A method for predicting miRNA-drug interactions based on route-guided dual attention and gated graph mask autoencoder
By constructing a miRNA-drug association prediction model using route-guided dual attention and gated graph mask autoencoder, the problem of insufficient capture of local structure and global dependency in existing methods under sparse and noisy conditions is solved, and efficient and accurate miRNA-drug association prediction is achieved.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- CHINA UNIV OF PETROLEUM (EAST CHINA)
- Filing Date
- 2026-05-11
- Publication Date
- 2026-06-16
AI Technical Summary
Existing miRNA-drug association prediction methods struggle to simultaneously capture global dependencies and local structures under sparse and noisy conditions, suffer from high computational complexity, insufficient retention of local discriminative information, and inadequate prediction robustness.
We employ a route-guided dual attention and gated graph mask autoencoder to learn local structural representations by constructing a miRNA-drug heterogeneity graph and combining it with global dependency representations to suppress noise and irrelevant connections. We then use a multilayer perceptron for association prediction.
It improves the performance and robustness of miRNA-drug association prediction, enabling accurate prediction of potential associations under sparse and noisy biological data conditions.
Smart Images
Figure FT_1
Abstract
Description
Technical Field
[0001] This invention relates to the fields of biomedical informatics, artificial intelligence-assisted drug development, non-coding RNA drug response analysis, and biological association prediction, and particularly to a miRNA-drug association prediction method based on route-guided dual attention and gated graph mask autoencoder. Background Technology
[0002] MicroRNAs, or miRNAs, are an important class of non-coding RNAs that participate in various biological processes by regulating gene expression. Numerous studies have shown that miRNAs are closely related to drug response, drug sensitivity, drug tolerance, and therapeutic efficacy. Therefore, identifying potential miRNA-drug associations is crucial for elucidating drug mechanisms of action, discovering new therapeutic targets, optimizing drug treatment strategies, and assisting in drug development.
[0003] Traditional experimental methods can verify the true association between miRNAs and drugs, but experimental validation is usually costly, time-consuming, and has limited throughput, making it difficult to meet the needs of large-scale candidate association screening. With the development of biomedical databases and artificial intelligence technologies, miRNA-drug association prediction based on computational models has gradually become an important research direction.
[0004] Existing computational methods mainly include matrix factorization-based methods, graph neural network-based methods, attention mechanism-based methods, and heterogeneous network propagation-based methods. While these methods have achieved some progress, they still have the following shortcomings:
[0005] First, some methods focus on global dependency modeling, which can capture the relationships between distant nodes, but they are prone to introducing irrelevant connections, weakening local discriminative evidence, and increasing the computational complexity of the model.
[0006] Second, some methods focus on local neighborhood modeling, which can utilize direct neighbor relationships and local topology, but the receptive field is limited, making it difficult to fully capture long-range dependencies in miRNA-drug heterogeneous networks.
[0007] Third, it is known that miRNA-drug association data are usually sparse, and biological data contain noise, missing and incomplete annotations, which makes the model susceptible to noise perturbation and lacks prediction robustness.
[0008] Fourth, existing methods often struggle to simultaneously achieve global context modeling, local structure preservation, noise suppression, and efficient computation within a unified framework.
[0009] Therefore, there is an urgent need for a new miRNA-drug association prediction method that can simultaneously learn global dependencies and local discriminative structures under sparse and noisy biological data conditions, thereby improving the performance and robustness of miRNA-drug association prediction. Summary of the Invention
[0010] This invention aims to provide a miRNA-drug association prediction method based on route-guided dual attention and gated graph mask autoencoder, to address the problems of existing methods, such as difficulty in simultaneously capturing global dependencies and local structures under sparse and noisy conditions, high computational complexity, insufficient retention of local discriminative information, and insufficient prediction robustness. The technical solution of this invention is as follows:
[0011] To achieve the above objectives, this invention provides a method for predicting miRNA-drug associations. This method first acquires miRNA sequences, drug structures, known miRNA-drug associations, and related database information to construct a miRNA-drug association matrix. Then, it constructs miRNA similarity matrices and drug similarity matrices separately. For miRNAs, it constructs a miRNA sequence similarity matrix and a miRNA Gaussian interaction spectrum similarity matrix; for drugs, it constructs a drug structure similarity matrix and a drug Gaussian interaction spectrum similarity matrix. Subsequently, the multi-source similarity matrices are fused to obtain a comprehensive miRNA similarity matrix and a comprehensive drug similarity matrix.
