Drug-disease association prediction method and model based on NMF and GT

Abstract

The invention relates to a drug-disease association prediction method and model based on NMF and GT. A drug-disease association prediction method based on NMF and GT comprises the steps of (1) obtaining a comprehensive similarity matrix of drugs and diseases according to drug fingerprint similarity, disease phenotype similarity and Gaussian kernel similarity of the drugs and the diseases; (2) extracting features of the comprehensive similarity matrix of the medicines and the diseases through non-negative matrix factorization to obtain characteristics of the medicines and the diseases; capturing a high-order structure relationship through a graph Transform to obtain similarity features of drugs and diseases; (3) respectively carrying out interactive fusion on the drug characteristics and the drug similarity characteristics and the disease characteristics and the disease similarity characteristics to obtain final drug and disease characteristic representations; and (4) carrying out association prediction according to the final characteristic representation of the medicine and the disease. The NMF and GT-based drug-disease association prediction method and model provided by the invention have good potential in the aspect of accurately predicting potential drugs.

Description

Technical Field

[0001] The present invention belongs to the field of biomedical data processing, and in particular relates to a drug-disease association prediction method and model based on NMF and GT. Background Art

[0002] Drug-disease association prediction plays a vital role in biomedical research, helping to identify potential drug therapeutic targets, accelerate the drug discovery process, and reduce the cost of clinical trials. With the development of big data and computing technology, computational methods are increasingly becoming a key tool for drug discovery. Progress in this field can not only promote the development of personalized medicine, but also provide new ideas for the treatment of complex diseases.

[0003] In recent years, many researchers have made significant progress in the field of drug-disease association prediction. Multiple research frameworks, such as DDAGDL and HINGRL by Zhao et al., DRWBNCF by Meng et al., DRHGCN by Cai et al., deepDR by Zeng et al., and HNet-DNN by Liu et al., have adopted different deep learning and graph representation techniques to improve the prediction accuracy of drug-disease association by integrating heterogeneous information networks, neighborhood interactions, and multimodal features. These studies provide theoretical support and practical basis for drug discovery and repositioning, and demonstrate the potential for broad application in the biomedical field.

[0004] Although the above studies have made significant progress, there are still some limitations. Existing drug-disease association prediction methods mainly focus on feature extraction of local or single similarity networks, and fail to fully integrate global information and high-order structural relationships from multiple networks. Although these methods can predict drug-disease associations to a certain extent, they have limitations in information fusion and modeling of complex structural relationships, resulting in limited improvement in prediction accuracy. Many studies have failed to effectively combine multiple feature extraction techniques, resulting in insufficient generalization ability of the model and difficulty in adapting to different diseases or drug types. In addition, some methods do not effectively utilize the biological knowledge of drugs and diseases, such as the chemical molecular structure information of drugs and the phenotypic information of diseases, which limits their performance improvement.

[0005] In order to alleviate the above challenges, the present invention proposes a drug-disease association prediction method and model based on NMF and GT, which is a DDA-NG model combining non-negative matrix factorization (NMF) and graph Transformer, which effectively alleviates the shortcomings of existing methods in information fusion and global structure modeling, and demonstrates its ability to discover potential drugs for diseases. Summary of the invention

[0006] The first invention objective of the present invention is to provide a drug-disease association prediction method based on NMF and GT, which has a good ability to discover potential drugs for diseases.

[0007] In order to achieve the above purpose, the technical solution adopted is:

[0008] A drug-disease association prediction method based on NMF and GT, comprising the following steps:

[0009] (1) Based on the drug fingerprint similarity, disease phenotype similarity, and Gaussian kernel similarity between drugs and diseases, similarity information is fused to obtain a comprehensive similarity matrix of drugs and a comprehensive similarity matrix of diseases;

[0010] (2) extracting latent features of the drugs and diseases from the comprehensive similarity matrix of the drugs and the comprehensive similarity matrix of the diseases by non-negative matrix decomposition to obtain drug features and disease features;

[0011] And through the graph transformer, we capture its high-order structural relationship to obtain drug similarity features and disease similarity features;

[0012] (3) After interactively fusing the drug features and drug similarity features, the disease features and disease similarity features, respectively, a final drug feature representation and a final disease feature representation are obtained;

[0013] (4) performing association prediction based on the final feature representation of the drug and the final feature representation of the disease.

