A drug-disease interaction prediction method fusing multi-neighborhood association information

Abstract

The application provides a drug-disease interaction prediction method fusing multi-neighborhood correlation information, first, drug and disease information in a drugomic database is preprocessed to construct a drug-disease correlation relationship network; second, node aggregation and linear fusion are introduced, and feature information of drugs and diseases in multiple neighborhood spaces is extracted in parallel; then, contrast learning is used to perform parallel fusion on single-domain features and multi-neighborhood features, and a hierarchical attention mechanism is used to aggregate feature representations to obtain general features of drugs and diseases; finally, the extracted feature information of drugs and diseases is sent into a classification model composed of a full connection neural network for training, a hybrid optimization strategy combining prediction loss and negative example generation loss is proposed to perform overall optimization on the model, the model is saved, and the relationship between drugs and diseases is predicted. The application has better performance when predicting the drug-disease correlation relationship, and reduces the dependence on prior knowledge.

Description

Technical Field

[0001] This invention relates to the fields of medical artificial intelligence, drug discovery, and natural language processing, specifically to a method for predicting drug-disease interactions by fusing multi-neighborhood association information. Background Technology

[0002] Drug development is a costly process. Due to inaccuracies in identifying drug response mechanisms, only 10% of drugs entering Phase I clinical trials are approved, with the remainder being rejected due to high toxicity or ineffectiveness. The high failure rate of clinical trials and the complexity of drug development not only prolong the development cycle but also increase costs. Improving the success rate and reducing costs are pressing issues for researchers. Drug repositioning, based on existing approved drugs, avoids complex clinical trials and can effectively improve the success rate, reduce costs, and shorten the development cycle. Drug-disease interaction studies play a crucial role in drug repositioning; researching the relationship between drugs and diseases is of great significance for new drug development and clinical guidance.

[0003] Currently, drug-disease interaction prediction has gradually evolved from traditional biological or clinical trial-based methods to computational methods. The accumulation of biological omics data has provided new prospects for computational drug-disease interaction prediction research. However, it is undeniable that existing computational research methods lack differentiation between data sources and data types in biological omics data, leading to the loss of some key information. Although some heterogeneous network-based methods have partitioned data types to some extent, these methods have also exposed many problems. For example, existing heterogeneous network-based methods only focus on data information in a single spatial domain, lacking information interaction between data domains, resulting in insufficient learning of data features and limiting model performance. In practical applications, these schemes rely on prior knowledge of drugs and diseases, resulting in poor model generalization ability and difficulty in guiding actual drug discovery. Therefore, designing drug relocation methods that can automatically integrate neighborhood features and have broad applicability is crucial. Summary of the Invention

[0004] The purpose of this invention is to propose a deep learning model based on multi-neighborhood spatial information, which automatically predicts the interaction between drugs and diseases, thereby improving prediction efficiency and reducing prediction costs.

[0005] To achieve the above objectives, the technical solution of this application is: a method for predicting drug-disease interactions by fusing multi-neighborhood association information, comprising:

[0006] Step 1: Preprocess drug and disease information from the omics database to obtain the drug-drug similarity network G. drug Disease-Disease Similarity Network G dis Heterogeneous network G of drug-disease associations;

[0007] Step 2: Introduce node aggregation and linear fusion to construct a single-neighborhood feature extraction module and a multi-neighborhood feature extraction module, and extract neighborhood feature information of drugs and diseases in multiple neighborhood spaces in parallel;

[0008] Step 3: Through contrastive learning, single-neighborhood features and multi-neighborhood features are fused in parallel, and a hierarchical attention mechanism is used to aggregate feature representations of the obtained drug and disease feature information;

[0009] Step 4: The drug and disease feature information is fed into a classification model composed of a fully connected neural network for training, and the model is optimized as a whole by a hybrid strategy that combines prediction loss and negative sample generation loss.

[0010] Step 5: Load the optimized model, input the drug and disease information to be predicted, predict the relationship between the drug and the disease, and output the prediction results.

