Protein classification method and system based on point cloud and graph neural network
Through the method of point cloud and graph neural network, the graph characteristics and overall spatial conformation characteristics of the protein structure diagram are extracted and fusion processed, which solves the problem of low protein classification accuracy in the existing technology, and achieves higher classification accuracy and voluntary Great.
Patent Information
- Application Number
- CN202411278996.2
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Priority Date
- 2024-08-20
- Filing Date
- 2024-09-12
- Publication Date
- 2024-12-17
AI Technical Summary
Existing protein classification methods fail to fully consider the overall structure and conformation information of the protein when constructing the protein map structure, resulting in low classification accuracy.
Using a method based on point cloud and graph neural network, the graph feature vectors of protein structure diagrams are extracted through geometric perception graph neural networks, and the structural diagrams are normalized to construct point clouds, and the overall spatial conformation feature vectors are extracted in combination with point cloud neural networks. , perform feature fusion, and finally use multi-perception layers for classification processing.
It enhances the expression ability of classification characteristics, improves the accuracy and robustness of protein classification, ensures the high accuracy of feature vectors, and solves the problem of low classification accuracy.
Smart Images
Figure CN119152944A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of data classification, and more specifically, to a protein classification method and system based on point cloud and graph neural network. Background Art
[0002] In recent years, with the rapid development of protein structure prediction algorithms, protein sequence and structure data have shown an exponential growth, and most of these data are now publicly available. This progress has greatly promoted the development of protein science, allowing researchers to use these rich data sets and make significant progress in multiple downstream tasks such as protein structure prediction and functional annotation through machine learning models.
[0003] At present, protein classification methods based on machine learning mainly focus on feature extraction of protein three-dimensional structural data. These methods construct protein graph structures through extracted features, and classify and predict the functions of proteins on this basis. Although these methods have achieved protein classification and function prediction to a certain extent, they do not fully consider the overall structure and conformation information of proteins. In the process of constructing protein graphs, some key information may be lost, which limits the accuracy of the classification algorithm and has the defect of low classification accuracy. Summary of the invention
[0004] In order to overcome the low precision defect of the existing protein classification technology, the present invention proposes the following technical solution:
[0005] In a first aspect, the present invention proposes a protein classification method based on point cloud and graph neural network, comprising:
[0006] Construct a protein structure diagram based on the three-dimensional structure of the protein.
[0007] The protein structure graph is subjected to feature extraction using a geometrically-aware graph neural network to obtain a graph feature vector.
[0008] The protein structure diagram is normalized to construct a protein point cloud.
[0009] The protein point cloud is input into a point cloud neural network for feature extraction to obtain an overall spatial conformation feature vector.
[0010] The graph feature vector and the overall spatial conformation feature vector are fused to obtain protein classification features.
[0011] The protein classification features are classified using multiple perception layers to obtain protein classification results.
[0012] As a preferred technical solution, the protein structure diagram is normalized to construct a protein point cloud, including:
[0013] Calculate the center point c of the point cloud, and its expression is as follows:
[0014]
[0015] in, Represents the i-th point in the point cloud, each point p i With three-dimensional coordinates (x i ,y i ,z i ), N is the number of points in the point cloud.
[0016] The point cloud is translated to the center point as the origin. The expression is as follows:
[0017] p′ i =p i -c
[0018] Among them, p′ i For each point p in the point cloud i The new point obtained after translation.
[0019] Calculate the maximum distance m from the point in the point cloud to the new origin after translation. The expression is as follows:
[0020]
[0021] For each point p′ in the point cloud i Zoom into the unit sphere and get the normalized point p″ i , whose expression is as follows:
[0022]
[0023] Using the normalized point p″ i Constructing protein point cloud
[0024] As a preferred technical solution, the point cloud neural network includes a first abstract layer, a second abstract layer and a third abstract layer. The protein point cloud is input into the point cloud neural network for feature extraction to obtain the overall spatial conformation feature vector, including:
[0025] A sampling point set S is constructed from the protein point cloud P′, and k points are selected from the sampling point set S as sampling points.
[0026] In the first abstraction layer, feature extraction is performed on the local area of each sampling point to obtain the first local area feature.
[0027] In the second abstract layer, the first local domain feature is compressed to obtain a second local domain feature.
[0028] In the third abstract layer, the second local domain features are integrated, and the integration results are feature encoded to obtain the overall spatial conformation feature vector.
[0029] As a preferred technical solution, the farthest point sampling method is used to construct a sampling point set S from the protein point cloud P′, including:
[0030] Select the first sampling point s0, whose expression is as follows:
[0031]
[0032] For the subsequent sampling points s i , perform iterative sampling until |S|=k, and obtain the sampling point set S. Among them, the subsequent sampling point s i The following iteration selection conditions are met:
[0033]
[0034] Among them, s j To remove s i For other sampling points outside the range, k is the preset number of sampling points, and d is the preset minimum distance threshold.
