A protein allosteric site identification system based on deep learning

Abstract

The present invention discloses a protein allosteric site identification system based on deep learning. The system comprises: a data import module, which loads the PDB file of proteins with allosteric effects and a CSV file containing their allosteric site location information and performs preprocessing to obtain the spatial structural data of the proteins; a feature extraction module, which extracts secondary structural features, contact network features, and residue microenvironmental physicochemical features from the spatial structural data of the proteins and fuses them to generate a feature matrix; a training module, which uses an improved dilated convolutional neural network to learn the local and global feature distribution of each allosteric protein from the fused feature matrix, captures the characteristic information of the allosteric site, learns the constraints of the allosteric site and the complex spatial dependencies, and ultimately obtains a trained improved model; and a recognition module, which uses the improved model to identify allosteric sites on the protein based on the protein's PDB file and provide identification interpretation. The present invention can effectively identify allosteric sites with high reliability.

Description

Technical Field

[0001] The present invention relates to the technical field of deep learning and biomedicine, and in particular to a protein allosteric site recognition system based on deep learning. Background Art

[0002] In modern drug development, identifying allosteric sites on proteins is crucial for developing highly effective, low-toxic novel drugs, particularly in the fields of targeted therapy and precision medicine. However, allosteric sites are not confined to specific proteins; rather, they are potentially widespread across proteins, meaning that almost all proteins may harbor them. However, only a few proteins have been identified with allosteric sites, and this scarcity of data severely limits the development of allosteric drugs. Therefore, identifying allosteric sites on proteins has become a pressing research topic.

[0003] At present, there are still some challenges in the identification of allosteric sites in proteins. Existing methods mainly rely on traditional experimental methods and machine learning models. However, the use of traditional experimental methods to find allosteric sites requires a lot of manpower, material resources and time, and the results are minimal. At the same time, traditional machine learning models are difficult to effectively capture the complex local and global dependencies in protein structure, and the identification results may have a high rate of false positives or false negatives. In addition, allosteric sites are often associated with long-range structural regions. Most methods only focus on the local characteristics of protein allosteric sites, ignoring the long-range dependencies and periodic characteristics that allosteric sites may involve, which further limits the accurate identification of allosteric sites. Summary of the Invention

[0004] The purpose of the present invention is to overcome the shortcomings and deficiencies of the existing technology and propose a protein allosteric site identification system based on deep learning. It adopts an improved void convolutional neural network to break through the limitations of traditional methods in capturing complex dependencies in protein structure, enhance the capture of periodic and long-range dependency features, and introduce the XGBoost model to reduce the risk of overfitting, while improving the fitting ability of complex data, further improving the accuracy and reliability of identifying allosteric sites.

[0005] To achieve the above objectives, the present invention provides a technical solution: a protein allosteric site recognition system based on deep learning, comprising:

[0006] The data import module is used to load the PDB file of proteins with allosteric effects and the CSV file of their allosteric site location information, and perform preprocessing to obtain the spatial structure data of the protein;

[0007] The feature extraction module uses the DSSP tool, residue contact network, and Feature software to extract secondary structure features, contact network features, and residue microenvironment physicochemical characteristics from the spatial structure data of proteins, and fuses the secondary structure features, contact network features, and residue microenvironment physicochemical characteristics to generate a feature matrix;

[0008] The training module, based on the improved dilated convolutional neural network (DCNN), learns the local and global feature distribution of each allosteric protein from the fused feature matrix, captures the feature information of the allosteric site, learns the constraints of the allosteric site and the complex spatial dependencies, and finally obtains a trained improved model. This improved model improves the feature capture module and prediction module of the dilated convolutional neural network (DCNN). The improvement to the feature capture module is as follows: three layers of parallel dilated convolution are introduced to extract features of large, medium and small receptive fields respectively, and at the same time, the frequency domain channel attention mechanism is combined to enhance the capture of periodic and long-distance dependency features through frequency domain feature conversion and global pooling. The improvement to the prediction module is as follows: the XGBoost model is introduced, and multiple decision trees are connected in series to improve the prediction accuracy and generalization ability of the model by weighting and regularizing the extracted key features.

[0009] The recognition module uses the trained improved model to identify the allosteric sites on the protein according to the protein's PDB file and provides identification explanations.

[0010] Furthermore, the data import module includes a data loading module and a data preprocessing module, wherein:

[0011] The data loading module uses the Bio.PDB and pandas toolkits in the Python environment to load the PDB file of the protein with allosteric effect and the CSV file of the allosteric site position information of the corresponding protein from the local, and parses the PDB file to obtain the three-dimensional structure information of the protein;

[0012] The data preprocessing module uses Bio.PDB and the pandas toolkit to retain the spatial structural information of the peptide chain where the allosteric site is located in the three-dimensional structural information based on the allosteric site position information provided by the CSV file, and generates spatial structural data of the protein only related to the allosteric site for subsequent feature extraction and model training.

[0013] Furthermore, the feature extraction module performs the following operations:

[0014] 1) Extraction of secondary structure features based on DSSP tools:

[0015] The DSSP tool analyzes the spatial structure data of the protein and identifies the secondary structure type of each residue and its corresponding sequence position and chain identity based on the secondary structure prediction method. First, the spatial structure data of the protein is input into the DSSP tool to generate a DSSP file containing the secondary structure characteristics of the protein. Then, by analyzing the spatial structure data of the protein and the secondary structure characteristics in the DSSP file, the secondary structure characteristics of the protein pocket region are extracted, including the secondary structure type and relative solvent accessibility RSA of the residue. Subsequently, the Bio.PDB.DSSP module in Biopython is used to parse the DSSP file, convert the secondary structure type of each residue into a one-hot encoding, and calculate and normalize its relative solvent accessibility RSA for subsequent feature matrix construction. Finally, the generated DSSP file will be retained for further extraction of secondary structure features in the residue microenvironment through the Feature software.

[0016] 2) Extract the atomic contact network characteristics based on the residue contact network, that is, the close-range interaction characteristics:

[0017] First, the contact network of the protein was generated using Biopython and NetworkX tools. The spatial coordinates of each residue were parsed from the PDB file of the protein, and the α carbon atom of each residue was selected as the representative atom. Based on the distance relationship in three-dimensional space, the distances were calculated according to the set thresholds. in It is a unit that describes the distance between atoms. It searches for contact residues and constructs the contact network of α-carbon atoms:

[0018]

[0019] Where, d k,c (R) is the residue contact density when the residue R at position k is set as the threshold, len(P) is the sequence length of protein P, and C k,c (R) is the total number of contacts when the residue R at position k is set as the threshold;

[0020] Next, according to the spatial coordinates of the α carbon atom of residue R, the residue R contacts are divided into the upper hemisphere and the lower hemisphere. The number of upper and lower hemisphere contacts of each residue at different thresholds is counted, and the exposure ratio is calculated:

[0021]

[0022] D k,c (R)=1-U k,c (R)

[0023] Where U k,c (R) is the upper hemisphere exposure ratio of residue R at position k with c as the threshold, D k,c(R) is the exposure ratio of the lower hemisphere when the residue R at position k is set as the threshold, C u,k,c (R) is the number of contacts in the upper hemisphere when the residue R at position k is set as the threshold, C k,c (R) is the total number of contacts in the upper and lower hemispheres when the residue R at position k is set as the threshold;

[0024] In the contact network, the local network properties of each residue are further analyzed, and the clustering coefficient and betweenness centrality of each residue are calculated:

[0025]

[0026] Where C c (R) is the clustering coefficient of residue R, T(R) is the number of triangles in the network passing through residue R, deg(R) is the degree of residue R in the network, is the theoretical maximum number of triangles for R residues;

[0027]

[0028] Where C B (R) is the betweenness centrality of residue R, V is the set of nodes, s and t are any two nodes in the network, σ(s,t) is the total number of shortest paths from node s to node t, and σ(s,t|R) is the number of shortest paths through node s, node t and residue R; in this network, it is defined as follows: if σ(s,t) = 0, then

[0029] The residue contact network is used to extract the characteristics of the close-range interactions between residues in the protein contact network. The residue contact density reveals the local contact of the residues, the hemispheric exposure ratio reflects the exposure characteristics of the residues in space, and the clustering coefficient and betweenness centrality characterize the local and global importance of the residues in the contact network from the perspective of network properties.