[0011] Furthermore, step 1 specifically includes:

[0012] Step 1.1: Extract drug-related information from databases such as DrugBank and CTD, and extract disease-related information from the OMIM database and previous studies;

[0013] Step 1.2: Based on the chemical structure of the drugs' SMILES, the Tanimoto score is calculated using the Chemical Development Kit to obtain the similarity matrix S between the drugs. drug This leads to the drug-drug similarity network G. drug Adjacency matrix A drug That is, if the drug j In the similarity matrix S drug Chinese medicine is a drug i The topk nearest neighbors, then otherwise

[0014] Step 1.3: Based on the disease phenotype, use MimMiner to calculate and obtain the similarity matrix S between diseases. dis This leads to the disease-disease similarity network G. dis Adjacency matrix A dis That is, if the disease is j In the similarity matrix S dis The middle is the disease disi The topk nearest neighbors, then otherwise

[0015] Step 1.4: Based on the association between drugs and diseases, construct a heterogeneous network G for drug-disease association, whose adjacency matrix is ​​represented as A∈{0,1}, that is, if the drug... i and disease j If there is a relationship, then A ij =1; otherwise A ij =0.

[0016] Furthermore, step 2 specifically includes:

[0017] Step 2.1: Through the drug-drug similarity network G drug Disease-Disease Similarity Network G dis The heterogeneous network G of drug-disease associations initializes drug and disease features as follows:

[0018]

[0019] Where H 0 ∈R (N+M)×(N+M) N and M represent the number of drugs and diseases, respectively.

[0020] Step 2.2: See Figure 2 The single-neighbor feature extraction module is used to extract single-neighbor features of drugs and diseases;

[0021] Step 2.3: See Figure 3 The multi-neighborhood feature extraction module is used to extract multi-neighborhood features of drugs and diseases.

[0022] Furthermore, step 2.2 specifically includes:

[0023] Step 2.2.1: Use GCN to extract single-neighborhood feature information of drugs or diseases. GCN is defined as follows:

[0024]

[0025] Where A is the adjacency matrix, W represents the weight matrix, and D = diag(∑ j A ij ), where σ(·) is the ReLU activation function.

[0026] Step 2.2.2: Connect the drug-drug similarity network G drug Disease-Disease Similarity Network G dis As input to GCN, the embedding representations of drugs and diseases in the single-neighborhood feature extraction module are obtained:

[0027]

[0028] in, It is a single-neighbor feature of the l-th layer, derived from the single-neighbor feature of the drug in the l-th layer. Single-neighborhood features of the disease in layer l Composition. W l This represents the weight matrix of the l-th layer.

[0029] Furthermore, step 2.3 specifically includes:

[0030] Step 2.3.1: Due to the varying structural importance of different neighboring nodes to the central node and the influence of multi-head attention, this invention designs a multi-neighborhood feature extraction module based on a multi-space neighborhood attention mechanism. This module learns significant neighbor features in multiple neighborhood spaces using an attention mechanism. The multi-neighborhood feature extraction module uses a drug-drug similarity network G... drug Disease-Disease Similarity Network G dis Using a heterogeneous network G of drug-disease associations as input, the following formula is used to obtain the disease dis in the neighborhood. i Importance of drugs:

[0031]

[0032] Where σ(·) is the nonlinear activation function, W represents the weight matrix, and N h (dis) represents the h-hop disease neighborhood of a drug in a drug-disease association network. Similarly, the following formula can be used to obtain the drug's... j The importance of disease:

[0033]

[0034] Step 2.3.2: Based on the GCN used in Step 2.2.1, a vector linear aggregation module was added to extract multi-neighborhood feature information of drugs and diseases. For drugs... i and disease j Obtain the corresponding multi-neighborhood features respectively. and

[0035]

[0036]

[0037] Where ⊙ represents the vector multiplication operation, and W l It is the weight matrix of the l-th layer in the multi-neighbor feature extraction module.

[0038] Furthermore, step 3 specifically includes:

[0039] Step 3.1: Through contrastive learning, the single-neighborhood features and multi-neighborhood features obtained in Steps 2.2 and 2.3 are fused in parallel to obtain neighborhood features. After fusion, the drug features of the l-th layer are:

[0040]

[0041] Similarly, the disease characteristics of layer l can be obtained:

[0042]

[0043] Step 3.2: Use a hierarchical attention mechanism to combine the neighborhood features of each hidden layer to obtain the final embedded representation of the drug and the disease:

[0044]

[0045] Where β l ∈R is automatically learned and initialized by the neural network. H represents the contribution of embeddings from different convolutional layers to the final embedding, where L is the number of hidden layers. drug It is the final embedded representation of the drug entity, H dis It is the final embedded representation of the disease entity.

