Drug-target protein association prediction method and system based on multi-source data embedding

Abstract

The invention discloses a drug-target protein association prediction method and system based on multi-source data embedding. The multi-source data embedding method is formed by a multi-kernel learning algorithm and a noise reduction auto-encoder. On the basis of a multi-kernel learning method, the robustness of the model is improved by integrating multiple kernel functions, excessive dependence on a single kernel function is relieved, weights are distributed to different similar networks, and finally a comprehensive similar matrix kernel is obtained. On the basis of a noise reduction self-encoding method, noise is introduced into input data, original data is tried to be restored from noise data, and the functions of noise removal and data restoration are achieved. According to the method, the potential association in the drug-disease association network can be accurately predicted simply according to association data of drugs and target proteins, and the effectiveness of the method is proved through experiments, so that the method has certain biological significance; the method has theoretical significance and reference value for drug relocation and promotion of drug research and development processes.

Description

Technical Field

[0001] The present invention belongs to the field of computational biology and relates to a drug-target protein association prediction method and system based on multi-source data embedding. Background Art

[0002] Traditional drug development usually requires a long process of drug discovery, preclinical research, clinical research, and finally drug marketing. This process usually takes 10-15 years, plus more than $1.5 billion in funding, and the overall development difficulty is relatively high. However, with such huge consumption, the probability of successful development is less than 10%. Obviously, such R&D consumption and success probability are unacceptable. How to find some new ways to reduce the time and capital investment costs in the traditional drug development process and increase the success probability of drug development is a question worthy of in-depth thinking. Drug repositioning, as a new R&D strategy, uses computer-assisted calculations to predict biological data including biological associations such as drug-target proteins, reduce initial screening costs, and obtain more related information in the drug development process, which can effectively solve the above problems.

[0003] As time goes by and science and technology develop, more and more biology-related data are available for researchers to use. For example, drug-related databases such as DrugBank and SIDER record the chemical structure, biological activity, pharmacology, toxicology and other information of drugs. Disease-related databases such as MalaCards provide detailed descriptions of the name, definition, pathogenesis, symptoms, clinical characteristics and other aspects of the disease. Protein-related databases such as HPRD and UniProt collect information on the structure, function, interaction and expression pattern of proteins in different tissues and cell types. The development of various biological information-related databases has created unlimited possibilities for association prediction based on computer calculation methods. The computational method uses the chemical fingerprint information of drugs, the amino acid sequence information of proteins, and the association data of various aspects such as drugs, diseases, proteins, and side effects as a medium to mine the deep information in the data, and finally calculate the possibility of association between target proteins, and then further verify the association through biological wet experiments. Computational biology methods conduct in-depth research on various biological factors at different levels, which greatly promotes the theoretical understanding of the pathogenic mechanism of complex diseases in organisms, and improves the speed of drug development and the efficiency of successful drug development.

[0004] Analyzing the complex biological association networks formed by biological molecules such as drugs, diseases, and target proteins is an important research topic in bioinformatics. With the development of computers, the use of machine learning, deep learning, and other computational methods to capture deep information between data is crucial for predicting the association between the two. Machine learning-based methods improve the performance of the model by obtaining global structural information between drugs and diseases. Although such methods take into account global structural information and effectively improve the predictive performance of the model, they often perform poorly on unbalanced data sets. Deep learning-based methods improve predictive performance by deeply fusing multi-source drug and protein data, but such methods usually require sufficient data for effective training, and involve a large amount of hyperparameter optimization, requiring iterative training to obtain good results; at the same time, outliers and noise data in the data may cause the performance of deep learning-based methods to decline.

[0005] However, the more data collected, the more relevant information it contains. How to retain valid information in the process of data information fusion, reduce the impact of noise data on the results, and further improve the predictive performance of the model is a question worth considering.

[0006] Therefore, it is necessary to design a drug-target protein association prediction method based on multi-source data embedding. Summary of the invention

[0007] The technical problem to be solved by the present invention is to provide a drug-target protein association prediction method and system based on multi-source data embedding. The method fuses multiple similarity matrices based on a multi-kernel learning algorithm to obtain a comprehensive similarity matrix, and uses a denoising autoencoder to reduce noise and dimension of feature data. It only needs to be based on the association matrix of drugs and target proteins to more accurately identify potential drug-target protein associations, which has certain biological significance.

