Drug repositioning method based on reinforcement symmetry metric learning and graph convolution network

Abstract

The application relates to the technical field of bioinformatics, in particular to a drug repositioning method based on reinforced symmetric metric learning and a graph convolution network, which comprises the following steps: a drug-disease heterogeneous network is constructed by integrating the correlation of drugs and diseases and biomedical information. The heterogeneous network comprises a drug-drug similarity network, a disease-disease similarity network and a drug-disease correlation network. A graph convolution network is applied to learn the node features of drugs and diseases, and potential drug-disease correlations are predicted to supplement the missing drug-disease correlation information. A reinforced symmetric metric learning method with adaptive margins is used to learn the potential vector representation of drugs and diseases, and the symmetric learning of drug-centered and disease-centered is considered. Based on the potential vector representation learned in the unified metric vector space, new drug-disease correlations are identified through a metric function. The application is simple and effective, and has good performance in drug repositioning prediction.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of bioinformatics, and particularly relates to a drug repositioning method based on reinforced symmetric metric learning and graph convolution network. BACKGROUND

[0002] With the increasing prevalence of complex diseases such as tumors, diabetes, and cardiovascular diseases in recent years, the traditional drug discovery process is time-consuming, costly, and labor-intensive, and cannot meet people's needs. Therefore, more accurately and effectively capturing the relevant indications of drugs is of great significance in drug development. Drug repositioning, i.e., old drugs for new uses, is an attractive approach to discovering new therapeutic potential for existing drugs approved by the FDA (Food and Drug Administration). Therefore, it has the advantages of reducing drug risk, shortening the clinical evaluation period, low cost, and high efficiency. Many computational drug repositioning methods have been proposed to identify candidate indications for drugs. These methods can be roughly divided into three categories, such as machine learning-based drug repositioning methods, network-based drug repositioning methods, and recommendation system-based drug repositioning methods.

[0003] Machine learning-based methods mainly apply support vector machines (SVM), logistic regression, naive Bayes, and random forests for classification and prediction tasks in drug repositioning. However, these traditional methods rely heavily on manually set features to represent the characteristics of drugs and diseases as input data, which leads to high implementation complexity. Deep learning, as an extension of machine learning, has the advantage of automatically capturing nonlinear features from raw data and has been widely used in drug repositioning in recent years. A network-based deep learning method deepDR is proposed, which uses a multi-modal deep autoencoder to learn high-order nonlinear features of drugs, and finally uses a collective variational autoencoder to infer new drug-disease associations. Network-based methods analyze the relationships between entities by information transmission in different paths constructed by multi-dimensional data, and have interpretability. A prioritization method based on heterogeneous networks is proposed to predict new diseases related to drugs. By simultaneously integrating information about drugs, diseases, and targets, this method uses a propagation flow algorithm to achieve drug-disease and disease-drug prioritization. A bi-random walk (BiRW) algorithm is used to identify potential drug-disease associations on a drug-disease heterogeneous network. Although network-based methods have certain effectiveness in link prediction tasks, they cannot achieve ideal performance.

[0004] The recommendation system-based method performs well in the recommendation field such as social networks, e-commerce platforms, personalized reading, and the like. Similar to recommending favorite items for users, the drug-disease association prediction problem can be modeled as recommending potential treatment diseases for given drugs. In recent years, the recommendation method based on matrix decomposition and matrix completion has been applied to drug repositioning and achieved great success. A drug repositioning recommendation system (DRRS) is proposed, which uses a singular value thresholding (SVT) algorithm to complete unknown entries in the drug-disease association matrix. The generalized matrix factorization (GMF) method is applied to the collaborative filtering process to discover the potential treatment relationship between drugs and diseases.

[0005] The matrix decomposition or matrix completion-based method can be flexibly applied, but it is inefficient for large-scale data due to complex matrix operations. In particular, the inner product operation of the matrix decomposition technique violates the triangle inequality rule, which can lead to suboptimal performance of the model recommendation. In addition, this simple linear combination method can only measure the drug-disease relationship, but ignores the modeling of drug-drug and disease-disease relationships. Therefore, metric learning is proposed to make up for the shortcomings of matrix decomposition to enhance the expression ability of the model, which is superior to matrix decomposition in the recommendation field. In recent studies, it has been introduced into the field of drug repositioning. For example, a collaborative metric learning method (CMLDR) is proposed for drug repositioning. CMLDR maps drugs and diseases to a unified vector space based on known drug-disease associations, and then predicts missing drug-disease pairs according to the learned vectors of metric learning. In addition, information related to drugs is used to recommend potential treatment diseases for new drugs. The current drug repositioning method based on metric learning only considers drug-centered learning, which learns the representation of drugs and diseases using drug-disease association information. However, these methods ignore the auxiliary role of disease-centered metric learning. In addition, drug-centered metric learning can alleviate the problem of matrix decomposition to some extent, but this way of modeling drug-disease relationships is biased. This one-way learning based on the push-pull mechanism cannot more effectively gather drugs with common treatment effects and diseases that can be treated by the same drug, and therefore cannot strengthen the establishment of drug-drug and disease-disease relationships, thus often leading to poor performance in drug repositioning prediction. SUMMARY

[0006] The summary part of the application is used to introduce the concepts in a brief form, which will be described in detail in the specific embodiment part. The summary part of the application is not intended to identify the key features or essential features of the claimed technical solution, nor is it intended to limit the scope of the claimed technical solution.

[0007] In order to solve the technical problem of poor performance in drug repositioning prediction, the present application proposes a drug repositioning method based on reinforcement symmetric metric learning and graph convolution network.

