Graph attention-based lncrna function prediction method, storage medium and device

Abstract

The lncRNA function prediction method based on graph attention, storage medium and equipment belong to the field of biological information technology. In order to solve the problem that the prior lncRNA function annotation data is little in the existing lncRNA function prediction method, the protein function annotation data needs to be borrowed to indirectly predict the function of lncRNA, resulting in low accuracy. The application first extracts various lncRNA similarity information by using a graph contrast learning method based on a cross attention mechanism, generates a comprehensive similarity matrix; then extracts the features of GO terms in the GO graph by using a knowledge graph embedding model, generates a GO semantic similarity matrix; finally, a graph representation learning model combining GCN and GAT is constructed to learn the features of lncRNA and GO, and a KAN classifier is used to predict the GO function annotation information of lncRNA. The method can effectively improve the accuracy of lncRNA function prediction.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of bioinformatics, and relates to an lncRNA function prediction method, a storage medium and equipment. BACKGROUND

[0002] Long non-coding RNA (lncRNA) is an RNA transcript with a length of more than 200 nucleotides, which lacks key characteristics such as protein-coding potential and sequence conservation, which are indispensable for functional roles. lncRNAs are widely involved in animal neural development, cell regulation, tumor occurrence and metastasis. In addition, human diseases and cancers are related to mutations and disorders of lncRNAs. In the whole genome analysis of lncRNA stability, it is found that the half-life of lncRNA varies widely, which all indicates that lncRNA has complex metabolism and extensive functions. Since lncRNA plays an important role in regulating biological activity, it is particularly important to determine the biological function of lncRNA.

[0003] Although a large number of lncRNAs have been identified so far, the functions of most lncRNAs are unknown compared with protein-coding genes. The gap between the discovery rate and the annotation rate of lncRNAs leads to limited understanding of the functions of lncRNAs. The methods for annotating the functions of lncRNAs are divided into two categories: biological experimental methods and calculation-based methods. The biological experimental method can accurately determine the function of lncRNA, but the experimental cost is high and time-consuming. Therefore, it is an urgent requirement for lncRNA research to develop a calculation alternative for predicting the function of lncRNA. The lncRNA function prediction method is divided into two kinds. The first kind is based on the calculation framework, using hypergeometric test, KATZ measurement, double random walk, matrix decomposition and other methods to calculate the association score of lncRNA and protein, and predicting the function of lncRNA according to the association score and known protein-GO annotation; the second kind of method is based on machine learning method, through various means including RWR, AvgSim, Bayesian network, LSTM, etc. to extract features, and then train the learned features into SVM, RF, etc. classifier model for prediction.

[0004] However, since the number of lncRNAs verified by experiments is small, most methods predict the GO annotation function of lncRNA by known protein-GO association, thereby causing the problem of low accuracy of lncRNA function prediction. SUMMARY

[0005] The purpose of the present invention is to solve the problem that in existing lncRNA function prediction methods, there is little prior lncRNA function annotation data, and protein function annotation data need to be borrowed to indirectly predict lncRNA function, resulting in low accuracy.

[0006] A lncRNA function prediction method based on graph attention, comprising the following steps:

[0007] S1. The similarity matrix S of the three lncRNAs was obtained based on the co-expression of lncRNA, lncRNA-protein interaction, and lncRNA-miRNA interaction. L 、S P and S M ; The embedded representation z of lncRNA is obtained by GCN model and graph contrast learning method based on cross attention mechanism respectively l 、z m and z p , based on z l 、z m and z p Get the lncRNA comprehensive similarity matrix M l ;

[0008] S2. Based on the GO semantic similarity calculation method in the GOGCN method, a knowledge graph embedding model is designed to learn the representation of GO terms; the semantic similarity between terms is calculated using the representation of the terms to obtain the GO similarity matrix M g ;

[0009] S3. Construct a training set and lncRNA-GO association matrix based on prior lncRNA-GO functional annotation data;

[0010] S4. Using hybrid graph representation learning combining GCN and GAT to learn lncRNA comprehensive similarity matrix M l , GO similarity matrix M g Processing is performed to obtain lncRNA-GO fusion features; the obtained lncRNA-GO fusion features are input into the KAN classifier to obtain the predicted lncRNA-GO functional annotations; the specific process includes the following steps:

[0011] S41, lncRNA comprehensive similarity matrix M l , GO similarity matrix M g Input them into GCN to get features H l (1) and H g (1) ;

[0012] S42, H l (1)and lncRNA similarity matrix M l Input into GAT to obtain the feature H optimized by the attention mechanism l (2) ; H g (1) and GO similarity matrix M g Input into GAT to obtain the feature H optimized by the attention mechanism g (2) ;

[0013] S43, H l (2) and M l Input into GCN to obtain the re-extracted feature H of the structural information after attention mechanism optimization l (3) ; H g (2) and M g Input into GCN to obtain the re-extracted feature H of the structural information after attention mechanism optimization g (3) ;

[0014] S44, H l (1) and H l (3) Input into CNN to get fusion feature X l , H g (1) and H g (3) Input into CNN to get fusion feature X g ; According to the fusion feature X l and X g Get the predicted preference matrix

[0015] The binary cross entropy loss function is used to measure the difference between U and the lncRNA-GO association matrix obtained in step S3; the final lncRNA feature representation X is obtained by minimizing the binary cross entropy loss function on the training set obtained in step S3. l and GO’s feature representation X g ;

[0016] S45, the lncRNA feature X obtained above l and GO feature X g The lncRNA-GO fusion features were concatenated and input into the KAN classifier to obtain the predicted lncRNA-GO functional annotations.

