A method and system for predicting herbal gene interactions
By employing dual hypergraph convolution and bidirectional gating attention mechanisms, the problem of high-order group synergistic effects of multiple components and targets in traditional Chinese medicine was solved, achieving high-precision prediction of herbal-gene interactions and improving the quality and robustness of cross-modal information fusion.
Patent Information
- Application Number
- CN202610857388.X
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2026-06-15
- Publication Date
- 2026-11-10
- Estimated Expiration
- 2046-06-15
AI Technical Summary
Existing technologies struggle to effectively model the high-order synergistic effects of multiple components and targets in traditional Chinese medicine, and lack adaptive feature fusion of cross-modal information, resulting in insufficient accuracy in predicting herb-gene interactions.
We employ a dual hypergraph convolution and bidirectional gating attention mechanism to construct a high-order population association model between herbs and genes, and achieve cross-modal semantic collaboration through bidirectional cross-attention alignment and adaptive feature fusion.
It significantly improves the accuracy of herbal gene interaction prediction, especially in scenarios with sparse data and cold start, showing excellent robustness, with prediction accuracy superior to existing methods, and has biological interpretability.
Smart Images

Figure CN122392617B_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of bioinformatics processing and intelligent drug screening technology, specifically relating to a herbal-gene interaction prediction method and system that uses hypergraph neural networks, bidirectional attention and adaptive gating mechanisms to predict the potential interaction between traditional Chinese herbal medicines and biological target genes. Background Technology
[0002] Traditional Chinese medicine has a long history of clinical application, and its therapeutic effects are usually achieved through complex synergistic interactions between various active compounds in herbs and multiple biological targets (genes and proteins) in the human body. Systematic identification of herb-gene interactions is of great significance for elucidating the molecular mechanisms of traditional Chinese medicine and accelerating drug discovery. However, traditional wet experimental validation methods are costly, time-consuming, and difficult to implement on a large scale.
[0003] In recent years, computational prediction methods have been extensively studied. Early computational methods mainly included similarity-based reasoning, network propagation algorithms, and traditional machine learning models. However, these methods have the following technical shortcomings: (1) They model interactions as independent pairwise relationships, which cannot capture the high-order, group-based dependency relationships of "multiple herbs-multiple targets" in traditional Chinese medicine prescriptions; (2) Herbs and genes belong to heterogeneous biological modalities with different semantic features, and traditional models lack effective cross-modal information integration mechanisms.
[0004] Although graph neural networks have been applied in the field of bioinformatics, they are inherently limited by their binary edge structure, making it difficult to mathematically model the "multi-component synergistic" action mode of traditional Chinese medicine. Hypergraph neural networks, through their hyperedge structure connecting any number of vertices, can theoretically overcome the above limitations. However, existing hypergraph-based methods still have the following technical bottlenecks: (1) lack of alignment mechanism for heterogeneous spaces between herbs and genes, leading to semantic mismatch; (2) the use of a unified message passing strategy, treating all nodes within the hyperedge equally, ignoring the differentiated functional contributions of different herbs in the formula; (3) rigid feature fusion methods (such as simple splicing or addition), which easily amplify the noise information of each modality, resulting in a significant decrease in prediction performance in data-sparse scenarios. Summary of the Invention
[0005] The purpose of this invention is to propose a herbal gene interaction prediction method and system that uses highly secure encryption technology and novel key storage technology to solve the problem of low security in current cloud storage systems.
[0006] This invention aims to provide a herbal gene interaction prediction method and system, which is constructed based on dual hypergraph convolution and bidirectional gated attention, in order to solve the technical problems in the prior art that it is impossible to model high-order group associations, difficult to align heterogeneous semantic spaces, and lack of adaptability in feature fusion.
[0007] To achieve the above objectives, the technical solution adopted by the present invention is as follows:
[0008] A method for predicting herbal gene interactions includes the following steps:
[0009] Step S1: Acquisition and Embedding Initialization of Biological Interaction Data
[0010] Obtain known interaction data between herbs and genes, and construct a binary adjacency matrix;
[0011] Step S2: Construction of the dual hypergraph and hypergraph convolution processing
[0012] Based on the binary adjacency matrix, herbal hypergraphs and gene hypergraphs are constructed respectively, and the herbal structure embedding matrix and gene structure embedding matrix are generated by processing through the hypergraph convolution module.