[0014] Furthermore, in step (1), the Gaussian kernel similarity r between the drugs i and r j It is expressed as follows:

[0015] RGSM(r i ,r j )=exp(-γ r ||IP(r i )-IP(r j )|| 2 );

[0016]

[0017] Among them, RGSM∈R m×m , IP(r i ) indicates drug r i The corresponding column in the incidence matrix, γ r ' is the normalized kernel bandwidth;

[0018] Gaussian kernel similarity d between diseases i and d jIt is expressed as follows:

[0019] GSM(d i ,d j )=exp(-γ d ||IP(d i )-IP(d j )|| 2 );

[0020]

[0021] Where DGSM∈R n×n , IP(d i ) indicates disease i The corresponding column in the incidence matrix, γ d '=1.

[0022] Furthermore, in step (1), the comprehensive similarity matrix IRSM (r i ,r j )∈R m×m and the comprehensive similarity matrix IDSM of diseases (d i ,d j )∈R n×n It is expressed as follows:

[0023]

[0024] Among them, RFSM(r i ,r j ) is the drug fingerprint similarity, DPSM(d i ,d j ) is the disease phenotype similarity, RGSM (r i ,r j ) is the drug Gaussian kernel similarity, DGSM(d i ,d j ) is the disease Gaussian kernel similarity.

[0025] Furthermore, in the step (2), the process of non-negative matrix decomposition is: after decomposing a non-negative matrix into two non-negative matrices U and V, an iterative method is used to perform alternating optimization to solve, U and V are gradually updated, and the difference between the product of the original matrix and the decomposed matrix is ​​minimized through an iterative optimization process to obtain drug characteristics and disease characteristics.

[0026] Furthermore, in step (2), the Transformer network is derived from the similarity network SN r and SN d Extract the embedding representation of the node;

[0027] The process of obtaining drug similarity features is as follows: To define the update process of node features of the pth GT layer, a similarity network SN is used. r , this process is expressed by the following formula:

[0028] f 0 =W 0 SN r +b 0 ;

[0029]

[0030] Where, d' gt represents the dimension of each attention head, j∈N i Represents the adjacent nodes of node i; And they represent f 0 The i-th and j-th columns of ; is the multi-head attention score, which captures both the similarity and correlation between Q and K;

[0031]

[0032] In the formula, represents a learnable matrix, passes the output of the attention mechanism to the feedforward network, and generates the final node representation of the (p+1)th layer; after the pth gt After the layer is updated, the drug similarity feature representation is obtained

[0033] Furthermore, in the step (3), the interactive fusion process is: superimposing the drug features and drug similarity features, the disease features and disease similarity features respectively, to obtain the initial drug embedding and the initial disease embedding, and then passing through a Transformer encoder with a multi-head attention mechanism to obtain the final drug feature representation and the final disease feature representation.

[0034] Furthermore, in the step (4), a three-layer MLP with a dropout layer and a ReLU activation function is used to output the predicted association score.

[0035] The second inventive object of the present invention is to provide a drug-disease association prediction model based on NMF and GT, which combines non-negative matrix factorization (NMF) and graph Transformer, can integrate information at different feature levels, improves the model's adaptability to diversified data, and has excellent classification capabilities.

[0036] In order to achieve the above purpose, the technical solution adopted is:

[0037] A drug-disease association prediction model based on NMF and GT, including:

[0038] Similarity information module: obtain drug fingerprint similarity, disease phenotype similarity, and Gaussian kernel similarity between drugs and diseases, perform similarity information fusion, and obtain the comprehensive similarity matrix of drugs and the comprehensive similarity matrix of diseases;

[0039] Non-negative matrix decomposition feature extraction module: extracts drug and disease features through non-negative matrix decomposition to obtain drug features and disease features;

[0040] Graph Transformer feature extraction module: captures high-order structural relationships through graph transformers to obtain drug similarity features and disease similarity features;

[0041] Interactive fusion module: Based on the attention mechanism, the drug and disease information obtained by the NMF feature extraction module and the graph Transformer feature extraction module are combined to obtain the final feature representation of the drug and the final feature representation of the disease;

[0042] Association prediction module: performs association prediction based on the final feature representation of the drug and the final feature representation of the disease.

