Protein structure characterization method, device, equipment and medium
By generating protein graph representations containing local and global features through a target Siamese neural network and solver, and combining the graph matching algorithm to optimize the representation, the problem that existing methods cannot capture global structural information is solved, and more efficient protein structure representation is achieved.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- SHENZHEN READLINE BIOTECH CO LTD
- Filing Date
- 2024-10-23
- Publication Date
- 2026-04-24
AI Technical Summary
Existing structure-based protein embedding methods mainly rely on graph neural networks, which cannot fully encompass global protein structural information, thus limiting their application in areas such as functional prediction.
A target twin neural network and a target solver are used to generate a graph representation containing local feature nodes and global representation nodes. By combining a graph matching algorithm and a replacement matrix, the local representation is optimized to reflect the global features of the protein structure.
It significantly enhances the ability to capture global information, improves the accuracy and efficiency of protein structure alignment, and generates characterizations that more comprehensively reflect the complexity and diversity of protein structures.
Smart Images
Figure CN121922189A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of bioinformatics, and in particular to methods, apparatus, equipment and media for protein structure characterization. Background Technology
[0002] In the field of bioinformatics, predicting protein function is a crucial task for understanding biological processes and drug design. To achieve this goal, proteins are typically represented as embedding vectors, i.e., their corresponding characterizations. These characterizations capture key protein features, supporting subsequent analysis and prediction. Currently, the main methods for protein embedding are sequence-based, such as ESM (Evolutionary Scale Modeling), which generate embedding vectors by analyzing the amino acid sequence of proteins. However, while sequence information provides some indication of protein function, protein structure contains much richer information. Protein structure determines its function and active sites; therefore, structure-based embedding methods hold promise for providing more accurate and comprehensive protein characterization.
[0003] However, existing structure-based embedding methods mainly rely on techniques such as graph neural networks to propagate information within a local domain. Although this method can capture some features of protein structure, it still cannot fully contain the global structural information of the protein. The lack of global information limits the application effect of these methods in protein function prediction and other fields.
[0004] In summary, how to generate a characterization that more comprehensively reflects the structural features of proteins is a problem that needs to be solved in this field. Summary of the Invention
[0005] In view of this, the purpose of this invention is to provide a method, apparatus, device, and medium for protein structure characterization, which generates a characterization that can more comprehensively reflect the structural features of proteins. The specific solution is as follows:
[0006] In a first aspect, this application discloses a method for characterizing protein structure, including:
[0007] The protein structure file to be processed is input into the target representation generation model; wherein, the target representation generation model includes a target Siamese neural network and a target solver;
[0008] The target Siamese neural network generates a target graph representation containing local feature nodes and global representation nodes based on the protein structure file to be processed, and obtains the target local representation of the target graph representation based on the global representation nodes.
[0009] The target solver uses a target replacement matrix corresponding to the target local representation and a graph matching algorithm to match each protein structure in the protein structure file to be processed to obtain the target matching result. Based on the target matching result, the target local representation is optimized to obtain the target representation of the protein structure file to be processed.
[0010] Optionally, a target graph representation containing local feature nodes and global representation nodes is generated based on the protein structure file to be processed, including:
[0011] Each amino acid in the protein structure file to be processed is identified as a local feature node, and local association edges between the local feature nodes are established based on the distance between them, so as to generate an initial graph representation containing the local feature nodes and the local association edges.
[0012] Average pooling is performed on each of the local feature nodes to obtain a global representation node, and global association edges are established between the global representation node and each of the local feature nodes. The global representation node and each of the global association edges are added to the initial graph representation to obtain the target graph representation.
[0013] Optionally, obtaining the target local representation of the target graph representation based on the global representation of the global representation node includes:
[0014] Generate each initial local representation of the target graph, and perform average pooling on each initial local representation to obtain the global representation of the global representation node;
[0015] The initial local representations are optimized based on the global representations of the global representation nodes to obtain the target local representations of the target graph.
[0016] Optionally, before inputting the protein structure file to be processed into the target characterization generation model, the method further includes:
[0017] Collect historical protein structure files and construct a historical graph representation containing historical global representation nodes based on each historical protein structure file;
[0018] Based on the aforementioned historical graph representations, a representation training dataset corresponding to the metric learning loss function is established;
[0019] An initial representation generation model, including an initial Siamese neural network and an initial solver, is constructed, and the initial representation generation model is trained using the representation training dataset and the metric learning loss function to obtain a target representation generation model including a target Siamese neural network and a target solver.