[0013] Step S3: Bidirectional cross-attention semantic alignment
[0014] The herbal structure embedding matrix and the gene structure embedding matrix are input together into the bidirectional cross-attention module. Through bidirectional alignment, the herbal embedding and gene embedding achieve cross-modal space mapping under the semantic guidance of each other.
[0015] Step S4: Bidirectional Gated Adaptive Fusion
[0016] Gated fusion operations are performed on the herbal medicine branch and the gene branch respectively, and the fusion structure embedding and attention features are used to generate fusion features;
[0017] Step S5: Residual Projection and Interactive Score Calculation
[0018] The final representation is generated by residual projection, and the interactive prediction score is calculated.
[0019] Step S6: Model Optimization and Output
[0020] The model is trained using Bayesian personalized ranking loss and outputs a list of candidate genes sorted by predicted scores.
[0021] Compared with the prior art, the beneficial effects of the present invention are as follows:
[0022] (1) High-order group association modeling ability: By constructing herbal hypergraphs and gene hypergraphs, and explicitly modeling the group synergy between "multiple herbs and multiple genes" in the form of mutual hyperedges, it breaks through the technical bottleneck that traditional graph neural networks can only handle binary relationships, and more accurately matches the pharmacological characteristics of "multiple components and multiple targets" of traditional Chinese medicine.
[0023] (2) Bidirectional alignment of heterogeneous semantic space: For the first time, a bidirectional cross-attention mechanism is introduced in the herbal-gene prediction task, enabling the embedding representations of herbs and genes to achieve semantic collaboration in mutual querying, effectively solving the "semantic gap" problem between the two heterogeneous biological modalities, and significantly improving the quality of cross-modal information fusion.
[0024] (3) Adaptive noise suppression and robust enhancement: Through a bidirectional gating fusion strategy, the model can dynamically adjust the weight ratio of structural topology information and cross-modal attention information according to the input features, and exhibits excellent robustness in real biomedical scenarios such as data sparsity and cold start. Ablation experiments show that the model performance decreases by about 5-8% after removing the gating mechanism.
[0025] (4) The prediction accuracy is significantly better than existing methods: On a real dataset containing 563 kinds of herbs, 2106 genes and 38002 known associations, the HR@10 index of this invention reaches 0.4909, which is a relative improvement of 7.02% compared with the current best baseline model MHGCN (0.4587). It surpasses all comparison methods in all indicators such as Top@1, 3, 5, 10 and 20.
[0026] (5) Strong biological interpretability: Through molecular docking verification of the case of Salvia miltiorrhiza, the core target PPARG predicted in this invention has a stable hydrogen bond with the active ingredient (binding energy -8.11 kcal / mol, hydrogen bond distance 1.9 Å), proving that the prediction results have a real physicochemical basis and can directly serve the verification of wet experiments. Attached Figure Description
[0027] Figure 1 This is a flowchart of a herbal gene interaction prediction method.
[0028] Figure 2 This is a bar chart comparing the performance of the embodiments of the present invention with nine baseline models on the Top@K metric.
[0029] Figure 3 This is a bar chart comparing the performance of ablation experiments on the Top@K metric.
[0030] Figure 4 This is a diagram showing the pathway enrichment analysis results in the case of Danshen (Salvia miltiorrhiza) in an embodiment of the present invention.
[0031] Figure 5 This is a graph showing the molecular docking verification results in the case of Danshen (Salvia miltiorrhiza) in an embodiment of the present invention. Detailed Implementation
[0032] like Figure 1As shown, this invention proposes a herbal gene interaction prediction method based on dual-hypergraph convolution and bidirectional gated attention, comprising the following steps:
[0033] Step S1: Acquisition and Embedding Initialization of Biological Interaction Data
[0034] Obtain known interaction data between herbs and genes, and construct a binary adjacency matrix R; where the set of herbs is denoted as . Gene set denoted as Matrix element R ij =1 indicates the herb h i With gene g j There is a known association, R ij =0 indicates that the association was not observed.
