CircRNA-disease associated prediction system
By designing a circRNA-disease association prediction system, using the autoencoder and Node2vec classifier to combine circRNA and disease characteristic vectors, the problem of low accuracy of circRNA-disease association prediction in the prior art is solved, and the accuracy of prediction and data utilization are improved efficiency.
Patent Information
- Application Number
- CN202410699812.3
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-05-31
- Publication Date
- 2024-11-08
AI Technical Summary
Existing circRNA-disease association prediction methods are difficult to mine deep information and cannot combine multiple data sources, resulting in low prediction accuracy.
Design a circRNA-disease association prediction system, including a circRNA-disease association network construction module, a disease semantic similarity score acquisition module, a circRNA functional similarity value matrix acquisition module, an integrated similarity value acquisition module, a positive and negative sample set acquisition module, The positive sample feature set acquisition module, the negative sample feature set acquisition module, and the circRNA-disease association prediction model acquisition module. The combination of the autoencoder and the Node2vec classifier was used to calculate the circRNA-circRNA similarity value and the semantic similarity value between diseases, and construct the circRNA and disease eigenvectors, integrate eigenvectors for prediction.
It improves the accuracy of data utilization and correlation calculation, effectively reduces the agnostic bias in prediction, and improves the accuracy of prediction of the association relationship between disease and circRNA.
Smart Images

Figure CN118918960A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of bioinformatics, and in particular to a circRNA-disease association prediction system. Background Art
[0002] Circular RNA (circRNA) is an endogenous non-coding RNA molecule with a stable circular conformation. In previous studies, it was considered to be noise generated during transcription. Now it is found that circRNA is involved in all aspects of the cell life cycle, including transcription, cell differentiation, cell transport, cell apoptosis, metabolic processes, etc. In addition, circRNA has also been found to be closely related to various human diseases including leukemia, diabetes, prostate cancer, lung cancer, colon cancer, cardiovascular disease, etc. Therefore, the prediction of the correlation between circRNA and disease can guide the prevention of major diseases.
[0003] Due to the complex relationship between circRNA and disease, conducting biological experiments related to circRNA costs a lot of money and time, and computer-assisted experiments have become an effective research method. Nowadays, high-throughput technology and artificial intelligence are developing rapidly, and a large number of biological tools and computer algorithms have been proposed to explore the relationship between circRNA and disease. The establishment of many related databases also provides a necessary prerequisite for computer algorithms to more accurately analyze the association relationship.
[0004] With the development of computing technology, a variety of methods for predicting the association between circRNA and diseases have emerged. Researchers have combined traditional machine learning methods, such as support vector machines, random forests, matrix decomposition and random walks, to build many effective computational models. However, these methods are difficult to mine deep information, cannot combine multiple data sources and cannot prevent overfitting, resulting in low accuracy in circRNA-disease association prediction. Summary of the invention
[0005] The purpose of the present invention is to solve the problem of low prediction accuracy in existing circRNA-disease association prediction, and propose a circRNA-disease association prediction system.
[0006] A circRNA-disease association prediction system, comprising: a circRNA-disease association network construction module, a disease semantic similarity score acquisition module, a circRNA functional similarity value matrix acquisition module, an integrated similarity value acquisition module, a positive and negative sample set acquisition module, a positive sample feature set acquisition module, a negative sample feature set acquisition module, a circRNA-disease association relationship prediction model acquisition module, and a circRNA-disease association relationship prediction module;
[0007] The circRNA-disease association network construction module is used to construct a circRNA-disease association matrix R according to known circRNA-disease association relationships, and send the diseases in the circRNA-disease association matrix R to the disease semantic similarity score acquisition module, send the circRNAs in the circRNA-disease association matrix R to the circRNA functional similarity value matrix acquisition module, and send the circRNA-disease association matrix R to the positive sample feature set acquisition module and the circRNA-disease association relationship prediction model acquisition module;
[0008] The disease semantic similarity score acquisition module is used to obtain the semantic similarity scores between diseases, and send the semantic similarity score matrix between diseases to the integrated similarity value acquisition module;
[0009] The circRNA function similarity value matrix acquisition module is used to obtain the function similarity value between circRNAs, and send the function similarity value matrix between circRNAs to the integrated similarity value acquisition module;
