A drug repositioning method and system based on multi-task learning and deep cross-domain

Abstract

The present application relates to the technical field of computational biology, and discloses a drug repositioning method and system based on multi-task learning and deep cross-domain, which comprises the following steps: S1. Collecting data with a "target node-drug node-disease node" ternary relationship; S2. Inputting target domain and auxiliary domain data into a feature extraction network for feature extraction; S3. Building a two-layer graph attention neural network for the target domain and the auxiliary domain respectively; S4. Fusing the enhanced deep target domain feature vector and the auxiliary domain feature vector; S5. Setting a loss function of the graph attention neural network of the target domain and the auxiliary domain, and performing multi-task learning on the graph attention neural network; and S6. Outputting a final predicted drug-disease association matrix to complete drug repositioning. The present application solves the problem that the prior art does not unify the prediction of drug-target interaction and the prediction of drug-disease association, and has the characteristics of accuracy and strong robustness.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of computational biology, and more particularly, to a drug repositioning method and system based on multi-task learning and deep cross-domain. BACKGROUND

[0002] Drug repositioning refers to the realization of "old drugs for new use" through the prediction of drug-target interaction and drug-disease association [1] Most of the current researches can be classified into "target-centric" and "disease-centric" methods. In the traditional pharmaceutical method, according to the statistics of the East Research Group (ERG), it usually takes 10-15 years and 8-1.5 billion dollars to develop a candidate drug, with an average success rate of less than 10% [2] . The method of drug repositioning can greatly improve the efficiency and safety, because the safety and pharmacokinetics of the drugs approved by the drug regulatory department are significantly reduced [3] . Therefore, the method of drug repositioning can significantly speed up the research process of drugs and save costs, which has important application value and research significance.

[0003] There is a "target-centric" drug repositioning method in the prior art. In the "target-centric" drug repositioning method, Luo et al. learn low-dimensional vector representations of drugs and targets from heterogeneous networks by using an unsupervised method, and then predict new drug-target association pairs by using a matrix completion method [4] . Liu et al. proposed a graph autoencoder method GADTI for drug-target interaction prediction using heterogeneous networks, which combines random walk, graph convolutional neural network and matrix decomposition [5] .

[0004] There is a "disease-centric" drug repositioning method in the prior art. In the "disease-centric" drug repositioning method, Jarada et al. improved the accuracy of drug-disease association prediction by integrating similarity measurement, similarity selection, similarity network fusion (SNF) and collective variational autoencoder (CAVE) for nonlinear analysis, which is called SNF-CAVE [6] . In addition, Xuan et al. proposed a drug repositioning method based on CNN and bidirectional LSTM, which uses a CNN module to learn the original representation of drugs and diseases from a drug-disease association matrix for predicting drug-disease association [7] .

[0005] However, the above models do not consider multi-task learning; that is, they do not unify the prediction of drug-target interactions and the prediction of drug-disease associations, but instead predict drug-target interactions or drug-disease associations separately. On the one hand, drugs interact with targets in cells to regulate target activity, thereby altering biological pathways to promote functional recovery and treat diseases. On the other hand, drug-disease prediction, drug-target prediction, and disease-gene prediction are three closely related tasks, and studying only one task is insufficient to comprehensively reveal the potential mechanisms of drug action. Furthermore, for the emergence of new drugs and diseases, studying only a single task is unlikely to achieve good predictive performance, as this is equivalent to the "cold start" problem in recommender systems.

[0006] In summary, given that existing technologies fail to unify the prediction of drug-target interactions and drug-disease associations, resulting in poor prediction performance, the urgent technical problem to be solved in this field is how to invent a drug relocation method that incorporates multi-task learning and graph neural networks to achieve unified prediction of drug-target interactions and drug-disease associations.

[0007] References:

[0008] [1]Jourdan JP, Bureau R, Rochais C, et al. Drug repositioning: a briefoverview[J]. Journal of Pharmacy and Pharmacology, 2020, 72(9): 1145-1151.

[0009] [2]Wong HH,Jessup A,Sertkaya A,et al.Examination of clinical trial costs and barriers for drug development final[J].Office of the AssistantSecretary for Planning and Evaluation,US Department of Health&Human Services,2014.

[0010] [3]Pushpakom S,Iorio F,Eyers PA,et al.Drug repurposing:progress ,challenges and recommendations[J].Nature reviews Drug discovery,2019,18(1):41-58.

[0011] [4]Luo Y,Zhao X,Zhou J,et al.A network integration approach fordrug—target interaction prediction and computational drug repositioning fromheterogeneous information[J].Nature communications,2017,8(1):1—13.

[0012] [5]Liu Z,Chen Q,Lan W,et al.GADTI:graph autoencoder approach for DTIprediction from heterogeneous network[J].Frontiers in Genetics,2021,12:650821.

[0013] [6]Jarada T N,Rokne J G,Alhajj R.SNF–CVAE:computational method topredict drug–disease interactions using similarity network fusion andcollective variational autoencoder[J].Knowledge-Based Systems,2021,212:106585.

[0014] [7]Xuan P,Ye Y,Zhang T,Zhao L,Sun C.Convolutional neural network andbidirectional long short—term memory—based method for predicting drug—disease associations.Cell.2019;8(7):705.

[0015] [8] Chen H, Cheng F, Li J. iDrug: Integration of drug repositioning and drug—target prediction via cross—network embedding[J]. PLoS computationalbiology, 2020, 16(7):e1008040.

[0016] [9] Zitnik, M. et al. BioSNAP Datasets: Stanford Biomedical Network Dataset Collection. http: / / snap.stanford.edu / biodata.2018. Summary of the Invention

[0017] To address the problem that existing technologies do not unify the two tasks of predicting drug-target interactions and drug-disease associations, this invention provides a drug relocation method and system that incorporates multi-task learning and graph neural networks, which is accurate and robust.

[0018] To achieve the above-mentioned objectives of this invention, the technical solution adopted is as follows:

[0019] A drug relocation method based on multi-task learning and deep cross-domain methods includes the following steps:

[0020] S1. Collect data with a ternary relationship of "target node - drug node - disease node" and construct a dataset; in the dataset, drug nodes are connected to disease nodes and target nodes respectively, forming a "drug-target" domain and a "drug-disease" domain; the "drug-disease" domain is used as the target domain and the "drug-target" domain is used as the auxiliary domain; in the target domain, construct the original drug similarity matrix and the original disease similarity matrix, and construct the drug-disease association matrix based on the known drug-disease association; in the auxiliary domain, construct the original drug similarity matrix and the original target similarity matrix, and construct the drug-target association matrix based on the known drug-target interaction; and construct the K-nearest neighbor similarity matrix of drug and disease nodes in the target domain and the K-nearest neighbor similarity matrix of drug and target nodes in the auxiliary domain.

