Drug relocation method based on multi-source information fusion and multi-level hypergraph learning

Abstract

The invention discloses a drug relocation method based on multi-source information fusion and multi-level hypergraph learning, and belongs to the technical field of drug relocation. On a model architecture, protein biological entity information is introduced for multi-source biological information fusion, and a global hypergraph network and a local hypergraph network are constructed; in the local hypergraph network, constructing a drug-disease local sub-network, a drug-protein local sub-network and a disease-protein local sub-network, and mining local feature information by adopting a targeted graph neural network; in the global hypergraph network, constructing a drug-disease-protein complex association network and directly mining global feature information; besides, the multi-level features and the original features are fused by adopting an attention weight strategy and a residual jumping strategy, so that a finally predicted drug-disease correlation matrix is obtained, and accurate prediction of drug-disease correlation can be realized.

Description

Technical Field

[0001] This invention relates to the field of drug relocation technology, specifically to a drug relocation method based on multi-source information fusion and multi-level hypergraph learning. Background Technology

[0002] To improve the success rate and speed of drug development, researchers have explored a range of new methods, among which drug repositioning has garnered significant attention due to its innovation and effectiveness. Drug repositioning refers to a research method that uses approved, discontinued, or abandoned drugs as research subjects, employing scientific methods to explore new indications for these drugs, thereby treating other diseases. The theoretical basis of drug repositioning lies primarily in two aspects: first, multiple targets can interact with the same drug; second, disease-related targets are often closely related to multiple biological processes.

[0003] Existing drug relocation methods primarily rely on mining the biological information of drugs and diseases to predict potential drug-disease associations. However, in the human biological environment, drug molecules and disease targets are not only related to each other but also interact in complex ways with various biological entities, such as proteins, genes, and enzymes. This biological association information can provide rich feature information for drug relocation, helping to improve the accuracy of model predictions.

[0004] In summary, given the increasing inadequacy of traditional drug development methods to meet the demands of modern medicine, drug repositioning methods have garnered significant attention from researchers due to their outstanding advantages of safety and efficiency. Numerous successful cases further highlight the application potential of this method. Against this research backdrop, this invention addresses the shortcomings of existing drug repositioning methods by focusing on a graph neural network-based approach. The aim is to achieve accurate prediction of drug-disease correlations, thereby providing an effective solution for drug development. Summary of the Invention

[0005] The technical problem to be solved by this invention is: how to construct a hypergraph network among multiple biological entities, how to extract features from different information in different hypergraph networks when the hypergraph network is obtained, and how to effectively integrate these features to generate more expressive features. This invention provides a drug relocation method based on multi-source information fusion and multi-level hypergraph learning.

[0006] The present invention solves the above-mentioned technical problems through the following technical solution, and the present invention includes the following steps:

[0007] S1: Data Collection

[0008] Drug similarity matrix, disease similarity matrix, drug-disease correlation matrix, drug-protein interaction matrix, disease-protein interaction matrix, and protein-protein interaction matrix are collected from existing databases to form a benchmark dataset. The benchmark dataset is then divided into training set and test set.

[0009] S2: Model Building

[0010] A drug relocation model is constructed, which has a multi-layer structure. Each layer includes a local hypergraph network and a global hypergraph network. In the local hypergraph network, multiple graph neural networks are used to obtain various local features, and an average pooling strategy is used to fuse these local features to obtain the output features of the local hypergraph network. In the global hypergraph network, a single graph neural network is used to obtain the output features of the global hypergraph network. A residual skipping strategy is used to introduce the output features of the previous layer in the drug relocation model into the hypergraph hierarchical feature fusion stage of the current layer to achieve hypergraph hierarchical feature fusion. Finally, a drug-disease interaction prediction matrix is ​​obtained. Except for the last layer, the output features of each layer in the model include the output features of the global hypergraph network and the output features of the local hypergraph network.

[0011] S3: Model Training

[0012] The constructed drug relocation model is trained using the training set data to obtain the trained drug relocation model;

[0013] S4: Interaction Prediction

[0014] The test set data is input into the trained drug relocation model, which outputs a drug-disease interaction prediction matrix.

[0015] Furthermore, in step S2, the local hypergraph network has three sub-views: a drug-disease local view, a drug-protein local view, and a disease-protein local view. In the drug-disease local view of the initial layer of the drug relocation model, the drug similarity matrix, the disease similarity matrix, and the drug-disease interrelation matrix are constructed into a similarity graph network, an association graph network, and a comprehensive graph network, respectively, and feature extraction is performed through graph neural networks. In the drug-protein local view of the initial layer of the drug relocation model, the drug-drug similarity matrix, the drug-protein interaction matrix, and the protein-protein interaction matrix are constructed into a drug-protein graph network, and feature extraction is performed through graph neural networks. In the disease-protein local view of the initial layer of the drug relocation model, the disease similarity matrix, the disease-protein interaction matrix, and the protein-protein interaction matrix are constructed into a disease-protein graph network, and feature extraction is performed through graph neural networks. The global hypergraph network only has a drug-disease-protein global view. In the drug-disease-protein global view, all the above matrices are constructed into a global hypergraph network, and feature extraction is performed through graph neural networks. Among them, the drug-disease interrelation matrix is ​​the drug-disease interaction matrix.

[0016] Furthermore, in the drug-disease local view, the construction methods for the drug similarity matrix, disease similarity matrix, and drug-disease interrelation matrix as similarity graph networks, association graph networks, and comprehensive graph networks are as follows:

[0017]

[0018] in, This represents the drug similarity matrix, where m is the number of drug types. Represents the disease similarity matrix, where n is the number of disease types; A rd ∈R mxn A rd This is a drug-disease correlation matrix. This is the transposed form of the drug-disease correlation matrix; and These represent the similarity graph network, association graph network, and comprehensive graph network constructed in the drug-disease local view, respectively.

