Prediction method of drug-target interaction based on graph attention network and multi-scale feature fusion

Abstract

The application belongs to the technical field of drug discovery. The application provides a prediction method for drug-target interaction based on a graph attention network and multi-scale feature fusion. In the disclosure, a large medical knowledge graph is introduced into a DTI prediction task, and rich biological information in the knowledge graph is used to mine potential relationships between drugs and targets. The sequence information of the drugs and targets, as well as the position information between the drugs and other medical entities, is considered, and the intrinsic characteristics and extrinsic connections are comprehensively considered. When the position information is obtained, the local position information between the drugs and targets, as well as the global position information of the drugs and targets in the knowledge graph, is simultaneously focused on, and a graph attention network is used to fuse the position information of different scales, so that the comprehensive position information is helpful to obtain high-quality drug and target protein features.

Description

Technical Field

[0001] This disclosure relates to the field of drug discovery technology, and more particularly to a method for predicting drug-target interactions based on graph attention networks and multi-scale feature fusion. Background Technology

[0002] Drug discovery refers to the process of discovering new drugs or developing existing drugs through systematic scientific research and experimentation. Its history can be traced back millions of years to the era when only herbs were used. Traditional new drug development is a complex system engineering project, requiring large-scale screening to discover lead compounds, followed by repeated in vitro, animal, and in vivo experiments by drug experts for testing and optimization. In the 1980s, with the rapid development of computers, related technologies began to extend to the field of drug discovery. Computer-aided drug design is a multidisciplinary research field that aims to use computer methods to find, develop, and analyze potential connections between drugs and related bioactive compounds. It encompasses a series of tasks and related computational methods, including virtual screening, lead compound optimization, and drug side effect prediction. This accelerates the early stages of chemical development, guides and speeds up the drug development process, effectively alleviates the time and cost limitations of traditional laboratory methods, provides a faster and more economical drug development pathway, and further meets the ever-changing needs for disease treatment.

[0003] Drug-target interaction (DTI) prediction is a hot topic in computer-aided drug design. This task aims to predict the interaction between drug molecules and their targets to reveal potential drug targets and accelerate the discovery and development of new drugs. Since most drugs achieve therapeutic effects through interaction with specific target molecules in vivo, such as enzymes, nuclear receptors, G-protein-coupled receptors (GPCRs), and ion channels, multiple areas of drug development, such as lead generation and optimization, drug repositioning, and multiplex pharmacology, involve the identification and prediction of DTIs. However, traditional laboratory methods have the following drawbacks: (1) traditional pharmacological methods for identifying DTIs are costly and time-consuming; (2) the spatial structures of proteins and drug molecules are complex, and the interaction mechanisms involve multiple levels. Given these drawbacks, researchers can use large-scale bioinformatics data and computer methods to predict potential DTIs, thereby narrowing the search space to supplement and assist subsequent wet experiments, thus accelerating the drug development process.

[0004] To address the above shortcomings, various machine learning-based methods have been proposed. However, existing computer-based DTI prediction methods still have some deficiencies, such as: (1) Many prediction models focus on the structural or sequence information of drugs and target proteins themselves, or focus on the local interaction relationships between a few key medical entities, while rarely paying attention to the global performance of drugs and target proteins in the medical knowledge graph, which to some extent limits the predictive ability of the models; (2) Many existing works start from single-scale features and only focus on the performance of drugs and target proteins at a certain level, while ignoring the potential synergistic effect between multi-scale information, which is not conducive to obtaining high-quality and expressive drug and target features. Summary of the Invention

[0005] To avoid the shortcomings of existing technologies, this application provides a drug-target interaction prediction method based on graph attention network and multi-scale feature fusion, which solves the problems of existing technologies that lack consideration of the global performance of drugs and targets and lack consideration of the potential synergistic effects between multi-scale information.

