A drug repositioning method and system based on implicit interaction adjacency matrix remodeling

Abstract

This invention belongs to the field of smart healthcare technology and discloses a drug relocation method and system based on implicit interaction adjacency matrix reshaping. The method includes the following steps: S1, constructing a fused drug-disease association matrix; S2, extracting drug domain features and disease domain features using a parameter-sharing dual-path self-attention mechanism; S3, mapping drug / disease features to the other's dimensional space using a trainable projection matrix to generate an implicit interaction prediction matrix; S4, dynamically adjusting the contribution of positive and negative samples using weight λ to address data sparsity. This invention achieves decoupling learning of drug domain and disease domain features through a dynamic adjacency matrix and a dual-path self-attention mechanism. It utilizes a trainable projection matrix to complete cross-dimensional implicit interaction prediction and combines iterative optimization of feature representation using a recurrent neural network. This effectively reduces the impact of noise from artificially similar data and demonstrates significant advantages in practical scenarios such as drug relocation for Parkinson's disease.

Description

Technical Field

[0001] This invention relates to the field of smart healthcare technology, specifically to a drug relocation method and system based on implicit interactive adjacency matrix reshaping. Background Technology

[0002] Traditional drug development faces challenges such as long development cycles and high costs. In recent years, drug repositioning has gained significant attention as an efficient drug development strategy. Drug repositioning, by utilizing existing drug data, can significantly shorten development cycles, reduce costs, and increase success rates. Currently, the mainstream methods for drug repositioning mainly include matrix factorization and deep learning. Matrix factorization-based methods treat drug repositioning as a recommender system problem, using matrix factorization techniques to complete the drug-disease association matrix to achieve drug repositioning. Commonly used techniques include singular value decomposition, principal component analysis, and nonnegative matrix factorization. Deep learning-based methods are typically combined with graph neural networks to complete new drug-disease association matrices, including graph convolutional networks (GCN), graph attention networks (GAT), and graph isomorphic networks (GIN). Building deep learning through graph neural networks can effectively mine the topological information of heterogeneous drug-disease networks, enabling efficient modeling of drugs and diseases for more accurate drug-disease associations. For example, LAGCN, an end-to-end model combining a drug-disease heterogeneous network and a graph convolutional neural network, integrates drug-disease association, drug similarity, and disease similarity networks into a single drug-disease heterogeneous network and utilizes a multi-layer graph convolutional network to embed drugs and diseases. REDDA, ​​based on a multi-layered attention mechanism of complex heterogeneous biological networks, constructs a network model by integrating five biological entities: drugs, proteins, genes, pathways, and diseases. It innovatively fuses node-level attention, topological subnet attention, and hierarchical attention mechanisms, sequentially employing modules such as node embedding, local topological structure learning, graph relational attention weighting, and multi-layer feature aggregation to ultimately achieve drug and disease feature representation. DRWBNCF effectively integrates neighbor interaction information to achieve drug and disease embedding through a weighted bilinear graph convolution operation. GCNGAT combines graph convolutional networks and graph attention networks, extracting specific subgraphs of drug-disease pairs to focus on local contextual information and improve prediction accuracy. MDGCN optimizes drug relocation performance by employing a multi-dependency graph convolutional network and dual contrastive learning (across views and across layers) to align embeddings of different graph structures. MPMB-DR significantly improves the accuracy of drug-disease association prediction by constructing a three-layer heterogeneous network of drugs, proteins, and diseases, combined with meta-path selection and similarity negative sampling strategies. KGRDR integrates multiple similarity matrices of drugs and diseases and introduces knowledge graph embedding to learn global topological features. It achieves high-precision prediction by fusing attribute and structural information through an attention mechanism. MGCNDDA utilizes random walks to enhance sparse data and improves model robustness by iteratively optimizing prediction results through a dual GCN model and nonlinear feature extraction.

[0003] In existing technologies, a drug relocation method using neighborhood information and a weighted fusion network disclosed in CN119920302A has limitations in explicit interaction. It predicts drug-disease associations directly through the connection between drug and disease nodes, failing to capture latent pathological associations (such as multi-target synergistic effects). The static adjacency matrix, based on predefined similarity weights, cannot dynamically correct data noise (such as Jaccard similarity errors). A drug-disease association prediction method based on a deep graph neural network disclosed in CN116343909A extracts drug and disease features independently through dual GCNs during independent feature learning, failing to achieve cross-domain information fusion. It cannot perform random walk updates, only enhancing sparse data coverage, and lacks a similarity matrix reshaping mechanism.

[0004] Therefore, although the existing technologies mentioned above can achieve good performance, they suffer from problems such as data noise and limited interaction patterns. The adjacency matrices they use consist only of single similarities between drugs or diseases, failing to accurately represent the adjacency information between drugs or diseases. This invention introduces a drug-disease association matrix to generate a new adjacency matrix together with the original similarity matrix. This reduces the incomplete information and noise caused by adjacency matrices constructed solely based on manually set similarities, thus enhancing the model's expressive power. Furthermore, existing methods focus more on the explicit interactions between drugs and diseases, neglecting implicit interaction information, which limits the accuracy of predictions and biological interpretability. Summary of the Invention

[0005] To overcome the shortcomings of existing technologies, this paper proposes a drug relocation method and system based on implicit interaction adjacency matrix reshaping. By constructing a dynamic adjacency matrix that integrates drug-disease association information, a dual-path self-attention mechanism is used to achieve decoupled learning of drug domain and disease domain features. A trainable projection matrix is ​​used to complete cross-dimensional implicit interaction prediction. Combined with recurrent neural network iterative optimization of feature representation, this method solves the key problems of existing technologies being unable to simultaneously handle data noise and implicit interactions. It effectively reduces the impact of artificial similarity data noise, significantly improves indicators such as AUC and AUPR, and shows significant advantages in practical scenarios such as Parkinson's disease drug relocation.