[0019] The similarity graph network uses the graph isomorphic neural network GIN for feature extraction, as shown in the following expression:

[0020]

[0021] in, The input is the feature matrix of the initial similarity graph network; This indicates the extraction of the feature matrix obtained by GIN; MLP represents multilayer perceptron; ∈ represents learnable parameters; This indicates that neighboring nodes are being aggregated;

[0022] The graph network uses a graph convolutional neural network (GCN) for feature extraction, as shown in the following expression:

[0023]

[0024] in, This represents the feature matrix of the initial correlation graph network input; denoted as the feature matrix extracted by GCN; A represents the adjacency matrix of the graph; W1 represents the learnable weight matrix.

[0025] The comprehensive graph network uses an improved graph attention neural network, GATv2, for feature extraction, as shown in the following expression:

[0026]

[0027] in, The input is the feature matrix of the initial synthetic graph network; is the feature matrix extracted by GATv2; A represents the adjacency matrix of the graph; W2 represents the learnable weight matrix.

[0028] Furthermore, in the drug-protein local view, the construction method for building the drug-protein graph network by combining the disease similarity matrix, disease-protein interaction matrix, and protein-protein interaction matrix is ​​as follows:

[0029]

[0030] in, This represents the drug similarity matrix, where m is the number of drug types; A rp ∈R mxl This is a drug-protein interaction matrix, where l represents the number of proteins; This is the transposed form of the drug-protein interrelation matrix; A pp ∈R lxl This is the protein-protein interaction matrix;

[0031] The drug-protein graph network (BGNN) is used for feature extraction using a bilinear graph neural network, as shown in the following expression:

[0032]

[0033] in, This represents the initial drug-protein local view feature matrix input. W3 represents the feature matrix extracted by BGNN, where A represents the adjacency matrix of the graph and W3 represents the learnable weight matrix.

[0034] Furthermore, in the disease-protein local view, the disease similarity matrix, disease-protein interaction matrix, and protein-protein interaction matrix are constructed into a disease-protein graph network as follows:

[0035]

[0036] in, Represents the disease similarity matrix, where n is the number of disease types; A dp ∈R nxl This is a disease-protein interaction matrix, where l represents the number of proteins; This is the transposed form of the disease-protein interrelation matrix; A pp ∈R lxl This is the protein-protein interaction matrix;

[0037] The disease-protein graph network (BGNN) is used for feature extraction using a bilinear graph neural network, as shown in the following expression:

[0038]

[0039] in, This represents the initial disease-protein local view feature matrix input. denoted as the feature matrix extracted by BGNN; A represents the adjacency matrix of the graph; W4 represents the learnable weight matrix.

[0040] Furthermore, in the global view of drugs, diseases, and proteins, the construction method for building all matrices into a global hypergraph network is as follows:

[0041]

[0042] in, This represents the drug similarity matrix, where m is the number of drug types. A represents a similarity matrix for diseases, where n is the number of disease types; pp ∈R lxl A is the protein-protein interaction matrix; rd ∈R mxn A rd A is a drug-disease correlation matrix; rp ∈R mxl A is the drug-protein interaction matrix; dp ∈R nxl Disease-protein interaction matrix;

[0043] A learnable graph convolutional attention neural network (L-CAT) is used to extract features from the global hypergraph network, resulting in the following expression for the output features of the global hypergraph network:

[0044]

[0045] in, This represents the initial global hypergraph network feature matrix as input; denoted as the feature matrix extracted by L-CAT; A represents the adjacency matrix of the graph; W5 represents the learnable weight matrix.

[0046] Furthermore, in the learnable graph convolutional attention neural network L-CAT, two learnable parameters are introduced before and after the attention coefficient calculation, as shown in the expression:

[0047]

[0048] in, This indicates that the convolution operation updates the node features. Let i represent the set of neighboring nodes. Represent the characteristics of neighboring nodes; W represents the attention coefficient between node i and node j, and W6 represents the learnable weight matrix.

[0049] Furthermore, in step S2, an average pooling strategy is used to fuse multiple local features obtained from the local hypergraph network, resulting in the following expression for the output features of the local hypergraph network:

[0050]

[0051] in, and A pp_1 This represents the drug-drug related features, disease-disease related features, and protein-protein related features obtained after local view feature extraction and average pooling operations; A rd_1 A rp_1 and A dp_1 These represent the extracted drug-disease interaction relationships, drug-protein interaction relationships, and disease-protein interaction relationships in the drug-disease local view, drug-protein local view, and disease-protein view, respectively.

[0052] Furthermore, in step S2, the expression for hypergraph hierarchical feature fusion is as follows:

[0053]

[0054] in, This represents the local hypergraph network output features extracted by the (l+1)th layer of the model; This represents the global hypergraph network output features extracted by the (l+1)th layer of the model; λ represents the original input of the (l+1)th layer of the network, i.e., the layer feature output by the lth layer; 12 ,λ 22 ,λ 32 ∈(0,1) are learnable weights, λ 12 +λ 22 +λ 32 =1.

[0055] Furthermore, in step S3, the loss function during model training is the polynomial loss function PolyLoss, with the following expression:

[0056]

[0057] Among them, y, These represent the model's actual value and predicted value, respectively. is the weight parameter for each polynomial term; p represents the highest order of the polynomial.

[0058] Compared with existing technologies, this invention has the following advantages: The drug relocation method based on multi-source information fusion and multi-level hypergraph learning, in terms of model architecture, introduces protein biological entity information for multi-source biological information fusion and constructs a global hypergraph network and a local hypergraph network; in the local hypergraph network, it constructs drug-disease, drug-protein, and disease-protein local sub-networks and uses targeted graph neural networks to mine local feature information; in the global hypergraph network, it constructs a complex drug-disease-protein association network and directly mines global feature information; furthermore, it uses attention weighting and residual skipping strategies to fuse multi-level features and original features, thereby obtaining the final predicted drug-disease interrelation matrix, which can achieve accurate prediction of drug-disease interrelation. Attached Figure Description

[0059] Figure 1 This is a flowchart illustrating the drug relocation method based on multi-source biological information fusion and multi-level hypergraph learning in an embodiment of the present invention.

