Drug-target interaction prediction method and device based on deep neural network

Abstract

The application provides a drug-target interaction prediction method and device based on a deep neural network, and relates to the technical field of drug-target interaction prediction. The method comprises the following steps: calculating the similarity between different elements in a drug-related network and a protein-related network to obtain an initial drug similarity matrix and an initial protein similarity matrix; preprocessing the initial drug similarity matrix and the initial protein similarity matrix to obtain a target drug similarity matrix and a target protein similarity matrix; splicing the target drug similarity matrix and the target protein similarity matrix to obtain a DPP; determining a feature adjacency matrix of the DPP and a topological adjacency matrix of the DPP according to the DPP; obtaining a feature matrix according to the initial drug similarity matrix and the initial protein similarity matrix; and obtaining a prediction result according to the feature adjacency matrix, the topological adjacency matrix and the feature matrix. The application can improve the network learning effect and better predict drug-target interactions.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of drug-target interaction prediction, in particular to a drug-target interaction prediction method and device based on deep neural network. BACKGROUND

[0002] Determining the interaction between known drugs and targets is an important content in the medical field, which helps drug discovery and repositioning. However, identifying drug-target interactions (DTI) through in vitro and in vivo experiments is time-consuming and expensive. Therefore, more and more researchers have begun to explore computational methods for DTI prediction, which not only can reduce the loss in the drug discovery process, but also can provide guidance for the research of drug repositioning, drug resistance and side effect prediction.

[0003] In recent years, there have been good achievements in drug-target binding affinity prediction, such as DeepDTA (Ozturk et al. 2018), DeepAffinity (Karimi et al. 2019), GraphDTA (Nguyen et al. 2021), DeepDTAF (Wang et al. 2021) and so on. Among them, DeepDTA adopts two deep convolutional neural networks (CNN) to encode drug and protein sequences respectively, which achieved good results at that time; DeepAffinity uses CNN to encode protein sequences, uses an unsupervised sequence-to-sequence model (Sequence to Sequence, Seq2Seq; Sutskever et al. 2014) to encode drugs, and then trains a regression model to predict drug-target binding affinity; GraphDTA uses graph neural networks (GNN) to encode drugs, and uses CNN to encode proteins, and then connects the encoded features to predict drug-target binding affinity; DeepDTAF uses protein-binding pockets to directly process target binding, and then uses dilated convolution method to process multi-scale DTP features, and then connects different scale features to predict drug-target binding affinity.

[0004] Chinese patent CN202210418625.4 discloses a drug target-affinity prediction method based on representation learning and graph neural network. CNN and GNN are used to extract features of the two-dimensional structure of drugs / proteins, respectively. GNN is used to extract features of the three-dimensional structure of drugs / proteins. After fusing the extracted feature information, the drug target affinity value is predicted through the fully connected layer. Chinese patent CN202110028123.6 discloses a drug-target binding affinity prediction method based on graph convolutional neural network. First, the Bi-GRU network is used to extract the context correlation feature of the drug. Then, the graph convolutional neural network (GCN), graph attention neural network (GAT), graph isomorphism network (GIN), and graph convolutional attention joint network (GCN-GAT) are used to model the drug molecule. Then, the long short-term memory (LSTM) model is used to extract the feature system of the protein. Finally, the fully connected layer (Fully connected layer) is used to predict the drug-target binding affinity. Chinese patent CN116825183A discloses a drug-target binding affinity prediction method based on TransVAE. First, drug information, protein information, and drug-target binding affinity information are extracted from the database. Through data preprocessing, the drug feature matrix and the protein feature matrix are obtained after noise reduction. Then, a TransVAE-based drug-target binding affinity prediction model is used, which mainly includes a data encoding module, a feature fusion module, and a prediction module. The model contains a variational autoencoder (Variational Autoencoder), a transformer, and an attention pooling to obtain the fused features. Finally, the fused features are input into the prediction module to predict the drug-target binding affinity.

[0005] However, the aforementioned research methods may cause loss of graph structure information due to layer-by-layer stacking, have low computational efficiency, and have limited processing capacity for large-scale data, resulting in insufficient generalization ability of the model, which may greatly affect the final performance. SUMMARY

[0006] The present application aims to overcome the deficiencies in the prior art and provide a drug-target interaction prediction method and device based on deep neural networks to improve the learning effect of deep neural networks and better predict drug-target interactions.

[0007] To achieve the above object, the technical scheme adopted by the embodiments of the present application is as follows:

[0008] In a first aspect, the embodiments of the present application provide a drug-target interaction prediction method based on a deep neural network, comprising: calculating the similarity between different elements in a drug-related network and a protein-related network to obtain an initial drug similarity matrix and an initial protein similarity matrix; preprocessing the initial drug similarity matrix and the initial protein similarity matrix to obtain a target drug similarity matrix and a target protein similarity matrix; splicing the target drug similarity matrix and the target protein similarity matrix to obtain a DPP; performing feature extraction on the DPP using a KNN algorithm to obtain a feature adjacency matrix of the DPP; processing the DPP according to a preset rule to obtain a topological adjacency matrix of the DPP; performing feature extraction on the initial drug similarity matrix and the initial protein similarity matrix using a graph embedding module to obtain multi-view information; performing feature extraction on the multi-view information using a feature extraction module to obtain a feature matrix; obtaining a feature graph and a topological graph according to the feature adjacency matrix, the topological adjacency matrix, and the feature matrix; inputting the feature graph and the topological graph into a graph attention network for feature extraction to obtain a first target feature and a second target feature; inputting the feature graph and the topological graph into a graph convolution network for feature extraction to obtain a third target feature; inputting the first target feature, the second target feature, and the third target feature into an attention network for feature fusion to obtain a fusion feature; and inputting the fusion feature into a fully connected layer to obtain a prediction result.

[0009] In an embodiment, the preprocessing of the initial drug similarity matrix and the initial protein similarity matrix to obtain a target drug similarity matrix and a target protein similarity matrix comprises: processing the initial drug similarity matrix and the initial protein similarity matrix using an RWR algorithm to obtain an intermediate drug similarity matrix and an intermediate protein similarity matrix; and performing denoising and dimensionality reduction on the intermediate drug similarity matrix and the intermediate protein similarity matrix using a DAE algorithm to obtain the target drug similarity matrix and the target protein similarity matrix.

[0010] In an embodiment, the preset rule is that if two DPPs have the same drug or protein, it is considered that the two DPPs have the same topological edge.

[0011] In an implementation, the feature extraction of the initial drug similarity matrix and the initial protein similarity matrix by the graph embedding module to obtain multi-view information comprises: feature extraction of the initial drug similarity matrix and the initial protein similarity matrix by the graph embedding module to obtain an embedding matrix; the graph embedding module comprises a convolution network composed of a plurality of convolution layers; the embedding matrix is learned to obtain corresponding weighted parameters to obtain the multi-view information.

