A method, model, device, and medium for predicting drug-target interactions

Abstract

The application discloses a kind of methods for predicting drug-target interaction, model, equipment and medium, method includes: generating drug-protein feature vector subgraph;Graph attention sampling is applied, and the center node of subgraph is updated;Adaptive iterative optimization of feature vector is carried out using NDLS;While using graph SAGE to update the feature vector of center node, obtain node multi-source information;Through double-layer GTN, node multi-source information is aggregated, and output prediction value.The application uses NDLS to carry out adaptive iterative optimization of node feature, uses graph SAGE to carry out deep sampling aggregation to neighborhood node, and pays attention to the receptive field of different levels of network node;More relevant node information is effectively aggregated from neighborhood;The method of the application considers the receptive field of different levels of network node, by fusing respective complementary advantages, so that more relevant node information is effectively aggregated from neighborhood, and the prediction ability of DTI prediction model is further improved.

Description

Technical Field

[0001] This invention relates to the field of drug-target interaction prediction technology, and in particular to a method, model, device and medium for predicting drug-target interactions. Background Technology

[0002] In the characterization of drugs and proteins, a large number of studies rely on the sequence information of drugs and proteins to infer potential drug-target interactions (DTIs). Potential drug-target interactions (DTIs) are a key component of drug discovery and new uses. This type of method has advantages such as simplicity, efficiency and wide applicability.

[0003] Existing sequence structures often learn from the hidden representations of a single drug, ignoring the valuable information provided by the interaction between two drugs for the substructure.

[0004] For example, invention application No. 202110382488.9 discloses a drug-target interaction prediction model method based on deep embedding learning of molecular graphs and sequences. This method establishes a graph neural network based on attention mechanism and an attention-guided bidirectional LSTM to predict interactions. On the one hand, in terms of drug molecules, molecular graphs can learn better spatial features; on the other hand, the large amount of protein sequence data can cover a larger protein space and improve generalization ability.

[0005] However, the published solution also has the following drawbacks: when extracting the sequence features of drugs and proteins, the feature information between drugs and proteins is relatively independent, which will ignore the hidden characterization information that interactions with other drugs and proteins can provide for the substructure.

[0006] Furthermore, while Graph Neural Networks (GNNs) are widely used in the DTI field, they still have other limitations. In existing methods, shallow GNNs with limited receptive fields can only aggregate incomplete information within the neighborhood, failing to consider the receptive fields of different levels of network nodes and thus not effectively aggregating more relevant node information from the neighborhood.

[0007] For example, invention application with application number 202211080659.3 discloses a drug and target prediction method based on graph attribute neural network. The scheme disclosed in this application uses graph attribute neural network to reduce the dependence of deep learning model for drug and target interaction prediction on training samples and improve prediction performance.

[0008] However, the proposed solution also has the following drawbacks: the receptive field is limited, GNN can only aggregate incomplete information within the neighborhood, it cannot take into account the receptive fields of different levels of network nodes, and it cannot effectively aggregate more relevant node information from the neighborhood.

[0009] Therefore, in reality, there is a need for a method to predict drug-target interactions that can dynamically extract potential interaction information from other drugs (targets); and take into account the receptive fields of different levels of network nodes, effectively aggregating more relevant node information from the neighborhood by integrating their complementary advantages, thereby improving the predictive ability of the DTI prediction model. Summary of the Invention

[0010] To address the aforementioned problems, the present invention aims to provide a method, model, device, and medium for predicting drug-target interactions, focusing on the receptive fields of different levels of graph nodes and performing adaptive iterative optimization of graph node features.

[0011] This invention provides a method, model, device, and medium for predicting drug-target interactions.

[0012] First aspect: A method for predicting drug-target interactions, comprising:

[0013] S1. Encode the sequence information of the drug and protein to generate feature vector subgraphs of the drug and protein, respectively;

[0014] S2. Apply graph attention sampling to the feature vector subgraph to obtain interaction information between drugs or proteins, and update the feature vectors of the central nodes of the drug and protein subgraphs.

