A drug target prediction method based on graph convolutional neural network

Abstract

The present application relates to the technical field of computer-aided drug design, and particularly relates to a drug target prediction method based on a graph convolutional neural network. The present application proposes a meta-path selection method based on a balanced spanning tree, which can adaptively select a suitable meta-path according to different path distances, so that the selected meta-path can cover more effective messages; an attention mechanism is used to fuse the neighborhood features obtained under different meta-paths, and the features of different types of nodes are differentially fused; the features of different near and far neighborhoods are obtained based on distance, which can better capture high-order topological structure features; and an MLP network considering inter-class confidence, path confidence and degree confidence is used for final drug target prediction, further fusing the information in the graph.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of computer-aided drug design, and in particular to a drug target prediction method based on a graph convolutional neural network. BACKGROUND

[0002] A drug target is a binding site of a drug and human body, and the drug target includes a receptor, a gene, an enzyme, an ion channel, a nucleic acid, a transporter, etc. A drug affects the change of a biological event by combining with the site, so as to realize the therapeutic effect of the drug. Identification of drug-target interactions (DTI) is the basis of modern drug discovery and development, and plays an important role in the process of drug discovery, drug side effect prediction, drug repositioning, and discovery of new targets interacting with existing drugs. However, traditional biochemical experiment methods for drug target discovery have problems such as long time consumption, expensive equipment, and large expenditure.

[0003] The emergence of large-scale biomedical data sets provides a data basis for using computer assistance to predict DTI, including PhID, Therapeutic Target Database (TTD), DrugBank, Binding DB, PharmGKB, ChEMBL, etc. Combined with computer technology, potential DTI can be quickly and low-cost identified.

[0004] A Chinese patent with publication number CN113053457A “A drug target prediction method based on a multi-channel graph convolutional neural network” proposes to use a multi-channel graph convolutional neural network, to obtain a PPMI matrix based on a random walk probability co-occurrence matrix, to use the PPMI matrix and a drug-drug with a self-loop as a convolution kernel, and to use the remaining subnetwork matrix as a feature for graph convolution, to capture graph network features from the global and local respectively.

[0005] A paper “NeoDTI: neural integration of neighbor information from a heterogeneous network for discovering new drug-target interactions Bioinformatics” uses data in a drug multi-source biomedical database to construct a heterogeneous network, and aggregates neighborhood information based on a heterogeneous graph convolutional neural network. In the aggregation process, different edges are processed by category, and edges of the same type are assigned weights by an edge weight mapping function in the aggregation process to differentiate learning of different features.

[0006] The above drug target prediction methods have made certain progress, but still have the following problems:

[0007] 1. Random walk algorithms are difficult to differentiate between different types of nodes in heterogeneous networks.

[0008] 2. It is difficult to capture higher-order topological features by using neighborhood information aggregation.

[0009] 3. Neighborhood aggregation is performed according to edge type, but the fusion between edge types is performed by summation, which cannot differentiate the characteristics of nodes of different types. Summary of the Invention

[0010] To address the aforementioned issues, this invention proposes a drug target prediction method based on graph neural networks. It extracts features from three levels: node fusion, type fusion, and distance fusion, fully extracting drug attribute features, disease attribute features, and spatial topological features of drug-disease association, thereby improving the accuracy of drug target prediction.

[0011] The technical solution of this invention is as follows:

[0012] A drug target prediction method based on graph convolutional neural networks includes the following steps:

[0013] S1. Construct a dataset based on known data, including:

[0014] Drug-Drug Attribute Similarity Matrix Where N d The number of drugs; drug-drug property similarity refers to the similarity between drugs in terms of chemical structure.

[0015] Drug Attribute Matrix Where g represents the dimension of the drug attribute vector, and each row is the attribute vector of a drug node;

[0016] Target-Target Attribute Similarity Matrix Where N t The number of targets is represented by the target-target attribute similarity, which is calculated using a normalized Smith-Waterman score.

[0017] Target attribute matrix h represents the dimension of the target attribute vector, and each row is the attribute vector of a target node.

[0018] Drug-target association Y(i,j) is defined as follows: when there is an interaction between the i-th drug and the j-th target, Y(i,j) = 1; when there is no interaction between the i-th drug and the j-th target, Y(i,j) = 0.

