A method and system for predicting drug-target interaction relationships

Abstract

The present invention discloses a method and system for predicting drug-target interaction relationships. The method comprises: obtaining three-dimensional drug structure data and three-dimensional target protein data as training sets; constructing a drug-target interaction prediction model; using the training set as input to the drug-target interaction prediction model, establishing a loss function based on the error between the predicted value output by the drug-target interaction prediction model and the true value; and training the drug-target interaction prediction model using a gradient descent method until the model accuracy meets the requirements, thereby obtaining an optimal drug-target interaction prediction model; and predicting the interaction relationship between the drug and the target protein using the optimal drug-target interaction prediction model. The drug-target interaction prediction method and system disclosed by the present invention can effectively extract and utilize spatial features, thereby capturing complex patterns in the data.

Description

Technical Field

[0001] The present invention relates to the technical field of drug target prediction, and in particular to a method and system for predicting drug-target interaction relationships. Background Art

[0002] Drug research and development refers to the process of introducing new drugs to the market through scientific methods and technologies, through a series of research and experimental activities, from drug discovery, drug design, drug development to drug marketing, to meet clinical needs, treat diseases or improve health conditions. Drug target prediction is of great significance in drug research and development. Accurate drug target prediction helps improve drug research and development efficiency, reduce costs, expand the scope of treatment, and thus promote the discovery of new drugs and progress in disease treatment. Drugs usually refer to small molecules, while targets usually refer to large molecules such as proteins. When a drug binds to a target, the drug may change the function of the biological target, which is called drug-target interaction (DTI). The prediction of DTI plays a key role in drug research and development.

[0003] In the early stages of drug development, researchers often rely on experimental techniques such as X-ray crystallography and nuclear magnetic resonance to obtain detailed three-dimensional structural information of drugs. These techniques reveal the spatial arrangement of drug molecules at the atomic level, providing important structural data for scientific research. With the development of Internet technology, more and more drug structure information has been integrated into public databases, making it easier for researchers to access and use this data. In terms of representing the three-dimensional structural information of drugs, researchers often use methods such as molecular descriptors or structural fingerprints. These methods can convert complex drug structures into computer-processable numerical forms, which can then be used to build predictive models or perform similarity searches. With the rise of personalized medicine, the three-dimensional structural information of drugs will also provide strong support for the development of precise medication and individualized treatment plans.

[0004] With the rapid development of bioinformatics and computational technology, methods for drug target prediction based on the three-dimensional structures of drugs and proteins have attracted much attention and demonstrated great potential and application value. Currently, more and more drug and target databases have been published, such as DrugBank (Drugbank 5.0: A major update to the drugbank database for 2018 [J]. Nucleic Acids Res, 2018, 46 (D1): D1074-D1082), UniProt (Consortium U P. The Universal Protein Resource (UniProt) in 2010 [J]. Nucleic Acids Research, 2010, 38 (1): 142-148.), and PDB (The Protein Data Bank [J]. Genetica, 1999, 106 (1): 149-158). These methods will help to discover patterns and unknown connections between data from massive amounts of data using computational methods. At the same time, investment in bioinformatics, computational chemistry and artificial intelligence continues to increase, and computer-assisted drug target protein research has gradually become a new research hotspot.

[0005] In recent years, with the widespread application of computational prediction models in processing complex data, many researchers have focused on using machine learning methods to predict drug-target interactions. For example, Yamanishi et al. (Drug-target interaction prediction from chemical, genomic and pharmacological data in an integrated framework [J]. Bioinformatics, 2010, 26 (12): i246-i254) integrated chemical information and genomic features with pharmacological data into the same framework, achieving large-scale prediction of potential drug-target interactions. Wang et al. (Computational methods for the prediction of drug-target interactions from drug fingerprints and protein sequences by stacked auto-encoder deep neural network [J]. International Symposium on Bioinformatics Research and Applications, 2017, 10 (30): 46-58) used a stacked autoencoder to take the sequence information of the target as input, and then imported the characteristics and fingerprint information of the drug into the encoding layer to predict potential drug-target interactions. Palma et al. (Drug-target interaction prediction using semantic similarity and edge partitioning [J]. International Semantic Web Conference, 2014, 8 (7): 131-146) proposed a prediction model called SemEP. They assumed that similar drugs are more likely to be associated with common targets. They incorporated semantic similarity and edge partitioning methods into the link prediction model framework. The former was used to measure the semantic relevance between nodes, and the latter was combined to define clear boundaries in the network to identify new drug-target interactions.Based on the interaction relationship between molecules, Lee et al. (Drug repositioning using drug-disease vectors based on an integrated network[J].BMC Bioinformatics,2018,19(1):1236-1286) constructed a directed network between protein and gene data to infer the shortest path between targets and genes.