[0010] The integrated similarity value acquisition module uses the semantic similarity scores between diseases to acquire the integrated similarity values between diseases, uses the functional similarity values between circRNAs to acquire the integrated similarity values between circRNAs, and sends the integrated similarity values between diseases and the integrated similarity values between circRNAs to the positive and negative sample set acquisition module;
[0011] The positive and negative sample set acquisition module constructs a disease feature vector using the integrated similarity values between diseases, constructs a circRNA feature vector using the integrated similarity values between circRNAs, and combines the disease feature vector and the circRNA feature vector into a circRNA-disease pair sample, sets a label of 0 or 1 for the circRNA-disease pair sample, and the circRNA-disease pairs with the label of 1 constitute the positive sample set, and the circRNA-disease pairs with the label of 0 constitute the negative sample set, and sends the positive sample set to the positive sample feature set acquisition module, and the negative sample set to the negative sample feature set acquisition module;
[0012] The positive sample feature set acquisition module samples the autoencoder, uses the positive sample set and the circRNA-disease association matrix R as a training set, uses the training set to train the autoencoder, obtains the trained autoencoder and positive sample feature set, and sends the trained autoencoder parameters to the negative sample feature set acquisition module, and sends the positive sample feature set to the circRNA-disease association relationship prediction model acquisition module;
[0013] The negative sample feature set acquisition module deploys an autoencoder and updates parameters according to the trained autoencoder parameters obtained by the positive sample feature set acquisition module, inputs the negative sample set into the trained autoencoder to obtain a negative sample feature set, and sends the negative sample feature set to the circRNA-disease association prediction module;
[0014] The circRNA-disease association prediction model acquisition module trains a Node2vec classifier using a positive sample feature set and a circRNA-disease association matrix R, uses the trained Node2vec classifier as a circRNA-disease association prediction model, and sends the circRNA-disease association prediction model parameters to the circRNA-disease association prediction module;
[0015] A Node2vec classifier is deployed on the circRNA-disease association prediction module, and the parameters of the Node2vec classifier are synchronously updated according to the parameters of the circRNA-disease association prediction model; the circRNA-disease association prediction module is used to obtain the circRNA-disease pairs to be predicted, and use the negative sample feature set to obtain the features of the circRNA-disease pairs to be predicted, and input the features of the circRNA-disease pairs to be predicted into the circRNA-disease association prediction model to obtain the association probability of the circRNA-disease pairs to be predicted.
[0016] Furthermore, the circRNA-disease association network construction module is used to construct a circRNA-disease association matrix R according to known circRNA-disease association relationships, specifically:
[0017] First, human circRNAs and diseases were obtained from the circR2Disease database, and the circRNA-disease association matrix R was constructed with circRNAs as rows and diseases as columns;
[0018] Then, the circRNA-disease pair association probabilities stored in the circR2Disease database are stored in the corresponding positions of the circRNA-disease association matrix R, and spaces are stored in other positions.
[0019] Furthermore, the disease semantic similarity score acquisition module is used to obtain the semantic similarity scores between diseases, specifically:
[0020]
[0021] Among them, SD(d j ,d j' ) is a disease j With disease j' The semantic similarity score of is the disease d in the disease map j The ancestral disease set, is the disease d in the disease map j' The ancestral disease set, is the disease d in the disease map j and diseases j' The intersection of the ancestral disease sets of , t is any disease, Is disease t to disease d j The semantic value of Is t for disease d j' The semantic value of , μ is the semantic contribution, Is disease t to disease d x The semantic value of N d Pick or or d x Take d j or j' , d' is d x Sub-diseases of the ancestral disease.
[0022] Furthermore, the circRNA functional similarity value matrix acquisition module is used to obtain the functional similarity values between circRNAs, specifically:
[0023]
[0024] Among them, FS(c i ,c i' ) is circRNAc i circRNAc i' The functional similarity value, D i is with c i Related disease groups, D i' is with c i' Related disease groups, q is related to c i The disease number associated with c i' Related disease number, d q Disease group D i Diseases in r Disease group D i' Diseases in the q ,D i' ) is a disease q Disease group D i' The similarity value between r ,D i ) is a disease r Disease group D i The similarity value between |D i| is disease group D i The total number of diseases in |D i' | is disease group D i' Total number of diseases in the q ,d r ) is a disease q With disease r The similarity between .
[0025] Furthermore, the semantic similarity scores between diseases are used to obtain the integrated similarity values between diseases, specifically:
[0026]
[0027] Among them, D sim (d j ,d j' ) is a disease j With disease j' The integrated similarity value, DGK(d j ,d j' ) is a disease j With disease j' The GIP kernel similarity value between them.