[0035] Randomly initialize the initial embedding representation vectors of each entity in the herb set and gene set respectively to obtain the initial embedding matrix of the herb. and the initial embedding matrix of genes , where d is the preset embedding dimension;
[0036] Step S2: Construction of the dual hypergraph and hypergraph convolution processing
[0037] Based on the binary adjacency matrix R, construct the herbal hypergraph and the gene hypergraph respectively:
[0038] Correlation matrix of herbal hypergraph Each gene node serves as a hyperedge, connecting all herb nodes that have known interactions with that gene.
[0039] Association matrix of gene hypergraph Each herb node serves as a hyperedge, connecting all gene nodes that have known interactions with that herb.
[0040] Initial embedding matrix of herbs Association matrix with herbal hypergraph Input the first hypergraph convolutional layer to generate the herbal structure embedding matrix. ; Initialize the gene embedding matrix Association matrix with gene hypergraph Input the second hypergraph convolutional layer to generate the gene structure embedding matrix. .
[0041] The message passing formula for hypergraph convolutional layers is: Where X is the input node feature matrix, H is the correlation matrix, and D is the input node feature matrix. v Let D be a diagonal matrix of vertex degrees. e Let be a hyper-diagonal matrix, W be a learnable linear transformation matrix, and b be a bias vector.
[0042] Step S3: Bidirectional cross-attention semantic alignment
[0043] Embedding herbal structures into a matrix With gene structure embedding matrix The input is fed into the bidirectional cross-attention module, which executes the following two parallel attention computation paths:
[0044] Herbal to Gene Alignment Path: After linear transformation, it becomes the query matrix Q. h ,by After linear transformation, it becomes the bond matrix K. g Sum matrix V g Calculate the multi-head attention output to obtain the herbal attention feature matrix. .
[0045] Gene-to-Herbal Alignment Path: After linear transformation, it becomes the query matrix Q. g ,by After linear transformation, it becomes the bond matrix K. h Sum matrix V h Calculate the multi-head attention output to obtain the gene attention feature matrix. .
[0046] Through bidirectional alignment, herbal embeddings and gene embeddings can achieve cross-modal space mapping under each other's semantic guidance;
[0047] Step S4: Bidirectional Gated Adaptive Fusion
[0048] Perform gated fusion operations on the herbal branch and the gene branch respectively:
[0049] For herbal branches, embedding herbal structures Herbal attention characteristics The components are concatenated and input into the first gating network to calculate the herbal fusion gating vector. and generate herbal fusion characteristics. .
[0050] For gene branches, the gene fusion gating vector is calculated symmetrically. and gene fusion characteristics .
[0051] Where || denotes the vector concatenation operation, σ is the Sigmoid activation function, ⊙ is the element-wise multiplication, and the gate vector is used to dynamically balance the fusion weights of structural information and attention information.
[0052] In the gating fusion operation, the herbal fusion gating vector and gene fusion gating vector It is generated by independently learnable parameters, and each element of the gating vector takes a value between (0,1), which is used to dynamically allocate the fusion weights of structural embedding and attention features in the feature dimension.
[0053] Step S5: Residual Projection and Interactive Score Calculation
[0054] Initial embedding of herbs Herbal fusion characteristics After concatenation, the herbal representation matrix is mapped back to the embedding dimension d through the first linear projection layer to obtain the final representation matrix of the herbs. .
[0055] Initial gene embedding Gene fusion characteristics After concatenation, the matrix is mapped back to the embedding dimension d through a second linear projection layer to obtain the final gene representation matrix. .
[0056] For any herb node h and gene node g, the interaction prediction score is calculated using the vector dot product. The candidate gene list is then sorted by score and output.
[0057] Step S6: Model Optimization and Output
[0058] The model is trained end-to-end using a Bayesian personalized ranking loss function, which makes the prediction score of positive sample pairs higher than that of negative sample pairs. After training, for any herb node to be predicted, the method outputs a list of candidate genes sorted from high to low prediction scores, which serves as the screening result for potential targets of the herb.
[0059] The model is trained end-to-end using a Bayesian personalized ranking loss function:
[0060]
[0061] For a given herb h, Genes with known positive associations Let denoted as a randomly sampled unobserved gene, D be the set of sampled triples, σ be the sigmoid function, Θ be all trainable parameters, and λ be the regularization coefficient.
