IncRNA and disease association prediction method based on deep learning

Abstract

The invention belongs to the technical field of bioinformatics, and particularly relates to an lncRNA and disease association prediction method based on deep learning, which comprises the following specific steps: firstly, according to known lncRNA-disease association information, disease-miRNA association information and lncRNA-miRNA association information, constructing an lncRNA-disease association matrix, a disease-miRNA association matrix and an lncRNA-miRNA association matrix; and constructing an lncRNA comprehensive similarity matrix LS, a disease comprehensive similarity matrix DS and a miRNA comprehensive similarity matrix MS. The biological sequence selective compression network adopted by the invention can effectively model a long-distance dependency relationship in an ultra-long sequence and dynamically distinguish key functional fragments from redundant fragments according to context.

Description

Technical Field

[0001] This invention relates to the field of bioinformatics technology, specifically to a method for predicting the association between lncRNA and disease based on deep learning. Background Technology

[0002] With the rapid development of high-throughput sequencing technology, the amount of biomedical data has exploded, leading to the identification of a large number of long non-coding RNAs (lncRNAs). Increasing evidence suggests that abnormal function of lncRNAs is closely related to the development and progression of various complex human diseases, including cancer and cardiovascular disease. However, identifying the association between lncRNAs and diseases through traditional biological experiments is costly and time-consuming, leaving the function of the vast majority of lncRNAs unknown. Therefore, developing efficient and accurate computational methods to predict potential lncRNA-disease associations has become a key challenge in the field of bioinformatics.

[0003] Existing lncRNA-disease association prediction models have the following main shortcomings: 1. Superficial lncRNA sequence feature extraction: Existing methods mostly rely on local fragment analysis such as K-mer, which can only capture short-range base motifs and cannot identify long-range cooperating motifs that determine function. At the same time, they indiscriminately process all sequence fragments, failing to selectively focus on key functional regions, resulting in low and redundant information density in the final feature representation.

[0004] 2. Rigid integration of media pathways and poor information fusion: Models often rely on single, fixed association pathways, ignoring other potential indirect pathways. When fusing multi-source information, they often use simple linear superposition or splicing, lacking an adaptive weighting mechanism to dynamically adjust the contribution of different biological pathway information according to the context.

[0005] 3. Conflicts in the fusion of homogeneous and heterogeneous features: Existing models, when fusing node features, often simply concatenate or statically weight homogeneous feature information (similarity between nodes of the same type) with heterogeneous feature information (association features obtained from lncRNAs for diseases, and association features obtained from lncRNAs for diseases). This approach cannot dynamically balance the contribution of these two information sources. When the two types of feature information conflict, the model struggles to fuse the feature information effectively, resulting in poor final feature fusion performance.

[0006] Based on the above, a deep learning-based method for predicting the association between lncRNA and disease is invented. Summary of the Invention

[0007] To solve the above-mentioned technical problems, the present invention provides the following technical solution: A deep learning-based method for predicting the association between lncRNAs and diseases includes the following specific steps: Step 1: Construct an lncRNA-disease association matrix based on known lncRNA-disease association information, disease-miRNA association information, and lncRNA-miRNA association information. Disease-miRNA Association Matrix lncRNA-miRNA association matrix Then, construct the lncRNA comprehensive similarity matrix LS, the disease comprehensive similarity matrix DS, and the miRNA comprehensive similarity matrix MS. Step 2: Obtain the lncRNA base sequence and use a biological sequence selective compression network to obtain the lncRNA latent function semantic feature matrix LFSF; Step 3: Based on , , LS, DS, MS, and the media-assisted signature generation method was used to generate lncRNA media-assisted signature LMHF and disease-mediated signature DMHF; Step 4: Based on LS, DS, LFSF, LMHF, and DMHF, obtain the multidimensional fusion feature of disease DF and the multidimensional fusion feature of lncRNA LF; Step 5: Based on DF, LF, LS, and DS, apply the similarity decoupling attention mechanism SDAM to enhance the feature interaction between the disease and lncRNA, and obtain the final disease characterization. and final lncRNA characterization ; Step 6: Based on , ,pass Obtaining the lncRNA-disease association prediction probability matrix using the Sigmoid activation function .

[0008] As a preferred embodiment of the deep learning-based lncRNA disease association prediction method described in this invention, the specific steps of Step 1 are as follows: Step 1.1: Construct a lncRNA-disease association matrix based on known lncRNA-disease association information. When lncRNA With disease When there is a relationship, let ,otherwise Construct a disease-miRNA association matrix based on known disease-miRNA association information. When disease With miRNA When there is a relationship, let ,otherwise Construct a lncRNA-miRNA association matrix based on known lncRNA-miRNA association information. When lncRNA With miRNA When there is a relationship, let ,otherwise Where ln, mn, and dn represent the number of lncRNAs, miRNAs, and diseases, respectively. For dimension Matrix; For dimension Matrix; For dimension Matrix; Step 1.2: Constructing the lncRNA comprehensive similarity matrix ,in For dimension Matrix; Step 1.3: Construct a comprehensive disease similarity matrix ,in For dimension Matrix; Step 1.4: Constructing the miRNA comprehensive similarity matrix ,in For dimension The matrix.

[0009] As a preferred embodiment of the deep learning-based lncRNA disease association prediction method described in this invention, the specific steps of Step 2 are as follows: Step 2.1: For each lncRNA base sequence in LSet, use a sliding window of size K and step size 1 to segment it, and obtain a K-mer list composed of K-mer strings; collect all the above K-mer lists to obtain a K-mer list set KS; Step 2.2: Based on KS, obtain the integer index tensor matrix IM; Step 2.3: Input the IM into the biological sequence selective compression network to obtain the lncRNA latent function semantic feature matrix LFSF.

