A drug repositioning method based on meta-pathway integration of multi-source biological data

Abstract

The present invention relates to the field of pharmaceutical technology, and in particular to a drug repositioning method for integrating multi-source biological data based on meta-paths, comprising obtaining an original data set; constructing a fused heterogeneous network, fusing the association matrices of drug, disease, and protein pairs and their respective similarity matrices to obtain a heterogeneous graph; designing a multi-level association path based on negative sampling of meta-paths and similarities, fusing similarity information of networks at the same level in the heterogeneous network based on the similarity principle, and selecting negative link samples; extracting drug features; constructing an SMS-MPF module, walking the meta-path through the heterogeneous network, extracting and fusing features between different meta-paths, and obtaining association features between drugs and diseases; and inputting the disease phenotype similarity matrix, drug features, and association features between drugs and diseases into an association prediction module to obtain a final node representation. The present invention addresses the significant defects in the selection of positive and negative samples in existing methods.

Description

Technical Field

[0001] The present invention relates to the field of pharmaceutical technology, and in particular to a drug repositioning method based on meta-path integration of multi-source biological data. Background Art

[0002] Drug repositioning, as an emerging research direction, is gaining increasing attention. Existing drug-disease association prediction tasks mainly include three modules: heterogeneous network construction, representation learning module, and prediction module. When constructing heterogeneous networks, the vision is not expanded to consider the intricacies of biological information. How to integrate more biological information data and filter out effective information to provide more assistance for model prediction is a challenge. In the representation learning module, due to the sparsity of similarity networks, improper selection of representation learning models may lead to the extraction of impurity features, which interferes with the prediction module to obtain effective results. In summary, existing methods mainly regard similarity information as the only feature of nodes in the association network, ignoring the importance of the essential characteristics of drugs and diseases, which may provide incorrect guidance for association prediction.

[0003] Recently, negative sampling techniques have gained attention in the field of association prediction, such as Wangy et al.'s "Predicting Potential Drug-Disease Associations Based on Hypergraph Learning with Subgraph Matching." This method suffers from flaws in the selection of positive and negative samples; links in the association network are considered positive samples for the model, while negative samples are randomly selected from unknown associations in equal quantities. This random selection method not only fails to provide effective feature condensation for model training, but also causes the model to overly focus on learning positive sample features. Given the generally sparse nature of association networks, this approach can lead to models with better predictions than reality, resulting in significant deviations from reality. Therefore, improving negative sample selection strategies is particularly important to ensure the comprehensiveness, accuracy, and practical applicability of model training. Summary of the Invention

[0004] To address the shortcomings of existing methods, this paper proposes a drug repositioning method based on meta-path integration of multi-source biological data. Its core lies in the deep integration and analysis of bioinformatics data. MPMB-DR improves the generalization of the model by introducing diverse bioinformatics data sources and appropriate feature extraction modules, thereby reducing the impact of noise from a single data source. A negative sampling strategy is designed that integrates meta-path and similarity principles, fully utilizing the rich meta-path information in a three-layer heterogeneous network to provide reliable negative samples for association prediction.

[0005] The technical solution adopted by the present invention is: a drug repositioning method based on meta-path integration of multi-source biological data includes the following steps:

[0006] Step 1: Obtain the original data set including drug, disease, and protein similarity matrices; drug-protein, drug-disease, and disease-protein adjacency matrices; disease comprehensive similarity matrix and drug comprehensive similarity matrix;

[0007] Step 2: Using the fusion heterogeneous network, the association matrices of the drug, disease, and protein pairs and their respective similarity matrices are fused to obtain a drug-protein-disease heterogeneous graph;

[0008] As a preferred embodiment of the present invention, the fused heterogeneous network is a three-layer heterogeneous network.

[0009] Step 3: Design multi-level association paths based on negative sampling of meta-paths and similarities, fuse similarity information of the same-level networks in heterogeneous networks based on the similarity principle, and select negative link samples;

[0010] As a preferred embodiment of the present invention, step three specifically includes:

[0011] According to the type and length of different meta-paths, the drug-disease adjacency matrix A dr-d Assign different weights to obtain the weighted drug-disease adjacency matrix A' dr-d ;

[0012] For drug m and disease n, the distance T(m,n) between them is measured based on the association matrix and their respective similarity matrices. The formula is:

[0013] T(m,n)=e -(TDD+TDC)

[0014] in, D m represents the set of diseases associated with drug m, C n represents the set of drugs associated with disease n.