[0012] Furthermore, for each miRNA and drug, initial feature vectors are constructed based on their comprehensive similarity to similar entities, and these miRNA and drug features are projected onto a unified latent space through linear mapping. On one hand, the mapped miRNA and drug features are used as node features, combined with known miRNA-drug associations to construct a miRNA-drug heterogeneity graph, which is then input into a gated graph mask autoencoder to learn local structural representations. On the other hand, the mapped features of candidate miRNA-drug pairs are organized into binary sequences and input into a routing-guided dual-attention module to learn global dependency representations.
[0013] The gated graph masking autoencoder first randomly masks the node features, then learns neighborhood information through graph aggregation layers, and sets a gating mechanism after each graph aggregation layer to adjust the propagation intensity of neighborhood information and suppress noise and irrelevant connections. Subsequently, the decoder reconstructs the masked node features, and the loss-constrained local structural representation is learned through masked feature reconstruction.
[0014] The routing-guided dual-attention module introduces a learnable routing matrix into the Transformer encoder, combining the expressive power of Softmax attention with the computational advantages of low-rank routing interactions. This module obtains routing value representations through routing-guided Softmax aggregation and generates global representations of candidate miRNA-drug pairs through low-rank interactions between the query matrix and the routing matrix.
[0015] Finally, the local representation obtained by the gated graph mask autoencoder and the global representation obtained by the routing-guided dual attention module are weighted and fused, and the fused representation is input into the multilayer perceptron prediction network to output the association probability score of candidate miRNA-drug pairs. Attached Figure Description
[0016] Figure 1 It is a miRNA-drug interaction prediction process based on route-guided dual attention and gated graph mask autoencoder; Detailed Implementation
[0017] To make the objectives, technical solutions, and advantages of this invention clearer, the miRNA-drug association prediction method of this invention will be further described in detail below with reference to the accompanying drawings. It should be understood that the following embodiments are for illustrative purposes only and are not intended to limit the scope of protection of this invention.
[0018] In the data preprocessing stage, miRNA sequence data, drug structure data, and known miRNA-drug association data from public databases are integrated to construct a miRNA-drug association matrix. This miRNA-drug association matrix indicates whether a known association exists between a miRNA and a drug, with a known association denoted as 1 and no known association denoted as 0.
[0019] In the multi-source similarity construction stage, miRNA similarity networks and drug similarity networks were constructed separately. For miRNAs, a miRNA sequence similarity network was constructed based on the edit distance between miRNA sequences; a miRNA Gaussian interaction spectrum similarity network was constructed based on the association contours of miRNAs in the miRNA-drug association matrix. For drugs, a drug structure similarity network was constructed based on the molecular fingerprint extracted from the drug SMILES representation and using Tanimoto coefficients; a drug Gaussian interaction spectrum similarity network was constructed based on the association contours of drugs in the miRNA-drug association matrix.
[0020] In the similarity fusion stage, the miRNA sequence similarity network and the miRNA Gaussian interaction spectrum similarity network are weighted and fused to obtain the miRNA comprehensive similarity network; similarly, the drug structure similarity network and the drug Gaussian interaction spectrum similarity network are weighted and fused to obtain the drug comprehensive similarity network. This step allows similarity information from different sources to complement each other, thereby improving the completeness and reliability of miRNA and drug feature representations.
[0021] In the unified feature mapping stage, based on the miRNA comprehensive similarity network, a similarity distribution vector between each miRNA and all miRNAs is constructed; based on the drug comprehensive similarity network, a similarity distribution vector between each drug and all drugs is constructed. Since the original dimensions of miRNA features and drug features are different, trainable linear mapping matrices are used to map miRNA features and drug features to a unified latent space of the same dimension.
[0022] In the heterogeneity graph construction phase, miRNAs and drugs are treated as different types of nodes, and known miRNA-drug associations are used as edges connecting miRNA and drug nodes to construct a miRNA-drug heterogeneity graph. The mapped miRNA and drug features are used as the initial node features of the heterogeneity graph for local structure representation learning.