[0010] As a preferred embodiment of the deep learning-based lncRNA disease association prediction method described in this invention, the specific steps of Step 2.2 are as follows: Step 2.2.1: Based on the lncRNA base alphabet {A, C, G, U}, construct a base containing all V= A K-mer dictionary of possible combinations, which maps each K-mer string to a unique integer index; Step 2.2.2: Using the K-mer dictionary, extract each K-mer string from KS. Convert to integer indices to obtain an integer index tensor matrix. .

[0011] As a preferred embodiment of the deep learning-based lncRNA disease association prediction method described in this invention, the specific steps of Step 2.3 are as follows: Step 2.3.1: Input the IM into a learnable biological motif embedding layer to construct the... ; Step 2.3.2: Use the biological sequence feature refinement module to refine Xbatch and obtain the enhanced feature matrix. ; Step 2.3.3: [The text appears to be incomplete and contains several grammatical errors. A more accurate translation Along its sequence length dimension Perform global average pooling to obtain the semantic feature matrix of lncRNA latent function. , For dimension The matrix.

[0012] As a preferred embodiment of the deep learning-based lncRNA disease association prediction method described in this invention, the specific steps of Step 2.3.2 are as follows: Step 2.3.2.1: For A bio-information selective gating system is generated through matrix multiplication. and a biological information vector ; Step 2.3.2.2: Update the long-range biological information compression carrier ; Step 2.3.2.3: Using a learnable output matrix right Projection is performed to obtain the global information fusion vector. ; Step 2.3.2.4: Put all Integrate and build .

[0013] As a preferred embodiment of the deep learning-based lncRNA disease association prediction method described in this invention, the specific steps of Step 3 are as follows: Step 3.1: Generate auxiliary features for disease vectors ,in For dimension Matrix; Step 3.2: Generate lncRNA-mediated auxiliary features ,in For dimension The matrix.

[0014] As a preferred embodiment of the deep learning-based lncRNA disease association prediction method described in this invention, the specific steps of Step 3.1 are as follows: Step 3.1.1: Generate first-order mediator features of the disease ; Step 3.1.2: Generate disease second-order mediator features ; Step 3.1.3: [The text appears to be incomplete and contains several grammatical errors. A more accurate translation would require the full context.] and Perform a weighted fusion operation to obtain ; The specific steps of Step 3.2 are as follows: Step 3.2.1: Generate lncRNA first-order mediator features ; Step 3.2.2: Generate lncRNA second-order mediator features ; Step 3.2.3: [The text appears to be incomplete and contains several grammatical errors. A more accurate translation would require the full context.] and Perform a weighted fusion operation to obtain .

[0015] As a preferred embodiment of the deep learning-based lncRNA disease association prediction method described in this invention, the specific steps of Step 4 are as follows: Step 4.1: Concatenate LS, LMHF, and LFSF along the feature dimension to obtain the multi-dimensional fusion feature of lncRNA. ; Step 4.2: Concatenate the DS and DMHF along the feature dimensions to obtain the multi-dimensional fusion features of the disease. .

[0016] As a preferred embodiment of the deep learning-based lncRNA disease association prediction method described in this invention, the specific steps of Step 5 are as follows: Step 5.1: Transform the DF into a disease query matrix through matrix multiplication. Disease Key Matrix Disease value matrix Transform LF into an lncRNA query matrix lncRNA bond matrix and lncRNA value matrix Through this matrix multiplication operation, , , , , , All feature dimensions are unified as pd; Step 5.2: Apply the content-related attention mechanism to obtain the output characteristics of the disease content pathway. lncRNA content pathway output characteristics ; Step 5.3: Apply the structural association attention mechanism to obtain disease structural enhancement features. Enhanced features of lncRNA structure ; Step 5.4: Utilize the content structure fusion mechanism to obtain disease-related collaborative attention characteristics. lncRNA co-attention features ; Step 5.5: Obtain the final disease phenotype through a gated residual refinement mechanism. With final lncRNA characterization ; The specific steps of Step 5.3 are as follows: Step 5.3.1: Construct a normalized disease similarity matrix Normalized lncRNA similarity matrix ; Step 5.3.2: Constructing a disease affinity matrix lncRNA affinity matrix ; Step5.3.3: Utilize polymerization Affinity information in the middle, using polymerization Affinity information in, and through , Project onto a specified dimension to generate and ; The specific steps of Step 5.4 are as follows: Step 5.4.1: Calculate the disease content structure fusion gate and lncRNA content structure fusion gate ; Step 5.4.2: Calculation and Obtaining and ; The specific steps of Step 5.5 are as follows: Step 5.5.1: Perform matrix multiplication to... Feature Dimension Transformation ,get ;Will Feature Dimension Transformation ,get ; Step 5.5.2: Obtain intermediate enhancement features of the disease through residual connections and layer normalization. Intermediate enhancement features of lncRNA ; Step 5.5.3: Obtain Disease Feature Update Gating lncRNA feature update gating ; Step 5.5.4: [The text appears to be incomplete and contains several grammatical errors. A more accurate translation would require the full context.] and , By performing weighted fusion, we obtain ;Will and , By performing weighted fusion, we obtain .

