A miRNA-disease association prediction method based on cross-attention and confidence fusion

Abstract

The application relates to a miRNA-disease association prediction method based on cross-attention and confidence fusion, which comprises the following steps: acquiring a miRNA-disease database and constructing a heterogeneous graph; encoding miRNA nodes and disease nodes in the heterogeneous graph to obtain miRNA and disease node embedding representations; introducing a cross-attention mechanism between miRNA and disease nodes according to the miRNA and disease node embedding representations to realize bidirectional interaction and obtain an interaction result; constructing multiple channels to obtain miRNA-disease pair features and channel-attached confidence; weighting and fusing the miRNA-disease pair features according to the channel-attached confidence to obtain fused miRNA-disease pair features; inputting the fused miRNA-disease pair features into a multilayer perception machine to output probability values of the miRNA-disease association. The application can realize feature extraction and prediction of the relationship between miRNA and diseases, and can introduce confidence estimation in the prediction process to improve the reliability of the result.

Description

Technical Field

[0001] This invention relates to the field of bioinformatics and artificial intelligence, and in particular to a miRNA-disease association prediction method based on cross-attention and confidence fusion. Background Technology

[0002] MicroRNAs (miRNAs) are a class of small, non-coding RNA molecules, approximately 22 nucleotides in length, that play a crucial role in post-transcriptional gene regulation. Numerous studies have shown that aberrant miRNA expression is closely associated with a variety of complex diseases, including cancer, cardiovascular disease, metabolic disorders, and neurological diseases. Therefore, systematically identifying the potential associations between miRNAs and diseases is of great significance for elucidating the molecular mechanisms of disease development, discovering biomarkers for early diagnosis, and developing new therapeutic strategies.

[0003] Currently, experimental methods are the main approach to studying miRNA-disease associations, such as high-throughput sequencing, microarray hybridization, and functional validation experiments. While these methods can provide direct evidence, they suffer from high costs, long cycles, and limited throughput, making it difficult to achieve comprehensive coverage within a large candidate space. To overcome these limitations, researchers have gradually developed a series of computational prediction methods to screen and optimize candidate miRNA-disease pairs before experiments.

[0004] Early computational methods were largely based on similarity and network propagation concepts. A typical approach involved calculating functional similarity between miRNAs and semantic similarity between diseases, combined with known association information, and using strategies such as matrix completion, random walks, or network propagation to infer new potential associations. These methods are simple to implement, but heavily rely on the accuracy of prior similarity measurements. Their predictive performance degrades significantly when data is sparse or noisy, and they struggle to handle "cold-start" entities lacking known associations.

[0005] As research has progressed, machine learning-based feature engineering methods have emerged. These methods extract sequence features, secondary structure features, and semantic or clinical features of miRNAs, and combine them with manually constructed network features to train classifiers such as support vector machines and random forests to make predictions. Although they improve prediction performance to some extent, they rely on manually designed features, making it difficult to capture complex nonlinear patterns and high-order interactions, and their generalization ability and scalability are limited.

[0006] In recent years, deep learning techniques, especially graph neural networks (GNNs), have been increasingly applied to miRNA-disease prediction tasks. These methods model miRNAs and diseases as different types of nodes in a heterogeneous graph, learning node representations through graph convolution or attention mechanisms, and further using them to predict potential associations. Compared to traditional methods, these models can automatically extract features from data, reducing reliance on manual feature engineering. However, existing graph neural network methods primarily focus on feature aggregation in local neighborhoods, making it difficult to capture global dependencies between distant nodes. Furthermore, they typically only output prediction scores, lacking a quantitative assessment of the reliability of the results, which is detrimental to researchers' reliable screening and decision-making in practical applications.

[0007] In summary, existing technologies still have significant shortcomings in miRNA-disease association prediction tasks, mainly manifested in limited cross-modal interaction modeling capabilities, insufficient capture of global dependencies, and a lack of credibility assessment for prediction results. These shortcomings largely limit the application effectiveness of existing methods in complex biological networks and also make it difficult for computational prediction to fully play its auxiliary role in disease research and clinical translation. Summary of the Invention

[0008] The purpose of this invention is to provide a miRNA-disease association prediction method based on cross-attention and confidence fusion, which can realize feature extraction and prediction of the relationship between miRNA and disease, and introduce confidence estimation in the prediction process to improve the reliability of the results.

[0009] To achieve the above objectives, the present invention provides the following solution:

[0010] A miRNA-disease association prediction method based on attention and confidence fusion includes:

[0011] Obtain the miRNA-disease database and construct a heterogeneity graph;

[0012] The miRNA nodes and disease nodes in the heterogeneous graph are encoded to obtain the miRNA and disease node embedding representations.

