Drug relocation system and method based on quantum graph Fourier convolution network
By constructing a quantum graph Fourier convolutional network and extracting spectral features using quantum coding and Fourier transform, the problem of ignoring multi-scale structural patterns in existing technologies is solved, and efficient prediction of drug relocation is achieved.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-11-28
- Publication Date
- 2026-04-07
AI Technical Summary
Existing drug relocation methods often neglect the influence of multi-scale structural patterns in biomedical interaction networks, resulting in insufficient predictive performance.
A quantum graph Fourier convolutional network is constructed. By acquiring local subgraphs in heterogeneous biological information networks, spectral features are extracted using quantum encoding and Fourier transform. Combined with multilayer perceptron, the correlation between drugs and diseases is predicted. A frequency domain contrastive regularization loss function is designed for model training.
It improves the discriminative and generalization capabilities of the drug repositioning system, and enhances the accuracy of predicting the association between drugs and diseases.
Smart Images
Figure CN121812007A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the interdisciplinary scientific research field of biomedicine, artificial intelligence, and quantum computing, and relates to a drug relocation system and method based on quantum graph Fourier convolutional networks. Background Technology
[0002] Drug retargeting refers to the use of marketed or clinically-stage drugs for new indications to explore their potential therapeutic value. In recent years, with the development of omics technologies and artificial intelligence, drug retargeting has become an important strategy for addressing bottlenecks such as long drug development cycles, high costs, and low success rates. Studies have shown that drug molecules can typically act on multiple targets, and their mechanisms of action in different diseases share certain commonalities. Therefore, conducting drug retargeting research helps to deepen the understanding of disease pathogenesis, promote the development of personalized treatment, and significantly accelerate the discovery of new therapies.
[0003] Driven by the demand for precision medicine, the biomedical field has accumulated massive amounts of data on the associations between drugs, targets, diseases, and various biomolecules, forming a complex network of biomolecular interactions. While biological experiments can verify the potential associations between drugs and diseases, the process is time-consuming, costly, and difficult to cover large-scale drug combinations. Artificial intelligence-based computational models offer new insights for predicting drug-disease associations, becoming an important tool in drug relocation research. In 2025, Li Yongliang et al. published an article in *Small & Microcomputer Systems* entitled "A Drug Relocation Method Combining Graph Convolution and Contrastive Learning," which predicted potential drug indications by constructing meta-paths and combining contrastive learning. In 2023, Qi Xin et al. published an article in *Journal of University of Electronic Science and Technology of China* entitled "Using Drug Relocation Strategies to Discover Drugs for Gastric Cancer Treatment," which screened drugs through association graph analysis and verified the strong binding ability of the candidate drug levonorgestrel to TIMP1 through molecular docking, providing support for the application of drug relocation in gastric cancer treatment. Meng Xiangmao et al. disclosed a drug relocation method based on subgraph perception and hybrid graph neural networks in patent application number 2025102234098. Lu Pengli et al. disclosed a drug relocation method based on multi-source biological information and graph neural networks in patent application number 2024119738393. Zhang Kaiyuan et al. disclosed a drug relocation method using neighborhood information and weighted fusion networks in patent application number 2024118596094.