[0015] Step 4: Construct a drug feature extraction module based on SMILE sequences, apply GCN to the molecular structure graph formed by the drug SMILE sequence, and extract drug features;

[0016] As a preferred embodiment of the present invention, the formula of drug characteristics is:

[0017] H DR =GCN(G ξ ,GCN(G ξ ,h ξ ))

[0018] Among them, Gξ is a self-circulating undirected graph, h ξ For drug characteristics.

[0019] Step 5: Construct the SMS-MPF module to walk the meta-paths through the heterogeneous network, extract and fuse the features between different meta-paths, and obtain the association features between drugs and diseases;

[0020] As a preferred embodiment of the present invention, the formula for the correlation characteristics between drugs and diseases is:

[0021]

[0022] Among them, MPF is the meta-path fusion operation, h v is the feature of node v, P φ(v) is a set of various meta-paths related to node v, represents a shared maximal subgraph based on the meta-path p.

[0023] Step 6: Input the disease phenotype similarity matrix, drug characteristics, and drug-disease association characteristics into the association prediction module to obtain the final node representation;

[0024] As a preferred embodiment of the present invention, the final node representation formula is:

[0025]

[0026] Among them, H v1 、H v2 are the association features of drugs and diseases, H DR is the final drug characteristic, S d is the disease phenotype similarity matrix.

[0027] As a preferred embodiment of the present invention, the association prediction module adopts a loss function of the sigmoid function, and the formula is:

[0028]

[0029] Where W is the learnable diagonal matrix parameter, σ(·) is the sigmoid function, Ω represents the positive link set, i.e., the observation graph; p n is the negative link of the sample (u,v - ) obeys the distribution.

[0030] As a preferred embodiment of the present invention, a drug repositioning system based on meta-path integration of multi-source biological data includes: a memory for storing instructions executable by a processor; and a processor for executing the instructions to implement a drug repositioning method based on meta-path integration of multi-source biological data.

[0031] As a preferred embodiment of the present invention, a computer readable medium stores computer program code, and when the computer program code is executed by a processor, the computer program code implements a drug repositioning method based on meta-pathway integration of multi-source biological data.

[0032] Beneficial effects of the present invention:

[0033] 1. Diversified integration of bioinformatics data: MPMB-DR incorporates more bioinformatics data to provide a more complex and realistic simulation environment. Because a single data source is one-sided for association prediction, this invention integrates more original drug SMILE sequences, protein similarity, and disease characteristics based on basic drug-disease similarity data to comprehensively analyze the complex associations between drugs and diseases.

[0034] 2. Design of a negative sampling module based on metapath and similarity principles: Design a new negative sampling module based on metapath and similarity principles. This module considers the multi-level associations between drugs, proteins, and diseases based on the structure of heterogeneous networks and the similarity between homogeneous networks, thereby improving the comprehensiveness and accuracy of association prediction.

[0035] 3. MPMB-DR fully utilizes the node characteristics and topological structure of the three-layer heterogeneous network to conduct in-depth analysis of the complex relationships between biological information;

[0036] 4. Experimental results show that MPMB-DR can effectively mine deep-level correlation features and its accuracy is better than most methods. BRIEF DESCRIPTION OF THE DRAWINGS

[0037] Figure 1 is a flow chart of the MPMB-DR prediction of potential drug-disease associations of the present invention;

[0038] Figure 2 Schematic diagram of the negative sampling module based on meta-path and similarity principle of the present invention;

[0039] Figure 3 10-CV ROC curves and PR curves of different models under the present invention (data set B);

[0040] Figure 4 10-CV ROC curves and PR curves of different models under the present invention (C data set);

[0041] Figure 5 10-CV ROC curves and PR curves of different models (F data set) of the present invention;

[0042] Figure 6 It is a comparison of AUC and AUPR of MPN and RN of the present invention when the positive and negative sample ratios are increased. DETAILED DESCRIPTION

[0043] The present invention will be further described below in conjunction with the accompanying drawings and embodiments. This figure is a simplified schematic diagram, which only illustrates the basic structure of the present invention in a schematic manner, and therefore only shows the components related to the present invention.

