A drug relocalization method based on multi-view stacked skip-level graph convolutional network

Abstract

The present invention belongs to the field of biomedical technology, specifically a drug relocalization method based on a multi-view stacked skip-level graph convolutional network, comprising the following specific steps: obtaining drug-disease association data, constructing an adjacency matrix A of drugs and diseases, an association graph G, and a similarity matrix S of drugs and diseases respectively. r 、S d The method uses a multi-layer graph convolutional encoder block to encode the drug-disease similarity matrix into a high-order feature representation. A two-level subgraph self-attention network is constructed, combining node-level and subgraph-level features for structural fusion with the graph topology. When using high-order node information, the present invention can extract contextual information about the high-order nodes, alleviating the inconsistency between local and global information. The present invention can more comprehensively utilize deep information, resolving the problem of the network being unable to distinguish between drug-disease nodes after multiple layers, which results in a decline in the model's predictive power.

Description

Technical Field

[0001] The present invention relates to the field of biomedical technology, and in particular to a drug relocalization method based on a multi-view stacked skip-level graph convolutional network. Background Art

[0002] Despite increasing investment in biomedicine over the past few decades, the number of new treatments and drugs approved by the U.S. Food and Drug Administration (FDA) has not kept pace. This situation has prompted researchers and pharmaceutical companies to seek more economical and rapid drug development pathways. Drug repositioning strategies have attracted attention for their potential to shorten drug development cycles and reduce costs.

[0003] Traditional drug repositioning methods rely primarily on experimental studies and clinical observations, which are time-consuming and costly. With the development of bioinformatics and computational technology, computational drug repositioning methods have emerged, including those based on text mining, network analysis, and machine learning.

[0004] Graph neural networks (GNNs), a type of deep learning algorithm, can efficiently process graph-structured data and capture the complex relationships and topological structures between nodes in drug-disease networks. They show potential in predicting potential drug-disease associations, but existing methods have limitations, as follows:

[0005] 1. Failure to effectively utilize the contextual information of long-distance nodes between drugs and diseases in the graph, resulting in a lack of consistency between global and local information in feature learning;

[0006] 2. Too much focus on information aggregation of connected nodes, ignoring global graph information and high-order dependencies. To improve prediction accuracy, the number of network layers is increased, but too many layers will lead to homogenization of node features of drugs and diseases, affecting prediction accuracy.

[0007] 3. The sparsity and imbalance of drug-disease connections further weaken the predictive ability.

[0008] Therefore, a drug relocalization method based on a multi-view stacked skip-level graph convolutional network was invented; by comprehensively understanding the complex relationship between drug-disease nodes on the graph from multiple perspectives, the accuracy of predicting the potential association between drugs and diseases is improved, which has important practical application value. Summary of the Invention

[0009] To solve the above technical problems, according to one aspect of the present invention, the present invention provides the following technical solutions:

[0010] A drug relocalization method based on a multi-view stacked skip-level graph convolutional network includes the following specific steps:

[0011] S1: Obtain drug-disease association data, and construct the drug-disease adjacency matrix A, association graph G, and drug-disease similarity matrix S. r 、S d , and uses a multi-layer graph convolutional encoder block to encode the similarity matrix of drugs and diseases into high-order feature representations;

[0012] S2: Construct a two-level subgraph self-attention network, combining the features of the node level and the subgraph level, and perform structural fusion with the graph topology;

[0013] S3: Construct a multi-view stacking skip-level sensing network to perform multi-view stacking learning on the association graph G, skip-level sensing deep feature information in multi-layer graph convolution, and learn the horizontal features of nodes;

[0014] S4: Fusion of a two-layer subgraph self-attention network and a multi-view stacked skip-level induction network, and design of a negative distance balancing method to enhance negative examples to optimize network performance;

[0015] S5: Use the optimized network to predict potential drug-disease interactions.

[0016] As a preferred solution of the drug relocalization method based on multi-view stacked skip-level graph convolutional network described in the present invention, the specific steps of S1 are as follows:

[0017] S11: Obtain drug-disease association data, and construct the drug-disease adjacency matrix A, association graph G, and drug-disease similarity matrix S. r 、S d ;

[0018] S12: S r 、S d Input to the multi-layer graph convolutional encoder block F encode , the specific form is as follows:

[0019]

[0020] Among them, σ is the activation function, v is the node, is the neighbor node of v, is the feature of node v, is the feature of neighbor node u, W l is the learnable weight matrix, are the degrees of node v and its neighbor node u respectively;

[0021] S13: Obtain drug high-order features and disease high-order features respectively, concatenate the two features together, and output drug-disease high-order features X.