[0021] S2. Input the original similarity matrices of the target domain and the auxiliary domain into the feature extraction network for feature extraction; specifically: reduce the dimensionality of the original drug similarity matrix and the original disease similarity matrix of the target domain through the feature extraction network to obtain the target domain feature matrix; reduce the dimensionality of the original drug similarity matrix and the original target similarity matrix of the auxiliary domain through the feature extraction network to obtain the auxiliary domain feature matrix;

[0022] S3. Based on the drug-disease association matrix and the K-nearest neighbor similarity matrix of drug and disease nodes in the target domain, the drug-target association matrix and the K-nearest neighbor similarity matrix of drug and target nodes in the auxiliary domain, construct the target domain adjacency matrix and the auxiliary domain adjacency matrix; based on the target domain adjacency matrix and the auxiliary domain adjacency matrix, build two-layer graph attention neural networks for the target domain and the auxiliary domain respectively; input the target domain feature matrix into the graph attention neural network of the target domain to extract deep-level target domain feature vectors; input the auxiliary domain feature matrix into the graph attention neural network of the auxiliary domain to extract deep-level auxiliary domain feature vectors.

[0023] S4. Perform data augmentation on the deep target domain feature vector and auxiliary domain feature vector, calculate weight parameters using a multi-head self-attention mechanism, and fuse the augmented deep target domain feature vector and auxiliary domain feature vector in a weighted summation manner to achieve deep cross-domain fusion.

[0024] S5. Set the loss function for the graph attention neural network in the target domain and auxiliary domain, perform multi-task learning on the graph attention neural network, and iteratively update the parameters of the graph attention neural network;

[0025] S6. After iterative update, the inner product decoder is used to decode the feature vector obtained after deep cross-domain fusion, and the final predicted drug-disease association matrix is ​​output to complete drug relocation.

[0026] Firstly, in step S1:

[0027] If the collected data includes the original drug similarity matrix, the original disease similarity matrix, and the original target similarity matrix, then use them directly.

[0028] If a similarity matrix is ​​not available, the calculation is performed as follows:

[0029] Based on the SMILES encoding information of the chemical structure of drugs in the dataset, the Tanimoto similarity calculation method is used to calculate and construct the original drug similarity matrix.

[0030] Based on the protein sequence information of the target in the dataset, the Smith-Waterman similarity calculation method was used to calculate and construct the original target similarity matrix;

[0031] Based on the semantic information of diseases in the dataset, the semantic similarity between MeSH terms describing diseases is calculated, and the original disease similarity matrix is ​​constructed.

[0032] Furthermore, in step S1, a K-nearest neighbor similarity matrix is ​​constructed for the target domain drug and disease nodes, and a K-nearest neighbor similarity matrix is ​​constructed for the auxiliary domain drug and target nodes, specifically as follows:

[0033] Construct K-nearest neighbor similarity matrices for drugs in the target domain, diseases in the auxiliary domain, and drugs and targets in the auxiliary domain, respectively.

[0034] Calculate the K-nearest neighbor similarity matrix S′ ij The method is as follows:

[0035]

[0036] Where N(i) represents the K nearest neighbor set of node i, S ij Let be the similarity matrix between node i and node j, where i and j are the node indices;

[0037] In step S3, based on the drug-disease association matrix, the K-nearest neighbor similarity matrix of the target domain drug and disease nodes, the drug-target association matrix, and the K-nearest neighbor similarity matrix of the auxiliary domain drug and target nodes, respectively, the target domain adjacency matrix Adj1 and the auxiliary domain adjacency matrix Adj2 are constructed.

[0038] Among them, Y D For the drug-disease association matrix, Y E This is a drug-target correlation matrix. Where m1 is the number of drug nodes in the target domain; n1 is the number of disease nodes; m2 is the number of drug nodes in the auxiliary domain; and n2 is the number of target nodes.

[0039] Furthermore, in step S2, the feature extraction network consists of two fully connected neural networks.

[0040] The first layer of the network has an input dimension equal to the dimension of the original features of the nodes, i.e., the number of columns in the similarity matrix, and an output dimension of 512, using the ReLU activation function for non-linear activation. The second layer of the network has an input dimension of 512 and an output dimension of 128. After feature extraction, the dimensionality-reduced feature matrices for the target domain drug, disease, auxiliary domain drug, and target point are obtained, respectively. (X) D ) MLP 、(X d )MLP As the feature matrix of the target domain; (X) E ) MLP 、(X e ) MLP As an auxiliary domain feature matrix.

[0041] Furthermore, step S3 specifically includes:

[0042] Let G1 be the two-layer graph attention neural network for the target domain; let G2 be the two-layer graph attention neural network for the auxiliary domain, and set the output feature dimension F. out And set the number N of multi-head attention. GAT The ELU activation function is used for non-linear activation; the graph network uses the adjacency matrices of the target domain and auxiliary domain as the graph storage, and uses the auxiliary domain feature matrix and the target domain feature matrix as the node feature input of the graph neural network; the graph neural network updates the node features to obtain deeper feature matrices for the target domain drugs and diseases, respectively. And deeper feature matrices of the auxiliary domain drug and target were obtained respectively.

[0043] Furthermore, the graph neural network updates the node features, specifically through the following steps:

[0044] S301. Calculate the attention coefficient α between node i and node j using a self-attention mechanism. ij :

[0045]

[0046] Where LeakyReLU is the activation function, || represents the concatenation operation, W is the weight matrix, a represents the parameterized deeper feature vector, and N(i) represents the neighboring nodes of node i.

[0047] S302. Based on the attention coefficient, the update formula for node features is obtained:

[0048]

[0049] in, This is for outputting deeper feature vectors;

[0050] S303. Uses a multi-head attention mechanism, using N GAT Each independent attention mechanism transforms the update formula for node features, thereby concatenating the node features to obtain a deeper feature vector in the final output:

[0051]

[0052] in, W represents the nth normalized attention coefficient calculated through the attention mechanism. n This represents the linear transformation matrix of the corresponding input.

[0053] Furthermore, in step S4, the enhanced deep target domain feature vector and auxiliary domain feature vector are fused through self-attention calculation to achieve deep cross-domain fusion, specifically as follows:

[0054] The target domain feature vector and the auxiliary domain feature vector are each mapped to the same distribution through a linear transformation.

