Inductive traditional Chinese medicine target relation discovery method and system based on multi-scale HGT

Abstract

The invention provides an inductive traditional Chinese medicine target relation discovery method and system based on multi-scale HGT, and the method comprises the steps: constructing a heterogeneous information graph based on the original data of traditional Chinese medicines, medicine molecules, targets and diseases, and carrying out the node initialization operation of the heterogeneous information graph; the node initialization features are input into topological robust local attention for local representation; the node initialization features are input into global attention for global node representation; inputting the local representation of the node and the global representation of the node into a feature fusion module for splicing operation and feature fusion in sequence; and inputting the final feature representation into a prediction model for prediction to obtain a prediction result. According to the method, rich semantics and multiple relationships among traditional Chinese medicines, targets, molecules and diseases extracted from a traditional Chinese medicine network pharmacology database are comprehensively represented by constructing a new traditional Chinese medicine-target heterograph.

Description

Technical Field

[0001] The present invention relates to the field of traditional Chinese medicine-target relationship prediction, and particularly to an inductive traditional Chinese medicine target relationship discovery method and system based on multi-scale HGT. Background Art

[0002] Predicting traditional Chinese medicine target interactions (HTIs) plays a key role in revealing the pharmacological mechanisms of traditional Chinese medicine (TCM) and accelerating new drug development. Although deep learning methods have shown great potential in this field, existing methods still face some challenges, including difficulties in modeling the complex relationships between traditional Chinese medicine entities, the problem of hidden biases, and poor generalization ability for unknown data. These challenges hinder the model's ability to effectively capture the complex interactions between biomedical entities and lead to poor performance in predicting newly discovered elements such as molecular or protein targets. Summary of the Invention

[0003] In view of the above situation, the main object of the present invention is to propose an inductive traditional Chinese medicine target relationship discovery method and system based on multi-scale HGT to solve the above technical problems.

[0004] The present invention proposes an inductive traditional Chinese medicine target relationship discovery method based on multi-scale HGT, and the method includes the following steps: Step 1: Construct local attention based on a topological robust local attention mechanism, construct global attention based on a multi-relation global cross-attention mechanism, construct a feature fusion module based on a feature fusion mechanism, and the local attention, global attention, feature fusion module, and prediction module constitute a prediction model; Among them, the global attention includes a feed-forward network, and the prediction module includes a linear layer; Step 2: Construct a heterogeneous information graph based on the original data of traditional Chinese medicine, drug molecules, targets, and diseases, and perform node initialization operations on the heterogeneous information graph to obtain node initialization features; Step 3: Input the node initialization features into the topological robust local attention for local representation to obtain the local representation of the nodes; Step 4: Input the node initialization features into the global attention for global node representation to obtain the global representation of the nodes; Step 5: Input the local representation of the nodes and the global representation of the nodes into the feature fusion module for concatenation operations and feature fusion processing in sequence to obtain the final feature representation; Step 6: Input the final feature representation into the linear layer for prediction to obtain a predicted value; Construct a binary cross-entropy loss based on the predicted value, and use the binary cross-entropy loss to optimize the prediction model to obtain an optimized prediction model; Use the optimized prediction model to obtain the final prediction result.

[0005] The present invention also provides an inductive traditional Chinese medicine target relationship discovery system based on multi-scale HGT, and the system includes: A construction module, configured to: Construct local attention based on a topological robust local attention mechanism, construct global attention based on a multi-relation global cross-attention mechanism, construct a feature fusion module based on a feature fusion mechanism, and the local attention, global attention, feature fusion module and prediction module constitute a prediction model; Among them, the global attention includes a feed-forward network, and the prediction module includes a linear layer; A data processing module, configured to: Construct a heterogeneous information graph based on the original data of traditional Chinese medicine, drug molecules, targets, and diseases, and perform node initialization operations on the heterogeneous information graph to obtain node initialization features; A feature extraction module, configured to: Input the node initialization features into the topological robust local attention for local representation to obtain the local representation of the nodes; Input the node initialization features into the global attention for global node representation to obtain the global representation of the nodes; A feature processing module, configured to: Input the local representation of the nodes and the global representation of the nodes into the feature fusion module to perform concatenation operations and feature fusion processing in sequence to obtain the final feature representation; A relationship prediction module, configured to: Input the final feature representation into the linear layer for prediction to obtain a predicted value; Construct a binary cross-entropy loss based on the predicted value, and use the binary cross-entropy loss to optimize the prediction model to obtain an optimized prediction model; Use the optimized prediction model to obtain the final prediction result.

