A drug-disease association prediction method
By extracting neighborhood subgraphs of drug-disease associations and using node labels and graph neural networks, the problem of missing information in drug-disease association prediction is solved, achieving higher prediction accuracy and greater network applicability.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- TIANJIN UNIV
- Filing Date
- 2023-03-07
- Publication Date
- 2026-05-12
AI Technical Summary
Existing drug-disease association prediction methods based on graph neural networks fail to work effectively or have low prediction accuracy when drug and disease node information is missing.
By extracting neighborhood subgraphs of drug-disease associations and using node labels as initial node information, end-to-end drug-disease association prediction is performed using graph neural networks. Feature extraction and prediction are performed using graph convolutional layers and multilayer perceptrons, and the loss function is optimized using gradient descent.
It significantly improves prediction accuracy even when drug and disease node information is missing, and can make effective predictions on larger-scale drug-disease networks.
Smart Images
Figure CN116168849B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of bioinformatics and is a drug-disease association prediction method based on graph neural networks. The method extracts neighborhood subgraphs of drug-disease associations and uses node labels as node information for these subgraphs. Subsequently, using these neighborhood subgraphs as training samples and the presence or absence of drug-disease associations as labels, an end-to-end drug-disease association prediction framework is implemented. Background Technology
[0002] Despite continuous advancements in genomics, life sciences, and technology over the past few decades, drug discovery from scratch has become extremely time-consuming and expensive. Given the high attrition rate, enormous costs, and slow pace of new drug discovery and development, repurposing existing drugs to treat common and rare diseases is increasingly becoming an attractive proposition, as it involves using de-risked compounds that have the potential to reduce overall development costs and shorten development time.
[0003] In fact, many drug discoveries are not based on systematic methods, but rather stem from an understanding of drug pharmacology, retrospective analysis of clinical effects, and unfocused screening. Therefore, exploring more efficient and systematic methods to effectively accelerate drug research is urgent and important. High-precision computational methods can effectively select potential candidate associations from drug-disease relationships, narrowing the drug search space, reducing unnecessary experimental costs, and improving drug development efficiency. Accurately predicting drug-disease associations is the core issue of computational methods.
[0004] With the improvement of drug databases and breakthroughs in related algorithms, more and more association prediction methods have been proposed. Recently, graph neural networks have shown convincing performance in biomedical network analysis. Yu et al.[1] integrated known drug-disease associations, drug-drug similarities and disease-disease similarities into a heterogeneous network and performed graph convolution operations on the network to learn the embeddings of drugs and diseases. The embeddings of multiple graph convolutional layers were combined using the attention mechanism, and the unobserved drug-disease associations were scored based on the integrated embeddings. Zhao et al.[2] used graph representation learning techniques to learn the node embeddings in heterogeneous information networks from the perspective of topology and biology, and used a random forest classifier to predict unknown drug-disease associations.
[0005] Existing prediction methods all use graph neural networks to represent the node embeddings of drugs and diseases in the drug-disease network, and then use the node embeddings to predict drug-disease associations. However, when drug and disease node information is missing, existing prediction methods either fail to work or have low prediction accuracy, thus having certain limitations.
[0006] 1. Yu Z, Huang F, Zhao X, et al. Predicting drug–disease associations through layer attention graph convolutional network. Brief. Bioinform. 2021; 22:bbaa243
[0007] 2. Zhao BW, Hu L, You ZH, et al. HINGRL: predicting drug–disease associations with graph representation learning on heterogeneous information networks. Brief. Bioinform. 2022; 23: bbab515 Summary of the Invention
[0008] This invention provides a drug-disease association prediction method based on graph neural networks. This method not only works when drug and disease node information is missing, but also achieves significant performance improvements compared to mainstream prediction methods. This invention is achieved through the following technical solution:
[0009] A method for predicting drug-disease associations includes the following steps:
[0010] Step 1: Extract the neighborhood subgraph of drug-disease association.
[0011] (1) Create a bipartite graph with drugs and diseases in the drug-disease association network as nodes and the relationship between drugs and diseases as edges, and construct the adjacency matrix of the association network.