[0006] According to embodiments of this disclosure, a method for predicting drug-target interactions based on graph attention networks and multi-scale feature fusion is provided, the method comprising:

[0007] Acquire drug sequence data, protein sequence information, and drug-target protein association data; wherein the drug-target protein association data includes DTI association data and knowledge graph data;

[0008] The Rdkit ensemble method was used to extract drug attribute information from the drug sequence data and target protein attribute information from the protein sequence information.

[0009] Based on the DTI-related data and the knowledge graph data, the global location features of the drug, the global location features of the target protein, the comprehensive location information of the drug, and the comprehensive location information of the target protein are obtained.

[0010] Multi-scale feature fusion is performed on the drug attribute information, the target protein attribute information, the comprehensive location information of the drug, and the comprehensive location information of the target protein to obtain a representation of the drug pair;

[0011] A predictor is constructed based on a three-layer fully connected network, and a network model is built based on the predictor.

[0012] The network model is trained using the representation of the drug pairs, and the association probability of the drug-target pairs is obtained.

[0013] Furthermore, the steps of acquiring drug sequence data, protein sequence information, and drug-target protein association data include:

[0014] The drug sequence data is obtained from the Drugbank database, the protein sequence information is obtained from the UniProt database, and the drug-target protein association data is obtained from the PrimeKG database; wherein, the drug-target protein association data includes the DTI association data and the knowledge graph data;

[0015] The DTI-related data is divided into a training set, a validation set, and a test set.

[0016] Furthermore, the step of extracting drug attribute information from the drug sequence data and target protein attribute information from the protein sequence information using the Rdkit ensemble method includes:

[0017] The Rdkit ensemble method is used to convert the drug sequence data into a molecular map corresponding to the drug.

[0018] The embedding of each atom is obtained from the molecular graph using an attribute masking method;

[0019] The embedding of the atoms is subjected to average pooling to obtain a representation of the molecular graph, and the representation of the molecular graph is used as drug property information;

[0020] Each protein sequence in the protein sequence information is converted into a pseudo-position-specific scoring matrix feature to encode the evolution and sequence information of proteins with different sequence lengths;

[0021] The target protein representation is obtained based on the features of the pseudo-position-specific scoring matrix, and the target protein representation is used as target protein attribute information.

[0022] Furthermore, the step of obtaining the global location features of the drug, the global location features of the target protein, the comprehensive location information of the drug, and the comprehensive location information of the target protein based on the DTI association data and the knowledge graph data includes:

[0023] The knowledge graph data is learned using knowledge graph embedding technology to obtain all entity embeddings. The entity embeddings of drugs in all the entity embeddings are used as the global position features of the drugs, and the entity embeddings of target proteins in all the entity embeddings are used as the global position features of the target proteins.

[0024] The training set in the DTI-related data is processed to generate positive and negative examples on a one-to-one basis;

[0025] Using the positive examples in the training set as edges, the global position features of the drug as the initial features of the drug node, and the global position features of the target protein as the initial features of the target protein, a bipartite graph of the drug and the target protein is constructed.

[0026] The bipartite graph is iteratively learned using a graph attention network, enabling each edge in the bipartite graph to adaptively learn weights, and the features of the nodes in the graph are updated through message passing to obtain updated drug features and updated target protein features.

[0027] The updated drug features are used as the comprehensive location information of the drug, and the updated target protein features are used as the comprehensive location information of the target protein.

[0028] Furthermore, the step of performing multi-scale feature fusion on the drug attribute information, the target protein attribute information, the comprehensive location information of the drug, and the comprehensive location information of the target protein to obtain a representation of the drug pair includes:

[0029] A vector concatenation operation is performed on the drug attribute information, the target protein attribute information, the comprehensive location information of the drug, and the comprehensive location information of the target protein to obtain a representation of the drug pair.

[0030] Furthermore, the step of constructing a predictor based on a three-layer fully connected network and building a network model based on the predictor includes:

[0031] A predictor is built based on a three-layer fully connected network, and the output dimension is set to 1 to ensure that the output of the network model is a number between 0 and 1.

[0032] Based on the predictor, a network model is constructed.

