Mirna-disease association prediction method and apparatus, electronic device, and storage medium

Abstract

An miRNA-disease association prediction method and apparatus, an electronic device, and a storage medium. The association prediction method comprises: acquiring initial association data of miRNAs and diseases; calculating miRNA similarity and disease similarity on the basis of the initial association data; on the basis of the miRNA similarity and the disease similarity, constructing a dynamic hypergraph, wherein the dynamic hypergraph comprises first node features of the miRNAs, first node features of the diseases, hyperedges connected to the first node features of the miRNAs, and hyperedges connected to the first node features of the diseases; obtaining second node features of the miRNAs and second node features of the diseases on the basis of the dynamic hypergraph; and obtaining an association prediction result of the miRNAs and the diseases on the basis of the second node features of the miRNAs and the second node features of the diseases. According to the prediction method, a hypergraph structure can be dynamically learned and updated, thereby better predicting the association between miRNAs and diseases, and improving prediction performance.

Description

Method, device, electronic device and storage medium for predicting association between miRNA and disease

[0001] This application claims priority to Chinese Patent Application No. 202410417442X filed on April 8, 2024, the disclosure of which is hereby incorporated by reference in its entirety into the present application for all purposes. TECHNICAL FIELD

[0002] Embodiments of the present disclosure relate to the technical field of biological association prediction, and in particular, to a method, device, electronic device and storage medium for predicting association between miRNA and disease. BACKGROUND

[0003] miRNA is an endogenous, short single-stranded, non-coding RNA with a length of 20-24 nt (nucleotide), which plays a key role in many life activities of organisms, such as differentiation, proliferation and apoptosis in cells of organisms, and transportation and transcription of genetic material. Therefore, in-depth study of the association between miRNA and disease helps to understand the pathogenesis of disease from the molecular level, so as to play a positive role in promoting the diagnosis and treatment of human diseases. The current research method is mainly biological wet experiment, which consumes a lot of time and cost. Therefore, people began to use computational methods to predict the association between miRNA and disease, which can greatly improve the efficiency of exploring the potential association between miRNA and disease. SUMMARY

[0004] At least one embodiment of the present disclosure provides a method for predicting association between miRNA and disease, comprising: obtaining initial association data of the miRNA and the disease; calculating similarity of the miRNA and similarity of the disease based on the initial association data; constructing a dynamic hypergraph based on the similarity of the miRNA and the similarity of the disease, wherein the dynamic hypergraph comprises first node features of the miRNA, first node features of the disease, and hyperedges connecting the first node features of the miRNA and hyperedges connecting the first node features of the disease, respectively; obtaining second node features of the miRNA and second node features of the disease based on the dynamic hypergraph; and obtaining a prediction result of the association between the miRNA and the disease based on the second node features of the miRNA and the second node features of the disease.

[0005] For example, the miRNA and disease association prediction method provided by at least one embodiment of the present disclosure, the similarity of the miRNA and the similarity of the disease are calculated based on the initial association data, including: based on the initial association data, calculating the functional similarity MFSM of the miRNA and the Gaussian interaction spectrum kernel similarity MGSM of the miRNA, and calculating the semantic similarity DSSM of the disease and the Gaussian interaction spectrum kernel similarity DGSM of the disease; and according to the functional similarity MFSM of the miRNA and the Gaussian interaction spectrum kernel similarity MGSM of the miRNA, the comprehensive similarity MISM of the miRNA is calculated, and according to the semantic similarity DSSM of the disease and the Gaussian interaction spectrum kernel similarity DGSM of the disease, the comprehensive similarity DISM of the disease is calculated.

[0006] For example, the miRNA and disease association prediction method provided by at least one embodiment of the present disclosure, the similarity of the miRNA and the similarity of the disease are calculated, and the method further includes: based on the functional similarity MFSM of the miRNA and the Gaussian interaction spectrum kernel similarity MGSM of the miRNA, the fusion similarity MSM of the miRNA is obtained by using the attention mechanism; based on the semantic similarity DSSM of the disease and the Gaussian interaction spectrum kernel similarity DGSM of the disease, the fusion similarity DSM of the disease is obtained by using the attention mechanism.

[0007] For example, the miRNA and disease association prediction method provided by at least one embodiment of the present disclosure, the dynamic hypergraph is constructed based on the similarity of the miRNA and the similarity of the disease, including: based on the fusion similarity MSM of the miRNA and the fusion similarity DSM of the disease, the initial node features of the miRNA and the initial node features of the disease are obtained; the initial node features of the miRNA and the initial node features of the disease are updated to obtain the first node features of the miRNA and the first node features of the disease.

[0008] For example, according to the miRNA and disease association prediction method provided by at least one embodiment of the present disclosure, the initial node features of the miRNAs and the initial nodes of the diseases are obtained based on the miRNA fusion similarity MSM and the disease fusion similarity DSM, and the method comprises: adjusting the dimension of the miRNA fusion similarity MSM to a first preset dimension by using a multilayer perceptron to obtain the miRNA fusion similarity features; obtaining the initial node features of the miRNAs based on the miRNA fusion similarity features and the random initialization features of the miRNAs; adjusting the dimension of the disease fusion similarity DSM to a second preset dimension by using the multilayer perceptron to obtain the disease fusion similarity features; and obtaining the initial node features of the diseases based on the disease fusion similarity features and the random initialization features of the diseases.

[0009] For example, according to the miRNA and disease association prediction method provided by at least one embodiment of the present disclosure, the initial node features of the miRNAs and the initial node features of the diseases are updated to obtain the first node features of the miRNAs and the first node features of the diseases, and the method comprises: calculating the aggregated interaction feature vector of the miRNAs based on the comprehensive similarity MISM of the miRNAs; calculating the aggregated interaction feature vector of the diseases based on the comprehensive similarity DISM of the diseases; calculating the interaction probability matrix of the miRNAs and the diseases based on the aggregated interaction feature vector of the miRNAs and the aggregated interaction feature vector of the diseases; and obtaining the miRNA-disease bipartite graph based on the interaction probability matrix of the miRNAs and the diseases and a preset negative connection value, wherein the miRNA-disease bipartite graph comprises the initial node features of the miRNAs and the initial node features of the diseases.

[0010] For example, according to the miRNA and disease association prediction method provided by at least one embodiment of the present disclosure, the initial node features of the miRNAs and the initial node features of the diseases are updated to obtain the first node features of the miRNAs and the first node features of the diseases, and the method comprises: calculating the adjacency matrix A G and the degree matrix D G of the miRNA-disease bipartite graph; and obtaining the first node features of the miRNAs and the first node features of the diseases based on the adjacency matrix A G and the degree matrix D G .

[0011] For example, according to the miRNA and disease association prediction method provided by at least one embodiment of the present disclosure, the first node features of the miRNAs and the first node features of the diseases are obtained according to the following formula:

[0012] wherein R represents a probability matrix of interaction between the miRNA and the disease, A G represents an adjacency matrix of the miRNA-disease bipartite graph, E (0) represents an initial value of the miRNA-disease bipartite graph, E *,(l) a first node feature of the miRNA in the lth layer and a first node feature of the disease in the lth layer, represents a corresponding first node feature of the miRNA in the lth layer, represents a corresponding first node feature of the disease in the lth layer, and l is a positive integer greater than 1, D G represents a degree matrix, and D G is a diagonal matrix, D G [i, i] represents a number of non-zero entries in the ith row of the adjacency matrix A G ; E (l-1) is equal to or E (0) , represents a second node feature of the miRNA in the l-1th layer, represents a second node feature of the disease in the l-1th layer.

[0013] For example, the method for predicting the correlation between miRNA and disease according to at least one embodiment of the present disclosure is based on the similarity of miRNA and the similarity of disease to construct a dynamic hypergraph, which comprises: using Gumbel-Softmax to construct a dynamic hypergraph structure of the miRNA and a dynamic hypergraph structure of the disease, wherein the lth layer hypergraph correlation matrix of the miRNA and the lth layer hypergraph correlation matrix of the disease are obtained according to the following formulas respectively:

[0014] wherein, and are learning matrices.

[0015] For example, the method for predicting the correlation between miRNA and disease according to at least one embodiment of the present disclosure is based on the dynamic hypergraph to obtain a second node feature of the miRNA and a second node feature of the disease, which comprises: performing hypergraph convolution operation on the lth layer hypergraph correlation matrix of the miRNA to obtain the second node feature of the miRNA; and performing hypergraph convolution operation on the lth layer hypergraph correlation matrix of the disease performing a hypergraph convolution operation to obtain the second node feature of the disease.

[0016] For example, the miRNA-disease association prediction method according to at least one embodiment of the present disclosure obtains the second node feature of the miRNA and the second node feature of the disease based on the dynamic hypergraph, and includes: obtaining the second node feature of the miRNA and the second node feature of the disease in a manner of jump connection, wherein the second node feature of the miRNA at the lth layer and the second node feature of the disease at the lth layer are obtained according to the following formula:

[0017] wherein, denotes the hyperedge degree matrix of the lth layer hypergraph association matrix of the miRNA . denotes the hyperedge degree matrix of the lth layer hypergraph association matrix of the disease .

[0018] For example, the miRNA-disease association prediction method according to at least one embodiment of the present disclosure obtains the miRNA-disease association prediction result based on the second node feature of the miRNA and the second node feature of the disease, and includes: performing splicing processing on the second node features in all layers of the miRNA to obtain the third node feature of the miRNA; performing splicing processing on the second node features in all layers of the disease to obtain the third node feature of the disease; and obtaining the miRNA-disease association matrix as the miRNA-disease association prediction result based on the third node feature of the miRNA and the third node feature of the disease.

[0019] For example, the miRNA-disease association prediction method according to at least one embodiment of the present disclosure further includes: obtaining a node feature learning loss function L based on the fusion similarity MSM of the miRNA, the fusion similarity DSM of the disease, and the interaction probability matrix of the miRNA and the disease sim ; obtaining an association prediction loss function L based on the third node feature of the miRNA and the third node feature of the disease BPR ; obtaining a node constraint loss L based on the third node feature of the miRNA and the third node feature of the disease const ; and obtaining a node feature learning loss function L sim , an association prediction loss function L BPR , and a node constraint loss L constobtaining a total loss function of the miRNA and the disease.

[0020] For example, the method for predicting the association between miRNA and disease according to at least one embodiment of the present disclosure, based on the similarity of miRNA and the similarity of disease, constructs a dynamic hypergraph, including: based on the comprehensive similarity MISM of the miRNA, taking each row in the comprehensive similarity MISM of the miRNA as the initial node feature of the miRNA, to construct the prior hypergraph of the miRNA; and based on the comprehensive similarity DISM of the disease, taking each row in the comprehensive similarity DISM of the disease as the initial node feature of the disease, to construct the prior hypergraph of the disease.

[0021] For example, the method for predicting the association between miRNA and disease according to at least one embodiment of the present disclosure, based on the similarity of miRNA and the similarity of disease, constructs a dynamic hypergraph, including: updating the initial node feature of the miRNA by HyperGAT to obtain the first node feature of the miRNA, and processing the first node feature of the miRNA by Gumbel-Topk and KMeans to construct the dynamic hypergraph of miRNA; updating the initial node feature of the disease by the HyperGAT to obtain the first node feature of the disease, and processing the first node feature of the disease by the Gumbel-Topk and the KMeans to construct the dynamic hypergraph of the disease.

[0022] For example, the method for predicting the association between miRNA and disease according to at least one embodiment of the present disclosure, based on the dynamic hypergraph, obtains the second node feature of miRNA and the second node feature of disease, including: performing node convolution operation and hyperedge convolution operation on the dynamic hypergraph of the miRNA to obtain the second node feature of the miRNA; and performing node convolution operation and hyperedge convolution operation on the dynamic hypergraph of the disease to obtain the second node feature of the disease.

[0023] For example, the method for predicting the association between miRNA and disease according to at least one embodiment of the present disclosure, based on the second node feature of miRNA and the second node feature of disease, obtains the association prediction result between the miRNA and the disease, including: performing splicing processing on the second node feature of the miRNA and the second node feature of the disease to obtain the miRNA-disease pair; performing at least one fully connected layer processing and softmax processing on the miRNA-disease pair to obtain the association prediction result between the miRNA and the disease.