[0035] As a preferred technical solution, a protein structure diagram is constructed according to the three-dimensional structure of the protein, including:
[0036] Extract the α carbon atom of the protein as a node of the protein structure graph
[0037] For each α carbon atom node, a sphere with a radius of 10 angstroms is constructed with the α carbon atom node as the center, and within the sphere, other α carbon atom nodes whose distance from the central α carbon atom node is greater than or equal to 5 angstroms and less than or equal to 10 angstroms are identified, and Euclidean distance edges are constructed between the central α carbon atom node and the other identified α carbon atom nodes.
[0038] According to the sequence arrangement of protein amino acids, a sequential edge is constructed between two amino acids that are adjacent or have a distance of no more than 2 amino acids in the sequence.
[0039] For each α carbon atom node, identify the 10 other α carbon atom nodes closest to it, and construct spatial neighboring edges with α carbon atom nodes whose distance is greater than or equal to 5 angstroms.
[0040] The α-carbon atom nodes, Euclidean distance edges, sequential edges, and spatial neighboring edges are integrated to obtain the protein structure graph.
[0041] As a preferred technical solution, the sequential edges include type 0 edges, type 1 edges, type -1 edges, type 2 edges and type -2 edges.
[0042] The O-type edge is constructed between two adjacent amino acids in the sequence of protein amino acids.
[0043] The type 1 edge is constructed between one amino acid relative to the next amino acid in the sequence of protein amino acids.
[0044] The -1 type edge is constructed between one amino acid relative to the previous amino acid in the sequence of protein amino acids.
[0045] The type 2 edge is constructed between one amino acid and the next two amino acids in the protein amino acid sequence.
[0046] The -2 type edge is constructed between one amino acid relative to the two preceding amino acids in the sequence of protein amino acids.
[0047] As a preferred technical solution, the geometric perception graph neural network includes sequentially connecting a plurality of graph convolution layers, using the geometric perception graph neural network to extract features of the protein structure graph to obtain a graph feature vector, including:
[0048] The protein structure graph is input into the geometric perception graph neural network, the node features and edge features of the protein structure graph are updated layer by layer through a plurality of graph convolution layers, and the node features and edge features output by the last graph convolution layer are fused to obtain a graph feature vector.
[0049] As a preferred technical solution, before updating the node features and edge features of the protein structure graph layer by layer through a plurality of graph convolution layers, the method further includes:
[0050] An edge feature graph is created, and the node features and the edge features are discretized in angle using the edge feature graph.
[0051] As a preferred technical solution, the protein classification features are classified using multiple perception layers to obtain protein classification results, including:
[0052] The protein classification features are input into a three-layer multi-perception layer. Each multi-perception layer performs linear transformation on the input protein classification features in turn and applies the ReLU activation function to classify the protein classification features to obtain the protein classification results.
[0053] In a second aspect, the present invention further proposes a protein classification system based on point cloud and graph neural network, which is applied to the protein classification method based on point cloud and graph neural network as described in any scheme of the first aspect, comprising:
[0054] The first building module is used to construct a protein structure diagram according to the three-dimensional structure of the protein.
[0055] The first extraction module is used to extract features from the protein structure graph using a geometrically-aware graph neural network to obtain a graph feature vector.
[0056] The second construction module is used to normalize the protein structure diagram to construct a protein point cloud.
[0057] The second extraction module is used to input the protein point cloud into a point cloud neural network for feature extraction to obtain an overall spatial conformation feature vector.
[0058] The fusion module is used to fuse the graph feature vector and the overall spatial conformation feature vector to obtain protein classification features.
[0059] The classification module is used to classify the protein classification features using multiple perception layers to obtain protein classification results.
[0060] The beneficial effects of the present invention include at least:
[0061] The present invention first uses a geometrically-aware graph neural network to extract features from a protein structure graph and obtains a graph feature vector. This process not only captures the local structural features of the protein, but also acquires topological and geometric properties through geometric perception capabilities. Secondly, the present invention normalizes the protein structure graph, constructs a protein point cloud, and applies a point cloud neural network to extract the overall spatial conformation feature vector, thereby capturing the conformational information of the protein in the overall space. The fusion of these two feature vectors complements each other with key information that may be lost in single feature extraction, enhances the expressive power of classification features, and ensures that the feature vectors used for downstream classification tasks have higher accuracy and robustness, thereby improving the accuracy of protein classification technology. BRIEF DESCRIPTION OF THE DRAWINGS
[0062] Figure 1 This is a schematic diagram of the process of the protein classification method based on point cloud and graph neural network provided in Example 1.
[0063] Figure 2 This is a framework diagram for implementing the protein classification method based on point cloud and graph neural network technology provided in Example 2.
[0064] Figure 3 The data flow diagram of the point cloud neural network provided in Example 2.
[0065] Figure 4 This is an architectural diagram of the protein classification system based on point cloud and graph neural network provided in Example 3. DETAILED DESCRIPTION
[0066] The following will describe the embodiments of the present invention with reference to the accompanying drawings and preferred technical solutions. Those skilled in the art can easily understand other advantages and effects of the present invention from the contents disclosed in this specification. The present invention can also be implemented or applied through other different specific embodiments, and the details in this specification can also be modified or changed in various ways based on different viewpoints and applications without departing from the spirit of the present invention. It should be understood that the preferred technical solutions are only for illustrating the present invention, not for limiting the scope of protection of the present invention.