[0055] For drug nodes that exist only in the target domain, their own feature vectors are retained and no feature fusion is performed. For drug nodes that exist in both the target domain and the auxiliary domain, the feature vectors of the drug nodes in the target domain are fused with the feature vectors of the drug nodes in the auxiliary domain by weighted summation to obtain the features after deep cross-domain fusion. The weights of the weighted summation are calculated by multi-head self-attention.

[0056] Furthermore, multi-head self-attention uses scaled dot product attention, specifically:

[0057] use Let $\mathbf{i}$ and $\mathbf{j}$ represent the feature vectors generated by the graph attention neural network after passing through the $i$-th drug in the target domain and the $j$-th drug in the auxiliary domain, respectively, where drug $j$ is the same as drug $i$. The feature vectors are then divided into $\mathbf{i}$ groups based on the number of multi-head attention nodes $N$. The form;

[0058] A1. The scaling dot product attention is calculated as follows:

[0059] Attention to target domain drugs and auxiliary domain drugs

[0060] Target domain drugs and their own attention

[0061] Where d is a scaling factor, d = F out *N GAT / N;

[0062] A2. The calculated attention coefficient of the i-th drug node in the target domain to itself. Attention coefficient for auxiliary domain drug nodes

[0063] The target domain drugs are weighted and combined to obtain the drug feature vector.

[0064] in,

[0065] A3. Update the target domain drug feature vector to X. D′ The features obtained after deep cross-domain fusion are obtained.

[0066] Furthermore, in step S5, the loss function of the graph attention neural network in the target domain and auxiliary domain is set, and multi-task learning is performed on the graph attention neural network. The specific steps are as follows:

[0067] S501. The loss function of the graph attention neural network with target and auxiliary domains simultaneously sets the loss for both the target and auxiliary domains, thereby enabling multi-task learning:

[0068] Set the loss function for the target domain as follows:

[0069]

[0070] Set the loss function for the auxiliary domain as follows:

[0071]

[0072] Where Tr(·) is the trace operation of the matrix, ||·|| F Let F be the norm, ⊙ be the element-wise multiplication of matrices, α be the regularization coefficient for network smoothing, and β be the regularization coefficient for eigenvectors; D S Let S be the degree matrix of the similarity matrix S;

[0073] S502. Target Domain Weight Matrix Auxiliary domain weight matrix The weights of negative samples are set to w, and the weights of positive samples are set to 1; the target domain weight matrix is... Auxiliary domain weight matrix The weights of negative samples are set to w, and the weights of positive samples are set to 1. The PageRank algorithm is used to measure the centrality of drug nodes, resulting in the importance coefficient matrix W′ of drug nodes, which is then normalized. The importance coefficient matrix of drug nodes in the target domain is also shown. Importance coefficient matrix of auxiliary domain drug nodes Assign different weights to different drug nodes, and transform the weight matrix into:

[0074] W1 = W1 * W′1, W2 = W2 * W′2

[0075] S503. Set small weights for the auxiliary domain loss in the loss function; simultaneously set the feature vectors of the same drugs in the target domain and auxiliary domain to be as similar as possible; the final loss function is set as follows:

[0076] min loss DD +τ·loss DE+γ·exp(-(cosine_similarity((X D′ ) com ,(X E ) com ))).

[0077] A drug relocation system based on multi-task learning and deep cross-domain includes a data input layer, a feature extraction layer, a graph attention neural network layer, a data augmentation and feature fusion layer, and a prediction layer;

[0078] The data input layer is used to collect data with a ternary relationship of "target node - drug node - disease node" and form a dataset. In the dataset, drug nodes are connected to disease nodes and target nodes respectively, forming a "drug-target" domain and a "drug-disease" domain. The "drug-disease" domain is used as the target domain, and the "drug-target" domain is used as the auxiliary domain. In the target domain, the original drug similarity matrix and the original disease similarity matrix are constructed, and a drug-disease association matrix is ​​constructed based on the known drug-disease associations. In the auxiliary domain, the original drug similarity matrix and the original target similarity matrix are constructed, and a drug-target association matrix is ​​constructed based on the known drug-target interactions. The K-nearest neighbor similarity matrix of drug and disease nodes in the target domain and the K-nearest neighbor similarity matrix of drug and target nodes in the auxiliary domain are also constructed.

[0079] The feature extraction layer is used to input the original similarity matrices of the target domain and the auxiliary domain into the feature extraction network for feature extraction; specifically, the feature extraction network reduces the dimensionality of the original drug similarity matrix and the original disease similarity matrix of the target domain to obtain the target domain feature matrix; the feature extraction network reduces the dimensionality of the original drug similarity matrix and the original target similarity matrix of the auxiliary domain to obtain the auxiliary domain feature matrix.

[0080] The graph attention neural network layer is used to construct the target domain adjacency matrix and the auxiliary domain adjacency matrix based on the drug-disease association matrix and the K-nearest neighbor similarity matrix of the target domain drug and disease nodes, the drug-target association matrix and the K-nearest neighbor similarity matrix of the auxiliary domain drug and target nodes; based on the target domain adjacency matrix and the auxiliary domain adjacency matrix, two-layer graph attention neural networks are built for the target domain and the auxiliary domain respectively; the target domain feature matrix is ​​input into the graph attention neural network of the target domain to extract deep-level target domain feature vectors; the auxiliary domain feature matrix is ​​input into the graph attention neural network of the auxiliary domain to extract deep-level auxiliary domain feature vectors; the loss function of the graph attention neural networks of the target domain and the auxiliary domain is set, and multi-task learning is performed on the graph attention neural networks to iteratively update the parameters of the graph attention neural networks;

[0081] The data augmentation and feature fusion layer is used to augment the deep target domain feature vector and auxiliary domain feature vector. It uses a multi-head self-attention mechanism to calculate weight parameters and fuses the augmented deep target domain feature vector and auxiliary domain feature vector in a weighted summation manner to achieve deep cross-domain fusion.

[0082] The prediction layer is used to decode the obtained deep cross-domain fusion feature vector using an inner product decoder, and outputs the final predicted drug-disease association matrix to complete drug relocation.

[0083] The beneficial effects of this invention are as follows:

[0084] This invention discloses a drug relocation method based on multi-task learning and deep cross-domain approaches. By unifying drug-target interaction prediction and drug-disease association prediction, it utilizes multi-task learning to achieve deep cross-domain integration, enabling the transfer of biological information between the two domains and thus improving model performance. Therefore, this invention addresses the problem in existing technologies that fail to unify the two tasks of drug-target interaction prediction and drug-disease association prediction, providing a drug relocation method and system that incorporates multi-task learning and graph neural networks, exhibiting accuracy and robustness. Attached Figure Description

[0085] Figure 1 This is a flowchart illustrating a drug relocation method based on multi-task learning and deep cross-domain approaches according to the present invention.

