Disease correlation prediction method based on twin network

Abstract

The present invention proposes a disease correlation prediction method based on a twin network, which solves the technical problem of a small range of disease correlation calculation in the prior art. The implementation steps are: (1) obtaining a disease data set; (2) constructing a training sample set and a test sample set; (3) constructing symptom characteristics and classification characteristics of the disease; (4) constructing a disease correlation prediction model based on a twin network; (5) iteratively training the disease correlation prediction model C based on the twin network; (6) obtaining the disease correlation prediction result. The present invention significantly expands the calculation range of disease correlation, and can be used to guide the classification of diseases, infer the pathogenesis of diseases, drug repositioning, and the development of new drugs.

Description

Technical Field

[0001] The present invention belongs to the field of bioinformatics and relates to a disease correlation prediction method, specifically to a disease correlation prediction method based on a twin network, which can be used to guide disease classification, infer disease pathogenesis, drug repositioning and new drug research and development. Background Art

[0002] The clinical phenotypes of different diseases vary greatly, but diseases with different clinical phenotypes may also have the same therapeutic target protein or the same pathogenic gene, which provides a new perspective for the study of disease mechanisms and drug development. Exploring the association between diseases helps to better understand the state of human diseases from multiple perspectives and has a certain guiding role in the classification of diseases. In addition, since similar disease pairs have a greater probability of sharing pathogenic genes or being regulated by the same microRNA and lncRNA, studying disease correlations helps to infer disease-related information such as pathogenic genes, microRNAs and lncRNAs. This promotes the design of new drugs, such as drugs that function by targeting pathogenic genes, microRNA and lncRNA drugs, and improves the diagnosis and treatment of diseases. At the same time, since similar diseases may have similar pathogenesis and the effects of drugs on similar diseases may also be similar, discovering disease correlations can also help with drug repositioning research and reduce time and cost in the process of new drug development.

[0003] Existing disease correlation calculation methods can be roughly divided into three categories according to the type of disease data used: methods based on disease semantic information, methods based on disease-related gene functions, and methods based on integrating semantic and related gene function information. Methods based on semantic information calculate the correlation between diseases based on biomedical texts describing diseases, disease ontology, or disease annotation information. Commonly used data include a variety of text data such as literature, case reports, and adverse drug reactions, as well as some databases such as medical subject heading databases and disease ontology databases. Methods based on disease-related gene functions are basically based on the assumption that similar diseases may have the same or similar functional pathogenic genes or gene products. These function-based methods convert the calculation of disease correlation into the calculation of the overlap between two disease-related gene sets, or calculate disease correlation through the adjacency relationship between disease-related gene sets and their corresponding proteins in the protein interaction network and the topological characteristics of the network. Some methods calculate disease correlation based on the overlap of disease-related functional pathways or disease modules. The method based on integrating semantic and related gene function information uses data from two disease types to calculate disease correlation. The integration process is divided into two types: simple integration through mathematical operations such as summation after using the semantic information-based method and the functional information-based method separately, and the joint use of these two types of disease information in the process of calculating disease correlation.

[0004] The disease correlation calculated based on the method of integrating semantics and related gene function information contains two types of disease-related information, which can more comprehensively represent the disease. Although the semantic information of the disease contains limited information about the molecular mechanism of the disease, it is widely available and easy to obtain, so it can be used to calculate the disease correlation on a larger set of diseases; while the functional information of the disease contains rich disease-related molecular mechanisms, but it is limited by experimental research. For example, in 2020, Peng et al. published a paper titled "Constructing Disease Similarity Networks Based on Disease Module Theory" in "IEEE-ACM TRANSACTIONS ON COMPUTATIONAL BIOLOGY AND BIOINFORMATICS", which disclosed a method for calculating disease correlation based on disease module theory. This method uses multiple semantic information of the disease to expand the disease-related gene set, and then uses the proteins corresponding to the gene set to construct the disease module in the protein interaction network, and calculates the disease correlation based on the overlap of the disease module. The calculation results of this method contain rich disease-related information, but due to incomplete disease function information, it limits its application in disease pathogenesis, drug repositioning and new drug development. Summary of the invention

[0005] The purpose of the present invention is to overcome the deficiencies of the above-mentioned prior art and to propose a disease correlation prediction method based on a twin network, which is used to solve the technical problem in the prior art that the disease correlation calculation range is small due to incomplete disease function information.