[0015] S3. Adaptive iterative optimization of drug and protein feature vectors is performed using NDLS; at the same time, graph SAGE is used to update the feature vector embedding information of the central node to obtain multi-source information of the node.

[0016] S4. Information from multiple sources at the nodes is aggregated through a two-layer GTN. A fully connected layer is used as the classification module to output predicted values. Based on the predicted values, it is determined whether there is an interaction between the drug and the protein.

[0017] Furthermore, in S2, when updating the feature vectors of the center nodes of the drug and protein subgraphs:

[0018] The feature vectors of drug and protein center nodes are updated by combining randomly sampled sets of nodes with similar attributes and weights.

[0019] Furthermore, the graph attention sampling in S2 is expressed by the following formula:

[0020] (1)

[0021] (2)

[0022] V1 is the central node, and V2 is a set of nodes with the same attribute.

[0023] Furthermore, in S3, the feature vector embedding information of the center node is updated using graph SAGE, including:

[0024] The first-order neighbor nodes of the drug and protein center nodes are randomly sampled, and then their first-order neighbors are randomly selected from these neighbor nodes. Finally, starting from the outermost node, the feature embedding information of the center node is updated layer by layer from the neighbor nodes inward.

[0025] Furthermore, in S3, the feature vector embedding information of the center node is updated using the graph SAGE, as expressed by the formula:

[0026] (3)

[0027] (4)

[0028] Where v is a node, For the set of neighboring nodes, This is an update embedding for the SAGE model layer.

[0029] Furthermore, in S3, NDLS is used for adaptive iterative optimization of drug and protein feature vectors, expressed by the following formula:

[0030] (11)

[0031] (12)

[0032] in, It is a node eigenvectors, It is a norm 2. express The i-th row represents the pairs of other nodes in the k-th layer. The distribution of influence This is the distance parameter.

[0033] Furthermore, in step S4, when aggregating multi-source information of nodes through a two-layer GTN, the following steps are included:

[0034] Multi-head attention is calculated for each edge from the distant node j to the source node i, as expressed by the formula:

[0035] (13)

[0036] (14)

[0037] (15)

[0038] (16)

[0039] After obtaining multi-head attention in the graph, information is aggregated from the distant node j to the source node i, as expressed by the formula:

[0040] (17)

[0041] (18)

[0042] in, Given a feature node, For trainable parameters, Let be the exponentially scaling dot product function, and d be the size of the hidden layers in each head. Features of the source node Features of distant nodes.

[0043] The second aspect: a model for predicting drug-target interactions, including:

[0044] The sequence information encoding module generates feature vector sub-graphs of drugs and proteins based on their sequence information encoding.

[0045] The graph attention sampling module applies graph attention sampling to the drug and protein subgraphs respectively to obtain the interaction information between drugs or proteins, and updates the feature vectors of the central nodes of the drug and protein subgraphs.

[0046] The NDLS and Graph SAGE modules employ NDLS for adaptive iterative optimization of drug and protein feature vectors, while simultaneously using Graph SAGE to update the feature vector embedding information of the central nodes and obtain multi-source information of the nodes.

[0047] The GTN fusion prediction module aggregates multi-source information from nodes through a two-layer GTN and uses a fully connected layer as the classification module to output predicted values.

[0048] Third aspect: An electronic device including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor, when executing the program, performs the steps of the method provided in the first aspect.

[0049] Fourth aspect: A non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the method provided in the first aspect.

[0050] The beneficial effects of this invention are:

[0051] This invention effectively improves model performance by adjusting sampling ratios and weight allocation. It employs the NDLS module to determine the specific propagation depth for all nodes in the graph, performing adaptive iterative optimization of node features. Simultaneously, it utilizes Graph SAGE for deep sampling and aggregation of neighboring nodes, focusing on the receptive fields of network nodes at different levels. Through GTN, it integrates the complementary advantages of each method, thereby effectively aggregating more relevant node information from the neighborhood. This invention's method considers the receptive fields of network nodes at different levels, effectively aggregating more relevant node information from the neighborhood by integrating their complementary advantages, further enhancing the predictive capability of the DTI prediction model. Attached Figure Description

[0052] Figure 1 This is a flowchart illustrating the method for predicting drug-target interactions according to the present invention.