[0013] Based on the miRNA and disease node embedding representation, a cross-attention mechanism is introduced between the miRNA and the disease node to perform bidirectional interaction and obtain the interaction results, wherein the interaction results include miRNA representation under disease context enhancement and disease representation under miRNA context enhancement.

[0014] The interaction results are fused into the original representation through a gated residual mechanism to obtain the final representation; multiple channels are constructed to obtain miRNA-disease pair features and channel-related confidence scores, wherein the channels include C1 and C2 channels based on the final representation, C3 channel based on heterogeneous graph encoding, and C4 channel based on basic feature operations;

[0015] The miRNA-disease pair features are weighted and fused according to the confidence level attached to the channel to obtain the fused miRNA-disease pair features;

[0016] The fused miRNA-disease pair features are input into a multilayer perceptron, which outputs the probability values ​​of miRNA association with disease.

[0017] Optionally, constructing the heterogeneous graph includes:

[0018] Based on the miRNA-disease database, known association information between miRNAs and diseases is obtained, and the heterogeneous graph is constructed by combining the similarity between miRNAs and between diseases. The heterogeneous graph includes miRNA nodes, disease nodes, miRNA internal edges, disease internal edges, and miRNA-disease association edges. The feature representation of the miRNA node is obtained by obtaining the functional information of the miRNA node and encoding the biological semantics related to the target gene. The feature representation of the disease node is obtained by encoding based on medical semantic information or ontology description.

[0019] Optionally, encoding the miRNA nodes and disease nodes in the heterogeneous graph includes:

[0020] An improved GraphSAGE encoder is used to encode the features of the miRNA node and the disease node. The improved GraphSAGE encoder obtains three types of aggregation results through a mean aggregator, a pool aggregator, and an LSTM aggregator. The three types of aggregation results are concatenated, and the concatenated aggregation results are processed by linear transformation, nonlinear activation function, and Dropout fusion.

[0021] Optionally, introducing cross-attention mechanisms for bidirectional interaction between miRNAs and disease nodes includes:

[0022] In the miRNA-to-disease direction, the miRNA embedding representation is used as the query and the disease embedding representation is used as the key and value. Weights are calculated and aggregated through an attention mechanism to obtain the miRNA representation under the disease context enhancement.

[0023] In the disease-to-miRNA direction, the disease embedding representation is used as the query and the miRNA embedding representation is used as the key and value. Weights are calculated and aggregated through an attention mechanism to obtain the disease representation under miRNA context enhancement.

[0024] Optionally, obtaining the miRNA representation enhanced in the disease context includes:

[0025] ;

[0026] = ;

[0027] in, For miRNA representation enhanced in the context of disease, This represents the value vector of disease node j. Let represent the attention weight of miRNA node i to disease node j, where i represents the i-th miRNA node and j represents the j-th disease node. Let be the query vector corresponding to the i-th miRNA node. Let represent the key vector corresponding to the j-th disease node, and w represent the dimension of each attention head. , , ,in , For learnable weights during the training process, and These are the miRNA node embedding representation and the disease node embedding representation output by the encoder, respectively.

[0028] Optionally, obtaining the final representation includes: fusing the miRNA representation enhanced in the disease context into the original representation using a gated residual mechanism.

[0029] ;

[0030] in, This represents the final representation of the miRNA after gated residual fusion. Indicates the original miRNA embedding; This represents the gating coefficient vector corresponding to the miRNA; ⊙ indicates element-wise multiplication;

[0031] The process of fusing the disease representation enhanced with miRNA context into the original representation via a gated residual mechanism includes:

[0032] ;

[0033] in, It is the final representation of the disease after gated residual fusion. Indicates the original embedding of the disease; This represents the gating coefficient vector corresponding to the disease.

[0034] Optionally, multiple channels can be constructed to obtain miRNA-disease pair characteristics and channel-related confidence scores, including:

[0035] The C1 channel splices the final representation of the gated residual fused miRNA with the disease node embedding representation to construct the miRNA-disease pair feature, and uses the entropy value of the attention distribution to calculate the confidence of the C1 channel;

[0036] The C2 channel concatenates the final representation of the disease after gating residual fusion with the miRNA node embedding representation to construct the miRNA-disease pair feature, and uses the entropy value of the attention distribution to calculate the confidence of the C2 channel;

[0037] The C3 channel concatenates the miRNA with the disease embedding representation and inputs it into the encoder to obtain the miRNA-disease pair features. MC-Dropout is used for multiple samplings, and variance transformation is used to obtain the confidence level of the C3 channel.

