A parameter fine-tuning method for interpretability of drug molecule graph structure

Abstract

The application belongs to the field of graph neural network explanation, and relates to a parameter fine-tuning method for explainability of drug molecule graph structure, comprising the following steps: obtaining drug molecule graph structure data and decomposing the drug molecule graph structure data into multiple subgraphs, embedding the multiple subgraphs into a space; using a greedy strategy to process the subgraphs in the embedding space to generate a candidate semantic subgraph set; inputting the candidate semantic subgraph set into a graph classification model to obtain a final semantic subgraph; generating a k-order induced graph for each node in the drug molecule graph structure data; setting a similarity threshold, calculating the similarity between the k-order induced graph and the final semantic subgraph, and regarding the structure composed of all nodes with a similarity greater than the similarity threshold as a key structure; adjusting the molecular weight parameters in the explanation process of the drug molecule graph structure according to the key structure by using a trained multilayer perceptron to obtain an optimal explainability result; and the application can find an explanation result that is easier to understand and assist personnel in the relevant field to make decisions.

Description

TECHNICAL FIELD

[0001] The application belongs to the field of graph neural network interpretation, and specifically relates to a parameter fine-tuning method for the interpretability of drug molecule graph structure. BACKGROUND

[0002] Graph neural networks have been widely applied in social network analysis, recommendation systems, bioinformatics, and other fields due to their excellent performance in processing graph-structured data. These applications take advantage of the characteristics of graphs to effectively capture the complex relationships between nodes. However, the "black box" nature of graph neural network models limits their practical applications in high-trust scenarios, as the decision-making process lacks transparency. The core of interpretability lies in identifying subgraphs in the input graph that have an impact on decision-making. This identification not only includes nodes and edges that directly affect decisions, but also considers their context and associated relationships. To achieve this goal, researchers have proposed various methods, such as attention mechanism-based interpretability frameworks, subgraph selection algorithms, and counterfactual-based explanation strategies. These methods can help users gain a deeper understanding of the model's behavior, providing effective explanations for specific tasks.

[0003] Accurate graph neural network model interpretation not only helps people understand the model's predictions, but also reveals potential flaws and improvement directions. For example, by analyzing the interpretation results of the model, researchers can find that certain specific features may lead to incorrect predictions, which provides a basis for subsequent model optimization. At the same time, good interpretability can also promote user trust in the model, promoting the widespread application of graph neural network models in critical areas.

[0004] Drug molecule graph structure and the chemical composition and spatial arrangement of drugs are crucial for the biological activity and pharmacokinetic properties of drugs; how to describe the interpretability of each chemical bond in the drug molecule graph structure through the graph neural network model to improve users' effective understanding of the properties of drug molecules is a technical problem that needs to be solved at present. SUMMARY

[0005] To solve the problems existing in the prior art, the present application provides a parameter fine-tuning method for interpretability of drug molecule graph structure, which comprises the following steps: obtaining drug molecule graph structure data; decomposing the drug molecule graph structure data into multiple subgraphs and embedding the multiple subgraphs in a space; processing the subgraphs in the embedding space by using a greedy strategy, and generating a candidate semantic subgraph set through a score function iteration; inputting the candidate semantic subgraph set into a trained graph classification model, and selecting the top R subgraphs with the maximum target class prediction value probability as the final semantic subgraphs; generating a k-order induced graph for each node in the drug molecule graph structure data; setting a similarity threshold, calculating the similarity between the k-order induced graph and the final semantic subgraph, and regarding the structure composed of all nodes with a similarity greater than the similarity threshold as a key structure; training a multilayer perceptron using the key structure; and adjusting the molecular weight parameters in the interpretation process of the drug molecule graph structure using the trained multilayer perceptron to obtain the optimal interpretability result.

[0006] The present application has the following advantages:

[0007] The present application calculates the similarity between the semantic subgraph and the induced graph of the original graph node, locates the key structure in the molecule graph, uses the multi-hot encoding of the located key structure as the target to train the multilayer perceptron, and adjusts the interpretation behavior of the graph neural network interpretation model using the multilayer perceptron, thereby improving the perception effect; the present application can be applied to various graph neural network interpretation models to enhance the interpretability of the interpretation; the present application can find a more understandable interpretation result to assist personnel in the relevant field to make decisions. BRIEF DESCRIPTION OF DRAWINGS

[0008] Figure 1 The present application has the following advantages:

[0009] Figure 2 The present application has the following advantages:

[0010] Figure 3 The present application has the following advantages: DETAILED DESCRIPTION

[0011] The technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, not all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor are within the scope of protection of the present application.