[0022] As a preferred solution of the drug relocalization method based on multi-view stacked skip-level graph convolutional network described in the present invention, the specific steps of S2 are as follows:

[0023] S21: Use the METIS software package to split the drug-disease association graph G into multiple equally divided subgraph sets P G , and use the high-order features X obtained by S1 to extract two-level features, namely the sub-graph level and the node level X sub 、X node , and use graph convolutional networks to extract structural features X strut ;

[0024] S22: Fuse the node-level features, sub-map-level features, and structure-level features at the structural level, and output the fused features X bi .

[0025] As a preferred solution of the drug relocalization method based on a multi-view stacked skip-level graph convolutional network described in the present invention, the specific steps of S21 are as follows:

[0026] S211: Get sub-layer level feature X sub ;

[0027] S212: Get node level feature X node , the specific steps are as follows:

[0028]

[0029] Among them, X is the high-order feature of drug-disease, d X is the drug-disease node vector dimension;

[0030] S213: Get structure-level feature X strut .

[0031] As a preferred solution of the drug relocalization method based on a multi-view stacked skip-level graph convolutional network described in the present invention, the specific steps of S211 are as follows:

[0032] S2111: Use the METIS tool to split the drug-disease association graph G into multiple equally divided subgraph sets P G ;

[0033] S2112: Pool the node features in each subgraph so that the nodes in each subgraph share a feature. The specific form is as follows:

[0034]

[0035] in, is the feature of the nodes in each subgraph, X sis the subgraph feature;

[0036] S2113: Using subgraph self-attention to extract subgraph level features:

[0037]

[0038] Among them, X s is the subgraph feature, is the subgraph vector dimension.

[0039] As a preferred solution of the drug relocalization method based on a multi-view stacked skip-level graph convolutional network described in the present invention, the specific steps of S213 are as follows:

[0040] S2131: Copy the subgraph hierarchical features and assign them to all nodes in the same subgraph:

[0041] X' sub =repeat(X sub )

[0042] S2132: Construct a graph convolutional network combination block GCN_F to extract structural features:

[0043]

[0044] X strut =GCN_F(X' sub +X node )

[0045] in, is the feature of the disease node associated with the drug, is the characteristic of the drug node, W is the learnable weight matrix, are the degrees of disease node d and drug node r, respectively.

[0046] As a preferred solution of the drug relocalization method based on a multi-view stacked skip-level graph convolutional network described in the present invention, the specific steps of S22 are as follows:

[0047] The node-level features X obtained in the above steps are node , sub-layer level feature X sub and structural features X strut Perform structure-level fusion and output the fused feature X bi :

[0048] X bi =fusion(X' sub ,X node ,X strut )

[0049] Among them, fusion is structural-level fusion, and the specific operation is to add the corresponding elements of each feature.

[0050] As a preferred solution of the drug repositioning method based on the multi-view stacked skip graph convolutional network described in the present invention, wherein: the specific steps of S3 are as follows:

[0051] S31: Generate multiple channels by combining a random mask and a graph convolutional network block;

[0052] S32: Define a domain skip induction mechanism to extract deep node information;

[0053] S33: Stack and splice all views, and output the feature X mv ;

[0054] The specific steps of S31 are as follows: <00002​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​

[0064] The specific steps of S32 are as follows:

[0065] S321: Using the characteristics of self-attention mechanism and graph convolutional network, design a domain skipping induction mechanism, including Layer i Perform pooling, calculate the hierarchical self-attention scores, and select the K highest scores as the aggregated levels:

[0066] P i =Pooling(Layer i )

[0067]

[0068] TopLayer=select(score topK (Layer i ))

[0069] Among them, P i To pool the node features of each layer, score is used to calculate the attention score, TopLayer is the feature set of the K layers before the score, d P is the level vector dimension;

[0070] S322: Fuse the selected skip-level graph convolutional network layers with the drug-disease high-level feature X described in S1:

[0071]

[0072] Among them, X is the high-order feature of drug disease, Σ k∈K TopLayer is the selected deep features and;

[0073] The specific steps of S33 are as follows:

[0074] Using Channel i As a foundation, we perform feature fusion by stacking views and S2 cross-domain features. To ensure the differences between views, we aggregate different features separately. The operation is as follows:

[0075] VIEW i =GCN_H(Channel i +VIEW i-1 +X bi )

[0076] VIEW i =GCN_H(Channel i +VIEW i-1 +X node +X sub )

[0077] Among them, X bi 、X node 、X sub Features extracted from the two-level subgraph self-attention network in S2;

[0078] Finally, multiple views are spliced ​​together as the output feature X mv :

[0079]

[0080] Among them, VIEW i For different views, A splicing symbol.