[0044] like Figure 1 As shown, a drug repositioning method based on meta-pathway integration of multi-source biological data includes the following steps:

[0045] Step 1: Obtain the original data set including drug, disease, and protein similarity matrices; drug-protein, drug-disease, and disease-protein adjacency matrices; disease comprehensive similarity matrix and drug comprehensive similarity matrix;

[0046] Get the drug-disease adjacency matrix A dr-d , protein and disease adjacency matrix A p-d , drug and protein adjacency matrix A dr-p , protein similarity matrix S p , drug fingerprint similarity matrix S dr and the disease phenotype similarity matrix S d ; Using the drug-disease adjacency matrix A dr-d Get the GIP similarity matrix GIP of drugs and diseases respectively dr 、GIP d ; Using GIP dr 、S dr Calculate the drug comprehensive similarity matrix SM dr ; Using GIP d 、S d Calculate the disease comprehensive similarity matrix SM d ;

[0047] That is, obtain A through the benchmark database dr-d 、S dr 、A dr-p 、S p 、A dr-d 、S d .

[0048] In the present invention, datasets B, C, and F will be used as benchmark datasets; each dataset contains three types of entities, namely drugs, diseases, and proteins, as well as the association information between them; among them, the drug-disease association network is obtained from the CTD database, the drug-protein association network is from the DrugBank database, and the diseases are collected from the Online Mendelian Inheritance in Man (OMIM) database; protein-related information is introduced from the DisGeNET database to expand these databases; the detailed information of the datasets is shown in Table 1; the similarity of drug molecular fingerprints is used to measure drug similarity.

[0049] Table 1. Details of the benchmark database

[0050]

[0051] Disease phenotype refers to the characteristics and behaviors exhibited by a disease in an individual; phenotypic similarity is a disease similarity metric calculated by performing text mining analysis on medical information derived from the OMIM database; among them, MimMiner is a commonly used method for calculating disease phenotypic similarity.

[0052] Obtain drug similarity matrix S from the dataset dr , disease similarity matrix S d and the protein similarity matrix S p ; Gaussian interaction spectrum kernel similarity (GIP) is introduced to enrich the similarity data of drugs and diseases; the GIP similarity matrices of drugs and diseases are represented as GIP dr and GIP d ; Drug dr in the drug comprehensive similarity matrix i and drug dr j The similarity is recorded as SM dr (dr i ,dr j ), the calculation formula is as follows:

[0053]

[0054] Similarly, the disease comprehensive similarity matrix SM d It is also calculated in a similar way.

[0055] Step 2: Construct a fusion heterogeneous network by fusing the association matrices of the drug, disease, and protein pairs in the original data and their respective similarity matrices to construct a drug-protein-disease heterogeneous graph G;

[0056] The fusion heterogeneous network is a three-layer heterogeneous network. The heterogeneous graph contains the associations between biological information of the same level and different types of biological information. The heterogeneous graph is defined as G(V,E); where V and E are the sets of nodes and edges in the graph respectively; the association matrices between drugs and proteins, proteins and diseases, and drugs and diseases are represented as A respectively. dr-p 、A p-d 、A dr-d , then the three-layer heterogeneous network of drug-protein-disease can be expressed as:

[0057]

[0058] Step 3: Design multi-level association paths based on negative sampling of meta-paths and similarities. Based on the similarity principle, similarity information of the same-level networks in heterogeneous networks is integrated to select balanced and high-quality negative link samples.

[0059] Existing methods select positive samples based on the principle of similarity and then filter out negative samples. This method, to a certain extent, filters out the second most likely positive samples to fill the sparse association network. However, this type of method does not fundamentally solve the problem: the model still learns features related to the positive samples from this type of sample data. The present invention is based on another idea of ​​similarity, that is, diseases that are dissimilar to each known / predicted target of a given drug are unlikely to become the target of the drug, and vice versa.

[0060] It is assumed that a disease that is dissimilar to every known / predicted disease targeted by a drug is unlikely to be the drug's target disease; on the other hand, a drug that is dissimilar to any known / predicted disease targeted by a drug is also unlikely to be the drug's target disease;

[0061] The negative sampling method based on the meta-path and similarity principle of the present invention is as follows: Figure 2 As shown;

[0062] According to the type and length of different meta-paths, the drug-disease adjacency matrix A dr-d Assign different weights, where the known drug-disease association pairs have the highest weight, thus obtaining the enriched association matrix A' dr-d ; Assume drug m and disease n, and measure the distance T(m,n) between them based on the association matrix and their respective similarity matrices.