[0012] An interpretable parameter fine-tuning method for drug molecule graph structure, the method comprising: obtaining drug molecule graph structure data; decomposing the drug molecule graph structure data into a plurality of subgraphs and embedding the plurality of subgraphs in a space; processing the subgraphs in the embedding space using a greedy strategy, generating a candidate semantic subgraph set through a score function iteration; inputting the candidate semantic subgraph set into a trained graph classification model, and selecting the top R subgraphs with the largest probability of predicting the target class as the final semantic subgraphs; generating a k-order induced graph for each node in the drug molecule graph structure data; setting a similarity threshold, calculating the similarity of the k-order induced graph and the final semantic subgraph, and taking the structure composed of all nodes with a similarity greater than the similarity threshold as the key structure; training a multilayer perceptron using the key structure; and adjusting the molecular weight parameters in the interpretation process of the drug molecule graph structure using the trained multilayer perceptron to obtain the optimal interpretability result.

[0013] An efficient parameter fine-tuning method for enhancing the interpretability of graph neural networks, as shown in Figure 1 , the method comprising:

[0014] S1: obtaining drug molecule graph structure data, the topological structure of the drug molecule graph structure data being represented as G=(V,E,X,A), wherein G represents the drug molecule graph structure; V represents a set of atomic nodes in G; E represents a set of edges in G, i.e., a set of chemical bonds between atoms in the molecule; X represents a feature matrix of the atoms in G; and A represents an adjacency matrix of G.

[0015] S2: decomposing the drug molecule into a plurality of subgraphs and embedding them in an ordered embedding space, generating a candidate semantic subgraph set through a score function iteration by a greedy strategy, inputting the semantic subgraph set into a graph classification model, and selecting the top R subgraphs with the largest probability of predicting the target class as the final explanation semantic subgraph.

[0016] G pattern ={g1,…,g R},f(g)=c,g∈G cand .

[0017] wherein topR represents a semantic subgraph predicted by a GNN as a target class by selecting the probability of the top R, c is the target class, and f represents a pre-trained GNN.

[0018] Generating the candidate semantic subgraph set comprises: sorting the subgraph set in the embedding space; setting the subgraph size, wherein the subgraph size is 6-8; selecting nodes from the graph sorted as C using a greedy strategy and iterating to generate subgraphs; repeating the iteration for 10 rounds to obtain the candidate semantic subgraph set.

[0019] The classification and sorting of the subgraph set of the embedding space includes calculating the score of the subgraph of the embedding space, and classifying and sorting according to the score. The calculation formula is:

[0020]

[0021] Wherein, G C represents a graph classified as C, Z c represents the corresponding embedding.

[0022] The framework structure of the graph classification model is shown in Figure 2 The expression of the node classification graph neural network model GCN is:

[0023]

[0024] Wherein, σ represents an activation function, I represents a unit matrix, represents a node degree matrix, W represents a parameter to be learned in the graph neural network; represents an adjusted new adjacency matrix, A represents an adjacency matrix, and X represents a feature matrix.

[0025] S3: Generate a k-order induced graph for each node in the molecular graph, calculate the similarity between the semantic subgraph and the k-order induced graph, and select the structure composed of nodes with higher similarity as the key structure. The value of k is generally the diameter of the semantic subgraph.

[0026] The similarity between the k-order induced graph and the final semantic subgraph is calculated as:

[0027]

[0028] Wherein, D is the embedding dimension, Δz d is the difference in d dimensions, represents the semantic subgraph embedding, represents the embedding of the induced graph of node u, and ∈ is the similarity threshold.

[0029] By minimizing the distance metric, a set of k-order induced graphs can be obtained to locate similar nodes.

[0030] S4: Train a multilayer perceptron with the goal of multi-hot encoding of the matched molecular graph, and optimize the multilayer perceptron by fitting the multi-hot encoding of the original graph and the matched molecular graph. The structure of the multilayer perceptron is shown in Figure 3 , that is, the edge embedding formed by connecting the corresponding embeddings of the nodes using the multilayer perceptron as the input, and the training formula is:

[0031] Z i,j =W(h i ||h j )+b

[0032] w ij =softmax(Z i,j )

[0033] where w ij represents the weight of the edge, h i and h j represent the embedding of the corresponding node, W is the connection weight, b is the bias coefficient, and Z i,j is the output of the hidden layer.

[0034] S5: Adjust the weights of the molecules in the interpretation process by using the trained multi-layer perceptron graph neural network interpretation model, and adjust by using the multi-layer perceptron:

[0035]

[0036] where M ij represents the edge mask between node i and node j, and the trained multi-layer perceptron can flexibly adjust the edge weight, so that the matched edge weight is increased and the unmatched edge weight is reduced.