[0017] Furthermore, in step S1, the similarity matrix S of the three lncRNAs is obtained according to the co-expression of lncRNA, lncRNA-protein interaction, and lncRNA-miRNA interaction. L 、S P and S M The process includes the following steps:

[0018] S11. Obtain lncRNA co-expression similarity data: extract lncRNA expression profiles and calculate the Pearson correlation coefficient between each pair of lncRNA expression profiles to represent the lncRNA co-expression similarity S L ;

[0019] S12. Obtain similarity data of lncRNA-miRNA interactions: Obtain a dataset of lncRNA-miRNA interaction data. After obtaining the association of each pair of lncRNA-miRNAs based on whether there is an association in the dataset, calculate the Gaussian kernel similarity matrix between lncRNA and miRNA to represent the similarity S of the lncRNA-miRNA interaction. M ;

[0020] S13. Obtain similarity data of lncRNA-protein interactions: Obtain a dataset of lncRNA-protein interactions. After obtaining the association of each pair of lncRNA-proteins based on whether a lncRNA-protein pair exists in the interaction dataset, calculate the Gaussian kernel similarity matrix between lncRNAs and proteins to represent the similarity S of the lncRNA-protein interaction. P .

[0021] Furthermore, the embedded representation z of lncRNA is obtained by the GCN model and the graph contrast learning method based on the cross attention mechanism in step S1. l 、z m 、z p The process includes the following steps:

[0022] S14. lncRNA co-expression similarity matrix S L , similarity matrix S of lncRNA and protein interactions P , similarity matrix S of lncRNA and miRNA interactions M Input them into the GCN model respectively, and the corresponding GCN generates the corresponding feature representation y l 、y p 、y m ;

[0023] Then, the cross attention calculation is performed on the output features of GCN, that is, the cross attention calculation is performed between the three features; l For example, first, l Applied with y m and y p The cross attention mechanism is used to input y m and y p The information in y is integrated into l In the above example, a new feature y′ is generated. l ;

[0024] The same method was used to obtain the characteristic y′ of lncRNA-miRNA m , lncRNA-protein characteristics y′ p ;

[0025] S15. After obtaining three features through the cross-attention mechanism, the feature pairs of the same node in different views are regarded as positive sample pairs to obtain the feature y′ l and feature y′ m Positive sample pairs, features y′ l and feature y′ p Positive sample pair, feature y′ m and feature y′ p There are three positive sample pairs; negative sample pairs are feature pairs between different nodes, according to y′ l , y′ m , y′ p The corresponding positive and negative sample pairs are used to obtain the lncRNA co-expression feature z using graph contrast learning. l , characteristics of lncRNA-miRNA m , lncRNA-protein characteristics p .

[0026] Furthermore, the graph contrast learning method was used to obtain the lncRNA co-expression feature z l In the process, the contrast loss used is as follows:

[0027]

[0028] Among them, 1 [k≠i] Represents the indicator function. When k≠i, the value of the indicator function is 1, which is used to calculate the similarity between node i and node k. When k=i, the value of the indicator function is 0, that is, the similarity between the same nodes is not calculated; y′ k represents y lnegative sample pair of [i]; N represents the number of lncRNAs; τ represents a temperature parameter, used to adjust the distribution smoothness of the similarity calculation; sim(·) represents a similarity function between two features.

[0029] Further, the process of obtaining the z l based on z m and z p in step S1 comprises the following steps:

[0030] S16, the feature z l of lncRNA co-expression is obtained m , the feature z p of lncRNA-miRNA is obtained l , and the feature z m of lncRNA-protein is obtained p .

[0031] S17, according to the feature representation Z of lncRNA, the cosine similarity of the feature representation Z of each two lncRNAs is calculated to obtain the comprehensive similarity matrix M l of lncRNA.

[0032] Further, the process of designing the knowledge graph embedding model to learn the representation of GO terms in step S2 comprises the following steps:

[0033] S21, the GO file is obtained;

[0034] S22, the GO graph is represented as G=(V, R, E, X, Z), wherein V is the entity set, representing the GO term; R is the relationship set; E represents the edge set; X represents the initialized entity representation, and Z represents the initialized relationship representation;

[0035] S23, to obtain the structure information of the GO graph, the encoder adds the propagation weight of the GCN layer on the basis of COMPGCN, and expands the edge and relationship with inverse edges and relationships;

[0036] S24, in the decoder, the GO graph is regarded as a triple, and a link prediction model is trained by using the corresponding representation of the triple; in the training stage, a 1-n scoring strategy is adopted, the corrupted triples are generated by destroying the head term or the tail term, ConvE is used as the scoring function, a higher score is provided for the real triple, and a lower score is provided for the corrupted triple, and the model reaches convergence to obtain the embedding representation of the GO term;

[0037] S25, after training, the cosine similarity is used to estimate the semantic similarity between entities according to the representation of the GO term, and then the GO similarity matrix M is obtainedg .