[0006] Although the use of three-dimensional structural information in drug-target prediction has great potential, it still faces a series of challenges. First, obtaining high-quality three-dimensional structural data of drugs and proteins requires expensive and complex experimental techniques, as well as sophisticated data processing processes. This process is not only time-consuming and labor-intensive, but may also be affected by factors such as experimental conditions and quality control, affecting the accuracy and reliability of the data. Secondly, for these massive amounts of three-dimensional structural data, how to extract key features from them and establish computational models that can effectively predict drug-target interactions is a very challenging task. Advanced machine learning and deep learning technologies are needed to convert complex structural information into computable features to achieve accurate prediction and analysis. Therefore, overcoming these challenges requires interdisciplinary collaboration and continuous exploration of innovative methods to promote progress and development in the field of drug research and development. Summary of the Invention

[0007] The technical problem to be solved by the present invention is to solve the problem of low accuracy in identifying the interaction relationship between drugs and target proteins.

[0008] In order to solve the above technical problems, the present invention provides the following technical solutions:

[0009] A method for predicting drug-target interaction relationships, comprising:

[0010] S10, obtaining drug 3D structure data and target protein 3D data as training sets;

[0011] S20, constructing a drug-target interaction prediction model;

[0012] S30, using the training set as input of the drug-target interaction prediction model, establishing a loss function according to the error between the predicted value output by the drug-target interaction prediction model and the true value, and training the drug-target interaction prediction model using a gradient descent method until the model accuracy meets the requirements, thereby obtaining an optimal drug-target interaction prediction model;

[0013] S40, predicting the interaction relationship between drugs and target proteins through the optimal drug-target interaction prediction model;

[0014] Among them, S20 builds a drug-target interaction prediction model through the following methods:

[0015] The framework of the drug-target interaction prediction model includes a drug graph convolutional network, a target graph convolutional network, a first fully connected layer, a feature fusion layer, and a second fully connected layer;

[0016] S21, converting the drug three-dimensional structure data and the target protein three-dimensional data into a drug molecular structure three-dimensional matrix and a target protein molecular structure three-dimensional matrix respectively;

[0017] S22, taking the three-dimensional matrix of drug molecular structure as the input of the drug graph convolutional network, and outputting the three-dimensional matrix of drug features;

[0018] S23, taking the target protein molecular structure three-dimensional matrix as the input of the target graph convolutional network, and outputting the target protein feature three-dimensional matrix;

[0019] S24, taking the drug feature three-dimensional matrix and the target protein feature three-dimensional matrix as the input of the first fully connected layer, and outputting the drug three-dimensional feature vector and the target protein three-dimensional feature vector

[0020] S25, inputting the drug three-dimensional feature vector and the target protein three-dimensional feature vector into the feature fusion layer for splicing and fusion to obtain a fused feature vector;

[0021] S26, introduces the fused feature vector into the second fully connected layer for dimensionality reduction, and outputs a one-dimensional correlation value, which represents the predicted value of the interaction strength between the drug and the target protein.

[0022] In one embodiment of the present invention, the three-dimensional structure data of the drug and the three-dimensional data of the target protein are obtained by:

[0023] S11, obtain the matched and screened drug data and target protein data from the DrugBank database;

[0024] S12, extracting drug three-dimensional structure data from the PubChem database based on drug data;

[0025] S13, extracting target protein three-dimensional data from the Alphafold database based on the target protein data.

[0026] In one embodiment of the present invention, the drug graph convolutional network and the target graph convolutional network both include three convolutional layers, and a ReLU activation function and a maximum pooling layer are introduced after each convolutional layer.

[0027] In one embodiment of the present invention, the feature extraction formulas of the drug graph convolutional network and the target graph convolutional network are both:

[0028]

[0029] Where H (l) is the node feature matrix of the lth layer, H (l+1) is the node feature matrix of the l+1th layer, is the normalized adjacency matrix, W (l) is the weight matrix of the lth layer, and σ is the activation function.

[0030] In one embodiment of the present invention, the three-dimensional matrix of drug molecular structure is obtained by:

[0031] Extract the coordinate values ​​and atom types of drug molecules from the three-dimensional structure data of drugs;

[0032] Traversing the coordinate values ​​of the spatial position of each atom in the drug molecule to determine the minimum coordinate value; subtracting the minimum coordinate value from the coordinate value corresponding to the atom to obtain the normalized coordinate value of the drug molecule;

[0033] According to different atom types, a unique value is assigned to each type, an all-zero matrix is ​​initialized, and the corresponding positions are filled with the values ​​of the corresponding atom types according to the normalized coordinate values ​​of the drug molecules to obtain the initial three-dimensional matrix of the drug molecule structure;

[0034] The size range of the initial drug molecule structure three-dimensional matrix is ​​adjusted to obtain the final drug molecule three-dimensional matrix.

[0035] In one embodiment of the present invention, the three-dimensional matrix of the target protein molecular structure is obtained by:

[0036] Extract protein structural information from target protein three-dimensional data;

[0037] Traverse each atom in the protein structure, obtain its coordinates and element type, and add them to the atomic coordinate list and atomic element type list respectively;

[0038] Convert the atomic coordinate list into a NumPy array. After preprocessing the NumPy array, obtain the minimum value of the protein structure on the three coordinate axes. The three-axis coordinate data in the NumPy array are divided by the corresponding multiples of the minimum value of the protein structure on the three coordinate axes and rounded up to obtain the resized atomic coordinate matrix.

[0039] Create an all-zero matrix that matches the adjusted atomic coordinate matrix and the matrix size to store the protein structure information. According to the adjusted atomic coordinate matrix and atomic element type list, add the corresponding atomic coordinate matrix and atomic element to the all-zero matrix to obtain the final three-dimensional matrix of the target protein molecular structure.