[0017] Compared with existing technologies: 1. Enhanced the characterization ability and information density of lncRNA sequence features: The biological sequence selective compression network used in this invention can effectively model long-distance dependencies in ultra-long sequences and dynamically distinguish key functional fragments from redundant fragments based on context. Compared with existing technologies such as K-mer, which can only capture local motifs, the sequence features extracted by this invention are more global, selective, and biologically significant, providing higher quality input for the model and thus improving the accuracy of prediction. 2. Enhanced fusion efficiency and intelligence of multi-source heterogeneous information: This invention solves the problem of rigid information fusion in existing technologies by using a media-assisted feature generation method; the model no longer simply superimposes all features linearly, but can dynamically and non-linearly weight and fuse them according to the importance of different information channels such as content, structure, and media channels; this intelligent fusion method enables the final feature representation to be dynamically adjusted, maximizing the synergistic advantages of various heterogeneous data. 3. Intelligent decoupling and dynamic balancing of homogeneous feature information (similarity between nodes of the same type) and heterogeneous feature information (association features obtained from lncRNAs and association features obtained from diseases by lncRNAs): The Similarity Decoupling Attention Mechanism (SDAM) designed in this invention innovatively sets up a content association attention mechanism and a structure association attention mechanism to handle homogeneous and heterogeneous feature information respectively. Compared with the existing technology that forcibly splices or statically fuses the two types of information, resulting in information conflict and noise amplification, the gated fusion mechanism and residual refinement mechanism in SDAM can dynamically weigh the contribution of different feature information. This not only protects the model from noise from a single information source, but also makes the final feature representation more accurate and robust because the model can adaptively select the optimal information ratio. Attached Figure Description

[0018] Figure 1 This is a flowchart of the technical solution of the present invention; Figure 2 This is a diagram of the biological sequence selective compression network structure of the present invention; Figure 3 This is a structural diagram of the similarity decoupling attention mechanism SDAM 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 deep learning-based method for predicting the association between lncRNAs and diseases. Please refer to [link to relevant documentation]. Figures 1-3 The specific steps are as follows: Step 1: Construct an lncRNA-disease association matrix based on known lncRNA-disease association information, disease-miRNA association information, and lncRNA-miRNA association information. Disease-miRNA Association Matrix lncRNA-miRNA association matrix Then, we constructed the lncRNA comprehensive similarity matrix LS, the disease comprehensive similarity matrix DS, and the miRNA comprehensive similarity matrix MS.

[0021] Example: In this approach, the lncRNA-disease association information was obtained from the LncRNADisease v2.0 dataset. This dataset is a widely recognized public database specifically designed to record the relationship between human lncRNAs and diseases. Its purpose is to provide experimentally validated or authoritative literature-reported data to support research on the function and mechanisms of lncRNAs in various diseases. The download link for this dataset is http: / / www.cuilab.cn / lncrnadisease.

[0022] The disease-miRNA association information was obtained from the HMDD v3.2 (Human microRNA Disease Database) dataset. This dataset is a comprehensive database focused on collecting and organizing associations between human miRNAs and diseases. Its purpose is to provide high-quality, experimentally supported association data for miRNA-related research, making it one of the core resources in the field. The dataset can be downloaded from http: / / www.cuilab.cn / hmdd. It contains a large number of disease-miRNA association entries.

[0023] The lncRNA-miRNA association information was obtained from the starBase v2.0 dataset. The starBase v2.0 dataset is designed to provide highly reliable evidence of intermolecular interactions directly validated by high-throughput experiments. The dataset can be downloaded from https: / / starbase.sysu.edu.cn / . It contains a list of experimentally validated lncRNA-miRNA interaction pairs.

[0024] The specific steps of Step 1 are as follows: Step 1.1: Construct a lncRNA-disease association matrix based on known lncRNA-disease association information. When lncRNA With disease When there is a relationship, let ,otherwise Construct a disease-miRNA association matrix based on known disease-miRNA association information. When disease With miRNA When there is a relationship, let ,otherwise Construct a lncRNA-miRNA association matrix based on known lncRNA-miRNA association information. When lncRNA With miRNA When there is a relationship, let ,otherwise Where ln, mn, and dn represent the number of lncRNAs, miRNAs, and diseases, respectively. For dimension Matrix; For dimension Matrix; For dimension Matrix; Step 1.2: Constructing the lncRNA comprehensive similarity matrix ,in For dimension Matrix; The specific steps of Step 1.2 are as follows: Step 1.2.1: For any two lncRNAs and ,calculate and Gaussian kernel similarity between ; The core idea of ​​Gaussian kernel similarity is that the similarity between two samples is inversely proportional to their Euclidean distance in the feature space; the closer the distance, the higher the similarity. This similarity is calculated using an exponentially decaying Gaussian function that takes the Euclidean distance as input.

[0025] Step 1.2.2: For any two lncRNAs and ,calculate and Cosine similarity between .

[0026] Cosine similarity is a metric for measuring the similarity between two vectors. It measures their similarity by calculating the cosine of the angle between the two vectors, focusing primarily on the direction of the vectors rather than their magnitude. This is a publicly available method and is not part of this patent invention.

[0027] Step 1.2.3: Based on any two lncRNAs and Gaussian kernel similarity and cosine similarity between them, construct .

[0028] The specific steps of Step 1.2.3 are as follows: Step 1.2.3.1: Calculation and Comprehensive similarity between :

[0029] Step 1.2.3.2: Integrate the pairwise similarities of all lncRNAs to construct the... :

[0030] Among them, comprehensive similarity is a concept used to measure the degree of similarity between two or more objects. It relies on similarity information from different sources to more comprehensively and accurately measure the degree of similarity between two or more objects. It allows information from different dimensions to complement each other, overcomes the bias that may exist in a single data source, and thus more stably and reliably assesses the similarity relationship between objects.

[0031] Step 1.3: Construct a comprehensive disease similarity matrix ,in For dimension Matrix; The specific steps of Step 1.3 are as follows: Step 1.3.1: For any two diseases and ,calculate and Gaussian kernel similarity between ; Step 1.3.2: For any two diseases and ,calculate and Cosine similarity between ; Step 1.3.3: Based on any two diseases and Using Gaussian kernel similarity and cosine similarity between the two pairs of samples, a comprehensive disease similarity matrix is ​​constructed. ; The specific steps of Step 1.3.3 are as follows: Step 1.3.3.1: Calculation and Comprehensive similarity between :

[0032] Step 1.3.3.2: Integrate the comprehensive similarity between all pairs of diseases to construct... :

[0033] Step 1.4: Constructing the miRNA comprehensive similarity matrix ,in For dimension The matrix.