[0038] The C4 channel performs basic feature operations on the miRNA and disease embedding representation to generate the miRNA-disease pair features, and sets the confidence level of the C4 channel to 1.

[0039] Optionally, weighted fusion of the miRNA-disease pair features based on the channel-attached confidence level includes:

[0040] = ;

[0041] in, This indicates the characteristics of the fused miRNA-disease pair. This represents the miRNA-disease pair characteristics of the k-th channel; This represents the confidence level of the k-th channel.

[0042] The beneficial effects of this invention are as follows:

[0043] 1. Bidirectional cross-attention mechanism enhances interaction modeling capabilities:

[0044] This invention introduces a bidirectional cross-attention mechanism between miRNAs and disease nodes, unlike traditional methods that can only capture feature relationships unidirectionally. This method can simultaneously achieve bidirectional interactive modeling of "miRNA→disease" and "disease→miRNA," allowing the representations of the two types of nodes to be dynamically updated through interaction. Through the design of a gated residual structure, original node features are preserved, while contextual information is injected in a controlled manner, avoiding excessive information perturbation. This mechanism ensures that the model not only captures global dependencies but also improves feature expressiveness while maintaining stability.

[0045] 2. Multi-channel fusion enhances the richness of pair representation:

[0046] The multi-channel fusion strategy proposed in this invention has significant advantages over existing technologies. Four channels construct feature representations of the miRNA-disease pair from different perspectives: C1 and C2 capture interaction information based on bidirectional cross-attention results, C3 preserves global topological and semantic features through a heterogeneous graph encoder, and C4 extracts simple and direct contrastive features from basic vector interactions. The design of these four channels is both independent and complementary, forming a multi-level, multi-faceted feature representation system. Through this design, the model avoids the limitations of relying on a single feature representation, effectively improving the diversity and expressive power of pair representations, thereby exhibiting stronger adaptability in complex biological networks.

[0047] 3. Introducing confidence estimation improves the stability and reliability of predictions:

[0048] This invention introduces a confidence calculation method based on entropy and variance during feature fusion, innovatively combining uncertainty estimation with a channel weighting mechanism. In channels C1 and C2, the entropy of the attention distribution is used to measure the reliability of the representation; lower entropy values ​​indicate that the model focuses more intently on a few key nodes, resulting in higher confidence. In channel C3, variance is calculated through multiple MC-Dropout samplings; smaller variance means the model is more stable in representing that pair, thus assigning higher confidence. The confidence scores of different channels are ultimately used as weighting factors in the fusion process, ensuring that high-confidence channels contribute more and low-confidence channels have a weaker impact. This mechanism improves the stability and reliability of prediction results under different data distributions and reduces the risk of misjudgment due to the uncertainty of a single channel.

[0049] 4. Provide interpretable results to enhance the value of scientific research applications:

[0050] This invention not only outputs the predicted probability of miRNA-disease associations but also the confidence scores for each channel. This design provides additional reference information for the prediction results, enabling researchers to assess the reliability and potential risks of the results based on the channel confidence scores. Unlike traditional black-box models that only provide a single score, this method can explain the information sources behind the predictions, i.e., whether a prediction mainly comes from cross-attention interactions, heterogeneous graph structure information, or support from basic features. This interpretability greatly enhances the model's application value in scientific research and clinical practice, helps improve the efficiency of experimental validation and the reliability of results, and provides more powerful tool support for subsequent miRNA functional research and disease mechanism exploration. Attached Figure Description

[0051] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0052] Figure 1 This is a flowchart of a miRNA-disease association prediction method based on cross-attention and confidence fusion according to an embodiment of the present invention;

[0053] Figure 2 The graph shows the prediction performance metrics of the MDA task in this embodiment of the invention on the HMDDv3.2 and HMDDv2.0 datasets.

[0054] Figure 3 This is a box plot showing the experimental verification of the five-fold crossover algorithm in an embodiment of the present invention. Detailed Implementation

[0055] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0056] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0057] This embodiment proposes a miRNA-disease association prediction method based on the fusion of attention and confidence, including:

[0058] Obtain the miRNA-disease database and construct a heterogeneity graph;

[0059] The miRNA nodes and disease nodes in the heterogeneous graph are encoded to obtain the miRNA and disease node embedding representations.

[0060] Based on the miRNA and disease node embedding representation, a cross-attention mechanism is introduced between the miRNA and the disease node to perform bidirectional interaction and obtain the interaction results, wherein the interaction results include miRNA representation under disease context enhancement and disease representation under miRNA context enhancement.

