A drug repositioning method fusing drug-target-disease correlation relationship
By constructing a drug-target-disease association graph and utilizing a graph convolution model with weighted attention, the drug-disease relationship is extracted, solving the problem of low success rate of drug relocation in existing technologies and achieving more efficient drug relocation.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- QINGDAO UNIV
- Filing Date
- 2024-03-25
- Publication Date
- 2026-07-21
Smart Images
Figure CN118155708B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of drug repositioning, and more specifically to a drug repositioning method that integrates drug-target-disease associations. Background Technology
[0002] Drug retargeting refers to the process of screening existing drugs suitable for other diseases and applying them to new diseases. Its advantages lie in significantly lower R&D costs and time compared to traditional drug development methods. This is because all existing drugs have already undergone extensive safety testing, and applying them to new diseases does not require the safety testing processes of Phase I and some Phase II clinical trials. Furthermore, with the development of technologies such as computer-aided drug design, network pharmacology, bioinformatics, and artificial intelligence, applying computer technology to drug retargeting research can pre-screen known drugs for their activity against other diseases, effectively improving drug development success rates, reducing costs, and accelerating the development process.
[0003] The affinity activity between drugs and targets can explain drug mechanisms and thus discover drugs. A large amount of experimental data provides accurate information on the relationship between drugs and targets. How to effectively use this information to assist in the prediction of drug-disease relationships, fully explore the existing relationship between drugs, targets and diseases, and use deep learning algorithms for drug relocation are also hot research topics.
[0004] Deep learning can avoid manual feature extraction to some extent, and therefore it has been applied to drug relocation studies based on drug target interactions. Hakime et al. proposed the DeepDTA model, which uses a deep neural network to perform end-to-end representation learning on one-dimensional sequence data of drugs and proteins. The WideDTA model represents drugs and proteins as biological words, and uses a convolutional neural network to process drug SMILES data and protein biological words. To extract drug target affinity activity features more efficiently, some graph convolutional network-based methods have also been attempted for affinity activity prediction. GraphDTA uses a graph convolutional network to process drug molecules, and still uses the sequence information of the target protein as input for one-dimensional convolution, and concatenates the convolution results of the two to predict affinity. Manoochehri et al. constructed a semi-bilateral graph based on drug target similarity and selected all nodes adjacent to the drug target to construct multiple subgraphs. After sorting the nodes in the graph, they were used as features input to a neural network for affinity activity prediction. However, relying solely on first-order neighbors cannot yield deeper information. Therefore, Tsubak et al. considered first-order and second-order neighbor information from the atomic graph, as well as protein sequence information, performing graph convolution and one-dimensional convolution respectively, while introducing an attention model for prediction. DTiGEMS+ defines multiple paths in the association graph for drug-target relationships, calculates the maximum and average values of these paths as features, and concatenates them to predict affinity activity. Sun et al. constructed drug similarity and target similarity matrices, fused the two matrices to construct a heterogeneous drug-target network graph, used a graph encoder to extract drug-target features separately, and optimized the model using a generative adversarial network.
[0005] However, existing deep learning methods have failed to delve into the role of target proteins in drug relocation methods, nor have they effectively utilized the interaction information of target proteins. Therefore, how to integrate protein information with drug and disease networks to construct multi-attribute and multi-connection networks to improve the success rate of drug relocation is also a problem that needs to be solved.
[0006] The relationship between drugs, targets, and diseases has not been effectively integrated. The affinity activity between drugs and targets can explain drug mechanisms and thus lead to drug discovery. Existing experimental data provides precise information on the relationship between drugs and targets. How to effectively utilize this information to assist in predicting drug-disease relationships, fully explore the existing relationships between drugs, targets, and diseases, and use deep learning algorithms for drug relocation are also current hot research topics.
[0007] Therefore, there is a need for a drug relocation method that can effectively integrate the relationship between drugs, targets, and diseases. Summary of the Invention
[0008] The main objective of this invention is to provide a drug repositioning method that integrates drug-target-disease associations to solve the problem that existing technologies cannot effectively integrate drug-target-disease associations.
[0009] To achieve the above objectives, the present invention provides a drug repositioning method that integrates drug-target-disease associations, specifically including the following steps: S1, protein feature extraction based on CNN-LSTM.
