IncRNA-cancer associated prediction method based on network motif consistency and carcinogenic coordinated regulation predictor

Abstract

The invention belongs to the technical field of bioinformatics and artificial intelligence crossing, and particularly relates to an lncRNA-cancer association prediction method based on a network motif consistency and carcinogenic coordinated regulation predictor, which comprises the following steps: firstly, acquiring lncRNA nucleotide sequence data in a public database, lncRNA and cancer association information and cancer CAG data in disease subject terms; the method comprises the following steps: respectively constructing an lncRNA sequence feature matrix, an lncRNAGIP similarity matrix, a cancer GIP similarity matrix and a cancer semantic similarity matrix; carrying out splicing operation on the lncRNA attribute feature matrix to obtain a comprehensive lncRNA attribute feature matrix; carrying out fusion operation on the cancer attribute feature matrix to obtain a comprehensive cancer attribute feature matrix; on the basis, potential positive samples and real negative samples can be fundamentally distinguished, and the problem that when a random negative sampling strategy is generally adopted in an existing method, unfound potential correlation is easily mismarked as the negative samples, and therefore model training is misguided is solved.

Description

Technical Field

[0001] This invention relates to the field of bioinformatics and artificial intelligence, specifically to a method for predicting lncRNA-cancer associations based on network motif consistency and oncogenic synergistic regulation predictors. Background Technology

[0002] Long non-coding RNAs (lncRNAs) are a class of functional RNA molecules widely distributed throughout the genome. They regulate gene expression at the epigenetic, transcriptional, and post-transcriptional levels through various mechanisms, participating in core life processes such as cell proliferation, differentiation, and apoptosis. In recent years, increasing research has shown that dysregulation or dysfunction of lncRNA expression is closely related to the occurrence, development, metastasis, and drug resistance of various complex human diseases, particularly cancers. Therefore, accurate and efficient identification of lncRNAs associated with specific cancers is crucial not only for elucidating the complex molecular mechanisms of cancer but also for providing key clues for developing novel cancer diagnostic biomarkers and precision therapeutic targets.

[0003] Traditional wet bioassay methods for identifying lncRNA-cancer associations, while the gold standard for final validation, face challenges such as long experimental cycles, high costs, and limited throughput, making them unsuitable for large-scale screening. With the accumulation of biomedical big data and the rapid development of artificial intelligence, developing computational biology-based methods to predict potential lncRNA-cancer associations has become a research hotspot and urgent need in this field. These computational methods can utilize existing multi-source biological data, learn potential patterns through deep learning models, and thus predict the most likely associations, providing high-value candidate targets for subsequent experimental validation and greatly improving research efficiency.

[0004] In the development of computational methods for lncRNA-cancer association prediction, early approaches primarily employed strategies based on network propagation or traditional machine learning. Network propagation methods, such as random walks or label propagation, infer unknown associations by spreading information through constructed similarity and association networks. However, the performance of these methods is highly dependent on the quality of network construction, easily affected by network sparsity, noise, and bias from known association data, and struggles to capture complex nonlinear dependencies between nodes.

[0005] In recent years, graph neural networks (GNNs), as a powerful deep learning tool, have become a mainstream technique for processing graph-structured data and mining biological network relationships. Models such as graph convolutional networks (GCNs) and graph attention networks (GATs) can directly perform end-to-end learning on lncRNA-cancer association networks, automatically capturing the topological neighborhood information and attribute features of nodes to generate low-dimensional embedding representations. Through its information propagation mechanism, GNNs can theoretically learn complex relationship patterns between nodes.

[0006] However, supervised learning models based on GNNs still face three major challenges: First, known lncRNA-cancer association data are extremely sparse, making it difficult to provide sufficient supervisory signals; second, biological experiments usually only verify the existence of associations, lacking clear evidence of non-association, leading to difficulties in selecting negative samples, which introduces huge noise and uncertainty; third, graph information aggregation is coarse, ignoring the inherent structure of biological networks, resulting in the dilution of signals at key nodes, and when judging specific pairs, simply concatenating the feature vectors of the two results in an overly simplistic association representation used for prediction.

[0007] Specifically, the existing technology has the following main drawbacks: (1) Model training relies excessively on extremely sparse known association labels: Many existing predictive models, especially GNN models using supervised learning paradigms, rely heavily on known, experimentally validated lncRNA-cancer associations. However, in the vast space of lncRNA and cancer combinations, the number of these validated associations is extremely small. Model training based on such sparse positive sample labels makes it difficult to fully learn complex and universal association patterns, resulting in limited generalization ability of the model, especially when predicting new, under-studied lncRNA or cancer associations.

[0008] (2) Obtaining positive and negative samples in biological experiments is very costly. Existing methods generally adopt a strategy of random sampling from unlabeled lncRNA-cancer pairs to construct negative sample sets, resulting in unreliable training signals. However, this random sampling strategy has a serious drawback: the unlabeled samples may contain a large number of undiscovered potential positive samples, called false negative samples. These false negative samples, which are incorrectly labeled as negative samples, will inject a large amount of noise and ambiguous signals into the model after being used for model training, misleading the optimization process and affecting the stability of the model and the accuracy of the prediction results.

[0009] (3) The graph information aggregation method is crude and the association representation is too simple: First, when using the node embeddings learned by GNN for final prediction, in order to obtain the overall representation for prediction from the graph, the commonly used global average pooling or max pooling operations compress all node information in a one-size-fits-all manner, which makes it easy to ignore the key structural information and the unique contribution of important nodes in the graph, resulting in information loss; Second, when judging the association strength of specific lncRNA-cancer pairs, existing methods usually only input the classifier after simple feature vector concatenation, which makes the model fail to effectively model the complex and high-order nonlinear interaction relationship that may exist between lncRNA embedding and cancer embedding. The representation of the association relationship is too simple, which limits the model's ability to capture deep association patterns, thus affecting the prediction accuracy. Summary of the Invention

[0010] To solve the above-mentioned technical problems, the present invention provides the following technical solution: A method for predicting lncRNA-cancer associations based on network motif consistency and oncogenic synergistic regulation predictors includes the following specific steps: S1. Obtain lncRNA nucleotide sequence data, lncRNA-cancer association information, and cancer CAG data from publicly available databases, and construct lncRNA sequence feature matrices accordingly. lncRNAGIP similarity matrix Cancer GIP Similarity Matrix Cancer semantic similarity matrix ; S2, will and The splicing operation was performed to obtain a comprehensive lncRNA attribute feature matrix. ;Will and A fusion operation is performed to obtain a comprehensive cancer attribute feature matrix. ; S3, will The first comprehensive lncRNA attribute feature matrix was obtained through two sets of random data augmentation operations. Second integrated lncRNA attribute feature matrix ;Will The first comprehensive cancer attribute feature matrix was obtained through two sets of random data augmentation operations. Second comprehensive cancer attribute feature matrix ;Will and Two sets of node encoding vectors are obtained through the GNN encoder. and ;Will and Two sets of node encoding vectors are obtained through the GNN encoder. and ; S4, based on and A network motif consistency sampling strategy was used to construct positive and negative sample pools for lncRNAs. ;based on and A network motif consistency sampling strategy was used to construct a positive and negative sample pool for cancer. ; S5, based on , as well as By performing contrastive learning and optimizing the results using a dual contrastive loss mechanism, a high-discriminative lncRNA feature matrix is ​​obtained. ;based on , as well as By performing contrastive learning and optimizing the results using a dual contrastive loss mechanism, a high-discriminative cancer feature matrix is ​​obtained. ; S6, based on and The lncRNA-cancer association prediction score was obtained through the oncogenic co-regulation predictor. .

