Training methods for drug-disease association prediction models

Abstract

This application proposes a training method for a drug-disease association prediction model, comprising: obtaining a training sample set; for any node among all drugs and all diseases involved in the training sample set, obtaining a fusion feature obtained by fusing the network topology features and similarity features of the node; based on the training sample set and the fusion feature of each node in the training sample set, performing semi-supervised adversarial training using a classifier and a discriminator until the classifier iterates to meet a predetermined prediction capability condition, at which point the classifier is determined as the drug-disease association prediction model. The technical solution of this application can construct a high-precision, high-generalization-ability drug-disease association prediction model.

Description

Technical Field

[0001] This application relates to the field of biomedical technology, and in particular to a training method for a drug-disease association prediction model. Background Technology

[0002] Drug development is a key driver for improving public health and advancing clinical medicine. Current drug development often involves a lengthy process, including target identification, virtual or high-throughput screening, preclinical research, and clinical trials. To address this, drug repositioning can be used to explore new indications for existing drugs. Therefore, drug repositioning plays an indispensable role in both biomedical research and new drug development.

[0003] While current drug relocation research has made some progress, its predictive methods still have significant limitations. On the one hand, most predictive methods rely solely on single similarity data, failing to fully integrate the multi-source characteristics of drugs and diseases in heterogeneous biological networks, resulting in insufficient depth in uncovering the potential associations between drugs and diseases. On the other hand, due to the limited amount of known association data and the large number of unlabeled samples with unknown associations, the generalization ability of predictive models in practical applications is limited.

[0004] Therefore, how to further improve the accuracy of predicting the association between drugs and diseases and enhance the generalization ability of drug-disease association prediction models has become an urgent technical problem to be solved. Summary of the Invention

[0005] This application provides a training method for a drug-disease association prediction model, aiming to solve the technical problems of insufficient accuracy of prediction results and limited generalization ability of drug-disease association prediction models in related technologies.

[0006] In a first aspect, embodiments of this application provide a method for training a drug-disease association prediction model, comprising:

[0007] Obtain a training sample set, wherein the training sample set includes a positive sample set and an unlabeled sample set, the positive sample set includes known drug-disease pairs that are associated, and the unlabeled sample set includes drug-disease pairs that are unknown whether they are associated.

[0008] For any node among all drugs and all diseases involved in the training sample set, obtain the fused feature obtained by fusing the network topology features and similarity features of the node;

[0009] Based on the training sample set and the fusion features of each node within the training sample set, a classifier and discriminator are used for semi-supervised adversarial training until the classifier iterates to meet a predetermined prediction ability condition. At this point, the classifier is determined as the drug-disease association prediction model.

[0010] The classifier is used to predict whether the drug and disease pair are related based on the fusion features of the drug and the fusion features of the disease in the drug and disease pair, and the discriminator is used to determine whether the drug and disease pair predicted by the classifier as related is a positive sample.

[0011] In one embodiment of this application, optionally, obtaining the fused feature obtained by fusing the network topology features and similarity features of any node among all drugs and all diseases involved in the training sample set includes:

[0012] Based on graph embedding algorithms, the network topology characteristics of the nodes are determined;

[0013] Based on known drug-disease association data, the similarity characteristics of the nodes are determined;

[0014] The network topology features and similarity features of the nodes are fused using a multi-head attention mechanism to obtain the fused features of the nodes.

[0015] In one embodiment of this application, optionally, determining the network topology features of the node based on the graph embedding algorithm includes:

[0016] Based on the graph embedding algorithm, multiple fixed-length biased random walks are performed in a pre-constructed biological heterogeneous network to obtain a set of random walk paths. The first parameter controls the probability of the walk returning to the previous node, and the second parameter controls the probability of the walk exploring new nodes.

[0017] By learning each random walk path in the set of random walk paths using the Skip-gram model, a 128-dimensional vector is obtained for each node in all drugs and all diseases involved in the training sample set, which serves as the network topology feature of each node.

[0018] Optionally, in one embodiment of this application, before the step of performing multiple fixed-length biased random walks in a pre-constructed biological heterogeneous network based on the graph embedding algorithm, the method further includes:

[0019] Based on known drug-disease association data, the biological heterogeneous network is constructed, wherein the nodes in the biological heterogeneous network include drugs, genes and diseases, and the biological heterogeneous network is used to reflect the interaction between drugs and targets, the association between proteins and diseases, protein-protein interaction relationships, and known associations between drugs and diseases.

[0020] In one embodiment of this application, optionally, determining the similarity features of the nodes based on known drug-disease association data includes:

[0021] For each pair of drugs in the training sample set, the drug similarity between the two drugs is determined based on their GIP kernel similarity and chemical structure similarity. A drug similarity matrix is ​​then constructed based on the drug similarity between each pair of drugs.

[0022] For each pair of diseases in the training sample set, the disease similarity between the two diseases is determined based on the GIP kernel similarity and MeSH similarity between the two diseases, and a disease similarity matrix is ​​constructed based on the disease similarity between each pair of diseases.

[0023] The drug similarity matrix and the disease similarity matrix are processed by a graph autoencoder to obtain the similarity features of each drug and the similarity features of each disease in the training sample set.

[0024] In one embodiment of this application, optionally, the method for determining the GIP kernel similarity of the two drugs includes:

[0025] For the first drug and the second drug in the two drugs, obtain the first interaction pattern vector of the first drug with all genes and the second interaction pattern vector of the second drug with all genes.

[0026] Based on the first interaction pattern vector, the second interaction pattern vector, and the Gaussian kernel function, the GIP kernel similarity of the two drugs is determined, wherein...

[0027] ,

[0028] This represents the GIP kernel similarity between the two drugs, where A represents the first drug and B represents the second drug. and These are the first interaction mode vector and the second interaction mode vector, respectively. It is a parameter used to adjust the bandwidth of the Gaussian kernel function;

[0029] The methods for determining the chemical structural similarity of the two drugs include:

[0030] Based on the MACCS fingerprint vectors of the two drugs respectively, the Tanimoto similarity between the two drugs is determined, which is used as the chemical structural similarity between the two drugs.

[0031] ,

[0032] This represents the Tanimoto similarity between the MACCS fingerprint vector of the first drug and the MACCS fingerprint vector of the second drug. This shows the number of 1s in the MACCS fingerprint vector of the first drug. This represents the number of 1s in the MACCS fingerprint vector of the second drug. This indicates that the MACCS fingerprint vector of the first drug and the MACCS fingerprint vector of the second drug have a total of 1 bits.