[0053] Figure 2 This is a schematic diagram of the structure of the drug-target interaction prediction model of the present invention;

[0054] Figure 3 This is a performance comparison graph of the ablation experiment of the model of the present invention;

[0055] Figure 4 This is a schematic diagram of the structure of the electronic device of the present invention. Detailed Implementation

[0056] Embodiments of the present invention are described in detail below. Examples of these embodiments are illustrated in the accompanying drawings, wherein the same or similar symbols denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and should not be construed as limiting the present invention.

[0057] When applying GNN models in the field of drug-target interaction (DTI), the learning of potential representations of drugs and targets is prone to oversmoothing problems, and shallow GNNs have limited receptive fields and can only aggregate incomplete information within the neighborhood.

[0058] To address the above problems, this invention provides a method for predicting drug-target interactions. Figure 1 This is a flowchart illustrating a method for predicting drug-target interactions provided in an embodiment of the present invention. The method includes:

[0059] S1. Encode the sequence information of the drug and protein to generate feature vector subgraphs of the drug and protein, respectively.

[0060] This invention uses two publicly available datasets, DrugBank and Davis, to train and evaluate the proposed DTI prediction model. The DrugBank dataset contains 6,708 drugs and 4,410 proteins, with 18,655 known DTI records; the Davis dataset contains 68 drugs and 379 proteins, with 7,320 DTI records.

[0061] like Figure 2 As shown, using the above dataset, the sequence information of drugs and proteins is encoded to generate feature vector subgraphs of drugs and proteins, respectively.

[0062] S2. Apply graph attention sampling to the drug and protein subgraphs respectively to obtain interaction information between drugs or proteins. Update the feature vectors of the central nodes of drugs and proteins by combining randomly sampled sets of nodes with similar attributes and weights. By customizing the sampling ratio and weight allocation, the experimental research has greater flexibility and scalability.

[0063] Graph attention sampling is a method that samples nodes with different attributes from heterogeneous graphs (feature vector subgraphs).

[0064] By applying the attention sampling method to the central node and considering the mutual information between it and nodes of the same type of attribute, we can explore the potential correlation information between nodes of various attributes in order to explore ways to optimize the prediction performance of the DTI prediction model.

[0065] In a heterogeneous graph, let the center node be V1, and the set of nodes with the same attribute containing a specified number of samples be denoted as V2. The graph attention sampling calculation formula is expressed as:

[0066] (1)

[0067] (2)

[0068] Using the above formula, the attention score is first calculated by matrix multiplication between the feature vector of the central node and the sampled node vectors of the same attribute. The attention score is then normalized using the softmax function to obtain the attention score of the sampled node.

[0069] Then, a new vector representation of the central node is obtained based on the weight combination method. The attention sampling of the graph nodes uses random sampling and weight combination to update the feature information of the central node. By customizing the sampling ratio and weight allocation, it has great flexibility and scalability.

[0070] S3. Adaptive iterative optimization of drug and protein feature vectors is performed using NDLS; at the same time, graph SAGE is used to update the feature vector embedding information of the central node to obtain multi-source information of the node.

[0071] GraphSAGE is used to randomly sample the first-order neighbor nodes of the drug and protein center nodes. Then, starting from these neighbors, their first-order neighbors are randomly selected. Finally, starting from the outermost nodes, the feature embedding information of the center nodes is updated layer by layer from the neighbor nodes inward. At the same time, NDLS is used to solve the over-smoothing or under-smoothing problem in the GNN propagation process. Based on the graph structure, the specific propagation depth is determined for all nodes in the graph, and adaptive iterative optimization of drug and protein feature vectors is performed to learn multi-neighborhood feature embeddings.

[0072] Among them, Graph SAGE (Graph Sample and Aggregate) is an inductive learning framework for learning graph nodes based on the feature information of neighboring nodes. Graph SAGE does not consider the complete k-hop neighborhood of the center node, but first samples the computation graph of the k-hop neighborhood to generate the feature representation of the center node.