[0023] In the local feature learning stage, the miRNA-drug heterogeneity graph is input into a gated graph mask autoencoder. First, a binary mask is randomly generated for the features of the heterogeneous graph nodes, and the features at the masked locations are set to invisible or zero, resulting in a corrupted node feature matrix. Then, the feature information of each node and its neighboring nodes is aggregated through a graph aggregation layer. Further, a gating mechanism is introduced after the graph aggregation layer, generating gating weights based on the aggregated node representation, and controlling the intensity of neighborhood information transmission through these gating weights. This approach preserves useful local structural information while suppressing interference from noisy neighbors and irrelevant connections on the node representation. Finally, the original node features are reconstructed using a decoder, and a locally robust representation is learned by constraining the model using cosine reconstruction loss at the mask locations.
[0024] In the global feature learning stage, for each candidate miRNA-drug pair, the corresponding latent miRNA features and latent drug features are combined into a binary input sequence and fed into a routing-guided dual attention module. This module first maps the input sequence into a query matrix, a key matrix, and a value matrix, while introducing a learnable routing matrix. The routing matrix interacts with the key matrix to generate routing attention weights, and then aggregates the value matrix to obtain the routing value representation. Subsequently, the query matrix and the routing matrix undergo a low-rank interaction to obtain the global attention output by overcoming the routing bottleneck. The multi-head routing attention output is then concatenated and linearly transformed, and finally pooled to obtain the global representation of the candidate miRNA-drug pair.
[0025] In the representation fusion stage, the local miRNA-drug pair representations learned by the gated graph mask autoencoder and the global miRNA-drug pair representations learned by the routing-guided dual attention module are weighted and fused. The weighted fusion coefficients are trainable parameters, enabling the model to adaptively adjust the contributions of local structural information and global dependency information based on the training data.
[0026] In the association prediction stage, the fused miRNA-drug pair representation is input into a multilayer perceptron prediction network. The multilayer perceptron includes an input layer, hidden layers, and an output layer. The hidden layers employ a non-linear activation function, and the output layer uses a sigmoid function to output association probability scores. For unknown miRNA-drug candidate pairs, the model sorts them from highest to lowest output score to obtain potential miRNA-drug association prediction results.
[0027] During model training, a joint optimization objective is employed. This joint optimization objective includes a miRNA-drug association prediction loss and a mask feature reconstruction loss from a gated graph mask autoencoder. The association prediction loss constrains the model to accurately predict known miRNA-drug associations, while the mask feature reconstruction loss constrains the model to recover the masked node features and learn robust local representations. By jointly optimizing the global attention module and the local graph reconstruction module, the model can simultaneously learn global contextual information and local structural information, thereby improving the miRNA-drug association prediction performance.
Claims
1. A miRNA-drug association prediction method based on route-guided dual attention and gated graph mask autoencoder, characterized in that, Includes the following steps: A. Obtain miRNA data, drug data, and known miRNA-drug association data, and construct a miRNA-drug association matrix; B. Construct a miRNA multi-source similarity matrix based on the miRNA data, construct a drug multi-source similarity matrix based on the drug data, and fuse the miRNA multi-source similarity matrix and the drug multi-source similarity matrix respectively to obtain a comprehensive miRNA similarity matrix and a comprehensive drug similarity matrix; C. Based on the miRNA comprehensive similarity matrix and the drug comprehensive similarity matrix, construct the initial feature vectors of miRNA and drug respectively, and map the initial feature vectors of miRNA and drug to a unified latent feature space; D. Based on the mapped miRNA features and drug features, construct a miRNA-drug isomorphism map, and use a gated graph mask autoencoder to perform local structural characterization learning on the miRNA-drug isomorphism map to obtain local miRNA-drug pair representations; E. For candidate miRNA-drug pairs, the mapped miRNA features and drug features are organized into binary input sequences, and a routing-guided dual attention mechanism is used to perform global dependency modeling to obtain a global miRNA-drug pair representation. F. Fuse the local miRNA-drug pair representation and the global miRNA-drug pair representation to obtain the fused miRNA-drug pair representation; G. Input the fused miRNA-drug pair representation into a multilayer perceptron prediction network and output the association probability score between miRNA and drug.