[0024] At least one embodiment of the present disclosure further provides an miRNA and disease association prediction device, comprising an acquisition unit, a similarity calculation unit, a construction unit, a hypergraph calculation unit and a prediction unit, the acquisition unit is configured to acquire initial association data of miRNA and disease; the similarity calculation unit is configured to calculate the similarity of the miRNA and the similarity of the disease based on the initial association data; the construction unit is configured to construct a dynamic hypergraph based on the similarity of the miRNA and the similarity of the disease, wherein the dynamic hypergraph comprises first node features of the miRNA, first node features of the disease, and hyperedges connecting the first node features of the miRNA and hyperedges connecting the first node features of the disease respectively; the hypergraph calculation unit is configured to obtain second node features of the miRNA and second node features of the disease based on the dynamic hypergraph; and the prediction unit is configured to obtain an association prediction result of the miRNA and the disease based on the second node features of the miRNA and the second node features of the disease.

[0025] At least one embodiment of the present disclosure further provides an electronic device, comprising a processor, a memory and one or more computer program modules, wherein the one or more computer program modules are stored in the memory and configured to be executed by the processor, and the one or more computer program modules comprise instructions for implementing the miRNA and disease association prediction method of any of the above embodiments.

[0026] At least one embodiment of the present disclosure further provides a storage medium, which non-transitorily stores computer readable instructions, when the computer readable instructions are executed by a computer, the miRNA and disease association prediction method according to any of the above embodiments can be performed. BRIEF DESCRIPTION OF DRAWINGS

[0027] In order to more clearly illustrate the technical solutions of the embodiments of the present disclosure, the drawings of the embodiments will be briefly introduced below. Obviously, the drawings in the following description only relate to some embodiments of the present disclosure, but not limit the present disclosure.

[0028] FIG. 1 is a schematic diagram of an miRNA and disease association prediction method according to at least one embodiment of the present disclosure.

[0029] FIG. 2 is a structural schematic diagram of an miRNA and disease association prediction model according to at least one embodiment of the present disclosure.

[0030] FIG. 3 is a schematic diagram of one example of step S200 shown in FIG. 1.

[0031] FIG. 4 is a schematic diagram of obtaining fusion similarity by using attention mechanism according to at least one embodiment of the present disclosure.

[0032] FIG. 5 is a schematic diagram of one example of step S300 shown in FIG. 1.

[0033] FIG. 6 is a schematic diagram of one example of step S320 shown in FIG. 5.

[0034] FIG. 7 is a schematic diagram of one example of step S400 shown in FIG. 1.

[0035] FIGS. 8-9 are schematic diagrams of global leave-one-out cross-validation results corresponding to the miRNA and disease association prediction method provided by at least one embodiment of the present disclosure and other methods.

[0036] FIG. 10 is a schematic diagram of the structure of another miRNA and disease association prediction model provided by at least one embodiment of the present disclosure.

[0037] FIG. 11 is a schematic diagram of another example of step S300 shown in FIG. 1.

[0038] FIG. 12A is a schematic diagram of the results of five-fold cross-validation corresponding to the prediction model shown in FIG. 10.

[0039] FIG. 12B is a schematic diagram of the results of five-fold cross-validation corresponding to the prediction method based on hypergraph (HFHLMDA).

[0040] FIG. 13 is a schematic block diagram of a miRNA and disease association prediction device provided by at least one embodiment of the present disclosure.

[0041] FIG. 14 is a schematic block diagram of another miRNA and disease association prediction device provided by at least one embodiment of the present disclosure.

[0042] FIG. 15 is a schematic block diagram of an electronic device provided by at least one embodiment of the present disclosure.

[0043] FIG. 16 is a schematic diagram of a storage medium provided by at least one embodiment of the present disclosure. DETAILED DESCRIPTION

[0044] In order to make the objects, technical solutions, and advantages of the embodiments of the present disclosure clearer, the technical solutions of the embodiments of the present disclosure will be described clearly and completely below with reference to the drawings of the embodiments of the present disclosure. Obviously, the described embodiments are some but not all of the embodiments of the present disclosure. Based on the described embodiments of the present disclosure, all other embodiments obtained by those of ordinary skill in the art without creative effort belong to the scope of the present disclosure.

[0045] Unless otherwise defined, technical terms or scientific terms used in the present disclosure shall have the meanings as commonly understood by one of ordinary skill in the art to which this disclosure belongs. The terms "first", "second", and similar terms do not denote any order, quantity, or importance, but are used to distinguish one element from another, and are used arbitrarily and merely for the purpose of description. Also, the terms "one", "a", or "the" and similar terms do not denote a quantity of particular noun, but denote the existence of at least one of the particular noun. The terms "include", "comprise", and similar terms do not exclude other elements or integers not listed in the terms. The terms "connected", "coupled", and similar terms do not exclude intermediate elements unless otherwise specified. The terms "above", "below", "left", "right", and similar terms do not denote a specific orientation, but denote relative positions to each other when the described object is placed in a certain orientation.

[0046] A hypergraph is a generalization of a graph, in which each hyperedge connects two or more nodes. A hypergraph can model high-order relationships between objects, and thus can express more complex and high-order relationships, thereby having higher flexibility.

[0047] The weighted k-neighborhood profile (WKNNP) is a method for processing a sparse correlation matrix, which can be used in the field of miRNA (microRNA) and disease association prediction. Due to the complexity of biological data and the limitation of experimental conditions, the correlation matrix between miRNA and disease is often very sparse, that is, most elements are unknown or missing. In order to compensate for the sparsity of the correlation matrix (for example, the correlation matrix of miRNA and disease), the WKNNP algorithm can find the first k miRNAs most similar to the miRNA according to the similarity of the miRNA. The corresponding rows of these miRNAs in the correlation matrix contain important information about the association between miRNA and disease. The WKNNP algorithm can use the information of their corresponding rows in the correlation matrix to integrate the first association probability of the current miRNA and disease. Similarly, the disease is also processed in a similar way, and the second association probability of the current disease and miRNA is obtained. Finally, the first association probability and the second association probability are averaged to obtain the final prediction result, while compensating for the sparsity of the correlation matrix.

[0048] In the research on the method of miRNA and disease association prediction based on graph learning, the inventors of the present application found that:

[0049] For example, network similarity-based methods, as introduced in reference 1 (You ZH, Huang ZA, Zhu Z, Yan GY, Li ZW, Wen Z, Chen X. PBMDA: A novel and effective path-based computational model for miRNA-disease association prediction. PLoS Computational Biology, 2017, 13(3): 4-8), integrate the functional similarity of miRNAs, the semantic similarity of diseases, and the Gaussian interaction profile kernel similarity (GIP kernel similarity), and obtain the feature representation of miRNAs and diseases. Then, the potential associations between miRNAs and diseases are inferred by a depth-first search algorithm on the heterogeneous graph. However, this similarity construction method simply combines multiple heterogeneous similarities, and does not fully utilize multi-source heterogeneous data.

[0050] For example, machine learning-based methods, as introduced in reference 2 (Liu D, Huang Y, Nie W, Zhang J, Deng L. SMALF: miRNA-disease associations prediction based on stacked autoencoder and XGBoost. BMC Bioinformatics, 2021, 22(1): 11-16), use a stacked autoencoder to learn the latent features of miRNAs and diseases directly from the association matrix, then integrate the functional similarity features of miRNAs and the semantic similarity features of diseases, and use an XGBoost classifier to predict the relevance. XGBoost is a powerful machine learning algorithm mainly used for classification and regression problems. The “X” stands for “Extreme”, emphasizing its efficiency; “GB” stands for “Gradient Boosting”, which is its underlying algorithm; and “oost” is the abbreviation of “boosting”, indicating that it is an algorithm based on the boosting idea. However, this method cannot preserve the topological information of the known miRNA-disease association network.

[0051] For example, based on the graph-based method as introduced in reference 3 (Li Z, Li J, Nie R, You ZH, Bao W, A graph auto-encoder model for miRNA-disease associations prediction, Briefings in Bioinformatics, 2021, 22(4): 3-6), the feature representation of miRNA and disease is obtained by aggregating the heterogeneous neighborhood features through Graph Convolutional Networks (GCN), and the connection between miRNA and disease is reconstructed using a bilinear decoder. This method can model the relationship data, but only considers the explicit pairwise relationship between nodes, and it is difficult to model the implicit non-pairwise high-order relationship.

[0052] For example, as introduced in reference 4 (Wang, Yu Tian, et al., MiRNA-disease association prediction via hypergraph learning based on high-dimensionality features, BMC Medical Informatics and Decision Making, 2021, 21. Suppl 1: 3-6), taking miRNA and disease pairs as nodes, a k-Nearest Neighbors (k-NN) algorithm is used to construct the hyperedges of the hypergraph, a matrix decomposition-based inductive hypergraph learning is used to define the objective function and obtain the mapping matrix through training, and then the association relationship between unknown miRNA and disease pairs is calculated.

[0053] However, although the introduction of hypergraph has solved the problem of difficulty in establishing implicit high-order relationships, the miRNA and disease association prediction method based on hypergraph mostly uses static hypergraph structure, i.e. heuristic hypergraph structure, which is easy to introduce noise, false negative connections (i.e. unknown associations) and task-irrelevant information, and cannot fully explore the real hypergraph structure.

[0054] Based on this, at least one embodiment of the present disclosure provides a miRNA and disease association prediction method, comprising: obtaining initial association data of miRNA and disease; calculating similarity of miRNA and similarity of disease based on the initial association data; constructing a dynamic hypergraph based on the similarity of miRNA and the similarity of disease, the dynamic hypergraph comprising first node features of miRNA, first node features of disease, and super edges connecting the first node features of miRNA and super edges connecting the first node features of disease; obtaining second node features of miRNA and second node features of disease based on the dynamic hypergraph; and obtaining an association prediction result of miRNA and disease based on the second node features of miRNA and the second node features of disease.

[0055] At least one embodiment of the present disclosure also provides a prediction device, an electronic device and a storage medium corresponding to the above-mentioned miRNA and disease association prediction method.

[0056] The miRNA and disease association prediction method provided by at least one embodiment of the present disclosure can adaptively and dynamically construct a hypergraph, the hypergraph structure is generated according to node features, and the hypergraph structure can be optimized simultaneously with the miRNA and disease association prediction task, so that the hypergraph structure can be adaptively changed according to the task and node characteristics to realize dynamic hypergraph structure learning, and update the representation of miRNA and disease, so that the implicit high-order relationship between miRNA and miRNA, and between diseases and diseases can be more fully explored, the noise in the hypergraph structure is reduced, the false negative connection is removed, and the information irrelevant to the prediction task is removed, so as to better predict the association between miRNA and disease, and improve the accuracy of miRNA-disease association prediction.

[0057] The embodiments of the present disclosure and some examples thereof will be described in detail below with reference to the accompanying drawings.

[0058] For example, the miRNA and disease association prediction method provided by at least one embodiment of the present disclosure can be implemented in the form of software, hardware, firmware or any combination thereof, and loaded and executed by a processor in a device such as a mobile phone, a digital camera, a tablet computer, a notebook computer, a desktop computer, a network server, etc.

[0059] For example, the miRNA and disease association prediction method is applicable to an association prediction device, which can be any electronic device having an association prediction function, such as a mobile phone, a digital camera, a notebook computer, a tablet computer, a desktop computer, a network server, etc., and can load and execute the miRNA and disease association prediction method. For example, the association prediction device can include a central processing unit (CPU), a graphics processing unit (GPU), or other forms of processing units having data processing and / or instruction execution capabilities, such as a digital signal processor (DSP), a storage unit, etc., and further installed with an operating system, an application programming interface (such as OpenGL (Open Graphics Library), Metal, etc.), etc., to realize the association prediction device method provided by the embodiments of the present disclosure by running codes or instructions.

[0060] FIG. 1 is a schematic diagram of a miRNA and disease association prediction method provided by at least one embodiment of the present disclosure. As shown in FIG. 1, the association prediction method includes steps S100 to S500.

[0061] Step S100: Obtain initial association data of miRNA and disease.

[0062] Step S200: Calculate the similarity of miRNA and the similarity of disease based on the initial association data.

[0063] Step S300: Construct a dynamic hypergraph based on the similarity of miRNA and the similarity of disease, wherein the dynamic hypergraph includes first node features of miRNA, first node features of disease, and hyperedges connecting the first node features of miRNA and hyperedges connecting the first node features of disease, respectively.

[0064] Step S400: Obtain second node features of miRNA and second node features of disease based on the dynamic hypergraph.

[0065] Step S500: Obtain an association prediction result of miRNA and disease based on the second node features of miRNA and the second node features of disease.