[0008] The technical solution provided by the present invention is as follows:

[0009] On the one hand, a drug-target protein association prediction method based on multi-source data embedding comprises the following steps:

[0010] Step 1), based on the data sources of drugs and target proteins, calculate the similarities between drugs and target proteins, and construct similarity networks of drugs and target proteins in each data source;

[0011] Step 2), the multi-kernel learning algorithm MKL is used to integrate the similarity networks of drugs and target proteins respectively to obtain the comprehensive similarity matrices SD and SP of drugs and target proteins;

[0012] Step 3), taking each row of data in the comprehensive similarity matrix SD and SP obtained in step 2) as the initial feature representation of the drug and the target protein, and introducing noise into the initial feature data through a denoising autoencoder; and restoring the initial feature data with the introduced noise;

[0013] , ;

[0014] in, represents randomly generated noise, represents the mean of the distribution, represents variance; Represents the feature data of the original input, express Feature data after adding noise;

[0015] The encoder and decoder used for restoration are as follows:

[0016] ;

[0017] ;

[0018] in, and denote the output of the encoder and decoder layer l, respectively. and They represent the weight matrix and bias matrix of the encoder layer l, respectively, and σ(*) represents the nonlinear activation function; Represents the reconstructed data finally output by the decoder, and Respectively represent the weight matrix and bias matrix of the decoder layer l;

[0019] Restoration is used to remove noise and restore data. After denoising, redundant or irrelevant information is removed, the core features of the data are retained, and it is more suitable for subsequent modeling tasks. This enhances the model's ability to learn potential patterns in the data and reduces the interference of noise on prediction accuracy.

[0020] In step 4), the restored drug and target protein feature data processed in step 3) are spliced ​​to obtain the feature vector of the drug-target protein pair, and the association between the feature vectors of the drug-protein pair is modeled using a multi-layer perceptron to obtain a drug-target protein association prediction model based on a multi-layer perceptron to predict the association between the drug and the target protein.

[0021] Furthermore, the prediction score of the association between drugs and target proteins Calculated by the following formula:

[0022] ;

[0023] in, is the feature vector of the input drug-target protein pair, , , and , , The corresponding weights and bias vectors determined in the drug-target protein association prediction model based on the multi-layer perceptron, The function calculation formula is: , e represents the natural exponent, and the Relu function calculation formula is , is the calculated variable of the input function.

[0024] The parameters in the drug-target protein association prediction model based on multi-layer perceptron are usually optimized using back propagation algorithm and gradient descent algorithm;

[0025] Furthermore, the construction process of the similarity network of drugs and target proteins in each data source is as follows:

[0026] Step 1.1: Construction of drug similarity network;

[0027] Obtain drug data sources, including at least the chemical structure of the drug, drug-drug association, drug-disease association, and drug-drug side effect data; for each data source, use the Jaccard similarity coefficient to calculate the similarity between drugs, and generate an independent drug similarity network for each data source, which is expressed in a matrix form, in which each element represents a similarity score between a pair of drugs;

[0028] The Jaccard similarity coefficient calculates the ratio of the intersection and union of two drugs on a specific data source to obtain a similarity score.

[0029] Step 1.2: Construction of target protein similarity network;

[0030] Obtain target protein data sources, including at least target protein sequence information, target protein-target protein interaction, and target protein-disease association; for each data source, use the Jaccard similarity coefficient to calculate the similarity between target proteins, and generate an independent target protein similarity network for each data source;

[0031] The Jaccard similarity coefficient calculates the ratio of the intersection and union of two target proteins on a specific data source to obtain a similarity score.

[0032] Furthermore, the multi-core learning algorithm MKL is used to integrate the similarity networks of drugs and target proteins respectively, and weights are assigned to each similarity network to obtain the comprehensive similarity matrices SD and SP of drugs and target proteins:

[0033] In the process of integrating the similarity network using the multi-core learning algorithm MKL, the objective function corresponding to the following formula is used to minimize the distance between the optimal kernel of the integrated similarity network and the ideal optimal kernel:

[0034] Objective function: ,

[0035] ,

[0036] , ,

[0037] ;

[0038] in, express The similarity network of the i-th express The weight vector of the similarity network, is the regularization parameter, yes The ideal optimal kernel of , β is a regularization parameter used to balance the similarity matrix fitting error and the weight of the weight regularization term; Indicates drug or target protein, Indicates drugs, indicates the target protein; It represents the association matrix of drug-target protein; and Respectively represent the integration of multiple The comprehensive similarity matrix generated after the similarity network is * yes The ideal optimal kernel matrix of express The number of similarity networks; represents the square of the Frobenius norm, Represents the integrated comprehensive similarity matrix SX and the ideal optimal kernel matrix * The difference between Represents the weight vector The regularization term is used to constrain the sparsity of full-time and avoid overfitting;

[0039] In order to avoid overfitting during training and improve the generalization ability of the model, making the system performance more robust, an L2 penalty is imposed on the weight vector to limit the growth of authority.

[0040] Weight solution formula:

[0041] ,

[0042] ,

[0043] ;

[0044] in: ,

[0045] ,

[0046] ;

[0047] Among them, trace represents the trace of the matrix;

[0048] The conversion formula is:

[0049]

[0050] ;

[0051] Represents drugs or target proteins, and the weights of each similarity network of drugs and target proteins are obtained according to the above formula.

[0052] The significance of multivariate data fusion, since the similarity between drugs and target proteins can be described from multiple perspectives (such as chemical structure, sequence information, functional association, etc.), a single data source may not be able to fully reflect their similarity. Therefore, by integrating the similarity network generated by multiple data sources, the complex relationship between drugs and target proteins can be more comprehensively captured, thereby improving the accuracy of subsequent prediction models.