[0011] As a preferred embodiment of the lncRNA-cancer association prediction method based on network motif consistency and oncogenic synergistic regulation predictor described in this invention, the specific steps of S1 are as follows: S11: Obtain lncRNA nucleotide sequence data and lncRNA-cancer association information, and construct a lncRNA-cancer association matrix. ; S12, Obtain the semantic hierarchical structure data of cancer and construct the directed acyclic graph (CAG) of cancer. S13, Constructing a lncRNA sequence feature matrix based on lncRNA nucleotide sequence data. ; S14, based on Calculate any two lncRNAs and GIP similarity between Thus, the lncRNAGIP similarity matrix was obtained. ; Calculate any two cancers and GIP similarity between This yields the cancer GIP similarity matrix. ; S15, based on CAG, calculates two cancers. and semantic similarity score between And from this, the semantic similarity matrix of cancer is obtained. , It is a dimension of The matrix.

[0012] As a preferred embodiment of the lncRNA-cancer association prediction method based on network motif consistency and oncogenic synergistic regulation predictor described in this invention, the specific steps of S2 are as follows: S21, for and Feature splicing operation is performed to obtain the lncRNA comprehensive attribute feature matrix. , It is a dimension of Matrix; S22, for and A weighted fusion operation is performed to obtain the comprehensive cancer attribute feature matrix. , It is a dimension of The matrix.

[0013] As a preferred embodiment of the lncRNA-cancer association prediction method based on network motif consistency and oncogenic synergistic regulation predictor described in this invention, the specific steps of S3 are as follows: S31, to Perform node feature masking and random edge discarding operations to obtain the comprehensive lncRNA attribute feature matrix. and ;right By performing node feature masking and random edge discarding operations, a comprehensive cancer attribute feature matrix is ​​obtained. and ; S32, based on and The node encoding vector is obtained using a GNN encoder. ;based on and The node encoding vector is obtained using a GNN encoder. ;based on and The node encoding vector is obtained using a GNN encoder. ;based on and The node encoding vector is obtained using a GNN encoder. .

[0014] As a preferred embodiment of the lncRNA-cancer association prediction method based on network motif consistency and oncogenic synergistic regulation predictor described in this invention, the specific steps of S4 are as follows: S41, Construct a globally learnable prototype matrix ,in It is a dimension of The matrix is ​​K, where K is the preset total number of network motif centers. It is the dimension of each vector; S42, for , and Using NMCS, a positive and negative sample pool for lncRNA was constructed. ; S43, to , and Using NMCS, construct positive and negative sample pools for cancer .

[0015] As a preferred embodiment of the lncRNA-cancer association prediction method based on network motif consistency and oncogenic synergistic regulation predictor described in this invention, the specific steps of S5 are as follows: S51, based on and For any lncRNA Calculate the dual-objective loss ; S52, based on and For any cancer Calculate the dual-target loss ; S53, based on and The learnable parameters are updated through the total loss function. ; S54, using a GNN encoder, for and Perform multi-layer graph convolution operations to obtain and .

[0016] As a preferred embodiment of the lncRNA-cancer association prediction method based on network motif consistency and oncogenic synergistic regulation predictor described in this invention, the specific steps of S6 are as follows: S61, for and The splicing operation was performed to obtain the lncRNA-cancer co-feature matrix. ; S62, for Perform a functional cluster weighted convergence operation to construct a global oncogenic regulatory landscape vector. ; S63, based on and Constructing deep carcinogenic fingerprints ; S64, based on and By jointly optimizing the total loss function Update the learnable parameter set ; S65, using Calculate lncRNA With cancer lncRNA-cancer association prediction score .

[0017] Compared with existing technologies: 1. This invention constructs a positive and negative sample pool for lncRNAs and a positive and negative sample pool for cancers with network motif consistency through the Network Motif Consistency Sampling (NMCS) strategy. By calculating the motif center affinity spectrum, it accurately identifies lncRNA nodes and cancer nodes with similar network motifs, thereby fundamentally distinguishing potential positive samples from true negative samples. This avoids the problem that existing methods, which generally use random negative sampling strategies, are prone to mislabeling undiscovered potential associations as negative samples, thus misleading model training. 2. This invention introduces relative potential boundary loss and motif homogenization loss through the dual contrast loss optimization mechanism DOCM, which optimizes the separation between network motifs and the homogenization within motifs in parallel, significantly enhancing the discriminative power of lncRNA-cancer synergistic features and avoiding the overfitting and poor generalization problems caused by the excessive reliance on extremely sparse known associated labels in existing supervised learning models. 3. The oncogenic synergistic regulation predictor SMP designed in this invention includes three levels of improvement: First, FCWA utilizes dynamic clustering and attention mechanisms to construct a global oncogenic regulation landscape vector, achieving accurate preservation of local functional structures and avoiding the information loss problem caused by traditional global average pooling operations; Second, MFIM deeply analyzes complex nonlinear regulation patterns through multi-view interactive computation, avoiding the problem of existing methods relying solely on simple feature concatenation, resulting in limited association representation capabilities and an inability to capture deep synergistic effects; Finally, MJO learns the deep interaction relationship between lncRNA and cancer through structured information convergence, avoiding the problem of features learned by the model on sparse data deviating from the real biological network topology, significantly improving the biological interpretability and robustness of the prediction results. Attached Figure Description

[0018] Figure 1 This is a flowchart of the technical solution of the present invention; Figure 2 For the present invention Build a flowchart; Figure 3 For the present invention Build a flowchart; Figure 4 This is a structural diagram of the carcinogenic synergistic regulation predictor of the present invention. Detailed Implementation

[0019] To make the objectives, technical solutions, and advantages of the present invention clearer, the embodiments of the present invention will be described in further detail below with reference to the accompanying drawings.