[0033] In one embodiment of this application, optionally, the method for determining the GIP kernel similarity of the two diseases includes:

[0034] For the first disease and the second disease, obtain the third interaction pattern vector of the first disease with all genes and the fourth interaction pattern vector of the second disease with all genes.

[0035] Based on the third interaction pattern vector, the fourth interaction pattern vector, and the Gaussian kernel function, the GIP kernel similarity of the two diseases is determined, wherein...

[0036] ,

[0037] This represents the GIP kernel similarity between the two diseases, where C represents the first disease and D represents the second disease. and These are the third interaction mode vector and the fourth interaction mode vector, respectively. It is a parameter used to adjust the bandwidth of the Gaussian kernel function;

[0038] The methods for determining the MeSH similarity of the two diseases include:

[0039] Obtain the first set of descriptors mapped to the first disease in the MeSH database, and obtain the second set of descriptors mapped to the second disease in the MeSH database;

[0040] Based on the first descriptor set and the second descriptor set, the MeSH similarity of the two diseases is determined, wherein if and Having a common ancestor

[0041] ,

[0042] For any first descriptor in the first descriptor set, For any second descriptor in the second descriptor set, express and similarity, yes The information content value in the MeSH tree, yes The information content value in the MeSH tree, represent and The information content value of the common ancestor in the MeSH tree;

[0043] like and No common ancestor It is zero;

[0044] ,

[0045] This indicates the MeSH similarity between the first disease and the second disease. This represents the first set of descriptors. This refers to the second set of descriptors.

[0046] In one embodiment of this application, optionally, determining the drug similarity of the two drugs based on their GIP core similarity and chemical structure similarity includes:

[0047] If the chemical structural similarity between the two drugs is zero, the drug similarity between the two drugs is determined to be the GIP core similarity between the two drugs;

[0048] If the chemical structural similarity of the two drugs is non-zero, the drug similarity of the two drugs is determined to be the average of the GIP core similarity and the chemical structural similarity of the two drugs.

[0049] The determination of disease similarity between the two diseases based on GIP kernel similarity and MeSH similarity includes:

[0050] If the MeSH similarity between the two diseases is zero, the disease similarity between the two diseases is determined to be the GIP kernel similarity between the two diseases;

[0051] If the MeSH similarity between the two diseases is non-zero, the disease similarity between the two diseases is determined to be the average of the GIP kernel similarity and the MeSH similarity between the two diseases.

[0052] In one embodiment of this application, optionally, the step of fusing the network topology features and similarity features of the node through a multi-head attention mechanism to obtain the fused features of the node includes:

[0053] Based on the network topology characteristics of the node, the enhanced network topology characteristics of the node are determined, wherein...

[0054] ,

[0055] H St S represents the enhanced network topology characteristics of the nodes. t S represents the network topology characteristics of the node. t ∈R X×L R represents the set of nodes in a single training iteration, X represents the number of nodes in a single training iteration, L represents the feature vector length of each node, MultiHead represents the multi-head attention mechanism, and W... St and b St represents the weights and biases of the fully connected layer after the network topology features are enhanced by the multi-head attention mechanism, respectively; LayerNorm is the layer normalization function, and LeakyReLU is the non-linear activation function;

[0056] Based on the similarity features of the nodes, the enhanced similarity features of the nodes are determined, wherein,

[0057] ,

[0058] H Si S represents the enhanced similarity feature of the nodes. i S represents the similarity feature of the nodes. i ∈R X×L W Si and b Si These represent the weights and biases of the fully connected layer after the similarity features are enhanced by the multi-head attention mechanism;

[0059] The enhanced network topology features and enhanced similarity features of the node are concatenated by channel to obtain the fused features of the node.

[0060] Optionally, in one embodiment of this application, the process of performing semi-supervised adversarial training using a classifier and a discriminator includes:

[0061] The classifier and the discriminator are compared using KL divergence as a metric to measure the difference in probability distributions. The classifier and the discriminator then engage in a game based on an adversarial objective function, which is:

[0062] ,

[0063] in, Indicates a positive sample. Indicates unlabeled samples. and These represent the classifier and the discriminator, respectively. It is the classification result of the classifier for the unlabeled sample. It is the discrimination result of the discriminator on the positive sample. It is the identification result of the discriminator on the unlabeled sample. It is a parameter that controls the impact of differences in data distribution. This represents the data distribution of the positive samples. This represents the data distribution of the unlabeled samples. This represents the expected distribution of the data for the positive samples. This represents the expected distribution of the data for the unlabeled samples. A quantified value representing the difference in probability distributions between the classifier and the discriminator.

[0064] Optionally, in one embodiment of this application, the process of performing semi-supervised adversarial training using a classifier and a discriminator includes:

[0065] When optimizing the discriminator, samples are taken from both the positive sample set and the unlabeled sample set, and the loss function of the discriminator is minimized using the stochastic gradient descent algorithm. The loss function of the discriminator is:

[0066] ,

[0067] in, Used to reflect the discrimination capability of the discriminator;

[0068] When optimizing the classifier, samples are taken from the unlabeled sample set, and the loss function of the classifier is minimized using the stochastic gradient descent algorithm, wherein the loss function of the classifier is:

[0069] ,

[0070] in, Used to reflect the classification accuracy of the classifier.

[0071] In a second aspect, embodiments of this application provide a computer device, including: at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor, the instructions being configured to perform the method described in the first aspect above.

[0072] Thirdly, embodiments of this application provide a computer-readable storage medium storing computer-executable instructions for performing the method described in the first aspect above.

[0073] The above technical solution addresses the technical problems of insufficient accuracy and limited generalization ability of drug-disease association prediction models in related technologies. By integrating the topological features and chemical or semantic features of drugs and diseases in the network, and utilizing semi-supervised adversarial training, it fully mines the potential information in known associated positive samples and unlabeled samples, and constructs a high-precision, high-generalization drug-disease association prediction model. This enables the rapid and accurate screening of candidate drugs with therapeutic potential for specific diseases (such as Alzheimer's disease and lung cancer) from a massive pool of existing drugs, providing an efficient and reliable data foundation for drug repositioning, helping to shorten the drug development cycle and reduce drug development costs. Attached Figure Description

[0074] To more clearly illustrate the technical solutions of the embodiments of this application, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0075] Figure 1 A flowchart illustrating a training method for a drug-disease association prediction model according to an embodiment of this application is shown;

[0076] Figure 2 A flowchart illustrating a method for obtaining node fusion features according to an embodiment of this application is shown;

[0077] Figure 3 A schematic diagram of a training method for a drug-disease association prediction model according to another embodiment of this application is shown;

[0078] Figure 4 A block diagram of a computer device according to one embodiment of this application is shown;

[0079] Figure 5 A block diagram of a computer device according to another embodiment of this application is shown. Detailed Implementation

[0080] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0081] Figure 1 A flowchart illustrating a training method for a drug-disease association prediction model according to an embodiment of this application is shown.