[0066] For example, for step S100, the initial association data of miRNA and disease can be the HMDD v2.0 human miRNA-disease association dataset, which can be downloaded from https: / / www.cuilab.cn / hmdd. For example, the initial association data of miRNA and disease contains 5430 associations between 495 miRNAs and 383 complex diseases that have been confirmed by biological experiments. In the Multiple Discriminant Analysis (MDA) prediction problem, known MDAs are considered to be positive samples, and unknown or non-existent MDAs are considered to be negative samples. For example, in some embodiments of the present disclosure, the initial association data of miRNA and disease can also be other datasets that can be used for association prediction, which is not limited here.

[0067] For example, at least one embodiment of the present disclosure provides an association prediction method that uses a binary matrix MD to describe the interaction between miRNA and disease, MD = R^(n m ×n d ), n m Indicates the number of miRNAs, n d Indicates the number of diseases. Each row of the matrix MD represents a miRNA, each column represents a disease, and the element value in the matrix MD is MD(m i ,d j For example, when miRNA mi is associated with disease j Related, then MD(m i ,d j )=1, otherwise MD(m i ,d j )=0, mi represents the i-th miRNA, d j represents the jth disease, i is not greater than n m A positive integer, j is not greater than n d For example, for the HMDD v2.0 human miRNA-disease association dataset, the elements of the matrix MD include 5430 1s and the rest are 0s.

[0068] For step S200 , the similarity of miRNAs may include the functional similarity MFSM of miRNAs, the Gaussian interaction spectrum kernel similarity MGSM of miRNAs, and the comprehensive similarity MISM of miRNAs.

[0069] If the functions of two miRNAs are similar, they are likely to be associated with similar diseases, based on which at least one embodiment of the present disclosure utilizes a machine inference code similarity method, i.e., the MISIM method, to calculate the function similarity score of miRNAs according to the semantic similarity of the disease sets corresponding to the miRNAs. For example, for a disease d t and a disease set D T , the overall semantic similarity value S(d t , D T ) between them can be obtained according to the following formula:

[0070] In the above formula, d t′ represents each disease in the disease set D T , and DS(d t , d t′ ) represents the specific semantic similarity value between the disease d t and each disease in the disease set D T .

[0071] For any two miRNAs m i and miRNA m j , first find the disease set D Ti and the disease set D Tj associated with them respectively, then the function similarity of the two miRNAs MFSM(m ) can be obtained according to the following formula:

[0072] In the above formula, represents an element in the disease set D Tj , represents an element in the disease set D Ti , a represents the number of diseases in the disease set D Ti , and b represents the number of diseases in the disease set D Tj . The value of an element in MFSM(m i , m j ) ranges from 0 to 1, which can represent the function similarity score between miRNA m i and miRNA m j . The larger the value of an element in MFSM(m i , m j ), the more similar the functions of miRNA m i and m j .

[0073] For example, in some embodiments of the present disclosure, the functional similarity scores of miRNAs can be downloaded directly from http: / / www.cuilab.cn / files / images / cuilab / misim.zip, and then the functional similarity of miRNAs can be constructed wherein, denotes a real set. For example, the construction process of the functional similarity of miRNAs MFSM can be determined by using conventional methods in the art, and embodiments of the present disclosure are not limited thereto.

[0074] However, due to the sparsity problem of the functional similarity of miRNAs MFSM, it is difficult to provide effective information, which may affect the prediction performance of the model. Therefore, according to the assumption that similar miRNAs are more likely to be associated with similar diseases, embodiments of the present disclosure use the known association of miRNAs and diseases in the binary matrix MD to calculate the Gaussian interaction profile kernel similarity of miRNAs

[0075] Specifically, BV(m i ) of the i-th row of the binary matrix MD is recorded as the interaction profile of miRNA m i , which is used to represent the association between miRNA m i and each disease. Therefore, the Gaussian interaction profile kernel similarity MGSM(m i ,m j ) between miRNA m i and miRNA m j can be obtained according to the following formula: MGSM(m i ,m j ) = exp(-γ m ||BV(m i )-BV(m j )|| 2 )

[0076] In the above formula, γ m represents a parameter for controlling the kernel bandwidth, the kernel function is a function for measuring the similarity between two data points, the kernel bandwidth represents the influence range or width of the kernel function, and is represented as:

[0077] Since MGSM is directly obtained from the binary matrix MD, it needs to be recalculated every time training and testing.

[0078] For example, the Gaussian interaction profile kernel similarity MGSM(mi ,m j ) to compensate for the sparsity of the MFSM, and the integrated similarity of the miRNAs is obtained according to the following formula

[0079] For example, the similarity of the diseases can include a semantic similarity of the diseases DSSM, a Gaussian interaction profile kernel similarity of the diseases DGSM, and an integrated similarity of the diseases DISM.

[0080] For example, the semantic similarity of the diseases (Deep Structured Semantic Model, DSSM) can be obtained by obtaining the relationship between different diseases from the medical subject headings (MeSH) database (https: / / www.ncbi.nlm.nih.gov / ) and performing calculation. In the MeSH database, the association between different diseases can be represented by a directed acyclic graph (DAG).

[0081] For example, the semantic similarity of the diseases (Deep Structured Semantic Model, DSSM) can be obtained by obtaining the relationship between different diseases from the medical subject headings (MeSH) database (https: / / www.ncbi.nlm.nih.gov / ) and performing calculation. In the MeSH database, the association between different diseases can be represented by a directed acyclic graph (DAG). i denotes a set composed of the self-node and the ancestor nodes of the disease d i denotes a set of direct edges from the ancestor nodes to the self-node of the disease d i According to the assumption that the more the shared parts of the DAGs of two diseases, the more similar the two diseases are, the first semantic similarity Sd i and the disease d j can be obtained according to the following formula:

[0082] In the above formula, denotes the semantic value of the disease d i denotes the contribution value of the disease d k to the semantic value of the disease d i , d k ' denotes the child node of d k , and Δ denotes the contribution factor of semantic decay, which is set to 0.5 for example. The contribution factor of the disease d i to itself is set to 1 for example.

[0083] ​​​​Since the diseases in the same DAG layer may appear at different times in the disease DAG, in order to more comprehensively calculate the semantic similarity of the disease, at least one embodiment of the present disclosure may also use the semantic similarity of another disease (i.e., the second semantic similarity of the disease). The calculation method can be obtained according to the following formula:

[0084] In the above formula, And indicates disease i The semantic value of And indicates disease k For disease i Contribution value of semantic value.

[0085] Finally, disease d i and disease j The disease semantic similarity between can be obtained according to the following formula:

[0086] Similar to the miRNA m i and miRNA m j Gaussian interaction spectrum kernel similarity MGSM(m i ,m j ) calculation method, disease d i and disease j Gaussian interaction spectrum kernel similarity between It can be obtained according to the following formula: DGSM(d i ,d j )=exp(-γ d ||BV(d i )-BV(d j )|| 2 )

[0087] In the above formula, BV(d i ) and BV(d j ) are the i-th and j-th columns of the matrix MD, γ d Represents the parameter used to control kernel bandwidth.

[0088] In at least one embodiment of the present disclosure, due to disease d i and disease j The Gaussian interaction spectrum kernel similarity DGSM between them is directly obtained from the binary matrix MD, so it also needs to be recalculated during each training and testing.

[0089] However, due to the sparsity problem of the semantic similarity of diseases DSSM, the prediction performance of the model may be affected. Therefore, at least one embodiment of the present disclosure utilizes the Gaussian interaction profile kernel similarity of diseases DGSM to make up for the semantic similarity of diseases DSSM, so that the comprehensive similarity of diseases For example, the fusion similarity of miRNAs MSM can be obtained according to the following formula:

[0090] For example, for step S200, the similarity of miRNAs also includes the fusion similarity of miRNAs MSM, and the similarity of diseases also includes the fusion similarity of diseases DSM. For example, in some cases, neither the single similarity nor the simple combination of multiple similarities (such as the similarity obtained by averaging operation, the comprehensive similarity of miRNAs MISM or the comprehensive similarity of diseases DISM in the above embodiments) can utilize the common information and complementary information of the similarity data from a deeper level. In addition, each similarity data has noise, and the above-mentioned way of obtaining a single similarity or a similarity by simple combination may have a problem of low signal-to-noise ratio. Therefore, at least one embodiment of the present disclosure learns the optimal similarity by adopting, for example, a channel attention mechanism, that is, obtains the fusion similarity of miRNAs MSM and the fusion similarity of diseases DSM, so as to solve this problem.

[0091] FIG. 2 is a structural schematic diagram of an miRNA and disease association prediction model provided by at least one embodiment of the present disclosure. As shown in FIG. 2, the prediction model includes a bipartite graph construction module, an miRNA node feature learning module and a disease node feature learning module, a graph convolution module, a dynamic hypergraph construction module, a hypergraph convolution module, and an association prediction module, each of which corresponds to a different operation step. For details, please refer to the related description of the following embodiments.

[0092] FIG. 3 is a schematic diagram of one example of step S200 shown in FIG. 1; and FIG. 4 is a schematic diagram of obtaining a fusion similarity by adopting an attention mechanism provided by at least one embodiment of the present disclosure.

[0093] For example, as shown in FIG. 3, in some examples, step S200 shown in FIG. 1 includes steps S210-S220.

[0094] Step S210: obtaining the fusion similarity of miRNAs MSM by adopting an attention mechanism based on the functional similarity of miRNAs MFSM and the Gaussian interaction profile kernel similarity of miRNAs MGSM.

[0095] Step S220: obtaining the fusion similarity of diseases DSM by adopting an attention mechanism based on the semantic similarity of diseases DSSM and the Gaussian interaction profile kernel similarity of diseases DGSM.

[0096] For example, for step S210, as shown in FIG. 2 and FIG. 4, for the miRNA functional similarity MFSM, each column can be regarded as each feature channel (C) of the miRNA, and the channel attention can be calculated as n m The importance degree of the miRNA. For example, the miRNA functional similarity MFSM can be compressed by using the average pooling method first to obtain the feature-based statistical information Z F , and As shown in the following formula:

[0097] Similarly, based on the miRNA Gaussian interaction spectrum kernel similarity MGSM, the feature-based statistical information Z G , and As shown in the following formula:

[0098] Then, the attention mechanism is used in the multi-layer perception (MLP) and combined with the Sigmoid activation function to calculate the attention weight of the miRNA functional similarity MFSM And the attention weight of the miRNA Gaussian interaction spectrum kernel similarity MGSM As shown in the following formula:

[0099] In the above formula, And are all learnable parameter matrices, and ratio represents a parameter for adjusting the number of feature channels.

[0100] By multiplying the miRNA functional similarity MFSM with its attention weight The new miRNA functional similarity By multiplying the miRNA Gaussian interaction spectrum kernel similarity MGSM with its corresponding attention weight The new miRNA Gaussian interaction spectrum kernel similarity As shown in the following formula:

[0101] Further, the fusion similarity MSM of the miRNA can be obtained according to the following formula:

[0102] For example, for step S220, by using a method similar to the calculation method of the fusion similarity MSM of the miRNA in the above example, the fusion similarity DSM of the disease can be obtained according to the following formula:

[0103] At least one embodiment of the present disclosure obtains the fusion similarity MSM of miRNA and the fusion similarity DSM of disease by adopting the channel attention mechanism, so that the public information and complementary information of each similarity data can be utilized from a deeper level, while the noise of the similarity data can be reduced, and the problem of low signal-to-noise ratio that may exist in the above-mentioned way of obtaining a single similarity or obtaining a similarity by simple combination can be solved.

[0104] FIG. 5 is a schematic diagram of one example of step S300 shown in FIG. 1.

[0105] For example, as shown in FIG. 5, in some examples, step S300 shown in FIG. 1 includes steps S310-S320.

[0106] Step S310: obtaining initial node features of miRNA and initial node features of disease based on the fusion similarity MSM of miRNA and the fusion similarity DSM of disease.

[0107] Step S320: updating the initial node features of miRNA and the initial node features of disease to obtain first node features of miRNA and first node features of disease.

[0108] For example, for step S310, it can include: adjusting the dimension of the fusion similarity MSM of miRNA to a first preset dimension by using a multi-layer perception (MLP) to obtain a fusion similarity feature of miRNA; and obtaining the initial node features of miRNA based on the fusion similarity feature of miRNA and a randomly initialized feature of miRNA.