[0062] The present invention will be further described in detail below with reference to the embodiments and accompanying drawings.
[0063] Example 1:
[0064] I. Experimental Data and Preprocessing
[0065] The dataset used in this embodiment contains 563 kinds of herbs, 2106 genes, and a total of 38002 known herb-gene associations. All associations are stored in the form of a binary matrix R, where This indicates a verified interaction between herb i and gene j.
[0066] The dataset was randomly divided into training, validation, and test sets in an 8:1:1 ratio. The training set was used for model parameter optimization, the validation set for hyperparameter selection, and the test set for final performance evaluation. All experiments were repeated 5 times, and the average results were reported.
[0067] II. Model Parameter Settings
[0068] In this embodiment, the key hyperparameters of the model are set as follows:
[0069] Embedding dimension d = 64 (balancing efficiency and effectiveness);
[0070] The number of hypergraph convolutional layers L = 4 (performance continues to improve in the range of 1-4 layers, but oversmoothing occurs in the 5th layer).
[0071] Multi-head attention count = 8;
[0072] The regularization coefficient λ in the Bayesian personalized ranking loss is 1e-4;
[0073] The optimizer used is Adam, with the learning rate initialized to 0.001 and the weight decay set to 1e-5.
[0074] The training epochs are 500, and an early stopping mechanism is used. Training stops when the validation set NDCG@10 shows no improvement for 20 consecutive epochs.
[0075] Batch Size = 2048.
[0076] III. Specific Implementation of Hypergraph Convolutional Layers
[0077] Taking the herb hypergraph as an example, given the initial embedding matrix of herbs... and the correlation matrix (Each column represents a gene hyperedge), the calculation process of the hypergraph convolutional layer is as follows:
[0078] First, calculate the vertex degree matrix. Its diagonal elements The hyperedge degree matrix is the number of hyperedges connected to herb i (i.e., the number of genes associated with that herb). Calculate the hyperedge degree matrix. Its diagonal elements , which is the number of herb nodes contained in the superedge e (that is, the number of herbs associated with the gene).
[0079] Then, normalized hypergraph message passing is performed:
[0080]
[0081] in The weight matrix is a learnable matrix. This is the bias vector. and These represent the negative 1 / 2 power and the inverse of the degree matrix, respectively, and are used to normalize the aggregated information to prevent gradient instability caused by differences in node degree.
[0082] Similarly, gene supergraph convolutional layers with and Input, Output The details are as follows:
[0083] Given the initial gene embedding matrix and the correlation matrix (Each column represents a gene hyperedge), the calculation process of the hypergraph convolutional layer is as follows:
[0084] First, calculate the vertex degree matrix. Its diagonal elements This refers to the number of hyperedges connected to gene j (i.e., the number of genes associated with that gene). Calculate the hyperedge degree matrix. Its diagonal elements , which is the number of gene nodes contained in the superedge e (that is, the number of genes associated with that gene).
[0085] Then, normalized hypergraph message passing is performed:
[0086]
[0087] in The weight matrix is a learnable matrix. This is the bias vector. and These represent the negative 1 / 2 power and the inverse of the degree matrix, respectively, and are used to normalize the aggregated information to prevent gradient instability caused by differences in node degree.
[0088] IV. Specific Implementation of Bidirectional Cross-Attention
[0089] The result obtained in the previous step and Input a multi-head attention layer. This embodiment uses 8 attention heads, each with a dimension of 64 / 8=8.
[0090] For the alignment path from herbs to genes:
[0091] First, a query, key, and value matrix is generated through linear transformation:
[0092] ,
[0093] in It is a learnable matrix.
[0094] Then, calculate the scaled dot product attention:
[0095]
[0096] in = 8 is the dimension of each attention head. The outputs of the 8 heads are concatenated, and then a linear transformation is applied to obtain the final herbal attention feature. .
[0097] For the alignment path from gene to herb, calculate symmetrically:
[0098]
[0099]
[0100] Through the bidirectional computation described above, semantic information from the gene space is incorporated into the herbal embedding, and semantic information from the herbal space is also incorporated into the gene embedding, achieving true cross-modal bidirectional alignment.
