Method and apparatus for predicting association relationship between mirna and diseases, and storage medium

Abstract

Provided are a method and apparatus for predicting an association relationship between miRNA and diseases, and a storage medium. The method comprises: acquiring a plurality of known associated positive samples and a plurality of unknown associated samples, and selecting from among the plurality of unknown associated samples some samples as first initial negative samples (101); and using the positive samples and second negative samples to iteratively train a hypergraph model and, after each iterative training, using the current hypergraph model to predict the unknown associated samples, the second negative samples used in the first iterative training being the first initial negative samples, and the second negative samples used in the remaining iterative training being some samples of the unknown associated samples predicted after the current iterative training (102).

Description

Methods, devices, and storage media for predicting the association between miRNAs and diseases. Technical Field

[0001] This disclosure relates to, but is not limited to, the field of biotechnology, and in particular to a method, apparatus, and storage medium for predicting the association between miRNA and disease. Background Technology

[0002] miRNAs are a class of endogenous small RNAs, approximately 20 to 24 nucleotides in length, that play a variety of important regulatory roles within cells. Increasing evidence suggests that miRNA regulation is closely related to a range of diseases. Identifying meaningful associations between miRNAs and diseases will contribute to understanding the molecular mechanisms underlying these diseases.

[0003] However, due to high costs and specialized requirements, traditional biological experiments can only infer a limited number of associations between miRNAs and diseases. Supervised learning-based methods require experimentally validated negative samples of unrelated miRNA-disease pairs, but these negative samples are often difficult to obtain due to a lack of interest in biomedical research. Related methods typically randomly select negative samples from unlabeled samples; this selection strategy may contain noise, leading to reduced reliability of predictive performance.

[0004] Summary of the Invention

[0005] The following is a summary of the subject matter described in detail herein. This summary is not intended to limit the scope of the claims.

[0006] This disclosure provides a method for predicting the association between miRNAs and diseases, including:

[0007] Obtain multiple positive samples with known associations and multiple samples with unknown associations, and select a portion of the samples with unknown associations as the first initial negative samples;

[0008] The hypergraph model is trained iteratively using the positive samples and the second negative samples. After each iteration of training, the current hypergraph model is used to predict the unknown associated samples. The second negative sample used in the first iteration of training is the first initial negative sample. The second negative sample used in the remaining iterations of training is a portion of the unknown associated samples predicted after the current iteration of training.

[0009] This disclosure also provides an apparatus for predicting the association between miRNA and disease, including a memory; and a processor connected to the memory, the memory being used to store instructions, the processor being configured to execute the steps of the method for predicting the association between miRNA and disease according to any embodiment of this disclosure based on the instructions stored in the memory.

[0010] This disclosure also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the method for predicting the association between miRNA and disease as described in any embodiment of this disclosure.

[0011] This disclosure also provides a program product including instructions that, when executed by a computer, perform a method for predicting miRNA-disease association as described in any embodiment of this disclosure.

[0012] This disclosure also provides a device for predicting the association between miRNA and disease, comprising: a data acquisition module and a first training module, wherein:

[0013] The data acquisition module is configured to acquire multiple known-associated positive samples and multiple unknown-associated samples, and select a portion of the samples from the multiple unknown-associated samples as the first initial negative samples.

[0014] The first training module is configured to iteratively train the hypergraph model using the positive samples and the second negative samples, and after each iteration of training, to predict the unknown associated samples using the current hypergraph model. The second negative sample used in the first iteration of training is the first initial negative sample, and the second negative sample used in the subsequent iterations of training is a portion of the unknown associated samples predicted after the current iteration of training.

[0015] After reading and understanding the accompanying diagrams and detailed descriptions, other aspects can be understood.

[0016] Summary of the Figures

[0017] The accompanying drawings are provided to further illustrate the technical solutions of this disclosure and form part of the specification. They are used together with the embodiments of this disclosure to explain the technical solutions of this disclosure and do not constitute a limitation on the technical solutions of this disclosure. The shapes and sizes of the components in the drawings do not reflect actual proportions and are only intended to illustrate the content of this disclosure.

[0018] Figure 1 is a flowchart illustrating a method for predicting the association between miRNA and disease according to an exemplary embodiment of this disclosure;

[0019] Figure 2 is a flowchart illustrating another method for predicting the association between miRNA and disease provided by an exemplary embodiment of this disclosure;

[0020] Figure 3 is a schematic diagram of a process for updating a first initial negative sample through a scoring model, provided by an exemplary embodiment of this disclosure;

[0021] Figure 4 is a schematic diagram of the iterative training process of a hypergraph model provided by an exemplary embodiment of this disclosure;

[0022] Figure 5 is a schematic diagram of the structure of a device for predicting the association between miRNA and disease provided in an exemplary embodiment of this disclosure;

[0023] Figure 6 is a schematic diagram of another miRNA-disease association prediction device provided by an exemplary embodiment of this disclosure.

[0024] Detailed Explanation

[0025] To make the objectives, technical solutions, and advantages of this disclosure clearer, the embodiments of this disclosure will be described in detail below with reference to the accompanying drawings. It should be noted that, unless otherwise specified, the embodiments and features described in this disclosure can be arbitrarily combined with each other.

[0026] Unless otherwise defined, the technical or scientific terms used in the embodiments of this disclosure shall have the ordinary meaning understood by one of ordinary skill in the art to which this disclosure pertains. The terms "first," "second," and similar terms used in the embodiments of this disclosure do not indicate any order, quantity, or importance, but are merely used to distinguish different components. Terms such as "comprising" or "including" indicate that the element or object preceding the word encompasses the elements or objects listed following the word and their equivalents, but do not exclude other elements or objects.

