Method for Predicting Drug Side Effects of Restricted Boltzmann Machine Based on Penalty Regularization Term

Abstract

The present invention discloses a method, device and equipment for predicting drug side effects based on a restricted Boltzmann machine with a penalty regular term, including: constructing a target drug similarity matrix, a target side effect similarity matrix and a drug-side effect association matrix; constructing a drug restricted Boltzmann machine model and a side effect restricted Boltzmann machine model; constructing a maximum likelihood function based on initial model parameters, the target drug similarity matrix, the target side effect similarity matrix and the drug-side effect association matrix, adding a penalty term to the maximum likelihood function, and performing iterative training until a preset number of iterations is reached to obtain the optimal model parameters; finally, using the trained model to perform prediction. The present invention trains the optimal parameters of the model through the gradient descent algorithm of the maximum likelihood function with a penalty term, avoids overfitting during the training process, improves the prediction accuracy of the model, and makes the prediction results more accurate.

Description

Technical Field

[0001] The present invention relates to the field of biomedical technologies, and particularly to a method, device, and equipment for predicting drug side effects based on a restricted Boltzmann machine with a penalty regular term. Background Art

[0002] Drugs bring health and hope to humans, but they are often accompanied by side effects. The identification of drug side effects is very important for the entire drug research and development process. Although many machine learning-based methods have been proposed for predicting drug side effects, there is still room for improvement.

[0003] Zhang et al. proposed an ensemble model for predicting drug side effects. This ensemble model uses the average score rule to fuse the prediction results of Pau's method, Liu's method, Cheng's method, RBMBM, and INBM methods, and achieved good results in predicting the potential side effects of approved drugs. However, Zhang's method cannot predict the side effects of new drugs. Regarding the problem of drug side effects of new drugs, Shabani-Mashcoo et al. proposed a networkand domain-based algorithm (NDDSA). The input of NDDSA only requires the chemical structure information of the drug, and it runs faster compared with other methods. However, the prediction results of NDDSA for the potential side effects of known drugs are not prominent. The above two methods for predicting drug side effects do not consider the overfitting of the model and the influence of introducing the original data on the prediction accuracy, and there is no prediction method that can simultaneously predict the side effects of known drugs and new drugs.

[0004] In summary, it can be seen that how to improve the prediction accuracy of drug side effects is a problem to be solved at present. Summary of the Invention

[0005] The object of the present invention is to provide a method for predicting drug side effects based on a restricted Boltzmann machine with a penalty regular term, which solves the influence of overfitting of the model and introducing the original data on the prediction accuracy in the prior art, improves the prediction accuracy of the model, and makes the prediction results of drug side effects more accurate.

[0006] To solve the above technical problems, the present invention provides a method for predicting drug side effects based on a restricted Boltzmann machine with a penalty regular term, including:

[0007] Constructing a plurality of drug similarity matrices and a plurality of side effect similarity matrices based on the collected drug side effect association information and drug chemical structure information;

[0008] Fuse and preprocess the multiple drug similarity matrices and the multiple side effect similarity matrices to obtain a target drug similarity matrix, a target side effect similarity matrix, and a drug-side effect association matrix;

[0009] Construct a drug restricted Boltzmann machine model and a side effect restricted Boltzmann machine model;

[0010] Based on the initial model parameters, the target drug similarity matrix, the target side effect similarity matrix, and the drug-side effect association matrix, construct a maximum likelihood function, add a penalty term to the maximum likelihood function, and perform iterative training until a preset number of iterations is reached to obtain the optimal model parameters;

[0011] Substitute the optimal parameters into the drug restricted Boltzmann machine model and the side effect restricted Boltzmann machine model to obtain a first drug-side effect prediction result and a second drug-side effect prediction result respectively;

[0012] Calculate a target drug-side effect prediction result according to the first drug-side effect prediction result and the second drug-side effect prediction result.

[0013] Preferably, the constructing of the multiple drug similarity matrices and the multiple side effect similarity matrices based on the collected drug-side effect association information and drug chemical structure information includes:

[0014] Construct a drug-side effect association matrix Y∈R n×m , where n is the number of drugs and m is the number of side effects. The element values in Y can be divided into two categories: Y ij =1 indicates that drug d i is associated with side effect s j ; Y ij =0 indicates that the association between drug d i and side effect s j is unknown;

[0015] Construct a matrix U∈R n×881 according to the drug chemical structure information, and U is obtained by fingerprint encoding of 881 chemical substructures;

[0016] Define as the association feature of drug d i , as the chemical structure feature of drug d i , denote the association feature of side effect s j ;

[0017] Calculate the multiple drug similarity matrices and the multiple side effect similarity matrices using multiple similarity evaluation methods.

[0018] Preferably, the fusion preprocessing of the multiple drug similarity matrices and the multiple side effect similarity matrices to obtain a target drug similarity matrix, a target side effect similarity matrix, and a drug-side effect association matrix includes:

[0019] Using the CKA-MKL algorithm to fuse the multiple drug similarity matrices and the multiple side effect similarity matrices into a target drug similarity matrix and a target side effect similarity matrix respectively;

[0020] Using the WKNKN algorithm to perform association preprocessing on the target drug similarity matrix and the target side effect similarity matrix to obtain the drug-side effect association matrix.

