Drug-disease association prediction method, device, electronic device and readable storage medium

Abstract

The present disclosure provides a drug-disease association prediction method, device, electronic device and readable storage medium. The drug-disease association prediction method includes: obtaining a first drug embedding between each first type node and a drug node, obtaining a first disease embedding between each second type node and a disease node; obtaining a second drug embedding according to the weight of the first drug embedding and the fusion feature of the first drug embedding, and obtaining a second disease embedding according to the weight of the first disease embedding and the fusion feature of the first disease embedding; updating the second drug embedding according to the preset drug-disease association relationship to obtain a third drug embedding, and updating the second disease embedding according to the preset drug-disease association relationship to obtain a third disease embedding; generating an association prediction result of a target drug corresponding to a drug node for a target disease corresponding to a disease node according to the third drug embedding and the third disease embedding. The disclosed embodiment can improve the processing speed for drug-disease association prediction.

Description

Technical Field

[0001] The embodiments of the present disclosure relate to the field of computer technology, and in particular, to a drug-disease association prediction method, device, electronic device, and readable storage medium. Background Art

[0002] Drug repositioning refers to the process of applying existing drugs to new diseases. Compared with traditional drug development methods, drug repositioning can significantly reduce costs. At the same time, since the repositioned drugs have passed a large number of safety tests, the safety of existing drugs is known, thereby reducing the risk of drug development failure. In related technologies, drug associations are mainly predicted by the similarities between drugs or diseases, which requires a large amount of calculations and a long time. Summary of the invention

[0003] The embodiments of the present disclosure provide a drug-disease association prediction method, device, electronic device, and readable storage medium to solve the problem that the existing method of drug-disease association prediction takes a long time.

[0004] To solve the above problems, the present disclosure is implemented as follows:

[0005] In a first aspect, the present disclosure provides a method for predicting drug-disease association, comprising the following steps:

[0006] Obtaining a first drug embedding between each first-type node and a drug node, and obtaining a first disease embedding between each second-type node and a disease node, wherein the first-type nodes are all nodes other than the drug nodes and the disease nodes, and at least some of the first-type nodes and the second-type nodes are the same;

[0007] updating the first drug embedding according to the weight of the first drug embedding and the fusion feature of the first drug embedding to obtain a second drug embedding, and updating the first disease embedding according to the weight of the first disease embedding and the fusion feature of the first disease embedding to obtain a second disease embedding, wherein the fusion feature of the first drug embedding is determined according to the first drug embedding and the association information between the drug node and the disease node, and the fusion feature of the first disease embedding is determined according to the first disease embedding and the association information between the drug node and the disease node;

[0008] updating the second drug embedding according to the preset drug-disease association relationship to obtain a third drug embedding, and updating the second disease embedding according to the preset drug-disease association relationship to obtain a third disease embedding;

[0009] An association prediction result of the target drug corresponding to the drug node with respect to the target disease corresponding to the disease node is generated according to the third drug embedding and the third disease embedding.

[0010] In some embodiments, updating the first drug embedding to obtain the second drug embedding according to the weight of the first drug embedding and the fusion feature of the first drug embedding includes:

[0011] determining a weight of each of the first drug embeddings;

[0012] Obtaining first update information embedded in the first drug according to the weight of each first drug embedding;

[0013] Acquire fusion features between the first drug embeddings;

[0014] generating second update information of the first drug embedding according to the fusion features between the first drug embeddings;

[0015] The first update information and the second update information are superimposed to obtain a second drug embedding.

[0016] In some embodiments, updating the first disease embedding according to the weight of the first disease embedding and the fusion feature of the first disease embedding to obtain the second disease embedding includes:

[0017] determining a weight for each of the first disease embeddings;

[0018] Obtaining third updated information of the first disease embedding according to the weight of each of the first disease embeddings;

[0019] Obtaining fusion features between the first disease embeddings;

[0020] generating fourth updated information of the first disease embedding according to the fusion features between the first disease embeddings;

[0021] The third update information and the fourth update information are superimposed to obtain a second disease embedding.

[0022] In some embodiments, the updating the second drug embedding to obtain the third drug embedding according to the preset drug-disease association relationship includes:

[0023] generating fifth update information according to the association relationship between the drug node and the disease node;

[0024] The fifth update information and the second drug embedding are superimposed to obtain a third drug embedding.

[0025] In some embodiments, the updating of the second disease embedding to obtain the third disease embedding according to the preset drug-disease association relationship includes:

[0026] generating sixth update information according to the association relationship between the drug node and the disease node;

[0027] The sixth update information and the second disease embedding are superimposed to obtain a third disease embedding.

[0028] In some embodiments, generating, according to the third drug embedding and the third disease embedding, an association prediction result of the target drug corresponding to the drug node with respect to the target disease corresponding to the disease node includes:

[0029] generating an association parameter between the target drug and the target disease according to the third drug embedding and the third disease embedding;

[0030] When the associated parameter is greater than or equal to a preset parameter threshold, generating a prediction result that the target drug is effective for the target disease;

[0031] When the associated parameter is less than a preset parameter threshold, a prediction result is generated that the target drug is ineffective for the target disease.

[0032] In some embodiments, the initialization dimension of the first drug embedding is equal to the number of nodes corresponding to the first drug embedding, wherein the nodes corresponding to the first drug embedding include the drug node and the first type of node; and / or

[0033] The initialization dimension of the first disease embedding is equal to the number of nodes corresponding to the first disease node, wherein the nodes corresponding to the first disease node include the disease node and the first type node.

[0034] In some embodiments, the first type of node includes one or more of a gene node, a target node, a gene ontology node, and a side effect node.

[0035] In some embodiments, the second type of node comprises one or more of a gene node, a target node, and a gene ontology node.

