Drug and microorganism association prediction method based on node category sensitive knowledge graph learning and gating enhanced meta-path semantic fusion

Abstract

The invention discloses a drug and microorganism association prediction method based on node category sensitive knowledge graph learning and gating enhanced meta-path semantic fusion. The method comprises the following steps: constructing a drug and microorganism entity knowledge graph; on the basis of a category-sensitive transform interactive attention mechanism, integrating various relations among the entity feature representations to obtain entity features; constructing a plurality of meta-paths based on biological semantics, learning specific semantic features of each meta-path through a recursive gating strategy, and fusing context relationships of the plurality of meta-paths to obtain node semantic features; and balancing entity features and node semantic features under multiple perspectives. According to the method, through a transformer cross attention mechanism sensitive to entity categories, a recursive gating strategy fusing context semantics of a plurality of meta-paths and a contrast learning strategy sensitive to entity categories, the feature consistency under the view angles of the knowledge graph and the meta-paths is enhanced, and the accuracy of drug and microorganism association prediction is improved.

Description

Technical Field

[0001] The present invention relates to the field of bioinformatics, and in particular to a method for predicting drug and microbial associations based on node category-sensitive knowledge graph learning and gated enhanced meta-path semantic fusion. Background Art

[0002] In recent years, deep neural network learning has achieved remarkable breakthroughs in computer vision, natural language processing, and speech recognition. At the same time, in key areas such as text processing, recommendation systems, and financial stock price prediction, natural language processing technology based on graph learning has demonstrated excellent performance. First, graph learning methods allow multiple types of relationships (such as semantic relationships) to be modeled simultaneously, while traditional NLP methods usually only process a single type of feature, resulting in information loss. Secondly, graph learning methods are able to link scattered information through a global graph structure to form a more complete semantic representation. Finally, graph learning can obtain information from neighbor nodes through a message passing mechanism to alleviate the problem of data sparsity, while traditional methods may perform poorly in sparse data scenarios.

[0003] At present, graph learning-based methods have been widely studied and applied to association prediction models. However, these graph learning models generally only model the association between nodes through simple adjacency relationships, which makes it difficult to capture complex relationships. In order to capture the complex connections between nodes, researchers have tried to introduce knowledge graphs to capture the complex connections between entities by building knowledge graphs to improve the accuracy of association prediction models. Combined with knowledge graphs, graph attention networks (GATs) give different weights to different neighbors to aggregate neighbor information, and are therefore widely used in various tasks. However, GATs ignore the heterogeneity of entities and the similarity and association between entities. In order to learn the contextual connections between different meta-paths, GCN networks are widely used. However, this method cannot achieve long-term and high-order spatial interactions between meta-paths. By introducing multi-view feature learning, the model can extract attribute information and semantic characteristics of nodes from different levels. However, the model fails to effectively balance the features under multiple views, resulting in feature redundancy and information duplication.

[0004] Therefore, there is an urgent need to develop a drug and microorganism association prediction method based on node category-sensitive knowledge graph learning and gated enhanced meta-path semantic fusion, which can aggregate the heterogeneity of entities, complex connections between entities, and tail entity information in the knowledge graph, as well as the spatial contextual connections between different meta-paths, and enhance the feature consistency between the knowledge graph perspective and the meta-path perspective, so as to effectively capture the complex association relationship between microorganism and drug nodes and improve the accuracy of association prediction. Summary of the invention

[0005] Based on the above background, the present invention provides a drug and microorganism association prediction method based on node category-sensitive knowledge graph learning and gated enhanced meta-path semantic fusion, which is used to solve the technical problems faced by current drug and microorganism association prediction methods when processing multi-view features, such as unbalanced feature distribution and inability to effectively capture the complex association relationship between microorganism and drug nodes.

[0006] The present invention provides a method for predicting drug and microbial association based on node category-sensitive knowledge graph learning and gated enhanced meta-path semantic fusion, comprising:

[0007] A knowledge graph is constructed based on the data source, wherein entities of the knowledge graph are classified into drugs and microorganisms, and feature representations of each entity are obtained based on the category feature vector and attribute vector of each entity;

[0008] The entity category-sensitive first network model built based on the Transformer model is used to integrate the multi-category connections between the feature representations of each entity to obtain the entity feature representations of drug and microbial entities.

[0009] According to the data source, a plurality of meta-paths are constructed with drugs and microorganisms as nodes respectively, and contextual connections and adjacency semantics between the plurality of meta-paths are encoded through a second network model with a meta-path semantic feature learning strategy with recursive gating to obtain semantic representations of the nodes;

[0010] Inputting the entity feature representations of the drug and microorganism entities and the semantic representations of the nodes into a third network model that is sensitive to node categories, enhancing the consistency of node features, and obtaining a drug-microorganism association model;

[0011] The drugs and microorganisms whose association relationships are to be predicted are input into the drug-microorganism association model to complete the association prediction.

[0012] Furthermore, the feature representation of each entity is obtained according to the category feature vector and the attribute vector of each entity, including:

[0013] According to the similarity matrix between drugs, the similarity matrix between microorganisms, and the association matrix between microorganisms and drugs, the tail entities corresponding to the entities and the relationship embedding between the head entity of any entity and the corresponding tail entity are obtained;

[0014] The attribute vector of each entity, the category feature vector, the attribute vector of the corresponding tail entity, and the relationship embedding between each entity and the corresponding tail entity constitute the feature representation of each entity in the knowledge graph.