[0043] The third inventive object of the present invention is to provide a computer device for drug-disease association prediction with good accuracy and adaptability.

[0044] In order to achieve the above purpose, the technical solution adopted is:

[0045] A computer device includes a processor and a memory, wherein the memory stores computer executable instructions that can be executed by the processor, and the processor executes the computer executable instructions to implement the above-mentioned drug-disease association prediction method.

[0046] The fourth invention object of the present invention is to provide a computer-readable storage medium that can be used for drug-disease association prediction

[0047] In order to achieve the above purpose, the technical solution adopted is:

[0048] A computer-readable storage medium stores computer-executable instructions. When the computer-executable instructions are called and executed by a processor, the computer-executable instructions prompt the processor to implement the above-mentioned drug-disease association prediction method.

[0049] Compared with the prior art, the present invention has the following beneficial effects:

[0050] The present invention proposes a drug-disease association prediction method and model based on NMF and GT, which is a DDA-NG model that combines non-negative matrix factorization (NMF) and graph Transformer. Among them, the NMF module extracts potential feature representations by decomposing the association matrix between drugs and diseases; the GT module uses a multi-head attention mechanism to extract high-order features from the similarity network and deeply explore complex relationships. The DDA-NG framework fuses the features from the two modules to generate richer feature representations and enhance the diversity of feature expressions. This interactive fusion process is achieved through the Transformer encoder, so that the association information between drugs and diseases can be fully utilized. Finally, the model outputs the association prediction score through a multi-layer perceptron (MLP), thereby providing a more accurate drug-disease association prediction. The design of the overall framework of the present invention aims to improve the prediction performance and provide effective support for drug discovery and disease diagnosis in the biomedical field. The advantages of this study are as follows:

[0051] (1) The present invention integrates multi-source data, especially considering drug molecular structure information and disease phenotype information, and comprehensively models the drug-disease association through two key modules, thereby improving the accuracy and robustness of the prediction.

[0052] (2) The present invention proposes a method that integrates NMF and improved graph Transformer, which enables the model to integrate information at different feature levels and improves the model's adaptability to diversified data.

[0053] (3) The performance of the DDA-NG framework proposed in this paper is significantly better than that of existing advanced models on three different datasets. BRIEF DESCRIPTION OF THE DRAWINGS

[0054] Figure 1 This is the overall framework diagram;

[0055] Figure 2 The ablation experiment results. DETAILED DESCRIPTION

[0056] In order to further illustrate the drug-disease association prediction method and model based on NMF and GT of the present invention and achieve the intended purpose of the invention, the following is a detailed description of the specific implementation method, structure, characteristics and efficacy of the drug-disease association prediction method and model based on NMF and GT proposed in accordance with the present invention in combination with the preferred embodiments. In the following description, different "one embodiment" or "embodiment" does not necessarily refer to the same embodiment. In addition, specific features, structures or characteristics in one or more embodiments may be combined in any suitable form.

[0057] The following will further introduce in detail a drug-disease association prediction method and model based on NMF and GT in conjunction with specific embodiments of the present invention:

[0058] In the biomedical field, predicting the relationship between drugs and diseases is of great significance. However, traditional wet lab experiments are time-consuming and expensive, and often fail to meet the strict requirements of prediction accuracy and efficiency. Therefore, the present invention proposes a method and model (DDA-NG) that integrates non-negative matrix factorization (NMF) and graph transformer (GT) networks to predict the association between drugs and diseases. The model extracts the characteristics of drugs and diseases through non-negative matrix decomposition, capturing the global structural information in the drug-disease relationship; then effectively fuses multiple similar information, further captures its high-order structural relationship through graph transformer, and finally fuses the features obtained from the two modules to generate a richer feature representation, thereby improving the diversity of feature expression and the accuracy of association prediction. A large number of experimental results show that DDA-NG has achieved significant performance improvements on multiple drug-disease association prediction datasets, demonstrating that the method has a good ability to discover potential drugs for diseases. Specific embodiments are as follows:

[0059] Example 1.