[0033] Furthermore, the step of training the network model using the representation of the drug pair and obtaining the association probability of the drug-target pair includes:

[0034] The network model is trained using the representation of the drug pairs to obtain the trained network model, and the association probability of drug-target pairs in the training set is output.

[0035] The trained network model is validated using the validation set.

[0036] The validated network model is tested using the test set to obtain the prediction results of the test set.

[0037] Furthermore, the objective function for optimizing the network model is the meta-cross-entropy loss function.

[0038] The technical solutions provided by the embodiments of this disclosure may include the following beneficial effects:

[0039] In the embodiments of this disclosure, a large medical knowledge graph is introduced into the DTI prediction task through the aforementioned drug-target interaction prediction method based on graph attention networks and multi-scale feature fusion. The rich biological information within the knowledge graph is used to mine potential relationships between drugs and targets. The method considers the sequence information of drugs and targets themselves, as well as their positional information with other medical entities, comprehensively taking into account both their inherent characteristics and external connections. When obtaining positional information, both local positional information between drugs and targets and their global positional information within the knowledge graph are considered. A graph attention network is used to fuse positional information at different scales, resulting in comprehensive positional information that helps obtain high-quality drug and target protein features. Attached Figure Description

[0040] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with this disclosure and, together with the description, serve to explain the principles of this disclosure. It is obvious that the drawings described below are merely some embodiments of this disclosure, and those skilled in the art can obtain other drawings based on these drawings without any inventive effort.

[0041] Figure 1 A flowchart illustrating the steps of a drug-target interaction prediction method based on graph attention networks and multi-scale feature fusion in an exemplary embodiment of this disclosure is shown.

[0042] Figure 2 This diagram illustrates message passing in a graph attention network in an exemplary embodiment of this disclosure.

[0043] Figure 3 A schematic diagram of the network model in an exemplary embodiment of this disclosure is shown. Detailed Implementation

[0044] Exemplary embodiments will now be described more fully with reference to the accompanying drawings. However, these exemplary embodiments can be implemented in many forms and should not be construed as limited to the examples set forth herein; rather, they are provided so that this disclosure will be more comprehensive and complete, and will fully convey the concept of the exemplary embodiments to those skilled in the art. The described features, structures, or characteristics may be combined in any suitable manner in one or more embodiments.

[0045] Furthermore, the accompanying drawings are merely illustrative diagrams of embodiments of this disclosure and are not necessarily drawn to scale. The same reference numerals in the drawings denote the same or similar parts, and therefore repeated descriptions of them will be omitted. Some block diagrams shown in the drawings are functional entities and do not necessarily correspond to physically or logically independent entities.

[0046] This example implementation provides a method for predicting drug-target interactions based on graph attention networks and multi-scale feature fusion. (Reference) Figure 1 As shown, the drug-target interaction prediction method based on graph attention network and multi-scale feature fusion may include steps S101 to S106.

[0047] Step S101: Obtain drug sequence data, protein sequence information, and drug-target protein association data; wherein, the drug-target protein association data includes DTI association data and knowledge graph data;

[0048] Step S102: Use the Rdkit ensemble method to extract drug attribute information from the drug sequence data and target protein attribute information from the protein sequence information;

[0049] Step S103: Based on the DTI association data and the knowledge graph data, obtain the global location features of the drug, the global location features of the target protein, the comprehensive location information of the drug, and the comprehensive location information of the target protein;

[0050] Step S104: Perform multi-scale feature fusion on the drug attribute information, the target protein attribute information, the comprehensive location information of the drug, and the comprehensive location information of the target protein to obtain a representation of the drug pair;

[0051] Step S105: Construct a predictor based on a three-layer fully connected network, and construct a network model based on the predictor;

[0052] Step S106: Train the network model using the representation of the drug pair and obtain the association probability of the drug-target pair.

