Correlation prediction model training method, correlation prediction method and system

Abstract

The application provides a correlation prediction model training method, a correlation prediction method and a system. The method comprises the following steps: obtaining a plurality of training samples, wherein the training samples are constructed by a first sample multi-source feature and a second sample multi-source feature with a correlation relationship label, the first sample multi-source feature is a multi-source feature corresponding to a sample micro ribonucleic acid, and the second sample multi-source feature is a multi-source feature corresponding to a sample physiological state type; inputting the plurality of training samples into a preset model, and training the preset model based on a vertex vector representation output by a fusion feature layer of the preset model, to obtain a trained correlation prediction model; and the fusion feature layer is used for fusing a global attention fusion feature representation and a local attention fusion feature representation to obtain the vertex vector representation. The application improves the correlation prediction accuracy of miRNA and the physiological state type.

Description

Technical Field

[0001] This invention relates to the field of bioinformatics association prediction technology, and in particular to an association prediction model training method, association prediction method and system. Background Technology

[0002] MicroRNAs (miRNAs) are a class of non-coding single-stranded RNA molecules, approximately 22 nucleotides in length, encoded by endogenous genes. They play a variety of important regulatory roles in living organisms, including cell differentiation, proliferation, apoptosis, transcription, and transport of genetic material.

[0003] Existing graph neural network-based miRNA and physiological state type (such as biological development, metabolism and other physiological processes) association prediction models typically only utilize a single information source. Furthermore, when fusing multiple perspective features from different vertices, they use the same set of attention coefficients, resulting in low accuracy of the final association prediction.

[0004] Therefore, there is an urgent need for a training method for association prediction models, an association prediction method, and a system to solve the above problems. Summary of the Invention

[0005] To address the problems existing in the prior art, this invention provides a method for training an association prediction model, an association prediction method, and a system.

[0006] This invention provides a method for training an association prediction model, comprising:

[0007] Multiple training samples are obtained. The training samples are constructed from the multi-source features of the first sample and the multi-source features of the second sample, which have related labels. The first multi-source features are the multi-source features corresponding to the sample microRNA, and the second multi-source features are the multi-source features corresponding to the sample physiological state type.

[0008] Multiple training samples are input into a preset model, and the preset model is trained based on the vertex vector representation output by the fusion feature layer of the preset model to obtain a trained association prediction model.

[0009] The fusion feature layer is used to fuse the global attention fusion feature representation output by the global attention fusion layer of the preset model and the local attention fusion feature representation output by the local attention fusion layer of the preset model to obtain the vertex vector representation.

[0010] According to the association prediction model training method provided by the present invention, the step of obtaining multiple training samples includes:

[0011] Obtain a first similarity network and a second similarity network, wherein the first similarity network is a similarity network between the microRNAs of the samples, and the second similarity network is a similarity network between the physiological state types of the samples;

[0012] Based on the network structure information of each node in the first similarity network and the second similarity network, a corresponding node aggregation embedding is constructed, wherein the network structure information is extracted based on a graph convolutional network;

[0013] Based on the node aggregation embedding corresponding to the first similarity network, the network node representation of the sample microRNA is obtained; based on the node aggregation embedding corresponding to the second similarity network, the network node representation of the sample physiological state type is obtained.

[0014] If it is determined that there is a correlation between the sample microRNA and the sample physiological state type, the network node representation of the sample microRNA is used as the first sample multi-source feature, and the network node representation of the sample physiological state type corresponding to the sample microRNA is used as the second sample multi-source feature, thereby constructing the training sample.

[0015] According to a method for training an association prediction model provided by the present invention, after inputting the plurality of training samples into a preset model, the method further includes:

[0016] Based on the preset activation function in the global attention fusion layer and the corresponding network node representation, a first attention coefficient and a second attention coefficient are calculated, wherein the first attention coefficient is the attention coefficient corresponding to the network node representation of the sample microRNA, and the second attention coefficient is the attention coefficient corresponding to the network node representation of the sample physiological state type.

[0017] Based on the first preset convolutional neural network in the global attention fusion layer, a first global attention fusion feature representation is calculated according to the first attention coefficient and the network node representation of the sample microRNA, wherein the first global attention fusion feature representation is the global attention fusion feature representation corresponding to the sample microRNA; based on the second attention coefficient and the network node representation of the sample physiological state type, a second global attention fusion feature representation is calculated, wherein the second global attention fusion feature representation is the global attention fusion feature representation corresponding to the sample physiological state type;

[0018] Based on the global attention fusion layer, the first global attention fusion feature representation and the second global attention fusion feature representation are input to the fusion feature layer.

[0019] According to a method for training an association prediction model provided by the present invention, after inputting the plurality of training samples into a preset model, the method further includes:

[0020] Based on the attention mechanism in the local attention fusion layer and the second preset convolutional neural network, the first local attention fusion feature representation is calculated according to the first node aggregation embedding, wherein the first node aggregation embedding is the node aggregation embedding corresponding to the node in the first similarity network.

[0021] Based on the attention mechanism in the local attention fusion layer and the second preset convolutional neural network, the second local attention fusion feature representation is calculated according to the second node aggregation embedding, wherein the second node aggregation embedding is the node aggregation embedding corresponding to the node in the second similarity network;

[0022] Based on the local attention fusion layer, the first local attention fusion feature representation and the second local attention fusion feature representation are input to the fusion feature layer.

[0023] According to the association prediction model training method provided by the present invention, the method further includes:

[0024] Based on the attention mechanism of the fusion feature layer, the first global attention fusion feature representation and the first local attention fusion feature representation are fused and embedded to obtain the first graph aggregation embedding, wherein the first graph aggregation embedding is the graph aggregation embedding corresponding to the node in the first similarity network;

[0025] Based on the attention mechanism of the fusion feature layer, the second global attention fusion feature representation and the second local attention fusion feature representation are fused and embedded to obtain the second graph aggregation embedding, wherein the second graph aggregation embedding is the graph aggregation embedding corresponding to the node in the second similarity network;

[0026] Based on the interaction embeddings of the first graph aggregation embedding and the second graph aggregation embedding, the interaction embeddings between nodes in the first similarity network and nodes in the second similarity network are calculated.