[0021] Preferably, the using the CKA-MKL algorithm to fuse the multiple drug similarity matrices and the multiple side effect similarity matrices into a target drug similarity matrix and a target side effect similarity matrix respectively includes:

[0022] Using the CKA-MKL algorithm to fuse the multiple drug similarity matrices and the multiple side effect similarity matrices into a target drug similarity matrix S final,d and the target side effect similarity matrix S final,s ;

[0023] The objective function of the CKA-MKL algorithm is defined as follows:

[0024]

[0025]

[0026] Where, I m ∈R m×m is the identity matrix, 1 m ∈R m is a vector with all elements being 1, ω i is the weight of the similarity matrix , S drug is the set of similarity matrices, and T is the transpose matrix.

[0027] Preferably, the using the WKNKN algorithm to perform association preprocessing on the target drug similarity matrix and the target side effect similarity matrix to obtain the drug-side effect association matrix includes:

[0028] S501: Initialize the drug association matrix Y d ∈R n×m to 0;

[0029] S502: Initialize the side effect association matrix Y s ∈Rm×n Initialize to 0;

[0030] S503: Let i = 1, and return the maximum nearest neighbor knn according to the similarity d = MaxSort(S final,d , i, K), knn s = MaxSort(S final,s , i, K), and determine whether i ≤ n. If it holds, then let i = i + 1 and calculate knn d and knn s ;

[0031] S504: If it does not hold, then stop the calculation. Let k = 1 and calculate w kd = μ k-1 S final,d [i, knn kd , w ks = μ k- 1 S final,s [i, knn ks and determine whether k ≤ K. If it holds, then let k = k + 1 and calculate w kd and w ks ;

[0032] S505: If it does not hold, then stop the calculation. Incorporate all the values of knn in S3 d and all the values of w in S4 kd into the side effect association matrix Y s ; Incorporate all the values of knn s and w ks into the side effect correlation matrix Y s ;

[0033] S506: Calculate the drug side effect association matrix Y d based on the drug association matrix Y s and the side effect correlation matrix Y ds = max((Y d + Y s T ) / 2, Y).

[0034] Preferably, the construction of the drug restricted Boltzmann machine model and the side effect restricted Boltzmann machine model includes:

[0035] The hidden layer and the visible layer of the drug restricted Boltzmann machine model and the side effect restricted Boltzmann machine model are respectively represented as h = {h1,..., h c} and v = {v1,..., v l}, and the joint probability distribution of v and h is:

[0036]

[0037] Among them, Z is a normalization constant function, and E(v, h) is an energy function:

[0038]

[0039] Hidden layer h j and visible layer v i have conditional probability distributions as follows:

[0040]

[0041]

[0042] Among them, σ(·) is the sigmoid function, a i and b j are the biases for the i-th visible layer node and the j-th hidden layer node, and W ij is the connection weight connecting the i-th visible layer node and the j-th hidden layer node;

[0043] Define the joint distribution of v as:

[0044]

[0045] Preferably, constructing a maximum likelihood function based on the initial model parameters, the target drug similarity matrix, the target side effect similarity matrix, and the drug side effect association matrix, adding a penalty term to the maximum likelihood function, and performing iterative training until a preset number of iterations is reached to obtain the optimal model parameters includes:

[0046] S701: Input the drug side effect association matrix the target drug similarity matrix S final,d a preset number of hidden nodes l [d] a preset hyperparameter λ [d] and θ [d] a preset learning rate α [d] ;

[0047] S702: Construct a maximum likelihood function based on the drug side effect association matrix and the drug similarity matrix as:

[0048] S703: Define the penalty term as:

[0049] Among them, ‖W [d] ‖ 2,1 is the l [d] -norm of W 2,1 and can be defined as:

[0050]

[0051] Matrix The elements in

[0052] are denoted as θ [d] tr(H T LH), where is the Laplacian regularization matrix, D is a diagonal matrix, and its i-th element is the sum of the i-th row of S final,d ;

[0053] S704: Substitute the penalty term into the maximum likelihood function to obtain

[0054] S705: Randomly initialize W [d] , a [d] , b [d] ;

[0055] S706: Initialize Let t = 1;

[0056] S707: Calculate the matrix i = 1,…,n; j = 1,…,l [d] ;

[0057] S708: Calculate i = 1,…,n; j = 1,…,m;

[0058] S709: Judge whether t ≥ T. If it holds, stop the calculation and output H * [i,j] and If it does not hold, let t = t + 1, and return to execute step S7, where T is the preset time;

[0059] S710: Calculate the matrix: i = 1,…,n; j = 1,…,l [d] ;

[0060] Calculate the weights:

[0061] Calculate the visible layer bias:

[0062] Calculate the hidden layer bias:

[0063] S711: Update the diagonal matrix G through the formula i = 1,…,m;

[0064] Update the weights △W = △W - λ [d] W[d] G;

[0065] Update weight

[0066] Update hidden layer bias

[0067] Update the model parameters W through the following formula [d] , a [d] , b [d] ;

[0068] W [d] = W [d] + α [d] △W

[0069] a [d] = a [d] + α [d] △a

[0070] b [d] = b [d] + α [d] △b

[0071] Stop iterating until the loss function converges and output W [d] , a [d] , b [d] , substitute the optimal parameters W [d] , a [d] , b [d] into the drug restricted Boltzmann machine model to obtain the trained drug restricted Boltzmann machine model;

[0072] Use the training process of steps S701 - S711 to train the side effect restricted Boltzmann machine model to obtain the trained side effect restricted Boltzmann machine model.