[0067] It should be noted that the illustrations provided in the following embodiments are only schematic illustrations of the basic concept of the present invention, and thus the drawings only show components related to the present invention rather than being drawn according to the number, shape and size of components in actual implementation. In actual implementation, the type, quantity and proportion of each component may be changed arbitrarily, and the component layout may also be more complicated.
[0068] In the following description, numerous details are discussed to provide a more thorough explanation of the embodiments of the present invention. However, it is obvious to those skilled in the art that the embodiments of the present invention can be implemented without these specific details. In other embodiments, well-known structures and devices are shown in the form of block diagrams rather than in detail to avoid making the embodiments of the present invention difficult to understand.
[0069] Example 1
[0070] This example proposes a protein classification method based on point cloud and graph neural network. Figure 1 As shown, Figure 1 A schematic diagram of a protein classification method based on point cloud and graph neural network provided in this embodiment, the method comprises the following steps:
[0071] S1: Construct a protein structure diagram based on the three-dimensional structure of the protein.
[0072] S2: Using a geometrically-aware graph neural network to extract features from the protein structure graph to obtain a graph feature vector.
[0073] S3: normalizing the protein structure diagram to construct a protein point cloud.
[0074] S4: Inputting the protein point cloud into a point cloud neural network for feature extraction to obtain an overall spatial conformation feature vector.
[0075] S5: fusing the graph feature vector and the overall spatial conformation feature vector to obtain protein classification features.
[0076] S6: Using multiple perception layers to classify the protein classification features to obtain protein classification results.
[0077] It is understandable that this embodiment first uses a geometric perception graph neural network to extract features from the protein structure graph to obtain a graph feature vector. This process not only captures the local structural features of the protein, but also obtains the topological and geometric properties through geometric perception capabilities. Secondly, the present invention normalizes the protein structure graph, constructs a protein point cloud, and applies a point cloud neural network to extract the overall spatial conformation feature vector, thereby capturing the conformation information of the protein in the overall space. The fusion of these two feature vectors complements each other with key information that may be lost in single feature extraction, enhances the expressive power of classification features, and ensures that the feature vectors used for downstream classification tasks have higher accuracy and robustness, thereby improving the accuracy of protein classification technology.
[0078] Example 2
[0079] like Figure 2 As shown, this embodiment provides a more detailed description based on the protein classification method based on point cloud and graph neural network technology proposed in Example 1, and preprocesses the public protein dataset for downstream evaluation tasks. This embodiment involves two tasks: Enzyme Commission (EC) number prediction and Gene Ontology (GO) term prediction.
[0080] For the EC number prediction task, the goal is to predict the EC numbers of different proteins, which describe the catalytic role of proteins in biochemical reactions. The protein sequence data used comes from the non-redundant protein data bank (nrPDB), which is a curated version of the protein data bank (PDB) in which highly similar sequences have been removed. The EC numbers are classified based on the third and fourth levels of the EC tree, forming a multi-label binary classification task with 538 labels.
[0081] Another task is GO term prediction, which aims to predict whether a protein belongs to a specific GO term. These terms classify proteins into three hierarchical ontology categories: molecular function (MF), biological process (BP), and cellular component (CC). Similar to the EC task, the protein sequences used in the GO task are also derived from the nrPDB database. The sequence data is processed by multi-label binary classification, and each ontology of MF, BP, and CC is used as an independent classification task, with the associated number of labels being 489, 1943, and 320, respectively.
[0082] In order to ensure the quality of the protein data set and the accuracy of the evaluation, this embodiment adopts a multi-truncation partitioning method to ensure that the sequence similarity between the PDB chain in the test set and the training set does not exceed 95%. The acquired data set is divided into three subsets: training set, validation set and test set to ensure the consistency of the data set and make the comparison between the classification models based on point cloud neural network and geometric perception graph neural network more accurate. Among them, the training set is used to train the classification model based on point cloud neural network and geometric perception graph neural network, and the classification model parameters are adjusted by letting the classification model learn the relationship between the three-dimensional structure of the protein and its corresponding EC number and GO term; the validation set is used to tune the classification model during the training process, and the hyperparameters are adjusted and overfitting is prevented by evaluating the performance of the classification model on unseen validation data; and the test set is used to finally evaluate the generalization ability of the classification model, and through independent data, ensure that the model can accurately predict the EC number and GO term of the protein in practical applications, thereby verifying the performance of the model in the real world.
[0083] Specifically, the dataset of the EC task covers 18,716 protein 3D structures in the nrPDB database, which are further divided into a training set (15,170 proteins), a validation set (1,686 proteins), and a test set (1,860 proteins). The dataset of the GO task includes 34,592 protein 3D structures, which are also divided into a training set (28,305 proteins), a validation set (3,139 proteins), and a test set (3,148 proteins). The division of the datasets for these two tasks is listed in Table 1.