[0015] Furthermore, the entity category-sensitive first network model built based on the Transformer model is used to integrate the multi-category connections between the feature representations of each entity to obtain the entity feature representations of drug and microbial entities, including:

[0016] Concatenate the attribute vectors and category feature vectors of each entity to obtain attribute embeddings of each entity;

[0017] Concatenating the attribute vector of each tail entity with the category feature vector of the corresponding entity to obtain the attribute embedding of each tail entity;

[0018] Determining a query vector of the first network model according to the attribute embedding of each entity and the relationship embedding;

[0019] Obtaining a key vector of the first network model according to the attribute embedding of each tail entity;

[0020] Obtaining entity feature representation of the first network model according to the query vector, the key vector, and the attribute embedding value vector of each tail entity;

[0021] The first network model is trained according to a first loss function, and entity feature representations of the entities are obtained using the first network model that has been trained completely.

[0022] Furthermore, the first loss function L co The calculation formula is:

[0023]

[0024] Among them, λ is a parameter used to balance the cross entropy loss and TransE loss, represents the set of training set triplets, y (h,r,y) represents the triple label, p (h,r,t) represents the predicted value of the triple, f (h,r,t) Represents the l2 norm of the triple.

[0025] Further, according to the data source, a second network model with a meta-path semantic feature learning strategy with recursive gating is used to encode the contextual connections and diverse adjacency semantics between multiple meta-paths, and learn the semantic representation of nodes, including:

[0026] Construct multiple metapaths, including a metapath with drugs as the core and a metapath with microorganisms as the core; wherein the metapath with drugs as the core includes: a metapath that expresses the similarity semantics between two drugs at the biological level Metapaths that express similarity semantics of indirectly connected drugs at the biological level Metapathways expressing multiple microbes that are biologically co-associated with either drug The meta-paths centered on microorganisms include: meta-paths that express the semantic similarity between two microorganisms at the biological level Metapaths that express the similarity semantics of indirectly connected microorganisms at the biological level Metapathways expressing multiple drugs that are biologically co-associated with any microorganism

[0027] aggregating neighbor information of the plurality of meta-paths according to the data source to obtain biological semantic representations of the corresponding meta-paths respectively;

[0028] The biological semantic representation of the meta-path is input into the second network model to learn and obtain a semantic representation matrix.

[0029] Furthermore, the entity feature representations of the drug and microorganism entities and the semantic representations of the nodes are input into a third network model that is sensitive to node categories to enhance the consistency of node features, thereby obtaining a drug-microorganism association model, including:

[0030] Inputting entity feature representations of the entities and semantic representations of the nodes into the third network model;

[0031] Calculate an adaptive node category-sensitive second loss function through the third network model;

[0032] According to the second loss function, the entity feature representation and semantic representation matrix of each entity are iteratively updated by back propagation of the third network model until a preset number of iterations is reached, thereby obtaining a drug-microorganism association model that enhances node feature consistency.

[0033] Furthermore, the second loss function is expressed as follows:

[0034]

[0035] Among them, N drug and N micr are the number of drugs and microorganisms, respectively, sim(·) is the cosine similarity function, τ represents the temperature parameter, and are randomly generated and normalized learnable parameters, Represent the entity feature representation and semantic representation matrix of the entity respectively.

[0036] The present invention also provides a drug and microorganism association prediction system based on node category-sensitive knowledge graph learning and gated enhanced meta-path semantic fusion, comprising:

[0037] A knowledge graph construction module is used to construct a knowledge graph based on a data source, wherein entities of the knowledge graph are classified into drugs and microorganisms, and feature representations of each entity are obtained based on the category feature vector and attribute vector of each entity;

[0038] A feature enhancement module is used to integrate multiple types of connections between feature representations of entities using the entity category-sensitive first network model built based on the Transformer model to obtain entity feature representations of drug and microbial entities;

[0039] A meta-path construction module is used to construct multiple meta-paths with drugs and microorganisms as nodes according to the data source, and encode the contextual connections and adjacency semantics between the multiple meta-paths through a second network model with a meta-path semantic feature learning strategy with recursive gating to obtain semantic representations of the nodes;

[0040] A contrastive learning module, used for inputting the entity feature representations of the drug and microorganism entities and the semantic representations of the nodes into a third network model that is sensitive to node categories, enhancing the consistency of node features, and obtaining a drug-microorganism association model;

[0041] The association prediction module is used to input the drugs and microorganisms whose association relationships are to be predicted into the drug-microorganism association model to complete the association prediction.

[0042] The present invention also provides an electronic device, including a processor and a memory, wherein a computer program is stored in the memory, and when the computer program is executed by the processor, the drug and microorganism association prediction method based on node category-sensitive knowledge graph learning and gated enhanced meta-path semantic fusion as described in any of the above technical solutions is implemented.

[0043] The present invention also provides a computer-readable storage medium, which stores a computer program. When the computer program is executed by a processor, it implements the drug and microorganism association prediction method based on node category-sensitive knowledge graph learning and gated enhanced meta-path semantic fusion as described in any of the above technical solutions.