[0081] As a preferred solution of the drug relocalization method based on multi-view stacked skip-level graph convolutional network described in the present invention, the specific steps of S4 are as follows:

[0082] S41: Fusion of a two-layer subgraph self-attention network and a multi-view stacked skip-level induction network;

[0083] S42: Design a negative example distance balancing method for unbalanced negative examples;

[0084] The specific steps of S41 are as follows:

[0085] Fusion of a two-layer subgraph self-attention network and a multi-view stacked skip-level induction network:

[0086] X final =αX mv +(1-α)X bi

[0087] Among them, α is a hyperparameter, X bi 、X mv Features extracted for the two networks respectively;

[0088] The specific steps of S42 are as follows:

[0089] S421: Calculate the mean of positive examples, the mean of negative examples, the cosine similarity between each negative example and the mean of positive examples, the cosine similarity between each negative example and the mean of negative examples, the Pearson correlation coefficient between each negative example and the mean of positive examples, and the Pearson correlation coefficient between each negative example and the mean of negative examples in the same training batch;

[0090] S422: In the cosine similarity calculation result of S421, if the cosine similarity between the negative example and the mean of the negative examples is less than the cosine similarity between the negative example and the mean of the positive examples, and the cosine similarity between the negative example and the mean of the positive examples is greater than 0, record it as true;

[0091] S423: In the result of calculating the Pearson correlation coefficient in S421, if the Pearson correlation coefficient between the negative example and the mean of the negative examples is less than the Pearson correlation coefficient between the negative example and the mean of the positive examples, and the Pearson correlation coefficient between the negative example and the mean of the positive examples is greater than 0, record it as true;

[0092] S424: If both S422 and S423 are true, distance balancing is used for negative examples. The specific operations are as follows:

[0093] Link' neg =Link neg +WLink pos_mean

[0094] Where W is the learnable weight matrix, Link pos_mean is the mean of positive examples, Link neg is a negative example.

[0095] As a preferred solution of the drug relocalization method based on multi-view stacked skip-level graph convolutional network described in the present invention, the specific steps of S5 are as follows:

[0096] Use trained positive and negative examples to predict potential interactions between drugs and diseases

[0097]

[0098] Pred=MLP(Link)

[0099] in, is a concatenation operation, MLP is a multi-layer perceptron, and Link is the positive and negative drug-disease samples under the same training batch.

[0100] Compared with existing technologies:

[0101] When using high-order node information, the present invention can extract the contextual information of high-order nodes, thereby alleviating the inconsistency between local information and global information; the present invention can more comprehensively utilize deep information, thereby solving the problem that the network cannot distinguish between drug-disease nodes after multiple layers, thereby causing the model's prediction ability to decline; the present invention can more accurately distinguish between positive and negative examples, thereby solving the problem of data imbalance, which causes the model to be unable to accurately identify positive and negative examples. BRIEF DESCRIPTION OF THE DRAWINGS

[0102] Figure 1 Schematic diagram of the main structure of the multi-view stacked skip-level graph convolutional network of the present invention;

[0103] Figure 2 This is a schematic diagram of the overall process of the multi-view stacked skip-level graph convolutional network of the present invention;

[0104] Figure 3This is a schematic diagram of the dual-level subgraph self-attention network process of the present invention;

[0105] Figure 4 This is a schematic diagram of the multi-view stacking hopping sensing network process of the present invention;

[0106] Figure 5 Schematic diagram of the negative distance balance method of the present invention. DETAILED DESCRIPTION

[0107] To make the objectives, technical solutions and advantages of the present invention more clear, the embodiments of the present invention will be described in further detail below with reference to the accompanying drawings.

[0108] This invention provides a drug relocalization method based on a multi-view stacked skip-level graph convolutional network for predicting potential interactions between drugs and diseases. It includes constructing a two-level subgraph self-attention network, a view stacked skip-level induction network, and designing a negative example distance balance method. Figure 1-Figure 5 , including the following specific steps:

[0109] S1: Obtain drug-disease association data, and construct the drug-disease adjacency matrix A, association graph G, and drug-disease similarity matrix S. r 、S d , and uses a multi-layer graph convolutional encoder block to encode the similarity matrix of drugs and diseases into high-order feature representation;

[0110] The specific steps of S1 are as follows:

[0111] S11: Obtain drug-disease association data, and construct the drug-disease adjacency matrix A, association graph G, and drug-disease similarity matrix S. r 、S d ;

[0112] S12: S r 、S d Input to the multi-layer graph convolutional encoder block F encode , the specific form is as follows:

[0113]

[0114] Among them, σ is the activation function, v is the node, is the neighbor node of v, is the feature of node v, is the feature of neighbor node u, W l is the learnable weight matrix, are the degrees of node v and its neighbor node u respectively;

[0115] S13: Obtain drug high-level features and disease high-level features respectively, concatenate the two features together, and output drug-disease high-level features X;

[0116] S1 includes but is not limited to the following embodiments:

[0117] In the experiment of the present invention, a B-dataset dataset was obtained, which consists of 269 drug nodes, 598 disease nodes, and 18,416 drug-disease association pairs. The similarity between drugs was calculated using molecular fingerprints and Gaussian interaction distribution (GIP) kernel similarity, and the average value was used as the drug similarity matrix. The similarity between diseases was calculated using phenotype and GIP similarity, and the average value was used as the disease similarity matrix. The drug similarity matrix had a scale of 269×269, and the disease similarity matrix had a scale of 598×598. The two similarity matrices were respectively passed through a constructed multi-layer graph convolution encoder block to extract their respective high-order features, and were spliced ​​together as the input for subsequent network learning. The scale was 867×128, that is, there were 867 drug and disease nodes in total, and each node had a 128-dimensional feature.

[0118] S2: Construct a two-level subgraph self-attention network, combine node-level and subgraph-level features, and perform structural fusion with the graph topology. This paper designs a two-level subgraph self-attention network to integrate local and global information of drug-disease nodes. The process is as follows: extract node-level features and subgraph-level features of drug-disease and perform structural fusion; integrate local and global multi-level features;

[0119] The specific steps of S2 are as follows:

[0120] S21: Use the METIS software package to split the drug-disease association graph G into multiple equally divided subgraph sets P G , and use the high-order features X obtained by S1 to extract two-level features, namely the sub-graph level and the node level X sub 、X node , and use graph convolutional networks to extract structural features X strut The high-level drug-disease features obtained from S1 are used to construct two-level features and extract structural features using graph convolutional networks.

[0121] The specific steps of S21 are as follows:

[0122] S211: Get sub-layer level feature X sub ;

[0123] The specific steps of S211 are as follows:

[0124] S2111: Use the METIS tool to split the drug-disease association graph G into multiple equally divided subgraph sets P G;

[0125] S2112: Pool the node features in each subgraph so that the nodes in each subgraph share a feature. The specific form is as follows:

[0126]

[0127] in, is the feature of the nodes in each subgraph, X s is the subgraph feature;

[0128] S2113: Using subgraph self-attention to extract subgraph level features:

[0129]

[0130] Among them, X s is the subgraph feature, is the subgraph vector dimension;

[0131] Among them, METIS (Multilevel Partitioning of Graphs and Matrices) is a widely used toolkit that focuses on solving the partitioning and graph partitioning problems of large-scale datasets;

[0132] S212: Get node level feature X node , the specific steps are as follows:

[0133]

[0134] Among them, X is the high-order feature of drug-disease, d X is the drug-disease node vector dimension;

[0135] S213: Get structure-level feature X strut ;

[0136] The specific steps of S213 are as follows:

[0137] S2131: Copy the subgraph hierarchical features and assign them to all nodes in the same subgraph:

[0138] X' sub =repeat(X sub )

[0139] S2132: Construct a graph convolutional network combination block GCN_F to extract structural features:

[0140]

[0141] X strut =GCN_F(X' sub +Xnode )

[0142] in, is the feature of the disease node associated with the drug, is the characteristic of the drug node, W l is the learnable weight matrix, are the degrees of disease node d and drug node r, respectively;

[0143] S22: Fuse the node-level features, sub-map-level features, and structure-level features at the structural level, and output the fused features X bi ;

[0144] The specific steps of S22 are as follows:

[0145] The node-level features X obtained in the above steps are node , sub-layer level feature X sub and structural features X strut Perform structure-level fusion and output the fused feature X bi :

[0146] X bi =fusion(X' sub ,X node ,X strut )

[0147] Among them, fusion is structure-level fusion, and the specific operation is to add the corresponding elements of each feature;

[0148] S2 includes but is not limited to the following embodiments:

[0149] The drug-disease high-order feature X obtained by S1 is used as input, with a scale of 867×128, representing a total of 867 nodes (including drug and disease nodes), and the feature of each node is 128-dimensional; first, node self-attention is used to extract node-level features, with a scale of 867×128. Secondly, METIS is used to divide the drug-disease association graph into 128 subgraphs, with the same number of nodes in each subgraph; the node features in the 128 subgraphs are pooled as the features of each subgraph, with a scale of 128×128; the subgraph self-attention mechanism is used to extract subgraph-level features, with a scale of 128×128; the subgraph-level features are copied to each node so that the nodes in the subgraph have the same features, with a scale of 867x128; the subgraph-level features and the node-level features are fused and then the structure-level features are extracted through a custom GCN_F. Finally, the structure-level features are structurally fused with the subgraph-level features and the node-level features as the output features of the current network;