[0034] The specific steps of Step 1.4 are as follows: Step 1.4.1: For any two miRNAs and ,calculate and Mixed Gaussian kernel similarity between :

[0035] in, , Represents the normalized bandwidth parameter; Indicates Euclidean distance; and These are two binary feature vectors constructed for miRNA m. The lncRNA binary association vector representing miRNA m, This represents the binary association vector of miRNA m with the disease. The binary association vector is a row vector consisting of 0s and 1s, where each element corresponds to a specific associated object, either a lncRNA or a disease. If miRNA m has a known association with the k-th associated object, the k-th element of the binary association vector is set to 1; otherwise, it is set to 0.

[0036] Among them, binary association vector is a data representation method for recording whether there is an association relationship between entities.

[0037] Step 1.4.2: For any two miRNAs and ,calculate and Mixed cosine similarity between :

[0038] in, This means that two vectors are multiplied element by element and then summed.

[0039] Step 1.4.3: Based on any two miRNAs and The mixture of Gaussian kernel similarity and mixture of cosine similarity between them is used to construct... : First calculate and Comprehensive similarity between The calculation formula is as follows:

[0040] Then, the pairwise similarity of all miRNAs was integrated to construct... :

[0041] Example: lncRNA-disease association information, disease-miRNA association information, and lncRNA-miRNA association information were obtained from authoritative databases. Through screening, 894 lncRNAs, 564 miRNAs, and 715 diseases were identified. An lncRNA-disease association matrix was constructed based on the association status of the elements. Disease-miRNA association matrix with dimensions of 894×715 715×564 lncRNA-miRNA association matrix The dimensions are 894×564.

[0042] Gaussian kernel similarity and cosine similarity among lncRNAs were calculated. The Gaussian kernel similarity and cosine similarity were first added together and then averaged to obtain the lncRNA comprehensive similarity. The comprehensive similarity of all lncRNAs was then integrated to obtain an 894×894 lncRNA comprehensive similarity matrix. Calculate the Gaussian kernel similarity and cosine similarity between diseases, add the Gaussian kernel similarity and cosine similarity, and then average them to obtain the comprehensive disease similarity. Integrate all the comprehensive disease similarities to obtain a comprehensive disease similarity matrix with a dimension of 715×715. The mixture of Gaussian kernel similarity and mixture of cosine similarity among miRNAs is calculated. The mixture of Gaussian kernel similarity and mixture of cosine similarity are then summed and averaged to obtain the comprehensive miRNA similarity. Finally, the comprehensive similarity of all miRNAs is integrated to obtain a 564×564 miRNA comprehensive similarity matrix. .

[0043] Step 2: Obtain the lncRNA base sequence and use a biological sequence selective compression network to obtain the lncRNA latent functional semantic feature matrix LFSF (Latent Functional Semantic Feature).

[0044] The specific steps of Step 2 are as follows: Step 2.1: For each lncRNA base sequence in LSet, use a sliding window of size K and step size 1 to segment it, and obtain a K-mer list composed of K-mer strings; collect all the above K-mer lists to obtain a K-mer list set KS; LSet is a list containing ln lncRNA base sequences, each of which is a string of different lengths. LSet is obtained from authoritative biological databases.

[0045] The disease-regulating function of lncRNAs often relies on specific short sequence motifs, such as protein- and DNA-binding motifs and methylation modification sites. Overlapping cleavage with a step size of 1 can completely preserve the continuity of these motifs. For example, in the cancer-associated lncRNA HOTAIR, the key motif regulating chromatin remodeling may consist of only 5-8 bases. Sliding windows can prevent the motif from being split, ensuring that subsequent models can recognize these core features directly related to the disease.

[0046] Wherein, K-mer refers to a continuous subsequence of length K that is decomposed from a lncRNA base sequence. The specific method for obtaining K-mer is to use a sliding window of length K, starting from the beginning of the lncRNA base sequence and moving one position backward one by one, and extracting all continuous and overlapping subsequences of length K in sequence.

[0047] Example: The lncRNA base sequences in this protocol were obtained from the NCBI GenBank dataset. This dataset is an authoritative gene sequence database established and maintained by the National Center for Biotechnology Information (NCBI) in the United States, used to store all publicly available DNA and RNA sequences worldwide. The access link for this dataset is https: / / www.ncbi.nlm.nih.gov / genbank / . It contains all the lncRNA base sequences involved in this protocol.

[0048] Step 2.2: Based on KS, obtain the integer index tensor matrix IM; The specific steps of Step 2.2 are as follows: Step 2.2.1: Based on the lncRNA base alphabet {A, C, G, U}, construct a base containing all V= A K-mer dictionary of possible combinations, which maps each K-mer string to a unique integer index; Step 2.2.2: Using the K-mer dictionary, extract each K-mer string from KS. Convert to integer indices to obtain an integer index tensor matrix. .

[0049] The construction form of IM is as follows:

[0050] in, The length of the largest K-mer list in KS; For dimension Matrix; Calculated using the following formula: It is the number of K-mer strings in the i-th K-mer list in KS; The K-mer string at the j-th position in the i-th K-mer list in KS; This refers to Integer index mapped to via the K-mer dictionary.

[0051] Step 2.3: Input the IM into the biological sequence selective compression network to obtain the lncRNA latent function semantic feature matrix LFSF.

[0052] Biological sequence selective compression network structure diagram as shown in the figure Figure 2 As shown, the main function of this network is to extract lncRNA base sequence features to facilitate subsequent prediction of lncRNA-disease associations. The feature vector matrix is ​​obtained by inputting the IM into the biological motif embedding layer. ;Will The input sequence feature refinement module generates bioinformatics selective gating and bioinformatics vectors. It then uses the bioinformatics selective gating and bioinformatics vectors to update the long-range bioinformatics compression carrier. After aggregation and projection, it outputs an enhanced feature matrix. Finally, through a global average pooling operation, the... Pooling yields LFSF.