[0006] To achieve the above objectives, the present invention proposes the following technical solution:

[0007] A drug relocation method based on implicit interaction adjacency matrix reshaping is characterized by the following steps:

[0008] S1. Construct a fused drug-disease association matrix A, and compare it with the original similarity matrix. S r / S dThe dynamic adjacency matrix is ​​optimized in real time through a recurrent neural network (RNN) iteratively.

[0009] in, S r A matrix representing the similarity between drugs. S d A similarity matrix representing diseases;

[0010] S2. A parameter-sharing dual-path self-attention mechanism (SAM) is used to extract drug domain features. H r(r) , H d(r) and disease domain characteristics H r(d) ,H d(d) And retain the original feature information through residual connections;

[0011] S3, using a trainable projection matrix P r / P d By mapping drug / disease features to each other's dimensional space, an implicit interaction prediction matrix is ​​generated;

[0012] S4. The weighted cross-entropy loss function is adopted, and the contribution of positive and negative samples is dynamically adjusted by weight λ to solve the data sparsity problem.

[0013] A drug relocation system based on implicit interactive adjacency matrix reshaping, used to perform the method, comprising:

[0014] Dynamic matrix generation module: used to perform adjacency matrix generation and dynamic adjacency matrix update operations;

[0015] Multimodal feature fusion module: integrates heterogeneous features of chemical structure fingerprint ECFP6, target protein sequence, and disease ontology MeSH;

[0016] Prediction output module: decodes the final embeddings of drugs and diseases, calculates the drug-disease association probability, and calculates the TOP-K candidate drugs;

[0017] Data input module: used to load drug molecular fingerprints, disease phenotype ontology, and known drug-disease association data;

[0018] Dual-path feature learning module: includes parameter-shared SAM layer, MLP and RNN iterative optimization unit;

[0019] Implicit interaction prediction module: trained by RNN P r / Pd Generate a drug-disease association probability matrix from the matrix.

[0020] Compared with the prior art, the method and system of the present invention have at least the following advantages:

[0021] 1. The present invention provides a drug relocation method and system based on implicit interaction adjacency matrix reshaping (DRIIAMR). DRIIAMR improves the drug and disease adjacency matrix information by introducing drug-disease association information and adaptively training a neural network model. It captures the potential association between drug and disease embeddings through implicit interaction and adopts a dynamic adjacency matrix reshaping and implicit projection interaction mechanism through the projection matrix (…). P r / P d This allows for the realization of implicit drug-disease interactions, uncovering deep biological mechanisms and unmanifested network relationships, thereby improving the accuracy of drug-disease association prediction.

[0022] 2. The method of this invention constructs a dynamic adjacency matrix that integrates drug-disease association information, employs a dual-path self-attention mechanism to decouple the learning of features from the drug domain and the disease domain, utilizes a trainable projection matrix to complete cross-dimensional implicit interaction prediction, and combines a recurrent neural network to iteratively optimize feature representation. Compared with existing technologies, this invention effectively reduces the impact of noise from artificially similar data, resulting in a significant improvement in metrics such as AUC and AUPR, demonstrating significant advantages in practical scenarios such as drug relocation in Parkinson's disease.

[0023] 3. The dual-path feature learning mechanism employed in this invention, by explicitly defining the embedding representations of drugs and diseases in different biological entity (drug and disease) spaces, adapts to the heterogeneous interaction characteristics between drugs and diseases, thereby capturing the implicit interactions between drugs and diseases. The dual-path intra-domain / cross-domain feature learning (H... (r) / H (d) It works in tandem with residual connections and parameter sharing SAM.

[0024] 4. In the adjacency matrix construction, this invention integrates the A matrix to dynamically reshape the drug / disease adjacency matrix. S r / S d Then, the adjacency matrix is ​​reshaped. By introducing drug-disease association information and adaptively training the neural network, the adjacency matrix information of drugs and diseases is improved. The adjacency matrix is ​​dynamically updated to reduce noise, and potential associations are mined through implicit interactions.

[0025] 5. The cyclic encoding used in this invention enables more accurate embedding of drugs and diseases through multiple iterations of training, while simultaneously training two projection matrices for implicit interactions.

[0026] 6. The implicit interaction between drugs and diseases realized in this invention captures the potential association between drugs and diseases through implicit interaction, explores deep biological mechanisms and unmanifested network relationships, thereby improving the accuracy of drug-disease association prediction.