[0004] Most of the aforementioned methods focus only on the topological structure or semantic features of drug-disease association networks, often neglecting the impact of the inherent multi-scale structural patterns in biomedical interaction networks on prediction performance. This invention first introduces the association information of proteins, genes, microorganisms, drugs, and diseases to construct a heterogeneous biological information network. Then, based on quantum coding theory and Fourier signal processing technology, a quantum graph Fourier convolutional network is designed to model the heterogeneous biological information network, mining the topological structural patterns and high-order semantic relationships of proteins, genes, microorganisms, drugs, metabolites, and diseases embedded in the spectral space to improve the discriminative and generalization capabilities of node representations in the biological information network, thereby enhancing the performance of drug relocation systems and methods. Summary of the Invention
[0005] To address the shortcomings of existing technologies, this invention proposes a drug relocation prediction method based on multi-source heterogeneous networks. First, a local subgraph corresponding to each pair of candidate drug-disease nodes is constructed based on a drug-disease heterogeneous network. Then, a frequency-domain convolutional neural network is used to perform spectral transformation and one-dimensional convolution operations on the node features in the local subgraphs to extract highly discriminative node embedding representations. Finally, the learned drug and disease embedding representations are input into a multilayer perceptron to predict potential drug-disease associations. Specifically, to achieve the above-mentioned objectives, the specific technical solution of this invention includes: 1. A drug relocation system and method using quantum graph Fourier convolutional networks, characterized in that: Step 1. Obtain datasets of proteins, genes, microorganisms, metabolites, drugs, and diseases, as well as interaction datasets, and construct a drug-disease heterogeneous biological information network. The set of nodes in this heterogeneous biological information network Nodes include proteins, genes, microorganisms, metabolites, drugs, and diseases; edge objects. This includes interaction information and associated edges between different nodes such as proteins, genes, microorganisms, metabolites, drugs, and diseases.
[0006] Step 2. Local Subgraph Generation. Initial features of protein, gene, microorganism, drug, metabolite, and disease nodes are calculated. Based on the heterogeneous biological information network of drugs and diseases, second-order neighbors of target drug nodes and target disease nodes are extracted to form local subgraphs; then, they are labeled according to the shortest path distance between neighbors and target drug nodes. Step 3. Constructing Quantum Graph Fourier Convolution: Based on the local subgraph generated in Step 2, and combining quantum encoding, Fourier transform, and graph convolution, a quantum graph convolution and graph Fourier convolution are designed to extract spectral features and topological information of protein, gene, microorganism, metabolite, drug, and disease nodes from heterogeneous biological information networks. This includes: first, using Fast Fourier Transform to transform node features to the frequency domain; second, decomposing the complex features after Fourier transform into real and imaginary parts and concatenating them into a real vector representation; then, using one-dimensional convolution to perform learnable linear transformations and nonlinear activations on the frequency domain features; and finally, reconstructing the processed features into complex numbers and transforming them back to the time domain using inverse Fourier transform to obtain an enhanced node representation.
[0007] Step 4. Construct a drug relocation predictor. Based on the drug and disease node embedding features extracted from heterogeneous networks, a multilayer perceptron is used to predict the probability of drug-disease association and identify potential relocationable drugs.
[0008] Step 5. Design a loss function that incorporates frequency domain contrast regularization.
[0009] Step 6. Drug relocation model training. The model is trained using a dataset consisting of proteins, genes, microorganisms, metabolites, drugs, and diseases. Dropout and regularization methods are used to optimize the model training process and obtain the optimal parameter settings. Attached Figure Description
[0010] Figure 1 Example diagram of the present invention. Detailed Implementation
[0011] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.
[0012] Figure 1 The method flowchart provided in the embodiments of the present invention is as follows: Figure 1 As shown, the method may include the following steps: 1. A drug relocation system and method using quantum graph Fourier convolutional networks, characterized in that: Step 1. Obtain datasets of proteins, genes, microorganisms, metabolites, drugs, and diseases, as well as interaction datasets, and construct a drug-disease heterogeneous information network. The set of nodes in this heterogeneous network Nodes include diseases, genes, metabolites, proteins, drugs, and microorganisms; edge objects. This includes the connections between different nodes such as diseases, genes, metabolites, proteins, drugs, and microorganisms.