[0020] Optionally, the metric learning loss function is a triplet loss function;
[0021] Accordingly, the step of establishing a representation training dataset corresponding to the metric learning loss function based on each of the historical graph representations includes:
[0022] Obtain the template modeling score among the historical graph representations, and construct a representation training dataset containing multiple triplet data based on the template modeling score.
[0023] Optionally, construct the current triplet data, including:
[0024] The target protein structure and positive and negative samples of the target protein structure are determined from each of the historical graph representations, and a current triplet data containing the target protein structure, the positive sample, and the negative sample is constructed; wherein the template modeling score between the positive sample and the target protein structure is greater than a preset threshold, and the template modeling score between the negative sample and the target protein structure is less than the preset threshold.
[0025] Optionally, the protein structure characterization method further includes:
[0026] The functional labels of each of the historical protein structure files are determined, and the historical representations of the historical protein structure files are obtained using the target representation generation model, so as to establish a prediction training dataset containing the historical representations and the corresponding functional labels.
[0027] The initial function prediction model is trained using the prediction training dataset to obtain the target function prediction model, so as to predict the target function corresponding to the target representation.
[0028] Secondly, this application discloses a protein structure characterization device, comprising:
[0029] The file input module is used to input the protein structure file to be processed into the target representation generation model; wherein, the target representation generation model includes a target Siamese neural network and a target solver;
[0030] The first processing module is used to generate a target graph representation containing local feature nodes and global representation nodes based on the protein structure file to be processed through the target Siamese neural network, and to obtain the target local representation of the target graph representation based on the global representation of the global representation nodes.
[0031] The second processing module is used to match each protein structure in the protein structure file to be processed using the target solver based on the target replacement matrix corresponding to the target local representation and using a graph matching algorithm to obtain the target matching result, and to optimize the target local representation based on the target matching result to obtain the target representation of the protein structure file to be processed.
[0032] Thirdly, this application discloses an electronic device, including:
[0033] Memory, used to store computer programs;
[0034] A processor is used to execute the computer program to implement the steps of the aforementioned disclosed protein structure characterization method.
[0035] Fourthly, this application discloses a computer-readable storage medium for storing a computer program; wherein, when the computer program is executed by a processor, it implements the steps of the aforementioned disclosed protein structure characterization method.
[0036] The beneficial effects of this application are as follows: This application inputs the protein structure file to be processed into a target representation generation model; wherein, the target representation generation model includes a target Siamese neural network and a target solver; the target Siamese neural network generates a target graph representation containing local feature nodes and global representation nodes based on the protein structure file to be processed, and obtains the target local representation of the target graph representation based on the global representation nodes; the target solver uses a target replacement matrix corresponding to the target local representation and a graph matching algorithm to match each protein structure in the protein structure file to be processed to obtain a target matching result, and optimizes the target local representation based on the target matching result to obtain the target representation of the protein structure file to be processed. Therefore, after the protein structure file to be processed is input into the target representation generation model, the target Siamese neural network of the target representation generation model generates a target graph representation containing local feature nodes and global representation nodes based on the protein structure file to be processed, and obtains the target local representation of the target graph representation based on the global representation nodes. In other words, by introducing global representation nodes and embedding them into the graph representation of the protein structure, this application significantly enhances the model's ability to capture global information, thereby integrating the information of the entire protein structure. Furthermore, the target solver of the target representation generation model uses a graph matching algorithm and combines it with the target replacement matrix to accurately match the protein structure, which can determine the best correspondence between protein structures, thereby further improving the accuracy and efficiency of structure comparison. The final generated target representation not only contains rich local feature information, but also integrates global structural information. This representation method can more comprehensively reflect the complexity and diversity of protein structures. Attached Figure Description
[0037] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort.
[0038] Figure 1 This is a flowchart of a protein structure characterization method disclosed in this application;
[0039] Figure 2 This is a schematic diagram of a specific representation generation model disclosed in this application;
[0040] Figure 3 This is a specific schematic diagram representing one aspect of the disclosure in this application;
[0041] Figure 4 This is a schematic diagram of a protein structure characterization device disclosed in this application.
[0042] Figure 5 This is a structural diagram of an electronic device disclosed in this application. Detailed Implementation
[0043] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.