[0006] Compared with the prior art, the beneficial effects of the present invention are as follows: 1. The present invention comprehensively characterizes the rich semantics and multiple relationships among traditional Chinese medicine, targets, molecules, and diseases extracted from the traditional Chinese medicine network pharmacology database by constructing a new traditional Chinese medicine-target heterogeneous graph; 2. The present invention processes diverse information on the traditional Chinese medicine-target heterogeneous graph through a graph neural network model, helps the model to more fully understand the meaning conveyed on the traditional Chinese medicine-target heterogeneous graph, and can more accurately obtain key information from the complex traditional Chinese medicine relationship graph; 3. The present invention can enhance feature extraction by proposing a heterogeneous graph transformer based on multi-scale relationships, which can capture different aspects of entities on the traditional Chinese medicine-target heterogeneous graph during the learning process. This mechanism effectively captures long-range node dependencies, reduces the impact of hidden biases, and promotes efficient prediction of traditional Chinese medicine targets in inductive learning scenarios; 4. The present invention fuses the local features and global features of nodes by means of a multi-level attention mechanism and a fusion mechanism, providing an effective feature representation for the generation of traditional Chinese medicine-target link prediction in the final relationship prediction layer.

[0007] Additional aspects and advantages of the present invention will be given in part in the following description, will become apparent in part from the following description, or will be understood through the embodiments of the present invention. Brief Description of the Drawings

[0008] Figure 1 It is a flowchart of the steps of an inductive traditional Chinese medicine target relationship discovery method based on multi-scale HGT proposed by the present invention.

[0009] Figure 2 It is a process diagram of obtaining node features based on local attention in an inductive traditional Chinese medicine target relationship discovery method based on multi-scale HGT proposed by the present invention.

[0010] Figure 3 It is a process diagram of obtaining node features based on global attention in an inductive traditional Chinese medicine target relationship discovery method based on multi-scale HGT proposed by the present invention.

[0011] Figure 4 It is a process diagram of feature fusion and relationship prediction in an inductive traditional Chinese medicine target relationship discovery method based on multi-scale HGT proposed by the present invention.

[0012] Figure 5 It is a schematic structural diagram of an inductive traditional Chinese medicine target relationship discovery system based on multi-scale HGT proposed by the present invention. Detailed Embodiments

[0013] The embodiments of the present invention are described in detail below. The examples of the embodiments are shown in the drawings, where the same or similar reference numerals are the same or similar elements or elements with the same or similar functions from beginning to end. The embodiments described below by referring to the drawings are exemplary and are only used to explain the present invention and should not be construed as a limitation to the present invention.

[0014] Referring to the following description and drawings, these and other aspects of the embodiments of the present invention will become clear. In these descriptions and drawings, some specific embodiments of the embodiments of the present invention are specifically disclosed as some ways to implement the principles of the embodiments of the present invention, but it should be understood that the scope of the embodiments of the present invention is not limited by this.

[0015] Please refer to Figure 1 , the embodiment of the present invention proposes an inductive traditional Chinese medicine target relationship discovery method based on multi-scale HGT. The method includes the following steps: Step 1: Construct local attention based on the topological robust local attention mechanism, construct global attention based on the multi-relation global cross-attention mechanism, and construct a feature fusion module based on the feature fusion mechanism. The local attention, global attention, and feature fusion module constitute a prediction model; Among them, the global attention includes a feed-forward network, and the prediction module includes a linear layer.

[0016] Step 2: Construct a heterogeneous information graph based on the original data of traditional Chinese medicine, drug molecules, targets, and diseases, and perform node initialization operations on the heterogeneous information graph to obtain node initialization features.

[0017] Step 3: Input the node initialization features into the topological robust local attention for local representation to obtain the local representation of the nodes.

[0018] Please refer to Figure 2 , in Step 3, input the node initialization features into the topological robust local attention, and perform local representation on the heterogeneous information graph to obtain the local representation of the nodes. The specific steps are as follows: Input the node initialization features into the topological robust local attention and perform a projection operation using the local weight matrix to obtain the features of the local attention nodes after projection; Use an edge to connect two adjacent nodes as a node pair, project the node pair to obtain the node representation of the node pair; Perform attention calculation on the node representation of the node pair to obtain the attention value between the node pairs; Use the attention value between the node pairs for aggregation operation to obtain aggregated heterogeneous neighbor information, and perform local representation on the aggregated heterogeneous neighbor information through the node representation of the node pair and in combination with the attention value between the node pairs to obtain the local representation of the nodes.