[0028] Furthermore, the disease j With disease j' The GIP kernel similarity value DGK(d j ,d j' ), specifically:
[0029] DGK(d j ,d j' )=exp(-λ'||R(d j )-R(d j' )|| 2 )
[0030]
[0031] Among them, λ' is the regularization parameter, R(d j ) is the jth column in R, R(d j' ) is the j'th column in R, N d is the total number of diseases.
[0032] Furthermore, the integrated similarity value between circRNAs is obtained by using the functional similarity value between circRNAs, specifically:
[0033]
[0034] Among them, C sim (c i ,ci' ) is circRNAc i With c i' The integrated similarity value of CGK(c i ,c i' ) is circRNAc i With circRNAc i' The GIP kernel similarity value between them.
[0035] Furthermore, circRNAc i With circRNAc i' The similarity value of GIP core between CGK(c i ,c i' ), specifically:
[0036] CGK(c i ,c i' )=exp(-λ||R(c i )-R(c i' )|| 2 )
[0037]
[0038] Where λ is the kernel broadband control parameter, R(c i ) is the i-th row in the matrix R, R(c i' ) is the i'th row in the matrix R, N c is the total number of circRNAs.
[0039] Furthermore, the positive and negative sample set acquisition module constructs a disease feature vector using the integrated similarity values between diseases, constructs a circRNA feature vector using the integrated similarity values between circRNAs, and combines the disease feature vector and the circRNA feature vector into a circRNA-disease pair sample, sets a label of 0 or 1 for the circRNA-disease pair sample, and the circRNA-disease pairs with a label of 1 constitute a positive sample set, and the circRNA-disease pairs with a label of 0 constitute a negative sample set, specifically:
[0040] B1. Use the integrated similarity values between diseases to construct disease feature vectors, and use the integrated similarity values between circRNAs to construct circRNA feature vectors;
[0041] The disease feature vector is the disease d j A vector of integrated similarity values with all other diseases;
[0042] The circRNA feature vector is circRNAc i A vector consisting of the integrated similarity values with all other circRNAs;
[0043] B2. According to the circRNA-disease pairs stored in R, the disease feature vector is combined with the corresponding circRNA feature vector to obtain circRNA-disease pair samples;
[0044] B3. Set the sample labels of circRNA-disease pairs corresponding to the positions with probability values in R to 1, and set the sample labels of circRNA-disease pairs corresponding to the blank positions in R to 0. The sample set of circRNA-disease pairs with label 1 is taken as the positive sample set, and the sample set of circRNA-disease pairs with label 0 is taken as the negative sample set.
[0045] Further, the autoencoder is an autoencoder comprising two hidden layers;
[0046] Use binary cross entropy as the loss function to train the autoencoder;
[0047] The hidden layer of the autoencoder outputs positive sample features, and all positive sample features constitute a positive sample feature set.
[0048] The beneficial effects of the present invention are:
[0049] The present invention first calculates the circRNA-circRNA similarity value and the semantic similarity value between diseases, thereby obtaining the integrated similarity value between circRNAs and the integrated similarity value between diseases, and then constructs circRNA and disease feature vectors respectively through the integrated similarity value between circRNAs and the integrated similarity value between diseases, integrates these feature vectors to construct circRNA-disease pair samples, and then assigns labels to the circRNA-disease pair samples to obtain training samples, inputs the training samples into a deep autoencoder to generate abstract features, and finally uses the generated abstract features and Node2vec classifier to infer the association probability of each disease with unknown probability and circRNA. The present invention gives full play to the role of limited data in the initial circRNA-disease association matrix R, improves the utilization of data and the accuracy of correlation calculation, and the present invention uses the autoencoder and Node2vec classifier to combine and finally infer the circRNA-disease association probability, effectively reducing the unknown deviation in the prediction, and further improving the accuracy of the prediction of the association relationship between the disease and circRNA. BRIEF DESCRIPTION OF THE DRAWINGS
[0050] Figure 1 It is the overall flow chart of the present invention;
[0051] Figure 2 Detailed flow chart of the present invention;
[0052] Figure 3 Schematic diagram of a deep autoencoder;
[0053] Figure 4 This is the ROC graph obtained by the model in the five-fold cross validation experiment;
[0054] Figure 5 This is the ROC graph obtained by the model in the ten-fold cross validation experiment. DETAILED DESCRIPTION
[0055] Specific implementation method 1: Figure 1-2 As shown, the present embodiment is a circRNA-disease association prediction system, comprising: a circRNA-disease association network construction module, a disease semantic similarity score acquisition module, a circRNA functional similarity value matrix acquisition module, an integrated similarity value acquisition module, a positive and negative sample set acquisition module, a positive sample feature set acquisition module, a negative sample feature set acquisition module, a circRNA-disease association relationship prediction model acquisition module, and a circRNA-disease association relationship prediction module;