[0053] The specific steps of Step 2.3 are as follows: Step 2.3.1: Input the IM into a learnable biological motif embedding layer to construct the... ; Functionally similar K-mers, such as different K-mers that are all involved in the regulation of inflammatory responses, will cluster in the biological motif embedding layer. This helps the model identify functionally equivalent disease-associated features. For example, different cancers may have different but functionally similar K-mers that all promote carcinogenesis by inhibiting the expression of tumor suppressor genes. The biological motif embedding layer can map these K-mers into similar vectors, improving the model's ability to generalize and predict lncRNA functions across cancer types.

[0054] First of all After processing by the biological motif embedding layer, the biological motif embedding query vector is obtained. :

[0055] in, The feature dimension is obtained after processing with the biological motif embedding layer; For dimension A one-dimensional vector; It is the learnable weight matrix of the biological motif embedding layer, which is used here as a lookup table; For dimension Matrix; For dimension The matrix.

[0056] Then, all of them Integrate and build :

[0057] The lookup table is the core mechanism by which the biological motif embedding layer achieves its functionality. Essentially, it's a learnable weight matrix where each row represents a dense feature vector for a specific class. When the model receives an integer index representing the class as input, it directly uses this index as the row number to look up the corresponding row vector in the weight matrix. This lookup method significantly improves the efficiency of processing discrete and high-dimensional categorical data.

[0058] Step 2.3.2: Use the biological sequence feature refinement module to refine Xbatch and obtain the enhanced feature matrix. ; The specific steps of Step 2.3.2 are as follows: Step 2.3.2.1: For A bio-information selective gating system is generated through matrix multiplication. and a biological information vector ;

[0059] in, , The weight matrix is ​​a learnable matrix; For dimension A one-dimensional vector; For dimension Matrix; This represents matrix multiplication.

[0060] lncRNA base sequences may contain a large number of non-functional regions. Bio-selective gating can automatically enhance key disease-related regions and weaken interference from irrelevant sequences. For example, in liver cancer-related lncRNAs, bio-selective gating can focus on the motif that binds to the p53 protein, ignoring other non-functional repetitive motifs, thus improving the accuracy of capturing key regions.

[0061] Step 2.3.2.2: Update the long-range biological information compression carrier The specific calculation process is as follows:

[0062] in, For element-wise multiplication; It is a learnable global state matrix; for The previous position is a biological long-range information compression carrier.

[0063] Step 2.3.2.3: Using a learnable output matrix right Projection is performed to obtain the global information fusion vector. ;

[0064] Step 2.3.2.4: Put all Integrate and build ;

[0065] Step 2.3.3: [The text appears to be incomplete and contains several grammatical errors. A more accurate translation Along its sequence length dimension Perform global average pooling to obtain the semantic feature matrix of lncRNA latent function. , For dimension The matrix; the specific construction process of LFSF is as follows:

[0066] The i-th line of LFSF is:

[0067] Global average pooling can filter redundant information and retain the global features most relevant to the disease, allowing subsequent prediction processes to focus more on core functional signals. For example, in predicting the association between lncRNAs and diabetes, pooled features can condense their overall regulatory effect on the insulin signaling pathway, reducing the interference of non-specific sequence fluctuations on the prediction.

[0068] Example: Step 2.1: lncRNA sequence acquisition and K-mer segmentation: K-mer set generation: K=3 was selected, and a sliding window with a step size of 1 was used for segmentation. Each lncRNA base sequence was segmented. For example, segmenting the 2273-base sequence of lncRNA HOTAIR yielded 2273-3+1=2271 K-mers. These 2271 K-mers were collected into a K-mer list. All K-mer lists obtained from lncRNA segmentation were collected to form a K-mer list set KS, where the maximum K-mer list length is 3754. KS is shown below:

[0069] Step 2.2: Constructing the Integer Index Tensor Matrix IM: Step 2.2.1: K-mer dictionary generation: Based on the lncRNA base alphabet {A,C,G,U}, construct a complete dictionary containing 64 3-mers in lexicographical order to achieve a unique mapping between K-mers and integer indices, for example, “AAA”→1, “AAC”→2, “AAG”→3, ..., “UUU”→64.

[0070] Step 2.2.2: Index Matrix Filling: Using the above dictionary, convert each K-mer in KS into an integer index, and construct an integer index tensor matrix IM with a dimension of 894×3754. For example, for lncRNA MEG3, its K-mer list length is 1593. Then, for the row corresponding to MEG3 in IM, columns 1-1593 are filled with the integer index of the corresponding K-mer dictionary, and columns 1594-3754 are filled with 0.

[0071]

[0072] Step 2.3: Execute the biological sequence selective compression network: Step 2.3.1: Biological motif embedding layer mapping: Set the embedding dimension to 256 and construct a learnable weight matrix. This is a learnable weight matrix for the biological motif embedding layer, which is used as a lookup table. The IM is input into the biological motif embedding layer, and each integer index is mapped to a 256-dimensional vector to generate the feature vector matrix. At this point, K-mers with similar functions, such as different K-mers that are all involved in the regulation of tumor suppressor genes, will form clusters in the embedding space.

[0073] Step 2.3.2: Execute the biological sequence feature refinement module: Step 2.3.2.1: Through learnable matrices , Perform matrix multiplication on each vector in Xbatch to obtain bio-information selective gating. and biological information vectors .

[0074] Step 2.3.2.2: Update the long-range biological information compression vector to achieve long-distance K-mer interaction capture, and obtain... .