[0060] Figure 2 This is a diagram of the original DRMBML input format in an embodiment of the present invention;

[0061] Figure 3 This is a DRMBML model framework diagram in an embodiment of the present invention;

[0062] Figure 4 This is the ROC curve of the ten-fold cross-validation experiment in this embodiment of the invention, where (a) is the T1 dataset and (b) is the T2 dataset;

[0063] Figure 5 This is a bar chart of AUROC values ​​from a potential drug prediction experiment for unknown diseases in this embodiment of the invention.

[0064] Figure 6 This is a bar chart of AUPR values ​​from a potential drug prediction experiment for an unknown disease in an embodiment of the present invention;

[0065] Figure 7 This is the ROC curve of the DRMBML ablation experiment in this embodiment of the invention. Detailed Implementation

[0066] The embodiments of the present invention are described in detail below. These embodiments are implemented based on the technical solution of the present invention, and provide detailed implementation methods and specific operation processes. However, the scope of protection of the present invention is not limited to the following embodiments.

[0067] like Figure 1 As shown, this embodiment provides a technical solution: a drug relocation method based on multi-source biological information fusion and multi-level hypergraph learning, comprising the following steps:

[0068] S1: Data Collection

[0069] This invention primarily uses benchmark datasets T1 and T2 to evaluate the performance of DRMBML (Drug Relocation Model Based on Multi-Source Biological Information Fusion and Multi-Level Hypergraph Learning). The two datasets contain: drug similarity matrix, disease similarity matrix, drug-disease interrelation matrix (drug-disease interaction matrix), drug-protein interaction matrix, disease-protein interaction matrix, and protein-protein interaction matrix.

[0070] Regarding drug and disease data, in the T1 dataset, the drug-disease interaction matrix was derived from drug-disease interaction information collected by Zhang et al. from the CTD database, including 15,630 interactions between 263 drugs and 480 diseases. The drug-drug similarity matrix and disease-disease similarity matrix were integrated from information on the similarity of related targets, channels, drug-drug interactions, and enzymes. In the T2 dataset, the drug-disease interaction matrix was derived from the repoDB database, containing interaction information between 850 drugs and 339 diseases. The drug-drug similarity matrix and disease-disease similarity matrix were integrated from relevant clinical similarities, drug side effect similarities, and chemical similarities.

[0071] For protein data, the drug-protein interaction matrix is ​​sourced from the DGIdb database, which contains information on genes and their known drug associations; the disease-protein interaction matrix is ​​sourced from the DisGeNET database, containing information on the interactions between diseases and proteins; and the protein-protein interaction matrix is ​​based on protein information research by Menche et al. The original DRMBML input format is as follows: Figure 2 As shown.

[0072] The specific details of the biological data contained in the benchmark datasets T1 and T2 are shown in Table 1 below:

[0073] Table 1. Composition of the benchmark dataset

[0074] Dataset T1 T2 Number of drug classes 263 850 Number of disease classes 480 339 Number of protein classes 6059 5105 Drug-disease interactions 15630 1921 Drug-protein interactions 5620 10872 Disease-protein interactions 20019 12625 Protein-protein interactions 49406 41903

[0075] S2: Hypergraph Network Construction

[0076] S2 includes the following steps:

[0077] S21: Construction of Local Hypergraph Network. The fundamental task of this invention is to predict potential drug-disease associations, and the drug and disease features mined from drug-disease association information are the most direct and primary source for achieving accurate predictions. In the local drug-disease view, DRMBML constructs the drug and disease input data (drug similarity matrix, disease similarity matrix, and drug-disease cross-association matrix) into a similarity graph network, an association graph network, and a comprehensive graph network, thereby fully mining drug and disease features. The specific construction method is shown in the following expression:

[0078]

[0079] in, This represents the drug similarity matrix, where m is the number of drug types. Represents the disease similarity matrix, where n is the number of disease types; A rd ∈R mxn A rd This is a drug-disease correlation matrix. This is the transposed form of the drug-disease correlation matrix; and These represent the similarity graph network, association graph network, and comprehensive graph network constructed in the drug-disease local view, respectively.

[0080] Drug-Protein Local View: The main purpose of introducing protein information in this invention is to enrich drug and disease-related features through drug-protein association information and disease-protein association, thereby providing auxiliary information for predicting potential drug-disease associations. Therefore, in the drug-protein local view, this invention constructs a drug-protein graph network from relevant information of drugs and proteins (drug-drug similarity matrix, drug-protein interaction matrix, and protein-protein interaction matrix) to mine drug and protein features. The construction method is as follows:

[0081]

[0082] in, This represents the drug similarity matrix, where m is the number of drug types; A rp ∈R mxl This is a drug-protein interaction matrix, where l represents the number of proteins; This is the transposed form of the drug-protein interrelation matrix; A pp ∈R lxl This is the protein-protein interaction matrix.

[0083] Disease-Protein Local View: Similar to the drug-protein local view, this invention constructs a disease-protein graph network from disease-protein related information (disease similarity matrix, disease-protein interaction matrix, and protein-protein interaction matrix) within the disease-protein local view, thereby extracting disease and protein information. The construction method of this graph network is as follows:

[0084]

[0085] in, Represents the disease similarity matrix, where n is the number of disease types; A dp ∈R nxl This is a disease-protein interaction matrix, where l represents the number of proteins; This is the transposed form of the disease-protein interrelation matrix; A pp ∈R lxl This is the protein-protein interaction matrix.

