A method for predicting the association between miRNA and disease based on heterogeneous graph transformer of biological entity graph

Abstract

The present invention discloses a method for predicting the association between miRNA and disease based on a heterogeneous graph transformer of a biological entity graph, including: selecting and establishing a miRNA-disease association (MDA) dataset, a biological entity dataset, and a biological entity association dataset; constructing a biological entity association graph; extracting and fusing miRNA and disease features; and constructing a deep learning model. The present invention achieves excellent performance under a five-fold cross-validation experiment, proving the effectiveness of the method for predicting the association between miRNA and disease based on a heterogeneous graph transformer of a biological entity graph. A case study on breast cancer further proves that the present invention has practical application functions in identifying new miRNA-disease associations.

Description

Technical Field

[0001] The present invention relates to the technical field of deep learning and bioinformatics, and more specifically, to a method for predicting the association between miRNA and disease based on a heterogeneous graph Transformer of a biological entity graph. Background Art

[0002] MicroRNAs (miRNAs) are non-coding RNAs that participate in the growth, differentiation and apoptosis of biological cells by regulating target genes, and play a vital role in regulating complex diseases, including prevention, prognosis, diagnosis and treatment. Abnormal expression of miRNAs is closely related to the occurrence of diseases. However, only a small number of miRNA-disease associations (MDAs) have been observed. Current computational methods for potential MDA prediction only focus on a few biological entities associated with miRNAs or diseases, but the human body is a highly complex system, and miRNAs, diseases and many types of small biological molecules are related.

[0003] In-depth study of the association between miRNAs and diseases can reveal the potential role of miRNAs in the pathogenesis of diseases and help explore potential biomarkers, which in turn provides new perspectives and methods for early disease diagnosis and personalized treatment. In addition, miRNAs can also serve as drug targets to open up new avenues for disease treatment. Therefore, studying the association between miRNAs and diseases has important theoretical and clinical significance.

[0004] With the widespread application of artificial intelligence technology, more and more machine learning methods are applied to the field of miRNA-disease association prediction. The present invention extracts graph structural elements of miRNAs and diseases and combines them with the attribute information of both to detect possible correlations. Summary of the invention

[0005] The purpose of the present invention is to provide a method for predicting the association between miRNA and disease based on a heterogeneous graph Transformer of a biological entity graph, so as to overcome the defects of the prior art.

[0006] In order to achieve the above object, the technical solution adopted by the present invention is:

[0007] A method for predicting the association between miRNA and disease based on a heterogeneous graph Transformer of a biological entity graph, comprising the following steps:

[0008] S1. Obtain miRNA-disease association data, obtain biological entity data, and obtain biological entity association data;

[0009] S2. Use biological entity data and biological entity association data to build a biological entity association graph, build a node mapping through different node types, build an edge index in heterogeneous data, and construct a node type label to represent the type of each node in the heterogeneous data.

[0010] S3. Calculate the sequence characteristics of miRNAs by comparing the sequences of each pair of miRNAs. Use sequence alignment to quantify the similarity between different miRNAs. And normalize the sequence similarity between miRNAs.

[0011] S4. Use directed acyclic graphs to represent the complex interactions between diseases;

[0012] S5. Calculate the Gaussian interaction profile kernel similarity (GIP) between miRNA and disease, including the GIP similarity between different diseases and the GIP similarity between different miRNAs;

[0013] S6, fusion of the original features of miRNA and disease and the corresponding GIP features to obtain more potential associations between miRNA and disease;

[0014] S7. Construct a heterogeneous graph Transformer, including heterogeneous mutual attention, heterogeneous message passing, and target-specific aggregation.

[0015] S8. Use a 5-fold cross-validation strategy to evaluate the generalization ability of the model.

[0016] Furthermore, in step S1, the latest miRNA-disease association (MDA) data is obtained from the HMDD v3.2 database, and biological entity data and biological entity association data are obtained from the Mesh Database, miRBase Database, mRNA Database, LncRNA Database, Drug Database, Protein Database and circRNA Database;

[0017] Furthermore, in step S2, the biological entity data and the biological entity association data are used to construct a biological entity association graph, a node mapping is constructed through different node types, an edge index is constructed in the heterogeneous data, and a node type label is constructed to indicate the type of each node in the heterogeneous data;

[0018] The biological entity graph consists of 8 biological entities and 16 biological associations, including circRNA-disease association, circRNA-miRNA association, disease-mRNA association, disease-microbe association, drug-disease association, drug-mRNA association, drug-microbe association, drug-protein association, LncRNA-disease association, LncRNA-mRNA association, LncRNA -miRNA association, LncRNA-protein association, miRNA-drug association, miRNA-mRNA association, miRNA-protein association and mRNA-protein association. The entity graph does not contain known miRNA-disease associations in the training set to avoid label leakage. First, the collected biological entities are input into the heterogeneous data as nodes, and then the node mapping is constructed through different node types to build edge indexes in the heterogeneous data. Finally, the node type label is constructed to represent the type of each node in the heterogeneous data.

