transpose convolution operation with a stride of 2, padding = 'valid', 256 kernels, and a kernel size of [2×2×1]. The tensor shape becomes (4×4×4×256). Then, it undergoes a PReLU activation operation, and the tensor shape remains unchanged.

[0111] (7.17) Perform a 3D convolution operation on the tensor output in step (7.12) with a stride of 1, padding = 'valid', 256 kernels, and a kernel size of [1×1×1] to obtain a tensor with a shape of (4×4×4×256). Then, perform regularization and PReLU activation operations to keep the tensor shape unchanged.

[0112] (7.18) Perform a 3D convolution operation on the tensor output in step (7.17) with a stride of 1, padding = 'same', 256 kernels, and a kernel size of [3×3×3] to obtain a tensor with a shape of (4×4×4×256). Then, perform regularization and PReLU activation operations to keep the tensor shape unchanged.

[0113] (7.19) Perform a 3D convolution operation on the tensor output in step (7.18) with a stride of 1, padding = 'valid', 256 kernels, and a kernel size of [1×1×1] to obtain a tensor with a shape of (4×4×4×256). Then perform a regularization operation. Finally, add the obtained tensor to the corresponding element values ​​of the tensor output in step (7.12). Finally, perform a PReLU activation operation. The shape of the tensor remains unchanged.

[0114] (7.20) Concatenate the tensor output in step (7.16) with the tensor output in step (7.19) to obtain a tensor with a shape of (4×4×4×512).

[0115] (7.21) Perform a 3D convolution operation on the tensor output in step (7.20) with a stride of 1, padding = 'same', 256 kernels, and a kernel size of [5×5×5], to output a tensor with a shape of (4×4×4×256). Then perform a PReLU activation operation, and the tensor shape remains unchanged. Then repeat the above convolution activation operation twice, and the tensor shape is still (4×4×4×256).

[0116] (7.22) Concatenate the tensor output in step (7.16) with the tensor output in step (7.21) to obtain a tensor with a shape of (4×4×4×512).

[0117] (7.23) Perform a 3D transpose convolution operation on the tensor output in step (7.20) with a stride of 2, padding = 'valid', 128 kernels, and kernel size of [3×3×3]. The tensor shape becomes (9×9×9×128). Then perform a PReLU activation operation, and the tensor shape remains unchanged.

[0118] (7.24) Perform a 3D convolution operation on the tensor output in step (7.9) with a stride of 1, padding = 'valid', 128 kernels, and kernel size of [1×1×1] to obtain a tensor with a shape of (9×9×9×128). Then perform regularization and PReLU activation operations to keep the tensor shape unchanged.

[0119] (7.25) Perform a 3D convolution operation on the tensor output in step (7.24) with a stride of 1, padding = 'same', 128 kernels, and a kernel size of [3×3×3] to obtain a tensor with a shape of (9×9×9×128). Then perform regularization and PReLU activation operations to keep the tensor shape unchanged.

[0120] (7.26) Perform a 3D convolution operation on the tensor output in step (7.25) with a stride of 1, padding = 'valid', 128 kernels, and kernel size of [1×1×1] to obtain a tensor with a shape of (9×9×9×128). Then perform a regularization operation. Finally, add the corresponding element values ​​of the tensor output in step (7.9) to the tensor. Finally, perform a PReLU activation operation. The shape of the tensor remains unchanged.

[0121] (7.27) Concatenate the tensor output in step (7.23) with the tensor output in step (7.26) to obtain a tensor with a shape of (9×9×9×256).

[0122] (7.28) Perform a 3D convolution operation on the tensor output in step (7.27) with a stride of 1, padding = 'same', 128 kernels, and a kernel size of [5×5×5], to output a tensor with a shape of (9×9×9×128). Then perform a PReLU activation operation, and the shape of the tensor remains unchanged. Then repeat the above convolution activation operation twice, and the shape of the tensor is still (9×9×9×128).

[0123] (7.29) Concatenate the tensor output in step (7.23) with the tensor output in step (7.28) to obtain a tensor with a shape of (9×9×9×256).

[0124] (7.30) Perform a 3D transpose convolution operation on the tensor output in step (7.29) with a stride of 2, padding = 'valid', 64 kernels, and kernel size of [2×2×1]. The tensor shape becomes (18×18×18×64). Then perform a PReLU activation operation, and the tensor shape remains unchanged.