[0086] S22: This invention constructs a global hypergraph network in addition to the local hypergraph network. Its construction method is as follows:

[0087]

[0088] in, This represents the drug similarity matrix, where m is the number of drug types. A represents a similarity matrix for diseases, where n is the number of disease types;pp ∈R lxl A is the protein-protein interaction matrix; rd ∈R mxn A rd A is a drug-disease correlation matrix; rp ∈R mxl A is the drug-protein interaction matrix; dp ∈R nxl This is the disease-protein interaction matrix. The DRMBML model framework diagram is shown below. Figure 3 As shown.

[0089] S3: Feature Extraction

[0090] The following steps are involved in S3:

[0091] S31: Local Hypergraph Network Feature Extraction. Drug-Disease Local Subview: Similarity Graph Network uses Graph Isomorphic Neural Network (GIN) for feature extraction of isomorphic graphs. The expression for GIN is as follows:

[0092]

[0093] in, The input is the feature matrix of the initial similarity graph network; This indicates the extraction of the feature matrix obtained by GIN; MLP stands for Multilayer Perceptron, used for nonlinear transformation; ∈ represents the learnable parameters; This indicates that neighboring nodes are being aggregated.

[0094] The association graph network uses a graph convolutional neural network (GCN) for feature extraction, mining direct association information in the drug-disease relationship in the data. Its expression is as follows:

[0095]

[0096] in, This represents the feature matrix of the initial correlation graph network input; denoted as the feature matrix extracted by GCN; A represents the adjacency matrix of the graph, used to describe the connection relationships between nodes; W1 represents the learnable weight matrix, used to control feature transformation.

[0097] The comprehensive graph network employs an improved graph attention neural network (GATv2), which solves the "static transformation problem" of the classic graph attention neural network (GAT) and exhibits superior performance on complex heterogeneous graph structures. Its expression is shown below:

[0098]

[0099] in, The input is the feature matrix of the initial synthetic graph network; is the feature matrix extracted by GATv2; A represents the adjacency matrix of the graph; W2 represents the learnable weight matrix.

[0100] GATv2 achieves its attention score dynamically adapts to complex combinations of node features by first performing a nonlinear transformation and then calculating the attention. It exhibits stronger classification ability and robustness when handling complex feature relationships in heterogeneous graphs and dynamic graphs. Therefore, this invention uses GATv2 as a feature extractor for comprehensive graph networks, enabling the exploration of potential associations between drugs and diseases from complex relationships. The specific formula is as follows:

[0101] a ij =LeakyReLU(a T ·[Wh i / / Wh j ])

[0102] a ij =LeakyReLU(a T ·W[h i / / h j ])

[0103] Among them, a T For a learnable single-layer feedforward neural network; W is a learnable weight matrix; h i and h j ...

[0104] Drug-protein local view and disease-protein local view feature extraction:

[0105] This invention enriches the association features between drugs and proteins in the hypergraph network by extracting features from two local subviews: drug-protein and disease-protein. This helps the hypergraph network fully learn the potential relationships between drugs and diseases from the drug-disease-protein bioinformatics network. The invention employs a bilinear graph neural network (BGNN) as the feature extractor for these two local subviews. Unlike traditional graph neural networks that only consider first-order neighbor information, BGNN combines bilinear mapping with graph neural networks to capture second-order relationships (i.e., interactions between nodes). This design allows BGNN to more finely characterize the dependencies between nodes, thus enabling it to handle more complex graph structures. Experiments have verified that BGNN improves the accuracy and robustness of models in various graph learning tasks, such as node classification, graph classification, and inter-relationship prediction. Its expression is as follows:

[0106]

[0107] in, This represents the initial drug-protein local view feature matrix input. This represents the feature matrix extracted by BGNN; This represents the initial disease-protein local view feature matrix input. W represents the feature matrix extracted by BGNN; A represents the adjacency matrix of the graph; W3 and W4 represent the learnable weight matrices.

[0108] S32: Feature Extraction from Global Hypergraph Networks

[0109] To directly mine and capture potential associations from the node features and interrelationships of the three entities—drug, disease, and protein—this invention constructs a complex heterogeneous graph network containing multiple types of nodes and various interaction relationships. When processing such highly complex graph networks, it is difficult to directly determine their graph data structure, thus compromising the performance of traditional graph neural networks. Therefore, this invention introduces a learnable graph convolutional attention network (L-CAT) for data analysis and feature extraction of the global hypergraph network. The feature extraction process is shown in the expression:

[0110]

[0111] in, This represents the initial global hypergraph network feature matrix as input; denoted as the feature matrix extracted by L-CAT; A represents the adjacency matrix of the graph; W5 represents the learnable weight matrix.

[0112] Studies have shown that data distribution and noise levels in graph data have a significant impact on the performance of graph neural networks. Applying convolutional operations before calculating the attention coefficients in GAT can effectively alleviate the problems caused by data distribution. This operation is called the Graph Convolutional Attention Layer (CAT), and its attention coefficient expression is shown below:

[0113]

[0114] in, This indicates that the convolution operation updates the node features. Let i represent the set of neighboring nodes. Represent the characteristics of neighboring nodes; W represents the attention coefficient between node i and node j, and W6 represents the learnable weight matrix.