[0082] like Figure 1 As shown, a method for training a drug-disease association prediction model according to an embodiment of this application includes:

[0083] Step 102: Obtain a training sample set, wherein the training sample set includes a positive sample set and an unlabeled sample set. The positive sample set includes known drug-disease pairs that are associated, and the unlabeled sample set includes drug-disease pairs that are unknown whether they are associated.

[0084] The training sample set includes the sample data used to train the drug-disease association prediction model. This sample data includes both known-associated drug-disease pairs and unknown-associated drug-disease pairs, i.e., it contains both positive and unlabeled samples. By obtaining a training sample set consisting of both positive samples with known associations and unlabeled samples with unknown associations, a robust data foundation is provided for subsequent semi-supervised learning, avoiding the problem of insufficient model generalization ability caused by using only a limited number of known-associated positive samples for model training.

[0085] Step 104: For any node in all drugs and all diseases involved in the training sample set, obtain the fused feature obtained by fusing the network topology features and similarity features of the node.

[0086] The network topology features of a node can reflect the global connectivity pattern of the drug or disease it represents in a biological heterogeneous network, while the similarity features of a node reflect the intrinsic functional similarity of the drug or disease it represents in terms of chemical structure or biological semantics. Combining network topology features and similarity features can comprehensively characterize the multidimensional information of a node. Based on this, fusing network topology features reflecting the global connectivity pattern of a node with similarity features revealing the intrinsic chemical or semantic attributes of a node can achieve the fusion of biological network structure and chemical or semantic attributes. This allows the feature representation of each node to simultaneously contain its structural performance in the biological heterogeneous network and its functional characteristics in terms of chemical structure or semantic attributes, providing a valid basis for subsequent accurate prediction of the association between drugs and diseases.

[0087] Step 106: Based on the training sample set and the fusion features of each node in the training sample set, a classifier and a discriminator are used for semi-supervised adversarial training until the classifier iterates to meet the predetermined prediction ability conditions, and then the classifier is determined as the drug-disease association prediction model.

[0088] The classifier predicts whether a drug-disease pair is associated based on the fusion features of the drugs and the diseases in the pair. The discriminator determines whether a drug-disease pair predicted as associated by the classifier is a positive sample. In other words, the classifier predicts the association between drug-disease pairs, while the discriminator judges the classifier's predictions as true or false. This constructs a semi-supervised adversarial training framework, allowing the classifier and discriminator to alternately optimize their parameters through adversarial game. Driven by the discriminator, the classifier continuously improves its ability to filter high-quality positive samples from unlabeled samples and gradually enhances its ability to identify potentially real drug-disease pairs during adversarial training, thus improving the generalization ability of the drug-disease association prediction model.

[0089] The above technical solution constructs a training sample set containing both known and unknown associations, and integrates the network topology and similarity features of nodes in drugs and diseases to form a fusion feature reflecting the multi-dimensional characteristics of nodes. Then, a semi-supervised adversarial training method is used with a classifier and discriminator to learn the fusion features of each drug-disease pair in the training sample set, ultimately obtaining a drug-disease association prediction model that meets the accuracy requirements. Thus, while training the drug-disease association prediction model with positive samples, a large number of unlabeled samples are added, allowing the model to fully explore the potential association information between drugs and diseases in a large number of samples during training. This avoids the problem of insufficient model generalization ability caused by training with only a small number of positive samples. Simultaneously, the adversarial training mechanism continuously optimizes the classifier's ability to identify the true association between drugs and diseases, improving the model's generalization performance, thereby achieving more accurate drug-disease association prediction in diverse biomedical scenarios.

[0090] exist Figure 1 Based on the illustrated embodiments, Figure 2 A flowchart illustrating a method for obtaining node fusion features according to an embodiment of this application is shown.

[0091] like Figure 2 As shown, a method for obtaining node fusion features according to an embodiment of this application includes:

[0092] Step 202: Determine the network topology characteristics of the nodes based on the graph embedding algorithm.

[0093] A node refers to any object among all drugs and diseases involved in the training sample set. Graph embedding algorithms are used to represent these nodes as low-dimensional dense vectors, thereby preserving their topological information for subsequent model training tasks. Optionally, the graph embedding algorithm is the Node2Vec algorithm (node-to-vector algorithm).

[0094] In one possible design, a graph embedding algorithm is used to perform multiple fixed-length biased random walks in a pre-constructed biological heterogeneous network to obtain a set of random walk paths. A first parameter controls the probability of a walk returning to the previous node, and a second parameter controls the probability of a walk exploring new nodes. Each random walk path in the set of random walk paths is learned using a Skip-gram model (a context prediction model based on the center word), resulting in a 128-dimensional vector for each node in all drugs and diseases covered in the training sample set, which serves as the network topology feature of each node.

[0095] By controlling the probability of returning to the previous node through the first parameter, the exploration of the local neighborhood can be enhanced. By controlling the probability of exploring new nodes through the second parameter, the learning of the global structure by the walking behavior can be promoted. By controlling the first and second parameters, it is easy to generate a set of random walking paths that takes into account both neighborhood exploration and global structure learning.

[0096] Furthermore, the Skip-gram model can capture the semantic and structural similarity within the network structure of a node by transforming node context relationships into vector representations. Based on this, each random walk path is treated as a sentence, and the nodes within each random walk path are treated as words in the sentence. The Skip-gram model learns the co-occurrence probability of a node and its context nodes in each random walk path, mapping each node to a low-dimensional vector that reflects its context relationship and co-occurrence probability with context nodes. Optionally, nodes can be converted into 128-dimensional vectors.