[0086] Figure 2 This is a schematic diagram of the system flow of a drug relocation system based on multi-task learning and deep cross-domain communication according to the present invention.

[0087] Figure 3 This is a comparative ablation experiment of a drug relocation method based on multi-task learning and deep cross-domain methods according to the present invention. Detailed Implementation

[0088] The present invention will now be described in detail with reference to the accompanying drawings and specific embodiments.

[0089] Example 1

[0090] like Figure 1 As shown, a drug relocation method based on multi-task learning and deep cross-domain methods includes the following steps:

[0091] S1. Collect data with a ternary relationship of "target node - drug node - disease node" and construct a dataset; in the dataset, drug nodes are connected to disease nodes and target nodes respectively, forming a "drug-target" domain and a "drug-disease" domain; the "drug-disease" domain is used as the target domain and the "drug-target" domain is used as the auxiliary domain; in the target domain, construct the original drug similarity matrix and the original disease similarity matrix, and construct the drug-disease association matrix based on the known drug-disease association; in the auxiliary domain, construct the original drug similarity matrix and the original target similarity matrix, and construct the drug-target association matrix based on the known drug-target interaction; and construct the K-nearest neighbor similarity matrix of drug and disease nodes in the target domain and the K-nearest neighbor similarity matrix of drug and target nodes in the auxiliary domain.

[0092] S2. Input the original similarity matrices of the target domain and the auxiliary domain into the feature extraction network for feature extraction; specifically: reduce the dimensionality of the original drug similarity matrix and the original disease similarity matrix of the target domain through the feature extraction network to obtain the target domain feature matrix; reduce the dimensionality of the original drug similarity matrix and the original target similarity matrix of the auxiliary domain through the feature extraction network to obtain the auxiliary domain feature matrix;

[0093] S3. Based on the drug-disease association matrix and the K-nearest neighbor similarity matrix of drug and disease nodes in the target domain, the drug-target association matrix and the K-nearest neighbor similarity matrix of drug and target nodes in the auxiliary domain, construct the target domain adjacency matrix and the auxiliary domain adjacency matrix; based on the target domain adjacency matrix and the auxiliary domain adjacency matrix, build two-layer graph attention neural networks for the target domain and the auxiliary domain respectively; input the target domain feature matrix into the graph attention neural network of the target domain to extract deep-level target domain feature vectors; input the auxiliary domain feature matrix into the graph attention neural network of the auxiliary domain to extract deep-level auxiliary domain feature vectors.

[0094] S4. Perform data augmentation on the deep target domain feature vector and auxiliary domain feature vector, calculate weight parameters using a multi-head self-attention mechanism, and fuse the augmented deep target domain feature vector and auxiliary domain feature vector in a weighted summation manner to achieve deep cross-domain fusion.

[0095] S5. Set the loss function for the graph attention neural network in the target domain and auxiliary domain, perform multi-task learning on the graph attention neural network, and iteratively update the parameters of the graph attention neural network;

[0096] S6. After iterative update, the inner product decoder is used to decode the feature vector obtained after deep cross-domain fusion, and the final predicted drug-disease association matrix is ​​output to complete drug relocation.

[0097] Example 2

[0098] More specifically, in step S1:

[0099] If the collected data includes the original drug similarity matrix, the original disease similarity matrix, and the original target similarity matrix, then use them directly.

[0100] If a similarity matrix is ​​not available, the calculation is performed as follows:

[0101] Based on the SMILES encoding information of the chemical structure of drugs in the dataset, the Tanimoto similarity calculation method is used to calculate and construct the original drug similarity matrix.

[0102] Based on the protein sequence information of the target in the dataset, the Smith-Waterman similarity calculation method is used to calculate and construct the original target similarity matrix;

[0103] Based on the semantic information of diseases in the dataset, the semantic similarity between MeSH terms describing diseases is calculated, and the original disease similarity matrix is ​​constructed.

[0104] In this embodiment, the constructed drug-disease association matrix is ​​Y. D The drug-target correlation matrix is ​​Y E The similarity matrix of drug nodes in the target domain is as follows: The disease similarity matrix is The auxiliary domain drug node similarity matrix is The target similarity matrix is A drug-to-drug matrix SMat was also constructed to record the correspondence between drug nodes in the target domain and the auxiliary domain. If the drugs in the two domains completely overlap, then... SMat is the identity matrix.

[0105] In one specific embodiment, step S1 involves constructing a K-nearest neighbor similarity matrix for the target domain drug and disease nodes, and a K-nearest neighbor similarity matrix for the auxiliary domain drug and target nodes, specifically as follows:

[0106] Construct K-nearest neighbor similarity matrices for drugs in the target domain, diseases in the auxiliary domain, and drugs and targets in the auxiliary domain, respectively.

[0107] Calculate the K-nearest neighbor similarity matrix S′ ij The method is as follows:

[0108]

[0109] Where N(i) represents the K nearest neighbor set of node i, S ij Let be the similarity matrix between node i and node j, where i and j are the node indices;

[0110] In step S3, based on the drug-disease association matrix, the K-nearest neighbor similarity matrix of the target domain drug and disease nodes, the drug-target association matrix, and the K-nearest neighbor similarity matrix of the auxiliary domain drug and target nodes, respectively, the target domain adjacency matrix Adj1 and the auxiliary domain adjacency matrix Adj2 are constructed.

[0111]

[0112] Among them, Y D For the drug-disease association matrix, Y E This is a drug-target correlation matrix. Where m1 is the number of drug nodes in the target domain; n1 is the number of disease nodes; m2 is the number of drug nodes in the auxiliary domain; and n2 is the number of target nodes.

[0113] In one specific embodiment, in step S2, the feature extraction network consists of two fully connected neural networks;

[0114] In this embodiment, to preserve node features to the greatest extent possible, the original similarity matrix is ​​used as the original features of the nodes. Since the original feature input dimension is too high, four two-layer fully connected neural networks are used to reduce the dimensionality of the original feature matrix. The input dimension of the first layer is the dimension of the original node features, i.e., the number of columns in the similarity matrix, and the output dimension is 512, using the ReLU activation function for non-linear activation. The input dimension of the second layer is 512, and the output dimension is 128. After feature extraction, the dimensionality-reduced feature matrices for the target domain drug, disease, auxiliary domain drug, and target point are obtained respectively. (X) D ) MLP 、(X d ) MLP As the feature matrix of the target domain; (X) E ) MLP 、(X e ) MLP As an auxiliary domain feature matrix.