[0056] The circRNA-disease association matrix construction module is used to construct a circRNA-disease association matrix R according to known circRNA-disease association relationships, and send the diseases in the circRNA-disease association matrix R to the disease semantic similarity score acquisition module, send the circRNAs in the circRNA-disease association matrix to the circRNA functional similarity value matrix acquisition module, and send the circRNA-disease association matrix R to the positive sample feature set acquisition module and the circRNA-disease association relationship prediction model acquisition module;
[0057] The disease semantic similarity score acquisition module is used to obtain the semantic similarity scores between diseases, and send the semantic similarity score matrix between diseases to the integrated similarity value acquisition module;
[0058] The circRNA function similarity value matrix acquisition module is used to obtain the function similarity value between circRNAs, and send the function similarity value matrix between circRNAs to the integrated similarity value acquisition module;
[0059] The integrated similarity value acquisition module uses the semantic similarity scores between diseases to acquire the integrated similarity values between diseases, uses the functional similarity values between circRNAs to acquire the integrated similarity values between circRNAs, and sends the integrated similarity values between diseases and the integrated similarity values between circRNAs to the positive and negative sample set acquisition module;
[0060] The positive and negative sample set acquisition module constructs a disease feature vector using the integrated similarity values between diseases, constructs a circRNA feature vector using the integrated similarity values between circRNAs, and combines the disease feature vector and the circRNA feature vector into a circRNA-disease pair sample, sets a label of 0 or 1 for the circRNA-disease pair sample, and the circRNA-disease pairs with the label of 1 constitute the positive sample set, and the circRNA-disease pairs with the label of 0 constitute the negative sample set, and sends the positive sample set to the positive sample feature set acquisition module, and the negative sample set to the negative sample feature set acquisition module;
[0061] The positive sample feature set acquisition module uses the positive sample set and the circRNA-disease association matrix R as a training set, uses the training set to train the autoencoder, obtains the trained autoencoder and positive sample feature set, and sends the trained autoencoder parameters to the negative sample feature set acquisition module, and the positive sample feature set is sent to the circRNA-disease association relationship prediction model acquisition module;
[0062] The hidden layer of the autoencoder outputs positive sample features; the autoencoder used in the present invention can be: VAE, CAE, SDA;
[0063] The negative sample feature set acquisition module deploys an autoencoder and updates parameters according to the trained autoencoder parameters obtained by the positive sample feature set acquisition module, inputs the negative sample set into the trained autoencoder to obtain a negative sample feature set, and sends the negative sample feature set to the circRNA-disease association prediction module;
[0064] The circRNA-disease association prediction model acquisition module trains a Node2vec classifier using a positive sample feature set and a circRNA-disease association matrix R, uses the trained Node2vec classifier as a circRNA-disease association prediction model, and sends the circRNA-disease association prediction model parameters to the circRNA-disease association prediction module;
[0065] A Node2vec classifier is deployed on the circRNA-disease association prediction module, and the parameters of the Node2vec classifier are synchronously updated according to the parameters of the circRNA-disease association prediction model; the circRNA-disease association prediction module is used to obtain the circRNA-disease pair to be predicted, and obtain the features of the circRNA-disease pair to be predicted by using the negative sample feature set, and input the features of the circRNA-disease pair to be predicted into the circRNA-disease association prediction model to obtain the association probability of the circRNA-disease pair to be predicted;
[0066] The method of using negative sample features to obtain the features of the circRNA-disease pair to be predicted is as follows:
[0067] First, the circRNA-disease pair features to be predicted are searched in the negative sample set. If the circRNA-disease pair features to be predicted are found, the circRNA-disease pair features to be predicted are directly input into the circRNA-disease association prediction model; if the circRNA-disease pair features to be predicted are not found, then:
[0068] The features of the circRNA-disease pairs to be predicted are added to the circRNA-disease association matrix R to update the matrix R, and the disease semantic similarity score acquisition module, the circRNA functional similarity value matrix acquisition module, and the integrated similarity value are sequentially executed using the matrix R to obtain the integrated similarity values between diseases and between circRNAs;
[0069] The integrated similarity values between diseases and the integrated similarity values between circRNAs are used to obtain the feature vector of the circRNA to be tested and the feature vector of the disease to be tested, respectively, so as to obtain the circRNA-disease pair samples to be tested, and at the same time, the circRNA-disease pair samples to be tested are added to the negative sample set and the negative sample set is updated;
[0070] Execute the negative sample feature set acquisition module and input the circRNA-disease pair samples to be tested into the trained autoencoder to obtain the features of the circRNA-disease pair to be tested.
