Chemical molecular diagram classification method and equipment based on structure enhancement and medium
The enhanced chemical molecular graph classification method addresses the limitations of existing methods by using multi-step random walks and structural coefficients to improve node and edge features, resulting in improved accuracy for chemical molecule classification.
Patent Information
- Application Number
- CN202411301664.1
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-09-18
- Publication Date
- 2025-01-03
AI Technical Summary
Existing graph neural network methods fail to fully capture the diverse structural information of chemical molecules, including ring structures, cross-linked structures, and mesh structures, limiting the accuracy of chemical molecule classification.
A chemical molecular graph classification method that enhances structural information by using multi-step random walk return probabilities to improve node features and structural coefficients for edge features, and incorporates subgraph representation and graph classification to capture rich contextual information.
Enhances the representation of chemical molecules, improving the accuracy of classification tasks by flexibly handling various graph structures and capturing important substructures, such as ring structures, leading to better prediction of molecular properties.
Smart Images
Figure CN119252377A_ABST
Abstract
Description
Technical Field
[0001] The present invention belongs to the technical field of graph machine learning, and specifically relates to a chemical molecule graph classification method, device and medium based on structure enhancement. Background Art
[0002] Classifying chemical molecules can determine whether they have properties such as mutagenicity, anticancer activity, and toxicity. By treating atoms and chemical bonds as nodes and edges in a graph, the structure of chemical molecules can be modeled as a topological graph. Using graph classification methods to classify chemical molecules helps to analyze and predict the biological activity and safety of molecules, thus playing an important role in drug development and chemical research. Graph classification methods can improve the efficiency of chemical molecule screening and provide a reliable theoretical basis for the discovery and development of new drugs.
[0003] A mainstream graph classification method is to use graph neural networks to extract the structural information of graphs and convert it into low-dimensional vector representations for graph classification. Existing graph neural network methods transmit and aggregate the root subtree composed of nodes and their neighborhoods as the structural information of nodes. However, the connection between atoms and bonds in chemical molecules is diverse, and the tree structure can only represent the chain structure, but cannot fully describe complex connection methods such as ring structure, cross-linked structure, mesh structure, cage structure, etc.
[0004] Designing effective graph neural network methods that can transform the diverse structural information in chemical molecule graphs into low-dimensional representations of graphs can significantly improve the accuracy of chemical molecule classification. This will have a profound impact on the research of chemical molecules and the development of new drugs. Summary of the invention
[0005] In order to solve the problem of insufficient utilization of structural information of chemical molecular graphs, the present invention provides a chemical molecular graph classification method, device and medium based on structure enhancement.
[0006] In order to achieve the above object, the present invention adopts the following technical solutions:
[0007] A chemical molecular graph classification method based on structure enhancement comprises the following steps:
[0008] Step 1, obtain chemical molecule data, use atoms as nodes and chemical bonds as edges, construct a chemical molecule graph for each chemical molecule, and obtain a chemical molecule data set;
[0009] Step 2: Construct a chemical molecular graph classification model based on structure enhancement, which includes three parts: structure enhancement, subgraph representation, and graph classification.
[0010] Step 3: Train and optimize the chemical molecular graph classification model based on structure enhancement, and conduct model performance evaluation and comparative analysis.
[0011] Furthermore, the construction of the chemical molecular graph classification model based on structure enhancement in step 2 is specifically:
[0012] For each node v in the chemical molecular graph G, the derived subgraph of its k-hop neighbor nodes is used as the root subgraph of node v Perform structural enhancement and subgraph representation on each root subgraph; pool the representations of all root subgraphs to obtain the representation of the chemical molecule graph and perform chemical molecule classification;
[0013] The structural enhancement part includes root subgraph Enhance from two aspects: node features and edge features;
[0014] The subgraph representation part includes the information within the aggregated subgraph and the information between aggregated subgraphs to the root subgraph. The nodes in are represented by average pooling, and the root subgraph is obtained by average pooling the node representations. Representation of;
[0015] The graph classification part includes summing and pooling all root subgraph representations of the chemical molecule graph G to obtain the vector representation of the chemical molecule graph G, using a graph classifier to classify the representation of the chemical molecule graph G, and predicting the properties of the corresponding chemical molecules.