[0073] Preferably, the calculating the target drug side effect prediction result according to the first drug side effect prediction result and the second drug side effect prediction result includes:

[0074] Obtain the first drug side effect prediction result Y according to the trained drug restricted Boltzmann machine model pre,drug ;

[0075] Obtain the second drug side effect prediction result Y according to the trained side effect restricted Boltzmann machine model pre,side-effect ;

[0076] Using the formula the first drug side effect prediction result Y pre,drug and the second drug side effect prediction result Y pre,side-effect, the target drug side effect prediction result Y is calculated pre,final .

[0077] The present invention also provides a device for predicting drug side effects based on a restricted Boltzmann machine with a penalty regularization term, including:

[0078] A similarity matrix construction module, configured to construct multiple drug similarity matrices and multiple side effect similarity matrices based on the collected drug side effect association information and drug chemical structure information;

[0079] A matrix preprocessing module, configured to perform fusion preprocessing on the multiple drug similarity matrices and the multiple side effect similarity matrices to obtain a target drug similarity matrix, a target side effect similarity matrix, and a drug side effect association matrix;

[0080] A prediction model construction module, configured to construct a drug restricted Boltzmann machine model and a side effect restricted Boltzmann machine model;

[0081] A training prediction model module, configured to construct a maximum likelihood function based on initial model parameters, a target drug similarity matrix, a target side effect similarity matrix, and a drug side effect association matrix, add a penalty term to the maximum likelihood function, and perform iterative training until a preset number of iterations is reached to obtain optimal model parameters;

[0082] A prediction result module, configured to substitute the optimal parameters into the drug restricted Boltzmann machine model and the side effect restricted Boltzmann machine model to obtain a first drug side effect prediction result and a second drug side effect prediction result respectively;

[0083] A calculation module, configured to calculate the prediction result of the target drug side effect according to the first drug side effect prediction result and the second drug side effect prediction result.

[0084] The present invention also provides a device for predicting drug side effects based on a restricted Boltzmann machine, including:

[0085] A memory, configured to store a computer program; a processor, configured to implement the steps of the method for predicting drug side effects based on a restricted Boltzmann machine with a penalty regularization term as described above when executing the computer program.

[0086] The method for predicting drug side effects based on a restricted Boltzmann machine with a penalty regular term provided by the present invention includes: constructing multiple drug similarity matrices and multiple side effect similarity matrices from known drug side effect associations and chemical structure information of drugs; and fusing them into a target drug similarity matrix and a target side effect similarity matrix; performing correlation preprocessing on the target drug similarity matrix and the target side effect similarity matrix to obtain a drug side effect correlation matrix; constructing a restricted Boltzmann machine model for drugs and side effects, and training with a gradient descent algorithm of a maximum likelihood function with a penalty term to obtain the optimal parameters of the restricted Boltzmann machine model for drugs and side effects. The training method with a penalty term is used to avoid overfitting during the training process, and the deviation is corrected at each iteration to prevent the estimated parameters from tending to infinity, so as to obtain the optimal parameters of the model. Finally, the optimal parameters are substituted into the model to predict drug prediction results and side effect prediction results, and the correlation probability value of drug side effects is calculated. By training the optimal parameters of the model with a gradient descent algorithm of a maximum likelihood function with a penalty term, the present invention avoids fitting during the training process, corrects the deviation, and prevents the estimated parameters from tending to infinity, improving the prediction accuracy of the model and making the prediction results more accurate. BRIEF DESCRIPTION OF THE DRAWINGS

[0087] In order to more clearly illustrate the technical solutions of the embodiments of the present invention or the prior art, the following will briefly introduce the drawings required for the description of the embodiments or the prior art. Obviously, the following drawings are only some embodiments of the present invention. For those of ordinary skill in the art, other drawings can be obtained based on these drawings without creative efforts.

[0088] Figure 1 It is a flowchart of the first specific embodiment of the method for predicting drug side effects based on a restricted Boltzmann machine with a penalty regular term provided by the present invention;

[0089] Figure 2 It is a flowchart of the second specific embodiment of the method for predicting drug side effects based on a restricted Boltzmann machine with a penalty regular term provided by the present invention;

[0090] Figure 3 It is a flowchart of the WKNKN algorithm provided by the embodiments of the present invention;

[0091] Figure 4 It is a flowchart of the training process of the drug RWM model provided by the embodiments of the present invention;

[0092] Figure 5 It is a structural block diagram of a device for predicting drug side effects based on a restricted Boltzmann machine provided by the embodiments of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0093] The core of the present invention is to provide a method for predicting drug side effects based on a restricted Boltzmann machine with a penalty regularization term. When training the restricted Boltzmann machine (RBM) model, the gradient descent algorithm of the maximum likelihood function with a penalty term is used to find the parameters of the model, avoiding overfitting during the training process and correcting the bias, thereby improving the prediction accuracy.

[0094] To enable those skilled in the art to better understand the solution of the present invention, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the protection scope of the present invention.