[0061] The interaction results are fused into the original representation through a gated residual mechanism to obtain the final representation; multiple channels are constructed to obtain miRNA-disease pair features and channel-related confidence scores, wherein the channels include C1 and C2 channels based on the final representation, C3 channel based on heterogeneous graph encoding, and C4 channel based on basic feature operations;

[0062] The miRNA-disease pair features are weighted and fused according to the confidence level attached to the channel to obtain the fused miRNA-disease pair features;

[0063] The fused miRNA-disease pair features are input into a multilayer perceptron, which outputs the probability values ​​of miRNA association with disease.

[0064] Furthermore, constructing the heterogeneous graph includes:

[0065] Based on the miRNA-disease database, known association information between miRNAs and diseases is obtained. Combining the similarity between miRNAs and between diseases, a heterogeneous graph is constructed. The heterogeneous graph includes miRNA nodes, disease nodes, internal edges of miRNAs, internal edges of diseases, and miRNA-disease association edges. The feature representation of miRNA nodes is obtained by obtaining the functional information of miRNA nodes and encoding the biological semantics related to target genes. The feature representation of disease nodes is obtained by encoding based on medical semantic information or ontology description.

[0066] Specifically, data preparation and graph structure construction include: First, collecting known miRNA-disease association information from public databases, and combining miRNA-miRNA similarity and disease-disease similarity to construct a heterogeneous graph. The input data can be formalized as: G=( ∪ , ∪ ∪ ).in, Represents a set of miRNA nodes. Represents a set of disease nodes. , , These represent miRNA-internal edges, disease-internal edges, and miRNA-disease association edges, respectively. For each miRNA node, the input feature vector can be derived from sequence information, functional annotations, or other prior computations; for each disease node, the input feature vector can be encoded based on medical semantic information or ontology descriptions. In this way, the data processing module outputs a graph structure containing node features and cross-modal edge relationships, laying the foundation for subsequent deep modeling.

[0067] Furthermore, encoding the miRNA nodes and disease nodes in the heterogeneous graph includes:

[0068] An improved GraphSAGE encoder is used to encode the features of miRNA nodes and disease nodes. The improved GraphSAGE encoder obtains three types of aggregation results through mean aggregator, pool aggregator and LSTM aggregator. The three types of aggregation results are concatenated and processed by linear transformation, nonlinear activation function and dropout fusion.

[0069] Specifically, high-quality representations are generated for miRNAs and disease nodes. An improved GraphSAGE encoder is employed, introducing three aggregation methods in parallel:

[0070] mean aggregation: calculates the mean of neighboring features;

[0071] pool aggregation: The pooling result that takes the features of neighbors after a non-linear transformation;

[0072] LSTM aggregation: learns the serialization information of neighbors through a loop structure.

[0073] After concatenating the three types of aggregation results, a stable and information-rich node embedding representation is formed through linear transformation, activation function, and Dropout fusion.

[0074] Furthermore, introducing cross-attention mechanisms for bidirectional interaction between miRNAs and disease nodes includes:

[0075] In the miRNA-to-disease direction, the miRNA embedding representation is used as the query and the disease embedding representation is used as the key and value. Weights are calculated and aggregated through an attention mechanism to obtain the miRNA representation under the disease context enhancement.

[0076] In the disease-to-miRNA direction, the disease embedding representation is used as the query and the miRNA embedding representation is used as the key and value. Weights are calculated and aggregated through an attention mechanism to obtain the disease representation under miRNA context enhancement.

[0077] Specifically, bidirectional cross-attention interaction aims to capture the bidirectional global interaction between miRNAs and diseases. After obtaining the miRNA and disease node embeddings, a bidirectional cross-attention mechanism is introduced:

[0078] (1) In the miRNA→disease direction, the attention weights are calculated using the miRNA node embedding as the query (Q) and the disease node embedding as the key (K) and value (V):

[0079] = ;

[0080] Where i represents the i-th miRNA node; j represents the j-th disease node; : The query vector corresponding to the i-th miRNA node; Let w represent the key vector corresponding to the j-th disease node, and w represent the dimension of each attention head, used for scaling to prevent the value from being too large; This represents the attention weight of miRNA node i to disease node j; , , ,in , For learnable weights during the training process, and These are the miRNA node embedding representation and the disease node embedding representation output by the encoder, respectively.

[0081] Then, by weighted summation, the miRNA representation enhanced in the disease context is obtained:

[0082] ;

[0083] in, This represents the value vector of disease node j.

[0084] (2) In the disease → miRNA direction, the process is similar, except that the roles of query and key value are swapped.

[0085] The attention weights are calculated using disease node embeddings as queries (Q) and miRNA node embeddings as keys (K) and values ​​(V):

[0086] = ;

[0087] : The query vector corresponding to the j-th disease node; represents the key vector corresponding to the i-th miRNA node, and w represents the dimension of each attention head, used for scaling to prevent the value from being too large; This represents the attention weight of disease node j on miRNA node i.