[0053] This method, based on graph attention networks and multi-scale feature fusion, introduces a large medical knowledge graph into the drug-target interaction (DTI) prediction task. It leverages the rich biological information within the knowledge graph to uncover potential relationships between drugs and targets. The method considers both the sequence information of drugs and targets, as well as their positional information with other medical entities, comprehensively taking into account both intrinsic characteristics and external connections. When obtaining positional information, it simultaneously focuses on the local positional information between drugs and targets, as well as their global positional information within the knowledge graph. A graph attention network is used to fuse positional information at different scales, resulting in comprehensive positional information that helps obtain high-quality drug and target protein features.

[0054] Below, we will refer to Figures 1 to 3 The steps of the drug-target interaction prediction method based on graph attention network and multi-scale feature fusion described in this example embodiment will be explained in more detail.

[0055] In step S101, the sequence data of the drug approved by the U.S. Food and Drug Administration is obtained from the Drugbank database. The sequence data used in this application is the Simplified Molecular-InputLine Entry System (SMILES).

[0056] The protein sequence information was obtained from the UniProt database. The protein sequence used in this application consists of multiple single letters, each letter representing a certain type of amino acid residue.

[0057] To avoid tag leakage, this application obtains drug and target protein association data from the large medical knowledge graph PrimeKG. The first part is all known DTI association data, which is used to construct training, validation and test sets. The second part is the knowledge graph data remaining after excluding DTI, which is used to mine global information about drugs and target proteins.

[0058] Step S102 comprises two parts: extraction of drug property information and extraction of target protein property information. The specific technical solution is as follows:

[0059] This application obtains drug attribute information from SMILES sequence data. First, step S101 yields a set of SMILES sequences for 2162 approved drugs. Then, to obtain the drug attribute information, the application uses the Rdkit ensemble tool to convert the SMILES sequences into molecular graphs corresponding to the drugs. Attribute masking is then used on the molecular graphs to obtain the embeddings of each atom. Finally, average pooling is performed on all atom embeddings to obtain a representation of the molecular graph, which is then used as the drug attribute information.

[0060] This application obtains the attribute information of target proteins from protein sequences. First, a set of 1296 protein sequences can be obtained through step S101. Then, each protein sequence is converted into a pseudo-position specific scoring matrix (PsePSSM) feature to encode the evolution and sequence information of proteins with different sequence lengths. The final target protein representation is then used as the attribute information of the target protein.

[0061] Step S102 comprises two parts: extraction of local location information of the drug and target protein, and extraction of global location information of the drug and target protein. The specific technical solution is as follows:

[0062] The knowledge graph data is learned using the knowledge graph embedding technology ComplEx. This application first obtains knowledge graph data (with DTI data removed) through step S101, and then uses the knowledge graph embedding technology ComplEx to learn this data, obtaining the embeddings of all entities. The embeddings of drug entities are used as the global positional features of the drug, and the embeddings of target protein entities are used as the global positional features of the target protein. This part of the data focuses on the global representation of drugs and target proteins within the entire medical knowledge graph.

[0063] This application divides DTI data and constructs a bipartite graph of drugs and target proteins. First, in step S101, all DTI data is obtained, and negative examples are generated one-to-one. To avoid label leakage, all DTI data is divided into training, validation, and test sets. This application uses positive examples from the training set and the obtained global location information to construct a bipartite graph of drugs and target proteins. Specifically: positive examples from the training set are used as edges in the bipartite graph; global location information of the drugs is used as the initial features of the drug nodes; and global location information of the target proteins is used as the initial features of the target proteins.

[0064] Graph Attention Networks (GATs) are used to update the node representations of drugs and target proteins in a bipartite graph. After obtaining the bipartite graph, this application uses a GAT to iteratively learn the graph. GAT enables each edge in the bipartite graph to adaptively learn weights and updates the features of nodes in the graph through message passing. It is important to note that GAT is a model designed for homogeneous graphs. To make GAT suitable for this application, relevant functions from the PyG library are used to transform it into a GAT model suitable for heterogeneous graphs. Finally, through GAT, we can obtain updated drug features as comprehensive drug location information, and updated target protein features as comprehensive target protein location information.