[0019] Input the node initialization features into the topological robust local attention and perform a projection operation using the local weight matrix to obtain the features of the local attention nodes after projection. The relational formula existing in the corresponding process is as follows: ; Among them, represents the features of node after projection in the local attention, represents the local weight matrix of node type , represents the initialization features of the nodes; In the step of connecting two adjacent nodes by an edge as a node pair, projecting the node pair, and obtaining the node representation of the node pair, the relational expressions existing in the corresponding process are as follows: ; Among them, represents the key representation of node . represents the query representation of node . represents the value representation of node . represents the parameter matrix for projecting the input of node into the key space. represents the parameter matrix for projecting the input of node into the query space. represents the parameter matrix for projecting the input of node into the value space. represents an edge, represents a key vector, represents a query vector, represents a value vector, represents a linear layer, represents a function for obtaining the relationship of edge ; In the step of calculating the attention of the node representation of the node pair to obtain the attention value between the node pairs, the relational expressions existing in the corresponding process are as follows: ; Among them, represents the attention value of node to node , represents a non-linear activation function, represents a row weight vector of a specific relationship, represents a weight matrix of a specific relationship, represents a concatenation operation; It should be noted that, maps the input value to the range (0, 1).

[0020] In the step of using the attention value between the node pairs for an aggregation operation to obtain aggregated heterogeneous neighbor information, and performing a local representation on the aggregated heterogeneous neighbor information by combining the node representation of the node pair with the attention value between the node pairs to obtain the local representation of the node, the relational expressions existing in the corresponding process are as follows: ; Among them, Represents a node The local representation of Represents the local output weight matrix related to the node type Related local output weight matrix Represents a linear layer Represents a node Through the relationship Set of neighbors Represents the set of relationship types

[0021] It should be noted that the relationship type refers to different types of edges between nodes

[0022] Furthermore, on the constructed heterogeneous information graph, it aims to capture the semantic and topological relationships between different node types. To achieve this goal, a simple and effective Topological Robust Local Attention (TRLA) is used to aggregate local information from source nodes to target nodes to capture fine-grained structures

[0023] Step 4: Input the initial node features into the global attention for global node representation to obtain the global representation of the node

[0024] Please refer to Figure 3 In step 4, input the initial node features into the global attention for global node representation to obtain the global representation of the node, which specifically includes the following steps Input the initial node features into the global attention for projection operation to obtain the features of the node after projection in the global attention Perform node partitioning operation on the features of the node after projection in the global attention to obtain the target node set and the source node set Use the target node set to construct the target node embedding sequence to obtain the constructed target node embedding sequence Use the source node set to construct the source node embedding sequence to obtain the constructed source node embedding sequence Perform self-attention mechanism enhancement processing and feed-forward network processing on the target node embedding sequence and the source node embedding sequence in sequence to obtain the enhanced target node representation and the enhanced source node representation respectively Perform cross-attention mechanism processing on the enhanced target node representation and the enhanced source node representation to obtain the global representation of the node

[0025] Input the initial node features into the global attention for projection operation to obtain the features of the node after projection in the global attention. The relational expressions existing in the corresponding process are as follows ; Among them Represents the features of the node after projection in the global attention Represents the node type The global weight matrix; In the step of performing a node partitioning operation on the features after node projection in global attention to obtain a target node set and a source node set, the relational expressions existing in the corresponding process are as follows: ; Among them, represents the target node set, represents all node sets in the heterogeneous information graph, represents a single node, is used to obtain the category of the node, represents the source node set, is used to obtain the node and the node the relationship category between them; In the step of constructing a target node embedding sequence by using the target node set to obtain the constructed target node embedding sequence, the relational expressions existing in the corresponding process are as follows: ; Among them, represents the embedding matrix of the target node, represents the first embedding of the target node, represents the second embedding of the target node, represents the th embedding of the target node, represents the transpose; In the step of constructing a source node embedding sequence by using the source node set to obtain the constructed source node embedding sequence, the relational expressions existing in the corresponding process are as follows: ; Among them, represents the embedding matrix of the source node, represents the th embedding of the source node.