[0012] In an implementation, the feature extraction of the initial drug similarity matrix and the initial protein similarity matrix by the graph embedding module to obtain an embedding matrix comprises: initialization of the initial drug similarity matrix and the initial protein similarity matrix by random initialization of weights and random initialization of biases to obtain an initialized matrix; updating of the initialized matrix by a SUM function to obtain an updated matrix; normalization processing, activation processing, and regularization processing of the updated matrix in sequence to obtain the embedding matrix.

[0013] In an implementation, the feature graph and the topology graph are obtained according to the feature adjacency matrix, the topology adjacency matrix, and the feature matrix, which comprises: combination of the feature adjacency matrix and the feature matrix to obtain the feature graph; combination of the topology adjacency matrix and the feature matrix to obtain the topology graph.

[0014] In an implementation, the similarity between different elements in the drug-related network and the protein-related network is calculated to obtain an initial drug similarity matrix and an initial protein similarity matrix, which comprises: calculation of the Jaccard similarity coefficient between different elements in the drug-related network and the protein-related network to obtain the initial drug similarity matrix and the initial protein similarity matrix.

[0015] In an implementation, the drug-related network comprises a drug-drug relationship network, a drug-related disease network, a drug-chemical structure similarity network, a drug-side effect network, and a drug-drug relationship network; the drug-drug relationship network is used to represent the interaction between different drugs; the drug-related disease network is used to represent the relationship between drugs and diseases; the drug-side effect network is used to represent the relationship between drugs and side effects; the drug-chemical structure similarity network is used to represent the relationship between the chemical structures of different drugs; the protein-related network comprises a protein-related disease network, a protein-protein relationship network, and a protein sequence similarity network; the protein-related disease network is used to represent the relationship between proteins and diseases; the protein-protein relationship network is used to represent the interaction between different proteins; and the protein sequence similarity network is used to represent the relationship between different protein sequences.

[0016] In an embodiment, the feature extraction module comprises an ECR network; the ECR network comprises an ECR input layer, an ECR convolution layer, an ECR pooling layer, an ECR CECJC layer, and an ECR output layer connected in sequence.

[0017] In a second aspect, the embodiments of the present application also provide a drug-target interaction prediction device based on a deep neural network, comprising: a preprocessing module configured to calculate similarities between different elements in a drug-related network and a protein-related network to obtain an initial drug similarity matrix and an initial protein similarity matrix; preprocessing the initial drug similarity matrix and the initial protein similarity matrix to obtain a target drug similarity matrix and a target protein similarity matrix; splicing the target drug similarity matrix and the target protein similarity matrix to obtain a DPP; configured to perform feature extraction on the DPP using a KNN algorithm to obtain a feature adjacency matrix of the DPP; processing the DPP according to a preset rule to obtain a topological adjacency matrix of the DPP; a graph embedding module configured to perform feature extraction on the initial drug similarity matrix and the initial protein similarity matrix using a graph embedding module to obtain multi-view information; a feature extraction module configured to perform feature extraction on the multi-view information using a feature extraction module to obtain a feature matrix; a prediction module configured to obtain a feature graph and a topological graph according to the feature adjacency matrix, the topological adjacency matrix, and the feature matrix; inputting the feature graph and the topological graph into a graph attention network for feature extraction to obtain a first target feature and a second target feature; inputting the feature graph and the topological graph into a graph convolution network for feature extraction to obtain a third target feature; inputting the first target feature, the second target feature, and the third target feature into an attention network for feature fusion to obtain a fusion feature; inputting the fusion feature into a fully connected layer to obtain a prediction result.

[0018] In an embodiment, the preprocessing module is configured to: process the initial drug similarity matrix and the initial protein similarity matrix using an RWR algorithm to obtain an intermediate drug similarity matrix and an intermediate protein similarity matrix; and perform denoising and dimensionality reduction on the intermediate drug similarity matrix and the intermediate protein similarity matrix using a DAE algorithm to obtain the target drug similarity matrix and the target protein similarity matrix.

[0019] In an embodiment, the preset rule is that if two DPPs have the same drug or protein, it is considered that the two DPPs have the same topological edge.

[0020] In an implementation, the graph embedding module is configured to: perform feature extraction on the initial drug similarity matrix and the initial protein similarity matrix by using the graph embedding module to obtain an embedding matrix; the graph embedding module comprises a convolutional network composed of a plurality of convolutional layers; and learn corresponding weighted parameters from the embedding matrix to obtain the multi-view information.

[0021] In an implementation, the graph embedding module is configured to: initialize the initial drug similarity matrix and the initial protein similarity matrix by using randomly initialized weights and randomly initialized biases to obtain an initialized matrix; update the initialized matrix by using a SUM function to obtain an updated matrix; and sequentially perform normalization processing, activation processing, and regularization processing on the updated matrix to obtain the embedding matrix.

[0022] In an implementation, the prediction module is configured to obtain a feature graph and a topology graph from the feature adjacency matrix, the topology adjacency matrix, and the feature matrix, comprising: combining the feature adjacency matrix and the feature matrix to obtain the feature graph; and combining the topology adjacency matrix and the feature matrix to obtain the topology graph.

[0023] In an implementation, the preprocessing module is configured to calculate Jaccard similarity coefficients between different elements in the drug-related network and the protein-related network to obtain the initial drug similarity matrix and the initial protein similarity matrix.

[0024] In an implementation, the drug-related network comprises a drug-drug relationship network, a drug-related disease network, a drug-side effect network, and a drug-chemical structure similarity network; the drug-drug relationship network is used to represent interactions between different drugs; the drug-related disease network is used to represent relationships between drugs and diseases; the drug-side effect network is used to represent relationships between drugs and side effects; and the drug-chemical structure similarity network is used to represent relationships between chemical structures of different drugs; the protein-related network comprises a protein-related disease network, a protein-protein relationship network, and a protein sequence similarity network; the protein-related disease network is used to represent relationships between proteins and diseases; the protein-protein relationship network is used to represent interactions between different proteins; and the protein sequence similarity network is used to represent relationships between different protein sequences.

[0025] In an implementation, the feature extraction module comprises an ECR network; the ECR network comprises an ECR input layer, an ECR convolutional layer, an ECR pooling layer, a CECJC layer, and an ECR output layer connected in sequence.

[0026] In a third aspect, an embodiment of the present application provides a computer device, comprising: a processor, a storage medium, and a bus, the storage medium storing program instructions executable by the processor, when the computer device is running, the processor and the storage medium communicate through the bus, and the processor executes the program instructions to perform the steps of any of the above methods.

[0027] In a fourth aspect, an embodiment of the present application provides a non-volatile computer readable storage medium, the storage medium storing a computer program, when the computer program is run by a processor, the steps of any of the above methods are performed.