[0036] In a second aspect, the present disclosure provides a drug-disease association prediction device, comprising:

[0037] an acquisition module, configured to acquire a first drug embedding between each first-type node and a drug node, and to acquire a first disease embedding between each second-type node and a disease node, wherein the first-type nodes are all nodes other than the drug nodes and the disease nodes, and at least some of the first-type nodes and the second-type nodes are the same;

[0038] A first updating module, configured to update the first drug embedding to obtain a second drug embedding according to a weight of the first drug embedding and a fusion feature of the first drug embedding, and to update the first disease embedding to obtain a second disease embedding according to a weight of the first disease embedding and a fusion feature of the first disease embedding, wherein the fusion feature of the first drug embedding is determined according to the first drug embedding and the association information between the drug node and the disease node, and the fusion feature of the first disease embedding is determined according to the first disease embedding and the association information between the drug node and the disease node;

[0039] A second updating module, configured to update the second drug embedding to obtain a third drug embedding according to a preset drug-disease association relationship, and to update the second disease embedding to obtain a third disease embedding according to the preset drug-disease association relationship;

[0040] A prediction module is used to generate an association prediction result of the target drug corresponding to the drug node with respect to the target disease corresponding to the disease node according to the third drug embedding and the third disease embedding.

[0041] In a third aspect, an embodiment of the present disclosure further provides an electronic device, comprising: a memory, a processor, and a program stored in the memory and executable on the processor; the processor is used to read the program in the memory to implement the steps in the method described in the first aspect above.

[0042] In a fourth aspect, an embodiment of the present disclosure further provides a readable storage medium for storing a program, wherein the program, when executed by a processor, implements the steps in the method described in the first aspect.

[0043] The embodiment of the present disclosure adopts a strategy of multi-channel information fusion, thereby avoiding the process of calculating similarity and reducing the time for calculating similarity. BRIEF DESCRIPTION OF THE DRAWINGS

[0044] In order to more clearly illustrate the technical solutions of the embodiments of the present disclosure, the drawings required for use in the description of the embodiments of the present disclosure will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present disclosure. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative labor.

[0045] Figure 1 is a flowchart of a drug-disease association prediction method provided by an embodiment of the present disclosure;

[0046] Figure 2 is a schematic diagram of node association in an embodiment of the present disclosure;

[0047] Figure 3is a schematic diagram of the processing flow of an embodiment of the present disclosure;

[0048] Figure 4A is another processing flow diagram provided by an embodiment of the present disclosure;

[0049] Figure 4B is another processing flow diagram provided by an embodiment of the present disclosure;

[0050] Figure 5 is another processing flow diagram provided by an embodiment of the present disclosure;

[0051] Figure 6 is a schematic diagram of a cascaded convolution processing flow provided by an embodiment of the present disclosure;

[0052] Figure 7 is a schematic diagram of the structure of a drug-disease association prediction device provided by an embodiment of the present disclosure;

[0053] Figure 8 It is a schematic diagram of the structure of an electronic device provided by an embodiment of the present disclosure. DETAILED DESCRIPTION

[0054] The following will be combined with the drawings in the embodiments of the present disclosure to clearly and completely describe the technical solutions in the embodiments of the present disclosure. Obviously, the described embodiments are part of the embodiments of the present disclosure, rather than all of the embodiments. Based on the embodiments in the present disclosure, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present disclosure.

[0055] The terms "first", "second" etc. in the disclosed embodiments are used to distinguish similar objects, and need not be used to describe a specific order or sequential order. In addition, the terms "include" and "have" and any variation thereof are intended to cover non-exclusive inclusions, for example, the process, method, system, product or equipment comprising a series of steps or units need not be limited to those steps or units clearly listed, but may include other steps or units that are not clearly listed or inherent to these processes, methods, products or equipment. In addition, "and / or" is used in the present application to represent at least one of the connected objects, such as A and / or B and / or C, indicating that A alone, B alone, C alone, and A and B all exist, B and C all exist, A and C all exist, and 7 situations in which A, B and C all exist.

[0056] The disclosed embodiments provide a method for predicting drug-disease association.

[0057] like Figures 1 to 6 As shown, in one embodiment, the method includes the following steps:

[0058] Step 101: Obtain a first drug embedding between each first type node and a drug node, and obtain a first disease embedding between each second type node and a disease node.

[0059] like Figure 2 and Figure 3 As shown, in this embodiment, a drug-disease multi-view heterogeneous information network between drugs and diseases is established to analyze the effectiveness of drugs against diseases.

[0060] In this embodiment, the constructed drug-disease multi-view heterogeneous information network G is recorded as G=(V, E). Wherein, V represents the node of the drug-disease multi-view heterogeneous information network, E represents the edge of the drug-disease multi-view heterogeneous information network, each node in the node set corresponds to a different type of object, and each edge in the edge set is used to represent the association relationship between different nodes.

[0061] In one embodiment, the network includes six different types of nodes, and the number of nodes of each type can be one or more.

[0062] In one embodiment, the nodes specifically include drug nodes, disease nodes, gene nodes, target nodes, Gene Ontology (abbreviated as GO) nodes and side effect nodes. Node V is a node in node set O, and node set O can be recorded as O={drug, disease, target, gene, GO, side effect}.

[0063] Gene ontology refers to a representation that contains the essence of the knowledge system in the field of biology. An ontology usually consists of a set of classes (or terms or concepts) with relationships between them.