[0027] The vertex vector representation is obtained by performing a connection operation on the first graph aggregation embedding, the interaction embedding, and the second graph aggregation embedding.

[0028] According to the association prediction model training method provided by the present invention, the step of training the preset model based on the vertex vector representation output by the fusion feature layer of the preset model to obtain a trained association prediction model includes:

[0029] Based on the fully connected layer in the preset model, and according to the vertex vector representation, the corresponding association prediction results between the multi-source features of the first sample and the multi-source features of the second sample in the training samples are predicted.

[0030] Based on the cross-entropy loss function, loss value analysis is performed on the association prediction results and association labels corresponding to the training samples to obtain the loss value;

[0031] If the loss value meets a preset threshold, the trained association prediction model is obtained.

[0032] The present invention also provides an association prediction method, comprising:

[0033] The multi-source features corresponding to the target microRNA and the multi-source features corresponding to the target physiological state type are input into the trained association prediction model to obtain the association prediction result between the target microRNA and the target physiological state type output by the trained association prediction model. The trained association prediction model is trained based on the above-mentioned association prediction model training method.

[0034] The present invention also provides a system for training an association prediction model, comprising:

[0035] The sample acquisition module is used to acquire multiple training samples. The training samples are constructed from the multi-source features of the first sample and the multi-source features of the second sample, which have related labels. The first multi-source features are the multi-source features corresponding to the sample microRNA, and the second multi-source features are the multi-source features corresponding to the sample physiological state type.

[0036] The model training module is used to input multiple training samples into a preset model and train the preset model based on the vertex vector representation output by the fusion feature layer of the preset model to obtain a trained association prediction model.

[0037] The fusion feature layer is used to fuse the global attention fusion feature representation output by the global attention fusion layer of the preset model and the local attention fusion feature representation output by the local attention fusion layer of the preset model to obtain the vertex vector representation.

[0038] The present invention also provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement any of the above-described association prediction model training methods, or the above-described association prediction methods.

[0039] The present invention also provides a non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the association prediction model training method as described above, or the association prediction method as described above.

[0040] The association prediction model training method, association prediction method and system provided by this invention obtain the final vector representation of the vertex through steps such as multi-view feature extraction, vertex global attention fusion feature representation acquisition, vertex local attention fusion feature representation acquisition and fusion feature integration. Based on the final vector representation of the vertex, the association relationship between potential miRNAs and physiological state types is predicted by training a fully connected network in a preset model, thereby improving the accuracy of miRNA and physiological state type association prediction. Attached Figure Description

[0041] To more clearly illustrate the technical solutions in this invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0042] Figure 1 This is a flowchart illustrating the association prediction model training method provided by the present invention.

[0043] Figure 2 This is a schematic diagram of the feature information aggregation structure based on graph neural networks provided by the present invention;

[0044] Figure 3 A schematic diagram of the structure of the global attention fusion layer provided by the present invention;

[0045] Figure 4 A schematic diagram of the structure of the local attention fusion layer provided by the present invention;

[0046] Figure 5 A schematic diagram of the structure of the fusion feature layer provided by the present invention;

[0047] Figure 6 This is a schematic diagram of the structure of the association prediction model training system provided by the present invention;

[0048] Figure 7 This is a schematic diagram of the structure of the electronic device provided by the present invention. Detailed Implementation

[0049] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of this invention. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without creative effort are within the scope of protection of this invention.

[0050] In the field of bioinformatics, predicting the association between miRNAs and physiological state types is an important and challenging task. In addition to associating existing physiological state types with miRNAs in the database, new miRNAs can also be associated with their corresponding physiological state types, which helps to reveal the role of miRNAs in the occurrence and development of physiological states.

[0051] Graph Neural Networks (GNNs) are widely used in bioinformatics due to their superior ability to process complex graph-structured data. However, a common limitation of current GNN-based miRNA and physiological state type association prediction models is that these existing techniques often rely on only a single information source for modeling, limiting further improvements in model prediction performance. Furthermore, existing multi-source information integration models generally only use traditional attention mechanisms to learn the global importance of features from different perspectives. While the same set of attention coefficients is used when fusing features from multiple perspectives at different vertices, the contributions of different perspectives to the feature representations of different vertices are not necessarily the same, which also affects the association prediction results.

[0052] To address the problems existing in the prior art, this invention provides a multi-view graph neural network model based on global and local attention fusion (Global and Local Attention-based Link Prediction, GLALP) for predicting the association between miRNAs and physiological state types.

[0053] Figure 1 This is a flowchart illustrating the association prediction model training method provided by the present invention, as shown below. Figure 1 As shown, the present invention provides a method for training an association prediction model, comprising:

[0054] Step 101: Obtain multiple training samples. The training samples are constructed from the multi-source features of the first sample and the multi-source features of the second sample, which have associated labels. The first multi-source features are the multi-source features corresponding to the sample microRNA, and the second multi-source features are the multi-source features corresponding to the sample physiological state type.

[0055] In graph structures, nodes represent entities (such as miRNAs and genes), and edges represent relationships between entities (such as the binding of miRNAs to target genes). In the field of bioinformatics, graph neural networks can effectively integrate and utilize these complex biological network data.

[0056] To train the graph neural network model, this invention first collects a large number of data sources, including multi-source feature data of miRNAs and multi-source feature data of physiological states. These data come from different biological databases or experimental measurements. The first sample of multi-source features (miRNA-related) may include miRNA expression levels, sequence characteristics, interaction networks with other molecules, and functional annotations. These features together constitute a multi-dimensional vector describing miRNA properties. The second sample of multi-source features (physiological state-related) may include biochemical indicators (such as blood test results), imaging findings, and genetic variations. These features reflect physiological states from different perspectives.

[0057] Furthermore, miRNAs and physiological state types are used as nodes in the graph structure, and edges are constructed based on their relationships (such as known miRNA-physiological state associations, correlations between miRNA expression changes and physiological states, etc.), thus forming a complex graph network containing miRNA nodes, physiological state nodes, and the relationships between them.