[0038] Furthermore, the process of constructing a training set based on prior lncRNA-GO functional annotation data in step S3 includes the following steps:

[0039] First, obtain the lncRNA functional annotation file predicted by the lnc-GFP method, and use the lncRNA-GO functional annotation pairs in the predicted lncRNA functional annotation file as positive samples;

[0040] S32. Randomly construct lncRNA-GO non-functional annotation pairs with the same number as lncRNA-GO functional annotation pairs as negative samples; lncRNA-GO functional annotation pairs and non-functional annotation pairs with the same number as lncRNA-GO functional annotation pairs form a training set.

[0041] Furthermore, the process of constructing the lncRNA-GO association matrix based on the prior lncRNA-GO functional annotation data in step S3 includes the following steps:

[0042] According to the predicted lncRNA functional annotation file, the lncRNA-GO association matrix was constructed, where 1 represents a lncRNA-GO functional annotation pair and 0 represents a lncRNA-GO non-functional annotation pair.

[0043] A computer storage medium stores at least one instruction, which is loaded and executed by a processor to implement the graph attention-based lncRNA function prediction method.

[0044] A graph attention-based lncRNA function prediction device comprises a processor and a memory, wherein the memory stores at least one instruction, and the at least one instruction is loaded and executed by the processor to implement the graph attention-based lncRNA function prediction method.

[0045] The beneficial effects of the present invention are:

[0046] This paper utilizes a cross-attention-based graph contrast learning method to extract lncRNA features, a knowledge graph embedding model to extract GO features, and a graph representation learning model combining GCN and GAT to predict lncRNA-GO functional annotations. Therefore, this paper effectively addresses the problem of limited prior lncRNA functional annotation data, requiring indirect prediction of lncRNA function using protein functional annotation data, resulting in low accuracy. Compared with other advanced models, this paper demonstrates superior prediction accuracy and model generalization, providing effective technical support for discovering potential lncRNA functional annotations. BRIEF DESCRIPTION OF THE DRAWINGS

[0047] Figure 1 Flow chart for lncRNA function prediction based on graph attention;

[0048] Figure 2 Ablation result graph for cross-attention-based graph contrastive learning method;

[0049] Figure 3 Ablation result PR curve graph for cross-attention-based graph contrastive learning method;

[0050] Figure 4 Ablation result graph for network model combining GCN and GAT;

[0051] Figure 5 PR curve for network model combining GCN and GAT;

[0052] Figure 6 Influence diagram of different classifiers on lncRNA function prediction performance;

[0053] Figure 7 GO term depth and F1 graph under different classifiers;

[0054] Figure 8 GO term IC and F1 graph under different classifiers;

[0055] Figure 9 PR curve comparison graph of the present application and related lncRNA function prediction models. DETAILED DESCRIPTION

[0056] In order to more clearly describe the present application, and introduce the purpose, technology of the present application, the following will be combined with relevant data, drawings and examples, further explain the present application. In particular, the description is a part of the specific examples of the present application, and is not a limitation on the scope of the invention.

[0057] DETAILED DESCRIPTION Figure 1 The present embodiment provides a lncRNA function prediction method based on graph attention, comprising the following steps S1-S5:

[0058] The present embodiment provides a lncRNA function prediction method based on graph attention, comprising the following steps S1-S5:

[0059] S1, obtain three lncRNA similarity matrices according to lncRNA co-expression, lncRNA-protein interaction, lncRNA-miRNA interaction, etc.; obtain the embedding representation of lncRNA by a graph contrastive learning method based on cross-attention mechanism; calculate the comprehensive similarity matrix of lncRNA. The specific process includes the following steps:

[0060] S11. Obtain lncRNA co-expression similarity data: Extract lncRNA expression profiles from the NONCODE database and calculate the Pearson correlation coefficient between each pair of lncRNA expression profiles to represent the lncRNA co-expression similarity S L .

[0061] S12. Obtain similarity data of lncRNA-miRNA interactions: Obtain lncRNA-miRNA interaction data through the starBase database. If there is an association in the dataset, the score is set to 1, otherwise it is set to 0. After obtaining the score of each lncRNA-miRNA pair, calculate the Gaussian kernel similarity matrix between lncRNA and miRNA to represent the similarity S of the lncRNA-miRNA interaction. M .

[0062] S13. Obtain similarity data of lncRNA-protein interactions: Obtain lncRNA-protein interactions through the NPInter database. If there is a lncRNA-protein pair in the interaction dataset, set their interaction score to 1, otherwise set it to 0. After obtaining the score of each lncRNA-protein pair, calculate the Gaussian kernel similarity matrix between lncRNA and protein to represent the similarity of lncRNA-protein interaction S P .

[0063] S14. lncRNA co-expression similarity matrix S L , similarity matrix S of lncRNA and protein interactions P , similarity matrix S of lncRNA and miRNA interactions M Input them into the GCN model respectively, and the corresponding GCN generates the corresponding feature representation y l 、y p 、y m :

[0064] y l =GCN(s l ,A l )

[0065] y m =GCN(s m ,A m )

[0066] y p =GCN(s p ,A p )

[0067] Among them, s l Represents the graph corresponding to the lncRNA co-expression similarity matrix, sp The graph corresponding to the similarity matrix of lncRNA and protein interactions, s m A graph showing the similarity matrix of lncRNA and miRNA interactions; l 、A m 、A p They are s l 、s m 、s p The adjacency matrix of .