[0073] For input network , where V is the set of nodes and E is the set of links or edges in the graph. Node feature matrix D is the dimension of the node features. For example, the adjacency matrix of graph A is represented as: If two nodes u and v are connected, then the value of A is 1; otherwise, it is 0.

[0074] The main purpose of using SAGE graph is to learn graph nodes. For each node Feature embeddings ∈V. For each node v, a set of neighboring nodes is created from its neighborhood. This ensures that each node has the same number of neighbors.

[0075] For each node v, the embeddings of its neighbors are aggregated using the AGGREGATE function, as shown in the formula:

[0076] (3)

[0077] (4)

[0078] In the aggregate embedding of the computational graph SAGE model Then, the embedding of node v is updated using the UPDATE function. It is an updated embedding of the SAGE model layer, and then the updated embedding is passed as the input to the next stage of the model.

[0079] Node-dependent Local Moothing (NDLS) is a strategy primarily designed to address oversmoothing and undersmoothing issues that may arise during the information propagation process in GNNs.

[0080] In existing graph neural networks, the neighbor order used for information aggregation is the same for all nodes. Due to differences in the local structure of nodes, the aggregation of neighbor information by different nodes may lead to undersmoothing or oversmoothing problems.

[0081] According to the classic GNN model, the feature representations of the drug and target at the k-th layer are denoted as... It can be obtained through feedforward recursion, and this propagation process is described as follows:

[0082] (5)

[0083] (6)

[0084] in, It is the degree matrix of the diagonal nodes of G. These are the convolution coefficients, and W is the trainable weight matrix of the k-th layer. It is an activation function.

[0085] The oversmoothing problem is mainly caused by and Caused by multiplication. For convenience... The derivation allows Let W be the identity function and the identity matrix, respectively. Then, the formula X can be rewritten as:

[0086] (7)

[0087] in, It is equivalent to the initial representation matrix of C. Through infinite depth... ∞ The propagation process After smoothing, the final representation of the drug and target can be obtained:

[0088] (8)

[0089] (9)

[0090] in, It is the final adjacency matrix of G, therefore express and Weights between .

[0091] Assumption yes The feature vector is also The i-th row, calculate hour Changes The extent of the impact. This allows us to determine... Value:

[0092] (10)

[0093] in, express The i-th row represents the pairs of other nodes in the k-th layer. The influence distribution is then determined. Then, an NDLS strategy is used to determine the minimum value of k specific to each node, the distance parameter. It is an arbitrarily small constant used to control the smoothing effect.

[0094] The formula for NDLS is expressed as follows:

[0095] (11)

[0096] in It is a norm 2. Once the learning is determined If k is the minimum value, an averaging operation will be applied to aggregate the values ​​from... Sufficient neighborhood information within k hops is obtained to obtain The update rule is expressed by the formula:

[0097] (12)

[0098] Using the above formula, we can obtain the feature vector of each node in G, and thus determine X.

[0099] For matrix Q, each of its elements, for example , representing the mutual influence between nodes i and j. Even when i and j are both drug nodes, it can still be obtained through the formula. The value of can be used to extract distinguishable representation features by determining the optimal propagation depth for each node, and further avoid oversmoothing. Furthermore, features within k hops are aggregated and then averaged to better capture neighborhood information.

[0100] S4. Information from multiple sources at the nodes is aggregated through a two-layer GTN. A fully connected layer is used as the classification module to output predicted values. Based on the predicted values, it is determined whether there is an interaction between the drug and the protein.

[0101] Graph Transformer Network (GTN) introduces the attention mechanism of the Transformer model into graph structure learning, which significantly improves the training speed of the model and can more effectively capture the relevant information between feature nodes and the whole graph.

[0102] Specifically, given feature nodes The multi-head attention calculation for each edge from the distant node j to the source node i is as follows:

[0103] (13)

[0104] (14)

[0105] (15)

[0106] (16)

[0107] in, It is an exponentially scaling dot product function, where d is the hidden layer size of each head. First, using different trainable parameters... Transform the source node features and distant node features into and Then, for edge features Encode it and add it as supplementary information for each layer. The attention of each edge from the distant node j to the source node i is calculated.