[0028] The beneficial effects of the present application are: based on the characteristics of drugs and proteins, drug-related networks (including drug-drug relationship network, drug-related disease network, drug and side effect network, drug-chemical structure similarity network) and protein-related networks (including protein-related disease network, protein-protein relationship network, protein sequence similarity network) are established, the feature vectors of the above seven networks are directly learned by using convolutional neural network, the original high-dimensional data is mapped to low-dimensional space to realize data dimension reduction, and the topology adjacency matrix and the feature adjacency matrix are obtained, which is of positive significance for finding the main dimension of data distribution; the multi-view inflation convolution, ResNet and K nearest neighbor algorithm (K Nearest Neighbors) in the graph embedding module are used to generate multi-view information, the complementary information and the relationship in the multi-view information can greatly improve the learning effect of the deep neural network, so that it is more convenient to obtain the key features affecting the drug-target interaction prediction; the ECR network in the feature extraction module is used to extract features from the multi-view information to obtain more comprehensive feature information (feature matrix); the graph convolution network and the graph attention network are designed to fuse the topology adjacency matrix, the feature adjacency matrix and the feature matrix, and then the prediction result of the drug-target interaction is obtained according to the fused features through the full connection layer, and the prediction effect of the drug-target interaction is improved. BRIEF DESCRIPTION OF DRAWINGS

[0029] In order to more clearly illustrate the technical solutions of the embodiments of the present application, the following will briefly introduce the drawings needed in the embodiments. It should be understood that the following drawings only show some embodiments of the present application, and therefore should not be regarded as a limitation on the scope, and for those skilled in the art, other related drawings can also be obtained without creative labor.

[0030] Figure 1 A flowchart of a drug-target interaction prediction method based on a deep neural network provided by an embodiment of the present application is shown in the figure.

[0031] Figure 2A flowchart illustrating a drug-target interaction prediction method based on deep neural networks, provided for embodiments of this application;

[0032] Figure 3 A flowchart illustrating a drug-target interaction prediction method based on deep neural networks, provided for embodiments of this application;

[0033] Figure 4 A flowchart illustrating a drug-target interaction prediction method based on deep neural networks, provided for embodiments of this application;

[0034] Figure 5 A schematic diagram illustrating a drug-target interaction prediction method based on a deep neural network, provided for an embodiment of this application;

[0035] Figure 6 A schematic diagram of a drug-target interaction prediction device based on a deep neural network, provided for an embodiment of this application;

[0036] Figure 7 This is a schematic diagram of the structure of a computer device provided in an embodiment of this application. Detailed Implementation

[0037] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are some embodiments of the present invention, but not all embodiments.

[0038] Therefore, the following detailed description of the embodiments of this application provided in the accompanying drawings is not intended to limit the scope of the claimed application, but merely to illustrate selected embodiments of the application. All other embodiments obtained by those skilled in the art based on the embodiments of this application without inventive effort are within the scope of protection of this application.

[0039] In the description of this application, it should be noted that if the terms "upper", "lower", etc. appear to indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings, or the orientation or positional relationship that the product of this application is usually placed in, it is only for the convenience of describing this application and simplifying the description, and does not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation, and therefore should not be construed as a limitation of this application.

[0040] Furthermore, the terms "first", "second", and the like, in the description and in the claims of the present application, as well as above-mentioned drawings, are used for distinguishing between similar objects, not necessarily described with a specific sequential or chronological order. It is to be understood that the data so described can be interchanged, where appropriate, so that one embodiment of the present application described herein could be implemented differently with respect to the order of steps described herein. Furthermore, the terms "comprising", "including", "containing", and any variations thereof, are intended to cover a non-exclusive inclusion, such that a process, method, system, product, or apparatus that comprises, includes, or contains an item or list of items that does not include all of the recited items is still considered to be within the scope of the present application.

[0041] It should be noted that the features of the embodiments of the present application can be combined with each other in the case of no conflict.

[0042] Figure 1 A flowchart of a drug-target interaction prediction method based on a deep neural network provided by an embodiment of the present application is shown in FIG. 1. As shown in FIG. 1, the method comprises the following steps. Figure 1

[0043] In step 110, similarities between different elements in a drug-related network and a protein-related network are calculated to obtain an initial drug similarity matrix and an initial protein similarity matrix. The initial drug similarity matrix and the initial protein similarity matrix are preprocessed to obtain a target drug similarity matrix and a target protein similarity matrix. The target drug similarity matrix and the target protein similarity matrix are spliced to obtain a DPP. A KNN algorithm is used to extract features of the DPP to obtain a feature adjacency matrix of the DPP. The DPP is processed according to a preset rule to obtain a topological adjacency matrix of the DPP.

[0044] The drug-related network comprises a drug-drug relationship network, a drug-related disease network, a drug-side effect network, and a drug-chemical structure similarity network. The drug-drug relationship network is used to represent interactions between different drugs. The drug-related disease network is used to represent relationships between drugs and diseases. The drug-side effect network is used to represent relationships between drugs and side effects. The drug-chemical structure similarity network is used to represent relationships between chemical structures of different drugs.

[0045] The protein-related network comprises a protein-related disease network, a protein-protein relationship network, and a protein sequence similarity network. The protein-related disease network is used to represent relationships between proteins and diseases. The protein-protein relationship network is used to represent interactions between different proteins. The protein sequence similarity network is used to represent relationships between different protein sequences.

[0046] ​The initial drug similarity matrix and the initial protein similarity matrix can be obtained by calculating the Jaccard similarity coefficient between different elements in the drug-related network and the protein-related network.

[0047] For example, the Jaccard similarity coefficient calculation formula is shown in the following formula (1):

[0048]

[0049] For the drug-drug relationship network, M i and M j represent drug i and drug j, respectively; J(M i , M j ) represents the interaction relationship between drug i and drug j, which includes the enhancement, weakening or adverse reaction of drug effect; |M i ∩M j | represents the size of the intersection of the two sets; and |M i ∪M j | represents the size of the union of the two sets.

[0050] For the drug-disease network, M i and M j represent drug i and disease j, respectively; J(M i , M j ) represents the relationship between drug i and disease j; |M i ∩M j | represents the size of the intersection of the two sets; and |M i ∪M j | represents the size of the union of the two sets.

[0051] For the drug-side effect network, M i and M j represent drug i and side effect j, respectively; J(M i , M j ) represents the relationship between drug i and side effect j; |M i ∩M j | represents the size of the intersection of the two sets; and |M i ∪M j | represents the size of the union of the two sets.

[0052] For the drug-chemical structure similarity network, M i and M j represent the chemical structure of drug i and the chemical structure of drug j, respectively; J(M i , M j ) represents the relationship between the chemical structure of drug i and the chemical structure of drug j; |M i ∩M j| represents the size of the intersection of two sets; | M i ∩M j | represents the size of the union of two sets.