[0053] The Multiple Kernel Learning (MKL) algorithm is used to integrate multiple similarity networks to generate a comprehensive similarity matrix kernel. Each similarity network can be regarded as a kernel function. The MKL algorithm reflects its importance in the comprehensive similarity matrix by assigning different weights to each kernel function. The optimal weight of each kernel function is determined by optimization algorithms (such as gradient descent or quadratic programming) so that the comprehensive similarity matrix can better reflect the true relationship between the drug and the target protein. Finally, the MKL algorithm weightedly fuses multiple similarity networks to generate a comprehensive similarity matrix kernel for subsequent drug-target protein association prediction models.

[0054] By using a multi-kernel learning algorithm, the model no longer relies on a single similarity network, but is able to comprehensively consider information from multiple data sources. This not only improves the model's prediction accuracy, but also enhances the model's robustness to noise and abnormal data, avoiding prediction errors caused by deviations from a single data source.

[0055] Furthermore, the drug comprehensive similarity matrix SD and the target protein comprehensive similarity matrix SP are calculated by the following formula:

[0056] ;

[0057] ;

[0058] in, Represent the number of drugs and target proteins, respectively. N D and N P are the number of similarity networks of drugs and target proteins, respectively, and R represents the real number domain.

[0059] Furthermore, the Softplus function and RMSProp algorithm are used to optimize the mean square error formula in the dimensionality reduction process. The mean square error calculation formula is as follows:

[0060] ;

[0061] Where L represents the loss function of the denoising autoencoder, Represents the reconstructed data finally output by the decoder, Represents the raw input data fed into the denoising autoencoder;

[0062] Reduce the loss of data information during dimensionality reduction;

[0063] Furthermore, the drug-target protein association prediction model based on the multi-layer perceptron uses the Adam algorithm and sets the learning rate to 0.001 to optimize the binary crossover loss:

[0064] ;

[0065] in, and They represent the positive and negative sample labels of drug-target protein pairs, respectively. Y is the true label, with a value of 0 or 1. N represents the number of drug-target protein pairs. Represents the prediction score of the association between drugs and target proteins, Represents the loss function of the drug-target protein association prediction model based on multi-layer perceptron.

[0066] In a second aspect, a system for predicting drug-target protein association based on the above-mentioned multi-source data embedding method comprises:

[0067] The similarity network construction module calculates the similarities between drugs and target proteins based on the data sources of drugs and target proteins, and constructs the similarity network of drugs and target proteins in each data source;

[0068] The similarity matrix acquisition module uses the multi-kernel learning algorithm MKL to integrate the similarity networks of drugs and target proteins respectively to obtain the comprehensive similarity matrices SD and SP of drugs and target proteins;

[0069] The initial feature restoration module uses each row of data in the obtained comprehensive similarity matrix SD and SP as the initial feature representation of the drug and target protein, and introduces noise into the initial feature data through a noise reduction autoencoder; and restores the initial feature data with introduced noise;

[0070] The association prediction module splices the processed and restored feature data of the drug and target protein to obtain the feature vector of the drug-target protein pair, and uses a multi-layer perceptron to model the association between the feature vectors of the drug-target protein pair, obtains a drug and target protein association prediction model based on a multi-layer perceptron, and predicts the association relationship between the drug and the target protein.

[0071] In a third aspect, a computer-readable storage medium stores a computer program, wherein the computer program is called by a processor to implement:

[0072] The above steps of the drug-target protein association prediction method based on multi-source data embedding.

[0073] In a fourth aspect, an electronic device includes:

[0074] one or more processors;

[0075] A storage device for storing one or more programs;

[0076] The one or more processors execute the one or more programs to implement the steps of the above-mentioned drug-target protein association prediction method based on multi-source data embedding.

[0077] Beneficial effects:

[0078] The present invention provides a drug-target protein association prediction method (MKLDAE) based on multi-source data embedding and a system. The method takes into account the differences in the amount of information contained in different association matrices. The sparser the association matrix, the less information it contains, and vice versa. A multi-core learning method is used to dynamically assign fusion weights to the association matrix to retain as much information as possible in the fused similarity matrix. In order to reduce the impact of noise in the data on the prediction results, a denoising autoencoder is used to perform denoising and dimensionality reduction on the feature data to improve the stability of the model. The prediction system has a simple structure and is easy to operate;

[0079] Compared with the existing drug-target protein association prediction methods, the MKLDAE method described in the present invention has the following advantages:

[0080] 1) Using the multi-core learning algorithm, the corresponding weights are dynamically assigned during the similarity matrix fusion process to maximize the amount of information contained in the fused similarity matrix.

[0081] 2) Use the denoising autoencoder to reduce the dimension and noise of the fused feature data, reduce the impact of noise data on performance, and improve the stability of the model.