[0071] Specific implementation method 2: The circRNA-disease association matrix construction module is used to construct a circRNA-disease association matrix R according to known circRNA-disease association relationships, specifically:
[0072] Experimentally confirmed circRNA-disease associations were obtained from the CircR2Disease database, which contains 739 verified relationships between 661 circRNAs and 100 diseases. The redundant data whose association restrictions were not human were deleted, and finally 650 associations between 585 circRNAs and 88 diseases were obtained. If the association probability of the circRNA-disease pair is known, it is stored in R in order. If the association probability of the circRNA-disease pair is unknown, the current position in R is a space.
[0073] Specific implementation method three: The disease semantic similarity score acquisition module is used to obtain the semantic similarity scores between diseases, specifically:
[0074] The present invention uses the "doSim" function in the DOSE software package to calculate the semantic similarity score between two diseases. The obtained similarity score ranges from 0 to 1, and the higher the value, the greater the similarity between the diseases:
[0075]
[0076] Among them, SD(d j ,d j' ) is a disease j With disease j' The semantic similarity score, SD(d j ,d j' ) ranges from 0 to 1. The larger the value, the greater the semantic similarity between diseases. is the disease d in the disease map j The ancestral disease set, is the disease d in the disease map j' The ancestral disease set, is the disease d in the disease map j and diseases j' The intersection of the ancestral disease sets of , t is any disease, Is disease t to disease d j The semantic value of Is t for disease d j' The semantic value of , μ = 0.5 is the semantic contribution, Is disease t to disease d x The semantic value of N d Pick or or d x Take d j or j' , d' is d x sub-diseases of the ancestral disease;
[0077] In this step, in formula (1) According to formula (2), first obtain a disease t, and in the numerator, judge whether the disease t is in If not in the set In The child node disease of each disease in is obtained as d' Since a disease has multiple child nodes, there are multiple Get the largest of them As current disease Similarly obtain If the disease t is in In and are all 1. In the denominator of formula (1), we can judge whether the disease t is in and In order to obtain the molecular and
[0078] Specific implementation method 4: The circRNA functional similarity value matrix acquisition module is used to obtain the functional similarity values between circRNAs, specifically:
[0079]
[0080] Among them, FS(c i ,c i' ) is circRNAc i circRNAc i' The functional similarity value, D i is with c i Related disease groups, D i' is with c i' Related disease groups, q is related to c i The disease number associated with c i' Related disease number, d q Disease group D i Diseases in r Disease group D i' Diseases in the q ,D i' ) is a disease q Disease group D i' The similarity value between r ,D i ) is a disease r Disease group D i The similarity value between |D i | is disease group D i The total number of diseases in |D i' | is disease group D i' Total number of diseases in the q ,d r ) is a disease q With disease r The similarity between q ,d r ) is obtained using formula (1);
[0081] Specific implementation method 5: The integrated similarity value acquisition module uses the semantic similarity between diseases to obtain the integrated similarity value between diseases, and uses the functional similarity value matrix between circRNAs to obtain the integrated similarity value between circRNAs, specifically:
[0082] A1. Obtain the GIP nuclear similarity value CGK (c i ,c i' ):
[0083] CGK(c i ,c i' )=exp(-λ||R(c i )-R(c i' )|| 2 )
[0084]
[0085] Among them, CGK(c i ,c i' ) is circRNAc i circRNAc i' The GIP kernel similarity value between them, λ is the kernel bandwidth control parameter, R(c i ) is the i-th row in the matrix R, R(c i' ) is the i'th row in the matrix R, N c is the total number of circRNAs;
[0086] The semantic similarity score of the disease will be,
[0087] A2. Obtain the GIP kernel similarity value DGK (dGK) between diseases based on the semantic similarity score between diseases j ,d j' ):
[0088] DGK(d j ,d j' )=exp(-λ'||R(d j )-R(d j' )|| 2 )
[0089]
[0090] Among them, DGK(d j ,d j' ) is a disease j With disease j' The GIP kernel similarity value between them, λ' is the regularization parameter, R(d j ) is the jth column in the matrix R, R(d j' ) is the j'th column in the matrix R, N d is the total number of diseases;
[0091] A3, using A1 to obtain the GIP nuclear similarity value CGK (ci ,c i' ) and the functional similarity between circRNAs to obtain the integrated similarity between circRNAs, and the GIP nuclear similarity between diseases obtained by A2 (DGK) j ,d j' ) and the semantic similarity score between diseases to obtain the integrated similarity value between diseases, specifically:
[0092]
[0093] Among them, C sim (c i ,c i' ) is circRNAc i With c i' The integrated similarity value, D sim (d j ,d j' ) is a disease j With disease j' The integrated similarity value of .