[0046] Furthermore, step 4 specifically includes:

[0047] Step 4.1: Input the final embedded representations of drugs and diseases obtained in Step 3.2 into a fully connected neural network, and use the sigmoid function as shown in the following formula to obtain the prediction score of the association between drugs and diseases.

[0048]

[0049] in It is a prediction probability score matrix, where the predicted scores for the association between drugs and diseases are composed of corresponding... Give;

[0050] Step 4.2: To obtain the best model performance, the model is optimized overall using a hybrid strategy that combines prediction loss and negative example generation loss from contrastive learning, and the best model (best_model) is saved as follows:

[0051]

[0052] Where (i,j) represents the drug i and disease j Yes, A +Let A represent the set of all known drug-disease association pairs. - H represents the set of all unknown drug-disease association pairs; H is the output of the current node. P and H N These are the average output vectors of the nodes in the positive and negative sets, respectively, in the contrastive learning process.

[0053] Furthermore, step 5 specifically includes:

[0054] Load the model best_model from step 4.2, input the drug-disease information from the test data into the model, determine whether there is an interaction relationship between the drug and the disease, and output the corresponding evaluation index.

[0055] By employing the above technical solutions, this invention achieves the following technical effects: It utilizes a deep learning model, leveraging information on drugs and diseases from a biological omics database, combined with their features in a multi-neighborhood space, to automatically predict drug-disease interactions. This effectively extracts common neighborhood feature information of drugs and diseases across multiple neighborhood spaces, resulting in higher accuracy in predicting drug-disease relationships, lower dependence on prior knowledge, and better robustness. This improves the efficiency and accuracy of drug-disease relationship prediction, effectively shortens the drug development cycle, significantly reduces the cost of new drug development, and provides an important foundation and guarantee for new drug development and clinical reuse. Attached Figure Description

[0056] Figure 1 Flowchart of a drug-disease interaction prediction method that integrates multi-neighborhood association information;

[0057] Figure 2 This is a structural diagram of the single-neighborhood feature extraction module;

[0058] Figure 3 Here is a structural diagram of the multi-neighborhood feature extraction module;

[0059] Figure 4 This is a graph showing the top 20 predicted Alzheimer's disease treatments in the examples. Detailed Implementation

[0060] The embodiments of the present invention are implemented under the premise of the technical solution of the present invention, and detailed implementation methods and specific operation processes are given. However, the protection scope of the present invention is not limited to the following embodiments.

[0061] Example 1

[0062] The present invention will be described in detail below with reference to embodiments, so that those skilled in the art can implement it after referring to this specification.

[0063] This embodiment uses Windows as the development environment, PyCharm as the development platform, and Python as the development language. It adopts the drug-disease interaction prediction method of the present invention that integrates multi-neighborhood association information. The following is an explanation using the prediction of potential treatment drugs for Alzheimer's disease as an example.

[0064] In this embodiment, a drug-disease interaction prediction method that integrates multi-neighborhood association information includes the following steps:

[0065] Given a disease name "Alzheimer's disease", remove all drug information related to Alzheimer's disease from the Gottlieb dataset, resulting in a total of 62 entries. Use the remaining known drug-disease associations as the training set and the unknown drug-disease associations as the candidate set.

[0066] For the data input model that does not contain Alzheimer's information, 10-fold cross-validation is used. 10% of the known related data and 10% of the unknown related data are randomly selected from the dataset as the test set, and the remaining data is used as the training set to train the model and save the model.

[0067] 124 pieces of information were randomly selected from unknown drug-disease associations and merged with 62 pieces of drug information related to Alzheimer's disease;

[0068] The merged data is used as input to load the saved model and obtain evaluation indicators and prediction scores for the interaction between drugs and Alzheimer's disease. The evaluation indicators include AUPR and AUROC.

[0069] Sort the predicted scores from step 4 in descending order to obtain the information on the top 20 drugs.

[0070] Based on the above steps, this invention compares the drug-disease relationship prediction performance with that of the NRLMF model, DRRS model, SCMFDD model, DRIMC model, LAGCN model, NIMCGCN model, and DRHGCN model. As shown in Table 1, the method proposed in this invention significantly outperforms other methods in terms of average AUPR and AUROC across multiple datasets.