[0044] Compared with the prior art, the advantages and beneficial effects of the present invention are as follows:

[0045] 1. The present invention constructs a knowledge graph that includes drug entities and microbial entities, which contains the relationship between the similarity of microorganisms and drugs and the fact that microorganisms are involved in the drug process. The knowledge graph can clearly represent entities, concepts and their relationships, and can capture complex connections through similarity and association relationships between entities, thereby enhancing the model's understanding ability.

[0046] 2. The present invention integrates multi-category connections between head and tail entities through the entity category-sensitive knowledge graph Transformer (ECST model). This strategy takes into account entity heterogeneity and the close connection between head entity features and multi-category tail entity features.

[0047] 3. Multiple metapaths were established, which reflect the biological characteristics of microorganisms that are more likely to be associated with drugs. Metapath-specific semantic feature learning strategies were designed to make it possible to encode and integrate diverse adjacency semantics. The semantics of multiple metapaths are contextually related, and the connections between multiple metapaths are integrated through a metapath semantic feature learning strategy with recursive gating (GUMP model).

[0048] 4. A node category-sensitive ratio learning strategy (CSCL model) is used to enhance the consistency between multi-angle features of target drug (microorganism) nodes.

[0049] In the process of predicting the microorganism-drug association task, the present invention makes full use of the heterogeneity and diversity of entities from the perspective of knowledge graphs, and can effectively capture the complex association relationship between microorganism and drug nodes. In addition, it also mines the contextual connection between multiple meta-paths from the perspective of meta-paths to construct a more comprehensive semantic representation. BRIEF DESCRIPTION OF THE DRAWINGS

[0050] Figure 1 A schematic diagram of the process of the drug and microorganism association prediction method based on node category-sensitive knowledge graph learning and gated enhanced meta-path semantic fusion provided by the present invention;

[0051] Figure 2 A graphical data flow diagram provided by the present invention;

[0052] Figure 3 A schematic diagram of data processing for the drug and microorganism association prediction method provided by the present invention;

[0053] Figure 4 A schematic diagram of the structure of the drug and microorganism association prediction system based on node category-sensitive knowledge graph learning and gated enhanced meta-path semantic fusion provided by the present invention;

[0054] Figure 5 The present invention provides a schematic structural diagram of an electronic device. DETAILED DESCRIPTION

[0055] The preferred embodiments of the present invention are described in detail below in conjunction with the accompanying drawings, wherein the accompanying drawings constitute a part of this application and are used together with the embodiments of the present invention to illustrate the principles of the present invention, but are not used to limit the scope of the present invention.

[0056] The present invention provides a method for predicting drug and microbial associations based on node category-sensitive knowledge graph learning and gated enhanced meta-path semantic fusion, which is described in detail below.

[0057] See also Figure 1 This embodiment provides a method for predicting drug and microbial association based on node category-sensitive knowledge graph learning and gated enhanced meta-path semantic fusion, including:

[0058] Step S101: construct a knowledge graph according to the data source, wherein the entities of the knowledge graph are classified into drugs and microorganisms, and a feature representation of each entity is obtained according to the category feature vector and attribute vector of each entity;

[0059] Step S102: using the entity category-sensitive first network model built based on the Transformer model, integrating multiple types of connections between feature representations of each entity to obtain entity feature representations of drug and microbial entities;

[0060] Step S103: constructing multiple meta-paths with drugs and microorganisms as nodes according to the data source, encoding the contextual connections and adjacency semantics between the multiple meta-paths through a second network model with a meta-path semantic feature learning strategy with recursive gating, and obtaining semantic representations of the nodes;

[0061] Step S104: inputting the entity feature representations of the drug and microorganism entities and the semantic representations of the nodes into a third network model that is sensitive to node categories, enhancing the consistency of node features, and obtaining a drug-microorganism association model;

[0062] Step S105: inputting the drugs and microorganisms whose association relationships are to be predicted into the drug-microorganism association model to complete association prediction.

[0063] The method of this embodiment makes full use of the heterogeneity and diversity of entities from the perspective of the knowledge graph in the process of the microorganism-drug association prediction task, and can effectively capture the complex association relationship between microorganism and drug nodes. In addition, this method also starts from the perspective of the meta-path, explores the contextual connections between multiple meta-paths, and constructs a more comprehensive semantic representation. Compared with previous methods, the present invention pays special attention to integrating entity features from the perspective of the knowledge graph and node semantic features from the perspective of the meta-path, and ensures the consistency of the two through a comparative learning strategy. This consistency construction can not only improve the accuracy of association prediction, but also effectively solve the problem of feature distribution imbalance faced by traditional methods when processing multi-perspective features.

[0064] As a preferred embodiment, in step S101, the feature representation of each entity is obtained according to the category feature vector and the attribute vector of each entity, including:

[0065] According to the similarity matrix between drugs, the similarity matrix between microorganisms, and the association matrix between microorganisms and drugs, the tail entities corresponding to the entities and the relationship embedding between the head entity of any entity and the corresponding tail entity are obtained;

[0066] The attribute vector of each entity, the category feature vector, the attribute vector of the corresponding tail entity, and the relationship embedding between each entity and the corresponding tail entity constitute the feature representation of each entity in the knowledge graph.

