Method and device for predicting correlation between lncrna and disease

Abstract

This invention discloses a method and apparatus for predicting the association between lncRNA and disease. The method includes: constructing a multi-representation heterogeneous map representing lncRNA-disease-miRNA using a pre-trained large RNA model and a medical subject thesaurus description of the disease; performing structural-level enhancement and meta-path-level enhancement processing on the multi-representation heterogeneous map to obtain a structural-level enhanced heterogeneous map and a meta-path-level enhanced heterogeneous map; using a negative sample filter to screen for predicted negative samples; training the association prediction model through multiple iterations until a preset number of iterations is reached; updating the training parameters of the association prediction model in each iteration using a heterogeneous map encoder, contrastive learning, diffusion model, ensemble representation learning, and association prediction; and obtaining the predicted value of the lncRNA-disease association using the trained association prediction model. This invention can solve the problems of insufficient association quantity, noise often introduced by negative sampling, and insufficient fusion of multi-view data.

Description

Technical Field

[0001] The embodiments of the present invention relate to the field of bioinformatics, and in particular to a system and method for predicting the association between lncRNA and disease. Background Technology

[0002] Long non-coding RNAs (lncRNAs) are a class of non-coding RNAs exceeding 200 nucleotides in length and lack protein-coding function. In recent years, numerous studies have demonstrated that lncRNAs play crucial roles in many human biological processes and are closely associated with various diseases. Exploring the association between lncRNAs and corresponding diseases is an important direction in biomolecular mechanism research. However, biological experimental methods for investigating such associations are generally costly and time-consuming. Therefore, computational methods for exploring the association between lncRNAs and corresponding diseases can obtain a list of candidate associations, effectively saving research time and costs.

[0003] In recent years, computational models used to explore the association between lncRNAs and corresponding diseases can be broadly divided into two categories: one based on random walks or various propagation algorithms, and the other based on machine learning and deep learning techniques. Among them, models based on random walks or various propagation algorithms are easily affected by the extent of data research and tend to focus on lncRNAs and corresponding diseases that have been studied more thoroughly, because these lncRNAs and diseases contribute more to the association network.

[0004] Models based on machine learning and deep learning typically perform feature extraction first, and then complete the association exploration through classifiers or matrix completion. In particular, graph neural network-based models have been widely used in this field in recent years. However, existing computational models still have many limitations, specifically: (1) The number of experimentally verified association samples between lncRNAs and corresponding diseases is relatively small, resulting in a sparse association network and thus limiting the performance of the model. Meanwhile, the potential association space between unverified lncRNAs and corresponding diseases is still very broad; (2) The negative samples required for model training are mostly obtained through random sampling. This sampling method easily introduces noisy and meaningless samples, and ignores samples with key value, affecting the model training effect; (3) The information interaction between different view data is not considered, resulting in insufficient effectiveness in processing multiple feature data, leading to insufficient node representation learning and poor fusion effect of multi-view data. Therefore, it is urgent to design a computational method for exploring the association between lncRNAs and corresponding diseases to solve the problems existing in the above-mentioned technologies.

[0005] The background section described above is merely a description made by the inventor based on their understanding, and the above content should not be regarded as evidence of prior art disclosed before the filing date of this application. Summary of the Invention

[0006] This invention provides a method for predicting the association between lncRNAs and diseases, addressing issues in existing technologies such as limited association data, noise introduced by random negative sampling, and insufficient node representation learning and multi-view data fusion. The method includes: Based on data on lncRNA-disease association, lncRNA-miRNA interaction, miRNA-disease association, lncRNA sequence, miRNA sequence, and medical subject headings of diseases, a multi-characteristic heterogeneity map is constructed using a pre-trained large RNA model. This multi-characteristic heterogeneity map is used to describe the initial characterization of lncRNA, the initial characterization of miRNA, and the initial characterization of disease. Based on feature constraints and reachability constraints, structural-level enhancement processing is performed on multi-representation heterogeneous graphs to obtain structurally enhanced heterogeneous graphs. Meta-path-level enhancement processing is performed on the multi-representation heterogeneous graph to obtain a meta-path-level enhanced heterogeneous graph; Based on lncRNA-disease association data, initial lncRNA characterization, and initial disease characterization, a negative sample filter is used to screen predicted negative samples. The negative sample filter is a reinforcement learning agent. The association prediction model is trained through multiple rounds of iterative training until a preset number of iterations is reached. In each iteration, based on the initial lncRNA representation, initial miRNA representation, initial disease representation, structural enhancement heterogeneity graph, and metapath-level enhancement heterogeneity graph, the total loss function value of contrastive learning and the total loss function value of the diffusion model are obtained through contrastive learning and diffusion model. The lncRNA-disease association pair representation is obtained through ensemble representation learning. Based on positive samples and predicted negative samples, the predicted value of the lncRNA-disease association pair representation is obtained, and the association prediction loss function value is calculated. Based on the total loss function value of contrastive learning, the total loss function value of the diffusion model, and the association prediction loss function value, the training parameters of the association prediction model are updated through backpropagation. The lncRNA-disease association pairs to be predicted are input into the trained association prediction model to obtain the predicted values ​​of the lncRNA-disease associations.

[0007] Another aspect of the present invention provides a device for predicting the association between lncRNAs and diseases, to address the problems existing in the prior art, such as the small number of associations, the noise typically introduced by random negative sampling, and insufficient node representation learning and multi-view data fusion. The device includes: The heterogeneous graph encoder is used to obtain lncRNA contextual information representation and disease contextual information representation in each iteration based on the current lncRNA representation, current disease representation, current miRNA representation, structural-level enhanced heterogeneous graph, and metapath-level enhanced heterogeneous graph. In the initial iteration, the current lncRNA representation is the initial lncRNA representation, the current disease representation is the initial disease representation, and the current miRNA representation is the initial miRNA representation. The contrastive learning module is used to perform contrastive learning on the lncRNA contextual information representation and the disease contextual information representation of the two views in each iteration, and obtain the total loss function value of the contrastive learning. The diffusion model is used to perform diffusion model analysis on the lncRNA context information representation and the disease context information representation of the two views in each iteration, and to obtain the total loss function value of the diffusion model. An integrated representation learning module is used to fuse the lncRNA contextual information representations of the two views in each iteration to obtain a fused lncRNA representation, and to fuse the disease contextual information representations of the two views to obtain a fused disease representation. The prediction module is used to perform element-wise multiplication of the lncRNA fusion representation and the disease fusion representation in each iteration to obtain the lncRNA-disease association pair representation; based on the positive samples and predicted negative samples, and the lncRNA-disease association pair representation, the prediction score of the lncRNA-disease association pair is obtained using a multilayer perceptron, and the association prediction loss function value is calculated. The iterative update module is used to obtain the global total loss function value in each iteration based on the total loss function value of the comparative learning, the total loss function value of the diffusion model, and the loss function value of the association prediction; and to update all training parameters through backpropagation based on the global total loss function value.

[0008] This invention also provides a computer device, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the above-described method for predicting the association between lncRNA and disease.

[0009] This invention also provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the above-described method for predicting the association between lncRNA and disease.

[0010] This invention also provides a computer program product, which includes a computer program that, when executed by a processor, implements the above-described method for predicting the association between lncRNA and disease.

[0011] The beneficial effects achieved by the embodiments of the present invention are as follows: 1. This invention addresses the problem of the limited number of known associations in the field of lncRNA-disease association prediction: The field of lncRNA-disease association prediction faces the challenge of a limited number of known associations, with the number of known associations (positive samples) being far less than the number of unknown associations. This invention alleviates the dependence of the pre-trained large RNA model on the limited number of known associations by using the pre-trained large RNA model RNAErnie, setting up structural-level heterogeneity graph enhancement and meta-path-level heterogeneity graph enhancement, and introducing contrastive learning and diffusion models, thereby enhancing the model's generalization ability.

[0012] 2. High-quality predicted negative samples are used for training the association prediction model: Existing association prediction models usually use random sampling to obtain predicted negative samples. However, random sampling usually introduces noisy and meaningless samples, while ignoring key and valuable samples. This invention proposes a reinforcement learning-based negative sampling strategy with active learning, dynamic optimization and intelligent sample quality assessment capabilities to build a relatively clean and reliable dataset, thereby improving the association prediction model's ability to learn more discriminative features.

[0013] 3. By introducing contrastive learning, diffusion models, and ensemble representation learning, information interaction and fusion between different views are fully realized, and the quality of node representations is enhanced. Attached Figure Description

[0014] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art 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. In the drawings: Figure 1 This is a flowchart of the method for predicting the association between lncRNA and disease in an embodiment of the present invention; Figure 2 This is a structural diagram of the negative sample filter in an embodiment of the present invention; Figure 3 This is a schematic diagram of the structure of the association prediction model in an embodiment of the present invention; Figure 4 This is a schematic diagram of the structure of the lncRNA-disease association prediction device in an embodiment of the present invention; Figure 5 This is a schematic diagram of a computer device in an embodiment of the present invention. Detailed Implementation

[0015] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the embodiments of the present invention will be further described in detail below with reference to the accompanying drawings. Here, the illustrative embodiments of the present invention and their descriptions are used to explain the present invention, but are not intended to limit the present invention.

[0016] Figure 1 This is a flowchart of the method for predicting the association between lncRNA and disease in an embodiment of the present invention, such as... Figure 1 As shown, the method includes: Step 101: Based on lncRNA-disease association data, lncRNA-miRNA interaction data, miRNA-disease association data, lncRNA sequence data, miRNA sequence data, and the medical subject heading list description of the disease, a multi-characteristic heterogeneity map is constructed using a pre-trained RNA large model. The multi-characteristic heterogeneity map is used to describe the initial characterization of lncRNA, the initial characterization of miRNA, and the initial characterization of the disease. Step 102: Based on feature constraints and reachability constraints, perform structural enhancement processing on the multi-representation heterogeneous graph to obtain a structurally enhanced heterogeneous graph; Step 103: Perform meta-path-level enhancement processing on the multi-representation heterogeneous graph to obtain a meta-path-level enhanced heterogeneous graph; Step 104: Based on lncRNA-disease association data, initial lncRNA characterization, and initial disease characterization, a negative sample filter is used to screen predicted negative samples. The negative sample filter is a reinforcement learning agent. Step 105: Train the association prediction model through multiple rounds of iteration until the preset number of iterations is reached. In each iteration, based on the initial lncRNA representation, initial miRNA representation, initial disease representation, structural enhancement heterogeneity graph, and metapath-level enhancement heterogeneity graph, obtain the total loss function value of contrastive learning and the total loss function value of the diffusion model through contrastive learning and diffusion model. Obtain the lncRNA-disease association pair representation through ensemble representation learning. Based on positive samples and predicted negative samples, obtain the predicted value of the lncRNA-disease association pair representation and calculate the association prediction loss function value. Based on the total loss function value of contrastive learning, the total loss function value of the diffusion model, and the association prediction loss function value, update the training parameters of the association prediction model through backpropagation. Step 106: Input the lncRNA-disease association pair to be predicted into the trained association prediction model to obtain the predicted value of the lncRNA-disease association.