[0044] In the field of bioinformatics, predicting protein function is a crucial task for understanding biological processes and drug design. To achieve this goal, proteins are typically represented as embedding vectors, i.e., their corresponding characterizations. These characterizations capture key protein features, supporting subsequent analysis and prediction. Currently, the main methods for protein embedding are sequence-based, such as ESM. These methods generate embedding vectors by analyzing the amino acid sequence of proteins. However, while sequence information provides some indication of protein function, protein structure contains much richer information. Protein structure determines its function and active sites. Therefore, structure-based embedding methods hold promise for providing more accurate and comprehensive protein characterization.
[0045] However, existing structure-based embedding methods mainly rely on techniques such as graph neural networks to propagate information within a local domain. Although this method can capture some features of protein structure, it still cannot fully contain the global structural information of the protein. The lack of global information limits the application effect of these methods in protein function prediction and other fields.
[0046] Therefore, this application provides a protein structure characterization scheme, which generates a characterization that can more comprehensively reflect the structural features of the protein.
[0047] See Figure 1 As shown in the embodiments of this application, a protein structure characterization method is disclosed, including:
[0048] Step S11: Input the protein structure file to be processed into the target representation generation model; wherein, the target representation generation model includes a target Siamese neural network and a target solver.
[0049] In this embodiment, before inputting the protein structure file to be processed into the target representation generation model, the method further includes: collecting each historical protein structure file, and constructing a historical graph representation containing historical global representation nodes based on each historical protein structure file; establishing a representation training dataset corresponding to the metric learning loss function based on each historical graph representation; constructing an initial representation generation model containing an initial Siamese neural network and an initial solver, and training the initial representation generation model using the representation training dataset and the metric learning loss function to obtain a target representation generation model including a target Siamese neural network and a target solver.
[0050] Before actually performing protein structure characterization, it is necessary to train a target characterization generation model, such as... Figure 2 The diagram shows a specific representation generation model. The target representation generation model is mainly divided into a target Siamese neural network and a target solver. The process of obtaining the target representation generation model requires iterative training of the initial representation generation model, which includes the initial Siamese neural network and the initial solver, using training data to obtain the target representation generation model.
[0051] During the acquisition of training data, historical protein structure files are first collected to store and describe the spatial structure of each protein molecule. A historical graph representation containing historical global representation nodes is then constructed based on these files. The graph representation includes nodes and the edges connecting them. The position of an amino acid is defined as the position of its α-carbon atom, and amino acids are represented as historical local feature nodes in the historical graph representation. An edge connects two amino acids when the distance between them is less than a preset threshold. For example... Figure 3The diagram illustrates a specific graph representation. Nodes can be represented by a tensor called node features, specifically using one-hot encoding. If the distance between historical local feature node 1 and historical local feature nodes 2 and 3 is less than a preset threshold, then an association edge is established between historical local feature node 1 and historical local feature nodes 2 and 3. If the distance between historical local feature node 2 and historical local feature node 3 is less than a preset threshold, then an association edge is established between historical local feature node 2 and historical local feature node 3. Alternatively, if the distance between historical local feature node 1 and historical local feature node 2 is not less than a preset threshold, then no association edge is needed between historical local feature node 1 and historical local feature node 2. After obtaining all historical local feature nodes and the association edges between them, average pooling is performed on all historical local feature nodes to obtain historical global representation nodes. Then, association edges are established between historical global representation nodes and each historical local feature node, thus obtaining the historical graph representation. It can be understood that the historical graph representation obtained at this point incorporates historical global representation nodes, so it can better learn global information during subsequent model training.
[0052] Based on each historical graph representation, a representation training dataset corresponding to the metric learning loss function is established. An initial representation generation model containing an initial Siamese neural network and an initial solver is constructed. The initial representation generation model is trained using the representation training dataset and the metric learning loss function. The model parameters are updated through the backpropagation algorithm. The Adam optimizer is used with a learning rate set to 0.01 to obtain a target representation generation model including a target Siamese neural network and a target solver.
[0053] Metric learning is a machine learning method that aims to learn distance or similarity metrics between data points. Traditional metrics (such as Euclidean distance) may not be accurate enough in high-dimensional spaces or complex data structures. Therefore, metric learning uses a data-driven approach to learn metric functions that are more suitable for specific tasks. Commonly used metric learning loss functions include contrastive loss and triplet loss, which are used to optimize the distance relationships between data points. The objectives of metric learning typically include the following two:
[0054] 1) Minimize the distance between samples of the same type: make data points of the same category closer together in the metric space;
[0055] 2) Maximize the distance between different classes of samples: This makes the data points of different classes more separated in the metric space.