[0067] As a preferred embodiment, in step S102, a first network model sensitive to entity categories constructed based on a Transformer model is used to integrate multiple types of connections between feature representations of various entities to obtain entity feature representations of drug and microbial entities, including:

[0068] Concatenate the attribute vectors and category feature vectors of each entity to obtain attribute embeddings of each entity;

[0069] Concatenating the attribute vector of each tail entity with the category feature vector of the corresponding entity to obtain the attribute embedding of each tail entity;

[0070] Determining a query vector of the first network model according to the attribute embedding of each entity and the relationship embedding;

[0071] Obtaining a key vector of the first network model according to the attribute embedding of each tail entity;

[0072] Obtaining an entity feature vector of the first network model according to the query vector, the key vector, and the attribute embedding value vector of each tail entity;

[0073] The first network model is trained according to a first loss function, and entity feature representations of the entities are obtained using the first network model that has been trained completely.

[0074] The following is a specific example of the detailed process of establishing the knowledge graph consisting of microorganisms and drug entities and learning the characteristics of each entity through the entity category-sensitive first network model built based on the Transformer model.

[0075] First, a knowledge graph is constructed based on the data source. The data source here refers to the prior data provided by various experimental records, literature, databases, etc. that study the relationship between drugs and microorganisms, the similarity of drug chemical substructures, and the similarity of microorganisms.

[0076] The constructed knowledge graph contains two entities, microorganisms and drugs, and the attributes of the two types of entities are heterogeneous. However, previous methods did not fully utilize the information of entity types. But in reality, there is a close connection between the attributes of drug or microorganism entities and their tail entities, so we designed an entity category-sensitive Transformer (i.e., the first network model, hereinafter abbreviated as ECST) to encode these connections.

[0077] Drug Entity i Take the update process of the relationship feature of the entity as an example. The adjacency matrix of the microorganism-drug node is B,

[0078]

[0079] Among them, I drug (or I micr ) represents the drug (or microorganism) similarity matrix, A drug-micr represents the microorganism-drug association matrix, A T represents the transpose of A, I drug The i-th row of , which contains drug d i The combined similarity with all drugs, Drugs recorded i Is there a link with microorganisms? represents the i-th row of B, which is composed of and composed of, Middle and d i Most similar N i drugs and all M associated with them i Microorganisms form k tail entity, φ k =M i +N i .

[0080] According to the entity and the corresponding tail entity, we get the set of triples:

[0081]

[0082] Among them, r j Represents drug entity d i To the end entity t j similarities or connections between them.

[0083] Considering d i heterogeneity, we will i The type of the character vector Connect to its attribute vector Entity di Attribute embedding

[0084]

[0085] Among them, || represents splicing, X and They are the entity attribute matrix and entity type feature matrix respectively. and is randomly initialized and is learnable during training, N e and N t They represent the dimensions of the attribute vector and the type vector respectively.

[0086] Similarly, the tail entity t j Attribute embedding By connecting j The attribute vector and type feature vector get,

[0087]

[0088] in, and is a learnable vector.

[0089] Through the above processing, we obtain the feature representation of each entity in the knowledge graph, including the attribute vector of each entity, the category feature vector, the attribute vector of the corresponding tail entity, and the relationship embedding between each entity and the corresponding tail entity.

[0090] The first network model ECST is built based on the Transformer model. In the Transformer model, Q, K, and V vectors are the core components of the self-attention mechanism. They are used to calculate the relationship and importance between each position in the sequence. The relationship between Q, K, and V vectors and the loss function of the model is realized through the self-attention mechanism and back propagation. Specifically, Q, K, and V vectors affect the calculation of self-attention, thereby affecting the output of the model and the final loss function.

[0091] We will entity d i The attribute vector and t j Relational embedding Fusion to form a query vector

[0092]

[0093] in, is the dimension of the query vector, WQ is the weight matrix, is randomly initialized, it is a learnable vector, N e Dimension representing the relation embedding.

[0094] key vector By embedding the attributes of the tail entity Performing linear transformation, we get

[0095]

[0096] Among them, W K represents the weight matrix.

[0097] Through the above and The product of the transpositions forms the tail entity t j Right i The attention score is α j , then φ k Attention score of the tail entity

[0098] φ k The feature matrix of the tail entity is obtained by stacking up and down After linear transformation, we get the value matrix

[0099]

[0100] Among them, W V is the weight matrix.

[0101] Will Normalized and multiplied by V to form d i The enhanced feature vector of

[0102]

[0103] Furthermore, in order to improve the expressiveness of the ECST model, due to d i The original attribute vector Contains more details, so and is formed by residual connection

[0104]

[0105] Similarly, we can obtain the relationship representation of all drug and microbial entities and form the entity feature matrix X kg , which is the entity feature representation of the entity.

[0106] As a preferred embodiment, the first loss function L co The calculation formula is:

[0107]

[0108] Among them, Ω represents the set of training set triplets, λ represents the parameter used to balance the cross entropy loss and TransE loss, λ is generally set to 0.1, y (h,r,t) represents the triple label, p (h,r,t) represents the predicted value of the triple, f (h,r,t) Represents the l2 norm of the triple.