[0008] The present application provides a drug repositioning method based on reinforcement symmetric metric learning and graph convolution network, which comprises:

[0009] Using the biomedical information of drugs and diseases, the drug similarity and disease similarity are calculated respectively, the drug similarity network and disease similarity network are constructed through the similarity information, and the drug-disease heterogeneous network is constructed by connecting with the known drug-disease association network;

[0010] Based on the constructed drug-disease heterogeneous network, the low-dimensional embedding features of drug and disease nodes are encoded by using the graph convolution network, the drug-disease association score is predicted by decoding, the input drug-disease association matrix is filled, and the improved drug-disease association matrix is obtained;

[0011] Based on the obtained drug-disease association matrix, the symmetric metric learning algorithm is adopted to learn the embedding representation of drugs and diseases in the unified metric vector space from the center of drug and disease, and finally the potential and new drug-disease association is obtained based on the Euclidean distance.

[0012] Further, based on the obtained drug-disease association matrix, the symmetric metric learning algorithm is adopted to learn the embedding representation of drugs and diseases in the unified metric vector space from the center of drug and disease, and finally the potential and new drug-disease association is obtained based on the Euclidean distance, which comprises:

[0013] Based on the known drug-disease association information, the drug-disease association can be represented as Γ={(r,d)|r∈A r ,d∈A d},(r,d) represents the known drug-disease pair, A r ={r1,r2,...r M} represents the set of M drugs, A d ={d1,d2,...d N} represents the set of N diseases, and the association matrix Y∈{0,1} M×N If there is an edge between drug r i and disease d j , then the entry y ij of the matrix Y is 1, otherwise y ij =0; N i + ={d j |d j ∈Γand yij =1} represents a disease set associated with drug r i ; =0} represents a disease set without recognized association with drug r i ; respectively selecting positive drug-disease pairs, negative drug-disease pairs and negative disease-drug pairs;

[0014] For a given triple (r,d)∈Γ represents a known drug-disease association, which is regarded as a positive sample, represents a negative sample, is a disease without association with drug r; drug-centered metric learning is performed to learn the distance between drug and positive disease and the distance between drug and negative disease based on a push-pull mechanism, a margin value m is set, and m>0, represents Euclidean distance;

[0015] For a given triple (d,r)∈Γ represents a known disease-drug association, which is regarded as a positive sample, represents a disease-drug negative sample, is a drug without association with disease d; based on the assumption that the distance metric similarity, d and are irrelevant, disease-centered metric learning is performed on the basis of drug-centered learning, a margin value n is set, and n>0, After disease-centered metric learning, drugs associated with and not associated with the disease are separated;

[0016] According to the inconsistency of drug-disease and disease-drug association strength, the margin personalized adaptive training is adapted to different drugs and diseases.

[0017] The present application has the following beneficial effects:

[0018] The application discloses a drug repositioning method based on reinforced symmetric metric learning and a graph convolutional network. BRIEF DESCRIPTION OF DRAWINGS

[0019] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, and the advantages thereof, below will briefly introduce the drawings needed in the embodiments or prior art description. Obviously, the drawings in the following description are only some embodiments of the present application, and for those skilled in the art, other drawings can also be obtained from these drawings without creative labor.

[0020] Figure 1 A flowchart of the drug repositioning method based on reinforced symmetric metric learning and a graph convolutional network according to the present application;

[0021] Figure 2 A flowchart of RSML-GCN according to the present application;

[0022] Figure 3 A schematic diagram of drug-centered metric learning according to the present application;

[0023] Figure 4 A schematic diagram of disease-centered metric learning according to the present application;

[0024] Figure 5 A result schematic diagram of recall value and precision value of top-k indications recommended for a drug on a ten-fold cross-validation on a data set Cdataset according to the present application;

[0025] Figure 6 A result schematic diagram of recall value and precision value of top-k indications recommended for a drug on a ten-fold cross-validation on a data set Fdataset according to the present application;

[0026] Figure 7 Figure 6 shows a schematic diagram of the recall and precision results of the top-k recommended indications for a drug according to the ablation experiments on the dataset Cdataset of the present application. DETAILED DESCRIPTION

[0027] To further clarify the technical means and effects taken by the present application to achieve the predetermined object, the specific embodiments, structures, features and effects of the technical solutions proposed by the present application are described in detail below in combination with the drawings and preferred embodiments. In the following description, different "one embodiment" or "another embodiment" do not necessarily refer to the same embodiment. In addition, the specific features, structures or characteristics in one or more embodiments can be combined in any suitable form.

[0028] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which the present application belongs.

[0029] The present application provides a drug repositioning method based on reinforced symmetric metric learning and graph convolutional network, which comprises the following steps:

[0030] Using the biomedical information of drugs and diseases, the drug similarity and disease similarity are calculated respectively, the drug similarity network and disease similarity network are constructed through the similarity information, and the drug-disease heterogeneous network is constructed by connecting the known drug-disease association network;

[0031] Based on the constructed drug-disease heterogeneous network, the low-dimensional embedding features of drug and disease nodes are encoded by using the graph convolutional network, the drug-disease association score is predicted by decoding, the input drug-disease association matrix is filled, and an improved drug-disease association matrix is obtained;

[0032] Based on the obtained drug-disease association matrix, a symmetric metric learning algorithm is used to learn the embedding representation of drugs and diseases in a unified metric vector space from the center of drugs and the center of diseases, and finally the potential and new drug-disease associations are obtained based on the Euclidean distance.