[0017] Depend on Figure 1As shown in the structural diagram, in this embodiment of the invention, the dependence of the pre-trained RNA large model (RNAErnie), the setting of structural-level heterogeneous graph enhancement and meta-path-level heterogeneous graph enhancement, and the introduction of contrastive learning and diffusion models alleviate the dependence of the pre-trained RNA large model on a small number of known associations, thereby enhancing the model's generalization ability. A reinforcement learning-based negative sampling strategy with active learning, dynamic optimization, and intelligent sample quality assessment capabilities is proposed to construct a relatively clean and reliable dataset, thereby improving the association prediction model's ability to learn more discriminative features. The introduction of contrastive learning, diffusion models, and ensemble representation learning fully realizes information interaction and fusion between different views, enhancing the quality of node representations.

[0018] In step 101, based on lncRNA-disease association data, lncRNA-miRNA interaction data, miRNA-disease association data, lncRNA sequence data, miRNA sequence data, and a medical subject heading list description of the disease, a multi-characteristic heterogeneity map is constructed using a pre-trained large RNA model. The multi-characteristic heterogeneity map is used to describe the initial characterization of lncRNA, the initial characterization of miRNA, and the initial characterization of the disease. In this embodiment, based on lncRNA-disease association data, lncRNA-miRNA interaction data, miRNA-disease association data, lncRNA sequence data, miRNA sequence data, and a medical subject heading list description of the disease, a multi-characteristic heterogeneity map is constructed using a pre-trained large RNA model, including: lncRNA sequence data and miRNA sequence data are input into a pre-trained large RNA model to extract initial lncRNA and miRNA representations, both of which are context-aware semantic representations. For each disease, a directed acyclic graph is constructed based on the medical subject headings (STB) description. Based on the STB description of each disease, an initial disease representation is generated. Based on the initial characterization of lncRNA, initial characterization of miRNA, initial characterization of disease, and data on the association between lncRNA and disease, the interaction between lncRNA and miRNA, and the association between miRNA and disease, a multi-characteristic heterogeneity map is constructed. The initial characterization of lncRNA, the initial characterization of miRNA, and the initial characterization of disease are the lncRNA node, miRNA node, and disease node in the multi-characteristic heterogeneity map.

[0019] During implementation, experimentally validated data on lncRNA-disease associations, lncRNA-miRNA interactions, miRNA-disease associations, lncRNA sequence data, and miRNA sequence data can be collected from existing public databases. The lncRNA-disease association data can be defined using an adjacency matrix (LD), where LD(i,j) = 1 indicates an association between the i-th lncRNA and the j-th disease, and LD(i,j) = 0 indicates that the association has not been experimentally validated; this could be a true negative sample or a potential positive sample. Similar adjacency matrix definitions can be used for lncRNA-miRNA interaction data and miRNA-disease association data.

[0020] A pre-trained large RNA model can be represented as RNAErnie.

[0021] In this embodiment, the pre-trained large RNA model is a Transformer-based RNA pre-trained model, and it employs a multi-layer multi-head Transformer module. During training, the pre-trained large RNA model incorporates base-level, subsequence-level, and motif-level random masked language models, and marks RNA types as stop words and appends them to the end of sequences.

[0022] In practice, the RNA types include lncRNA and miRNA, and the training data comes from a large number of RNA sequences in known public databases. The multi-layer multi-head Transformer module has 768-dimensional hidden states. By using RNA motifs as biological priors and introducing base-level, subsequence-level, and motif-level random masking language models, RNAErnie captures knowledge at the subsequence and motif levels during pre-training, thereby improving pre-training performance and enriching the representation capabilities of RNA sequences. During the pre-training phase, RNAErnie marks RNA types as stop words and appends them to the end of sequences, enabling the model to distinguish the unique features of different RNA types. For example, the extracted lncRNA context-aware semantic representation and miRNA context-aware semantic representation can be 768-dimensional.

[0023] definition Indicates disease A directed acyclic graph, express and its ancestor nodes, right Contribution Represented as:

[0024] In this embodiment, an initial disease characterization is generated based on the directed acyclic graph of each disease, including: Calculate the semantic similarity between any two diseases based on the directed acyclic graph of all diseases; The semantic similarity of each disease is integrated into a vector, which serves as the initial representation of each disease.

[0025] The semantic similarity is shown in the following formula:

[0026] In practice, the dimension of the initial characterization of each disease is the number of diseases, and the dimension of the initial characterization of each lncRNA and miRN is 768.

[0027] When constructing the multi-representation heterogeneous graph, three types of nodes were constructed to represent lncRNA (L), miRNA (M), and disease (D). The initial representation of each node was attached as a node attribute. According to the three types of associated data, corresponding edges were added in sequence: first LD edges were added, then LM and MD edges were added, ensuring that the edge connections were completely consistent with the experimental verification results. The node type and edge type were labeled to form a complete "lncRNA-disease-miRNA heterogeneous graph". This graph contains both the feature information of the nodes and the topological association information between the three types of biological entities, providing a foundation for subsequent heterogeneous graph enhancement and feature extraction.

[0028] In step 102, based on feature constraints and reachability constraints, the multi-representation heterogeneous graph is subjected to structural-level enhancement processing to obtain a structural-level enhanced heterogeneous graph; In practice, since there are few known associations, heterogeneous graphs with multiple representations often contain noise. Therefore, feature constraints and reachability constraints are used to reduce the impact of noise.

[0029] In this embodiment, based on feature constraints and reachability constraints, a structural-level enhancement process is performed on the multi-representation heterogeneous graph to obtain a structurally enhanced heterogeneous graph, including: Perform the following reachability constraint logic judgment: based on the adjacency matrix of a multi-representation heterogeneous graph. The reachability matrix is ​​constructed by exponentiation, and the reachability matrix is ​​used to reflect the number of reachable paths between any two nodes; the set of node pairs formed by all node pairs in the reachability matrix is ​​traversed, and the first set of node pairs with a number greater than a preset threshold for the number of reachable paths between nodes is selected. Perform the following feature constraint logic judgment: calculate the cosine similarity between any two nodes in the multi-representation heterogeneous graph, and filter out the set of second node pairs that are greater than the cosine similarity threshold; Take the intersection of the first set of node pairs and the second set of node pairs, and add the connections formed by all node pairs in the intersection to the multi-representation heterogeneous graph to form a structural-level enhanced heterogeneous graph.

[0030] In a specific implementation, the cosine similarity between nodes can be used to measure their association strength, and a threshold is set to retain connections between nodes with high similarity.

[0031] The original heterogeneous graph Graw=(V,Eraw,T): The constructed lncRNA-disease-miRNA heterogeneous graph, where V=VL∪VD∪VM (VL is the set of lncRNA nodes, VD is the set of disease nodes, and VM is the set of miRNA nodes), Eraw is the original edge set (containing three types of experimental validation edges: lncRNA-disease, lncRNA-miRNA, and miRNA-disease), and T is the node / edge type label.

[0032] lncRNA node characterization matrix HLraw: 768 dimensions (RNAErnie extraction); Disease node representation matrix HDraw: Dimensions are consistent with the number of diseases (generated based on semantic similarity); miRNA node characterization matrix HMraw: 768 dimensions (RNAErnie extraction); The representation dimensions of the three nodes are unified by a linear layer network to obtain HL, HD and HM respectively. The full node representation matrix H=[HL;HD;HM] is then merged (rows correspond to nodes and columns correspond to feature dimensions).

[0033] The adjacency matrix Araw of a heterogeneous graph has a dimension of N×N (N is the total number of nodes, i.e., |VL| + |VD| + |VM|). Araw(i,j)=1 indicates that there is an original edge (edge ​​in Eraw) between node i and node j, and Araw(i,j)=0 indicates that there is no original edge.

[0034] In step 103, the multi-characteristic heterogeneous graph is subjected to meta-path-level enhancement processing to obtain a meta-path-level enhanced heterogeneous graph; In this embodiment, a meta-path-level enhancement process is performed on the multi-representation heterogeneous graph to obtain a meta-path-level enhanced heterogeneous graph, including: Based on the multi-characteristic heterogeneity diagram, a first-class pathway representing lncRNA-disease-lncRNA and a second-class pathway representing lncRNA-miRNA-lncRNA were constructed for lncRNA; Based on the multi-characteristic heterogeneity diagram, a third-order pathway representing disease-lncRNA-disease and a fourth-order pathway representing disease-miRNA-disease are constructed for diseases. Construct the first-element path reachability graph, second-element path reachability graph, third-element path reachability graph, and fourth-element path reachability graph respectively for the first-element path, second-element path, third-element path, and fourth-element path; The graph limits of the first-order path reachability graph and the second-order path reachability graph are obtained by using the general singular value thresholding method and then mixed to obtain the enhanced graph limit matrix corresponding to lncRNA. The meta-path reachability graph of lncRNA synthesis is sampled from the enhanced graph limit matrix corresponding to lncRNA. The graph limits of the third-order and fourth-order path reachable graphs are obtained by using the general singular value thresholding method and then mixed to obtain the enhanced graph limit matrix corresponding to the disease. The synthetic meta-path reachable graph of the disease is then sampled from the enhanced graph limit matrix corresponding to the disease. Edges from the metapath reachability graphs of lncRNA synthesis and disease synthesis are added to the multi-characteristic heterogeneity graph to obtain metapath-level enhancement heterogeneity graphs at the lncRNA level and at the disease level, respectively.