[0060] The specific steps are as follows:

[0061] A. Materials and Methods

[0062] 1. Overall framework

[0063] The DDA-NG model framework proposed in this paper adopts an integrated multi-level structure to effectively predict the association between drugs and diseases. It mainly includes three modules: NMF feature extraction module, graph Transformer feature extraction module and association prediction module. The overall framework is shown in Figure 1 . Specifically:

[0064] Drug-disease association prediction models, including:

[0065] Similarity information module: obtain drug fingerprint similarity, disease phenotype similarity, and Gaussian kernel similarity between drugs and diseases, perform similarity information fusion, and obtain the comprehensive similarity matrix of drugs and the comprehensive similarity matrix of diseases;

[0066] Non-negative matrix decomposition feature extraction module: extracts drug and disease features through non-negative matrix decomposition to obtain drug features and disease features;

[0067] Graph Transformer feature extraction module: captures high-order structural relationships through graph transformers to obtain drug similarity features and disease similarity features;

[0068] Interactive fusion module: Based on the attention mechanism, the drug and disease information obtained by the NMF feature extraction module and the graph Transformer feature extraction module are combined to obtain the final feature representation of the drug and the final feature representation of the disease;

[0069] Association prediction module: performs association prediction based on the final feature representation of the drug and the final feature representation of the disease.

[0070] 2. Dataset

[0071] In this experiment, three datasets are used to evaluate the performance. B-dataset, C-dataset, and F-dataset are collected from previous studies. The positive samples in the experiment are known association pairs, and the negative samples are randomly generated from unconfirmed association pairs. The statistics of the datasets are shown in Table 1.

[0072] Table 1 Dataset statistics

[0073]

[0074] 3. Methods

[0075] The present invention proposes a drug-disease association prediction method based on NMF and GT, which combines the DDA-NG model framework and uses non-negative matrix factorization (NMF) and graph transformer. NMF decomposes the association matrix between drugs and diseases to extract potential feature representations; GT uses a multi-head attention mechanism to extract high-order features from similarity networks and deeply explore complex relationships. The DDA-NG framework fuses features from non-negative matrix factorization (NMF) and graph transformers to generate richer feature representations and enhance the diversity of feature expression. Specifically:

[0076] A drug-disease association prediction method based on NMF and GT, comprising the following steps:

[0077] S10: Based on the drug fingerprint similarity, disease phenotype similarity, and Gaussian kernel similarity between drugs and diseases, similarity information is fused to obtain a comprehensive similarity matrix of drugs and a comprehensive similarity matrix of diseases;

[0078] S20: extracting potential features of the drugs and diseases from the comprehensive similarity matrix of the drugs and the comprehensive similarity matrix of the diseases by non-negative matrix decomposition to obtain drug features and disease features;

[0079] And through the graph transformer, we capture its high-order structural relationship to obtain drug similarity features and disease similarity features;

[0080] S30: After interactively fusing the drug features and drug similarity features, the disease features and disease similarity features, respectively, a final drug feature representation and a final disease feature representation are obtained;

[0081] S40: performing association prediction based on the final characteristic representation of the drug and the final characteristic representation of the disease.

[0082] The key points of the drug-disease association prediction method and model in this embodiment are as follows:

[0083] 1. Similarity Information

[0084] (1) Drug fingerprint similarity

[0085] In order to more comprehensively express the similarity relationship between drugs, this embodiment uses the molecular chemical structure of the drug and obtains fingerprint similarity based on RDKit and Tanimoto. This embodiment represents the drug fingerprint similarity as RFSM∈R m×m , where m represents the number of drugs. The more consistent the chemical structures of two drugs are, the higher their similarity score is.

[0086] (2) Similarity of disease phenotype

[0087] Disease phenotypic similarity is calculated based on the analysis and mining of disease-related medical texts in the OMIM database, aiming to measure the similarity of different diseases in phenotypic characteristics. Tools such as MimMiner are usually used to extract and quantify the similarity of different diseases in phenotypic characteristics. In this example, disease phenotypic similarity is represented as DPSM∈R n×n , where n represents the number of diseases, and this information is used to reveal the similarities in phenotypic characteristics of different diseases.