[0125] (7.31) Perform a 3D convolution operation on the tensor output in step (7.6) with a stride of 1, padding = 'valid', 64 kernels, and a kernel size of [1×1×1] to obtain a tensor with a shape of (18×18×18×64). Then, perform regularization and PReLU activation operations to keep the tensor shape unchanged.

[0126] (7.32) Perform a 3D convolution operation on the tensor output in step (7.31) with a stride of 1, padding = 'same', 64 kernels, and a kernel size of [3×3×3] to obtain a tensor with a shape of (18×18×18×64). Then perform regularization and PReLU activation operations to keep the tensor shape unchanged.

[0127] (7.33) Perform a 3D convolution operation on the tensor output in step (7.32) with a stride of 1, padding = 'valid', 64 kernels, and a kernel size of [1×1×1] to obtain a tensor with a shape of (18×18×18×64). Then perform a regularization operation. Finally, add the obtained tensor to the corresponding element values ​​of the tensor output in step (7.9). Finally, perform a PReLU activation operation. The shape of the tensor remains unchanged.

[0128] (7.34) Concatenate the tensor output in step (7.30) with the tensor output in step (7.33) to obtain a tensor with a shape of (18×18×18×128).

[0129] (7.35) Perform a 3D convolution operation on the tensor output in step (7.34) with a stride of 1, padding = 'same', 64 kernels, and a kernel size of [5×5×5], to output a tensor with a shape of (18×18×18×64). Then perform a PReLU activation operation, and the shape of the tensor remains unchanged. Then repeat the above convolution activation operation, and the shape of the tensor is still (18×18×18×64).

[0130] (7.36) Concatenate the tensor output in step (7.30) with the tensor output in step (7.35) to obtain a tensor with a shape of (18×18×18×128).

[0131] (7.37) Perform a 3D transpose convolution operation on the tensor output in step (7.36) with a stride of 2, padding = 'valid', 32 kernels, and kernel size of [2×2×1]. The tensor shape becomes (36×36×36×32). Then perform a PReLU activation operation, and the tensor shape remains unchanged.

[0132] (7.38) Concatenate the tensor output in step (7.35) with the tensor output in step (7.37) to obtain a tensor with a shape of (36×36×36×64).

[0133] (7.39) Perform a 3D convolution operation on the tensor output in step (7.38) with a stride of 1, padding = 'valid', 1 kernel, and kernel size of [1×1×1]. Then activate it with the sigmoid function. Finally, output tensor data with shape (36×36×36×1), denoted as yi,jk.

[0134] (7.40) Calculate the loss function C(y,t) based on yi,jk and ti,jk. If it does not meet the requirements, perform backpropagation and update the parameter weights of each layer of the model. Repeat the above steps and iterate continuously until the model fits.

[0135] (7.41) Based on the evaluation metrics of the model, select the model with the highest scores across all metrics as the optimal model, and save the optimal model as the prediction model for the protein binding pocket. It should be understood that the evaluation metrics include, but are not limited to: accuracy, precision, recall, F1 score, etc.

[0136] It should be noted that the F-measure is a statistic, also known as the F-score. The F-measure is a weighted harmonic mean of precision and recall, and is a commonly used evaluation metric in the field of information retrieval (IR), often used to assess the performance of a model. In the F-measure function, when the parameter α = 1, F1 combines the results of precision and recall; a higher F1 indicates that the experimental method is more effective. In this embodiment, the F1 measurement value is the F1-measure, an evaluation index that is a comprehensive assessment based on both precision and recall.

[0137] (8) Determine the predicted protein binding pocket. The protein binding pocket determination unit predicts the protein binding pocket based on the three-dimensional voxel representation of the protein. Based on the three-dimensional voxel representation of the protein binding pocket, it calculates the similarity between the protein binding pocket and the protein binding pocket predicted by the protein binding pocket prediction model. Based on the similarity, it comprehensively evaluates the accuracy of the protein binding pocket prediction model to determine the final predicted protein binding pocket.

[0138] (9) Process the predicted protein binding pockets. The predicted protein binding pocket processing output unit performs screening processing on the predicted protein binding pockets in step (8) and converts them into a specified format to obtain the final predicted protein binding pockets.