[0019] Furthermore, in step S3, the sequence characteristics of miRNAs are calculated by comparing the sequences of each pair of miRNAs.

[0020] Specifically, the higher the sequence similarity between two miRNAs, the higher the possibility that they have the same function. 901 miRNA sequences were downloaded from miRBase, and the sequence characteristics of miRNAs were calculated by comparing the sequences of each pair of miRNAs. Based on previous studies, sequence alignment was used to quantify miRNA m a and miRNA m b In addition, the min-max normalization function is used to normalize the sequence similarity between miRNAs. The calculation formula is as follows:

[0021]

[0022] Among them, Score max and Score min Represent the maximum and minimum similarity scores of all miRNA sequence pairs, respectively.

[0023] Furthermore, in step S4, a directed acyclic graph is used to represent the complex interactions between diseases;

[0024] Specifically, MeSH descriptors of multiple diseases were obtained from the National Library of Medicine, and a directed acyclic graph (DAG) was used to represent the complex interactions between diseases. Each MeSH descriptor in the DAG is connected to another through an edge from the parent node to the child node. Each MeSH descriptor has one or more tree structures and is stored in the DAG as a numerical value. For example, the DAG of disease A is represented as DAG(A) = (D(A), E(A)), where D(A) represents the disease and its ancestor, and E(A) represents the connection between two nodes. The semantic similarity of the disease is calculated according to the above definition. The semantic contribution of a disease term t in DAG(A) to disease A is defined as follows:

[0025]

[0026] Among them, Δ represents the semantic contribution attenuation factor, which decreases as the distance between the disease and its ancestor node increases. Therefore, the semantic value of the disease at the first level is 1. In order to distinguish the semantic contribution of diseases at different levels to disease A, the semantic contribution value is obtained by multiplying the contribution of diseases at different levels by the semantic contribution attenuation factor.

[0027] In addition, the semantic value of a disease is defined as follows:

[0028]

[0029] Therefore, the semantic similarity between SCA and SCB of different diseases is defined as follows:

[0030]

[0031] Furthermore, in step S5, the Gaussian interaction spectrum kernel similarity (GIP) between miRNA and disease is calculated, including the GIP similarity between different diseases and the GIP similarity between different miRNAs;

[0032] First, calculate the disease d a With d b Similarities between GIP d (d a , d b ):

[0033] GIP d (d a ,d b )=exp(-γ d *||IP(d a )-IP(d b )|| 2 )

[0034] Where IP(·) represents a binary vector, γ drepresents the kernel bandwidth, and the original parameter γ′ d The normalization calculation is as follows:

[0035]

[0036] Similarly, the miRNAGIP similarity m(a, b) between miRNA a and miRNA b is calculated as follows

[0037] GIP m (m a ,m b )=exp(-γ m *||V(m a )-V(m b )|| 2 )

[0038]

[0039] Furthermore, in step S6, the original features of miRNA and disease and the corresponding GIP features are fused to obtain more potential associations between miRNA and disease;

[0040] For miRNA, construct SM (m a , m b ) as the fused miRNA similarity matrix. It is described as follows:

[0041]

[0042] Similarly, construct SD(d a , d b ) as the fused disease similarity matrix. It is described as follows:

[0043]

[0044] Furthermore, in step S7, a heterogeneous graph transformer (HGT) is constructed, and the entire HGT is divided into three modules: heterogeneous mutual attention, heterogeneous message passing, and target-specific aggregation.

[0045] For heterogeneous mutual attention, unlike the traditional Transformer model, the HGT model assigns a unique set of projection weights W to each meta-relation. A linear mapping technique is applied, namely Convert the target node t into the i-th query vector:

[0046]

[0047] Where ATT-head represents the i-th attention head, K(s) represents the i-th key vector to which the source node s is projected, Q(t) represents the i-th query vector to which the target node t is projected, e represents the relationship between the source node s and the target node t, and μ(·) is the attention adaptive scaling method, representing the general meaning of each relationship triple. The operation of Attention(·) is mainly to connect h ATT-heads to obtain the attention vector for each node pair (s, t). On this basis, the process involves performing another softmax function to generate a probability distribution specific to each target node t, which is generated using the attention vectors accumulated from the neighboring nodes N(t).