[0088] (3) Gaussian kernel similarity between drugs and diseases

[0089] In the known drug-disease network, similar diseases may have similar association patterns, and diseases with similar properties are often closely related to each other. Therefore, using this feature, the Gaussian function is applied to the topological association network between biological information nodes, and the kernel method is used to establish the kernel function from the feature vector to achieve the purpose of extracting association features. Therefore, the present invention introduces the GIP kernel similarity to solve its inherent sparsity and potential lack of information. The GIP kernel similarity between drugs is calculated by i and r j Expressed as:

[0090] RGSM(r i ,r j )=exp(-γ r ||IP(r i )-IP(r j )||2 ) (1)

[0091]

[0092] Among them, RGSM∈R m×m , IP(r i ) indicates drug r i The corresponding column in the incidence matrix, γ r ' is the normalized kernel bandwidth, which determines the density of samples in the feature space and controls the width of the similarity function from the similarity point of view, and its value is usually set to 1. k (1≤k≤m) is the characteristic representation of the drug, r i ,r j are the feature vectors of two different drugs. r is the Gaussian kernel parameter, which is used to control the decay rate of drug similarity and determines the shape and width of the similarity function.

[0093] Similarly, the GIP nuclear similarity between diseases d i and d j Expressed as:

[0094] DGSM(d i ,d j )=exp(-γ d ||IP(d i )-IP(d j )|| 2 ) (3)

[0095]

[0096] Where DGSM∈R n×n , IP(d i ) indicates disease i The corresponding column in the incidence matrix, γ r ' is the normalized kernel bandwidth, γ d ' = 1. d k (1≤k≤n) is the characteristic representation of the disease, d i ,d j are the feature vectors of two different diseases. d is the Gaussian kernel parameter, which is used to control the decay speed of disease similarity and determines the shape and width of the similarity function.

[0097] (4) Similarity Information Fusion

[0098] In this embodiment, the various similarity information obtained above is integrated through the following formula to obtain the comprehensive similarity matrix IRSM (r i ,r j )∈Rm×m and the comprehensive similarity matrix IDSM of diseases (d i ,d j )∈R n×n .

[0099]

[0100]

[0101] Among them, RFSM(r i ,r j ) is the drug fingerprint similarity, DPSM(d i ,d j ) is the disease phenotype similarity, RGSM (r i ,r j ) is the drug Gaussian kernel similarity, DGSM(d i ,d j ) is the disease Gaussian kernel similarity.

[0102] 2. NMF feature extraction

[0103] As a powerful tool for extracting information, non-negative matrix factorization (NMF) provides a very convenient way to decompose matrices. By decomposing matrices into non-negative factor matrices, potential feature representations can be mined. Especially in association prediction, it can effectively capture the structure and pattern of data. Therefore, the present invention introduces NMF to extract potential features of drugs and diseases. The core idea is to convert a non-negative matrix A into m×n Decomposed into two non-negative matrices U m×k and V k×n , where A m×n ≈U m×k V k×n , so that U m×k and V k×n The product of can approximately restore the original matrix A m×n , where U m×k is the basis matrix, V k×n is the coefficient matrix.

[0104] In the matrix decomposition process, the matrix A m×n The column vector of can be expressed as the matrix U m×k The column vectors of are weighted by the linear combination of weight coefficients, and these weights are represented by the matrix V k×n The elements of the corresponding column vector in are determined. Again, m here represents the number of drugs, n represents the number of diseases, k is the dimension of the low-dimensional space, and k satisfies k<m and (m+n)*k<mn. In order to ensure the smoothness of U and V, Tikhonov regularization L2 is used.

[0105]

[0106] Among them, A represents the association matrix between drugs and diseases, W is the same size as A, and × represents the product of the corresponding elements of the two matrices. Expressed as the matrix Frobenius norm. Regularization coefficient λ 1 and λ 2 Both are set to 0.01, and the dimension k of the decomposed latent space is 200.

[0107] NMF is usually solved by alternating optimization through an iterative method, gradually updating U and V, and minimizing the difference between the product of the original matrix and the decomposed matrix through an iterative optimization process. and is the Lagrange multiplier, then the Lagrangian function Y of the optimization problem can be constructed as:

[0108]

[0109] Among them, T represents the transpose operation, and Tr represents the trace operation, that is, calculating the sum of the elements on the diagonal of a matrix.