[0040] The present invention also provides a drug-target interaction relationship prediction system, which applies the above-mentioned drug-target interaction relationship prediction method and comprises:

[0041] The data module is used to obtain the three-dimensional structure data of drugs and target proteins as training sets;

[0042] Prediction model building module, used to build drug-target interaction prediction models;

[0043] A training module is used to use the training set as the input of the drug-target interaction prediction model, establish a loss function based on the error between the predicted value output by the drug-target interaction prediction model and the true value, and train the drug-target interaction prediction model using the gradient descent method until the model accuracy meets the requirements, thereby obtaining the optimal drug-target interaction prediction model;

[0044] A prediction module is used to predict the interaction relationship between drugs and target proteins through the optimal drug-target interaction prediction model;

[0045] Among them, the drug-target interaction prediction model is constructed in the following ways:

[0046] The framework of the drug-target interaction prediction model includes a drug graph convolutional network, a target graph convolutional network, a first fully connected layer, a feature fusion layer, and a second fully connected layer;

[0047] S21, converting the drug three-dimensional structure data and the target protein three-dimensional data into a drug molecular structure three-dimensional matrix and a target protein molecular structure three-dimensional matrix respectively;

[0048] S22, taking the three-dimensional matrix of drug molecular structure as the input of the drug graph convolutional network, and outputting the three-dimensional matrix of drug features;

[0049] S23, taking the target protein molecular structure three-dimensional matrix as the input of the target graph convolutional network, and outputting the target protein feature three-dimensional matrix;

[0050] S24, taking the drug feature three-dimensional matrix and the target protein feature three-dimensional matrix as the input of the first fully connected layer, and outputting the drug three-dimensional feature vector and the target protein three-dimensional feature vector

[0051] S25, inputting the drug three-dimensional feature vector and the target protein three-dimensional feature vector into the feature fusion layer for splicing and fusion to obtain a fused feature vector;

[0052] S26, introduces the fused feature vector into the second fully connected layer for dimensionality reduction, and outputs a one-dimensional correlation value, which represents the predicted value of the interaction strength between the drug and the target protein.

[0053] In one embodiment of the present invention, the feature extraction formulas of the drug graph convolutional network and the target graph convolutional network are both:

[0054]

[0055] Where H (l) is the node feature matrix of the lth layer, H (l+1) is the node feature matrix of the l+1th layer, is the normalized adjacency matrix, W (l) is the weight matrix of the lth layer, and σ is the activation function.

[0056] In one embodiment of the present invention, the drug graph convolutional network and the target graph convolutional network both include three convolutional layers, and a ReLU activation function and a maximum pooling layer are introduced after each convolutional layer.

[0057] Compared with the existing technology, the beneficial effects of the present invention are: the present invention introduces three-dimensional structural information as input, overcomes the limitations of traditional methods, provides richer spatial relationships and stereochemical characteristics, enables the model to more accurately identify the interaction between drugs and targets, and enables the model to understand and analyze data from a higher dimension and deeper level.

[0058] The present invention uses graph convolutional networks to create models, which have advantages in processing images and three-dimensional data. Through convolution operations, the model can effectively extract and utilize spatial features, thereby capturing complex patterns in the data. BRIEF DESCRIPTION OF THE DRAWINGS

[0059] Figure 1 Schematic diagram of a method for predicting drug-target interaction relationships according to an embodiment of the present invention.

[0060] Figure 2 Schematic diagram of the framework of the drug-target interaction prediction model according to an embodiment of the present invention.

[0061] Figure 3 Schematic diagram of the drug graph convolutional network and target graph convolutional network according to an embodiment of the present invention.

[0062] Figure 4 Schematic diagram of comparison results based on different models according to an embodiment of the present invention.

[0063] Figure 5 This is a block diagram of a drug-target interaction relationship prediction system according to an embodiment of the present invention. DETAILED DESCRIPTION

[0064] To facilitate those skilled in the art to understand the technical solution of the present invention, the technical solution of the present invention is further described with reference to the accompanying drawings.

[0065] The terms "first" and "second" are used for descriptive purposes only and should not be understood to indicate or imply relative importance or implicitly specify the number of the technical features indicated. Therefore, a feature specified as "first" or "second" may explicitly or implicitly include one or more of the features. In the description of this application, "plurality" means two or more, unless otherwise specifically defined.

[0066] See also Figure 1 As shown, the present invention provides a method for predicting drug-target interaction relationships, comprising:

[0067] S10, obtaining drug three-dimensional structure data and target protein three-dimensional data as training sets.

[0068] In one embodiment of the present invention, the three-dimensional structure data of the drug and the three-dimensional data of the target protein are obtained by:

[0069] S11, obtain the matched drug data and target protein data from the DrugBank database.

[0070] Both drug data and target protein data come from the DrugBank database, originally created by medical researchers at the University of Alberta in Canada. It is a drug database containing detailed information on various drugs. The database provides data on drug chemical structures, pharmacological properties, biological activity, pharmacokinetics, toxicity information, and clinical applications. DrugBank also provides detailed information on drug interactions and drug metabolic pathways. Furthermore, it provides information on drug therapeutic classes, indications, contraindications, and side effects, supporting researchers in comprehensive analysis and evaluation during drug design and screening.