[0115] When faced with graph data of varying distribution and structure, traditional graph neural networks (GCN, GAT, and CAT) may each have their own strengths and weaknesses, and there is no absolutely optimal choice. Therefore, L-CAT uses interpolation to adjust the information aggregation method of the graph neural network. This allows L-CAT to learn the graph structure data and automatically adjust itself during model training, thereby achieving efficient analysis and extraction of real-world data. Specifically, L-CAT introduces two learnable parameters before and after the attention coefficient calculation, as shown in the following expression:

[0116]

[0117] Here, λ1 and λ2 are the introduced learnable parameters; when λ1 = 0, L-CAT transforms into a traditional Graph Convolutional Neural Network (GCN); when λ1 = 1 and λ2 = 0, L-CAT transforms into a traditional Graph Attention Neural Network (GAT); and when λ1 = 1 and λ2 = 1, L-CAT transforms into a Convolutional Attention Neural Network (CAT). As L-CAT analyzes and learns from the data, the model automatically adjusts the learnable parameters during training, thus adjusting the information aggregation method of L-CAT between GCN, GAT, and CAT.

[0118] S4: Feature Fusion

[0119] The following steps are involved in S4:

[0120] S41: Feature Fusion in Local Hypergraph Networks. Local hypergraph networks not only mine pairwise interaction information between biological entities but also provide drug, disease, and protein features from multiple sub-views. This multi-view biological entity information is crucial for enriching the feature representation of the hypergraph network and improving model prediction performance. Therefore, this invention employs an average pooling strategy to fuse drug, disease, and protein features from multiple sub-views within the local hypergraph network and reconstructs them as the output features of the local hypergraph network, facilitating subsequent feature fusion at the hypergraph level. The specific construction method is shown in the following expression:

[0121]

[0122] in, and A pp_1 This represents the drug-drug related features, disease-disease related features, and protein-protein related features obtained after local view feature extraction and average pooling operations; A rd_1 A rp_1 and A dp_1 These represent the extracted drug-disease interaction relationships, drug-protein interaction relationships, and disease-protein interaction relationships in the drug-disease local view, drug-protein local view, and disease-protein view, respectively.

[0123] S42: Hypergraph Network Hierarchical Feature Fusion. DRMBML provides multi-level hypergraph output features for predicting potential drug-disease associations by designing local and global hypergraph networks. Furthermore, graph neural networks extract deep features through multi-layer network stacking, but shallow feature loss and gradient explosion may occur during network layer deepening. Therefore, this invention employs a residual skip strategy to introduce the previous layer output into the hypergraph hierarchical feature fusion stage of the current layer. To fully fuse these multi-level output features and improve the model's predictive performance, this invention designs an attention weight strategy. Specifically, the method assigns attention coefficients to the global hypergraph network output features, local hypergraph network output features, and previous layer output features. Through model training, these attention coefficients are automatically adjusted, thereby achieving efficient fusion of layer output features. Its expression is as follows:

[0124]

[0125] in, This represents the local hypergraph network output features extracted by the (l+1)th layer of the model; This represents the global hypergraph network output features extracted by the (l+1)th layer of the model; λ represents the original input of the (l+1)th layer of the network, i.e., the layer feature output by the lth layer; 12 ,λ 22,λ 32 ∈(0,1) are learnable weights, λ 12 +λ 22 +λ 32 =1, which can be automatically adjusted through backpropagation during model training.

[0126] S5: Model Training

[0127] This invention employs a polynomial loss function (PolyLoss) as the model optimization strategy. Inspired by the Taylor expansion approximation principle, Leng et al. proposed PolyLoss in 2022. The basic idea of ​​this loss function is to calculate the error between the model's predicted value and the true value in polynomial form. In drug relocation tasks that require handling complex biological entity relationships, PolyLoss, by introducing higher-order polynomials, can better capture complex nonlinear relationships, enabling the model to maintain superior performance when facing imbalanced samples and sparse data. Furthermore, PolyLoss effectively reduces overfitting by regularizing the polynomial in the loss function. Its expression is shown below:

[0128]

[0129] Among them, y, These represent the model's actual value and predicted value, respectively. These are the weight parameters for each polynomial term; p represents the highest order of the polynomial. This invention employs the Adam optimizer and Dropout technique to improve the training efficiency of the model while effectively preventing overfitting. Furthermore, the model also uses L2 regularization and a recurrent learning rate scheduler.

[0130] This embodiment operates on a server running on a Linux system, equipped with an A6000 graphics card. The server has two GPUs, each with 40GB of dedicated video memory. Specific parameters are shown in Table 2.

[0131] Table 2 Experimental Environment Configuration

[0132] Configuration name Parameter / version number CPU Intel Xeon Gold 6248R GPU Nvidia RTX A6000 Pytorch 2.1.2 CUDA 12.2 Python 3.9.19

[0133] In this embodiment, the specific parameter settings for DRMBML are shown in Table 3:

[0134] Table 3 Model Parameter Settings

[0135] Parameter name Parameter value Number of network layers 3 Training epochs 600 Optimizer Adam Feature dimension 256 Dropout 0.6 Learning rate 0.05

[0136] S6: Experimental Results and Evaluation

[0137] This invention uses AUROC and AUPR as the main performance metrics to measure the predictive performance of DRMBML. During the comparative experiments on the benchmark datasets T1 and T2, the comparison models were configured with the required parameters.

[0138] In the comparative experiments, the AUROC and AUPR of ten 10-fold cross-validation experiments are shown in Tables 4 and 5. Furthermore, this invention also plotted the ROC curves for the 10-fold cross-validation experiments, as shown below. Figure 4 As shown, this further examines the predictive performance of DRMBML.

[0139] Table 4. AUROC values ​​from ten 10-fold cross-validation experiments.

[0140]

[0141]

[0142] Table 5. AUPR values ​​of the ten-fold cross-validation experiment.