[0053] For the protein-disease association network, M i and M j represent protein i and disease j, respectively; J(M i , M j ) represents the relationship between protein i and disease j; | M i ∩M j | represents the size of the intersection of two sets; | M i ∪M j | represents the size of the union of two sets.

[0054] For the protein-protein relationship network, M i and M j represent protein i and protein j, respectively; J(M i , M j ) represents the interaction relationship between protein i and protein j, which includes direct physical contact, functional association, and interaction revealed by genetic evidence; | M i ∩M j | represents the size of the intersection of two sets; | M i ∪M j | represents the size of the union of two sets.

[0055] For the protein sequence similarity network, M i and M j represent the sequence of protein i and the sequence of protein j, respectively; J(M i , M j ) represents the relationship between the sequence of protein i and the sequence of protein j; | M i ∩M j | represents the size of the intersection of two sets; | M i ∪M j | represents the size of the union of two sets.

[0056] The initial drug similarity matrix obtained in step 110 includes the initial similarity matrices corresponding to the four networks of drug-drug relationship network, drug-related disease network, drug-side effect network, and drug-chemical structure similarity network; the initial protein similarity matrix includes the initial similarity matrices corresponding to the three networks of protein-related disease network, protein-protein relationship network, and protein sequence similarity network; therefore, step 110 actually obtains seven initial similarity matrices.

[0057] The preprocessing includes integration, denoising, dimensionality reduction, etc. of the above seven similarity matrices.

[0058] The target drug similarity matrix is a preprocessed similarity matrix for drugs; the target protein similarity matrix is a preprocessed similarity matrix for proteins. After preprocessing, the target drug similarity matrix and the target protein similarity matrix are both single matrices.

[0059] A DPP is a drug-protein pair formed by a combination of a drug and a protein. The connection of a drug p and a protein q in a DPP node i is shown in the following formula (2):

[0060]

[0061] The KNN algorithm is shown in the following formula (3):

[0062]

[0063] Wherein, A and B are a certain row in the target drug similarity matrix and the target protein similarity matrix in the DPP, respectively, both of which are vectors; A i and B i are the ith elements of the vector A and the vector B; n is the number of elements in the vector A and the vector B.

[0064] The preset rule is that if two DPPs have the same drug or protein, it is considered that the two DPPs have the same topological edge.

[0065] Step 120, using a graph embedding module to extract features from the initial drug similarity matrix and the initial protein similarity matrix to obtain multi-view information.

[0066] In actual operation, this step can be realized through the following process: first, using a convolutional network in the graph embedding module to extract features from the initial drug similarity matrix and the initial protein similarity matrix to obtain a plurality of embedding matrices; second, learning weighting parameters for the plurality of embedding matrices to obtain a plurality of weighted embedding matrices (i.e., the above multi-view information). Here, the convolutional network can be single-layer or multi-layer, which is not limited here.

[0067] Step 130, using a feature extraction module to extract features from the multi-view information to obtain a feature matrix.

[0068] The feature extraction module comprises an ECR network; the ECR network comprises an ECR input layer, an ECR convolution layer, an ECR pooling layer, a CECJC layer and an ECR output layer connected in sequence. The ECR network combines the advantages of spatial dilated convolution and skip connection, captures local, medium and long distance neighborhood information through three adaptive circle layer dilated convolution kernels, and realizes comprehensive coverage of multi-scale information. This design not only avoids the problem of computational complexity caused by excessive parameters, but also enhances the model's ability to capture key detail information through specially designed skip connection layers. The ECR network is specifically shown in the following formula (4), formula (5) and formula (6):

[0069] x conv = ReLU (Conv2d (x, kernelsize, padding, dilation)) (4)

[0070] x pooling = Interpolate (ReLU (Conv2d (AdaptiveAvPool2d (x, 1)))) (5)

[0071] x CECJC = Dropout (ReLU (Conv2d (Concat (x conv ,x pooling ,x), 1))) (6)

[0072] In formula (4), x conv is a convolution layer formula, x is the original input, kernelsize is the convolution kernel size, padding is padding, dilation is the dilation rate, Conv2d represents ordinary two-dimensional convolution, and ReLU represents the activation function; in formula (5), x pooling is a pooling layer formula, AdaptiveAvPool2d represents adaptive average pooling, and Interpolate represents sampling using interpolation method; in formula (6), x CECJC is a formula of the CECJC layer, Concat is the merging of channels, and Dropout represents using Dropout to prevent overfitting.

[0073] Through the ECR network, we learn the context representation and semantic representation of DPP from the drug feature information and the protein feature information. Specifically,

[0074] Step 140, according to the feature adjacency matrix, the topology adjacency matrix, the feature matrix, the feature graph and the topology graph are obtained; the feature graph and the topology graph are input into the graph attention network for feature extraction to obtain the first target feature, the second target feature; the feature graph and the topology graph are input into the graph convolution network for feature extraction to obtain the third target feature; the first target feature, the second target feature and the third target feature are input into the attention network for feature fusion to obtain the fusion feature; the fusion feature is input into the full connection layer to obtain the prediction result.

[0075] Wherein, the feature graph and the topology graph can be different matrix combinations, specifically, the feature adjacency matrix and the feature matrix can be combined to obtain the feature graph; the topology adjacency matrix and the feature matrix can be combined to obtain the topology graph.

[0076] Graph Attention Networks (GAT) is a variant of Graph Neural Networks (GNN), which introduces an attention mechanism to weight the aggregation of neighboring nodes, thereby achieving adaptive allocation of different neighbor weights. This mechanism enables GAT to better capture the relationships between nodes and improve the model's expressive power. The core idea of GAT is that the hidden state of each node can be calculated by focusing on its neighbor nodes. This self-attention mechanism allows the model to assign different weights to each neighbor node, rather than treating all neighbors equally as in traditional Graph Convolutional Networks (GCN).

[0077] The related calculations of the graph attention network are as follows:

[0078] First, given N node features, the set of all node features is F is the number of feature dimensions for each node; h i is the feature of node i; then the attention coefficient e ij of node i relative to node j is as shown in the following formula (7):

[0079]

[0080] Where, W are shared learnable parameters.

[0081] Second, the attention coefficient is normalized as shown in the following formula (8):

[0082]

[0083] LeakyReLU (Leaky Rectified Linear Unit) is an activation function commonly used in neural networks in deep learning. It is a variant of the ReLU (Rectified Linear Unit) activation function designed to address the "dead neurons" problem that may occur during training.

[0084] Again, LeakyReLU (i.e., f(x)) is shown in the following equation (9):

[0085]

[0086] where a is a constant, usually set to 0.01, to control the slope of negative inputs.

[0087] Next, the node is updated, as shown in the following equation (10):

[0088]

[0089] where, is the new feature of node i after fusing neighborhood information; σ is the activation function; W is the shared learnable parameter.