[0035] Metapaths can capture semantic information in heterogeneous biological networks. Metapath-level enhancement utilizes graph limits and aims to address errors and discrepancies in metapath-based sampling by creating new synthetic metapaths without requiring additional biological information. Two metapaths are employed for lncRNAs: the first metapath. Secondary path The first-order meta-pathway means that two lncRNAs are semantically associated through a shared disease; the second-order meta-pathway means that two lncRNAs are semantically associated through a co-acting miRNA. Two meta-pathways are used for disease-related data: a third-order meta-pathway. and the fourth path The third metapath signifies a semantic association between two diseases through a shared lncRNA, while the fourth metapath signifies a semantic association between two diseases through a shared miRNA. Metapaths are semantic paths based on node type sequences, used to capture higher-order semantic associations between nodes in heterogeneous networks. For example, "LDL" represents the semantic association "lncRNA - disease - lncRNA" (two lncRNAs are associated through a common disease). A metapath reachability graph is constructed, which is an adjacency graph of node reachability for a specific metapath. Edges in the graph represent the semantic association between node pairs that satisfy the metapath. For example, the steps for constructing a first-order metapath reachability graph can be: Traverse the target node set: Traverse all lncRNA node pairs (li, lj) (li, lj ∈ VL, i ≠ = j); Meta-path matching verification: Check whether there exists at least one disease node dk∈VD such that there are edges (li,dk) and (dk,lj) in the original heterogeneous graph (i.e., li and lj satisfy the "LDL" path through dk); Reachability graph edge markers: If validation passes, the reachability graph G is the lncRNA metapath corresponding to Ω1. L,Ω1 In the algorithm, the edge (li, lj) is marked as 1 (indicating reachability); otherwise, it is marked as 0.

[0036] Two types of reachability graphs for lncRNAs: Following the logic above, construct first-order path reachability graphs G respectively. L,Ω1 (Corresponding to Ω1=LDL) and the second-order path reachable graph G L,Ω2（ Corresponding to Ω2=LML), both are adjacency matrices of dimension |VL|×|VL|. Construct G D,Ω3 (Corresponding to Ω3=DLD) and G D,Ω4 (corresponding to Ω4=DMD), all are adjacency matrices of dimension |VD|×|VD|.

[0037] Next, graph limits for lncRNA are obtained using the Universal Singular Value Thresholding (USVT) method based on two metapath reachability graphs and then mixed to generate new enhanced graph limits for lncRNA. The synthesized metapath reachability graph is then sampled from these enhanced graph limits. Graph limits refer to the ideal graph structure obtained by matrix decomposition and reconstruction of multiple metapath reachability graphs using USVT, resulting in "averaging + denoising," which reduces the error of single metapath sampling. The operation for diseases is similar. Edges from the synthesized metapath reachability graph are added as enhancing edges to the original multi-representation heterogeneous graph, resulting in metapath-level enhanced heterogeneous graphs at the lncRNA level and at the disease level.

[0038] In step 104, based on lncRNA-disease association data, initial lncRNA characterization, and initial disease characterization, a negative sample filter is used to screen negative samples. The negative sample filter is a reinforcement learning agent. In this embodiment, the negative sample filter is used to: determine positive samples and unlabeled samples based on the adjacency matrix corresponding to lncRNA and disease association data; and perform multiple iterations based on the positive samples and unlabeled samples to obtain predicted negative samples.

[0039] Figure 2 This is a structural diagram of the negative sample filter in an embodiment of the present invention. Figure 2 The negative sample filter includes: Iterative controller 201 is used for: Based on the adjacency matrix corresponding to lncRNA and disease association data, positive samples and unlabeled samples are identified; the iteration rounds corresponding to the stable period are set, and the number of times each unlabeled sample is predicted as a positive sample during the stable period is calculated. When the ratio of the number of predictions to the iteration rounds corresponding to the stable period is less than a set threshold, the unlabeled sample is retained as a predicted negative sample. Sample selector 202 is used for: In each iteration, based on positive samples, unlabeled samples, initial lncRNA characterization, and initial disease characterization, a binary classification decision is made for unlabeled samples, outputting four categories of candidate samples. The four categories of candidate samples include unlabeled samples judged as potentially positive, unlabeled samples judged as candidate negative samples, positive samples judged as positive, and positive samples judged as negative. Sample inspector 203 is used for: In each iteration, positive samples, unlabeled samples identified as potential positives, and unlabeled samples identified as candidate negatives are combined into an evaluation set. Calculate the degree of separation between positive samples in the assessment set and unlabeled samples identified as potential positives and unlabeled samples identified as candidate negatives; A reward signal is generated based on the degree of separation; Update the policy network parameters of the sample selector based on the reward signal; The cross-entropy loss function is used as the loss function to update the network parameters of the sample checker.

[0040] In practice, positive samples are lncRNA-disease association pairs with LD(i,j)=1 in the adjacency matrix corresponding to lncRNA and disease association data, and unlabeled samples are lncRNA-disease association pairs with LD(i,j)=0. When an unlabeled sample is judged as a potentially positive unlabeled sample ω1, the reinforcement learning agent believes that the association pair may be associated and does not include it as a negative sample. When an unlabeled sample is judged as a candidate negative sample ω2, the reinforcement learning agent believes that the association pair is likely not associated and is a candidate negative sample. When a positive sample ω3 is judged as positive, the reinforcement learning agent believes that the positive sample has been correctly identified and the judgment is accurate. When a positive sample ω4 is judged as negative, the agent mistakenly classifies the positive sample as a negative sample and the judgment is incorrect.

[0041] Constructing the evaluation dataset: Combine the positive sample sets Spos, ω1, and ω2 into an evaluation set; Calculate the evaluation metric: Train the sample checker on the evaluation set and calculate the AUC value (measures the discrimination of the screening results: the degree of separation between ω2 and the sets of Spos and ω1). Generate a reward signal re (the reward for the e-th iteration):

[0042] like ≥0: The current screening result is better than the historical best, the reward signal is positive, and the current strategy is encouraged; like <0: The current screening result has deteriorated, the reward signal is negative, and the current strategy is penalized.

[0043] Based on the reward signal, the objective function of the policy network of the sample selector is set as follows when updating the policy network parameters of the sample selector:

[0044] in Indicates the first Each sample is in state Select action The probability of.

[0045] The sample checker uses cross-entropy as the loss function to update the network parameters of the sample checker; After multiple rounds of training by the sample selector and sample checker, the iteration termination condition is the iteration round corresponding to the stable period. Based on the adjacency matrix corresponding to lncRNA and disease association data, positive samples and unlabeled samples are determined. The iteration round corresponding to the stable period is set, and the number of times each unlabeled sample is predicted as a positive sample during the stable period is calculated. When the ratio of the number of predictions to the iteration round corresponding to the stable period is less than a set threshold, the unlabeled sample is retained as a predicted negative sample.

[0046] In step 105, the association prediction model is trained through multiple rounds of iteration until a preset number of iterations is reached. In each iteration, based on the initial lncRNA representation, initial miRNA representation, initial disease representation, structural enhancement heterogeneity graph, and metapath-level enhancement heterogeneity graph, the total loss function value of contrastive learning and the total loss function value of the diffusion model are obtained through contrastive learning and diffusion model. The lncRNA-disease association pair representation is obtained through ensemble representation learning. Based on positive samples and predicted negative samples, the predicted value of the lncRNA-disease association pair representation is obtained, and the association prediction loss function value is calculated. Based on the total loss function value of contrastive learning, the total loss function value of the diffusion model, and the association prediction loss function value, the training parameters of the association prediction model are updated through backpropagation.

[0047] Figure 3 This is a schematic diagram of the structure of the association prediction model in an embodiment of the present invention. In this embodiment, the association prediction model includes: The heterogeneous graph encoder 301 is used to obtain lncRNA contextual information representation and disease contextual information representation based on the current lncRNA representation, current disease representation, current miRNA representation, structural-level enhanced heterogeneous graph, and metapath-level enhanced heterogeneous graph in each iteration. In the initial iteration, the current lncRNA representation is the initial lncRNA representation, the current disease representation is the initial disease representation, and the current miRNA representation is the initial miRNA representation. The contrastive learning module 302 is used to perform contrastive learning on the lncRNA contextual information representation of the two views and the disease contextual information representation of the two views in each iteration, and obtain the total loss function value of the contrastive learning. Diffusion model 303 is used to perform diffusion model analysis on the dual-view lncRNA context information representation and the dual-view disease context information representation in each iteration to obtain the total loss function value of the diffusion model. The integrated representation learning module 304 is used to fuse the lncRNA context information representations of the two views in each iteration to obtain the lncRNA fusion representation, and to fuse the disease context information representations of the two views to obtain the disease fusion representation. The prediction module 305 is used to perform element-wise multiplication of the lncRNA fusion characterization and the disease fusion characterization in each iteration to obtain the lncRNA-disease association pair characterization; based on the positive samples and predicted negative samples, and the lncRNA-disease association pair characterization, the prediction score of the lncRNA-disease association pair is obtained using a multilayer perceptron, and the association prediction loss function value is calculated. The iterative update module 306 is used to obtain the global total loss function value in each iteration based on the total loss function value of the comparative learning, the total loss function value of the diffusion model, and the loss function value of the association prediction; and to update all training parameters through backpropagation based on the global total loss function value.

[0048] In a specific implementation, the heterogeneous graph encoder is a Heterogeneous Graph Transformer (HGT), a type of deep learning model specifically designed for heterogeneous graphs (where there are multiple types of node and edge). Based on the attention mechanism of the Transformer architecture, it can adaptively capture the "type specificity" and "association complexity" of nodes and edges in heterogeneous graphs, and finally generate node representations that integrate local and global semantic information.

[0049] Compared to traditional graph neural networks (such as GCN and GAT), its core difference lies in the fact that traditional models treat heterogeneous graphs in a homogeneous way. They either use a uniform transformation for all node types or, while learning node importance, ignore relation types and cannot fully adapt to the semantic differences between different types of nodes / edges. In contrast, the heterogeneous graph Transformer automatically learns the importance weights of different types of nodes / edges through a type-aware attention mechanism, making it more suitable for handling complex biological heterogeneous networks such as lncRNA-disease-miRNA that contain multiple entities and multiple associations.

[0050] In the embodiments, the dual-view lncRNA context information representation includes a first lncRNA context information representation and a second lncRNA context information representation, and the dual-view disease context information representation includes a first disease context information representation and a second disease context information representation. Heterogeneous graph encoders are used for: In each iteration, a first lncRNA contextual information representation is obtained based on the current lncRNA representation, the current disease representation, the current miRNA representation, and the structural enhancement isomorphism. A second lncRNA contextual information representation is obtained based on the current lncRNA representation, the current disease representation, the current miRNA representation, and the metapath-level enhancement isomorphism at the lncRNA level. A first disease contextual information representation is obtained based on the current lncRNA representation, the current disease representation, the current miRNA representation, and the structural enhancement isomorphism. A second disease contextual information representation is obtained based on the current lncRNA representation, the current disease representation, the current miRNA representation, and the metapath-level enhancement isomorphism at the lncRNA level.