[0143] AUPR DRRS BNNR DRIMC SCPMF DRHGCN DRMBML T1 0.460±0.002 0.552±0.002 0.512±0.004 0.550±0.002 0.558±0.002 0.576±0.001 T2 0.381±0.004 0.498±0.001 0.462±0.003 0.494±0.003 0.491±0.001 0.510±0.002 Average 0.421 0.525 0.537 0.522 0.524 0.543

[0144] The fundamental task of drug relocation is to explore potential therapeutic drugs for unknown diseases and provide effective treatment options for new diseases. Therefore, to further evaluate the prediction table performance of DRMBML in real-world applications, this invention conducted experiments on predicting potential drugs for unknown diseases. Specifically, the experiment used all drug-disease interactions (drug-disease association matrices) for 20% of the diseases in the dataset as the test set, and the drug-disease interactions for the remaining 80% of the diseases as the training set. This means that the model, without any known drug-disease associations for the unknown disease, mines the features of drugs and diseases to predict potential drugs for the unknown disease. This design, in principle, simulates the process of finding potential drugs for new diseases in real-world drug relocation tasks. To ensure the stability and validity of the validation results, this invention conducted five random partitioning experiments on the benchmark datasets T1 and T2 respectively and took the average value. The experimental results are as follows: Figure 5 and Figure 6 As shown.

[0145] The experimental results show that, compared to the 10-fold cross-validation experiment, all models exhibited a decline in performance in predicting potential drugs for unknown diseases, particularly on the benchmark dataset T2, which has a higher degree of positive-negative sample imbalance. This is mainly because the models can only analyze and calculate the probability of association between various drugs and unknown diseases based on existing biological association information, leading to a decrease in predictive ability. Although DRMBML's predictive performance was also affected by this phenomenon, thanks to the rich drug and disease features mined from global and local hypergraph networks, DRMBML can still predict potential drugs for unknown diseases relatively accurately. Its average AUROC and average AUPR on the benchmark datasets T1 and T2 outperform existing state-of-the-art models, surpassing the second-best performing DRHGCN model by 0.104 and 0.051, respectively.

[0146] Ablation Experiment: Thanks to the fusion of multi-source biological information and the construction of multi-level hypergraph networks, DRMBML can deeply mine drug and disease feature information in complex biological associations of drugs, diseases and proteins, thereby achieving efficient prediction of potential drug-disease associations. To verify the effectiveness of this design, the present invention conducted an ablation experiment on the benchmark dataset T1 and designed several variant models of DRMBML. These include: (1) DRMBML_NP: This model only uses drug and disease features for feature extraction, which is essentially the same as the drug-disease subview in the local hypergraph network. Therefore, this model only retains this view to complete the prediction task; (2) DRMBML_LOCAL: This model only uses the local hypergraph network to mine drug and disease feature information; (3) DRMBML_GLOBAL: This model only uses the global hypergraph network to mine drug and disease feature information; (4) DRMBML_ADD: This model uses a simple addition strategy to achieve multi-level feature fusion.

[0147] Ablation test results as follows Figure 7As shown, when DRMBML uses only drug and disease information for potential drug-association prediction, its AUROC value is 0.910. However, after introducing protein information, the prediction performance of DRMBML_LOCAL and DRMBML_GLOBAL, which use only a single hypergraph network, improves, with AUROC values ​​of 0.923 and 0.916, respectively. This indicates that the drug-disease-protein bioinformatics network constructed by introducing additional protein information can provide rich biological features for the model to predict potential drug-disease interactions. Furthermore, DRMBML_ADD achieves an AUROC of 0.934, surpassing the prediction performance of DRMBML_LOCAL and DRMBML_GLOBAL. This demonstrates that each layer of the hypergraph network provides unique and crucial biological features for DRMBML, thus contributing to improved model prediction performance. However, compared to the simple additive strategy adopted by DRMBML_ADD, the attention weight strategy in DRMBML better preserves and fuses features from multiple views, helping DRMBML achieve optimal prediction performance with an AUROC value of 0.937.

[0148] Case Study: To further evaluate the predictive performance of DRMBML, this invention selected two representative diseases and their corresponding predictive drugs from the DRMBML prediction results on the benchmark dataset T2, and retrieved evidence of a correlation between the diseases and the predictive drugs from multiple authoritative biological databases. Furthermore, this invention conducted molecular docking experiments to calculate the molecular binding energy between the predictive drugs and the disease target proteins. The experimental details are described below.

[0149] Parkinson's disease (PD) is a progressive neurodegenerative disease characterized by symptoms such as resting tremor, rigidity, and bradykinesia, significantly impacting patients' lives. With an aging population, the number of patients is rapidly increasing, projected to double globally by 2040, making the need for effective treatments increasingly urgent. As shown in Table 6, 80% of the predicted candidate drugs for DRMBML have records related to Parkinson's disease in multiple databases. Among them, the predicted drug dantrolin can reduce oxidative stress and neuroinflammation by inhibiting Ryanodine receptors, potentially alleviating the neurodegenerative process of Parkinson's disease. The predicted drug mannitol can reduce α-synuclein aggregation, potentially offering neuroprotective effects for Parkinson's patients.

[0150] Table 6 Case Studies of the Top 10 Candidate Drugs for Predicting Parkinson's Disease

[0151] Ranking Candidate drugs (DrugBank IDs) Case source 1 Carbidopa (DB00190) CTD, ClinicalTrials.gov, drugcentral.org 2 Phosphor levodopa (DB16171) CTD, ClinicalTrials.gov 3 Dantrolene (DB01219) Unconfirmed 4 Apomorphine (DB00714) CTD, ClinicalTrials.gov, drugcentral.org 5 Itracurium (DB11757) CTD, ClinicalTrials.gov, drugcentral.org 6 Hydroxyurea (DB01005) Unconfirmed 7 Perphenazine (DB00850) CTD, drugcentral.org 8 Isopropylamide (DB00443) drugcentral.org 9 Mannitol (DB00742) CTD, drugcentral.org 10 Biperiden (DB00810) CTD, drugcentral.org

[0152] Liver neoplasms (LN) are tumors originating in the liver. Malignant LN tumors typically develop rapidly, accompanied by symptoms such as abdominal distension, jaundice, or hepatomegaly. In advanced stages, liver failure may also occur. China has a high incidence of liver cancer globally, with approximately 410,000 new cases in 2020, accounting for about half of all liver cancer cases worldwide. As shown in Table 7, the prediction accuracy rate among the top ten candidate drugs in the DRMBML prediction results was 70%. Among them, the predicted drug Ceritinib showed anti-tumor activity in the treatment of liver neoplasms, and the predicted drug Etoposide may enhance anti-cancer effects when combined with chemotherapy.