[0084] Table 1 Information of two task datasets
[0085]
[0086]
[0087] In this embodiment, a protein structure diagram is constructed based on the three-dimensional structure of the protein, including:
[0088] Extract the α carbon atom of the protein as a node of the protein structure graph
[0089] For each α carbon atom node, a sphere with a radius of 10 angstroms is constructed with the α carbon atom node as the center, and within the sphere, other α carbon atom nodes whose distance from the central α carbon atom node is greater than or equal to 5 angstroms and less than or equal to 10 angstroms are identified, and Euclidean distance edges are constructed between the central α carbon atom node and the other identified α carbon atom nodes.
[0090] According to the sequence arrangement of protein amino acids, a sequential edge is constructed between two amino acids that are adjacent or have a distance of no more than 2 amino acids in the sequence.
[0091] For each α carbon atom node, identify the 10 other α carbon atom nodes closest to it, and construct spatial neighboring edges with α carbon atom nodes whose distance is greater than or equal to 5 angstroms.
[0092] The α-carbon atom nodes, Euclidean distance edges, sequential edges, and spatial neighboring edges are integrated to obtain the protein structure graph.
[0093] In this embodiment, the sequential edges include type 0 edges, type 1 edges, type -1 edges, type 2 edges and type -2 edges.
[0094] The O-type edge is constructed between two adjacent amino acids in the sequence of protein amino acids.
[0095] The type 1 edge is constructed between one amino acid relative to the next amino acid in the sequence of protein amino acids.
[0096] The -1 type edge is constructed between one amino acid relative to the previous amino acid in the sequence of protein amino acids.
[0097] The type 2 edge is constructed between one amino acid and the next two amino acids in the protein amino acid sequence.
[0098] The -2 type edge is constructed between one amino acid relative to the two preceding amino acids in the sequence of protein amino acids.
[0099] In the specific implementation process, in order to achieve a more accurate characterization of the protein structure, this embodiment adopts a method of reorganizing the connections between the residues within the protein. First, by extracting the protein structure diagram containing only the α carbon in the residue and removing the side chain information, the model is simplified and focused on the main chain structure of the protein. In this step, all the original edge information and connections are removed, laying the foundation for the subsequent edge construction.
[0100] Next, a series of steps are implemented to redefine the connection relationship between residues. First, based on the spatial coordinates of each α-carbon, a sphere with a radius of 10 angstroms is drawn, and other α-carbons within the sphere that are more than 5 angstroms away from the central α-carbon are identified to construct edges based on Euclidean distance. Secondly, according to the order of amino acids in the sequence, sequential edges are formed for amino acids with a distance of less than 2, and are subdivided into five types according to the sequential distance: -2, -1, 0, 1, and 2. In addition, based on the spatial neighbor relationship of α-carbons, the 10 nearest α-carbons are found for each α-carbon, and α-carbons with a distance less than 5 angstroms are excluded, and the remaining α-carbons form edges with the central α-carbon. Finally, the 7 types of edges obtained in these steps are integrated into the first protein structure diagram to form a new, more accurate protein structure diagram, which provides a structural basis for subsequent graph neural network analysis.
[0101] In this embodiment, the geometric perception graph neural network includes connecting a plurality of graph convolution layers in sequence, and using the geometric perception graph neural network to extract features of the protein structure graph to obtain a graph feature vector, including:
[0102] The protein structure graph is input into the geometric perception graph neural network, the node features and edge features of the protein structure graph are updated layer by layer through a plurality of graph convolution layers, and the node features and edge features output by the last graph convolution layer are fused to obtain a graph feature vector.
[0103] In this embodiment, before updating the node features and edge features of the protein structure graph layer by layer through a plurality of graph convolution layers, the method further includes:
[0104] An edge feature graph is created, and the node features and the edge features are discretized in angle using the edge feature graph.
[0105] In the specific implementation process, this embodiment conducts in-depth characterization learning on the new protein structure graph. First, each protein P is represented as a relationship graph G P , which is composed of (V, E, R). V represents a node set, each node in the node set represents a residue and contains the type of the amino acid residue and its coordinate information in three-dimensional space. E represents the edge set between nodes, whose type is R. The edge set defines the relationship between nodes, including the connection between residues based on the protein sequence and the geometric proximity based on the spatial coordinates. In order to further enrich the structural information of the graph, this embodiment introduces an edge feature graph. The edge feature graph of the protein is represented as a relationship graph G E , by (V E ,E E ,R E ) is composed of. Node set V E Represents the edges between residues in the protein. The set of edges between vertices E E It can be considered as the angle between the vertices of two edges. The size of this angle is determined by R E Indicates that , are also included in this set.
[0106] Next, this embodiment constructs a network containing multiple graph convolutional layers, which are responsible for updating the features of nodes and edges layer by layer. Each hidden layer corresponds to a graph convolutional layer, which is responsible for processing the feature information of nodes and edges. In the first stage of forward propagation, the input graph structure and node features are fed into the first convolutional layer, and the edge feature graph is used for angle discretization to calculate the first edge features. In each subsequent layer, the graph convolutional layer updates the node features and edge features, and calculates new node features by combining the updated edge features with the original node features. This process is repeated in all layers of the network until the last layer, whose output node features are used as the final graph feature vector of the protein.