[0065] Steps S104 and S105 consist of two parts: the fusion of multi-scale features of the drug and target protein, and the construction of the predictor. The specific technical solutions are as follows:

[0066] Fusion of multi-scale features of drugs and target proteins. By concatenating four vectors—drug attribute information, drug comprehensive location information, target protein attribute information, and target protein comprehensive location information—a representation of a drug pair can be obtained.

[0067] Predictor Construction. In this application, the vector representation of the drug-target pair needs to be passed through a predictor to obtain the association probability of the drug-target pair. This application uses a three-layer fully connected network as the predictor and sets the output dimension to 1 to ensure that the model outputs a number between 0 and 1. The higher the probability, the greater the association between the drug-target pair, and vice versa.

[0068] Step S105 comprises two parts: model training and model testing. The specific technical solution is as follows:

[0069] Model Training. In this application, data from the training set is used for model training, and the objective function for optimization is a common binary classification loss function, namely the binary cross-entropy loss function. Iterative training using labeled data from the training set is used to determine suitable network parameters to minimize the loss function. To avoid overfitting, early stopping is also employed in this application, where model performance is tested on the validation set to determine whether early stopping is necessary.

[0070] Model testing. After training the model, a trained network is obtained. Then, unlabeled data in the test set is used to evaluate the model's final performance. During testing, the network parameters are not iteratively updated. Comparing the predicted drug-target pairs in the test set with the actual results can be used to evaluate the model's final performance.

[0071] like Figure 2 The diagram shown is a schematic representation of the network model.

[0072] In one specific embodiment, sequence data of FDA-approved drugs are obtained from the Drugbank database. This application first downloads the file `approved_structure links.csv` from the latest Drugbank database, which contains all currently FDA-approved drugs and related information, and then extracts the SMILES data from it as the drug sequence data.

[0073] This application obtains protein sequence information from the UniProt database.

[0074] The knowledge graph data was processed. This application first obtained triplet data (kg.csv) and node information (node.csv) from the PrimeKG dataset. Then, the triplet data was divided into two parts: one part contained all known DTI association data, and the remaining knowledge graph data was used to obtain global information on drugs and target proteins. To ensure that all drugs and target proteins involved in the invention could obtain attribute and location information, the DTI data was filtered, resulting in 9520 DTI records between 2162 approved drugs and 1298 target proteins. It should be noted that the representation of target proteins in PrimeKG uses IDs from the NCBI database. To ensure that these correspond to the UniProt IDs and protein sequences downloaded in this application, the HGNC database was used for mapping.

[0075] This paper extracts the attribute information of a drug from its SMILES sequence data. First, the molecular graph of the drug, the original features of the nodes in the molecular graph (closely related to attributes, including information such as atom type and chirality), and the edge features are obtained through Rdki t. Then, some nodes are replaced with masked nodes, i.e., their original features are replaced with special values. The node embeddings are then iteratively updated using a graph neural network. The iterative result of node v in the k-th layer of the graph neural network can be represented as:

[0076]

[0077] Where N(v) represents the set of nodes adjacent to node v, and e represents all edges connected to node v, where (v,v) represents the self-loop edge of the node. Then, based on the updated node embeddings, a linear model is applied to predict the attributes of the masked nodes, ensuring that the updated representation of the masked nodes approximates its original features as closely as possible. After training, this graph neural network can effectively aggregate the attributes of neighborhood nodes, and the resulting atom embeddings can well reflect the attribute information of the atom. Finally, this application performs average pooling on the embeddings of all atoms to obtain the representation of the molecular graph, i.e.:

[0078]

[0079] Where h G It can be viewed as a molecular representation that aggregates all atomic information. MEAN represents the average pooling operation, K represents the last layer of GNN (i.e., k = K), and V represents the set of nodes in the molecular graph.

[0080] This application extracts attribute information of target proteins from protein sequences. For target protein sequences with L amino acid residues, a position-specific scoring matrix (PSSM) is used as its descriptor, which represents a protein as a PSSM of size L×20.