[0027] 7. The method and system of the present invention are applicable to at least one of the following scenarios: screening of drug targets for Parkinson's disease (sparse association scenario with SP≤0.3); repositioning of multiple drug effects of antibiotics (large-scale drug library scenario with n≥1000); prediction of drug interaction in combination therapy for tumors (complex pathological association scenario with m≥50), which significantly expands its application scope. Attached Figure Description

[0028] Figure 1 This is a schematic diagram of the process of drug repositioning using DRIIAMR according to an embodiment of the present invention;

[0029] Figure 2 This is a schematic diagram of the model architecture for drug relocation implemented by DRIIAMR according to an embodiment of the present invention;

[0030] Figure 3 This is a schematic diagram of a 10 CV experiment on Gdataset according to an embodiment of the present invention, where a is the ROC curve of different models and b is the PR curve of different models;

[0031] Figure 4 This is a schematic diagram of a 10 CV experiment on Cdataset according to an embodiment of the present invention, where a is the ROC curve of different models and b is the PR curve of different models;

[0032] Figure 5 This is a schematic diagram of a 10 CV experiment on Ldataset according to an embodiment of the present invention, where a is the ROC curve of different models and b is the PR curve of different models;

[0033] Figure 6 This is a schematic diagram of a 10 CV experiment on LRSSL according to an embodiment of the present invention, where a is the ROC curve of different models and b is the PR curve of different models. Detailed Implementation

[0034] To further illustrate the technical means and effects adopted by the present invention to achieve its intended purpose, the following is combined with... Figures 1 to 6 The present invention will be described in conjunction with preferred embodiments.

[0035] Basic Implementation

[0036] See Figure 1 The drug relocation method based on implicit interactive adjacency matrix reshaping provided in this embodiment of the invention includes the following steps:

[0037] S1. Construct a fused drug-disease association matrix A, and compare it with the original similarity matrix. S r / S d The dynamic adjacency matrix is ​​optimized in real time through a recurrent neural network (RNN) iteratively; including:

[0038] S1-1 constructs an initial adjacency matrix based on known drug and disease similarity networks, and initializes drug and disease features; the construction methods for the original drug / disease similarity matrix include:

[0039] Drug similarity matrix S r The disease similarity matrix was calculated using the Jaccard distance between the chemical structural fingerprint ECFP6 and the target protein sequence. S d Semantic similarity calculation based on medical subject headings (MeSH);

[0040] S1-2 constructs four types of neighbor information aggregation based on the self-attention mechanism (SAM) according to four different network relationships;

[0041] S1-3 Original Similarity Matrix S r / S d The mathematical expression for the dynamic adjacency matrix update is given by Equations 1 and 2 below:

[0042] A r (t) = RNN( A , A r(t−1) , P r Formula 1

[0043] A d (t) = RNN ( A T , A d (t−1) , P d Equation 2

[0044] In the formula, t is the current loop level, and RNN represents the total function of reshaping the adjacency matrix in a loop (see Appendix). Figure 2 The RNN process includes two stages: encoding and reshaping. A r (t) ∈ {0,1} n*n and A d (t) ∈ {0,1} m*m This represents the adjacency matrix of drugs and diseases input in each loop. A r (t) ∈ {0,1} n*n and A d (t) ∈ {0,1} m*m This represents the remodeled drug-disease adjacency matrix in each cycle. P r ∈ R k*n With P d ∈ R k*m These represent the projection matrices that project the low-dimensional features of the drug and disease onto the drug dimension and disease dimension, respectively; where n, m Indicates the quantity of drugs and diseases ，k It is the feature embedding dimension;

[0045] S2. A parameter-sharing dual-path self-attention mechanism (SAM) is used to extract drug domain features. H r(r) and disease domain characteristics H r(d) And retain the original feature information through residual connections;

[0046] Specifically, SAM and MLP are used to map drug and disease features to the drug domain or disease domain, so that they can achieve semantic alignment in the drug or disease feature space, so as to facilitate the subsequent reshaping of the adjacency matrix and association prediction.

[0047] The dual-path self-attention mechanism SAM and different MLPs are used to extract drug domain features and disease domain features respectively. The calculation formulas are as follows: Equation 3 and Equation 4:

[0048] H r(r) , H r(d) = MLP r (SAM(A,S r ,S d ), MLP d (SAM(A,S r ,Sd ) ) Formula 3

[0049] H d(r) , H d(d) = MLP r (SAM(A,S r ,S d ), MLP d (SAM(A,S r ,S d ) ) Formula 4

[0050] In the formula, H r(r) ∈ R n*k , H r(d) ∈ R n*k These represent the drug domain and disease domain characteristics of a drug, respectively. ，H d(r) ∈R m*k , H d(d) ∈R m*k These represent the drug domain and disease domain characteristics of a disease, respectively. MLP r With MLP d These represent multilayer perceptrons that embed drug and disease features into the drug domain and disease domain, respectively. SAM(A,S r ,S d ) This describes the process of aggregating drug and disease neighbor features using the correlation matrix and different adjacency matrices. ；

[0051] S3, using a trainable projection matrix P r / P d Mapping drug / disease features to each other's dimensional space to generate an implicit interaction prediction matrix includes the following steps:

[0052] S3-1 Using the projection matrix P r The drug domain features of a drug are mapped to the drug space to obtain the drug remodeled adjacency matrix: H r(r) Pr ,use P d Mapping the disease domain features of a disease to the disease space yields the drug-reshaped adjacency matrix: H d(d) P d ;

[0053] S3-2 iteratively optimizes the reshaped adjacency matrix process based on a recurrent neural network (RNN). It dynamically updates the embedding features of drugs and diseases through multiple rounds of training, and optimizes the features obtained in each iteration. H r(d) as well as H d(r) Perform average fusion to obtain a stable final embedding vector;

[0054] S3-3 utilizes the two projection matrices mentioned above to map the final drug embedding representation to the disease space and the final disease embedding representation to the drug space, performing implicit interactions between drugs and diseases, and simultaneously completing drug-disease association prediction. The specific calculation is shown in Equation 5 below:

[0055] Formula 5

[0056] In the formula, The drug-disease association prediction score, The activation function is sigmoid. H r ∈R n*k and H d ∈R m*k This represents the final embedding of drugs and diseases after multiple rounds of RNN cyclic encoding.