[0037] In this embodiment, the open source medical field data set Mutagenicity is used, which contains 4337 compound samples, each sample has a binary label indicating whether it is a mutagenic compound. The Mutagenicity data set is divided into training set, validation set and test set according to 7:2:1, and the data set is trained by GNN. The method of the application generates a batch of semantic subgraphs on the data set classified by GNN as mutagenic compound category, and applies the semantic subgraphs to the drug molecules to obtain new drug molecules. The new drug molecules are input into the trained GNN for verification, and the probability value of being classified as a mutagenic class is very high. The molecular structure of the mutagenic compound category in the Mutagenicity data set is greatly understood, and a reference is provided for designing new mutagenic compound category molecular compounds.

[0038] In the present embodiment, the present application adopts the disclosed biochemical experimental dataset BBBP for drug development research, which contains the amino acid sequences and secondary structure information of 2039 protein molecules. Each sample has a binary label indicating whether the protein can penetrate the blood-brain barrier. The BBBP dataset is divided into training set, validation set and test set according to 7:2:1, and the dataset is trained by GNN. The method of the present application generates a batch of new molecules on the GNN classified as penetrating the blood-brain barrier dataset, which can help scientists understand what typical structural features the molecules penetrating the blood-brain barrier have, and at the same time, the generated molecules are connected to a certain molecule, which can make the connected molecule be classified as penetrating the blood-brain barrier by GNN, helping to design new molecules with the characteristic of penetrating the blood-brain barrier.

[0039] The above examples further illustrate the objects, technical solutions and advantages of the present application. It should be understood that the above examples are only preferred embodiments of the present application and are not intended to limit the present application. Any modifications, equivalent replacements, improvements, etc. made to the present application within the spirit and principles of the present application shall be included in the protection scope of the present application.

Claims

1. A parameter fine-tuning method for the interpretability of drug molecule graph structures, characterized in that, The method comprises the following steps: acquiring drug molecule graph structure data; decomposing the drug molecule graph structure data into multiple subgraphs and embedding the multiple subgraphs in a space; adopting a greedy strategy to process the subgraphs in the embedding space, generating a candidate semantic subgraph set through a score function iteration, inputting the candidate semantic subgraph set into a trained graph classification model, and selecting the top R subgraphs with the largest probability of predicting the target class as the final semantic subgraphs; generating a k-order induced graph for each node in the drug molecule graph structure data; where z q represents the embedding of graph q, z u represents the embedding of graph u, P represents a set of graph embeddings of positive examples, N represents a set of graph embeddings of negative examples, a represents a hyperparameter, and E(z q ,z u ) is an embedding space; setting a similarity threshold, calculating the similarity between the k-order induced graph and the final semantic subgraph, and regarding the structure composed of all nodes with a similarity greater than the similarity threshold as a key structure; 2. The method of claim 1, wherein, training a multilayer perceptron using the key structure; 3. The method of claim 1, wherein, adjusting the molecular weight parameters in the interpretation process of the drug molecule graph structure using the trained multilayer perceptron to obtain the optimal interpretability result; 4. The method of claim 1, wherein, The target function expression of embedding the multiple subgraphs in the space is: where D is the embedding dimension, Δz d is the difference in d dimensions, denotes the semantic subgraph embedding, denotes the embedding of the induced graph of node u, and ∈ is the similarity threshold.

5. The method of claim 1, wherein, Generating the candidate semantic subgraph set comprises: sorting the subgraph set in the embedding space; setting the subgraph size, randomly selecting nodes from the graphs classified as C using the greedy strategy, and iterating to generate subgraphs; repeating the iteration for 10 rounds to obtain the candidate semantic subgraph set. Z i,j = W(h i || h j ) + b w ij = softmax(Z i,j ) where w ij represents the weight of the edge, h i and h j represent the embedding of the corresponding nodes, W is the connection weight, b is the bias coefficient, and Z i,j is the output of the hidden layer.

6. The method of claim 1, wherein, The topological structure of the drug molecule graph structure data is represented as G=(V,E,X,A); wherein G represents the drug molecule graph structure; V represents the set of atomic nodes in G; E represents the edge set of G, i.e., the set of chemical bonds between atoms in the molecule; X represents the feature matrix of the atoms in G; and A represents the adjacency matrix of G. The graph classification model is a node classification graph neural network model GCN. The similarity between the k-order induced graph and the final semantic subgraph is calculated as: Training the multilayer perceptron comprises: using the edge embedding formed by connecting the corresponding embeddings of the nodes using the multilayer perceptron as input, and the training formula is: Adjusting the molecular weight parameters using the trained multilayer perceptron comprises: where M ij represents the edge mask between node i and node j, w ij is the weight of each edge.

Images