[0081]

[0082] Where j represents one of the 20 natural amino acids, E i →j represents the score for the mutation of the residue at position i in the amino acid sequence into the j-th amino acid residue. It can be obtained by searching the Swiss-Prot database using PSI-BLAST. A positive score indicates that the mutation frequency of the corresponding residue is higher than expected, while a negative score indicates that it is lower than expected. Then, each element Ei→j in the matrix is ​​normalized to the (0,1) interval:

[0083]

[0084] Because different target proteins have different L values, the number of rows in the matrix also varies. To make the PSS M descriptor a uniform representation, a unified PSSM can be used, i.e.:

[0085]

[0086] Where T represents the transpose operator, and E i→j The method for calculating the average score of residues in a protein during evolution (i.e., the amino acid type mutates from i to j) is as follows:

[0087]

[0088] However, if a unified PSSM is used to represent proteins, the sequence information of amino acids in the target protein will be lost. To avoid this deficiency, PsePSSM incorporates the concept of "pseudo-amino acid composition" and makes the following changes based on the unified PSSM:

[0089]

[0090] in:

[0091]

[0092] in The relevant factor represents the j-th amino acid, L refers to the number of amino acids in the target protein, which can also be understood as the maximum continuous distance of the protein sequence, and λ∈[0,L] is a hyperparameter representing the continuous distance along the protein sequence. In this application, λ is set to 10. In general, PsePSSM can convert protein sequences of different lengths in the dataset into vectors of the same dimension (i.e., 20+20×λ) after feature extraction. For each protein sequence, its attribute information can be obtained using PsePSSM.

[0093] The specific process of step S103 is as follows:

[0094] Extraction of global location information. For knowledge graph data, this application uses the knowledge graph embedding technology ComplEx to embed entities, using the embedding of drug entities as the global location feature of the drug and the embedding of target protein entities as the global location feature of the target protein.

[0095] This application constructs a bipartite graph of drugs and target proteins by combining local location information. Specifically, it uses positive examples from the training set and global location information to construct the bipartite graph. The positive examples from the training set are used as edges in the bipartite graph, the global location information of the drug is used as the initial feature of the drug node, and the global location information of the target protein is used as the initial feature of the target protein.

[0096] Global Attention (GAT) is used to fuse global and local location information to obtain comprehensive location information for the drug and target protein. GAT first uses a self-attention mechanism to calculate the attention weights between two nodes, i.e.:

[0097] α ij =attention(h i ,h j ),α ij ∈R 1 (9)

[0098] Where attention represents the attention mechanism, α ij This represents the influence of node j on node i. In the convolutional layers of GAT, the attention mechanism is implemented through a single-layer neural network and controlled by a weight matrix. Parameterize the function, and then use the LeakyReLu function to perform a nonlinear transformation, i.e.:

[0099]

[0100] Where T represents the matrix transpose operation, and || represents the vector concatenation operation.

[0101] Typically, without considering structural information, the model needs to calculate the weights of all nodes j with respect to node i. To utilize the graph's structural information, this application performs masked attention based on the input adjacency matrix, considering only the node itself and its neighbors when calculating the attention score. Then, the attention weights α need to be adjusted. ij Perform a normalization operation, that is:

[0102]

[0103] After obtaining the normalized attention score, this application computes a linear combination of node features through message passing and outputs the final aggregated feature for each node, i.e.:

[0104]

[0105] Where σ represents a nonlinear operation, this application uses the softmax function, h′ i W represents the final feature of a node, and is a weight matrix used for dimension transformation.

[0106] like Figure 3 The diagram shown illustrates message passing in a graph attention network.

[0107] The specific processes of steps S104 and S105 are as follows:

[0108] First, the final representations of the drug and target protein are linked together as the representation of the DTI pair, i.e.:

[0109] X pair =Concat(AI) D ,PI D AI T ,PI T (13)

[0110] Where AID represents drug attribute information, PID represents comprehensive drug location information, AIT represents target protein attribute information, and GIT represents comprehensive target protein location information. The DTI pair representations are then input into a three-layer fully connected network to obtain the predicted probabilities, i.e.:

[0111]

[0112] Here, MLP represents a fully connected network, and Xpair represents a DTI pair representation. σ represents the probability that a DTI pair is associated, and σ represents the sigmoid activation function.