[0115] In one specific embodiment, step S3 specifically includes:

[0116] Since graph-structured data often contains noise and the importance of neighbors varies, this invention addresses this issue by introducing a graph attention neural network to further extract node features. Let G1 be the two-layer graph attention neural network for the target domain; let G2 be the two-layer graph attention neural network for the auxiliary domain; and set the output feature dimension F. out And set the number N of multi-head attention.GAT The ELU activation function is used for non-linear activation; the graph network uses the adjacency matrices of the target domain and auxiliary domain as the graph storage, and uses the auxiliary domain feature matrix and the target domain feature matrix as the node feature input of the graph neural network; the graph neural network updates the node features to obtain deeper feature matrices for the target domain drugs and diseases, respectively. We obtained deeper feature matrices for the auxiliary domain drug and the target, respectively.

[0117] In one specific embodiment, the graph neural network updates node features, and the specific steps are as follows:

[0118] S301. Calculate the attention coefficient α between node i and node j using a self-attention mechanism. ij :

[0119]

[0120] Where LeakyReLU is the activation function, || represents the concatenation operation, W is the weight matrix, a represents the parameterized deeper feature vector, and N(i) represents the neighboring nodes of node i.

[0121] S302. Based on the attention coefficient, the update formula for node features is obtained:

[0122]

[0123] in, This is for outputting deeper feature vectors;

[0124] S303. Furthermore, graph attention neural networks utilize a multi-head attention mechanism to further enhance the model's learning ability, that is, using N... GAT Each independent attention mechanism transforms the update formula for node features, thereby concatenating the node features to obtain a deeper feature vector in the final output:

[0125]

[0126] in, W represents the nth normalized attention coefficient calculated through the attention mechanism. n This represents the linear transformation matrix of the corresponding input.

[0127] In one specific embodiment, in step S4, the enhanced deep target domain feature vector and auxiliary domain feature vector are fused through self-attention calculation to achieve deep cross-domain fusion, specifically as follows:

[0128] The target domain feature vector and the auxiliary domain feature vector are each mapped to the same distribution through a linear transformation.

[0129] For drug nodes that exist only in the target domain, their own feature vectors are retained and no feature fusion is performed. For drug nodes that exist in both the target domain and the auxiliary domain, the feature vectors of the drug nodes in the target domain are fused with the feature vectors of the drug nodes in the auxiliary domain by weighted summation to obtain the features after deep cross-domain fusion. The weights of the weighted summation are calculated by multi-head self-attention.

[0130] In this embodiment, whether two domain drug nodes coexist is determined by the corresponding matrix SMat.

[0131] In one specific embodiment, multi-head self-attention uses scaled dot product attention, specifically:

[0132] use Let $\mathbf{i}$ and $\mathbf{j}$ represent the feature vectors generated by the graph attention neural network after passing through the $i$-th drug in the target domain and the $j$-th drug in the auxiliary domain, respectively, where drug $j$ is the same as drug $i$. The feature vectors are then divided into $\mathbf{i}$ groups based on the number of multi-head attention nodes $N$. The form;

[0133] A1. The scaling dot product attention is calculated as follows:

[0134] Attention to target domain drugs and auxiliary domain drugs

[0135] Target domain drugs and their own attention

[0136] Where d is a scaling factor, d = F out *N GAT / N;

[0137] A2. The calculated attention coefficient of the i-th drug node in the target domain to itself. Attention coefficient for auxiliary domain drug nodes

[0138] The target domain drugs are weighted and combined to obtain the drug feature vector.

[0139] in,

[0140] A3. Update the target domain drug feature vector to X. D′ The features obtained after deep cross-domain fusion are obtained.

[0141] In one specific embodiment, step S5 involves setting the loss function for the graph attention neural network in the target domain and the auxiliary domain, and performing multi-task learning on the graph attention neural network. The specific steps are as follows:

[0142] S501. The loss function of the graph attention neural network with target and auxiliary domains simultaneously sets the loss for both the target and auxiliary domains, thereby enabling multi-task learning:

[0143] Set the loss function for the target domain as follows:

[0144]

[0145] Set the loss function for the auxiliary domain as follows:

[0146]

[0147] Where Tr(·) is the trace operation of the matrix, ||·|| F Let F be the norm, ⊙ be the element-wise multiplication of matrices, α be the regularization coefficient for network smoothing, and β be the regularization coefficient for eigenvectors; D S Let S be the degree matrix of the similarity matrix S;

[0148] The first term of the loss function is the loss between the predicted reconstruction matrix and the correlation matrix; the second term is the network smoothing term, which makes the feature vectors of similar nodes more similar; and the third term is the regularization term, which prevents the model from overfitting.

[0149] S502. Considering that positive samples have been proven to have a definite correlation, they are considered more important. Furthermore, setting a weight matrix W can avoid poor model performance due to an imbalance between positive and negative samples. Wherein, the target domain weight matrix... Auxiliary domain weight matrix The weight of negative samples is set to w, and the weight of positive samples is set to 1. Furthermore, considering the cross-domain information integration through drug nodes, more important drug nodes have a greater impact on the entire network. Therefore, the PageRank algorithm is used to measure the centrality of drug nodes, obtaining the importance coefficient matrix W′ of drug nodes, which is then normalized; the importance coefficient matrix of drug nodes in the target domain... Importance coefficient matrix of auxiliary domain drug nodes Assign different weights to different drug nodes, and transform the weight matrix into:

[0150] W1 = W1 * W′1, W2 = W2 * W′2

[0151] S503. To optimize the target domain better, the auxiliary domain loss is given a small weight in the loss function. Simultaneously, to make cross-domain information fusion more effective, the feature vectors of the same drug in both domains need to be as similar as possible. Therefore, the final loss function is set as follows:

[0152] minloss DD +τ·loss DE +γ·exp(-(cosine_similarity((X D′ ) com ,(X E ) com ))).

[0153] In this embodiment, during step S5, when iteratively updating the parameters of the graph attention neural network, the present invention uses the Adam optimizer to update the parameters and sets weight decay. The weight decay rate is set to 0.0001, and the learning rate is set to 0.001. The learning rate is adjusted by adjusting the learning rate at equal intervals, making the training of the model more stable.

[0154] In this embodiment, in step S6, an inner product decoder is used to output the final predicted drug-disease association matrix. The specific decoding method is as follows:

[0155]

[0156] After decoding, the output prediction correlation matrix P is normalized. ij This represents the predicted probability that drug i and disease j are associated. The higher the probability, the more likely the association is to exist.