[0033] The above steps are described in detail as follows:

[0034] Reference Figure 1 Figure 1 shows the flow of some embodiments of a drug repositioning method based on reinforced symmetric metric learning and graph convolutional network according to the present application. The RSML-GCN flowchart of the present application is shown in Figure 1. The drug repositioning method based on reinforced symmetric metric learning and graph convolutional network comprises the following steps: Figure 2

[0035] ​Step S1, using biomedical information of drugs and diseases, respectively calculating drug similarity and disease similarity, constructing drug similarity network and disease similarity network through similarity information, and connecting with known drug-disease association network to construct drug-disease heterogeneous network.

[0036] In some embodiments, biomedical information of drugs and diseases can be used to calculate drug similarity and disease similarity respectively, construct drug similarity network and disease similarity network through similarity information, and connect with known drug-disease association network to construct drug-disease heterogeneous network.

[0037] It should be noted that the data set applied by the present application includes a drug set, a disease set and known drug-disease associations. The present application adopts two benchmark data sets. The first is Cdataset, the verified drug-disease association data in this data set is extracted from CTD (The Comparative Toxicogenomics Database), CTD is a publicly available database that integrates a large number of drug, disease, gene and functional phenotype interaction data. Cdataset contains 269 drugs and 598 diseases, of which there are 18416 drug-disease pairs that have been confirmed to be associated. In addition, the target information of the drug is collected from the DrugBank database as the characteristics of the drug, and the characteristics of the disease are defined based on the medical subject headings (MeSH, Medical Subject Headings), MeSH is a subject heading table compiled by the National Library of Medicine (National Library of Medicine) as a biomedical indexing, which provides semantic feature descriptors for diseases. The second data set is Fdataset, which contains 593 drugs collected from the DrugBank database and 1933 known drug-disease associations between 313 diseases obtained from the OMIM (Online Mendelian Inheritance in Man) database. The detailed statistical data of the data set is summarized in Table 1.

[0038] Table 1

[0039] Dataset Drug Disease Known drug-disease associations Cdataset 269 598 18416 Fdataset 593 313 1933

[0040] Drug similarity calculation can include the following steps:

[0041] According to the target feature descriptors of the drug, a drug can be encoded as a binary feature vector, and 1 in the vector indicates that the drug is associated with the corresponding target, and 0 indicates that the drug is not associated with the corresponding target. Based on the target characteristics of the drug, the Jaccard similarity coefficient is used to calculate the drug-drug similarity, two binary feature vectors ai and a j respectively represent the target characteristics of drug i and drug j, and the formula corresponding to the Jaccard similarity coefficient can be:

[0042]

[0043] where |a i ∩a j | represents the number of elements that are 1 at the corresponding positions of a i and a j , |a i ∪a j | represents the number of elements of a i or a j that are 1, and S ij ∈[0,1] represents the similarity value between drug i and drug j.

[0044] Disease similarity calculation can include the following steps:

[0045] Calculate the semantic similarity of disease-disease using MeSH information. Each disease can be regarded as a hierarchical directed acyclic graph (DAG, Directed Acyclic Graph), and its nodes represent disease MeSH descriptors, and edges represent the relationship between the current node and its ancestor node. For disease O, its DAG is composed of disease O and all its ancestor diseases, which can be represented as DAG(O) = (N(O), E(O)), where N(O) is the set of all ancestor nodes (including itself) of O, and E(O) is the edge set of all relationships between O and its ancestor nodes. The semantic value contributed by a disease pair to disease O can be represented by the formula:

[0046]

[0047] where C O (d) is the semantic value, representing the contribution of a disease d in N(O) to disease O. λ∈[0,1] is a contribution factor, which is set to 0.5 here, and the total semantic value contributed by disease d to disease O is defined as DV(O) = ∑ d∈N(O) C O (d). According to the assumption that diseases with more common ancestors in the DAG tend to have higher semantic similarity, the semantic similarity S O,P between disease O and disease P can be represented as:

[0048]

[0049] Here C P (d) is the semantic value of disease d related to disease P in N(P), and DV(P) is the total semantic value contributed by disease d to disease P.

[0050] The construction of drug-disease heterogeneous networks may include the following steps:

[0051] Based on drug similarity and disease similarity, a drug similarity network R and a disease similarity network D can be constructed, with the edge weights set to similarity values. Where A... r ={r1,r2,...r M} represents a set of M drugs, A d ={d1,d2,...d N} represents a set of N diseases. The adjacency matrix represents the drug similarity network. This represents the adjacency matrix of a disease similarity network. Based on known associations between drugs and diseases, a drug-disease association network S can be constructed. rd If drug r i and disease d j If there is a known correlation between them, then r i and d j There exists an edge between them. In S rd The corresponding binary incidence matrix Y∈{0,1} M×N In the middle, if drug r i and disease d j If there exists an edge between them, then the entries y in matrix Y are... ij If y is 1, otherwise y ij =0, but this does not mean that drug r i and disease d j There is no correlation between them, or there may be a potential correlation that has not yet been discovered. This study's work involves analyzing each drug r... i , discovered with r i Potentially related diseases. Finally, by fusing the drug-drug similarity network, the disease-disease similarity network, and the drug-disease association network, a drug-disease heterogeneous network was constructed.

[0052] Step S2: Based on the constructed drug-disease heterogeneous network, the low-dimensional embedding features of drug and disease nodes are encoded using a graph convolutional network. The drug-disease association score is predicted by decoding and then filled into the input drug-disease association matrix to obtain an improved drug-disease association matrix.

[0053] In some embodiments, based on the constructed drug-disease heterogeneous network, a graph convolutional network can be used to encode the low-dimensional embedding features of drug and disease nodes. By decoding and predicting the drug-disease association score, the input drug-disease association matrix can be filled to obtain an improved drug-disease association matrix.

