Graph neural network-based interpretable molecular property prediction method

Abstract

The invention belongs to the field of graph neural network interpretation, and particularly relates to an interpretable molecular property prediction method based on a graph neural network, which comprises the following steps: acquiring graph data G, and iterating the graph data to generate K key sub-graphs; obtaining a graph neural network graph classification model, and obtaining node embedding according to the graph neural network graph classification model; according to node embedding, obtaining an edge mask matrix of the graph; inputting the K key sub-graphs and the mask matrix into a perceptron to score the importance of each key sub-graph to obtain a score matrix W and an affiliation relation matrix L between each edge in the graph G and a key sub-graph mode; inputting the score matrix W and the affiliation relation matrix into a mask updater to obtain a final mask matrix M; selecting topK edges with the highest mask values according to the mask matrix M to generate interpretation sub-graphs as interpretation of the graph neural network graph classification model; according to the method, an important sub-graph mode in the data is extracted by adopting a frequent sub-graph mining algorithm, so that the interpretable accuracy is improved.

Description

Technical Field

[0001] The present invention belongs to the field of graph generation and graph neural network interpretation, and specifically relates to an interpretable molecular property prediction method based on graph neural network. Background Art

[0002] With the widespread use of the internet and the rapid development of data technologies, people are able to collect vast amounts of data. Unlike traditional neural networks, which process vector or matrix data, graph neural networks (GNNs) can directly process graph-structured data, modeling nodes and edges. They can learn feature representations of nodes and edges from graph-structured data and apply these representations to various graph applications. Due to the widespread availability of graph-structured data, GNNs have achieved excellent performance in many fields, such as social networks, recommendation systems, and protein structure prediction.

[0003] While graph neural networks (GNNs) outperform some traditional machine learning algorithms on graph-structured data, understanding how they make decisions is often difficult, leading to their being considered a black box. This limits their use in risk-sensitive and security-critical scenarios, such as finance and bioinformatics. Therefore, exploring how GNNs make decisions is of great research interest.

[0004] However, in real-world graph data, structure plays a significant role in determining graph properties. For example, in chemical molecules, molecules with benzene rings possess a certain degree of stability, molecules with amino groups exhibit certain alkaline properties, and molecules with aldehyde groups exhibit a certain degree of water solubility. Because post-hoc interpretation methods based on masks only consider the influence of edges and nodes, the resulting interpretation subgraphs are often disorganized and lack semantic information. Summary of the Invention

[0005] In order to solve the problems existing in the above-mentioned prior art, the present invention proposes an interpretable molecular property prediction method based on graph neural network, which includes: obtaining a chemical molecular structure graph G, iterating the chemical molecular structure graph using the MiCaM algorithm, and generating K key subgraph patterns; obtaining a trained graph neural network model for classifying the molecular structure graph, and obtaining the embedding of each node in the molecular structure graph according to the graph neural network model; obtaining the edge mask matrix of the graph according to the node embedding; inputting the K key subgraphs and the edge mask matrix into the perceptron to score the importance of each key subgraph, and obtaining a score matrix W and an attribution relationship matrix L of each edge and key subgraph pattern in the graph G; inputting the score matrix W and the attribution relationship matrix into the mask updater to obtain the final mask matrix According to the mask matrix The topK edges with the highest mask values ​​are selected to generate an explanatory subgraph, and the explanatory subgraph is used as the explanatory result of the graph neural network graph classification model; the chemical properties of the chemical molecular structure graph are predicted based on the explanatory results of the graph neural network graph classification model.

[0006] Beneficial effects of the present invention:

[0007] The present invention simultaneously considers the impact of edges and subgraph patterns on the prediction of graph neural networks from a multi-granularity perspective; considering that subgraph patterns may have a significant impact on the prediction of the model, in order to avoid the limitations of manually selecting specific subgraph patterns, a frequent subgraph mining algorithm is used to extract important subgraph patterns in the data. The present invention proposes a multi-granularity structure perception unit, which learns the importance score of the extracted important subgraph patterns to the model decision. This score reflects the degree of positive or negative impact of the subgraph pattern on the decision. Based on the learned score, the present invention further maps the impact of the subgraph pattern on the model decision to the mask matrix of the interpreter, thereby generating an explanation subgraph, and using it as the explanation of the model. Through this method, the knowledge requirements for manually defined subgraph patterns are avoided, and the noise problem caused by automatically learning multi-granularity subgraph patterns is avoided, thereby optimizing and improving the performance and robustness of graph neural networks when processing complex structured data. BRIEF DESCRIPTION OF THE DRAWINGS

[0008] Figure 1 This is a flow chart of the graph neural network interpretation method for subgraph pattern masking of the present invention;