[0153] Table 7 Case Studies of the Top 10 Candidate Drugs for Predicting Liver Tumors

[0154] Ranking Candidate drugs (DrugBank IDs) Case source 1 Sorafenib (DB09063) CTD, ClinicalTrials.gov 2 Calcium folinate (DB00650) CTD, ClinicalTrials.gov, drugcentral.org 3 Etoposide (DB00773) CTD, ClinicalTrials.gov, drugcentral.org 4 Irinotecan (DB00762) CTD, ClinicalTrials.gov, drugcentral.org 5 Mosapride (DB11675) Unconfirmed 6 Goserelin (DB00014) CTD, ClinicalTrials.gov 7 Vandetanib (DB05294) CTD, ClinicalTrials.gov, drugcentral.org 8 Fulvestrant (DB00947) CTD, ClinicalTrials.gov 9 Voglibose (DB04878) Unconfirmed 10 Melatonin (DB01065) Unconfirmed

[0155] To further investigate the validity of DRMBML's predictions for candidate drugs, this invention, in addition to the aforementioned case study, selected Parkinson's disease from the case study for molecular binding energy calculation. Specifically, this invention first selected five target proteins of Parkinson's disease (PDB numbers: 1SOA, 2OR3, 3CY6, 3SF8, and 4OQ4) from the RCSB Protein Data Bank (RCSB PDB) protein database as research objects. Secondly, docking experiments were conducted with the candidate drugs predicted by DRMBML in AutodockVina software, and the molecular binding energy was calculated to measure the strength of the molecular affinity between the candidate drugs and Parkinson's disease. The docking experiment results are shown in Table 8. The absolute values ​​of the molecular binding energies between the top 10 predicted drugs and the five target proteins were all at a high level, indicating a strong interaction between the predicted drugs and Parkinson's disease. This is an important basis for the predicted drugs to be considered as potential therapeutic agents for further research.

[0156] Table 8. Predicted molecular binding energies of candidate drugs to Parkinson's disease target proteins using DRMBML.

[0157]

[0158] Although embodiments of the present invention have been shown and described above, it is understood that the above embodiments are exemplary and should not be construed as limiting the present invention. Those skilled in the art can make changes, modifications, substitutions and variations to the above embodiments within the scope of the present invention.

Claims

1. A drug relocation method based on multi-source information fusion and multi-level hypergraph learning, characterized in that, Includes the following steps: S1: Data Collection Drug similarity matrix, disease similarity matrix, drug-disease correlation matrix, drug-protein interaction matrix, disease-protein interaction matrix, and protein-protein interaction matrix are collected from existing databases to form a benchmark dataset. The benchmark dataset is then divided into training set and test set. S2: Model Building A drug relocation model is constructed, which has a multi-layer structure. Each layer includes a local hypergraph network and a global hypergraph network. In the local hypergraph network, multiple graph neural networks are used to obtain various local features, and an average pooling strategy is used to fuse these local features to obtain the output features of the local hypergraph network. In the global hypergraph network, a single graph neural network is used to obtain the output features of the global hypergraph network. A residual skipping strategy is used to introduce the output features of the previous layer in the drug relocation model into the hypergraph hierarchical feature fusion stage of the current layer to achieve hypergraph hierarchical feature fusion. Finally, a drug-disease interaction prediction matrix is ​​obtained. Except for the last layer, the output features of each layer in the model include the output features of the global hypergraph network and the output features of the local hypergraph network. S3: Model Training The constructed drug relocation model is trained using the training set data to obtain the trained drug relocation model; S4: Interaction Prediction The test set data is input into the trained drug relocation model, which outputs a drug-disease interaction prediction matrix.

2. The drug relocation method based on multi-source information fusion and multi-level hypergraph learning according to claim 1, characterized in that, In step S2, the local hypergraph network has three sub-views: a drug-disease local view, a drug-protein local view, and a disease-protein local view. In the drug-disease local view of the initial layer of the drug relocation model, the drug similarity matrix, the disease similarity matrix, and the drug-disease correlation matrix are constructed into a similarity graph network, a correlation graph network, and a comprehensive graph network, and feature extraction is performed through graph neural networks respectively. In the drug-protein local view of the initial layer of the drug relocation model, the drug-drug similarity matrix, drug-protein interaction matrix and protein-protein interaction matrix are constructed into a drug-protein graph network, and feature extraction is performed through graph neural network; In the initial layer of the drug relocation model, the disease-protein local view is constructed by combining the disease similarity matrix, the disease-protein interaction matrix, and the protein-protein interaction matrix into a disease-protein graph network, and feature extraction is performed using a graph neural network. The global hypergraph network only has a global view of the drug-disease-protein relationship. In the global view of the drug-disease-protein relationship, all the above matrices are constructed into a global hypergraph network, and feature extraction is performed using a graph neural network. Among them, the drug-disease correlation matrix is ​​the drug-disease interaction matrix.