[0108] After obtaining the attention of the graph, information is aggregated from the distant node j to the source node i:

[0109] (17)

[0110] (18)

[0111] This invention also provides a model for predicting drug-target interactions, such as... Figure 2 As shown, the model includes:

[0112] The sequence information encoding module generates feature vector sub-graphs of drugs and proteins based on their sequence information encoding.

[0113] The graph attention sampling module applies graph attention sampling to the drug and protein subgraphs respectively to obtain the interaction information between drugs or proteins, and updates the feature vectors of the central nodes of the drug and protein subgraphs.

[0114] The NDLS and Graph SAGE modules employ NDLS for adaptive iterative optimization of drug and protein feature vectors, while simultaneously using Graph SAGE to update the feature vector embedding information of the central nodes and obtain multi-source information of the nodes.

[0115] The GTN fusion prediction module aggregates multi-source information from nodes through a two-layer GTN and uses a fully connected layer as the classification module to output predicted values.

[0116] The DTI prediction model constructed in this invention encodes the sequence information of drugs and proteins using a sequence information encoding module, generating corresponding feature vectors for each. A graph attention sampling module applies attention sampling to the nodes of the drug and protein subgraphs to account for the interaction information between them. The feature vectors of the central nodes of drugs and proteins are updated using a combination of randomly sampled node sets and weights. Custom sampling ratios and weight allocations provide greater flexibility and scalability for experimental research. Then, NDLS and graph SAGE modules are used. The NDLS strategy addresses the over-smoothing or under-smoothing problem in the GNN propagation process, determining the specific propagation depth for all nodes based on the graph structure, and performing adaptive iterative optimization of the drug and protein feature vectors. Simultaneously, graph SAGE is used to randomly sample the first-order neighbor nodes of the central nodes of drugs and proteins, and then randomly selects their first-order neighbors from these neighbors. Finally, starting from the outermost nodes, the feature embedding information of the central nodes is updated layer by layer from the neighbor nodes inwards. Finally, the GTN fusion prediction module is used to fuse multi-source information through a two-layer GTN. A fully connected layer is used as the classification module, and the model loss function is selected as binary cross-entropy. The output is a predicted value between 0 and 1. The presence of an interaction between the drug and protein is determined based on the predicted value.

[0117] The model of this invention can be evaluated using a five-fold cross-validation method, with AUC and AUPR metrics used to assess model performance. The AUC curve is plotted with FPR on the horizontal axis and TPR on the vertical axis; AUC represents the area under the ROC curve. The AUPR curve is plotted with Recall on the horizontal axis and Precision on the vertical axis; AUPR is the area under the PR curve. Where TP represents the number of true positive samples, TN represents the number of true negative samples, FP represents the number of false positive samples, and FN represents the number of false negative samples, expressed by the following formula:

[0118] TCP(Recall) = TP / (TP + FN)

[0119] FPR = FP / (FP+TN)

[0120] Precision = TP / (TP+FP)

[0121] To investigate the impact of the number of attention samples on the experimental results, a parameter range of 100 to 500 samples with a data interval of 50 was selected on the DrugBank dataset for performance evaluation. The experimental results are shown in Table 1.

[0122] Table 1 Performance comparison of attention sampling number

[0123]

[0124] As can be seen from the results in Table 1, before the number of samples is 350, the performance of the model generally shows an upward trend as the number of samples increases. After that, the prediction performance gradually decreases as the number of samples increases. Therefore, the model achieves the best performance when the number of attention samples is 350.

[0125] To verify the effectiveness of the GraphSAGE and NDLS modules for DTI prediction, ablation experiments were conducted on the DrugBank dataset. Specifically, the performance of each new model was evaluated using the module elimination method, and the importance of different modules on model performance was explored by combining different attention sampling numbers. The experimental results are shown in Table 2.

[0126] Table 2 Performance comparison of ablation experiments

[0127]

[0128] As shown in Table 2, the AUC and AUPR values ​​decreased after removing the GraphSAGE or NDLS modules across all ranges of attention sampling number.