[0016] Furthermore, in step 2, the root subgraph The node features to be enhanced are:
[0017] Compute root subgraph The D1-step random walk return probability of node a It is expressed as:
[0018]
[0019] in, Represents the root subgraph The probability that node a in starts to return to a after d steps of random walk;
[0020] Then use a linear layer to move node a to the root subgraph The D1-step random walk return probability on Encoded as node a corresponding to the root subgraph The hidden representation vector
[0021] Similarly, the D2-step random walk of node a on the original graph G returns the probability code Then the hidden representation vector of node a corresponding to the original chemical molecule graph G is Will With the initial node feature x a Splice to get the root subgraph Enhanced node feature representation of midpoint a Enhance the root subgraph by returning probabilities using multi-step random walks The node characteristics of the chemical molecule can measure the influence of the ring structure.
[0022] Furthermore, in step 2, the root subgraph The edge features are enhanced specifically as follows:
[0023] Root subgraph The features of edge ab are enhanced, and the L-step structural coefficient on edge ab is calculated as the edge weight. L-step structural coefficient of middle edge ab The calculation formula is:
[0024]
[0025] in, Represented in the root subgraph The l-th step structural coefficient of the edge ab in, Represents the root subgraph The common l-hop neighbors of endpoints a and b, express Zhongyou The edge set corresponding to the derived subgraph, λ>0.
[0026] Furthermore, the information in the aggregated subgraph in step 2 is specifically:
[0027] First, in each root subgraph of the chemical molecular graph G Add a node to make its vertex set The edge set remains unchanged, and the feature of the newly added node w is If the newly added node is an isolated point, then the root subgraph The adjacency matrix A v for:
[0028]
[0029] A v Normalize row by row For each node The feature vector of the t+1th layer is expressed as:
[0030]
[0031] Among them, γ (t) is a learnable scalar parameter, MLP (t+1) represents the multilayer perceptron at layer t+1, Represents the root subgraph The neighbor set of node a.
[0032] Furthermore, the information between the aggregated subgraphs in step 2 is specifically:
[0033] Let the node set with the shortest path distance k from the root node v in the chemical molecule graph G be Node The root graph for A neighbor subgraph of , then the root subgraph The information between neighbor subgraphs aggregated by node a is expressed as:
[0034]
[0035] By aggregating the intra-subgraph information and inter-subgraph information, we get the representation of node a after the t+1th iteration:
[0036]
[0037] After T iterations, Midpoint Representation Perform average pooling to obtain sub-images Representation:
[0038]
[0039] Furthermore, in the graph classification part of step 2, all root subgraph representations of the chemical molecule graph G are summed and pooled to obtain a vector representation of the chemical molecule graph G:
[0040]
[0041] Using graph classifier to represent the vector h of chemical molecular graph G G Classify and predict the properties of corresponding chemical molecules.
[0042] Furthermore, the training and optimization of the chemical molecular graph classification model based on structure enhancement in step 3 specifically includes the following steps:
[0043] Step 3.1, divide the chemical molecule dataset into 10 parts equally, take 1 part as the test set each time, and use the remaining 9 parts as the training set;
[0044] Step 3.2, input the chemical molecule graphs in the chemical molecule training set into the chemical molecule graph classification model based on structure enhancement, obtain the classification prediction value of each chemical molecule graph, and use the cross entropy loss function to calculate the difference between the prediction value and the label value. The calculation formula of the cross entropy loss function Loss is as follows:
[0045]
[0046] Among them, m represents the number of chemical molecules, n represents the number of chemical molecule categories, and yji represents the i-th component of the true label value of the j-th chemical molecule, represents the i-th component of the predicted value of the j-th chemical molecule;
[0047] After obtaining the training loss, use the Adam optimizer to optimize the model parameters;
[0048] Step 3.3, after training, fix the model parameters, use the test set for prediction, and calculate the accuracy of the graph classification.
[0049] A processing device, comprising:
[0050] at least one memory for storing one or more programs;
[0051] At least one processor can execute one or more programs stored in the memory. When the one or more programs are executed by the processor, the processor can implement the above method.