[0090] Finally, use the multi-head attention mechanism to learn the different meanings of DPP nodes from different representation subspaces, and get the K-head set of node i as shown in the following equation (11):

[0091]

[0092] From the feature map G f = (A f ,X DPP ) of DPP, get the K-head set of node i as shown in the following equation (12):

[0093]

[0094] Graph Convolutional Networks (GCN) is a neural network structure for processing graph data. By performing convolution operations on graphs, it can effectively capture the complex relationships between nodes and the structural characteristics of graphs. The core idea of GCN is to update the representation of each node by aggregating the feature information of the node and its neighbor nodes. This approach mimics the process of extracting local features through filters in traditional Convolutional Neural Networks (CNN), but is suitable for graph data structures that are irregular.

[0095] The relevant calculations for graph convolutional networks are as follows:

[0096] First, let Z ct and Z cf Let represent the common representations learned from the topology graph and feature graph, respectively; the topology space and feature space output by the I layer of the GCN model are shown in Equation (13) and Equation (14), respectively:

[0097]

[0098] Where A is the adjacency matrix of the input graph; D is the degree matrix of the input graph; It is the weight matrix of the I-th layer of the input graph.

[0099] Secondly, the input of the first layer is defined as (·) indicates a replaceable representation, which can be one of the above. or The output of the GCN model learned in the topological and feature spaces is the Z-intercalation of the common embedding. ct and Z cf ;

[0100] Finally, the average of the two representations is taken to obtain the common representation Z of the DPP nodes. C As shown in the following formula (15):

[0101]

[0102] For a DPP node i, the three representations of node i can be used as follows: and Z iC Let vector a i Their attention vectors can be represented by the attention function Att(Zt, Zf, Zc) as shown in the following formula (16):

[0103]

[0104] in, They are respectively The corresponding attention score.

[0105] Attention networks are a type of deep learning model whose core idea is to mimic the human attention mechanism, which focuses on the most relevant parts of the current task when processing information. This mechanism allows the model to dynamically assign different processing weights to different parts of the input data, thereby improving the model's ability to process complex data.

[0106] The fully connected layers are the last few layers (usually one or a few) in a neural network that convert the feature maps from the previous layer into a fixed-length feature vector through matrix multiplication and classification or regression through an activation function. In classification tasks, the activation function converts the feature vector into a probability distribution representing the predicted probability of each class.

[0107] The drug-target interaction prediction method based on the deep neural network provided in the embodiments of the present application firstly calculates the similarity between different elements in the drug-related network and the protein-related network to obtain an initial drug similarity matrix and an initial protein similarity matrix; secondly, the initial drug similarity matrix and the initial protein similarity matrix are preprocessed to obtain a target drug similarity matrix and a target protein similarity matrix; the target drug similarity matrix and the target protein similarity matrix are spliced to obtain a DPP; thirdly, the KNN algorithm is used to extract features of the DPP to obtain a feature adjacency matrix of the DPP; the DPP is processed according to a preset rule to obtain a topological adjacency matrix of the DPP; then, the initial drug similarity matrix and the initial protein similarity matrix are extracted by a graph embedding module to obtain multi-view information; then, the multi-view information is extracted by a feature extraction module to obtain a feature matrix; then, a feature graph and a topological graph are obtained according to the feature adjacency matrix, the topological adjacency matrix and the feature matrix; the feature graph and the topological graph are respectively input into a graph attention network for feature extraction to obtain a first target feature and a second target feature; the feature graph and the topological graph are input into a graph convolution network for feature extraction to obtain a third target feature; finally, the first target feature, the second target feature and the third target feature are input into an attention network for feature fusion to obtain a fusion feature; and the fusion feature is input into a fully connected layer to obtain a prediction result. In this way, based on the characteristics of drugs and proteins, a drug-related network (including a drug-drug relationship network, a drug-related disease network, a drug-chemical structure similarity network, a drug-side effect network) and a protein-related network (including a protein-related disease network, a protein-protein relationship network, a protein sequence similarity network) are established, the convolutional neural network is used to directly learn the feature vectors of the above seven networks, the original high-dimensional data is mapped to a low-dimensional space to realize data dimension reduction, and the topological adjacency matrix and the feature adjacency matrix are obtained, which is of positive significance for finding the main dimension of data distribution; the multi-view inflation convolution, the ResNet and the K Nearest Neighbors in the graph embedding module are used to generate multi-view information, and the complementary information and the relationship in the multi-view information can greatly improve the learning effect of the deep neural network, so that it is more convenient to obtain the key features affecting the drug-target interaction prediction; the ECR network in the feature extraction module is used to extract features from the multi-view information to obtain more comprehensive feature information (a feature matrix); the graph convolution network and the graph attention network are designed to fuse the topological adjacency matrix, the feature adjacency matrix and the feature matrix, and then the fusion feature is input into the fully connected layer to obtain the prediction result of the drug-target interaction, thereby improving the prediction effect of the drug-target interaction.

[0108] The process of preprocessing the initial drug similarity matrix and the initial protein similarity matrix in the above embodiments to obtain the target drug similarity matrix and the target protein similarity matrix will be explained by examples as follows. Figure 2 The flowchart of the method for predicting drug-target interaction based on deep neural network provided by the embodiments of the present application is shown in Figure 1. Figure 2 As shown in Figure 2, the step 120 can include the following steps 210 and 220:

[0109] In step 210, the RWR algorithm is used to process the initial drug similarity matrix and the initial protein similarity matrix to obtain an intermediate drug similarity matrix and an intermediate protein similarity matrix.

[0110] The RWR algorithm is shown in the following formula (17):

[0111] r = cW + (1-c)e (17) ij r = cW + (1-c)e (17)

[0112] Wherein, the size of c is between [0, 1]; W ij is the transition probability from node i to node j; e is the starting vector; r is the ending vector.

[0113] Since the initial drug similarity matrix includes four initial similarity matrices related to drugs, and the initial protein similarity matrix includes three initial similarity matrices related to proteins; therefore, the initial similarity matrices processed by the RWR algorithm in this step are also seven, including four initial similarity matrices related to drugs and three initial similarity matrices related to proteins. Further, in actual operation, the initial similarity matrices processed by the RWR algorithm can be integrated, that is, the four initial similarity matrices related to drugs processed by the RWR algorithm are spliced to obtain the above-mentioned intermediate drug similarity matrix, and the three initial similarity matrices related to proteins processed by the RWR algorithm are spliced to obtain the above-mentioned intermediate protein similarity matrix. In this way, the obtained intermediate drug similarity matrix and the intermediate protein similarity matrix are both single matrices.

[0114] In step 220, the DAE algorithm is used to denoise and reduce the dimension of the intermediate drug similarity matrix and the intermediate protein similarity matrix to obtain a target drug similarity matrix and a target protein similarity matrix.

[0115] Wherein, the denoising autoencoder (denoising autoencoder, DAE) is a deep learning model, which adds noise to the input data, and then trains the model to recover the original data from the noise data, thereby realizing denoising. At the same time, the DAE can also realize the dimension reduction of data through its encoding and decoding process.