[0054] As an example, the present step can include the following steps:

[0055] First, in order to solve the problem of imbalance between positive and negative samples in drug-disease association data in drug repositioning, the related information of drugs and diseases can be used to predict the potential indications of drugs to supplement the drug-disease association data. GCN is to learn the low-dimensional feature representation of nodes from irregular graph structure, each layer of which aggregates the neighbor node information of the target node, taking the output of the last layer as the input of the next layer, which is a process of recursive aggregation of neighbor features. In this work, GCN is introduced to learn the low-dimensional features of nodes by using the similarity and association information of drugs and diseases, which are used to predict new drug-disease associations, which improves the drug-disease association matrix from the perspective of biological network, which will be used as a pre-training step to predict the possibility of drug-disease association.

[0056] The association matrix G corresponding to the drug-disease heterogeneous network can be:

[0057]

[0058] and are the normalized drug similarity and disease similarity matrices, respectively, where ( or is the (i, j) entry of the similarity matrix) are the degree matrices of the drug and disease similarity matrices, respectively.

[0059] Given an adjacency matrix G, the general process of convolution operation based on GCN encoder can be described as:

[0060]

[0061] Here, H l+1 represents the embedding of nodes encoded at the l+1 layer, E is the degree matrix of the adjacency matrix G (E = diag(∑ j G ij )), H l represents the embedding of nodes encoded at the l layer, which serves as the input feature of the l+1 layer, W is a learnable weight matrix, and σ is a nonlinear activation function (such as the RELU activation function).

[0062] In the process of learning based on GCN structure, it is not enough to consider only the drug-disease association. The use of similarity information can enhance the learning of drug-disease association information, and the introduction of appropriate similarity information contribution degree can better learn the embedding representation of drugs and diseases. Therefore, in order to control the contribution degree of drug and disease similarity information, a similarity penalty factor is introduced, which can be represented as Then, the adjacency matrix of the input graph Y in GCN is represented by the following:

[0063]

[0064] follow The GCN recursively learns node features according to the rules, and the use of exponential linear activation units (ELUs) in graph convolutional layers not only prevents gradient explosion and vanishing but also accelerates the learning process. After L iterations (L = 1, 2, ... L), the GCN captures different structural information of heterogeneous networks at different layers. In the first layer, it captures information about neighbors directly connected to the target node; in the second layer, it captures information about second-order neighbors; and so on, recursively capturing information about higher-order neighbors. To enable the GCN to fully learn node features, this invention uses an attention mechanism to connect the embeddings of different layers learned by the GCN. Different attention weights are set according to the degree of attention given to different layers. The embeddings of the first layer are given the greatest attention, and the attention gradually decreases as the number of layers increases. The final embeddings of drugs and diseases are represented as follows: Here, β l It was initially initialized to 1 / (l+1). and These represent the final feature embeddings of drugs and diseases, respectively.

[0065] To supplement the drug-disease association matrix, the final drug and disease embeddings are fed into a bilinear decoder for predicting drug-disease links. Therefore, the reconstruction of the drug-disease association matrix can be used... This indicates that ρ is the sigmoid activation function, and W... ′ It is a trainable matrix. Entries y i ′ j Indicates drug r i and disease d j The predicted scores between.

[0066] Finally, the binary cross-entropy loss function is used as the objective function to continuously optimize the drug-disease association.

[0067]

[0068] (i,j) represents a drug-disease pair. This represents the ratio of positive to negative drug-disease pairs, used to balance the positive and negative sample data.

[0069] The pre-training is implemented by using GCN on the drug-disease heterogeneous network to complement the drug-disease association information to alleviate the sparsity problem of the data. The drug-disease association matrix models the relationship between drugs and diseases, that is, it implies whether a drug can treat a disease. If the entry of the matrix is 1, it means that the disease is the indication of the drug and is a known association that has been clinically proven. Conversely, for the entry of 0, it means that there may be a certain degree of potential association that has not yet been found. Here, GCN is used to preprocess unknown drug-disease associations to obtain more promising association information for the subsequent prediction task. A threshold θ is set to screen the highly confident drug indications. Specifically, if the prediction score of drug-disease is greater than or equal to θ, the original value is retained, otherwise it is set to 0, because a larger value means that there is a stronger association between the drug and the disease. After that, a preprocessed and completed drug-disease association matrix is obtained.

[0070] In some embodiments, based on the obtained drug-disease association matrix, a symmetric metric learning algorithm can be used to learn the embedding representation of drugs and diseases in a unified metric vector space from the perspective of drugs and diseases, and finally obtain potential new drug-disease associations based on the Euclidean distance.

[0071] In some embodiments, based on the obtained drug-disease association matrix, a symmetric metric learning algorithm can be used to learn the embedding representation of drugs and diseases in a unified metric vector space from the perspective of drugs and diseases, and finally obtain potential new drug-disease associations based on the Euclidean distance.

[0072] As an example, this step can include the following steps:

[0073] First, based on the known drug-disease association information, the drug-disease association can be represented as Γ = {(r, d) | r ∈ A r , d ∈ A d}, (r, d) represents a known drug-disease pair, A r = {r1, r2,... r M} represents a set of M drugs, A d = {d1, d2,... d N} represents a set of N diseases, and the association matrix Y ∈ {0, 1} M×N If there is an edge between drug r i and disease d j , the entry y ij of the matrix Y is 1, otherwise y ij = 0; N i + = {d j | d jand y ij =1} represents the disease set associated with drug r i =1} represents the disease set associated with drug r =1} represents the disease set associated with drug r i =1} represents the disease set associated with drug r

[0074] Second, for a given triple (r,d)∈Γ represents a known drug-disease association, which is regarded as a positive sample, represents a negative sample, is a disease that has no association with drug r. Drug-centered metric learning is used to learn the distance between drug and positive disease and the distance between drug and negative disease based on the push-pull mechanism. A margin value m is set, and m>0, represents the Euclidean distance.