[0027] As shown in Figure 1, this disclosure provides a method for predicting the association between miRNAs and diseases, including:

[0028] Step 101: Obtain multiple positive samples with known associations and multiple samples with unknown associations, and select a portion of the samples with unknown associations as the first initial negative samples;

[0029] Step 102: Iteratively train the hypergraph model using positive samples and the second negative sample. After each iteration of training, use the current hypergraph model to predict samples with unknown associations. The second negative sample used in the first iteration of training is the first initial negative sample. The second negative sample used in the remaining iterations of training is a portion of the samples with unknown associations predicted after the current iteration of training.

[0030] The method for predicting the association between miRNAs and diseases provided in this disclosure uses a hypergraph model that is trained iteratively based on positive and second negative samples. The second negative sample used in the first iteration of training is the first initial negative sample, and the second negative samples used in the subsequent iterations of training are a portion of the samples with unknown associations predicted after the current iteration of training. This method can predict the association between a large number of miRNAs and diseases. Furthermore, since the negative samples generated by the scoring model are relatively reliable, the model prediction performance has high reliability.

[0031] In some exemplary embodiments, prior to the step of iteratively training the hypergraph model using positive samples and a second negative sample, the method further includes:

[0032] The scoring model is trained using positive samples and the first initial negative sample. The trained scoring model is then used to score multiple samples with unknown associations. Finally, a subset of the scored samples with unknown associations is selected to update the first initial negative sample.

[0033] As shown in Figure 2, the prediction method provided in this embodiment can be roughly divided into two stages. In the first stage, positive samples and randomly sampled initial negative samples are obtained. A scoring model is trained using the positive samples and the first initial negative samples. The trained scoring model scores multiple samples with unknown associations. The first initial negative samples are updated based on the scoring results. The first initial negative samples generated at this time are relatively reliable (the first initial negative samples generated by this method are used to train a better initial hypergraph model). In the second stage, the generated first initial negative samples are used as initial second negative samples. The hypergraph model is iteratively trained based on the positive samples and the second negative samples. Each iteration includes: generating an initial feature vector based on the positive samples and / or the second negative samples; encoding the initial feature vector through a second dual-tower model with shared parameters to obtain the local dependency relationship between miRNA and disease; dynamically constructing an attention hypergraph based on distance; performing hypergraph convolution on the attention hypergraph and combining it with the low-dimensional embedding representation output by the second dual-tower model to obtain the high-order dependency relationship between miRNA and disease; using the hypergraph model to predict samples with unknown associations; and updating the second negative samples with the predicted samples.

[0034] Traditional correlation prediction methods typically consider only limited biological associations, rely on low-order interaction information from conventional graphs, and often use static graph structures, resulting in limited model prediction performance. In contrast, the prediction method of this disclosure dynamically constructs a hypergraph, performs attention hypergraph convolution, and combines it with the low-dimensional embedding representation of a second dual-tower model to obtain high-order dependencies between miRNAs and diseases, thus better identifying potential miRNA-disease associations.

[0035] For example, the benchmark dataset used in this disclosure is human miRNA-disease association data obtained from the HMDD database. This benchmark dataset contains 495 miRNAs and 383 diseases, of which 5430 miRNAs have been experimentally verified to be associated with diseases. This disclosure uses an adjacency matrix. To describe the association between miRNAs and disease pairs, where N m =495, N d =383. If miRNAm i With disease d j If there is a correlation, then A(m)i ,d j ) = 1, otherwise A(m) i ,d j = 0. Therefore, the adjacency matrix A has 5430 elements assigned a value of 1, and the remaining elements are assigned a value of 0. In step 101, 5430 experimentally verified miRNA-disease pairs can be used as positive samples.

[0036] In some exemplary embodiments, in step 101, selecting a subset of samples from a plurality of unknown-related samples as first initial negative samples includes:

[0037] From a number of samples with unknown associations, k samples are randomly selected as the first initial negative samples, and the size of k is the same as the number of positive samples with known associations.

[0038] For example, in this embodiment of the disclosure, 5430 miRNA-disease pairs are randomly sampled from 184,155 unknown association miRNA-disease pairs as the first initial negative sample. In this embodiment of the disclosure, unknown association means that it is unknown whether there is an association between the miRNA and the disease; correspondingly, known association means that it is known that there is an association between the miRNA and the disease.

[0039] In some exemplary embodiments, the method further includes, prior to:

[0040] Initial feature vectors for multiple miRNAs and initial feature vectors for various diseases are generated.

[0041] In some exemplary embodiments, initial feature vectors for multiple miRNAs and initial feature vectors for multiple diseases are generated, including:

[0042] Establish miRNA functional similarity matrix (FM) and disease semantic similarity matrix (DM);

[0043] Calculate the Gaussian interaction spectrum kernel similarity matrix KM of miRNAs and the Gaussian interaction spectrum kernel similarity matrix KD of diseases;

[0044] The miRNA functional similarity matrix FM and the miRNA Gaussian interaction spectrum kernel similarity matrix KM are integrated into the miRNA similarity representation matrix IM. Each row vector of the miRNA similarity representation matrix IM is the initial feature vector of a miRNA.

[0045] The disease semantic similarity matrix DM and the disease Gaussian interaction spectrum kernel similarity matrix KD are integrated into a disease similarity representation matrix ID. Each row vector of the disease similarity representation matrix ID is an initial feature vector of a disease.

[0046] Generally, if the target gene sets of two miRNAs overlap, the two miRNAs are considered to have functional similarity. Based on functional similarity, miRNAs often act on similar diseases. Some researchers have designed the MISIM method to calculate the functional similarity of miRNAs. Therefore, this embodiment uses the MISIM method to establish a 495×495 miRNA functional similarity matrix FM, where FM(m i ,m j The larger ) is, the more m i and m j The more similar they are.