[0116] The DAE algorithm is shown in the following formula (18):

[0117]

[0118] wherein, is a noise distribution; x is an original input; is an input after adding noise; W1 and b1 are respectively a weight and a bias of encoding; W2 and b2 are respectively a weight and a bias of decoding; h1 represents a hidden layer after encoding; y represents a reconstruction output; σ e is an activation function used in an encoding process; σ d is an activation function used in a decoding process.

[0119] The method for predicting drug-target interaction based on the deep neural network provided in the embodiments of the present application firstly processes the initial drug similarity matrix and the initial protein similarity matrix by using the RWR algorithm to obtain an intermediate drug similarity matrix and an intermediate protein similarity matrix; secondly, the DAE algorithm is used to denoise and reduce dimensions of the intermediate drug similarity matrix and the intermediate protein similarity matrix to obtain a target drug similarity matrix and a target protein similarity matrix. In this way, based on the characteristics of drugs and proteins, a drug-related network (including a drug-drug relationship network, a drug-related disease network, a drug and side effect network, and a drug-chemical structure similarity network) and a protein-related network (including a protein-related disease network, a protein-protein relationship network, and a protein sequence similarity network) are established, the convolutional neural network is used to directly learn the feature vectors of the above seven networks, the original high-dimensional data is mapped to a low-dimensional space to realize data dimension reduction, and a topology adjacency matrix and a feature adjacency matrix are obtained, which is of positive significance for finding the main dimension of data distribution.

[0120] The process of extracting features from the initial drug similarity matrix and the initial protein similarity matrix by using the graph embedding module to obtain multi-view information in the above embodiments will be explained by examples as follows. Figure 3 The flowchart of the method for predicting drug-target interaction based on the deep neural network provided in the embodiments of the present application is shown in FIG. 1. Figure 3 As shown in FIG. 1, the step 140 can include the following step 310 and step 320:

[0121] The step 310 extracts features from the initial drug similarity matrix and the initial protein similarity matrix by using the graph embedding module to obtain an embedding matrix.

[0122] The graph embedding module includes a convolutional network, and the convolutional network is composed of multiple convolutional layers.

[0123] The convolutional network is preferably composed of multiple convolutional layers, but can also be composed of one convolutional layer, which is not limited here.

[0124] Specifically, as shown in Figure 4 The step 310 can include the following steps 410 to 430.

[0125] Step 410, the initial drug similarity matrix and the initial protein similarity matrix are initialized by using random initialization weight and random initialization bias to obtain an initialization matrix.

[0126] The initialization process is shown in the following formula (19):

[0127] r = W0x + b0 (19)

[0128] Wherein, W0 is the random initialization weight; b0 is the random initialization bias; x is the input; r is the initialization matrix.

[0129] Step 420, the initialization matrix is updated by using the SUM function to obtain an updated matrix.

[0130] The SUM function is shown in the following formula (20):

[0131] y i = W i x + b i (20)

[0132] Wherein, W i is the weight; b i is the bias.

[0133] The updated matrix r' is shown in the following formula (21):

[0134] r' = r + y i (21)

[0135] Step 430, the updated matrix is sequentially normalized, activated and regularized to obtain an embedding matrix.

[0136] The normalization, activation and regularization process is shown in the following formula (22):

[0137] r = Dropout(ReLU(BathNorm(r))) (22)

[0138] Wherein, Dropout represents the regularization process; ReLU represents the activation process; BathNorm represents the normalization process.

[0139] Taking the lth layer as an example, the forward propagation of the graph embedding module is shown in the following formula (23):

[0140] x l =ConvLayer l (x l-1 )(twenty three)

[0141] Among them, ConvLayer l This indicates the l-th convolutional layer.

[0142] Update input x = x l .

[0143] Step 320: Learn the corresponding weighting parameters for the embedding matrix to obtain multi-view information.

[0144] Since the initial drug similarity matrix includes four drug-related initial similarity matrices and the initial protein similarity matrix includes three protein-related initial similarity matrices, the embedding matrix also includes four drug-related embedding matrices and three protein-related embedding matrices.

[0145] In practice, each embedding matrix in this step corresponds to a weighting parameter. That is, each embedding matrix learns its corresponding weighting parameter to obtain its corresponding weighted embedding matrix. Finally, these seven weighted embedding matrices form the aforementioned multi-view information.

[0146] The learning process for this step is shown in the following formula (24):

[0147]

[0148] Where x1, x2, x3, x4, x5, x6, and x7 are four drug-related embedding matrices and three protein-related embedding matrices, respectively; beta1, beta2, beta3, beta4, beta5, beta6, and beta7 are the weighting parameters of the four drug-related embedding matrices and the three protein-related embedding matrices, respectively, which are automatically learned by the network based on the loss function; x1′, x2′, x3′, x4′, x5′, x6′, and x7′ are the four drug-related weighted embedding matrices and the three protein-related weighted embedding matrices, respectively.

[0149] like Figure 5 The diagram shown is a schematic representation of a drug-target interaction prediction method based on a deep neural network, provided in an embodiment of this application. In the diagram, the spatial information is the aforementioned topological adjacency matrix; the spectral information is the aforementioned feature adjacency matrix.

[0150] The method for predicting drug-target interaction based on a deep neural network provided by the embodiments of the present application first extracts features of an initial drug similarity matrix and an initial protein similarity matrix by using a graph embedding module to obtain an embedding matrix; second, learns corresponding weighted parameters from the embedding matrix to obtain multi-view information. In this way, the multi-view information is generated by using multi-view dilated convolution, ResNet and K Nearest Neighbors in the graph embedding module, and the complementary information and relationship in the multi-view information can greatly improve the learning effect of the deep neural network, making it more convenient to obtain key features affecting the prediction of drug-target interaction.

[0151] After introducing the method for predicting drug-target interaction based on a deep neural network of the exemplary embodiments of the present application, next, with reference to Figure 6 The device 600 for predicting drug-target interaction based on a deep neural network of the exemplary embodiments of the present application is described.

[0152] With reference to Figure 6 The device 600 for predicting drug-target interaction based on a deep neural network includes: a preprocessing module 610 configured to calculate the similarity between different elements in a drug-related network and a protein-related network to obtain an initial drug similarity matrix and an initial protein similarity matrix; to preprocess the initial drug similarity matrix and the initial protein similarity matrix to obtain a target drug similarity matrix and a target protein similarity matrix; to splice the target drug similarity matrix and the target protein similarity matrix to obtain a DPP; configured to extract features of the DPP by using a KNN algorithm to obtain a feature adjacency matrix of the DPP; to process the DPP according to a preset rule to obtain a topological adjacency matrix of the DPP; a graph embedding module 620 configured to extract features of the initial drug similarity matrix and the initial protein similarity matrix by using the graph embedding module 620 to obtain multi-view information; a feature extraction module 630 configured to extract features of the multi-view information by using the feature extraction module 630 to obtain a feature matrix; a prediction module 640 configured to obtain a feature graph and a topological graph according to the feature adjacency matrix, the topological adjacency matrix and the feature matrix; to input the feature graph and the topological graph into a graph attention network for feature extraction to obtain a first target feature and a second target feature; to input the feature graph and the topological graph into a graph convolution network for feature extraction to obtain a third target feature; to input the first target feature, the second target feature and the third target feature into an attention network for feature fusion to obtain a fusion feature; and to input the fusion feature into a fully connected layer to obtain a prediction result.