[0056] The Siamese Neural Network (SiN) is primarily used for embedding graph-structured data. It consists of two graph neural networks (GNNs) sharing parameters, used to process two protein structure graphs. The GNNs generate node feature embeddings, including global node embeddings, through multi-layer graph convolutional networks (GCNs). The Siamese Neural Network's input layer receives node features and edge information from the protein structure graphs. The hidden layers consist of multiple layers of graph convolutional layers, each updating the node's feature representation using information from neighboring nodes. The Siamese Neural Network's output layer generates the feature embedding for each node, including global node embeddings.
[0057] Furthermore, the solver part uses the node embeddings obtained from the Siamese neural network to calculate the substitution matrix. It then employs a linear solution method to calculate the optimal match between two protein structures based on the substitution matrix, using this as a similarity measure. The aim is to optimize the substitution matrix so that the generated structural representations more accurately reflect the similarities and differences between biochemical structures. It's important to note that when calculating the substitution matrix of the protein structure map, the embeddings of global representation nodes need to be removed, considering only the substitution cost between local feature nodes. The specific formula is shown below:
[0058] ;
[0059] In the formula, G1 and G2 represent the replacement cost between local feature nodes u and v, respectively, and represent two protein structures. The representation function is used to process deep networks;
[0060] When calculating the optimal match between two protein structures, the Hungarian algorithm from graph matching can be used, as shown in the following formula:
[0061] ;
[0062] In the formula, π represents all possible node matches, and C G1，G2 It is a replacement matrix. The edit distance represents the predicted edit distance. In other words, the Hungarian algorithm is used to calculate the best match between node features to determine the edit distance between protein structure graphs. The edit distance represents the minimum number of edit operations required to transform one graph into another. The edit distance can represent the similarity between protein structures. The larger the template modeling score (tm-score) value, the stronger the similarity between the graphs and the smaller their edit distance. The edit distance can be considered the predicted value, and the tm-score the true value. During training, the edit distance calculated based on the replacement matrix obtained by the model is required to be positively correlated with the tm-score value. Furthermore, the node replacement cost and triplet loss function value should be minimized during iterative training to improve the accuracy of protein structure embedding and alignment.
[0063] In this embodiment, the metric learning loss function is the triplet loss function. Correspondingly, the step of establishing a representation training dataset corresponding to the metric learning loss function based on each historical graph representation includes: obtaining template modeling scores between each historical graph representation, and constructing a representation training dataset containing multiple triplet data based on the template modeling scores. If the metric learning loss function is selected as the triplet loss function, then the template modeling scores between protein structure graphs are calculated, the template modeling scores between each historical graph representation are obtained, and a representation training dataset containing multiple triplet data is constructed based on the template modeling scores. The triplet loss function ensures that the embedding distances of similar protein structure graphs are relatively close, while the embedding distances of different types of protein structure graphs are relatively far apart. That is, it minimizes the distance between embeddings of similar nodes and maximizes the distance between embeddings of different types of nodes, thereby optimizing the embedding representation effect. The specific formula is as follows:
[0064] Loss=max(0,d(a,p)−d(a,n)+α);
[0065] In the formula, p and n represent positive and negative samples, respectively, d(a, p) represents the distance between the anchor point and the positive sample, d(a, n) represents the distance between the anchor point and the negative sample, and α is a hyperparameter called margin, which is used to ensure that the distance between positive samples is a certain distance smaller than the distance between negative samples.
[0066] In this embodiment, constructing the current triplet data includes: determining the target protein structure and positive and negative samples of the target protein structure from each of the historical graph representations, and constructing the current triplet data containing the target protein structure, the positive samples, and the negative samples; wherein, the template modeling score between the positive sample and the target protein structure is greater than a preset threshold, and the template modeling score between the negative sample and the target protein structure is less than the preset threshold. Each triplet data set is constructed, and each triplet data set includes the target protein structure (target), a positive sample (negative) of the target protein structure, and a negative sample (positive). First, the target protein structure is selected, and then positive samples with a template modeling score greater than a preset threshold and negative samples with a template modeling score less than a preset threshold are determined, for example, the preset threshold is 0.5.
[0067] Step S12: Generate a target graph representation containing local feature nodes and global representation nodes based on the protein structure file to be processed using the target Siamese neural network, and obtain the target local representation of the target graph representation based on the global representation of the global representation nodes.