[0109] On the other hand, since there are similarities and associated connection semantics between drugs and microorganisms, and these semantics can reflect the biological significance of drug and microorganism nodes. As a preferred embodiment, in step S103, according to the data source, the contextual connections and diverse adjacency semantics between multiple meta-paths are encoded through a second network model with a meta-path semantic feature learning strategy with recursive gating, and the semantic representation of the learning node is learned, including:

[0110] Construct multiple metapaths, including a metapath with drugs as the core and a metapath with microorganisms as the core; wherein the metapath with drugs as the core includes: a metapath that expresses the similarity semantics between two drugs at the biological level Metapaths that express similarity semantics of indirectly connected drugs at the biological level Metapathways expressing multiple microbes that are biologically co-associated with either drug The meta-paths centered on microorganisms include: meta-paths that express the semantic similarity between two microorganisms at the biological level Metapaths that express the similarity semantics of indirectly connected microorganisms at the biological level Metapathways expressing multiple drugs that are biologically co-associated with any microorganism

[0111] aggregating neighbor information of the plurality of meta-paths according to the data source to obtain biological semantic representations of the corresponding meta-paths respectively;

[0112] The biological semantic representation of the meta-path is input into the second network model to learn and obtain a semantic representation matrix.

[0113] The second network model (GUMP model) with a meta-path semantic feature learning strategy with recursive gating is used to learn meta-path semantic features. and The biological semantic learning process is similar. As a specific example, we use the metapath For example, describe the drug d i The learning process of biological semantic representation.

[0114] For meta path From any drug i Departure from the node d it can reach j The node neighbor set is Target drug i Similarity vector between its second-order drug neighbors It can be expressed as:

[0115]

[0116] Among them, N drug is the amount of drug, Contains the similarity of all first-order drug neighbors. Based on metapath Aggregating the meta-path neighbors of all second-order drugs, we obtain the semantic representation under this meta-path

[0117]

[0118] Among them, B drug The first N of B drug OK, is the drug feature matrix, N micr Indicates the number of microorganisms.

[0119] In order to get i For more representative feature vectors, we target the meta-path Establish its unique fully connected neural network. is input into the neural network to learn d i The low-dimensional representation of

[0120]

[0121] in, and are the weight matrix and bias vector respectively, N ph Represents the dimension of the feature space.

[0122] Will and as well as and As the input of the second network model, in order to fuse d i Three biological semantic representations of and It is stacked in the spatial dimension and its channel number is increased to 2C through 1×1 convolution.

[0123] Z is divided into C / 4, C / 4, C / 2 and C at the channel level to obtain and We encode via recursive gated convolution and Contextual connection. After the gated convolution is recursively performed, the contextual connection enhanced d i The eigenvector of

[0124]

[0125] in, and are the weight matrix and the bias vector respectively. After three meta-path contextual enhancements, d i The semantic representation vector of is,

[0126]

[0127] Among them, W c and b c are the weight matrix and bias vector respectively.

[0128] With d i The semantic representation learning process is similar, and we can obtain the semantic representation matrix P of all drug and microbial nodes. se .

[0129] As a preferred embodiment, the entity feature representation of the drug and microorganism entity and the semantic representation of the node are input into a node category-sensitive third network model (CSCL model) to enhance the consistency of node features and obtain a drug-microorganism association model, including:

[0130] Inputting entity feature representations of the entities and semantic representations of the nodes into the third network model;

[0131] Calculate an adaptive node category-sensitive second loss function through the third network model;

[0132] According to the second loss function, the entity feature representation and semantic representation matrix of each entity are iteratively updated by back propagation of the third network model until a preset number of iterations is reached, thereby obtaining a drug-microorganism association model that enhances node feature consistency.

[0133] As a preferred embodiment, the second loss function is expressed by the formula:

[0134]

[0135] Among them, N drug and N micr are the number of drugs and microorganisms, respectively, sim(·) is the cosine similarity function, τ represents the temperature parameter, and are randomly generated and normalized learnable parameters, Represent the entity feature representation and semantic representation matrix of the entity respectively.

[0136] like Figure 2 As shown, Figure 2 A complete schematic flow chart of the method of the present invention is shown.

[0137] In order to more clearly illustrate the complete processing flow of the present invention, Figure 3 The actual computer implementation process is demonstrated, and the specific steps are as follows:

[0138] S1, input original data and the first, second and third network models f ECST ,f GUMP ,f CSCL ;

[0139] S2. Build a knowledge graph, including the head entity d i , relationship j and tail entity t j , and obtain the same i Most similar N i drugs and all M associated with them i Microorganisms form k tail entity, φ k =N i +M i ;

[0140] S3, initialize feature vector, including entity original embedding Relation Embedding and entity category embedding

[0141] S4, start training according to the given number of iterations T, initialize the initial value t = 0, the original embedding

[0142] S5. Embed the original header entity Relation Embedding and φ k The original embedding of the tail entity and entity category embedding As a machine learning model f ECST Input, calculate the first loss function L co , and then back-propagation is performed to update the feature vector of the target head entity in Represents the target head entity d when the number of training rounds is t i The eigenvector of

[0143] S6. End the loop after training T times and obtain the entity features of the target head Similarly, get the feature matrix X of all head entities kg ;

[0144] S7. Construct multiple meta-pathways, including drug-centered and microorganism-centered

[0145] S8. Take the original data, including the drug similarity matrix D, the microbial similarity matrix M and the drug-microbial association matrix A, as input, aggregate the neighbor information according to the constructed multiple meta-paths, and obtain the biological semantic representation of the corresponding meta-paths, including and as well as and where i=1,2,…,N drug and j = 1, 2, ..., N micr ;

[0146] S9, will and as well as and As machine learning GUMP Input, learn the semantic representation matrix P se ;

[0147] S10, the entity feature matrix X kg and the semantic representation matrix P se As machine learning CSCL Input, calculate the adaptive node category sensitive second loss function L cl , and then back propagate to update the entity feature matrix X kg and the semantic representation matrix P se ;

[0148] S11. After T rounds of training, the entity feature matrix X is returned. kg and the semantic representation matrix P se .