[0075] Steps 2.3.2.3-2.3.2.4: The enhanced feature matrix is ​​obtained by projecting and integrating the learnable output matrix. .

[0076] Step 2.3.3: Global Average Pooling: Pooling along the length of the maximum K-mer list. A global average pooling operation is performed to calculate the 256-dimensional feature vector for each lncRNA, ultimately yielding the latent functional semantic feature matrix. .

[0077] Step 3: Based on , , LS, DS, MS, and media-assisted feature generation methods were used to generate lncRNA media-assisted features LMHF (LncRNA Media-Harmonized Features) and disease media-assisted features DMHF (Disease Media-Harmonized Features).

[0078] Mediator-assisted signature generation aims to generate signature representations of diseases or lncRNAs by using miRNAs as the medium for information transmission.

[0079] The specific steps of Step 3 are as follows: Step 3.1: Generate auxiliary features for disease vectors ,in For dimension Matrix; The specific steps of Step 3.1 are as follows: Step 3.1.1: Generate first-order mediator features of the disease ;

[0080] The purpose of this step is to incorporate the similarity information of miRNAs directly related to the disease into disease characteristics.

[0081] Step 3.1.2: Generate disease second-order mediator features ;

[0082] in, For dimension Matrix; The purpose of this step is to supplement disease characteristics by introducing similarity information of lncRNAs that are indirectly related to the disease through the association between the disease, miRNA, and lncRNA.

[0083] Step 3.1.3: [The text appears to be incomplete and contains several grammatical errors. A more accurate translation would require the full context.] and Perform a weighted fusion operation to obtain ;

[0084] in and All are multilayer perceptron modules. The input layer takes a dimension of The matrix, the output layer outputs a matrix of dimension 1. Matrix; The input layer takes a dimension of The matrix, the output layer outputs a matrix of dimension 1. Matrix; , The weights are learnable parameters; dmfn is... The feature dimensions.

[0085] Among them, the Multilayer Perceptron (MLP) is a classic feedforward artificial neural network, consisting of an input layer, one or more hidden layers, and an output layer. Data enters from the input layer, undergoes complex feature extraction and nonlinear transformation through the hidden layers, and finally outputs the results by the output layer.

[0086] Step 3.2: Generate lncRNA-mediated auxiliary features ,in For dimension The matrix.

[0087] The specific steps of Step 3.2 are as follows: Step 3.2.1: Generate lncRNA first-order mediator features ;

[0088] The purpose of this step is to incorporate the similarity information of miRNAs directly related to lncRNAs into the lncRNA signature.

[0089] Step 3.2.2: Generate lncRNA second-order mediator features ;

[0090] The purpose of this step is to supplement the lncRNA features by introducing disease similarity information indirectly related to lncRNA through the association between lncRNA, miRNA, and disease, using miRNA.

[0091] Step 3.2.3: [The text appears to be incomplete and contains several grammatical errors. A more accurate translation would require the full context.] and Perform a weighted fusion operation to obtain ;

[0092] in , It is a multilayer perceptron module. The input layer takes a dimension of The matrix, the output layer outputs a matrix of dimension 1. Matrix; The input layer takes a dimension of The matrix, the output layer outputs a matrix of dimension 1. Matrix; , The weights are learnable parameters; lmfn is... The feature dimensions.

[0093] Example: calculate Multiplying the matrix by MS yields a matrix of dimension 715×564. ;calculate , , The matrix product yields a matrix of dimension 715×894. ;Will enter Map the feature dimension to 256, enter The feature dimension is mapped to 256, and the two are weighted and fused to obtain a matrix DMHF with a dimension of 715×256.

[0094] calculate Multiplying the matrix by MS yields a matrix of dimension 894×564. ;calculate , , The matrix product yields a matrix of dimension 894×715. ;Will enter Map the feature dimension to 256, enter The feature dimension is mapped to 256, and the two are weighted and fused to obtain a matrix LMHF with a dimension of 894×256.

[0095] Step 4: Based on LS, DS, LFSF, LMHF, and DMHF, obtain the multidimensional fusion feature DF of the disease and the multidimensional fusion feature LF of lncRNA.

[0096] The specific steps of Step 4 are as follows: Step 4.1: Concatenate LS, LMHF, and LFSF along the feature dimension to obtain the multi-dimensional fusion feature of lncRNA. ;

[0097] Here, Concat is the matrix concatenation operation. For dimension The matrix, for Feature dimensions; Step 4.2: Concatenate the DS and DMHF along the feature dimensions to obtain the multi-dimensional fusion features of the disease. ;

[0098] in, For dimension The matrix, for The feature dimensions.

[0099] Step 5: Based on DF, LF, LS, and DS, apply the Similarity Decoupling Attention Mechanism (SDAM) to enhance the feature interaction between the disease and lncRNA, obtaining the final disease characterization. and final lncRNA characterization ; The specific steps of Step 5 are as follows: Step 5.1: Transform the DF into a disease query matrix through matrix multiplication. Disease Key Matrix Disease value matrix Transform LF into an lncRNA query matrix lncRNA bond matrix and lncRNA value matrix Through this matrix multiplication operation, , , , , , All feature dimensions are unified as pd; , , , ,

[0100] in , , , , , The weight matrix is ​​a learnable matrix; For dimension Matrix; For dimension The matrix.

[0101] DF and LF come from different data sources and have significant dimensional differences. By unifying the dimension to pd, the feature association between lncRNAs and diseases can be directly calculated through QK matching, avoiding the loss of association information due to dimensional incompatibility. For example, the functional matching degree between insulin resistance signals in disease features and PI3K pathway regulatory motifs in lncRNA features can be accurately calculated after dimension unification, providing a quantitative basis for predicting the association between lncRNAs and type 2 diabetes.