[0097] Therefore, by employing a graph embedding algorithm to perform a controllable biased random walk in a biological heterogeneous network, and using the Skip-gram model to map the nodes in the path to low-dimensional vectors, the contextual representation and global structure of drug and disease nodes in the biological heterogeneous network can be uniformly quantified, providing effective input information that combines local feature representation and global structure representation for subsequent prediction tasks.

[0098] In addition, prior to step 202, the biological heterogeneous network is constructed based on known drug-disease association data. The nodes in the biological heterogeneous network include drugs, genes, and diseases. The biological heterogeneous network is used to reflect the interaction between drugs and targets, the association between proteins and diseases, protein-protein interaction relationships, and known associations between drugs and diseases.

[0099] In a real-world scenario, 1201 positive samples of drug-disease pairs were collected. Based on these positive samples, a biological heterogeneous network containing 721 drugs, 1894 genes, and 431 diseases was constructed. This biological heterogeneous network includes four types of associations among the three biological entities: drugs, genes, and diseases. These include 8832 drug-target interactions, 23080 protein-disease associations, 14927 protein interactions, and 1201 drug-disease associations.

[0100] Step 204: Based on known drug-disease association data, determine the similarity features of the nodes.

[0101] Known drug-disease association data includes, but is not limited to, interaction pattern vectors between drugs and all genes, chemical structures of drugs, interaction pattern vectors between diseases and all genes, and MeSH (Medical Subject Headings) descriptors for diseases. By utilizing known drug-disease association data to extract node similarity features, complex information such as the chemical structures and MACCS (Molecular Access System) fingerprint vectors between drugs and the semantic relationships between diseases can be converted into dense vector representations, i.e., similarity features. These similarity features can reflect the intrinsic functional association between drugs and diseases. It should be noted that some English expressions in the context are conventional descriptions in the field and do not have a unified Chinese interpretation. The Chinese interpretations provided in parentheses in this application are for illustrative purposes only.

[0102] In one possible design, for every two drugs in the training sample set, the drug similarity between the two drugs is determined based on their GIP (Gaussian Interaction Spectrum) nuclear similarity and chemical structure similarity, and a drug similarity matrix is ​​constructed based on the drug similarity between every two drugs; the drug similarity matrix is ​​then processed by a graph autoencoder to obtain the similarity features of each drug in the training sample set.

[0103] The GIP core similarity between two drugs reflects their functional similarity at the biological activity level. Specifically, it can be calculated by analyzing the interaction patterns of the two drugs with all genes. If two drugs have highly similar interaction spectra, it indicates that they may affect similar sets of targets, and thus their GIP core similarity will be high. In other words, two drugs with high GIP core similarity exert their effects in vivo through similar pathways, and therefore may have similar therapeutic effects or be used to treat similar diseases.

[0104] In one possible design, the method for determining the GIP kernel similarity of the two drugs includes: for the first drug and the second drug, obtaining a first interaction pattern vector between the first drug and all genes and a second interaction pattern vector between the second drug and all genes; and determining the GIP kernel similarity of the two drugs based on the first interaction pattern vector, the second interaction pattern vector, and a Gaussian kernel function, wherein...

[0105] ,

[0106] This represents the GIP kernel similarity between the two drugs, where A represents the first drug and B represents the second drug. and These are the first interaction mode vector and the second interaction mode vector, respectively. It is a parameter used to adjust the bandwidth of the Gaussian kernel function.

[0107] The similarity of the chemical structures of two drugs reflects the degree of structural similarity at the molecular level. If two drugs have highly similar chemical structures, it indicates that they are similar in physicochemical properties and have the potential to interact with biological targets through similar molecular mechanisms, thus potentially possessing similar pharmacological activities. Therefore, two drugs with high chemical structural similarity may have similar therapeutic effects or can be used to treat similar diseases.

[0108] In one possible design, the method for determining the chemical structural similarity of the two drugs includes: determining the Tanimoto similarity of the two drugs based on their respective MACCS fingerprint vectors, as the chemical structural similarity of the two drugs, wherein...

[0109] ,

[0110] This represents the Tanimoto similarity between the MACCS fingerprint vector of the first drug and the MACCS fingerprint vector of the second drug. This shows the number of 1s in the MACCS fingerprint vector of the first drug. This represents the number of 1s in the MACCS fingerprint vector of the second drug. This indicates that the MACCS fingerprint vector of the first drug and the MACCS fingerprint vector of the second drug have a total of 1 bits.

[0111] Therefore, the drug similarity matrix reflects the distribution of comprehensive similarity relationships between all drugs in terms of chemical structure and biological activity. This matrix reveals the potential structural and functional correlations between drugs, providing a structured data foundation for subsequent graph autoencoder extraction of low-dimensional features.

[0112] In one possible design, for every two diseases in the training sample set, the disease similarity between the two diseases is determined based on the GIP kernel similarity and MeSH similarity between the two diseases, and a disease similarity matrix is ​​constructed based on the disease similarity between every two diseases; the disease similarity matrix is ​​processed by a graph autoencoder to obtain the similarity features of each disease in the training sample set.

[0113] The GIP core similarity between two diseases reflects their functional similarity at the pathological mechanism level. Specifically, the GIP core similarity can be calculated by analyzing the association patterns between these two diseases and all genes. If two diseases have highly similar gene association profiles, it indicates that they are significantly associated with similar biomolecules, pathways, or functional modules, and are likely to share some biological basis, pathogenesis, or genetic background. Therefore, if the GIP core similarity between two diseases is high, known treatment strategies or drugs for one disease may also have potential therapeutic value for the other disease.

[0114] In one possible design, the method for determining the GIP kernel similarity of the two diseases includes: for the first disease and the second disease, obtaining the third interaction pattern vector of the first disease with all genes and the fourth interaction pattern vector of the second disease with all genes; and determining the GIP kernel similarity of the two diseases based on the third interaction pattern vector, the fourth interaction pattern vector, and the Gaussian kernel function, wherein...

[0115] ,

[0116] This represents the GIP kernel similarity between the two diseases, where C represents the first disease and D represents the second disease. and These are the third interaction mode vector and the fourth interaction mode vector, respectively. It is a parameter used to adjust the bandwidth of the Gaussian kernel function.

[0117] MeSH similarity between two diseases reflects their semantic relevance at the medical conceptual level. Specifically, it can be calculated by analyzing the hierarchical relationship and shared ancestor node information of the two diseases in the tree structure of the medical subject heading list. The closer the common ancestor node of the two diseases is to the nodes of the two diseases in the MeSH classification system, the closer the medical definitions of the two diseases are in terms of pathophysiology, anatomical location, or clinical manifestations, and the higher their semantic similarity.