[0030] 3) Extraction of the physicochemical characteristics of the residue microenvironment based on Feature software:

[0031] The microenvironment sampling is performed by the featurize module of the Feature software. The protein structure is analyzed at the atomic level. The microenvironment sampling is performed on the area centered on the α carbon atom of each residue. In the microenvironment sampling, the spatial range of the microenvironment is defined as a radius of and spherical shell;

[0032] Using the Atomselect module of the Feature software, we selected the α-carbon atoms of the target residues and characterized the microenvironment of each residue one by one. The characterization included the physical, chemical, and structural properties within the sphere and shell:

[0033] Extract the number of various elements in the spherical shell or sphere according to the element level, including: the number of any element ELEMENT_IS_ANY, the number of carbon elements ELEMENT_IS_C, the number of nitrogen elements ELEMENT_IS_N, the number of oxygen elements ELEMENT_IS_O, the number of sulfur elements ELEMENT_IS_S, and the number of other elements except carbon, nitrogen, oxygen, and sulfur ELEMENT_IS_OTHER;

[0034] Extract the number of various atoms in a shell or sphere at the atomic level, including:

[0035] The number of carbon atoms connected to oxygen atoms ATOM_TYPE_IS_C, the number of terminal carbon atoms on the side chain ATOM_TYPE_IS_CT, the number of carbon atoms connected to the amino carbon atom ATOM_TYPE_IS_CA, the number of nitrogen atoms on the amino group ATOM_TYPE_IS_N, the number of atoms with the structural atom identifier N2 in the PDB file ATOM_TYPE_IS_N2, the number of atoms with the structural atom identifier N3 in the PDB file ATOM_TYPE_IS_N3, the number of nitrogen atoms connected to the α carbon atom ATOM_TYPE_IS_NA, the number of double-bonded oxygen atoms connected to the α carbon atom ATOM_TYPE_IS_O , the number of double bond oxygen atoms of the carboxyl group of the α carbon atom closest to the α carbon atom on the side chain ATOM_TYPE_IS_O2, the number of hydrogen atoms of the hydroxyl group connected to the α carbon atom ATOM_TYPE_IS_OH, the number of sulfur atoms ATOM_TYPE_IS_S, the number of hydrogen atoms on sulfur ATOM_TYPE_IS_SH, the number of all atoms except the above atoms ATOM_TYPE_IS_OTHER;

[0036] Extract various charges in the spherical shell or sphere according to the residue level, including: atomic partial charge PARTIAL_CHARGE, negative charge NEG_CHARGE, positive charge POS_CHARGE, total charge considering histidine CHARGE_WITH_HIS, and total charge excluding histidine CHARGE;

[0037] Extract the number of various residues in a spherical shell or sphere at the atomic level, including:

[0038] The number of alanine RESIDUE_NAME_IS_ALA, the number of arginine RESIDUE_NAME_IS_ARG, the number of asparagine RESIDUE_NAME_IS_ASN, the number of aspartic acid RESIDUE_NAME_IS_ASP, the number of cysteine ​​RESIDUE_NAME_IS_CYS, the number of glutamine RESIDUE_NAME_IS_GLN, the number of glutamic acid RESIDUE_NAME_IS_GLU, the number of glycine RESIDUE_NAME_IS_GLY, the number of histidine RESIDUE_NAME_IS_HIS, the number of isoleucine RESIDUE_NAME_IS_ILE, the number of leucine RESIDUE_NAME_IS_LEU, the number of lysine RESIDUE_NAME_IS_LYS, the number of methionine RESIDUE_NAME_IS_MET, the number of phenylalanine RESIDUE_NAME_IS_PHE, the number of proline RESIDUE_NAME_IS_PRO, the number of serine RESIDUE_NAME_IS_SER 、The number of threonine RESIDUE_NAME_IS_THR 、The number of tryptophan RESIDUE_NAME_IS_TRP 、The number of tyrosine RESIDUE_NAME_IS_TYR 、The number of valine RESIDUE_NAME_IS_VAL 、The residue name does not belong to all the above residue names RESIDUE_NAME_IS_OTHER;

[0039] Based on the various residues and atomic properties in the microenvironment, the overall physicochemical properties within the sphere or shell are calculated, including:

[0040] The number of hydrophobic residues RESIDUE_CLASS1_IS_HYDROPHOBIC, the number of charged residues RESIDUE_CLASS1_IS_CHARGED, the number of polar residues RESIDUE_CLASS1_IS_POLAR, and the number of residues that are neither hydrophobic nor polar and are uncharged

[0041] RESIDUE_CLASS1_IS_UNKNOWN, the number of non-polar residues RESIDUE_CLASS2_IS_NONPOLAR, the number of hydrophobic residues that differ from CLASS1 RESIDUE_CLASS2_IS_POLAR, the number of basic residues RESIDUE_CLASS2_IS_BASIC, the number of acidic residues RESIDUE_CLASS2_IS_ACIDIC, the number of residues that do not belong to the non-polar class and are not acidic or basic RESIDUE_CLASS2_IS_UNKNOWN;

[0042] Combined with the secondary structure features provided by the DSSP tool, the number of various secondary structures contained in the microenvironment is counted, including:

[0043] The number of secondary structures with 3-turn α-helices SECONDARY_STRUCTURE1_IS_3HELIX, the number of secondary structures with 4-turn α-helices SECONDARY_STRUCTURE1_IS_4HELIX, the number of secondary structures with 5-turn α-helices SECONDARY_STRUCTURE1_IS_5HELIX, the number of secondary structures with short segments connecting α-helices and β-helices SECONDARY_STRUCTURE1_IS_BRIDGE, the number of secondary structures with β-strands and anti-β-sheets SECONDARY_STRUCTURE1_IS_STRAN, the number of secondary structures with turns SECONDARY_STRUCTURE1_IS_TURN, the number of secondary structures with regular bends SECONDARY_STRUCTURE1_IS_BEND, the number of secondary structures that are not strand, helix, bend, bridge, turn, or heterocycle SECONDARY_STRUCTURE1_IS_COIL, the number of heteroatoms SECONDARY_STRUCTURE1_IS_HET , the number of secondary structures that cannot be recognized by DSSP SECONDARY_STRUCTURE1_IS_UNKNOWN, the number of secondary structures of type α helix SECONDARY_STRUCTURE2_IS_HELIX, the number of secondary structures of type β helix including BRIDGE, BEND, and TURN SECONDARY_STRUCTURE2_IS_BETA, the number of secondary structures of type random bend SECONDARY_STRUCTURE2_IS_COIL, the number of secondary structures of type heterocyclic structure SECONDARY_STRUCTURE2_IS_HET, the number of secondary structures that do not belong to HELIX, BETA, COIL, or HET SECONDARY_STRUCTURE2_IS_UNKNOWN;

[0044] Finally, the special structures and functional groups in the spherical shell or sphere are counted, including hydroxyl HYDROXYL, amide AMIDE, amine AMINE, carbonyl CARBONYL, ring system RING_SYSTEM, peptide PEPTIDE; as well as the physical properties at the atomic level: van der Waals volume VDW_VOLUME, hydrophobicity HYDROPHOBICITY, mobility MOBILITY, solvent accessibility SOLVENT_ACCESSIBILITY;

[0045] Feature software analyzes the structures of functionally similar proteins to correlate protein function with their structure. The core concept of Feature software is to analyze protein structure at the atomic level, sampling the small spherical volume surrounding each atom or a specified set of points. These volumes are called microenvironments. A microenvironment is represented by a real vector of eigenvalues, which contains information about the physical and chemical characteristics within the sphere or shell. By extracting microenvironment information, complex three-dimensional protein structural data can be converted into numerical features that can be used in machine learning models.

[0046] 4) Using the pandas toolkit in the Python environment, the secondary structure features, contact network features, and residue microenvironment physicochemical characteristics obtained in steps 1), 2, and 3) were combined to obtain a protein feature matrix; through the feature extraction module, 258 feature descriptions were generated for each residue in the protein.

[0047] Furthermore, the training module performs the following operations:

[0048] 1) Input the feature matrix into the embedding layer:

[0049] The shape of the input feature matrix X for each protein is (L, 258, 1), where L is the number of residues in the protein, 258 represents the number of features extracted for each residue, and 1 represents single-channel input. A two-layer fully connected neural network is used to embed the 258 features of each residue to reduce dimensionality and introduce nonlinearity. The input dimension of the first fully connected neural network layer is 258, and the output dimension is 128. The input dimension of the second fully connected neural network layer is 128, and the output dimension is 64. Both fully connected neural networks use the ReLU function as the activation function. The shape of the feature matrix after embedding is (L, 64).

[0050] 2) Extract local and global features of the embedded feature matrix through the dilated convolutional layer:

[0051] A dilated convolutional neural network (DCNN) consisting of three layers of parallel dilated convolution is applied to the embedded feature matrix to extract local and global features. The shape of the input feature matrix is ​​(L, 64, 1). The mathematical expression of the dilated convolution layer is as follows:

[0052]

[0053] Where O(i,j′) is the value of the output feature matrix at position (i,j′) after the dilated convolution, where i and j′ represent the i-th row and j′th column of the output feature matrix, W(m,n) is the weight of the convolution kernel at position (m,n), where m and n represent the m-th row and n-th column of the convolution kernel, I(i+m·r,j′+n·r) is the value of the input feature matrix at position (i+m·r,j′+n·r), r is the dilation rate, and M is the weight of the convolution kernel at position (m,n). c and N c Represents the size of the convolution kernel; three parallel convolutional layers with different dilation rates are used, with dilation rates r of 4, 2, and 1 respectively. The dilated convolution layer with r of 4 is responsible for extracting features with a large receptive field, and the dilated convolution layer with r of 2 is responsible for extracting features with a medium receptive field. When r = 1, the dilated convolution degenerates into a normal convolution. When r > 1, gaps are inserted between the elements of the convolution kernel. The convolution kernel size of each convolution layer is (3, 1), the stride is 1, and the padding method is the same as the padding method that keeps the output feature map size consistent with the input size. The number of output channels of each convolution layer is 64, and the activation function is the ReLU function.