[0095] Please refer to Figure 1 , Figure 1 which is a flowchart of the first specific embodiment of the method for predicting drug side effects based on a restricted Boltzmann machine with a penalty regularization term provided by the present invention; the specific operation steps are as follows:

[0096] Step S101: Construct a plurality of drug similarity matrices and a plurality of side effect similarity matrices based on the collected drug side effect association information and drug chemical structure information;

[0097] Step S102: Perform fusion preprocessing on the plurality of drug similarity matrices and the plurality of side effect similarity matrices to obtain a target drug similarity matrix, a target side effect similarity matrix, and a drug side effect association matrix;

[0098] Step S103: Construct a drug restricted Boltzmann machine model and a side effect restricted Boltzmann machine model;

[0099] Step S103: Construct a maximum likelihood function based on the initial model parameters, the target drug similarity matrix, the target side effect similarity matrix, and the drug side effect association matrix, add a penalty term to the maximum likelihood function, and perform iterative training until a preset number of iterations is reached to obtain the optimal model parameters;

[0100] Step S104: Substitute the optimal parameters into the drug restricted Boltzmann machine model and the side effect restricted Boltzmann machine model to obtain a first drug side effect prediction result and a second drug side effect prediction result respectively;

[0101] Step S105: Calculate the target drug side effect prediction result according to the first drug side effect prediction result and the second drug side effect prediction result.

[0102] In this embodiment, the known drug side effect association information and drug chemical structure information are collected and constructed into multiple drug similarity matrices and side effect similarity matrices, and then they are fused into a target drug similarity matrix and a target side effect similarity matrix. Then, the target drug similarity matrix and the target side effect similarity matrix are preprocessed for association to obtain a drug side effect association matrix. The gradient descent algorithm of the maximum likelihood function with a penalty term is used to train the constructed restricted Boltzmann machine model of drugs and side effects, avoiding overfitting during training, and correcting the deviation at each step of iteration to prevent the estimated parameter region from being infinitely large until the preset number of iterations is reached to obtain the optimal parameters. Based on the optimal parameters, drug prediction results and side effect prediction results are obtained, and finally the drug side effect prediction results are calculated. By training the optimal parameters of the model with the gradient descent algorithm of the maximum likelihood function with a penalty term, the present invention avoids fitting during the training process, corrects the deviation, prevents the estimated parameters from tending to infinity, improves the prediction accuracy of the model, and makes the prediction results more accurate.

[0103] Please refer to Figure 2 , Figure 2 which is the flowchart of the method for predicting drug side effects based on the restricted Boltzmann machine with a penalty regular term provided by the present invention; the specific operation steps are as follows:

[0104] Step S201: Construct a similarity matrix of multiple drugs and a similarity matrix of multiple side effect matrices;

[0105] The known drug side effect association matrix is Y ∈ R nh×m , where n is the number of drugs and m is the number of side effects. The element values in Y can be divided into two categories: Y ij = 1 indicates that drug d i and side effect s j are associated; Y ij = 0 indicates that the association between drug d i and side effect s j is unknown. Define the matrix U ∈ R n×881 to indicate the chemical structure information of the drug. U is obtained by fingerprint encoding of 881 chemical substructures.

[0106] We define and to represent the association feature and chemical structure feature of drug d i , where and are the i-th row of Y and the i-th row of U respectively. represents the association feature of side effect s j , where is the j-th column of Y. Taking the drug association feature as an example, we adopt the following four similarity evaluation methods:

[0107] (1) Gaussian Interaction Profile (GIP):

[0108]

[0109] Among them, γ = 1 represents the kernel width.

[0110] (2) Cosine Similarity (COS):

[0111]

[0112] (3) Correlation coefficient (Corr):

[0113]

[0114] (4) Normalized Mutual Information (NMI):

[0115]

[0116] Among them, is and 's interaction information. and are respectively and 's entropy.

[0117] Similarly, S GIP-link,s , S COS-link,s , S Corr-link,s and S NMI-link,s can be calculated according to the associated characteristics of side effects, and S GIP-chem,d , S COS-chem,d , S Corr-chem,d and S NMI-chen,d can be calculated from the chemical structure characteristics of the drug.

[0118] Step S202: Use the multi-kernel learning algorithm to fuse multiple drug similarity matrices and side effect similarity matrices respectively;

[0119] We use the CKA-MKL algorithm to fuse multiple drug similarity matrices and side effect similarity matrices respectively.

[0120] Taking the fusion of the drug similarity matrix as an example. First, define the set of drug similarity matrices: S drug .

[0121] Secondly, the objective function of the CKA-MKL algorithm is defined as follows:

[0122]

[0123]

[0124] Among them, I m ∈R m×m is the identity matrix, and 1 m ∈R m is a vector with elements all being 1. ω i is the similarity matrix weight. Formula 1.5 can be transformed into a quadratic programming problem. By solving Formula 1.5, the optimal weight vector ω can be obtained.

[0125] Finally, we obtain the fused drug similarity matrix:

[0126]

[0127] Similarly, we use the CKA-MKL algorithm to fuse multiple side-effect similarity matrices and finally obtain S final,s .

[0128] Step S203: Preprocess the drug side-effect association matrix using the WKNKN algorithm to obtain the drug side-effect association matrix;

[0129] The flowchart of the steps of the WKNKN algorithm is as Figure 3 shown.