[0013] Step 2. Local subgraph generation. Calculate the sequence features of each gene. and expression spectrum features drug sequence characteristics Structural features Gaussian kernel function similarity Microbial functional gene spectrum and evolutionary marker gene characteristics Evolutionary characteristics of proteins Structural features and sequence features Molecular fingerprint similarity of metabolites Similarity to Gaussian kernel function semantic similarity of diseases Similarity to Gaussian kernel function Gene characteristics were obtained by fusing them separately. Protein characteristics Drug characteristics Microbial characteristics Disease characteristics Metabolite characteristics From the drug-disease heterogeneous information network obtained in step 1, extract the second-order neighbors of the target drug node and the target disease node to form a local subgraph. Then, label them according to the shortest path distance between the neighbors and the target drug node. Specifically, let the target drug node be... The target disease node is In the heterogen diagram Extract its second-order neighbor node set Furthermore, a node labeling strategy is employed to distinguish node types: the target drug node is labeled... The label is set to 0, and the target disease node is... The target node's label is set to 1, and other neighboring nodes are labeled according to their shortest path distance to the target node. Finally, all labels are one-hot encoded to obtain the input feature representation of each node. .
[0014] Step 3. Constructing a quantum graph Fourier convolutional network: Based on the local subgraphs generated in Step 2, a quantum graph convolutional network is proposed to extract the spectral and topological features of protein, gene, microorganism, drug, metabolite, and disease nodes in heterogeneous biological information networks, and obtain discriminative embedding representations of drugs and diseases.
[0015] ① Quantum bit encoding is employed to encode node features in heterogeneous biological information networks into quantum states via quantum gates. To enhance the feature learning model's ability to model sparse, low-frequency graph patterns, a 4-qubit quantum circuit is introduced. This circuit receives an input vector. x It also outputs the Pauli-Z expectation values for a set of qubits. The Pauli-Z operator is: In quantum circuit computation, each qubit is initialized to the Hadamard ground state, and the computation method is as follows: For each qubit, a rotation gate (RY) is applied to map classical features to a quantum state:
[0016] in, These are the encoding parameters related to the input features ( i= (1,2,3,4), generated by quantum feature mapping. In the random quantum layer, the parameter range of a single-layer random parameterized quantum gate is set to [0, 2π], enhancing the nonlinear expressive power.
[0017] ② For the desired measurement value, the circuit output is the expected value of the Pauli-Z operator for each qubit. The measurement results will be stored in a tensor. x The 4-dimensional classical features of the node after passing through this quantum circuit module are: The quantum state information of proteins, genes, microorganisms, drugs, diseases, and metabolites obtained through the quantum convolution module is denoted as follows: .
[0018] ③ Input features to nodes x Perform a forward discrete Fourier transform to convert the time-domain position coding information to the frequency domain, obtaining the information for each node. i In the k Complex representation over each frequency component:
[0019] in d Represents the feature dimension. This represents the imaginary unit. Subsequently, the frequency domain characteristics... Decompose the matrix by its real and imaginary parts and concatenate them according to their dimensions to obtain the frequency domain feature matrix in pure real form: Introducing a linear transformation matrix Combining batch normalization and ReLU activation function Transform the frequency domain features:
[0020] in, and These are learnable parameters, learned through gradient descent. This represents the mean. Represents variance. To prevent division by zero of extremely small constants. The transformed eigenvectors. Reassembled into a frequency domain representation in complex form:
[0021] ④ Perform an inverse discrete Fourier transform on the frequency domain features to restore the node representation to the time domain, thus obtaining the node... i After frequency domain transformation, the first n 3D eigenvalues Combining As the final output feature: Step 4. Construct a drug relocation predictor. Based on the node frequency domain features extracted by Fourier convolution in Step 3. A graph-level representation is obtained through global sorting pooling. The transformation process is as follows: .in It is based on node features Characteristic matrices arranged in descending norm order K The preset fixed pooling length is used; a fully connected neural network classifier is used to perform nonlinear transformation to predict the drug-disease association, obtaining a prediction score indicating a therapeutic relationship between the drug and the disease. The calculation process is as follows:
[0022] Step 5. Design a loss function that incorporates frequency domain contrast regularization. Train and optimize the model based on the cross-entropy loss function, introducing a term into the entropy loss. The cross-entropy loss calculation is used to assess the loss between the output of the drug relocation association prediction model and the actual data. The calculation process of the loss function is as follows: , This is the regularization weight hyperparameter, which controls the importance of the contrast loss. M represents the batch size. express loss function Anchor point sample Frequency domain characteristics, These are the frequency domain features of positive samples. It is the frequency domain feature of the negative sample.