[0054] The outputs of the three layers of parallel dilated convolution are concatenated in the channel dimension to form a joint feature matrix X concat , whose shape is (L, 192, 1), and the dimensionality of the concatenated feature matrix is ​​reduced by a 1×1 convolution layer to obtain the global feature matrix X gap , its shape is (L,64);

[0055] 3) Use the frequency domain channel attention mechanism in the feature capture module to capture the global feature matrix X gap Implicit periodicity and long-range dependence in:

[0056] For the global feature matrix X gap Apply fast Fourier transform to convert it from spatial domain to frequency domain and obtain frequency domain eigenvector F:

[0057]

[0058] Where F(u,v) is the value of the frequency domain feature matrix at position (u,v), where u and v represent the uth row and vth column of the frequency domain feature matrix, and X gap (x,y) is the value of the global feature matrix at position (x,y), where x and y represent the xth row and yth column of the global feature matrix, H and W are the height and width of the feature matrix, and j is the imaginary unit;

[0059] Extract the amplitude spectrum A from the frequency domain feature F:

[0060] A(u,v)=|F(u,v)|

[0061] Where A(u,v) is the value of the frequency domain amplitude spectrum at position (u,v);

[0062] Perform global pooling on the frequency domain amplitude spectrum A to obtain the channel-level frequency domain features S:

[0063]

[0064] 4) Input the frequency domain feature S into the two-layer fully connected neural network to generate the attention weight Attention(S), and apply the generated channel attention weight Attention(S) to the original feature:

[0065] Attention(S)=σ(W2·ReLU(W1·S))

[0066] F'=Attention(S)·F

[0067] Where Attention(S) is the attention weight of the frequency domain feature S, W1 and W2 are the weight matrices of the two-layer fully connected neural network, ReLU is the activation function, and σ is the Sigmoid function; the input dimension of the first layer of the fully connected neural network is 64, the output dimension is 32, and the activation function is the ReLU function, which is used to introduce nonlinearity. The input dimension of the second layer of the fully connected neural network is 32, the output dimension is 64, and the activation function is the Sigmoid function, which is used to limit the output to the range of [0,1]. F' is the weighted channel feature, and its shape is (L,64);

[0068] 5) Use the frequency domain variational encoder to perform feature compression and latent variable modeling on the weighted channel feature F', and generate the latent variable matrix Z:

[0069] The encoder of the frequency domain variation encoder adopts a two-layer fully connected neural network. The input dimension of the first layer of the fully connected neural network is 64, the output dimension is 32, and the activation function is the ReLU function. The input dimension of the second layer of the fully connected neural network is 32, the output dimension is 16, and the activation function is the Linear function. The encoder of the frequency domain variation encoder performs a nonlinear transformation on the channel feature F' to extract its potential low-dimensional representation. After being processed by each layer of the encoder of the frequency domain variation encoder, the mean and variance are generated, and then the latent variable matrix Z is obtained by reparameterization, whose shape is (L, 16). The decoder of the frequency domain variation encoder also adopts a two-layer fully connected neural network. The input dimension of the first layer of the fully connected neural network is 16, the output dimension is 32, and the activation function is the ReLU function. The input dimension of the second layer of the fully connected neural network is 32, the output dimension is 64, and the activation function is the Linear function. The loss function of the encoder and decoder of the frequency domain variation encoder consists of reconstruction loss and KL divergence loss:

[0070] Ltotal =L recon +0.1 L KL

[0071] Where, L total Represents the total loss function, L recon represents the difference between F' and the reconstructed feature matrix generated by the decoder, calculated using the mean square error, L KL represents KL divergence loss;

[0072] 6) Residue level identification:

[0073] The latent variable matrix Z generated by the frequency domain variational encoder is input into the XGBoost model as the feature of each residue for residue-level recognition. The recognition target value of the XGBoost model is a binary label vector Y with a length equal to the number of amino acids in the protein currently being trained. θ , where Y θ The value range is {0,1}, when Y θ = 0, indicating that the θth residue is a non-allosteric site. θ =1, it means that the θth residue is an allosteric site; the parameter settings of the XGBoost model are: maximum tree depth of 6, learning rate of 0.1, number of trees of 100, subsample ratio of 0.8, regularization parameters of γ = 0.1, λ = 1; finally, the prediction module determines whether the residue is an allosteric site based on the probability value output by the XGBoost model. If the probability of a residue being predicted as an allosteric site is greater than 0.5, the prediction module will identify it as an allosteric site.

[0074] Furthermore, the identification module performs the following operations:

[0075] 1) Read the input protein PDB file and parse it using the Bio.PDB toolkit in the Python environment to extract the protein's three-dimensional structure information; store the parsed protein's three-dimensional structure information in the variable protein_structure;

[0076] 2) Using the data import module, retain the peptide chain spatial structure information related to the allosteric site in protein_structure obtained in step 1) to generate spatial structure data of the protein related to the allosteric site; store the processed protein spatial structure data in the variable processed_protein_data as input for feature extraction;

[0077] 3) Use the feature extraction module to extract features from the spatial structure data of the protein stored in the variable processed_protein_data in step 2), call the DSSP tool, residue contact network, and Feature software to extract secondary structure features, contact network features, and residue microenvironment physicochemical features, and store them in the variables dssp_feature, residue_connect_feature, and mirco_env_feature, respectively. Use the pandas toolkit to splice dssp_feature, residue_connect_feature, and mirco_env_feature to generate the protein feature matrix feature_matrix;

[0078] 4) The generated feature matrix feature_matrix is ​​input into the trained improved model for prediction. The model combines local and global feature distributions, captures the spatial dependence and constraints of the residues, and obtains the probability P of each residue being an allosteric site. The value range of P is [0,1], which indicates the possibility of the residue being an allosteric site. The allosteric site probability of each residue is stored in the prediction probability list prediction_probability_list;

[0079] 5) For each residue in the prediction probability list prediction_probability_list, the probability threshold T = 0.5 is set to classify each residue: if the probability P of a residue is greater than or equal to T, it is classified as an allosteric site with a label of 1; if the probability P of a residue is greater than or equal to T, it is classified as a non-allosteric site with a label of 0; the classified labels are stored in the prediction label list prediction_list; finally, whether the residue is an allosteric site is interpreted based on the probability value of the residue in the prediction probability list prediction_probability_list and the label in the prediction label list prediction_list.

[0080] Compared with the prior art, the present invention has the following advantages and beneficial effects:

[0081] 1. Through systematic feature extraction using DSSP tools, residue contact networks, and Feature software, secondary structure features, contact network features, and residue microenvironment physicochemical characteristics are obtained from protein structures, significantly improving the ability to characterize allosteric site features.

[0082] 2. The multi-dimensional feature matrix of proteins is analyzed through an improved dilated convolutional network, and multi-dilation rate convolutional layers are introduced to extract features of large, medium, and small receptive fields respectively, thereby enhancing the ability to capture structural information at different scales.

[0083] 3. By introducing the frequency-domain channel attention mechanism, we enhance the capture of periodic and long-distance dependency features. This improvement not only improves the comprehensiveness of feature capture, enabling it to learn more complex spatial dependencies, but also enhances the model's ability to recognize long-range structural associations.

[0084] 4. The multi-layer structure design of the frequency-domain variational encoder is conducive to compressing weighted features and modeling latent variables, which can better extract potential low-dimensional representations. It also improves the effectiveness of feature representation and reduces the impact of data noise by optimizing reconstruction loss and KL divergence loss.

[0085] 5. By introducing the XGBoost model for final prediction and using the method of connecting multiple decision trees in series, the extracted key features are weighted and regularized, providing accurate and reliable allosteric site identification results, and improving the prediction accuracy and generalization ability of the model.

[0086] 6. The accurate identification of protein allosteric sites by the present invention can not only provide key target information for new drug research and development, but also provide an important theoretical basis for understanding the allosteric mechanism of proteins and designing allosteric regulators, thereby promoting the development of targeted drugs. BRIEF DESCRIPTION OF THE DRAWINGS

[0087] Figure 1 Schematic diagram of the relationship between the various modules of the system of the present invention.