[0102] Step 5.2: Apply the content-related attention mechanism to obtain the output characteristics of the disease content pathway. lncRNA content pathway output characteristics ;

[0103] in, For dimension Matrix; For dimension The matrix.

[0104] It is a normalized activation function that can flexibly adjust the sparsity of the output distribution, possessing both... It has the ability to assign weights to all items, and also possesses The ability to set the weight of minor items to zero:

[0105] in, These are learnable weight parameters.

[0106] Step 5.3: Apply the structural association attention mechanism to obtain disease structural enhancement features. Enhanced features of lncRNA structure ; The specific steps of Step 5.3 are as follows: Step 5.3.1: Construct a normalized disease similarity matrix Normalized lncRNA similarity matrix :

[0107] Here, `row_softmax` refers to performing a softmax operation row by row. This operation transforms each row of the matrix into a probability distribution, ensuring that the sum of all elements in the row is 1.

[0108] Step 5.3.2: Constructing a disease affinity matrix lncRNA affinity matrix ;

[0109] Step5.3.3: Utilize polymerization Affinity information in the middle, using polymerization Affinity information in, and through , Project onto a specified dimension to generate and ;

[0110] in and This is a learnable weight matrix.

[0111] Affinity matrices quantify the overall similarity between disease and lncRNA features. By combining these with similarity matrices, common regulatory pathways between different diseases and lncRNAs can be extracted. For example, many autoimmune diseases, such as rheumatoid arthritis and systemic lupus erythematosus, exhibit abnormalities in immune inflammatory pathways. Structural association attention can capture the common associations between lncRNAs regulating the NF-κB inflammatory pathway and these diseases, improving the model's generalization ability to predict disease associations with similar lncRNAs.

[0112] Step 5.4: Utilize the content structure fusion mechanism to obtain disease-related collaborative attention characteristics. lncRNA co-attention features ; The specific steps of Step 5.4 are as follows: Step 5.4.1: Calculate the disease content structure fusion gate and lncRNA content structure fusion gate ;

[0113] Where σ represents the Sigmoid activation function. and The weight matrix is ​​a learnable matrix; Step 5.4.2: Calculation and Obtaining and ;

[0114] Step 5.5: Obtain the final disease phenotype through a gated residual refinement mechanism. With final lncRNA characterization .

[0115] The specific steps of Step 5.5 are as follows: Step 5.5.1: Perform matrix multiplication to... Feature Dimension Transformation ,get ;Will Feature Dimension Transformation ,get ;

[0116] in, , The weight matrix is ​​a learnable matrix; For dimension Matrix; For dimension Matrix; Step 5.5.2: Obtain intermediate enhancement features of the disease through residual connections and layer normalization. Intermediate enhancement features of lncRNA :

[0117] Residual connections can preserve key biological information in the original features, avoiding the loss of core associations due to feature transformation during attention enhancement.

[0118] in: Layer normalization is an important normalization technique in deep learning. It normalizes all features of the training samples within the training samples themselves.

[0119] Step 5.5.3: Obtain Disease Feature Update Gating lncRNA feature update gating ;

[0120] in , The weight matrix is ​​a learnable matrix; For dimension Matrix; For dimension Matrix; Step 5.5.4: [The text appears to be incomplete and contains several grammatical errors. A more accurate translation would require the full context.] and , By performing weighted fusion, we obtain ;Will and , By performing weighted fusion, we obtain ;

[0121] Example: Based on DF and LF, a matrix of dimension 715×256 is obtained through matrix multiplication. , , A matrix with dimensions of 894×256 , , The disease content pathway output features were calculated using a content-related attention mechanism. lncRNA content pathway output characteristics By applying the structural association attention mechanism, we obtain... and Dynamically integrate content and structural features through a content structure fusion mechanism. and Finally, a gated residual refining mechanism is applied to generate... and .

[0122] Step 6: Based on , ,pass Obtaining the lncRNA-disease association prediction probability matrix using the Sigmoid activation function .

[0123]

[0124] in The weight matrix is ​​a learnable matrix; For dimension The matrix; P is the lncRNA-disease association prediction probability matrix, where any element Indicates disease With lncRNA The probability of association between them The higher the value, the more likely it is to indicate a disease. With lncRNA The greater the probability of association between them; The smaller the value, the more likely it is to indicate a disease. With lncRNA The smaller the probability of association between them.

[0125] 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 deep learning-based method for predicting lncRNA-disease associations, characterized in that, The specific steps include the following: Step 1: According to known lncRNA-disease association information, disease-miRNA association information, lncRNA-miRNA association information, construct lncRNA-disease association matrix , disease-miRNA association matrix , lncRNA-miRNA association matrix ; then construct lncRNA comprehensive similarity matrix LS, disease comprehensive similarity matrix DS, miRNA comprehensive similarity matrix MS; Step 2: Obtain the lncRNA base sequence, and obtain the lncRNA potential function semantic feature matrix LFSF through a biological sequence selective compression network; Step3: Based on , , , LS, DS, MS, generate lncRNA media auxiliary feature LMHF and disease media auxiliary feature DMHF by applying media auxiliary feature generation method. Step 4: Obtain disease multi-dimensional fusion features DF and lncRNA multi-dimensional fusion features LF based on LS, DS, LFSF, LMHF and DMHF; Step5: Based on DF, LF, LS, DS, apply similarity decoupling attention mechanism SDAM to enhance the feature interaction between diseases and lncRNAs to obtain final disease representation and final lncRNA representation ; Step6: Based on , , by and Sigmoid activation function to obtain the lncRNA-disease association prediction probability matrix . 2.The method of claim 1, wherein, The specific steps of Step 1 are as follows: Step 1.1: Constructing lncRNA-disease association matrix according to known lncRNA-disease association information When lncRNA has association with disease , let , otherwise ; Constructing disease-miRNA association matrix according to known disease-miRNA association information When disease has association with miRNA , let , otherwise ; Constructing lncRNA-miRNA association matrix according to known lncRNA-miRNA association information When lncRNA has association with miRNA , let , otherwise ; Where ln, mn, dn are the number of lncRNA, miRNA, disease respectively; is a matrix with dimension ; is a matrix with dimension ; is a matrix with dimension ; Step 1.2: Constructing the lncRNA comprehensive similarity matrix wherein is a matrix of dimension . Step 1.3: Constructing the disease composite similarity matrix wherein is a matrix of dimension . Step 1.4: Constructing the miRNA integrated similarity matrix wherein is a matrix of dimension . 3.The method of claim 1, wherein the method comprises: The specific steps of Step 2 are as follows: Step 2.1: For each lncRNA base sequence in LSet, a sliding window with a size of K and a step of 1 is used for segmentation to obtain a K-mer list composed of K-mer strings; all K-mer lists are collected to obtain a K-mer list set KS; Step 2.2: Obtain an integer index tensor matrix IM based on KS; Step 2.3: Input IM into the biological sequence selective compression network to obtain the lncRNA potential function semantic feature matrix LFSF.