[0051] Heterogeneous graph encoders can obtain the first lncRNA contextual information representation. and the first disease context information representation Second lncRNA contextual information characterization Second disease context information representation .

[0052] In this embodiment, the contrastive learning module is used for: In each iteration, multilayer perceptron mapping is performed on the lncRNA context information representation and the disease context information representation of the two-view image to obtain the lncRNA mapping features and the disease mapping features of the two-view image. For all lncRNA nodes in the multi-characteristic heterogeneity map, construct a set of positive lncRNA samples and a set of negative lncRNA samples; and for all disease nodes, construct a set of positive disease samples and a set of negative disease samples. Based on the set of positive and negative lncRNA samples for each lncRNA node, and the lncRNA mapping features of the two views, calculate the inter-view contrastive learning loss function value and the intra-view contrastive learning loss function value for each lncRNA node; calculate the sum of the inter-view contrastive learning loss function values ​​and the intra-view contrastive learning loss function values ​​for all lncRNA nodes to obtain the total lncRNA contrastive learning loss function value. Based on the set of positive and negative disease samples for each disease node, and the disease mapping features of the two views, calculate the inter-view contrast learning loss function value and the intra-view contrast learning loss function value for each disease node; calculate the sum of the inter-view contrast learning loss function values ​​and the intra-view contrast learning loss function values ​​for all disease nodes to obtain the total disease contrast learning loss function value. The total contrastive learning loss function value is obtained by adding the sum of the lncRNA contrastive learning loss function values ​​to the sum of the disease contrastive learning loss function values.

[0053] In this embodiment, the positive samples of each lncRNA node are the first number of lncRNA nodes that have a feature similarity greater than the similarity threshold with the lncRNA node and share the most diverse path with the lncRNA node, and the negative sample set of each lncRNA node is the set of lncRNA nodes other than the positive samples of the lncRNA node. The positive samples of each disease node are the first number of disease nodes whose feature similarity to the disease node is greater than the similarity threshold and which share the most diverse path with the disease node. The negative sample set of each disease node is the set of disease nodes other than the positive samples of the disease node.

[0054] In specific implementation, for example, it can be used for and Mapped using a multilayer perceptron and .

[0055] lncRNA For example, feature similarity greater than And with Sharing the most diverse paths Each lncRNA node is considered a positive sample.

[0056] lncRNA The loss function for inter-view comparison learning is shown below:

[0057] in For temperature coefficient, This represents the cosine similarity.

[0058] After obtaining the inter-view contrastive learning loss function value for a single lncRNA, the inter-view contrastive learning loss function values ​​for all lncRNAs can be obtained by summing them. Similarly, the in-view contrastive learning loss function values ​​for all lncRNAs This can be obtained. Then, the two are added together to obtain the sum of the lncRNA contrastive learning loss function values. Similarly, the sum of the disease contrastive learning loss function values ​​can be obtained. .Will and The sum is the total loss function value for contrastive learning. .

[0059] In the embodiments, the diffusion model is used for: In each iteration, Gaussian noise is gradually added to the contextual information representation of the first lncRNA during the forward noise addition process, and the noise is gradually eliminated during the reverse noise removal process to obtain the restored representation of the first lncRNA. The first expected value of the Euclidean distance between the restored representation of the first lncRNA and the contextual information representation of the second lncRNA is calculated. Gaussian noise is gradually added to the contextual information representation of the second lncRNA during the forward noise addition process, and the noise is gradually eliminated during the reverse noise removal process to obtain the restored representation of the second lncRNA. The second expected value of the Euclidean distance between the restored representation of the second lncRNA and the contextual information representation of the first lncRNA is calculated. Calculate the loss function value of lncRNA diffusion process based on the first and second expected values; In the forward noise addition process, Gaussian noise is gradually added to the first disease context information representation, and in the reverse noise removal process, the noise is gradually eliminated to obtain the first disease recovery representation. The third expectation value of the Euclidean distance between the first disease recovery representation and the second disease context information representation is calculated. Gaussian noise is gradually added to the second disease context information representation during the forward noise addition process, and the noise is gradually eliminated during the reverse noise removal process to obtain the second disease recovery representation. The fourth expectation value of the Euclidean distance between the second disease recovery representation and the first disease context information representation is calculated. Calculate the loss function value for the disease spread process based on the third and fourth expected values; The total loss function value of the diffusion model is obtained by adding the loss function value of the lncRNA diffusion process and the loss function value of the disease diffusion process.

[0060] This invention uses a diffusion model to bridge the gap between representations learned from two different views and achieve information sharing through a generative approach. The diffusion model used in this invention is a denoising diffusion probability model. Taking lncRNA as an example, Gaussian noise is progressively added during the forward denoising process. Next, a neural network is trained to reverse this process, progressively eliminating noise during the reverse denoising process. Assume... The corresponding first lncRNA recovery characterization is ,calculate and The first expected value of the Euclidean distance between them. Similarly, calculate... and The second expected value of the Euclidean distance between them. The loss function for the diffusion process of lncRNA is shown below:

[0061] Similarly, the loss function for the disease spread process can be obtained. Total loss function value of the diffusion model pass and Add them together to get the result.

[0062] In this embodiment, the integrated representation learning module is used for: In each iteration, the first query vector is calculated based on the first lncRNA context information, and the first bond vector and the first value vector are calculated based on the second lncRNA context information. The first attention score is obtained by performing a dot product operation on the first query vector and the first key vector, followed by normalization using the softmax function. The first attention score is then weighted and summed with the first value vector to obtain the enhanced lncRNA structural view representation guided by the metapath view. A second query vector is calculated based on the second lncRNA contextual information representation, and a second key vector and a second value vector are calculated based on the first lncRNA contextual information representation. The second attention score is then obtained by performing a dot product operation on the second query vector and the second key vector, followed by normalization using the softmax function. Finally, the second attention score is weighted and summed with the second value vector to obtain the enhanced lncRNA metapath view representation guided by the structural view. The enhanced lncRNA structural view guided by the metapath view and the enhanced lncRNA metapath view guided by the structural view are spliced ​​together to obtain the lncRNA fusion characterization, which is used as the current lncRNA characterization for the next iteration. The third query vector is calculated based on the first disease context information representation, and the third key vector and third value vector are calculated based on the second disease context information representation. The third attention score is obtained by performing a dot product operation on the third query vector and the third key vector, combined with softmax normalization. The third attention score is then weighted and summed with the third value vector to obtain the enhanced representation of the disease structure view guided by the meta-path view. The fourth query vector is calculated based on the second disease context information representation, and the fourth key vector and fourth value vector are calculated based on the first disease context information representation. The fourth attention score is obtained by performing a dot product operation on the fourth query vector and the fourth key vector, combined with softmax normalization. The fourth attention score is then weighted and summed with the fourth value vector to obtain the enhanced representation of the disease meta-path view guided by the structure view. The enhanced representation of the disease structure view guided by the meta-path view and the enhanced representation of the disease meta-path view guided by the structure view are concatenated to obtain the disease fusion representation, which is then used as the current disease representation for the next iteration.

[0063] In each iteration, the prediction module performs element-wise multiplication of the lncRNA fusion representation and the disease fusion representation to obtain the lncRNA-disease association pair representation. Based on the positive samples, predicted negative samples, and lncRNA-disease association pair representation, a multilayer perceptron is used to obtain the prediction score of the lncRNA-disease association pair, and the association prediction loss function value is calculated.

[0064] In this embodiment, the association prediction loss function uses the binary cross-entropy loss function, as shown below:

[0065] in This represents the total number of positive samples and predicted negative samples. The true label representing a positive sample or a predicted negative sample.

[0066] In this embodiment, the iterative update module is specifically used for: In each iteration, the auxiliary task loss function value is determined based on the total loss function value of the contrastive learning and the total loss function value of the diffusion model. The auxiliary task loss function value is added to the associated prediction loss function value to obtain the global total loss function value. The gradient of the global total loss with respect to all training parameters is calculated through backpropagation. All training parameters are updated once by the optimizer along the gradient descent direction.

[0067] Both the contrastive learning process and the diffusion model process are auxiliary tasks containing loss functions. In this invention, the loss function value of the auxiliary task is weighted by these two parts, as shown below:

[0068] The total global loss function value is obtained by adding the correlation prediction loss function value and the auxiliary task loss function value, as shown below:

[0069] In step 106, the lncRNA-disease association pair to be predicted is input into the trained association prediction model to obtain the predicted value of the lncRNA-disease association.

[0070] The predicted values ​​of the above-mentioned lncRNA-disease associations can be used as a probability value for doctors to refer to in order to help determine the disease type.

[0071] This invention also provides a device for predicting the association between lncRNA and disease, as described in the following embodiments. Since the principle by which this device solves the problem is similar to the method for predicting the association between lncRNA and disease, the implementation of this method can be found in the implementation of the method for predicting the association between lncRNA and disease; repeated details will not be elaborated further.

[0072] Figure 4 This is a schematic diagram of the structure of the lncRNA-disease association prediction device in an embodiment of the present invention, as shown below. Figure 4 As shown, the device includes: The heterogeneity graph construction module 401 is used to construct a multi-characteristic heterogeneity graph based on lncRNA-disease association data, lncRNA-miRNA interaction data, miRNA-disease association data, lncRNA sequence data, miRNA sequence data, and a medical subject thesaurus description of the disease, using a pre-trained large RNA model. The multi-characteristic heterogeneity graph is used to describe the initial characterization of lncRNA, the initial characterization of miRNA, and the initial characterization of the disease. The structural enhancement processing module 402 is used to perform structural enhancement processing on multi-representation heterogeneous graphs based on feature constraints and reachability constraints to obtain structurally enhanced heterogeneous graphs. Meta-path-level enhancement processing module 403 is used to perform meta-path-level enhancement processing on multi-representation heterogeneous graphs to obtain meta-path-level enhanced heterogeneous graphs. The negative sample screening module 404 is used to screen predicted negative samples based on lncRNA-disease association data, initial lncRNA characterization and initial disease characterization, using a negative sample screener, wherein the negative sample screener is a reinforcement learning agent. The association prediction model training module 405 is used to train the association prediction model through multiple rounds of iteration until a preset number of iterations is reached. In each round of iteration, based on the initial lncRNA representation, initial miRNA representation, initial disease representation, structural enhancement heterogeneity graph, and metapath-level enhancement heterogeneity graph, the total loss function value of contrastive learning and the total loss function value of the diffusion model are obtained through contrastive learning and diffusion model. The lncRNA-disease association pair representation is obtained through ensemble representation learning. Based on positive samples and predicted negative samples, the predicted value of the lncRNA-disease association pair representation is obtained, and the association prediction loss function value is calculated. Based on the total loss function value of contrastive learning, the total loss function value of the diffusion model, and the association prediction loss function value, the training parameters of the association prediction model are updated through backpropagation. The association prediction model application module 406 is used to input the lncRNA-disease association pair to be predicted into the trained association prediction model to obtain the predicted value of the lncRNA-disease association.