[0058] Furthermore, one or more association labels are assigned to each training sample. These labels represent specific associations or classification results (such as a certain metabolic state) between miRNAs and physiological states. These labels are the targets of model training and are used to guide the model to learn the mapping relationship between miRNA features and physiological states.

[0059] Step 102: Input multiple training samples into a preset model, and train the preset model based on the vertex vector representation output by the fusion feature layer of the preset model to obtain a trained association prediction model.

[0060] The fusion feature layer is used to fuse the global attention fusion feature representation output by the global attention fusion layer of the preset model and the local attention fusion feature representation output by the local attention fusion layer of the preset model to obtain the vertex vector representation.

[0061] In this invention, multiple training samples containing known correlations are input into a preset model. These training samples contain data from different sources, such as multi-source features of sample microRNAs (miRNAs) and multi-source features of sample physiological state types. There are correlations between them that the model needs to learn and predict.

[0062] One of the core components of the preset model in this invention is its fusion feature layer. This layer is designed to integrate feature information from different levels, that is, to fuse the feature representations output by the global attention fusion layer and the local attention fusion layer in order to more comprehensively understand the complex relationships between data. Figure 2 This is a schematic diagram of the feature information aggregation structure based on graph neural networks provided by the present invention, which can be referred to. Figure 2 As shown, this invention adaptively learns the importance contribution of different perspectives on multi-view features, thereby obtaining the vector representation of the vertex.

[0063] In this invention, the global attention fusion layer focuses on capturing and integrating feature information across the entire dataset or graph to generate a global attention fusion feature representation. This global perspective helps the model understand macro-trends and global structure within the data. In contrast, the local attention fusion layer focuses more on the local neighborhood information surrounding a node. It generates a local attention fusion feature representation by considering the node's direct or closer neighbors. This local perspective helps the model capture local dependencies and interaction details between nodes.

[0064] In this invention, the fusion feature layer combines global attention fusion feature representations and local attention fusion feature representations through weighted summation, concatenation, or nonlinear transformation to generate the final vertex vector representation. This vertex vector representation contains both global macroscopic information and local microscopic details, enabling a more comprehensive reflection of the features and attributes of nodes (or samples). Based on these vertex vector representations, a preset model is trained. The goal of the training is to optimize the model's parameters so that the model can accurately predict the association relationships in the training samples. For example, the model parameters are adjusted by comparing the model's prediction results with the true labels (such as cross-entropy loss). After multiple iterations of training, the model gradually converges to a stable and well-performing state, resulting in a well-trained association prediction model. This model can accurately predict the associations of new and unseen samples using the patterns and features learned from the training data.

[0065] The association prediction model training method provided by this invention obtains the final vector representation of the vertex through steps such as multi-view feature extraction, vertex global attention fusion feature representation acquisition, vertex local attention fusion feature representation acquisition, and fusion feature integration. Based on the final vector representation of the vertex, the method trains a fully connected network in a preset model to predict the association between potential miRNAs and physiological state types, thereby improving the accuracy of miRNA and physiological state type association prediction.

[0066] Based on the above embodiments, obtaining multiple training samples includes:

[0067] Obtain a first similarity network and a second similarity network, wherein the first similarity network is a similarity network between the microRNAs of the samples, and the second similarity network is a similarity network between the physiological state types of the samples;

[0068] Based on the network structure information of each node in the first similarity network and the second similarity network, a corresponding node aggregation embedding is constructed, wherein the network structure information is extracted based on a graph convolutional network;

[0069] Based on the node aggregation embedding corresponding to the first similarity network, the network node representation of the sample microRNA is obtained; based on the node aggregation embedding corresponding to the second similarity network, the network node representation of the sample physiological state type is obtained.

[0070] If it is determined that there is a correlation between the sample microRNA and the sample physiological state type, the network node representation of the sample microRNA is used as the first sample multi-source feature, and the network node representation of the sample physiological state type corresponding to the sample microRNA is used as the second sample multi-source feature, thereby constructing the training sample.

[0071] In this invention, multi-source features of sample miRNAs and sample physiological state types can be extracted separately through a multi-view feature extraction layer in a preset model, namely, first sample multi-source features and second sample multi-source features, thereby constructing training samples. This invention first obtains the similarity networks corresponding to sample miRNAs and sample physiological state types respectively. In the graph network structure, nodes represent sample miRNAs or sample physiological state types, and edges represent the similarity between sample miRNAs or between sample physiological state types. For example, functional similarity, sequence similarity, and Gaussian interaction similarity between sample miRNAs; or semantic similarity, target similarity, and Gaussian interaction association similarity between physiological state types. In this invention, existing similarity calculation methods can be selected to calculate the similarity networks of sample miRNAs and sample physiological state types, providing a data foundation for subsequent miRNA and sample physiological state type association prediction methods.

[0072] Specifically, after similarity calculation, a similarity network of sample miRNAs (i.e., the first similarity network) and a similarity network of sample physiological state types (i.e., the second similarity network) are obtained. The first similarity network can be specifically represented as:

[0073]

[0074] in, , and , where is the adjacency matrix of sample miRNA nodes to the associated network in the graph network structure, and represents the functional similarity, sequence similarity, and Gaussian interaction similarity between sample miRNAs, respectively.

[0075] The second similarity network can be specifically represented as:

[0076]

[0077] in, , and , is the adjacency matrix of sample physiological state type nodes to the associated network in the graph network structure. Taking the graph network structure constructed from the physiological state type information in the current database as an example, , respectively represent the semantic similarity between physiological state types, the target similarity between physiological state types, and the Gaussian interaction association similarity between physiological state types.