[0026] Performing self-attention mechanism enhancement processing and feed-forward network processing on the target node embedding sequence and the source node embedding sequence in sequence to obtain an enhanced target node representation and an enhanced source node representation respectively, the relational expressions existing in the corresponding process are as follows: ; Among them, represents the representation of the target node after self-attention enhancement, represents the representation of the source node after self-attention enhancement, represents the self-attention mechanism of the node, represents the feed-forward network; In the step of performing cross-attention mechanism processing on the enhanced target node representation and the enhanced source node representation to obtain the global representation of the target node, the relational expressions existing in the corresponding process are as follows: ; Among them, represents the global representation matrix of the target node, represents the query matrix, represents the first weight matrix, represents the key matrix, represents the second weight matrix, represents the value matrix, represents the third weight matrix, represents the normalization function, represents the square root of the dimension of the embedding vector, represents the key matrix transpose of , represents the global representation of node , represents mapping the identifier of node to the specific row number in the matrix.

[0027] It should be noted that by jointly utilizing self-attention and cross-attention in the global graph, the global representations of all node types can be calculated.

[0028] Furthermore, on the constructed heterogeneous information graph, newly arrived nodes are usually isolated, which makes it difficult to learn comprehensive node representations through message passing strategies. For this reason, a new multi-relational global attention (MRGA) is proposed, which can automatically capture long-range dependencies in the entire heterogeneous information graph.

[0029] Step 5: Input the local representation and the global representation of the target node into the feature fusion module to perform concatenation operation and feature fusion processing in sequence to obtain the final feature representation.

[0030] Please refer to Figure 4 , in Step 5, input the local representation and the global representation of the node into the feature fusion module to perform concatenation operation and feature fusion processing in sequence to obtain the final feature representation, and the relational expressions existing in the corresponding process are as follows: ; Among them, represents the concatenation of the local representation of node and the global representation of node , and then perform fusion calculation through a specific type of transformation matrix to obtain node The final feature representation.

[0031] Step 6: Input the final feature representation into the linear layer for prediction to obtain the predicted value; Construct a binary cross-entropy loss based on the predicted value, and use the binary cross-entropy loss to optimize the prediction model to obtain the optimized prediction model; Use the optimized prediction model to obtain the final prediction result.

[0032] In the said Step 6, when constructing the binary cross-entropy loss based on the predicted value, the existing relational expressions in the corresponding process are as follows: ; Among them, represents the binary cross-entropy loss, represents the true label, represents the logarithmic function, represents the generated predicted interaction probability.

[0033] Furthermore, first take the traditional Chinese medicine node as the target node and process it through local attention and global attention in turn, and the local feature of traditional Chinese medicine and the global feature of traditional Chinese medicine can be obtained. Then, fuse the obtained local feature of traditional Chinese medicine and the global feature of traditional Chinese medicine in the fusion layer of the feature fusion module to obtain the final representation of the traditional Chinese medicine node; in the same way, take the target node as the target node and go through the same process, and respectively obtain the local feature of the target and the global feature of the target. Then, fuse the local feature of the target and the global feature of the target into the final traditional Chinese medicine feature and perform feature fusion in the fusion layer to obtain the final representation of the target node. Finally, input the final representation of the traditional Chinese medicine node and the final representation of the target node into the prediction module including three fully connected layers for prediction.

[0034] Please refer to Figure 5 The present invention also proposes an inductive traditional Chinese medicine target relationship discovery system based on multi-scale HGT. The system includes: A construction module, used for: Construct local attention based on the topological robust local attention mechanism, construct global attention based on the multi-relation global cross-attention mechanism, construct a feature fusion module based on the feature fusion mechanism. The local attention, global attention, feature fusion module and prediction module constitute the prediction model; Among them, the global attention includes a feed-forward network, and the prediction module includes a linear layer; A data processing module, used for: Construct a heterogeneous information graph based on the original data of traditional Chinese medicine, drug molecules, targets, and diseases, and perform node initialization operations on the heterogeneous information graph to obtain node initialization features; A feature extraction module, used for: Input the initial features of the nodes into the topologically robust local attention for local representation to obtain the local representations of the nodes; Input the initial features of the nodes into the global attention for global node representation to obtain the global representations of the nodes; Feature processing module, configured to: Input the local representations of the nodes and the global representations of the nodes into the feature fusion module to perform concatenation operations and feature fusion processing in sequence to obtain the final feature representation; Relationship prediction module, configured to: Input the final feature representation into a linear layer for prediction to obtain a predicted value; Construct a binary cross-entropy loss based on the predicted value, and use the binary cross-entropy loss to optimize the prediction model to obtain an optimized prediction model; Use the optimized prediction model to obtain the final prediction result.