[0073] In this embodiment, the pre-trained large RNA model is a Transformer-based RNA pre-trained model, and it employs a multi-layer multi-head Transformer module. During training, the pre-trained large RNA model incorporates base-level, subsequence-level, and motif-level random masked language models, and marks RNA types as stop words and appends them to the end of sequences.

[0074] In this embodiment, the heterogeneous graph generation module is used for: lncRNA sequence data and miRNA sequence data are input into a pre-trained large RNA model to extract initial lncRNA and miRNA representations, both of which are context-aware semantic representations. For each disease, a directed acyclic graph is constructed based on the medical subject headings (STB) description. Based on the STB description of each disease, an initial disease representation is generated. Based on the initial characterization of lncRNA, initial characterization of miRNA, initial characterization of disease, and data on the association between lncRNA and disease, the interaction between lncRNA and miRNA, and the association between miRNA and disease, a multi-characteristic heterogeneity map is constructed. The initial characterization of lncRNA, the initial characterization of miRNA, and the initial characterization of disease are the lncRNA node, miRNA node, and disease node in the multi-characteristic heterogeneity map.

[0075] In this embodiment, the heterogeneous graph construction module is used for: Calculate the semantic similarity between any two diseases based on the directed acyclic graph of all diseases; The semantic similarity of each disease is integrated into a vector, which serves as the initial representation of the disease.

[0076] In this embodiment, the structural enhancement processing module is used for: Perform the following reachability constraint logic judgment: based on the adjacency matrix of a multi-representation heterogeneous graph. The reachability matrix is ​​constructed by exponentiation, and the reachability matrix is ​​used to reflect the number of reachable paths between any two nodes; the set of node pairs formed by all node pairs in the reachability matrix is ​​traversed, and the first set of node pairs with a number greater than a preset threshold for the number of reachable paths between nodes is selected. Perform the following feature constraint logic judgment: calculate the cosine similarity between any two nodes in the multi-representation heterogeneous graph, and filter out the set of second node pairs that are greater than the cosine similarity threshold; Take the intersection of the first set of node pairs and the second set of node pairs, and add the connections formed by all node pairs in the intersection to the multi-representation heterogeneous graph to form a structural-level enhanced heterogeneous graph.

[0077] In this embodiment, the meta-path-level enhancement processing module is used for: Based on the multi-characteristic heterogeneity diagram, a first-class pathway representing lncRNA-disease-lncRNA and a second-class pathway representing lncRNA-miRNA-lncRNA were constructed for lncRNA; Based on the multi-characteristic heterogeneity diagram, a third-order pathway representing disease-lncRNA-disease and a fourth-order pathway representing disease-miRNA-disease are constructed for diseases. Construct the first-element path reachability graph, second-element path reachability graph, third-element path reachability graph, and fourth-element path reachability graph respectively for the first-element path, second-element path, third-element path, and fourth-element path; The graph limits of the first-order path reachability graph and the second-order path reachability graph are obtained by using the general singular value thresholding method and then mixed to obtain the enhanced graph limit matrix corresponding to lncRNA. The meta-path reachability graph of lncRNA synthesis is sampled from the enhanced graph limit matrix corresponding to lncRNA. The graph limits of the third-order and fourth-order path reachable graphs are obtained by using the general singular value thresholding method and then mixed to obtain the enhanced graph limit matrix corresponding to the disease. The synthetic meta-path reachable graph of the disease is then sampled from the enhanced graph limit matrix corresponding to the disease. Edges from the metapath reachability graphs of lncRNA synthesis and disease synthesis are added to the multi-characteristic heterogeneity graph to obtain metapath-level enhancement heterogeneity graphs at the lncRNA level and at the disease level, respectively.

[0078] In this embodiment, the negative sample filter is used to: determine positive samples and unlabeled samples based on the adjacency matrix corresponding to lncRNA and disease association data; and perform multiple iterations based on the positive samples and unlabeled samples to filter and predict negative samples.

[0079] In this embodiment, the negative sample filter includes: Iterative controllers are used for: Based on the adjacency matrix corresponding to lncRNA and disease association data, positive samples and unlabeled samples are identified; the iteration rounds corresponding to the stable period are set, and the number of times each unlabeled sample is predicted as a positive sample during the stable period is calculated. When the ratio of the number of predictions to the iteration rounds corresponding to the stable period is less than a set threshold, the unlabeled sample is retained as a predicted negative sample. The sample selector is used for: In each iteration, based on positive samples, unlabeled samples, initial lncRNA characterization, and initial disease characterization, a binary classification decision is made for unlabeled samples, outputting four categories of candidate samples. The four categories of candidate samples include unlabeled samples judged as potentially positive, unlabeled samples judged as candidate negative samples, positive samples judged as positive, and positive samples judged as negative. The sample inspector is used for: In each iteration, positive samples, unlabeled samples identified as potential positives, and unlabeled samples identified as candidate negatives are combined into an evaluation set. Calculate the degree of separation between positive samples in the assessment set and unlabeled samples identified as potential positives and unlabeled samples identified as candidate negatives; A reward signal is generated based on the degree of separation; Update the policy network parameters of the sample selector based on the reward signal; The cross-entropy loss function is used as the loss function to update the network parameters of the sample checker.

[0080] In this embodiment, the association prediction model includes: The heterogeneous graph encoder is used to obtain lncRNA contextual information representation and disease contextual information representation in each iteration based on the current lncRNA representation, current disease representation, current miRNA representation, structural-level enhanced heterogeneous graph, and metapath-level enhanced heterogeneous graph. In the initial iteration, the current lncRNA representation is the initial lncRNA representation, the current disease representation is the initial disease representation, and the current miRNA representation is the initial miRNA representation. The contrastive learning module is used to perform contrastive learning on the lncRNA contextual information representation and the disease contextual information representation of the two views in each iteration, and obtain the total loss function value of the contrastive learning. The diffusion model is used to perform diffusion model analysis on the lncRNA context information representation and the disease context information representation of the two views in each iteration, and to obtain the total loss function value of the diffusion model. An integrated representation learning module is used to fuse the lncRNA contextual information representations of the two views in each iteration to obtain a fused lncRNA representation, and to fuse the disease contextual information representations of the two views to obtain a fused disease representation. The prediction module is used to perform element-wise multiplication of the lncRNA fusion representation and the disease fusion representation in each iteration to obtain the lncRNA-disease association pair representation; based on the positive samples and predicted negative samples, and the lncRNA-disease association pair representation, the prediction score of the lncRNA-disease association pair is obtained using a multilayer perceptron, and the association prediction loss function value is calculated. The iterative update module is used to obtain the global total loss function value in each iteration based on the total loss function value of the comparative learning, the total loss function value of the diffusion model, and the loss function value of the association prediction; and to update all training parameters through backpropagation based on the global total loss function value.

[0081] In the embodiments, the dual-view lncRNA context information representation includes a first lncRNA context information representation and a second lncRNA context information representation, and the dual-view disease context information representation includes a first disease context information representation and a second disease context information representation. Heterogeneous graph encoders are used for: In each iteration, a first lncRNA contextual information representation is obtained based on the current lncRNA representation, the current disease representation, the current miRNA representation, and the structural enhancement isomorphism. A second lncRNA contextual information representation is obtained based on the current lncRNA representation, the current disease representation, the current miRNA representation, and the metapath-level enhancement isomorphism at the lncRNA level. A first disease contextual information representation is obtained based on the current lncRNA representation, the current disease representation, the current miRNA representation, and the structural enhancement isomorphism. A second disease contextual information representation is obtained based on the current lncRNA representation, the current disease representation, the current miRNA representation, and the metapath-level enhancement isomorphism at the lncRNA level.

[0082] In this embodiment, the contrastive learning module is used for: In each iteration, multilayer perceptron mapping is performed on the lncRNA context information representation and the disease context information representation of the two-view image to obtain the lncRNA mapping features and the disease mapping features of the two-view image. For all lncRNA nodes in the multi-characteristic heterogeneity map, construct a set of positive lncRNA samples and a set of negative lncRNA samples; and for all disease nodes, construct a set of positive disease samples and a set of negative disease samples. Based on the set of positive and negative lncRNA samples for each lncRNA node, and the lncRNA mapping features of the two views, calculate the inter-view contrastive learning loss function value and the intra-view contrastive learning loss function value for each lncRNA node; calculate the sum of the inter-view contrastive learning loss function values ​​and the intra-view contrastive learning loss function values ​​for all lncRNA nodes to obtain the total lncRNA contrastive learning loss function value. Based on the set of positive and negative disease samples for each disease node, and the disease mapping features of the two views, calculate the inter-view contrast learning loss function value and the intra-view contrast learning loss function value for each disease node; calculate the sum of the inter-view contrast learning loss function values ​​and the intra-view contrast learning loss function values ​​for all disease nodes to obtain the total disease contrast learning loss function value. The total contrastive learning loss function value is obtained by adding the sum of the lncRNA contrastive learning loss function values ​​to the sum of the disease contrastive learning loss function values.

[0083] In this embodiment, the positive samples of each lncRNA node are the first number of lncRNA nodes that have a feature similarity greater than the similarity threshold with the lncRNA node and share the most diverse path with the lncRNA node, and the negative sample set of each lncRNA node is the set of lncRNA nodes other than the positive samples of the lncRNA node. The positive samples of each disease node are the first number of disease nodes whose feature similarity to the disease node is greater than the similarity threshold and which share the most diverse path with the disease node. The negative sample set of each disease node is the set of disease nodes other than the positive samples of the disease node.