[0078] Furthermore, based on graph convolutional networks, this invention extracts network structure information of various nodes in the similarity network and constructs node aggregation embeddings on different similarity networks, which can be obtained from the following node aggregation embedding formula:

[0079] , ;

[0080]

[0081]

[0082]

[0083]

[0084]

[0085]

[0086]

[0087]

[0088]

[0089]

[0090]

[0091]

[0092] in, , , , where is the initial feature vector of the sample miRNA node in the graph network structure and its association with the network. This represents the initial feature vector of the sample physiological state type node in the graph network structure that is associated with the network. Represents the identity matrix; yes The degree matrix; yes The degree matrix; and The graph convolutional network model represents the first... Layer node aggregation and embedding; and This represents the corresponding learnable weight matrix. It is a non-linear activation function.

[0093] In this invention, a two-layer graph convolutional network model is constructed for predicting the association between sample miRNA and sample physiological state type. Using the aforementioned node aggregation embedding formula, the corresponding node aggregation embeddings are obtained, resulting in two types of node representations: sample miRNA and sample physiological state type.

[0094] ;

[0095] ;

[0096] in, The network node representation of the sample's microRNA, i.e., the multi-source features of the first sample; The network node representation of the physiological state type of the sample is the second sample multi-source feature.

[0097] Based on the above embodiments, after inputting the plurality of training samples into the preset model, the method further includes:

[0098] Based on the preset activation function in the global attention fusion layer and the corresponding network node representation, a first attention coefficient and a second attention coefficient are calculated, wherein the first attention coefficient is the attention coefficient corresponding to the network node representation of the sample microRNA, and the second attention coefficient is the attention coefficient corresponding to the network node representation of the sample physiological state type.

[0099] Based on the first preset convolutional neural network in the global attention fusion layer, a first global attention fusion feature representation is calculated according to the first attention coefficient and the network node representation of the sample microRNA, wherein the first global attention fusion feature representation is the global attention fusion feature representation corresponding to the sample microRNA; based on the second attention coefficient and the network node representation of the sample physiological state type, a second global attention fusion feature representation is calculated, wherein the second global attention fusion feature representation is the global attention fusion feature representation corresponding to the sample physiological state type;

[0100] Based on the global attention fusion layer, the first global attention fusion feature representation and the second global attention fusion feature representation are input to the fusion feature layer.

[0101] In this invention, based on a global attention fusion layer and a convolutional neural network, a preset model adaptively learns the importance contribution of different perspectives on multi-view features to obtain a global attention fusion feature representation of the vertex. Figure 3 A schematic diagram of the global attention fusion layer provided by this invention can be referred to. Figure 3 As shown, after the training samples are input into the preset model, the global attention fusion layer calculates the similarity of sample miRNAs and sample physiological state type nodes in different types and at different view levels. The formula for calculating the attention coefficient is as follows:

[0102]

[0103]

[0104] in, Indicates the first attention coefficient; Indicates the second attention coefficient; It is the global average pooling corresponding to the sample miRNA. It is the global average pooling corresponding to the physiological state type of the sample; and It is a non-linear activation function; , , , The weight matrix is ​​trainable. , , , It is the offset vector;

[0105] Furthermore, based on the first preset convolutional neural network, the global attention fusion feature representations corresponding to the sample miRNA and the sample physiological state type are calculated respectively. The calculation formula for the global attention fusion feature representation is as follows:

[0106]

[0107]

[0108] in, This represents the first global attention fusion feature representation; This represents the second global attention fusion feature representation; and It is a two-dimensional convolutional neural network used to integrate different similarity perspectives and graph neural network layers, namely the first preset convolutional neural network; It is a multiplication operation between element-level attention coefficients and feature representations from different perspectives.

[0109] Based on the above embodiments, after inputting the plurality of training samples into the preset model, the method further includes:

[0110] Based on the attention mechanism in the local attention fusion layer and the second preset convolutional neural network, the first local attention fusion feature representation is calculated according to the first node aggregation embedding, wherein the first node aggregation embedding is the node aggregation embedding corresponding to the node in the first similarity network.

[0111] Based on the attention mechanism in the local attention fusion layer and the second preset convolutional neural network, the second local attention fusion feature representation is calculated according to the second node aggregation embedding, wherein the second node aggregation embedding is the node aggregation embedding corresponding to the node in the second similarity network;

[0112] Based on the local attention fusion layer, the first local attention fusion feature representation and the second local attention fusion feature representation are input to the fusion feature layer.

[0113] In this invention, based on the attention mechanism and convolutional neural network in the local attention fusion layer, a preset model learns the attention coefficients of different vertices when fusing multi-view features, and obtains the local attention fusion feature representation. Figure 4 A schematic diagram of the local attention fusion layer provided by the present invention can be referred to. Figure 4 As shown, based on the network node representations of sample microRNA and sample physiological state types in the above embodiments, the node aggregation embeddings of each corresponding node are obtained. Then, these node aggregation embeddings are applied to the attention mechanism in the local attention fusion layer. The corresponding local attention fusion feature representation is calculated using the local attention fusion feature representation calculation formula, as follows:

[0114]

[0115] ,

[0116] ,

[0117]

[0118]

[0119] in, The node aggregation embedding corresponding to the i-th node in the network node representation of the sample microRNA is the first node aggregation embedding. The node aggregation embedding corresponding to the j-th node in the network node representation representing the physiological state type of the sample is the second node aggregation embedding. and The weight matrix is ​​trainable. and It is the normalization coefficient; and It is a one-dimensional convolutional neural network used to integrate different similarities and graph neural network layer perspectives, namely the second preset convolutional neural network.

[0120] Based on the above embodiments, the method further includes:

[0121] Based on the attention mechanism of the fusion feature layer, the first global attention fusion feature representation and the first local attention fusion feature representation are fused and embedded to obtain the first graph aggregation embedding, wherein the first graph aggregation embedding is the graph aggregation embedding corresponding to the node in the first similarity network;

[0122] Based on the attention mechanism of the fusion feature layer, the second global attention fusion feature representation and the second local attention fusion feature representation are fused and embedded to obtain the second graph aggregation embedding, wherein the second graph aggregation embedding is the graph aggregation embedding corresponding to the node in the second similarity network;

[0123] Based on the interaction embeddings of the first graph aggregation embedding and the second graph aggregation embedding, the interaction embeddings between nodes in the first similarity network and nodes in the second similarity network are calculated.