[0068] Then, the cross attention calculation is performed on the output features of GCN, that is, the cross attention calculation is performed between the three features. l For example, first, l Applied with y m and y p The cross attention mechanism is used to input y m and y p The information in y is integrated into l In the above example, a new feature y′ is generated. l .

[0069] The same method was used to obtain the characteristic y′ of lncRNA-miRNA m , lncRNA-protein characteristics y′ p .

[0070] y′ l =CrossAttention(y l ,y m ,y p )

[0071] y′ m =CrossAttention(y m ,y p ,y l )

[0072] y′ p =CrossAttention(y p ,y l ,y m )

[0073] S15. After obtaining three features through the cross-attention mechanism, the feature pairs of the same node in different views are regarded as positive sample pairs, and the feature y′ can be obtained. l and feature y′ m Positive sample pairs, features y′ l and feature y′ p Positive sample pair, feature y′ m and feature y′ pThere are three types of positive sample pairs. Negative sample pairs are feature pairs between different nodes, regardless of whether these nodes are from the same view or different views. l , y′ m , y′ p The corresponding positive and negative sample pairs are used to obtain the lncRNA co-expression feature z based on the contrast loss function NT-Xent. l , characteristics of lncRNA-miRNA m , lncRNA-protein characteristics p .

[0074] Based on the lncRNA co-expression comparison loss function L l Obtain lncRNA co-expression characteristics l The process includes:

[0075] In the calculation of lncRNA co-expression contrast loss, for node i, the positive sample pair includes y′ l [i] and y′ m [i],y′ l [i] and y′ p [i], negative sample pairs represent feature pairs between different nodes. By maximizing the similarity of positive sample pairs and minimizing the similarity of negative sample pairs, the feature z of lncRNA co-expression is finally obtained after training. l .

[0076]

[0077] Among them, when k≠i, the indicator function 1 [k≠i] The value of y′ is 1, which is used to calculate the similarity between node i and node k. When k=i, the value of the indicator function is 0, that is, the similarity between the same nodes is not calculated, regardless of whether the nodes are from the same view; k represents y′ l [i] is the sample corresponding to the negative sample pair; N is the number of lncRNAs; τ is the temperature parameter used to adjust the distribution smoothness of the similarity calculation; sim(·) represents the similarity function between two features, which is expressed by the following formula:

[0078]

[0079] Among them, ‖y' j ‖ represents the two-norm.

[0080] The same method was used to obtain the characteristics of lncRNA-miRNA. m , lncRNA-protein characteristics p .

[0081] After comparing the loss function NT-Xent, three features z are obtained. l 、z m 、z p .

[0082] S16. Obtained the characteristics of lncRNA co-expression l , characteristics of lncRNA-miRNA m , lncRNA-protein characteristics p After feature fusion, the three features are concatenated into a feature representation Z = [z l ||z m ||z p ].

[0083] S17. Based on the lncRNA feature representation Z, calculate the cosine similarity of the feature representation Z of each two lncRNAs to obtain the lncRNA comprehensive similarity matrix M l .

[0084] S2. Based on the GO semantic similarity calculation method in the GOGCN method, a knowledge graph embedding model is designed to learn the representation of GO terms. The semantic similarity between terms is calculated using the term representations. The GO semantic similarity calculation method in the GOGCN method can be implemented using "GOGCN: Graph Convolutional Network on Gene Ontology for Functional Similarity Analysis of Genes" by Tian Z et al. The process of designing a knowledge graph embedding model to learn the representation of GO terms includes the following steps:

[0085] S21. GO files can be downloaded from the Gene Ontology database.

[0086] S22. Represent the GO graph as G = (V, R, E, X, Z), where V is the entity set, representing GO terms; R is the relationship set, representing the two relationships "is_a" and "part_of"; E is the edge set; X represents the initialized entity representation, and Z represents the initialized relationship representation.

[0087] To obtain the structural information of the GO graph, the encoder adds the propagation weights of the GCN layer based on COMPGCN and uses inverse edges and relations to represent the extended edges and relations. The encoder adds the propagation weights of the GCN layer based on COMPGCN based on Vashishth S et al.'s "Composition-based multi-relational graph convolutional networks" to implement this.

[0088] In S24, the decoder treats the Go graph as triplets, and uses the corresponding triple representations to train the link prediction model. During the training phase, a 1-n scoring strategy is used to generate corrupted triplets by corrupting the head or tail term. ConvE is used as the scoring function, giving higher scores to true triplets and lower scores to corrupted triplets. The model converges to obtain embedded representations of Go terms.

[0089] S25. After training, the cosine similarity is used to estimate the semantic similarity between entities for the representation of GO terms, and then the GO similarity matrix M is obtained. g .

[0090] S3. Collect and organize prior lncRNA-GO functional annotation data to construct a training set and lncRNA-GO association matrix. The specific process includes the following steps:

[0091] S31. Obtain the lncRNA function annotation file predicted by the lnc-GFP method, and use the lncRNA-GO function annotation pair in the predicted lncRNA function annotation file as a positive sample; the lncRNA function annotation file predicted by the lnc-GFP method was obtained from "Long non-coding RNAs function annotation: a global prediction method based on bi-colored networks" by GuoX et al.