[0052] A readable storage medium stores a computer program, which can implement the above method when the computer program is executed by a processor.
[0053] Compared with the prior art, the present invention has the following advantages:
[0054] By injecting multi-step structural coefficients into the neighborhood aggregation framework, the local structure around the node can be flexibly quantified, and different categories of local structural information can be captured, including information of different circle structures; circle structure is an important substructure in chemical molecules. By using multi-step random walk return probability to enhance node features, the influence of circle structure and local substructure information and global information of nodes can be effectively quantified; subgraph GNNs can be designed to flexibly handle subgraphs of different sizes and shapes, which enables them to adapt to a variety of different graph structures and application scenarios; since the same node in the original graph will appear in different subgraphs, the neighbor subgraph can be used to capture the view of the same node from different subgraph perspectives. We can make full use of this rich contextual information to further enhance the representation of the subgraph, achieve a better representation of chemical molecules, and thus improve the accuracy of chemical molecule classification. BRIEF DESCRIPTION OF THE DRAWINGS
[0055] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings required for use in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other accompanying drawings can be obtained based on these accompanying drawings without paying creative work.
[0056] Figure 1 A flow chart of a chemical molecular graph classification method based on structure enhancement provided in an embodiment of the present invention.
[0057] Figure 2 A schematic diagram of selecting a sub-image from an original image provided in an embodiment of the present invention.
[0058] Figure 3 A schematic diagram of an extended subgraph node provided in an embodiment of the present invention.
[0059] Figure 4 A subgraph node update framework diagram provided by an embodiment of the present invention.
[0060] Figure 5 A schematic diagram of information between aggregated subgraphs provided in an embodiment of the present invention. DETAILED DESCRIPTION
[0061] In order to gain a deeper understanding of the present invention, we will provide a comprehensive and detailed description of the present invention. However, the present invention has multiple implementations and is not limited to the specific examples listed herein. The presentation of these examples is intended to deepen the comprehensive understanding of the disclosure of the present invention.
[0062] Example 1
[0063] like Figure 1 As shown, this embodiment provides a chemical molecular graph classification method based on structure enhancement, which specifically includes the following steps:
[0064] Step 1, obtain chemical molecule data, use atoms as nodes and chemical bonds as edges, construct a chemical molecule graph for each chemical molecule, and obtain a chemical molecule dataset. For example, the MUTAG dataset consists of 188 chemical compounds, which are divided into two categories according to their mutagenic effects on bacteria, containing 125 and 63 compounds respectively. Atoms are modeled as nodes in the graph, and the atom types are carbon (C), nitrogen (N), oxygen (O), fluorine (F), iodine (I), chlorine (Cl), and bromine (Br). The corresponding node labels are 0, 1, 2, 3, 4, 5, and 6. Chemical bonds are modeled as edges in the graph. Chemical bonds can be divided into four types: aromatic bonds, single bonds, double bonds, and triple bonds. The corresponding edge labels are 0, 1, 2, and 3, realizing the modeling of chemical molecules as graphs.
[0065] Step 2: Construct a chemical molecular graph classification model based on structure enhancement, including three parts: structure enhancement, subgraph representation, and graph classification;
[0066] like Figure 2 As shown, first, for each node v in the chemical molecular graph G, the derived subgraph of its k-hop neighbor nodes is used as the root subgraph of node v Perform structural enhancement and subgraph representation on each root subgraph; pool the representations of all root subgraphs to obtain the representation of the chemical molecule graph and perform chemical molecule classification;
[0067] The structural enhancement part includes root subgraph Enhance from two aspects: node features and edge features;
[0068] The subgraph representation part includes the information within the aggregated subgraph and the information between aggregated subgraphs to the root subgraph. The nodes in are represented by average pooling, and the root subgraph is obtained by average pooling the node representations. Representation of;
[0069] The graph classification part includes summing and pooling all root subgraph representations of the chemical molecule graph G to obtain the vector representation of the chemical molecule graph G, using a graph classifier to classify the representation of the chemical molecule graph G, and predicting the properties of the corresponding chemical molecules.