[0124] The vertex vector representation is obtained by performing a connection operation on the first graph aggregation embedding, the interaction embedding, and the second graph aggregation embedding.

[0125] In this invention, based on the global attention fusion feature representation and the local attention fusion feature representation obtained in the above embodiments, the final vector representation of the vertex is obtained, namely the vertex vector representation. Figure 5 The schematic diagram of the fusion feature layer provided by the present invention can be referred to. Figure 5 As shown, by using the attention mechanism of the fusion feature layer, the global attention fusion feature representation and the local attention fusion feature representation are combined. The corresponding fusion embedding calculation formula is as follows:

[0126]

[0127]

[0128] in, This represents the graph aggregation embedding corresponding to the i-th node in the first similarity network, i.e., the first graph aggregation embedding; This represents the graph aggregation embedding corresponding to the j-th node in the second similarity network, i.e., the second graph aggregation embedding. These are trainable parameters;

[0129] Furthermore, the interaction embedding between the two types of nodes (i.e., the nodes corresponding to the sample microRNA and the nodes corresponding to the sample physiological state type) is calculated using element-wise multiplication. The formula for calculating the interaction embedding is:

[0130]

[0131] in, This represents the interaction embedding between the i-th node in the first similarity network and the j-th node in the second similarity network;

[0132] Based on the connection operation, and combining the two types of nodes and interaction embeddings, the final representation, namely the vertex vector representation, is obtained:

[0133]

[0134] in, Represented by vertex vectors. For connection operations.

[0135] Based on the above embodiments, the step of training the preset model using the vertex vector representation output by the fusion feature layer of the preset model to obtain a trained association prediction model includes:

[0136] Based on the fully connected layer in the preset model, and according to the vertex vector representation, the corresponding association prediction results between the multi-source features of the first sample and the multi-source features of the second sample in the training samples are predicted.

[0137] Based on the cross-entropy loss function, loss value analysis is performed on the association prediction results and association labels corresponding to the training samples to obtain the loss value;

[0138] If the loss value meets a preset threshold, the trained association prediction model is obtained.

[0139] In this invention, a multilayer perceptron (MLP) decoder is applied for link prediction. Then, using a loss function and the vertex vector representation obtained in the above embodiments, the loss value between the association prediction result and the association label is analyzed. Thus, after determining that the preset training conditions are met, a trained association prediction model is obtained. This model can predict the association between potential miRNAs and physiological state types. Specifically, the cross-entropy loss function used in this invention is:

[0140]

[0141]

[0142] in, This is the association prediction result (association probability) between sample miRNA and sample physiological state type obtained through fully connected layer prediction. For a trainable weight matrix, This represents the loss value calculated using the cross-entropy loss function; Labels indicating the association between sample miRNAs and sample physiological state types.

[0143] In one embodiment, the performance of the association prediction model training method provided by the present invention is verified to outperform eight existing association prediction models applicable to miRNA and physiological state types by measuring five common evaluation metrics: AUC (Area Under Curve), precision and recall area under the curve (AUPR), accuracy, F1 score, precision and recall.

[0144] Specifically, in this embodiment, the association process between miRNA and physiological state type data is described. The miRNA-physiological state type association data comes from the MicroRNA database (HMDD v3.2). HMDD v3.2 provides the latest and most comprehensive collection of miRNA-physiological state type associations, identifying 12,446 experimentally verified associations involving 853 miRNAs and 591 physiological state types. Positive samples are designated as known miRNA-physiological state type associations (represented as 1), while negative samples indicate that there is no discernible association between miRNA and physiological state type, represented as 0.

[0145] Furthermore, among the samples labeled 0, an equal number of negative samples are randomly selected as positive samples to maintain sample balance. The dataset is then divided into training and test sets, with the test set comprising 20%. To ensure robust generalization performance, 5x cross-validation is used within the training set to fine-tune the model's hyperparameters. Specifically, all samples in the dataset are divided into five equal-sized subsets, and this process is repeated five times. Each subset serves as the test set, while the other four subsets serve as the training set. The model is then tested on the independent test set and compared with other baseline models.

[0146] In this embodiment, GLALP was implemented based on Torch, and the trainable model variables were optimized using Adam. To avoid overfitting, standard dropout was employed. During training, hidden units were randomly dropped after the final fully connected layers. Based on 5-fold cross-validation, the learning rate was set to 0.0005, the dropout rate to 0.5, the number of GNN layers L = 2, and the embedding dimension k = 128. The model outputs a probability between 0 and 1 for each potential edge of a miRNA-physiological state type pair, using 0.5 as the threshold for link prediction. If the probability is greater than or equal to 0.5, the pair is predicted to be linked (i.e., it is a positive pair); otherwise, the pair is unlinked.

[0147] Furthermore, five common evaluation metrics were used to measure performance: AUC, Area Under the Precision and Recall Curve (AUPR), Precision, F1 Score, and Recall. AUC describes the probability that the model ranks positive samples before negative samples; AUPR is particularly effective for imbalanced datasets and situations where the positive class receives more attention. The other three evaluation metrics are F1 Score, Precision, and Recall.

[0148] Then, GLALP was compared with eight other link prediction methods, including NIMCGCN, MMGCN, ERMDA, HGANMDA, AGAEMD, MINIMDA, MAGCN, and AMHMDA. All experiments were conducted on the HMDD v3.2 dataset, and the parameters of each method were kept in their original state.