[0101] V. Specific Implementation of Two-Way Gating Fusion
[0102] Taking herbal medicine branches as an example, the implementation details of gating fusion are as follows:
[0103] First, the structural embedding and attention features are concatenated along the feature dimension:
[0104] The spliced dimensions are 128.
[0105] Then, the gate vector is computed through a fully connected layer with a Sigmoid activation function:
[0106]
[0107] in Gating vector Each element takes a value between (0,1), representing the weight proportion of the feature in the corresponding dimension from the attention branch.
[0108] Finally, perform weighted fusion:
[0109]
[0110] when When the value is close to 1, the model tends to use the attention-aligned semantic features; when... When the value approaches 0, the model tends to use the original hypergraph structure features. This gating mechanism is entirely data-driven and can adaptively adjust the fusion strategy based on the characteristics of different herb-gene pairs.
[0111] The computation of gene branching is completely symmetric, using independent gating parameters. and The implementation details are as follows:
[0112] First, the structural embedding and attention features are concatenated along the feature dimension:
[0113] The spliced dimensions are 128.
[0114] Then, the gate vector is computed through a fully connected layer with a Sigmoid activation function:
[0115]
[0116] in Gating vector Each element takes a value between (0,1), representing the weight proportion of the feature in the corresponding dimension from the attention branch.
[0117] Finally, perform weighted fusion:
[0118]
[0119] when When the value is close to 1, the model tends to use the attention-aligned semantic features; when... When the value is close to 0, the model tends to use the original hypergraph structure features.
[0120] VI. Residual Projection and Score Prediction
[0121] To preserve the original information in the initial embedding, a residual-connected projection layer is used:
[0122] First, the initial embedded and fused features are concatenated:
[0123]
[0124] Then it is compressed back to 64 dimensions through a linear projection layer:
[0125]
[0126] in .
[0127] For gene branches, calculate symmetrically. .
[0128] Interaction scores are calculated using the inner product:
[0129]
[0130] in, This refers to the prediction score of herbs and genes.
[0131] VII. Loss Function and Training
[0132] This method employs Bayesian personalized ranking loss, the core idea of which is: for a given herb h, to make it correlated with known positive sample genes. The score was higher than that of unobserved genes. The score.
[0133] For each training batch, triples (h, ...) are sampled from the training set. , ), where (h, ) is a known positive correlation, (h, () represents randomly sampled unrelated pairs. The loss function is defined as:
[0134]
[0135] Where σ is the Sigmoid function, D is the set of sampled triples, Θ represents all trainable parameters (including embedding matrix, hypergraph convolutional layer parameters, attention parameters, gating parameters, and projection layer parameters), and λ=1e-4 is the regularization coefficient.
[0136] VIII. Experimental Results
[0137] 1. Under the above dataset and parameter settings, the method of this invention is comprehensively compared with nine baseline methods (including CN, Jaccard, AA, RA, PRINCE, heNetRW, BPR-MF, HTINET2, and MHGCN). The evaluation metrics used are Hit Ratio (HR@K) and Normalized Discounted Cumulative Gain (NDCG@K), with K values of 1, 3, 5, 10, and 20.
[0138] Experimental results are as follows Figure 2 As shown, the method proposed in this invention significantly outperforms various existing methods on real datasets, demonstrating excellent performance.
[0139] 2. Ablation test
[0140] In addition, to verify the contribution of each core module, three sets of ablation experiments were designed:
[0141] w / o Gate_fusion: Remove gated fusion and replace it with simple splicing followed by direct projection;
[0142] w / o Gate & Hyper: Simultaneously remove gated fusion and hypergraph convolutional layers (replace with ordinary GCN);
[0143] w / o Hyper & Attention: Simultaneously removes hypergraph convolution and bidirectional attention (only retains gated fusion).
[0144] like Figure 3 As shown, the ablation experiment shows that: (1) after removing the gating fusion, the performance drops by about 5-8%, indicating that adaptive fusion is crucial for suppressing noise; (2) after removing the hypergraph structure, the performance drops the most significantly (about 35%), proving that high-order group association modeling is the core technical contribution of this invention; (3) after removing both the hypergraph and attention, the performance is close to that of traditional methods, indicating that the two work together rather than simply superimposed.