[0071] In this example, 7,800 drug data sets and 4,920 target protein data sets were extracted from the DrugBank database. After matching and screening the two data sets, the final data set was obtained: 5,945 drug data sets and 4,920 target protein data sets. The data information is shown in Table 1. In Table 1, drugbank_id represents drug data, and uniprot_id represents target protein data.

[0072] Table 1 Drug-target protein data information

[0073]

[0074] S12, extracting drug three-dimensional structure data from the PubChem database based on the drug data.

[0075] In this embodiment, the drug three-dimensional structure data is extracted from the PubChem database based on the final drug data obtained in step S11 and saved in .sdf format.

[0076] The PubChem database, maintained by the National Center for Biotechnology Information (NCBI), a subsidiary of the National Institutes of Health (NIH), is a database of small organic molecule bioactivity data. It contains a wide range of molecular material information, including the physicochemical properties, biological activity, safety, and toxicity of chemical substances. The PubChem database is rich in information, allowing users to search for related compounds by name, molecular formula, structure, and other identifiers, and access patent information, literature citations, and other data.

[0077] S13, extracting target protein three-dimensional data from the Alphafold database based on the target protein data.

[0078] In this embodiment, the target protein three-dimensional data is extracted from the Alphafold database based on the target protein data obtained in step S11 and saved in a .pdb format.

[0079] The AlphaFold database is a large-scale protein structure prediction database developed by DeepMind. Based on the AlphaFold2 model, it can predict the three-dimensional structure of proteins from their amino acid sequences with extremely high accuracy. The database contains structural information for over 200 million proteins, covering nearly all known protein sequences.

[0080] S20, construct a drug-target interaction prediction model.

[0081] In one embodiment of the present invention, the framework of the drug-target interaction prediction model includes a drug graph convolutional network, a target graph convolutional network, a first fully connected layer, a feature fusion layer, and a second fully connected layer. The drug-target interaction prediction model is constructed in the following manner:

[0082] S21, converting the drug three-dimensional structure data and the target protein three-dimensional data into a drug molecular structure three-dimensional matrix and a target protein molecular structure three-dimensional matrix respectively.

[0083] In this embodiment, the complex three-dimensional structural data of drugs and targets are not suitable for machine learning models. In order to fully utilize the data information, the present invention converts the structural data obtained in step S10 into a three-dimensional matrix, so that it can apply advanced algorithms such as convolutional networks to predict drug-target interaction relationships.

[0084] In this embodiment, the three-dimensional matrix of drug molecular structure is obtained by the following method:

[0085] Extract drug molecular structure information: Use cheminformatics tools such as the Chem module of the RDKit library to extract the coordinate information and atom types of drug molecules from the .sdf file of the drug's three-dimensional structure data. This data describes the spatial position of each atom in the molecule and the type of chemical element to which it belongs.

[0086] Coordinate normalization: Traverse each atom in the drug molecule and obtain the coordinate information of its spatial position. Determine the minimum value of each coordinate (the smallest coordinate value on the x, y, and z axes) and subtract the corresponding minimum value from all coordinates. This will shift all coordinates to the non-negative range to ensure correct display in the matrix.

[0087] Assign and fill the matrix: assign a unique value to each type according to different atom types, initialize an all-zero matrix, and fill the corresponding positions with the values ​​of the corresponding atom types according to the coordinate information of the drug molecular structure to obtain the initial three-dimensional matrix of the drug molecular structure;

[0088] Adjusting the matrix size: adjusting the size range of the initial three-dimensional matrix of the drug molecule structure to ultimately obtain a three-dimensional matrix of the drug molecule.

[0089] In this embodiment, the three-dimensional matrix of the target protein molecular structure is obtained by the following method:

[0090] File parsing and information extraction: Use the PDBParser module in the Biopython library to parse the .pdf file containing the target protein's 3D data and retrieve the protein's structural information, including the model, chain, residue, and atomic hierarchies. Iterate over each atom in the protein structure, obtain its coordinates and element type, and add them to the coordinate list and element type list, respectively.

[0091] Coordinate processing and matrix resizing: Convert the coordinate list to a NumPy array and preprocess it to resize the coordinates to non-negative integers by subtracting the minimum coordinate value. From the preprocessed NumPy array, obtain the minimum value of the protein structure on the three coordinate axes, divide it by the corresponding multiple, and round it up to obtain the resized protein coordinate matrix.

[0092] Create and fill the matrix: Create an all-zero matrix of the same size as the resized protein coordinate matrix to store the protein structure information. Add the corresponding values ​​to the matrix according to the adjusted coordinates of the atoms and the element type. The value corresponding to the atomic element type can be obtained through a pre-defined dictionary (the dictionary key is the representation of the atomic element type, and the value is the specific numerical value associated with the atomic element type). If there is no corresponding value in the dictionary, it defaults to 0.

[0093] Return result: Returns the generated three-dimensional matrix of the target protein molecular structure, in which specific values ​​are used to represent the existence and positional relationship of different atoms, and finally obtains the three-dimensional matrix of the target protein molecular structure.