[0118] In one possible design, determining the MeSH similarity of the two diseases includes: obtaining a first set of descriptors mapped to the first disease in a MeSH database, and obtaining a second set of descriptors mapped to the second disease in a MeSH database; determining the MeSH similarity of the two diseases based on the first set of descriptors and the second set of descriptors, wherein, if and If they have a common ancestor, then

[0119] ,

[0120] For any first descriptor in the first descriptor set, For any second descriptor in the second descriptor set, express and similarity, yes The information content value in the MeSH tree, yes The information content value in the MeSH tree, represent and The information content value of the common ancestor in the MeSH tree; if and No common ancestor It is zero;

[0121] ,

[0122] This indicates the MeSH similarity between the first disease and the second disease. This represents the first set of descriptors. This refers to the second set of descriptors.

[0123] Therefore, the disease similarity matrix reflects the similarity distribution of different diseases at the level of pathological mechanism and medical semantics. It is a structured expression that can simultaneously characterize the biological characteristics of diseases and the hierarchical relationship of knowledge. Thus, the disease similarity matrix can provide multi-dimensional quantitative basis for mining potential treatment strategy associations between diseases.

[0124] Step 206: The network topology features and similarity features of the node are fused using a multi-head attention mechanism to obtain the fused features of the node.

[0125] Multi-head attention (M&A) is a method that captures complex dependencies between features from different dimensions by computing multiple sets of attention weights in parallel. By fusing network topology and similarity features of nodes, M&A can simultaneously capture the complex relationships between the structural features of nodes in heterogeneous biological networks and their chemical or semantic functions, thus more comprehensively exploring the feature representation of nodes. Further enhancements can be made through residual connections and feedforward networks to achieve nonlinear adaptive weighted fusion of topology and chemical or semantic features. This results in node fusion features that retain the original topology and chemical or semantic characteristics while enhancing the expression of inter-feature relationships, providing multi-dimensional learning support for subsequent drug-disease association prediction tasks.

[0126] It should be added that, the determination of drug similarity between the two drugs based on GIP core similarity and chemical structure similarity includes: if the chemical structure similarity between the two drugs is zero, the drug similarity between the two drugs is determined to be the GIP core similarity between the two drugs; if the chemical structure similarity between the two drugs is non-zero, the drug similarity between the two drugs is determined to be the average of the GIP core similarity and chemical structure similarity between the two drugs.

[0127] When calculating drug similarity, for cases where the chemical structure similarity is zero, the GIP kernel similarity can be directly used as the final result. This preserves the functional correlation between drugs at the biological activity level and avoids underestimation of functional similarity due to significant differences in drug chemical structures. For cases where the chemical structure similarity is non-zero, the average of the GIP kernel similarity and the chemical structure similarity can be taken. This balances the combined influence of molecular structural characteristics and biological functional characteristics between drugs, improving the robustness of drug similarity calculation.

[0128] The step of determining the disease similarity of the two diseases based on the GIP kernel similarity and MeSH similarity includes: if the MeSH similarity of the two diseases is zero, determining the disease similarity of the two diseases as the GIP kernel similarity of the two diseases; if the MeSH similarity of the two diseases is non-zero, determining the disease similarity of the two diseases as the average of the GIP kernel similarity and MeSH similarity of the two diseases.

[0129] When calculating disease similarity, for cases where MeSH semantic similarity is zero, GIP kernel similarity can be directly used as the final result. This ensures that the correlation between diseases at the gene function level is preserved, avoiding the underestimation of biological similarity due to complete semantic non-overlap. For cases where MeSH semantic similarity is non-zero, the average of the GIP kernel similarity and MeSH similarity of the two diseases can be taken, considering both the molecular function and medical semantic correlations of the diseases, thus improving the accuracy of disease similarity calculation.

[0130] In one possible design, step 206 specifically includes: determining the enhanced network topology characteristics of the node based on the network topology characteristics of the node, wherein...

[0131] ,

[0132] H St S represents the enhanced network topology characteristics of the nodes. t S represents the network topology characteristics of the node. t ∈R X×L R represents the set of nodes in a single training iteration, X represents the number of nodes in a single training iteration, L represents the feature vector length of each node, MultiHead represents the multi-head attention mechanism, and W... St and b St represents the weights and biases of the fully connected layer after the network topology features are enhanced by the multi-head attention mechanism, respectively; LayerNorm is the layer normalization function, and LeakyReLU is the nonlinear activation function; based on the similarity features of the nodes, the enhanced similarity features of the nodes are determined, wherein,

[0133] ,

[0134] H Si S represents the enhanced similarity feature of the nodes. i S represents the similarity feature of the nodes. i ∈R X×L W Si and b SiThe weights and biases of the fully connected layer after the similarity features are enhanced by the multi-head attention mechanism are represented respectively. The enhanced network topology features and enhanced similarity features of the node are concatenated by channel to obtain the fused features of the node.

[0135] By enhancing the network topology and similarity features of nodes using a multi-head attention mechanism, and then concatenating the enhanced features into channels, deep fusion of multi-source features of nodes is achieved simultaneously with enhancement. This effectively captures the complex relationships between the topology and functional attributes of nodes, ultimately generating fused features that combine structural and chemical / semantic representations, providing effective input data for the subsequent classifier training process.

[0136] Figure 3 A schematic diagram of a training method for a drug-disease association prediction model according to another embodiment of this application is shown.

[0137] like Figure 3 As shown, the model input first includes positive drug-disease association samples and unlabeled drug-disease samples. Next, feature extraction is performed through two parallel channels: on one hand, the network topology features of each node in the drug and disease associations are extracted based on the Node2Vec graph embedding algorithm; on the other hand, the similarity features of the nodes are extracted based on GAE (Graph Autoencoder). Then, the network topology features and similarity features of the nodes are respectively input into an attention network for feature enhancement and fusion using a multi-head attention mechanism.

[0138] The fused features are simultaneously fed into both the discriminator and the classifier. The discriminator analyzes the features to determine whether there is a drug-disease association or not, while the classifier is responsible for identifying potential drug-disease associations or not. The discriminator and classifier optimize each other through an adversarial training mechanism, ultimately making the classifier a predictive model capable of accurately identifying potential drug-disease associations.