[0075] For example, for a given triple (r,d)∈Γ represents a known drug-disease association, which is regarded as a positive sample, represents a negative sample, is a disease that has no association with drug r. In metric learning, distance is used to measure similarity, and closer distance means that two entities are more similar, indicating that the degree of association between drug and disease is higher. Drug-centered metric learning is used to learn the distance between drug and positive disease and the distance between drug and negative disease based on the push-pull mechanism. In order to learn the position of drug and disease in the metric space, a margin value m is set, and m>0, to ensure represents the Euclidean distance.

[0076] Third, for a given triple (d,r)∈Γ represents a known disease-drug association, which is regarded as a positive sample, represents a disease-drug negative sample, is a drug that has no association with disease d. According to the assumption that distance measures similarity, d and are not related. Based on drug-centered learning, disease-centered metric learning is performed, a margin value n is set, and n>0, After disease-centered metric learning, drugs associated with and not associated with the disease are separated.

[0077] For example, for a given triple (d,r)∈Γ represents a known disease-drug association, which is regarded as a positive sample, represents a disease-drug negative sample, is a drug that is not associated with disease d. According to the assumption of distance measure similarity, d and are not associated, therefore, based on drug-centered learning, disease-centered metric learning is performed, a margin value n is set, and n > 0 is ensured After learning in the disease-centered metric, drugs associated and not associated with the disease are separated, and drugs associated with the disease are closer in the metric space, while drugs not associated with the disease are farther away.

[0078] Fourthly, according to the inconsistency of drug-disease and disease-drug association strength, marginal personalized adaptive training is adapted to different drugs and diseases.

[0079] For example, according to the inconsistency of all drug-disease and disease-drug association strengths, marginal personalized adaptive training is adapted to different drugs and diseases. During learning, the margin value tends to be larger for less interactive drugs or diseases, thereby pushing non-interactive diseases or drugs farther away to prevent false recommendations.

[0080] It should be noted that traditional metric learning methods mainly consider drug-centered metrics, ignoring modeling drug-disease relationships from the perspective of disease, which leads to biased learning of drug and disease latent vector representations, limiting the predictive performance of the model. Therefore, drug and disease-centered metrics are considered in the enhanced symmetric metric learning algorithm of the present method, which not only considers the relationship between drugs and diseases, but also implicitly establishes the relationship between drugs and diseases, thereby enhancing the latent vector representation of drugs and diseases.

[0081] The goal of metric learning is to learn a metric function that pulls similar entities closer together and pushes dissimilar entities farther apart. For example, in a recommendation system that identifies items that a user is likely to like, metric learning assigns a smaller distance to users and items that have had interactions, and a larger distance to users and items that have unknown interactions. Similarly, it can be applied to the problem of predicting a drug's potential indications. The metric learning algorithm maps drugs and diseases to a unified vector space based on the association of drugs with different diseases, encoding the latent vector representation of drugs and diseases. In this way, the distance between drugs and diseases with known associations is closer, and the distance between drugs and diseases with no or unknown associations is farther apart. The likelihood of drug-disease association is measured by the position of drugs and diseases in the unified metric vector space. For a given drug, unverified diseases are sorted in descending order by prediction score, thereby achieving top-k disease recommendations for the drug.

[0082] The problem definition can include:

[0083] Based on the complemented drug-disease associations, the metric learning maps drugs and diseases into a unified n-dimensional latent vector space. In the unified metric vector space, is the latent vector representation of drug r, is the latent vector representation of disease d. The probability of the association between drug r and disease d is measured by the simple and effective Euclidean distance, defined as:

[0084]

[0085] where ||*||2 represents the L2 regularization, and it is known that the Euclidean distance of the drug-disease association calculation should be smaller than the distance without known associations.

[0086] The drug-centered metric learning can include the following steps:

[0087] Based on the complemented drug-disease association matrix, the drug-centered metric learning is defined for a given triple (r, d) e G represents a known association, which is considered as a positive sample, represents a negative sample, which is an unknown drug-disease pair, randomly selected. The metric learning is a distance-based similarity measure, and the distance is closer means the two entities are more similar. In drug-disease association prediction, the distance and the relevance are two opposite concepts, and the distance is closer means the drug-disease association is higher, and the drug has a more likely treatment behavior on the disease. To ensure better learning of the latent vector representation of drugs and diseases, a margin value (safety distance) m is set, and m > 0. The distance between drug r and negative disease is greater than the distance between drug r and positive disease d is guaranteed by the following formula:

[0088]

[0089] Figure 3 The drug-centered metric learning method in two-dimensional space is described, and the margin value is set to separate the positive drug-disease pairs and the negative drug-disease pairs. Specifically, the drug and the disease are represented as a latent vector in a drug-disease metric space, and if the drug and the disease are predicted to be associated, the gradient direction moves inward to limit the disease within the safety margin, otherwise the gradient direction moves outward to make the disease away from the drug until it exceeds the safety margin. Note that the positive disease is within the sphere centered on drug r, however, the negative disease is outside the sphere centered on drug r. This guarantees that the distance between the drug and the positive disease is smaller than the distance between the drug and the negative disease, and maximizes the relevance between the drug and the relevant disease.

[0090] Therefore, for drug-centered metric learning, the triplet loss is used as the objective function, which can be:

[0091]

[0092] In the formula [x] + =max(x,0) represents the standard hinge loss, which is a loss function widely used in the field of recommender systems.