[0139] Specifically, the predicted size of the connecting pocket is determined, and the minimum size of the custom connecting pocket is set to... Pocket size smaller than The process involves deleting the predicted binding pockets, creating new binding pocket molecules, iterating through each value in turn based on the predicted binding pockets, adding coordinate information to the atoms in the molecule according to the index of each value of the binding pocket, and finally outputting the predicted binding pockets according to the specified output format.

[0140] In this embodiment, when the user performs protein binding pocket prediction, such as Figure 3 As shown, the protein structure information to be predicted needs to be provided first. Then, the obtained candidate protein structures are cleaned, the data format is converted to mol2 format, the positional and atomic features of the protein data are extracted, and the protein data is saved as an hdf file again.

[0141] Opening the processed protein 3D structure file with PyMol software reveals that water molecules and other impurities have been removed from the protein structure, and only 20 standard amino acids remain.

[0142] Furthermore, the protein mol2 format is converted into a three-dimensional voxel representation, and the coordinates of each atom in the protein are mapped to the coordinates of each grid point in the voxel grid, constructing a 36×36×36 grid file. The extracted atomic features are mapped to the positional features in the voxel grid, finally generating a 36×36×36×18 feature vector, which is then input into the protein binding pocket prediction model to predict the binding pocket.

[0143] The prediction is performed using a deep learning algorithm. The previously trained model is loaded, and the input 36×36×36×18 protein feature vector is combined with pocket prediction. The training steps are the same as the deep learning protein pocket prediction model training method described above.

[0144] The predicted protein binding pockets were screened by shape and size, and then converted into mol2 file format for output, ultimately determining the information of the predicted protein binding pockets. The predicted binding pocket files and their corresponding 3D protein structure files were opened using PyMol software, allowing for visualization of the predicted structures. This clearly shows the specific locations of the binding pockets and proteins in 3D space, as well as their spatial relationships.

[0145] The technical solution of the present invention will now be further described in conjunction with the accompanying drawings and embodiments.

[0146] Example 1:

[0147] Step 1: Download the public dataset scPDB of proteins and their binding pockets. The dataset contains the three-dimensional structure files of proteins and their corresponding ligands, binding cavities and binding sites, and they correspond one-to-one.

[0148] Step 2: Following the data processing method described above, such as... Figure 4 As shown, water molecules and impurities are removed from the protein file, the amino acid residues in the protein are examined and 20 standard amino acids are retained, and the data is saved as an hdf file.

[0149] Step 3: Extract the features of the processed protein file and any of its combined pocket files, including the 3D position information and atomic information, to form 36×36×36×18 and 36×36×36×1 feature vectors respectively, and construct the training dataset accordingly.

[0150] Step 4: Feed the training dataset into the pre-built protein binding pocket prediction model in batches of ten data points for training. Iterate continuously as required until the model reaches its optimal performance. The model is as follows: Figure 5 As shown;

[0151] Step 5: Save the trained protein binding pocket prediction model to predict protein binding pockets. Obtain the 3D structure file of the protein whose pocket needs to be predicted, and perform the same data processing as described above to obtain its 36×36×36×18 feature vector. Input this protein feature vector into the saved model, and the model will output multiple 36×36×36×1 feature vectors, which are the predicted binding pockets. Then delete those smaller than 18. The pocket is then iterated through each value in the eigenvector that meets the requirements, and coordinate information is added to the atoms in the molecule according to their index. Finally, the output is in the format of a mol2 file.

[0152] Step 6: Visualize the prediction results. Open the binding pocket file and its protein 3D structure file predicted by the model using PyMol software. You can clearly see the specific location of the binding pocket and protein in 3D space, as well as their spatial relationship, such as... Figure 6 As shown.

[0153] Example 2:

[0154] Step 1: Select the protein to be predicted. Protein 6BQH is one of the receptors for serotonin (5-HT), belonging to the 5-HT2 subtype of the serotonin receptor. 5-HT2A receptors are distributed throughout the central nervous system (CNS). This is largely due to the abundance of serotonin terminals in areas including the cerebral cortex (primarily the prefrontal, parietal, and somatosensory cortex) and the olfactory tubercle region. In particular, this high concentration of receptors on the top dendrites of pyramidal cells in layer 5 of the cerebral cortex can regulate cognitive processes, working memory, and interact with 5-HT1A, GABAA, adenosine A1, AMPA, mGlu5, and OX2 receptors by enhancing glutamate release. It is also highly present in the peripheral nervous system, platelets, many cardiovascular cells, fibroblasts, and neurons.