[0139] Alternatively, during the model training phase, the classifier and discriminator are trained alternately in an adversarial game. The classifier attempts to filter out candidate instances with positive sample features from unlabeled samples to confuse the discriminator, while the discriminator continuously adjusts its parameters to accurately identify potential positive samples, thereby strengthening its discriminative ability. Through continuous adversarial training, the classifier gradually narrows the representational difference between real positive samples and high-quality potential positive samples, while simultaneously enhancing its ability to distinguish unlabeled samples.

[0140] In one possible design, the semi-supervised adversarial training process using a classifier and a discriminator includes: measuring the difference in probability distributions between the classifier and the discriminator based on KL (Kullback-Leibler Divergence, relative entropy) divergence; the classifier and the discriminator engaging in a game through an adversarial objective function, wherein the adversarial objective function is:

[0141] ,

[0142] in, Indicates a positive sample. express, and These represent the classifier and the discriminator, respectively. It is the classification result of the classifier for the unlabeled sample. It is the discrimination result of the discriminator on the positive sample. It is the identification result of the discriminator on the unlabeled sample. It is a parameter that controls the impact of differences in data distribution. This represents the data distribution of the positive samples. This represents the data distribution of the unlabeled samples. This represents the expected distribution of the data for the positive samples. This represents the expected distribution of the data for the unlabeled samples. This represents the quantified value of the difference in probability distributions between the classifier and the discriminator. Training ends when the adversarial objective function converges, the fluctuations in the classifier loss and discriminator loss stabilize, and there are no longer any systematic increases or decreases—that is, when the classifier and discriminator enter a state of oscillating equilibrium.

[0143] Therefore, by introducing an adversarial objective function based on KL divergence, the probability distribution difference between the classifier and discriminator on positive and unlabeled samples is quantified. This effectively utilizes a small amount of labeled data and a large amount of unlabeled samples for joint optimization. By accurately measuring the distribution difference, the model's ability to identify potential drug-disease associations and its generalization performance in data-sparse scenarios are improved, thereby enhancing the reliability of the prediction results.

[0144] In optimizing the discriminator, samples are taken from both the positive sample set and the unlabeled sample set, and the loss function of the discriminator is minimized using a stochastic gradient descent algorithm. The loss function of the discriminator is:

[0145] ,

[0146] in, Used to reflect the discrimination capability of the discriminator.

[0147] The loss function of the discriminator is minimized by using the stochastic gradient descent algorithm. This loss function maximizes the ability to distinguish the distribution differences between real and potential samples. Therefore, this optimization process can force the discriminator to strengthen the discrimination boundary of sample source, thereby improving the model's recognition accuracy in adversarial games.

[0148] When optimizing the classifier, samples are taken from the unlabeled sample set, and the loss function of the classifier is minimized using the stochastic gradient descent algorithm, wherein the loss function of the classifier is:

[0149] ,

[0150] in, Used to reflect the classification accuracy of the classifier.

[0151] When optimizing the classifier, samples are taken from an unlabeled sample set to simulate real-world application scenarios. Similarly, the stochastic gradient descent algorithm can be used to minimize the classifier's loss function. This loss function aims to improve classification accuracy by optimizing the classifier's ability to filter unlabeled samples, enabling it to gradually and accurately identify potential positive samples in the unlabeled data. Finally, by alternately optimizing the classifier and discriminator, the classifier and discriminator achieve co-evolution, thereby effectively improving the overall model's generalization performance under sparse data conditions.

[0152] The above technical solutions, by integrating the topological and chemical or semantic features of drugs and diseases in the network and utilizing semi-supervised adversarial training, fully explore the potential information in known associated positive samples and unlabeled samples to construct a high-precision, high-generalization drug-disease association prediction model. This enables the rapid and accurate screening of candidate drugs with therapeutic potential for specific diseases (such as Alzheimer's disease and lung cancer) from a vast pool of existing drugs, providing an efficient and reliable data foundation for drug repositioning, helping to shorten the drug development cycle and reduce drug development costs.

[0153] In another embodiment, this application provides a computer device, which may be a server, and its internal structure diagram may be as follows. Figure 4As shown, the computer device includes a processor, memory, network interface, and database connected via a system bus. The processor provides computing and control capabilities. The memory includes non-volatile and / or volatile storage media and internal memory. The non-volatile storage media stores the operating system, computer programs, and database. The internal memory provides an environment for the operation of the operating system and computer programs in the non-volatile storage media. The network interface is used to communicate with external clients via a network connection. When the computer program is executed by the processor, it can implement the methods described in any of the above embodiments.

[0154] In one embodiment, this application also provides a computer device, which can be a client, and its internal structure diagram can be as follows: Figure 5 As shown, the computer device includes a processor, memory, network interface, display screen, and input device connected via a system bus. The processor provides computing and control capabilities. The memory includes a non-volatile storage medium and internal memory. The non-volatile storage medium stores an operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs in the non-volatile storage medium. The network interface is used to communicate with an external server via a network connection. When the computer program is executed by the processor, it can implement the methods described in any of the above embodiments.

[0155] Any of the computer devices described in the embodiments of this application exist in various forms, including but not limited to:

[0156] (1) Mobile communication devices: These devices are characterized by their mobile communication capabilities and primarily aim to provide voice and data communication. These terminals include smartphones, multimedia phones, feature phones, and low-end phones.

[0157] (2) Ultra-mobile personal computer devices: These devices fall under the category of personal computers, have computing and processing functions, and generally also have mobile Internet access capabilities.

[0158] (3) Portable entertainment devices: These devices can display and play multimedia content. This category includes: audio and video players, handheld game consoles, e-books, as well as smart toys, wearable devices, and portable car navigation devices.

[0159] (4) Server: A device that provides computing services. The components of a server include a processor, hard disk, memory, system bus, etc. Servers are similar to general computer architectures, but because they need to provide highly reliable services, they have higher requirements in terms of processing power, stability, reliability, security, scalability, and manageability.

[0160] (5) Other electronic devices with data interaction functions.

[0161] Additionally, embodiments of this application provide a computer-readable storage medium storing computer-executable instructions, which are used to perform the following steps:

[0162] Obtain a training sample set, wherein the training sample set includes a positive sample set and an unlabeled sample set, the positive sample set includes known drug-disease pairs that are associated, and the unlabeled sample set includes drug-disease pairs that are unknown whether they are associated.