[0071] Table 1. Comparison of drug-disease relationship prediction results of different models

[0072]

[0073] Simultaneously, the method of this invention was used to predict potential treatments for Alzheimer's disease, and the results were as follows: Figure 4As shown, four of the top five drugs have been proven to treat or alleviate Alzheimer's disease. Levodopa is the number one predicted drug for Alzheimer's treatment, confirmed in databases such as PubChem and CTD. Previous literature has confirmed that Amantadine can improve cellular activity in Alzheimer's patients. Previous studies have indicated that Memantine hydrochloride is an effective treatment option for Alzheimer's disease, Parkinson's disease, and other conditions. Rivastigmine tartrate has also been shown in relevant literature to effectively alleviate Alzheimer's symptoms.

[0074] The foregoing description of specific exemplary embodiments of the invention is for illustrative and explanatory purposes. These descriptions are not intended to limit the invention to the precise forms disclosed, and it will be apparent that many changes and variations can be made in accordance with the foregoing teachings. The exemplary embodiments were chosen and described in order to explain the specific principles of the invention and its practical application, thereby enabling those skilled in the art to implement and utilize various different exemplary embodiments of the invention, as well as various different choices and variations. The scope of the invention is intended to be defined by the claims and their equivalents.

Claims

1. A method for predicting drug-disease interactions by fusing multi-neighborhood association information, characterized in that, include: Step 1: Preprocess drug and disease information from the omics database to obtain the drug-drug similarity network G. drug Disease-Disease Similarity Network G dis Heterogeneous network G of drug-disease associations; Step 2: Introduce node aggregation and linear fusion to construct a single-neighborhood feature extraction module and a multi-neighborhood feature extraction module, and extract neighborhood feature information of drugs and diseases in multiple neighborhood spaces in parallel; Step 3: Through contrastive learning, single-neighborhood features and multi-neighborhood features are fused in parallel, and a hierarchical attention mechanism is used to aggregate feature representations of the obtained drug and disease feature information; Step 4: The drug and disease feature information is fed into a classification model composed of a fully connected neural network for training, and the model is optimized as a whole by a hybrid strategy that combines prediction loss and negative sample generation loss. Step 5: Load the optimized model, input the drug and disease information to be predicted, predict the relationship between the drug and the disease, and output the prediction results.

2. The drug-disease interaction prediction method fusion of multi-neighborhood association information according to claim 1, characterized in that, Preprocessing of drug and disease information from biological omics databases to obtain drug-drug similarity network G drug Disease-Disease Similarity Network G dis The heterogeneous network G of drug-disease associations specifically includes: Step 1.1: Obtain relevant information about the medication and the disease; Step 1.2: Based on the chemical structures of the drugs' SMILES, the Tanimoto score is calculated using the Chemical Development Kit to obtain the similarity matrix S between the drugs. drug This leads to the drug-drug similarity network G. drug Adjacency matrix A drug That is, if the drug j In the similarity matrix S drug Chinese medicine is a drug i The top k nearest neighbors, then otherwise Step 1.3: Based on the disease phenotype, use MimMiner to calculate and obtain the similarity matrix S between diseases. dis This leads to the disease-disease similarity network G. dis Adjacency matrix A dis That is, if the disease is j In the similarity matrix S dis The middle is the disease dis i The top k nearest neighbors, then otherwise Step 1.4: Based on the association between drugs and diseases, construct a heterogeneous network G for drug-disease association, whose adjacency matrix is ​​represented as A∈{0,1}, that is, if the drug... i and disease j If there is a relationship, then A ij =1; otherwise A ij =0.

3. The drug-disease interaction prediction method fusing multi-neighborhood association information according to claim 1, characterized in that, By introducing node aggregation and linear fusion, a single-neighborhood feature extraction module and a multi-neighborhood feature extraction module are constructed to extract neighborhood feature information of drugs and diseases in multiple neighborhood spaces in parallel, specifically including: Step 2.1: Through the drug-drug similarity network G drug Disease-Disease Similarity Network G dis The heterogeneous network G of drug-disease associations initializes drug and disease features as follows: Where H 0 ∈R (N+M)×(N+M) N and M represent the number of drugs and diseases, respectively; Step 2.2: Extract single-neighborhood features of drugs and diseases using the single-neighborhood feature extraction module; Step 2.3: Use the multi-neighborhood feature extraction module to extract multi-neighborhood features of drugs and diseases.