[0153] In an implementation, the preprocessing module 610 is configured to: adopt the RWR algorithm to process the initial drug similarity matrix and the initial protein similarity matrix to obtain an intermediate drug similarity matrix and an intermediate protein similarity matrix; and adopt the DAE algorithm to denoise and reduce dimensions of the intermediate drug similarity matrix and the intermediate protein similarity matrix to obtain a target drug similarity matrix and a target protein similarity matrix.

[0154] In an implementation, the preset rule is that if two DPPs have the same drug or protein, it is considered that the two DPPs have the same topological edge.

[0155] In an implementation, the graph embedding module 620 is configured to: adopt the graph embedding module 620 to extract features from the initial drug similarity matrix and the initial protein similarity matrix to obtain an embedding matrix; the graph embedding module 620 comprises a convolution network composed of multiple convolution layers; and the embedding matrix is learned to obtain corresponding weighted parameters to obtain multi-view information.

[0156] In an implementation, the graph embedding module 620 is configured to: adopt random initialization weights and random initialization biases to initialize the initial drug similarity matrix and the initial protein similarity matrix to obtain an initialization matrix; adopt a SUM function to update the initialization matrix to obtain an updated matrix; and sequentially perform normalization processing, activation processing, and regularization processing on the updated matrix to obtain an embedding matrix.

[0157] In an implementation, the prediction module 640 is configured to: obtain a feature graph and a topology graph according to the feature adjacency matrix, the topology adjacency matrix, and the feature matrix, comprising: combining the feature adjacency matrix and the feature matrix to obtain the feature graph; and combining the topology adjacency matrix and the feature matrix to obtain the topology graph.

[0158] In an implementation, the preprocessing module 610 is configured to: calculate the Jaccard similarity coefficient between different elements in the drug-related network and the protein-related network to obtain the initial drug similarity matrix and the initial protein similarity matrix.

[0159] In an implementation, the drug-related network includes a drug-drug relationship network, a drug-related disease network, a drug-side effect network, and a drug-chemical structure similarity network; the drug-drug relationship network is used to represent the interaction between different drugs; the drug-related disease network is used to represent the relationship between drugs and diseases; the drug-side effect network is used to represent the relationship between drugs and side effects; and the drug-chemical structure similarity network is used to represent the relationship between the chemical structures of different drugs.

[0160] In an implementation, the feature extraction module 630 includes an ECR network; the ECR network includes an ECR input layer, an ECR convolution layer, an ECR pooling layer, a CECJC layer, and an ECR output layer connected in sequence.

[0161] The apparatus described above is configured to execute the method provided by the foregoing embodiments, and has similar implementation principles and technical effects, which will not be described here.

[0162] The above modules can be one or more integrated circuits configured to implement the above method, for example, one or more application specific integrated circuits (ASICs), or one or more microprocessors, or one or more field programmable gate arrays (FPGAs), etc. For another example, when a certain module above is implemented in the form of a processing element scheduling code, the processing element can be a general-purpose processor, such as a central processing unit (CPU) or other processor that can invoke program code. For another example, the modules can be integrated together to implement in the form of a system on a chip (SOC).

[0163] Figure 7 A schematic diagram of a computer device provided by the embodiments of the present application is shown in FIG. 7. The device can be integrated in a terminal device or a chip of the terminal device, and the terminal device can be a computing device with a data processing function.

[0164] The device includes a processor 701, a storage medium 702, and a bus 703.

[0165] The storage medium 702 stores program instructions executable by the processor 701. When the computer device 700 is running, the processor 701 communicates with the storage medium 702 through the bus 703. The processor 701 executes the program instructions to perform the above method embodiments. The specific implementation manners and technical effects are similar, and will not be described here.

[0166] Optionally, the present application also provides a program product, for example, a computer readable storage medium, comprising a program which, when executed by a processor, is used to perform the above method embodiments.

[0167] In several embodiments provided by the present application, it should be understood that the disclosed apparatus and method can be implemented in other manners. For example, the described embodiments of the apparatus are merely schematic, and the division of the units is merely logical function division, and there can be another division manner in actual implementation. For example, a plurality of units or components can be combined or integrated into another system, or some features can be ignored or not executed. In addition, the displayed or discussed mutual couplings or direct couplings or communication connections between the units can be indirect couplings or communication connections through some interfaces, devices or units, and can be electrical, mechanical or in other forms.

[0168] The units described as separate components can or can not be physically separate, and the components shown as units can or can not be physical units, that is, can be located in one place, or can be distributed on a plurality of network units. Some or all of the units can be selected according to actual needs to achieve the purpose of the embodiments.

[0169] In addition, each functional unit in the various embodiments of the present application can be integrated into one processing unit, or each unit can exist physically, or two or more units can be integrated into one unit. The integrated unit can be implemented in the form of hardware, or in the form of hardware plus software function units.

[0170] The integrated unit implemented in the form of software function units can be stored in a computer readable storage medium. The software function unit stored in the storage medium includes a plurality of instructions for causing a computer device (which can be a personal computer, a server, or a network device, etc.) or a processor (English: processor) to execute part of the steps of the method described in the various embodiments of the present application. The foregoing storage medium includes: a U disk, a mobile hard disk, a read-only memory (English: Read-Only Memory, abbreviated as: ROM), a random access memory (English: Random Access Memory, abbreviated as: RAM), a magnetic disk or an optical disk, and various media that can store program codes.

[0171] The above merely provides the specific implementation of the present application, but the protection scope of the present application is not limited thereto, any person skilled in the art can easily think of the changes or replacements within the technical range disclosed by the present application, which should be covered within the protection scope of the present application. Therefore, the protection scope of the present application should be subject to the protection scope of the claims.