[0150] S3: Construct a multi-view stacked skip induction network to perform multi-view stacked learning on the association graph G, skip and induce deep feature information in multi-layer graph convolution, and learn the horizontal features of nodes. The present invention designs a multi-view stacked skip induction network for extracting deep feature information from multiple perspectives of nodes. Its steps include generating multiple channels using a random mask and a graph convolution network combination block; defining a domain skip induction mechanism to extract deep node information; stacking and concatenating all views, and outputting the feature X mv ;

[0151] The specific steps of S3 are as follows:

[0152] S31: Generate multiple channels using a random mask and a graph convolution network combination block; the generated channels extract the horizontal features of drug-disease nodes from multiple perspectives;

[0153] The specific steps of S31 are as follows:

[0154] S311: Perform random masking on the drug-disease high-order feature X in S1 for the following different views:

[0155] X' = qX

[0156] where 0 < q < 1 is the random masking ratio, and X' is the masked high-order feature;

[0157] S312: Construct GCN_H as a graph convolution network combination block for learning deep information of the graph network:

[0158]

[0159] where, is the feature of the disease node associated with the drug, is the feature of the drug node, W l is a learnable weight matrix, are the degrees of the disease node d and the drug node r respectively;

[0160] S313: Design the hierarchical representation for skip induction in the channel, and repeat L times to generate L layers:

[0161] Layer i = σ(BatchNorm(GraphSizeNorm((GCN_H(GCN_H(X'))))))

[0162] where GraphSizeNorm is graph normalization, specifically by dividing the feature of each node by the square root of the number of nodes in the graph, and BatchNorm is batch normalization, specifically by normalizing the feature using the mean and variance in the current batch;

[0163] S32: Define domain skipping sensing mechanism to extract deep node information;

[0164] The specific steps of S32 are as follows:

[0165] S321: Using the characteristics of self-attention mechanism and graph convolutional network, design a domain skipping induction mechanism, including Layer i Perform pooling, calculate the hierarchical self-attention scores, and select the K highest scores as the aggregated levels:

[0166] P i =Pooling(Layer i )

[0167]

[0168] TopLayer=select(score topK (Layer i ))

[0169] Among them, P i To pool the node features of each layer, score is used to calculate the attention score, TopLayer is the feature set of the K layers before the score, d P is the level vector dimension;

[0170] S322: Fuse the selected skip-level graph convolutional network layers with the drug-disease high-level feature X described in S1:

[0171]

[0172] Among them, X is the high-order feature of drug-disease, ∑ k∈K TopLayer is the selected deep features and;

[0173] S33: Stack the views and concatenate all views to output feature X mv ;

[0174] The specific steps of S33 are as follows:

[0175] Using Channel i As a foundation, we perform feature fusion by stacking views and S2 cross-domain features. To ensure the differences between views, we aggregate different features separately. The operation is as follows:

[0176] VIEW i =GCN_H(Channel i +VIEW i-1 +X bi )

[0177] VIEW i =GCN_H(Channel i +VIEW i-1 +X node +X sub )

[0178] Among them, X bi 、X node 、X sub Features extracted from the two-level subgraph self-attention network in S2;

[0179] Finally, multiple views are spliced ​​together as the output feature X mv :

[0180]

[0181] Among them, VIEW i For different views, is the splicing symbol;

[0182] S3 includes but is not limited to the following embodiments:

[0183] The high-order features X obtained by S1 are randomly masked at a ratio of 0.2 and distributed into 4 channels. Each channel extracts deep features from 6 layers of GCN_H. Each layer undergoes GCN_H, graph normalization, and batch normalization. The features of each layer are retained and the importance of each layer feature is calculated using self-attention. The top 3 scores are used as the number of layers for skip-level sensing. Therefore, the final scale is 4×867×32, representing 4 channels, 867 nodes, and 32 as the feature dimension. Each channel aggregates different sensing levels. Each channel fuses the features extracted by S2 and the features of the previous channel to generate a view stack. Finally, the 4 channels are spliced ​​together to a scale of 867×128 as the output features of this network.