[0019] S2. Construct a heterogeneous network G(E,V,X) based on the dataset in S1. d ,X t), where V is a vertex set including two types of nodes of drugs and targets, E is an edge set based on S d , S t , Y structure, including three types of drug-drug, drug-target, target-target;

[0020] In the heterogeneous network G, two nodes are connected by different meta-paths, a meta-path Φ is defined as a path composed of R l represents the relationship between nodes A l and A l+1 ;

[0021] S3, learning node embedding based on graph convolutional neural network, specifically:

[0022] S31, for a drug node i in the heterogeneous network G, the set of all neighbor nodes connected by the drug node i through the k-distance meta-path Φ is denoted as N i Φ The balanced tree is generated with i as the root node, and the path from the root to the leaf node of the balanced tree is recorded as the alternative meta-path, then the set of meta-path neighbor nodes of the drug node i is defined as m represents the meta-path number, and the K-layer meta-path screening method is adopted, then the set of meta-path neighbor nodes of the drug node i is further defined as represents the mth meta-path with a distance of k, 1≤k≤K, so that the neighbor feature aggregation of the drug node i based on the meta-path is obtained as:

[0023]

[0024] wherein represents the learned embedding feature, d i is the feature vector corresponding to the drug node i, is a node-level attention vector, representing the importance between the neighbor node l and the drug node i under the meta-path

[0025]

[0026] wherein σ represents an activation function, is a weight vector, and || represents a splicing operation;

[0027] In the kth layer, given the meta-path set ​where 1≤m≤M, M is the total number of meta-paths, for drug node i, in the heterogeneous network G, the meta-paths of drug node i include two types of meta-paths connecting drug nodes-drug nodes and meta-paths connecting drug nodes-target nodes, the meta-paths connecting drug nodes-drug nodes are denoted as , and the meta-paths connecting drug nodes-target nodes are denoted as , the embedding of drug node i in each meta-path is obtained, and two types of features of node i in different meta-paths are obtained, respectively and

[0028] S32, the obtained two types of features are fused by using the average pooling method:

[0029]

[0030]

[0031] wherein represents the number of meta-paths connecting drug nodes-drug nodes in the kth layer, represents the number of meta-paths connecting drug nodes-target nodes in the kth layer.

[0032] the final feature of the kth layer is obtained by using the splicing method:

[0033]

[0034] the final feature set of drug node i based on distance is obtained

[0035] S33, the attention mechanism is used to fuse the feature set based on distance, and the final embedding vector of drug node i is obtained:

[0036]

[0037] wherein β k is an attention parameter, the importance ω k is obtained by normalization:

[0038]

[0039] the importance ω k is obtained by using a single-layer MLP to convert the hierarchical embedding:

[0040]

[0041] wherein q represents a mapping vector, tanh is a nonlinear activation function, W is a weight parameter of a single-layer MLP, and b is a bias;

[0042] S34, obtaining the embedding vector p corresponding to the target point node j by using the method of S31-S33 j ;

[0043] S4, performing binary classification by using the confidence-based MLP network, the input of the confidence-based MLP network being h obtained in S3 i and p j obtained in S1, regarding the known drug target associations with 1 in Y as positive examples and randomly selecting the same number of unknown drug target associations as negative examples, calculating the confidence vector c between drug i and target point j, and splicing the embedding vector h i and p j and the confidence vector c as the input of the MLP network, and the prediction result being represented as:

[0044] pro ji = g (h i , p j , c, Θ)

[0045] wherein pro ji represents the probability that the target point j is the target of the drug i, and Θ represents the parameters of the MLP network;

[0046] The confidence vector c includes inter-class transition confidence, path confidence and degree confidence, wherein the inter-class transition confidence is obtained according to the known drug target associations between different categories of drugs and different categories of target points, the path confidence is obtained by comparing the shortest path and the meta path between drug i and target point j, and the degree confidence is obtained by calculating the degrees of drug i and target point j in the graph;

[0047] The confidence-based MLP network is trained by using the obtained data, and the loss function uses cross-entropy loss:

[0048]

[0049] y ij represents whether there is an association between drug i and target point j, the positive class is 1, the negative class is 0, N represents the number of known drug target association pairs, and λ is a hyperparameter for controlling the strength of the regularization term;

[0050] Finally, a trained network model is obtained;

[0051] S5, performing drug target prediction by using the trained network model.