[0109] For example, as shown in FIG. 2, the dimension of the fusion similarity MSM of miRNA can be adjusted to the first preset dimension by using a multi-layer perception, for example, the dimension of the fusion similarity MSM of miRNA can be reduced to the first preset dimension f. For example, the first preset dimension f can be flexibly determined according to actual operation conditions, and embodiments of the present disclosure are not limited thereto. The fusion similarity feature of miRNA The fusion similarity feature of miRNA can be obtained according to the following formula:

[0110] In the above formula, represents a learnable parameter matrix, and

[0111] Further, the initial node features of miRNA E m The initial node features of miRNA E

[0112] In the above formula, a random initialization feature of the miRNA, The initial value of the random initialization feature of the miRNA can be flexibly determined according to actual operation conditions, and embodiments of the present disclosure do not limit this.

[0113] For example, for step S320, the following can be included: adjusting the dimension of the fusion similarity DSM of the disease to a second preset dimension f by using a multi-layer perception machine to obtain a fusion similarity feature of the disease; and obtaining an initial node feature of the disease based on the fusion similarity feature of the disease and the random initialization feature of the disease.

[0114] For example, as shown in FIG. 2, the dimension of the fusion similarity DSM of the disease can be adjusted to the second preset dimension f by using a multi-layer perception machine, for example, the dimension of the fusion similarity DSM of the disease can be reduced to the second preset dimension f. For example, the second preset dimension f can be flexibly determined according to actual operation conditions, and embodiments of the present disclosure do not limit this. For example, the second preset dimension f can be equal to the first preset dimension f, or can not be equal, which can be determined according to actual conditions, and embodiments of the present disclosure do not limit this. Here, in order to facilitate understanding, the second preset dimension uses the same symbol as the first preset dimension. The fusion similarity feature of the disease The fusion similarity feature of the disease can be obtained according to the following formula:

[0115] In the above formula, denotes a learnable parameter matrix, and

[0116] Further, the initial node feature E of the disease d The initial node feature E of the disease can be obtained according to the following formula:

[0117] In the above formula, denotes a random initialization feature of the disease, The initial value of the random initialization feature of the disease can be flexibly determined according to actual operation conditions, and embodiments of the present disclosure do not limit this.

[0118] FIG. 6 is a schematic diagram of one example of step S320 shown in FIG. 5.

[0119] For example, as shown in FIG. 6, in some examples, step S320 shown in FIG. 5 includes steps S321-S324. For example, steps S321-S324 can be used to construct a miRNA-disease bipartite graph.

[0120] Step S321: calculating an aggregated interaction feature vector of the miRNA based on the comprehensive similarity MISM of the miRNA.

[0121] Step S322: calculating the aggregated interaction feature vector of the disease based on the comprehensive similarity DISM of the disease.

[0122] Step S323: calculating the miRNA-disease interaction probability matrix based on the aggregated interaction feature vector of the miRNA and the aggregated interaction feature vector of the disease.

[0123] Step S324: obtaining the miRNA-disease bipartite graph based on the miRNA-disease interaction probability matrix and the preset negative connection value, the miRNA-disease bipartite graph comprising the initial node features of the miRNAs and the initial node features of the diseases.

[0124] For example, for step S321, the interaction feature vector of the miRNA m i in the i-th row of the binary matrix MD is denoted as MD(m i ), and denote the interaction features between the miRNA m i and other miRNAs, respectively. The interaction feature vector of the disease d j in the j-th column of the binary matrix MD is denoted as MD(d j ), and denote the interaction features between the disease d j and other diseases, respectively.

[0125] For example, in the WKNNP algorithm, for the miRNA m i , the first K most similar miRNAs to the miRNA m i can be obtained according to the comprehensive similarity MISM of the miRNAs, and the interaction feature vectors of these similar miRNAs are aggregated to obtain the aggregated interaction feature vector MD i (m m ) of the miRNA m i , as follows:

[0126] In the above formula, m1 to m k denote a plurality of miRNAs ranked in descending order according to the comprehensive similarity of the miRNA m i ; ω k denotes a decay factor, and ω k = α k-1 MISM(m k , m i ), the value of ω k is higher if m k is more similar to m i ; α ∈ [0, 1]; Q m = ∑1≤k≤K MISM(m k ,m i ), and Q m is a regularization term.

[0127] For example, for step S322, similar to the method of calculating the aggregation interaction feature vector of miRNA in the above example, for disease d j , its aggregation interaction feature vector MD d (d j ) can be obtained according to the following formula:

[0128] In the above formula, d1 to d k represent multiple diseases arranged in descending order according to the comprehensive similarity of disease d j ; ω k = α k-1 DISM(d k , d j ), and ω k represents a decay factor; Q d = ∑ 1≤k≤K DISM(d k , d j ), and Q d is a regularization term. In the above formula, the values of K and α can be determined according to actual operation, for example, K can be 24 and α can be 0.8, but embodiments of the present disclosure are not limited thereto.

[0129] For example, for step S323, the initial miRNA-disease interaction probability matrix R1 can be obtained according to the following formula:

[0130] For example, for step S324, taking miRNA and disease as nodes and known associations between miRNA and disease as positive connections. For example, for the i-th miRNA and the j-th disease, MD(m i , d j ) = 1. In addition, at least one embodiment of the present disclosure randomly extracts an equal number of unknown associations from unknown associations other than test samples as negative connections, for example, setting their values to 0.5. In this way, the extracted unknown associations can be distinguished from positive connections and cases where no association is determined, that is, distinguished from positive edges and unconnected edges, thereby reducing the dependence on verified miRNA and disease associations.

[0131] Further, based on the initial miRNA-disease interaction probability matrix R1 and the negative correlation, the final miRNA-disease interaction probability matrix R is obtained, as shown in the following formula:

[0132] For example, according to the above formula and FIG. 2, it can be known that the miRNA-disease bipartite graph G includes initial node features of miRNAs and initial node features of diseases.

[0133] For example, for step S320, further comprising: calculating an adjacency matrix A of the miRNA-disease bipartite graph G G and a degree matrix D G ; and obtaining first node features of miRNAs and first node features of diseases based on the adjacency matrix A G and the degree matrix D G . For example, the process can use a graph convolution operation.

[0134] For example, the adjacency matrix A of the miRNA-disease bipartite graph G G can be obtained according to the following formula:

[0135] wherein, and D G is a diagonal matrix, D G [i, i] represents the number of non-zero entries in the ith row of the adjacency matrix A G .

[0136] For example, the miRNA-disease bipartite graph G can encode the features of miRNAs or diseases by showing local structures. For example, in the initialization process of the miRNA-disease bipartite graph G, the embedding of miRNAs is the initial node feature of miRNAs, and the embedding of diseases is the initial node feature of diseases.

[0137] Based on a light graph convolution network (Light Graph Convolution Network, LightGCN), at least one embodiment of the present disclosure uses a graph convolution that removes a learnable matrix of linear transformation and a nonlinear activation function, for example, a rectified linear unit (Rectified Linear Unit, ReLU) and the like, to capture local topological relationships between nodes, so that the model can be made more lightweight and easy to train.

[0138] For example, in the explicit propagation process on the miRNA-disease bipartite graph G, the first node features of miRNAs and the first node features of diseases can be obtained as follows:

[0139] In the above formula, E (0)an initial value of the miRNA-disease bipartite graph G, E *,(l) a first node feature corresponding to the miRNA of the lth layer and a first node feature corresponding to the disease of the lth layer, i.e., a node feature of the updated miRNA-disease bipartite graph G, a first node feature corresponding to the miRNA of the lth layer, a first node feature corresponding to the disease of the lth layer, and l is a positive integer greater than 1.

[0140] or E (0) , a second node feature of the miRNA of the l-1th layer, a second node feature of the disease of the l-1th layer. The second node feature E m of the miRNA and the second node feature E d of the disease are features obtained after hypergraph convolution. For details of the process of hypergraph convolution, please refer to the description of the following embodiments. Thus, when the miRNA-disease bipartite graph G is propagated each round, the nodes in the miRNA-disease bipartite graph G can be explicitly updated by the local topological relationship of the nodes, so as to prepare for the next step of constructing a dynamic hypergraph.

[0141] For example, for step S300, a hypergraph H can be represented, and H=(V, ε, w), V represents a set of nodes, ε represents a set of hyperedges, and the positive weight of each hyperedge e is w(e). The hypergraph H can be represented by an incidence matrix I H of |V|x|ε|, and the entries are defined as:

[0142] For a node v in V and a hyperedge e in ε, the node degree d(v) and the hyperedge degree δ(e) can be obtained according to the following formula:

[0143] In addition, the diagonal matrix of the node degree d(v) is D v , and the diagonal matrix of the hyperedge degree δ(e) is D e .

[0144] For example, as shown in FIG. 2, for miRNA, at least one embodiment of the present disclosure takes miRNA as a node, and learns the hyperedge of miRNA based on the node representation after graph convolution (e.g., the first node feature of miRNA). For disease, at least one embodiment of the present disclosure takes disease as a node, and learns the hyperedge of disease based on the node representation after graph convolution (e.g., the first node feature of disease). The node representation after graph convolution is the first node feature of miRNA and the first node feature of disease. In this way, the message passing restriction of fixed heuristic hypergraph structure is broken, and the implicit high-order relationship between nodes can be better learned.

[0145] For example, for the l-th layer, at least one embodiment of the present disclosure adopts MLP to calculate the hyperedge assignment of miRNA and disease, to obtain the hypergraph association matrix of miRNA and the hypergraph association matrix of disease and k m and k d are the number of hyperedges. For example, the calculation process can adopt a "unbiased" manner, that is, no additional bias term or bias parameter is introduced, so that the calculation process can be simplified. For example, the process of constructing a hypergraph can be regarded as a clustering process, and the number of hyperedges is the number of clusters. K-Means clustering is an unsupervised clustering algorithm, and the number of hyperedges k in the hypergraph association matrix of miRNA m and the number of hyperedges k in the hypergraph association matrix of disease d can be determined according to the elbow method in k-Means clustering, but embodiments of the present disclosure are not limited thereto.

[0146] For example, at least one embodiment of the present disclosure adopts Gumbel-Softmax to construct the dynamic hypergraph structure of miRNA and the dynamic hypergraph structure of disease, and the l-th layer hypergraph association matrix of miRNA and the l-th layer hypergraph association matrix of disease are obtained according to the following formulas respectively:

[0147] In the above formulas, and are learnable matrices, and In embodiments of the present disclosure, the learnable matrix represents a matrix whose value can be automatically adjusted during the training process.

[0148] Gumbel-Softmax is a reparameterization trick for discrete random variables, which can be used to optimize models involving discrete choices by gradient descent. In traditional neural networks, the data and gradients are usually continuous, but when discrete choices are involved (such as class selection in classification problems), direct gradient propagation becomes difficult because discrete choices are not differentiable. Gumbel-Softmax provides a way to convert discrete choices into a continuous and differentiable form, so that gradients can be backpropagated.

[0149] Therefore, in a hypergraph, a hyperedge can connect multiple nodes, and Gumbel-Softmax can achieve a differentiable way to "softly" assign nodes to different hyperedges, and optimize model parameters through backpropagation. Through multiple iterations and optimization processes, the hypergraph structure will learn an optimal node assignment strategy. For example, "softly" here means non-deterministically. Thus, by using Gumbel-Softmax, a node can be assigned to a certain hyperedge with a high probability and to other hyperedges with a low probability, thereby ensuring the exploratory nature of the model. In addition, by using Gumbel-Softmax, the value of the hyperedge can be represented as a probability distribution between 0 and 1, so that the normalization operation can be omitted in the subsequent hypergraph convolution process (see the related description of the embodiments below), which is beneficial to simplify the hypergraph convolution process.

[0150] For example, as shown in FIG. 2, the dynamic hypergraph includes a miRNA dynamic hypergraph and a disease dynamic hypergraph, the miRNA dynamic hypergraph includes first node features of miRNAs and hyperedges connecting the first node features of miRNAs, and the disease dynamic hypergraph includes first node features of diseases and hyperedges connecting the first node features of diseases.

[0151] FIG. 7 is a schematic diagram of one example of step S400 shown in FIG. 1.

[0152] For example, as shown in FIG. 7, in some examples, step S400 shown in FIG. 1 includes steps S410-S420. For example, steps S410-S420 can be operations of performing hypergraph convolution on the hypergraph structure constructed in the above embodiments, and correspond to the hypergraph convolution module in FIG. 2.

[0153] Step S410: performing hypergraph convolution on the lth layer hypergraph association matrix of miRNA to obtain the second node feature of miRNA.

[0154] Step S420: performing hypergraph convolution on the lth layer hypergraph association matrix of disease to obtain the second node feature of disease.