[0057] S4. The weighted cross-entropy loss function is adopted, and the contribution of positive and negative samples is dynamically adjusted by weight λ to solve the data sparsity problem. The specific calculation is as follows: Equation 6:

[0058] Formula 6

[0059] in λ=|y - | / |y + | To balance the weights for positive and negative samples, |y + | and y - | s represents the number of positive and negative samples respectively. ij For the predicted drugs i diseasej The correlation score.

[0060] Drug similarity matrix S r The disease similarity matrix was calculated using the Jaccard distance between the chemical structural fingerprint ECFP6 and the target protein sequence. S d Semantic similarity calculation based on medical subject terms (MeSH).

[0061] A drug relocation system based on implicit interactive adjacency matrix reshaping, used to perform the method, comprising:

[0062] Data input module: used to load drug molecular fingerprints, disease phenotype ontology, and known drug-disease association data;

[0063] Dynamic matrix generation module: used to perform adjacency matrix generation and dynamic adjacency matrix update operations;

[0064] Dual-path feature learning module: includes a parameter-shared SAM layer, an MLP with residual connections, and an RNN iterative optimization unit;

[0065] Implicit interaction prediction module: trained by RNN P r / P d Matrix generation of drug-disease association probability matrix;

[0066] Prediction output module: decodes the final embeddings of drugs and diseases, calculates the drug-disease association probability, and calculates the TOP-K candidate drugs.

[0067] The embodiments of the present invention make full use of the implicit interaction information between drugs and diseases. The direct explicit interaction between drugs and diseases can effectively express the correlation information. In particular, the implicit interaction between drugs and diseases can further reveal complex correlation patterns, thereby improving the accuracy of prediction and biological explanatory power.

[0068] This embodiment can be applied to scenarios such as drug target screening for Parkinson's disease and location-disease association prediction. The following are detailed descriptions using several specific embodiments.

[0069] Example 1

[0070] The drug relocation method and system based on implicit interactive adjacency matrix reshaping provided in this invention embodiment are based on the basic embodiment, and are described and verified in detail using Parkinson's disease drug target screening as an example. Parkinson's disease (PD) is a chronic progressive neurodegenerative disease, ranking second in global neurodegenerative disease incidence, second only to Alzheimer's disease. This invention embodiment conducts a case study on Parkinson's disease in Gdataset to predict new candidate drugs, and uses the Comparative Toxicology Genomics Database (CTD) as a validation source to confirm the performance of DRIIAMR in practical applications, predicting the top 10 candidate drugs for Parkinson's disease (as shown in Table 6).

[0071] See Figure 2 This embodiment is a drug relocation method based on implicit interactive adjacency matrix reshaping, including an RNN cyclic encoding stage and a drug-disease association prediction and decoding stage. In addition, the RNN cyclic encoding stage involves feature embedding of drugs and diseases and adjacency matrix reshaping, and further includes the following content based on the basic embodiment:

[0072] S1. Based on relevant research, construct an initial adjacency matrix according to known drug similarity and disease similarity networks, and initialize drug and disease features;

[0073] In step S1, based on relevant research data, initial data is collected, including drug-disease association data, drug-drug similarity data, and disease-disease similarity data, and the number of drugs is set to [number missing]. The number of diseases is The data includes a drug-disease association matrix. , Indicates drug With disease There is a connection. Indicates drug and disease No association or association not confirmed. Additionally, the drug similarity matrix... , Indicates drug With drugs The similarity between them, and the disease similarity matrix. , Indicates disease With disease The similarity between them.

[0074] S11. Using a drug-disease similarity matrix, The nearest neighbor algorithm extracts the drug (or disease) that is most similar to it. Each drug (or disease) has 10 neighbors and itself, forming a closed neighborhood. This generates the drug adjacency matrix and the disease adjacency matrix, which are respectively... The neighboring relationship between drugs and diseases;

[0075] S12. Based on the above three initial data, first initialize the feature matrices of drugs and diseases in different biological entity spaces, specifically as follows:

[0076]

[0077] in The initial feature matrix represents the drug, including the initial features of the drug in the drug domain. and the initial characteristics of drugs in the disease domain , The initial feature matrix representing the disease includes the initial features of the disease in the drug domain. and the initial characteristics of disease in the disease domain .