[0047] In this embodiment, a directed acyclic graph (DAG) constructed from biological terms can be obtained from the MeSH database. Semantic similarity between diseases is then calculated within the DAG; terms that are further apart in the DAG have lower semantic similarity. The similarity between two diseases can be calculated based on their semantic contributions to their ancestor nodes. After calculating the semantic similarity between all diseases, a 383×383 disease semantic similarity matrix (DM) is constructed.

[0048] Due to the sparsity problem of the miRNA functional similarity matrix FM and the disease semantic similarity matrix DM, sparse matrices are unlikely to provide more effective information, which will affect the prediction performance. Therefore, based on the validated miRNA-disease association matrix in the training data, the Gaussian interaction spectrum kernel similarity matrix KM of miRNAs and the Gaussian interaction spectrum kernel similarity matrix KD of diseases are calculated. Wherein, KM(m i ,m j ) indicates miRNAm i and miRNAm j The Gaussian interaction spectrum kernel similarity is a value that indicates the greater similarity two miRNAs are likely to have if they are associated with more diseases. Similarly, KD(d i ,d j ) indicates disease d i and disease d j The Gaussian interaction spectrum kernel similarity value indicates that if two diseases have more miRNA associations, they may have greater similarity.

[0049] The miRNA functional similarity matrix FM and the miRNA Gaussian interaction spectrum kernel similarity matrix KM are integrated into the final miRNA similarity representation matrix IM; the disease semantic similarity matrix DM and the disease Gaussian interaction spectrum kernel similarity matrix KD are integrated into the disease similarity representation matrix ID, thereby solving the sparsity problem of the miRNA functional similarity matrix FM and the disease semantic similarity matrix DM. In this embodiment, each miRNA is represented as a 495-dimensional initial feature vector IM(mi )=(m1,m2,…,m 495 ),IM(m i ) is the i-th row vector of IM, representing miRNAm i Similarity with all miRNAs. Similarly, each disease is represented as a 383-dimensional initial feature vector ID(d i )=(d1,d2,…,d 383 ), ID(d i ) is the i-th row vector of ID, representing disease d. i Similarity to all diseases.

[0050] In some exemplary embodiments, the miRNA functional similarity matrix FM and the miRNA Gaussian interaction spectrum kernel similarity matrix KM are integrated into a miRNA similarity representation matrix IM according to the following formula:

[0051] The disease semantic similarity matrix DM and the disease Gaussian interaction spectrum kernel similarity matrix KD are integrated into the disease similarity representation matrix ID according to the following formula:

[0052] However, this disclosure is not limited in this respect. In other examples, other formulas can be used to integrate the miRNA similarity representation matrix IM and the disease similarity representation matrix ID, and the initial feature vectors for each miRNA and each disease can be obtained based on the integrated miRNA similarity representation matrix IM and the disease similarity representation matrix ID.

[0053] In some exemplary implementations, the scoring model is any one of the following: a first dual-tower model, a convolutional neural network (CNN), a recurrent neural network (RNN), and a long short-term memory network (LSTM), etc.

[0054] In some exemplary embodiments, the scoring model includes a first dual-tower model; the hypergraph model includes a second dual-tower model, the input of the first dual-tower model includes an initial feature vector of miRNA and an initial feature vector of disease, and the input of the second dual-tower model includes an initial feature vector of miRNA and an initial feature vector of disease.

[0055] The dual-tower model is a deep learning model used to measure the similarity or distance between two input samples. It consists of two identical neural networks, each encoding one input sample and outputting its feature representation. These feature representations can be used to calculate the similarity or distance between the two samples. The two neural networks have the same network structure and parameters, but they are independent.

[0056] In some exemplary embodiments, the first dual-tower model includes: a first encoding layer, a second encoding layer, a similarity calculation layer, and a training optimization layer;

[0057] The first and second coding layers are set in parallel and share parameters. The first coding layer encodes the initial feature vector of the miRNA into the feature representation of the miRNA, and the second coding layer encodes the initial feature vector of the disease into the feature representation of the disease.

[0058] The similarity calculation layer calculates the similarity between the feature representations of miRNAs and the feature representations of diseases;

[0059] The training optimization layer calculates the loss function during the training process and updates the model parameters to gradually optimize the model.

[0060] In some exemplary embodiments, the first coding layer and the second coding layer each include three fully connected layers. In some examples, as shown in FIG3, the output dimension of the first fully connected layer is 128, the output dimension of the second fully connected layer is 128, and the output dimension of the third fully connected layer is 100. However, this disclosure does not limit this.

[0061] In some exemplary embodiments, the similarity calculation layer can calculate the similarity between the feature representation of miRNA and the feature representation of disease using Euclidean distance, cosine similarity, and Manhattan distance. For example, in this embodiment of the disclosure, cosine similarity can be used to calculate the similarity between the feature representation of miRNA and the feature representation of disease.

[0062] In some exemplary implementations, the training optimization layer can use optimization algorithms such as Stochastic Gradient Descent (SGD), Adaptive Moment Estimation (Adam), and RMSprop (Root Mean Square prop) to update the model parameters.

[0063] In some exemplary implementations, the training optimization layer uses a contrastive loss function to progressively optimize the model. The goal of the contrastive loss function is to minimize the distance between samples of the same class and maximize the distance between samples of different classes.

[0064] As shown in Figure 3, the network structures of the first and second encoding layers on both sides of the first dual-tower model are identical. The network input consists of initial feature vectors of miRNA and disease. The encoding layers are three fully connected layers that encode the feature representations of miRNA and disease. Then, a cosine metric function is used to calculate the similarity between the feature representations of miRNA and disease. The scoring model uses ReLU as the activation function, with a dropout value of 0.5. The network is trained using the Adam optimizer, and the loss function is a contrastive loss function.