[0082] The present invention is simple to implement, and only needs to use the association data of drugs and target proteins to more accurately predict the association between potential drugs and target proteins. Experimental verification shows that the method MKLDAE described in the present invention can effectively predict the association between drugs and target proteins. At the same time, by comparing with other methods, the AUROC and AUPR values ​​of MKLDAE both obtained the highest values, and the performance is better. For specific experimental result comparison and analysis, please refer to the embodiment. BRIEF DESCRIPTION OF THE DRAWINGS

[0083] Figure 1 It is a schematic diagram of a process of a drug-target protein association prediction method (MKLDAE) based on multi-source data embedding provided by an embodiment of the present invention;

[0084] Figure 2 is a ROC curve diagram of the prediction method (MKLDAE) provided by an embodiment of the present invention and other comparison methods;

[0085] Figure 3 It is a PR curve diagram of the prediction method (MKLDAE) provided in the embodiment of the present invention and other comparison methods. DETAILED DESCRIPTION

[0086] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0087] Embodiment 1:

[0088] A drug-target protein association prediction method based on multi-source data embedding, such as Figure 1 As shown, the following steps are included:

[0089] exist Figure 1 In the table, Drug Similarity Networks represents drug similarity networks, Protein Similarity Networks represents protein similarity networks, Multiple Kernels Learning represents multi-kernel learning, Fusionsimilarity network represents fusion similarity network, DAE represents denoising autoencoder, Loss represents loss, Decode represents decoding, Encode represents encoding, Drug representation represents drug representation, Protein representation represents target protein representation, Fully Connected Classification represents fully connected classification, Concatenating represents concatenation, and Prediction Score represents prediction score;

[0090] Step 1), based on the data sources of drugs and target proteins, calculate the similarities between drugs and target proteins, and construct similarity networks of drugs and target proteins in each data source;

[0091] The construction process of the similarity network of drugs and target proteins in each data source is as follows:

[0092] Step 1.1: Construction of drug similarity network;

[0093] Obtain drug data sources, including at least the chemical structure of the drug, drug-drug association, drug-disease association, and drug-drug side effect data; for each data source, use the Jaccard similarity coefficient to calculate the similarity between drugs, and generate an independent drug similarity network for each data source, which is expressed in a matrix form, in which each element represents a similarity score between a pair of drugs;

[0094] The similarity score is obtained by calculating the ratio of the intersection and union of two drugs on a specific data source.

[0095] Step 1.2: Construction of target protein similarity network;

[0096] Obtain target protein data sources, including at least target protein sequence information, target protein-target protein interaction, and target protein-disease association; for each data source, use the Jaccard similarity coefficient to calculate the similarity between target proteins, and generate an independent target protein similarity network for each data source;

[0097] The similarity score is obtained by calculating the ratio of the intersection and union of two target proteins on a specific data source.

[0098] Step 2), the multi-kernel learning algorithm MKL is used to integrate the similarity networks of drugs and target proteins respectively to obtain the comprehensive similarity matrices SD and SP of drugs and target proteins;

[0099] In the process of integrating the similarity network using the multi-core learning algorithm MKL, the objective function corresponding to the following formula is used to minimize the distance between the optimal kernel of the integrated similarity network and the ideal optimal kernel:

[0100] Objective function: ,

[0101] ,

[0102] , ,

[0103] ;

[0104] in, express The similarity network of the i-th express The weight vector of the similarity network, is the regularization parameter, yes The ideal optimal kernel of , β is a regularization parameter used to balance the similarity matrix fitting error and the weight of the weight regularization term; Indicates drug or target protein, When it indicates medicine, indicates the target protein; It represents the association matrix of drug-target protein; Indicates the integration of multiple The comprehensive similarity matrix generated after the similarity network is * yes The ideal optimal kernel matrix of express The number of similarity networks; represents the square of the Frobenius norm, Represents the integrated comprehensive similarity matrix SX and the ideal optimal kernel matrix * The difference between Represents the weight vector The regularization term is used to constrain the sparsity of full-time and avoid overfitting;

[0105] Weight solution formula:

[0106] ,

[0107] ,

[0108] ;

[0109] in: ,

[0110] ,

[0111] ;

[0112] Among them, trace represents the trace of the matrix;

[0113] The conversion formula is:

[0114]

[0115] ;

[0116] Represents drugs or target proteins, and the weights of each similarity network of drugs and target proteins are obtained according to the above formula.

[0117] In order to avoid overfitting during training and improve the generalization ability of the model, making the system performance more robust, an L2 penalty is imposed on the weight vector to limit the growth of authority.

[0118] The significance of multivariate data fusion, since the similarity between drugs and target proteins can be described from multiple perspectives (such as chemical structure, sequence information, functional association, etc.), a single data source may not be able to fully reflect their similarity. Therefore, by integrating the similarity network generated by multiple data sources, the complex relationship between drugs and target proteins can be more comprehensively captured, thereby improving the accuracy of subsequent prediction models.

[0119] The Multiple Kernel Learning (MKL) algorithm is used to integrate multiple similarity networks to generate a comprehensive similarity matrix kernel. Each similarity network can be regarded as a kernel function. The MKL algorithm reflects its importance in the comprehensive similarity matrix by assigning different weights to each kernel function. The optimal weight of each kernel function is determined by optimization algorithms (such as gradient descent or quadratic programming) so that the comprehensive similarity matrix can better reflect the true relationship between the drug and the target protein. Finally, the MKL algorithm weightedly fuses multiple similarity networks to generate a comprehensive similarity matrix kernel for subsequent drug-target protein association prediction models.