[0184] S4: Fusion of a two-layer subgraph self-attention network and a multi-view stacked skip-level induction network. A negative distance balancing method is designed to enhance negative examples and optimize network performance. Features extracted by the S3 and S4 networks are fused, and the degree of deviation of negative examples is measured by cosine similarity and Pearson correlation coefficient to balance negative examples.

[0185] The specific steps of S4 are as follows:

[0186] S41: Fusion of a two-layer subgraph self-attention network and a multi-view stacked skip-level induction network;

[0187] S42: Design a negative distance balancing method for unbalanced negative examples; the operation of the negative distance balancing method is as follows Figure 5 As shown;

[0188] The specific steps of S41 are as follows:

[0189] Fusion of a two-layer subgraph self-attention network and a multi-view stacked skip-level induction network:

[0190] X final =αX mv +(1-α)X bi

[0191] Among them, α is a hyperparameter, X bi 、X mv Features extracted for the two networks respectively;

[0192] The specific steps of S42 are as follows:

[0193] S421: Calculate the mean of positive examples, the mean of negative examples, the cosine similarity between each negative example and the mean of positive examples, the cosine similarity between each negative example and the mean of negative examples, the Pearson correlation coefficient between each negative example and the mean of positive examples, and the Pearson correlation coefficient between each negative example and the mean of negative examples in the same training batch;

[0194] S422: In the cosine similarity calculation result of S421, if the cosine similarity between the negative example and the mean of the negative examples is less than the cosine similarity between the negative example and the mean of the positive examples, and the cosine similarity between the negative example and the mean of the positive examples is greater than 0, record it as true;

[0195] S423: In the result of calculating the Pearson correlation coefficient in S421, if the Pearson correlation coefficient between the negative example and the mean of the negative examples is less than the Pearson correlation coefficient between the negative example and the mean of the positive examples, and the Pearson correlation coefficient between the negative example and the mean of the positive examples is greater than 0, record it as true;

[0196] S424: If both S422 and S423 are true, distance balancing is used for negative examples. The specific operations are as follows:

[0197] Link' neg =Link neg +WLink pos_mean

[0198] Where W is the learnable weight matrix, Link pos_mean is the positive example mean, Link neg is a negative example;

[0199] S4 includes but is not limited to the following embodiments:

[0200] First, the features of the two networks are fused, and α is set to 0.7 to obtain drug-disease features with a scale of 867×128. There are 18416 associations in the drug-disease association graph. According to the drug-disease association graph, features are obtained for each positive and negative example. Associated drug-disease pairs are used as positive examples, and unassociated drug-disease pairs are used as negative examples, and finally 18416×128 positive example features and 18416×128 negative example features are obtained. The mean of all negative example features and the mean of positive example features are calculated, and the cosine similarity and Pearson correlation coefficient of each negative example with the negative example mean and the positive example mean are compared. If the cosine similarity and Pearson correlation coefficient of the negative example and the positive example mean are both greater than 0, and the cosine similarity and Pearson correlation coefficient of the negative example and the positive example mean are both greater than the cosine similarity and Pearson correlation coefficient of the negative example and the negative example mean, then the positive example mean of the learnable weight matrix is ​​fused for this negative example to expose the imbalance of negative examples.

[0201] Among them, in this network, positive examples are samples with a connection between drugs and diseases, and negative examples are samples without a connection between drugs and diseases;

[0202] S5: Use the optimized network to predict potential interactions between drugs and diseases;

[0203] The specific steps of S5 are as follows:

[0204] Use trained positive and negative examples to predict potential interactions between drugs and diseases

[0205]

[0206] Pred=MLP(Link)

[0207] in, is a concatenation operation, MLP is a multi-layer perceptron, and Link is the positive and negative drug-disease samples under the same training batch.

[0208] Based on the above, this paper designs a multi-view stacked skip-level graph convolutional network, whose structure includes two networks and one method: a two-level subgraph self-attention network, a multi-view stacked skip-level induction network, and a negative example distance balancing method;

[0209] The multi-view stacked skip-level graph convolutional network is used to analyze the complex relationships and topological structures between nodes in the drug-disease network; the two-level subgraph self-attention network extracts node-level features, subgraph-level features, and structure-level fusion to capture the local and global semantic information of drugs and diseases on the graph and the contextual information of long-distance nodes, effectively alleviating the inconsistency of global and local information; the multi-view stacked skip-level induction network stacks multiple views and simultaneously skips to aggregate different high-order neighbor information, effectively utilizing the deep information of nodes from multiple aspects, solving the problem of homogenization of node features of drugs and diseases due to too many network layers, and improving prediction accuracy; the negative example distance balance method adjusts the weights through the relationship between negative and positive examples, alleviating the sparsity and imbalance of drug and disease connections and improving the network's ability to distinguish samples.