[0149] NIMCGCN and MMGCN use graph convolutional networks as the primary method for extracting node embeddings. NIMCGCN extracts embeddings from a single feature dimension, while MMGCN applies a multi-view GCN to extract multi-channel features. ERMDA employs a resampling strategy and uses ensemble learning and feature selection to integrate various data sources to predict the association between miRNAs and physiological state types. HGANMDA is a deep learning model that utilizes hierarchical graph attention networks to predict the association between miRNAs and physiological state types. It focuses on constructing a heterogeneous graph containing associations between miRNAs, physiological state types, and lncRNAs, and applies node-level and semantic-level attention to prioritize connections to improve prediction accuracy. GAEMD utilizes a node-level attention autoencoder within the miRNA physiological state type framework to collect data and reconstruct the miRNA physiological state type connection network. MINIMDA develops various multimodal networks and synthesizes embedding representations of miRNAs and physiological state types by integrating various high-order neighborhood data from these networks, aiming to infer the association between miRNAs and physiological state types. MAGCN utilizes graph convolutional networks with multi-channel attention mechanisms and convolutional neural network combiners to process the constructed lncRNA-miRNA and miRNA-physiological state type networks, and then employs a bilinear decoder to reconstruct the miRNA-physiological state type associations. AMHMDA utilizes attention-aware multi-view similarity networks and hypergraph learning, combined with graph convolutional networks and attention mechanisms, to predict the association between miRNAs and physiological state types.

[0150] Table 1 shows the comparison results of GLALP with other model experiments. Please refer to Table 1 for details.

[0151] Table 1 Comparison results of GLALP with other model experiments

[0152]

[0153] Among them, GLALP's AUC, AUPR, accuracy, F1-score, recall, and precision were 0.9451, 0.9459, 0.8727, 0.8674, and 0.8781, respectively. It can be seen that these six evaluation metrics are the highest among all methods. Specifically, GLALP's F1-score is 0.75% higher than AMHMDA's, and its recall is 1.25% higher than AMHMDA's.

[0154] Furthermore, global attention feature fusion targets individual layers within the graph neural network, while local attention feature fusion focuses on single miRNA and drug nodes. Both fusion methods employ different feature embeddings. In this embodiment, ablation experiments were conducted to evaluate the impact of combining local and global feature fusion methods. Table 2 shows the performance metrics for global and local feature fusion.

[0155] Table 2. Index results of global and local feature fusion

[0156]

[0157] In Table 2, the integration of local and global strategies produced the most favorable results across the five metrics. It can be seen that the fusion of local and global attention showed quite good performance in the miRNA and physiological state type association prediction task. Although the combined strategy did not significantly outperform the other two methods, further research highlights its superior performance on selected datasets.

[0158] Furthermore, several experiments were conducted using AMHMDA on selected datasets, with only 20% of the HMDD v3.2 data used for testing, while the remaining 80% was used to train the AMHMDA model. Using a similar approach, this embodiment evaluated GLALP on various datasets. For model training, 80% of the HMDDv3.2 data was split into ratios of 1:1, 1:3, 1:5, 1:7, and 1:9, with the remaining 20% ​​reserved for the final prediction task. Table 3 shows the experimental performance comparison results between GLALP and AMHMDA, which can be referred to in Table 3:

[0159] Table 3. Comparison of experimental performance between GLALP and AMHMDA

[0160]

[0161] Table 3 presents a performance comparison between GLALP and AMHMDA. Under various class imbalances, GLALP outperforms the existing AMHMDA across all metrics, with the most significant differences observed in AUC and F1 score, where GLALP consistently shows higher values. When the data is split 1:1, GLALP's AUC is 10% higher than AMHMDA's, and GLALP's F1 score is 11% higher than AMHMDA's. GLALP outperforms AMHMDA by at least 1% in all other metrics. These results indicate that the classifier performs better in terms of overall ranking ability (AUC) and the balance between precision and recall (F1 score). The trend shows that as class imbalance increases (from 1:1 to 1:9), the performance of both methods declines across all metrics, which is common in imbalanced classes.

[0162] The present invention also provides an association prediction method, comprising:

[0163] The multi-source features corresponding to the target microRNA and the multi-source features corresponding to the target physiological state type are input into the trained association prediction model to obtain the association prediction result between the target microRNA and the target physiological state type output by the trained association prediction model. The trained association prediction model is trained based on the association prediction model training method described in the above embodiments.

[0164] In this invention, for miRNAs and physiological states, it is necessary to collect characteristic information from multiple sources. These characteristics may include, but are not limited to, gene expression levels, protein-protein interactions, metabolite concentrations, environmental factors (such as diet and lifestyle), and clinical manifestations. The use of multi-source characteristics helps to more comprehensively understand the relationship between miRNAs and physiological states.

[0165] An association prediction model is a pre-trained machine learning or deep learning model used to predict the association between miRNAs and specific physiological states. This model establishes prediction rules by learning from a large amount of data on known miRNA-physiological state associations. After the model is trained, multi-source features of new, unknown-association miRNAs and multi-source features of the target physiological state can be input into the model. These input features will serve as the model's input variables to generate prediction results.

[0166] Furthermore, based on the multi-source features of the input, the model outputs a prediction result that represents the degree or probability of association between the target miRNA and the target physiological state type. This result can be quantitative (such as probability value or score) or qualitative (such as "associated" or "not associated").

[0167] The association prediction method provided by this invention obtains the final vector representation of the vertex through steps such as multi-view feature extraction, vertex global attention fusion feature representation acquisition, vertex local attention fusion feature representation acquisition, and fusion feature integration. Based on the final vector representation of the vertex, the method trains a fully connected network in a preset model to predict the association between potential miRNAs and physiological state types, thereby improving the accuracy of miRNA and physiological state type association prediction.

[0168] The association prediction model training system provided by the present invention is described below. The association prediction model training system described below can be referred to in correspondence with the association prediction model training method described above.

[0169] Figure 6 This is a schematic diagram of the structure of the association prediction model training system provided by the present invention, as shown below.Figure 6 As shown, the present invention provides an association prediction model training system, including a sample acquisition module 601 and a model training module 602. The sample acquisition module 601 acquires multiple training samples, which are constructed from first sample multi-source features and second sample multi-source features with association labels. The first sample multi-source features are multi-source features corresponding to sample microRNAs, and the second sample multi-source features are multi-source features corresponding to sample physiological state types. The model training module 602 inputs the multiple training samples into a preset model and trains the preset model based on the vertex vector representation output by the fusion feature layer of the preset model to obtain a trained association prediction model. The fusion feature layer fuses the global attention fusion feature representation output by the global attention fusion layer of the preset model and the local attention fusion feature representation output by the local attention fusion layer of the preset model to obtain the vertex vector representation.