[0009] Figure 2 It is the overall framework structure diagram of the present invention. DETAILED DESCRIPTION

[0010] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0011] A method for predicting interpretable molecular properties based on graph neural networks, comprising: obtaining a chemical molecular structure graph G, iterating the chemical molecular structure graph using the MiCaM algorithm to generate K key subgraph patterns; obtaining a trained graph neural network model for classifying the molecular structure graph, and obtaining embeddings of each node in the molecular structure graph based on the graph neural network model; obtaining an edge mask matrix of the graph based on the node embeddings; inputting the K key subgraphs and the edge mask matrix into a perceptron to score the importance of each key subgraph, obtaining a score matrix W and an attribution relationship matrix L for each edge and key subgraph pattern in the graph G; inputting the score matrix W and the attribution relationship matrix into a mask updater to obtain a final mask matrix According to the mask matrix The topK edges with the highest mask values ​​are selected to generate an explanatory subgraph, and the explanatory subgraph is used as the explanatory result of the graph neural network graph classification model; the chemical properties of the chemical molecular structure graph are predicted based on the explanatory results of the graph neural network graph classification model.

[0012] The present invention proposes an interpretable molecular property prediction method based on graph neural network, such as Figure 1 , specifically including the following: In the graph dataset In the above example, the MiCaM algorithm is used to iteratively generate K key subgraph patterns. in represents the set of edges in the original graph G contained in the i-th common subgraph.

[0013] Train a graph neural network graph classification model as the model to be explained, and from the model to be explained we can get the embedding Z of the node i Taking PGEXplainer as an example, the edge mask matrix M is obtained by embedding nodes and using the reparameterization technique. The structure of the graph neural network graph classification model is as follows Figure 2 shown.

[0014] The obtained mask matrix M and the first K sub-image patterns generated Train our structured perceptron to score the importance of each subgraph pattern, and obtain the score matrix W and each edge and key subgraph pattern in the graph G. The ownership relationship matrix L.

[0015] The obtained subgraph pattern score matrix W and relationship matrix L are input into the mask updater to finally obtain the mask matrix According to the mask matrix The topK edges with the highest mask values ​​are selected to generate an explanation subgraph as the explanation of the model to be explained.

[0016] In this example, the design of the MiCaM algorithm: In, use Represents each graph data. For each graph G, use Track the merge state, which represents the graph fragments and their connections. In each merge graph G M In each node Represents a segment of graph data. Edge set Representing the connection relationship between two fragments, each merged graph in the original graph is initialized by treating each node as a single fragment and inheriting the connection of G, that is, Define an operation "⊕" to combine two fragments and Merge together to create a new snippet:

[0017]

[0018] Newly acquired Contains from All nodes and edges of , and all edges between them. The k-th merge operation is defined as: if Then and Merge them together. Update them to Repeat K iterations to obtain a sequence of merge operations in represents the i-th key subgraph, represents the set of edges in the original graph G contained in the i-th common subgraph.

[0019] Design of Graph Neural Network GCN:

[0020]

[0021] Where σ represents the activation function, A represents the adjacency matrix, D represents the degree matrix, and I represents the identity matrix. express The degree matrix of H. (l) Represents the node features of the lth layer in the graph neural network. When l = 0, H (l) =X is the original feature of the node, W (l) Represents the parameters that need to be learned in the lth layer of the graph neural network.

[0022] Design of Structure Perceptron:

[0023]

[0024] W uv =σ(MLP(concat(Z u |Z v )))∈R 1×k

[0025] L u,v ={l1,l2,…,l k}∈R n×n×k

[0026] where Q = {q1,q2,…,q n}∈R 1×k , represents the average score of each subgraph pattern. ε is the set of edges in the original graph G contained in all common subgraphs, is the set of edges in the original graph G contained in the i-th common subgraph, M (u,v) is the original mask value of the edge where node u and node v are located, W uv is the importance of the edge between node u and node v in the k most common subgraph patterns, σ is the activation function, Z u is the node embedding of node u, Z v is the node embedding of node v, R 1×k is a 1×k-dimensional vector, L u,v Is the edge between node u and node v belonging to the k most common subgraphs, l k When it is 1, it means that the edge between node u and node v belongs to the kth most common subgraph, R n×n×k is an n×n×k dimensional vector. uv represents the score of edge (u,v) on the k subgraph patterns it belongs to. L represents the relationship matrix between edges and subgraph patterns.