[0145] 3. Cold start experiment
[0146] In real-world applications, newly discovered herbs or genes often lack historical interaction data (the cold start problem). To simulate this scenario, nodes with a degree ≤20 in the test set were divided into a cold start subset. Experimental results (shown in Table 1) show that the HR@10 of this invention reaches 0.4821 in the herb cold start scenario, a 16.9% improvement compared to MHGCN (0.4125); it also maintains the best performance in the gene cold start scenario. This demonstrates that the dual-hypergraph structure and gated fusion can effectively utilize high-order similarity to "warm up" new nodes.
[0147] Table 1 Cold start experiment results
[0148] method Herbal Cold Start (HR) Herbal Cold Start (NDCG) Gene cold start (HR) Cold start gene (NDCG) CN 0.1429 0.0464 0.0001 0.0001 Jaccard 0.1071 0.0290 0.0055 0.0016 AA 0.1250 0.0420 0.0055 0.0004 RA 0.1250 0.0381 0.0331 0.0114 heNetRW 0.1429 0.0436 0.0166 0.0062 BPR-MF 0.0179 0.0113 0.0110 0.0023 HITNET2 0.3929 0.2435 0.0351 0.0409 MHGCN 0.4125 0.2698 0.0436 0.0441 OURS 0.4821 0.2942 0.0473 0.0493
[0149] Example 2
[0150] Biological Case Validation
[0151] To verify the biological interpretability of this invention, target prediction was performed using Salvia miltiorrhiza (a traditional Chinese medicine used to treat cardiovascular diseases) as an example. During the training phase, all known associations of Salvia miltiorrhiza were masked, and the model was tested to see if it could predict the targets supported by the literature.
[0152] like Figure 4 As shown, the prediction results indicate that 8 out of the top 10 genes have been validated in the literature. Among them, PPARG (peroxisome proliferator-activated receptor γ), which ranks first, has been confirmed by multiple studies as a core target of tanshinone compounds, inhibiting tumor progression through the PPARγ / RXRα signaling pathway.
[0153] like Figure 5 As shown, further molecular docking experiments revealed that the active ingredient tanshinone IIA of tanshinone binds to PPARG protein at a rate of -8.11 kcal / mol and forms a strong hydrogen bond of 1.9 Å at the key residue ASP-475, verifying the physicochemical rationality of the predicted results.
[0154] In summary, this invention discloses a method and system for predicting herbal-gene interactions, which can be integrated into a traditional Chinese medicine new drug development platform, a drug relocation system, or a traditional Chinese medicine knowledge graph analysis tool for:
[0155] Rapidly screen potential active targets of herbs to guide wet experimental design;
[0156] Predicting new indications for known drugs (drug repositioning);
[0157] Analyzing the multi-target synergistic mechanism of compound traditional Chinese medicine;
[0158] Assess the potential off-target effects and toxicological risks of herbal ingredients.
[0159] Compared with purely experimental methods, this invention can narrow down the candidate target screening range from thousands of genes to the top 10-20, reducing experimental verification costs by more than 95%, and has significant industrial application value.