[0163] For any node among all drugs and all diseases involved in the training sample set, obtain the fused feature obtained by fusing the network topology features and similarity features of the node;

[0164] Based on the training sample set and the fusion features of each node within the training sample set, a classifier and a discriminator are used for semi-supervised adversarial training until the classifier iterates to meet a predetermined prediction capability condition. At this point, the classifier is determined as a drug-disease association prediction model. The classifier is used to predict whether the drug and disease pair are associated based on the fusion features of the drug and the fusion features of the disease in the drug-disease pair. The discriminator is used to determine whether the drug-disease pair predicted as associated by the classifier is a positive sample.

[0165] It should be noted that the functions or steps that can be implemented by the computer-readable storage medium or computer device described above can be referred to the relevant descriptions in the foregoing method embodiments. To avoid repetition, they will not be described one by one here.

[0166] The technical solution of this application has been described in detail above with reference to the accompanying drawings. This technical solution constructs a training sample set containing both known and unknown associations, and integrates the network topology features and similarity features of each node in the drug and disease sets to form a fusion feature reflecting the multi-dimensional characteristics of the nodes. Then, a classifier and discriminator are used for semi-supervised adversarial training to learn the fusion features of each drug-disease pair in the training sample set, ultimately obtaining a drug-disease association prediction model that meets the accuracy requirements. Therefore, while training the drug-disease association prediction model with positive samples, a large number of unlabeled samples are added, allowing the model to fully explore the potential association information between drugs and diseases in a large number of samples during training. This avoids the problem of insufficient model generalization ability caused by training with only a small number of positive samples. Simultaneously, the adversarial training mechanism continuously optimizes the classifier's ability to identify the true association between drugs and diseases, improving the model's generalization performance, thereby achieving more accurate drug-disease association prediction in diverse biomedical scenarios.

[0167] It should be understood that although the terms "first," "second," etc., may be used to describe drugs in the embodiments of this application, these drugs should not be limited to these terms. These terms are only used to distinguish drugs from each other. For example, without departing from the scope of the embodiments of this application, a first drug may also be referred to as a second drug, and similarly, a second drug may also be referred to as a first drug.

[0168] Depending on the context, the word "if" as used here can be interpreted as "when," "when," "in response to determination," or "in response to detection." Similarly, depending on the context, the phrase "if determination" or "if detection (of the stated condition or event)" can be interpreted as "when determination," "in response to determination," "when detection (of the stated condition or event)," or "in response to detection (of the stated condition or event)."

[0169] The terminology used in the embodiments of this application is for the purpose of describing particular embodiments only and is not intended to be limiting of this application. The singular forms “a,” “the,” and “the” used in the embodiments of this application and the appended claims are also intended to include the plural forms unless the context clearly indicates otherwise.

[0170] In the several embodiments provided in this application, it should be understood that the disclosed systems, apparatuses, and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces; the indirect coupling or communication connection between apparatuses or units may be electrical, mechanical, or other forms.

[0171] Furthermore, the functional units in the various embodiments of this application can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or in a combination of hardware and software functional units.

[0172] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium, and when executed, it can include the processes of the embodiments of the above methods. Any references to memory, storage, databases, or other media used in the embodiments provided in this application can include non-volatile and / or volatile memory. Non-volatile memory can include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory can include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in various forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), dual data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), Rambus direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM), etc.

[0173] The above-described embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention, and should all be included within the protection scope of the present invention.

Claims

1. A method for training a drug-disease association prediction model, characterized in that, The method comprises the following steps: obtaining a training sample set, wherein the training sample set comprises a positive sample set and an unlabeled sample set, the positive sample set comprises known drug and disease pairs having correlations, and the unlabeled sample set comprises drug and disease pairs unknown whether having correlations; obtaining a fusion feature of a network topology structure feature and a similarity feature of any node in all drugs and all diseases involved in the training sample set; based on the training sample set and the fusion feature of each node in the training sample set, performing semi-supervised adversarial training by using a classifier and a discriminator until the classifier meets a predetermined prediction ability condition, and determining the classifier as a drug-disease correlation prediction model, wherein the classifier is used to predict whether a drug and disease pair has a correlation based on the fusion feature of the drug and the fusion feature of the disease in the drug and disease pair, and the discriminator is used to judge whether a drug and disease pair predicted as having a correlation by the classifier is a positive sample; the step of obtaining the fusion feature of the network topology structure feature and the similarity feature of any node in all drugs and all diseases involved in the training sample set comprises the following steps: determining the network topology structure feature of the node based on a graph embedding algorithm; determining the similarity feature of the node based on known drug-disease correlation data; and fusing the network topology structure feature and the similarity feature of the node by using a multi-head attention mechanism to obtain the fusion feature of the node.

2. The method of claim 1, wherein, The step of determining the network topology structure feature of the node based on the graph embedding algorithm comprises the following steps: performing multiple fixed-length biased random walks in a pre-constructed biological heterogeneous network by using the graph embedding algorithm to obtain a random walk path set, wherein a first parameter is used to control the probability of returning to a previous node, and a second parameter is used to control the probability of exploring a new node; learning each random walk path in the random walk path set by using a Skip-gram model to obtain a 128-dimensional vector of each node in all drugs and all diseases involved in the training sample set as the network topology structure feature of each node.

3. The method of claim 2, wherein, Before the step of performing multiple fixed-length biased random walks in a pre-constructed biological heterogeneous network by using the graph embedding algorithm, the method further comprises the following step: constructing the biological heterogeneous network based on known drug-disease correlation data, wherein the nodes in the biological heterogeneous network comprise drugs, genes and diseases, and the biological heterogeneous network is used to reflect the interaction relationship between drugs and target points, the correlation relationship between proteins and diseases, the protein interaction relationship, and the known correlation relationship between drugs and diseases.

4. The method of claim 1, wherein, The step of determining the similarity feature of the node based on the known drug-disease correlation data comprises the following steps: for each two drugs in the training sample set, determining a drug similarity of the two drugs based on the GIP core similarity and the chemical structure similarity of the two drugs, and constructing a drug similarity matrix based on the drug similarity of each two drugs, and For each two diseases in the training sample set, based on the GIP kernel similarity and the MeSH similarity of the two diseases, a disease similarity of the two diseases is determined, and based on the disease similarity of each two diseases, a disease similarity matrix is constructed; The drug similarity matrix and the disease similarity matrix are processed by a graph auto-encoder respectively to obtain similarity features of each drug and similarity features of each disease in the training sample set.