[0130] S1: Initialize the drug association matrix Y d ∈R n×m to 0;

[0131] S2: Initialize the side-effect association matrix Y s ∈R m×n to 0;

[0132] S3: Let i = 1, and return the maximum neighbor knn d = MaxSort(S final,d , i, K), knn s = MaxSort(S final,s , i, K), and judge whether i ≤ n. If it holds, then let i = i + 1 and calculate knn d and knn s ;

[0133] S4: If it does not hold, then stop the calculation, let k = 1, and calculate w kd = μ k-1 S final,d [i, knn kd , w ks= μ k-1 S final,s [i, knn ks ; Determine whether k ≤ K. If true, then set k = k + 1 and calculate w kd and w ks , where μ = 0.9 is the descent parameter;

[0134] S5: If not true, then stop the calculation and incorporate all the knn d from S3 and all the w kd values from S4 into the side effect association matrix Y s , and incorporate all the knn s and w ks values into the side effect correlation matrix Y s ;

[0135] S6: Based on the drug association matrix Y d and the side effect correlation matrix Y s calculate the drug side effect association matrix Y dt = max((Y d + Y s T ) / 2, Y).

[0136] Step S204: Construct a drug restricted Boltzmann machine (RBM) model and a side effect restricted Boltzmann machine (RBM) model for prediction respectively, and conduct training;

[0137] The input of the restricted Boltzmann machine (RBM) needs to be 0 or 1. Therefore, we designed an activation function to convert continuous values into 0 or 1:

[0138]

[0139] Construct a drug restricted Boltzmann machine (RBM) model and a side effect restricted Boltzmann machine (RBM) model respectively. The restricted Boltzmann machine (RBM) is a classical collaborative filtering model, which is often used in data denoising and recommendation systems. In a restricted Boltzmann machine (RBM) model, the visible layer nodes and the hidden layer nodes are c and l respectively. The hidden layer and the visible layer are represented as h = {h1,..., h c} and v = {v1,..., v l}. Then the joint probability distribution of v and h is:

[0140]

[0141] where Z is the normalization constant function and E(v, h) is the energy function:

[0142]

[0143] h j and v i The conditional probability distributions are:

[0144]

[0145]

[0146] where σ(·) is the sigmoid function, a i and b j is the bias of the i-th visible layer node and the j-th hidden layer node. ij is the connection weight connecting the i-th visible layer node and the j-th hidden layer node.

[0147] Define the joint distribution of v as:

[0148]

[0149] In order to train the parameters (W [d] ,a [d] ,b [d] ) and (W [s] ,a [s] ,b [s] ), using the maximum likelihood function gradient descent algorithm with a penalty term. Taking the training of the drug restricted Boltzmann machine (RBM) model as an example, the input is: Y new , S final,d , the number of hidden nodes is l [d] . Its maximum likelihood function is:

[0150]

[0151] Adding the penalty term, we get:

[0152]

[0153] The penalty term is defined as:

[0154]

[0155] where λ [d] and θ [d] is a non-negative hyperparameter. ‖W [d] ‖ 2,1 It's about W [d] l 2,1 -norm, can be defined as:

[0156]

[0157] ‖W [d] ‖ 2,1 is used to avoid overfitting during training.

[0158] The elements in the matrix are

[0159] The second term in Equation 1.15 can be expressed as θ [d] tr(H T LH), where is the Laplacian regularization matrix, D is a diagonal matrix, and its i-th element is the sum of the i-th row of S final,d This regularization term can preserve the structural information of the hidden layer data in the hidden layer.

[0160] Taking the derivatives of ι(W [d] , a [d] , b [d] ) with respect to W [d] , a [d] , b [d] respectively, we get:

[0161]

[0162] where <·> data denotes the expectation of the observed data, and <·> mod el denotes the expectation learned by the restricted Boltzmann machine (RBM) model. Therefore, it is difficult to calculate <·> mod el during the training process.

[0163] Using Gibbs sampling to obtain <·> T to approximate <·> mod el . T is set to 1.

[0164] Taking the derivative of λ [d] ‖W [d] ‖ 2,1 with respect to W [d] we get:

[0165]

[0166] where G is a diagonal matrix, and its diagonal elements are:

[0167]

[0168] Taking the derivatives of tr(H T LH) with respect to W [d] , b [d] respectively, we get:

[0169]

[0170]

[0171] Similarly, the input of the side-effect restricted Boltzmann machine (RBM) model is and S final,s , the number l of hidden nodes [s] .

[0172] After training the parameters of the drug restricted Boltzmann machine (RBM) model and the side-effect restricted Boltzmann machine (RBM) model, we obtain the prediction result by reconstructing the visible layer:

[0173]

[0174]

[0175] Similarly, the prediction result Y of the side-effect restricted Boltzmann machine (RBM) model can be obtained pre,side-effect .

[0176] The training process diagram of the drug restricted Boltzmann machine (RBM) model is as Figure 4 shown.

[0177] Step S205: Calculate the final prediction result using the average scoring rule.

[0178] According to the trained drug restricted Boltzmann machine (RBM) model, the first drug side-effect prediction result Y pre,drug ;

[0179] According to the trained side-effect restricted Boltzmann machine (RBM) model, the second drug side-effect prediction result Y pre,side-effect ;

[0180] Using the formula the first drug side-effect prediction result Y pre,drug and the second drug side-effect prediction result Y pre,side-effect , calculate the target drug side-effect prediction result Y pre,final .