[0094] S22, the three-dimensional matrix of drug molecular structure is used as the input of the drug graph convolutional network, and the three-dimensional matrix of drug features is obtained as the output.

[0095] S23, takes the target protein molecular structure three-dimensional matrix as the input of the target graph convolutional network, and outputs the target protein feature three-dimensional matrix.

[0096] In this embodiment, the drug graph convolutional network and the target graph convolutional network both include three convolutional layers, and a ReLU activation function and a maximum pooling layer are introduced after each convolutional layer.

[0097] The basic principle of graph convolutional networks is to extract feature information and topological relationships between nodes by performing convolution operations on graph structured data. In graph convolutional networks, each node is regarded as a data processing unit, and the connection relationship between nodes constitutes the data propagation path. Through layer-by-layer convolutional networks, the neighborhood information of nodes can be gradually aggregated, thereby achieving the capture and expression of global features. Specifically, the feature extraction formulas for drug graph convolutional networks and target graph convolutional networks are:

[0098]

[0099] Where H (l) is the node feature matrix of the lth layer, H (l+1) is the node feature matrix of the l+1th layer, is the normalized adjacency matrix, W (l) is the weight matrix of the lth layer, and σ is the activation function.

[0100] In this example, the model takes the three-dimensional matrix of the drug molecular structure and the three-dimensional matrix of the target protein molecular structure as input, and performs convolution on each matrix using a three-layer three-dimensional convolutional neural network to achieve an accurate feature extraction process. The specific process is as follows:

[0101] Extracting local features from a three-dimensional matrix: The three-dimensional matrix of the drug molecular structure and the three-dimensional matrix of the target protein molecular structure are used as input. The input channel is 1, indicating that each position has only one eigenvalue. A 3*3*3 convolution kernel is used to scan the data in three-dimensional space for local area feature extraction, extracting preliminary local spatial features. The padding size is 2*2*2 to prevent the loss of matrix edge information during the convolution process, thereby preserving the edge feature information. After the convolution operation, 16 different feature maps are generated for each position. This means that the network will learn 16 different feature representations, which will be used to capture multiple spatial patterns in the input data, that is, the output channel is 16.

[0102] Further process the features extracted by the first layer: use the 16 feature maps output by the previous step as input data, with 16 input channels, use a convolution kernel of size 3*3*3, padding size of 2*2*2 to retain edge information, and output channels of 32.

[0103] Key feature extraction: The 32 feature maps output from the previous step are used as input data. A 3*3*3 convolution kernel is used to further extract higher-level features and integrate these features into a single feature representation. Padding size is 2*2*2 to preserve edge information. After three convolutional layers, the output channel is 1.

[0104] In order to make feature extraction more accurate, the maximum pooling layer and ReLU activation function are introduced after each convolution layer.

[0105] The maximum pooling layer is used to reduce the dimension of the feature map while retaining important features. The pooling kernel size is 2*2*2. By sampling in three-dimensional space, the maximum value in the local area is extracted, thereby reducing the size of the feature map.

[0106] Use the ReLU activation function to introduce nonlinear characteristics and improve the representation ability of the model. In the ReLU activation function, neurons with inputs less than zero will not be activated, which means that in some cases, only a part of the neurons will be activated, making the network more sparse, helping to reduce calculations and improve efficiency. The formula is as follows:

[0107] ReLU(x)=max(0,x);

[0108] Among them, x represents the input value, ReLU(x) represents the specific activation function, and max represents the maximum value.

[0109] After feature extraction through three convolutional layers, the three-dimensional matrix of drug features and the three-dimensional matrix of target protein features are obtained.

[0110] In this embodiment, the three-dimensional matrix of drug features and the three-dimensional matrix of target protein features are introduced into the first fully connected layer. The first fully connected layer is a three-layer fully connected layer structure. The feature vectors of the two are extracted and fused to obtain a fused feature vector. The second fully connected layer is then introduced. The second fully connected layer is also a three-layer fully connected layer. The fused feature vector is converted into a one-dimensional correlation value, which represents the interaction strength between the drug and the target protein. The specific process is as follows:

[0111] S24, taking the drug feature three-dimensional matrix and the target protein feature three-dimensional matrix as the input of the first fully connected layer, and outputting the obtained drug three-dimensional feature vector and target protein three-dimensional feature vector.

[0112] In this embodiment, the first fully connected layer in the first fully connected layer converts the input vector into a 2048-dimensional vector. It performs preliminary conversion and high-dimensional representation on the input features, increases the dimension of the feature space, and enables the network to capture more complex features and patterns.

[0113] The second fully connected layer in the first fully connected layer: The input vector is converted from 2048 dimensions to 1024 dimensions, further compressing the feature space while retaining key information. By reducing the dimension, redundant information is removed and computational efficiency is improved.

[0114] The third fully connected layer in the first fully connected layer converts the input vector from 1024 dimensions to 128 dimensions, generating a compact low-dimensional representation for representing the key features of proteins or drug molecules.

[0115] S25, inputting the three-dimensional feature vector of the drug and the three-dimensional feature vector of the target protein into the feature fusion layer for splicing and fusion to obtain a fused feature vector.