4. The drug-disease interaction prediction method fusing multi-neighborhood association information according to claim 3, characterized in that, The implementation process of the single-neighborhood feature extraction module specifically includes: Step 2.2.1: Use GCN to extract single-neighborhood feature information of drugs or diseases, wherein the GCN is defined as follows: Where A is the adjacency matrix, W represents the weight matrix, and the variable D = diag(∑ j A ij ), where σ(·) is the ReLU activation function; Step 2.2.2: Connect the drug-drug similarity network G drug Disease-Disease Similarity Network G dis As input to GCN, the embedding representations of drugs and diseases in the single-neighborhood feature extraction module are obtained: in, It is a single-neighbor feature of the l-th layer, derived from the single-neighbor feature of the drug in the l-th layer. Single-neighborhood features of the disease in layer l Composition; W l This represents the weight matrix of the l-th layer.

5. The drug-disease interaction prediction method fusing multi-neighborhood association information according to claim 3, characterized in that, The implementation process of the multi-neighborhood feature extraction module specifically includes: Step 2.3.1: Connect the drug-drug similarity network G drug Disease-Disease Similarity Network G dis Using a heterogeneous network G of drug-disease associations as input, the following formula is used to obtain the disease dis in the neighborhood. i Importance of drugs: Where σ(·) is the nonlinear activation function, W represents the weight matrix, and N h (dis) represents the h-hop disease neighborhood of a drug in a heterogeneous network of drug-disease associations; then, the following formula is used to obtain the drug... j The importance of disease: Step 2.3.2: For drugs i and disease j Obtain the corresponding multi-neighborhood features respectively. and Where ⊙ represents the vector multiplication operation, and W l It is the weight matrix of the l-th layer in the multi-neighbor feature extraction module.

6. The drug-disease interaction prediction method according to claim 1, characterized in that, By employing contrastive learning, single-neighborhood features and multi-neighborhood features are fused in parallel. A hierarchical attention mechanism is used to aggregate feature representations of the acquired drug and disease feature information, specifically including: Step 3.1: Through contrastive learning, single-neighborhood features and multi-neighborhood features are fused in parallel to obtain neighborhood features. After fusion, the drug features of the l-th layer are: The disease characteristics of layer l are: Step 3.2: Use a hierarchical attention mechanism to combine the neighborhood features of each hidden layer to obtain the final embedded representation of the drug and the disease: Where β l ∈R is automatically learned and initialized by the neural network. The contribution of embeddings from different convolutional layers to the final embedding is represented by L, where L is the number of hidden layers; H represents the number of hidden layers. drug It is the final embedded representation of the drug entity, H dis It is the final embedded representation of the disease entity.

7. The drug-disease interaction prediction method according to claim 1, characterized in that, The drug and disease feature information is fed into a classification model composed of a fully connected neural network for training. The model is then optimized overall using a hybrid strategy combining prediction loss and negative example generation loss, specifically including: Step 4.1: Input the final embedded representations of drugs and diseases into a fully connected neural network, and use the sigmoid function shown in the following formula to obtain the prediction score of the association between drugs and diseases; in It is a prediction probability score matrix, where the predicted scores for the association between drugs and diseases are composed of corresponding... Give; Step 4.2: Optimize the model overall using a hybrid strategy that combines prediction loss and negative example generation loss from contrastive learning, and save the best model (best_model) as follows: Where (i,j) represents the drug i and disease j Yes, A + Let A represent the set of all known drug-disease association pairs. - H represents the set of all unknown drug-disease association pairs; H is the output of the current node. P and H N These are the average output vectors of the nodes in the positive and negative sets, respectively, in the contrastive learning process.

8. The drug-disease interaction prediction method according to claim 1, characterized in that, Load the optimized model, input the drug and disease information to be predicted, predict the relationship between the drug and disease, and output the prediction results, specifically including: Load the model best_model, input the drug-disease information from the test data into the model, determine whether there is an interaction relationship between the drug and the disease, and output the corresponding evaluation index.

Images