[0120] By using a multi-kernel learning algorithm, the model no longer relies on a single similarity network, but is able to comprehensively consider information from multiple data sources. This not only improves the model's prediction accuracy, but also enhances the model's robustness to noise and abnormal data, avoiding prediction errors caused by deviations from a single data source.

[0121] The drug comprehensive similarity matrix SD and the target protein comprehensive similarity matrix SP are calculated by the following formula:

[0122] ;

[0123] ;

[0124] in, Represent the number of drugs and target proteins, respectively. N D and N P are the number of similarity networks of drugs and target proteins, respectively, and R represents the real number domain.

[0125] Step 3), taking each row of data in the comprehensive similarity matrix SD and SP obtained in step 2) as the initial feature representation of the drug and the target protein, and introducing noise into the initial feature data through a denoising autoencoder; and restoring the initial feature data with the introduced noise;

[0126] Restoration is used to remove noise and restore data. After denoising, redundant or irrelevant information is removed, the core features of the data are retained, and it is more suitable for subsequent modeling tasks. This enhances the model's ability to learn potential patterns in the data and reduces the interference of noise on prediction accuracy.

[0127] The process of using each row of data in the comprehensive similarity matrix SD and SP obtained in step 2) as the initial feature representation of the drug and target protein, and introducing noise into the feature data through the denoising autoencoder is as follows:

[0128] First, the denoising autoencoder randomly generates noise data through Gaussian normal distribution, and adds the randomly generated noise to the initial input features as the initial data for dimensionality reduction of the denoising autoencoder, as shown in the following formula:

[0129] ;

[0130] ;

[0131] in, represents the mean of the distribution, represents variance; Represents the feature data of the original input, express The feature data after adding noise, represents the corresponding random noise;

[0132] Then, a denoising autoencoder is used to Reconstruct the data by reconstructing the data dimension Perform dimensionality reduction and reduce the reconstructed data by reducing the difference between the reconstructed data and the original input data. The loss between layers is used to continuously optimize the denoising autoencoder loss. The L-layer encoder is defined as follows:

[0133] ,

[0134] ,

[0135] …

[0136] ;

[0137] in represents the output of the encoder layer l, represents the raw input data of the first layer of the encoder, and Represent the weight matrix and bias matrix of the encoder layer l, respectively. represents a nonlinear activation function;

[0138] The L-layer decoder is defined as follows:

[0139] ,

[0140] ,

[0141] …

[0142] .

[0143] in, represents the output of the lth layer of the decoder, Represents the reconstructed data finally output by the decoder, and denote the weight matrix and bias matrix of the decoder layer l, respectively. represents a non-linear activation function.

[0144] The Softplus function and RMSProp algorithm are used to optimize the mean square error formula in the dimensionality reduction process. The mean square error calculation formula is as follows:

[0145] ;

[0146] Where L represents the loss function of the denoising autoencoder, represents the reconstructed data finally output by the decoder, and x represents the original input data of the denoising autoencoder;

[0147] Reduce the loss of data information during dimensionality reduction;

[0148] In step 4), the restored drug and target protein feature data processed in step 3) are spliced ​​to obtain the feature vector of the drug-target protein pair, and the association between the feature vectors of the drug-target protein pair is modeled using a multi-layer perceptron to obtain a drug and target protein association prediction model based on a multi-layer perceptron to predict the association between the drug and the target protein.

[0149] Prediction score of the association between drugs and target proteins Calculated by the following formula:

[0150] ;

[0151] in, is the feature vector of the input drug-target protein pair, , , and , , The corresponding weights and bias vectors determined in the drug-target protein association prediction model based on the multi-layer perceptron, The function calculation formula is: , e represents the natural exponent, and the Relu function calculation formula is , is the calculated variable of the input function.

[0152] The parameters in the drug-target protein association prediction model based on multi-layer perceptron are usually optimized using back propagation algorithm and gradient descent algorithm;

[0153] The drug-target protein association prediction model based on multi-layer perceptron uses the Adam algorithm and sets the learning rate to 0.001 to optimize the binary crossover loss:

[0154] ;

[0155] in, and They represent the positive and negative sample labels of drug-target protein pairs, respectively. Y is the true label, with a value of 0 or 1. N represents the number of drug-target protein pairs. Represents the prediction score of the association between drugs and target proteins, Represents the loss function of the drug-target protein association prediction model based on multi-layer perceptron.

[0156] Validation of the drug-target protein association prediction method based on multi-source data embedding

[0157] In order to verify the effectiveness of the MKLDAE method, the MKLDAE method provided by the technical solution example of the present invention is applied to a set of drug and target protein related data sets, including 708 drugs, 1512 proteins, 5603 diseases, 4192 drug side effects, and drug-drug, drug-disease, drug-drug side effects, drug-target protein, target protein-disease, target protein-drug association data, and the specific data information is shown in Table 1.

[0158] Table 1 Drug and protein related data information: .