[0063] For any disease l The set of diseases related to drug m is denoted as n l ∈D m , consider the drugs associated with drug m to measure the distance between m and n: Similarly, for any drug m l The set of drugs related to disease n is denoted as m l ∈C n , considering disease n and each known / predicted drug m l The similarity between them is calculated, and the weighted score is obtained.

[0064] So, for drug m and disease n, the distance between m and n is defined as follows:

[0065] T(m,n)=e -(TDD+TDC) (3)

[0066] Among them, T(m,n) represents the possibility that drug m is not associated with disease n. The larger T(m,n) is, the less likely it is that the two are associated. The scores are sorted from large to small, and the drug-disease pairs with the highest scores constitute the negative sample candidate set.

[0067] In order to improve the accuracy of screening drugs targeting specific diseases, it is expected that the characteristics of the targeted diseases are as different as possible, so that candidates similar to known targets can be more effectively excluded; by evaluating the characteristic differences of candidate negative samples, variance analysis is used to confirm whether the characteristic differences of these samples are significantly higher than the overall difference, thereby enhancing the credibility of the screening process.

[0068] Step 4: Construct a drug feature extraction module based on SMILE sequences, apply GCN to the molecular structure graph formed by the drug SMILE sequence, and extract drug features;

[0069] Drug characteristics are obtained through the heterogeneous network in step three.

[0070] In order to mine bioinformatics features to explore candidate diseases for drugs, the SMILES sequence of drugs is introduced; SMILES (Simplified Molecular Input Line Entry System) is a linear representation for describing chemical molecules, which uses ASCII characters to represent the structure of molecules.

[0071] In order to obtain drug features, the SMILES string of a drug ξ is first converted into a two-dimensional molecular graph G ξ =(V ξ ,E ξ ); where V ξ and E ξ Represent the vertex set (i.e., atoms) and edge set (i.e., chemical bonds) respectively; in order to characterize G ξ For each node in the dataset, the DGL-LifeSci software package is first used to initialize each node according to its chemical properties. To uniformly represent drug molecules, each atomic node is represented as a 74-dimensional vector containing integer values. Since different drugs contain different numbers of atoms, a maximum number of atoms, N, is set to convert them into tensors of the same dimension. When the number of atoms of a drug is less than N, the missing nodes are filled with zero values, thereby unifying the data structure of all drugs into a tensor of size N×74. This processing method ensures that all drugs have the same dimension, which is convenient for batch input into the model for calculation and training.

[0072] Therefore, each drug can be expressed as By simple linear transformation It can be further used as the input of GCN, where N ξ It is the potential spatial dimension of the drug.

[0073] As a neural network method, GCN propagates from the lth layer to the (l+1th layer) in the following way:

[0074]

[0075] in, and represents the characteristics of the drug ξ in the lth layer and the (l+1)th layer, Represents an undirected graph G with self-loops ξ The corresponding adjacency matrix, A and I represent G ξ The adjacency matrix and the corresponding identity matrix, W l represents the learnable parameter matrix of layer l, and σ(·) is the nonlinear activation function.

[0076] The two-layer GCN is used to learn the graph representation features of drugs. The final drug features can be expressed as:

[0077] H DR =GCN(G ξ ,GCN(G ξ ,h ξ )) (5)

[0078] Step 5: Construct a meta-path fusion module based on shared maximum subgraph (SMS-MPF). This module walks multiple meta-paths through a three-layer heterogeneous network, extracts and fuses features between different meta-paths, and obtains the association features between drugs and diseases.

[0079] Avoid methods such as HAN, GTN and MAGCN from repeatedly fusing meta-paths in heterogeneous networks, which will cause data redundancy and hinder the effective extraction of meta-path features.

[0080] The present invention is based on sharing the largest subgraph and aggregating meta-paths to form node features;

[0081] SMS-MPF includes: feature preprocessing, construction of shared maximum subgraph on meta-path, feature extraction of specific meta-path and feature fusion of multiple meta-paths.

[0082] In the constructed three-layer heterogeneous network, the feature dimensions of different types of nodes are inconsistent. In order to facilitate the message passing and aggregation operations in the graph, linear transformation is used to project the features of different types of nodes into potential vectors with the same dimension d. The projection representation of node v It can be obtained in the following ways:

[0083]

[0084] in represents the original features of node v, and are the learnable weights and biases respectively.