[0088] The disease representation enhanced by miRNA context is obtained by weighted summation:

[0089] ;

[0090] in, This represents the value vector of miRNA node i.

[0091] (3) The final interaction result is fused into the original representation through a gated residual mechanism, that is, while preserving the original features, contextual information is selectively injected. This not only captures global dependencies but also avoids excessive interference with the original node features. The formula is as follows:

[0092] ;

[0093] in Indicates the original miRNA embedding; This represents the gating coefficient vector corresponding to the miRNA, with values ​​in [0,1], controlling the context injection ratio; ⊙ represents element-wise multiplication. It is the final representation of miRNA after gated residual fusion.

[0094] ;

[0095] in, It is the final representation of the disease after gated residual fusion. Indicates the original embedding of the disease; This represents the gating coefficient vector corresponding to the disease.

[0096] Furthermore, multiple channels were constructed to obtain miRNA-disease pair characteristics and channel-related confidence scores, including:

[0097] The C1 channel splices the final representation of the gated residual fused miRNA with the disease node embedding representation to construct the miRNA-disease pair feature, and uses the entropy value of the attention distribution to calculate the confidence of the C1 channel;

[0098] The C2 channel concatenates the final representation of the disease after gating residual fusion with the miRNA node embedding representation to construct the miRNA-disease pair feature, and uses the entropy value of the attention distribution to calculate the confidence of the C2 channel;

[0099] The C3 channel concatenates the miRNA with the disease embedding representation and inputs it into the encoder to obtain the miRNA-disease pair features. MC-Dropout is used for multiple samplings, and variance transformation is used to obtain the confidence level of the C3 channel.

[0100] The C4 channel performs basic feature operations on the miRNA and disease embedding representation to generate the miRNA-disease pair features, and the confidence level of the C4 channel is set to 1 by default.

[0101] Specifically, MDA-BiTri is designed with four channels to construct the representation of miRNA-disease pairs, i.e., miRNA-disease pair features, with each channel accompanied by a confidence level. MDA-BiTri consists of three key components: data processing, the MDA-BiTri model (composed of an encoder, an interaction module, and a multi-channel fusion module), and prediction output.

[0102] C1 channel (miRNA→disease): Construct pair features based on the miRNA→disease interaction representation, and calculate confidence based on the entropy of the attention distribution.

[0103] =1- ,

[0104] Where H(·): information entropy, representing the uncertainty of attention distribution; Indicates the number of disease nodes; The confidence level of the C1 channel is as close to 1 as it is more reliable.

[0105] The pair feature based on miRNA-disease interaction representation is obtained by splicing the final miRNA representation after gating residual fusion with the disease node embedding representation one by one.

[0106] C2 channel (disease → miRNA: symmetric to C1, constructing pair features based on the interaction representation of disease → miRNA, and calculating confidence based on the entropy of the attention distribution):

[0107] =1- ,

[0108] in : Number of miRNA nodes.

[0109] The pair feature based on disease → miRNA representation is obtained by splicing the final representation of the disease after gating residual fusion with the miRNA disease node embedding representation one by one.

[0110] C3 channel (Heterogeneous Graph Encoding): This channel directly generates paired features using a heterogeneous graph encoder, specifically an improved GraphSAGE encoder, and employs multiple random samplings (MC-Dropout) during the training phase, using variance as a measure of uncertainty.

[0111] = ;

[0112] Where Var(c3) represents the feature variance obtained from multiple samplings of the same pair; a large variance indicates a low confidence level, and vice versa, a small variance indicates a high confidence level.

[0113] C4 channel (basic features): Pair features are generated using simple methods such as concatenation, differencing, and Hadamard product, with a fixed confidence level. =1.

[0114] The final outputs of the four channels are weighted and fused according to confidence level:

[0115] = ;

[0116] in This represents the pair feature of the k-th channel; This represents the confidence level of the k-th channel.

[0117] The fused pair represents the input to a multilayer perceptron (MLP) and the output probability of miRNA-disease association:

[0118] score=σ(W· ),

[0119] Where W,b represents the weights and bias parameters of the MLP; The pair representation vector obtained by fusion; σ represents the Sigmoid function, ensuring the output is in [0,1]; score: the potential association probability between miRNAs and diseases.

[0120] At the same time, the model will also output the values ​​of each channel. , , , Used as a reference for prediction reliability.