[0159] The information of drugs, target proteins, diseases, and drug side effects were collected from the DrugBank (Version 3.0) database, the HPRD (Release 9) database, the Comparative Toxicogenomics database, and the SIDER (Version 2) database, respectively.

[0160] 1. Comparative experiments to verify the effectiveness of the algorithm:

[0161] In order to further verify the performance of MKLDAE, the model was compared with four drug-target protein association prediction models in recent years: MolTrans, AEFS, DTINet, and DTICNN. The four comparison models used their own optimal parameters, and evaluated the AUROC and AUPR values ​​of the models through ten-fold cross validation. Considering the differences in data information processing among different models, for the sake of fairness, MKLDAE, DTINet, and DTICNN used the same dataset, AEFS used the optimal dataset in its own experiment, and MolTrans used the BIOSNAP dataset. The specific comparison results are shown in Table 2. Figure 2 and 3 The ROC curves and PR curves of the five models are shown in Figure 2. The AUROC values ​​of MKLDAE, DTICNN, DTINet, AEFS, and MolTrans are 0.9528, 0.9390, 0.9147, 0.9081, and 0.8886, respectively, and the AUPR values ​​are 0.9595, 0.9445, 0.9323, 0.3696, and 0.9004, respectively. MKLDAE achieved the best results in the comparative experiment, proving its superiority over other methods in the process of drug-target protein association prediction.

[0162] Among them, MKLDAE is the method proposed in this scheme, MolTrans is the method disclosed in the literature Molecular Interaction Transformer for drug–target interaction prediction, AEFS is the method disclosed in the literature Autoencoder-based drug–target interaction prediction by preserving the consistency of chemical properties and functions of drugs; DTINet is the method disclosed in the literature A network integration approach for drug-target interaction prediction and computational drug repositioning from heterogeneous information; DTICNN is the method disclosed in the literature A learning-target based method for drug-interaction prediction based on feature representation learning and deep neural network.

[0163] Table 2 Comparative experiment of five methods corresponding to AUROC and AUPR values ​​and data set sources:

[0164] .

[0165] 2. Case analysis experiment to verify the effectiveness of the results:

[0166] In order to further demonstrate the performance of the MKLDAE model in discovering drug-target protein associations, a case analysis experiment was used for verification. By deleting the association information between amitriptyline (DB00321: Amitriptyline) and quetiapine (DB01224: Quetiapine) and all target proteins in the drug-target protein association dataset, the association between the remaining drugs and target proteins was used as a training set to train the model, and the model was evaluated by the accuracy of the top 20 data predicted in the test set. Tables 3 and 4 list in detail the comparison between the predicted labels of amitriptyline, quetiapine and target proteins predicted by the MKLDAE model and the true labels, and it is assumed that there is an association between the top 20 drugs and target proteins.

[0167] The accuracy of the association between amitriptyline and the top 20 target proteins reached 86.7%, and the prediction accuracy of quetiapine reached 100%. The case analysis experiment effectively proved that the MKLDAE model has real effect in predicting the association between drugs and target proteins.

[0168] Table 3 Information of the top 20 predicted target proteins of amitriptyline:

[0169] .

[0170] Table 4 Information on the top 20 predicted target proteins of quetiapine:

[0171] .

[0172] Example 2

[0173] A system for predicting drug-target protein association based on the above-mentioned multi-source data embedding method comprises:

[0174] The similarity network construction module calculates the similarities between drugs and target proteins based on the data sources of drugs and target proteins, and constructs the similarity network of drugs and target proteins in each data source;

[0175] The similarity matrix acquisition module uses the multi-kernel learning algorithm MKL to integrate the similarity networks of drugs and target proteins respectively to obtain the comprehensive similarity matrices SD and SP of drugs and target proteins;

[0176] The initial feature restoration module uses each row of data in the obtained comprehensive similarity matrix SD and SP as the initial feature representation of the drug and target protein, and introduces noise into the initial feature data through a noise reduction autoencoder; and restores the initial feature data with introduced noise;

[0177] The association prediction module splices the processed and restored feature data of the drug and target protein to obtain the feature vector of the drug-protein pair, and uses a multi-layer perceptron to model the association between the feature vectors of the drug-target protein pair, obtains a drug and target protein association prediction model based on a multi-layer perceptron, and predicts the association relationship between the drug and the target protein.

[0178] It should be understood that the specific implementation process of each module please refer to the above method content, the present invention will not be repeated here, and the division of the above functional modules is only for example illustration. In some embodiments, some functional modules can be merged, and some functional modules can be split. Each functional module can be implemented in software or hardware or a combination of software and hardware. Among them, the hardware and software equipment includes but is not limited to general-purpose computer equipment, programmable gate arrays, digital signal processors, microprocessors and their corresponding programming or burning software.

[0179] Example 3

[0180] A computer-readable storage medium stores a computer program, wherein the computer program is called by a processor to implement:

[0181] The above steps of the drug-target protein association prediction method based on multi-source data embedding.

[0182] For the specific implementation process of each step, please refer to the description of the above method.