[0093] Disease-centric metrics may include the following steps:

[0094] Drug-centered metric learning considers drug-disease interactions from the perspective of the drug itself, thus placing diseases associated with the targeted drug closer to them and unassociated diseases further away. However, accurately locating drugs and diseases in a unified metric vector space from only the drug's perspective to obtain their latent vector representations is insufficient. Therefore, we introduce a disease-centered metric from the disease's perspective to explore the relationship between drugs and diseases. In the unified metric vector space, similar to the drug-centered metric, for a targeted disease, drugs with known associations are located closer to it, and otherwise farther away. Based on the assumption of distance metric similarity, d and They are unrelated, therefore they should not be placed closer together, and it should be ensured that... Similarly, a marginal value n is set, and n > 0. The following formula is used to guarantee that the disease d is negative for the drug. The formula for determining whether the distance between two points is greater than the distance between the disease d and the positive drug r can be:

[0095]

[0096] Because Euclidean distance is symmetric, disease-centered learning can be replaced by... Figure 4 This paper demonstrates a symmetric metric learning method centered on drugs and diseases within an explicit treatment relationship. In summary, disease-centered metrics are symmetric to drug-centered metrics, but their mechanisms of action are exactly opposite. Disease-centered metrics predict related drugs from the perspective of the disease, utilizing a safety margin for gradient learning. The introduction of this method reinforces the push-pull learning mechanism, better clustering behaviorally similar drugs or diseases for more effective recommendations. Clearly, the goal of symmetric metric learning is to push unrelated drugs or diseases out of the sphere and pull related or potentially related drugs or diseases into the sphere. Therefore, the distance between known drug-disease pairs is smaller than the distance between unknown pairs.

[0097] Finally, the objective function for disease-centered metric learning is defined as follows:

[0098]

[0099] The purpose of the present application is to determine the relationship between drugs and diseases from the perspective of drugs and diseases, rather than directly using drug-centered metric learning.

[0100] The adaptive margin value can include the following steps:

[0101] Previous studies have found that one drug can treat multiple diseases, and one disease can also be treated by multiple drugs. Considering the inconsistency of all drug-disease and disease-drug association strengths, different margins are introduced to adapt to drugs and diseases individually. In order to better simulate the complex drug-disease relationship, we introduce personalized margins by adaptive learning training. In the process of learning, m r and n d are set as the margin values of drugs and diseases, respectively. In order to reduce variance, the margins m r and n d are biased to be as large as possible, especially for drugs or diseases with less interaction. Greater margins should be given to avoid overfitting, so as to push diseases or drugs with no interaction further away and prevent false recommendations. The adaptive margin constraint in the objective function can be expressed as:

[0102]

[0103] Optimization can include the following steps:

[0104] The number of unknown associations in the related data of drugs and diseases is significantly higher than the number of known associations. Therefore, the present method uses a negative sampling method to optimize the model. For each drug (disease), based on the known drug-disease association, P disease (drug) with no association with it is randomly selected as a negative sample in the training process, and P is the minimum value of the number of drugs and the number of diseases. By combining the drug-centered and disease-centered metric learning loss, the final loss function of RSML-GCN is:

[0105]

[0106] That is:

[0107]

[0108] Where m r ∈(0,1]. n d ∈(0,1].

[0109] l is used to prevent the marginal value from being too large to affect the performance of the prediction. In addition, to prevent the dimension explosion caused by the too wide distribution of data points, the l2 norm clipping is applied to the latent vector representation learned for drugs and diseases, which is limited within the Euclidean sphere with the size of l and ). The learning rate of the objective function is controlled using the AdaGrad optimizer to update the latent vectors until convergence. After the training process is completed, the Euclidean distance is used to calculate the association probability between drugs and diseases.

[0110] Experimental verification can include:

[0111] To verify the effectiveness of the present application, the present application uses ten-fold cross-validation to conduct experiments to test the prediction performance of the RSML-GCN method.

[0112] In the present application, ten-fold cross-validation is implemented to evaluate the performance of the model. All the verified known drug-disease associations are randomly divided into 10 approximately equal subsets. Each subset is in turn used as test data, while the remaining nine subsets are used as training data, and 10% of them are randomly selected as validation data. In each fold, a prediction model is constructed based on known associations on the training data, its parameters are adjusted by using the validation data, and then the prediction is implemented on the test data. The ten-fold cross-validation process is iterated 10 times, each time producing different test data. To ensure that the results of cross-validation are unbiased, 10 independent ten-fold cross-validation is implemented, and the final average result is used to measure the performance of the model.

[0113] To evaluate the accuracy of the RSML-GCN model, the Receiver Operating Characteristic Curve (ROC) is used, which is plotted by two variables of false positive rate and true positive rate, and thus the Area Under Curve (AUC) value is obtained, which can be used in binary classification problems and is widely used in previous studies. Because AUC cannot fully summarize the performance of the model, a second evaluation index (AUPR, Area Under PR) is introduced, which is used to accurately reflect the actual performance of the prediction model using the precision-recall (PR) curve. Because there are much more drug-disease pairs without associations than known drug-disease associations, for such a class-imbalanced data set, AUPR is used as the main measure here. Because the number of correctly predicted true positives reflects the ability of the model to identify positive and negative samples, especially when the number of positive samples is much smaller than that of negative samples, precision and recall are also used as measures of model performance.