[0094] Specific implementation method six: The positive and negative sample set acquisition module uses the integrated similarity values between diseases to construct a disease feature vector, uses the integrated similarity values between circRNAs to construct a circRNA feature vector, and combines the disease feature vector and the circRNA feature vector into a circRNA-disease pair sample, sets a label of 0 or 1 for the circRNA-disease pair sample, and the circRNA-disease pairs with a label of 1 constitute a positive sample set, and the circRNA-disease pairs with a label of 0 constitute a negative sample set, specifically:
[0095] B1. Use the integrated similarity values between diseases to construct disease feature vectors, and use the integrated similarity values between circRNAs to construct circRNA feature vectors;
[0096] The disease feature vector is the disease d j A vector of integrated similarity values with all other diseases;
[0097] The circRNA feature vector is circRNAc i A vector consisting of the integrated similarity values with all other circRNAs;
[0098] B2. According to the circRNA-disease pairs in the matrix R, the disease feature vector is combined with the corresponding circRNA feature vector to obtain circRNA-disease pair samples;
[0099] B3. Set the sample labels of circRNA-disease pairs with known association probabilities to 1, and set the sample labels of circRNA-disease pairs with unknown association probabilities to 0. The sample set of circRNA-disease pairs with label 1 is taken as the positive sample set, and the sample set of circRNA-disease pairs with label 0 is taken as the negative sample set.
[0100] Since there is coexistence N c ×N d Sample features, each sample has a corresponding label. If there is a known connection between circRNA and disease, the sample label is 1; otherwise, it is 0; samples with label 1 constitute the positive set, and unverified circRNA-disease associations are randomly selected to generate a negative set, which is equal to the size of the positive set. There is a possibility of unconfirmed positive samples in the generated negative sample set, but the probability is negligible compared to the entire unknown samples in the data set. Finally, 1,300 samples were obtained in the training set, with equal positive and negative samples;
[0101] Specific implementation method seven: the autoencoder is a neural network including two hidden layers, and the number of neurons in the input layer and the output layer is the same;
[0102] The autoencoder learns the identity function as:
[0103] Q'=h w,b (Q)≈Q
[0104] Where h is the hidden layer, w and b are weight and bias variables respectively, Q is the autoencoder input, and Q' is the autoencoder output;
[0105] The autoencoder first encodes the input Q into a compact form as follows:
[0106]
[0107] Where x is a variable;
[0108]
[0109] This process will be iterated until Q ≈ Q ', and the output of the final encoder will represent the nonlinear features of the circRNA-disease pair in a compact, high-level form;
[0110] The encoder dimension of the autoencoder is set to 128, binary cross entropy is used as the loss function, and Adadelta is used as the optimization algorithm;
[0111] Among them, the hidden layer outputs positive sample features, and all positive sample features constitute the positive sample feature set.
[0112] An autoencoder is an artificial neural network that learns implicit patterns in input data in an unsupervised manner. A deep autoencoder with two hidden layers is used to generate a compact nonlinear representation of the features of each circRNA-disease pair in the dataset, such as Figure 3 shown.
[0113] Specific implementation method eight: The circRNA-disease association prediction model acquisition module uses the positive sample feature set and the circRNA-disease association matrix R to train the Node2vec classifier, and uses the trained Node2vec classifier as the circRNA-disease association prediction model. The probability of the Node2vec classifier accessing the next node is as follows:
[0114]
[0115] π vx' =α p'q' (t',x')·w vx'
[0116]
[0117] where π vx' is an unstandardized transition probability, Z is a standardization constant, α p'q' (t',x') is given by d t'x' Determine the transition probability, d t'x' is the distance between node t' and node x', p' is the backoff parameter, q' is the input-output parameter, v is the current node, E is the set of all edges, P(c i'+1 =x'|c i' =v) represents the current node c i 'If it is v, the next node is c i'+1 is the probability of x', w vx' is the weight of the edge (v,x').
[0118] If q'>1, the random walk tends to visit nodes that are close to node t (breadth-first search). If q'<1, the random walk tends to visit nodes that are far away from node t (depth-first search). p' controls the probability of revisiting a node that has just been visited. The larger p' is, the lower the probability of visiting a node that has just been visited, and vice versa.