[0183] The readable storage medium is a computer-readable storage medium, which may be an internal storage unit of the software and hardware device described in any of the aforementioned embodiments, such as a hard disk or memory of a controller. The readable storage medium may also be an external storage device of the controller, such as a plug-in hard disk, a smart memory card (Smart Media Card, SMC), a secure digital (Secure Digital, SD) card, a flash card (Flash Card), etc. equipped on the controller. Furthermore, the readable storage medium may also include both an internal storage unit of the controller and an external storage device. The readable storage medium is used to store the computer program and other programs and data required by the controller. The readable storage medium may also be used to temporarily store data that has been output or is to be output.

[0184] Example 4

[0185] According to an embodiment of the present invention, the present invention further provides an electronic device, including:

[0186] one or more processors;

[0187] a storage device for storing one or more programs,

[0188] When the one or more programs are executed by the one or more processors, the one or more processors implement the aforementioned method.

[0189] In specific use, users can interact with servers, which are also electronic devices, through electronic devices as terminal devices and based on the network to achieve functions such as receiving or sending messages. Terminal devices are generally various electronic devices equipped with display devices and used based on human-computer interfaces, including but not limited to smart phones, tablet computers, laptops and desktop computers. Various specific application software can be installed on the terminal devices as needed, including but not limited to web browser software, instant messaging software, social platform software, shopping software, etc.

[0190] The server is a network service end for providing various services. The location identification method provided in this embodiment is generally executed by the server. In actual application, under the necessary conditions, the terminal device can also directly perform location identification. Accordingly, the location identification device can be set on the server. Similarly, under the necessary conditions, the location identification device can also be set in the terminal device.

[0191] Based on this understanding, the technical solution of the present invention is essentially or the part that contributes to the prior art, or all or part of the technical solution can be embodied in the form of a software product, which is stored in a storage medium and includes several instructions for enabling a computer device (which can be a personal computer, a server, or a network device, etc.) to perform all or part of the steps of the method described in each embodiment of the present invention. The aforementioned readable storage medium includes: U disk, mobile hard disk, read-only memory (ROM, Read-Only Memory), random access memory (RAM, Random Access Memory), disk or optical disk, etc. Various media that can store program codes.

[0192] Those skilled in the art will appreciate that the embodiments of the present application can be provided as methods, systems, or computer program products. Therefore, the present application can adopt the form of a complete hardware embodiment, a complete software embodiment, or an embodiment combining software and hardware. Moreover, the present application can adopt the form of a computer program product implemented on one or more computer-readable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program codes. The present application is a device for realizing the function specified in one flow or multiple flows of the flow chart and / or one or multiple blocks of the block diagram with reference to the instructions executed by the processor according to the method, device (system) and computer program product of the embodiment of the present application. These computer program instructions can also be stored in a computer-readable memory that can guide a computer or other programmable data processing device to work in a specific manner, so that the instructions stored in the computer-readable memory produce a product including an instruction device, which realizes the function specified in one flow or multiple flows of the flow chart and / or one or multiple blocks of the block diagram. These computer program instructions may also be loaded onto a computer or other programmable data processing device so that a series of operational steps are executed on the computer or other programmable device to produce a computer-implemented process, whereby the instructions executed on the computer or other programmable device provide steps for implementing the functions specified in one or more processes in the flowchart and / or one or more boxes in the block diagram.

[0193] It should be emphasized that the examples described in the present invention are illustrative rather than restrictive, and therefore the present invention is not limited to the examples described in the specific embodiments. Any other embodiments derived by those skilled in the art based on the technical solution of the present invention that do not depart from the purpose and scope of the present invention, whether modified or replaced, also fall within the scope of protection of the present invention.

Claims

1. A drug-target protein association prediction method based on multi-source data embedding, characterized in that: The following steps are involved: Step 1), based on the data sources of drugs and target proteins, calculate the similarities between drugs and target proteins, and construct similarity networks of drugs and target proteins in each data source; Step 2), the multi-kernel learning algorithm MKL is used to integrate the similarity networks of drugs and target proteins respectively to obtain the comprehensive similarity matrices SD and SP of drugs and target proteins; Step 3), using each row of data in the comprehensive similarity matrix SD and SP obtained in step 2) as the initial feature representation of the drug and the target protein, and introducing noise into the initial feature data through a denoising autoencoder, and restoring the initial feature data with the introduced noise; ， ； in, represents randomly generated noise, represents the mean of the distribution, represents variance; Represents the feature data of the original input, express Feature data after adding noise; The encoder and decoder used for restoration are as follows: ； ； in, and denote the output of the encoder and decoder layer l, respectively. and Represent the weight matrix and bias matrix of the encoder layer l, respectively. represents a nonlinear activation function; Represents the reconstructed data finally output by the decoder, and Respectively represent the weight matrix and bias matrix of the decoder layer l; In step 4), the restored drug and target protein feature data processed in step 3) are spliced ​​to obtain the feature vector of the drug-protein pair, and the association between the feature vectors of the drug-protein pair is modeled using a multi-layer perceptron to obtain a drug-target protein association prediction model based on a multi-layer perceptron to predict the association between the drug and the target protein.