2. The method according to claim 1, characterized in that, In step B, the miRNA multi-source similarity matrix includes a miRNA sequence similarity matrix and a miRNA Gaussian interaction spectrum similarity matrix; wherein, the miRNA sequence similarity matrix is obtained by calculating the edit distance or Levenshtein distance between any two miRNA sequences, and the miRNA Gaussian interaction spectrum similarity matrix is calculated based on the association contour vector of miRNA in the miRNA-drug association matrix.
3. The method according to claim 1, characterized in that, In step B, the drug multi-source similarity matrix includes a drug structure similarity matrix and a drug Gaussian interaction spectrum similarity matrix; wherein, the drug structure similarity matrix is obtained by extracting molecular fingerprints through drug SMILES representation and calculating the structural similarity between any two drugs based on the Tanimoto coefficient, and the drug Gaussian interaction spectrum similarity matrix is calculated based on the drug association contour vector in the miRNA-drug association matrix.
4. The method according to claim 1, characterized in that, In steps C and D, for any miRNA, an initial miRNA feature vector is constructed based on its similarity to all miRNAs; for any drug, an initial drug feature vector is constructed based on its similarity to all drugs; and the initial miRNA feature vector and the initial drug feature vector are mapped to a unified latent feature space of the same dimension through independent linear mapping matrices; the miRNA-drug heterogeneity graph includes miRNA nodes, drug nodes, and edges formed by known miRNA-drug associations.
5. The method according to claim 1, characterized in that, In step D, the gated graph mask autoencoder includes a feature mask module, a graph encoder, a gated message passing module, a graph decoder, and a mask feature reconstruction loss module. The feature mask module is used to randomly generate a binary mask matrix for the node features in the miRNA-drug heterogeneity graph, and obtain the destroyed node feature matrix based on the binary mask matrix. The graph encoder includes at least one GraphSAGE aggregation layer for aggregating the local structural information of the target node and its neighboring nodes. The gated message passing module is used to generate gate weights for each layer of graph aggregation output, and use the gate weights to adjust the propagation intensity of local neighborhood information. The graph decoder is used to reconstruct the original node features based on the latent node representation output by the graph encoder, and calculate the mask feature reconstruction loss based on the reconstruction error at the mask position.
6. The method according to claim 1, characterized in that, In step E, the routing-guided dual attention mechanism includes a routing-guided Softmax aggregation stage and a routing-guided low-rank interaction stage. The routing-guided dual attention mechanism introduces a learnable routing matrix, uses this matrix to interact with a key matrix to obtain routing value representations, and uses a query matrix to interact with the learnable routing matrix, completing low-rank attention calculations through routing bottlenecks. The routing-guided dual attention mechanism is set in the Transformer encoder and obtains global interaction representations of candidate miRNA-drug pairs through multi-head attention.
7. The method according to claim 1, characterized in that, In steps F and G, the local miRNA-drug pair representation and the global miRNA-drug pair representation are weighted and fused using trainable weight parameters; the multilayer perceptron prediction network includes an input layer, at least one hidden layer, and an output layer, wherein the hidden layer uses the ReLU activation function, and the output layer uses the sigmoid function to output the miRNA-drug association probability score; the model training objectives include association prediction loss and mask feature reconstruction loss, wherein the association prediction loss includes a weighted combination of mean squared error loss and mean absolute error loss, and the mask feature reconstruction loss is a cosine reconstruction loss based on mask features.
8. A miRNA-drug association prediction system based on route-guided dual attention and gated graph mask autoencoder, characterized in that, include: The data construction module is used to acquire miRNA data, drug data, and known miRNA-drug association data, and to construct a miRNA-drug association matrix; The similarity building module is used to construct miRNA multi-source similarity matrices and drug multi-source similarity matrices; A similarity fusion module is used to fuse the miRNA multi-source similarity matrix and the drug multi-source similarity matrix to obtain a comprehensive miRNA similarity matrix and a comprehensive drug similarity matrix; The unified feature mapping module is used to map miRNA features and drug features to a unified latent feature space; The local structure learning module is used to learn the local structure representation in miRNA-drug isomorphism graphs based on gated graph mask autoencoders. The global dependency learning module is used to learn the global interaction representation of candidate miRNA-drug pairs based on a route-guided dual attention mechanism; The feature fusion module is used to fuse local structural representations and global interaction representations; The association prediction module is used to output the miRNA-drug association probability score.