[0092] S32. In order to balance the number of positive and negative samples in the training set, randomly construct lncRNA-GO non-functional annotation pairs equal to the number of lncRNA-GO functional annotation pairs as negative samples; the lncRNA-GO functional annotation pairs and the non-functional annotation pairs equal to the number of lncRNA-GO functional annotation pairs form the training set.

[0093] S33. Based on the predicted lncRNA functional annotation file, a lncRNA-GO association matrix was constructed, where 1 represents a lncRNA-GO functional annotation pair and 0 represents a lncRNA-GO non-functional annotation pair.

[0094] S4. In order to fully learn the feature representation of lncRNA and GO, a hybrid graph representation combining GCN and GAT is used to learn the lncRNA comprehensive similarity matrix M l , GO similarity matrix M g Processing is performed to obtain lncRNA-GO fusion features; the obtained lncRNA-GO fusion features are input into the KAN classifier to obtain the predicted lncRNA-GO functional annotations; the specific process includes the following steps:

[0095] S41, input the lncRNA similarity matrix M l , the GO similarity matrix M g into the GCN respectively to obtain the features H l (1) and H g (1) .

[0096] S42, input H l (1) and the lncRNA similarity matrix M l into the GAT to obtain the features H l (2) ; input H g (1) and the GO similarity matrix M g into the GAT to obtain the features H g (2) .

[0097] S43, input H l (2) and M l into the GCN to obtain the re-extracted features H l (3) of the structure information optimized by the attention mechanism; input H g (2) and M g into the GCN to obtain the re-extracted features H g (3) .

[0098] Take the feature acquisition of lncRNA as an example, H l (0) represents the lncRNA node features input by the lncRNA similarity network. Then the GCN and the GAT are introduced alternately to obtain lncRNA feature representations of different levels.

[0099]

[0100] wherein, M l is the adjacency matrix of the lncRNA similarity network.

[0101] S44, input H l (1) and H l (3) into the CNN to obtain the fusion features X l , input H g (1) and H g(3) Input into the CNN to obtain the fusion feature X g ; according to the fusion feature X l and X g Obtain the predicted preference matrix U:

[0102]

[0103] The difference between U and the known lncRNA-GO association matrix is measured using a binary cross-entropy loss function (i.e. a BCE loss function). By minimizing the BCE loss function on the training set (the training set obtained in S3), the final feature representation X l and the feature representation X g of GO can be obtained.

[0104] S45, connect the above-obtained lncRNA feature X l and the GO feature X g to form an lncRNA-GO fusion feature, which is input into a KAN classifier to obtain a predicted lncRNA-GO functional annotation.

[0105] S5, perform comparative experiments and ablation experiments on commonly used public test data sets, collect and organize the experimental results; use AUPR (area under the PR curve), precision (Precision), recall (Recall), F1 score (F1-score), and accuracy (Accuracy) and other indicators to analyze the experimental results from multiple angles, thereby evaluating the performance of the new method.

[0106] S51, compare the performance of the present application on the training set and the independent test set lncRNA2GO-55 from the graph contrast learning method without cross-attention, the graph contrast learning method without GCN, the network model without GCN, the network model without GAT, and the present method, respectively. The evaluation indicators include AUPR (area under the PR curve), precision (Precision), recall (Recall), F1 score (F1-score), and accuracy (Accuracy).

[0107] S52, compare the performance of the present application and the related models HLPFN, KATZLGO, BIRWLGO, DNGRGO, and PLNRGO on the independent test set lncRNA2GO-55. The evaluation indicators include precision (Precision), recall (Recall), and F1 score (F1-score).

[0108] The specific calculation formula is as follows:

[0109]

[0110] Among them, TP is true positive, which indicates the number of correctly predicted lncRNA-GO associations; FP is false positive, which indicates the number of incorrectly predicted lncRNA-GO associations; TN is true negative, which indicates the number of correctly predicted non-lncRNA-GO associations; FN is false negative, which indicates the number of incorrectly predicted non-lncRNA-GO associations.

[0111] The recognition effect of the present invention is further described below with a set of specific experimental examples.

[0112] In order to evaluate the impact of graph contrast learning based on the cross-attention mechanism on the performance of lncRNA function prediction, this example established a prediction method that directly integrates three similarity matrices to represent the similarity of lncRNAs (lncRNA-TSIM), a prediction method that only uses graph contrast learning to extract lncRNA features (lncRNA-GraphCL), and a prediction method that uses graph contrast learning based on the cross-attention mechanism to extract lncRNA features (lncRNA-CA-GraphCL). The performance on an independent test set was compared. Figures 2-3 As shown ( Figure 2 The order of the five columns in each method from left to right is Accuracy, Precision, Recall, F1, AUPRC, Figure 3 (representing the PR curve), the performance of the proposed method on the independent test set is better than that of other models. This shows that the graph contrast learning method based on the cross-attention mechanism can more effectively extract lncRNA features.