[0088] Figure 2 Flowchart of training and recognition of the system of the present invention.

[0089] Figure 3 Schematic diagram of the structure of the improved dilated convolutional neural network used in the system of the present invention. DETAILED DESCRIPTION

[0090] The present invention will be further described below with reference to specific embodiments.

[0091] This embodiment discloses a protein allosteric site recognition system based on deep learning. It is a protein allosteric site recognition system based on deep learning developed in Python and can be run on Windows devices. The relationship between the modules of the system is as follows: Figure 1 As shown, the system training and recognition process is as follows Figure 2 shown.

[0092] It includes:

[0093] The data import module is used to load the PDB file of proteins with allosteric effects and the CSV file of their allosteric site location information, and perform preprocessing to obtain the spatial structure data of the protein;

[0094] The feature extraction module uses the DSSP tool, residue contact network, and Feature software to extract secondary structure features, contact network features, and residue microenvironment physicochemical characteristics from the spatial structure data of proteins, and fuses the secondary structure features, contact network features, and residue microenvironment physicochemical characteristics to generate a feature matrix;

[0095] The training module, based on the improved dilated convolutional neural network (DCNN), learns the local and global feature distribution of each allosteric protein from the fused feature matrix, captures the feature information of the allosteric site, learns the constraints of the allosteric site and the complex spatial dependencies, and finally obtains a trained improved model (i.e., the optimal model); Figure 3 As shown in the figure, the improved model improves the feature capture module and prediction module of the dilated convolutional neural network (DCNN). The improvement to the feature capture module is as follows: three layers of parallel dilated convolution are introduced to extract features of large, medium and small receptive fields respectively, while combining the frequency domain channel attention mechanism to enhance the capture of periodic and long-distance dependency features through frequency domain feature conversion and global pooling; the improvement to the prediction module is as follows: the XGBoost model is introduced, and multiple decision trees are connected in series to weight and regularize the extracted key features to improve the prediction accuracy and generalization ability of the model.

[0096] The recognition module uses the trained improved model to identify the allosteric sites on the protein according to the protein's PDB file and provides identification explanations.

[0097] Specifically, the data import module includes a data loading module and a data preprocessing module, wherein:

[0098] The data loading module uses the Bio.PDB and pandas toolkits in the Python environment to load the PDB file of the protein with allosteric effect and the CSV file of the allosteric site position information of the corresponding protein from the local, and parses the PDB file to obtain the three-dimensional structure information of the protein;

[0099] The data preprocessing module uses Bio.PDB and the pandas toolkit to retain the spatial structural information of the peptide chain where the allosteric site is located in the three-dimensional structural information based on the allosteric site position information provided by the CSV file, and generates spatial structural data of the protein only related to the allosteric site for subsequent feature extraction and model training.

[0100] Specifically, the feature extraction module performs the following operations:

[0101] 1) Extraction of secondary structure features based on DSSP tools:

[0102] The DSSP tool analyzes the spatial structure data of the protein and identifies the secondary structure type of each residue and its corresponding sequence position and chain identity based on the secondary structure prediction method. First, the spatial structure data of the protein is input into the DSSP tool to generate a DSSP file containing the secondary structure characteristics of the protein. Then, by analyzing the spatial structure data of the protein and the secondary structure characteristics in the DSSP file, the secondary structure characteristics of the protein pocket region are extracted, including the secondary structure type and relative solvent accessibility RSA of the residue. Subsequently, the Bio.PDB.DSSP module in Biopython is used to parse the DSSP file, convert the secondary structure type of each residue into a one-hot encoding, and calculate and normalize its relative solvent accessibility RSA for subsequent feature matrix construction. Finally, the generated DSSP file will be retained for further extraction of secondary structure features in the residue microenvironment through the Feature software.

[0103] 2) Extract the atomic contact network characteristics based on the residue contact network, that is, the close-range interaction characteristics:

[0104] First, the contact network of the protein was generated using Biopython and NetworkX tools. The spatial coordinates of each residue were parsed from the PDB file of the protein, and the α carbon atom of each residue was selected as the representative atom. Based on the distance relationship in three-dimensional space, the distances were calculated according to the set thresholds. in It is a unit that describes the distance between atoms. It searches for contact residues and constructs the contact network of α-carbon atoms:

[0105]

[0106] Where, d k,c (R) is the residue contact density when the residue R at position k is set as the threshold, len(P) is the sequence length of protein P, and C k,c (R) is the total number of contacts when the residue R at position k is set as the threshold;

[0107] Next, according to the spatial coordinates of the α carbon atom of residue R, the residue R contacts are divided into the upper hemisphere and the lower hemisphere. The number of upper and lower hemisphere contacts of each residue at different thresholds is counted, and the exposure ratio is calculated:

[0108]

[0109] D k,c (R)=1-U k,c (R)

[0110] Where U k,c (R) is the upper hemisphere exposure ratio of residue R at position k with c as the threshold, D k,c(R) is the exposure ratio of the lower hemisphere when the residue R at position k is set as the threshold, C u,k,c (R) is the number of contacts in the upper hemisphere when the residue R at position k is set as the threshold, C k,c (R) is the total number of contacts in the upper and lower hemispheres when the residue R at position k is set as the threshold;

[0111] In the contact network, the local network properties of each residue are further analyzed, and the clustering coefficient and betweenness centrality of each residue are calculated:

[0112]

[0113] Where C c (R) is the clustering coefficient of residue R, T(R) is the number of triangles in the network passing through residue R, deg(R) is the degree of residue R in the network, is the theoretical maximum number of triangles for R residues;

[0114]

[0115] Where C B (R) is the betweenness centrality of residue R, V is the set of nodes, s and t are any two nodes in the network, σ(s,t) is the total number of shortest paths from node s to node t, and σ(s,t|R) is the number of shortest paths through node s, node t and residue R; in this network, it is defined as follows: if σ(s,t) = 0, then

[0116] The residue contact network is used to extract the characteristics of the close-range interactions between residues in the protein contact network. The residue contact density reveals the local contact of the residues, the hemispheric exposure ratio reflects the exposure characteristics of the residues in space, and the clustering coefficient and betweenness centrality characterize the local and global importance of the residues in the contact network from the perspective of network properties.

[0117] 3) Extraction of the physicochemical characteristics of the residue microenvironment based on Feature software:

[0118] The microenvironment sampling is performed by the featurize module of the Feature software. The protein structure is analyzed at the atomic level. The microenvironment sampling is performed on the area centered on the α carbon atom of each residue. In the microenvironment sampling, the spatial range of the microenvironment is defined as a radius of and spherical shell;

[0119] Using the Atomselect module of the Feature software, we selected the α-carbon atoms of the target residues and characterized the microenvironment of each residue one by one. The characterization included the physical, chemical, and structural properties within the sphere and shell:

[0120] Extract the number of various elements in the spherical shell or sphere according to the element level, including: the number of any element ELEMENT_IS_ANY, the number of carbon elements ELEMENT_IS_C, the number of nitrogen elements ELEMENT_IS_N, the number of oxygen elements ELEMENT_IS_O, the number of sulfur elements ELEMENT_IS_S, and the number of other elements except carbon, nitrogen, oxygen, and sulfur ELEMENT_IS_OTHER;

[0121] Extract the number of various atoms in a shell or sphere at the atomic level, including:

[0122] The number of carbon atoms connected to oxygen atoms ATOM_TYPE_IS_C, the number of terminal carbon atoms on the side chain ATOM_TYPE_IS_CT, the number of carbon atoms connected to the amino carbon atom ATOM_TYPE_IS_CA, the number of nitrogen atoms on the amino group ATOM_TYPE_IS_N, the number of atoms with the structural atom identifier N2 in the PDB file ATOM_TYPE_IS_N2, the number of atoms with the structural atom identifier N3 in the PDB file ATOM_TYPE_IS_N3, the number of nitrogen atoms connected to the α carbon atom ATOM_TYPE_IS_NA, the number of double-bonded oxygen atoms connected to the α carbon atom ATOM_TYPE_IS_O , the number of double bond oxygen atoms of the carboxyl group of the α carbon atom closest to the α carbon atom on the side chain ATOM_TYPE_IS_O2, the number of hydrogen atoms of the hydroxyl group connected to the α carbon atom ATOM_TYPE_IS_OH, the number of sulfur atoms ATOM_TYPE_IS_S, the number of hydrogen atoms on sulfur ATOM_TYPE_IS_SH, the number of all atoms except the above atoms ATOM_TYPE_IS_OTHER;

[0123] Extract various charges in the spherical shell or sphere according to the residue level, including: atomic partial charge PARTIAL_CHARGE, negative charge NEG_CHARGE, positive charge POS_CHARGE, total charge considering histidine CHARGE_WITH_HIS, and total charge excluding histidine CHARGE;

[0124] Extract the number of various residues in a spherical shell or sphere at the atomic level, including:

[0125] The number of alanine RESIDUE_NAME_IS_ALA, the number of arginine RESIDUE_NAME_IS_ARG, the number of asparagine RESIDUE_NAME_IS_ASN, the number of aspartic acid RESIDUE_NAME_IS_ASP, the number of cysteine ​​RESIDUE_NAME_IS_CYS, the number of glutamine RESIDUE_NAME_IS_GLN, the number of glutamic acid RESIDUE_NAME_IS_GLU, the number of glycine RESIDUE_NAME_IS_GLY, the number of histidine RESIDUE_NAME_IS_HIS, the number of isoleucine RESIDUE_NAME_IS_ILE, the number of leucine RESIDUE_NAME_IS_LEU, the number of lysine RESIDUE_NAME_IS_LYS, the number of methionine RESIDUE_NAME_IS_MET, the number of phenylalanine RESIDUE_NAME_IS_PHE, the number of proline RESIDUE_NAME_IS_PRO, the number of serine RESIDUE_NAME_IS_SER 、The number of threonine RESIDUE_NAME_IS_THR 、The number of tryptophan RESIDUE_NAME_IS_TRP 、The number of tyrosine RESIDUE_NAME_IS_TYR 、The number of valine RESIDUE_NAME_IS_VAL 、The residue name does not belong to all the above residue names RESIDUE_NAME_IS_OTHER;

[0126] Based on the various residues and atomic properties in the microenvironment, the overall physicochemical properties within the sphere or shell are calculated, including:

[0127] The number of hydrophobic residues RESIDUE_CLASS1_IS_HYDROPHOBIC, the number of charged residues RESIDUE_CLASS1_IS_CHARGED, the number of polar residues RESIDUE_CLASS1_IS_POLAR, and the number of residues that are neither hydrophobic nor polar and are uncharged

[0128] RESIDUE_CLASS1_IS_UNKNOWN, the number of non-polar residues RESIDUE_CLASS2_IS_NONPOLAR, the number of hydrophobic residues that differ from CLASS1 RESIDUE_CLASS2_IS_POLAR, the number of basic residues RESIDUE_CLASS2_IS_BASIC, the number of acidic residues RESIDUE_CLASS2_IS_ACIDIC, the number of residues that do not belong to the non-polar class and are not acidic or basic RESIDUE_CLASS2_IS_UNKNOWN;

[0129] Combined with the secondary structure features provided by the DSSP tool, the number of various secondary structures contained in the microenvironment is counted, including:

[0130] The number of secondary structures with 3-turn α-helices SECONDARY_STRUCTURE1_IS_3HELIX, the number of secondary structures with 4-turn α-helices SECONDARY_STRUCTURE1_IS_4HELIX, the number of secondary structures with 5-turn α-helices SECONDARY_STRUCTURE1_IS_5HELIX, the number of secondary structures with short segments connecting α-helices and β-helices SECONDARY_STRUCTURE1_IS_BRIDGE, the number of secondary structures with β-strands and anti-β-sheets SECONDARY_STRUCTURE1_IS_STRAN, the number of secondary structures with turns SECONDARY_STRUCTURE1_IS_TURN, the number of secondary structures with regular bends SECONDARY_STRUCTURE1_IS_BEND, the number of secondary structures that are not strand, helix, bend, bridge, turn, or heterocycle SECONDARY_STRUCTURE1_IS_COIL , the number of heteroatoms SECONDARY_STRUCTURE1_IS_HET, the number of secondary structures that cannot be recognized by DSSP SECONDARY_STRUCTURE1_IS_UNKNOWN, the number of secondary structures of type α helix SECONDARY_STRUCTURE2_IS_HELIX, the number of secondary structures of type β helix including BRIDGE, BEND, and TURN SECONDARY_STRUCTURE2_IS_BETA, the number of secondary structures of type random bend SECONDARY_STRUCTURE2_IS_COIL, the number of secondary structures of type heterocyclic structure SECONDARY_STRUCTURE2_IS_HET, the number of secondary structures that are not HELIX, BETA, COIL, or HET SECONDARY_STRUCTURE2_IS_UNKNOWN;

[0131] Finally, the special structures and functional groups in the spherical shell or sphere are counted, including hydroxyl HYDROXYL, amide AMIDE, amine AMINE, carbonyl CARBONYL, ring system RING_SYSTEM, peptide PEPTIDE; as well as the physical properties at the atomic level: van der Waals volume VDW_VOLUME, hydrophobicity HYDROPHOBICITY, mobility MOBILITY, solvent accessibility SOLVENT_ACCESSIBILITY;

[0132] Feature software analyzes the structures of functionally similar proteins to correlate protein function with their structure. The core concept of Feature software is to analyze protein structure at the atomic level, sampling the small spherical volume surrounding each atom or a specified set of points. These volumes are called microenvironments. A microenvironment is represented by a real vector of eigenvalues, which contains information about the physical and chemical characteristics within the sphere or shell. By extracting microenvironment information, complex three-dimensional protein structural data can be converted into numerical features that can be used in machine learning models.

[0133] 4) Using the pandas toolkit in the Python environment, the secondary structure features, contact network features, and residue microenvironment physicochemical characteristics obtained in steps 1), 2, and 3) were combined to obtain a protein feature matrix; through the feature extraction module, 258 feature descriptions were generated for each residue in the protein.

[0134] Specifically, if Figure 3 As shown, the training module performs the following operations:

[0135] 1) Input the feature matrix into the embedding layer:

[0136] The shape of the input feature matrix X for each protein is (L, 258, 1), where L is the number of residues in the protein, 258 represents the number of features extracted for each residue, and 1 represents single-channel input. A two-layer fully connected neural network is used to embed the 258 features of each residue to reduce dimensionality and introduce nonlinearity. The input dimension of the first fully connected neural network layer is 258, and the output dimension is 128. The input dimension of the second fully connected neural network layer is 128, and the output dimension is 64. Both fully connected neural networks use the ReLU function as the activation function. The shape of the feature matrix after embedding is (L, 64).

[0137] 2) Extract local and global features of the embedded feature matrix through the dilated convolutional layer:

[0138] A dilated convolutional neural network (DCNN) consisting of three layers of parallel dilated convolution is applied to the embedded feature matrix to extract local and global features. The shape of the input feature matrix is ​​(L, 64, 1). The mathematical expression of the dilated convolution layer is as follows:

[0139]

[0140] Where O(i,j′) is the value of the output feature matrix at position (i,j′) after the dilated convolution, where i and j′ represent the i-th row and j′th column of the output feature matrix, W(m,n) is the weight of the convolution kernel at position (m,n), where m and n represent the m-th row and n-th column of the convolution kernel, I(i+m·r,j′+n·r) is the value of the input feature matrix at position (i+m·r,j′+n·r), r is the dilation rate, and M is the weight of the convolution kernel at position (m,n). c and N c Represents the size of the convolution kernel; three parallel convolutional layers with different dilation rates are used, with dilation rates r of 4, 2, and 1 respectively. The dilated convolution layer with r of 4 is responsible for extracting features with a large receptive field, and the dilated convolution layer with r of 2 is responsible for extracting features with a medium receptive field. When r = 1, the dilated convolution degenerates into a normal convolution. When r > 1, gaps are inserted between the elements of the convolution kernel. The convolution kernel size of each convolution layer is (3, 1), the stride is 1, and the padding method is to keep the output feature map size consistent with the input size (same). The number of output channels of each convolution layer is 64, and the activation function is the ReLU function.

[0141] The outputs of the three layers of parallel dilated convolution are concatenated in the channel dimension to form a joint feature matrix X concat , whose shape is (L, 192, 1), and the dimensionality of the concatenated feature matrix is ​​reduced by a 1×1 convolution layer to obtain the global feature matrix X gap , its shape is (L,64);

[0142] 3) Use the frequency domain channel attention mechanism in the feature capture module to capture the global feature matrix X gap Implicit periodicity and long-range dependence in:

[0143] For the global feature matrix X gap Apply fast Fourier transform to convert it from spatial domain to frequency domain and obtain frequency domain eigenvector F:

[0144]

[0145] Where F(u,v) is the value of the frequency domain feature matrix at position (u,v), where u and v represent the uth row and vth column of the frequency domain feature matrix, and X gap (x,y) is the value of the global feature matrix at position (x,y), where x and y represent the xth row and yth column of the global feature matrix, H and W are the height and width of the feature matrix, and j is the imaginary unit;

[0146] Extract the amplitude spectrum A from the frequency domain feature F:

[0147] A(u,v)=|F(u,v)|

[0148] Where A(u,v) is the value of the frequency domain amplitude spectrum at position (u,v);

[0149] Perform global pooling on the frequency domain amplitude spectrum A to obtain the channel-level frequency domain features S:

[0150]

[0151] 4) Input the frequency domain feature S into the two-layer fully connected neural network to generate the attention weight Attention(S), and apply the generated channel attention weight Attention(S) to the original feature:

[0152] Attention(S)=σ(W2·ReLU(W1·S))

[0153] F'=Attention(S)·F

[0154] Where Attention(S) is the attention weight of the frequency domain feature S, W1 and W2 are the weight matrices of the two-layer fully connected neural network, ReLU is the activation function, and σ is the Sigmoid function; the input dimension of the first layer of the fully connected neural network is 64, the output dimension is 32, and the activation function is the ReLU function, which is used to introduce nonlinearity. The input dimension of the second layer of the fully connected neural network is 32, the output dimension is 64, and the activation function is the Sigmoid function, which is used to limit the output to the range of [0,1]. F' is the weighted channel feature, and its shape is (L,64);

[0155] 5) Use the frequency domain variational encoder to perform feature compression and latent variable modeling on the weighted channel feature F', and generate the latent variable matrix Z:

[0156] The encoder of the frequency domain variational encoder uses a two-layer fully connected neural network. The input dimension of the first layer of the fully connected neural network is 64, the output dimension is 32, and the activation function is the ReLU function. The input dimension of the second layer of the fully connected neural network is 32, the output dimension is 16, and the activation function is the Linear function. The encoder of the frequency domain variational encoder performs a nonlinear transformation on the channel feature F' to extract its potential low-dimensional representation. After being processed by each layer of the encoder of the frequency domain variational encoder, the mean and variance are generated, and then the latent variable matrix Z is obtained by reparameterization, whose shape is (L, 16). The decoder also uses a two-layer fully connected neural network. The input dimension of the first layer of the fully connected neural network is 16, the output dimension is 32, and the activation function is the ReLU function. The input dimension of the second layer of the fully connected neural network is 32, the output dimension is 64, and the activation function is the Linear function. The loss function of the encoder and decoder of the frequency domain variational encoder consists of reconstruction loss and KL divergence loss:

[0157] Ltotal =L recon +0.1 L KL

[0158] Where, L total Represents the total loss function, L recon represents the difference between F' and the reconstructed feature matrix generated by the decoder, calculated using the mean square error, L KL represents KL divergence loss;

[0159] 6) Residue level identification:

[0160] The latent variable matrix Z generated by the frequency domain variational encoder is input into the XGBoost model as the feature of each residue for residue-level recognition. The recognition target value of the XGBoost model is a binary label vector Y with a length equal to the number of amino acids in the protein currently being trained. θ , where Y θ The value range is {0,1}, when Y θ = 0, indicating that the θth residue is a non-allosteric site. θ =1, it means that the θth residue is an allosteric site; the parameter settings of the XGBoost model are: maximum tree depth of 6, learning rate of 0.1, number of trees of 100, subsample ratio of 0.8, regularization parameters of γ = 0.1, λ = 1; finally, the prediction module determines whether the residue is an allosteric site based on the probability value output by the XGBoost model. If the probability of a residue being predicted as an allosteric site is greater than 0.5, the prediction module will identify it as an allosteric site.

[0161] Specifically, the identification module performs the following operations:

[0162] 1) Read the input protein PDB file and parse it using the Bio.PDB toolkit in the Python environment to extract the protein's three-dimensional structure information; store the parsed protein's three-dimensional structure information in the variable protein_structure;

[0163] 2) Using the data import module, retain the peptide chain spatial structure information related to the allosteric site in protein_structure obtained in step 1) to generate spatial structure data of the protein related to the allosteric site; store the processed protein spatial structure data in the variable processed_protein_data as input for feature extraction;

[0164] 3) Use the feature extraction module to extract features from the spatial structure data of the protein stored in the variable processed_protein_data in step 2), call the DSSP tool, residue contact network, and Feature software to extract secondary structure features, contact network features, and residue microenvironment physicochemical features, and store them in the variables dssp_feature, residue_connect_feature, and mirco_env_feature, respectively. Use the pandas toolkit to splice dssp_feature, residue_connect_feature, and mirco_env_feature to generate the protein feature matrix feature_matrix;

[0165] 4) The generated feature matrix feature_matrix is ​​input into the trained improved model for prediction. The model combines local and global feature distributions, captures the spatial dependence and constraints of the residues, and obtains the probability P of each residue being an allosteric site. The value range of P is [0,1], which indicates the possibility of the residue being an allosteric site. The allosteric site probability of each residue is stored in the prediction probability list prediction_probability_list;

[0166] 5) For each residue in the prediction probability list prediction_probability_list, the probability threshold T = 0.5 is set to classify each residue: if the probability P of a residue is greater than or equal to T, it is classified as an allosteric site with a label of 1; if the probability P of a residue is less than or equal to T, it is classified as a non-allosteric site with a label of 0; the classified labels are stored in the prediction label list prediction_list; finally, whether the residue is an allosteric site is interpreted based on the probability value of the residue in the prediction probability list prediction_probability_list and the label in the prediction label list prediction_list.

[0167] The above embodiments are preferred implementation modes of the present invention, but the implementation modes of the present invention are not limited to the above embodiments. Any other changes, modifications, substitutions, combinations, and simplifications that do not deviate from the spirit and principles of the present invention should be considered as equivalent replacement methods and are included in the scope of protection of the present invention.