[0078] S2. Construct four types of neighbor information aggregation based on self-attention mechanism (SAM) according to four different network relationships;

[0079] In S2, the neighbor information aggregation process aggregates the feature information of drug and disease neighbors. By integrating the correlation data between drugs and diseases and the feature information of their neighbor nodes, four types of neighbor information aggregation are achieved: drug-needle neighbor information aggregation, drug-disease neighbor information aggregation, disease-needle neighbor information aggregation, and disease-needle neighbor information aggregation. This invention utilizes a self-attention mechanism (SAM) to assign different attention weights to different neighbors to achieve neighbor information aggregation. Four hybrid features are obtained through SAM aggregation. The implementation of the four processes can be summarized as follows:

[0080] The aggregation of drug neighbor information is as follows:

[0081]

[0082] The disease neighbor information of drugs is aggregated as follows:

[0083]

[0084] Aggregation of drug neighbor information for diseases:

[0085]

[0086] Disease neighbor information aggregation:

[0087]

[0088] In the formula, Indicates the process The mixed characteristic 1 of the polymerized drug implies information from different feature spaces. Similarly, Indicates the process Mixing characteristics of the polymerized drug 2, and This represents the mixed features of the disease after aggregation via a self-attention mechanism. The learned parameters are shared from the perspective of the aggregated features. and The learning parameters are different.

[0089] S3. Use different multilayer perceptrons to learn dual-path features for drugs and diseases, and learn the embedding representations of drugs and diseases in different semantic spaces respectively.

[0090] Specifically, different multilayer perceptrons (MLPs) are used to learn embeddings for the drug domain and the disease domain, achieving decoupling and representation optimization of the feature space. The aim is to separate the feature representations of the drug domain and the disease domain from the mixed features, including the following steps:

[0091] S31. In step S2, and The polymerized characteristics are all drugs, therefore, respectively... and Perform the same dual-path feature learning:

[0092] Drug domain embedding feature 1: ;

[0093] Drug domain embedding feature 1 of the disease: ;

[0094] Disease domain embedding feature of drugs 1: ;

[0095] Disease domain embedding feature 1: ;

[0096] in and These represent multilayer perceptrons used to learn feature representations for the drug domain and disease domain, respectively, and the embeddings processed when the aggregated features are drugs.

[0097] S32, in step S2 and The aggregated features are all diseases, therefore, respectively... and Perform dual-path feature learning:

[0098] Drug domain embedding feature 2: ;

[0099] Drug domain embedding feature 2 of the disease: ;

[0100] Disease domain embedding feature of drugs 2: ;

[0101] Disease domain embedding feature 2: ;

[0102] In the formula, and These represent multilayer perceptrons used to learn feature representations for the drug domain and disease domain, respectively, and the embeddings processed when the aggregated features are drugs.

[0103] S33, S31, and S32 will yield two embedding features: the drug (or disease) embeddings in the drug space and the disease space, respectively. Next, features from the same feature space will be fused, and a weighted embedding of the initial features will be introduced through a residual connection, resulting in:

[0104] Drug domain characteristics of drugs: ;

[0105] Disease domain characteristics of drugs: ;

[0106] Drug domain characteristics of diseases: ;

[0107] Disease domain characteristics: ;

[0108] in, and It is a trainable parameter matrix used as a residual connection. By implementing the initial feature weighting operation, it effectively preserves the original feature information, alleviates the information loss problem in the feature learning process, and thus improves the robustness and accuracy of the model in predicting drug-disease association.

[0109] S4. Using two projection matrices, the drug domain (or disease domain) features of the drug are mapped to the drug (or disease) dimension, so that they can achieve semantic alignment in the drug (or disease) feature space, and the reshaped adjacency matrix is ​​obtained.

[0110] In step S33, the drug domain characteristics of the drug are obtained. Disease domain characteristics of diseases These are the feature embeddings of drugs in the drug feature space and diseases in the disease space, respectively. Both of these feature embeddings are learned by incorporating drug-disease association information. By mapping these two feature embeddings to the drug dimension and disease dimension respectively, a more accurate drug similarity matrix and disease similarity matrix can be achieved. The calculation method is as follows:

[0111]

[0112] In the formula, and It is a trainable parameter matrix, which is used as a reshaping matrix to map the drug domain features of the drug to the drug dimension and the disease domain features of the disease to the disease dimension.

[0113] S5. The above reshaping process is iteratively optimized based on recurrent neural networks (RNN). The embedding features of drugs and diseases are dynamically updated through multiple rounds of training, and the feature representations obtained in each round of iteration are averaged and fused to obtain a stable final embedding vector.

[0114] Specifically, based on the new information obtained in step S4... and Repeat steps S1-S4 to refine the drug and disease adjacency matrix information; each repeated calculation will output a disease domain feature of a drug in step S33: Drug domain characteristics of diseases Both of these features implicitly contain drug-disease association information. The final drug embedding feature is obtained by averaging and fusing the features from each layer. and the final embedded features of the disease .

[0115] S6. Using the two projection matrices mentioned above, the final drug embedding representation is mapped to the disease space, and the final disease embedding representation is mapped to the drug space to perform implicit interactions between drugs and diseases, while simultaneously completing drug-disease association prediction.

[0116] in, In step S4, this serves as a mapping of drug characteristics. In step S4, these two projection matrices are used as mappings for disease features to create the final drug embedding features. and disease final embedding features The mapping is used to achieve implicit interaction between drug and disease features, thereby completing the prediction of drug-disease association. The specific calculation method is as follows:

[0117]

[0118] in It is the final predicted drug-disease association score matrix. The activation function is sigmoid.