[0052] In the above scheme, the key point of the present application is that the present application proposes a meta-path selection method based on balanced spanning tree, which can adaptively select appropriate meta-paths according to different path distances, so that the selected meta-paths can cover more effective messages; the attention mechanism is used to fuse the neighborhood features obtained under different meta-paths, and the features of different types of nodes are differentially fused; the features of different near and far neighborhoods are obtained based on distance, which can better capture high-order topological structure features; the MLP network considering inter-class confidence, path confidence and degree confidence is used for final drug target prediction, further fusing the information in the graph.

[0053] The beneficial effects of the present application are:

[0054] (1) The present application uses a confidence-based MLP network to predict drug target correlation, introduces the concepts of path confidence, degree confidence and inter-class confidence, and further improves the model prediction effect.

[0055] (2) The present application screens the meta-paths by the balanced spanning tree method, saves the computing resources while ensuring that the obtained meta-paths are key meta-paths containing more potential information, which reduces the computational complexity of the model and improves the effect of the model.

[0056] (3) The present application extracts meta-path features under different distances based on distance levels, which can better capture high-order drug target topological structure features, improve the effectiveness of the learned embedding of the convolutional neural network, and improve the final prediction effect. BRIEF DESCRIPTION OF DRAWINGS

[0057] Figure 1 The present application is a process flowchart.

[0058] Figure 2 The present application is a network model diagram.

[0059] Figure 3 The present application is a meta-path screening example schematic diagram. DETAILED DESCRIPTION

[0060] The present application will be described in detail below in conjunction with the drawings and examples.

[0061] Example

[0062] The data used in this example comes from the Cdataset dataset, including 2353 known drug-disease associations, involving 663 FDA-approved drugs in the United States Food and Drug Administration, 409 registered diseases, drug chemical structure, pfam protein domain annotation, and gene semantic annotation information obtained from the DrugBank database, and disease phenotype data obtained from the OMIM database.

[0063] As Figure 1The specific steps of this example are as follows:

[0064] 1. Model input and network construction

[0065] Drug-drug chemical structure similarity: Drug-drug attribute similarity data is obtained from the Yamanishi dataset using the SIMCOMP tool. Given two drugs d i and d j , representing the chemical structure of the i-th and j-th drugs in the drug list, respectively, as a 2D graph with atoms as vertices and covalent bonds as edges, the similarity between the two drugs is calculated using the Jaccard similarity:

[0066]

[0067] |d i ∩d j | represents the number of atoms in the maximum common subgraph between d i and d j , |d i ∪d j | represents the sum of the number of d i and d j minus the number of atoms in the maximum common subgraph between d i and d j , i.e., the union of d i and d j . In order to facilitate computer input, all similarities between N d drugs are calculated, and the drug-drug attribute similarity matrix S is used to represent S d (i,j)∈[0,1) represents the structural attribute similarity between the i-th drug and the j-th drug, and the higher the value, the greater the similarity between the two drugs.

[0068] Drug attribute matrix N d represents the number of different drugs, and g represents the dimension of the drug attribute vector. Each row is an attribute vector of a drug node. Drug attributes include SMILES, molecular weight, octanol-water partition coefficient, number of H bond donors, number of H bond acceptors, half-life, melting point, boiling point, water solubility, pathway, and indication.

[0069] Target-target sequence similarity: Target-target attribute similarity data is obtained from the Yamanishi dataset, and the normalized Smith-Waterman score is used to calculate the similarity between two targets as follows:

[0070]

[0071] where ti and t j denote the ith and jth target in the target list, and SW(·,·) denotes the original Smith-Waterman score. In order to facilitate computer input, the similarity between N t targets is calculated in total, and a target-target attribute similarity matrix S is used, where S t (i,j) represents the sequence attribute similarity between the ith target and the jth target.

[0072] Target attribute matrix N t denotes the number of different drugs, and h represents the dimension of the target attribute vector. Each row is an attribute vector of a target node. The target attributes include species, source, proteome, action class, molecular formula, molecular weight, function, catalytic activity, activity regulation, GO molecular function, and GO biological process.