2. The method according to claim 1, characterized in that Prediction score of the association between drugs and target proteins Calculated by the following formula: ； in, is the feature vector of the input drug-target protein pair, , , and , , The corresponding weights and bias vectors determined in the drug-target protein association prediction model based on the multi-layer perceptron, The function calculation formula is: , e represents the natural exponent, and the Relu function calculation formula is , is the calculated variable of the input function.

3. The method according to claim 1, characterized in that: The construction process of the similarity network of drugs and target proteins in each data source is as follows: Step 1.1: Construction of drug similarity network; Obtain drug data sources, including at least the chemical structure of the drug, drug-drug association, drug-disease association, and drug-drug side effect data; for each data source, use the Jaccard similarity coefficient to calculate the similarity between drugs, and generate an independent drug similarity network for each data source, which is expressed in a matrix form, in which each element represents a similarity score between a pair of drugs; Step 1.2: Construction of target protein similarity network; Obtain target protein data sources, including at least target protein sequence information, target protein-target protein interactions, and target protein-disease associations; For each data source, the similarity between target proteins was calculated using the Jaccard similarity coefficient, and each data source generated an independent target protein similarity network.

4. The method according to claim 1, characterized in that The multi-core learning algorithm MKL is used to integrate the similarity networks of drugs and target proteins respectively, and weights are assigned to each similarity network to obtain the comprehensive similarity matrices SD and SP of drugs and target proteins: In the process of integrating the similarity network using the multi-core learning algorithm MKL, the objective function corresponding to the following formula is used to minimize the distance between the optimal kernel of the integrated similarity network and the ideal optimal kernel: Objective function: , ， ， ， ； in, express The similarity network of the i-th, express The weight vector of the similarity network, is the regularization parameter, yes The ideal optimal kernel of , β is a regularization parameter used to balance the similarity matrix fitting error and the weight of the weight regularization term; Indicates drug or target protein, Indicates drugs, indicates the target protein; It represents the association matrix of drug-target protein; Indicates the integration of multiple The comprehensive similarity matrix generated after the similarity network is * yes The ideal optimal kernel matrix of express The number of similarity networks; represents the square of the Frobenius norm, Represents the integrated comprehensive similarity matrix SX and the ideal optimal kernel matrix * The difference between Represents the weight vector The regularization term is used to constrain the sparsity of full-time and avoid overfitting; Weight solution formula: ， ， ； in: , ， ； Among them, trace represents the trace of the matrix; The conversion formula is: ； According to the above formula, the weights of each similarity network of drugs and target proteins are obtained.

5. The method according to claim 4, characterized in that The drug comprehensive similarity matrix SD and the target protein comprehensive similarity matrix SP are calculated by the following formula: ， ； in, Represent the number of drugs and target proteins, respectively. N D and N P are the number of similarity networks of drugs and target proteins, respectively, and R represents the real number domain.

6. The method according to claim 5, characterized in that The Softplus function and RMSProp algorithm are used to optimize the mean square error formula in the dimensionality reduction process. The mean square error calculation formula is as follows: ； Where L represents the loss function of the denoising autoencoder, Represents the reconstructed data finally output by the decoder, Represents the raw input data fed into the denoising autoencoder.

7. The method according to claim 1, characterized in that The drug-target protein association prediction model based on multi-layer perceptron uses the Adam algorithm and sets the learning rate to 0.001 to optimize the binary crossover loss: ； in, and They represent the positive and negative sample labels of drug-target protein pairs, respectively. Y is the true label, with a value of 0 or 1. N represents the number of drug-target protein pairs. Represents the prediction score of the association between drugs and target proteins, Represents the loss function of the drug-target protein association prediction model based on multi-layer perceptron.

8. A system for predicting drug-target protein association based on the multi-source data embedding method according to any one of claims 1 to 7, characterized in that: include: The similarity network construction module calculates the similarities between drugs and target proteins based on the data sources of drugs and target proteins, and constructs the similarity network of drugs and target proteins in each data source; The similarity matrix acquisition module uses the multi-kernel learning algorithm MKL to integrate the similarity networks of drugs and target proteins respectively to obtain the comprehensive similarity matrices SD and SP of drugs and target proteins; The initial feature restoration module uses each row of data in the obtained comprehensive similarity matrix SD and SP as the initial feature representation of the drug and target protein, and introduces noise into the initial feature data through a denoising autoencoder; And restore the initial feature data with introduced noise; The association prediction module splices the processed and restored feature data of the drug and target protein to obtain the feature vector of the drug-target protein pair, and uses a multi-layer perceptron to model the association between the feature vectors of the drug-target protein pair, obtains a drug and target protein association prediction model based on a multi-layer perceptron, and predicts the association relationship between the drug and the target protein.

9. A computer-readable storage medium, characterized in that: A computer program is stored, which is called by a processor to implement: The steps of the method according to any one of claims 1 to 7.

10. An electronic device, comprising: one or more processors; A storage device for storing one or more programs; It is characterized in that the one or more processors execute the one or more programs to implement the method according to any one of claims 1-7.

Images