[0048] For heterogeneous message passing, edge meta-relationships are introduced into the message passing mechanism of this module to alleviate the differences in the distribution of different types of nodes and edges.

[0049] Message HGT (s,e,t)=|| i∈[1,h] MSG-head i (s,e,t)

[0050]

[0051] For target-specific aggregation, in this module, the above heterogeneous mutual attention and heterogeneous message passing are first aggregated from the source node to the target node as follows:

[0052]

[0053] Then the above results are input into the ELU activation layer, the output is linearly transformed and the residual is connected. The above steps are as follows:

[0054]

[0055] Finally, the feature embedding and extraction of the biological entity graph part are completed.

[0056] Furthermore, in step S8, a 5-fold cross-validation strategy is used to evaluate the generalization ability of the model (HGTMDA), proving the high accuracy and robustness of the model.

[0057] Compared with the prior art, the advantages of the present invention are:

[0058] 1. The present invention collects 8 types of biological entities and their relationships to construct a relatively complete heterogeneous biological entity graph.

[0059] 2. The present invention develops a heterogeneous graph Transformer to extract the graph structural features of miRNA and diseases, and combines their inherent attribute features to identify the potential associations between them.

[0060] 3. The present invention outperforms the most advanced methods in predicting potential miRNA-disease associations, and the results of case studies prove that the model is also effective in practical applications. The present invention uses computational methods including machine learning and deep learning neural networks, and its results are competitive with all the most advanced prediction tools. BRIEF DESCRIPTION OF THE DRAWINGS

[0061] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0062] Figure 1 The present invention is a flow chart of a method for predicting the association between miRNA and disease based on a heterogeneous graph Transformer of a biological entity graph.

[0063] Figure 2 The results of the present invention and other computational methods in simultaneously predicting the association between miRNA and disease are compared. DETAILED DESCRIPTION

[0064] The preferred embodiments of the present invention are described in detail below in conjunction with the accompanying drawings so that the advantages and features of the present invention can be more easily understood by those skilled in the art, thereby making a clearer and more definite definition of the protection scope of the present invention.

[0065] Embodiment 1

[0066] See also Figure 1 As shown, a method for predicting the association between miRNA and disease based on a heterogeneous graph Transformer of a biological entity graph includes the following steps:

[0067] Step S1, obtaining the latest miRNA-disease association (MDA) data, obtaining biological entity data, and obtaining biological entity association data;

[0068] Step S2, constructing a biological entity association graph using biological entity data and biological entity association data, constructing a node mapping through different node types, constructing an edge index in the heterogeneous data, and constructing a node type label to indicate the type of each node in the heterogeneous data;

[0069] Step S3, calculating the sequence characteristics of miRNAs by comparing the sequences of each pair of miRNAs, using sequence alignment to quantify the similarity between different miRNAs, and normalizing the sequence similarity between miRNAs;

[0070] Step S4, using a directed acyclic graph to represent the complex interactions between diseases;

[0071] Step S5, calculating the Gaussian interaction profile kernel similarity (GIP) between miRNA and disease, including the GIP similarity between different diseases and the GIP similarity between different miRNAs;

[0072] Step S6, fusing the original features of miRNA and disease and the corresponding GIP features to obtain more potential associations between miRNA and disease;

[0073] Step S7: construct a heterogeneous graph Transformer, including heterogeneous mutual attention, heterogeneous message passing, and target-specific aggregation.

[0074] Step S8: Use a five-fold cross-validation strategy to evaluate the generalization ability of the model.

[0075] Embodiment 2

[0076] In order to better illustrate the effect of the prediction method of the present invention, we implemented a strict procedure and adopted a cross-validation step to obtain an unbiased evaluation of the effectiveness of the present invention. Table 1 lists the results of the miRNA-disease association prediction of this example under the 5-fold cross-validation method:

[0077] Table 1 Results of the present invention for predicting miRNA-disease association under five-fold cross validation (Accuracy: Sensitive: Specificity: Precision: MCC: Seuss correlation coefficient)

[0078]

[0079] It can be seen from the table that this embodiment has achieved high evaluation scores in multiple experimental results. This result shows that the present invention can not only predict miRNA-disease associations, but also has excellent performance under multiple indicators at the same time.

[0080] Embodiment 3

[0081] In order to test the ability of the present invention to predict miRNA-disease association, we compared the performance of HGTMDA with ERMDA, AGAEMD, PATMDA, NIMCGCN and MAGCN on the same data set. Six evaluation indicators were compared from different perspectives. Table 2 lists the results of miRNA-disease association prediction by this embodiment and the baseline comparison method, where AUC is the area under the ROC curve and AUPRC is the area under the PR curve. The results show that the present invention has higher prediction performance than the latest computational model and is better than other models in comprehensive performance.