[0155] Step 2: Obtain the 3D structure file of protein 6BQH. The 3D structure data in PDB format can be obtained from the protein database PDB. Water molecules, heteroatoms, and non-standard amino acid residues are then removed. Finally, the PDB format is converted to Mol2 format.

[0156] Step 3: Extract the protein's feature information and convert it into a feature vector for the input model. Read the protein file information, retain its three-dimensional position information and atomic feature information, and use this to generate a 36×36×36×18 feature vector. See details below. Figures 7-9 .

[0157] Step 4: Input the processed feature vector into the protein pocket prediction model trained in Example 1 above to obtain the model's prediction result. The output is a 36×36×36×1 tensor, which can be visualized as follows: Figure 10 .

[0158] Step 5: Convert the output of the above model into a format that can be opened by PyMOL software, such as PDB or MOLU2. The visualization image of the binding pocket of protein 6BQH is shown below. Figure 6 As shown.

[0159] This invention establishes a workflow for processing protein 3D structure data and using deep learning algorithms to build protein pocket prediction models to predict potential drug-binding pockets. Applying deep learning methods to protein 3D structure binding pocket prediction significantly improves the accuracy and precision of traditional protein binding pocket detection methods based on geometry, energy, and biochemistry. The protein pocket prediction device of this invention can predict the binding pocket of a protein based on its positional and atomic features in three-dimensional space, providing reliable molecular target binding information for drug design.

[0160] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.< / tripos> < / tripos> < / tripos> < / tripos>

Claims

1. A device for predicting protein binding pockets, comprising a storage unit and a control unit, characterized in that, The storage unit includes: Protein structure information storage unit, used to store protein structure information related to the protein's structure; and A protein binding pocket structure information storage unit is used to store protein binding pocket structure information related to the protein binding pocket structure. The control unit includes: The protein structure information extraction unit is used to obtain the structure information of candidate proteins as candidate proteins from the protein structure information storage unit. The candidate proteins are those predicted as candidates for protein binding pockets. The protein structure information extraction unit converts the three-dimensional structure file of the protein in pdb format into a mol2 format file of the protein structure. The protein binding pocket information extraction unit is used to obtain the protein binding pocket structure information from the protein binding pocket information storage unit; the protein binding pocket information extraction unit converts the three-dimensional structure file of the binding pocket in pdb format into a mol2 format file of the protein binding pocket; The protein feature extraction unit obtains the candidate protein based on the protein structure information extraction unit, and extracts the structural information features and atomic information features of the candidate protein to determine the final three-dimensional voxel representation of the protein to be submitted into the protein binding pocket prediction model. The protein feature extraction unit performs feature extraction processing on the mol2 format file of the candidate protein, extracting the structural information features and atomic information features of the protein, including atom type, PyBel atomic features, and SMARTS features. The protein feature extraction unit converts the mol2 format file of the protein into a three-dimensional voxel representation of the protein, i.e., generating a 36×36×36×18 feature vector. The 18 channels of this feature vector correspond to 9 atom types, 4 PyBel atomic features, and 5 SMARTS features. The protein binding pocket feature extraction unit obtains the protein binding pocket based on the protein binding pocket structure information extraction unit, and performs protein binding pocket information feature extraction on the protein binding pocket to determine the final three-dimensional voxel representation of the protein binding pocket to be submitted into the protein binding pocket prediction model; the protein binding pocket feature extraction unit converts the mol2 format file of the protein binding pocket into the three-dimensional voxel representation of the protein binding pocket. The protein binding pocket prediction unit predicts protein binding pockets using a deep learning method based on the protein's three-dimensional voxel representation. It then calculates the similarity between the predicted protein binding pocket and the protein binding pocket predicted by the protein binding pocket prediction model, and comprehensively evaluates the accuracy of the prediction model based on this similarity to determine the final predicted protein binding pocket. The protein binding pocket prediction processing output unit performs screening processing on the predicted protein binding pockets and converts them into a specified format to obtain the final predicted protein binding pockets. The protein binding pocket prediction model includes an encoder and a decoder. The protein binding pocket prediction model is configured to receive a 36×36×36×18 feature vector input. The encoder includes 16 3D convolution operations, 16 PReLU activation operations, and 5 residual connection operations. The decoder includes 4 3D transposed convolution operations, 19 3D convolution operations, 21 PReLU activation operations, 9 regularization operations, and 1 sigmoid activation operation. The protein binding pocket prediction model is built using a deep learning algorithm. Processed training data is input, and the model is iteratively trained using mini-batch gradient descent. Based on a loss function, a set similarity metric is used to calculate the similarity between the predicted and actual values ​​of the protein binding pocket prediction model. The loss function is: In the formula, Let be the loss function, y be the model prediction, t be the true value, and i, j, k represent the three-dimensional coordinates of the atom. It is an interference factor.