[0064] Gene Ontology describes the understanding of biological fields from three aspects (GO domains):

[0065] 1. Molecular Function (MF): The activity of a single gene product (including protein and RNA) or a complex of multiple gene products at the molecular level, such as "catalysis" and "transport";

[0066] 2. Cellular Component (CC): The location of the cell structure where the gene product is located when performing its function, such as in mitochondria, ribose;

[0067] 3. Biological Process (BP) A biological process that is completed through a variety of molecular activities. Broad examples are DNA repair or signal transduction. More specific examples are pyrimidine nucleoside biosynthesis or glucose transport across membranes.

[0068] Please continue reading Figure 2 and Figure 3 , an edge represents the relationship between two nodes, Figure 2 A double-arrow line represents an edge between two nodes.

[0069] The network includes eight types of edges. Specifically, the edge set R can be recorded as R = {drug-disease association, drug-target association, drug-gene association, drug-GO association, drug-side effect association, disease-target association, disease-gene association, disease-GO association}.

[0070] In this embodiment, the first drug embedding includes edges related to the drug node and unrelated to the disease node, specifically, edges representing the association relationship between the drug node and the first type of node.

[0071] The first disease embedding includes edges related to the disease node and unrelated to the drug node, specifically, edges representing the association relationship between the disease node and the second type node.

[0072] The first type of nodes and the second type of nodes are nodes other than drug nodes and disease nodes, and at least some of the first type of nodes and the second type of nodes are the same.

[0073] In some embodiments, the first type of node includes one or more of a gene node, a target node, a gene ontology node, and a side effect node, and correspondingly, the first drug embedding includes one or more of a drug-target association, a drug-gene association, a drug-GO association, and a drug-side effect association.

[0074] In some embodiments, since the side effects are side effects of the drug, they are only associated with the drug and have no association with the disease. Therefore, the second type of node includes one or more of a gene node, a target node, and a gene ontology node. Correspondingly, the first disease embedding includes one or more of a disease-target association, a disease-gene association, and a disease-GO association.

[0075] It can be understood that, in this embodiment, according to the relationship between different types of nodes, the drug-disease multi-view heterogeneous information network is decomposed into three types of bipartite networks, where the bipartite network refers to a network containing two types of nodes.

[0076] like Figure 3 , Figure 4A and Figure 4B As shown, the first type of bipartite network includes drug nodes and non-disease nodes, that is, the first drug embedding mentioned above; the second type of bipartite network includes disease nodes and non-drug nodes, and does not include side effect nodes, that is, the first disease embedding mentioned above; the third type of network includes drug nodes and disease nodes.

[0077] Step 102: Update the first drug embedding according to the weight of the first drug embedding and the fusion feature of the first drug embedding to obtain the second drug embedding, and update the first disease embedding according to the weight of the first disease embedding and the fusion feature of the first disease embedding to obtain the second disease embedding.

[0078] In this embodiment, importance analysis is performed on the obtained multiple first drug insertions to determine the importance of each first drug insertion.

[0079] During implementation, the features of each first drug embedding may be extracted, and then the importance difference of each first drug embedding may be determined by analyzing the impact of each first drug embedding on the output result when it is used as an input.

[0080] Specifically, the features of each network included in the first drug embedding protection can be lifted through a convolution operation, and then the importance between different first drug embeddings can be determined through an attention mechanism, and then different weight coefficients can be assigned to different first drug embeddings.

[0081] In one embodiment, the step of updating the first drug embedding to obtain the second drug embedding according to the weight of the first drug embedding and the fusion feature of the first drug embedding includes:

[0082] determining a weight of each of the first drug embeddings;

[0083] Obtaining first update information embedded in the first drug according to the weight of each first drug embedding;

[0084] Acquire fusion features between the first drug embeddings;

[0085] generating second update information of the first drug embedding according to the fusion features between the first drug embeddings;

[0086] The first update information and the second update information are superimposed to obtain a second drug embedding.

[0087] like Figure 5 and Figure 6 As shown, let {G rm |m=1,2,…,M} is the first drug embedding set after partitioning. Let the number of nodes in the mth bipartite network in the first drug embedding be S m , where the number of drug nodes is S rm , the number of first type nodes is but

[0088] Furthermore, Indicates that {G rm} corresponding to the binary adjacency matrix, that is, the association relationship between the drug node and the first type of node in the first drug embedding.

[0089] If a drug node has an association with a first-type node, then A rm (i,j)=1, otherwise, A rm (i,j)=0; let It is the adjacency matrix between all nodes. Because the similarity relationship between drugs and the relationship between the first type of nodes are not considered, it is filled with a 0 matrix.

[0090] Network G rm The initialization node embedding is It is expressed as follows:

[0091]

[0092] In this way, the initialization dimension of the constructed first drug embedding is equal to the number of nodes corresponding to the first drug embedding. Here, the nodes corresponding to the first drug embedding include drug nodes and first type nodes. That is, S m G rm The number of nodes in is also the initialization feature dimension of each node.

[0093] GCN (Graph Convolutional Network) is a multi-layer connected neural network architecture used to learn low-dimensional representations of nodes in graph-structured data. Each layer of GCN directly aggregates the information of connected neighbors through the graph and uses the reconstructed embedding as the input of the next layer. The spectral graph convolution theorem defines convolution based on the normalized graph Laplacian operator in the Fourier domain:

[0094]

[0095] In the above formula, I rm is the identity matrix, To represent the degree matrix, since the degree distribution of nodes in HIN (Hetegeneous Information Network) may vary greatly, the asymmetric matrix Replace the symmetrical L rm , represents the transition probability matrix.

[0096] In this embodiment, the transition probability matrix P is used respectively rm Each resulting network is convolved as a Fourier basis.