4. The lncRNA-disease association prediction method based on deep learning according to claim 3, characterized in that, The specific steps of Step 2.2 are as follows: Step 2.2.1: Construct a K-mer dictionary containing all possible combinations of V= kind according to the lncRNA base alphabet {A, C, G, U}, which can map each K-mer string to a unique integer index; Step 2.2.2: Convert each K-mer string in KS to an integer index using the K-mer dictionary, obtaining an integer index tensor matrix .​ 5. The lncRNA-disease association prediction method based on deep learning according to claim 4, characterized in that, The specific steps of Step 2.3 are as follows: Step 2.3.1 : IM is input to a learnable biological motif embedding layer, resulting in ; Step 2.3.2: Refine Xbatch using the biological sequence feature refinement module to obtain an enhanced feature matrix ; Step2.3.3: Obtain the lncRNA potential function semantic feature matrix by performing global average pooling operation on the lncRNA potential function semantic feature matrix along its sequence length dimension along its sequence length dimension is a matrix with dimension .​​ 6. The lncRNA-disease association prediction method based on deep learning according to claim 5, characterized in that, The specific steps of Step 2.3.2 are as follows: Step 2.3.2.1 : For , generate a bio-information selection gate and a bio-information vector by matrix multiplication operation; Step 2.3.2.2: Update the biological long-range information compression carrier ; Step 2.3.2.3: Projection through a learnable output matrix to obtain a global information fusion vector ;​ Step2.3.2.4: All Integration was performed to construct .

7. The lncRNA-disease association prediction method based on deep learning according to claim 1, characterized in that, The specific steps of Step 3 are as follows: Step 3.1: Generating disease mediator auxiliary features wherein is a matrix of dimension . Step 3.2: Generating lncRNA-mediated auxiliary features wherein is a matrix of dimension . 8.The method of claim 7, wherein, The specific steps of Step 3.1 are as follows: Step 3.1.1: Generating disease first-order mediator features ; Step 3.1.2: Generating disease second-order medium features ; Step3.1.3: performing a weighted fusion operation on and to obtain ; The specific steps of Step 3.2 are as follows: Step 3.2.1 : Generating lncRNA first-order mediator features ; Step 3.2.2: Generating lncRNA second-order mediator features ; Step3.2.3: Perform a weighted fusion operation on and to obtain . 9.The method of claim 1, wherein, The specific steps of Step 4 are as follows: Step 4.1: Concatenate LS, LMHF and LFSF along the feature dimension to obtain the lncRNA multi-dimensional fusion feature ; Step 4.2: Concatenate DS and DMHF along the feature dimension to obtain disease multi-dimensional fusion features .

10. The lncRNA-disease association prediction method based on deep learning according to claim 1, characterized in that, The specific steps of Step 5 are as follows: Step 5.1: Transforming DF into disease query matrix by matrix multiplication operation , disease key matrix , and disease value matrix Step 5.2: Transforming LF into lncRNA query matrix by matrix multiplication operation , lncRNA key matrix , and lncRNA value matrix ; Through the matrix multiplication operation, , , , , , The feature dimensions of the matrix multiplication operation are unified to pd. Step 5.2: applying the content-related attention mechanism to obtain disease content pathway output features , lncRNA content pathway output features ; Step 5.3: applying the structure-related attention mechanism to obtain disease structure-enhanced features and the lncRNA structure-enhanced features ; Step 5.4: Obtain disease collaborative attention feature by using content structure fusion mechanism , lncRNA collaborative attention feature ; Step 5.5: Obtain final disease characterization by gating residual refinement mechanism with final lncRNA characterization ; The specific steps of Step 5.3 are as follows: Step 5.3.1 : Constructing the normalized disease similarity matrix , the normalized lncRNA similarity matrix ; Step 5.3.2: Constructing the disease affinity matrix , lncRNA affinity matrix ; Step5.3.3: Utilize affinity information in the aggregation affinity information in the aggregation , project to the specified dimension, generating with ; The specific steps of Step 5.4 are as follows: Step 5.4.1 : Calculate disease content structure fusion gate and lncRNA content structure fusion gate ; Step 5.4.2: Calculation With ; The specific steps of Step 5.5 are as follows: Step 5.5.1: Transform the feature dimension into , by a matrix multiplication operation, to get ; Transform the feature dimension into , by a matrix multiplication operation, to get ; Transform the feature dimension into , by a matrix multiplication operation, to get ; Step 5.5.2: Obtain disease intermediate enhanced features by residual connection and layer normalization With lncRNA intermediate enhanced features ; Step 5.5.3: Obtain disease feature update gate , IncRNA feature update gate ; Step5.5.4: Weighted fusion of and , yields ; Weighted fusion of and , yields .

Images