Claims

1. A protein allosteric site recognition system based on deep learning, characterized by: include: The data import module is used to load the PDB file of proteins with allosteric effects and the CSV file of their allosteric site location information, and perform preprocessing to obtain the spatial structure data of the protein; The feature extraction module uses the DSSP tool, residue contact network, and Feature software to extract secondary structure features, contact network features, and residue microenvironment physicochemical characteristics from the spatial structure data of proteins, and fuses the secondary structure features, contact network features, and residue microenvironment physicochemical characteristics to generate a feature matrix; The training module, based on an improved dilated convolutional neural network (DCNN), learns the local and global feature distribution of each allosteric protein from the fused feature matrix, captures the characteristic information of the allosteric sites, learns the constraints of the allosteric sites and the complex spatial dependencies, and ultimately obtains a trained improved model. This improved model improves the feature capture module and prediction module of the dilated convolutional neural network (DCNN). The improvements to the feature capture module include the introduction of three layers of parallel dilated convolution to extract features of large, medium, and small receptive fields, respectively. Furthermore, the frequency-domain channel attention mechanism is combined to enhance the capture of periodic and long-range dependency features through frequency-domain feature conversion and global pooling. The improvements to the prediction module are: introducing the XGBoost model, using the method of connecting multiple decision trees in series, and weighting and regularizing the extracted key features to improve the prediction accuracy and generalization ability of the model; The recognition module uses the trained improved model to identify the allosteric sites on the protein according to the protein's PDB file and provides identification explanations; The feature extraction module performs the following operations: 1) Extraction of secondary structure features based on DSSP tools: The DSSP tool analyzes the spatial structure data of the protein and identifies the secondary structure type of each residue and its corresponding sequence position and chain identity based on the secondary structure prediction method. First, the spatial structure data of the protein is input into the DSSP tool to generate a DSSP file containing the secondary structure characteristics of the protein. Then, by analyzing the spatial structure data of the protein and the secondary structure characteristics in the DSSP file, the secondary structure characteristics of the protein pocket region are extracted, including the secondary structure type and relative solvent accessibility RSA of the residue. Subsequently, the Bio.PDB.DSSP module in Biopython is used to parse the DSSP file, convert the secondary structure type of each residue into a one-hot encoding, and calculate and normalize its relative solvent accessibility RSA for subsequent feature matrix construction. Finally, the generated DSSP file will be retained for further extraction of secondary structure features in the residue microenvironment through the Feature software. 2) Extract the atomic contact network characteristics based on the residue contact network, that is, the close-range interaction characteristics: First, the contact network of the protein was generated using Biopython and NetworkX tools. The spatial coordinates of each residue were parsed from the PDB file of the protein, and the α carbon atom of each residue was selected as the representative atom. Based on the distance relationship in three-dimensional space, the distances were calculated according to the set thresholds. in It is a unit that describes the distance between atoms. It searches for contact residues and constructs the contact network of α-carbon atoms: Where, d k,c (R) is the residue contact density when the residue R at position k is set as the threshold, len(P) is the sequence length of protein P, and C k,c (R) is the total number of contacts when the residue R at position k is set as the threshold; Next, according to the spatial coordinates of the α carbon atom of residue R, the residue R contacts are divided into the upper hemisphere and the lower hemisphere. The number of upper and lower hemisphere contacts of each residue at different thresholds is counted, and the exposure ratio is calculated: D k,c (R)=1-U k,c (R) Where U k,c (R) is the upper hemisphere exposure ratio of residue R at position k with c as the threshold, D k,c (R) is the exposure ratio of the lower hemisphere when the residue R at position k is set as the threshold, C u,k,c (R) is the number of contacts in the upper hemisphere when the residue R at position k is set as the threshold, C k,c (R) is the total number of contacts in the upper and lower hemispheres when the residue R at position k is set as the threshold; In the contact network, the local network properties of each residue are further analyzed, and the clustering coefficient and betweenness centrality of each residue are calculated: Where C c (R) is the clustering coefficient of residue R, T(R) is the number of triangles in the network passing through residue R, deg(R) is the degree of residue R in the network, is the theoretical maximum number of triangles for R residues; Where C B (R) is the betweenness centrality of residue R, V is the set of nodes, s and t are any two nodes in the network, σ(s,t) is the total number of shortest paths from node s to node t, and σ(s,t|R) is the number of shortest paths through node s, node t and residue R; in this network, it is defined as follows: if σ(s,t) = 0, then The residue contact network is used to extract the characteristics of the close-range interactions between residues in the protein contact network. The residue contact density reveals the local contact of the residues, the hemispheric exposure ratio reflects the exposure characteristics of the residues in space, and the clustering coefficient and betweenness centrality characterize the local and global importance of the residues in the contact network from the perspective of network properties. 3) Extraction of the physicochemical characteristics of the residue microenvironment based on Feature software: The microenvironment sampling is performed by the featurize module of the Feature software. The protein structure is analyzed at the atomic level. The microenvironment sampling is performed on the area centered on the α carbon atom of each residue. In the microenvironment sampling, the spatial range of the microenvironment is defined as a radius of and spherical shell; Using the Atomselect module of the Feature software, we selected the α-carbon atoms of the target residues and characterized the microenvironment of each residue one by one. The characterization included the physical, chemical, and structural properties within the sphere and shell: Extract the number of various elements in the spherical shell or sphere according to the element level, including: the number of any element ELEMENT_IS_ANY, the number of carbon elements ELEMENT_IS_C, the number of nitrogen elements ELEMENT_IS_N, the number of oxygen elements ELEMENT_IS_O, the number of sulfur elements ELEMENT_IS_S, and the number of other elements except carbon, nitrogen, oxygen, and sulfur ELEMENT_IS_OTHER; Extract the number of various atoms in a shell or sphere at the atomic level, including: The number of carbon atoms connected to oxygen atoms ATOM_TYPE_IS_C, the number of terminal carbon atoms on the side chain ATOM_TYPE_IS_CT, the number of carbon atoms connected to the amino carbon atom ATOM_TYPE_IS_CA, the number of nitrogen atoms on the amino group ATOM_TYPE_IS_N, the number of atoms with the structural atom identifier N2 in the PDB file ATOM_TYPE_IS_N2, the number of atoms with the structural atom identifier N3 in the PDB file ATOM_TYPE_IS_N3, the number of nitrogen atoms connected to the alpha carbon atom ATOM_TYPE_IS_NA, the number of double-bonded oxygen atoms connected to the α carbon atom ATOM_TYPE_IS_O, the number of double-bonded oxygen atoms of the carboxyl group of the α carbon atom closest to the α carbon atom on the side chain ATOM_TYPE_IS_O2, the number of hydroxyl hydrogen atoms connected to the α carbon atom ATOM_TYPE_IS_OH, the number of sulfur atoms ATOM_TYPE_IS_S, the number of hydrogen atoms on sulfur ATOM_TYPE_IS_SH, the number of all atoms except the above atoms ATOM_TYPE_IS_OTHER; Extract various charges in the spherical shell or sphere according to the residue level, including: atomic partial charge PARTIAL_CHARGE, negative charge NEG_CHARGE, positive charge POS_CHARGE, total charge considering histidine CHARGE_WITH_HIS, and total charge excluding histidine CHARGE; Extract the number of various residues in a spherical shell or sphere at the atomic level, including: The number of alanine RESIDUE_NAME_IS_ALA, the number of arginine RESIDUE_NAME_IS_ARG, the number of asparagine RESIDUE_NAME_IS_ASN, the number of aspartic acid RESIDUE_NAME_IS_ASP, the number of cysteine ​​RESIDUE_NAME_IS_CYS, the number of glutamine RESIDUE_NAME_IS_GLN, the number of glutamic acid RESIDUE_NAME_IS_GLU, the number of glycine RESIDUE_NAME_IS_GLY, the number of histidine RESIDUE_NAME_IS_HIS, the number of isoleucine RESIDUE_NAME_IS_ILE, the number of leucine RESIDUE_NA ME_IS_LEU, the number of lysines RESIDUE_NAME_IS_LYS, the number of methionines RESIDUE_NAME_IS_MET, the number of phenylalanines RESIDUE_NAME_IS_PHE, the number of prolines RESIDUE_NAME_IS_PRO, the number of serines RESIDUE_NAME_IS_SER, the number of threonines RESIDUE_NAME_IS_THR, the number of tryptophans RESIDUE_NAME_IS_TRP, the number of tyrosines RESIDUE_NAME_IS_TYR, the number of valines RESIDUE_NAME_IS_VAL, the residue name does not belong to any of the above residue names RESIDUE_NAME_IS_OTHER; Based on the various residues and atomic properties in the microenvironment, the overall physicochemical properties within the sphere or shell are calculated, including: The number of hydrophobic residues RESIDUE_CLASS1_IS_HYDROPHOBIC, the number of charged residues RESIDUE_CLASS1_IS_CHARGED, the number of polar residues RESIDUE_CLASS1_IS_POLAR, and the number of residues that are neither hydrophobic nor polar and are uncharged RESIDUE_CLASS1_IS_UNKNOWN, the number of non-polar residues RESIDUE_CLASS2_IS_NONPOLAR, the number of hydrophobic residues that differ from CLASS1 RESIDUE_CLASS2_IS_POLAR, the number of basic residues RESIDUE_CLASS2_IS_BASIC, the number of acidic residues RESIDUE_CLASS2_IS_ACIDIC, the number of residues that do not belong to the non-polar class and are not acidic or basic RESIDUE_CLASS2_IS_UNKNOWN; Combined with the secondary structure features provided by the DSSP tool, the number of various secondary structures contained in the microenvironment is counted, including: The number of secondary structures with 3 turns of α-helix SECONDARY_STRUCTURE1_IS_3HELIX, the number of secondary structures with 4 turns of α-helix SECONDARY_STRUCTURE1_IS_4HELIX, the number of secondary structures with 5 turns of α-helix SECONDARY_STRUCTURE1_IS_5HELIX, the number of secondary structures with short segments connecting α-helix and β-helix SECONDARY_STRUCTURE1_IS_BRIDGE, the number of secondary structures with β-strand and anti-β-sheet SECONDARY_STRUCTURE1_IS_STRAN, the number of secondary structures with turns SECONDARY_STRUCTURE1_IS_TURN, the number of secondary structures with regular bends SECONDARY_STRUCTURE1_IS_BEND, the secondary structures are not strand, helix, bend, bridge, turn, or heterocycle The number of SECONDARY_STRUCTURE1_IS_COIL, the number of heteroatoms SECONDARY_STRUCTURE1_IS_HET, the number of secondary structures that cannot be recognized by DSSP SECONDARY_STRUCTURE1_IS_UNKNOWN, the number of secondary structures of type α helix SECONDARY_STRUCTURE2_IS_HELIX, the number of secondary structures of type β helix including BRIDGE, BEND, and TURN SECONDARY_STRUCTURE2_IS_BETA, the number of secondary structures of type random bend SECONDARY_STRUCTURE2_IS_COIL, the number of secondary structures of type heterocyclic structures SECONDARY_STRUCTURE2_IS_HET, the number of secondary structures that are not HELIX, BETA, COIL, or HET SECONDARY_STRUCTURE2_IS_UNKNOWN; Finally, the special structures and functional groups in the spherical shell or sphere are counted, including hydroxyl HYDROXYL, amide AMIDE, amine AMINE, carbonyl CARBONYL, ring system RING_SYSTEM, peptide PEPTIDE; as well as the physical properties at the atomic level: van der Waals volume VDW_VOLUME, hydrophobicity HYDROPHOBICITY, mobility MOBILITY, solvent accessibility SOLVENT_ACCESSIBILITY; Feature software analyzes the structures of functionally similar proteins to correlate protein function with their structure. The core concept of Feature software is to analyze protein structure at the atomic level, sampling the small spherical volume surrounding each atom or a specified set of points. These volumes are called microenvironments. A microenvironment is represented by a real vector of eigenvalues, which contains information about the physical and chemical characteristics within the sphere or shell. By extracting microenvironment information, complex three-dimensional protein structural data can be converted into numerical features that can be used in machine learning models. 4) Using the pandas toolkit in the Python environment, the secondary structure features, contact network features, and residue microenvironment physicochemical characteristics obtained in steps 1), 2, and 3) were combined to obtain a protein feature matrix; through the feature extraction module, 258 feature descriptions were generated for each residue in the protein.