[0149] This method not only learns the entity features and node semantics of drugs and microorganisms from the knowledge graph perspective and meta-path perspective, but also balances the feature information from multiple perspectives through a node category-sensitive contrastive learning strategy.

[0150] In order to verify the actual effect of the present invention, as shown in Table 1, firstly, an ablation experiment was performed using a data set consisting of 1,373 drugs to calculate the AUC and AUPR values ​​on this data set to evaluate the performance of the model in the drug prediction task.

[0151] Ablation experiments can help us systematically evaluate the role of each component or module in the model. ECST, GUMP, and CSCL in Table 1 represent the first, second, and third network models in the above scheme, respectively. From the experimental results in Table 1, it can be seen that ECST contributes the most to the improvement of model performance. The reason is that the entity category-sensitive Transformer model we designed can effectively integrate the dependencies between multiple types of entity attributes.

[0152] Table 1 Comparison of ablation experiment results

[0153]

[0154] Table 2 shows the comparison between the association model proposed in the present invention (hereinafter abbreviated as KNDM) and five advanced drug-microorganism association prediction methods, including GCNMD, EGATMDA, GSAMDA, SCSMDA, GACNNMDA, NGMDA, MFLP and DHDMP.

[0155] Table 2 Comparison of prediction results of KNDM and other models

[0156]

[0157] It can be seen that KNDM achieved the best average AUC (AUC = 0.969), which is 6.6%, 2.9%, 6.7%, 5.3%, 12.6%, 2.5%, 2.3% and 1.0% higher than GCNMDA, EGATMDA, GACNNMDA, GSAMDA, SCSMDA, NGMDA, MFLP and DHDMP respectively. KNDM achieved the highest average AUPR of 86.9%, which is 4.6% higher than the second best DHDMP, 67.3% better than the worst GACNNMDA, and 55.4%, 56.2%, 62.2%, 52.9%, 14.1% and 10.2% higher than GCNMDA, EGATMDA, GSAMDA, SCSMDA, NGMDA and MFLP respectively.

[0158] Experimental results show that the improvement of KNDM prediction performance is mainly due to encoding the semantic connection of meta-path context, integrating the close connection between entity features, and enhancing the consistency of features under multiple views.

[0159] Furthermore, we continued to compare the results of the Wilcoxon test using KNDM and other models, and compared the AUCs and AUPRs of 1,373 drugs. The experimental results are shown in Table 3. From Table 3, we can see that the performance of KNDM is significantly higher than that of the compared methods, because the p-value is less than 0.05.

[0160] Table 3 Comparison of Wilcoxon test results

[0161]

[0162] Compared with traditional association prediction methods, the present invention can further demonstrate the application of our strategy in screening potential drug-related microbial candidates through case studies of curcumin, green tea extract and ciprofloxacin. This strategy makes full use of the heterogeneity and diverse facts of entities from the perspective of knowledge graphs, as well as the contextual connections between multiple meta-paths from the perspective of meta-paths. In addition, previous methods do not integrate the consistency of entity features from the perspective of knowledge graphs and node semantic features from the perspective of meta-paths.

[0163] like Figure 4 As shown, the embodiment of the present invention further provides a drug and microorganism association prediction system 400 based on node category-sensitive knowledge graph learning and gated enhanced meta-path semantic fusion, including:

[0164] The knowledge graph construction module 401 is used to construct a knowledge graph according to the data source, wherein the entities of the knowledge graph are classified into drugs and microorganisms by category, and the feature representation of each entity is obtained according to the category feature vector and attribute vector of each entity;

[0165] A feature enhancement module 402 is used to integrate multiple types of connections between feature representations of entities using a first network model that is sensitive to entity categories and is constructed based on a Transformer model, so as to obtain entity feature representations of drug and microbial entities;

[0166] A meta-path construction module 403 is used to construct multiple meta-paths with drugs and microorganisms as nodes according to the data source, and encode the contextual connections and adjacency semantics between the multiple meta-paths through a second network model with a meta-path semantic feature learning strategy with recursive gating to obtain semantic representations of the nodes;

[0167] A comparative learning module 404 is used to input the entity feature representations of the drug and microorganism entities and the semantic representations of the nodes into a third network model that is sensitive to node categories, enhance the consistency of node features, and obtain a drug-microorganism association model;

[0168] The association prediction module 405 is used to input the drugs and microorganisms whose association relationships are to be predicted into the drug-microorganism association model to complete the association prediction.

[0169] like Figure 5 As shown, the above-mentioned drug and microorganism association prediction method based on node category sensitive knowledge graph learning and gated enhanced meta-path semantic fusion, the present invention also provides an electronic device 500 accordingly, which can be a computing device such as a mobile terminal, a desktop computer, a notebook, a palm computer and a server. The electronic device includes a processor 501, a memory 502 and a display 503.