[0155] For example, for step S410, based on the obtained miRNA's 1st layer hypergraph association matrix The implicit high-order relationship between miRNAs can be learned in a spectral hypergraph convolution manner, and node features are aggregated to hyperedge using node convolution, and hyperedge features are aggregated to centroid node using hyperedge convolution. The spectral hypergraph convolution is a graph convolution method based on spectral representation of graph (i.e. eigen-decomposition of Laplacian matrix of graph), and its core concept is similar to spectral graph convolution, and it is applied to hypergraph structure. The basic idea of spectral hypergraph convolution is to convert node features of hypergraph into spectral domain (i.e. eigenvector space of Laplacian matrix) for processing, and then convert back to spatial domain. At least one embodiment of the present disclosure can accurately capture global information in hypergraph structure by adopting spectral hypergraph convolution manner.

[0156] For example, the miRNA's 1st layer second node feature The miRNA's 1st layer second node feature can be obtained according to the following formula:

[0157] In the above formula, represents the miRNA's 1st layer hypergraph association matrix The hyperedge degree matrix of is used for normalization, and node convolution is realized by , and hyperedge convolution is realized by

[0158] According to the above formula, the miRNA's 1st layer corresponding first node feature As the last term in the above formula, the process of obtaining the miRNA's 1st layer second node feature is realized in a skip connection manner to retain local topological information obtained by graph convolution and explicit interaction information between miRNAs, which is beneficial to avoid gradient disappearance.

[0159] For example, similar to the way of obtaining the miRNA's 1st layer second node feature , for step S420, based on the obtained disease's 1st layer hypergraph association matrix The implicit high-order relationship between diseases can be learned in a spectral hypergraph convolution manner.

[0160] For example, the disease's 1st layer second node feature The disease's 1st layer second node feature can be obtained according to the following formula:

[0161] In the above formula, represents the disease's 1st layer hypergraph association matrix the superedge incidence matrix of the bipartite graph, and the node convolution is realized by the superedge incidence matrix of the bipartite graph, and the node convolution is realized by . and as the node features of the miRNA-disease bipartite graph in the graph convolution operation to continue learning.

[0162] According to the above formula, the first node feature of the lth layer of the disease is As the last term in the above formula, the second node feature of the lth layer of the disease is obtained The process is realized in the form of a skip connection to preserve the local topological information obtained by graph convolution and the explicit interaction information between diseases, which is beneficial to avoid gradient disappearance.

[0163] For step S500, as shown in FIG. 2, it can include: splicing the second node features in all layers of the miRNA to obtain the third node features of the miRNA; splicing the second node features in all layers of the disease to obtain the third node features of the disease; and obtaining the miRNA-disease association matrix based on the third node features of the miRNA and the third node features of the disease as the association prediction result of the miRNA and the disease.

[0164] For example, the steps corresponding to the graph convolution module, the dynamic supergraph construction module, and the supergraph convolution module can be stacked in multiple layers, for example, L layers, and each layer can output a second node feature corresponding to the miRNA and a second node feature corresponding to the disease, and the output of each layer contains semantic information of different levels. Thus, the third node features E M of the miRNA and the third node features E D of the disease can be obtained according to the following formula:

[0165] In the above formula, concat represents a function for splicing, respectively represent the second node features corresponding to the miRNA output by each layer, respectively represent the second node features corresponding to the disease output by each layer.

[0166] For example, the third node features E M of the miRNA and the third node features E D of the disease can be subjected to inner product operation to obtain the miRNA-disease association matrix, as shown in the following formula:

[0167] The miRNA-disease association matrix can be used as the association prediction result of the miRNA and the disease to represent the association between the miRNA and the disease.

[0168] For example, the method for predicting the association between miRNA and disease provided by at least one embodiment of the present disclosure further comprises calculating a total loss function of miRNA and disease. By calculating the total loss function of miRNA and disease, the accuracy of the prediction result of the association between miRNA and disease can be better controlled.

[0169] For example, the total loss function L of miRNA and disease total The total loss function L of miRNA and disease can be obtained according to the following formula:

[0170] In the above formula, L BPR represents the association prediction loss function, L sim represents the node feature learning loss function, L const represents the node constraint loss, represents a regularization term. By adding the regularization term, the model can be made more concise and the risk of overfitting can be reduced. β, ρ, μ all represent hyperparameters.

[0171] For example, based on the third node feature of miRNA and the third node feature of disease, the association prediction loss function L BPR The association prediction loss function L can be obtained according to the following formula:

[0172] In the above formula, is a pair of positive and negative sample training data, represents a positive sample pair, represents a negative sample pair. The positive sample pair refers to a pair of miRNA and disease that are known to be associated, and the negative sample pair refers to a pair of miRNA and disease that are not associated.

[0173] For example, the score of the area under the curve (Area Under the Curve, AUC) represents the ability of the classification model to rank the positive sample pair ahead of the negative sample pair. The closer the value is to 1, the better the performance of the classification model. The association prediction loss function L BPR By comparing the ranking relationship of the positive and negative sample pairs, the purpose of optimizing the model can be achieved. The association prediction loss function L BPR is a pairwise loss function for a recommendation system. The core is to maximize the difference in predicted ranking of positive and negative samples, so that the score of the associated miRNA and disease is as large as possible compared to the score of the unassociated miRNA and disease. Therefore, the association prediction loss function L BPR can optimize the score of the area under the curve (Area Under the Curve, AUC).

[0174] For example, based on the miRNA fusion similarity MSM, the disease fusion similarity DSM, and the interaction probability matrix R of miRNA and disease, a node feature learning loss function L sim The node constraint loss L

[0175] In the above formula, the miRNA fusion similarity MSM(m i ,m j ) and the disease fusion similarity DSM(d i ,d j ) can be obtained according to the method described in the foregoing embodiments, which will not be repeated here. R T represents the first-order reachable matrix of miRNA, which can represent the number of interaction diseases of each miRNA pair, R T R represents the first-order reachable matrix of disease, represents the Frobenius norm. According to the above formula, when the miRNA fusion similarity MSM is closer to the first-order reachable matrix RR T of miRNA, and when the disease fusion similarity DSM is closer to the first-order reachable matrix R T of disease, it is more conducive to reducing the node feature learning loss function L Sim .

[0176] For example, the process of constructing a hypergraph is equivalent to a clustering process, each hyperedge is equivalent to a clustering center, and the process of connecting a hyperedge to a node is equivalent to the process of dividing nodes with the same class into the same cluster. In order to make the nodes divided into the same hyperedge more similar, at least one embodiment of the present disclosure adds an unsupervised constraint to the node feature. Since the Gumbel-softmax is used to generate the distribution of nodes in different hyperedges when constructing the hypergraph structure, all nodes in the same hyperedge are involved, and therefore the problem can be simplified to calculating the similarity of all node pairs.

[0177] Based on the third node feature of miRNA and the third node feature of disease, the node constraint loss L cOnst can be obtained by the following formula:

[0178] In the above formula, d(·) represents the Euclidean distance, which can evaluate the similarity of two nodes; represents the first constraint loss for miRNA; represents the second constraint loss for disease.

[0179] Table 1 shows the prediction results of the miRNA-disease association prediction method provided by at least one embodiment of the present disclosure corresponding to other methods. For example, the miRNA-disease association prediction method corresponds to the prediction model shown in FIG2 .

[0180] For example, the miRNA-disease association prediction method based on dynamic hypergraph learning (DHLMDA) adopts an end-to-end training method. During training, the batch size, i.e., Batch size, is set to 240; the number of times the entire training dataset is completely traversed by the neural network is epoch, and 100 epochs are trained; the Adaptive Moment Estimation (Adam) algorithm is used to optimize the model; the initial learning rate Lr is set to 0.001; the multi-step learning rate attenuator (MultiStepLR) attenuation strategy is adopted, and the learning rate decreases by 0.01 times at epoch20 and epoch40 respectively. For example, based on experience, when calculating the total loss function Lr of miRNA and disease, total The hyperparameters β = 0.5, ρ = 0.1, and μ = 0.1 for the Gumbel-Softmax algorithm. The temperature coefficient τ represents the interpolation of the Gumbel-Softmax distribution between the discrete one-hot encoding categorical distribution and the continuous categorical density. For example, at low temperatures, τ ranges from 0.1 to 0.5, and the expected value of the Gumbel-Softmax random variable approaches the expected value of a categorical random variable with the same logits. As the temperature increases, τ ranges from 0.1 to 10.0, and the expected value of the Gumbel-Softmax random variable covers the entire categorical range. At higher temperatures, Gumbel-Softmax samples are no longer one-hot; that is, they no longer choose a single category with high probability, but instead choose multiple categories with a certain probability distribution. Furthermore, as τ approaches infinity (i.e., τ → ∞), the Gumbel-Softmax distribution becomes uniform. During training, τ is set to 1. When calculating the attention weight, ratio=3, the number of stacking layers L=2, the node feature dimension f=100, and according to the elbow method in k-Means clustering, k is selected. m =8,k d = 16. For example, the settings of the above parameters may be determined according to specific experimental results, and the embodiments of the present disclosure are not limited thereto.

[0181] To ensure the fairness of the validation results, at least one embodiment of the present disclosure uses 5-fold cross validation (5-CV) and Leave-One-Out Cross-Validation (LOOCV) to evaluate the performance of DHLMDA based on the aforementioned 5430 experimentally verified associations in the HMDD v2.0 human miRNA-disease association dataset. The value of the corresponding position in the binary matrix MD of the test sample is set to 0. Considering that the calculation results of the Gaussian interaction profile kernel similarity of miRNA MGSM and the Gaussian interaction profile kernel similarity of disease DGSM depend on the known miRNA-disease association, therefore, it is necessary to recalculate according to the matrix MD with the different test samples each time.

[0182] In 5-fold cross validation, in order to construct a balanced dataset, 5430 experimentally verified associations in the HMDD v2.0 human miRNA-disease association dataset are taken as positive samples. All known associations are divided into 5 groups, and each time 1 group is taken as a test sample set and the remaining 4 groups are taken as a training sample set. This is done five times, and each time a different group is taken as a test sample set and the remaining 4 groups are taken as a training sample set. For test positive samples, the same number of unknown associations are randomly selected from the pool of unknown associations as test negative samples. For training positive samples, 20 diseases (or miRNAs) unknown to the miRNA (or disease) are randomly selected from the remaining unknown associations outside the test negative samples as negative samples to join the training set, to form a positive and negative sample pair. Repeat the 5-fold cross validation multiple times (for example, 10 times), and finally take the average of the results obtained as the final result.

[0183] In global leave-one-out cross-validation, each known association of miRNA and disease is taken in turn as a test sample. In addition to the association selected as the test sample, other known associations are taken as training samples, and other unknown associations not involved in the training process are taken as candidate samples. After the model training is completed, the test sample is used together with all the candidate samples. For example, a score or probability is assigned to each sample (including test samples and candidate samples), indicating the likelihood of an association between them. Then, all samples are ranked according to these scores or probabilities. Finally, the predictive performance of the model is evaluated by checking the position of the test sample in the ranking. If the test sample is located in an earlier position in the ranking, it indicates that the model can better predict this hidden association, thereby verifying the effectiveness of the model.

[0184] On this basis, at least one embodiment of the present disclosure measures the prediction performance of the DHLMDA, for example, by using 6 indicators, including Accuracy, Precision, Recall, F1-score, AUC and AUPR. The Accuracy represents the accuracy, which can represent the ratio of the number of samples correctly classified by the classification model to the total number of samples; the Precision represents the precision, which can represent the proportion of true positive samples in all samples predicted to be positive; the Recall represents the recall, which can represent the proportion of samples predicted to be positive in the actual positive samples; the F1-score represents the F1 score, which is the harmonic mean of the precision and the recall, and the higher the F1 score, the better the classification model performs in terms of precision and recall; the AUC, as described in the above embodiment, the closer the AUC value is to 1, the better the prediction performance of the model; and the AUPR represents the area under the precision-recall curve, which measures the precision performance of the model at different recall levels, and the higher the AUPR, the better the model can maintain high precision while maintaining high recall.

[0185] Table I

[0186] As shown in Table I, several other prediction methods are used as comparative examples, which are MERWMDA, SAEMDA, SFGAE and HFHLMDA, representing four types of calculation methods based on network similarity, machine learning, ordinary graph and hypergraph, respectively. These methods are used to obtain the comparative prediction effect. As can be seen from Table I, the calculation method based on the ordinary graph corresponds to the second best prediction result of the prediction method for the correlation between miRNA and disease provided by at least one embodiment of the present disclosure. The prediction result corresponding to the prediction method for the correlation between miRNA and disease provided by at least one embodiment of the present disclosure has the highest value in the above 6 indicators, and the results corresponding to the Accuracy, Precision, Recall, F1-score, AUC and AUPR are 1.41%, 2.39%, 0.07%, 1.23%, 1% and 1.47% higher than the results corresponding to SFGAE, respectively.