[0084] In the embodiments, the diffusion model is used for: In each iteration, Gaussian noise is gradually added to the contextual information representation of the first lncRNA during the forward noise addition process, and the noise is gradually eliminated during the reverse noise removal process to obtain the restored representation of the first lncRNA. The first expected value of the Euclidean distance between the restored representation of the first lncRNA and the contextual information representation of the second lncRNA is calculated. Gaussian noise is gradually added to the contextual information representation of the second lncRNA during the forward noise addition process, and the noise is gradually eliminated during the reverse noise removal process to obtain the restored representation of the second lncRNA. The second expected value of the Euclidean distance between the restored representation of the second lncRNA and the contextual information representation of the first lncRNA is calculated. Calculate the loss function value of lncRNA diffusion process based on the first and second expected values; In the forward noise addition process, Gaussian noise is gradually added to the first disease context information representation, and in the reverse noise removal process, the noise is gradually eliminated to obtain the first disease recovery representation. The third expectation value of the Euclidean distance between the first disease recovery representation and the second disease context information representation is calculated. Gaussian noise is gradually added to the second disease context information representation during the forward noise addition process, and the noise is gradually eliminated during the reverse noise removal process to obtain the second disease recovery representation. The fourth expectation value of the Euclidean distance between the second disease recovery representation and the first disease context information representation is calculated. Calculate the loss function value for the disease spread process based on the third and fourth expected values; The total loss function value of the diffusion model is obtained by adding the loss function value of the lncRNA diffusion process and the loss function value of the disease diffusion process.

[0085] In this embodiment, the integrated representation learning module is used for: In each iteration, the first query vector is calculated based on the first lncRNA context information, and the first bond vector and the first value vector are calculated based on the second lncRNA context information. The first attention score is obtained by performing a dot product operation on the first query vector and the first key vector, followed by normalization using the softmax function. The first attention score is then weighted and summed with the first value vector to obtain the enhanced lncRNA structural view representation guided by the metapath view. A second query vector is calculated based on the second lncRNA contextual information representation, and a second key vector and a second value vector are calculated based on the first lncRNA contextual information representation. The second attention score is then obtained by performing a dot product operation on the second query vector and the second key vector, followed by normalization using the softmax function. Finally, the second attention score is weighted and summed with the second value vector to obtain the enhanced lncRNA metapath view representation guided by the structural view. The enhanced lncRNA structural view guided by the metapath view and the enhanced lncRNA metapath view guided by the structural view are spliced ​​together to obtain the lncRNA fusion characterization, which is used as the current lncRNA characterization for the next iteration. The third query vector is calculated based on the first disease context information representation, and the third key vector and third value vector are calculated based on the second disease context information representation. The third attention score is obtained by performing a dot product operation on the third query vector and the third key vector, combined with softmax normalization. The third attention score is then weighted and summed with the third value vector to obtain the enhanced representation of the disease structure view guided by the meta-path view. The fourth query vector is calculated based on the second disease context information representation, and the fourth key vector and fourth value vector are calculated based on the first disease context information representation. The fourth attention score is obtained by performing a dot product operation on the fourth query vector and the fourth key vector, combined with softmax normalization. The fourth attention score is then weighted and summed with the fourth value vector to obtain the enhanced representation of the disease meta-path view guided by the structure view. The enhanced representation of the disease structure view guided by the meta-path view and the enhanced representation of the disease meta-path view guided by the structure view are concatenated to obtain the disease fusion representation, which is then used as the current disease representation for the next iteration.

[0086] In this embodiment, the iterative update module is specifically used to: determine the auxiliary task loss function value based on the total loss function value of the contrastive learning and the total loss function value of the diffusion model in each iteration; add the auxiliary task loss function value to the associated prediction loss function value to obtain the global total loss function value; calculate the gradient of the global total loss with respect to all training parameters through backpropagation; and update all training parameters at once along the gradient descent direction using the optimizer.

[0087] This invention also provides a computer device. Figure 5 This is a schematic diagram of a computer device in an embodiment of the present invention. The computer device 500 includes a memory 510, a processor 520, and a computer program 530 stored in the memory 510 and executable on the processor 520. When the processor 520 executes the computer program 530, it implements the above-mentioned method for predicting the association between lncRNA and disease.

[0088] This invention also provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the above-described method for predicting the association between lncRNA and disease.

[0089] This invention also provides a computer program product, which includes a computer program that, when executed by a processor, implements the above-described method for predicting the association between lncRNA and disease.

[0090] The beneficial effects achieved by the method and apparatus proposed in the embodiments of the present invention are as follows: This invention addresses the problem of the limited number of known associations in the field of lncRNA-disease association prediction: The field of lncRNA-disease association prediction faces the challenge of a limited number of known associations, with the number of known associations (positive samples) being far less than the number of unknown associations. This invention alleviates the dependence of the pre-trained large RNA model on the limited number of known associations by using the pre-trained large RNA model RNAErnie, setting up structural-level heterogeneity graph enhancement and meta-path-level heterogeneity graph enhancement, and introducing contrastive learning and diffusion models, thereby enhancing the model's generalization ability.

[0091] High-quality predicted negative samples are used for training the association prediction model: Existing association prediction models usually use random sampling to obtain predicted negative samples. However, random sampling usually introduces noisy and meaningless samples, while ignoring key and valuable samples. This invention proposes a reinforcement learning-based negative sampling strategy with active learning, dynamic optimization and intelligent sample quality assessment capabilities to build a relatively clean and reliable dataset, thereby improving the association prediction model's ability to learn more discriminative features.

[0092] By introducing contrastive learning, diffusion models, and association prediction models, the quality of node representation is enhanced, thereby fully realizing information interaction and fusion between different views.

[0093] The multi-layered, multi-headed Transformer module provides deep semantic capture capabilities for lncRNAs and miRNAs. Its 768-dimensional hidden states can accommodate rich sequence features, enhancing the information density of the initial representation. Triple random masks at the base, subsequence, and motif levels enable the model to learn multi-scale biological knowledge of RNA sequences (from single bases to functional motifs), strengthening the biological significance of the representation. Adding stop words to the end of the sequence to label RNA types allows the model to accurately distinguish the unique features of different RNA types such as lncRNAs and miRNAs, avoiding cross-type feature confusion and improving the specificity of the representation.

[0094] By integrating three core association data types (lncRNA-disease, lncRNA-miRNA, and miRNA-disease), a heterogeneous graph is constructed that comprehensively covers the regulatory pathways between biomolecules, compensating for the information deficiencies of single data types. Context-aware semantic representations are extracted using a pre-trained large RNA model, allowing the initial representations of lncRNAs and miRNAs to contain long-range dependencies and functional patterns, outperforming traditional sequence similarity-based feature extraction methods. A directed acyclic graph (DAG) is constructed using disease MESH descriptions to generate initial disease representations, fully mining the semantic association information of diseases and making disease representations more consistent with biological classification logic, providing high-quality node feature support for the heterogeneous graph.

[0095] Semantic similarity is calculated based on disease DAG, and the similarity is integrated into a vector as the initial representation of the disease, directly quantifying the semantic association strength between diseases and giving the representation biological meaning.

[0096] Reachability constraints construct a reachability matrix by raising the adjacency matrix to the nth power, preserving connections with a high number of paths between nodes, strengthening the topological structural relevance of the heterogeneous graph, and alleviating the problem of insufficient structural information caused by the sparsity of known associations. Feature constraints select high-strength connections based on the cosine similarity of node representations, ensuring that enhanced edges conform to the semantic associations of node features and reducing the risk of introducing noisy edges. The enhanced edges selected by dual constraints possess both structural rationality and feature relevance, and the optimized structurally enhanced heterogeneous graph improves the robustness and generalization ability of subsequent feature extraction.

[0097] Two types of dedicated metapaths (LDL, LML, DLD, DMD) were designed for lncRNA and disease respectively to accurately capture high-order semantic associations in heterogeneous graphs, overcoming the limitation of traditional heterogeneous graphs that only focus on direct connections. A universal singular value thresholding (USVT) method was used to generate graph limits and mix them, addressing errors and discrepancies in metapath sampling, and generating high-quality synthesized metapaths without requiring additional biological information. Edges of the synthesized metapaths can be added to the multi-representation heterogeneous graphs, enriching the semantic connections of the multi-representation heterogeneous graphs and improving the semantic richness of the representations.

[0098] The adjacency matrix clearly defines the boundary between positive and unlabeled samples, ensuring the accuracy of sample classification and laying the foundation for negative sample selection. A multi-round iterative convergence mechanism gradually brings the negative sample selection closer to the "high-quality" goal, avoiding noise residue caused by a single selection and ensuring the credibility of the final predicted negative samples. Only candidate negative samples after iterative convergence are used as the final negative samples, effectively eliminating meaningless and potential positive samples, improving the purity of the training sample set, and enhancing the model's discriminative ability.

[0099] The binary classification decision design of the sample selector enables refined classification of unlabeled samples, accurately distinguishing potential positives, candidate negatives, and misclassified samples, providing a clear direction for subsequent optimization. The sample checker generates a reward signal by calculating the degree of sample separation, forming a reinforcement learning closed loop of "screening-evaluation-feedback," driving continuous optimization of the sample selector's policy network and improving screening accuracy. An iteration cycle corresponding to the stabilization period is set, calculating the number of times each unlabeled sample is predicted as a positive sample within the stabilization period. When the ratio of this number to the iteration cycle corresponding to the stabilization period is less than a set threshold, the unlabeled sample is retained as a predicted negative sample, ensuring that predicted negative samples are as "clean and noise-free" as possible, reducing interference during model training from the source.

[0100] The heterogeneous graph encoder dynamically updates node representations in each iteration, achieving collaborative optimization of "representation-graph structure" by combining dual-enhanced heterogeneous graphs, thus improving the quality of contextual information representation. The comparative learning, diffusion model, and association prediction loss function are collaboratively optimized, allowing the model to simultaneously achieve the three goals of "view alignment," "representation sharing," and "accurate association prediction," avoiding biases caused by single-task optimization. The integrated representation learning module fuses dual-view representations, and the iterative update module performs a one-time global parameter update, ensuring the stability and efficiency of model training and improving the accuracy and generalization ability of the final prediction.

[0101] Two types of contextual information representations are extracted based on structural-level enhanced heterogeneous graphs and meta-path-level enhanced heterogeneous graphs, respectively, allowing lncRNAs and diseases to simultaneously acquire structural topological features and high-order semantic features, achieving feature complementarity. The independent extraction and subsequent collaborative optimization of dual-view representations provide differentiated feature inputs for contrastive learning and diffusion models, strengthen the information interaction between views, and improve the discriminativeness and consistency of representations.