5. The method of claim 4, wherein, The manner of determining the GIP kernel similarity of the two drugs comprises: For a first drug and a second drug in the two drugs, a first interaction pattern vector of the first drug with all genes and a second interaction pattern vector of the second drug with all genes are obtained; Based on the first interaction pattern vector, the second interaction pattern vector and a Gaussian kernel function, the GIP kernel similarity of the two drugs is determined, wherein, ， denotes the GIP kernel similarity of the two drugs, A represents the first drug, B represents the second drug, and are the first and second interaction mode vectors, respectively, is a parameter for adjusting the bandwidth of the Gaussian kernel function; The manner of determining the chemical structure similarity of the two drugs comprises: Based on the MACCS fingerprint vectors of the two drugs respectively, a Yubin similarity of the two drugs is determined as the chemical structure similarity of the two drugs, wherein, ， a Tanimoto similarity of the MACCS fingerprint vector of the first drug to the MACCS fingerprint vector of the second drug, a number of bits in the MACCS fingerprint vector of the first drug that are 1, a number of bits in the MACCS fingerprint vector of the second drug that are 1, a number of bits that are 1 in both the MACCS fingerprint vector of the first drug and the MACCS fingerprint vector of the second drug.

6. The method of claim 4, wherein, The manner of determining the GIP kernel similarity of the two diseases comprises: For a first disease and a second disease in the two diseases, a third interaction pattern vector of the first disease with all genes and a fourth interaction pattern vector of the second disease with all genes are obtained; Based on the third interaction pattern vector, the fourth interaction pattern vector and a Gaussian kernel function, the GIP kernel similarity of the two diseases is determined, wherein, ， represents a GIP kernel similarity between the two diseases, C represents the first disease, D represents the second disease, and are the third and fourth interaction mode vectors, respectively, is a parameter for adjusting the bandwidth of the Gaussian kernel function; The manner of determining the MeSH similarity of the two diseases comprises: A first descriptor set mapped by the first disease in the MeSH database is obtained, and a second descriptor set mapped by the second disease in the MeSH database is obtained; determining a MeSH similarity of the two diseases based on the first descriptor set and the second descriptor set, wherein if and have a common ancestor, ， for any first descriptor in the first set of descriptors, for any second descriptor in the second set of descriptors, represents the similarity of to is the information content value in the MeSH tree, is the information content value in the MeSH tree, represents and the information content value in the MeSH tree of the common ancestor of If and no common ancestor, is zero; ， represents a MeSH similarity of the first disease to the second disease, represents the first set of descriptors, represents the second set of descriptors.

7. The method of claim 4, wherein, The determination of the drug similarity of the two drugs based on the GIP kernel similarity and the chemical structure similarity of the two drugs comprises: If the chemical structure similarity of the two drugs is zero, the drug similarity of the two drugs is determined as the GIP kernel similarity of the two drugs; If the chemical structure similarity of the two drugs is non-zero, the drug similarity of the two drugs is determined as the average of the GIP kernel similarity and the chemical structure similarity of the two drugs; The determination of the disease similarity of the two diseases based on the GIP kernel similarity and the MeSH similarity of the two diseases comprises: If the MeSH similarity of the two diseases is zero, the disease similarity of the two diseases is determined as the GIP kernel similarity of the two diseases; If the MeSH similarity of the two diseases is non-zero, the disease similarity of the two diseases is determined as the average of the GIP kernel similarity and the MeSH similarity of the two diseases.

8. The method according to any one of claims 1 to 7, characterized in that, The fusion of the network topology structure features and the similarity features of the node by the multi-head attention mechanism to obtain the fusion features of the node comprises: Based on the network topology structure features of the node, enhanced network topology structure features of the node are determined, wherein, ， H St represents the enhanced network topology feature of the node, St represents the network topology feature of the node, St∈R X×L , R represents a node set for single training, X represents the number of nodes in single training, L represents the length of the feature vector of each node, MultiHead represents a multi-head attention mechanism, W St and b St respectively represent the weight and bias of the full connection layer after the network topology feature is enhanced by the multi-head attention mechanism, LayerNorm is a layer normalization function, and LeakyReLU is a nonlinear activation function; Determine an enhanced similarity feature of the node based on the similarity feature of the node, wherein ， H Si denotes the enhanced similarity feature of the node, S1denotes the similarity feature of the node, S1∈R X×L , W Si and b Si respectively denote the weight and bias of the full connection layer after the similarity feature is enhanced by the multi-head attention mechanism Concatenate the enhanced network topology feature and the enhanced similarity feature of the node by channel to obtain a fusion feature of the node.

9. The method of claim 1, wherein, In the process of semi-supervised adversarial training by using the classifier and the discriminator, comprising: Based on the KL divergence measure of the probability distribution difference of the classifier and the discriminator, the classifier and the discriminator are played through the adversarial objective function, wherein the adversarial objective function is: ， wherein, represents a positive sample, represents an unlabeled sample, and respectively represent the classifier and the discriminator, is a classification result of the unlabeled sample by the classifier, is a discrimination result of the positive sample by the discriminator, is a discrimination result of the unlabeled sample by the discriminator, is a parameter for controlling the influence of data distribution difference, represents a data distribution of the positive sample, represents a data distribution of the unlabeled sample, represents an expectation of the data distribution of the positive sample, represents an expectation of the data distribution of the unlabeled sample, represents a quantization value of the probability distribution difference of the classifier and the discriminator.

10. The method of claim 9, wherein, In the process of semi-supervised adversarial training by using the classifier and the discriminator, comprising: When optimizing the discriminator, respectively sampling from the positive sample set and the unlabeled sample set, minimizing the loss function of the discriminator by the stochastic gradient descent algorithm, wherein the loss function of the discriminator is: ， wherein, to reflect the discrimination ability of the discriminator; When optimizing the classifier, sampling from the unlabeled sample set, minimizing the loss function of the classifier by the stochastic gradient descent algorithm, wherein the loss function of the classifier is: ， wherein, for reflecting the classification accuracy of the classifier.

11. A computer device, comprising: Comprising: At least one processor; And the memory connected with the at least one processor in communication; Wherein, the memory has instructions executable by the at least one processor, the instructions are set to execute the method of any one of claims 1 to 10.

12. A computer-readable storage medium, characterized in that, Computer executable instructions are stored, and the computer executable instructions are used to execute the method of any one of claims 1 to 10.

Images