[0065] After training the scoring model, it can be used to score multiple known-association positive samples and unknown-association samples. Based on the scoring results of multiple positive samples, some samples from the scored unknown-association samples are selected to update the first initial negative sample.

[0066] In some exemplary implementations, selecting a subset of samples from the scored, unassociated samples to update the first initial negative samples includes:

[0067] The trained scoring model is used to score multiple known associated positive samples, and the mean score of the multiple known associated positive samples is calculated as threshold_up.

[0068] Sort the samples with unknown associations in descending order of their scores, and take the first k samples with scores less than or equal to (threshold_up-a) to update the first initial negative samples, where a is between 0.05 and 0.2, and k is the same as the number of positive samples with known associations.

[0069] For example, 'a' can be 0.1. However, this disclosure does not limit this, and the size of 'a' can be set as needed.

[0070] For example, the scoring model trained above is used to score all samples with known and unknown associations. The output results of the scoring model are shown in Table 1 below:

[0071] Table 1

[0072] The mean score of all known associated samples is calculated and denoted as threshold_up. The scores of unknown associated samples are sorted in descending order of predicted scores. The first k samples with score < (threshold_up - 0.1) are taken to update the first initial negative sample. In this embodiment, k = 5430.

[0073] In other exemplary embodiments, selecting a subset of samples from the scored, unassociated samples to update the first initial negative sample includes:

[0074] Multiple samples with unknown associations are sorted in descending order of their scores. The first k samples with unknown associations whose scores are less than or equal to a preset score threshold are selected to update the first initial negative samples. The preset score threshold is between 0.55 and 0.75, and the value of k is the same as the number of positive samples with known associations.

[0075] For example, sort multiple samples with unknown associations in descending order of their scores, and select the first k samples with unknown associations whose scores are less than or equal to 0.6 to update the first initial negative sample.

[0076] In other exemplary embodiments, samples with scores within a preset threshold range can be directly selected to update the first initial negative sample. For example, k samples with unknown associations whose scores are between 0.6 and 0.7 can be selected to update the first initial negative sample.

[0077] As shown in Figure 4, in the second stage, the 5430 miRNA-disease pairs that were experimentally verified to be related were still used as positive samples, and the first initial negative samples updated in the first stage were used as the initial second negative samples. Initial feature vectors IZ of the miRNAs were generated based on the positive and second negative samples, respectively. m =(m1,m2,...,m 495 The initial feature vector IZ of the disease d =(d1,d2,...,d 383 ).

[0078] In some exemplary embodiments, the second dual-tower model includes: a third coding layer and a fourth coding layer;

[0079] The third and fourth coding layers are set in parallel and share parameters. The third coding layer encodes the initial feature vector of the miRNA into the feature representation of the miRNA, and the fourth coding layer encodes the initial feature vector of the disease into the feature representation of the disease.

[0080] This disclosure obtains the local dependency relationship between miRNA and disease through a second dual-tower model with shared parameters. The network structure of the coding layer of the second dual-tower model can be the same as that of the coding layer of the first dual-tower model (trained separately, therefore, the model parameters may be different). In other exemplary embodiments, the network structure of the coding layer of the second dual-tower model can also be different from that of the first dual-tower model. The low-dimensional representation Z of the miRNA is obtained after encoding by the second dual-tower model. m Low-dimensional representation of disease Z d .

[0081] In some exemplary embodiments, the hypergraph model further includes: a hypergraph construction module and a hypergraph convolution module, wherein:

[0082] The second dual-tower model transforms the initial feature vector of miRNA into a low-dimensional representation of miRNA, and the initial feature vector of disease into a low-dimensional representation of disease.

[0083] The hypergraph construction module constructs a miRNA hypergraph using miRNA as nodes and low-dimensional representation of miRNA as a feature, and constructs a disease hypergraph using epidemic as nodes and low-dimensional representation of disease as a feature;

[0084] The hypergraph convolution module performs hypergraph convolution on the miRNA hypergraph and the disease hypergraph respectively to obtain the embedding representations of miRNA and disease.

[0085] In this embodiment, the hypergraph construction module constructs two hypergraphs: a miRNA hypergraph and a disease hypergraph. The miRNA hypergraph uses miRNAs as nodes, with the low-dimensional representation of miRNAs being Zi. m Using cosine similarity as a feature, the K-Nearest Neighbors (KNN) algorithm is used to generate hyperedges. Specifically, for each vertex, its K nearest neighbors are searched to form a hyperedge. Here, K is a hyperparameter. Similarly, the disease hypergraph uses diseases as nodes, with the disease's low-dimensional representation Z... d Using the features, the KNN algorithm based on cosine similarity is used to generate hyperedges.

[0086] In some exemplary embodiments, the hypergraph construction module also calculates the miRNA hypergraph association matrix and the disease hypergraph association matrix through attention scores; the miRNA embedding representation combines the low-dimensional representation information of miRNA output by the second twin-tower model with the hyperedge information obtained by convolving the miRNA hypergraph association matrix, and the disease embedding representation combines the low-dimensional representation information of disease output by the second twin-tower model with the hyperedge information obtained by convolving the disease hypergraph association matrix.

[0087] This disclosure utilizes attention scores to enrich the representation of the association matrix. Taking a miRNA hypergraph as an example, for vertex v... i and super edge v j Let vertex v i The low-dimensional embedding is represented as Vertex v j The low-dimensional embedding is represented as miRNA hypergraph association matrix H m Obtained by calculating the attention score:

[0088] Where α(.) is a nonlinear activation function, N i For node v i Let W be the set of K nearest neighbors, W be the learnable weight matrix, w be the learnable weight vector, and sim(.) be the function to calculate the similarity between vertices.

[0089] Similarly, the disease hypergraph correlation matrix H can be obtained using the above formula. d .