[0129] In addition, to more intuitively demonstrate the impact of module ablation, data will also be statistically analyzed in... Figure 3 The results show that, as the figure illustrates, NDLS is the most significant factor affecting model performance, and its removal results in the most pronounced performance decline. Similarly, the removal of the GraphSAGE module also leads to varying degrees of performance degradation depending on the selected number of attention samples. From a model design perspective, the ablation experiments demonstrate the effectiveness of combining NDLS and GraphSAGE modules. Furthermore, the models achieve optimal performance with a sampling number of 350, consistent with the conclusions of the parametric experiments. This is likely due to the optimal feature complementarity between NDLS and GraphSAGE modules at this sampling number.

[0130] To further verify the effectiveness of GTN in feature fusion, feature fusion modules such as GCN, TAG, GAT, and SuperGAT were selected for comparison. The experimental results are shown in Table 3.

[0131] Table 3. Performance comparison of different feature fusion modules in AUC and AUPR

[0132]

[0133] As shown in Table 3, compared with other advanced graph neural network methods, using GTN can effectively integrate complementary information from the advantages of GraphSAGE and NDLS modules, enabling the model to achieve the best performance.

[0134] To further evaluate model performance, six advanced DTI prediction methods were selected: MMDG-DTI, DeepConv-DTI, MCANet, TransformerCPI, Moltrans, and HyperAttentionDTI. Comparative experiments were conducted on the DrugBank and Davis public datasets using five-fold cross-validation. Tables 4 and 5 show the AUC and AUPR values ​​obtained by the model of this invention after five-fold cross-validation with other models.

[0135] Table 4 Comparison with existing methods on the DrugBank dataset

[0136]

[0137] Table 5 Comparison with existing methods on the Davis dataset

[0138]

[0139] As can be seen from the results in Tables 4 and 5, the model of the present invention has the best performance in the DTI prediction task. Compared with the other six models, the model of the present invention is more effective than other state-of-the-art methods in the DTI prediction task.

[0140] This invention proposes a DTI prediction model based on graph attention sampling and multi-neighborhood interactive attention fusion to predict drug-target interactions. For GNN-based DTI prediction models, shallow GNNs with limited receptive fields can only aggregate incomplete information within their neighborhood, exhibiting strong structural bias and noise; deep GNNs suffer from oversmoothing, potentially aggregating a large amount of irrelevant information. This invention employs an NDLS module to determine the specific propagation depth for nodes across the entire graph, performing adaptive iterative optimization of node features. Simultaneously, GraphSAGE is used for deep sampling and aggregation of neighborhood nodes, focusing on the receptive fields of different network node levels. Finally, GTN is used to fuse the complementary advantages of each approach, effectively aggregating more relevant node information from the neighborhood.

[0141] Furthermore, this invention explores the role of different drug sequence encoding feature information on the feature information encoded by a single drug sequence; similarly, protein sequence features also incorporate feature information from other proteins. Parameter analysis and ablation experiments confirm the positive effects of the NDLS and GraphSAGE modules on the model. Simultaneously, the sampling ratio and weight allocation effectively improve model performance, reflecting to some extent that drugs (proteins) can acquire beneficial predictive feature information from other drugs (proteins).

[0142] The present invention also provides an electronic device, Figure 4 This is a schematic diagram of the structure of an electronic device provided in an embodiment of the present invention, such as... Figure 4 As shown, the electronic device may include a processor, a communications interface, memory, and a communication bus, wherein the processor, communications interface, and memory communicate with each other via the communication bus. The processor can invoke logical instructions from the memory, for example, to execute the following method:

[0143] S1. Encode the sequence information of the drug and protein to generate feature vector subgraphs of the drug and protein, respectively;

[0144] S2. Apply graph attention sampling to the feature vector subgraph to obtain interaction information between drugs or proteins, and update the feature vectors of the central nodes of the drug and protein subgraphs.

[0145] S3. Adaptive iterative optimization of drug and protein feature vectors is performed using NDLS; at the same time, graph SAGE is used to update the feature vector embedding information of the central node to obtain multi-source information of the node.

[0146] S4. Information from multiple sources at the nodes is aggregated through a two-layer GTN. A fully connected layer is used as the classification module to output predicted values. Based on the predicted values, it is determined whether there is an interaction between the drug and the protein.