[0210] Although the present invention has been described above with reference to embodiments, various modifications may be made thereto and equivalent components may be substituted without departing from the scope of the present invention. In particular, as long as there are no structural conflicts, the various features of the embodiments disclosed herein may be combined with each other in any manner, and the omission of an exhaustive description of such combinations in this specification is solely for the sake of space and resource conservation. Therefore, the present invention is not limited to the specific embodiments disclosed herein, but includes all technical solutions falling within the scope of the claims.

Claims

1. A drug relocalization method based on multi-view stacked skip-level graph convolutional network, characterized in that: The specific steps are as follows: S1: Obtain drug-disease association data, and construct the drug-disease adjacency matrix A, association graph G, and drug-disease similarity matrix S. r 、S d , and uses a multi-layer graph convolutional encoder block to encode the similarity matrix of drugs and diseases into high-order feature representation; S2: Construct a two-level subgraph self-attention network, combine the features of the node level and the sub-layer level, and perform structural-level fusion with the graph topology structure; S3: Construct a multi-view stacked skip-level induction network, perform multi-view stacked learning on the associated graph G, skip-level sense deep feature information in multi-layer graph convolution, and learn the horizontal features of nodes; S4: Integrate the two-level subgraph self-attention network and the multi-view stacked skip-level induction network, and design a negative example distance balancing method to enhance negative examples to optimize the network performance; S5: Use the optimized network to predict the potential interactions between drugs and diseases.

2. The drug relocalization method based on multi-view stacked skip-level graph convolutional network according to claim 1, characterized in that: The specific steps of S1 are as follows: S11: Obtain drug-disease association data, and construct the drug-disease adjacency matrix A, association graph G, and drug-disease similarity matrix S. r 、S d ; S12: S r 、S d Input to the multi-layer graph convolutional encoder block F encode , the specific form is as follows: Among them, σ is the activation function, v is the node, is the neighbor node of v, is the feature of node v, is the feature of neighbor node u, W l is the learnable weight matrix, are the degrees of node v and its neighbor node u respectively; S13: Obtain drug high-order features and disease high-order features respectively, concatenate the two features together, and output the drug-disease high-order feature X.

3. The drug relocalization method based on multi-view stacked skip-level graph convolutional network according to claim 1, characterized in that: The specific steps of S2 are as follows: S21: Use the METIS software package to split the drug-disease association graph G into multiple equally divided subgraph sets P G , and use the high-order features X obtained by S1 to extract two-level features, namely the sub-graph level and the node level X sub 、X node , and use graph convolutional networks to extract structural features X strut ; S22: Fuse the node-level features, sub-map-level features, and structure-level features at the structural level, and output the fused features X bi .

4. The drug relocalization method based on multi-view stacked skip-level graph convolutional network according to claim 3, characterized in that: The specific steps of S21 are as follows: S211: Get sub-layer level feature X sub ; S212: Get node level feature X node , the specific steps are as follows: Among them, X is the high-order feature of drug-disease, d X is the drug-disease node vector dimension; S213: Get structure-level feature X strut .

5. The drug relocalization method based on multi-view stacked skip-level graph convolutional network according to claim 4, characterized in that: The specific steps of S211 are as follows: S2111: Use the METIS tool to split the drug-disease association graph G into multiple equally divided subgraph sets P G ; S2112: Pool the node features in each subgraph so that the nodes within each subgraph share one feature. The specific form is as follows: in, is the feature of the nodes in each subgraph, X s is the subgraph feature; S2113: Use subgraph self-attention to extract sub-layer level features: Among them, X s is the subgraph feature, is the subgraph vector dimension.

6. The drug relocalization method based on multi-view stacked skip-level graph convolutional network according to claim 5, characterized in that: The specific steps of S213 are as follows: S2131: Copy the sub-layer level features and assign them to all nodes under the same subgraph: X' sub =repeat(X sub ) S2132: Construct a graph convolutional network combination block GCN_F to extract structural-level features: X strut =GCN_F(X' sub +X node ) in, is the feature of the disease node associated with the drug, is the characteristic of the drug node, W l is the learnable weight matrix, are the degrees of disease node d and drug node r, respectively.

7. The drug relocalization method based on multi-view stacked skip-level graph convolutional network according to claim 6, characterized in that: The specific steps of S22 are as follows: The node-level features X obtained in the above steps are node , sub-layer level feature X sub and structural features X strut Perform structure-level fusion and output the fused feature Xbi: X bi =fusion(X' sub ,X node ,X strut ) Among them, fusion is structural-level fusion, and the specific operation is to add the corresponding elements of each feature.