[0121] The method of this embodiment will be further described below with reference to the accompanying drawings:

[0122] This embodiment integrates graph neural networks, bidirectional cross-attention mechanisms, and multi-channel fusion strategies, enabling effective modeling of the global interaction between miRNAs and diseases in heterogeneous biological networks. MDA-BiTri comprises three key components: data processing, the MDA-BiTri model (consisting of an encoder, an interaction module, and a multi-channel fusion module), and prediction output.

[0123] like Figure 1 As shown, a miRNA-disease association prediction method based on cross-attention and confidence fusion includes:

[0124] 1. Dataset preparation:

[0125] This embodiment uses the Human MicroRNA Disease Database (HMDDv3.2) as the experimental data source. The HMDD database contains a large amount of experimentally validated association information between miRNAs and diseases, covering various disease types such as cancer, cardiovascular diseases, and neurological diseases. It is one of the most commonly used authoritative benchmark datasets for miRNA-disease association prediction tasks. The data preprocessing stage specifically includes the following operations:

[0126] (1) miRNA feature representation: The initial features of miRNA nodes are derived from functional information, etc. The miRNA feature representation is obtained by encoding the biological semantics related to its target genes, so that the model can capture local sequence patterns and global semantic information at the same time.

[0127] (2) Disease feature representation: Based on the Medical Subject Headings (MeSH) ontology, disease semantic vectors are constructed. By calculating the semantic similarity between diseases and performing low-dimensional embedding, semantically consistent disease feature representations are obtained.

[0128] (3) Heterogeneous graph construction: miRNA nodes and disease nodes are uniformly incorporated into the heterogeneous graph. The edges include three types: miRNA-disease known association edges; miRNA-miRNA similarity edges (calculated based on sequence or functional information); and disease-disease similarity edges (calculated based on MeSH semantic similarity). The resulting heterogeneous graph contains node attributes and cross-modal relationships, which can reflect both the intrinsic characteristics of biological entities and their interactions.

[0129] 2. Model Training:

[0130] This embodiment uses the MDA-BiTri model proposed in this invention, and the main steps are as follows:

[0131] (1) Encoding module: The miRNA and the node features of the disease are input into the encoder. The encoder adopts the GraphSAGE three-aggregator structure (mean, pool, lstm). The mean aggregator extracts the statistical average information of the neighborhood features; the pool aggregator can highlight the most significant features in the neighborhood; and the lstm aggregator models the order dependency between neighboring nodes.

[0132] After concatenating the three types of aggregation results, a linear transformation and a nonlinear activation function (such as ELU) are applied, followed by Dropout to prevent overfitting, ultimately resulting in a stable and rich node embedding representation.

[0133] (2) Interaction module: A bidirectional cross-attention mechanism is introduced based on the encoding results.

[0134] In the miRNA→disease direction, the miRNA node embedding representation is used as the query (Q), and the disease node embedding representation is used as the key (K) and value (V). The weights are calculated and aggregated through the attention mechanism to obtain the miRNA representation under the disease context enhancement.

[0135] In the disease→miRNA direction, the roles are reversed, resulting in a disease representation enhanced by the miRNA context.

[0136] The attention calculation results are fused into the original representation through a gating residual mechanism, which preserves the node's own information while selectively introducing contextual information to avoid excessive information interference.

[0137] (3) Fusion module: Four complementary channels were designed to generate pair (miRNA-disease pair) representations:

[0138] Channel C1: Constructs paired features based on the miRNA→disease interaction results, and calculates confidence using the entropy value of the attention distribution; Channel C2: Constructs paired features based on the disease→miRNA interaction results, and similarly estimates confidence using entropy; Channel C3: Concatenates miRNA and disease node embeddings and inputs them into a heterogeneous graph paired encoder, then performs multiple samplings using MC-Dropout. The variance reflects the uncertainty, and the confidence is then calculated; Channel C4: Generates auxiliary features by concatenating miRNA and disease node embeddings, using difference and Hadamard product methods, etc., with a fixed confidence of 1. Finally, the output features of the four channels are weighted according to confidence during fusion, with high-confidence features contributing more and low-confidence features contributing less, thus obtaining a robust paired representation.

[0139] (4) Prediction module: The fused pair represents the input to the multilayer perceptron (MLP) classifier, and the final output is the predicted probability of miRNA-disease association. The output range is [0,1], and the higher the probability, the higher the confidence of the potential association between the miRNA and the disease.