[0102] Multilayer perceptrons map dual-view representations to obtain mapped features, adapting to the similarity calculation requirements of comparative learning and improving the computational rationality of the loss function. Positive and negative sample sets are constructed separately for lncRNA and disease, maximizing the similarity between positive samples and minimizing the similarity between negative samples at this stage. Hierarchical calculation and summation of inter-view and intra-view losses comprehensively constrain the consistency and discriminativeness of representations, ensuring that node representations satisfy cross-view alignment while accurately distinguishing semantic differences between different nodes.

[0103] Positive sample selection considers both "feature similarity" and "the number of shared metapaths," ensuring that positive samples possess both structural and semantic relevance, providing high-quality positive examples to support contrastive learning. The negative sample set is defined as nodes of the same type outside of the positive samples, avoiding semantic overlap between negative and positive samples, strengthening the distinction between them, and improving the training effect of contrastive learning. A unified selection criterion applies to both lncRNA and disease nodes, ensuring consistency in the contrastive learning process, unifying the representation optimization logic for both types of nodes, and improving the stability of model training.

[0104] The forward noise addition and backward denoising processes allow the representation of one type of view to approximate the representation of another type of view through a generative method, bridging the gap between the two view representations and achieving cross-view information sharing. Loss is calculated based on the expected value of Euclidean distance, accurately quantifying the difference between the recovered representation and the target representation, providing a clear gradient signal for model optimization. Diffusion losses are calculated and summed separately for lncRNA and disease, ensuring the comprehensiveness of the dual-view information fusion, making the representations of both types of nodes view-compatible, and improving the quality of subsequent fused representations.

[0105] The bidirectional cross-attention mechanism guides the optimization of the structure view and the meta-path view, achieving "precise matching and weighted fusion of information between views," which is more effective at uncovering feature complementarity than simple concatenation. Attention scores are calculated based on QK dot products and softmax normalization, accurately quantifying the correlation strength between views and focusing the fusion process on highly correlated features while avoiding interference from invalid information. The fused representation serves as the current representation for the next iteration, forming a closed loop of "fusion-iteration-optimization," allowing node representations to continuously improve information density and consistency across multiple training rounds.

[0106] A weighted fusion of auxiliary task loss (contrastive learning + diffusion model) and association prediction loss enables multi-task collaborative optimization, allowing the model to simultaneously consider representation quality and prediction accuracy, avoiding the one-sidedness of optimizing a single loss. A one-time parameter update driven by the global total loss avoids parameter oscillations caused by multiple rounds of separate updates, ensuring synchronous optimization of all trainable parameters (heterogeneous graph models, MLPs, attention mechanisms, etc.) and improving training stability. Unified optimization of the gradient descent direction allows model parameters to be collaboratively adjusted towards the goal of "minimizing the global loss," improving model convergence speed and final prediction performance.

[0107] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0108] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0109] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0110] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1The steps of the function specified in one or more boxes.

[0111] The specific embodiments described above further illustrate the purpose, technical solution, and beneficial effects of the present invention. It should be understood that the above descriptions are merely specific embodiments of the present invention and are not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for predicting the association between lncRNA and disease, characterized in that, include: Based on data on lncRNA-disease association, lncRNA-miRNA interaction, miRNA-disease association, lncRNA sequence, miRNA sequence, and medical subject headings of diseases, a multi-characteristic heterogeneity map is constructed using a pre-trained large RNA model. This multi-characteristic heterogeneity map is used to describe the initial characterization of lncRNA, the initial characterization of miRNA, and the initial characterization of disease. Based on feature constraints and reachability constraints, structural-level enhancement processing is performed on multi-representation heterogeneous graphs to obtain structurally enhanced heterogeneous graphs. Meta-path-level enhancement processing is performed on the multi-representation heterogeneous graph to obtain a meta-path-level enhanced heterogeneous graph; Based on lncRNA-disease association data, initial lncRNA characterization, and initial disease characterization, a negative sample filter is used to screen predicted negative samples. The negative sample filter is a reinforcement learning agent. The correlation prediction model is trained through multiple rounds of iterative training until the preset number of iterations is reached. In each iteration, based on the initial lncRNA characterization, initial miRNA characterization, initial disease characterization, structural enhancement isomorphism map, and metapath-level enhancement isomorphism map, the total loss function value of contrastive learning and the total loss function value of diffusion model are obtained through contrastive learning and diffusion model. The lncRNA-disease association pair characterization is obtained through ensemble characterization learning. Based on positive samples and predicted negative samples, the predicted value of the lncRNA-disease association pair characterization is obtained, and the association prediction loss function value is calculated. Based on the total loss function value of contrastive learning, the total loss function value of diffusion model, and the loss function value of association prediction, the training parameters of association prediction model are updated through backpropagation; The lncRNA-disease association pairs to be predicted are input into the trained association prediction model to obtain the predicted values ​​of the lncRNA-disease associations.

2. The method as described in claim 1, characterized in that, The pre-trained large RNA model is a Transformer-based RNA pre-trained model, and it uses a multi-layer multi-head Transformer module. During training, the pre-trained large RNA model incorporates base-level, subsequence-level, and motif-level random masked language models, and marks RNA types as stop words and appends them to the end of sequences.

3. The method as described in claim 1, characterized in that, Based on data on lncRNA-disease associations, lncRNA-miRNA interactions, miRNA-disease associations, lncRNA sequence data, miRNA sequence data, and medical subject headings (MDTs) describing diseases, a multi-characteristic heterogeneity map is constructed using a pre-trained large-scale RNA model, including: lncRNA sequence data and miRNA sequence data are input into a pre-trained large RNA model to extract initial lncRNA and miRNA representations, both of which are context-aware semantic representations. For each disease, a directed acyclic graph is constructed based on the medical subject headings (STB) description. Based on the STB description of each disease, an initial disease representation is generated. Based on the initial characterization of lncRNA, initial characterization of miRNA, initial characterization of disease, and data on the association between lncRNA and disease, the interaction between lncRNA and miRNA, and the association between miRNA and disease, a multi-characteristic heterogeneity map is constructed. The initial characterization of lncRNA, the initial characterization of miRNA, and the initial characterization of disease are the lncRNA node, miRNA node, and disease node in the multi-characteristic heterogeneity map.

4. The method as described in claim 3, characterized in that, Based on the directed acyclic graph of each disease, generate the initial disease representation, including: Calculate the semantic similarity between any two diseases based on the directed acyclic graph of all diseases; The semantic similarity of each disease is integrated into a vector, which serves as the initial representation of each disease.

5. The method as described in claim 1, characterized in that, Based on feature constraints and reachability constraints, structural enhancement processing is performed on multi-representation heterogeneous graphs to obtain structurally enhanced heterogeneous graphs, including: Perform the following reachability constraint logic judgment: based on the adjacency matrix of a multi-representation heterogeneous graph. The reachability matrix is ​​constructed by exponentiation, and the reachability matrix is ​​used to reflect the number of reachable paths between any two nodes; the set of node pairs formed by all node pairs in the reachability matrix is ​​traversed, and the first set of node pairs with a number greater than a preset threshold for the number of reachable paths between nodes is selected. Perform the following feature constraint logic judgment: calculate the cosine similarity between any two nodes in the multi-representation heterogeneous graph, and filter out the set of second node pairs that are greater than the cosine similarity threshold; Take the intersection of the first set of node pairs and the second set of node pairs, and add the connections formed by all node pairs in the intersection to the multi-representation heterogeneous graph to form a structural-level enhanced heterogeneous graph.

6. The method as described in claim 1, characterized in that, Metapath-level enhancement processing is performed on multi-representation heterogeneous graphs to obtain metapath-level enhanced heterogeneous graphs, including: Based on the multi-characteristic heterogeneity diagram, a first-class pathway representing lncRNA-disease-lncRNA and a second-class pathway representing lncRNA-miRNA-lncRNA were constructed for lncRNA; Based on the multi-characteristic heterogeneity diagram, a third-order pathway representing disease-lncRNA-disease and a fourth-order pathway representing disease-miRNA-disease are constructed for diseases. Construct the first-element path reachability graph, second-element path reachability graph, third-element path reachability graph, and fourth-element path reachability graph respectively for the first-element path, second-element path, third-element path, and fourth-element path; The graph limits of the first-order path reachability graph and the second-order path reachability graph are obtained by using the general singular value thresholding method and then mixed to obtain the enhanced graph limit matrix corresponding to lncRNA. The meta-path reachability graph of lncRNA synthesis is sampled from the enhanced graph limit matrix corresponding to lncRNA. The graph limits of the third-order and fourth-order path reachable graphs are obtained by using the general singular value thresholding method and then mixed to obtain the enhanced graph limit matrix corresponding to the disease. The synthetic meta-path reachable graph of the disease is then sampled from the enhanced graph limit matrix corresponding to the disease. Edges from the metapath reachability graphs of lncRNA synthesis and disease synthesis are added to the multi-characteristic heterogeneity graph to obtain metapath-level enhancement heterogeneity graphs at the lncRNA level and at the disease level, respectively.

7. The method as described in claim 1, characterized in that, The negative sample filter is used to: determine positive samples and unlabeled samples based on the adjacency matrix corresponding to lncRNA and disease association data; and perform multiple iterations based on positive samples and unlabeled samples to filter and predict negative samples.

8. The method as described in claim 7, characterized in that, The negative sample filter includes: Iterative controllers are used for: Based on the adjacency matrix corresponding to lncRNA and disease association data, positive samples and unlabeled samples are identified; the iteration rounds corresponding to the stable period are set, and the number of times each unlabeled sample is predicted as a positive sample during the stable period is calculated. When the ratio of the number of predictions to the iteration rounds corresponding to the stable period is less than a set threshold, the unlabeled sample is retained as a predicted negative sample. The sample selector is used for: In each iteration, based on positive samples, unlabeled samples, initial lncRNA characterization, and initial disease characterization, a binary classification decision is made for unlabeled samples, outputting four categories of candidate samples. The four categories of candidate samples include unlabeled samples judged as potentially positive, unlabeled samples judged as candidate negative samples, positive samples judged as positive, and positive samples judged as negative. The sample inspector is used for: In each iteration, positive samples, unlabeled samples identified as potential positives, and unlabeled samples identified as candidate negatives are combined into an evaluation set. Calculate the degree of separation between positive samples in the assessment set and unlabeled samples identified as potential positives and unlabeled samples identified as candidate negatives; Based on the degree of separation, a reward signal is generated; Update the policy network parameters of the sample selector based on the reward signal; The cross-entropy loss function is used as the loss function to update the network parameters of the sample checker.