[0085] With node v as the center, the meta-path p∈P φ(v) , the subgraph formed by all meta-paths of type p is It is important to note that a predetermined meta-path length K is used to automatically select the meta-path p; any meta-path of length less than K must be a subsequence of a meta-path of length K; the context of the shorter meta-path will be completely contained in the subgraph of the context of the longer meta-path; this avoids redundant computation of long and short meta-paths.

[0086] In constructing the meta-path shared maximal subgraph Afterwards, average graph pooling is used to learn the embedding In the process of layer-by-layer propagation, not only local information is extracted, but also neighborhood features under different relationships are aggregated in each layer of pooling, which helps to build a richer node representation; therefore, the features of a specific meta-path p in layer l It can be expressed as:

[0087]

[0088] in, is the feature of node v in the l-1th layer, Represents a shared maximum subgraph The number of nodes, u is Nodes in .

[0089] Subsequently, multiple meta-path encodings of node v are fused to form a unified node representation. Different meta-paths have different importance to node v due to their different types. Here, a one-dimensional convolution kernel is used to obtain the representation of node v:

[0090]

[0091] in, It is a learnable parameter.

[0092] Then, further projection is applied to obtain the final node vector of the desired output dimension, as follows:

[0093]

[0094] in, and are the weight matrix and bias term respectively, and σ(·) is the RELU activation function.

[0095] The above meta-path fusion operation is uniformly referred to as MPF. A two-layer meta-path feature extraction is used for this. The final meta-path fusion feature of node v is Among them, h v is the feature of node v, P φ(v) It is a set of various meta-paths related to node v; node v includes drug node v1 and disease node v2.

[0096] Step 6: Disease phenotype similarity matrix S d , drug characteristics H DR and the association characteristics between drugs and diseases H v Input the association prediction module to obtain the final representation of node v;

[0097] The formula for representing node v is:

[0098]

[0099] The loss function based on the sigmoid function is expressed as:

[0100]

[0101] Where W is the learnable diagonal matrix parameter, σ(·) is the sigmoid function, Ω represents the positive link set, i.e., the observation graph; p n is the negative link of the sample (u,v - ) obeys the distribution.

[0102] Evaluation indicators:

[0103] To quantify the performance of the model, a receiver operating characteristic (ROC) curve was plotted to visually demonstrate the effectiveness of the model, with the false positive rate (FPR) as the X-axis, the true positive rate (TPR) as the Y-axis, and the area under the curve (AUC) as the evaluation indicator; the higher the AUC, the better its predictive ability; when the model exhibits perfect predictive ability, the AUC value is 1; if the model only has random predictive ability, the AUC value is 0.5.

[0104] Another widely used evaluation metric is AUPR (Area Under the Precision-Recall curve), which is based on the area under the Precision-Recall curve. The horizontal axis represents the recall rate, and the vertical axis represents the accuracy rate. The AUPR value range is 0 to 1, and the higher the value, the better the model performance. AUPR can help understand the accuracy performance of the model at different recall rates.

[0105] In order to more comprehensively demonstrate the performance of the model, a variety of indicators are used to evaluate the model, including accuracy, precision, recall, F1-score and Matthews correlation coefficient (MCC).

[0106] Comparison with other methods:

[0107] Ten-fold cross-validation (10-CV) was used to verify the accuracy of MPMB-DR; in the ten-fold cross-validation, all established associations were randomly divided into ten equivalent subsets; in each iteration, one subset was designated as the test sample, and the remaining subsets together constituted the training sample; in this process, all unidentified associations were considered candidate samples.

[0108] At the same time, MPMB-DR was compared and analyzed with the current most advanced drug-disease association prediction models: HNet-DNN, DRHGCN, HINGRL, DRWBNCF, DDAGDL, and AMDGT; the comparison results are shown in Tables 2, 3, and 4, and these tables also give the performance based on basic biological information (only containing drug similarity, disease similarity, and association information).

[0109] Table 210-Comparison of performance of different methods under CV dataset B

[0110]

[0111] Table 310-Comparison of performance of different methods under C dataset under CV

[0112]

[0113]

[0114] Table 410-Comparison of performance between different methods under F dataset under CV

[0115]

[0116] Figure 3 、 Figure 4 and Figure 5 The ROC curves and PR curves of MPMB-DR and the above methods under ten-fold cross validation on the B dataset, C dataset, and F dataset were plotted respectively; it can be seen from the figure that the method of the present invention is superior to most drug-disease association prediction methods.