[0068] In this embodiment, generating a target graph representation containing local feature nodes and global representation nodes based on the protein structure file to be processed includes: identifying each amino acid in the protein structure file to be processed as a local feature node, and establishing local association edges between the local feature nodes according to the distance between the local feature nodes to generate an initial graph representation containing the local feature nodes and the local association edges; performing average pooling on the local feature nodes to obtain global representation nodes, establishing global association edges between the global representation nodes and the local feature nodes, and adding the global representation nodes and the global association edges to the initial graph representation to obtain the target graph representation.
[0069] Understandably, during training, the Siamese neural network learns to generate graph representations and embeddings. Therefore, in the process of representing the protein structure file to be processed, the target Siamese neural network is also used to generate graph representations and embeddings. The specific process of generating the target graph representation is as follows:
[0070] 1) Identify each amino acid in the protein structure file to be processed as a local feature node; use the α carbon atom in each amino acid as the position of each local feature node;
[0071] 2) Establish local association edges between local feature nodes based on the distances between them to generate an initial graph representation containing each local feature node and each local association edge; calculate the distance between the residues of the current local feature node and the residues of other local feature nodes. If there are other local feature nodes whose distance to the current local feature node is less than a preset threshold, then establish local association edges between the current local feature node and the other local feature node to obtain the initial graph representation. For example, if the distance between local feature node 1 and local feature node 2 is less than the preset threshold, and the distance between local feature node 1 and local feature node 3 is not less than the preset threshold, then it is necessary to establish a local association edge between local feature node 1 and local feature node 2, but it is not necessary to establish a local association edge between local feature node 1 and local feature node 3.
[0072] 3) Perform average pooling on each local feature node to obtain a global representation node. In this way, the global representation node can perceive the overall features of the entire protein structure.
[0073] 4) Establish global association edges between global representation nodes and local feature nodes, and add global representation nodes and global association edges to the initial graph representation to obtain the target graph representation. It should be noted that when establishing global association edges, it is necessary to establish global association edges between global representation nodes and all local feature nodes.
[0074] In this embodiment, obtaining the target local representation of the target graph representation based on the global representation of the global representation node includes: generating initial local representations of the target graph representation, and performing average pooling on each initial local representation to obtain the global representation of the global representation node; optimizing each initial local representation based on the global representation of the global representation node to obtain the target local representation of the target graph representation. It can be understood that embedding the target graph representation, i.e., generating initial local representations of the target graph representation, and further performing average pooling on each initial local representation, yields the global representation of the global representation node. Optimizing each initial local representation based on the global representation of the global representation node to obtain the target local representation of the target graph representation means that after generating the global representation of the global representation node, the global representation of the global representation node influences the initial local representations, thereby obtaining the target local representation. Therefore, when subsequently calculating the replacement matrix of the target local representation, even if the embedding of the global representation node is removed, and only the replacement cost between local feature nodes is considered, the replacement cost is still influenced by the global representation node.
[0075] Step S13: The target solver uses the target replacement matrix corresponding to the target local representation and a graph matching algorithm to match each protein structure in the protein structure file to be processed to obtain the target matching result. The target local representation is then optimized based on the target matching result to obtain the target representation of the protein structure file to be processed.
[0076] During training, the solver learns substitution matrix computation, protein matching, and embedding optimization. Therefore, in representing the protein structure files, the target solver is used for substitution matrix computation, protein matching, and embedding optimization. The substitution matrix defines the cost of replacing nodes or edges, considering only the cost of replacing local nodes. These costs reflect the "price" of replacing one node or edge with another, thus affecting the optimality of the match. Graph matching is a widely used technique in graph theory and computer science to find the optimal correspondence between two graphs. This correspondence is usually determined based on some similarity metric, with the goal of minimizing the differences between the matched graph structures. Specific applications include:
[0077] 1) Node matching: Find the one-to-one correspondence between nodes in the graph so that similar nodes are matched together;
[0078] 2) Edge matching: Not only should node matching be considered, but the edges between matched nodes should also be matched as much as possible.
[0079] In this embodiment, the method further includes: determining the functional labels of each of the historical protein structure files, and using the target representation generation model to obtain the historical representations of the historical protein structure files, so as to establish a prediction training dataset containing the historical representations and the corresponding functional labels; using the prediction training dataset to train the constructed initial function prediction model to obtain a target function prediction model, so as to use the target function prediction model to predict the target function corresponding to the target representation.