[0006] To achieve the above object, the technical solution adopted by the present invention includes the following steps:

[0007] (1) Obtain disease dataset:

[0008] Get M symptoms E = {e1, e2, ..., e m ,...,e M} and K disease types Y={y1,y2,...,y k ,...,y K The set D of N diseases is {d1,d2,...,d n ,...,d N}, where M ≥ 300, e m represents the mth symptom, K≥24, y k represents the kth disease type, N ≥ 200, d n represents the nth disease;

[0009] (2) Construct training sample set and test sample set:

[0010] Arrange the correlation scores of every two diseases with gene association in the disease set D in descending order to form a correlation score set A = {A1, A2, ..., A s ,...,A S}, and select the first R association scores as the positive sample set Ω, and the remaining T association scores as the negative sample set Ω - , and then select more than half of the positive samples in Ω, and Ω - The number of negative samples in Ω that is the same as the number of selected positive samples constitutes the training sample set Tr, and the remaining positive samples in Ω and Ω - The negative samples with the same number of positive samples remaining in Ω form the test sample set Te, where R<T, S=R+T;

[0011] (3) Constructing symptom characteristics and classification characteristics of the disease:

[0012] Extract each symptom e in symptom set E m The number of relevant literatures of N diseases is calculated by the number of relevant literatures of M symptoms. DI={g1,g2,...,g n ,...,g N}, and perform one-hot encoding on the disease type set Y to obtain the classification features of N diseases F = {h1,h2,...,h n ,...,h N}, g n Indicates disease n Symptoms and characteristics, g n =(s n,1 ,s n,2 ,...s n,m ,...,s n,M ), s n,m Indicates disease n and symptoms m The correlation score between n For disease n The classification features, h n =(c n,1 ,c n,2 ,...,c n,k ,...,c n,K ), c n,k ∈{0,1},c n,k =1 means d n Belongs to disease category k , c n,k =0 means d n Not y k ;

[0013] (4) Constructing a disease correlation prediction model C based on a twin network:

[0014] Construct a disease correlation prediction model C including a sequentially connected twin network, a vector concatenation layer and a multi-layer perceptron MLP; wherein the twin network includes two sub-networks CS1 and CS2 that share weight and bias parameters, and both sub-networks include multiple fully connected layers; the MLP includes an input layer, a hidden layer and an output layer;

[0015] (5) Iteratively train the disease correlation prediction model C based on the twin network:

[0016] (5a) Initialize the number of iterations j, the maximum number of iterations is J, J ≥ 200, and the current disease correlation prediction model is C j , and let j = 0, C j =C;

[0017] (5b) Diseases d with associated scores in the training sample set Tr are a and d b Symptoms and characteristics a , g b , and the classification feature h a 、h bAs a disease-related prediction model C j The input is forward propagated to obtain the disease pair d a and d b Predicted value of association score

[0018] (5c) Using the loss function L, and passing and the disease d in the training sample set Tr a and d b The correlation score Calculate the model C for the jth iteration j The loss of L j , and then use the back propagation method and pass L j Calculate C j The parameter gradient Finally, the gradient descent algorithm is used to C j The weight parameter Update to obtain the disease correlation prediction model after the jth iteration;

[0019] (5d) Determine whether j ≥ J or the model loss does not decrease in consecutive Q iterations. If so, obtain the trained disease correlation prediction model C′. Otherwise, set j = j + 1 and execute step (4b), where 2 ≤ Q ≤ 5;

[0020] (6) Obtaining disease correlation prediction results:

[0021] The disease d with association score in the test sample set Te is p and d q Symptoms and characteristics p , g q , and the classification feature h p 、h q Input the disease correlation prediction model C′ for forward propagation to obtain the disease correlation d p and d q Predicted value of association score

[0022] Compared with the prior art, the present invention has the following advantages:

[0023] The disease correlation prediction model constructed by the present invention includes a twin network composed of two subnetworks sharing weights and bias parameters. In the process of training the model and obtaining the disease correlation prediction results, the two subnetworks of the twin network respectively perform nonlinear transformations on the symptom characteristics of the two diseases; because the model uses the symptom characteristics and classification characteristics of the disease to predict the disease association score containing the semantic information and functional information of the disease, the parameters of the two subnetworks of the twin network are updated through the back propagation process of the model training, so that the output vector after the nonlinear transformation of the symptom characteristics contains the characteristics of the disease functional information. Compared with the functional information of disease-related genes, the subnetwork output vector of the twin network also includes the symptom characteristics of the disease. Therefore, by calculating the correlation based on the twin network model, compared with the prior art, the limitation of incomplete disease functional information on the calculation of disease correlation is overcome, and the calculation range of disease correlation is expanded. BRIEF DESCRIPTION OF THE DRAWINGS

[0024] Figure 1 It is an implementation flow chart of the present invention. DETAILED DESCRIPTION

[0025] The present invention is further described in detail below in conjunction with the accompanying drawings and specific embodiments. It should be noted that the present invention does not belong to the subject matter that is not granted with a patent right as stipulated in Article 25 of the Patent Law, and also complies with the provisions of Article 2, Paragraph 2 of the Patent Law:

[0026] Reference Figure 1 , this example includes the following steps:

[0027] Step 1, obtain the disease dataset:

[0028] Obtain M symptoms E={e1,e2,...,e m ,...,e M} and K disease types Y={y1,y2,...,y k ,...,y K The set D of N diseases is {d1,d2,...,d n ,...,d N}, where M = 385, e m represents the mth symptom, K = 24, y k represents the kth disease type, N = 4268, d n represents the nth disease;

[0029] Step 2: Construct training sample set and test sample set:

[0030] From the data published in the paper “Uncovering disease-disease relationships through the incomplete interactome” by Menche et al. in Science, the correlation scores of every two diseases with genetic associations in the disease set D are arranged in descending order to form a correlation score set A = {A1, A2, ..., A s ,...,A S}, and select the first R association scores as the positive sample set Ω, and the remaining T association scores as the negative sample set Ω - , and then select more than half of the positive samples in Ω, and Ω - The number of negative samples in Ω that is the same as the number of selected positive samples constitutes the training sample set Tr, and the remaining positive samples in Ω and Ω - The negative samples with the same number of positive samples remaining in Ω form the test sample set Te, where R<T, S=R+T;

[0031] Step 3: Construct symptom characteristics and classification characteristics of the disease:

[0032] E-utilities tools were used to search the literature in PubMed database and count the disease d n and symptoms m The number of documents that can be indexed simultaneously W n,m , and adopt the TF-IDF method, according to W n,m Calculate d n and e m The correlation score between n,m , and get d n Symptoms and characteristics n =(s n,1 ,s n,2 ,...s n,m ,...,s n,M ), then use g n The symptom characteristics of N diseases DI = {g1, g2, ..., g n ,...,g N},in:

[0033]

[0034] Where N m For use symptoms m The disease set D counted by W n,m >μ, μ=0; N m Represents the disease set D and symptom e m Number of associated diseases, Nm The larger the value, the more diseases the symptom is associated with. n The weaker the specific correlation, the smaller the association score between the disease and the symptom through the TF-IDF method;