[0181] The present invention can be applied to the prediction of potential side effects of marketed drugs and can also be used for the prediction of side effects of new drugs. We detect the effect of this method through 5Cross Validation (5-CV) and Local5Cross Validation (Local 5-CV) on three drug side-effect datasets of Liu, Pau and Miz, as shown in Tables 1 and 2.

[0182] Table 1. Comparison of the results of the method in this paper with other methods in 5-CV.

[0183]

[0184]

[0185] Table 2. Comparison of the results of the method in this paper with other methods in Local 5-CV.

[0186]

[0187]

[0188] As can be seen from Table 1, in the CV-5 experiment, our method achieved the optimal AUPR and AUC results on three datasets. In Table 2, our method achieved the optimal AUPR and F score results on the Pau and Miz datasets. Drug side effect prediction can be regarded as a classification problem of imbalanced samples. Therefore, evaluation metrics such as AUPR, Spec, and F score are more valuable than other metrics (AUC, SN, Precision, and ACC). Generally speaking, compared with other methods, our method has higher prediction accuracy.

[0189] In this embodiment, the WKNKN algorithm is disclosed to preprocess the association matrix. The matrix of new drugs and side effects is preprocessed by the WKNKN algorithm to obtain the drug side effect association matrix. The model can also predict the side effects of drugs, and the steps of the WKNKN algorithm are described in detail. Then, the gradient descent algorithm of the maximum likelihood function with a penalty term is used to train the drug RBM model and the side effect RBM model to avoid overfitting during the training process, correct the deviation in each iteration, and obtain the optimal model parameters. Then, the drug prediction result and the side effect prediction result are obtained through the drug RBM model and the side effect RBM model. Finally, the final prediction result is calculated by the average score rule. The present invention uses the WKNKN algorithm so that the present invention can not only detect the side effects of known drugs, but also predict the side effects of new drugs. The optimal parameters of the model are trained by the gradient descent algorithm of the maximum likelihood function with a penalty term to avoid overfitting during the training process, correct the deviation, prevent the estimated parameters from tending to infinity, improve the prediction accuracy of the model, and make the prediction result more accurate.

[0190] Please refer to Figure 5 , Figure 5 which is the structural block diagram of the device for predicting drug side effects based on the restricted Boltzmann machine with a penalty regular term provided by the embodiment of the present invention; the specific device may include:

[0191] The similarity matrix construction module 100 is used to construct multiple drug similarity matrices and multiple side effect similarity matrices based on the collected drug side effect association information and drug chemical structure information;

[0192] The matrix preprocessing module 200 is used to perform fusion preprocessing on the multiple drug similarity matrices and the multiple side effect similarity matrices to obtain a target drug similarity matrix, a target side effect similarity matrix, and a drug side effect association matrix;

[0193] The prediction model construction module 300 is used to construct a drug restricted Boltzmann machine model and a side effect restricted Boltzmann machine model;

[0194] The prediction model training module 400 is used to construct a maximum likelihood function based on the initial model parameters, the target drug similarity matrix, the target side effect similarity matrix, and the drug side effect association matrix, add a penalty term to the maximum likelihood function, and perform iterative training until a preset number of iterations is reached to obtain the optimal model parameters;

[0195] The prediction result module 500 is used to substitute the optimal parameters into the drug restricted Boltzmann machine model and the side effect restricted Boltzmann machine model to obtain a first drug side effect prediction result and a second drug side effect prediction result respectively;

[0196] The calculation module 600 is used to calculate the prediction result of the target drug side effect according to the first drug side effect prediction result and the second drug side effect prediction result.

[0197] An apparatus for predicting drug side effects based on a restricted Boltzmann machine in this embodiment is used to implement the foregoing method for predicting drug side effects based on a restricted Boltzmann machine with a penalty regular term. Therefore, the specific implementation manners in an apparatus for predicting drug side effects based on a restricted Boltzmann machine can be seen in the embodiment part of the foregoing method for predicting drug side effects based on a restricted Boltzmann machine with a penalty regular term. For example, the similarity matrix construction module 100, the matrix preprocessing module 200, the prediction model construction module 300, the prediction model training module 400, the prediction result module 500, and the calculation module 600 are respectively used to implement steps S101, S102, S103, S104, S105, and S106 in the foregoing method for predicting drug side effects based on a restricted Boltzmann machine with a penalty regular term. Therefore, the specific implementation manners can refer to the descriptions of the corresponding individual embodiments and will not be elaborated here.

[0198] A specific embodiment of the present invention also provides a device for predicting drug side effects based on a restricted Boltzmann machine, including: a memory for storing a computer program; a processor for implementing the steps of the above method for predicting drug side effects of a restricted Boltzmann machine based on a penalty regular term when executing the computer program.

[0199] A specific embodiment of the present invention also provides a computer-readable storage medium, on which a computer program is stored, and when the computer program is executed by a processor, the steps of the above method for predicting drug side effects of a restricted Boltzmann machine based on a penalty regular term are implemented.

[0200] In this specification, the various embodiments are described in a progressive manner. Each embodiment focuses on the differences from other embodiments. For the same or similar parts among the various embodiments, reference can be made to each other. For the devices disclosed in the embodiments, since they correspond to the methods disclosed in the embodiments, the description is relatively simple, and reference can be made to the description in the method part for related parts.