[0107] In this embodiment, the protein structure diagram is normalized to construct a protein point cloud, including:
[0108] Calculate the center point c of the point cloud, and its expression is as follows:
[0109]
[0110] in, Represents the i-th point in the point cloud, each point p i With three-dimensional coordinates (x i ,y i ,z i ), N is the number of points in the point cloud.
[0111] The point cloud is translated to the center point as the origin. The expression is as follows:
[0112] p′ i =p i -c
[0113] Among them, p′ i For each point p in the point cloud i The new point obtained after translation.
[0114] Calculate the maximum distance m from the point in the point cloud to the new origin after translation. The expression is as follows:
[0115]
[0116] For each point p′ in the point cloud i Zoom into the unit sphere and get the normalized point p″ i , whose expression is as follows:
[0117]
[0118] Using the normalized point p″ i Constructing protein point cloud
[0119] In this embodiment, the point cloud neural network includes a first abstraction layer, a second abstraction layer and a third abstraction layer. The protein point cloud is input into the point cloud neural network for feature extraction to obtain an overall spatial conformation feature vector, including:
[0120] A sampling point set S is constructed from the protein point cloud P′, and k points are selected from the sampling point set S as sampling points.
[0121] In the first abstraction layer, feature extraction is performed on the local area of each sampling point to obtain the first local area feature.
[0122] In the second abstract layer, the first local domain feature is compressed to obtain a second local domain feature.
[0123] In the third abstract layer, the second local domain features are integrated, and the integration results are feature encoded to obtain the overall spatial conformation feature vector.
[0124] In this embodiment, the farthest point sampling method is used to construct a sampling point set S from the protein point cloud P′, including:
[0125] Select the first sampling point s0, whose expression is as follows:
[0126]
[0127] For the subsequent sampling points s i , perform iterative sampling until |S|=k, and obtain the sampling point set S. Among them, the subsequent sampling point s i The following iteration selection conditions are met:
[0128]
[0129] Among them, s j To remove s i For other sampling points outside the range, k is the preset number of sampling points, and d is the preset minimum distance threshold.
[0130] In the specific implementation process, Figure 3 As shown, in this embodiment, a point cloud neural network is designed to process the original protein structure diagram and extract the three-dimensional spatial configuration feature vector therefrom. The network architecture consists of three abstract layers: the first abstract layer, the second abstract layer and the third abstract layer. Each abstract layer is responsible for extracting and processing the local neighborhood features of the sampling points. The first abstract layer first uses the farthest point sampling (FPS) method to select a set of sampling points from the point cloud to ensure that these points are evenly distributed throughout the point cloud. The FPS algorithm selects through iterative selection, each time selecting the point farthest from all points in the selected point set until the preset number of sampling points k is reached. This step not only extracts the key feature points in the protein point cloud, but also ensures the comparability of the features by maintaining the consistency of the number of sampling points in all protein point clouds.
[0131] Subsequently, the second abstraction layer further compresses the number of points in the point cloud, refining the feature representation by reducing the number of points required to describe the overall structure of the protein. The third abstraction layer no longer performs sampling, but uses all sampling points to encode and integrate local neighborhood features through a series of multi-layer perceptron (MLP) convolution units, and finally outputs a feature vector representing the overall structure of the protein. This progressive feature extraction method from local to global is similar to protein structure analysis from the atomic to subunit level, and can fully capture the spatial conformational characteristics of proteins.
[0132] In this embodiment, the protein classification features are classified using multiple perception layers to obtain protein classification results, including:
[0133] The protein classification features are input into a three-layer multi-perception layer. Each multi-perception layer performs linear transformation on the input protein classification features in turn and applies the ReLU activation function to classify the protein classification features to obtain the protein classification results.
[0134] In the specific implementation process, the fused protein classification feature vector is input into a three-layer multi-layer perception layer (MLP) for further processing. In this process, each layer of the MLP performs a linear transformation on the input features and applies the ReLU activation function. This layer-by-layer processing method helps to deeply extract and transform features and enhance the model's understanding of protein functional characteristics. After layer-by-layer refinement of the MLP, the final output is the functional classification result of the protein. This method of combining graph structural features and spatial configuration features effectively improves the accuracy of protein classification because it can capture the key features of proteins and provide a more accurate decision-making basis for classification.
[0135] Example 3
[0136] This example conducts a comparative test based on the protein classification method and data set preprocessing based on point cloud and graph neural network technology proposed in the first two examples. This study aims to verify the efficiency of this method in protein classification and the effectiveness of the point cloud neural network module by testing the classification performance of 12 methods on the EC dataset and the classification performance of the two most efficient methods on the GO dataset.