[0035] At the same time, the disease type set Y is one-hot encoded, and the discrete categorical variables of N diseases are converted into discrete vectors that can distinguish N diseases and can be used by the machine learning model, and the classification features F = {h1, h2, ..., h n ,...,h N}, h n For disease n The classification features, h n =(c n,1 ,c n,2 ,...,c n,k ,...,c n,K ), c n,k ∈{0,1},c n,k =1 means d n Belongs to disease category k , c n,k =0 means d n Not y k ;

[0036] Step 4: Construct a disease correlation prediction model C based on the twin network:

[0037] Step 4a) Construct the structure of the disease correlation prediction model C based on the twin network:

[0038] Construct a disease correlation prediction model C including a sequentially connected twin network, a vector concatenation layer and a multi-layer perceptron MLP; wherein the twin network includes two sub-networks CS1 and CS2 that share weight and bias parameters, and both sub-networks include multiple fully connected layers; the MLP includes an input layer, a hidden layer and an output layer;

[0039] Step 4b) Loss function L of the disease correlation prediction model C based on the twin network:

[0040]

[0041] in For disease a and d b The correlation score, is the value of the MLP output layer, i.e., the prediction model C for disease d a and d b The predicted value of the association score, θ Cis the weight parameter of the disease correlation prediction model C, λ is the coefficient of the L2 regular term, λ=0.001; the loss function L uses the binary cross entropy loss function to calculate d a and d b The difference between the association score predicted by the model and the association score in the sample set is taken into account, and the L2 regular term is added to L to reduce the overfitting of the model during training and enhance the generalization ability of the model;

[0042] Step 5: Iteratively train the disease correlation prediction model C based on the twin network:

[0043] Step 5a) Initialize the number of iterations j, the maximum number of iterations is J, J = 200, and the current disease correlation prediction model is C j , and let j = 0, C j =C;

[0044] Step 5b) The diseases d with associated scores in the training sample set Tr are a and d b Symptoms and characteristics a , g b , and the classification feature h a 、h b As a disease-related prediction model C j The input is forward propagated;

[0045] First, the sub-network CS1 of the twin network inputs the symptom feature g a Perform nonlinear transformation and output vector g′ a , while sub-network CS2 is used to analyze the symptom feature g b Perform nonlinear transformation and output vector g′ b , when predicting the disease relevance that integrates the semantic information and functional information of the disease, the parameters of the two sub-networks are updated through back propagation. At the same time, since the two sub-networks of the twin network have the same structure and share weights, for the input paired disease symptom features, the sub-network of the twin network establishes an association between the disease symptom features and the disease relevance that integrates the semantic information and functional information, so that the output vector g′ a and g′ b Contains information about symptoms and functions of the disease;

[0046] Then the vector concatenation layer concatenates the output vector g′ of the Siamese network a and g′ b , and diseases a The classification feature h a and diseases b The classification feature h b , get the splicing vector H concat =[g′ a ,g′b ,h a ,h b ], [·] is a concatenated activation function. By using the concatenation layer, when using the input vector of the twin network to predict the disease association score containing the disease semantic information and functional information, the classification information of the disease is supplemented to guide the calculation of the disease correlation; the obtained concatenation vector contains the symptom information, functional information and classification information of the disease;

[0047] Finally, the input layer of the multi-layer perceptron MLP receives the concatenated vector H concat , the hidden layer pair H concat Perform multiple nonlinear transformations, and the output vector H of the u-th hidden layer to the u-1 hidden layer u-1 Perform nonlinear transformation and output vector H u , the output vector H of the last hidden layer of the output layer U Perform nonlinear transformation and output disease pairs d a and d b Predicted value of association score U is the number of hidden layers, U=5;

[0048] Step 5c) Use the loss function L and pass and the disease d in the training set Tr a and d b The correlation score Calculate the model C for the jth iteration j The loss of L j , and then use the back propagation method and pass L j Calculate C j The parameter gradient Finally, the gradient descent algorithm is used to C j The weight parameter Update to obtain the disease correlation prediction model of this iteration;

[0049] Step 5d) Determine whether j ≥ J and the model loss does not decrease in consecutive Q iterations. If so, obtain the trained disease correlation prediction model C′. Otherwise, set j = j + 1 and execute step (4b), where Q = 5. By setting the Earlystopping threshold Q, the model can terminate the iteration early during the training process to alleviate the overfitting problem of the model.