2. A protein allosteric site recognition system based on deep learning according to claim 1, characterized in that: The data import module includes a data loading module and a data preprocessing module, wherein: The data loading module uses the Bio.PDB and pandas toolkits in the Python environment to load the PDB file of the protein with allosteric effect and the CSV file of the allosteric site position information of the corresponding protein from the local, and parses the PDB file to obtain the three-dimensional structure information of the protein; The data preprocessing module uses Bio.PDB and the pandas toolkit to retain the spatial structural information of the peptide chain where the allosteric site is located in the three-dimensional structural information based on the allosteric site position information provided by the CSV file, and generates spatial structural data of the protein only related to the allosteric site for subsequent feature extraction and model training.

3. A protein allosteric site recognition system based on deep learning according to claim 2, characterized in that: The training module performs the following operations: 1) Input the feature matrix into the embedding layer: The shape of the input feature matrix X for each protein is (L, 258, 1), where L is the number of residues in the protein, 258 represents the number of features extracted for each residue, and 1 represents single-channel input. A two-layer fully connected neural network is used to embed the 258 features of each residue to reduce dimensionality and introduce nonlinearity. The input dimension of the first fully connected neural network layer is 258, and the output dimension is 128. The input dimension of the second fully connected neural network layer is 128, and the output dimension is 64. Both fully connected neural networks use the ReLU function as the activation function. The shape of the feature matrix after embedding is (L, 64). 2) Extract local and global features of the embedded feature matrix through the dilated convolutional layer: A dilated convolutional neural network (DCNN) consisting of three layers of parallel dilated convolution is applied to the embedded feature matrix to extract local and global features. The shape of the input feature matrix is ​​(L, 64, 1). The mathematical expression of the dilated convolution layer is as follows: Where O(i,j′) is the value of the output feature matrix at position (i,j′) after the dilated convolution, where i and j′ represent the i-th row and j′th column of the output feature matrix, W(m,n) is the weight of the convolution kernel at position (m,n), where m and n represent the m-th row and n-th column of the convolution kernel, I(i+m·r,j′+n·r) is the value of the input feature matrix at position (i+m·r,j′+n·r), r is the dilation rate, and M is the weight of the convolution kernel at position (m,n). c and N c Represents the size of the convolution kernel; three parallel convolutional layers with different dilation rates are used, with dilation rates r of 4, 2, and 1 respectively. The dilated convolution layer with r of 4 is responsible for extracting features with a large receptive field, and the dilated convolution layer with r of 2 is responsible for extracting features with a medium receptive field. When r = 1, the dilated convolution degenerates into a normal convolution. When r > 1, gaps are inserted between the elements of the convolution kernel. The convolution kernel size of each convolution layer is (3, 1), the stride is 1, and the padding method is the same as the padding method that keeps the output feature map size consistent with the input size. The number of output channels of each convolution layer is 64, and the activation function is the ReLU function. The outputs of the three layers of parallel dilated convolution are concatenated in the channel dimension to form a joint feature matrix X concat , whose shape is (L, 192, 1), and the dimensionality of the concatenated feature matrix is ​​reduced by a 1×1 convolution layer to obtain the global feature matrix X gap , its shape is (L,64); 3) Use the frequency domain channel attention mechanism in the feature capture module to capture the global feature matrix X gap Implicit periodicity and long-range dependence in: For the global feature matrix X gap Apply fast Fourier transform to convert it from spatial domain to frequency domain and obtain frequency domain eigenvector F: Where F(u,v) is the value of the frequency domain feature matrix at position (u,v), where u and v represent the uth row and vth column of the frequency domain feature matrix, and X gap (x,y) is the value of the global feature matrix at position (x,y), where x and y represent the xth row and yth column of the global feature matrix, H and W are the height and width of the feature matrix, and j is the imaginary unit; Extract the amplitude spectrum A from the frequency domain feature F: A(u,v)=|F(u,v)| Where A(u,v) is the value of the frequency domain amplitude spectrum at position (u,v); Perform global pooling on the frequency domain amplitude spectrum A to obtain the channel-level frequency domain features S: 4) Input the frequency domain feature S into the two-layer fully connected neural network to generate the attention weight Attention(S), and apply the generated channel attention weight Attention(S) to the original feature: Attention(S)=σ(W2·ReLU(W1·S)) F'=Attention(S)·F Where Attention(S) is the attention weight of the frequency domain feature S, W1 and W2 are the weight matrices of the two-layer fully connected neural network, ReLU is the activation function, and σ is the Sigmoid function; the input dimension of the first layer of the fully connected neural network is 64, the output dimension is 32, and the activation function is the ReLU function, which is used to introduce nonlinearity. The input dimension of the second layer of the fully connected neural network is 32, the output dimension is 64, and the activation function is the Sigmoid function, which is used to limit the output to the range of [0,1]. F' is the weighted channel feature, and its shape is (L,64); 5) Use the frequency domain variational encoder to perform feature compression and latent variable modeling on the weighted channel feature F', and generate the latent variable matrix Z: The encoder of the frequency domain variation encoder adopts a two-layer fully connected neural network. The input dimension of the first layer of the fully connected neural network is 64, the output dimension is 32, and the activation function is the ReLU function. The input dimension of the second layer of the fully connected neural network is 32, the output dimension is 16, and the activation function is the Linear function. The encoder of the frequency domain variation encoder performs a nonlinear transformation on the channel feature F' to extract its potential low-dimensional representation. After being processed by each layer of the encoder of the frequency domain variation encoder, the mean and variance are generated, and then the latent variable matrix Z is obtained by reparameterization, whose shape is (L, 16). The decoder of the frequency domain variation encoder also adopts a two-layer fully connected neural network. The input dimension of the first layer of the fully connected neural network is 16, the output dimension is 32, and the activation function is the ReLU function. The input dimension of the second layer of the fully connected neural network is 32, the output dimension is 64, and the activation function is the Linear function. The loss function of the encoder and decoder of the frequency domain variation encoder consists of reconstruction loss and KL divergence loss: L total =L recon +0.1 L KL Where, L total Represents the total loss function, L recon represents the difference between F' and the reconstructed feature matrix generated by the decoder, calculated using the mean square error, L KL represents KL divergence loss; 6) Residue level identification: The latent variable matrix Z generated by the frequency domain variational encoder is input into the XGBoost model as the feature of each residue for residue-level recognition. The recognition target value of the XGBoost model is a binary label vector Y with a length equal to the number of amino acids in the protein currently being trained. θ , where Y θ The value range is {0,1}, when Y θ = 0, indicating that the θth residue is a non-allosteric site. θ =1, it means that the θth residue is an allosteric site; the parameter settings of the XGBoost model are: maximum tree depth of 6, learning rate of 0.1, number of trees of 100, subsample ratio of 0.8, regularization parameters of γ = 0.1, λ = 1; finally, the prediction module determines whether the residue is an allosteric site based on the probability value output by the XGBoost model. If the probability of a residue being predicted as an allosteric site is greater than 0.5, the prediction module will identify it as an allosteric site.

4. A protein allosteric site recognition system based on deep learning according to claim 3, characterized in that: The identification module performs the following operations: 1) Read the input protein PDB file and parse it using the Bio.PDB toolkit in the Python environment to extract the protein's three-dimensional structure information; store the parsed protein's three-dimensional structure information in the variable protein_structure; 2) Using the data import module, retain the peptide chain spatial structure information related to the allosteric site in protein_structure obtained in step 1) to generate spatial structure data of the protein related to the allosteric site; store the processed protein spatial structure data in the variable processed_protein_data as input for feature extraction; 3) Use the feature extraction module to extract features from the spatial structure data of the protein stored in the variable processed_protein_data in step 2), call the DSSP tool, residue contact network, and Feature software to extract secondary structure features, contact network features, and residue microenvironment physicochemical features, and store them in the variables dssp_feature, residue_connect_feature, and mirco_env_feature, respectively. Use the pandas toolkit to splice dssp_feature, residue_connect_feature, and mirco_env_feature to generate the protein feature matrix feature_matrix; 4) The generated feature matrix feature_matrix is ​​input into the trained improved model for prediction. The model combines local and global feature distributions, captures the spatial dependence and constraints of the residues, and obtains the probability P of each residue being an allosteric site. The value range of P is [0,1], which indicates the possibility of the residue being an allosteric site. The allosteric site probability of each residue is stored in the prediction probability list prediction_probability_list; 5) For each residue in the prediction probability list prediction_probability_list, the probability threshold T = 0.5 is set to classify each residue: if the probability P of a residue is greater than or equal to T, it is classified as an allosteric site with a label of 1; if the probability P of a residue is less than or equal to T, it is classified as a non-allosteric site with a label of 0; the classified labels are stored in the prediction label list prediction_list; finally, whether the residue is an allosteric site is interpreted based on the probability value of the residue in the prediction probability list prediction_probability_list and the label in the prediction label list prediction_list.

Images