[0082] Table 2 Results of miRNA-disease association prediction by this embodiment and baseline comparison method (Accuracy: Precision: Recall: F1-score: F1 score)

[0083]

[0084] Figure 2 The comparison results of the present invention and two groups of ablation experiments are shown. The first group is to remove the attribute features and only use the biological entity graph to extract the features of the heterogeneous graph transformer to predict the miRNA-disease association. The second group is to remove the biological entity graph and only use the attribute features containing the fused GIP to predict the miRNA-disease association. The last group is the result of the present invention, and the height of the bar graph represents the prediction performance of the method. Through comparison, it can be seen that the present invention, that is, HGTMDA that integrates the attribute features and the structural features of the biological entity graph, can better predict the potential miRNA-disease association.

Claims

1. A method for predicting the association between miRNA and disease based on a heterogeneous graph Transformer of a biological entity graph, characterized in that: The following steps are involved: S1. Obtain miRNA-disease associated MDA data, biological entity data, and biological entity associated data; S2. Constructing a biological entity graph using biological entity data and biological entity association data, including: inputting biological entity data as nodes into heterogeneous data, constructing node mapping through different node types, constructing edge indexes in heterogeneous data, and constructing node type labels to represent the type of each node in heterogeneous data; wherein the biological entity graph consists of 8 biological entities and 16 biological associations, including circRNA-disease association, circRNA-miRNA association, disease-mRNA association, disease-microbe association, drug-disease association, drug-mRNA association, drug-microbe association, drug-protein association, LncRNA-disease association, LncRNA-mRNA association, LncRNA-miRNA association, LncRNA-protein association, miRNA-drug association, miRNA-mRNA association, miRNA-protein association, and mRNA-protein association; S3. Calculating the sequence characteristics of miRNAs by comparing the sequences of each pair of miRNAs, including: using sequence alignment to quantify the similarity between different miRNAs and normalizing the similarity between miRNAs; S4. Obtain MeSH descriptors for multiple diseases and use directed acyclic graph DAG(A)=(D(A),E(A)) to represent the complex interactions between diseases; where A represents the disease, D(A) represents the disease and its ancestor, and E(A) represents the connection between two nodes; Define the semantic contribution of a disease term t in DAG(A) to disease A as follows: Among them, Δ represents the semantic contribution decay factor, which decreases as the distance between the disease and its ancestor node increases; Compute semantic similarity between different diseases: Among them, SC A is the semantic value of disease A, calculated as: SC B is the semantic value of disease B; S5, respectively calculating the Gaussian interaction spectrum GIP core similarity between different diseases and the Gaussian interaction spectrum GIP core similarity between different miRNAs; S6, sequence features of fused miRNAs and their GIP features, semantic similarity features of fused diseases and their GIP features; S7, constructing a heterogeneous graph Transformer, including heterogeneous mutual attention, heterogeneous message passing and target aggregation, with the input being the biological entity graph constructed in step S2; wherein, for heterogeneous mutual attention, the heterogeneous graph Transformer model assigns a unique set of projection weights to each meta-relation, and uses linear mapping to convert the target node into a query vector; for heterogeneous message passing, the meta-relation of edges is introduced into the message passing mechanism to alleviate the differences in the distribution of different types of nodes and edges; for target aggregation, the heterogeneous mutual attention and heterogeneous message passing are first aggregated from the source node to the target node, and then the aggregation result is input into the activation layer, the output of the activation layer is linearly transformed and the residual is connected to complete the feature embedding and extraction of the biological entity graph; S8. Input the fused miRNA features and fused disease features obtained in step S6, and the biological entity graph features extracted in step S7 into the constructed multi-layer perceptron to predict the association between miRNA and disease.

2. According to claim 1, a method for predicting the association between miRNA and disease based on a heterogeneous graph Transformer of a biological entity graph, characterized in that: In the step S1, the latest miRNA-disease data is obtained from the database HMDD v3.2, and biological entity data and biological entity associated data are obtained from the databases Mesh Database, miRBase Databasebas e, mRNA Database, LncRNA Database, Drug Database, Protein Database and circRNA Database.

3. According to claim 1, a method for predicting the association between miRNA and disease based on a heterogeneous graph Transformer of a biological entity graph, characterized in that: In step S4, MeSH descriptors of multiple diseases are obtained from the National Library of Medicine, and each MeSH descriptor in the DAG is connected to another through an edge from a parent node to a child node; each MeSH descriptor has one or more tree structures and is stored in the DAG as a number.

Images