[0020] This invention provides a method for predicting lncRNA-cancer associations based on network motif consistency and oncogenic synergistic regulation predictors. Please refer to [link to related documentation]. Figures 1-4 The specific steps are as follows: S1. Obtain lncRNA nucleotide sequence data, lncRNA-cancer association information, and cancer CAG data from publicly available databases, and construct lncRNA sequence feature matrices accordingly. lncRNAGIP similarity matrix Cancer GIP Similarity Matrix Cancer semantic similarity matrix .

[0021] The specific steps of S1 are as follows: S11: Obtain lncRNA nucleotide sequence data, lncRNA-cancer association information, and cancer semantic hierarchy data to construct a lncRNA-cancer association matrix. Disease directed acyclic graph (CAG); Nucleotide sequence data of lncRNAs and information on the association between lncRNAs and cancer were obtained from public databases to construct a lncRNA-cancer association matrix. :

[0022] in, Indicates lncRNA With cancer Is there a connection?

[0023] Example: In one embodiment of the present invention, information on the association between lncRNA and cancer is obtained from the Lnc2Cancer 3.0 database, which is an authoritative database focused on recording experimentally verified associations between lncRNA and human cancer; lncRNA nucleotide sequence data is obtained from the GENCODE database, which is a high-quality annotation database for identifying human genome features.

[0024] S12, Obtain the semantic hierarchical structure data of cancer and construct the directed acyclic graph (CAG) of cancer. Example: In one embodiment of the present invention, semantic hierarchical structure data of cancer is obtained from the MeSH database. The MeSH database is an authoritative thesaurus maintained and published by the U.S. National Library of Medicine, providing a rigorous hierarchical structure system. A CAG can be directly constructed using a recursive parsing algorithm based on MeSH tree coding.

[0025] S13, Constructing a lncRNA sequence feature matrix based on lncRNA nucleotide sequence data. ; Based on lncRNA nucleotide sequence data, the k-mer algorithm is applied to decompose each lncRNA sequence into a series of overlapping short nucleotide fragments. Then, a fixed-dimensional frequency vector is generated according to the frequency of all fragment combinations in each lncRNA sequence. Finally, all vectors are stacked to obtain the lncRNA sequence feature matrix. ,in It is a dimension of The matrix, Indicates the number of lncRNAs, This represents the dimension of the frequency vector.

[0026] Among them: lncRNA nucleotide sequence data is a raw data set obtained from public databases. Any element in this set represents a lncRNA sequence, and the lncRNA sequence is a base string representing a single lncRNA molecule.

[0027] The k-mer algorithm is a well-known fundamental algorithm in bioinformatics used for processing and analyzing biological macromolecule sequences. Its core operating mechanism involves setting a fixed length... The algorithm uses a sliding window to perform a step-by-step scan of the nucleotide sequence from beginning to end, thereby breaking down long sequences into a series of continuous and overlapping short subsequences. This algorithm can discretize complex linear sequence information into feature units containing local structural information.

[0028] S14, based on Calculate any two lncRNAs and GIP similarity between Thus, the lncRNAGIP similarity matrix was obtained. ; Calculate any two cancers and GIP similarity between This yields the cancer GIP similarity matrix. ; Among them, GIP kernel similarity (Graphical Interface Profile Kernel Similarity) is a commonly used and well-known technique in the fields of bioinformatics and graph data mining to infer the similarity of cancer nodes based on known associations. Its core principle is to measure similarity by quantifying the differences in the association contours of cancer nodes in the association network.

[0029] The general formula for GIP similarity is:

[0030] in: Representing entities The interaction spectrum, Representing entities The interaction spectrum, Representing entities The interaction spectrum, This represents the square of the Euclidean distance. It is a normalization parameter used to control the bandwidth of the kernel function.

[0031] In this step, the general formula for GIP similarity is used to calculate lncRNA. and GIP similarity between and cancer and GIP similarity between : (1) For Calculation based on: For any two lncRNAs and ,use and In the general formula for replacing GIP similarity and ,use Interaction profile with any lncRNA Replace the general formula for GIP similarity in and It can be calculated ; (2) For Calculation based on: For any two cancers and ,use and In the general formula for replacing GIP similarity and ,use Interaction spectrum with any cancer Replace the general formula for GIP similarity in and You can get .

[0032] S15, based on CAG, calculates two cancers. and semantic similarity score between And from this, the semantic similarity matrix of cancer is obtained. , It is a dimension of Matrix; Among them, semantic similarity calculation based on CAG is a mature technology in the field of bioinformatics used to quantify the degree of semantic association between biological entities. Its core principle is to utilize the inherent hierarchical structure of CAG in ontology, and organize biological entities layer by layer from abstract to concrete according to semantic categories through directed edges. Then, by measuring the overlap, distance or information content of the entities in the CAG, the semantic similarity can be quantitatively calculated.

[0033] Example: In the experiments of this invention, a method containing... lncRNAs and A dataset of cancers was generated by obtaining the sequences of the aforementioned 861 lncRNAs and MeSH CAG information for 253 cancers, constructing an 861×253-dimensional association matrix. This includes 4517 known lncRNA-cancer associations with a value of 1. Set in S13. The 861×64 dimensional sequence feature matrix was calculated. In S14, based on the correlation matrix The 861×861-dimensional lncRNA GIP similarity matrix was calculated. and a 253×253 dimensional cancer GIP similarity matrix In S15, based on MeSH CAG data, a 253×253 dimensional cancer semantic similarity matrix was calculated using a graph topology-based semantic similarity algorithm. .

[0034] S2, will and The splicing operation was performed to obtain a comprehensive lncRNA attribute feature matrix. ;Will and A fusion operation is performed to obtain a comprehensive cancer attribute feature matrix. .

[0035] The specific steps of S2 are as follows: S21, for and Feature splicing operation is performed to obtain the lncRNA comprehensive attribute feature matrix. , It is a dimension of Matrix; S22, for and A weighted fusion operation is performed to obtain the comprehensive cancer attribute feature matrix. , It is a dimension of The matrix.

[0036] Weighted fusion is a common multimodal information integration strategy. It balances the contributions of different information sources to the final task by linearly combining them using adjustable weight parameters. Its mathematical definition is: , To balance the hyperparameters.

[0037] Example: In the experiment of the present invention, S21 is executed, and... and Horizontal splicing was performed to obtain the lncRNA comprehensive attribute feature matrix. Its dimensions are 861×(64 + 861), which is 861×925. Execute S22 to set the hyperparameters. = 0.5, for and Perform weighted summation and fusion, that is The comprehensive attribute feature matrix of cancer was obtained. Its dimensions are 253×253.