[0097] Specifically, let Λ rm and Φ rm They are P rm The resulting convolution on each network is defined as follows:

[0098]

[0099]

[0100] Give a signal In the Fourier domain of the graph with the filter The product of represents the output of graph convolution. It's a signal To perform convolution on the local neighbors of the target node, define As a K-order polynomial filter,

[0101]

[0102] θ rm ∈R K is a vector of polynomial coefficients. Therefore,

[0103]

[0104] From the above, specifically, given the network G rm A with the corresponding adjacency matrix Yes, G. rm The convolution on depends only on nodes that are at most K steps away from the target node. In other words, the output signal after the convolution operation is defined by the K-order approximation of the local spectral filter on the network. The filter parameters are Can be used throughout the network rm Shared on. Extended to the general form, network G rm The convolution operation is defined as:

[0105]

[0106] in, and They represent the first layer respectively, d is the output node embedding dimension, σ(·) is the activation function, and ReLU(·) function is adopted.

[0107]

[0108] in, and are the trainable weight matrix and node embedding of the l+1th layer, respectively, so that convolution can be performed separately on each individual network of the first drug embedding.

[0109] For the first drug embedding in this embodiment, there are four types of meta-paths between drugs, namely:

[0110] There are four types: drug-target-drug, drug-gene-drug, drug-GO-drug, and drug-side effect-drug.

[0111] A meta-path refers to a connection path between two similar nodes in a heterogeneous network, for example, drug-disease-drug and drug-target-drug.

[0112] Different meta-paths represent different semantic information. For example, drug-disease-drug represents the association between different drugs and the same disease, while drug-target-drug represents the interaction between two drugs and the same target. Given a meta-path, each node has a set of neighbors based on the meta-path, which can reveal various structural information and rich semantic information in the heterogeneous graph.

[0113] In this embodiment, the output of M drug-related bipartite networks based on the attention mechanism is obtained through the following process.

[0114] Taking as input the M sets of semantically specific first drug embeddings learned from M bipartite networks, each meta-path The learning weights can be expressed as follows:

[0115]

[0116] Here sem Represents a method for performing semantic-level attention. is the feature matrix of the drug node learned based on the node-level attention level under M different meta-paths. It shows that the attention at the semantic level can obtain various types of semantic information contained in various meta-paths in heterogeneous networks. To understand the importance of each meta-path, a nonlinear transformation is first performed:

[0117]

[0118] Extract The operation process of the drug node in Figure 5 The C3 process in are all drug nodes in the mth bipartite network.

[0119]

[0120] Where W is the weight matrix, b is the bias vector, and q T is a semantically trainable weight vector used to measure the similarity between the embedding representations under multiple meta-paths. is the characteristic representation of drug node i under meta-path m. mis the drug node in the mth bipartite network. After obtaining the importance of each meta-path, it is normalized by the softmax function. By normalizing the importance of all meta-paths using the softmax function, the weight under the mth meta-path can be obtained:

[0121]

[0122] The weights can be interpreted as the contributions of different meta-paths. Obviously, The higher the value, the more important the meta-path m is. As coefficients, the different low-dimensional feature representations of drugs under these different meta-paths can be fused to obtain the low-dimensional embedding representation of the final drug node, as shown below:

[0123]

[0124] In order to utilize the homologous and heterogeneous information of heterogeneous information networks for embedding, the vectors of convolution signals are concatenated in sequence to obtain the final output signal of each node according to the network to which it belongs.

[0125] For nodes that are not network elements, a zero vector is used to represent the corresponding output signal. Represents G m The cascade convolution signal of the middle node is defined as G rm The layered convolution on is as follows:

[0126] It is the node feature matrix cascaded and spliced ​​based on the mth bipartite network in the first type of network set.

[0127] The specific splicing process is as follows: Figure 6 See Figure 6 In the C1 process, the blank area is the adjacency matrix area between non-drug nodes. Since this scheme does not use the interaction relationship and association relationship data between non-drug nodes, the relevant area is replaced by 0. For example, in the first row, the part below H2 to HT is a blank area, and in the second row, the part below H1 to HT except H2 is a blank area.

[0128]

[0129]

[0130]

[0131] in Represent the input signal, output signal and trainable weight matrix of the lth layer respectively,

[0132] Then, after the convolution processing corresponding to the C2 process and the Extract operation corresponding to the C3 process, the drug nodes in each binary network in the first type of network set are extracted, as shown in the figure, and expressed as:

[0133]

[0134]

[0135] Finally, concatenate the drug matrices of each binary network and pass through the fully connected neural network FC to obtain the final first-class network set, which corresponds to Figure 5 The C4 process in the drug node matrix, The formula is as follows:

[0136]

[0137]

[0138] Add the output of each layer of the first update information obtained by the attention mechanism operation and the second update information after the cascade convolution operation to obtain That is, the second drug insertion is obtained.

[0139] In this embodiment, the second disease embedding can be obtained by a method similar to the above process.

[0140] In one embodiment, the step of updating the first disease embedding to obtain the second disease embedding according to the weight of the first disease embedding and the fusion feature of the first disease embedding includes:

[0141] determining a weight for each of the first disease embeddings;

[0142] Obtaining third updated information of the first disease embedding according to the weight of each of the first disease embeddings;

[0143] Obtaining fusion features between the first disease embeddings;

[0144] generating fourth updated information of the first disease embedding according to the fusion features between the first disease embeddings;

[0145] The third update information and the fourth update information are superimposed to obtain a second disease embedding.

[0146] Specifically, let {G dn |n=1,2,…,N} is the first disease embedding set after partitioning. Let the number of nodes in the nth bipartite network in the first disease embedding be S n , where the number of disease nodes is S dn , the number of second type nodes is but

[0147] Furthermore, Indicates that {G dn}, that is, the association relationship between the disease node and the second type node in the first disease embedding.