[0113] In step S105, the model training process is as follows:

[0114] To train the model, this application assigns a binary class label y∈{0,1} to each drug-target protein pair, representing the presence or absence of interaction. All samples are divided into training data, validation data, and test data. Samples from the training data are used for information transfer and model parameter training. The model's performance on the validation data is used to determine whether to stop early. Finally, the test data is used to obtain the final performance results. During model training, this application minimizes the true label y and the predicted probability. The binary cross-entropy loss between the two variables is used to optimize the model parameters, i.e.:

[0115]

[0116] Hyperparameter values ​​during the experiment:

[0117] The choice of hyperparameters is closely related to model performance. The four hyperparameters affecting the model performance in this application are the number of layers k in the graph attention network, the hidden layer dimension d in the fully connected network of the predictor, the learning rate lr, and the batch size BS. In this application, k=2, d=1024, lr=0.001, and BS=1024 are used as the hyperparameter values ​​of the model.

[0118] This method, based on graph attention networks and multi-scale feature fusion, introduces a large medical knowledge graph into the drug-target interaction (DTI) prediction task. It leverages the rich biological information within the knowledge graph to uncover potential relationships between drugs and targets. The method considers both the sequence information of drugs and targets, as well as their positional information with other medical entities, comprehensively taking into account both intrinsic characteristics and external connections. When obtaining positional information, it simultaneously focuses on the local positional information between drugs and targets, as well as their global positional information within the knowledge graph. A graph attention network is used to fuse positional information at different scales, resulting in comprehensive positional information that helps obtain high-quality drug and target protein features.

[0119] In the description of this specification, the references to terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of this disclosure. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples. In addition, those skilled in the art can combine and integrate the different embodiments or examples described in this specification.

[0120] Other embodiments of this disclosure will readily occur to those skilled in the art upon consideration of the specification and practice of the invention disclosed herein. This application is intended to cover any variations, uses, or adaptations of this disclosure that follow the general principles of this disclosure and include common knowledge or customary techniques in the art not disclosed herein. The specification and examples are to be considered exemplary only, and the true scope and spirit of this disclosure are indicated by the appended claims.

Claims

1. A method for predicting drug-target interactions based on graph attention networks and multi-scale feature fusion, characterized in that, The method includes: Acquire drug sequence data, protein sequence information, and drug-target protein association data; among which, drug-target protein association data includes DTI association data and knowledge graph data; The Rdkit ensemble method was used to extract drug property information from drug sequence data and target protein property information from protein sequence data. Based on DTI-related data and knowledge graph data, we obtain the global location features of the drug, the global location features of the target protein, the comprehensive location information of the drug, and the comprehensive location information of the target protein. Multi-scale feature fusion is performed on drug attribute information, target protein attribute information, comprehensive drug location information, and comprehensive target protein location information to obtain a representation of drug pairs; A predictor is built based on a three-layer fully connected network, and a network model is constructed based on the predictor. The network model is trained using drug-target pair representations, and the association probabilities of drug-target pairs are obtained; where, The steps for extracting drug attribute information from drug sequence data and target protein attribute information from protein sequence information using the Rdkit ensemble method include: The Rdkit ensemble method was used to convert drug sequence data into molecular maps corresponding to the drugs. The embedding of each atom is obtained by using the attribute masking method in the molecular diagram; Average pooling is performed on the atomic embeddings to obtain a molecular graph representation, and the molecular graph representation is used as drug property information. Each protein sequence in the protein sequence information is converted into a pseudo-position-specific scoring matrix feature to encode the evolution and sequence information of proteins with different sequence lengths; The target protein representation is obtained based on the characteristics of the pseudo-position-specific scoring matrix, and the target protein representation is used as the target protein attribute information. The specific steps for obtaining attribute information of a target protein from a protein sequence include: For a protein sequence with L amino acid residues, a position-specific scoring matrix is ​​used as its descriptor, representing a protein as a PSSM of size L×20: in, It represents one of the 20 natural amino acids. Represents the first amino acid in the sequence The residue variation at the position is the _th The score for each amino acid residue is calculated, with a positive score indicating that the mutation frequency of the corresponding residue is higher than expected, and a negative score indicating that it is lower than expected. Each element of the matrix Standardize to the (0,1) interval: This makes the PSSM descriptor a unified representation, that is: Where T represents the transpose operator. The formula used to calculate the average score of residues in a protein over its evolutionary process is as follows: Based on the pseudo-amino acid composition, the unified PSSM is converted to: in, Representing the Factors related to various amino acids; ∈[0,L] is a hyperparameter; representing the continuous distance along the protein sequence; The amino acid sequence number The position shifts backward. After the first position, the second The residue variation at the position is the _th Standardized scores of amino acid residues.