[0012] (2) Extract the h-hop neighborhood of relevant drug nodes and disease nodes;
[0013] (3) Merge the h-hop neighborhoods of drug nodes and disease nodes into a drug-disease related h-hop neighborhood subgraph;
[0014] (4) Extract the neighborhood subgraph corresponding to each association as positive samples for model training, and at the same time randomly select the same number of unrelated drug-disease pairs to generate negative sample test data, and divide the training set and test set.
[0015] Step 2: Construct the initial node features of the neighborhood subgraph
[0016] (1) Assign an integer label to each node in the extracted domain subgraph and distinguish the center node from other nodes;
[0017] (2) The central node is labeled with a 1 to represent drugs and diseases. For other nodes, the node is labeled using the DRN method based on dual-radius node labeling.
[0018] (3) After assigning labels to each node, construct the initial node features of the neighborhood subgraph;
[0019] Step 3, Learning the Graph Neural Network
[0020] (1) For drug node u and disease node v, assign the initial node features to the neighborhood subgraph G. h (u,v) extracts the representation h of each node in the subgraph through multiple graph convolutional layers. i For each node i, it is represented as follows:
[0021]
[0022] Among them, W (k) Let N(i) represent the trainable weight matrix of a certain layer, N(i) represent the set of neighboring nodes of node i, σ(·) represent the activation function, and h represent the training weight matrix of a certain layer. (k) h is the activation vector of the k-th layer. (0) =x0, where x0 represents the initial node feature;
[0023] (2) The neighborhood subgraph G h The node representations of each node in (u,v) are aggregated to obtain the representation h of the neighborhood subgraph. G As shown below:
[0024] h G =f({h i :i∈N h (u)∪N h (v)})
[0025] In the formula, f represents the join aggregation function, h i It is the representation of each node, N h (u) and N h (v) represents the h-hop neighborhood of drug u and disease v;
[0026] (3) The obtained neighborhood subgraph is input into a multilayer perceptron (MLP), and the predicted probability is output to predict the association between the drug and the disease, as shown below:
[0027] y (i,j) =w T σ(Wh G )
[0028] Among them, y (i,j) ∈(0,1) represents the probability of association, where w and W are parameters of the MLP, and the neighborhood subgraph h is... GThe representation is mapped to the predicted probability;
[0029] (4) Use gradient descent algorithm for iterative optimization. In each iteration, use label information and weighted binary cross-entropy loss function to calculate the loss of the neural network and update the neural network parameters according to the loss using gradient descent.
[0030] Furthermore, the one-hot encoded vectors of the node labels are used to construct the initial node features of the neighborhood subgraph.
[0031] Furthermore, the binary cross-entropy loss function is expressed as follows:
[0032]
[0033] Where n and m are the number of drugs and diseases, respectively, and λ is the penalty weight for a positive sample being judged as incorrect. It is the predicted probability score matrix, s + and s - These represent the number of positive and negative samples, respectively.
[0034] The beneficial effects of this invention are as follows: This invention extracts neighborhood subgraphs of drug-disease associations and uses node labels as the initial node information of these neighborhood subgraphs. Subsequently, using the neighborhood subgraphs as training samples and the presence or absence of drug-disease associations as labels, an end-to-end drug-disease association prediction framework is implemented. The subgraph extraction method fully learns the structural and hidden features of the surrounding nodes of the drug-disease association. Compared with mainstream prediction models, this invention can significantly improve prediction accuracy even when drug and disease node information features are missing. Furthermore, because it does not rely on drug and disease node information, this invention can perform predictions on larger-scale drug-disease networks. Attached Figure Description
[0035] Figure 1 Flowchart of implementation steps. Detailed Implementation
[0036] Example 1
[0037] Complex network relationships may exist between multiple drugs and multiple diseases; a single drug may cure multiple diseases, while a single disease may be cured by multiple drugs. Therefore, this invention is primarily used to predict new indications for a particular drug or corresponding drugs that cure a particular disease. The specific steps of this invention are as follows:
[0038] 1. Extracting the neighborhood subgraph of drug-disease associations
[0039] 1) For a bipartite graph of drug-disease association G=(U,V,E), where U={u1,…,u m} is the set of drug nodes, V = {v1, ..., v n Let} represent the set of disease nodes, and E be the set of drug-disease associations in the bipartite graph, where each association e(u,v)∈E. The adjacency matrix of the drug-disease associations in the bipartite graph is A. The total number of drug nodes is m, and the total number of disease nodes is n.