[0090] In this embodiment of the disclosure, the hypergraph convolution module follows the hypergraph messaging paradigm to construct hypergraph convolutional layers that capture global high-order dependencies on the hypergraph. The hypergraph convolution executes a vertex-hyperedge-vertex pattern, where vertex convolution is used to aggregate vertex features onto the hyperedges containing these vertices, and hyperedge convolution aggregates hyperedge features onto the centroid vertex.

[0091] miRNA hypermap H m For example, hypergraph convolution is learned using the following formula:

[0092] Where δ is the LeakyReLU activation function, D e Let D be the degree of the hyperedge. v W is the degree of the vertex. l Z is a learnable weight matrix. l Let l be the embedding representation of the miRNA output from the coding layer of the dual-tower model, where l is the number of convolutional layers, and l = 2.

[0093] In layer l+1, the miRNA node embedding representation combines information from both the dual-tower model and hypergraph convolution, as detailed below:

[0094] Similarly, the above process is used to obtain an updated embedding representation of the disease.

[0095] This disclosure trains a hypergraph model using an iterative training method. First, positive samples and an initial second negative sample (i.e., the first initial negative sample generated based on the aforementioned scoring model) are used as the training set to train the hypergraph model. Then, the hypergraph model trained in this iteration is used to predict data with unknown labels (i.e., the aforementioned multiple samples with unknown associations) to generate pseudo-labels. The generated pseudo-label data is then used to update the second negative sample in the training set, and the updated training set is used to iteratively train the hypergraph model again. During the training process, this disclosure uses the same hypergraph model for multiple iterative training iterations, as shown in Figure 4. Through multiple iterative training iterations, this disclosure allows the generated pseudo-labels to be more realistic, thereby improving the model's predictive performance.

[0096] In this embodiment, the structure of the miRNA hypergraph and the disease hypergraph dynamically change during different training iterations. Since the second negative sample is updated after each training iteration using a subset of samples from the unknown associations predicted by the hypergraph model, meaning the second negative sample changes with each iteration, the structures of the miRNA and disease hypergraphs constructed by the hypergraph construction module also dynamically change during each training iteration, significantly improving the model's predictive performance.

[0097] To optimize the model, this disclosure uses the cross-entropy loss function to calculate the loss when training the hypergraph model:

[0098] Where U is the training dataset, y′ ij =Z m .Z d Z m and Z d y represents the final embedding representation of miRNA and disease after all convolutional layers are executed, respectively, with . representing the dot product symbol. ij This is a truth label; if the corresponding miRNA and the disease are related, then y... ij If the two are unrelated, then y is 1. ij It is 0.

[0099] In this embodiment, a threshold number of training iterations or a threshold loss can be preset to set the termination condition for iterative training. When a threshold number of training iterations is set, iterative training of the hypergraph model stops when the threshold number of training iterations is reached. When a threshold loss is set, iterative training of the hypergraph model stops when the calculated loss is lower than the threshold loss once or multiple times. Based on the prediction results of the last trained hypergraph model for all unknown-label data, the association between different miRNAs and different diseases can be determined.

[0100] In some exemplary implementations, the association between miRNAs and diseases is updated based on the prediction results of the hypergraph model generated in the last iteration of training for samples with unknown associations.

[0101] The prediction method disclosed herein can predict the associations between a large number of miRNAs and diseases, and the model's prediction performance is highly reliable because the negative samples generated by the scoring model are relatively reliable. Furthermore, this disclosure obtains higher-order dependencies between miRNAs and diseases by dynamically constructing an attention hypergraph and performing hypergraph convolution on it, further improving the model's prediction performance.

[0102] As shown in Figure 5, this embodiment of the present disclosure also provides a device for predicting the association between miRNA and disease, including a data acquisition module 510 and a first training module 520, wherein:

[0103] The data acquisition module 510 is configured to acquire multiple known-associated positive samples and multiple unknown-associated samples, and select a portion of the samples from the multiple unknown-associated samples as first initial negative samples.

[0104] The first training module 520 is configured to iteratively train the hypergraph model using the positive samples and the second negative samples. After each iteration of training, the current hypergraph model is used to predict the samples of the unknown association. The second negative sample used in the first iteration of training is the first initial negative sample. The second negative sample used in the remaining iterations of training is a portion of the samples of the unknown association predicted after the current iteration of training.

[0105] In some exemplary embodiments, the prediction device further includes a second training module configured to train a scoring model using the positive samples and the first initial negative samples, score a plurality of the unknown associated samples using the trained scoring model, and select a subset of samples from the scored unknown associated samples to update the first initial negative samples.

[0106] In some exemplary embodiments, the prediction device further includes an initial feature generation module, wherein the initial feature generation module is configured to generate initial feature vectors for multiple miRNAs and initial feature vectors for multiple diseases.

[0107] In some exemplary embodiments, the scoring model is any one of the following: a first dual-tower model, a convolutional neural network, a recurrent neural network, and a long short-term memory network.

[0108] In some exemplary embodiments, the scoring model includes a first dual-tower model; the hypergraph model includes a second dual-tower model, wherein the input of the first dual-tower model includes an initial feature vector of miRNA and an initial feature vector of disease, and the input of the second dual-tower model includes an initial feature vector of miRNA and an initial feature vector of disease.

[0109] In some exemplary embodiments, the initial feature generation module is specifically configured as follows:

[0110] Establish miRNA functional similarity matrix (FM) and disease semantic similarity matrix (DM);

[0111] Calculate the Gaussian interaction spectrum kernel similarity matrix KM of miRNAs and the Gaussian interaction spectrum kernel similarity matrix KD of diseases;

[0112] The miRNA functional similarity matrix FM and the miRNA Gaussian interaction spectrum kernel similarity matrix KM are integrated into a miRNA similarity representation matrix IM, where each row vector of the miRNA similarity representation matrix IM is an initial feature vector of a miRNA.