[0113] In order to evaluate the effect of the graph representation learning model combining GCN and GAT on the performance of lncRNA function prediction, this example established a graph representation learning model with only GCN (GraphGCN), a graph representation learning model with only GAT (GraphGAT), and a graph representation learning model with GCN and GAT (GraphGCN-GAT, this model), and compared their performance on an independent test set. Figures 4-5 As shown ( Figure 4 The order of the five columns in each method from left to right is Accuracy, Precision, Recall, F1, AUPRC, Figure 5(The PR curves for each method are shown in Figure 2). The performance of the present invention on the independent test set is superior to that of other models. The F1 of the present invention is slightly higher than that of other methods, but the accuracy, precision, and AUPRC of the present invention are significantly higher than those of other methods. This demonstrates that the graph representation learning model combining GCN and GAT can effectively integrate representations at all levels and effectively improve the representation learning capabilities of lncRNA and GO.

[0114] In order to determine the most suitable classifier for the present invention, various machine learning-based models were tested, such as random forest (RF), support vector machine (SVM), decision tree (DT), logistic regression (LR), adaptive boosting (AdaBoost), multilayer perceptron (MLP), K nearest neighbor algorithm (KNN), Kolmogorov-Arnold Networks (KAN). Different classifiers were used to verify their performance on independent test sets to predict lncRNA-GO functional annotations. Figure 6 As shown in Figure 3, the model using KAN as the final classifier significantly outperformed the other methods in terms of the distribution of AUPRC values. This indicates that the KAN classifier is more suitable for the present invention than the other tested algorithms for predicting lncRNA-GO association.

[0115] To explain the results in more depth, Figures 7-8 The F1-GO depth and F1-IC of different classifiers on an independent test set are shown. Because GO terms have depth and breadth, the closer to the root node a GO term is, the less meaningful it is. Therefore, the IC value for each GO term needs to be calculated. The IC value is calculated as follows:

[0116]

[0117] Here, depth(term) represents the depth of the term in the GO graph, max_depth is the maximum depth of a term in the GO graph, descendants(term) is the number of a term and all its descendant terms, and total_terms is the total number of all terms in the GO graph. It can be seen that the KAN classifier performs better at almost all levels of prediction.

[0118] In order to verify the performance superiority of the present invention, the lncRNA function prediction task was performed on the test data set using the methods of the present invention, HLPFN, KATZLGO, BIRWLGO, DNGRGO, PLNRGO, etc., and the performance was obtained after analyzing their prediction results. Figure 9 The present invention outperforms similar methods and has better prediction performance.

[0119] To verify the application effect of the present application in practical cases, three IncRNAs, NEAT1, GHET1, and HOTAIRM1, are used as examples to further prove the prediction performance and show the application of our method. This case study mainly focuses on the prediction of the functional associations of the three IncRNAs established. The top 20 functions of NEAT1 predicted by the model are shown in Table 1. This study predicted 48 GO terms, successfully predicting 37 functional associations of NEAT1, of which there are 38 GO term annotations of NEAT1. The top 20 functions of GHET1 predicted are shown in Table 2. This study predicted 34 GO terms, successfully predicting 21 functional associations of GHET1, of which there are 23 GO terms of this IncRNA. The top 20 functions of HOTAIRM1 predicted are shown in Table 3. This study predicted 22 GO terms, successfully predicting 10 functional associations of HOTAIRM1, of which there are 10 GO terms of this IncRNA.

[0120] Table 1 Verification of the top 20 GOs predicted by the present application to be associated with NEAT1

[0121]

[0122]

[0123] Table 2 Verification of the top 20 GOs predicted by the present application to be associated with GHET1

[0124]

[0125] Table 3 Verification of the top 20 GOs predicted by the present application to be associated with HOTAIRM1

[0126]

[0127] Specific embodiment two:

[0129] The present embodiment is a computer storage medium, wherein at least one instruction is stored in the storage medium, and the at least one instruction is loaded and executed by a processor to realize the IncRNA function prediction method based on graph attention.

[0130] It should be understood that the instructions include computer program products, software, or computerized methods corresponding to any method described in the present invention; the instructions can be used to program a computer system or other electronic device. Computer storage media may include readable media on which instructions are stored, and may include but are not limited to magnetic storage media, optical storage media; magneto-optical storage media include read-only memory ROM, random access memory RAM, erasable programmable memory (e.g., EPROM and EEPROM) and flash memory layers, or other types of media suitable for storing electronic instructions. Specific implementation method three:

[0132] This embodiment is a device for predicting lncRNA functions based on graph attention, which includes a processor and a memory. It should be understood that the device includes any device including a processor and a memory described in the present invention, and may also include other units and modules that perform display, interaction, processing, control, and other functions through signals or instructions;

[0133] The memory stores at least one instruction, and the at least one instruction is loaded and executed by the processor to implement the lncRNA function prediction method based on graph attention.

[0134] Those skilled in the art will appreciate that at least one instruction stored is a computer program product corresponding to the method or system. Therefore, the present application may take the form of a complete hardware embodiment, a complete software embodiment, or an embodiment combining software and hardware. Moreover, the present application may take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code. The solutions in the embodiments of the present application can be implemented in various computer languages, for example, object-oriented programming language Java and interpreted scripting language JavaScript, etc.

[0135] The present application is described with reference to the flowcharts and / or block diagrams of the methods, systems, and computer program products according to the embodiments of the present application, and can also be used for corresponding devices. It should be understood that each process and / or box in the flowchart and / or block diagram, as well as the combination of the processes and / or boxes in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowchart and / or block diagram. Figure One a process or multiple processes and / or boxes Figure One A device that provides the functions specified in a block or multiple blocks.