[0119] Example: In order to verify the beneficial effects of the present invention, the present invention uses a 5-fold and 10-fold cross-validation method to evaluate a circRNA-disease association prediction system of the present invention. The ROC curve image obtained based on the 5-fold and 10-fold cross-validation is as follows: Figure 4 and Figure 5 shown.
Claims
1. A circRNA-disease association prediction system, characterized in that The system comprises: a circRNA-disease association network construction module, a disease semantic similarity score acquisition module, a circRNA functional similarity value matrix acquisition module, an integrated similarity value acquisition module, a positive and negative sample set acquisition module, a positive sample feature set acquisition module, a negative sample feature set acquisition module, a circRNA-disease association relationship prediction model acquisition module, and a circRNA-disease association relationship prediction module; The circRNA-disease association network construction module is used to construct a circRNA-disease association matrix R according to known circRNA-disease association relationships, and send the diseases in the circRNA-disease association matrix R to the disease semantic similarity score acquisition module, send the circRNAs in the circRNA-disease association matrix R to the circRNA functional similarity value matrix acquisition module, and send the circRNA-disease association matrix R to the positive sample feature set acquisition module and the circRNA-disease association relationship prediction model acquisition module; The disease semantic similarity score acquisition module is used to obtain the semantic similarity scores between diseases, and send the semantic similarity score matrix between diseases to the integrated similarity value acquisition module; The circRNA function similarity value matrix acquisition module is used to obtain the function similarity value between circRNAs, and send the function similarity value matrix between circRNAs to the integrated similarity value acquisition module; The integrated similarity value acquisition module uses the semantic similarity scores between diseases to acquire the integrated similarity values between diseases, uses the functional similarity values between circRNAs to acquire the integrated similarity values between circRNAs, and sends the integrated similarity values between diseases and the integrated similarity values between circRNAs to the positive and negative sample set acquisition module; The positive and negative sample set acquisition module constructs a disease feature vector using the integrated similarity values between diseases, constructs a circRNA feature vector using the integrated similarity values between circRNAs, and combines the disease feature vector and the circRNA feature vector into a circRNA-disease pair sample, sets a label of 0 or 1 for the circRNA-disease pair sample, and the circRNA-disease pairs with the label of 1 constitute the positive sample set, and the circRNA-disease pairs with the label of 0 constitute the negative sample set, and sends the positive sample set to the positive sample feature set acquisition module, and the negative sample set to the negative sample feature set acquisition module; The positive sample feature set acquisition module uses an autoencoder, uses the positive sample set and the circRNA-disease association matrix R as a training set, uses the training set to train the autoencoder, obtains the trained autoencoder and positive sample feature set, and sends the trained autoencoder parameters to the negative sample feature set acquisition module, and sends the positive sample feature set to the circRNA-disease association relationship prediction model acquisition module; The negative sample feature set acquisition module deploys an autoencoder and updates parameters according to the trained autoencoder parameters obtained by the positive sample feature set acquisition module, inputs the negative sample set into the trained autoencoder to obtain a negative sample feature set, and sends the negative sample feature set to the circRNA-disease association prediction module; The circRNA-disease association prediction model acquisition module trains a Node2vec classifier using a positive sample feature set and a circRNA-disease association matrix R, uses the trained Node2vec classifier as a circRNA-disease association prediction model, and sends the circRNA-disease association prediction model parameters to the circRNA-disease association prediction module; A Node2vec classifier is deployed on the circRNA-disease association prediction module, and the parameters of the Node2vec classifier are synchronously updated according to the parameters of the circRNA-disease association prediction model; the circRNA-disease association prediction module is used to obtain the circRNA-disease pairs to be predicted, and use the negative sample feature set to obtain the features of the circRNA-disease pairs to be predicted, and input the features of the circRNA-disease pairs to be predicted into the circRNA-disease association prediction model to obtain the association probability of the circRNA-disease pairs to be predicted.
2. A circRNA-disease association prediction system according to claim 1, characterized in that: The circRNA-disease association network construction module is used to construct a circRNA-disease association matrix R according to the known circRNA-disease association relationship, specifically: First, human circRNAs and diseases were obtained from the circR2Disease database, and the circRNA-disease association matrix R was constructed with circRNAs as rows and diseases as columns; Then, the circRNA-disease pair association probabilities stored in the circR2Disease database are stored in the corresponding positions of the circRNA-disease association matrix R, and spaces are stored in other positions.