Claims
1. A method for predicting herbal gene interactions, characterized in that, Includes the following steps: Step S1: Acquisition and Embedding Initialization of Biological Interaction Data Obtain known interaction data between herbs and genes, and construct a binary adjacency matrix; Step S2: Construction of the dual hypergraph and hypergraph convolution processing Based on a binary adjacency matrix, a herbal hypergraph and a gene hypergraph are constructed respectively. The hypergraph convolution module is then used to generate the herbal structure embedding matrix and the gene structure embedding matrix; specifically: Based on the binary adjacency matrix R, a herbal hypergraph and a gene hypergraph are constructed respectively. The herbal hypergraph contains an association matrix. Each gene node serves as a hyperedge connecting all herb nodes that have known interactions with that gene; the association matrix of the gene hypergraph Each herb node serves as a hyperedge connecting all gene nodes that have known interactions with that herb; Initial embedding matrix of herbs Association matrix with herbal hypergraph Input the first hypergraph convolutional layer to generate the herbal structure embedding matrix. ; Initialize the gene embedding matrix Association matrix with gene hypergraph Input the second hypergraph convolutional layer to generate the gene structure embedding matrix. ; Step S3: Bidirectional cross-attention semantic alignment The herbal structure embedding matrix and the gene structure embedding matrix are jointly input into a bidirectional cross-attention module. Through bidirectional alignment, the herbal embeddings and gene embeddings achieve cross-modal space mapping under each other's semantic guidance; specifically: Embedding herbal structures into a matrix With gene structure embedding matrix The input is fed into a bidirectional cross-attention module, which executes the herb-to-gene alignment path and the gene-to-herb alignment path, respectively generating herb attention feature matrices. and gene attention feature matrix ; In the herbal-to-gene alignment pathway, After linear transformation, it becomes the query matrix. ,by After linear transformation, it becomes the key matrix. Sum matrix ; In the gene-to-herb alignment pathway, After linear transformation, it becomes the query matrix. ,by After linear transformation, it becomes the key matrix. Sum matrix ; Step S4: Bidirectional Gated Adaptive Fusion Gated fusion operations are performed on the herbal medicine branch and the gene branch respectively, and the fusion structure embedding and attention features are used to generate fusion features; specifically: Perform gated fusion operations on the herbal medicine branch and the gene branch separately: For the herbal medicine branch, calculate the herbal medicine fusion gate vector. and generate herbal fusion characteristics. For gene branches, the gene fusion gating vector is calculated symmetrically. and gene fusion characteristics ; Step S5: Residual Projection and Interactive Score Calculation The final representation is generated by residual projection, and the interactive prediction score is calculated. Step S6: Model Optimization and Output The model is trained using Bayesian personalized ranking loss and outputs a list of candidate genes sorted by predicted scores.
2. The herbal-gene interaction prediction method according to claim 1, characterized in that, In step S4, the herbal fusion gate vector and gene fusion gating vector It is generated by independent learnable parameters, and each element of the gating vector takes a value between (0,1), which is used to dynamically allocate the fusion weights of structural embedding and attention features in the feature dimension.
3. The herbal-gene interaction prediction method according to claim 1, characterized in that, In step S5, the herbs are initially embedded. Herbal fusion characteristics After being concatenated and mapped through the first linear projection layer, the final representation matrix of the herbs is obtained. Initial embedding of genes Gene fusion characteristics After splicing, the final gene representation matrix is obtained by mapping through a second linear projection layer. ; For any herb node h and gene node g, the interaction prediction score is calculated using the vector dot product. The candidate gene list is then sorted by score and output.
4. The herbal-gene interaction prediction method according to claim 3, characterized in that, In step S6, the model is trained end-to-end using the Bayesian personalized ranking loss function: For a given herb h, For genes that are known to be positively associated, Let denoted as a randomly sampled unobserved gene, D be the set of sampled triples, σ be the sigmoid function, Θ be all trainable parameters, and λ be the regularization coefficient.
5. A herbal-gene interaction prediction system, based on the herbal-gene interaction prediction method according to any one of claims 1-4, characterized in that, include: The data acquisition module is used to acquire known interaction data between herbs and genes and construct a binary adjacency matrix; A dual-hypergraph construction module is used to construct herb hypergraphs and gene hypergraphs based on binary adjacency matrices; The hypergraph convolution module includes a first hypergraph convolution layer and a second hypergraph convolution layer, which are used to generate herbal structure embeddings and gene structure embeddings, respectively. A bidirectional cross-attention module is used to perform bidirectional semantic alignment between herbs and genes, generating attention features; A bidirectional gated fusion module is used to adaptively fuse structural embeddings and attention features to generate fused features; The projection prediction module is used to generate the final representation through residual projection and calculate the interactive prediction score. The output module is optimized to train the model using Bayesian personalized ranking loss and output a list of candidate genes sorted by predicted scores.
6. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the steps of the method according to any one of claims 1 to 4.
7. A computer-readable storage medium having a computer program stored thereon, characterized in that, When executed by a processor, the program implements the steps of the method described in any one of claims 1 to 4.
Citation Information
Patent Citations
Drug-disease association prediction method, device and electronic device based on hypergraph
CN119786064A
Social recommendation method based on conditional diffusion denoising
CN121614681A