[0136] These computer program instructions can also be stored in a computer readable memory that can direct a computer or other programmable data processing apparatus to function in a particular manner, such that the instructions stored in the computer readable memory produce an article of manufacture including instructions which implement the flow Figure One one or more flowcharts and / or blocks Figure One one or more blocks or steps.

[0137] These computer program instructions can also be loaded onto a computer or other programmable data processing apparatus to cause a series of operational steps to be performed on the computer or other programmable apparatus to produce a computer implemented process such that the instructions which execute on the computer or other programmable apparatus provide steps for implementing the functions specified in the flow Figure One one or more flowcharts and / or blocks Figure One one or more blocks or steps.

[0138] Although preferred embodiments of the application have been described, those skilled in the art will recognize that additional modifications and variations are possible in light of the above teachings. It is therefore intended that the appended claims cover all such modifications and changes as fall within the scope of the application.

[0139] Obviously, numerous modifications and variations of the present application are possible in light of the above teachings. It is therefore intended that the present application be covered by all such modifications and changes as fall within the scope of the claims and their equivalents.

[0140] The above examples of the present application are merely illustrative of the computing model and the computing process of the present application, and are not intended to limit the embodiments of the present application. Based on the above description, other different forms of changes or modifications can be made by those skilled in the art, and it is impossible to exhaust all the embodiments herein. Any obvious changes or modifications derived from the technical solutions of the present application are still within the protection scope of the present application.

Claims

1. A lncRNA function prediction method based on graph attention, characterized in that: The following steps are involved: S1. The similarity matrix S of the three lncRNAs was obtained based on the co-expression of lncRNA, lncRNA-protein interaction, and lncRNA-miRNA interaction. L 、S P and S M ; The embedded representation z of lncRNA is obtained by GCN model and graph contrast learning method based on cross attention mechanism respectively l 、z m and z p , based on z l 、z m and z p Get the lncRNA comprehensive similarity matrix M l ; S2. Based on the GO semantic similarity calculation method in the GOGCN method, a knowledge graph embedding model is designed to learn the representation of GO terms; the semantic similarity between terms is calculated using the representation of the terms to obtain the GO similarity matrix M g ; S3. Construct a training set and lncRNA-GO association matrix based on prior lncRNA-GO functional annotation data; S4. Using hybrid graph representation learning combining GCN and GAT to learn lncRNA comprehensive similarity matrix M l , GO similarity matrix M g Processing is performed to obtain lncRNA-GO fusion features; the obtained lncRNA-GO fusion features are input into the KAN classifier to obtain the predicted lncRNA-GO functional annotations; The specific process includes the following steps: S41, lncRNA comprehensive similarity matrix M l , GO similarity matrix M g Input them into GCN to get features H l (1) and H g (1) ; S42, H l (1) and lncRNA similarity matrix M l Input into GAT to obtain the feature H optimized by the attention mechanism l (2) ; H g (1) and GO similarity matrix M g Input into GAT to obtain the feature H optimized by the attention mechanism g (2) ; S43, H l (2) and M l Input into GCN to obtain the re-extracted feature H of the structural information after attention mechanism optimization l (3) ; H g (2) and M g Input into GCN to obtain the re-extracted feature H of the structural information after attention mechanism optimization g (3) ; S44, H l (1) and H l (3) Input into CNN to get fusion feature X l , H g (1) and H g (3) Input into CNN to get fusion feature X g ; According to the fusion feature X l and X g Get the predicted preference matrix U=X l T X g ; The binary cross entropy loss function is used to measure the difference between U and the lncRNA-GO association matrix obtained in step S3; the final lncRNA feature representation X is obtained by minimizing the binary cross entropy loss function on the training set obtained in step S3. l and GO’s feature representation X g ; S45, the lncRNA feature X obtained above l and GO feature X g The lncRNA-GO fusion features were concatenated and input into the KAN classifier to obtain the predicted lncRNA-GO functional annotations.

2. The lncRNA function prediction method based on graph attention according to claim 1, characterized in that: The similarity matrix S of the three lncRNAs is obtained according to the co-expression of lncRNA, lncRNA-protein interaction, and lncRNA-miRNA interaction described in step S1. L 、S P and S M The process includes the following steps: S11. Obtain lncRNA co-expression similarity data: extract lncRNA expression profiles and calculate the Pearson correlation coefficient between each pair of lncRNA expression profiles to represent the lncRNA co-expression similarity S L ; S12. Obtain similarity data of lncRNA-miRNA interactions: Obtain a dataset of lncRNA-miRNA interaction data. After obtaining the association of each pair of lncRNA-miRNAs based on whether there is an association in the dataset, calculate the Gaussian kernel similarity matrix between lncRNA and miRNA to represent the similarity S of the lncRNA-miRNA interaction. M ; S13. Obtain similarity data of lncRNA-protein interactions: Obtain a dataset of lncRNA-protein interactions. After obtaining the association of each pair of lncRNA-proteins based on whether a lncRNA-protein pair exists in the interaction dataset, calculate the Gaussian kernel similarity matrix between lncRNAs and proteins to represent the similarity S of the lncRNA-protein interaction. P .