[0140] 3. Experimental Results and Analysis:

[0141] (1) On the HMDD dataset, this embodiment uses five-fold cross-validation for performance evaluation. The evaluation metrics include: AUC (area under the ROC curve); AUPR (area under the PR curve). For example... Figure 3The diagram shows the AUC (Area Under Curve) distribution of the model in 5-Fold Cross-Validation, used to evaluate the performance stability of the model at different training folds. The horizontal axis (Train) represents each fold in the 5-Fold Cross-Validation (5fCV_1 ~ 5fCV_5). The vertical axis (AUC) represents the AUC value of the model on the training set for each fold, ranging from approximately 0.92 to 0.98. The gray box represents the distribution range of the AUC values ​​for that fold (box plot), with the thick horizontal line in the middle representing the median, and the upper and lower boundaries of the box representing the upper and lower quartiles (Q1, Q3). Whiskers represent the range of non-outliers; values ​​exceeding this range are outliers. The gray scatter plot represents the specific AUC value obtained for each training iteration. The AUC value is a primary indicator of model accuracy; a more concentrated and larger value is better.

[0142] Depend on Figure 3 It can be seen that: (1) The median AUC of each fold is above 0.96, indicating that the model has high classification performance under different data partitions. (2) The box height (interquartile range) of each fold is relatively small, indicating that the performance fluctuation is small and the model is stable. (3) A small number of low outliers (below 0.94) may correspond to unstable performance in some training batches, but the overall impact is not significant. (4) The AUC of the 4th fold (5fCV_4) is slightly higher and more concentrated, indicating that the training effect of this fold is the best.

[0143] Experimental results show that the proposed method achieves AUC 0.97 and AUPR 0.97 on HMDD3.2. Even with a high proportion of negative samples, the proposed method still maintains stable performance, and the introduction of confidence-weighted multi-channel fusion effectively reduces the false positive rate.

[0144] like Figure 2 As shown, the horizontal axis represents six commonly used evaluation metrics for classification tasks, including: Area Under the Curve (AUC): measures the model's ability to distinguish between positive and negative samples. Area Under the Precision-Recall Curve (AUPR): focuses more on the prediction quality of positive samples. Accuracy: the proportion of correctly predicted samples out of the total number of samples. F1 score: the harmonic mean of precision and recall. Recall: the proportion of positive samples that are correctly identified. Precision: the proportion of truly positive samples out of those predicted as positive. The vertical axis represents the numerical values ​​of each metric (ranging from 0 to 1), with higher values ​​indicating better model performance. Figure 2The markers in the graph are: ○ circle: HMDD v2.0 dataset results; △ triangle: HMDD v3.2 dataset results; data labels: indicating the specific values ​​of each metric on the corresponding dataset. Interpreting the curves, overall, HMDDv3.2 slightly outperforms HMDDv2.0 on almost all metrics, especially showing a more significant advantage in AUC, AUPR, and Recall; Accuracy, F1 score, and Precision are also slightly higher, but the differences are relatively small. This indicates that the model's overall prediction performance on the v3.2 dataset is better than that on the v2.0 dataset, but the trends of both are basically consistent, indicating high accuracy.

[0145] (2) The ablation experiment further verified the importance of each component:

[0146] ① After removing the bidirectional cross-attention mechanism, the model's ability to capture cross-modal relationships decreases, with the AUC decreasing by an average of about 4%;

[0147] ②After removing the confidence mechanism, the model's performance in predicting small sample diseases became unstable, with AUPR decreasing by approximately 6%;

[0148] ③ After removing the C3 channel (MC-Dropout uncertainty modeling), the overall prediction accuracy decreased, and the prediction performance for highly heterogeneous diseases fluctuated significantly.

[0149] These experimental results demonstrate that this method effectively captures the global interaction between miRNA and disease through bidirectional cross-attention, and improves the robustness and reliability of prediction results through a confidence mechanism.

[0150] The embodiments described above are merely preferred embodiments of the present invention and are not intended to limit the scope of the present invention. Various modifications and improvements made to the technical solutions of the present invention by those skilled in the art without departing from the spirit of the present invention should fall within the protection scope defined by the claims of the present invention.