[0070] In the node feature part of the structure enhancement, the root subgraph is Enhance the node features;
[0071] Compute root subgraph The D1-step random walk return probability of node a It is expressed as:
[0072]
[0073] in, Represents the root subgraph The probability that node a in starts to return to a after d steps of random walk;
[0074] Then use a linear layer to move node a to the root subgraph The D1-step random walk return probability on Encoded as node a corresponding to the root subgraph The hidden representation vector Similarly, the D2-step random walk of node a on the original graph G returns the probability code Then the hidden representation vector of node a corresponding to the original chemical molecule graph G is Will With the initial node feature x a Splice to get the root subgraph Enhanced node feature representation of midpoint a Enhance the root subgraph by returning probabilities using multi-step random walks The node characteristics of the chemical molecule can measure the influence of the ring structure.
[0075] Edge feature part in structural enhancement:
[0076] Root subgraph The features of edge ab are enhanced, and the L-step structural coefficient on edge ab is calculated as the edge weight. L-step structural coefficient of middle edge ab The calculation formula is:
[0077]
[0078] in, Represented in the root subgraph The l-th step structural coefficient of the edge ab in, Represents the root subgraph The common l-hop neighbors of endpoints a and b, express Zhongyou The edge set corresponding to the derived subgraph, λ>0.
[0079] In the field of chemistry, the more chemical bonds and atoms an atom shares with its neighboring atoms, the larger its multi-step structure coefficient is, indicating that the neighboring atom is more important.
[0080] Aggregate subgraph information in subgraph representation, such as Figure 3 As shown, first, in each root subgraph of the chemical molecular graph G Add a node to make its vertex set The edge set remains unchanged, and the feature of the newly added node w is If the newly added node is an isolated point, then the root subgraph The adjacency matrix A v for:
[0081]
[0082] A v Normalize row by row For each node The feature vector of the t+1th layer is expressed as:
[0083]
[0084] Among them, γ (t) is a learnable scalar parameter, MLP (t+1) represents the multilayer perceptron at layer t+1, Represents the root subgraph The neighbor set of node a.
[0085] Aggregate subgraph information in subgraph representation. Let the node set with the shortest path distance k from the root node v in the chemical molecule graph G be Node The root graph for A neighbor subgraph of Figure 5 As shown, the root subgraph The information between neighbor subgraphs aggregated by node a is expressed as:
[0086]
[0087] like Figure 4 As shown in Figure 2, by aggregating the intra-subgraph information and inter-subgraph information, we get the representation of node a after the t+1th iteration:
[0088]
[0089] After T iterations, Midpoint Representation Perform average pooling to obtain sub-images Representation:
[0090]
[0091] In the graph classification part, all root subgraph representations of the chemical molecule graph G are summed and pooled to obtain the vector representation of the chemical molecule graph G:
[0092]
[0093] The graph classifier consists of two linear layers, and the vector representation h of the chemical molecule graph G is G Classify and predict the properties of corresponding chemical molecules.
[0094] Step 3: Train and optimize the graph classification model built in step 2, and perform model performance evaluation and comparative analysis. This includes the following steps:
[0095] Step 3.1, divide the chemical molecule dataset into 10 parts equally, take 1 part as the test set each time, and use the remaining 9 parts as the training set;
[0096] Step 3.2, input the chemical molecule graphs in the chemical molecule training set into the chemical molecule graph classification model based on structure enhancement, obtain the classification prediction value of each chemical molecule graph, and use the cross entropy loss function to calculate the difference between the prediction value and the label value. The calculation formula of the cross entropy loss function Loss is as follows:
[0097]
[0098] Among them, m represents the number of chemical molecules, n represents the number of chemical molecule categories, and y ji represents the i-th component of the true label value of the j-th chemical molecule, represents the i-th component of the predicted value of the j-th chemical molecule;
[0099] After obtaining the training loss, use the Adam optimizer to optimize the model parameters;
[0100] Step 3.3, after training, fix the model parameters, use the test set for prediction, and calculate the accuracy of the graph classification.