[0148] If a disease node has an association with a second type node, then A dn (i,j)=1, otherwise, A dn (i,j)=0; let It is the adjacency matrix between all nodes. Because the similarity relationship between diseases and the relationship between non-disease nodes are not considered, it is filled with 0 matrix.

[0149] Network G dn The initialization node embedding is It is expressed as follows:

[0150]

[0151] The initialization dimension of the constructed first disease embedding is equal to the number of nodes corresponding to the first disease embedding, S n G dn The number of nodes in is also the initial feature dimension of each node. The derivation process of the convolution part is the same as the convolution operation of the first type of network set mentioned above, and will not be repeated here.

[0152] Given a network G dn A with the corresponding adjacency matrix Yes, G. dn The convolution on depends only on nodes that are at most K steps away from the target node. In other words, the output signal after the convolution operation is defined by the K-order approximation of the local spectral filter on the network. The filter parameters are Can be used throughout the network dn Shared on. Extended to the general form, network G dn The convolution operation is defined as:

[0153]

[0154] in, and They represent the first layer respectively, d is the output node embedding dimension, σ(·) is the activation function, and ReLU(·) function is adopted.

[0155]

[0156] in, and are the trainable weight matrix and node embedding of the l+1th layer, respectively, so that convolution can be performed separately on each individual network of the first disease embedding.

[0157] Consistent with the first drug embedding process, regarding the first disease embedding in this embodiment, there are three meta-paths between diseases, namely: disease-target-disease, disease-gene-disease, and disease-GO-disease.

[0158] In this embodiment, the outputs of N disease-related bipartite networks based on the attention mechanism are further obtained based on the following process.

[0159] Taking N sets of semantically specific disease node embeddings learned from N bipartite networks as input, each meta-path The learning weights can be expressed as follows:

[0160]

[0161] Here sem Represents a method for performing semantic-level attention. It is the feature matrix of the disease node learned based on the node-level attention level under N different meta-paths. It shows that the attention at the semantic level can obtain various types of semantic information contained in various meta-paths in heterogeneous networks.

[0162] To understand the importance of each meta-path, a non-linear transformation is first performed.

[0163]

[0164] Extract The operation process of the disease node in are all disease nodes in the nth bipartite network.

[0165]

[0166] Where W is the weight matrix, b is the bias vector, and q T is a semantically trainable weight vector used to measure the similarity between the embedding representations under multiple meta-paths. V is the feature representation of disease node i under meta-path n. n is the disease node in the nth bipartite network. After obtaining the importance of each meta-path, it is normalized by the softmax function. By normalizing the importance of all meta-paths using the softmax function, the weight under the nth meta-path can be obtained:

[0167]

[0168] can be interpreted as the contribution of different meta-paths. Obviously, The higher the value, the more important the meta-path n is. As coefficients, the different low-dimensional feature representations of the disease under these different meta-paths can be fused to obtain the low-dimensional embedding representation of the final disease node, as shown below:

[0169]

[0170] In order to utilize the homogeneous and heterogeneous information of heterogeneous information networks for embedding, the vectors of convolution signals are concatenated in sequence to obtain the final output signal of each node according to the network to which it belongs. For nodes that are not network elements, zero vectors are used to represent the corresponding output signals.

[0171] set up Represents G dn The cascade convolution signal of the middle node, in this embodiment, is defined as G dn The layered convolution on is as follows:

[0172] It is the node feature matrix of the first type of network set based on the cascaded splicing of the nth bipartite network. The specific splicing process is shown in the figure. It is the first concatenation process in the figure. The white area is the adjacency matrix area between non-disease nodes. Because this scheme does not use the interaction relationship and association relationship data between non-disease nodes, the relevant area is replaced by 0.

[0173]

[0174]

[0175]

[0176] in Represent the input signal, output signal and trainable weight matrix of the lth layer respectively,

[0177] Then, after the Extract operation, the disease nodes in each bipartite network in the second type of network set are extracted, as shown in the figure, expressed as:

[0178]

[0179]

[0180] Finally, concatenate the disease matrices of each binary network and pass them through the fully connected neural network FC to obtain the node matrix of the disease in the final first-class network set. The formula is as follows:

[0181]

[0182]

[0183] Add the third update information obtained by the attention mechanism operation and the fourth update information obtained by the cascade convolution operation to the output of each layer, and we get In this way, the second disease is embedded.

[0184] Step 103: updating the second drug embedding according to the preset drug-disease association relationship to obtain a third drug embedding, and updating the second disease embedding according to the preset drug-disease association relationship to obtain a third disease embedding.

[0185] In this embodiment, the preset drug-disease association relationship refers to the association relationship between known drugs and diseases. Since the known association between drugs and diseases has been verified, its reliability is relatively high. Therefore, the accuracy of the prediction results can be improved based on the known association between drugs and diseases.

[0186] In some embodiments, updating the second drug embedding to obtain the third drug embedding according to the preset drug-disease association relationship includes:

[0187] generating fifth update information according to the association relationship between the drug node and the disease node;

[0188] The fifth update information and the second drug embedding are superimposed to obtain a third drug embedding.

[0189] generating fifth update information according to the association relationship between the drug node and the disease node;

[0190] The fifth update information and the second drug embedding are superimposed to obtain a third drug embedding.

[0191] Next, fifth update information needs to be determined, and the fifth update information is determined based on the fusion feature of the first drug embedding, and the fusion feature of the first drug embedding is determined based on the first drug embedding and the association information between the drug node and the disease node.