[0187] FIGS. 8-9 are schematic diagrams of global leave-one-out cross-validation results corresponding to the prediction method for the correlation between miRNA and disease provided by at least one embodiment of the present disclosure and other methods. For example, in FIG. 8, the horizontal axis represents the false positive rate, and the vertical axis represents the true positive rate. In FIG. 9, the horizontal axis represents the recall rate, and the vertical axis represents the precision.

[0188] As shown in FIG. 8, the result of AUC corresponding to the miRNA and disease association prediction method provided by at least one embodiment of the present disclosure is 0.9471, and the results of AUC corresponding to MERWMDA, SAEMDA, SFGAE and HFHLMDA are 0.8764, 0.9219, 0.9316 and 0.9213 respectively. Therefore, the AUC corresponding to the miRNA and disease association prediction method provided by at least one embodiment of the present disclosure is higher than the results of AUC corresponding to the above four methods, so that the association prediction method provided by the embodiment of the present disclosure has good association prediction ability.

[0189] For example, as shown in FIG. 9, the result of AUPR corresponding to the miRNA and disease association prediction method provided by at least one embodiment of the present disclosure is 0.9593, and the results of AUPR corresponding to MERWMDA, SAEMDA, SFGAE and HFHLMDA are 0.9170, 0.9447, 0.9462 and 0.9375 respectively. Therefore, the AUPR corresponding to the miRNA and disease association prediction method provided by at least one embodiment of the present disclosure is higher than the results of AUPR corresponding to the above four methods, so that the association prediction method provided by the embodiment of the present disclosure can enable the model to maintain high recall rate while also maintaining high precision.

[0190] In order to further verify the ability of DHLMDA to predict the association between miRNA and disease, at least one embodiment of the present disclosure also predicts common human diseases respectively.

[0191] For example, the first case study is used to evaluate the ability of DHLMDA to predict the association between potential miRNA and disease. The known association between miRNA and disease in the above HMDD v2.0 database is used as input data to train the model, and the target disease is given, and the miRNA without known association with it is used as candidate miRNA. Then, according to the descending order of the score, the top 50 miRNAs are verified by using two miRNA-disease association databases dbDEMC (http: / / www.picb.ac.cn / dbDEMC / ) and miR2Disease. The database miR2Disease provides detailed association information between miRNA and various human diseases, for example, which diseases are associated with miRNA can be found by searching the ID of miRNA, the name of disease or the name of target gene. In the first case study, kidney tumor and thyroid tumor are used as research objects.

[0192] Table II shows the top 50 miRNAs associated with kidney tumor predicted by the miRNA and disease association prediction method provided by at least one embodiment of the present disclosure.

[0193] Some studies show that kidney tumors rank second in the most common urogenital system tumors in humans, next to bladder tumors, and the incidence is higher in North America and Northern Europe. As the first case study of the first type, the top 50 miRNAs associated with kidney tumors were all successfully verified by the dbDEMC database, and among the 50 predicted associations, 24 were verified by both the dbDEMC database and the miR2Disease database.

[0194] Table III shows the top 50 miRNAs associated with thyroid tumors predicted by the miRNA and disease association prediction method provided by at least one embodiment of the present disclosure.

[0195] Some studies show that thyroid tumors are one of the most common malignant tumors of the endocrine system, ranking ninth in the world in terms of incidence, and are more common in women. As the second case study of the first type, 49 of the top 50 miRNAs associated with kidney tumors were successfully verified by the dbDEMC database, and 1 was verified by both the dbDEMC database and the miR2Disease database, as shown in Table III.

[0196] For example, in the second case study, in order to simulate the prediction ability of DHLMDA in the absence of known miRNA associations, known associations related to the target disease were deleted during training, and then the top 50 predicted potential miRNAs were still verified by the dbDEMC database and the miR2Disease database. In the second case study, the research object is cervical cancer.

[0197] Some studies show that cervical cancer is one of the top four cancers in the world, with more than 500,000 cases diagnosed each year, and about 300,000 people die from cervical cancer each year. miRNAs play an important role in the occurrence, invasion, metastasis and recurrence of cervical cancer. As the second case study of the second type, 47 of the top 50 miRNAs associated with cervical cancer were verified by the dbDEMC database, of which 3 were verified by both the dbDEMC database and the miR2Disease database, as shown in Table IV.

[0198] According to the verification results in Tables I-IV and Figs. 8-9, the miRNA-disease association prediction method (e.g., the prediction model shown in Fig. 2) provided by at least one embodiment of the present disclosure can dynamically and adaptively construct a hypergraph, and the hypergraph structure can be optimized simultaneously with the miRNA-disease association prediction task, to achieve dynamic hypergraph structure learning, so that the implicit high-order relationships between miRNAs and between diseases can be more fully explored, and the association between miRNAs and diseases can be better predicted, thereby improving the accuracy of miRNA-disease association prediction.

[0199] Table II

[0200] Table III

[0201] Table IV

[0202] Fig. 10 is a structural schematic diagram of another miRNA-disease association prediction model provided by at least one embodiment of the present disclosure. As shown in Fig. 10, the model includes a prior hypergraph construction module, a node updating module, a dynamic hypergraph construction module, a hypergraph convolution module, and an association prediction module, each of which corresponds to a different operation, and specific details can be found in the related descriptions of the following embodiments.

[0203] For example, as shown in Figs. 1 and 10, for steps S100 and S200, the data acquisition method, the miRNA functional similarity MFSM, the miRNA Gaussian interaction profile kernel similarity MGSM, the disease semantic similarity DSSM, the disease Gaussian interaction profile kernel similarity DGSM, the miRNA comprehensive similarity MISM, the disease comprehensive similarity DISM, the miRNA fusion similarity MSM, and the disease fusion similarity DSM of the prediction model shown in Fig. 10 are the same as those of the prediction model shown in Fig. 2 in the above embodiments, and will not be repeated here.

[0204] For example, as shown in Figs. 1 and 10, in the prior hypergraph construction module, for step S300, it can include: based on the miRNA comprehensive similarity MISM, taking each row in the miRNA comprehensive similarity MISM as the initial node feature of the miRNA, to construct a prior hypergraph of the miRNA; and based on the disease comprehensive similarity DISM, taking each row in the disease comprehensive similarity DISM as the initial node feature of the disease, to construct a prior hypergraph of the disease.

[0205] For example, as shown in FIG. 10, a prior hypergraph of miRNAs can be constructed based on the binary matrix MD, with miRNAs as nodes and diseases as hyperedges, when miRNA m i is related to disease d j , then the corresponding node belongs to disease d i and the corresponding hyperedge belongs to miRNA m j . For example, for the HMDD v2.0 human miRNA-disease association dataset, since 5430 associations between 495 miRNAs and 383 complex diseases have been confirmed by biological experiments, the dataset has 383 hyperedges. The binary matrix MD is the association matrix of the prior hypergraph of miRNAs, each row of the binary matrix MD represents a miRNA, and each column represents a disease. Thus, the rows of the binary matrix MD represent nodes, and the columns represent hyperedges. After obtaining the comprehensive similarity of miRNAs MISM, each row in the comprehensive similarity of miRNAs MISM is taken as the initial node feature of the miRNA to construct the prior hypergraph of miRNAs.

[0206] For example, as shown in FIG. 10, a prior hypergraph of diseases can be constructed based on the transpose matrix MDT of the binary matrix MD, with diseases as nodes and miRNAs as hyperedges, when disease d i is related to miRNA m j , then the corresponding node belongs to miRNA m i and the corresponding hyperedge belongs to disease d j . Thus, the HMDD v2.0 human miRNA-disease association dataset has 495 hyperedges. The transpose matrix MDT of the binary matrix MD is the association matrix of the prior hypergraph of diseases, and the rows of the transpose matrix represent nodes and the columns represent hyperedges. After obtaining the comprehensive similarity of diseases DISM, each row in the comprehensive similarity of diseases DISM is taken as the initial node feature of the disease to construct the prior hypergraph of diseases.

[0207] FIG. 11 is a schematic diagram of another example of step S300 shown in FIG. 1. For example, in some examples, for step S300, after the construction of the prior hypergraph of miRNAs and the prior hypergraph of diseases is completed, steps S301-S302 are further included.

[0208] Step S301: updating the initial node features of miRNAs to obtain first node features of miRNAs by using HyperGAT, and processing the first node features of miRNAs by using Gumbel-Topk and KMeans to construct a dynamic hypergraph of miRNAs.

[0209] Step S302: updating the initial node features of the diseases by HyperGAT to obtain first node features of the diseases, and processing the first node features of the diseases by Gumbel-Topk and KMeans to construct a dynamic hypergraph of the diseases.

[0210] For example, HyperGAT (Hypergraph Attention Networks) is a kind of graph neural network based on attention mechanism, which can be applied to hypergraph. For example, for each node in the hypergraph, first, all the hyperedges connected to it are determined according to HyperGAT. Since each node in the hyperedge has different contributions to the hyperedge, HyperGAT can use the attention mechanism to calculate the contribution of each node hyperedge and obtain the attention weight of each node. In this process, the introduction of the attention mechanism can highlight the nodes that are meaningful to the hyperedge, and then according to the attention weight, the features of all nodes in the hyperedge are weighted and aggregated. Subsequently, the features of the nodes are updated again using the hyperedge features. Since different hyperedges have different contributions to the update of the node features, HyperGAT can again use the attention mechanism to calculate the attention weight of each hyperedge when updating the node features, so as to highlight those hyperedges that are meaningful to the node, and then according to the attention weight, the features of all hyperedges connected to the node are weighted and aggregated. In this way, the features of the node are updated.

[0211] Therefore, for step S301 and step S302, as shown in FIG. 10, in the node updating module, the initial node features of the miRNAs are updated by HyperGAT, and the initial nodes of the diseases are updated. The first node features of the miRNAs are the nodes obtained after the initial nodes in the prior hypergraph of the miRNAs are updated, and the first node features of the diseases are the nodes obtained after the initial nodes in the prior hypergraph of the diseases are updated. By using the attention mechanism to assign different attention weights to different hyperedges and nodes, the association between the miRNAs and the diseases can be more accurately obtained.

[0212] For example, “dropout” (random dropout) is an effective regularization technique. In some embodiments of the present disclosure, during the process of updating the initial node features of the miRNAs and the initial nodes of the diseases by HyperGAT, the “dropout” (random dropout) method can also be used, which can prevent overfitting and improve the stability and generalization ability of the model.

[0213] For example, the process of updating the initial node features of miRNAs and updating the initial nodes of diseases by HyperGAT can be repeated multiple times to more comprehensively and accurately extract the features of the initial nodes of miRNAs and the initial nodes of diseases, and to enhance the stability of model training.

[0214] For example, for step S301, Gumbel-TopK can be used to differentially perceive the local feature information of nodes in the dynamic hypergraph construction module. For example, the P (P is a positive integer) miRNA nodes closest to each miRNA node can be found in the hypergraph of miRNAs (i.e., the hypergraph in which the first node features of miRNAs are located), and the closer the distance, the more similar the nodes are, and these nodes are constructed in a hyperedge. However, since the TopK operation itself is not differentiable, the model cannot be trained by backpropagation. Therefore, in order to better update the model, at least one embodiment of the present disclosure also uses the Gumbel reparameterization trick, that is, by introducing Gumbel noise and a softmax function, the non-differentiable discrete selection process is converted into a differentiable, probabilistic selection process, so that the model can be trained and optimized by backpropagation.

[0215] For example, in the hypergraph structure, KMeans can perceive the global feature information of nodes. For each miRNA, the S-1 nearest clusters (except the cluster in which it is located) can be found, and then all miRNA nodes in these clusters are constructed in the same hyperedge. This method constructs hyperedges based on global feature information, which can capture the global similarity between miRNA nodes.

[0216] Similarly, for step S302, the local feature information of the first node features of diseases is differentially perceived by Gumbel-Topk. For example, the P (P is a positive integer) disease nodes closest to each disease node can be found in the hypergraph of diseases (i.e., the hypergraph in which the first node features of diseases are located), and these nodes are constructed in a hyperedge. For example, in some embodiments, the Gumbel reparameterization trick can be used to enable the model to be trained and optimized by backpropagation.