9. The method as described in claim 6, characterized in that, Association prediction models include: The heterogeneous graph encoder is used to obtain lncRNA contextual information representation and disease contextual information representation in each iteration based on the current lncRNA representation, current disease representation, current miRNA representation, structural-level enhanced heterogeneous graph, and metapath-level enhanced heterogeneous graph. In the initial iteration, the current lncRNA representation is the initial lncRNA representation, the current disease representation is the initial disease representation, and the current miRNA representation is the initial miRNA representation. The contrastive learning module is used to perform contrastive learning on the lncRNA contextual information representation and the disease contextual information representation of the two views in each iteration, and obtain the total loss function value of the contrastive learning. The diffusion model is used to perform diffusion model analysis on the lncRNA context information representation and the disease context information representation of the two views in each iteration, and to obtain the total loss function value of the diffusion model. An integrated representation learning module is used to fuse the lncRNA contextual information representations of the two views in each iteration to obtain a fused lncRNA representation, and to fuse the disease contextual information representations of the two views to obtain a fused disease representation. The prediction module is used to perform element-wise multiplication of the lncRNA fusion representation and the disease fusion representation in each iteration to obtain the lncRNA-disease association pair representation; based on the positive samples and predicted negative samples, and the lncRNA-disease association pair representation, the prediction score of the lncRNA-disease association pair is obtained using a multilayer perceptron, and the association prediction loss function value is calculated. The iterative update module is used to obtain the global total loss function value in each iteration based on the total loss function value of the comparative learning, the total loss function value of the diffusion model, and the loss function value of the association prediction; and to update all training parameters through backpropagation based on the global total loss function value.

10. The method as described in claim 9, characterized in that, The dual-view lncRNA contextual information representation includes a first lncRNA contextual information representation and a second lncRNA contextual information representation; the dual-view disease contextual information representation includes a first disease contextual information representation and a second disease contextual information representation. Heterogeneous graph encoders are used for: In each iteration, the first lncRNA context information representation is obtained based on the current lncRNA representation, the current disease representation, the current miRNA representation, and the structural enhancement isomorphism map. Based on the current lncRNA characterization, current disease characterization, current miRNA characterization, and metapath-level enhancement isomorphism at the lncRNA level, the second lncRNA contextual information characterization is obtained. Based on the current lncRNA characterization, current disease characterization, current miRNA characterization, and structural enhancement isomorphism map, obtain the first disease context information characterization; Based on the current lncRNA characterization, current disease characterization, current miRNA characterization, and metapath-level enhancement heterogeneity map at the lncRNA level, a second disease contextual information characterization is obtained.

11. The method as described in claim 9, characterized in that, The contrastive learning module is used for: In each iteration, multilayer perceptron mapping is performed on the lncRNA context information representation and the disease context information representation of the two-view image to obtain the lncRNA mapping features and the disease mapping features of the two-view image. For all lncRNA nodes in the multi-characteristic heterogeneity map, construct a set of positive lncRNA samples and a set of negative lncRNA samples; for all disease nodes, construct a set of positive disease samples and a set of negative disease samples. Based on the set of positive and negative lncRNA samples for each lncRNA node, and the lncRNA mapping features of the two views, calculate the inter-view contrastive learning loss function value and the intra-view contrastive learning loss function value for each lncRNA node. Calculate the sum of the inter-view contrastive learning loss function value and the intra-view contrastive learning loss function value for all lncRNA nodes to obtain the total lncRNA contrastive learning loss function value; Based on the set of positive and negative disease samples for each disease node, and the disease mapping features of the two views, calculate the inter-view contrast learning loss function value and the intra-view contrast learning loss function value for each disease node; calculate the sum of the inter-view contrast learning loss function values ​​and the intra-view contrast learning loss function values ​​for all disease nodes to obtain the total disease contrast learning loss function value. The total contrastive learning loss function value is obtained by adding the sum of the lncRNA contrastive learning loss function values ​​to the sum of the disease contrastive learning loss function values.

12. The method as described in claim 9, characterized in that, The positive samples of each lncRNA node are the first number of lncRNA nodes that have a feature similarity greater than the similarity threshold with the lncRNA node and share the most diverse path with the lncRNA node. The negative sample set of each lncRNA node is the set of lncRNA nodes other than the positive samples of the lncRNA node. The positive samples of each disease node are the first number of disease nodes whose feature similarity to the disease node is greater than the similarity threshold and which share the most diverse path with the disease node. The negative sample set of each disease node is the set of disease nodes other than the positive samples of the disease node.

13. The method as described in claim 9, characterized in that, The diffusion model is used for: In each iteration, Gaussian noise is gradually added to the contextual information representation of the first lncRNA during the forward noise addition process, and the noise is gradually eliminated during the reverse noise removal process to obtain the restored representation of the first lncRNA. The first expected value of the Euclidean distance between the restored representation of the first lncRNA and the contextual information representation of the second lncRNA is calculated. Gaussian noise is gradually added to the contextual information representation of the second lncRNA during the forward noise addition process, and the noise is gradually eliminated during the reverse noise removal process to obtain the restored representation of the second lncRNA. The second expected value of the Euclidean distance between the restored representation of the second lncRNA and the contextual information representation of the first lncRNA is calculated. Calculate the loss function value of lncRNA diffusion process based on the first and second expected values; In the forward noise addition process, Gaussian noise is gradually added to the first disease context information representation, and in the reverse noise removal process, the noise is gradually eliminated to obtain the first disease recovery representation. The third expectation value of the Euclidean distance between the first disease recovery representation and the second disease context information representation is calculated. Gaussian noise is gradually added to the second disease context information representation during the forward noise addition process, and the noise is gradually eliminated during the reverse noise removal process to obtain the second disease recovery representation. The fourth expectation value of the Euclidean distance between the second disease recovery representation and the first disease context information representation is calculated. Calculate the loss function value for the disease spread process based on the third and fourth expected values; The total loss function value of the diffusion model is obtained by adding the loss function value of the lncRNA diffusion process and the loss function value of the disease diffusion process.

14. The method as described in claim 9, characterized in that, The integrated representation learning module is used for: In each iteration, the first query vector is calculated based on the first lncRNA context information, and the first bond vector and the first value vector are calculated based on the second lncRNA context information. The first attention score is obtained by performing a dot product operation on the first query vector and the first key vector, combined with normalization using the softmax function. The first attention score is weighted and summed with the first value vector to obtain the enhanced representation of the lncRNA structure view guided by the metapath view. The second query vector is calculated based on the contextual information of the second lncRNA, and the second bond vector and the second value vector are calculated based on the contextual information of the first lncRNA. The second attention score is obtained by performing a dot product operation on the second query vector and the second key vector, combined with normalization using the softmax function. The second attention score is weighted and summed with the second value vector to obtain the enhanced lncRNA metapath view representation guided by the structural view. The enhanced lncRNA structural view guided by the metapath view and the enhanced lncRNA metapath view guided by the structural view are spliced ​​together to obtain the lncRNA fusion characterization, which is used as the current lncRNA characterization for the next iteration. The third query vector is calculated based on the first disease context information representation, and the third key vector and third value vector are calculated based on the second disease context information representation; The third attention score is obtained by performing a dot product operation between the third query vector and the third key vector, combined with normalization using the softmax function. The third attention score and the third value vector are weighted and summed to obtain the enhanced representation of the disease structure view after the meta-path view is guided by the meta-path view; the fourth query vector is calculated based on the second disease context information representation, and the fourth key vector and the fourth value vector are calculated based on the first disease context information representation. The fourth attention score is obtained by performing a dot product operation on the fourth query vector and the fourth key vector, combined with normalization using the softmax function. The fourth attention score and the fourth value vector are weighted and summed to obtain the disease metapath view enhancement representation guided by the structural view. The enhanced representation of the disease structure view guided by the meta-path view and the enhanced representation of the disease meta-path view guided by the structure view are concatenated to obtain the disease fusion representation, which is then used as the current disease representation for the next iteration.

15. The method as described in claim 9, characterized in that, The iterative update module is specifically used for: in each iteration, determining the auxiliary task loss function value based on the total loss function value of the contrastive learning and the total loss function value of the diffusion model; adding the auxiliary task loss function value to the associated prediction loss function value to obtain the global total loss function value; calculating the gradient of the global total loss with respect to all training parameters through backpropagation; and updating all training parameters once by the optimizer along the gradient descent direction.

16. A device for predicting the association between lncRNA and disease, characterized in that, include: The heterogeneity graph construction module is used to construct a multi-characteristic heterogeneity graph based on lncRNA-disease association data, lncRNA-miRNA interaction data, miRNA-disease association data, lncRNA sequence data, miRNA sequence data, and a medical subject thesaurus description of the disease, using a pre-trained large RNA model. The multi-characteristic heterogeneity graph is used to describe the initial characterization of lncRNA, the initial characterization of miRNA, and the initial characterization of the disease. The structural enhancement processing module is used to perform structural enhancement processing on multi-representation heterogeneous graphs based on feature constraints and reachability constraints to obtain structurally enhanced heterogeneous graphs. The meta-path-level enhancement processing module is used to perform meta-path-level enhancement processing on multi-representation heterogeneous graphs to obtain meta-path-level enhanced heterogeneous graphs. The negative sample screening module is used to screen predicted negative samples based on lncRNA-disease association data, initial lncRNA characterization, and initial disease characterization. The negative sample screening module is a reinforcement learning agent. The association prediction model training module is used to train the association prediction model by performing multiple rounds of iterations until a preset number of iterations is reached. In each iteration, based on the initial lncRNA characterization, initial miRNA characterization, initial disease characterization, structural enhancement isomorphism map, and metapath-level enhancement isomorphism map, the total loss function value of contrastive learning and the total loss function value of diffusion model are obtained through contrastive learning and diffusion model. The lncRNA-disease association pair characterization is obtained through ensemble characterization learning. Based on positive samples and predicted negative samples, the predicted value of the lncRNA-disease association pair characterization is obtained, and the association prediction loss function value is calculated. Based on the total loss function value of contrastive learning, the total loss function value of diffusion model, and the loss function value of association prediction, the training parameters of association prediction model are updated through backpropagation; The association prediction model application module is used to input the lncRNA-disease association pairs to be predicted into the trained association prediction model to obtain the predicted values ​​of the lncRNA-disease associations.

17. A computer 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 computer program, it implements the method of any one of claims 1 to 15.

18. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that, when executed by a processor, implements the method of any one of claims 1 to 15.

19. A computer program product, characterized in that, The computer program product includes a computer program that, when executed by a processor, implements the method of any one of claims 1 to 15.

Images