3. A circRNA-disease association prediction system according to claim 2, characterized in that: The disease semantic similarity score acquisition module is used to obtain the semantic similarity scores between diseases, specifically: Among them, SD(d j ,d j' ) is a disease j With disease j' The semantic similarity score of is the disease d in the disease map j The ancestral disease set, is the disease d in the disease map j' The ancestral disease set, is the disease d in the disease map j and diseases j' The intersection of the ancestral disease sets of , t is any disease, Is disease t to disease d j The semantic value of Is t for disease d j' The semantic value of , μ is the semantic contribution, Is disease t to disease d x The semantic value of N d Pick or or d x Take d j or j' , d' is d x Sub-diseases of the ancestral disease.
4. A circRNA-disease association prediction system according to claim 3, characterized in that: The circRNA functional similarity value matrix acquisition module is used to obtain the functional similarity values between circRNAs, specifically: Among them, FS(c i ,c i' ) is circRNAc i With circRNAc i' The functional similarity value, D i is with c i Related disease groups, D i' is with c i' Related disease groups, q is related to c i The disease number associated with c i' Related disease number, d q Disease group D i Diseases in r Disease group D i' Diseases in q ,D i' ) is a disease q Disease group D i' The similarity value between r ,D i ) is a disease r Disease group D i The similarity value between |D i | is disease group D i The total number of diseases in |D i' | is disease group D i' Total number of diseases in the q ,d r ) is a disease q With disease r The similarity between .
5. A circRNA-disease association prediction system according to claim 4, characterized in that: The method of using the semantic similarity scores between diseases to obtain the integrated similarity values between diseases is specifically: Among them, D sim (d j ,d j' ) is a disease j With disease j' The integrated similarity value, DGK(d j ,d j' ) is a disease j With disease j' The GIP kernel similarity value between them.
6. A circRNA-disease association prediction system according to claim 5, characterized in that: Disease j With disease j' The GIP kernel similarity value DGK(d j ,d j' ), specifically: DGK(d j ,d j' )=exp(-λ'||R(d j )-R(d j' )|| 2 ) Among them, λ' is the regularization parameter, R(d j ) is the jth column in R, R(d j' ) is the j'th column in R, N d is the total number of diseases.
7. A circRNA-disease association prediction system according to claim 6, characterized in that: The method of using the functional similarity values between circRNAs to obtain the integrated similarity values between circRNAs is specifically as follows: Among them, C sim (c i ,c i' ) is circRNAc i With c i' The integrated similarity value of CGK(c i ,c i' ) is circRNAc i With circRNAc i' The GIP kernel similarity value between them.
8. A circRNA-disease association prediction system according to claim 7, characterized in that: circRNAc i With circRNAc i' The similarity value of GIP core between CGK(c i ,c i' ), specifically: CGK(c i ,c i' )=exp(-λ||R(c i )-R(c i' )|| 2 ) Where λ is the kernel broadband control parameter, R(c i ) is the i-th row in the matrix R, R(c i' ) is the i'th row in the matrix R, N c is the total number of circRNAs.
9. A circRNA-disease association prediction system according to claim 8, characterized in that: The positive and negative sample set acquisition module constructs a disease feature vector using the integrated similarity values between diseases, constructs a circRNA feature vector using the integrated similarity values between circRNAs, and combines the disease feature vector and the circRNA feature vector into a circRNA-disease pair sample, sets a label of 0 or 1 for the circRNA-disease pair sample, and the circRNA-disease pairs with a label of 1 constitute a positive sample set, and the circRNA-disease pairs with a label of 0 constitute a negative sample set, specifically: B1. Use the integrated similarity values between diseases to construct disease feature vectors, and use the integrated similarity values between circRNAs to construct circRNA feature vectors; The disease feature vector is the disease d j A vector of integrated similarity values with all other diseases; The circRNA feature vector is circRNAc i A vector consisting of the integrated similarity values with all other circRNAs; B2. According to the circRNA-disease pairs stored in R, the disease feature vector is combined with the corresponding circRNA feature vector to obtain circRNA-disease pair samples; B3. Set the sample labels of circRNA-disease pairs corresponding to the positions with probability values in R to 1, and set the sample labels of circRNA-disease pairs corresponding to the blank positions in R to 0. The sample set of circRNA-disease pairs with label 1 is taken as the positive sample set, and the sample set of circRNA-disease pairs with label 0 is taken as the negative sample set.
10. A circRNA-disease association prediction system according to claim 9, characterized in that: The autoencoder is an autoencoder comprising two hidden layers; Use binary cross entropy as the loss function to train the autoencoder; The hidden layer of the autoencoder outputs positive sample features, and all positive sample features constitute a positive sample feature set.