[0114] To demonstrate the effectiveness of the proposed method in predicting drug-disease associations, RSML-GCN is compared with five state-of-the-art GCN and recommendation system-based drug repositioning methods: GRGMF, DRWBNCF, LAGCN, DRHGCN, and CMLDR. GRGMF builds a generalized matrix factorization model that obtains latent representations of each node by adaptively learning the neighborhood information of each node and introduces external similarity information to facilitate the prediction of latent links; DRWBNCF is a neural collaborative filtering method that proposes a new weighted bilinear graph convolution operation to integrate the information of known drug-disease associations, drug and disease neighborhoods, and neighborhood interactions into a unified representation to infer new latent drug-disease associations; LAGCN is a layer-attention GCN that uses GCN to learn drug and disease embeddings from drug-disease heterogeneous networks. It integrates the learned embeddings through an attention mechanism to predict new associations; DRHGCN is based on different network topology information of drugs and diseases in different domains and uses GCN to extract intra-domain and inter-domain feature information of drugs and diseases to discover new drug indications; CMLDR is a collaborative metric learning algorithm that predicts the probability of drug-disease associations by applying metric learning. The latent vector representations of drugs and diseases are learned based on known drug and disease related information, which are used to identify candidate drug-disease associations.

[0115] The comparison with other methods can include:

[0116] Comparative experiments were conducted on the dataset, and all experiments were performed 10 times with 10-fold cross-validation and the mean was taken to ensure smaller deviation of experimental results. The specific experimental results are listed in Table 2:

[0117] Table 2

[0118]

[0119] As shown in Table 2, the prediction results of multiple methods under 10 times 10-fold cross-validation are shown. The experimental results show that the RSML-GCN model shows good performance in the metrics and is superior to other methods. In terms of the main metric AUPR, RSML-GCN reaches the highest average value of 0.7941, which is 33.7% higher than GRGMF, and the average AUPR values of DRWBNCF, LAGCN, DRHGCN and CMLDR are 0.4992, 0.1562, 0.5480 and 0.2607, respectively. In addition, in terms of AUC value, RSML-GCN still performs best (the average AUC value is 0.9077), which is 0.20% higher than the second best method DRHGCN, while the average AUC values of DRWBNCF, GRGMF, LAGCN and CMLDR are 0.8642, 0.8994, 0.7874 and 0.7999, respectively. The drug-disease association prediction problem is regarded as a top-k recommendation problem, which involves recommending potential treatment diseases for a specific drug, so the top-k prediction results are used as an evaluation index. The evaluation metric methods precision (p@k) and recall (r@k) are widely used in recommendation systems.

[0120] The recall (a) and precision (b) values of different methods for recommending top-k drug indications on the Cdataset can be shown as Figure 5 Figure 5 The performance of different models in predicting top-k drug-disease associations on the Cdataset is shown. In terms of recall and precision, the RSML-GCN proposed in the application is superior to other models in r@5, r@10, p@5 and p@10. In addition, the recall (a) and precision (b) values of different methods for recommending top-k drug indications on the Fdataset can be shown as Figure 6 Figure 6 ​​In the results, it can be found that the recall and precision values of the top-k predictions of RSML-GCN on Fdataset also obtain the best performance, which is much better than the collaborative filtering, GCN-based and metric learning-based methods. The indicators of LAGCN are generally lower than those of other methods in these results, which may be due to the over-smoothing phenomenon of GCN caused by data imbalance. The matrix decomposition method is a common method for drug repositioning, and previous studies have shown that it has good performance. The prediction results of GRGMF are lower than those of RSML-GCN, which indicates that the metric learning method can make up for the defects of matrix decomposition to a large extent. For CMLDR, the results are significantly lower than those of RSML-GCN, which indicates that the metric learning method of increasing disease-centered auxiliary reuse is useful for improving drug-centered metrics. These results fully prove the effectiveness of the application in identifying drug-disease associations.

[0121] The ablation experiment can include:

[0122] To evaluate the performance of the model, in this part, the variant RSML of RSML-GCN is set. In RSML, only the reinforced symmetric metric learning method is used to predict the drug-disease association score, and the pre-training step of using GCN to reconstruct the drug-disease association matrix is removed. In order to check the contribution of the pre-training component, here RSML-GCN is compared with RSML based on Cdataset.

[0123] Based on the drug-disease association matrix, RSML respectively maps drugs and diseases to the metric vector space based on drugs and diseases, and learns the latent vector representation of drugs and diseases based on the push-pull mechanism. The Euclidean distance is used to obtain the latent treatment probability between drugs and diseases. Similarly, 10 ten-fold cross-validation is implemented in the variant RSML, all parameter settings remain unchanged, and the final average results are compared with RSML-GCN. The results of the ablation experiment under 10 ten-fold cross-validation can be shown in Table 3.

[0124] Table 3

[0125] Methods RSML RSML-GCN AUPR 0.8060 0.8580 AUC 0.9220 0.9308

[0126] As can be seen from Table 3, the addition of GCN pre-training in RSML-GCN supplements the drug-disease association data, which improves the prediction performance of the model, and the average AUPR value is 6.45% higher than that of RSML, the AUC value is almost the same, and other top-k prediction evaluation indicators are also significantly improved, such as Figure 7 The recall (a) and precision (b) values of RSML and RSML-GCN in recommending top-k drug indications can be shown in Figure 7The results show that the RSML-GCN model can effectively predict the drug-disease associations. The reason is that the GCN model learns a more comprehensive representation of drugs and diseases based on drug-disease associations, integrates the similarity information of drugs and diseases, obtains more highly confident drug-disease association information, and thus alleviates the imbalance problem between positive and negative samples, and improves the prediction potential of the downstream task. This indicates the reliability of the RSML-GCN model in predicting drug-related diseases.