[0010] S2, establish the association graph structure of drug-target-disease.
[0011] S3 analyzes the interaction between different nodes in the association graph structure, proposes a graph convolution model based on weighted attention to extract the drug-disease relationship, and fuses the two relationships for drug relocation.
[0012] Furthermore, step S1 specifically includes the following steps: S1.1 employs a dense encoding method, which maps each amino acid triplet into a dense vector through training and constructs an amino acid triplet vocabulary.
[0013] S1.2 uses a convolutional neural network (CNN) and a long short-term memory (LSTM) model to extract protein features. For each row of the protein feature matrix, features are first obtained through a one-dimensional convolutional neural network, and then captured by an LSTM to capture the correlation between the amino acid triples before and after each row. The outputs of all LSTM units are then aggregated as the final protein features.
[0014] Furthermore, step S2 specifically includes the following steps: S2.1, firstly constructs a binary group through the drug-disease relationship. This links each drug to a characteristic of the disease, among which... Represents drug nodes, Represents disease nodes.
[0015] S2.2, Based on the drug-target and disease-target relationships, find all [targets / targets]. Related target set : And perform average pooling calculation on all target features in the set to obtain the features. and with splicing to form triple nodes ,in, In drug-target association, the relationship between drugs Connected target sets, In the disease-target association, the disease A set of connected target points.
[0016] S2.3 defines three types of connection relationships: strong connection, weak connection, and no connection. Weights are assigned to different types of connections using formula (1), and a new adjacency matrix of the drug-target-disease association graph is constructed based on the weights. Each node's features are formed by concatenating the drug, the disease, and the average of all associated protein features for that node, thus creating a feature matrix. ; (1).
[0017] Furthermore, step S3 specifically includes the following steps: S3.1, For network graphs, graph convolution is used for feature extraction. The convolution process of a single layer is defined as follows: ; (2); in, For a uniform adjacency matrix, Let I be the degree matrix and I be the identity matrix. For the first The feature matrix of the layer, For the parameters to be learned in each layer, This is the activation function.
[0018] S3.2, for the node feature input of each layer F is the feature dimension. An attention mechanism is introduced during the convolution process, and the relationship between the two nodes is calculated using formula (3): (3); in, For learnable parameters, This indicates the importance of node j to node i.
[0019] S3.3, normalize the attention coefficients and introduce connection weights into the attention calculation. The calculation formula is as follows: (4); in, It is an adjacency matrix.
[0020] S3.4, Features of each node Update using formula (5); (5); in, Represents the characteristics of neighboring nodes.
[0021] S3.5, for all node features after passing through the weighted attention map convolution model A multilayer perceptron is used to determine whether the included drugs and diseases are related. Finally, the drug-disease association prediction based on weighted graph regularized matrix factorization is fused with the drug-disease association prediction based on target information to perform drug relocation.
[0022] The present invention has the following beneficial effects: This invention first studies the problem of representing target protein sequence features, and proposes a protein representation method based on convolutional neural network and long short-term memory network (CNN-LSTM). A feature dictionary is constructed through amino acid sequence, and convolutional neural network is introduced into feature representation. LSTM is used to ensure that the features can cover sequence information of different scales.
[0023] This invention studies the construction of drug-target-disease association relationships and proposes a triple-based association graph construction method. It pools the protein features involved in drugs and diseases and concatenates them with drug and disease features to construct triples. It then integrates drug-disease, drug-target, and disease-target networks to calculate the connection relationships in the triple graph.
[0024] This invention analyzes the interactions between different nodes in the association graph and proposes a graph convolution model based on weighted attention. It calculates the contribution of all neighboring nodes of each node to the convolution process, thereby accurately extracting the drug-disease relationship and fusing the two association relationships to achieve drug relocation. Attached Figure Description
[0025] To more clearly illustrate the specific embodiments of the present invention or the technical solutions in the prior art, the drawings used in the description of the specific embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort. In the drawings: Figure 1 The diagram illustrates the feature extraction process of step S1 in a drug relocation method that integrates drug-target-disease associations according to the present invention.
[0026] Figure 2 The diagram showing the triplet node generation in step S2 of the present invention is illustrated.
[0027] Figure 3 The structure of the drug-target-disease association diagram in step S2 of the present invention is shown.