[0116] In this embodiment, the three-dimensional feature vector of the drug and the three-dimensional feature vector of the target protein are spliced, and the ReLU activation function is introduced, and the regularization technology Dropout layer is added to reduce overfitting, so that the 128-dimensional feature vectors of the drug and the target protein are fused into a 256-dimensional fused feature vector.

[0117] S26, introduces the fused feature vector into the second fully connected layer for dimensionality reduction, and outputs a one-dimensional correlation value, which represents the predicted value of the interaction strength between the drug and the target protein.

[0118] The fused feature vector is converted into a one-dimensional correlation value through three fully connected layers, which represents the interaction strength between the drug and the target protein.

[0119] S30, using the training set as the input of the drug-target interaction prediction model, establishing a loss function based on the error between the predicted value output by the drug-target interaction prediction model and the true value, and using the gradient descent method to train the drug-target interaction prediction model until the model accuracy meets the requirements, thereby obtaining the optimal drug-target interaction prediction model.

[0120] In this embodiment, a cross entropy loss function suitable for drug target prediction tasks is used to obtain the model's calculation accuracy by calculating the difference between the model's prediction results and the true labels. Specifically, the formula of the loss function is:

[0121]

[0122] Where, is the loss function, C is the number of categories, y iis the indicator function of the true label (1 if the sample belongs to class i, 0 otherwise), It is expressed as the predicted probability of the drug-target interaction prediction model for class i. The function can clearly indicate the difference between the model's prediction and the true label in terms of value. Log is the logarithmic function.

[0123] By using gradient descent to optimize the loss function, the gradient (partial derivative) of the loss function with respect to the model parameters is calculated, that is, the rate at which the loss function changes with the parameters. Then, the model parameters are adjusted according to the direction and magnitude of the gradient to reduce the value of the loss function. The parameter update formula is:

[0124]

[0125] Where θ represents the parameters of the drug-target interaction prediction model, η is the learning rate (a small positive number), is the loss function, the gradient of L with respect to the parameters θ.

[0126] In this embodiment, accuracy, recall, precision, and F1 score are used as evaluation indicators. These indicators can evaluate the optimal drug-target interaction prediction model from different perspectives.

[0127] A five-fold cross-validation was performed on the drug three-dimensional structure data and the target protein three-dimensional data, and the final experimental results of the present invention were obtained by removing the average value.

[0128] Accuracy is an intuitive indicator for evaluating whether the model prediction is correct or not. It indicates the proportion of samples that are correctly predicted as positive or negative in all predictions. The formula is as follows:

[0129]

[0130] Recall is the ability of a model to identify all positive samples. In the context of drug-target interaction prediction, recall directly affects whether the model can discover all potential drug targets as much as possible. A model with a high recall rate means it misses fewer true drug targets. The formula is as follows:

[0131]

[0132] The purpose of precision is to provide a way to measure the accuracy of the model in predicting the positive class. Its formula is as follows:

[0133]

[0134] The F1 score is a comprehensive evaluation indicator of precision and recall. It aims to find a balance point so that the model can have a good recall rate while maintaining a high accuracy rate. Its formula is as follows:

[0135]

[0136] Among them, TP represents the number of samples correctly classified as positive; TN represents the number of samples correctly classified as negative; FP represents the number of samples incorrectly classified as positive; FN represents the number of samples incorrectly classified as negative.

[0137] Comparative experiments were performed to verify the advantages and effectiveness of the optimal drug-target interaction prediction model:

[0138] Drug-target interaction prediction models based on traditional machine learning have been introduced. For example, support vector machines (SVMs) achieve binary classification by finding the optimal splitting hyperplane, and random forests, based on decision tree ensembles, construct multiple trees through bootstrap sampling and random feature selection, and then aggregate the outputs of each tree. These models have been shown to have some predictive effectiveness in previous studies, but they generally require extensive feature engineering to extract effective input features.

[0139] In addition, a deep neural network model (DNN) that has achieved remarkable results in the field of bioinformatics in recent years was selected for comparison. This model performs well in processing image sequences and time series data, but may face challenges in processing complex drug and protein three-dimensional structure data.

[0140] The experimental results based on the above method are shown in Table 2:

[0141] Table 2 Comparison results of different models

[0142] Accuracy Recall Precision F1 Support Vector Machine 0.45 0.40 0.38 0.39 Random Forest 0.54 0.55 0.52 0.53 DNN 0.50 0.55 0.47 0.51 Graph Convolutional Networks 0.63 0.45 0.77 0.57

[0143] In order to see the results more clearly, the experimental results are shown in a bar graph, as shown in the attached figure. Figure 4 shown.

[0144] By comparing these models with the graph convolutional network proposed in this paper, it is found that the graph convolutional network performs best in most indicators, especially in accuracy and precision. This paper can more accurately evaluate the unique advantages and potential of graph convolutional networks in drug target prediction tasks.