2. The method for predicting drug-target interactions based on graph attention networks and multi-scale feature fusion according to claim 1, characterized in that, The steps for obtaining drug sequence data, protein sequence information, and drug-target protein association data include: Drug sequence data was obtained from the Drugbank database, protein sequence information was obtained from the UniProt database, and drug-target protein association data was obtained from the PrimeKG database; among which, drug-target protein association data included DTI association data and knowledge graph data. The DTI-related data is divided into training set, validation set and test set.

3. The method for predicting drug-target interactions based on graph attention networks and multi-scale feature fusion according to claim 2, characterized in that, The steps for obtaining the global location features of the drug, the global location features of the target protein, the comprehensive location information of the drug, and the comprehensive location information of the target protein based on DTI-linked data and knowledge graph data include: Knowledge graph embedding technology is used to learn knowledge graph data to obtain all entity embeddings. The entity embeddings of drugs in all entity embeddings are used as the global location features of drugs, and the entity embeddings of target proteins in all entity embeddings are used as the global location features of target proteins. Process the training set in the DTI correlated data to generate positive and negative examples on a one-to-one basis; Using the positive examples in the training set as edges, the global position features of the drug as the initial features of the drug node, and the global position features of the target protein as the initial features of the target protein, a bipartite graph of the drug and the target protein is constructed. The graph attention network is used to iteratively learn the weights of each edge in the bipartite graph, and the features of the nodes in the graph are updated through message passing to obtain the updated drug features and updated target protein features. The updated drug features are used as the comprehensive location information of the drug, and the updated target protein features are used as the comprehensive location information of the target protein.

4. The method for predicting drug-target interactions based on graph attention networks and multi-scale feature fusion according to claim 3, characterized in that, The step of performing multi-scale feature fusion on drug attribute information, target protein attribute information, comprehensive drug location information, and comprehensive target protein location information to obtain a representation of the drug pair includes: A vector concatenation operation is performed on drug attribute information, target protein attribute information, comprehensive drug location information, and comprehensive target protein location information to obtain a representation of the drug pair.

5. The method for predicting drug-target interactions based on graph attention networks and multi-scale feature fusion according to claim 4, characterized in that, The steps for constructing a predictor based on a three-layer fully connected network and building a network model from the predictor include: A predictor is built based on a three-layer fully connected network, and the output dimension is set to 1 to ensure that the output of the network model is a number between 0 and 1. Based on the predictor, a network model is built.

6. The method for predicting drug-target interactions based on graph attention networks and multi-scale feature fusion according to claim 5, characterized in that, The steps of training the network model using drug-target pair representations and obtaining the association probabilities of drug-target pairs include: The network model is trained using the drug-target pair representation to obtain a trained network model, and the association probability of drug-target pairs in the training set is output. The trained network model is validated using a validation set; The validated network model is tested using a test set to obtain the prediction results for the test set.

7. The method for predicting drug-target interactions based on graph attention networks and multi-scale feature fusion according to claim 6, characterized in that, The objective function for optimizing the network model is the meta-cross-entropy loss function.

Images