[0192] Specifically, the location of the drug-disease association network is denoted as G rd Here, the drug-disease association network refers to the third type of network mentioned above. Let the number of drug nodes be u, the number of disease nodes be v, A rd is the adjacency matrix of the drug-disease association network. If a drug node is associated with a disease node, then A rd (i, j) = 1, otherwise, A rd(i, j) = 0, then we can get:

[0193]

[0194] set up It is the adjacency matrix between all nodes. Because the similarity relationship between diseases and the relationship between non-disease nodes are not considered, it is filled with 0 matrix.

[0195] Network G rd The initialization node embedding is It is expressed as follows:

[0196]

[0197] in,

[0198] For the initialized drug and disease nodes, a variety of methods can be used to learn their low-dimensional node embeddings, exemplarily including matrix decomposition, random walk, neural network and other methods.

[0199] In this embodiment, GCN is used as an example.

[0200]

[0201] in, and They represent the initial trainable weight matrix and the node embedding of the first layer respectively, d is the output node embedding dimension, σ(·) is the activation function, and ReLU(·) function is adopted.

[0202]

[0203]

[0204] in, and They are the node embedding of the lth layer, the trainable weight matrix of the l+1th layer, and the node embedding of the l+1th layer. In this way, the fifth update information is obtained.

[0205] Next, the fifth update information and the second drug embedding are superimposed to obtain a third drug embedding.

[0206] Specifically, the processing result of the first drug embedding of each layer is superimposed with the processing result of the first drug embedding of the previous layer and the processing result of the drug-disease association network of the corresponding layer, so that the following can be obtained:

[0207]

[0208] Among them, since the initial state has no prior information as input, the result of the first layer does not include the output result of the 0th layer, and is obtained by superimposing the processing result of the first drug embedding of the corresponding layer and the processing result of the drug-disease association network.

[0209] Furthermore, since the contributions of different first drug embeddings in different layers are different, in implementation, these first drug embeddings can be superimposed through the attention matrix to obtain the third drug embedding:

[0210]

[0211] H R ∈R u×d ;

[0212] Here, α l It is obtained through automatic learning and can be initialized to 1 / (l+1), l=1,2,...,L.

[0213] In some embodiments, updating the second disease embedding to obtain the third disease embedding according to the preset drug-disease association relationship includes:

[0214] generating sixth update information according to the association relationship between the drug node and the disease node;

[0215] The sixth update information and the second disease embedding are superimposed to obtain a third disease embedding.

[0216] In some embodiments, the sixth update information is obtained in a similar manner to the fifth update information. The third disease embedding can be determined by the similar method described above, as follows:

[0217]

[0218]

[0219] H R ∈R u×d ,H D ∈R v×d ;

[0220] Similarly, β l It is obtained through automatic learning and is also initialized to 1 / (l+1), l=1,2,...,L.

[0221] Step 104: Generate an association prediction result of the target drug corresponding to the drug node with respect to the target disease corresponding to the disease node according to the third drug embedding and the third disease embedding.

[0222] In some embodiments, step 104 specifically includes:

[0223] generating an association parameter between the target drug and the target disease according to the third drug embedding and the third disease embedding;

[0224] When the associated parameter is greater than or equal to a preset parameter threshold, generating a prediction result that the target drug is effective for the target disease;

[0225] When the associated parameter is less than a preset parameter threshold, a prediction result is generated that the target drug is ineffective for the target disease.

[0226] In this embodiment, the association parameter between the target drug and the target disease can be determined by a decoder. In some embodiments, the following decoders can be selected:

[0227]

[0228] A′∈R u×v , is the prediction probability score matrix. Target drug r i and target diseases j The predicted score of the association between ij Item is given.

[0229] Exemplarily, in some embodiments, the preset parameter threshold is 1. If the association parameter between the target drug and the target disease is equal to 1, the target drug is considered to be effective for the target disease. If the association parameter between the target drug and the target disease is less than 1, the target drug is considered to be ineffective for the target disease.

[0230] Exemplarily, in some embodiments, the preset parameter threshold is 0.8. If the association parameter between the target drug and the target disease is greater than or equal to 0.8, the target drug is considered to be effective for the target disease. If the association parameter between the target drug and the target disease is less than 0.8, the target drug is considered to be ineffective for the target disease.

[0231] In some embodiments, the step of optimizing the prediction results is also included.

[0232] Since the known associations between drugs and diseases have been verified, their reliability is relatively high, which is very important for improving prediction performance. However, the number of known drug-disease associations is far less than the number of unknown or unobserved drug-disease pairs. Therefore, in this example, the parameters are learned by minimizing the weighted binary cross entropy loss, as shown below:

[0233]

[0234] Where (i, j) represents drug r i and diseases j Yes, S.+ represents the set of all known drug-disease association pairs, S - represents the set of all unknown or unobserved drug-disease association pairs. Balancing factor Used to reduce the impact of data imbalance, where |S + | and |S - ||S + and S - The model is optimized using the Adam optimizer.

[0235] The disclosed embodiment uses the graph neural network in the field of drug repositioning based on cascade convolution operations. Through cascade convolution, the features under multiple meta-paths can be learned, which is convenient for processing results using multiple information, avoiding the decrease in the generalization ability of the results caused by the limitations and particularity of single information. The interactive information between different types of nodes and the hidden semantic information between nodes can be integrated.

[0236] Furthermore, the embodiment of the present disclosure adopts a strategy of multi-channel information fusion, thereby avoiding the process of calculating similarity and reducing the time for calculating similarity.

[0237] This embodiment also provides a drug-disease association prediction device.