[0117] Ablation experiment:

[0118] Different degrees of positive and negative samples are introduced to simulate the imbalance in the real environment; when constructing negative samples, meta-paths are combined to capture the multiple semantic relationships between nodes, providing a multi-dimensional perspective for the similarity measurement of negative links; Figure 6 As shown in the figure, random sampling (RN) outperforms the negative sample module (MPN) only under a significant positive-negative sample imbalance (1:10), and MPN still performs well under this ratio; as the association network becomes sparse, the information extracted by the meta-path may become messy, affecting the effectiveness of the features, but at a smaller sample ratio, MPN can effectively extract negative sample features; by adopting a more targeted negative sample selection strategy (such as MPN), the model's bias is effectively reduced, and its generalization ability and prediction accuracy are improved.

[0119] Metapaths describe the relationships between different types of nodes in a heterogeneous network and can effectively capture and understand the complex interactions between nodes. However, the selection of metapaths must strike a balance between path length and information richness. If the metapath is too short, it may not fully capture the necessary semantic information and network structural features, limiting the associations learned by the model and, in turn, affecting prediction accuracy. Conversely, if the metapath is too long, it may introduce excessive noise and irrelevant information, making the prediction model complex and difficult to optimize. Taking dataset B as an example, Table 5 shows that a fixed metapath length of K = 3 is the most effective. This length not only captures direct relationships between key nodes but also reveals indirect connections between nodes, thereby providing rich semantic information. In this case, the path complexity is moderate, reflecting the structural characteristics of the network without significantly degrading model performance due to information overload.

[0120] Table 5 Performance comparison of MPMB-DR under different meta-path lengths

[0121]

[0122] Sensitivity test:

[0123] To investigate the impact of data noise in the association network on the model, 20% and 30% of the nodes were randomly deleted from the dataset for sensitivity analysis. To mitigate random errors, the experiment used 10-CV iterations for 20 times. As can be seen from Table 6, under different degrees of node deletion, the proposed model still exhibited satisfactory performance and showed good robustness compared to the original baseline performance.

[0124] Table 6 AUC comparison based on different degrees of node removal

[0125]

[0126] Case Study:

[0127] To evaluate the practical utility of the method of the present invention, a case study of a specific disease was conducted based on the F dataset. When investigating a specific disease, it is necessary to exclude known associations with the specific disease and treat it as a new entity. The top 10 most relevant drugs were predicted using MPMB-DR, and it was observed how many of them had been validated by published experimental literature.

[0128] The first case study is Alzheimer's disease (AD), a common neurodegenerative disorder that progressively impairs brain function, leading to declines in memory, thinking, and behavior. Promising drug treatments are in late-stage clinical trials, including anti-amyloid β, anti-tau, and anti-inflammatory strategies. The top 10 drugs predicted by MPMB-DR for the treatment of AD are shown in Table 7. Seven of these drugs have been validated in the relevant medical literature. For example, ropinirole, a dopamine agonist known for its role in treating conditions such as restless legs syndrome, can reverse the effects of neuroinflammation and cell death by downregulating the MARK4-NFκβ signaling system in Alzheimer's disease. Lixi's research has shown that the autophagy enhancer carbamazepine can alleviate memory deficits and brain amyloid β pathology in a mouse model of Alzheimer's disease.

[0129] Table 7 Ranking of the top 10 potential drugs for treating Alzheimer's disease identified by MPMB-DR

[0130]

[0131] Parkinson's disease (PD) is one of the most common neurodegenerative diseases. It is a progressive, multiorgan proteinopathy caused by misfolded α-synuclein (αSyn), resulting in various motor and non-motor disorders due to a widespread process of synaptic and neuronal loss in the nervous system. Table 8 identifies eight potential drugs with literature support for effective treatment of PD. For example, levodopa, predicted in the table, is the most common drug used as a first-line treatment for Parkinson's disease. At low concentrations and in the presence of glial cells, levodopa can protect cultured dopaminergic neurons and upregulate antioxidant and anti-apoptotic proteins. Furthermore, in the 1960s, amantadine, a stable, colorless, crystalline, symmetrical amine, was discovered to have inhibitory effects against various influenza viruses. Previous studies have confirmed the efficacy of amantadine in treating levodopa-induced movement disorders in patients with Parkinson's disease.