[0038] S3, will The first comprehensive lncRNA attribute feature matrix was obtained through two sets of random data augmentation operations. Second integrated lncRNA attribute feature matrix ;Will The first comprehensive cancer attribute feature matrix was obtained through two sets of random data augmentation operations. Second comprehensive cancer attribute feature matrix ;Will and Two sets of node encoding vectors are obtained through the GNN encoder. and ;Will and Two sets of node encoding vectors are obtained through the GNN encoder. and .

[0039] The specific steps of S3 are as follows: S31, to Perform node feature masking and random edge discarding operations to obtain the comprehensive lncRNA attribute feature matrix. and ;right By performing node feature masking and random edge discarding operations, a comprehensive cancer attribute feature matrix is ​​obtained. and ; Among them, node feature masking and random edge dropping are two well-known data augmentation techniques widely used in graph self-supervised contrastive learning. Their purpose is to generate feature matrices that have the same semantics as the original data but different representations. This invention constructs two feature matrices with heterogeneous sources by applying node feature masking and random edge dropping to the same feature matrix respectively, driving the model to learn robustness to feature loss and topological changes simultaneously.

[0040] S32, based on and The node encoding vector is obtained using a GNN encoder. ;based on and The node encoding vector is obtained using a GNN encoder. ;based on and The node encoding vector is obtained using a GNN encoder. ;based on and The node encoding vector is obtained using a GNN encoder. .

[0041] S321: Yes Sparsification was performed to obtain the lncRNA adjacency matrix. ,in It is a dimension of The matrix; for After sparsification, the cancer adjacency matrix is ​​obtained. ,in It is a dimension of Matrix; S322: Using a GNN encoder, for and Perform multi-layer graph convolution operations to obtain node encoding vectors. ; Will The input matrix of the 0th layer of the GNN encoder ,in It is a dimension of The matrix, Input dimension; set The graph convolutional layer is used to update the node encoding vector, where the first layer is the first layer. The layer update process can be represented as:

[0042] in: yes An adjacency matrix with self-loops was added. yes The inverse of the degree matrix; It is the first The learnable weight matrix of the layer, For the first The input dimension of the layer; It is the first Layer bias vector; It is a non-linear activation function; go through After layer propagation, the final high-order node matrix of lncRNA is obtained. That is, node encoding vector d is the output dimension; S323: Using a GNN encoder, for and Perform multi-layer graph convolution operations, the same process as S322, to obtain node encoding vectors. ; S324: Using a GNN encoder, for and Perform multi-layer graph convolution operations to obtain node encoding vectors. : Will As the input matrix of the 0th layer of the GNN ,in It is a dimension of The matrix, For input dimensions; In the GNN encoder, set The graph convolutional layer is used to update the node encoding vector, where the first layer is the first layer. The layer update process can be represented as:

[0043] in: yes An adjacency matrix with self-loops was added. yes The inverse of the degree matrix; It is the first The learnable weight matrix of the layer; It is the first Layer bias vector; It is a non-linear activation function.

[0044] go through After layer propagation, the final high-order node matrix of the cancer is obtained. That is, node encoding vector .

[0045] S325: Using a GNN encoder, for and Perform multi-layer graph convolution operations, the same process as S324, to obtain node encoding vectors. ; Among them, the Graph Neural Network (GNN) encoder is a well-known technique in the field of deep learning for processing graph-structured data. It updates the lncRNA node encoding vector and cancer node encoding vector by aggregating neighborhood information, and can effectively learn topological structure and attribute features.

[0046] Example: In the experiment of this invention, S31 is executed: The first set of data augmentation operations uses attribute perturbation: setting the feature mask probability. .right , Obtain by performing feature masking , The second set of data augmentation operations uses structural perturbation: setting the edge drop probability. .right , Obtain by discarding edges , .

[0047] Execute S321, based on Build a sparse adjacency matrix by retaining the 10 neighbors with the highest similarity in each row. , dimension ;based on Build a sparse adjacency matrix by retaining the 10 neighbors with the highest similarity in each row. , dimension Execute steps S322-S325, setting the number of GNN encoder layers for both lncRNA and cancer to be [number missing]. Set the output dimension of the first layer of the GNN. The output dimension of the second layer In S322 and S323, the learnable parameters are initialized, and the first layer weights are set. Second layer weights In S324 and S325, the learnable parameters are initialized, and the first layer weights are set. Second layer weights .Will and The input is a GNN encoder, which passes through two convolutional layers to obtain the node encoding. , All dimensions ;Will and The input is a GNN encoder, which passes through two convolutional layers to obtain the node encoding. , All dimensions .

[0048] S4, based on and A network motif consistency sampling strategy was used to construct positive and negative sample pools for lncRNAs. ;based on and A network motif consistency sampling strategy was used to construct a positive and negative sample pool for cancer. .

[0049] This invention designs a Network Motif Consistency Sampling (NMCS) strategy, which is a sample construction strategy that incorporates the principle of network motif consistency to construct positive and negative lncRNA sample pools. and cancer positive and negative sample pools The technical advantage of NMCS lies in solving the problem of misjudging false negative samples, providing the model with more reliable and biologically consistent self-supervised training signals.

[0050] The specific steps of S4 are as follows: S41, Construct a globally learnable prototype matrix ,in It is a dimension of The matrix is ​​K, where K is the preset total number of network motif centers. It is the dimension of each vector;

[0051] Where: any , , is the A learnable network motif center vector Its initial value is generated by random initialization using a normal distribution.

[0052] in: In subsequent S42 and S43, it serves as a shared global network motif reference frame. Although lncRNAs and cancers are heterogeneous biological entities with completely different properties, from a network structure perspective, they share common network motifs.

[0053] S42, for , and Using NMCS, a positive and negative sample pool for lncRNA was constructed. ; Construct based on NMCS strategy The execution process is as follows Figure 2 As shown.

[0054] S421: For any lncRNA Calculate the query vector :

[0055] in, It is a learnable nonlinear transformation used to map the node encoding vector of lncRNA to... 3D space; It is lncRNA Node encoding vector; and These are learnable parameters; S422: Calculate attention score :

[0056] S423: Calculate the affinity spectrum vector of lncRNA motif center : based on Use the Softmax function to Normalize the attention scores to obtain lncRNAs. motif center affinity spectrum vector ,in It is a dimension of The matrix:

[0057] in, Indicates lncRNA Belonging to the The probability of a structural motif center; It is a temperature hyperparameter used to control the sharpness of the distribution; Wherein: the Softmax function is a well-known normalization function, and its mathematical formula is: The attention mechanism here differs from GAT; it calculates the similarity between lncRNA node attributes and the global prototype, rather than neighborhood attention between lncRNA nodes.

[0058] S424: For all Perform steps S421-S423 to obtain the motif center affinity vectors for all lncRNAs. .