[0101] Example 2
[0102] To verify the effectiveness of the present invention, MUTAG, PROTEINS, IMDB-M, IMDB-B, NCI1 and COX2 in the graph classification benchmark dataset TUDataset are selected for graph classification tasks, and the information of the dataset is shown in Table 1. For IMDB-B and IMDB-M, the initial node features are encoded as one-hot vectors of node degrees, and the initial node features of MUTAG, PROTEINS, NCI1 and COX2 are encoded as one-hot vectors of node labels.
[0103] Table 1: Graph classification dataset statistics
[0104]
[0105] All experiments in this embodiment are implemented based on Pytorch and executed on NVIDIA GeForce GTX 1080Ti GPU. The experiments on TUDataset dataset adopt the same experimental settings as GIN. The dataset is divided into standard parts, with 90% training set and 10% test set, and sent to the chemical molecular graph classification model based on structure enhancement constructed in step 2 for 10-fold cross validation. In this embodiment, TUDataset uses classification accuracy as the evaluation index, averages the test accuracy of 10-fold cross validation, and selects the epoch with the highest average accuracy. Then, the average accuracy and standard deviation are calculated according to the selected epoch, and the final result is reported. The model is trained using Adam optimizer, with the learning rate set to 0.01, decayed by 0.5 every 50 epochs, the batch size set to 128, and the hidden layer dimension set to 64. The model of the present invention is compared with nine existing advanced model algorithms, namely WL, GraphSAGE, GIN, GraphSNN, PPGN, ID-GNN, ESAN, GIN-AK, and GSN. The performance comparison between the proposed model and the existing model is shown in Table 2 below, which shows the graph classification experimental results obtained on TUDataset, and shows the average and standard deviation of the ten-fold cross-validation accuracy in percentage.
[0106] Table 2 Experimental results of graph classification on TUDataset
[0107]
[0108] The implementation of all or part of the processes in the above-mentioned embodiments can be completed by instructing the relevant hardware through a program, and the program can be stored in a computer-readable storage medium. When the program is executed, it can include the processes of the embodiments of the above-mentioned methods. The storage medium can be a disk, an optical disk, a read-only memory (ROM) or a random access memory (RAM).
[0109] The contents not described in detail in the specification of the present invention belong to the prior art known to the professional and technical personnel in the field. Although the illustrative specific embodiments of the present invention are described above to facilitate the understanding of the present invention by the technical personnel in the field, it should be clear that the present invention is not limited to the scope of the specific embodiments. For the ordinary technical personnel in the field, as long as various changes are within the spirit and scope of the present invention defined and determined by the attached claims, these changes are obvious, and all inventions and creations using the concept of the present invention are protected.
Claims
1. A chemical molecular graph classification method based on structure enhancement, characterized in that: Step 1, obtain chemical molecule data, use atoms as nodes and chemical bonds as edges, construct a chemical molecule graph for each chemical molecule, and obtain a chemical molecule data set; Step 2: Construct a chemical molecular graph classification model based on structure enhancement, which includes three parts: structure enhancement, subgraph representation, and graph classification. Step 3: Train and optimize the chemical molecular graph classification model based on structure enhancement, and conduct model performance evaluation and comparative analysis.
2. A chemical molecular graph classification method based on structure enhancement according to claim 1, characterized in that: The construction of the chemical molecular graph classification model based on structure enhancement in step 2 is specifically as follows: For each node v in the chemical molecular graph G, the derived subgraph of its k-hop neighbor nodes is used as the root subgraph of node v Perform structural enhancement and subgraph representation on each root subgraph; pool the representations of all root subgraphs to obtain the representation of the chemical molecule graph and perform chemical molecule classification; The structural enhancement part includes root subgraph Enhance from two aspects: node features and edge features; The subgraph representation part includes the information within the aggregated subgraph and the information between aggregated subgraphs to the root subgraph. The nodes in are represented by average pooling, and the root subgraph is obtained by average pooling the node representations. ' The graph classification part includes summing and pooling all root subgraph representations of the chemical molecule graph G to obtain the vector representation of the chemical molecule graph G, using a graph classifier to classify the representation of the chemical molecule graph G, and predicting the properties of the corresponding chemical molecules.