2. The protein binding pocket prediction device according to claim 1, characterized in that, The protein binding pocket structural information includes the protein binding site, the protein binding cavity, and the protein-binding ligand.

3. The protein binding pocket prediction device according to claim 1, characterized in that, The protein structure information extraction unit preprocesses the protein data, removing water molecules and other impurities, and retaining 20 standard amino acids to obtain the candidate protein structure information, which is a three-dimensional structure file of the protein in PDB format.

4. The protein binding pocket prediction device according to claim 1, characterized in that, The protein binding pocket structure information is a three-dimensional structure file of the binding pocket in pdb format.

5. A method for predicting protein binding pockets, said prediction method using the protein binding pocket prediction device according to any one of claims 1-4, characterized in that, The prediction method includes the following steps: (1) Extracting protein structure information: The protein structure information extraction unit obtains the candidate protein structure information, i.e., the three-dimensional structure file of the protein in pdb format, from the protein structure information storage unit, preprocesses the protein data, removes water molecules and impurities from the protein molecules, examines the amino acid residues in the protein and retains 20 standard amino acids. (2) Extracting protein binding pocket information, wherein the protein binding pocket information extraction unit obtains the protein binding pocket structure information from the protein binding pocket information storage unit, namely, a three-dimensional structure file of the protein binding pocket in pdb format; (3) The protein structure information extraction unit extracts the atomic information of the protein according to the three-dimensional structure file of the protein in pdb format and converts it into a mol2 format file of the protein structure; the protein binding pocket information extraction unit extracts the atomic information of the binding pocket according to the three-dimensional structure file of the binding pocket in pdb format and converts it into a mol2 format file of the protein binding pocket. (4) Extracting protein features: The protein feature extraction unit further processes the mol2 format file of the candidate protein obtained after processing by the protein structure information extraction unit in step (3), and extracts the structural information features and atomic information features of the candidate protein respectively, so as to determine the three-dimensional voxel representation of the protein to be finally submitted into the protein binding pocket prediction model. (5) Extract protein binding pocket features. The protein binding pocket feature extraction unit further processes the mol2 format file of the protein binding pocket obtained after processing by the protein binding pocket structure information extraction unit in step (3) to extract the protein binding pocket structure information features, so as to determine the three-dimensional voxel representation of the protein binding pocket that is finally submitted to the protein binding pocket prediction model. (6) Construct a training dataset and save the processed proteins and their binding pocket data in a folder in a one-to-one correspondence manner; convert the protein data into a 36×36×36×18 feature vector and its binding pocket into a 36×36×36×1 feature vector. (7) Use deep learning algorithms to build a protein binding pocket prediction model, input the processed training data, and iteratively train the model using mini-batch gradient descent to determine the optimal protein binding pocket prediction model. (8) Determine the predicted protein binding pocket. The predicted protein binding pocket determination unit predicts the protein binding pocket based on the three-dimensional voxel representation of the protein. Based on the three-dimensional voxel representation of the protein binding pocket, it calculates the similarity between the protein binding pocket and the protein binding pocket predicted by the protein binding pocket prediction model. Based on the similarity, it comprehensively evaluates the accuracy of the protein binding pocket prediction model to determine the finally predicted protein binding pocket. (9) Process the predicted protein binding pockets. The predicted protein binding pocket processing output unit performs screening processing based on the predicted protein binding pockets in step (8) and converts them into a specified format to obtain the final predicted protein binding pockets.

Images