Claims

1. A method for predicting drug-target interactions based on deep neural networks, characterized in that, include: Calculate the similarity between different elements in the drug-related network and the protein-related network to obtain the initial drug similarity matrix and the initial protein similarity matrix; The initial drug similarity matrix and the initial protein similarity matrix are preprocessed to obtain the target drug similarity matrix and the target protein similarity matrix; the target drug similarity matrix and the target protein similarity matrix are then concatenated to obtain the DPP. The KNN algorithm is used to extract features from the DPP to obtain the feature adjacency matrix of the DPP. The DPP is processed according to preset rules to obtain the topological adjacency matrix of the DPP; The graph embedding module is used to extract features from the initial drug similarity matrix and the initial protein similarity matrix to obtain multi-view information; The feature extraction module is used to extract features from the multi-view information to obtain a feature matrix; Based on the feature adjacency matrix, the topological adjacency matrix, and the feature matrix, a feature map and a topological map are obtained. The feature map and the topological map are then input into a graph attention network for feature extraction to obtain a first target feature and a second target feature. The feature map and the topological map are then input into a graph convolutional network for feature extraction to obtain a third target feature. The first target feature, the second target feature, and the third target feature are then input into an attention network for feature fusion to obtain a fused feature. Finally, the fused feature is input into a fully connected layer to obtain a prediction result. The feature extraction module includes an ECR network; the ECR network includes an ECR input layer, an ECR convolutional layer, an ECR pooling layer, a CECJC layer, and an ECR output layer connected in sequence; the ECR network combines the advantages of spatial dilated convolution and skip connections, and captures neighborhood information at three different scales—local, intermediate, and long distance—through three adaptive convolutional kernels. The ECR network is shown in the following formulas (4), (5), and (6): （4） （5） （6） In formula (4) This is the formula for a convolutional layer, where x is the original input. It is the kernel size. For filling, It is the expansion rate. This represents a regular two-dimensional convolution. In formula (5), the activation function is represented. It is the pooling layer formula. Indicates adaptive average pooling. This indicates that interpolation is used for sampling; in formula (6) It is the formula for the CECJC layer. It is a merging of channels. Indicates use Prevent overfitting.

2. The method according to claim 1, characterized in that, The preprocessing of the initial drug similarity matrix and the initial protein similarity matrix to obtain the target drug similarity matrix and the target protein similarity matrix includes: The RWR algorithm is used to process the initial drug similarity matrix and the initial protein similarity matrix to obtain the intermediate drug similarity matrix and the intermediate protein similarity matrix; The DAE algorithm is used to denoise and reduce the dimensionality of the intermediate drug similarity matrix and the intermediate protein similarity matrix to obtain the target drug similarity matrix and the target protein similarity matrix.

3. The method according to claim 1, characterized in that, The preset rule is: if two DPPs have the same drug or protein, then the two DPPs are considered to have the same topological edge.

4. The method according to claim 1, characterized in that, The graph embedding module is used to extract features from the initial drug similarity matrix and the initial protein similarity matrix to obtain multi-view information, including: The graph embedding module is used to extract features from the initial drug similarity matrix and the initial protein similarity matrix to obtain an embedding matrix; the graph embedding module includes a convolutional network, which consists of multiple convolutional layers; The weighted parameters corresponding to the embedding matrix are learned to obtain the multi-view information.

5. The method according to claim 4, characterized in that, The graph embedding module is used to extract features from the initial drug similarity matrix and the initial protein similarity matrix to obtain an embedding matrix, including: The initial drug similarity matrix and the initial protein similarity matrix are initialized using random initialization weights and random initialization biases to obtain the initialization matrix; The initialization matrix is ​​updated using the SUM function to obtain the updated matrix; The update matrix is ​​then subjected to normalization, activation, and regularization processes in sequence to obtain the embedding matrix.

6. The method according to claim 1, characterized in that, The step of obtaining the feature graph and the topological graph based on the feature adjacency matrix, the topological adjacency matrix, and the feature matrix includes: The feature map is obtained by combining the feature adjacency matrix and the feature matrix. The topological adjacency matrix and the feature matrix are combined to obtain the topological graph.

7. The method according to claim 1, characterized in that, The calculation of similarities between different elements in the drug-related network and the protein-related network to obtain the initial drug similarity matrix and the initial protein similarity matrix includes: Calculate the Jaccard similarity coefficients between different elements in the drug-related network and the protein-related network to obtain the initial drug similarity matrix and the initial protein similarity matrix.

8. The method according to claim 1, characterized in that, The drug-related networks include drug-drug relationship networks, drug-related disease networks, drug-side effect networks, and drug-chemical structure similarity networks; the drug-drug relationship networks are used to represent the interactions between different drugs; the drug-related disease networks are used to represent the relationships between drugs and diseases; and the drug-side effect networks are used to represent the relationships between drugs and side effects. The drug-chemical structure similarity network is used to represent the relationship between the chemical structures of different drugs; The protein-related network includes a protein-related disease network, a protein-protein relationship network, and a protein sequence similarity network; the protein-related disease network is used to represent the relationship between proteins and diseases; the protein-protein relationship network is used to represent the interaction between different proteins; and the protein sequence similarity network is used to represent the relationship between different protein sequences.

9. A drug-target interaction prediction device based on deep neural networks, characterized in that, include: The preprocessing module is configured to calculate the similarity between different elements in the drug-related network and the protein-related network to obtain the initial drug similarity matrix and the initial protein similarity matrix. The initial drug similarity matrix and the initial protein similarity matrix are preprocessed to obtain the target drug similarity matrix and the target protein similarity matrix; the target drug similarity matrix and the target protein similarity matrix are then concatenated to obtain the DPP. It is configured to use the KNN algorithm to extract features from the DPP, thereby obtaining the feature adjacency matrix of the DPP; The DPP is processed according to preset rules to obtain the topological adjacency matrix of the DPP; The graph embedding module is configured to extract features from the initial drug similarity matrix and the initial protein similarity matrix to obtain multi-view information. The feature extraction module is configured to extract features from the multi-view information to obtain a feature matrix; The prediction module is configured to obtain a feature map and a topological map based on the feature adjacency matrix, the topological adjacency matrix, and the feature matrix; input the feature map and the topological map into a graph attention network for feature extraction to obtain a first target feature and a second target feature; input the feature map and the topological map into a graph convolutional network for feature extraction to obtain a third target feature; input the first target feature, the second target feature, and the third target feature into an attention network for feature fusion to obtain a fused feature; and input the fused feature into a fully connected layer to obtain a prediction result. The feature extraction module includes an ECR network; the ECR network includes an ECR input layer, an ECR convolutional layer, an ECR pooling layer, a CECJC layer, and an ECR output layer connected in sequence; the ECR network combines the advantages of spatial dilated convolution and skip connections, and captures neighborhood information at three different scales—local, intermediate, and long distance—through three adaptive convolutional kernels. The ECR network is shown in the following formulas (4), (5), and (6): （4） （5） （6） In formula (4) This is the formula for a convolutional layer, where x is the original input. It is the kernel size. For filling, It is the expansion rate. This represents a regular two-dimensional convolution. In formula (5), the activation function is represented. It is the pooling layer formula. Indicates adaptive average pooling. This indicates that interpolation is used for sampling; in formula (6) It is the formula for the CECJC layer. It is a merging of channels. Indicates use Prevent overfitting.

Images