[0110] According to the Karush-Kuhn-Tucker condition φ ik u ik = 0 and φ kj v kj = 0, and the update rules for U and V are as follows:

[0111]

[0112]

[0113] U and V are iteratively updated 1000 times, and U' and V' are obtained as the characteristic matrices of the drug, V' T is the feature matrix of the disease. In this embodiment, the drug features and disease features obtained by this module are respectively expressed as and

[0114] 3. Graph transformer extracts high-level features

[0115] The graph Transformer network mainly learns the representation of nodes through a multi-head attention mechanism. This paper adopts an improved graph Transformer network to transform the IRSM (r i ,r j )∈R m×m and IDSM(d i ,d j )∈R n×n Construct graph networks separately to obtain similarity networks SNr and SN d , and extract the embedding representation of the node from it.

[0116] (1) Extracting high-level features of drugs

[0117] To define the update process of node features of the p-th GT layer, a similar network SN is used r , this process is expressed by formulas (11), (13)-(15):

[0118] f 0 =W 0 SN r +b 0 (11)

[0119] f 0 =W 0 SN d +b 0 (12)

[0120]

[0121]

[0122] Among them, d' gt represents the dimension of each attention head; j∈N i Represents the adjacent nodes of node i; And they represent f 0 The i-th and j-th columns of ; is the multi-head attention score, which captures both the similarity and correlation between Q and K; W 0 is the weight matrix, b 0 is the bias vector.

[0123]

[0124] in, represents a learnable matrix, passing the output of the attention mechanism to the feedforward network to generate the final node representation of the (p+1)th layer. gt After the layer is updated, the drug similarity feature representation is obtained

[0125] (2) Extracting high-level disease features

[0126] Similarly, the same is true for extracting high-order features of diseases. To define the update process of node features of the p-th GT layer, a similar network SN is used. d This process is expressed by formula (12)-formula (15), and finally the disease similarity feature representation is obtained

[0127] 4. Interactive Integration

[0128] Most previous studies used direct concatenation. In this paper, an interactive fusion module based on the attention mechanism is designed to combine the drug and disease information obtained from different modules. This module uses the Transformer encoder as the core unit of fusion and interaction to capture effective associations in two different feature representations.

[0129] First, the two drug characteristics obtained above are expressed as and Perform superposition to generate initial drug embedding Similarly, the two disease characteristics obtained above are expressed as and Overlay to generate initial disease embedding The initial drug and disease embeddings are then passed through the Transformer encoder to obtain the final feature representation of the drug. and disease final characterization The multi-head attention mechanism is the key module of the Transformer encoder, and its function can be expressed as the following formula:

[0130]

[0131]

[0132] Where 1≤k≤f t , d' t =d t / f t .f t The dimension of the feature vector representing the drug or disease. d' t is the dimension of the attention head.

[0133] 5. Association Prediction Module

[0134] The multi-source feature representation is obtained through the above method, and finally a three-layer MLP with a dropout layer and a ReLU activation function is used to output the predicted association score. The cross entropy loss function is used for optimization, which is expressed as follows:

[0135]

[0136] where Y + and Y - Represent positive training samples and negative training samples respectively. (i, j) represents a given pair of drug i and disease j. ij and y' ij are the true labels and predicted scores, respectively.

[0137] Experiment B

[0138] 1. Experimental setup

[0139] (1) Baseline

[0140] This example compares the DDA-NG model with the DDAGDL, DRWBNCF, DRHGCN, deepDR, HNet-DNN, and HINGRL models to demonstrate the competitiveness and robustness of DDA-NG.

[0141] 1) DDAGDL: Predicting drug associations by using geometric deep learning on heterogeneous information networks.

[0142] 2) DRWBNCF: A novel neighborhood and neighborhood interaction-based neural collaborative filtering approach to infer new potential drugs for treating diseases.

[0143] 3) DRHGCN: A deep learning framework that designs inter-domain and intra-domain feature extraction modules by performing graph convolution operations on the network to learn embeddings of drugs and diseases to discover potential drugs for a certain disease.

[0144] 4) deepDR: A network-based deep learning approach for in silico drug repurposing by integrating 10 networks.