3. A chemical molecular graph classification method based on structure enhancement according to claim 2, characterized in that: In step 2, the root subgraph The node features to be enhanced are: Compute root subgraph The D1-step random walk return probability of node a It is expressed as: in, Represents the root subgraph The probability that node a in starts to return to a after d steps of random walk; Then use a linear layer to move node a to the root subgraph The return probability of the D1-step random walk on Encoded as node a corresponds to the root subgraph The hidden representation vector Similarly, the D2-step random walk of node a on the original graph G returns the probability code Then the hidden representation vector of node a corresponding to the original chemical molecule graph G is Will With the initial node feature x a Splice and get the root subgraph Enhanced node feature representation of midpoint a Enhance the root subgraph by returning probabilities using multi-step random walks The node characteristics of the chemical molecule can measure the influence of the ring structure.
4. The method for chemical molecular graph classification based on structure enhancement according to claim 2, characterized in that: In step 2, the root subgraph The edge features are enhanced specifically as follows: Root subgraph The features of edge ab are enhanced, and the L-step structural coefficient on edge ab is calculated as the edge weight. L-step structural coefficient of middle edge ab The calculation formula is: in, Represented in the root subgraph The l-th step structural coefficient of the edge ab in, Represents the root subgraph The common l-hop neighbors of endpoints a and b, express Zhongyou The edge set corresponding to the derived subgraph, λ>
0.
5. The method for chemical molecular graph classification based on structure enhancement according to claim 2, characterized in that: The information in the aggregated subgraph in step 2 is specifically: First, in each root subgraph of the chemical molecular graph G Add a node to make its vertex set The edge set remains unchanged, and the feature of the newly added node w is If the newly added node is an isolated point, then the root subgraph The adjacency matrix A v for: A v Normalize row by row For each node The feature vector of the t+1th layer is expressed as: Among them, γ (t) is a learnable scalar parameter, MLP (t+1) represents the multilayer perceptron at layer t+1, Represents the root subgraph The neighbor set of node a.
6. The method for chemical molecular graph classification based on structure enhancement according to claim 2, characterized in that: The information between the aggregated subgraphs in step 2 is specifically: Let the node set with the shortest path distance k from the root node v in the chemical molecule graph G be Node The root graph for A neighbor subgraph of , then the root subgraph The information between neighbor subgraphs aggregated by node a is expressed as: By aggregating the intra-subgraph information and inter-subgraph information, we get the representation of node a after the t+1th iteration: After T iterations, Midpoint Representation Perform average pooling to obtain sub-images Representation:
7. The method for chemical molecular graph classification based on structure enhancement according to claim 2, characterized in that: In the graph classification part of step 2, all root subgraph representations of the chemical molecule graph G are summed and pooled to obtain a vector representation of the chemical molecule graph G: Using graph classifier to represent the vector h of chemical molecular graph G G Classify and predict the properties of corresponding chemical molecules.
8. The method for chemical molecular graph classification based on structure enhancement according to claim 1, characterized in that: The training and optimization of the chemical molecular graph classification model based on structure enhancement in step 3 specifically includes the following steps: Step 3.1, divide the chemical molecule dataset into 10 parts equally, take 1 part as the test set each time, and use the remaining 9 parts as the training set; Step 3.2, input the chemical molecule graphs in the chemical molecule training set into the chemical molecule graph classification model based on structure enhancement, obtain the classification prediction value of each chemical molecule graph, and use the cross entropy loss function to calculate the difference between the prediction value and the label value. The calculation formula of the cross entropy loss function Loss is as follows: Among them, m represents the number of chemical molecules, n represents the number of chemical molecule categories, and y ji represents the i-th component of the true label value of the j-th chemical molecule, represents the i-th component of the predicted value of the j-th chemical molecule; After obtaining the training loss, use the Adam optimizer to optimize the model parameters; Step 3.3, after training, fix the model parameters, use the test set for prediction, and calculate the accuracy of the image classification.
9. A processing device, characterized in that: include: at least one memory for storing one or more programs; At least one processor is capable of executing one or more programs stored in the memory, and when the one or more programs are executed by the processor, the processor is capable of implementing the method according to any one of claims 1 to 8.
10. A readable storage medium, characterized in that: A computer program is stored, and when the computer program is executed by a processor, the method described in any one of claims 1 to 8 can be implemented.