[0132] Table 8 Ranking of the top 10 potential drugs for the treatment of Parkinson's disease identified by MPMB-DR

[0133]

[0134] Traditional drug repositioning methods usually rely on manual screening of clinical drug databases, which is time-consuming and costly. With the continuous accumulation of high-throughput genomics and proteomics data, artificial intelligence-based computational methods have gradually attracted attention and can provide potential research objects for biological experiments, thereby effectively saving experimental time and costs. However, when constructing heterogeneous networks, only drug similarity networks, disease similarity networks and drug-disease association information are usually considered, while other closely related biomolecular information is ignored. How to integrate more bioinformatics data and filter out effective information to improve the accuracy of model predictions is a challenge.

[0135] To overcome this challenge, the present invention proposes the MPMB-DR method, which deeply integrates diverse bioinformation data sources in terms of topological structure and node feature fusion, designs a negative sampling module based on metapath and similarity principles, and fuses original bioinformation features. Starting from the essential characteristics of bioinformation and the mined association features, the complex relationships between bioinformation are deeply analyzed. Experimental results show that the MPMB-DR method can effectively mine deep-level association features with an accuracy rate that is better than most existing methods. In short, MPMB-DR provides a new solution for the prediction of complex associations between drugs and diseases. The proposed method is versatile and can be applied to the prediction of associations between other bioinformation, and is expected to provide effective guidance for biological experiments.

[0136] With the above-described preferred embodiments of the present invention as a guide, and with reference to the above description, relevant personnel are fully capable of making various changes and modifications without departing from the technical scope of this invention. The technical scope of this invention is not limited to the contents of the specification and must be determined according to the scope of the claims.

Claims

1. A drug repositioning method based on meta-pathway integration of multi-source biological data, characterized in that: The following steps are involved: Step 1: Obtain drug, disease, and protein similarity matrices; drug-protein, drug-disease, and disease-protein adjacency matrices; The original datasets of disease comprehensive similarity matrix and drug comprehensive similarity matrix; Step 2: Using the fusion heterogeneous network, the association matrices of the drug, disease, and protein pairs and their respective similarity matrices are fused to obtain a drug-protein-disease heterogeneous graph; Step 3: Design multi-level association paths based on negative sampling of meta-paths and similarities, fuse similarity information of the same-level networks in heterogeneous networks based on the similarity principle, and select negative link samples; Step three specifically includes: According to the type and length of different meta-paths, the drug-disease adjacency matrix A dr-d Assign different weights to obtain the weighted drug-disease adjacency matrix A ’ dr-d ; For drugs and diseases , based on the association matrix and their respective similarity matrices to measure the distance between the two , the formula is: in, , , Indications and drugs A collection of related diseases, Expression and disease associated drug collections; Step 4: Construct a drug feature extraction module based on SMILE sequences, apply GCN to the molecular structure graph formed by the drug SMILE sequence, and extract drug features; The formula for drug characteristics is: in, is a self-circulating undirected graph, For drug characteristics; Step 5: Construct the SMS-MPF module to walk the meta-paths through the heterogeneous network, extract and fuse the features between different meta-paths, and obtain the association features between drugs and diseases; The formula for the association characteristics between drugs and diseases is: in, MPF is the meta-path fusion operation, is a node Features, Is with the node A collection of related meta-paths. Represents a meta-path based The shared maximum subgraph of Step 6: Input the disease phenotype similarity matrix, drug characteristics, and drug-disease association characteristics into the association prediction module to obtain the final node representation; The final node representation formula is: in, 、 are the association features of drugs and diseases, For the final drug characteristics, S d is the disease phenotype similarity matrix; The association prediction module uses the sigmoid function loss function, the formula is: in, are learnable diagonal matrix parameters, is the sigmoid function, represents the set of positive links, i.e., the observation graph; is the negative link of the sampling The distribution it obeys.

2. The drug repositioning method based on meta-path integration of multi-source biological data according to claim 1, characterized in that: The fused heterogeneous network is a three-layer heterogeneous network.

3. The drug repositioning method based on meta-path integration of multi-source biological data according to claim 1, characterized in that: The heterogeneous network in step five is a three-layer heterogeneous network.

4. A drug repositioning system based on meta-path integration of multi-source biological data, characterized by: include: a memory for storing instructions executable by the processor; A processor, configured to execute instructions to implement the drug repositioning method based on meta-pathway integration of multi-source biological data as described in any one of claims 1 to 3.

5. A computer-readable medium storing computer program code, characterized in that When the computer program code is executed by a processor, the computer program code implements the drug repositioning method based on meta-pathway integration of multi-source biological data according to any one of claims 1 to 3.

Images