[0119] Loss function design:

[0120] Typically, known drug-disease association matrices are sparse matrices with imbalanced positive and negative sample data. The method in this embodiment of the invention uses weighted cross-entropy as the loss function.

[0121]

[0122] In the formula, and These represent the quantities of drugs and diseases, respectively. This represents the known set of positive samples, i.e., the validated drug-disease association pairs; This represents the negative sample set, i.e., unrelated drug-disease pairs;

[0123] Indicates predictive drug and disease The correlation score; This is used to balance the contribution weights of positive and negative samples. The introduction of this method enhances the model's robustness to class imbalance problems. This invention can dynamically balance the weights of positive and negative samples in a sparse matrix, effectively improving the model accuracy and cross-dataset generalization ability for drug-disease association prediction.

[0124] The benchmark dataset used in this embodiment is described below:

[0125] To evaluate the DRIIAMR method of this invention, experiments were conducted on four benchmark datasets: Gataset, Cdataset, Ldataset, and LRSSL. The Gdataset contains 593 drugs from DrugBank and 313 diseases from the Online Human Mendelian Inheritance Database (OMIM), with a total of 1,933 drug-disease associations. The Cdataset includes 663 drugs from DrugBank, 409 diseases from OMIM, and 2,532 known associations. The Ldataset, from the Comparative Toxicology Genomics Database (CTD), covers 269 drugs, 598 diseases, and 18,416 associations. The LRSSL dataset contains 763 drugs, 681 diseases, and 3,051 validation associations. These datasets cover a wide range of drug categories and disease types, providing a comprehensive foundation for model evaluation. Drug similarity was determined based on standardized SMILES symbols and DrugBank chemical structures, while disease similarity was calculated by analyzing disease phenotypes in the OMIM database using MimMiner. Statistical data for the four datasets are shown in Table 1 (statistical information for the benchmark dataset).

[0126] Table 1

[0127]

[0128] The DRIIAMR was compared with five advanced drug repositioning methods, including:

[0129] lLAGCN: This method utilizes multi-layer GCNs to propagate neighbor information on heterogeneous networks and introduces a layer attention mechanism.

[0130] lDRHGCN: This model uses GCN and bilinear aggregators to design inter-domain and intra-domain embeddings to predict drug-disease associations.

[0131] lDRWBNCF: This model focuses on the interaction information between neighbors and uses generalized matrix factorization collaborative filtering to predict drug-disease associations.

[0132] lHDGAT: This method combines graph convolutional neural networks with bidirectional long short-term memory networks and integrates layer attention mechanisms and residual connections.

[0133] lDRGBCN: This method uses a layered attention graph convolutional network to encode drugs and diseases, and uses a bilinear attention network module to enhance the complex relationships between them.

[0134] Evaluation indicators:

[0135] This invention employs a 10-fold cross-validation experimental method and uses multiple metrics to comprehensively evaluate model performance, including Accuracy, Precision, Recall, F1 Score, AUROC, and AUPR. Accuracy measures the overall classification accuracy, Precision represents the proportion of correctly predicted positive classes, and Recall represents the proportion of correctly predicted positive classes. The harmonic mean that balances precision and recall. AUROC (Area Under the Receiver Operating Characteristic) is calculated by measuring the true positive rate (...). ) and false positive rate ( The area under the precision-recall curve is used to measure model performance. AUPR (Area Under the Precision-Recall Curve) is based on precision (… ) and recall rate ( The comprehensive evaluation of model performance is shown in Table 2 (performance comparison of DRIIAMR with other methods on G dataset under 10-fold cross-validation), Table 3 (performance comparison of DRIIAMR with other methods on C dataset under 10-fold cross-validation), Table 4 (performance comparison of DRIIAMR with other methods on L dataset under 10-fold cross-validation), and Table 5 (performance comparison of DRIIAMR with other methods on LRSSL under 10-fold cross-validation). The prediction results of candidate drugs for Parkinson's disease using DRIIAMR are shown in Table 6.

[0136] Table 2

[0137]

[0138] Table 3

[0139]

[0140] Table 4

[0141]

[0142] Table 5

[0143]

[0144] Table 6

[0145]

[0146] In this embodiment of the invention, the drug relocation method based on implicit interactive adjacency matrix reshaping (DRIIAMR) focuses on the drug similarity matrix. S r Disease similarity matrix S d An initial adjacency matrix is ​​constructed using the drug-disease association matrix A; four types of information fusion are achieved through self-attention mechanism (SAM): intra-drug domain aggregation, cross-drug domain aggregation, cross-disease domain aggregation, and intra-disease domain aggregation; then, a dual-path multilayer perceptron is used to generate drug domain features H. (r) and disease domain features H (d) Then through the projection matrix P r / P d H (r) / H (d) Mapping to the other's dimensional space, the drug / disease adjacency matrix is ​​dynamically updated; finally, feature representation is optimized through iterative training using RNN, and drug-disease association prediction is achieved through implicit interactive projection. Its advantages include:

[0147] Dynamic adjacency matrix reshaping: By introducing a drug-disease association matrix (A) and the original similarity matrix (A) S r / S d The method integrates and uses a recurrent neural network to iteratively generate an optimized drug / disease adjacency matrix, reducing the noise impact of manually set similarity data.

[0148] Dual-path feature learning: A dual-path embedding (H^(r) / H^(d)) is adopted, combining parameter-sharing multilayer perceptron (MLP) and self-attention mechanism (SAM) to achieve decoupled learning of heterogeneous feature spaces.