[0157] In this embodiment, the specific steps of the present invention are shown in Table 1:

[0158]

[0159] In this embodiment, data collection: The present invention conducts experiments on three "target-drug-disease" datasets, namely DTINet. [4] Dataset, idrug [8] Dataset and the processed BioSNAP [9] Datasets. Because the targets in the drug-target domain and the diseases in the drug-disease domain of the original BioSNAP dataset are not entirely correlated, to ensure the effectiveness of cross-domain fusion, only correlated targets and diseases were retained during data processing. Descriptions of the three datasets are shown in Table 2 below:

[0160]

[0161] Table 2

[0162] In this embodiment, the model is built by following the steps described above and inputting the association matrix and similarity matrix into the network model.

[0163] This embodiment uses a five-fold cross-validation method. Since cross-domain fusion information is mainly reflected in the ability to utilize information from auxiliary domains when a new node appears, this invention randomly selects certain rows from the association matrix for testing, that is, sets all elements of these rows of the association matrix to 0 to simulate the scenario of a new node.

[0164] In this embodiment, the parameter setting method for the graph attention neural network in this invention is to use a grid search method with fixed random seeds to tune the parameters and find the optimal parameters through cross-validation. The parameters that need to be grid searched are: the number of K-nearest neighbor similarity matrices K: {3, 5, 10}; the dimension of the graph attention neural network output feature F_out: {16, 32, 64}; the number of multi-head attention points N_GAT: {2, 4, 8}; the number of cross-domain fusion multi-head self-attention points N: {1, 4, 8, 16}; the loss function hyperparameters α: {0.001, 0.01, 0.1}, β: {0.001, 0.01, 0.1}, γ: {0.001, 0.01, 0.1}, τ: {0.1, 0.3, 0.5}, and the weight coefficients w of the weight matrix: {0.1, 0.2, 0.3, 0.5}. In the scenario of cross-validation at new nodes, AUC and AUPR are used as evaluation metrics, and the parameters that achieve the highest AUC and AUPR are selected as the optimal parameters for the model on this dataset.

[0165] In this embodiment, the drug-target interaction prediction problem and the drug-disease association prediction problem in drug relocation are treated as binary classification problems. For the binary classification problem, the classification result is saved as a confusion matrix, as shown in Table 2:

[0166]

[0167] Table 2

[0168] Where TP represents a positive sample being predicted as positive, FN represents a positive sample being predicted as negative, FP represents a negative sample being predicted as positive, and TN represents a negative sample being predicted as negative. Based on the confusion matrix, accuracy, precision, recall, F1-score, area under the ROC curve (AUC), and area under the PR curve (AUPR) can be calculated. The formulas for the first four evaluation metrics are as follows:

[0169]

[0170] Among the evaluation metrics above, the larger the value, the better the model's performance.

[0171] like Figure 3 The results show a comparison of the performance of models with and without a cross-domain multi-task learning module. The presented results are the comparison of models with the "drug-disease" domain as the target domain in the new node scenario across three datasets, with AUC as the evaluation metric. The new node scenario involves randomly selecting rows from the association matrix for testing, setting all elements in the selected rows to 0. CDRGAT represents the model with the cross-domain multi-task learning module, and CDRGAT- represents the model without it.

[0172] In this embodiment, an ablation experiment was conducted. The multi-task learning cross-domain module of this invention was removed before comparison, and AUC was selected as the evaluation metric. The experimental results are as follows: Figure 3 As shown, this demonstrates that the cross-domain model designed in this invention can significantly improve the model's performance for scenarios with new nodes.

[0173] This invention discloses a drug relocation method based on multi-task learning and deep cross-domain approaches. By unifying drug-target interaction prediction and drug-disease association prediction, it utilizes multi-task learning to achieve deep cross-domain integration, enabling the transfer of biological information between the two domains and thus improving model performance. Therefore, this invention addresses the problem in existing technologies that fail to unify the two tasks of drug-target interaction prediction and drug-disease association prediction, providing a drug relocation method and system that incorporates multi-task learning and graph neural networks, exhibiting accuracy and robustness.

[0174] Example 3

[0175] like Figure 2 As shown, a drug relocation system based on multi-task learning and deep cross-domain learning includes a data input layer, a feature extraction layer, a graph attention neural network layer, a data augmentation and feature fusion layer, and a prediction layer.

[0176] The data input layer is used to collect data with a ternary relationship of "target node - drug node - disease node" and form a dataset. In the dataset, drug nodes are connected to disease nodes and target nodes respectively, forming a "drug-target" domain and a "drug-disease" domain. The "drug-disease" domain is used as the target domain, and the "drug-target" domain is used as the auxiliary domain. In the target domain, the original drug similarity matrix and the original disease similarity matrix are constructed, and a drug-disease association matrix is ​​constructed based on the known drug-disease associations. In the auxiliary domain, the original drug similarity matrix and the original target similarity matrix are constructed, and a drug-target association matrix is ​​constructed based on the known drug-target interactions. The K-nearest neighbor similarity matrix of drug and disease nodes in the target domain and the K-nearest neighbor similarity matrix of drug and target nodes in the auxiliary domain are also constructed.

[0177] The feature extraction layer is used to input the original similarity matrices of the target domain and the auxiliary domain into the feature extraction network for feature extraction; specifically, the feature extraction network reduces the dimensionality of the original drug similarity matrix and the original disease similarity matrix of the target domain to obtain the target domain feature matrix; the feature extraction network reduces the dimensionality of the original drug similarity matrix and the original target similarity matrix of the auxiliary domain to obtain the auxiliary domain feature matrix.

[0178] The graph attention neural network layer is used to construct the target domain adjacency matrix and the auxiliary domain adjacency matrix based on the drug-disease association matrix and the K-nearest neighbor similarity matrix of the target domain drug and disease nodes, the drug-target association matrix and the K-nearest neighbor similarity matrix of the auxiliary domain drug and target nodes; based on the target domain adjacency matrix and the auxiliary domain adjacency matrix, two-layer graph attention neural networks are built for the target domain and the auxiliary domain respectively; the target domain feature matrix is ​​input into the graph attention neural network of the target domain to extract deep-level target domain feature vectors; the auxiliary domain feature matrix is ​​input into the graph attention neural network of the auxiliary domain to extract deep-level auxiliary domain feature vectors; the loss function of the graph attention neural networks of the target domain and the auxiliary domain is set, and multi-task learning is performed on the graph attention neural networks to iteratively update the parameters of the graph attention neural networks;

[0179] The data augmentation and feature fusion layer is used to augment the deep target domain feature vector and auxiliary domain feature vector. It uses a multi-head self-attention mechanism to calculate weight parameters and fuses the augmented deep target domain feature vector and auxiliary domain feature vector in a weighted summation manner to achieve deep cross-domain fusion.