[0127] The example analysis can include:

[0128] The effectiveness of the present application in predicting drug-disease associations has been demonstrated by ten-fold cross-validation. We will apply the present application to the prediction of unknown drug-disease associations. The present application selects three representative drugs, Atorvastatin Calcium, Etoposide, and Riluzole, for analysis. Atorvastatin Calcium is a commonly used lipid-lowering drug in clinical practice, mainly used for the treatment of mixed hyperlipidemia and hypercholesterolemia. These diseases have a high incidence and are difficult to detect and treat, and can potentially induce cardiovascular and cerebrovascular diseases. Therefore, the analysis of Atorvastatin Calcium is of great significance. Etoposide is a cell cycle-specific antitumor drug, mainly used for small cell lung cancer, acute leukemia, and malignant lymphoma. The common characteristic of tumor diseases is complexity and difficulty to cure, so it is very valuable to analyze whether Etoposide can treat other similar diseases in drug reuse. Riluzole is a central nervous system drug that plays a crucial role in the treatment of neurotrauma or degenerative diseases such as Alzheimer's disease, Parkinson's disease, and brain injury. These diseases can have a serious impact on patients, so it is necessary to analyze the new therapeutic potential of this drug that can treat multiple neurodegenerative diseases. In particular, the RSML-GCN is used to predict candidate diseases for the three drugs, Atorvastatin Calcium, Etoposide, and Riluzole. For each of the three drugs, all the predicted candidate disease scores are ranked in priority, and then we exclude all known drug-disease associations from the main dataset to generate a new drug-disease association top-ranking list. Finally, we use highly reliable sources and clinical trials (i.e., DrugBank (DB), CTD, PubChem, DrugCentral, and ClinicalTrials) as references to check the predicted drug-disease associations. The results of the case analysis are shown in Table 4.

[0129] Table 4

[0130]

[0131]

[0132] Table 4 lists the top 10 predicted diseases for the three drugs. Atorvastatin Calcium can also be shown to treat pulmonary diseases, and left ventricular dysfunction, and is also associated with renal failure, which is supported by CTD, ClinicalTrials and DrugCentral. The discovery of Etoposide can be verified in multiple clinical trials, which shows that Etoposide not only has good therapeutic effect on multiple tumors, but also can be used to treat skin rash and drug eruption. Riluzole is checked to be also associated with heart failure, drug-induced liver injury and arrhythmia. In summary, most of the predictions of RSML-GCN can be verified by reliable sources and clinical trials, and the case study illustrates the effectiveness of RSML-GCN in detecting new drug-disease associations.

[0133] In summary, the present application uses symmetry metric learning and graph convolution network to predict drug-disease associations. By using the known biomedical information of drugs and diseases, the graph convolution network is applied to the heterogeneous network to obtain some highly confident potential drug-disease associations, and through the supplement of association information, the symmetry metric learning algorithm is used to learn the relationship between drugs and diseases centered on drugs and diseases. The experimental results show that the drug repositioning prediction method can effectively predict new associated diseases of drugs.

[0134] The above examples are only used to illustrate the technical solutions of the present application, but not limit it; although the present application has been described in detail with reference to the foregoing examples, those skilled in the art should understand that they can still modify the technical solutions recorded in the foregoing examples, or make equivalent replacement for part of the technical features; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the scope of the technical solutions of the embodiments of the present application, and should be included in the protection scope of the present application.

Claims

1. A drug relocation method based on reinforcement symmetric metric learning and graph convolutional networks, characterized in that, Includes the following steps: Using biomedical information about drugs and diseases, drug similarity and disease similarity are calculated separately. Drug similarity networks and disease similarity networks are constructed based on the similarity information, and then connected with known drug-disease association networks to construct a drug-disease heterogeneous network. Based on the constructed drug-disease heterogeneous network, the low-dimensional embedding features of drug and disease nodes are encoded using graph convolutional networks. By decoding and predicting the drug-disease association score, the input drug-disease association matrix is ​​filled to obtain an improved drug-disease association matrix. Based on the obtained drug-disease association matrix, a symmetric metric learning algorithm is used to learn the embedding representation of drugs and diseases in a unified metric vector space with drugs as the center and diseases as the center. Finally, potential new drug-disease associations are obtained based on Euclidean distance. Based on the obtained drug-disease association matrix, a symmetric metric learning algorithm is used to learn the embedding representations of drugs and diseases in a unified metric vector space, centered on the drugs and diseases respectively. Finally, potential new drug-disease associations are obtained based on Euclidean distance, including: Based on known drug-disease association information, drug-disease associations can be represented as follows: , This indicates a known drug-disease pair. express A collection of drugs, express A set of diseases, association matrix If the drug and disease If there exists an edge between them, then the matrix Entries =1, otherwise =0; Indicates related to drugs A set of related diseases; This indicates that no drugs were identified. A set of related diseases; select positive drug-disease pairs, negative drug-disease pairs, and negative disease-disease pairs respectively; For a given triple , A known drug-disease association is considered a positive sample. To represent a negative sample, It is related to drugs Unrelated diseases; drug-centric metric learning, based on a push-pull mechanism to learn the distance between a drug and positive diseases and the distance between a drug and negative diseases, setting a marginal value. And make , , Represents Euclidean distance; For a given triple , A known disease-drug association is considered a positive sample. This represents a negative sample of a disease-drug pair. Is with disease Unrelated drugs; based on the assumption of similarity by distance measurement, and It is unrelated; based on drug-centered learning, disease-centered metric learning is performed, with a marginal value set. and make , After learning from disease-centric metrics, drugs associated with and unassociated with the disease will be separated. Based on the inconsistency in the strength of drug-disease and disease-disease associations, marginal personalized adaptive training learns to adapt to different drugs and diseases.

Images