[0113] The disease semantic similarity matrix DM and the disease Gaussian interaction spectrum kernel similarity matrix KD are integrated into a disease similarity representation matrix ID, where each row vector of the disease similarity representation matrix ID is an initial feature vector of a disease.

[0114] In some exemplary embodiments, the initial feature generation module integrates the miRNA functional similarity matrix FM and the miRNA Gaussian interaction spectrum kernel similarity matrix KM into a miRNA similarity representation matrix IM according to the following formula:

[0115] The disease semantic similarity matrix DM and the disease Gaussian interaction spectrum kernel similarity matrix KD are integrated into a disease similarity representation matrix ID according to the following formula:

[0116] In some exemplary embodiments, the first dual-tower model includes: a first encoding layer, a second encoding layer, a similarity calculation layer, and a training optimization layer;

[0117] The first coding layer and the second coding layer are set in parallel and share parameters. The first coding layer encodes the initial feature vector of the miRNA into the feature representation of the miRNA, and the second coding layer encodes the initial feature vector of the disease into the feature representation of the disease.

[0118] The similarity calculation layer calculates the similarity between the feature representation of miRNA and the feature representation of disease;

[0119] The training optimization layer calculates the loss function during the training process and updates the model parameters to gradually optimize the model.

[0120] In some exemplary embodiments, the second dual-tower model includes: a third coding layer and a fourth coding layer;

[0121] The third and fourth coding layers are configured in parallel and share parameters. The third coding layer encodes the initial feature vector of the miRNA into a feature representation of the miRNA, and the fourth coding layer encodes the initial feature vector of the disease into a feature representation of the disease.

[0122] In some exemplary embodiments, the hypergraph model further includes: a hypergraph construction module and a hypergraph convolution module;

[0123] The second dual-tower model transforms the initial feature vector of the miRNA into a feature representation of the miRNA, and transforms the initial feature vector of the disease into a feature representation of the disease;

[0124] The hypergraph construction module constructs a miRNA hypergraph using miRNA as nodes and miRNA feature representations as features, and constructs a disease hypergraph using epidemics as nodes and disease feature representations as features.

[0125] The hypergraph convolution module performs hypergraph convolution on the miRNA hypergraph and the disease hypergraph respectively to obtain the embedding representation of miRNA and the embedding representation of disease.

[0126] In some exemplary embodiments, the loss function of the hypergraph model is:

[0127] Where, y′ ij =Z m .Z d Z m and Z d These represent the embedding representations of miRNAs and the disease, respectively, obtained after performing hypergraph convolution; y ij As the truth label, y ij =1 or 0, where U is the training data set.

[0128] In some exemplary embodiments, the hypergraph construction module further calculates the miRNA hypergraph association matrix and the disease hypergraph association matrix through attention scores; the miRNA embedding representation combines the feature representation information of the miRNA output by the second dual-tower model with the hyperedge information obtained by convolving the miRNA hypergraph association matrix, and the disease embedding representation combines the feature representation information of the disease output by the second dual-tower model with the hyperedge information obtained by convolving the disease hypergraph association matrix.

[0129] In some exemplary embodiments, the second training module selects a subset of samples from the scored, unassociated samples to update the first initial negative samples, including:

[0130] The trained scoring model is used to score multiple known associated positive samples, and the mean score of the multiple known associated positive samples is calculated as threshold_up.

[0131] The samples with unknown associations are sorted in descending order of their scores. The first initial negative samples are updated by taking the first k samples with scores less than or equal to (threshold_up-a), where a is between 0.05 and 0.2, and k is the same as the number of positive samples with known associations.

[0132] In some exemplary embodiments, the second training module selects a subset of samples from the scored, unassociated samples to update the first initial negative samples, including:

[0133] The first initial negative sample is updated by selecting k samples with scores within a preset score threshold range that are of unknown association, wherein the preset score threshold range is between 0.55 and 0.75, and the value of k is the same as the number of samples with known association.

[0134] In some exemplary embodiments, the second training module selects a subset of samples from the scored, unassociated samples to update the first initial negative samples, including:

[0135] The samples with unknown associations are sorted in descending order of their scores. The first k samples with scores less than or equal to a preset score threshold are selected to update the first initial negative samples. The preset score threshold is between 0.55 and 0.75, and the value of k is the same as the number of positive samples with known associations.

[0136] In some exemplary embodiments, the data acquisition module 510 selects a subset of samples from a plurality of the unknown associated samples as first initial negative samples, including:

[0137] From the plurality of samples with unknown associations, k samples are randomly selected as the first initial negative samples, and the size of k is the same as the number of positive samples with known associations.

[0138] In some exemplary embodiments, the first training module 520 is further configured to update the association between miRNA and disease based on the prediction results of the hypergraph model generated in the last iteration of training for the unknown associated samples.

[0139] This disclosure also provides an apparatus for predicting the association between miRNA and disease, including a memory; and a processor connected to the memory, the memory being used to store instructions, the processor being configured to perform the steps of the method for predicting the association between miRNA and disease as described in any embodiment of this disclosure based on the instructions stored in the memory.

[0140] As shown in Figure 6, in one example, the device for predicting the association between miRNA and disease may include: a processor 610, a memory 620, a bus system 630, and a transceiver 640. The processor 610, memory 620, and transceiver 640 are connected via the bus system 630. The memory 620 stores instructions, and the processor 610 executes the instructions stored in the memory 620 to control the transceiver 640 to transmit and receive signals. Specifically, under the control of the processor 610, the transceiver 640 can acquire multiple known-association positive samples and multiple unknown-association samples. The processor 610 selects a portion of the samples with unknown associations as first initial negative samples. The hypergraph model is iteratively trained using the positive samples and the second negative samples. After each iteration of training, the current hypergraph model is used to predict the samples with unknown associations. The second negative sample used in the first iteration of training is the first initial negative sample, and the second negative sample used in subsequent iterations of training is a portion of the samples with unknown associations predicted after the current iteration of training.