[0201] Those skilled in the art can further realize that the units and algorithm steps of the examples described in combination with the embodiments disclosed herein can be implemented by electronic hardware, computer software, or a combination of the two. To clearly illustrate the interchangeability of hardware and software, the components and steps of the examples have been generally described according to their functions in the above description. Whether these functions are executed in a hardware or software manner depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered to exceed the scope of the present invention.

[0202] The steps of the method or algorithm described in combination with the embodiments disclosed herein can be directly implemented by hardware, a software module executed by a processor, or a combination of the two. The software module can be placed in a random access memory (RAM), internal memory, read-only memory (ROM), electrically programmable ROM, electrically erasable programmable ROM, register, hard disk, removable disk, CD-ROM, or any other form of storage medium well-known in the technical field.

[0203] The above has introduced in detail the method, device, and equipment for predicting drug side effects of a restricted Boltzmann machine based on a penalty regular term provided by the present invention. Specific examples are used herein to elaborate on the principle and implementation manner of the present invention. The description of the above embodiments is only used to help understand the method and its core idea of the present invention. It should be noted that for those of ordinary skill in the art in this technical field, without departing from the principle of the present invention, several improvements and modifications can be made to the present invention, and these improvements and modifications also fall within the protection scope of the claims of the present invention.

Claims

1. A method for predicting drug side effects based on a restricted Boltzmann machine with a penalty regularization term, characterized in that Including: Constructing a plurality of drug similarity matrices and a plurality of side effect similarity matrices based on the collected drug side effect association information and drug chemical structure information; Performing fusion preprocessing on the plurality of drug similarity matrices and the plurality of side effect similarity matrices to obtain a target drug similarity matrix, a target side effect similarity matrix, and a drug side effect association matrix; Constructing a drug restricted Boltzmann machine model and a side effect restricted Boltzmann machine model; Constructing a maximum likelihood function based on the initial model parameters, the target drug similarity matrix, the target side effect similarity matrix, and the drug side effect association matrix, adding a penalty term to the maximum likelihood function, and performing iterative training until a preset number of iterations is reached to obtain the optimal model parameters; Substituting the optimal parameters into the drug restricted Boltzmann machine model and the side effect restricted Boltzmann machine model to obtain a first drug side effect prediction result and a second drug side effect prediction result respectively; Calculating a target drug side effect prediction result according to the first drug side effect prediction result and the second drug side effect prediction result.

2. The method according to claim 1, characterized in that, The constructing a plurality of drug similarity matrices and a plurality of side effect similarity matrices based on the collected drug side effect association information and drug chemical structure information includes: Construct a drug side effect association matrix \(Y\in R\) based on the collected drug side effect association information n×m , where \(n\) is the number of drugs and \(m\) is the number of side effects; the element values in \(Y\) can be divided into two categories: \(Y ij = 1 indicates that drug \(d i is associated with side effect \(s j ; \(Y ij = 0 indicates that the association between drug \(d i and side effect \(s j is unknown; Construct a matrix U ∈ R according to the above medicinal chemical structure information n×881 , where U is encoded by the fingerprints of 881 chemical substructures; Definition is the associated feature of drug d i ; is the chemical structure feature of drug d i ; represents the associated feature of side effect s j ; Calculating the plurality of drug similarity matrices and the plurality of side effect similarity matrices by using a variety of similarity evaluation methods.

3. The method according to claim 1, characterized in that, The performing fusion preprocessing on the plurality of drug similarity matrices and the plurality of side effect similarity matrices to obtain a target drug similarity matrix, a target side effect similarity matrix, and a drug side effect association matrix includes: Using the CKA-MKL algorithm to fuse the plurality of drug similarity matrices and the plurality of side effect similarity matrices into the target drug similarity matrix and the target side effect similarity matrix respectively; Using the WKNKN algorithm to perform association preprocessing on the target drug similarity matrix and the target side effect similarity matrix to obtain the drug side effect association matrix.

4. The method according to claim 3, wherein The using the CKA-MKL algorithm to fuse the plurality of drug similarity matrices and the plurality of side effect similarity matrices into a target drug similarity matrix and a target side effect similarity matrix respectively includes: Fusing the multiple drug similarity matrices and the multiple side effect similarity matrices into a target drug similarity matrix S respectively by using the CKA-MKL algorithm final，d and the target side effect similarity matrix S final，s ; The objective function of the CKA-MKL algorithm is defined as follows: Among them, is the identity matrix, 1 m ∈R m is a vector with elements all being 1, ω i is a similarity matrix is the weight of, S drug is the set of similarity matrices, and T is the transpose matrix.

5. The method according to claim 4, wherein The using the WKNKN algorithm to perform association preprocessing on the target drug similarity matrix and the target side effect similarity matrix to obtain a drug side effect association matrix includes: S501: Initialize the drug association matrix Y d ∈R n×m to 0; S502: Initialize the side effect correlation matrix Y s ∈R m×n to 0; S503: Let i = 1, and return the maximum nearest neighbor knn according to the similarity d = MaxSort(S final，d , i, K), knn s = MaxSort(S final，s , i, K), and determine whether i ≤ n. If it holds, then let i = i + 1 and calculate knn d and knn s ; S504: If not established, stop the calculation, set k = 1, and calculate w kd = μ k-1 S final，d [i, knn kd , w ks = μ k-1 S final，s [i, knn ks ; Determine whether k ≤ K. If established, set k = k + 1 and calculate w kd and w ks ; S505: If not, stop the calculation and incorporate all the knn in S3 d and all the w in S4 kd into the side effect association matrix Y s In this way, incorporate all the knn s and w ks into the side effect association matrix Y s ; S506: Based on the drug association matrix Y d and the side effect association matrix Y s Calculate the drug side effect association matrix Y ds = max((Y d + Y s T ) / 2, Y).