[0027] Design of mask updater:

[0028]

[0029] in, is the updated mask value of the edge where node u and node v are located, M uv is the original mask value of the edge where node u and node v are located, Q is the score of all subgraph patterns, ⊙ is the dot product, L uv Whether the edge between node u and node v belongs to the k most common subgraphs, is the transpose of the importance matrix of the edge between node u and node v in the k most common subgraph patterns, X s To explain the subgraph G S Node features, A S To explain the subgraph G S The adjacency matrix of is the updated mask matrix, A is the adjacency matrix of the original image, To explain the subgraph G S The prediction of G S represents the explanation subgraph generated by the mask matrix, MI represents the mutual information, represents sparsity, and Y represents the loss function.

[0030] The above embodiments further illustrate the purpose, technical solutions and advantages of the present invention in detail. It should be understood that the above embodiments are only preferred implementation plans of the present invention and are not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made to the present invention within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. An interpretable molecular property prediction method based on graph neural network, characterized in that: include: Obtain a chemical molecular structure graph G, iterate the chemical molecular structure graph using the MiCaM algorithm to generate K key subgraph patterns; obtain a trained graph neural network model for classifying the molecular structure graph, and obtain the embedding of each node in the molecular structure graph based on the graph neural network model; obtain the edge mask matrix of the graph based on the node embedding; input the K key subgraphs and the edge mask matrix into the perceptron to score the importance of each key subgraph, and obtain the score matrix W and the attribution relationship matrix L of each edge and key subgraph pattern in the graph G; Input the score matrix W and the attribution matrix into the mask updater to obtain the final mask matrix According to the mask matrix The topK edges with the highest mask values ​​are selected to generate an explanatory subgraph, and the explanatory subgraph is used as the explanatory result of the graph neural network graph classification model; the chemical properties of the chemical molecular structure graph are predicted based on the explanatory results of the graph neural network graph classification model.

2. The method for predicting molecular properties based on graph neural network according to claim 1, characterized in that: The MiCaM algorithm is used to iterate the graph data, including: use Track the merge status and get the fragments; merge the two fragments and Merge to get new fragments; the merged graph at the kth iteration In the example, each edge represents a pair of segments. They are adjacent in the original graph; traverse all merged graphs All edges in statistics frequency, and the most frequent Expressed as All Merge to get the final sequence, and the final training is used as the key subgraph pattern.

3. The method for predicting molecular properties based on graph neural network according to claim 1, characterized in that: The graph neural network graph classification model is: Among them, σ represents the activation function, I represents the identity matrix, express The degree matrix of , X is the original feature of the node, and W is the parameter to be learned in the graph neural network.

4. The method for predicting molecular properties based on graph neural network according to claim 1, characterized in that: The node embedding is: Z = GCN(A, X, W) Among them, Z represents the embedding of the node, A represents the adjacency matrix, X represents the original features of the node, and W represents the parameters that need to be learned in the graph neural network.

5. The method for predicting molecular properties based on graph neural network according to claim 1, characterized in that: The edge mask matrix of the graph obtained based on the node embedding includes: W ij =ConCat(Z i ,Z j ) Among them, Z i , Z j represents the embedding of node i and node j, W ij Represents the mask value of the edge connecting node i and node j.

6. The method for predicting molecular properties based on graph neural network according to claim 1, characterized in that: The perceptron is: IN uv =σ(MLP(concat(Z u |From v )))∈R 1×k L u，v ={l1, l2,..., l k }∈R n×n×k Among them, q i is the score of the ith subgraph pattern, ε is the set of edges in the original graph G contained in all common subgraphs, is the set of edges in the original graph G contained in the i-th common subgraph, M (u，v) is the original mask value of the edge where node u and node v are located, W uv is the importance of the edge between node u and node v in the k most common subgraph patterns, σ is the activation function, Z u is the node embedding of node u, Z v is the node embedding of node v, R 1×k is a 1×k-dimensional vector, L u，v Whether the edge between node u and node v belongs to the k most common subgraphs, l k When it is 1, it means that the edge between node u and node v belongs to the kth most common subgraph, R n×n×k is an n×n×k dimensional vector.

7. The method for predicting molecular properties based on graph neural network according to claim 1, characterized in that: The mask updater is: in, is the updated mask value of the edge where node u and node v are located, M uv is the original mask value of the edge where node u and node v are located, Q is the score of all subgraph patterns, ⊙ is the dot product, L uv Whether the edge where node u and node v are located belongs to the k most common subgraphs, is the transpose of the importance matrix of the edge between node u and node v in the k most common subgraph patterns, X s To explain the subgraph G s The node characteristics of A S To explain the subgraph G S The adjacency matrix of is the updated mask matrix, A is the adjacency matrix of the original image, To explain the subgraph G s The prediction of G S represents the explanation subgraph generated by the mask matrix, MI represents mutual information, Indicates sparsity.

Images