[0050] Step 6: Obtain disease correlation prediction results:

[0051] The disease d with association score in the test sample set Te is p and d q Symptoms and characteristics p , gq , and the classification feature h p 、h q Input the disease correlation prediction model C′ for forward propagation to obtain the disease correlation d p and d q Predicted value of association score

[0052] The technical effect of the present invention is further illustrated by simulation experiments as follows:

[0053] 1. Simulation conditions and contents:

[0054] The simulation experiments were carried out on an Intel(R) Core(TM) i7-4710MQ CPU with a main frequency of 2.50 GHz, 8 GB of memory, and Python 3.6.5 combined with TensorFlow 2.1.0.

[0055] The disease correlation ranges of the present invention and the existing disease correlation calculation method based on disease module theory are compared and simulated, and the results are shown in Table 1.

[0056] 2. Analysis of simulation results:

[0057] The comparison results of the disease correlation prediction range calculated by the present invention and the prior art are shown in Table 1.

[0058] Table 1

[0059] method Number of diseases Number of disease associations Prior art 697 14671 The present invention 697 242556

[0060] From Table 1, it can be seen that compared with the prior art, under the premise of the same number of diseases, the number of disease correlations that can be calculated in the present invention is expanded by about 15.5 times, and the prediction range of disease correlations is much larger than that of the prior art, which proves that the method of the present invention effectively improves the calculation range of disease correlations.

Claims

1. A disease correlation prediction method based on twin networks, characterized in that: The steps include: (1) Obtain disease dataset: Get M symptoms E = {e1, e2, ..., e m ,...,e M } and K disease types Y={y1,y2,...,y k ,...,y K The set D of N diseases is {d1,d2,...,d n ,...,d N }, where M ≥ 300, e m represents the mth symptom, K≥24, y k represents the kth disease type, N ≥ 200, d n represents the nth disease; (2) Construct training sample set and test sample set: Arrange the correlation scores of every two diseases with gene association in the disease set D in descending order to form a correlation score set A = {A1, A2, ..., A s ,...,A S }, and select the first R association scores as the positive sample set Ω, and the remaining T association scores as the negative sample set Ω - , and then select more than half of the positive samples in Ω, and Ω - The number of negative samples in Ω that is the same as the number of selected positive samples constitutes the training sample set Tr, and the remaining positive samples in Ω and Ω - The negative samples with the same number of positive samples remaining in Ω form the test sample set Te, where R<T, S=R+T; (3) Constructing symptom characteristics and classification characteristics of the disease: Extract each symptom e in symptom set E m The number of relevant literatures of N diseases is calculated by the number of relevant literatures of M symptoms. DI={g1,g2,...,g n ,...,g N }, and perform one-hot encoding on the disease type set Y to obtain the classification features of N diseases F = {h1,h2,...,h n ,...,h N }, g n Indicates disease n Symptoms and characteristics, g n =(s n,1 ,s n,2 ,...s n,m ,...,s n,M ), s n,m Indicates disease n and symptoms m The correlation score between n For disease n The classification features, h n =(c n,1 ,c n,2 ,...,c n,k ,...,c n,K ), c n,k ∈{0,1},c n,k =1 means d n Belongs to disease category k , c n,k =0 means d n Not y k ; (4) Constructing a disease correlation prediction model C based on a twin network: Construct a disease correlation prediction model C including a sequentially connected twin network, a vector concatenation layer and a multi-layer perceptron MLP; wherein the twin network includes two sub-networks CS1 and CS2 that share weight and bias parameters, and both sub-networks include multiple fully connected layers; the MLP includes an input layer, a hidden layer and an output layer; (5) Iteratively train the disease correlation prediction model C based on the twin network: (5a) Initialize the number of iterations j, the maximum number of iterations is J, J ≥ 200, and the current disease correlation prediction model is C j , and let j = 0, C j =C; (5b) Diseases d with associated scores in the training sample set Tr are a and d b Symptoms and characteristics a , g b , and the classification feature h a 、h b As a disease-related prediction model C j The input is forward propagated to obtain the disease pair d a and d b Predicted value of association score (5c) Using the loss function L, and passing and the disease d in the training sample set Tr a and d b The correlation score Calculate the model C for the jth iteration j The loss of L j , and then use the back propagation method and pass L j Calculate C j The parameter gradient Finally, the gradient descent algorithm is used to C j The weight parameter Update to obtain the disease correlation prediction model after the jth iteration; (5d) Determine whether j ≥ J or the model loss does not decrease in consecutive Q iterations. If so, obtain the trained disease correlation prediction model C′. Otherwise, set j = j + 1 and execute step (4b), where 2 ≤ Q ≤ 5; (6) Obtaining disease correlation prediction results: The disease d with association score in the test sample set Te is p and d q Symptoms and characteristics p , g q , and the classification feature h p 、h q Input the disease correlation prediction model C′ for forward propagation to obtain the disease correlation d p and d q Predicted value of association score 2. The disease correlation prediction method based on twin network according to claim 1, characterized in that: The symptom characteristics of N diseases calculated in step (3) are DI = {g1, g2, ..., g n ,...,g N }, the implementation steps are: Diseases in the Statistical Literature Database n and symptoms m The number of documents that can be indexed simultaneously W n,m ; and using the TF-IDF method, through W n,m Calculate d n and e m The correlation score between n,m , get the disease d n Symptoms and characteristics n =(s n,1 ,s n,2 ,...s n,m ,...,s n,M ), then use g n The symptom characteristics of N diseases DI = {g1, g2, ..., g n ,...,g N },in: Where N m For use symptoms m and W from disease set D n,m >μ, 0≤μ≤10.