[0080] A functional prediction model is established to predict function when the target representation is obtained. A dataset including historical protein structure files and corresponding functional labels is collected; each data entry in this dataset has the structure (protein structure file, functional label). The historical protein structure files in this dataset are embedded, i.e., the historical protein structure files are represented using the target representation generation model to obtain historical representations of the historical protein structure files. This allows the creation of a prediction training dataset containing historical representations and corresponding functional labels; each data entry in the prediction training dataset has the structure (protein structure embedding, functional label). The initial functional prediction model is then trained using the prediction training dataset to obtain the target functional prediction model. It is understood that a prediction validation dataset and a prediction test dataset with the structure (protein structure embedding, functional label) can also be established to validate and test the target functional prediction model, determining its functional prediction performance. If the functional prediction performance meets the preset requirements, the target functional prediction model can be used to predict the target function corresponding to the target representation.
[0081] The beneficial effects of this application are as follows: This application inputs the protein structure file to be processed into a target representation generation model; wherein, the target representation generation model includes a target Siamese neural network and a target solver; the target Siamese neural network generates a target graph representation containing local feature nodes and global representation nodes based on the protein structure file to be processed, and obtains the target local representation of the target graph representation based on the global representation nodes; the target solver uses a target replacement matrix corresponding to the target local representation and a graph matching algorithm to match each protein structure in the protein structure file to be processed to obtain a target matching result, and optimizes the target local representation based on the target matching result to obtain the target representation of the protein structure file to be processed. Therefore, after the protein structure file to be processed is input into the target representation generation model, the target Siamese neural network of the target representation generation model generates a target graph representation containing local feature nodes and global representation nodes based on the protein structure file to be processed, and obtains the target local representation of the target graph representation based on the global representation nodes. In other words, by introducing global representation nodes and embedding them into the graph representation of the protein structure, this application significantly enhances the model's ability to capture global information, thereby integrating the information of the entire protein structure. Furthermore, the target solver of the target representation generation model uses a graph matching algorithm and combines it with the target replacement matrix to accurately match the protein structure, which can determine the best correspondence between protein structures, thereby further improving the accuracy and efficiency of structure comparison. The final generated target representation not only contains rich local feature information, but also integrates global structural information. This representation method can more comprehensively reflect the complexity and diversity of protein structures.
[0082] See Figure 4 As shown in the figure, this application discloses a protein structure characterization device, comprising:
[0083] The file input module 11 is used to input the protein structure file to be processed into the target representation generation model; wherein, the target representation generation model includes a target Siamese neural network and a target solver;
[0084] The first processing module 12 is used to generate a target graph representation containing local feature nodes and global representation nodes based on the protein structure file to be processed through the target Siamese neural network, and to obtain the target local representation of the target graph representation based on the global representation of the global representation nodes.
[0085] The second processing module 13 is used to match each protein structure in the protein structure file to be processed using the target solver based on the target replacement matrix corresponding to the target local representation and using a graph matching algorithm to obtain the target matching result, and to optimize the target local representation based on the target matching result to obtain the target representation of the protein structure file to be processed.
[0086] The beneficial effects of this application are as follows: This application inputs the protein structure file to be processed into a target representation generation model; wherein, the target representation generation model includes a target Siamese neural network and a target solver; the target Siamese neural network generates a target graph representation containing local feature nodes and global representation nodes based on the protein structure file to be processed, and obtains the target local representation of the target graph representation based on the global representation nodes; the target solver uses a target replacement matrix corresponding to the target local representation and a graph matching algorithm to match each protein structure in the protein structure file to be processed to obtain a target matching result, and optimizes the target local representation based on the target matching result to obtain the target representation of the protein structure file to be processed. Therefore, after the protein structure file to be processed is input into the target representation generation model, the target Siamese neural network of the target representation generation model generates a target graph representation containing local feature nodes and global representation nodes based on the protein structure file to be processed, and obtains the target local representation of the target graph representation based on the global representation nodes. In other words, by introducing global representation nodes and embedding them into the graph representation of the protein structure, this application significantly enhances the model's ability to capture global information, thereby integrating the information of the entire protein structure. Furthermore, the target solver of the target representation generation model uses a graph matching algorithm and combines it with the target replacement matrix to accurately match the protein structure, which can determine the best correspondence between protein structures, thereby further improving the accuracy and efficiency of structure comparison. The final generated target representation not only contains rich local feature information, but also integrates global structural information. This representation method can more comprehensively reflect the complexity and diversity of protein structures.