Claims

1. A miRNA-disease association prediction method based on cross-attention and confidence fusion, characterized in that, include: Obtain the miRNA-disease database and construct a heterogeneity graph; The miRNA nodes and disease nodes in the heterogeneous graph are encoded to obtain the miRNA and disease node embedding representations. Based on the miRNA and disease node embedding representation, a cross-attention mechanism is introduced between the miRNA and the disease node to perform bidirectional interaction and obtain the interaction results, wherein the interaction results include miRNA representation under disease context enhancement and disease representation under miRNA context enhancement. The interaction results are fused into the original representation through a gated residual mechanism to obtain the final representation; multiple channels are constructed to obtain miRNA-disease pair features and channel-related confidence scores, wherein the channels include C1 and C2 channels based on the final representation, C3 channel based on heterogeneous graph encoding, and C4 channel based on basic feature operations; Obtaining the final representation includes: fusing the miRNA representation enhanced under the disease context into the original representation through a gated residual mechanism. ； in, This represents the final representation of the miRNA after gated residual fusion. Indicates the original miRNA embedding; This represents the gating coefficient vector corresponding to the miRNA; ⊙ indicates element-wise multiplication; miRNA representation enhanced in the context of disease; The process of fusing the disease representation enhanced with miRNA context into the original representation via a gated residual mechanism includes: ； in, It is the final representation of the disease after gated residual fusion. Indicates the original embedding of the disease; This represents the gating coefficient vector corresponding to the disease. For disease representation enhanced in the miRNA context, ⊙ indicates element-wise multiplication; Multiple channels were constructed to obtain miRNA-disease pair features and channel-related confidence scores, including: The C1 channel splices the final representation of the gated residual fused miRNA with the disease node embedding representation to construct the miRNA-disease pair feature, and uses the entropy value of the attention distribution to calculate the confidence of the C1 channel; The C2 channel concatenates the final representation of the disease after gating residual fusion with the miRNA node embedding representation to construct the miRNA-disease pair feature, and uses the entropy value of the attention distribution to calculate the confidence of the C2 channel; The C3 channel concatenates the miRNA with the disease embedding representation and inputs it into the encoder to obtain the miRNA-disease pair features. MC-Dropout is used for multiple samplings, and variance transformation is used to obtain the confidence level of the C3 channel. The C4 channel performs basic feature operations on the miRNA and disease embedding representation to generate the miRNA-disease pair features, and sets the confidence level of the C4 channel to 1; The miRNA-disease pair features are weighted and fused according to the confidence level attached to the channel to obtain the fused miRNA-disease pair features; The fused miRNA-disease pair features are input into a multilayer perceptron, which outputs the probability values ​​of miRNA association with disease.

2. The miRNA-disease association prediction method based on cross-attention and confidence fusion according to claim 1, characterized in that, Constructing the heterogeneous graph includes: Based on the miRNA-disease database, known association information between miRNAs and diseases is obtained, and the heterogeneous graph is constructed by combining the similarity between miRNAs and between diseases. The heterogeneous graph includes miRNA nodes, disease nodes, miRNA internal edges, disease internal edges, and miRNA-disease association edges. The feature representation of the miRNA node is obtained by obtaining the functional information of the miRNA node and encoding the biological semantics related to the target gene. The feature representation of the disease node is obtained by encoding based on medical semantic information or ontology description.

3. The miRNA-disease association prediction method based on cross-attention and confidence fusion according to claim 1, characterized in that, Encoding the miRNA nodes and disease nodes in the heterogeneous graph includes: An improved GraphSAGE encoder is used to encode the features of the miRNA node and the disease node. The improved GraphSAGE encoder obtains three types of aggregation results through a mean aggregator, a pool aggregator, and an LSTM aggregator. The three types of aggregation results are concatenated, and the concatenated aggregation results are processed by linear transformation, nonlinear activation function, and Dropout fusion.

4. The miRNA-disease association prediction method based on cross-attention and confidence fusion according to claim 1, characterized in that, Introducing cross-attention mechanisms for bidirectional interaction between miRNAs and disease nodes includes: In the miRNA-to-disease direction, the miRNA embedding representation is used as the query and the disease embedding representation is used as the key and value. Weights are calculated and aggregated through an attention mechanism to obtain the miRNA representation under the disease context enhancement. In the disease-to-miRNA direction, the disease embedding representation is used as the query and the miRNA embedding representation is used as the key and value. Weights are calculated and aggregated through an attention mechanism to obtain the disease representation under miRNA context enhancement.

5. The miRNA-disease association prediction method based on cross-attention and confidence fusion according to claim 4, characterized in that, Obtaining the miRNA representation in the disease context enhancement includes: ； ； in, This represents the value vector of disease node j. Let represent the attention weight of miRNA node i to disease node j, where i represents the i-th miRNA node and j represents the j-th disease node. Let be the query vector corresponding to the i-th miRNA node. Let represent the key vector corresponding to the j-th disease node, and w represent the dimension of each attention head. , , ,in , For learnable weights during the training process, and These are the miRNA node embedding representation and the disease node embedding representation output by the encoder, respectively.

6. The miRNA-disease association prediction method based on cross-attention and confidence fusion according to claim 1, characterized in that, Weighted fusion of miRNA-disease pair features based on the channel-attached confidence includes: = ； in, This indicates the characteristics of the fused miRNA-disease pair. This represents the miRNA-disease pair characteristics of the k-th channel; This represents the confidence level of the k-th channel.

Images