[0180] The prediction layer is used to decode the obtained deep cross-domain fusion feature vector using an inner product decoder, and outputs the final predicted drug-disease association matrix to complete drug relocation.

[0181] Obviously, the above embodiments of the present invention are merely examples for clearly illustrating the present invention, and are not intended to limit the implementation of the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the protection scope of the claims of the present invention.

Claims

1. A drug repositioning method based on multi-task learning and deep cross-domain, characterized in that: The method comprises the following steps: S1. Collecting data with "target point node-drug node-disease node" ternary relationship and forming a data set; in the data set, the drug node is connected with the disease node and the target point node respectively, forming a "drug-target" domain and a "drug-disease" domain; taking the "drug-disease" domain as the target domain and the "drug-target" domain as the auxiliary domain; in the target domain, an original drug similarity matrix, an original disease similarity matrix, and a drug-disease association matrix are constructed according to the known association of drugs and diseases; in the auxiliary domain, an original drug similarity matrix, an original target point similarity matrix, and a drug-target association matrix are constructed according to the known interaction of drugs and target points; and a K-nearest neighbor similarity matrix of the drug and disease nodes in the target domain and a K-nearest neighbor similarity matrix of the drug and target point nodes in the auxiliary domain are constructed; S2. Inputting the original similarity matrices of the target domain and the auxiliary domain into a feature extraction network for feature extraction; Specifically: dimension reduction is performed on the original drug similarity matrix and the original disease similarity matrix of the target domain by the feature extraction network to obtain a target domain feature matrix; dimension reduction is performed on the original drug similarity matrix and the original target point similarity matrix of the auxiliary domain by the feature extraction network to obtain an auxiliary domain feature matrix; S3. Constructing a target domain adjacency matrix and an auxiliary domain adjacency matrix according to the drug-disease association matrix, the K-nearest neighbor similarity matrix of the drug and disease nodes in the target domain, the drug-target association matrix, and the K-nearest neighbor similarity matrix of the drug and target point nodes in the auxiliary domain; According to the target domain adjacency matrix and the auxiliary domain adjacency matrix, a two-layer graph attention neural network is built for the target domain and the auxiliary domain respectively; The target domain feature matrix is input into the graph attention neural network of the target domain to extract a deep target domain feature vector; the auxiliary domain feature matrix is input into the graph attention neural network of the auxiliary domain to extract a deep auxiliary domain feature vector; S4. Data augmentation is performed on the deep target domain feature vector and the deep auxiliary domain feature vector, weight parameters are calculated using a multi-head self-attention mechanism, and the augmented deep target domain feature vector and the deep auxiliary domain feature vector are fused in a weighted summation manner to realize deep cross-domain fusion; S5. Setting a loss function of the graph attention neural network of the target domain and the auxiliary domain, performing multi-task learning on the graph attention neural network, and iteratively updating the parameters of the graph attention neural network; S6. After iteration and update, the obtained deep cross-domain fused feature vector is decoded using an inner product decoder to output a final predicted drug-disease association matrix, and drug repositioning is completed.

2. The method of claim 1, wherein the method is based on multi-task learning and deep cross-domain drug repositioning. In the step S1: If the collected data includes the original drug similarity matrix, the original disease similarity matrix, and the original target point similarity matrix, they are directly used; If the data does not contain the similarity matrix, the following method is used to calculate: According to the SMILES encoding information of the chemical structure of the drug in the data set, the original drug similarity matrix is constructed by using the Tanimoto similarity calculation method; According to the protein sequence information of the target points in the data set, a similarity matrix of the target points is constructed by using a Smith-Waterman similarity calculation method; According to the semantic information of the diseases in the data set, the semantic similarity between the MeSH terms describing the diseases is calculated to construct an original disease similarity matrix.

3. The method of claim 1, wherein the method is based on multi-task learning and deep cross-domain drug repositioning. In the step S1, a K-neighbor similarity matrix of the target domain drug and disease nodes and a K-neighbor similarity matrix of the auxiliary domain drug and target nodes are constructed, specifically as follows: K-neighbor similarity matrices of the target domain drug, disease, auxiliary domain drug and target are constructed respectively Computing a K-nearest neighbor similarity matrix S i ′ j The method is as follows: where N(i) represents the K-neighbor set of node i, S ij is the similarity matrix of node i and node j, i and j are node numbers; In the step S3, a target domain adjacency matrix and an auxiliary domain adjacency matrix are constructed according to the drug-disease association matrix, the K-neighbor similarity matrix of the target domain drug and disease nodes and the drug-target association matrix, the K-neighbor similarity matrix of the auxiliary domain drug and target nodes, to obtain a target domain adjacency matrix Adj1 and an auxiliary domain adjacency matrix Adj2: wherein Y D is a drug-disease association matrix, Y E is a drug-target association matrix, wherein m1 is the number of target domain drug nodes; n1 is the number of disease nodes, m2 is the number of auxiliary domain drug nodes, and n2 is the number of target nodes.

4. The method of claim 3, wherein the method is based on multi-task learning and deep cross-domain for drug repositioning. In the step S2, the feature extraction network is composed of two layers of fully connected neural networks; The input dimension of the first layer network is the dimension of the original feature of the node, that is, the column number of the similarity matrix, the output dimension is 512, and a ReLU activation function is selected for nonlinear activation; the input dimension of the second layer network is 512, and the output dimension is 128; after feature extraction, the feature matrix of the target domain drug, disease, auxiliary domain drug and target after dimension reduction is obtained respectively (X D ) MLP , (X d ) MLP as the target domain feature matrix; (X E ) MLP 、(X e ) MLP as an auxiliary domain feature matrix.

5. The method of claim 4, wherein the method is based on multi-task learning and deep cross-domain for drug repositioning. In the step S3, the specific steps are as follows: Set the two-layer graph attention neural network of the target domain as G1; set the two-layer graph attention neural network of the auxiliary domain as G2, and set the output feature dimension F out , and set the number of multi-head attentions N GAT , and use the elu activation function for nonlinear activation; the graph network uses the adjacency matrix of the target domain and the auxiliary domain as the storage of the graph, and inputs the auxiliary domain feature matrix and the target domain feature matrix as the node feature of the graph neural network; the graph neural network updates the node feature to obtain the deeper-level feature matrix of the target domain drug and disease respectively, and obtains the deeper-level feature matrix of the auxiliary domain drug and target 6. The method of claim 5, wherein the method is based on multi-task learning and deep cross-domain for drug repositioning. The graph neural network updates the node features, and the specific steps are as follows: S301. Calculate the attention coefficient a between node i and node j using self-attention mechanism ij : wherein LeakyReLU is an activation function, || represents a splicing operation, W is a weight matrix, a represents a parameterized deeper feature vector, and N(i) represents the neighbor nodes of the node i; S302. According to the attention coefficient, an updating formula of the node feature is obtained: wherein, is an output deeper feature vector; S303. Using the multi-head attention mechanism, the update formula of the node feature is converted using N GAT independent attention mechanisms to connect the node features to obtain a deeper feature vector of the final output: wherein, denotes the normalized attention coefficient calculated by the attention mechanism for the nth, W n denotes the linear transformation matrix of the corresponding input.