[0087] Furthermore, embodiments of this application also provide an electronic device. Figure 5 This is a structural diagram of an electronic device 20 according to an exemplary embodiment. The content of the diagram should not be construed as limiting the scope of this application.
[0088] Figure 5 This is a schematic diagram of the structure of an electronic device provided in an embodiment of this application. Specifically, it may include: at least one processor 21, at least one memory 22, a power supply 23, a communication interface 24, an input / output interface 25, and a communication bus 26. The memory 22 stores a computer program, which is loaded and executed by the processor 21 to implement the relevant steps in the protein structure characterization method performed by the electronic device disclosed in any of the foregoing embodiments.
[0089] In this embodiment, the power supply 23 is used to provide operating voltage for various hardware devices on the electronic device; the communication interface 24 can create a data transmission channel between the electronic device and external devices, and the communication protocol it follows can be any communication protocol applicable to the technical solution of this application, and is not specifically limited here; the input / output interface 25 is used to acquire external input data or output data to the outside world, and its specific interface type can be selected according to specific application needs, and is not specifically limited here.
[0090] The processor 21 may include one or more processing cores, such as a quad-core processor or an octa-core processor. The processor 21 may be implemented using at least one hardware form selected from DSP (Digital Signal Processing), FPGA (Field-Programmable Gate Array), and PLA (Programmable Logic Array). The processor 21 may also include a main processor and a coprocessor. The main processor, also known as a CPU (Central Processing Unit), is used to process data in the wake-up state; the coprocessor is a low-power processor used to process data in the standby state. In some embodiments, the processor 21 may integrate a GPU (Graphics Processing Unit), which is responsible for rendering and drawing the content to be displayed on the screen. In some embodiments, the processor 21 may also include an AI (Artificial Intelligence) processor, which is used to handle computational operations related to machine learning.
[0091] In addition, the memory 22, as a carrier for resource storage, can be a read-only memory, random access memory, disk or optical disk, etc. The resources stored on it include operating system 221, computer program 222 and data 223, etc., and the storage method can be temporary storage or permanent storage.
[0092] The operating system 221 manages and controls the various hardware devices and computer programs 222 on the electronic device to enable the processor 21 to perform calculations and processing on the massive amounts of data 223 in the memory 22. The operating system can be Windows, Unix, Linux, etc. The computer program 222, in addition to including a computer program capable of performing the protein structure characterization method executed by the electronic device as disclosed in any of the foregoing embodiments, may further include computer programs capable of performing other specific tasks. The data 223 may include data received by the electronic device from external devices, as well as data collected by its own input / output interface 25.
[0093] Furthermore, this application also discloses a computer-readable storage medium for storing a computer program; wherein, when the computer program is executed by a processor, it implements the aforementioned protein structure characterization method. Specific steps of this method can be found in the corresponding content disclosed in the foregoing embodiments, and will not be repeated here.
[0094] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on its differences from other embodiments. Similar or identical parts between embodiments can be referred to interchangeably. For the apparatus disclosed in the embodiments, since it corresponds to the method disclosed in the embodiments, the description is relatively simple; relevant parts can be referred to in the method section.
[0095] Those skilled in the art will further recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, computer software, or a combination of both. To clearly illustrate the interchangeability of hardware and software, the components and steps of the various examples have been generally described in terms of functionality in the foregoing description. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application. The steps of the methods or algorithms described in conjunction with the embodiments disclosed herein can be implemented directly in hardware, software modules executed by a processor, or a combination of both. The software module may be located in random access memory (RAM), memory, read-only memory (ROM), electrically programmable programmable read-only memory (EPROM), electrically erasable programmable read-only memory (EEPROM), register, hard disk, removable disk, CD-ROM (Compact Disc Read-Only Memory), or any other form of storage medium known in the art.
[0096] Finally, it should be noted that in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.
[0097] The foregoing has provided a detailed description of a protein structure characterization method, apparatus, device, and medium provided by the present invention. Specific examples have been used to illustrate the principles and implementation methods of the present invention. The descriptions of the above embodiments are only for the purpose of helping to understand the method and core ideas of the present invention. At the same time, for those skilled in the art, there will be changes in the specific implementation methods and application scope based on the ideas of the present invention. Therefore, the content of this specification should not be construed as a limitation of the present invention.