[0149] Implicit interactive projection mechanism: By using a trainable projection matrix (Pr / Pd), the mapping between drugs and diseases in each other's dimensional space is realized, and implicit feature interaction prediction is completed, breaking through the limitations of traditional explicit interaction.

[0150] Recurrent Encoding Framework: Dynamic feature fusion is achieved through iterative training of RNNs, and residual connections are used to preserve the original feature information, thereby improving the robustness of the model.

[0151] Example 2

[0152] The drug relocation method and system based on implicit interactive adjacency matrix reshaping provided in this invention embodiment are based on the basic embodiment and embodiment 1, and further illustrated using the Gdataset dataset (containing 593 drugs / 313 diseases) as an example. Each step is basically the same as in embodiment 1, except that:

[0153] Implicit interaction weights: projection matrix dimension k=128, attention head number K=8, MLP hidden layer dimension is 128;

[0154] Input feature matrix H init The dimension is 906×906 (n+m=593+313);

[0155] The final predicted AUC is 0.952, an improvement of 13.7% compared to CN119920302A;

[0156] Efficiency Improvement: The parameter sharing mechanism reduces the number of model parameters by 35% and shortens the training time to 70% of that in comparison file 1;

[0157] The method and system of this invention have been deployed on a pharmaceutical AI prediction platform that is being piloted under confidential conditions, and five potential treatments for Parkinson's disease have been successfully screened (as shown in Table 6), which can significantly accelerate the drug development process.

[0158] The embodiments of this invention, through the synergy of dynamic reshaping and implicit interaction, solve the problems of noise in artificial similarity data and the limitations of explicit interaction. Experimental data (Tables 2-5) show that DRIIAMR achieves an average improvement of 8%-15% on AUROC / AUPR, with significant advantages, especially on sparse datasets (such as LRSSL). Specifically, adjacency matrix reshaping reduces noise in artificial similarity data (e.g., validated in Parkinson's disease cases); the implicit interaction mechanism significantly improves the ability to capture complex pathological associations (see...). Figures 3-6 (AUC comparison); the parameter sharing mechanism reduces the number of model parameters and improves training efficiency. Some embodiments can prove that training time is reduced by 30%.

[0159] The embodiments of this invention primarily address the problems of previous drug relocation methods, such as ignoring implicit drug-disease interactions and relying solely on manually set similarity data for adjacency matrix construction. A novel drug relocation method, DRIIAMR, based on implicit drug-disease interactions for adjacency matrix reshaping is proposed. This method introduces drug-disease association data and combines iterative training with a recurrent neural network to dynamically reshape the drug and disease adjacency matrices, thereby improving adjacency information and reducing the noise impact of manually set similarity data. This method integrates "adjacency matrix reshaping" with "implicit interaction modeling," using a dual-path feature learning mechanism to learn the features of drugs and diseases in different biological entity spaces. During the reshaping stage, the learned drug features in the drug domain (…) are utilized… ) and the characteristics of diseases in the disease domain ( ), through a trainable reshaping parameter matrix ( , This generates a more accurate drug-disease similarity matrix; in the implicit interaction modeling stage, the same reshaping parameter matrix is ​​used to represent the average features of the drug in the disease domain. ) and the average characteristics of diseases in the drug domain ( By restoring the drug to its original dimension and performing interactive computation decoding, the complex implicit interaction information between drugs and diseases can be effectively captured and utilized to uncover more valuable predictive signals, thereby significantly improving the accuracy and interpretability of drug repositioning.

[0160] The key features of the above embodiments of the present invention are: constructing a dynamic adjacency matrix by fusing drug-disease association information; using a dual-path self-attention mechanism to achieve decoupled learning of drug domain and disease domain features; using a trainable projection matrix to complete cross-dimensional implicit interaction prediction; and combining recurrent neural networks to iteratively optimize feature representation. This effectively reduces the impact of noise from artificial similarity data, which can improve indicators such as AUC and AUPR by more than 10%, and shows significant advantages in practical scenarios such as drug relocation in Parkinson's disease.

[0161] It should be noted that, within the scope of the present invention described above, other technical solutions employing similar steps, algorithms, data, etc., can achieve the technical effects described in the present invention, and therefore will not be listed one by one.

[0162] The above description is merely a preferred embodiment of the present invention and does not constitute any limitation on the technical scope of the present invention. Therefore, any technical features that are the same as or similar to those in the above embodiments of the present invention are within the protection scope of the present invention.