[0035] It should be understood that each part of the present invention can be implemented by hardware, software, firmware, or a combination thereof. In the above embodiments, multiple steps or methods can be implemented by software or firmware stored in a memory and executed by a suitable instruction execution system. For example, if implemented by hardware, as in another embodiment, any one or a combination of the following techniques well known in the art can be used: discrete logic circuits having logic gate circuits for implementing logic functions on data signals, application-specific integrated circuits having appropriate combinational logic gate circuits, programmable gate arrays (PGAs), field programmable gate arrays (FPGAs), etc.

[0036] In the description of this specification, the description with reference to terms such as "one embodiment", "some embodiments", "example", "specific example", or "some examples" means that the specific features, structures, materials, or characteristics described in connection with the embodiment or example are included in at least one embodiment or example of the present invention. In this specification, the schematic representations of the above terms do not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials, or characteristics described can be combined in a suitable manner in any one or more embodiments or examples.

[0037] The above-described embodiments merely represent several implementation manners of the present invention, and their descriptions are relatively specific and detailed, but should not be construed as limiting the scope of the patent of the present invention. It should be noted that for those of ordinary skill in the art, without departing from the concept of the present invention, several modifications and improvements can still be made, and these all belong to the protection scope of the present invention. Therefore, the protection scope of the patent of the present invention should be subject to the appended claims.

Claims

1. An inductive TCM target relationship discovery method based on multi-scale HGT, characterized in that: The method comprises the following steps: Step 1: construct local attention based on the topological robust local attention mechanism, construct global attention based on the multi-relation global cross attention mechanism, and construct a feature fusion module based on the feature fusion mechanism. The local attention, global attention, feature fusion module and prediction module constitute a prediction model; Among them, the global attention includes a feed-forward network, and the prediction module includes a linear layer; Step 2: construct a heterogeneous information graph based on the original data of traditional Chinese medicine, drug molecules, targets, and diseases, perform node initialization operations on the heterogeneous information graph, and obtain node initialization features; Step 3: Input the node initialization features into the topological robust local attention for local representation to obtain the local representation of the node; Step 4: Input the node initialization features into the global attention to perform global node representation and obtain the global representation of the node; Step 5: Input the local representation of the node and the global representation of the node into the feature fusion module for concatenation and feature fusion processing in sequence to obtain the final feature representation; Step 6: Input the final feature representation into the linear layer for prediction to obtain the predicted value; Based on the predicted value, a binary cross entropy loss is constructed, and the prediction model is optimized using the binary cross entropy loss to obtain an optimized prediction model; The optimized prediction model is used to obtain the final prediction results.

2. The inductive TCM target relationship discovery method based on multi-scale HGT according to claim 1, characterized in that: In step 3, the node initialization feature is input into the topological robust local attention for local representation to obtain the local representation of the node, which specifically includes the following steps: Input the node initialization features into the topological robust local attention, use the local weight matrix to perform the projection operation, and obtain the projected features of the local attention node; Connect two adjacent nodes with an edge as a node pair, project the node pair, and obtain the node representation of the node pair; Perform attention calculation on the node representation of the node pair to obtain the attention value between the node pairs; Aggregation operation is performed using the attention values ​​between node pairs to obtain aggregated heterogeneous neighbor information, and the aggregated heterogeneous neighbor information is locally represented by the node representation of the node pairs and combined with the attention values ​​between the node pairs to obtain the local representation of the node.