[0170] The association prediction model training system provided by this invention obtains the final vector representation of vertices through steps such as multi-view feature extraction, vertex global attention fusion feature representation acquisition, vertex local attention fusion feature representation acquisition, and fusion feature integration. Based on the final vector representation of vertices, the system trains a fully connected network in a preset model to predict the association between potential miRNAs and physiological state types, thereby improving the accuracy of miRNA and physiological state type association prediction.

[0171] The system provided by this invention is used to execute the above-described method embodiments. For specific processes and details, please refer to the above embodiments, which will not be repeated here.

[0172] Figure 7 This is a schematic diagram of the structure of the electronic device provided by the present invention, such as... Figure 7As shown, the electronic device may include: a processor 701, a communications interface 702, a memory 703, and a communication bus 704, wherein the processor 701, the communications interface 702, and the memory 703 communicate with each other through the communication bus 704. The processor 701 can call logical instructions in the memory 703 to execute an association prediction model training method. This method includes: acquiring multiple training samples, which are constructed from first sample multi-source features and second sample multi-source features with association labels, wherein the first sample multi-source features are multi-source features corresponding to sample microRNAs, and the second sample multi-source features are multi-source features corresponding to sample physiological state types; inputting the multiple training samples into a preset model, and training the preset model based on the vertex vector representation output by the fusion feature layer of the preset model to obtain a trained association prediction model; wherein the fusion feature layer is used to fuse the global attention fusion feature representation output by the global attention fusion layer of the preset model and the local attention fusion feature representation output by the local attention fusion layer of the preset model to obtain the vertex vector representation.

[0173] Alternatively, an association prediction method may be performed, comprising: inputting multi-source features corresponding to the target microRNA and multi-source features corresponding to the target physiological state type into a trained association prediction model, and obtaining an association prediction result between the target microRNA and the target physiological state type output by the trained association prediction model, wherein the trained association prediction model is trained based on the association prediction model training method described in the above embodiments.

[0174] Furthermore, the logical instructions in the aforementioned memory 703 can be implemented as software functional units and, when sold or used as independent products, can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, essentially, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0175] On the other hand, the present invention also provides a computer program product, the computer program product comprising a computer program stored on a non-transitory computer-readable storage medium, the computer program comprising program instructions, wherein when the program instructions are executed by a computer, the computer is able to execute the association prediction model training method provided by the above methods, the method comprising: acquiring multiple training samples, the training samples being constructed from first sample multi-source features and second sample multi-source features having association relationship labels, wherein the first sample multi-source features are multi-source features corresponding to sample microRNAs, and the second sample multi-source features are multi-source features corresponding to sample physiological state types; inputting the multiple training samples into a preset model, and training the preset model based on the vertex vector representation output by the fusion feature layer of the preset model, to obtain a trained association prediction model; wherein the fusion feature layer is used to fuse the global attention fusion feature representation output by the global attention fusion layer of the preset model and the local attention fusion feature representation output by the local attention fusion layer of the preset model to obtain the vertex vector representation;

[0176] Alternatively, an association prediction method may be performed, comprising: inputting multi-source features corresponding to the target microRNA and multi-source features corresponding to the target physiological state type into a trained association prediction model, and obtaining an association prediction result between the target microRNA and the target physiological state type output by the trained association prediction model, wherein the trained association prediction model is trained based on the association prediction model training method described in the above embodiments.

[0177] In another aspect, the present invention also provides a non-transitory computer-readable storage medium storing a computer program thereon. When executed by a processor, the computer program implements the association prediction model training method provided in the above embodiments. The method includes: acquiring multiple training samples, wherein the training samples are constructed from first sample multi-source features and second sample multi-source features with association labels, wherein the first sample multi-source features are multi-source features corresponding to sample microRNAs, and the second sample multi-source features are multi-source features corresponding to sample physiological state types; inputting the multiple training samples into a preset model, and training the preset model based on the vertex vector representation output by the fusion feature layer of the preset model to obtain a trained association prediction model; wherein the fusion feature layer is used to fuse the global attention fusion feature representation output by the global attention fusion layer of the preset model and the local attention fusion feature representation output by the local attention fusion layer of the preset model to obtain the vertex vector representation;

[0178] Alternatively, an association prediction method may be performed, comprising: inputting multi-source features corresponding to the target microRNA and multi-source features corresponding to the target physiological state type into a trained association prediction model, and obtaining an association prediction result between the target microRNA and the target physiological state type output by the trained association prediction model, wherein the trained association prediction model is trained based on the association prediction model training method described in the above embodiments.

[0179] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Those skilled in the art can understand and implement this without any creative effort.

[0180] Through the above description of the embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus necessary general-purpose hardware platforms, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solutions, in essence or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods described in the various embodiments or some parts of the embodiments.