[0217] In the process of constructing the dynamic hypergraph of miRNA and the dynamic hypergraph of disease, the Gumbel-TopK and KMeans are combined to consider the local similarity between nodes and capture the global structure of data, thereby adaptively supplementing the global and local high-order relationships and improving the ability of the model to predict the association between miRNA and disease, so that the hypergraph can comprehensively represent the complex relationship and interaction mode between miRNA and disease. In addition, since the dynamic hypergraph of miRNA and the dynamic hypergraph of disease can be reconstructed in the process of iterating the model through Gumbel-Topk and KMeans, the high-order relationship between miRNAs and the high-order relationship between diseases can be utilized in the operation process of the model, so that the dynamic hypergraph of miRNA and the dynamic hypergraph of disease can better adapt to the association prediction task, and the prediction model can fully explore the dynamic hypergraph structure of miRNA and disease, thereby better predicting the association between miRNA and disease.

[0218] For example, as shown in FIG. 1 and FIG. 10, step S400 can be implemented as: performing node convolution operation and hyperedge convolution operation on the dynamic hypergraph of miRNA to obtain the second node feature of miRNA in the hypergraph convolution module; and performing node convolution operation and hyperedge convolution operation on the dynamic hypergraph of disease to obtain the second node feature of disease.

[0219] For example, as shown in FIG. 10, the hypergraph convolution required by the dynamic hypergraph of miRNA and the dynamic hypergraph of disease includes node convolution and hyperedge convolution. For example, for miRNA, through node convolution, the first node feature of miRNA can be aggregated into the hyperedge connected thereto, and then through hyperedge convolution, the adjacent hyperedge feature is aggregated back to the centroid node. For example, for disease, through node convolution, the first node feature of disease can be aggregated into the hyperedge connected thereto, and then through hyperedge convolution, the adjacent hyperedge feature is aggregated back to the centroid node.

[0220] For example, the process of node convolution operation of the dynamic hypergraph of miRNA can be implemented according to the following formula: T = MLP(X v ) x e = conv(T MLP(X v ))

[0221] In the above formula, X v represents the first node feature of miRNA, and the matrix T represents the transformation matrix learned by MLP in the first node feature X v of miRNA.

[0222] For example, in the operation process of node convolution, first, the first node feature Xv The MLP is used to learn the transformation matrix T to arrange and weight the features to realize the information flow between nodes and channels. Then, one-dimensional convolution is used to compress the transformed features to obtain the hyperedge features x e .

[0223] For example, the process of hyperedge convolution operation of the dynamic hypergraph of miRNA adopts an attention mechanism, which can be implemented according to the following formula: w = softmax(x e W + b)

[0224] In the above formula, ε v represents the hyperedge set containing the node x v ; W represents a weight matrix used to transform x e from the original feature space to another space; b represents a bias vector used to increase the flexibility of the model; and w represents the attention weight obtained by converting the unnormalized weight score of the hyperedge through the softmax function. x v represents the centroid node, that is, the feature of the centroid node is the weighted sum of the hyperedge features.

[0225] Similarly, the process of node convolution operation of the dynamic hypergraph of the disease can also be implemented according to the above formula. Through the hypergraph convolution module, the first node features of miRNA and the first node features of the disease are updated to obtain the second node features of miRNA and the second node features of the disease. For example, as shown in FIG. 10, the dynamic hypergraph construction module and the hypergraph convolution module can be stacked in one or more layers, which is not limited by the embodiments of the present disclosure.

[0226] For example, as shown in FIG. 1 and FIG. 10, for step S500, it can include: in the association prediction module, splicing the second node features of miRNA and the second node features of the disease to obtain the miRNA-disease pair; and performing at least one fully connected layer processing and softmax processing on the miRNA-disease pair to obtain the association prediction results of miRNA and the disease.

[0227] For example, as shown in FIG. 10, the second node features of miRNA can be spliced by using a splicing function, and the second node features of the disease can be spliced by using a splicing function, for example, the process can refer to the above-mentioned embodiments for the way of splicing the second node features of miRNA and splicing the second node features of the disease in FIG. 2, which is not limited by the embodiments of the present disclosure.

[0228] For example, as shown in FIG. 10, after obtaining the miRNA-disease pairs, the non-linear combination of features is further learned by inputting the miRNA-disease pairs into the fully connected layer, so as to capture the complex relationship between the miRNA and the disease. The output of the fully connected layer is transmitted to the softmax layer, and the softmax function can convert the output into a probability distribution, which represents the possibility of the association and non-association of the miRNA-disease pairs. For example, the result corresponding to the association of the miRNA and the disease pairs is 1, and the result corresponding to the non-association of the miRNA and the disease pairs is 0, and the output result of the softmax layer is taken as the prediction result of the prediction model.

[0229] FIG. 12A is a schematic diagram of the result of five-fold cross-validation corresponding to the prediction model shown in FIG. 10; and FIG. 12B is a schematic diagram of the result of five-fold cross-validation corresponding to the prediction method based on the hypergraph (HFHLMDA).

[0230] For example, in the five-fold cross-validation (5-CV) experiment, the known association matrix of the miRNA and the disease is taken as the input, and the prediction result of the association of each miRNA and disease pair is output by the model. In order to construct a balanced data set, 5430 known associations in the HMDD v2.0 human miRNA-disease association data set which have been verified by biological experiments are taken as the positive samples. All the known associations are divided into 5 groups, and each time one group is taken as the test sample set, and the remaining 4 groups are taken as the training sample set. This is carried out for five times, and each time a different group is taken as the test sample set, and the remaining 4 groups are taken as the training sample set. For the test positive samples, the same number of unknown associations are randomly selected from the unknown association pool as the test negative samples.

[0231] For example, the result obtained by using the prediction method described in the above embodiments for the prediction model shown in FIG. 10 is shown in FIG. 12A, and the result obtained by using the prediction method based on the hypergraph (HFHLMDA) is shown in FIG. 12B. According to FIG. 12A and FIG. 12B, the AUC result obtained by using the prediction method described in the above embodiments for the prediction model shown in FIG. 10 is higher than that obtained by using the prediction method based on the hypergraph (HFHLMDA), which indicates that the prediction model shown in FIG. 10 has good prediction ability.

[0232] According to the verification results of FIGS. 12A-12B, the miRNA-disease association prediction method (e.g., the prediction model corresponding to FIG. 10) provided by at least one embodiment of the present disclosure can adaptively and dynamically construct a hypergraph, and the hypergraph structure can be optimized simultaneously with the miRNA-disease association prediction task. The high-order relationships between miRNAs and the high-order relationships between diseases can be utilized in the operation process of the model, so that the dynamic hypergraph of miRNAs and the dynamic hypergraph of diseases can better adapt to the association prediction task, and the prediction model can fully explore the dynamic hypergraph structure of miRNAs and diseases, thereby better predicting the association between miRNAs and diseases and improving the accuracy of miRNA-disease association prediction.

[0233] FIG. 13 is a schematic block diagram of an miRNA-disease association prediction device provided by at least one embodiment of the present disclosure. As shown in FIG. 13, the miRNA-disease association prediction device 1000 includes an acquisition unit 1010, a similarity calculation unit 1020, a construction unit 1030, a hypergraph calculation unit 1040, and a prediction unit 1050. These units can be implemented by hardware (e.g., circuit) modules or software modules, etc. For example, these units can be implemented by a central processing unit (CPU), a graphics processing unit (GPU), a tensor processing unit (TPU), a field programmable logic gate array (FPGA), or other forms of processing units with data processing capability and / or instruction execution capability, and corresponding computer instructions, and embodiments of the present disclosure do not limit this.

[0234] As shown in FIG. 13, the acquisition unit 1010 is configured to acquire initial association data of miRNAs and diseases. The specific implementation can refer to the related description of step S100, which will not be repeated here.

[0235] The similarity calculation unit 1020 is configured to calculate the similarity of miRNAs and the similarity of diseases based on the initial association data. The specific implementation can refer to the related description of step S200, which will not be repeated here.

[0236] The construction unit 1030 is configured to construct a dynamic hypergraph based on the similarity of miRNAs and the similarity of diseases. The dynamic hypergraph includes first node features of miRNAs, first node features of diseases, and super edges connecting the first node features of miRNAs and super edges connecting the first node features of diseases, respectively. The specific implementation of the construction unit 1030 can refer to the related description of step S300, which will not be repeated here.

[0237] The supergraph computing unit 1040 is configured to obtain the second node features of the miRNAs and the second node features of the diseases based on the dynamic supergraph. For the specific implementation of the supergraph computing unit 1040, refer to the related description of step S400, which will not be repeated here.

[0238] The prediction unit 1050 is configured to obtain the prediction result of the association between the miRNAs and the diseases based on the second node features of the miRNAs and the second node features of the diseases. For the specific implementation of the prediction unit 1050, refer to the related description of step S500, which will not be repeated here.

[0239] It should be noted that in the embodiments of the present disclosure, the miRNA and disease association prediction device can include more or fewer circuits or units, and the connection relationship between the circuits or units is not limited, and can be determined according to actual needs. The specific implementation of each circuit is not limited, which can be constituted by an analog device according to the circuit principle, or can be constituted by a digital chip, or constituted in other applicable ways.

[0240] FIG. 14 is a schematic block diagram of another miRNA and disease association prediction device provided by at least one embodiment of the present disclosure. For example, as shown in FIG. 14, the miRNA and disease association prediction device 2000 includes a processor 2100, a memory 2200, and one or more computer program modules 2210.

[0241] For example, the processor 2100 and the memory 2200 are connected through a bus system 2300. For example, the one or more computer program modules 2210 are stored in the memory 2200. For example, the one or more computer program modules 2210 include instructions for executing the method for predicting the association between the miRNAs and the diseases provided by any embodiment of the present disclosure. For example, the instructions in the one or more computer program modules 2210 can be executed by the processor 2100. For example, the bus system 2300 can be a commonly used serial, parallel communication bus, etc., and the embodiments of the present disclosure are not limited thereto.

[0242] For example, the processor 2100 can be a central processing unit (CPU), a digital signal processor (DSP), a graphics processing unit (GPU), or other forms of processing units with data processing and / or instruction execution capabilities, which can be general-purpose processors or special-purpose processors, and can control other components in the data processing device 2000 to perform desired functions. For example, in the embodiments of the present disclosure, the processor is taken as an example of a graphics processing unit (GPU).

[0243] The memory 2200 can include one or more computer program products that can include various forms of computer-readable storage media, such as volatile memory and / or non-volatile memory. The volatile memory, for example, can include random access memory (RAM), cache memory, and / or the like. The non-volatile memory, for example, can include read-only memory (ROM), hard disk, flash memory, and / or the like. One or more computer program instructions can be stored on the computer-readable storage media, and the processor 2100 can execute the program instructions to implement the functions (implemented by the processor 2100) in the embodiments of the present disclosure and / or other desired functions, such as the miRNA and disease association prediction method, and the like. Various application programs and various data, such as initial association data, similarity, node features, and various data used and / or generated by the application programs, and the like, can also be stored in the computer-readable storage media.

[0244] It should be noted that, for the sake of clarity and simplicity, the present embodiments do not give all the component units of the miRNA and disease association prediction apparatus 2000. To achieve the necessary functions of the miRNA and disease association prediction apparatus 2000, those skilled in the art can provide and set other unshown component units according to specific needs, and the embodiments of the present disclosure do not limit this.

[0245] The miRNA and disease association prediction method or apparatus according to the embodiments of the present disclosure can also be implemented by means of the architecture of an exemplary electronic device 3000 as shown in FIG. 15. As shown in FIG. 15, the electronic device 3000 can include a bus 3010, one or more central processing units (CPUs) or graphics processing units (GPUs) 3020, read-only memory (ROM) 3030, random access memory (RAM) 3040, a communication port connected to a network 3050, input / output components 3060, a hard disk 3070, and the like. The storage devices in the electronic device 3000, such as the ROM 3030 or the hard disk 3070, can store various data or files required for processing and / or communication of the method provided by the present disclosure and program instructions executed by the CPU or GPU. The electronic device 3000 can also include a user interface 3080. Of course, the architecture shown in FIG. 15 is only exemplary, and when implementing different devices, one or more components in the electronic device shown in FIG. 15 can be omitted according to actual needs.

[0246] The miRNA and disease association prediction method provided in any embodiment of the present disclosure can be executed by a computer when the computer readable instructions are read by the computer.