[0059] S425: Based on , and Calculate any lncRNA and Adaptive motif cross-correlation similarity between :

[0060] in: It is the basic order cross-correlation matrix. These are learnable parameters. It is the identity matrix; S426: Based on Constructing positive and negative sample pools for lncRNA : (1) If Above the high threshold Then Included Positive sample set:

[0061] in, Indicates in In the middle, the first The node encoding vector of each lncRNA node; This indicates the current training batch.

[0062] (2) If Below the low threshold Then Included Negative sample set:

[0063] in: Between and The node encoding vectors between them will be treated as fuzzy samples and ignored to further improve the quality of negative samples.

[0064] For all and By performing the above judgment operations, a positive lncRNA sample pool is obtained. and lncRNA negative sample pool Finally, positive and negative sample pools of lncRNA were obtained. ; Where: threshold and It is a key hyperparameter for controlling signal quality. It should be set to a higher value to ensure that the lncRNA nodes entering the positive sample pool are compatible with... It has an extremely high confidence level in structural similarity; It should be set to a low value to ensure that the lncRNA nodes entering the negative sample pool are structurally identical to... The differences are significant.

[0065] S43, to , and Using NMCS, construct positive and negative sample pools for cancer Building based on NMCS strategy The execution process is as follows Figure 3 As shown.

[0066] S431: For any cancer Calculate the query vector :

[0067] in, It is a learnable nonlinear transformation used to map the node encoding vector of cancer to... 3D space; It's cancer. Node encoding vector; and These are learnable parameters; S432: Calculate attention score :

[0068] Among them: in S421 and S431, through and By mapping the node-coding vectors of lncRNAs and cancer cells to the same dimensional space, the same node-coding vector can be used. .

[0069] S433: Calculate the affinity spectrum vector of cancer motif centers : based on Use the Softmax function to Normalize the attention scores to obtain cancer motif center affinity spectrum vector ,in It is a dimension of The matrix:

[0070] in, Indicates cancer Belonging to the The probability of a structural motif center; S434: For all Execute steps S431~S433 to obtain the motif center affinity spectrum vectors for all cancers. ; S435: Based on , and Calculate any cancer and Adaptive motif cross-correlation similarity between :

[0071] in: These are learnable parameters; S436: Based on Constructing positive and negative sample pools for cancer : (1) If Above the high threshold Then Included Positive sample set:

[0072] in, Indicates in In the middle, the first The node encoding vector of each cancer node; This indicates the current training batch.

[0073] (2) If Below the low threshold Then Included Negative sample set:

[0074] in: Between and The node encoding vectors between them will be treated as fuzzy samples and will be ignored to further improve the quality of negative samples.

[0075] Example: and The value needs to be determined based on the experimental conditions. In the experimental scheme of this invention, the value is taken as... 0.9, .

[0076] For all and By performing the above judgment operations, a positive sample pool of cancer samples is obtained. and cancer negative sample pool Ultimately, positive and negative sample pools for cancer were obtained. ; Wherein: In this invention, all learnable parameters in S4 are denoted as .

[0077] Example: In the experiment of this invention, step S41 is executed to set the number of structural motif centers. motif space mapping dimension Initialize the prototype matrix. Execute S42, for ,Will Mapped to query vector Set temperature The motif center affinity spectrum was obtained after Softmax normalization. In the batch, there were also... and Calculations yielded and .set up , Initialize learnable parameters .by For example: Calculation ;calculate ;calculate Finally, the following was constructed: , Perform the above operations on all lncRNA nodes to obtain the final lncRNA positive and negative sample pools. Execute S43 for cancer. Perform the same operation logic as S42 and initialize the learnable parameters. and using independent mapping parameters , construct cancer positive and negative sample sets and Perform the above operations on all cancer nodes to obtain the final positive and negative cancer sample pool. .

[0078] S5, based on , as well as By performing contrastive learning and optimizing the results using a dual contrastive loss mechanism, a high-discriminative lncRNA feature matrix is ​​obtained. ;based on , as well as By performing contrastive learning and optimizing the results using a dual contrastive loss mechanism, a high-discriminative cancer feature matrix is ​​obtained. .

[0079] This invention designs a Dual-Objective Contrastive Optimization Mechanism (DOCM), a self-supervised loss function based on the contrastive learning paradigm. It collaboratively optimizes the inter-motif separation and intra-motif homogenization in the feature space by introducing the Relative Potential Boundary Loss (RPML) and the Intra-Motif Homogenization Loss (MCUL). In DOCM, leveraging… , To optimize updatable parameters, RPML is used to separate network motifs in space, aligning different network motifs, while MCUL is used to achieve intra-motif homogenization. This allows the learning of lncRNA and cancer features that combine motif consistency with individual discriminative power. DOCM significantly enhances the discriminative power and homogenization of the lncRNA and cancer feature matrices, reduces dependence on sparse labels, and exhibits higher generalization ability.

[0080] The specific steps of S5 are as follows: S51, based on and For any lncRNA Calculate the dual-target loss ;

[0081] in, yes The relative potential energy boundary loss; yes Intra-motif homogenization loss; It is a positive hyperparameter used to balance the two operations of inter-site separation and intra-site homogenization in the network; The feature space constructed by DOCM allows the model to accurately map lncRNAs with few tags to the correct cancer functional clusters based on the similarity of their structural motifs, significantly improving the ability to discover potential novel lncRNA biomarkers.

[0082] (1) Calculate the relative potential energy boundary loss :

[0083] in, It is a positive hyperparameter, representing the safety boundary; Where: If If the loss is 0, no gradient is generated; if The loss is positive. The model will generate gradients, reducing At the same time improve until the boundary is satisfied. For example, in complex lncRNA-cancer association networks, there are numerous non-functional weak associations. By setting... The model forces positive samples to have a significantly higher affinity than negative samples, enabling it to define clear discrimination boundaries in a high-dimensional feature space. This is analogous to setting strict screening thresholds in biological experiments, filtering out spurious associations caused by data sparsity and ensuring the model captures robust interaction patterns with strong biological significance.

[0084] yes The positive sample set attracting potential energy:

[0085] in, It is the natural logarithm function; for The number of neutral samples; for Any positive sample vector of lncRNA in the dataset; It is a lncRNA anchor. and Elastic Magnitude-Weighted Cosine Similarity (EMWCS) between them:

[0086] in, Represents the Euclidean norm; It is a positive temperature hyperparameter; It is the hyperbolic tangent activation function; yes Negative sample set repulsion potential:

[0087] in, for The number of negative samples; for Any negative sample vector of lncRNA in the dataset; It is a lncRNA anchor. and The similarity between them was calculated using EMWCS:

[0088] Among them, anchor points express exist The node encoding vector in.

[0089] To maintain symmetry, during model training, it is usually also based on... Using this as an anchor point, repeat the above process to calculate another set of losses, and incorporate them into the total loss function. middle.