[0145] 5) HNet-DNN: A method that uses deep neural networks to predict new drug-disease associations based on features extracted from drug-disease heterogeneous networks.

[0146] 6) HINGRL: A novel model based on heterogeneous information networks to accurately identify new indications for drugs based on graph representation learning technology.

[0147] (2) Evaluation indicators

[0148] This example uses AUC, AUPR, Accuracy, Precision, Recall, F1-score, MCC7 standards and five-fold cross validation to evaluate the model in order to more comprehensively understand the effectiveness and reliability of the DDA-NG model in drug-disease association prediction. The evaluation index formula is as follows:

[0149]

[0150]

[0151]

[0152]

[0153]

[0154]

[0155]

[0156] C Experimental Results and Analysis

[0157] 1. Main Experiment Results

[0158] The results in Table 2 show that the DDA-NG model shows significant advantages in the three data sets, especially in key indicators such as AUC, AUPR, accuracy, F1-score and MCC, which are better than other comparison models. In particular, the recall is significantly higher than deepDR and DRWBNCF, which shows that these two models tend to identify known DDA as negative. DDA-NG shows excellent classification ability and can effectively balance precision and recall. This enables DDA-NG to accurately identify the association between drugs and diseases while reducing the occurrence of false positives and false negatives, thereby improving the reliability of predictions. The present invention attributes the excellent performance of DDA-NG to the implementation of low-rank decomposition, multi-source information network and graph attention fusion architecture.

[0159] Table 2 Association prediction results

[0160]

[0161] 2. Ablation experiment

[0162] In order to verify the effectiveness of each module of DDA-NG, this example conducts ablation experiments on each module. Only_NMF is the feature representation extracted by NMF only, and only_GT is the feature representation extracted by graph transformer only. Each module is used alone and in combination for comparison. Figure 2 The results show that the only_NMF module has some gaps in AUC and AUPR indicators, which shows that using NMF alone has limitations in complex relationship modeling. only_GT can better handle complex graph structure information, showing its advantages in capturing complex relationships, but the performance of the model is further improved when the two are combined. The DDA-NG model performs best in multiple indicators, especially on C-dataset and F-dataset, with AUC reaching 0.9648 and 0.9541 respectively, which is significantly better than the effect of using a single module. This verifies the synergy of NMF and graph transformer modules, giving full play to their respective advantages and improving the overall performance of drug-disease association prediction.

[0163] Combined with the examples, it can be seen that the DDA-NG model proposed in the present invention effectively integrates non-negative matrix factorization (NMF) and graph transformer (GT) technology, and demonstrates its excellent performance in the task of drug-disease association prediction. Through the precise extraction and fusion of drug and disease features, DDA-NG not only enhances the richness of feature representation, but also significantly improves the model's ability to capture complex relationships. The experimental results verify the superior performance of the model on multiple data sets, prove its effectiveness in accurately predicting potential drugs, and show that this method has good potential. This method provides new ideas for future drug discovery and disease treatment.

[0164] The above is only a preferred embodiment of the embodiment of the present invention, and does not impose any form of limitation on the embodiment of the present invention. Any simple modification, equivalent changes and modifications made to the above embodiment based on the technical essence of the embodiment of the present invention are still within the scope of the technical solution of the embodiment of the present invention.

Claims

1. A drug-disease association prediction method based on NMF and GT, characterized in that: The following steps are involved: (1) Based on the drug fingerprint similarity, disease phenotype similarity, and Gaussian kernel similarity between drugs and diseases, similarity information is fused to obtain a comprehensive similarity matrix of drugs and a comprehensive similarity matrix of diseases; (2) The comprehensive similarity matrix of the drugs and the comprehensive similarity matrix of the diseases are used to extract potential features of the drugs and diseases through non-negative matrix decomposition to obtain drug features and disease features; And through the graph transformer, we capture its high-order structural relationship to obtain drug similarity features and disease similarity features; (3) After interactively fusing the drug features and drug similarity features, the disease features and disease similarity features, respectively, a final drug feature representation and a final disease feature representation are obtained; (4) performing association prediction based on the final characteristic representation of the drug and the final characteristic representation of the disease.