[0247] For example, the storage medium can be any combination of one or more computer readable storage media, such as one computer readable storage medium containing computer readable program code for obtaining initial association data of miRNAs and diseases, another computer readable storage medium containing computer readable program code for calculating the similarity of miRNAs and the similarity of diseases based on the initial association data, another computer readable storage medium containing computer readable program code for constructing a dynamic hypergraph based on the similarity of miRNAs and the similarity of diseases, and causing the dynamic hypergraph to include first node features of miRNAs, first node features of diseases, and hyperedges connecting the first node features of miRNAs and hyperedges connecting the first node features of diseases, respectively, another computer readable storage medium containing computer readable program code for obtaining second node features of miRNAs and second node features of diseases based on the dynamic hypergraph, and another computer readable storage medium containing computer readable program code for obtaining an association prediction result of miRNAs and diseases based on the second node features of miRNAs and the second node features of diseases. For example, when the program code is read by the computer, the computer can execute the program code stored in the computer storage medium to execute, for example, the miRNA and disease association prediction method provided in any embodiment of the present disclosure.

[0248] For example, the storage medium can include a memory card of a smart phone, a storage component of a tablet computer, a hard disk of a personal computer, a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM), a portable compact disc read-only memory (CD-ROM), a flash memory, or any combination of the above storage media, and can also be other applicable storage media.

[0249] For the above disclosure, the following points need to be explained.

[0250] (1) The drawings of the embodiments of the present disclosure only relate to the structures involved in the embodiments of the present disclosure, and other structures can be referred to the general design.

[0251] (2) In the case of no conflict, the embodiments of the present disclosure and the features in the embodiments can be combined with each other to obtain new embodiments.

[0252] The above merely describes a specific implementation of the present disclosure, but the protection scope of the present disclosure is not limited thereto, and the protection scope of the present disclosure should be subject to the protection scope of the claims.

Claims

1. A method for predicting the association between miRNA and disease, comprising: Obtaining initial association data between the miRNA and the disease; Calculating the similarity of the miRNA and the similarity of the disease based on the initial association data; Based on the similarity of the miRNA and the similarity of the disease, constructing a dynamic hypergraph, wherein the dynamic hypergraph includes a first node feature of the miRNA, a first node feature of the disease, and a hyperedge connecting the first node feature of the miRNA and a hyperedge connecting the first node feature of the disease; Obtaining a second node feature of the miRNA and a second node feature of the disease based on the dynamic hypergraph; as well as A prediction result of the association between the miRNA and the disease is obtained based on the second node feature of the miRNA and the second node feature of the disease.

2. The method for predicting the association between miRNA and disease according to claim 1, wherein The calculating the similarity of the miRNA and the similarity of the disease based on the initial association data includes: Based on the initial association data, calculating the functional similarity MFSM of the miRNA and the Gaussian interaction spectrum kernel similarity MGSM of the miRNA, and calculating the semantic similarity DSSM of the disease and the Gaussian interaction spectrum kernel similarity DGSM of the disease; and The comprehensive similarity MISM of the miRNA is calculated based on the functional similarity MFSM of the miRNA and the Gaussian interaction spectrum kernel similarity MGSM of the miRNA, and the comprehensive similarity DISM of the disease is calculated based on the semantic similarity DSSM of the disease and the Gaussian interaction spectrum kernel similarity DGSM of the disease.

3. The method for predicting the association between miRNA and disease according to claim 2, wherein: The calculation of miRNA similarity and disease similarity also includes: Based on the functional similarity MFSM of the miRNA and the Gaussian interaction spectrum kernel similarity MGSM of the miRNA, an attention mechanism is used to obtain the fusion similarity MSM of the miRNA; Based on the semantic similarity DSSM of the disease and the Gaussian interaction spectrum kernel similarity DGSM of the disease, the attention mechanism is used to obtain the fused similarity DSM of the disease.

4. The method for predicting the association between miRNA and disease according to claim 3, wherein: The dynamic hypergraph is constructed based on the similarity of miRNAs and the similarity of diseases, including: Obtaining the initial node features of the miRNA and the initial node features of the disease based on the fusion similarity MSM of the miRNA and the fusion similarity DSM of the disease; The initial node feature of the miRNA and the initial node feature of the disease are updated to obtain the first node feature of the miRNA and the first node feature of the disease.

5. The method for predicting the association between miRNA and disease according to claim 4, wherein: The method of obtaining the initial node features of the miRNA and the initial node of the disease based on the fusion similarity MSM of the miRNA and the fusion similarity DSM of the disease includes: Using a multi-layer perceptron, the dimension of the miRNA fusion similarity MSM is adjusted to a first preset dimension to obtain the miRNA fusion similarity feature; Obtaining an initial node feature of the miRNA based on the fusion similarity feature of the miRNA and the random initialization feature of the miRNA; Using the multi-layer perceptron, the dimension of the fused similarity DSM of the disease is adjusted to a second preset dimension to obtain a fused similarity feature of the disease; and The initial node features of the disease are obtained based on the fused similarity features of the disease and the randomly initialized features of the disease.

6. The method for predicting the association between miRNA and disease according to claim 5, wherein: The updating of the initial node feature of the miRNA and the initial node feature of the disease to obtain the first node feature of the miRNA and the first node feature of the disease includes: Calculating the aggregated interaction feature vector of the miRNA based on the comprehensive similarity MISM of the miRNA; Calculating the aggregated interaction feature vector of the disease based on the comprehensive similarity DISM of the disease; Calculating an interaction probability matrix between the miRNA and the disease based on the aggregated interaction feature vector of the miRNA and the aggregated interaction feature vector of the disease; and Based on the interaction probability matrix between the miRNA and the disease and a preset negative connection value, a miRNA-disease bipartite graph is obtained, wherein the miRNA-disease bipartite graph includes initial node features of the miRNA and initial node features of the disease.

7. The method for predicting the association between miRNA and disease according to claim 6, wherein: Updating the initial node feature of the miRNA and the initial node feature of the disease to obtain the first node feature of the miRNA and the first node feature of the disease, including: Calculate the adjacency matrix A of the miRNA-disease bipartite graph G Sum degree matrix D G ;as well as Based on the adjacency matrix A G and the degree matrix D G , obtain the first node feature of the miRNA and the first node feature of the disease.

8. The method for predicting the association between miRNA and disease according to claim 7, wherein: The first node feature of the miRNA and the first node feature of the disease are obtained according to the following formula: Wherein, R represents the interaction probability matrix between the miRNA and the disease, A G represents the adjacency matrix of the miRNA-disease bipartite graph, E (0) represents the initial value of the miRNA-disease bipartite graph, E *,(l) represents the first node feature of the miRNA at the lth layer and the first node feature of the disease at the lth layer, represents the corresponding first node feature of the lth layer of the miRNA, represents the first node feature corresponding to the lth layer of the disease, and l is a positive integer greater than 1, D G represents the degree matrix, and D G is a diagonal matrix, D G [i,i] represents the adjacency matrix A G The number of non-zero entries in the i-th row of E; (l-1) equal or E (0) , represents the second node feature of the l-1th layer of the miRNA, The second node feature of the l-1th layer represents the disease.

9. The method for predicting the association between miRNA and disease according to claim 8, wherein: A dynamic hypergraph is constructed based on the similarity of miRNAs and diseases, including: Gumbel-Softmax is used to construct the dynamic hypergraph structure of the miRNA and the dynamic hypergraph structure of the disease, wherein the l-th layer hypergraph association matrix of the miRNA and the l-th layer hypergraph association matrix of the disease According to the following formulas: in, and And they are all learning matrices.

10. The method for predicting the association between miRNA and disease according to claim 9, wherein: Based on the dynamic hypergraph, the second node feature of the miRNA and the second node feature of the disease are obtained, including: The first layer hypergraph association matrix of the miRNA Performing a hypergraph convolution operation to obtain a second node feature of the miRNA; The hypergraph association matrix of the lth layer for the disease A hypergraph convolution operation is performed to obtain a second node feature of the disease.

11. The method for predicting the association between miRNA and disease according to claim 10, wherein: The obtaining of the second node feature of the miRNA and the second node feature of the disease based on the dynamic hypergraph includes: The second node feature of the miRNA and the second node feature of the disease are obtained by skip connection. Among them, the second node feature of the first layer of the miRNA and the second node characteristics of the disease at level l According to the following formula: in, Represents the l-th layer hypergraph association matrix of the miRNA The hyperedge degree matrix of Representing the l-th level hypergraph incidence matrix of the disease The hyperedge degree matrix of .

12. The method for predicting the association between miRNA and disease according to any one of claims 6 to 11, wherein: The obtaining of the association prediction result between the miRNA and the disease based on the second node feature of the miRNA and the second node feature of the disease includes: performing a splicing process on the second node features in all layers of the miRNA to obtain a third node feature of the miRNA; performing concatenation processing on the second node features in all layers of the disease to obtain a third node feature of the disease; Based on the third node feature of the miRNA and the third node feature of the disease, a miRNA-disease association matrix is ​​obtained as a prediction result of the association between the miRNA and the disease.

13. The method for predicting the association between miRNA and disease according to claim 12, wherein: The method further comprises: Based on the fusion similarity MSM of the miRNA, the fusion similarity DSM of the disease, and the interaction probability matrix between the miRNA and the disease, the node feature learning loss function L is obtained. sim ; Based on the third node feature of the miRNA and the third node feature of the disease, the association prediction loss function L is obtained. BPR ; Based on the third node feature of the miRNA and the third node feature of the disease, the node constraint loss L is obtained. const ;as well as Based on the node feature learning loss function L sim , the association prediction loss function L BPR And the node constraint loss L const , and obtain the total loss function of the miRNA and the disease.

14. The method for predicting the association between miRNA and disease according to claim 2, wherein: A dynamic hypergraph is constructed based on the similarity of miRNAs and diseases, including: Based on the integrated similarity MISM of the miRNA, each row in the integrated similarity MISM of the miRNA is used as an initial node feature of the miRNA to construct a priori hypergraph of the miRNA; and Based on the comprehensive similarity DISM of the disease, each row in the comprehensive similarity DISM of the disease is used as the initial node feature of the disease to construct a priori hypergraph of the disease.

15. The method for predicting the association between miRNA and disease according to claim 14, wherein: A dynamic hypergraph is constructed based on the similarity of miRNAs and diseases, including: Using HyperGAT to update the initial node features of the miRNA to obtain the first node features of the miRNA, and using Gumbel-Topk and KMeans to process the first node features of the miRNA to construct a dynamic hypergraph of the miRNA; The initial node features of the disease are updated using the HyperGAT to obtain the first node features of the disease, and the first node features of the disease are processed using the Gumbel-Topk and the KMeans to construct a dynamic hypergraph of the disease.

16. The method for predicting the association between miRNA and disease according to claim 15, wherein: Based on the dynamic hypergraph, the second node feature of the miRNA and the second node feature of the disease are obtained, including: Performing a node convolution operation and a hyperedge convolution operation on the dynamic hypergraph of the miRNA to obtain a second node feature of the miRNA; and A node convolution operation and a hyperedge convolution operation are performed on the dynamic hypergraph of the disease to obtain a second node feature of the disease.

17. The method for predicting the association between miRNA and disease according to claim 16, wherein: Obtaining a prediction result of the association between the miRNA and the disease based on the second node feature of the miRNA and the second node feature of the disease, including: performing splicing processing on the second node feature of the miRNA and the second node feature of the disease to obtain a miRNA-disease pair; The miRNA-disease pair is processed with at least one fully connected layer and a softmax process to obtain an association prediction result between the miRNA and the disease.

18. A device for predicting the association between miRNA and disease, comprising: an acquisition unit, configured to acquire initial association data between miRNAs and diseases; a similarity calculation unit configured to calculate the similarity of the miRNA and the similarity of the disease based on the initial association data; A construction unit is configured to construct a dynamic hypergraph based on the similarity of the miRNA and the similarity of the disease, wherein the dynamic hypergraph includes a first node feature of the miRNA, a first node feature of the disease, and a hyperedge connecting the first node feature of the miRNA and a hyperedge connecting the first node feature of the disease; a hypergraph computing unit configured to obtain a second node feature of the miRNA and a second node feature of the disease based on the dynamic hypergraph; and The prediction unit is configured to obtain a prediction result of the association between the miRNA and the disease based on the second node feature of the miRNA and the second node feature of the disease.

19. An electronic device comprising: processor; Memory; One or more computer program modules, wherein the one or more computer program modules are stored in the memory and configured to be executed by the processor, and the one or more computer program modules include instructions for executing the method for predicting the association between miRNA and disease according to any one of claims 1 to 17.

20. A storage medium non-transitorily storing computer-readable instructions, which, when executed by a computer, can execute the method for predicting the association between miRNA and disease according to any one of claims 1 to 17.

Images