3. The disease correlation prediction method based on twin network according to claim 1, characterized in that: In step (5b), the ath disease d in the disease set D is a Symptoms and characteristics a and disease b b Symptoms and characteristics b As the input of the disease correlation prediction model C, forward propagation is performed. The implementation steps are: (5b1) The symptom feature g of the input of the sub-network CS1 of the twin network a Perform nonlinear transformation and output vector g′ a , while sub-network CS2 is used to analyze the symptom feature g b Perform nonlinear transformation and output vector g′ b ; (5b2) The vector concatenation layer outputs the vector g′ of CS1 a and the vector g′ output by CS2 b , and diseases a The classification feature h a and diseases b The classification feature h b Perform splicing to obtain the splicing vector H concat =[g′ a ,g′ b ,h a ,h b ], [] are concatenated activation functions; (5b2) The vector concatenation layer concatenates the output vector g′ of the twin network of model C a and g′ b , and diseases a The classification feature h a and diseases b The classification feature h b , get the splicing vector H concat =[g′ a ,g′ b ,h a ,h b ], [·] is the concatenated activation function; (5b3) The input layer of the multi-layer perceptron MLP receives the concatenated vector H concat , the hidden layer pair H concat Perform multiple nonlinear transformations, and the output vector H of the u-th hidden layer to the u-1 hidden layer u-1 Perform nonlinear transformation and output vector H u , the output vector H of the last hidden layer of the output layer U Perform nonlinear transformation and output disease pairs d a and d b Predicted value of association score U is the number of hidden layers, U≥5.

4. The disease correlation prediction method based on twin network according to claim 1, characterized in that: The loss function L described in step (5c) is expressed as: in For disease a and d b The correlation score, is the value of the MLP output layer, i.e., the prediction model C for disease d a and d b The predicted value of the association score, θ C is the weight parameter of the disease correlation prediction model C, and λ is the coefficient of the L2 regular term.

Images