3. The lncRNA function prediction method based on graph attention according to claim 1, characterized in that: The embedded representation z of lncRNA is obtained by the GCN model and the graph contrast learning method based on the cross attention mechanism as described in step S1. l 、z m 、z p The process includes the following steps: S14. lncRNA co-expression similarity matrix S L , similarity matrix S of lncRNA and protein interactions P , similarity matrix S of lncRNA and miRNA interactions M Input them into the GCN model respectively, and the corresponding GCN generates the corresponding feature representation y l 、y p 、y m ; Then, the cross attention calculation is performed on the output features of GCN, that is, the cross attention calculation is performed between the three features; l For example, first, l Applied with y m and y p The cross attention mechanism is used to input y m and y p The information in y is integrated into l In the above example, a new feature y′ is generated. l ; The same method was used to obtain the characteristic y′ of lncRNA-miRNA m , lncRNA-protein characteristics y′ p ; S15. After obtaining three features through the cross-attention mechanism, the feature pairs of the same node in different views are regarded as positive sample pairs to obtain the feature y′ l and feature y′ m Positive sample pairs, features y′ l and feature y′ p Positive sample pair, feature y′ m and feature y′ p There are three positive sample pairs; negative sample pairs are feature pairs between different nodes, according to y′ l , y′ m , y′ p The corresponding positive and negative sample pairs are used to obtain the lncRNA co-expression feature z using graph contrast learning. l , characteristics of lncRNA-miRNA m , lncRNA-protein characteristics p .

4. The lncRNA function prediction method based on graph attention according to claim 3, characterized in that: The lncRNA co-expression feature z is obtained by graph contrast learning. l In the process, the contrast loss used is as follows: Among them, 1 [k≠i] Represents the indicator function. When k≠i, the value of the indicator function is 1, which is used to calculate the similarity between node i and node k. When k=i, the value of the indicator function is 0, that is, the similarity between the same nodes is not calculated; y′ k represents y l N represents the number of lncRNAs; τ represents the temperature parameter used to adjust the distribution smoothness of the similarity calculation; sim(·) represents the similarity function between two features.

5. The lncRNA function prediction method based on graph attention according to claim 1, characterized in that: Step S1 is based on z l 、z m and z p The process of obtaining the lncRNA comprehensive similarity matrix includes the following steps: S16. Obtained the characteristics of lncRNA co-expression l , characteristics of lncRNA-miRNA m , lncRNA-protein characteristics p After feature fusion, the three features are concatenated into a feature representation Z = [z l ||z m ||z p ]; S17. Based on the lncRNA feature representation Z, calculate the cosine similarity of the feature representation Z of each two lncRNAs to obtain the lncRNA comprehensive similarity matrix M l .

6. The lncRNA function prediction method based on graph attention according to claim 1, characterized in that: The process of designing a knowledge graph embedding model to learn the representation of GO terms in step S2 includes the following steps: S21, obtain GO file; S22. Represent the GO graph as G = (V, R, E, X, Z), where V is the entity set, representing GO terms; R is the relationship set; E is the edge set; X is the initialized entity representation, and Z is the initialized relationship representation; S23. To obtain the structural information of the GO graph, the encoder adds the propagation weight of the GCN layer on the basis of COMPGCN, and uses the inverse edge and relationship to represent the extended edge and relationship; In S24, in the decoder, the GO graph is treated as a triple, and the corresponding representation of the triple is used to train the link prediction model; in the training phase, a 1-n scoring strategy is adopted to generate corrupted triplets by destroying the head term or the tail term. ConvE is used as the scoring function to provide higher scores for true triplets and lower scores for corrupted triplets. The model converges to obtain the embedded representation of the GO term; S25. After training, the cosine similarity is used to estimate the semantic similarity between entities for the representation of GO terms, and then the GO similarity matrix M is obtained. g .

7. The lncRNA function prediction method based on graph attention according to any one of claims 1 to 6, characterized in that: The process of constructing a training set based on prior lncRNA-GO functional annotation data described in S3 includes the following steps: First, obtain the lncRNA functional annotation file predicted by the lnc-GFP method, and use the lncRNA-GO functional annotation pairs in the predicted lncRNA functional annotation file as positive samples; S32. Randomly construct lncRNA-GO non-functional annotation pairs with the same number as lncRNA-GO functional annotation pairs as negative samples; lncRNA-GO functional annotation pairs and non-functional annotation pairs with the same number as lncRNA-GO functional annotation pairs form a training set.

8. The lncRNA function prediction method based on graph attention according to claim 7, characterized in that: The process of constructing the lncRNA-GO association matrix based on prior lncRNA-GO functional annotation data described in S3 includes the following steps: According to the predicted lncRNA functional annotation file, the lncRNA-GO association matrix was constructed, where 1 represents a lncRNA-GO functional annotation pair and 0 represents a lncRNA-GO non-functional annotation pair.

9. A computer storage medium, characterized in that The storage medium stores at least one instruction, and the at least one instruction is loaded and executed by the processor to implement the lncRNA function prediction method based on graph attention as described in any one of claims 1 to 8.

10. A lncRNA function prediction device based on graph attention, characterized in that: The device includes a processor and a memory, wherein the memory stores at least one instruction, and the at least one instruction is loaded and executed by the processor to implement a graph attention-based lncRNA function prediction method according to any one of claims 1 to 8.

Images