Claims
1. A method for characterizing protein structure, characterized in that, include: The protein structure file to be processed is input into the target representation generation model; wherein, the target representation generation model includes a target Siamese neural network and a target solver; The target Siamese neural network generates a target graph representation containing local feature nodes and global representation nodes based on the protein structure file to be processed, and obtains the target local representation of the target graph representation based on the global representation nodes. The target solver uses a target replacement matrix corresponding to the target local representation and a graph matching algorithm to match each protein structure in the protein structure file to be processed to obtain the target matching result. Based on the target matching result, the target local representation is optimized to obtain the target representation of the protein structure file to be processed.
2. The protein structure characterization method according to claim 1, characterized in that, Based on the protein structure file to be processed, a target graph representation containing local feature nodes and global representation nodes is generated, including: Each amino acid in the protein structure file to be processed is identified as a local feature node, and local association edges between the local feature nodes are established based on the distance between them, so as to generate an initial graph representation containing the local feature nodes and the local association edges. Average pooling is performed on each of the local feature nodes to obtain a global representation node, and global association edges are established between the global representation node and each of the local feature nodes. The global representation node and each of the global association edges are added to the initial graph representation to obtain the target graph representation.
3. The protein structure characterization method according to claim 2, characterized in that, Based on the global representation of the global representation node, the target local representation of the target graph is obtained, including: Generate each initial local representation of the target graph, and perform average pooling on each initial local representation to obtain the global representation of the global representation node; The initial local representations are optimized based on the global representations of the global representation nodes to obtain the target local representations of the target graph.
4. The protein structure characterization method according to claim 1, characterized in that, Before inputting the protein structure file to be processed into the target characterization generation model, the process also includes: Collect historical protein structure files and construct a historical graph representation containing historical global representation nodes based on each historical protein structure file; Based on the aforementioned historical graph representations, a representation training dataset corresponding to the metric learning loss function is established; An initial representation generation model, including an initial Siamese neural network and an initial solver, is constructed, and the initial representation generation model is trained using the representation training dataset and the metric learning loss function to obtain a target representation generation model including a target Siamese neural network and a target solver.
5. The protein structure characterization method according to claim 4, characterized in that, The metric learning loss function is the triplet loss function; Accordingly, the step of establishing a representation training dataset corresponding to the metric learning loss function based on each of the historical graph representations includes: Obtain the template modeling score among the historical graph representations, and construct a representation training dataset containing multiple triplet data based on the template modeling score.
6. The protein structure characterization method according to claim 5, characterized in that, Construct the current triplet data, including: The target protein structure and positive and negative samples of the target protein structure are determined from each of the historical graph representations, and a current triplet data containing the target protein structure, the positive sample, and the negative sample is constructed; wherein the template modeling score between the positive sample and the target protein structure is greater than a preset threshold, and the template modeling score between the negative sample and the target protein structure is less than the preset threshold.
7. The protein structure characterization method according to any one of claims 1 to 6, characterized in that, Also includes: The functional labels of each of the historical protein structure files are determined, and the historical representations of the historical protein structure files are obtained using the target representation generation model, so as to establish a prediction training dataset containing the historical representations and the corresponding functional labels. The initial function prediction model is trained using the prediction training dataset to obtain the target function prediction model, so as to predict the target function corresponding to the target representation.
8. A protein structure characterization device, characterized in that, include: The file input module is used to input the protein structure file to be processed into the target representation generation model; wherein, the target representation generation model includes a target Siamese neural network and a target solver; The first processing module is used to generate a target graph representation containing local feature nodes and global representation nodes based on the protein structure file to be processed through the target Siamese neural network, and to obtain the target local representation of the target graph representation based on the global representation of the global representation nodes. The second processing module is used to match each protein structure in the protein structure file to be processed using the target solver based on the target replacement matrix corresponding to the target local representation and using a graph matching algorithm to obtain the target matching result, and to optimize the target local representation based on the target matching result to obtain the target representation of the protein structure file to be processed.
9. An electronic device, characterized in that, include: Memory, used to store computer programs; A processor for executing the computer program to implement the steps of the protein structure characterization method as described in any one of claims 1 to 7.
10. A computer-readable storage medium, characterized in that, Used to store a computer program; wherein, when executed by a processor, the computer program implements the steps of the protein structure characterization method as described in any one of claims 1 to 7.