7. The method of claim 5, wherein the method is based on multi-task learning and deep cross-domain for drug repositioning. In the step S4, the enhanced deep target domain feature vector and the auxiliary domain feature vector are fused by self-attention calculation to realize deep cross-domain fusion, specifically as follows: The target domain feature vector and the auxiliary domain feature vector are respectively mapped to the same distribution through a linear transformation; For the drug nodes existing only in the target domain, the feature vectors thereof are retained without feature fusion; for the drug nodes existing in both the target domain and the auxiliary domain, the feature vectors of the drug nodes in the target domain are fused with the feature vectors of the drug nodes in the auxiliary domain by weighted summation to obtain the feature after deep cross-domain fusion; wherein the weight of the weighted summation is calculated by multi-head self-attention.

8. The method of claim 7, wherein the method is based on multi-task learning and deep cross-domain for drug repositioning. The multi-head self-attention uses scaled dot-product attention, specifically as follows: With respectively represent the feature vectors generated by the target domain i-th drug and the auxiliary domain j-th drug after passing through the graph attention neural network, and drug j and drug i are the same drug; The feature vectors are divided into the form of A1. The calculation method of the scaled dot-product attention is as follows: Attention of target domain drugs and auxiliary domain drugs Target domain drug and its own attention where d is a scaling factor, d = F out *N GAT / N; A2. The attention coefficient of the i-th drug node in the target domain calculated The attention coefficient of the drug node in the auxiliary domain The target domain drug is combined by weighting to obtain the feature vector of the drug wherein A3. Update the target domain drug feature vector to X D′ , to obtain the fused features across domains.

9. The method of claim 8, wherein the method is based on multi-task learning and deep cross-domain drug repositioning. In the step S5, the loss function of the graph attention neural network of the target domain and the auxiliary domain is set to perform multi-task learning on the graph attention neural network, and the specific steps are as follows: S501. The loss function of the graph attention neural network of the target domain and the auxiliary domain is set to include the loss of the target domain and the auxiliary domain, so as to perform multi-task learning: The loss function of the target domain is set as follows: The loss function of the auxiliary domain is set as follows: where Tr(·) is the trace operation of a matrix, ||·||F is the F-norm, and is the element-wise multiplication of matrices, and are the regularization coefficients for the network smoothing term and the eigenvector, respectively; D is the degree matrix of the similarity matrix S. F S is the degree matrix of the similarity matrix S.​ S502. The centrality of the drug nodes is measured using the PageRank algorithm to obtain the importance coefficient matrix W' of the drug nodes and normalized; the importance coefficient matrix of the target domain drug nodes The importance coefficient matrix of the auxiliary domain drug nodes Different weights are given to different drug nodes, and the weight matrix becomes: W1=W1*W′1,W2=W2*W2 S503. A small weight is set for the auxiliary domain loss in the loss function; meanwhile, the features of the same drugs in the target domain and the auxiliary domain are set to be as similar as possible; and the final loss function is set as follows: min loss DD +τ·loss DE +γ·exp(-(cosine_similarity((X D′ ) com ,(X E ) com )))。 10.A drug repositioning system based on multi-task learning and deep cross-domain, characterized in that: The method comprises a data input layer, a feature extraction layer, a graph attention neural network layer, a data enhancement and feature fusion layer, and a prediction layer. The data input layer is used for collecting data with a "target point node-drug node-disease node" ternary relationship and forming a data set; in the data set, the drug node is connected with the disease node and the target point node respectively, forming a "drug-target point" domain and a "drug-disease" domain; the "drug-disease" is taken as a target domain, and the "drug-target point" domain is taken as an auxiliary domain; in the target domain, an original drug similarity matrix, an original disease similarity matrix, and a drug-disease association matrix are constructed according to known drug-disease associations; in the auxiliary domain, an original drug similarity matrix, an original target point similarity matrix, and a drug-target point association matrix are constructed according to known drug-target point interactions; and a K-nearest neighbor similarity matrix of the target domain drug and disease nodes and a K-nearest neighbor similarity matrix of the auxiliary domain drug and target point nodes are constructed. The feature extraction layer is used for inputting the original similarity matrices of the target domain and the auxiliary domain into a feature extraction network for feature extraction. Specifically, the original drug similarity matrix and the original disease similarity matrix of the target domain are dimensionally reduced by the feature extraction network to obtain a target domain feature matrix; the original drug similarity matrix and the original target point similarity matrix of the auxiliary domain are dimensionally reduced by the feature extraction network to obtain an auxiliary domain feature matrix. The target domain adjacency matrix and the auxiliary domain adjacency matrix are constructed according to the drug-disease association matrix, the K-nearest neighbor similarity matrix of the target domain drug and disease nodes, the drug-target point association matrix, and the K-nearest neighbor similarity matrix of the auxiliary domain drug and target point nodes. The two-layer graph attention neural networks of the target domain and the auxiliary domain are respectively built according to the target domain adjacency matrix and the auxiliary domain adjacency matrix. The target domain feature matrix is input into the graph attention neural network of the target domain to extract a deep target domain feature vector; the auxiliary domain feature matrix is input into the graph attention neural network of the auxiliary domain to extract a deep auxiliary domain feature vector; the loss function of the graph attention neural network of the target domain and the auxiliary domain is set, the graph attention neural network is subjected to multi-task learning, and the parameters of the graph attention neural network are iteratively updated. The data enhancement and feature fusion layer is used for data enhancement of the deep target domain feature vector and the deep auxiliary domain feature vector, weighted parameter calculation by using a multi-head self-attention mechanism, fusion of the enhanced deep target domain feature vector and the deep auxiliary domain feature vector in a weighted summation manner, and realization of deep cross-domain fusion. The prediction layer is used for decoding the obtained deep cross-domain fused feature vector by using an inner product decoder, outputting a final predicted drug-disease association matrix, and completing drug repositioning.

Images