[0137] This example uses Fmax and the area under the micro-average precision-recall curve (AUPRC) as the main evaluation indicators. AUPRC comprehensively measures the prediction ability of the model in each category by calculating the area under the precision-recall curve, which is particularly suitable for the evaluation of unbalanced data sets in multi-label classification tasks. Fmax combines the harmonic mean of precision and recall to select the balance point between precision and recall. It is usually used in application scenarios that require a balance between precision and recall, such as classification tasks in information retrieval and bioinformatics.
[0138] The PointNet-GearNet-Edge model structure is divided into two parts. The first part is the geometric edge model part, which is responsible for processing the protein main chain feature information. In this part, the protein main chain features are defined, the protein graph structure is reconstructed, and then the feature extraction is performed through a six-layer geometric convolutional graph neural network. The second part is to extract the overall structural features of the protein. First, the three-dimensional structure of the protein is extracted, the coordinates of each point are extracted to construct a point cloud, and only the overall structural features are retained, and the features of each node are discarded. Then normalization is performed, and the overall structural features of the protein are obtained through three levels of abstraction. Finally, the structural and overall structural features of the protein graph obtained from the two parts are merged into the MLP layer, and the probability value of each multi-label binary classification task is finally output.
[0139] For the hyperparameter setting, the number of layers of the gearnet-edge model in the first part is set to 6, and the hidden layer dimension size is 512. In the point cloud neural network in the second part, the configuration details of the three set abstraction layers (sa1, sa2, sa3) are as follows: the sa1 layer samples 128 points from the original point cloud data, and samples 4, 8, and 16 neighborhood points in spherical areas with radii of 0.2, 0.4, and 0.6, respectively. The sa2 layer samples 64 points from the 323-dimensional point features based on the output of sa1, and samples 4, 8, and 16 neighborhood points in spherical areas with radii of 0.4, 0.6, and 0.8, respectively. After the neighborhood points pass through the MLP encoder, the number of feature channels becomes 128, 256, and 256, respectively. The sa3 layer does not perform point sampling, but directly uses the MLP encoder on the 643-dimensional point features output by sa2 to output 512-dimensional global features.
[0140] For the training strategy of the PointNet-GearNet-Edge model, the AdamW optimizer, BCE loss function, and AUPRCmicro and Fmax evaluation indicators are used. The learning rate is 0.0001 and the weight decay is 0. Every 5 epochs, if the indicator does not improve, the learning rate is reduced by 0.6 times. A total of 200 epochs are trained with a batch size of 8 and a seed setting of 1024. The training strategy and seed setting are consistent with the gearnet-edge model (previous work) for easy comparison and reproduction.
[0141] In this example, the previously proposed model is compared with many existing protein representation learning methods, including four sequence-based models (CNN, ResNet, LSTM, and Transformer), and seven structure-based models (GCN, GAT, GVP, 3DCNN, MQA, GraphQA, New IEConv, and Gearnet-Edge). The comparison results show that the PointNet-GearNet-Edge model performs best in terms of classification accuracy, highlighting the effectiveness of integrating the point cloud neural network with the graphical framework of the geometry-aware neural network. The features extracted by the point cloud neural network mainly represent the overall spatial structural characteristics of the protein. Each point corresponds to an atom in the protein, retaining only the position information and ignoring other details. This method of integrating the overall spatial structural information improves the performance of the model and highlights the necessity of integrating the overall spatial structural information in protein representation.
[0142] Table 2 Important parameters for each model implementation
[0143]
[0144]
[0145] Table 3 Predicted Fmax scores in the test set of the EC dataset
[0146] Method EC CNN (2020) 0.619 ResNet (2019) 0.613 LSTM (2019) 0.425 Transformer (2019) 0.238 GCN (2017) 0.320 GAT(2018) 0.368 GVP(2021) 0.489 3DCNN MQA (2018) 0.077 GraphQA(2021) 0.509 New IEConv(2022) 0.735 GearNet-Edge(2023) 0.814 PointNet-GearNet-Edge 0.829
[0147] This example specifically selected the GearNet-Edge and PointNet-GearNet-Edge models and conducted in-depth tests on the GO dataset. GO (Gene Ontology) terms are widely used to describe the properties of proteins, which include three main categories: Molecular Function (MF), Biological Process (BP), and Cellular Component (CC). This example carefully evaluated the performance of the model on these three categories and recorded the test results, as shown in Table 4.
[0148] The test results clearly demonstrate the significant advantages of integrating point cloud neural network technology into the GearNet-Edge model. This integration significantly enhances the model's ability to capture and utilize complex spatial and structural information of proteins. Specifically, the PointNet-GearNet-Edge model, through its advanced point cloud processing capabilities, is able to more accurately identify and characterize the spatial configuration of proteins, which is crucial for understanding the molecular function of proteins.
[0149] In addition, this enhanced characterization capability is not limited to a single protein structure, but can also be extended to the level of entire biological processes and cellular components. This means that by using the PointNet-GearNet-Edge model, this embodiment can more comprehensively understand the various functions and mechanisms of action of proteins in organisms.
[0150] Ultimately, these findings not only provide a powerful new tool for the field of bioinformatics, but also have important implications for advancing our understanding of protein function and their role in health and disease. These properties of the PointNet-GearNet-Edge model make it a promising tool that will help accelerate biomedical research and may provide new perspectives and approaches for disease treatment and drug development.