[0147] Furthermore, the logical instructions in the aforementioned memory can be implemented as software functional units and sold or used as independent products, and can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0148] This invention also provides a non-transitory computer-readable storage medium storing a computer program thereon, which, when executed by a processor, is implemented to perform the methods provided in the above embodiments, including, for example:

[0149] S1. Encode the sequence information of the drug and protein to generate feature vector subgraphs of the drug and protein, respectively;

[0150] S2. Apply graph attention sampling to the feature vector subgraph to obtain interaction information between drugs or proteins, and update the feature vectors of the central nodes of the drug and protein subgraphs.

[0151] S3. Adaptive iterative optimization of drug and protein feature vectors is performed using NDLS; at the same time, graph SAGE is used to update the feature vector embedding information of the central node to obtain multi-source information of the node.

[0152] S4. Information from multiple sources at the nodes is aggregated through a two-layer GTN. A fully connected layer is used as the classification module to output predicted values. Based on the predicted values, it is determined whether there is an interaction between the drug and the protein.

[0153] The model embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Those skilled in the art can understand and implement this without any creative effort.

[0154] Through the above description of the embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus necessary general-purpose hardware platforms, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solutions, in essence or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods described in the various embodiments or some parts of the embodiments.

[0155] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for predicting drug-target interactions, characterized in that, include: S1. Encode the sequence information of the drug and protein to generate feature vector subgraphs of the drug and protein, respectively; S2. Apply graph attention sampling to the feature vector subgraph to obtain interaction information between drugs or proteins, and update the feature vectors of the central nodes of the drug and protein subgraphs. S3. Adaptive iterative optimization of drug and protein feature vectors is performed using NDLS; at the same time, graph SAGE is used to update the feature vector embedding information of the central node to obtain multi-source information of the node. S4. Information from multiple sources at the nodes is aggregated through a two-layer GTN. A fully connected layer is used as the classification module to output predicted values. Based on the predicted values, it is determined whether there is an interaction between the drug and the protein.

2. The method according to claim 1, characterized in that, When updating the feature vectors of the center nodes of the drug and protein subgraphs in S2: The feature vectors of drug and protein center nodes are updated by combining randomly sampled sets of nodes with similar attributes and weights.

3. The method according to claim 2, characterized in that, The attention sampling in S2 is expressed by the following formula: （1） （2） V1 is the central node, and V2 is a set of nodes with the same attribute.

4. The method according to claim 1, characterized in that, The S3 step involves updating the feature vector embedding information of the center node using the graph SAGE, including: The first-order neighbor nodes of the drug and protein center nodes are randomly sampled, and then their first-order neighbors are randomly selected from these neighbor nodes. Finally, starting from the outermost node, the feature embedding information of the center node is updated layer by layer from the neighbor nodes inward.

5. The method according to claim 4, characterized in that, In step S3, the feature vector embedding information of the central node is updated using the graph SAGE, as expressed by the formula: （3） （4） Where v is a node, For the set of neighboring nodes, This is an update embedding for the SAGE model layer.

6. The method according to claim 1, characterized in that, In S3, NDLS is used for adaptive iterative optimization of drug and protein feature vectors, expressed by the following formula: （11） （12） in, It is a node eigenvectors, It is a norm 2. express The i-th row represents the pairs of other nodes in the k-th layer. The distribution of influence This is the distance parameter.

7. The method according to claim 1, characterized in that, When performing information aggregation on multi-source information of nodes through a two-layer GTN in S4, it includes: Multi-head attention is calculated for each edge from the distant node j to the source node i, as expressed by the formula: （13） （14） （15） （16） After obtaining multi-head attention in the graph, information is aggregated from the distant node j to the source node i, as expressed by the formula: （17） （18） in, Given a feature node, For trainable parameters, Let be the exponentially scaling dot product function, and d be the size of the hidden layers in each head. Features of the source node Features of distant nodes.

8. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the steps of the method for predicting drug-target interactions as described in any one of claims 1 to 7.

9. A non-transitory computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the steps of the method for predicting drug-target interactions as described in any one of claims 1 to 7.

Images