[0040] 2) For a drug node u, u is called the center node, and there are edges connecting the diseases to node u. The set of nodes whose distance to u (denoted as d(u,x)) is no greater than h is defined as the h-hop neighborhood of node u. Using N... h (u) represents the h-hop neighborhood of node u. Similarly, the h-hop neighborhood N of disease node v can be obtained. h (v), where the disease node v is also called the central node.
[0041] 3) Merge the closed subgraphs consisting of the h-hop neighborhoods of drug u and disease v into an h-hop neighborhood subgraph of the drug-disease association e(u,v), denoted as G. h (u,v). Formally, for the neighborhood subgraph G h For any node y in (u,v), d(y,u)≤h or d(y,v)≤h should be satisfied.
[0042] 4) Randomly extract some drug-disease associations as positive sample neighborhood subgraphs for model training. At the same time, in order to reduce the impact of data imbalance, randomly select the same number of unassociated drug-disease pairs to generate negative sample neighborhood subgraphs as negative test data, and divide the training set and test set into a ratio of 9:1.
[0043] 2. Constructing the initial node features of the neighborhood subgraph
[0044] 1) Transform the neighborhood subgraph G h Before inputting the graph neural network model (u,v), assign an integer label to each node in the subgraph.
[0045] 2) First, label the central node with 1, representing medicine and disease. For any other node i, label it using the Double Radius Node Labeling (DRNL) method, as follows:
[0046] f i (i)=1+min(d u ,d v )+(d / 2)[(d / 2)+(d%2)-1]
[0047] Here, f i (i) represents the label of node i, d u d is the distance from node i to drug u. vIt is the distance from node i to drug v, d = d u +d v d / 2 and d%2 are the integer quotient and the remainder when d is divided by 2, respectively. For nodes where d(i,u) = ∞ or d(i,v) = ∞, assign them an empty label 0.
[0048] 3) After assigning a label to each node, the one-hot encoded vector of the node label is used to construct the neighborhood subgraph G. h Initial node characteristics of (u,v).
[0049] 3. Learning Graph Neural Networks
[0050] 1) First, assign the initial node features to the neighborhood subgraph G. h (u,v) extracts the representation h of each node in the subgraph through multiple graph convolutional layers. i For each node i, it is represented as follows:
[0051]
[0052] Among them, W (k) Let N(i) represent the trainable weight matrix of a certain layer, N(i) represent the set of neighboring nodes of node i, and σ(·) represent the activation function, such as ReLU(·) = max(0,·). (k) h is the activation vector of the k-th layer. (0) =x0, where x0 represents the initial node characteristics.
[0053] 2) Transform the neighborhood subgraph G h The node representations of each node in (u,v) are aggregated to obtain the representation of the neighborhood subgraph, as shown below:
[0054] h G =f({h i :i∈N h (u)∪N h (v)})
[0055] Here, h G It is a neighborhood subgraph G h The representation of (u,v), f represents the join aggregation function, h i It is the representation of each node, N h (u) and N h (v) represents the h-hop neighborhood of drug u and disease v.
[0056] 3) Subsequently, the representation h of the obtained neighborhood subgraph G The input is a multilayer perceptron (MLP), and the output is a predicted probability to predict the association between the drug and the disease, as shown below:
[0057] y(i,j) =w T σ(Wh G )
[0058] Among them, y (i,j) ∈(0,1) represents the probability of the association existing. Here, w and W are parameters of the MLP, which represent the neighborhood subgraph h. G Mapped to predicted probabilities.
[0059] 4) Finally, gradient descent is used for iterative optimization. In each iteration, the loss of the neural network is calculated using the label information and a weighted binary cross-entropy loss function, and the neural network parameters are updated using gradient descent according to the loss. The specific loss function is as follows:
[0060]
[0061] Where m and n are the number of drugs and diseases, respectively, and λ is the penalty weight for a positive sample being incorrectly identified. Furthermore, This is the predicted probability score matrix. λ should be as close as possible to s. - / s + s + and s - These represent the number of positive and negative samples, respectively.