[0028] Figure 4 A flowchart of step S3 of the present invention is shown. Detailed Implementation
[0029] The technical solution of the present invention will now be clearly and completely described 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.
[0030] Since the process by which a drug acts in the body is a direct interaction between the drug and the target, introducing the target into the drug relocation method can improve the accuracy of drug relocation. Therefore, this patent utilizes the correlation between drug-target and disease-target, and integrates it with the drug-disease relationship to discover new diseases corresponding to the drug and achieve drug relocation prediction.
[0031] A drug repositioning method that integrates drug-target-disease associations includes the following steps: S1, protein feature extraction based on CNN-LSTM.
[0032] S2, establish the association graph structure of drug-target-disease.
[0033] S3 analyzes the interaction between different nodes in the association graph structure, proposes a graph convolution model based on weighted attention to extract the drug-disease relationship, and fuses the two relationships for drug relocation.
[0034] Specifically, step S1 includes the following steps: S1.1 Since the structures of many proteins remain unknown, extracting protein features through amino acid sequences ensures a more comprehensive coverage of protein nodes in the network. An amino acid sequence is a character sequence composed of 20 unique amino acid codes. This invention combines any three amino acids to form amino acid triplets, totaling 20. 3 =8000. However, encoding these 8000 triples using one-hot vectors would result in excessively sparse encoding, limiting the generalization ability of the encoding. Therefore, a dense encoding method is adopted, which maps each amino acid triple to a dense vector through training and constructs an amino acid triple vocabulary. Based on this vocabulary, a sequence of length N+2 can be encoded into a vector with N indices.
[0035] S1.2, to ensure that local and global information of the amino acid sequence is not lost, while preserving the relevance of the sequence context, a Convolutional Neural Network (CNN) and a Long Short-Term Memory (LSTM) model are used for protein feature extraction. For each row of the protein feature matrix, features are first obtained through a one-dimensional CNN, and then passed through an LSTM to capture the correlation between the amino acid triples before and after each row. The outputs of all LSTM units are then aggregated to obtain the final protein features. The feature extraction process is as follows: Figure 1 As shown.
[0036] After obtaining target features, this invention constructs a network graph structure based on three relationships: drug-disease (GD), drug-target (GT), and disease-target (DT) to achieve drug relocation.
[0037] Specifically, step S2 includes the following steps: S2.1, firstly constructs a binary group through the drug-disease relationship. This links each drug to a characteristic of the disease, among which... Represents drug nodes, Represents disease nodes.
[0038] S2.2, Based on the drug-target and disease-target relationships, find all [targets / targets]. Related target set : And perform average pooling calculation on all target features in the set to obtain the features. and with splicing to form triple nodes ,in, In drug-target association, the relationship between drugs Connected target sets, In the disease-target association, the disease A set of connected target points. For example... Figure 2 As shown. Therefore, the number of nodes in the network graph is Where M represents the total number of all drugs and N represents the total number of all diseases.
[0039] S2.3 defines three types of connection relationships: strong connection, weak connection, and no connection, as shown in Table 1.
[0040] Table 1 Description of Connection Relationships Weights are assigned to different types of connections using formula (1), and an adjacency matrix of a new drug-target-disease association graph is constructed based on these weights. Each node's features are formed by concatenating the drug, the disease, and the average of all associated protein features for that node, thus creating a feature matrix. , to obtain Figure 3 The structure of the drug-target-disease association diagram shown; (1).
[0041] After obtaining the association graph structure, in order to ensure that the convolution process can fully extract the information of adjacent nodes, this invention proposes a graph convolution model based on weighted attention to extract node features.
[0042] Specifically, step S3 includes the following steps: S3.1, for the adjacency matrix of the network graph, graph convolution is used for feature extraction. The single-layer convolution process is defined as follows: ; (2); in For a uniform adjacency matrix, Let I be the degree matrix and I be the identity matrix. For the first The feature matrix of the layer, For the parameters to be learned in each layer, This is the activation function.
[0043] S3.2, To ensure the robustness of the model's feature extraction, for each layer's node feature input... F is the feature dimension. An attention mechanism is introduced during the convolution process, and the relationship between the two nodes is calculated using formula (3): (3); in, For learnable parameters, This indicates the importance of node j to node i.