3. The drug relocation method based on multi-source information fusion and multi-level hypergraph learning according to claim 2, characterized in that, In the drug-disease local view, the construction methods for the drug similarity matrix, disease similarity matrix, and drug-disease interrelation matrix as similarity graph network, association graph network, and comprehensive graph network are as follows: in, This represents the drug similarity matrix, where m is the number of drug types. Represents the disease similarity matrix, where n is the number of disease types; A rd ∈R mxn A rd This is a drug-disease correlation matrix. This is the transposed form of the drug-disease correlation matrix; and These represent the similarity graph network, association graph network, and comprehensive graph network constructed in the drug-disease local view, respectively. The similarity graph network uses the graph isomorphic neural network GIN for feature extraction, as shown in the following expression: in, The input is the feature matrix of the initial similarity graph network; This indicates the extraction of the feature matrix obtained by GIN; MLP represents multilayer perceptron; ∈ represents learnable parameters; This indicates that neighboring nodes are being aggregated; The graph network uses a graph convolutional neural network (GCN) for feature extraction, as shown in the following expression: in, This represents the feature matrix of the initial correlation graph network input; denoted as the feature matrix extracted by GCN; A represents the adjacency matrix of the graph; W1 represents the learnable weight matrix. The comprehensive graph network uses an improved graph attention neural network, GATv2, for feature extraction, as shown in the following expression: in, The input is the feature matrix of the initial synthetic graph network; is the feature matrix extracted by GATv2; A represents the adjacency matrix of the graph; W2 represents the learnable weight matrix.

4. The drug relocation method based on multi-source information fusion and multi-level hypergraph learning according to claim 2, characterized in that, In the drug-protein local view, the construction method for the drug-protein graph network, which combines the disease similarity matrix, disease-protein interaction matrix, and protein-protein interaction matrix, is as follows: in, This represents the drug similarity matrix, where m is the number of drug types; A rp ∈R mxl This is a drug-protein interaction matrix, where l represents the number of proteins; This is the transposed form of the drug-protein interrelation matrix; A pp ∈R lxl This is the protein-protein interaction matrix; The drug-protein graph network (BGNN) is used for feature extraction using a bilinear graph neural network, as shown in the following expression: in, This represents the initial drug-protein local view feature matrix input. W3 represents the feature matrix extracted by BGNN, where A represents the adjacency matrix of the graph and W3 represents the learnable weight matrix.

5. The drug relocation method based on multi-source information fusion and multi-level hypergraph learning according to claim 2, characterized in that, In the disease-protein local view, the disease similarity matrix, disease-protein interaction matrix, and protein-protein interaction matrix are constructed into a disease-protein graph network as follows: in, Represents the disease similarity matrix, where n is the number of disease types; A dp ∈R nxl This is a disease-protein interaction matrix, where l represents the number of proteins; This is the transposed form of the disease-protein interrelation matrix; A pp ∈R lxl This is the protein-protein interaction matrix; The disease-protein graph network (BGNN) is used for feature extraction using a bilinear graph neural network, as shown in the following expression: in, This represents the initial disease-protein local view feature matrix input. denoted as the feature matrix extracted by BGNN; A represents the adjacency matrix of the graph; W4 represents the learnable weight matrix.

6. The drug relocation method based on multi-source information fusion and multi-level hypergraph learning according to claim 2, characterized in that, In the global view of drugs, diseases, and proteins, the construction method for building a global hypergraph network of all matrices is as follows: in, This represents the drug similarity matrix, where m is the number of drug types. A represents a similarity matrix for diseases, where n is the number of disease types; pp ∈R lxl A is the protein-protein interaction matrix; rd ∈R mxn A rd A is a drug-disease correlation matrix; rp ∈R mxl A is the drug-protein interaction matrix; dp ∈R nxl Disease-protein interaction matrix; A learnable graph convolutional attention neural network (L-CAT) is used to extract features from the global hypergraph network, resulting in the following expression for the output features of the global hypergraph network: in, This represents the initial global hypergraph network feature matrix as input; denoted as the feature matrix extracted by L-CAT; A represents the adjacency matrix of the graph; W5 represents the learnable weight matrix.

7. The drug relocation method based on multi-source information fusion and multi-level hypergraph learning according to claim 6, characterized in that, In the learnable graph convolutional attention neural network L-CAT, two learnable parameters are introduced before and after the attention coefficient calculation, as shown in the expression: in, This indicates that the convolution operation updates the node features. Let i represent the set of neighboring nodes. Represent the characteristics of neighboring nodes; W represents the attention coefficient between node i and node j, and W6 represents the learnable weight matrix.

8. The drug relocation method based on multi-source information fusion and multi-level hypergraph learning according to claim 6, characterized in that, In step S2, an average pooling strategy is used to fuse multiple local features obtained from the local hypergraph network, resulting in the following expression for the output features of the local hypergraph network: in, and A pp_1 This represents the drug-drug related features, disease-disease related features, and protein-protein related features obtained after local view feature extraction and average pooling operations; A rd_1 A rp_1 and A dp_1 These represent the extracted drug-disease interaction relationships, drug-protein interaction relationships, and disease-protein interaction relationships in the drug-disease local view, drug-protein local view, and disease-protein view, respectively.

9. The drug relocation method based on multi-source information fusion and multi-level hypergraph learning according to claim 8, characterized in that, In step S2, the expression for hypergraph hierarchical feature fusion is as follows: in, This represents the local hypergraph network output features extracted by the (l+1)th layer of the model; This represents the global hypergraph network output features extracted by the (l+1)th layer of the model; λ represents the original input of the (l+1)th layer of the network, i.e., the layer feature output by the lth layer; 12 ,λ 22 ,λ 32 ∈(0,1) are learnable weights, λ 12 +λ 22 +λ 32 =1.

10. The drug relocation method based on multi-source information fusion and multi-level hypergraph learning according to claim 1, characterized in that, In step S3, the loss function during model training is the polynomial loss function PolyLoss, with the following expression: Among them, y, These represent the model's actual value and predicted value, respectively. is the weight parameter for each polynomial term; p represents the highest order of the polynomial.

Images