6. The method according to claim 1, wherein The constructing a drug restricted Boltzmann machine model and a side effect restricted Boltzmann machine model includes: The hidden layer and the visible layer of the said drug restricted Boltzmann machine model and the said side effect restricted Boltzmann machine model are respectively represented as h = {h1,..., h c} and v = {v1,..., v l}, and the joint probability distribution of v and h is: Where Z is a normalization constant function and E(v, h) is an energy function: Hidden layer h j and visible layer v i have conditional probability distributions as follows: where σ(·) is the sigmoid function, a i and b j are the biases of the i-th visible layer node and the j-th hidden layer node, and W ij is the connection weight connecting the i-th visible layer node and the j-th hidden layer node; Defining the joint distribution of v as:

7. The method according to claim 6, wherein The constructing a maximum likelihood function based on the initial model parameters, the target drug similarity matrix, the target side effect similarity matrix, and the drug side effect association matrix, adding a penalty term to the maximum likelihood function, and performing iterative training until a preset number of iterations is reached to obtain the optimal model parameters includes: S701: Input the drug side effect association matrix The target drug similarity matrix S final，d , the preset number of hidden nodes is l [d] , the preset hyperparameter λ [d] and θ [d] , the preset learning rate α [d] ; S702: Construct the maximum likelihood function based on the drug side effect association matrix and the drug similarity matrix as follows: S703: Define the penalty term as: Among them, ||W [d] || 2，1 is the l [d] -norm with respect to W 2，1 and can be defined as: Matrix The elements in Denoted as θ [d] tr(H T LH), where is the Laplacian regularization matrix, D is a diagonal matrix, and its ith element is the sum of the ith row of S final，d ; S704: Substitute the penalty term into the maximum likelihood function to obtain S705: Randomly initialize W [d] 、a [d] 、b [d] ; S706: Initialization Set t = 1; S707: Calculate matrix S708: Calculate S709: Determine if t≥T. If true, stop the calculation and output H * [i, j] and If not, let t = t + 1, and return to execute step S7, where T is the preset time S710: Calculate matrix: Calculate weights: Calculate the bias of the display layer: Calculate the hidden layer bias: S711: Update the diagonal matrix G through the formula ; Updated weight ΔW = ΔW - λ [d] W [d] G; Update weights Update hidden layer bias Update the model parameters W through the following formula [d] , a [d] , b [d] ; W [d] = W [d] + α [d] ΔW a [d] = a [d] + α [d] Δa b [d] = b [d] + α [d] Δb Stop iterating until the loss function converges and output W [d] 、a [d] 、b [d] Substitute the optimal parameters W [d] 、a [d] 、b [d] into the described drug restricted Boltzmann machine model to obtain the trained drug restricted Boltzmann machine model; Training the side effect restricted Boltzmann machine model by using the training process of steps S701 - S711 to obtain the trained side effect restricted Boltzmann machine model.

8. The method according to claim 1, characterized in that, Calculating the predicted result of the target drug side effect according to the first predicted result of drug side effect and the second predicted result of drug side effect includes: Based on the trained drug restricted Boltzmann machine model, the first drug side effect prediction result Y is obtained pre，drug ; Based on the trained side-effect restricted Boltzmann machine model, the second drug side-effect prediction result Y is obtained pre，side-effect ; Using the formula the first drug side effect prediction result Y pre，drug and the second drug side effect prediction result Y pre，side-effect to calculate the target drug side effect prediction result Y pre，final .

9. A device for predicting drug side effects based on a restricted Boltzmann machine with a penalty regularization term, characterized in that, Including: A similarity matrix construction module, configured to construct a plurality of drug similarity matrices and a plurality of side effect similarity matrices based on the collected drug side effect association information and drug chemical structure information; A matrix preprocessing module, configured to perform fusion preprocessing on the plurality of drug similarity matrices and the plurality of side effect similarity matrices to obtain a target drug similarity matrix, a target side effect similarity matrix, and a drug side effect association matrix; A prediction model construction module, configured to construct a drug restricted Boltzmann machine model and a side effect restricted Boltzmann machine model; A prediction model training module, configured to construct a maximum likelihood function based on initial model parameters, a target drug similarity matrix, a target side effect similarity matrix, and a drug side effect association matrix, add a penalty term to the maximum likelihood function, and perform iterative training until a preset number of iterations is reached to obtain optimal model parameters; A prediction result module, configured to substitute the optimal parameters into the drug restricted Boltzmann machine model and the side effect restricted Boltzmann machine model to respectively obtain a first predicted result of drug side effect and a second predicted result of drug side effect; A calculation module, configured to calculate a predicted result of the target drug side effect according to the first predicted result of drug side effect and the second predicted result of drug side effect.

10. A device for predicting drug side effects based on a restricted Boltzmann machine with a penalty regularization term, characterized in that, Including: A memory, configured to store a computer program; A processor, configured to implement the steps of the method for predicting drug side effects based on a restricted Boltzmann machine with a penalty regularization term according to any one of claims 1 to 8 when executing the computer program.

Images