[0073] Known drug-target associations N d denotes the total number of drugs, and N t denotes the total number of targets. Y(i,j) represents the interaction relationship between the ith drug and the jth target. When the ith drug and the jth target have an interaction relationship, Y(i,j) = 1; when the ith drug and the jth target do not have an interaction relationship, Y(i,j) = 0.

[0074] A heterogeneous network G(E,V,X d ,X t ) is constructed using the drug-drug attribute similarity matrix S d , the target-target attribute similarity matrix S t , and the drug-target association matrix Y, where E is the edge set of the heterogeneous network, containing drug-drug, drug-target, and target-target edges, V is the vertex set of the heterogeneous network, containing drug and target nodes, X d denotes the drug attribute matrix, and X t denotes the target attribute matrix.

[0075] 2. Learning node embeddings based on graph convolutional neural networks

[0076] 2.1 Node feature extraction

[0077] In the heterogeneous graph G, two nodes can be connected through different semantic paths, called meta-paths. A meta-path Φ is defined as a path composed of R l denotes the nodes A l and A l+1The relationship between node i and its neighbor nodes based on the meta-path Φ. i Φ The set of all neighbor nodes connected by the meta-path Φ starting from node i, including node i itself.

[0078] The meta-path filtering method based on balanced spanning tree is used for k-layer meta-path filtering: for the kth layer, the filtered meta-paths with a distance equal to k are selected for node feature extraction. Taking the 2nd layer as an example, for any node u in the graph, its 2-hop neighborhood is obtained, and a balanced tree is generated with u as the root node. The paths from the root to the leaf nodes of the balanced tree are recorded as candidate meta-paths. All candidate meta-paths are sorted according to their frequencies, and the top ∝ are selected as the selected meta-paths for subsequent embedding learning, where ∝ is a hyperparameter that controls the number of meta-paths at each layer.

[0079] The neighbor feature of drug node i based on the meta-path is aggregated as follows:

[0080]

[0081] where represents the embedding feature learned by node i in the meta-path with a distance of k, m represents the meta-path number, d i is the feature vector corresponding to node i, is the node-level attention vector, which represents the importance between the neighbor node l and the center node i in the meta-path is calculated as follows:

[0082]

[0083] where σ represents the activation function, and relu is selected here, is a weight vector, and || represents the concatenation operation, represents the neighbor node set of node i in the meta-path .

[0084] At the kth layer, given the meta-path set After embedding the drug node i under each meta-path, the features of the node under different meta-paths can be obtained.

[0085] 2.2 Meta-path feature fusion

[0086] After obtaining the features extracted based on different meta-paths at the kth layer, these different types of features are fused. First, the meta-paths are classified. For drug node i, all meta-paths are divided into two categories. One category is the meta-path connecting drug nodes-drug nodes, denoted as denote a class of meta-paths connecting drug nodes-target nodes, and denote the features extracted from the same class of meta-paths, and the average pooling method is used for fusion.

[0087]

[0088]

[0089] wherein denotes the number of meta-paths connecting drug nodes-drug nodes in the kth layer, denotes the number of meta-paths connecting drug nodes-target nodes in the kth layer.

[0090] For different types of features after pooling, the final features of this layer are obtained by using the splicing method:

[0091]

[0092] After the above processing of each layer, the feature set of drug node i based on distance at different levels is obtained

[0093] 2.3 Distance Level Feature Fusion

[0094] The attention mechanism is used to fuse the spatial features extracted based on different distances obtained in the previous step to obtain the final embedding vector representation of drug node i:

[0095]

[0096] wherein β k is an attention parameter, and the importance ω k is normalized to obtain:

[0097]

[0098] The importance ω k is obtained by using a single-layer MLP to convert the hierarchical embedding:

[0099]

[0100] wherein q represents a mapping vector, tanh is a nonlinear activation function, W is a weight parameter of a single-layer MLP, and b is a bias.

[0101] Finally, the embedding vector h i corresponding to each drug node i is obtained, and similarly, the embedding vector p i corresponding to each target node j is obtained.

[0102] 3. Drug target prediction model based on confidence MLP network

[0103] The drug target prediction is regarded as a binary classification task. In order to improve the accuracy and introduce more external knowledge, a confidence-based MLP network is used for binary classification. After obtaining the embedding vectors of drugs and targets, the embedding vectors of drug targets will be input, and the confidence factors containing external knowledge will also be input, so as to train the confidence MLP network as a binary classifier g for drug target prediction.