[0151] Table 4 Fmax and AUPRC scores in GO dataset prediction
[0152]
[0153] Example 4
[0154] like Figure 4 As shown, this embodiment proposes a protein classification system based on point cloud and graph neural network, which is applied to the protein classification method based on point cloud and graph neural network as described in the above embodiment, including: a first building module 100, a first extraction module 200, a second building module 300, a second extraction module 400, a fusion module 500 and a classification module 600.
[0155] Among them, the first construction module 100 is used to construct a protein structure graph according to the three-dimensional structure of the protein. The first extraction module 200 is used to use a geometric perception graph neural network to extract features from the protein structure graph to obtain a graph feature vector. The second construction module 300 is used to normalize the protein structure graph and construct a protein point cloud. The second extraction module 400 is used to input the protein point cloud into the point cloud neural network for feature extraction to obtain an overall spatial conformation feature vector. The fusion module 500 is used to fuse the graph feature vector and the overall spatial conformation feature vector to obtain a protein classification feature. The classification module 600 is used to classify the protein classification feature using multiple perception layers to obtain a protein classification result.
[0156] It should be noted that the aforementioned explanation of the embodiment of the protein classification method based on point cloud and graph neural network is also applicable to the protein classification system based on point cloud and graph neural network in this embodiment, and will not be repeated here.
[0157] In the description of this specification, the description with reference to the terms "one embodiment", "some embodiments", "example", "specific example", or "some examples" etc. means that the specific features, structures, materials or characteristics described in conjunction with the embodiment or example are included in at least one embodiment or example of the present application. In this specification, the schematic representations of the above terms are not necessarily directed to the same embodiment or example. Moreover, the specific features, structures, materials or characteristics described may be combined in any one or N embodiments or examples in a suitable manner. In addition, those skilled in the art may combine and combine the different embodiments or examples described in this specification and the features of the different embodiments or examples, without contradiction.
[0158] In addition, the terms "first" and "second" are used for descriptive purposes only and should not be understood as indicating or implying relative importance or implicitly indicating the number of the indicated technical features. Therefore, the features defined as "first" and "second" may explicitly or implicitly include at least one of the features. In the description of this application, the meaning of "N" is at least two, such as two, three, etc., unless otherwise clearly and specifically defined.
[0159] Any process or method description in a flowchart or otherwise described herein may be understood to represent a module, fragment or portion of code comprising one or more executable instructions for implementing the steps of a custom logical function or process, and the scope of the preferred embodiments of the present application includes alternative implementations in which functions may not be performed in the order shown or discussed, including performing functions in a substantially simultaneous manner or in reverse order depending on the functions involved, which should be understood by technicians in the technical field to which the embodiments of the present application belong.
[0160] It should be understood that the various parts of the present application can be implemented by hardware, software, firmware or a combination thereof. In the above-mentioned embodiment, the N steps or methods can be implemented by software or firmware stored in a memory and executed by a suitable instruction execution system. For example, if implemented by hardware, as in another embodiment, it can be implemented by any one of the following technologies known in the art or their combination: a discrete logic circuit having a logic gate circuit for implementing a logic function for a data signal, a dedicated integrated circuit having a suitable combination of logic gate circuits, a programmable gate array, a field programmable gate array, etc.
[0161] A person skilled in the art may understand that all or part of the steps in the above-mentioned embodiment method may be completed by instructing related hardware through a program, and the program may be stored in a computer-readable storage medium, which, when executed, includes one or a combination of the steps of the method embodiment.
[0162] Obviously, the above embodiments of the present invention are merely examples for clearly illustrating the present invention, and are not intended to limit the embodiments of the present invention. For those skilled in the art, other different forms of changes or modifications can be made based on the above description. It is not necessary and impossible to list all the embodiments here. Any modifications, equivalent substitutions and improvements made within the spirit and principles of the present invention should be included in the protection scope of the claims of the present invention.
Claims
1. A protein classification method based on point cloud and graph neural network, characterized in that: include: Construct a protein structure diagram based on the three-dimensional structure of the protein; Using a geometrically-aware graph neural network to extract features from the protein structure graph to obtain a graph feature vector; Normalizing the protein structure diagram to construct a protein point cloud; Inputting the protein point cloud into a point cloud neural network for feature extraction to obtain an overall spatial conformation feature vector; fusing the graph feature vector and the overall spatial conformation feature vector to obtain protein classification features; The protein classification features are classified using multiple perception layers to obtain protein classification results.