[0141] It should be understood that processor 610 can be a central processing unit (CPU), or it can be other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), off-the-shelf programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. A general-purpose processor can be a microprocessor or any conventional processor.

[0142] Memory 620 may include read-only memory and random access memory, and provides instructions and data to processor 610. A portion of memory 620 may also include non-volatile random access memory. For example, memory 620 may also store device type information.

[0143] In addition to the data bus, the bus system 630 may also include a power bus, a control bus, and a status signal bus. However, for clarity, all buses are labeled as bus system 630 in Figure 6.

[0144] In implementation, the processing performed by the processing device can be accomplished through integrated logic circuits in the hardware of the processor 610 or through software instructions. That is, the method steps of this embodiment can be executed by a hardware processor, or by a combination of hardware and software modules within the processor. The software modules can reside in random access memory, flash memory, read-only memory, programmable read-only memory, electrically erasable programmable memory, registers, or other storage media. This storage medium is located in memory 620, and the processor 610 reads information from memory 620 and, in conjunction with its hardware, completes the steps of the aforementioned method. To avoid repetition, further details are omitted here.

[0145] This disclosure also provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the method for predicting the association between miRNA and disease as described in any embodiment of this disclosure. The method for predicting the association between miRNA and disease driven by executing executable instructions is essentially the same as the method for predicting the association between miRNA and disease provided in the above embodiments of this disclosure, and will not be described in detail here.

[0146] In some possible implementations, various aspects of the miRNA-disease association prediction method provided in this disclosure can also be implemented as a program product comprising program code that, when run on a computer device, causes the computer device to perform the steps in the miRNA-disease association prediction method according to various exemplary embodiments of this disclosure as described above. For example, the computer device can perform the miRNA-disease association prediction method described in the embodiments of this disclosure.

[0147] The program product may employ any combination of one or more readable media. A readable medium may be a readable signal medium or a readable storage medium. A readable storage medium may be, for example, but not limited to, an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus, or device, or any combination thereof. More specific examples of readable storage media (a non-exhaustive list) include: an electrical connection having one or more wires, a portable disk, a hard disk, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fiber, portable compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination thereof.

[0148] It will be understood by those skilled in the art that all or some of the steps, systems, or apparatuses disclosed above, and their functional modules / units, can be implemented as software, firmware, hardware, or suitable combinations thereof. In hardware implementations, the division between functional modules / units mentioned above does not necessarily correspond to the division of physical components; for example, a physical component may have multiple functions, or a function or step may be performed collaboratively by several physical components. Some or all components may be implemented as software executed by a processor, such as a digital signal processor or microprocessor, or as hardware, or as an integrated circuit, such as an application-specific integrated circuit (ASIC). Such software may be distributed on a computer-readable medium, which may include computer storage media (or non-transitory media) and communication media (or transient media). As is known to those skilled in the art, the term computer storage media includes volatile and non-volatile, removable and non-removable media implemented in any method or technology for storing information (such as computer-readable instructions, data structures, program modules, or other data). Computer storage media include, but are not limited to, RAM, ROM, EEPROM, flash memory or other memory technologies, CD-ROM, digital versatile disc (DVD) or other optical disc storage, magnetic cartridges, magnetic tape, disk storage or other magnetic storage devices, or any other medium that can be used to store desired information and can be accessed by a computer. Furthermore, it is well known to those skilled in the art that communication media typically contain computer-readable instructions, data structures, program modules, or other data in modulated data signals such as carrier waves or other transmission mechanisms, and may include any information delivery medium.

[0149] It should be noted that the above embodiments or implementation methods are merely exemplary and not restrictive. Therefore, this disclosure is not limited to the content specifically shown and described herein. Various modifications, substitutions, or omissions can be made to the form and details of the implementations without departing from the scope of this disclosure.

Claims

1. A method for predicting the association between miRNA and disease, comprising: Obtain multiple positive samples with known associations and multiple samples with unknown associations, and select a portion of the samples with unknown associations as the first initial negative samples; The hypergraph model is trained iteratively using the positive samples and the second negative samples. After each iteration of training, the current hypergraph model is used to predict the unknown associated samples. The second negative sample used in the first iteration of training is the first initial negative sample. The second negative sample used in the remaining iterations of training is a portion of the unknown associated samples predicted after the current iteration of training.

2. The method according to claim 1, prior to the step of iteratively training the hypergraph model using the positive samples and the second negative samples, the method further includes: The scoring model is trained using the positive samples and the first initial negative samples. The trained scoring model is then used to score multiple samples with unknown associations. A subset of the scored samples with unknown associations is then selected to update the first initial negative samples.

3. The method according to claim 2, wherein the scoring model is any one of the following: a first dual-tower model, a convolutional neural network, a recurrent neural network, and a long short-term memory network.

4. The method according to claim 3, wherein, The first dual-tower model includes: a first encoding layer, a second encoding layer, a similarity calculation layer, and a training optimization layer; The first coding layer and the second coding layer are set in parallel and share parameters. The first coding layer encodes the initial feature vector of the miRNA into the feature representation of the miRNA, and the second coding layer encodes the initial feature vector of the disease into the feature representation of the disease. The similarity calculation layer calculates the similarity between the feature representation of miRNA and the feature representation of disease; The training optimization layer calculates the loss function during the training process and updates the model parameters to gradually optimize the model.

5. The method according to claim 2, wherein, The step of selecting a subset of samples from the scored, unassociated samples to update the first initial negative sample includes: The first initial negative sample is updated by selecting k samples with scores within a preset score threshold range that are of unknown association, wherein the preset score threshold range is between 0.55 and 0.75, and the value of k is the same as the number of samples with known association.