[0145] See also Figure 5 As shown, the present invention also provides a drug-target interaction relationship prediction system, which applies the above-mentioned drug-target interaction relationship prediction method, including:

[0146] The data module is used to obtain the three-dimensional structure data of drugs and target proteins as training sets;

[0147] Prediction model building module, used to build drug-target interaction prediction models;

[0148] A training module is used to use the training set as the input of the drug-target interaction prediction model, establish a loss function based on the error between the predicted value output by the drug-target interaction prediction model and the true value, and train the drug-target interaction prediction model using the gradient descent method until the model accuracy meets the requirements, thereby obtaining the optimal drug-target interaction prediction model;

[0149] A prediction module is used to predict the interaction relationship between drugs and target proteins through the optimal drug-target interaction prediction model;

[0150] Among them, the drug-target interaction prediction model is constructed in the following ways:

[0151] The framework of the drug-target interaction prediction model includes a drug graph convolutional network, a target graph convolutional network, a first fully connected layer, a feature fusion layer, and a second fully connected layer;

[0152] S21, converting the drug three-dimensional structure data and the target protein three-dimensional data into a drug molecular structure three-dimensional matrix and a target protein molecular structure three-dimensional matrix respectively;

[0153] S22, taking the three-dimensional matrix of drug molecular structure as the input of the drug graph convolutional network, and outputting the three-dimensional matrix of drug features;

[0154] S23, taking the target protein molecular structure three-dimensional matrix as the input of the target graph convolutional network, and outputting the target protein feature three-dimensional matrix;

[0155] S24, taking the drug feature three-dimensional matrix and the target protein feature three-dimensional matrix as the input of the first fully connected layer, and outputting the drug three-dimensional feature vector and the target protein three-dimensional feature vector

[0156] S25, inputting the drug three-dimensional feature vector and the target protein three-dimensional feature vector into the feature fusion layer for splicing and fusion to obtain a fused feature vector;

[0157] S26, introduces the fused feature vector into the second fully connected layer for dimensionality reduction, and outputs a one-dimensional correlation value, which represents the predicted value of the interaction strength between the drug and the target protein.

[0158] In one embodiment of the present invention, the feature extraction formulas of the drug graph convolutional network and the target graph convolutional network are both:

[0159]

[0160] Where H (l) is the node feature matrix of the lth layer, H (l+1) is the node feature matrix of the l+1th layer, is the normalized adjacency matrix, W (l) is the weight matrix of the lth layer, and σ is the activation function.

[0161] Both the drug graph convolutional network and the target graph convolutional network include three convolutional layers, and the ReLU activation function and the maximum pooling layer are introduced after each convolutional layer.

[0162] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above and that the invention can be embodied in other specific forms without departing from the spirit or essential characteristics of the invention. Therefore, the embodiments should be considered in all respects as illustrative and non-restrictive, and the scope of the invention is defined by the appended claims rather than the foregoing description. It is intended that all variations within the meaning and range of equivalents of the claims be embraced herein, and any reference signs in the claims should not be construed as limiting the claims to which they relate.

[0163] The above-mentioned embodiments merely represent the implementation methods of the invention. The protection scope of the present invention is not limited to the above-mentioned embodiments. For those skilled in the art, several variations and improvements can be made without departing from the concept of the present invention, which all fall within the protection scope of the present invention.

Claims

1. A method for predicting drug-target interaction relationships, characterized in that: include: S10, obtaining drug 3D structure data and target protein 3D data as training sets; S20, constructing a drug-target interaction prediction model; S30, using the training set as input of the drug-target interaction prediction model, establishing a loss function according to the error between the predicted value output by the drug-target interaction prediction model and the true value, and training the drug-target interaction prediction model using a gradient descent method until the model accuracy meets the requirements, thereby obtaining an optimal drug-target interaction prediction model; S40, predicting the interaction relationship between drugs and target proteins through the optimal drug-target interaction prediction model; Among them, S20 builds a drug-target interaction prediction model through the following methods: The framework of the drug-target interaction prediction model includes a drug graph convolutional network, a target graph convolutional network, a first fully connected layer, a feature fusion layer, and a second fully connected layer; S21, converting the drug three-dimensional structure data and the target protein three-dimensional data into a drug molecular structure three-dimensional matrix and a target protein molecular structure three-dimensional matrix respectively; S22, taking the three-dimensional matrix of drug molecular structure as the input of the drug graph convolutional network, and outputting the three-dimensional matrix of drug features; S23, taking the target protein molecular structure three-dimensional matrix as the input of the target graph convolutional network, and outputting the target protein feature three-dimensional matrix; S24, taking the drug feature three-dimensional matrix and the target protein feature three-dimensional matrix as the input of the first fully connected layer, and outputting the drug three-dimensional feature vector and the target protein three-dimensional feature vector S25, inputting the drug three-dimensional feature vector and the target protein three-dimensional feature vector into the feature fusion layer for splicing and fusion to obtain a fused feature vector; S26, introduces the fused feature vector into the second fully connected layer for dimensionality reduction, and outputs a one-dimensional correlation value, which represents the predicted value of the interaction strength between the drug and the target protein.

2. The method for predicting drug-target interaction relationships according to claim 1, wherein: The three-dimensional structure data of drugs and target proteins are obtained through the following methods: S11, obtain the matched and screened drug data and target protein data from the DrugBank database; S12, extracting drug three-dimensional structure data from the PubChem database based on drug data; S13, extracting target protein three-dimensional data from the Alphafold database based on the target protein data.