[0104] The known drug target association of 1 in Y is regarded as a positive example, and an equal amount of unknown drug target association is randomly selected as a negative example. For drug i and target j, the confidence vector between them is calculated, and the embedding vectors h i and p j and the confidence vector c are spliced as the input of the MLP network, and the prediction result is represented as:

[0105] pro ji =g(h i ,p j ,c,Θ)

[0106] Where pro ji represents the probability that target j is the target of drug i, and Θ represents the parameters of the MLP network.

[0107] The confidence vector c is composed of three parts, including inter-class conversion confidence, path confidence and degree confidence.

[0108] The inter-class conversion confidence is calculated according to the known drug target association between different categories of drugs and different categories of targets. The drugs are sorted according to categories, and are divided into 14 categories according to ATC (Anatomical, Therapeutic, and Chemical Classification System) classification code, in the following order:

[0109] 1. Digestive tract and metabolism (A); 2. Blood and blood forming organs (B); 3. Cardiovascular system (C); 4. Skin diseases (D); 5. Reproductive urinary system and sex hormones (G); 6. Systemic hormone preparations (H); 7. Systemic anti-infective drugs (J); 8. Anti-tumor and immunomodulators (L); 9. Musculoskeletal system (M); 10. Nervous system (M); 11. Antiparasitic drugs (P); 12. Respiratory system (R); 13. Sensory organs (S); 14. Other drugs (V). A small amount of drugs correspond to two or more ATC codes, then the first one in the DrugBank website is used.

[0110] Proteins are classified by four different criteria, including classification by constituent components (classified into simple proteins, binding proteins, and derivative proteins), classification by molecular shape (globular proteins, fibrous proteins), classification by structure (monomeric proteins, oligomeric proteins, and polymeric proteins), and classification by function (structural proteins, regulatory proteins, contractile proteins, and antibody proteins), and the inter-class transition confidence is calculated respectively. The target points are classified according to the corresponding proteins.

[0111] Taking the inter-class transition confidence of the target points classified by function as an example, the inter-class transition confidence calculation includes the following processes: the drugs are classified into 14 categories according to the ATC classification standard, the target points are classified into 4 categories according to the function, and the inter-class edge count matrix T is established according to the existing drug target point association, T ∈ R 14×4 , T ij represents the number of existing associations between i-class drugs and j-class target points, and the following process is performed for each row to obtain the inter-class correlation matrix T':

[0112]

[0113] Each item in the matrix is the inter-class transition confidence value of the drug target points based on the function classification. Similarly, the inter-class transition confidence values of the drug target points based on the constituent components, the molecular shape, and the structural classification can be obtained, and the four values are spliced to obtain the inter-class transition confidence vector.

[0114] The path confidence is calculated by comparing the shortest path and the meta path between the drug i and the target j. The number of meta paths contained in the shortest path is calculated, and the path confidence value is obtained by weighted summation of the contained meta paths through attention value. If there are multiple shortest paths, all of them are selected and weighted average is performed.

[0115] The degree confidence is obtained by calculating the degree of the drug i and the target j in the graph. The more edges a node has, the more complete its information is, so the degree confidence is selected as one of the components of the confidence vector.

[0116] The obtained node embedding data is used as training data to train the MLP network, and the loss function uses cross-entropy loss:

[0117]

[0118] y ij represents whether there is an association between the drug i and the target j, the positive class is 1, the negative class is 0, N represents the number of known drug target association pairs, and λ is a hyperparameter used to control the strength of the regularization term.

[0119] Thus, a trained network model is obtained, and the trained model is used for drug target prediction.