[0238] like Figure 7 As shown, in one embodiment, the drug-disease association prediction device 700 includes:

[0239] An acquisition module 701 is used to acquire a first drug embedding between each first-type node and a drug node, and to acquire a first disease embedding between each second-type node and a disease node, wherein the first-type nodes are all nodes other than the drug nodes and the disease nodes, and at least some of the first-type nodes and the second-type nodes are the same;

[0240] A first updating module 702 is configured to update the first drug embedding to obtain a second drug embedding according to the weight of the first drug embedding and the fusion feature of the first drug embedding, and to update the first disease embedding to obtain a second disease embedding according to the weight of the first disease embedding and the fusion feature of the first disease embedding, wherein the fusion feature of the first drug embedding is determined according to the first drug embedding and the association information between the drug node and the disease node, and the fusion feature of the first disease embedding is determined according to the first disease embedding and the association information between the drug node and the disease node;

[0241] A second updating module 703, configured to update the second drug embedding to obtain a third drug embedding according to a preset drug-disease association relationship, and to update the second disease embedding to obtain a third disease embedding according to the preset drug-disease association relationship;

[0242] The prediction module 704 is used to generate an association prediction result of the target drug corresponding to the drug node with respect to the target disease corresponding to the disease node according to the third drug embedding and the third disease embedding.

[0243] In some embodiments, the first updating module 702 is specifically used to:

[0244] determining a weight of each of the first drug embeddings;

[0245] Obtaining first update information embedded in the first drug according to the weight of each first drug embedding;

[0246] Acquire fusion features between the first drug embeddings;

[0247] generating second update information of the first drug embedding according to the fusion features between the first drug embeddings;

[0248] The first update information and the second update information are superimposed to obtain a second drug embedding.

[0249] In some embodiments, the first updating module 702 is specifically used to:

[0250] determining a weight for each of the first disease embeddings;

[0251] Obtaining third updated information of the first disease embedding according to the weight of each of the first disease embeddings;

[0252] Obtaining fusion features between the first disease embeddings;

[0253] generating fourth updated information of the first disease embedding according to the fusion features between the first disease embeddings;

[0254] The third update information and the fourth update information are superimposed to obtain a second disease embedding.

[0255] In some embodiments, the second updating module 703 is specifically used to:

[0256] generating fifth update information according to the association relationship between the drug node and the disease node;

[0257] The fifth update information and the second drug embedding are superimposed to obtain a third drug embedding.

[0258] In some embodiments, the second updating module 703 is specifically used to:

[0259] generating sixth update information according to the association relationship between the drug node and the disease node;

[0260] The sixth update information and the second disease embedding are superimposed to obtain a third disease embedding.

[0261] In some embodiments, the prediction module 704 is specifically used to:

[0262] generating an association parameter between the target drug and the target disease according to the third drug embedding and the third disease embedding;

[0263] When the associated parameter is greater than or equal to a preset parameter threshold, generating a prediction result that the target drug is effective for the target disease;

[0264] When the associated parameter is less than a preset parameter threshold, a prediction result is generated that the target drug is ineffective for the target disease.

[0265] In some embodiments, the initialization dimension of the first drug embedding is equal to the number of nodes corresponding to the first drug embedding, wherein the nodes corresponding to the first drug embedding include the drug node and the first type of node; and / or

[0266] The initialization dimension of the first disease embedding is equal to the number of nodes corresponding to the first disease node, wherein the nodes corresponding to the first disease node include the disease node and the first type node.

[0267] In some embodiments, the first type of node includes one or more of a gene node, a target node, a gene ontology node, and a side effect node.

[0268] In some embodiments, the second type of node comprises one or more of a gene node, a target node, and a gene ontology node.

[0269] The drug-disease association prediction device 700 of this embodiment can implement each step of the above method embodiment and can achieve basically the same technical effects, which will not be described in detail here.

[0270] The present disclosure also provides an electronic device. Figure 8 , the electronic device may include a processor 801, a memory 802, and a program 8021 stored in the memory 802 and executable on the processor 801.

[0271] When program 8021 is executed by processor 801, any steps in the above method embodiment can be implemented and the same beneficial effects can be achieved, which will not be described in detail here.

[0272] Those skilled in the art will appreciate that all or part of the steps of implementing the above-mentioned embodiment method can be completed by hardware associated with program instructions, and the program can be stored in a readable medium.

[0273] The embodiments of the present disclosure also provide a readable storage medium, on which a computer program is stored. When the computer program is executed by a processor, any step in the above method embodiment can be implemented and the same technical effect can be achieved. To avoid repetition, it will not be repeated here.

[0274] The storage medium may be a read-only memory (ROM), a random access memory (RAM), a magnetic disk or an optical disk.

[0275] It should be noted that it should be understood that the division of the above modules is only a division of logical functions. In actual implementation, they can be fully or partially integrated into one physical entity, or they can be physically separated. And these modules can all be implemented in the form of software called by processing elements; they can also be all implemented in the form of hardware; some modules can also be implemented in the form of software called by processing elements, and some modules can be implemented in the form of hardware. For example, the determination module can be a separately established processing element, or it can be integrated in a chip of the above-mentioned device for implementation. In addition, it can also be stored in the memory of the above-mentioned device in the form of program code, and called and executed by a processing element of the above-mentioned device. The implementation of other modules is similar. In addition, these modules can be fully or partially integrated together, or they can be implemented independently. The processing element described here can be an integrated circuit with signal processing capabilities. In the implementation process, each step of the above method or each module above can be completed by an integrated logic circuit of hardware in the processor element or instructions in the form of software.

[0276] For example, each module, unit, sub-unit or sub-module may be one or more integrated circuits configured to implement the above method, such as: one or more application specific integrated circuits (ASIC), or one or more microprocessors (digital signal processor, DSP), or one or more field programmable gate arrays (FPGA), etc. For another example, when a module above is implemented in the form of a processing element scheduling program code, the processing element may be a general-purpose processor, such as a central processing unit (CPU) or other processor that can call program code. For another example, these modules can be integrated together and implemented in the form of a system-on-a-chip (SOC).