Claims

1. A drug relocation method based on implicit interactive adjacency matrix reshaping, characterized in that, Includes the following steps: S1. Construct a fused drug-disease association matrix A, and compare it with the original similarity matrix. S r / S d The dynamic adjacency matrix is ​​optimized in real time through a recurrent neural network (RNN) iteratively. in, S r A matrix representing the similarity between drugs. S d A similarity matrix representing diseases; S2. A parameter-sharing dual-path self-attention mechanism (SAM) is used to extract drug domain features. H r(r) , H d(r) and disease domain characteristics H r(d) ,H d(d) And retain the original feature information through residual connections; S3, using a trainable projection matrix P r / P d By mapping drug / disease features to each other's dimensional space, an implicit interaction prediction matrix is ​​generated; S4. The weighted cross-entropy loss function is adopted, and the contribution of positive and negative samples is dynamically adjusted by weight λ to solve the data sparsity problem. Step S1 includes: S1-1 constructs an initial adjacency matrix based on known drug similarity and disease similarity networks, and initializes drug and disease features; S1-2 constructs four neighbor information aggregations based on the self-attention mechanism (SAM) according to four different network relationships; the four neighbor information aggregations are: drug-needle neighbor information aggregation, drug-disease neighbor information aggregation, disease-needle neighbor information aggregation, and disease-disease neighbor information aggregation. S1-3 Original Similarity Matrix S r / S d The mathematical expression for the dynamic adjacency matrix update is given by Equations 1 and 2 below: A r (t) = RNN( A , A r(t-1) , P r Formula 1 A d (t) = RNN ( A T , A d (t-1) , P d Equation 2 In the formula, t is the current loop layer number, and RNN represents the total function of reshaping the adjacency matrix; the RNN process includes two stages: encoding and reshaping. A r (t) ∈ {0,1} n*n and A d (t) ∈ {0,1} m*m This represents the adjacency matrix of drugs and diseases input in each loop. P r ∈ R k*n With P d ∈R k*m These represent the projection matrices that project the low-dimensional features of the drug and disease onto the drug dimension and disease dimension, respectively; where n, m Indicates the quantity of drugs and diseases ，k It is the feature embedding dimension; Step S2 involves using SAM and MLP to map drug and disease features to the drug domain or disease domain, achieving semantic alignment in the drug or disease feature space, so as to reshape the adjacency matrix and predict associations in the future; the dual-path self-attention mechanism SAM and different MLPs are used to extract drug domain features and disease domain features of drugs and diseases respectively. The calculation formulas are as follows: Equations 3 and 4: H r(r) , H r(d) = MLP r (SAM(A,S r ,S d ), MLP d (SAM(A,S r ,S d ) ) Formula 3 H d(r) , H d(d) = MLP r (SAM(A,S r ,S d ), MLP d (SAM(A,S r ,S d ) ) Formula 4 In the formula, H r(r) ∈ R n*k , H r(d) ∈ R n*k These represent the drug domain and disease domain characteristics of a drug, respectively. ，H d(r) ∈R m*k , H d(d) ∈R m*k These represent the drug domain and disease domain characteristics of a disease, respectively. MLP r With MLP d These represent multilayer perceptrons that embed drug and disease features into the drug domain and disease domain, respectively. SAM(A,S r ,S d ) This describes the process of aggregating drug and disease neighbor features using the correlation matrix and different adjacency matrices. 。 2. The method according to claim 1, characterized in that, Step S3 includes: S3-1 Using the projection matrix P r The drug features are mapped to the drug space to obtain the drug remodeled adjacency matrix: H r(r) P r ,use P d Mapping disease features to the disease space yields the drug-reshaped adjacency matrix: H d(d) P d ; S3-2 iteratively optimizes the reshaped adjacency matrix process based on a recurrent neural network (RNN). It dynamically updates the embedding features of drugs and diseases through multiple rounds of training, and optimizes the features obtained in each iteration. H r(d) as well as H d(r) Perform average fusion to obtain a stable final embedding vector; S3-3 utilizes the two projection matrices mentioned above to map the final drug embedding representation to the disease space and the final disease embedding representation to the drug space, performing implicit interactions between drugs and diseases, and simultaneously completing drug-disease association prediction. The specific calculation is shown in Equation 5 below: Formula 5 In the formula, The drug-disease association prediction score, The activation function is sigmoid. H r ∈ R n*k and H d ∈ R m*k This represents the final embedding of drugs and diseases after multiple rounds of RNN cyclic encoding.

3. The method according to claim 1, characterized in that, The model training uses a weighted cross-entropy loss function, which is given by equation 6 below: Formula 6 In the formula λ=|y - | / |y + | To balance the weights for positive and negative samples, |y + | and |y - | s represents the number of positive and negative samples respectively. ij For the predicted drugs i disease j The correlation score.

4. The method according to claim 1, characterized in that, The methods for constructing the original drug / disease similarity matrix include: Drug similarity matrix S r The disease similarity matrix was calculated using the Jaccard distance between the chemical structural fingerprint ECFP6 and the target protein sequence. S d Semantic similarity calculation based on medical subject terms (MeSH).

5. A drug relocation system based on implicit interactive adjacency matrix reshaping, characterized in that, The method for performing any one of claims 1 to 4 comprises: Dynamic matrix generation module: used to perform adjacency matrix generation and dynamic adjacency matrix update operations; Multimodal feature fusion module: integrates heterogeneous features of chemical structure fingerprint ECFP6, target protein sequence, and disease ontology MeSH; Prediction output module: decodes the final embeddings of drugs and diseases, calculates the drug-disease association probability, and calculates the TOP-K candidate drugs.

6. The drug relocation system based on implicit interactive adjacency matrix reshaping according to claim 5, characterized in that, It further includes: Data input module: used to load drug molecular fingerprints, disease phenotype ontology, and known drug-disease association data; Dual-path feature learning module: includes parameter-shared SAM layer, MLP and RNN iterative optimization unit; Implicit interaction prediction module: trained by RNN P r / P d Generate a drug-disease association probability matrix from the matrix.

Images