Claims

1.A method for drug target prediction based on graph convolutional neural network, characterized in that, The method comprises the following steps: S1, constructing a data set based on known data, comprising: Drug-drug attribute similarity matrix where N d is the number of drugs, and drug-drug attribute similarity refers to the similarity between drugs in chemical structure; Drug property matrix where g represents the dimension of the drug property vector, and each row is a property vector of a drug node. Target-target attribute similarity matrix where N t is the number of targets, and target-target attribute similarity is computed by normalizing Smith-Waterman scores; Target point attribute matrix h represents the target point attribute vector dimension, and each row is an attribute vector of a target point node. Drug-target associations Y(i,j) = 1 if there is an interaction between the ith drug and the jth target; Y(i,j) = 0 if there is no interaction between the ith drug and the jth target. S2, constructing a heterogeneous network G(E, V, X based on the dataset in S1 d , X t ), wherein V is a vertex set, the vertex set includes two types of nodes of drugs and target points, E is an edge set, the edge set is constructed based on S d , S t , Y, including three types of drug-drug, drug-target, and target-target. In a heterogeneous network G, two nodes are connected by different meta-paths, a meta-path Φ is defined as a path consisting of R l represents the relationship between nodes A l and A l+1 ; S3, learning node embedding based on graph convolutional neural network, specifically: S31, for a drug node i in the heterogeneous network G, the set of all neighbor nodes connected by the drug node i through the meta-path Φ is denoted as N i Φ A balanced tree is generated with i as the root node, and the path from the root to the leaf node of the balanced tree is recorded as an alternative meta-path, then the set of meta-path neighborhood nodes of the drug node i is defined as m represents the meta-path number, and the K-layer meta-path screening method is adopted, then the set of meta-path neighborhood nodes of the drug node i is further defined as denotes the mth meta-path with a distance of k, 1≤k≤K, so that the neighborhood feature aggregation of the drug node i based on the meta-path is: wherein denotes the learned embedding features, d i is the feature vector corresponding to drug node i, is the node-level attention vector, denoting the importance between neighbor node l and drug node i in meta-path . where σ denotes an activation function, is a weight vector, and || denotes a concatenation operation; At the k-th layer, given a set of meta-paths After, 1≤m≤M, M is the total number of meta-paths, for drug node i, in the heterogeneous network G, the meta-paths of drug node i include two types of meta-paths connecting drug nodes-drug nodes and meta-paths connecting drug nodes-target nodes, the meta-paths connecting drug nodes-drug nodes are denoted as , and the meta-paths connecting drug nodes-target nodes are denoted as , the embedding of drug node i is performed under each meta-path, and two types of features of node i under different meta-paths are obtained, which are and S32, the obtained two types of features are fused by using average pooling: wherein represents the number of meta-paths connecting drug nodes-target nodes in the kth layer, represents the number of meta-paths connecting drug nodes-target nodes in the kth layer, The final feature of the kth layer is obtained by using splicing fusion: The final drug node i all levels of distance-based feature set is obtained S33, the feature set based on distance is fused by using attention mechanism, and the final embedding vector of the drug node i is: where β k is an attention parameter, normalized by the importance ω k of each layer: Importance ω k is obtained by transforming the metric using a single-layer MLP on the hierarchical embedding: Where q represents a mapping vector, tanh is a nonlinear activation function, W is a weight parameter of a single-layer MLP, and b is a bias; S34, adopt the method as S31-S33, obtain the embedding vector p corresponding to the target point node j j ; S4, binary classification using confidence-based MLP network, the input of the confidence-based MLP network is h i and p j , and the confidence factor, according to Y obtained in S1, the known drug target association of 1 in Y is regarded as a positive example, and the unknown drug target association is randomly selected as a negative example, for drug i and target j, the confidence vector between them is calculated, and the embedding vector h i and p j and the confidence vector c are spliced as the input of the MLP network, and the prediction result is represented as: pro ji = g(h i , p j , c, Θ) where pro ji denotes the probability that target j is a target of drug i, and Θ denotes the parameters of the MLP network. The confidence vector c includes inter-class conversion confidence, path confidence, and degree confidence, wherein the inter-class conversion confidence is obtained according to known drug-target associations between different categories of drugs and different categories of target points, the path confidence is obtained by comparing the shortest path and the meta path between the drug i and the target point j, and the degree confidence is obtained by calculating the degree of the drug i and the target point j in the graph; The confidence-based MLP network is trained by using the obtained data, and a cross-entropy loss is used as a loss function: y ij denotes whether there is an association between drug i and target j, 1 for positive class and 0 for negative class, N denotes the number of known drug-target association pairs, and λ is a hyperparameter used to control the strength of the regularization term. Finally, a trained network model is obtained; S5, using the trained network model to predict drug targets.

Images