[0170] In some embodiments, the memory 502 may be an internal storage unit of the computer device, such as a hard disk or memory of the computer device. In other embodiments, the memory 502 may also be an external storage device of the computer device, such as a plug-in hard disk, a smart media card (SMC), a secure digital (SD) card, a flash card, etc. equipped on the computer device. Further, the memory 502 may also include both an internal storage unit of the computer device and an external storage device. The memory 502 is used to store application software and various types of data installed in the computer device, such as program codes installed in the computer device. The memory 502 may also be used to temporarily store data that has been output or is to be output. In one embodiment, a drug and microorganism association prediction method program 504 based on node category-sensitive knowledge graph learning and gated enhanced meta-path semantic fusion is stored on the memory 502. The drug and microorganism association prediction method program 504 based on node category-sensitive knowledge graph learning and gated enhanced meta-path semantic fusion can be executed by the processor 501, thereby realizing a drug and microorganism association prediction method based on node category-sensitive knowledge graph learning and gated enhanced meta-path semantic fusion according to each embodiment of the present invention.

[0171] In some embodiments, the processor 501 can be a central processing unit (CPU), a microprocessor or other data processing chip, which is used to run the program code or process data stored in the memory 502, such as executing a drug and microorganism association prediction method program based on node category-sensitive knowledge graph learning and gated enhanced meta-path semantic fusion.

[0172] In some embodiments, the display 503 may be an LED display, a liquid crystal display, a touch-sensitive liquid crystal display, an OLED (Organic Light-Emitting Diode) touch device, etc. The display 503 is used to display information on the computer device and to display a visual user interface. The components 501-503 of the computer device communicate with each other through a system bus.

[0173] This embodiment also provides a computer-readable storage medium, which stores a computer program. When the computer program is executed by a processor, it implements the drug and microorganism association prediction method based on node category-sensitive knowledge graph learning and gated enhanced meta-path semantic fusion as described in any of the above technical solutions.

[0174] The computer-readable storage medium and computing device provided according to the above embodiments of the present invention can be implemented with reference to the specific description of the method for predicting drug and microorganism association based on node category-sensitive knowledge graph learning and gated enhanced meta-path semantic fusion as described above according to the present invention, and have similar beneficial effects as the method for predicting drug and microorganism association based on node category-sensitive knowledge graph learning and gated enhanced meta-path semantic fusion as described above, which will not be repeated here.

[0175] In summary, the drug and microorganism association prediction method based on node category-sensitive knowledge graph learning and gated enhanced meta-path semantic fusion provided by the present invention firstly establishes a knowledge graph composed of microorganism and drug entities, secondly, designs an entity category-sensitive Transformer (ECST) to learn the features of each entity, and thirdly, designs a meta-path semantic feature learning model (GUMP) with recursive gating, and learns the biological semantics unique to the meta-path through a meta-path-specific semantic feature learning strategy. For the target drug (or microorganism) node, there is a contextual connection between the features learned from multiple meta-paths. Finally, a node category-sensitive contrastive learning strategy (CSCL) is designed to adaptively enhance the consistency of entity features and semantic features.

[0176] By integrating entity features from the knowledge graph perspective and node semantic features from the meta-path perspective, and using a contrastive learning strategy to ensure the consistency between the two, this consistency construction can not only improve the accuracy of association prediction, but also effectively solve the problem of feature distribution imbalance faced by traditional methods when dealing with multi-view features.

[0177] The above description is only a preferred specific implementation manner of the present invention, but the protection scope of the present invention is not limited thereto. Any changes or substitutions that can be easily conceived by any technician familiar with the technical field within the technical scope disclosed by the present invention should be covered within the protection scope of the present invention.

Claims

1. A drug and microbial association prediction method based on node category-sensitive knowledge graph learning and gated enhanced meta-path semantic fusion, characterized in that: include: A knowledge graph is constructed based on the data source, wherein entities of the knowledge graph are classified into drugs and microorganisms, and feature representations of each entity are obtained based on the category feature vector and attribute vector of each entity; The entity category-sensitive first network model built based on the Transformer model is used to integrate the multi-category connections between the feature representations of each entity to obtain the entity feature representations of drug and microbial entities. According to the data source, a plurality of meta-paths are constructed with drugs and microorganisms as nodes respectively, and contextual connections and adjacency semantics between the plurality of meta-paths are encoded through a second network model with a meta-path semantic feature learning strategy with recursive gating to obtain semantic representations of the nodes; Inputting the entity feature representations of the drug and microorganism entities and the semantic representations of the nodes into a third network model that is sensitive to node categories, enhancing the consistency of node features, and obtaining a drug-microorganism association model; The drugs and microorganisms whose association relationships are to be predicted are input into the drug-microorganism association model to complete the association prediction.

2. The drug and microbial association prediction method based on node category-sensitive knowledge graph learning and gated enhanced meta-path semantic fusion according to claim 1 is characterized in that: The feature representation of each entity is obtained according to the category feature vector and the attribute vector of each entity, including: According to the similarity matrix between drugs, the similarity matrix between microorganisms, and the association matrix between microorganisms and drugs, the tail entities corresponding to the entities and the relationship embedding between the head entity of any entity and the corresponding tail entity are obtained; The attribute vector of each entity, the category feature vector, the attribute vector of the corresponding tail entity, and the relationship embedding between each entity and the corresponding tail entity constitute the feature representation of each entity in the knowledge graph.