[0090] (2) Calculate the homogenization loss within the sequence :

[0091] in, This represents the square of the Euclidean distance between two vectors; It is a positive bandwidth hyperparameter; It is a normalization factor, i.e., lncRNA. The total number of positive sample pairs in the positive sample set.

[0092] Calculated All positive lncRNA sample nodes represent the sum of the repulsive potential energies between pairs. Minimize This is equivalent to maximizing the average distance between positive lncRNA sample nodes within the network motif to combat model collapse. In biology, even a group of lncRNAs regulating the same type of cancer have different specific pathways of action. If only the distance is minimized, it is easy to map all related lncRNAs to the same point, i.e., model collapse. By introducing a repulsive potential energy, the subtle differences between positive lncRNA sample nodes are forcibly preserved, so that the prediction results can not only determine whether there is a correlation, but also preserve the uniqueness of different lncRNAs in terms of molecular mechanisms, thereby improving the model's ability to generalize and recognize different subtypes of cancer.

[0093] S52, based on and For any cancer Calculate the dual-target loss ;

[0094] in, yes The relative potential energy boundary loss; yes Intra-motif homogenization loss; It is a positive hyperparameter used to balance the two operations of inter-site separation and intra-site homogenization in the network; (1) Calculate the relative potential energy boundary loss :

[0095] in, It is a positive hyperparameter, representing the safety boundary; Where: If If the loss is 0, no gradient is generated; if The loss is positive. The model will generate gradients, reducing At the same time improve until the boundary is satisfied. ; yes The positive sample set attracting potential energy:

[0096] in, It is the natural logarithm function; for The number of neutral samples; for Any positive sample vector in the vector; Cancer anchor and The similarity between them was calculated using EMWCS: yes Negative sample set repulsion potential:

[0097] in, for The number of negative samples; for Any negative sample vector in the vector; It is a cancer anchor. and The similarity between them was calculated using EMWCS:

[0098] Among them, anchor points express exist The node encoding vector in; (2) Calculate the homogenization loss within the sequence :

[0099] in, This represents the square of the Euclidean distance between two vectors; It is a positive bandwidth hyperparameter; It is a normalization factor, i.e., cancer. The total number of positive sample pairs in the positive sample set; Calculated All positive cancer sample nodes represent the sum of the repulsive potential energy between pairs. Minimize This is equivalent to maximizing the average distance between positive cancer sample nodes within the network motif to combat model collapse.

[0100] S53, based on and The learnable parameters are updated through the total loss function. ;

[0101] in, It includes , and A set; It is the learning rate, a preset positive hyperparameter; This represents the gradient operator, used to calculate the loss function relative to the parameters. The direction of the partial derivative; This indicates an assignment operation, which assigns the calculated updated value to... This completes one iteration update; The total loss is due to the dual objectives:

[0102] in, This represents the set of all lncRNA samples sampled in the current training iteration; This represents the set of all cancer samples sampled in the current training iteration; Specifically: After completing the parameter iteration update and making the model converge, freeze all learnable parameters of the GNN encoder.

[0103] S54, using a GNN encoder, for and Perform multi-layer graph convolution operations to obtain and .

[0104]

[0105] in, , It is a dimension of Matrix; , It is a dimension of The matrix, This is the output dimension of the GNN encoder.

[0106] Example: In the experiment of this invention, the output of Example S4 (lncRNA anchor) is used. Positive sample set and negative sample pool Setting hyperparameters , , , Execute S51, first calculate Compared with positive samples The EMWCS similarity between them are respectively ,through Scaling Calculations yielded ; then calculate With negative samples The EMWCS similarity between them are respectively ,through Scaling Calculations yielded Next, calculate. ;because There are 3 elements, hence the normalization factor. ,and and The squares of the distances are respectively Calculations yielded The final total loss was obtained. Perform steps S52 and S53 to calculate and sum the aforementioned losses for all lncRNAs and cancer anchors within the batch, yielding the results. Then update using backpropagation. ;Execute S54, and The input is obtained by the GNN encoder. , dimension and , dimension .

[0107] S6, based on and The lncRNA-cancer association prediction score was obtained through the oncogenic co-regulation predictor. .

[0108] This invention designs a carcinogenic synergistic regulation predictor (SMP), which is an association prediction learner based on a multi-module collaborative architecture. It innovatively introduces functional cluster weighted convergence (FCWA) and multi-view feature interaction (MFIM) operations. The process involves first constructing a context vector from the FCWA convergence graph, then building a deep carcinogenic association fingerprint using MFIM, and finally using a multi-task joint optimization mechanism (MJO) to learn the deep interaction relationship between lncRNA and cancer at the prediction layer through structured information convergence. SMP solves the problems of coarse information convergence and simple association representation construction, and enhances the model's generalization ability and stability. The structure of SMP is as follows: Figure 4 As shown.

[0109] The specific steps of S6 are as follows: S61, for and The splicing operation was performed to obtain the lncRNA-cancer co-feature matrix. ;

[0110] in, It is a matrix concatenation operation; It is a dimension of The matrix, ; S62, for Perform a functional cluster weighted convergence operation to construct a global oncogenic regulatory landscape vector. ; S621: Using the K-means algorithm, for Perform functional cluster partitioning to obtain the functional cluster member set. : (1) Random selection lncRNA-cancer co-feature vectors Each serves as a member of its respective functional cluster. The center of mass, and make each Including only the centroid, we get ; (2) For any lncRNA-cancer co-feature vector ,calculate The Euclidean distances to each centroid, then... Assigned to the nearest centroid Corresponding functional cluster members ; (3) Update the centroid of each functional cluster member. Through calculation The mean vector of all lncRNA-cancer co-evaluative feature vectors within the dataset was calculated as follows:

[0111] (4) Repeat steps (2) and (3) until the convergence condition is met, and the functional cluster member set can be obtained. ; Example: In the experiment of this invention, the convergence condition includes the following two conditions, and meeting either one is considered to have reached the convergence condition: (1) the number of iterations reaches 100; (2) the change in the new centroid is less than 1 / 2 for 3 consecutive iterations. The change is obtained by subtracting the absolute value of the centroid from the centroid of the previous iteration from the new centroid generated in this iteration; Wherein: the centroid refers to a representative feature vector dynamically calculated for each functional cluster using the K-means algorithm. In each iteration, it serves as the geometric center of the functional cluster members, is the mean of all vectors within the functional cluster members, measures the distance between a vector and the functional cluster members, and serves as the allocation benchmark for the next round of iterations.

[0112] The K-means clustering algorithm is a classic unsupervised partitioning clustering technique in data mining and machine learning. Its core idea is to divide the dataset into subgroups through iterative optimization. The samples are arranged in non-overlapping clusters, such that samples within a cluster are as close together as possible, while samples between clusters are as far apart as possible.