2. The protein classification method based on point cloud and graph neural network according to claim 1, characterized in that: The protein structure diagram is normalized to construct a protein point cloud, including: Calculate the center point c of the point cloud, and its expression is as follows: in, Represents the i-th point in the point cloud, each point p i With three-dimensional coordinates (x i ,y i ,z i ), N is the number of points in the point cloud; The point cloud is translated to the center point as the origin. The expression is as follows: p′ i =p i -c Among them, p′ i For each point p in the point cloud i The new point obtained after translation; Calculate the maximum distance m from the point in the point cloud to the new origin after translation. The expression is as follows: For each point p′ in the point cloud i Zoom into the unit sphere and get the normalized point p″ i , whose expression is as follows: Using the normalized point p″ i Constructing protein point cloud 3. The protein classification method based on point cloud and graph neural network according to claim 2, characterized in that: The point cloud neural network includes a first abstraction layer, a second abstraction layer and a third abstraction layer. The protein point cloud is input into the point cloud neural network for feature extraction to obtain an overall spatial conformation feature vector, including: Construct a sampling point set S from the protein point cloud P′, and select k points from the sampling point set S as sampling points; In the first abstraction layer, feature extraction is performed on the local area of each sampling point to obtain the first local area feature; In the second abstract layer, the first local domain feature is compressed to obtain a second local domain feature; In the third abstract layer, the second local domain features are integrated, and the integration results are feature encoded to obtain the overall spatial conformation feature vector.
4. The protein classification method based on point cloud and graph neural network according to claim 3, characterized in that: Using the farthest point sampling method, the sampling point set S is constructed from the protein point cloud P′, including: Select the first sampling point s0, whose expression is as follows: For the subsequent sampling points s i , iterative sampling is performed until |S| = k, and a sampling point set S is obtained; among which, the subsequent sampling point s i The following iteration selection conditions are met: Among them, s j To remove s i For other sampling points outside the range, k is the preset number of sampling points, and d is the preset minimum distance threshold.
5. The protein classification method based on point cloud and graph neural network according to claim 1, characterized in that: According to the three-dimensional structure of the protein, construct a protein structure diagram, including: Extract the α carbon atom of the protein as a node of the protein structure graph For each α carbon atom node, a sphere with a radius of 10 angstroms is constructed with the α carbon atom node as the center, and within the sphere, other α carbon atom nodes whose distances from the central α carbon atom node are greater than or equal to 5 angstroms and less than or equal to 10 angstroms are identified, and Euclidean distance edges are constructed between the central α carbon atom node and the identified other α carbon atom nodes; According to the sequence arrangement of protein amino acids, a sequential edge is constructed between two amino acids that are adjacent or have a distance of no more than 2 amino acids in the sequence; For each α carbon atom node, identify the 10 other α carbon atom nodes closest to it, and construct spatial neighboring edges with α carbon atom nodes whose distance is greater than or equal to 5 angstroms; The α-carbon atom nodes, Euclidean distance edges, sequential edges, and spatial neighboring edges are integrated to obtain the protein structure graph.
6. The protein classification method based on point cloud and graph neural network according to claim 5, characterized in that: The sequential edges include type 0 edges, type 1 edges, type -1 edges, type 2 edges and type -2 edges; The 0-type edge is constructed between two adjacent amino acids in the sequence of protein amino acids; The type 1 edge is constructed between one amino acid and the next amino acid in the sequence of protein amino acids; The -1 type edge is constructed between one amino acid relative to the previous amino acid in the sequence of protein amino acids; The type 2 edge is constructed between one amino acid and the two amino acids following it in the sequence of protein amino acids; The -2 type edge is constructed between one amino acid relative to the two preceding amino acids in the sequence of protein amino acids.
7. The protein classification method based on point cloud and graph neural network according to claim 5, characterized in that: The geometric perception graph neural network includes sequentially connecting a plurality of graph convolution layers, and using the geometric perception graph neural network to extract features of the protein structure graph to obtain a graph feature vector, including: The protein structure graph is input into the geometric perception graph neural network, the node features and edge features of the protein structure graph are updated layer by layer through a plurality of graph convolution layers, and the node features and edge features output by the last graph convolution layer are fused to obtain a graph feature vector.
8. The protein classification method based on point cloud and graph neural network according to claim 7, characterized in that: Before updating the node features and edge features of the protein structure graph layer by layer through a plurality of graph convolutional layers, the method further includes: An edge feature graph is created, and the node features and the edge features are discretized in angle using the edge feature graph.
9. The protein classification method based on point cloud and graph neural network according to claim 1, characterized in that: The protein classification features are classified using multiple perception layers to obtain protein classification results, including: The protein classification features are input into a three-layer multi-perception layer. Each multi-perception layer performs linear transformation on the input protein classification features in turn and applies the ReLU activation function to classify the protein classification features to obtain the protein classification results.
10. A protein classification system based on point cloud and graph neural network, characterized in that: include: The first building module is used to construct a protein structure diagram according to the three-dimensional structure of the protein; A first extraction module is used to extract features from the protein structure graph using a geometric perception graph neural network to obtain a graph feature vector; A second construction module is used to normalize the protein structure diagram to construct a protein point cloud; A second extraction module is used to input the protein point cloud into a point cloud neural network for feature extraction to obtain an overall spatial conformation feature vector; A fusion module, used for fusing the graph feature vector and the overall spatial conformation feature vector to obtain protein classification features; The classification module is used to classify the protein classification features using multiple perception layers to obtain protein classification results.