[0044] To accurately reflect the structure of the diagram, Only all neighboring nodes of node i are calculated. Since the adjacency matrix is a matrix with weighted properties, any two nodes are considered neighbors if there is a strong or weak connection between them.
[0045] S3.3, to ensure the consistency of attention parameters, the attention coefficients are normalized, and the connection weights are introduced into the attention calculation. The calculation formula is as follows: (4).
[0046] in, It is an adjacency matrix.
[0047] S3.4, the features of each node are updated using formula (5); (5).
[0048] in, Represents the characteristics of neighboring nodes.
[0049] S3.5, for all node features after passing through the weighted attention map convolution model A multilayer perceptron is used to determine whether the included drugs are associated with the disease. The process is as follows: Figure 4 As shown, the drug-disease association prediction based on weighted graph regularized matrix factorization is finally fused with the drug-disease association prediction based on target information to realize a deep learning-based drug relocation method.
[0050] Of course, the above description is not intended to limit the present invention, and the present invention is not limited to the examples given above. Any changes, modifications, additions or substitutions made by those skilled in the art within the scope of the present invention should also fall within the protection scope of the present invention.
Claims
1. A drug repositioning method integrating drug-target-disease associations, characterized in that, Specifically, the steps include the following: S1, protein feature extraction based on CNN-LSTM; S2, Establish the association graph structure of drug-target-disease; S3 analyzes the interaction between different nodes in the association graph structure, proposes a graph convolution model based on weighted attention to extract the drug-disease relationship, and fuses the two association relationships for drug relocation; Step S3 specifically includes the following steps: S3.1, For network graphs, graph convolution is used for feature extraction. The convolution process of a single layer is defined as follows: （2）； in, For a uniform adjacency matrix, This is the adjacency matrix of the drug-target-disease association graph. Let I be the degree matrix and I be the identity matrix. For the first The feature matrix of the layer, For the parameters to be learned in each layer, For activation functions; S3.2, for the node feature input of each layer F is the feature dimension. An attention mechanism is introduced during the convolution process, and the relationship between the two nodes is calculated using formula (3): （3）； in, For learnable parameters, This indicates the importance of node j to node i; S3.3, normalize the attention coefficients and introduce connection weights into the attention calculation. The calculation formula is as follows: （4）； in, It is an adjacency matrix; S3.4, Features of each node Update using formula (5); （5）； in, Represents the characteristics of neighboring nodes; S3.5, for all node features after passing through the weighted attention map convolution model A multilayer perceptron is used to determine whether the included drugs and diseases are related. Finally, the drug-disease association prediction based on weighted graph regularized matrix factorization is fused with the drug-disease association prediction based on target information to perform drug relocation.
2. The drug repositioning method based on the fusion of drug-target-disease associations according to claim 1, characterized in that, Step S1 specifically includes the following steps: S1.1 employs a dense encoding method, which maps each amino acid triplet into a dense vector through training and constructs an amino acid triplet vocabulary. S1.2 uses a convolutional neural network (CNN) and a long short-term memory (LSTM) model to extract protein features. For each row of the protein feature matrix, features are first obtained through a one-dimensional convolutional neural network, and then captured by an LSTM to capture the correlation between the amino acid triples before and after each row. The outputs of all LSTM units are then aggregated as the final protein features.
3. The drug repositioning method based on the fusion of drug-target-disease associations according to claim 1, characterized in that, Step S2 specifically includes the following steps: S2.1, firstly constructs a binary group through the drug-disease relationship. This links each drug to a characteristic of the disease, among which... Represents drug nodes, Represents disease nodes; S2.2, Based on the drug-target and disease-target relationships, find all [targets / targets]. Related target set : And perform average pooling calculation on all target features in the set to obtain the features. and with splicing to form triple nodes ;in, In drug-target association, the relationship between drugs Connected target sets, In the disease-target association, the disease A set of connected target points; S2.3 defines three types of connection relationships: strong connection, weak connection, and no connection. Weights are assigned to different types of connections using formula (1), and a new adjacency matrix of the drug-target-disease association graph is constructed based on the weights. Each node's features are formed by concatenating the drug, the disease, and the average of all associated protein features for that node, thus creating a feature matrix. ; （1）。