[0113] In the complex mechanisms of cancer pathogenesis, different lncRNAs and cancers often cluster to form specific functional modules. Through dynamic clustering, the model can adaptively identify potential biological pathways and classify heterogeneous lncRNA-cancer co-operating nodes into functional subspaces with similar pathogenic logic. This makes subsequent operations no longer a blind global average, but a precise aggregation based on specific functional backgrounds, effectively solving the problem of large differences in pathogenesis among different cancer subtypes and the difficulty in unified modeling.

[0114] S622: Based on Construct a cluster summary vector set :

[0115] in, This is the cluster summary vector of the k-th cluster, obtained by weighted attention fusion and summation:

[0116] in, Indicates to All The attention scores are summed exponentially, and the result is a normalization factor. yes Intra-cluster attention score:

[0117] in, , and These are learnable intra-cluster attention parameters; It is the hyperbolic tangent activation function; S623: Yes and Cross-cluster attention convergence processing is performed to obtain the global oncogenic regulatory landscape vector. :

[0118] in, yes The value vector, by After linear transformation, we obtain:

[0119] in, It is a learnable value weight matrix; It is the dimension of the value vector; For all Perform a linear transformation on each cluster summary vector to construct a set of value vectors. ; It is the first Key vectors of functional clusters Attention score:

[0120] in, It is a learnable query weight matrix. yes The dimension; Depend on After linear transformation, we obtain:

[0121] in, It is a learnable key weight matrix. It is a dimension of Matrix; For all Perform a linear transformation on each cluster summary vector to construct a set of key vectors. ; S63, based on and Constructing deep carcinogenic fingerprints ; S631: lncRNA for the predicted node and cancer Feature extraction was performed to obtain the features of the lncRNA nodes to be predicted. and features of cancer nodes to be predicted ; S632: Yes and Multi-perspective interactive computation was performed to obtain lncRNA-cancer semantic co-expression features. Carcinogenic synergistic coupling characteristics and the differential distribution characteristics of lncRNA-cancer functions : (1) To and Perform vector concatenation to obtain lncRNA-cancer semantic co-expression features. :

[0122] in, It's a vector concatenation operation, which combines two vectors... Concatenating dimensional vectors into one dimensional vector; Vector concatenation is a data processing technique used to combine feature vectors from different sources.

[0123] (2) To and Calculate the Hadamard product to obtain the carcinogenic co-coupling characteristics. :

[0124] in, It represents the Hadamardi (or Hadama) stack; The Hadamard product is a vector operation that is often used to model multiplicative interactions between features.

[0125] (3) To and The absolute difference was calculated to obtain the differential distribution characteristics of lncRNA-cancer functions. :

[0126] in, This indicates taking the absolute value of each element; S633: Yes , and as well as Perform feature fusion operations to construct a deep carcinogenic association fingerprint. , :

[0127] in, It is a feature fusion operation; Multi-perspective interactive computation aims to comprehensively capture the complex nonlinear relationship patterns between lncRNA and cancer.

[0128] S64, based on and By jointly optimizing the total loss function Update the learnable parameter set ; S641: Based on Calculate the prediction loss of carcinogenic association. :

[0129] in, This represents the set of lncRNA-cancer sample pairs sampled in the current training batch; yes The total number of sample pairs in the sample; It is lncRNA and cancer The probability scores of the associations between them are used by a carcinogenic association predictor. get:

[0130] in, , yes Learnable weights It is the dimension of the hidden layer; , yes The bias parameters; These are hidden layer activation features; It is a modified linear unit activation function used to extract nonlinear features; It is the Sigmoid activation function, used to map the output to... The interval is used as the probability value; Among them, the Multilayer Perceptron (MLP) is a classic feedforward neural network model in the field of deep learning and artificial neural networks. An MLP consists of an input layer, at least one hidden layer, and an output layer. Its core computational mechanism involves linearly transforming the input feature vector through fully connected layers and then progressively stacking nonlinear activation functions.

[0131] S642: Based on Calculate the function cluster to assist in predicting loss :

[0132] in, It is the collection of all lncRNAs and cancer nodes in the current training batch; It is a lncRNA-cancer signature node The actual functional cluster The corresponding one-hot label vector, It is a dimension of The matrix, Represents the first one of the one-hot label vectors. Dimensional elements; yes Belonging to The predicted probability distribution of each functional cluster is obtained through a functional cluster discriminator. get:

[0133] in, , , , All Learnable parameters; These are hidden layer activation features; Among them, the One-hot Label Vector is a well-known data structure in machine learning and data mining used to represent discrete category information. Its encoding rule is: for a given label vector containing... A classification task involving mutually exclusive categories, where a one-hot vector is a vector with dimensions of 1. A sparse vector. If a sample belongs to the th... The first category, then the first category in the vector. The element value of dimension 1 is 1, while the element values ​​of all other dimensions are 0.

[0134] Introducing functional clusters to assist the prediction task aims to impose biological consistency constraints during model training. Because validated association data between lncRNAs and cancer are extremely sparse, a single primary prediction task can easily lead to model overfitting to a small number of known samples. By forcing the model to simultaneously predict the functional clusters to which lncRNA-cancer feature nodes belong, this invention essentially introduces a regularization mechanism, requiring that the learned features conform to the biological logic of clustering similar features.

[0135] S643: Yes and The backpropagation algorithm is used to update the learnable parameter set. :

[0136] in, This represents the set of all learnable parameters in S6 that are yet to be updated; The learning rate; For gradient operators; This indicates an assignment operation, which assigns the calculated updated value to... This completes one iteration update; It is a joint optimization of the total loss function, through the... and We obtain the following using a weighted summation operation:

[0137] in, It is a positive hyperparameter; Among them: after the joint optimization of the total loss function converges during training, accurate predictions can be obtained. .

[0138] S65, using Calculate lncRNA With cancer lncRNA-cancer association prediction score ;

[0139] Example: In the experiment of this invention, step S61 is executed to generate the lncRNA feature matrix. With cancer feature matrix By concatenating the components, we obtain the dimension as follows: LncRNA-Cancer Co-signature Matrix ; Execute S62 to set the number of function clusters. The K-means algorithm is used to... It is dynamically divided into 50 functional clusters. For each cluster... Calculate the attention score for each lncRNA-cancer co-feature node within the cluster. Then, a weighted sum is performed to obtain 50 dimensions. Cluster summary vector , construct the set Next, set the query / key dimensions. Value Dimension .according to and The calculation formula uses 50 prototype vectors. Mapped to 50 dimensions respectively key vector and 50 dimensions value vector ;according to The convergence formula is used to calculate the attention score and apply it to a vector of 50 values. After performing a weighted summation, the final dimension is... of ; Execute S63 for the lncRNA to be predicted. and cancer ,extract and All dimensions are 128. Through multi-perspective interactive computation: (1) lncRNA-cancer semantic co-expression features (2) Carcinogenic synergistic coupling characteristics (3) LncRNA-Cancer Functional Distribution Differences The dimension is 128. (The rest of the text appears to be incomplete and requires further context.) , , and Fusion to construct a deep carcinogenic fingerprint Its ultimate dimension ; Execute S64, set The structure is , The structure is Balancing hyperparameters Learning rate .Will enter Obtain the predicted score and calculate .Will enter A 50-dimensional class probability distribution is obtained. And calculate the true cluster labels obtained by k-means partitioning. Calculate the total loss. And use backpropagation to update all learnable parameters in S6.