3. The drug and microbial association prediction method based on node category-sensitive knowledge graph learning and gated enhanced meta-path semantic fusion according to claim 2 is characterized in that: The entity category-sensitive first network model built based on the Transformer model is used to integrate the multi-category connections between the feature representations of each entity to obtain the entity feature representations of drug and microbial entities, including: Concatenate the attribute vectors and category feature vectors of each entity to obtain attribute embeddings of each entity; Concatenating the attribute vector of each tail entity with the category feature vector of the corresponding entity to obtain the attribute embedding of each tail entity; Determining a query vector of the first network model according to the attribute embedding of each entity and the relationship embedding; Obtaining a key vector of the first network model according to the attribute embedding of each tail entity; Obtaining an entity feature representation value vector of the first network model according to the query vector, the key vector, and the attribute embedding value vector of each tail entity; The first network model is trained according to a first loss function, and entity feature representations of the entities are obtained using the first network model that has been trained completely.

4. The drug and microbial association prediction method based on node category-sensitive knowledge graph learning and gated enhanced meta-path semantic fusion according to claim 3 is characterized in that: The first loss function L co The calculation formula is: Among them, λ is a parameter used to balance the cross entropy loss and TransE loss, represents the set of training set triplets, y (h,r,t) represents the triple label, p (h,r,t) represents the predicted value of the triple, f (h,r,t) Represents a triple Norm.

5. The drug and microbial association prediction method based on node category-sensitive knowledge graph learning and gated enhanced meta-path semantic fusion according to claim 1 is characterized in that: According to the data source, a second network model with a meta-path semantic feature learning strategy with recursive gating is used to encode contextual connections and diverse adjacency semantics between multiple meta-paths, and learn semantic representations of nodes, including: Construct multiple metapaths, including a metapath with drugs as the core and a metapath with microorganisms as the core; wherein the metapath with drugs as the core includes: a metapath that expresses the similarity semantics between two drugs at the biological level Metapaths that express similarity semantics of indirectly connected drugs at the biological level Metapathways expressing multiple microbes that are biologically co-associated with either drug The meta-paths with microorganisms as the core include: meta-paths that express the semantic similarity between two microorganisms at the biological level Metapaths that express the similarity semantics of indirectly connected microorganisms at the biological level Metapathways expressing multiple drugs that are biologically co-associated with any microorganism aggregating neighbor information of the plurality of meta-paths according to the data source to obtain biological semantic representations of the corresponding meta-paths respectively; The biological semantic representation of the meta-path is input into the second network model to learn and obtain a semantic representation matrix.

6. The drug and microbial association prediction method based on node category-sensitive knowledge graph learning and gated enhanced meta-path semantic fusion according to claim 1 is characterized in that: The entity feature representations of the drug and microorganism entities and the semantic representations of the nodes are input into a third network model that is sensitive to node categories, the consistency of node features is enhanced, and a drug-microorganism association model is obtained, including: Inputting entity feature representations of the entities and semantic representations of the nodes into the third network model; Calculate an adaptive node category-sensitive second loss function through the third network model; According to the second loss function, the entity feature representation and semantic representation matrix of each entity are iteratively updated by back propagation of the third network model until a preset number of iterations is reached, thereby obtaining a drug-microorganism association model that enhances node feature consistency.

7. The method for drug and microbial association prediction based on node category-sensitive knowledge graph learning and gated enhanced meta-path semantic fusion according to claim 6 is characterized in that: The second loss function is expressed as: Among them, N drug and N micr are the number of drugs and microorganisms, respectively, sim(·) is the cosine similarity function, τ represents the temperature parameter, and are randomly generated and normalized learnable parameters, Represent the entity feature representation and semantic representation matrix of the entity respectively.

8. A drug and microbial association prediction system based on node category-sensitive knowledge graph learning and gated enhanced meta-path semantic fusion, characterized in that: include: A knowledge graph construction module is used to construct a knowledge graph based on a data source, wherein entities of the knowledge graph are classified into drugs and microorganisms, and feature representations of each entity are obtained based on the category feature vector and attribute vector of each entity; A feature enhancement module is used to integrate multiple types of connections between feature representations of entities using the entity category-sensitive first network model built based on the Transformer model to obtain entity feature representations of drug and microbial entities; A meta-path construction module is used to construct multiple meta-paths with drugs and microorganisms as nodes according to the data source, and encode the contextual connections and adjacency semantics between the multiple meta-paths through a second network model with a meta-path semantic feature learning strategy with recursive gating to obtain semantic representations of the nodes; A contrastive learning module, used for inputting the entity feature representations of the drug and microorganism entities and the semantic representations of the nodes into a third network model that is sensitive to node categories, enhancing the consistency of node features, and obtaining a drug-microorganism association model; The association prediction module is used to input the drugs and microorganisms whose association relationships are to be predicted into the drug-microorganism association model to complete the association prediction.

9. An electronic device, characterized in that: It includes a processor and a memory, wherein a computer program is stored in the memory, and when the computer program is executed by the processor, a drug and microorganism association prediction method based on node category-sensitive knowledge graph learning and gated enhanced meta-path semantic fusion as described in any one of claims 1 to 7 is implemented.

10. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores a computer program, and when the computer program is executed by the processor, the method for predicting drug and microbial association based on node category-sensitive knowledge graph learning and gated enhanced meta-path semantic fusion as described in any one of claims 1-7 is implemented.

Images