2. The drug-disease association prediction method according to claim 1, characterized in that: In the step (1), the Gaussian kernel similarity r between the drugs i and r j It is expressed as follows: RGSM(r i ,r j )=exp(-γ r ||IP(r i )-IP(r j )|| 2 ); Among them, RGSM∈R m×m , IP(r i ) indicates drug r i The corresponding column in the incidence matrix, γ r ' is the normalized kernel bandwidth; The Gaussian kernel similarity d between diseases i and d j It is expressed as follows: DGSM(d i ,d j )=exp(-γ d ||IP(d i )-IP(d j )|| 2 ); Where DGSM∈R n×n , IP(d i ) indicates disease i The corresponding column in the incidence matrix, γ d '=1.

3. The drug-disease association prediction method according to claim 2, characterized in that: In the step (1), the comprehensive similarity matrix IRSM (r i ,r j )∈R m×m and the comprehensive similarity matrix IDSM of diseases (d i ,d j )∈R n×n It is expressed as follows: Among them, RFSM(r i ,r j ) is the drug fingerprint similarity, DPSM(d i ,d j ) is the disease phenotype similarity, RGSM (r i ,r j ) is the drug Gaussian kernel similarity, DGSM(d i ,d j ) is the disease Gaussian kernel similarity.

4. The drug-disease association prediction method according to claim 1, characterized in that: In the step (2), the process of non-negative matrix decomposition is: after decomposing a non-negative matrix into two non-negative matrices U and V, an iterative method is used to perform alternating optimization to solve, U and V are gradually updated, and the difference between the product of the original matrix and the decomposed matrix is ​​minimized through an iterative optimization process to obtain drug characteristics and disease characteristics.

5. The drug-disease association prediction method according to claim 1, characterized in that: In step (2), the graph Transformer network is derived from the similarity network SN r and SN d Extract the embedding representation of the node in The process of obtaining drug similarity features is as follows: To define the update process of node features of the pth GT layer, the similarity network SN is used r , this process is expressed by the following formula: f 0 =W 0 SN r +b 0 ; Where, d' gt represents the dimension of each attention head, j∈N i represents the adjacent nodes of node i; And they represent f 0 The i-th and j-th columns of ; is the multi-head attention score, which captures both the similarity and correlation between Q and K; In the formula, represents a learnable matrix, passes the output of the attention mechanism to the feedforward network, and generates the final node representation of the (p+1)th layer; after the pth gt After the layer is updated, the drug similarity feature representation is obtained 6. The drug-disease association prediction method according to claim 1, characterized in that: In the step (3), the interactive fusion process is: superimposing the drug features and drug similarity features, the disease features and disease similarity features respectively, to obtain the initial drug embedding and the initial disease embedding, and then passing through the Transformer encoder using a multi-head attention mechanism to obtain the final drug feature representation and the final disease feature representation.

7. The drug-disease association prediction method according to claim 1, characterized in that: In the step (4), a three-layer MLP with a dropout layer and a ReLU activation function is used to output the predicted association score.

8. A drug-disease association prediction model based on NMF and GT, characterized in that: include: Similarity information module: obtain drug fingerprint similarity, disease phenotype similarity, and Gaussian kernel similarity between drugs and diseases, perform similarity information fusion, and obtain the comprehensive similarity matrix of drugs and the comprehensive similarity matrix of diseases; Non-negative matrix decomposition feature extraction module: extracts drug and disease features through non-negative matrix decomposition to obtain drug features and disease features; Graph Transformer feature extraction module: captures high-order structural relationships through graph transformers to obtain drug similarity features and disease similarity features; Interactive fusion module: Based on the attention mechanism, the drug and disease information obtained by the NMF feature extraction module and the graph Transformer feature extraction module are combined to obtain the final feature representation of the drug and the final feature representation of the disease; Association prediction module: performs association prediction based on the final feature representation of the drug and the final feature representation of the disease.

9. A computer device, characterized in that: The method comprises a processor and a memory, wherein the memory stores computer executable instructions that can be executed by the processor, and the processor executes the computer executable instructions to implement the drug-disease association prediction method according to any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores computer-executable instructions. When the computer-executable instructions are called and executed by the processor, the computer-executable instructions prompt the processor to implement the drug-disease association prediction method according to any one of claims 1 to 7.

Images