6. The method according to claim 2, wherein, The step of selecting a subset of samples from the scored, unassociated samples to update the first initial negative sample includes: The trained scoring model is used to score multiple known associated positive samples, and the mean score of the multiple known associated positive samples is calculated as threshold_up. The samples with unknown associations are sorted in descending order of their scores. The first initial negative samples are updated by selecting the k samples with scores less than or equal to (threshold_up-a), where a is between 0.05 and 0.2, and k is the same as the number of positive samples with known associations.

7. The method according to claim 2, wherein, The step of selecting a subset of samples from the scored, unassociated samples to update the first initial negative sample includes: The samples with unknown associations are sorted in descending order of their scores. The first k samples with scores less than or equal to a preset score threshold are selected to update the first initial negative samples. The preset score threshold is between 0.55 and 0.75, and the value of k is the same as the number of positive samples with known associations.

8. The method according to claim 1, wherein, The hypergraph model includes a second dual-tower model, which includes a third coding layer and a fourth coding layer. The third and fourth coding layers are configured in parallel and share parameters. The third coding layer encodes the initial feature vector of the miRNA into a feature representation of the miRNA, and the fourth coding layer encodes the initial feature vector of the disease into a feature representation of the disease.

9. The method according to claim 8, wherein, The hypergraph model also includes: a hypergraph construction module and a hypergraph convolution module; The second dual-tower model transforms the initial feature vector of the miRNA into a feature representation of the miRNA, and transforms the initial feature vector of the disease into a feature representation of the disease; The hypergraph construction module constructs a miRNA hypergraph using miRNA as nodes and miRNA feature representations as features, and constructs a disease hypergraph using epidemics as nodes and disease feature representations as features. The hypergraph convolution module performs hypergraph convolution on the miRNA hypergraph and the disease hypergraph respectively to obtain the embedding representation of miRNA and the embedding representation of disease.

10. The method according to claim 9, wherein, The structure of the miRNA hypergraph changes dynamically during different training iterations, as does the structure of the disease hypergraph.

11. The method according to claim 9, wherein, The loss function of the hypergraph model is: Where, y′ ij =Z m .Z d Z m and Z d These represent the embedding representations of miRNAs and the disease, respectively, obtained after performing hypergraph convolution; y ij As the truth label, y ij =1 or 0, where U is the training data set.

12. The method according to claim 9, wherein, The hypergraph construction module also calculates the miRNA hypergraph association matrix and the disease hypergraph association matrix through attention scores; the miRNA embedding representation combines the miRNA feature representation information output by the second dual-tower model with the hyperedge information obtained by convolving the miRNA hypergraph association matrix, and the disease embedding representation combines the disease feature representation information output by the second dual-tower model with the hyperedge information obtained by convolving the disease hypergraph association matrix.

13. The method according to claim 1, further comprising, prior to the method: Initial feature vectors for multiple miRNAs and initial feature vectors for various diseases are generated.

14. The method according to claim 13, wherein, The generation of initial feature vectors for multiple miRNAs and initial feature vectors for multiple diseases includes: Establish miRNA functional similarity matrix (FM) and disease semantic similarity matrix (DM); Calculate the Gaussian interaction spectrum kernel similarity matrix KM of miRNAs and the Gaussian interaction spectrum kernel similarity matrix KD of diseases; The miRNA functional similarity matrix FM and the miRNA Gaussian interaction spectrum kernel similarity matrix are used to... KM is integrated into a miRNA similarity representation matrix IM, where each row vector of the miRNA similarity representation matrix IM is an initial feature vector of a miRNA; The disease semantic similarity matrix DM and the disease Gaussian interaction spectrum kernel similarity matrix KD are integrated into a disease similarity representation matrix ID, where each row vector of the disease similarity representation matrix ID is an initial feature vector of a disease.

15. The method according to claim 13, wherein, The miRNA functional similarity matrix FM and the miRNA Gaussian interaction spectrum kernel similarity matrix KM are integrated into a miRNA similarity representation matrix IM according to the following formula: The disease semantic similarity matrix DM and the disease Gaussian interaction spectrum kernel similarity matrix KD are integrated into a disease similarity representation matrix ID according to the following formula:

16. The method according to claim 1, wherein, The step of selecting a subset of samples from the plurality of unknown associated samples as the first initial negative samples includes: From the plurality of samples with unknown associations, k samples are randomly selected as the first initial negative samples, and the size of k is the same as the number of positive samples with known associations.

17. The method according to claim 1, further comprising: The association between miRNAs and the disease is updated based on the prediction results of the hypergraph model generated by the last iteration of training for the samples with unknown associations.

18. An apparatus for predicting the association between miRNA and disease, comprising a memory; and a processor connected to the memory, the memory for storing instructions, the processor being configured to perform the steps of the method for predicting the association between miRNA and disease as claimed in any one of claims 1 to 17 based on the instructions stored in the memory.

19. A computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the method for predicting the association between miRNA and disease as described in any one of claims 1 to 17.

20. A computer program product comprising instructions that, when executed by a computer, perform a method for predicting the association between miRNA and disease as claimed in any one of claims 1 to 17.

21. A device for predicting the association between miRNA and disease, comprising a data acquisition module and a first training module, wherein: The data acquisition module is configured to acquire multiple known-associated positive samples and multiple unknown-associated samples, and select a portion of the samples from the multiple unknown-associated samples as the first initial negative samples. The first training module is configured to iteratively train the hypergraph model using the positive samples and the second negative samples. After each iteration, the module predicts the samples with unknown associations using the current hypergraph model. The second negative sample used in the first iteration is the first initial negative sample, and the second negative sample used in subsequent iterations is the current negative sample. A subset of samples from the unknown associations predicted after each iteration of training.

Images