[0140] Execute S65, and after training, use the optimized... Output the final association probability of any lncRNA-cancer pair. Taking MALAT1-non-small cell lung cancer as an example, input it into... Finally obtained This is consistent with clinical understanding.

[0141] Although the present invention has been described above with reference to embodiments, various modifications can be made and components can be replaced with equivalents without departing from the scope of the invention. In particular, as long as there is no structural conflict, the features in the disclosed embodiments can be combined with each other in any manner. The lack of an exhaustive description of these combinations in this specification is merely for the sake of brevity and resource conservation. Therefore, the present invention is not limited to the specific embodiments disclosed herein, but includes all technical solutions falling within the scope of the claims.

Claims

1. A method for predicting lncRNA-cancer associations based on network motif consistency and oncogenic synergistic regulation predictors, characterized in that, The specific steps are as follows: S1. Obtain lncRNA nucleotide sequence data, lncRNA-cancer association information, and cancer CAG data from publicly available databases, and construct lncRNA sequence feature matrices accordingly. lncRNAGIP similarity matrix Cancer GIP Similarity Matrix Cancer semantic similarity matrix ; S2, will and The splicing operation was performed to obtain a comprehensive lncRNA attribute feature matrix. ;Will and A fusion operation is performed to obtain a comprehensive cancer attribute feature matrix. ; S3, will The first comprehensive lncRNA attribute feature matrix was obtained through two sets of random data augmentation operations. Second integrated lncRNA attribute feature matrix ;Will The first comprehensive cancer attribute feature matrix was obtained through two sets of random data augmentation operations. Second comprehensive cancer attribute feature matrix ;Will and Two sets of node encoding vectors are obtained through the GNN encoder. and ;Will and Two sets of node encoding vectors are obtained through the GNN encoder. and ; S4, based on and A network motif consistency sampling strategy was used to construct positive and negative sample pools for lncRNAs. ;based on and A network motif consistency sampling strategy was used to construct a positive and negative sample pool for cancer. ; S5, based on , as well as By performing contrastive learning and optimizing the results using a dual contrastive loss mechanism, a high-discriminative lncRNA feature matrix is ​​obtained. ;based on , as well as By performing contrastive learning and optimizing the results using a dual contrastive loss mechanism, a high-discriminative cancer feature matrix is ​​obtained. ; S6, based on and The lncRNA-cancer association prediction score was obtained through the oncogenic co-regulation predictor. .

2. The lncRNA-cancer association prediction method based on network motif consistency and oncogenic synergistic regulation predictor as described in claim 1, characterized in that, The specific steps of S1 are as follows: S11: Obtain lncRNA nucleotide sequence data and lncRNA-cancer association information, and construct a lncRNA-cancer association matrix. ; S12, Obtain semantic hierarchical structure data of cancer and construct a directed acyclic graph (CAG) of cancer; S13, Constructing a lncRNA sequence feature matrix based on lncRNA nucleotide sequence data. ; S14, based on Calculate any two lncRNAs and GIP similarity between Thus, the lncRNA GIP similarity matrix was obtained. ; Calculate any two cancers and GIP similarity between This yields the cancer GIP similarity matrix. ; S15, based on CAG, calculates two cancers. and semantic similarity score between And from this, the semantic similarity matrix of cancer is obtained. , It is a dimension of The matrix.

3. The lncRNA-cancer association prediction method based on network motif consistency and oncogenic synergistic regulation predictor according to claim 1, characterized in that, The specific steps of S2 are as follows: S21, for and Feature splicing operation is performed to obtain the lncRNA comprehensive attribute feature matrix. , It is a dimension of Matrix; S22, for and A weighted fusion operation is performed to obtain the comprehensive cancer attribute feature matrix. , It is a dimension of The matrix.

4. The lncRNA-cancer association prediction method based on network motif consistency and oncogenic synergistic regulation predictor as described in claim 1, characterized in that, The specific steps of S3 are as follows: S31, for Perform node feature masking and random edge discarding operations to obtain the comprehensive lncRNA attribute feature matrix. and ;right By performing node feature masking and random edge discarding operations, a comprehensive cancer attribute feature matrix is ​​obtained. and ; S32, based on and The node encoding vector is obtained using a GNN encoder. ;based on and The node encoding vector is obtained using a GNN encoder. ;based on and The node encoding vector is obtained using a GNN encoder. ;based on and The node encoding vector is obtained using a GNN encoder. .

5. The lncRNA-cancer association prediction method based on network motif consistency and oncogenic synergistic regulation predictor according to claim 1, characterized in that, The specific steps of S4 are as follows: S41, Construct a globally learnable prototype matrix ,in It is a dimension of The matrix is ​​K, where K is the preset total number of network motif centers. It is the dimension of each vector; S42, for , and Using NMCS, a positive and negative sample pool for lncRNA was constructed. ; S43, to , and Using NMCS, construct positive and negative sample pools for cancer .

6. The lncRNA-cancer association prediction method based on network motif consistency and oncogenic synergistic regulation predictor according to claim 1, characterized in that, The specific steps of S5 are as follows: S51, based on and For any lncRNA Calculate the dual-target loss ; S52, based on and For any cancer Calculate the dual-target loss ; S53, based on and The learnable parameters are updated through the total loss function. ; S54, using a GNN encoder, for and Perform multi-layer graph convolution operations to obtain and .

7. The lncRNA-cancer association prediction method based on network motif consistency and oncogenic synergistic regulation predictor according to claim 1, characterized in that, The specific steps of S6 are as follows: S61, for and The splicing operation was performed to obtain the lncRNA-cancer co-feature matrix. ; S62, for Perform a functional cluster weighted convergence operation to construct a global oncogenic regulatory landscape vector. ; S63, based on and Constructing deep carcinogenic fingerprints ; S64, based on and By jointly optimizing the total loss function Update the learnable parameter set ; S65, using Calculate lncRNA With cancer lncRNA-cancer association prediction score .

Images