3. The method for predicting drug-target interaction relationships according to claim 1, wherein: Both the drug graph convolutional network and the target graph convolutional network include three convolutional layers, and the ReLU activation function and the maximum pooling layer are introduced after each convolutional layer.

4. The method for predicting drug-target interaction relationships according to claim 3, wherein: The feature extraction formulas of drug graph convolutional network and target graph convolutional network are both: Where H (l) is the node feature matrix of the lth layer, H (l+1) is the node feature matrix of the l+1th layer, is the normalized adjacency matrix, W (l) is the weight matrix of the lth layer, and σ is the activation function.

5. The method for predicting drug-target interaction relationships according to claim 1, wherein: The three-dimensional matrix of drug molecular structure is obtained by: Extract the coordinate values ​​and atom types of drug molecules from the three-dimensional structure data of drugs; Traversing the coordinate values ​​of the spatial position of each atom in the drug molecule to determine the minimum coordinate value; subtracting the minimum coordinate value from the coordinate value corresponding to the atom to obtain the normalized coordinate value of the drug molecule; According to different atom types, a unique value is assigned to each type, an all-zero matrix is ​​initialized, and the corresponding positions are filled with the values ​​of the corresponding atom types according to the normalized coordinate values ​​of the drug molecules to obtain the initial three-dimensional matrix of the drug molecule structure; The size range of the initial drug molecule structure three-dimensional matrix is ​​adjusted to obtain the final drug molecule three-dimensional matrix.

6. The method for predicting drug-target interaction relationships according to claim 1, wherein: The three-dimensional matrix of the target protein molecular structure is obtained by the following methods: Extract protein structural information from target protein three-dimensional data; Traverse each atom in the protein structure, obtain its coordinates and element type, and add them to the atomic coordinate list and atomic element type list respectively; Convert the atomic coordinate list into a NumPy array. After preprocessing the NumPy array, obtain the minimum value of the protein structure on the three coordinate axes. The three-axis coordinate data in the NumPy array are divided by the corresponding multiples of the minimum value of the protein structure on the three coordinate axes and rounded up to obtain the resized atomic coordinate matrix. Create an all-zero matrix that matches the adjusted atomic coordinate matrix and the matrix size to store the protein structure information. According to the adjusted atomic coordinate matrix and atomic element type list, add the corresponding atomic coordinate matrix and atomic element to the all-zero matrix to obtain the final three-dimensional matrix of the target protein molecular structure.

7. The method for predicting drug-target interaction relationships according to claim 1, wherein: The loss function is obtained by the following formula: Where, is the loss function, C is the number of categories, y i is the indicator function of the true label, It is expressed as the predicted probability of the drug-target interaction prediction model for class i, and log is the logarithmic function.

8. A drug-target interaction relationship prediction system, using the drug-target interaction relationship prediction method according to any one of claims 1 to 7, characterized in that: include: The data module is used to obtain the three-dimensional structure data of drugs and target proteins as training sets; Prediction model building module, used to build drug-target interaction prediction models; A training module is used to use the training set as the input of the drug-target interaction prediction model, establish a loss function based on the error between the predicted value output by the drug-target interaction prediction model and the true value, and train the drug-target interaction prediction model using the gradient descent method until the model accuracy meets the requirements, thereby obtaining the optimal drug-target interaction prediction model; A prediction module is used to predict the interaction relationship between drugs and target proteins through the optimal drug-target interaction prediction model; Among them, the drug-target interaction prediction model is constructed in the following ways: The framework of the drug-target interaction prediction model includes a drug graph convolutional network, a target graph convolutional network, a first fully connected layer, a feature fusion layer, and a second fully connected layer; S21, converting the drug three-dimensional structure data and the target protein three-dimensional data into a drug molecular structure three-dimensional matrix and a target protein molecular structure three-dimensional matrix respectively; S22, taking the three-dimensional matrix of drug molecular structure as the input of the drug graph convolutional network, and outputting the three-dimensional matrix of drug features; S23, taking the target protein molecular structure three-dimensional matrix as the input of the target graph convolutional network, and outputting the target protein feature three-dimensional matrix; S24, taking the drug feature three-dimensional matrix and the target protein feature three-dimensional matrix as the input of the first fully connected layer, and outputting the drug three-dimensional feature vector and the target protein three-dimensional feature vector S25, inputting the drug three-dimensional feature vector and the target protein three-dimensional feature vector into the feature fusion layer for splicing and fusion to obtain a fused feature vector; S26, introduces the fused feature vector into the second fully connected layer for dimensionality reduction, and outputs a one-dimensional correlation value, which represents the predicted value of the interaction strength between the drug and the target protein.

9. The drug-target interaction relationship prediction system according to claim 8, characterized in that: The feature extraction formulas of drug graph convolutional network and target graph convolutional network are both: Where H (l) is the node feature matrix of the lth layer, H (l+1) is the node feature matrix of the l+1th layer, is the normalized adjacency matrix, W (l) is the weight matrix of the lth layer, and σ is the activation function.

10. The drug-target interaction relationship prediction system according to claim 8, characterized in that: Both the drug graph convolutional network and the target graph convolutional network include three convolutional layers, and the ReLU activation function and the maximum pooling layer are introduced after each convolutional layer.

Images