[0277] The above is a preferred implementation of the embodiment of the present disclosure. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the principles described in the present disclosure. These improvements and modifications should also be regarded as the scope of protection of the present disclosure.

Claims

1. A drug-disease association prediction method, characterized in that: The following steps are involved: Obtaining a first drug embedding between each first-type node and a drug node, and obtaining a first disease embedding between each second-type node and a disease node, wherein the first-type nodes are all nodes other than the drug nodes and the disease nodes, and at least some of the first-type nodes and the second-type nodes are the same; updating the first drug embedding according to the weight of the first drug embedding and the fusion feature of the first drug embedding to obtain a second drug embedding, and updating the first disease embedding according to the weight of the first disease embedding and the fusion feature of the first disease embedding to obtain a second disease embedding, wherein the fusion feature of the first drug embedding is determined according to the first drug embedding and the association information between the drug node and the disease node, and the fusion feature of the first disease embedding is determined according to the first disease embedding and the association information between the drug node and the disease node; updating the second drug embedding according to the preset drug-disease association relationship to obtain a third drug embedding, and updating the second disease embedding according to the preset drug-disease association relationship to obtain a third disease embedding; An association prediction result of the target drug corresponding to the drug node with respect to the target disease corresponding to the disease node is generated according to the third drug embedding and the third disease embedding.

2. The method according to claim 1, characterized in that The updating the first drug embedding to obtain the second drug embedding according to the weight of the first drug embedding and the fusion feature of the first drug embedding includes: determining a weight of each of the first drug embeddings; Obtaining first update information embedded in the first drug according to the weight of each first drug embedding; Acquire fusion features between the first drug embeddings; generating second update information of the first drug embedding according to the fusion features between the first drug embeddings; The first update information and the second update information are superimposed to obtain a second drug embedding.

3. The method according to claim 1 or 2, characterized in that The updating the first disease embedding according to the weight of the first disease embedding and the fusion feature of the first disease embedding to obtain the second disease embedding includes: determining a weight for each of the first disease embeddings; Obtaining third updated information of the first disease embedding according to the weight of each of the first disease embeddings; Obtaining fusion features between the first disease embeddings; generating fourth updated information of the first disease embedding according to the fusion features between the first disease embeddings; The third update information and the fourth update information are superimposed to obtain a second disease embedding.

4. The method according to claim 1, characterized in that The updating of the second drug embedding to obtain the third drug embedding according to the preset drug-disease association relationship includes: generating fifth update information according to the association relationship between the drug node and the disease node; The fifth update information and the second drug embedding are superimposed to obtain a third drug embedding.

5. The method according to claim 1 or 4, characterized in that The updating of the second disease embedding to obtain the third disease embedding according to the preset drug-disease association relationship includes: generating sixth update information according to the association relationship between the drug node and the disease node; The sixth update information and the second disease embedding are superimposed to obtain a third disease embedding.

6. The method according to claim 1, characterized in that The generating, according to the third drug embedding and the third disease embedding, an association prediction result of the target drug corresponding to the drug node with respect to the target disease corresponding to the disease node comprises: generating an association parameter between the target drug and the target disease according to the third drug embedding and the third disease embedding; When the associated parameter is greater than or equal to a preset parameter threshold, generating a prediction result that the target drug is effective for the target disease; When the associated parameter is less than a preset parameter threshold, a prediction result is generated that the target drug is ineffective for the target disease.

7. The method according to claim 1, characterized in that The initialization dimension of the first drug embedding is equal to the number of nodes corresponding to the first drug embedding, wherein the nodes corresponding to the first drug embedding include the drug node and the first type of node; and / or The initialization dimension of the first disease embedding is equal to the number of nodes corresponding to the first disease node, wherein the nodes corresponding to the first disease node include the disease node and the first type node.

8. The method according to claim 1, characterized in that The first type of nodes includes one or more of gene nodes, target nodes, gene ontology nodes and side effect nodes.

9. The method according to claim 1 or 7, characterized in that: The second type of nodes includes one or more of gene nodes, target nodes and gene ontology nodes.

10. A drug-disease association prediction device, characterized in that: include: an acquisition module, configured to acquire a first drug embedding between each first-type node and a drug node, and to acquire a first disease embedding between each second-type node and a disease node, wherein the first-type nodes are all nodes other than the drug nodes and the disease nodes, and at least some of the first-type nodes and the second-type nodes are the same; a first updating module, configured to update the first drug embedding to obtain a second drug embedding according to a weight of the first drug embedding and a fusion feature of the first drug embedding, and to update the first disease embedding to obtain a second disease embedding according to a weight of the first disease embedding and a fusion feature of the first disease embedding, wherein the fusion feature of the first drug embedding is determined according to the first drug embedding and the association information between the drug node and the disease node, and the fusion feature of the first disease embedding is determined according to the first disease embedding and the association information between the drug node and the disease node; A second updating module, configured to update the second drug embedding to obtain a third drug embedding according to a preset drug-disease association relationship, and to update the second disease embedding to obtain a third disease embedding according to the preset drug-disease association relationship; A prediction module is used to generate an association prediction result of the target drug corresponding to the drug node with respect to the target disease corresponding to the disease node according to the third drug embedding and the third disease embedding.

11. An electronic device, comprising: A memory, a processor, and a program stored in the memory and executable on the processor; wherein the processor is used to read the program in the memory to implement the steps in the drug-disease association prediction method as described in any one of claims 1 to 9.

12. A readable storage medium for storing a program, characterized in that: When the program is executed by a processor, the steps of the drug-disease association prediction method according to any one of claims 1 to 9 are implemented.

Images