[0023] Step 6. Drug relocation model training. The model is trained on protein, gene, microbial, drug, metabolite, and disease datasets. Dropout and regularization strategies are introduced during training to reduce the risk of overfitting during the training phase.
Claims
1. A drug relocation system and method using quantum graph Fourier convolutional networks, characterized in that: Step 1. Construct a heterogeneous biological information network of drugs and diseases. The set of nodes in this heterogeneous biological information network Nodes including proteins, genes, microorganisms, metabolites, drugs, and diseases, and edge objects. This includes interaction information and associated edges between different nodes such as proteins, genes, microorganisms, metabolites, drugs, and diseases; Step 2. Local Subgraph Generation: Based on the heterogeneous biological information network of drug-disease, for each candidate drug-disease relationship pair, the corresponding drug and disease nodes are selected as the central nodes. Second-order neighbor nodes are extracted based on the labeling strategy of node type and depth to form a local subgraph; the initial features of protein, gene, microorganism, metabolite, drug, and disease nodes are calculated. Step 3. Constructing Quantum Graph Fourier Convolution: Based on Fourier transform, the node features in the biological information network are mapped to the spectral space. The importance of information of different frequency components is adaptively modeled through quantum graph convolution network, effectively capturing the global structural patterns and high-order semantic dependency information of protein, gene, microorganism and metabolite nodes, and generating discriminative drug and disease node embedding representations. Step 4. Construct a drug relocation predictor. Based on the drug and disease node embedding features extracted from heterogeneous networks, a multilayer perceptron is used to predict their association probability and identify potential therapeutic drugs for the disease; Step 5. Design a loss function that incorporates frequency domain contrast regularization; Step 6. Drug relocation model training. The model is trained using a dataset consisting of proteins, genes, microorganisms, metabolites, drugs, and diseases. Dropout and regularization methods are used to optimize the model training process and obtain the optimal parameter settings.
2. The drug relocation system and method using a quantum graph Fourier convolutional network according to claim 1, characterized in step 1: acquiring datasets of proteins, genes, microorganisms, drugs, metabolites, and diseases, as well as interaction and correlation datasets, to construct a multi-source heterogeneous biological information network. .
3. The drug relocation system and method using a quantum graph Fourier convolutional network according to claim 1, step 2 is characterized in that: the second-order neighbors of the target drug node and the target disease node are extracted to form a local subgraph. Then, the neighbors are labeled according to the shortest path distance between them and the target drug node; the features of protein, gene, microorganism, drug, metabolite, and disease nodes are calculated.
4. The drug relocation system and method of quantum graph Fourier convolutional network according to claim 1, wherein step 3 is characterized in that: based on Fourier transform, node features are mapped to the spectral space, and the contributions of different frequency components are adaptively modeled through quantum graph convolution, effectively capturing the structural pattern information and high-order semantic dependency features of nodes in the biological information network, generating the embedding representation features of nodes, and obtaining discriminative representations of drugs and diseases.
5. The drug relocation system and method based on a quantum graph Fourier convolutional network according to claim 1, wherein step 4 is characterized in that: the discriminative features learned in step 3 are fused using a multilayer perceptron and the relationship between the drug and the disease is predicted.
6. The drug relocation system and method based on quantum graph Fourier convolutional networks according to claim 1, step 5 is characterized in that: by combining binary cross-entropy loss and frequency domain contrast regularization, an effective loss function is designed for the drug relocation task, which effectively improves prediction performance.
7. The drug relocation system and method using a quantum graph Fourier convolutional network according to claim 1, characterized in that, In step 6, the model is trained on datasets of proteins, genes, microorganisms, drugs, metabolites, and diseases. The Dropout mechanism and regularization strategy are combined to suppress overfitting, thereby optimizing the model training process and obtaining the optimal parameter configuration.