[0181] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for training an association prediction model, characterized in that, The method comprises: obtaining a plurality of training samples, the training samples being constructed from first sample multi-source features and second sample multi-source features with an association relationship label, wherein the first sample multi-source features and the second sample multi-source features are extracted by a multi-view feature extraction layer in a preset model; the first sample multi-source features are multi-source features corresponding to sample micro ribonucleic acids, and the second sample multi-source features are multi-source features corresponding to sample physiological state types; inputting the plurality of training samples into the preset model, and training the preset model based on a vertex vector representation output by a fusion feature layer of the preset model to obtain a trained association prediction model; wherein the fusion feature layer is configured to fuse global attention fusion feature representations output by a global attention fusion layer of the preset model and local attention fusion feature representations output by a local attention fusion layer of the preset model to obtain the vertex vector representation; the method further comprises: obtaining a first similarity network and a second similarity network, wherein the first similarity network is a similarity network between the sample micro ribonucleic acids, and the second similarity network is a similarity network between the sample physiological state types; constructing corresponding node aggregation embeddings according to network structure information of each node in the first similarity network and the second similarity network, wherein the network structure information is extracted based on a graph convolution network; obtaining network node representations of the sample micro ribonucleic acids according to the node aggregation embeddings corresponding to the first similarity network, and obtaining network node representations of the sample physiological state types according to the node aggregation embeddings corresponding to the second similarity network; in a case where it is determined that the sample micro ribonucleic acids and the sample physiological state types have the association relationship, constructing the training samples by taking the network node representations of the sample micro ribonucleic acids as the first sample multi-source features and taking the network node representations of the sample physiological state types corresponding to the sample micro ribonucleic acids as the second sample multi-source features; the method further comprises: based on a fully connected layer in the preset model, predicting an association prediction result corresponding to the first sample multi-source features and the second sample multi-source features in the training samples according to the vertex vector representation; based on a cross-entropy loss function, performing loss value analysis on the association prediction result corresponding to the training samples and the association relationship label to obtain a loss value; in a case where it is determined that the loss value meets a preset threshold, obtaining the trained association prediction model. 2.The method of claim 1, wherein, after the plurality of training samples are input into the preset model, the method further comprises: The first attention coefficient is an attention coefficient corresponding to a network node representation of the sample micro ribonucleic acid, and the second attention coefficient is an attention coefficient corresponding to a network node representation of the sample physiological state type. The first global attention fusion feature representation is a global attention fusion feature representation corresponding to the sample micro ribonucleic acid, and the second global attention fusion feature representation is a global attention fusion feature representation corresponding to the sample physiological state type. The first global attention fusion feature representation and the second global attention fusion feature representation are input into the fusion feature layer based on the global attention fusion layer. 3.The method of claim 2, wherein, After the plurality of training samples are input into the preset model, the method further comprises: The first local attention fusion feature representation is calculated based on the attention mechanism and the second preset convolutional neural network in the local attention fusion layer according to the first node aggregated embedding, wherein the first node aggregated embedding is a node aggregated embedding corresponding to a node in the first similarity network. The second local attention fusion feature representation is calculated based on the attention mechanism and the second preset convolutional neural network in the local attention fusion layer according to the second node aggregated embedding, wherein the second node aggregated embedding is a node aggregated embedding corresponding to a node in the second similarity network. The first local attention fusion feature representation and the second local attention fusion feature representation are input into the fusion feature layer based on the local attention fusion layer. 4.The method of Claim 3, wherein, The method further comprises: The first graph aggregated embedding is obtained by fusion embedding of the first global attention fusion feature representation and the first local attention fusion feature representation based on the attention mechanism of the fusion feature layer, wherein the first graph aggregated embedding is a graph aggregated embedding corresponding to a node in the first similarity network. The second graph aggregated embedding is obtained by fusion embedding of the second global attention fusion feature representation and the second local attention fusion feature representation based on the attention mechanism of the fusion feature layer, wherein the second graph aggregated embedding is a graph aggregated embedding corresponding to a node in the second similarity network. The interaction embedding between the nodes in the first similarity network and the nodes in the second similarity network is calculated according to the interaction embedding of the first graph aggregated embedding and the second graph aggregated embedding. The first graph aggregation embedding, the interaction embedding, and the second graph aggregation embedding are processed by a connection operation to obtain the vertex vector representation.

5. A method of association prediction, characterized by, Comprise: Input the target micro ribonucleic acid corresponding multi-source feature and the target physiological state type corresponding multi-source feature into the trained association prediction model to obtain the association prediction result between the target micro ribonucleic acid and the target physiological state type output by the trained association prediction model, wherein the trained association prediction model is trained based on the association prediction model training method in any one of claims 1 to 4. 6.A correlation prediction model training system, characterized in that, Comprise: The sample acquisition module is configured to acquire a plurality of training samples, wherein the training samples are constructed from first sample multi-source features and second sample multi-source features that have an association relationship label, and the first sample multi-source features and the second sample multi-source features are extracted by a multi-view feature extraction layer in a preset model; the first sample multi-source features are sample micro ribonucleic acid corresponding multi-source features, and the second sample multi-source features are sample physiological state type corresponding multi-source features; The model training module is configured to input the plurality of training samples into a preset model and train the preset model based on vertex vector representations output by a fusion feature layer of the preset model to obtain a trained association prediction model; The fusion feature layer is configured to fuse global attention fusion feature representations output by a global attention fusion layer of the preset model and local attention fusion feature representations output by a local attention fusion layer of the preset model to obtain the vertex vector representations; The sample acquisition module is specifically configured to: acquire a first similarity network and a second similarity network, wherein the first similarity network is a similarity network between the sample micro ribonucleic acids, and the second similarity network is a similarity network between the sample physiological state types; construct corresponding node aggregation embeddings according to network structure information of each node in the first similarity network and the second similarity network, wherein the network structure information is extracted based on a graph convolution network; obtain network node representations of the sample micro ribonucleic acids according to the node aggregation embeddings corresponding to the first similarity network, and obtain network node representations of the sample physiological state types according to the node aggregation embeddings corresponding to the second similarity network; in a case where it is determined that the sample micro ribonucleic acids and the sample physiological state types have the association relationship, construct the training samples by taking network node representations of the sample micro ribonucleic acids as the first sample multi-source features and taking network node representations of the sample physiological state types corresponding to the sample micro ribonucleic acids as the second sample multi-source features; The model training module is specifically configured to: predict an association prediction result corresponding to the first sample multi-source features and the second sample multi-source features in the training samples based on the vertex vector representations and a fully connected layer in the preset model. Based on a cross-entropy loss function, loss value analysis is performed on the association prediction result corresponding to the training sample and the association relationship label, to obtain a loss value; In a case where it is determined that the loss value meets a preset threshold, the trained association prediction model is obtained.

7. An electronic device comprising a memory, a processor, and a computer program stored on the memory and executable on the processor, characterized in that, The processor implements the association prediction model training method according to any one of claims 1 to 4, or the association prediction method according to claim 5 when executing the computer program.

8. A non-transitory computer-readable storage medium having stored thereon a computer program, characterized in that, The computer program is executed by the processor to implement the association prediction model training method according to any one of claims 1 to 4, or the association prediction method according to claim 5.

Images