8. The drug relocalization method based on multi-view stacked skip-level graph convolutional network according to claim 1, characterized in that: The specific steps of S3 are as follows: S31: Use random masking and a graph convolutional network combination block to generate multiple channels; S32: Define a domain skip-level induction mechanism to extract node deep information; S33: Stack the views and concatenate all views to output feature X mv ; The specific steps of S31 are as follows: S311: Perform random masking on the drug-disease high-order feature X in S1 for the following different views: X' = qX Where 0 < q < 1 is the random masking ratio, and X' is the masked high-order feature; S312: Construct GCN_H as a graph convolutional network combination block to learn the deep information of the graph network: in, is the feature of the disease node associated with the drug, is the characteristic of the drug node, W l is the learnable weight matrix, are the degrees of disease node d and drug node r, respectively; S313: Design a hierarchical representation for skip-level induction in the channel, and repeat it L times to generate L layers: Layer i =σ(BatchNorm(GraphSizeNorm((GCN_H(GCN_H(X')))))) Among them, GraphSizeNorm is graph normalization. Specifically, divide the feature of each node by the square root of the number of nodes in the graph. BatchNorm is batch normalization. Specifically, normalize the feature using the mean and variance in the current batch; The specific steps of S32 are as follows: S321: Using the characteristics of self-attention mechanism and graph convolutional network, design a domain skipping induction mechanism, including Layer i Perform pooling, calculate the hierarchical self-attention scores, and select the K highest scores as the aggregated levels: P i =Pooling(Layer i ) TopLayer=select(score topK (Layer i )) Among them, P i To pool the node features of each layer, score is used to calculate the attention score, TopLayer is the feature set of the K layers before the score, d P is the level vector dimension; S322: Integrate the selected hierarchical levels of the skip-level graph convolutional network with the drug-disease high-order feature X described in S1: Among them, X is the high-order feature of drug-disease, ∑ k∈K TopLayer is the selected deep features and; The specific steps of S33 are as follows: Using Channel i As a foundation, we perform feature fusion by stacking views and S2 cross-domain features. To ensure the differences between views, we aggregate different features separately. The operation is as follows: VIEW i =GCN_H(Channel i +VIEW i-1 +X bi ) VIEW i =GCH_H(Channel i +VIEW i-1 +X node +X sub ) Among them, X bi 、X node 、X sub Features extracted from the two-level subgraph self-attention network in S2; Finally, multiple views are spliced ​​together as the output feature X mv : Among them, VIEW i For different views, A splicing symbol.

9. The drug relocalization method based on multi-view stacked skip-level graph convolutional network according to claim 1, characterized in that: The specific steps of S4 are as follows: S41: Integrate the two-level subgraph self-attention network and the multi-view stacked skip-level induction network; S42: Design a negative example distance balancing method for unbalanced negative examples; The specific steps of S41 are as follows: Integrate the two-level subgraph self-attention network and the multi-view stacked skip-level induction network: X final =αX mv +(1-a)X bi Among them, α is a hyperparameter, X bi 、X mv Features extracted for the two networks respectively; The specific steps of S42 are as follows: S421: Calculate the mean of positive examples, the mean of negative examples, the cosine similarity between each negative example and the mean of positive examples, the cosine similarity between each negative example and the mean of negative examples, the Pearson correlation coefficient between each negative example and the mean of positive examples, and the Pearson correlation coefficient between each negative example and the mean of negative examples in the same training batch; S422: In the cosine similarity calculation result of S421, if the cosine similarity between the negative example and the mean of the negative examples is less than the cosine similarity between the negative example and the mean of the positive examples, and the cosine similarity between the negative example and the mean of the positive examples is greater than 0, record it as true; S423: In the result of calculating the Pearson correlation coefficient in S421, if the Pearson correlation coefficient between the negative example and the mean of the negative examples is less than the Pearson correlation coefficient between the negative example and the mean of the positive examples, and the Pearson correlation coefficient between the negative example and the mean of the positive examples is greater than 0, record it as true; S424: If both S422 and S423 are true, distance balancing is used for negative examples. The specific operations are as follows: Link' neg =Link neg +WLink pos_mean Where W is the learnable weight matrix, Link pos_mean is the mean of positive examples, Link neg is a negative example.

10. The drug relocalization method based on multi-view stacked skip-level graph convolutional network according to claim 1, characterized in that: The specific steps of S5 are as follows: Use trained positive and negative examples to predict potential interactions between drugs and diseases Pred=MLP(Link) in, is a concatenation operation, MLP is a multi-layer perceptron, and Link is the positive and negative drug-disease samples under the same training batch.

Images