[0062] 4. Output prediction results
[0063] 1) Extract the neighborhood subgraphs associated with candidate drugs and all diseases.
[0064] 2) Assign node labels to all candidate associated neighborhood subgraphs, and transform all node labels into initial node features of the candidate associated neighborhood subgraphs.
[0065] 3) Input all candidate association neighborhood subgraphs into the trained drug-disease association prediction model, output the probability of association existence, and then sort these prediction results. The higher the ranking, the greater the probability of association existence, so as to facilitate subsequent experimental verification.
[0066] Given a known drug-disease network, this invention uses drugs and diseases as nodes, their associations as edges, and initial node information generated from node labels to construct a graph structure. A graph neural network is then used for iterative model training. Furthermore, this invention can still predict drug-disease associations even when node information is missing, achieving AUCs of 94.64%, 93.34%, 93.39%, and 91.86% on four datasets in the drug-disease association prediction domain (Cdataset, DNdataset, Fdataet, and LRSSL), respectively.
Claims
1. A method for predicting drug-disease association, comprising the following steps: Step 1: Extract the neighborhood subgraph of drug-disease association. (1) Create a bipartite graph with drugs and diseases in the drug-disease association network as nodes and the relationship between drugs and diseases as edges, and construct the adjacency matrix of the association network. (2) Extract the h-hop neighborhood of relevant drug nodes and disease nodes; (3) Merge the h-hop neighborhoods of drug nodes and disease nodes into a drug-disease related h-hop neighborhood subgraph; (4) Extract the neighborhood subgraph corresponding to each association as positive samples for model training, and at the same time randomly select the same number of unrelated drug-disease pairs to generate negative sample test data. Step 2: Construct the initial node features of the neighborhood subgraph (1) Assign an integer label to each node in the extracted domain subgraph and distinguish the center node from other nodes; (2) The central node is labeled with a 1 to represent drugs and diseases. For other nodes, the node is labeled using the DRN method based on dual-radius node labeling. (3) After assigning labels to each node, construct the initial node features of the neighborhood subgraph; Step 3, Learning the Graph Neural Network (1) For drug node u and disease node v, assign the initial node features to the neighborhood subgraph G. h (u,v) extracts the representation h of each node in the subgraph through multiple graph convolutional layers. i For each node i, it is represented as follows: Among them, W (k) Let N(i) represent the trainable weight matrix of a certain layer, N(i) represent the set of neighboring nodes of node i, σ(·) represent the activation function, and h represent the training weight matrix of a certain layer. (k) h is the activation vector of the k-th layer. (0) =x0, where x0 represents the initial node feature; (2) The neighborhood subgraph G h The node representations of each node in (u,v) are aggregated to obtain the representation h of the neighborhood subgraph. G As shown below: h G =f({h i :i∈N h (u)∪N h (v)}) In the formula, f represents the join aggregation function, h i It is the representation of each node, N h (u) and N h (v) represents the h-hop neighborhood of drug u and disease v; (3) The obtained neighborhood subgraph is input into a multilayer perceptron (MLP), and the predicted probability is output to predict the association between the drug and the disease, as shown below: y (i,j) =w T σ(Wh G ) Among them, y (i,j) ∈(0,1) represents the probability of association, where w and W are parameters of the MLP, and the neighborhood subgraph h is... G The representation is mapped to the predicted probability; (4) Use gradient descent algorithm for iterative optimization. In each iteration, use label information and weighted binary cross-entropy loss function to calculate the loss of the neural network and update the neural network parameters according to the loss using gradient descent.
2. The drug-disease association prediction method according to claim 1, characterized in that, Use the one-hot encoded vectors of node labels to construct the initial node features of the neighborhood subgraph.
3. The drug-disease association prediction method according to claim 1, characterized in that, The binary cross-entropy loss function is expressed as follows: Where n and m are the number of drugs and diseases, respectively, and λ is the penalty weight for a positive sample being judged as incorrect. It is the predicted probability score matrix, s + and s - These represent the number of positive and negative samples, respectively.