3. The inductive Chinese medicine target relationship discovery method based on multi-scale HGT according to claim 2, characterized in that: The node initialization features are input into the topological robust local attention, and the projection operation is performed using the local weight matrix to obtain the features after the local attention node projection. The corresponding relationship in the process is as follows: ； in, Represents the local attention center node The projected features are Indicates the node type The local weight matrix, Represents the initialization characteristics of the node; In the step of connecting two adjacent nodes with an edge as a node pair, projecting the node pair, and obtaining the node representation of the node pair, the corresponding process has the following relationship: ； in, Representation Node The key indicates that Representation Node The query indicates that Representation Node The value of indicates that Indicates the node Input The parameter matrix projected into the key space, Indicates the node Input The parameter matrix projected into the query space, Indicates the node Input The parameter matrix projected into the value space, represents the edge, represents the key vector, represents the query vector, represents a value vector, represents a linear layer, Get edge The function of the relationship between In the step of calculating the attention value between the node pairs by the node representation of the node pair, the relationship between the corresponding process is as follows: ； in, Representation Node For Node The attention value, represents a nonlinear activation function, A vector of row weights representing a particular relationship, represents the weight matrix of a specific relationship, Represents a splicing operation; In the step of performing an aggregation operation using the attention values ​​between node pairs to obtain aggregated heterogeneous neighbor information, and locally representing the aggregated heterogeneous neighbor information through node representation of node pairs and combining the attention values ​​between node pairs to obtain local representation of nodes, the corresponding process has the following relationship: ； in, Representation Node The local representation of Representation and Node Type The associated local output weight matrix, represents a linear layer, Representation Node Through relationships Neighbor Set, Represents a collection of relationship types.

4. The inductive Chinese medicine target relationship discovery method based on multi-scale HGT according to claim 3, characterized in that: In step 4, the node initialization feature is input into the global attention to perform global node representation, and a global representation of the node is obtained, which specifically includes the following steps: Input the node initialization features into the global attention for projection operation to obtain the features of the node after projection in the global attention; Perform node partitioning on the features after node projection in the global attention to obtain the target node set and the source node set; Using the target node set to construct a target node embedding sequence to obtain a constructed target node embedding sequence; Constructing a source node embedding sequence using the source node set to obtain a constructed source node embedding sequence; The target node embedding sequence and the source node embedding sequence are sequentially subjected to self-attention mechanism enhancement processing and feedforward network processing to obtain enhanced target node representation and enhanced source node representation respectively; The enhanced target node representation and the enhanced source node representation are processed with a cross-attention mechanism to obtain a global representation of the node.

5. The inductive Chinese medicine target relationship discovery method based on multi-scale HGT according to claim 4, characterized in that: Input the node initialization features into the global attention for projection operation to obtain the features of the node after projection in the global attention. The relationship between the corresponding process is as follows: ； in, represents the feature after node projection in global attention, Indicates the node type The global weight matrix of In the step of performing node partitioning operation on the features after node projection in the global attention to obtain the target node set and the source node set, the relationship between the corresponding process is as follows: ； in, represents the target node set, Represents the set of all nodes in the heterogeneous information graph, Represents a single node, Indicates the category used to obtain the node. represents the source node set, Indicates the node to be obtained With Node the types of relationships between them; In the step of constructing a target node embedding sequence using a target node set to obtain a constructed target node embedding sequence, the relationship between the corresponding processes is as follows: ； in, represents the embedding matrix of the target node, represents the first embedding of the target node, represents the second embedding of the target node, The target node embedding, represents transpose; In the step of constructing a source node embedding sequence using a source node set to obtain a constructed source node embedding sequence, the relationship between the corresponding processes is as follows: ； in, represents the embedding matrix of the source node, The source node embed.

6. The inductive Chinese medicine target relationship discovery method based on multi-scale HGT according to claim 5, characterized in that: The target node embedding sequence and the source node embedding sequence are sequentially subjected to self-attention mechanism enhancement processing and feedforward network processing to obtain enhanced target node representation and enhanced source node representation, respectively. The relationship between the corresponding processes is as follows: ； in, represents the self-attention enhanced representation of the target node, represents the self-attention enhanced representation of the source node, represents the self-attention mechanism of the node, represents a feed-forward network; In the step of performing cross-attention processing on the enhanced target node representation and the enhanced source node representation to obtain the global representation of the node, the relationship between the corresponding process is as follows: ； in, represents the global representation matrix of the target node, represents the query matrix, represents the first weight matrix, represents the key matrix, represents the second weight matrix, represents the value matrix, represents the third weight matrix, represents the normalization function, represents the square root of the dimension of the embedding vector, Represents the key matrix Transpose , Representation Node The global representation of Indicates that the node The identifiers are mapped to specific row numbers in the matrix.

7. The inductive TCM target relationship discovery method based on multi-scale HGT according to claim 6, characterized in that: In step 5, the local representation of the node and the global representation of the node are input into the feature fusion module for concatenation and feature fusion processing in sequence to obtain the final feature representation. The relationship between the corresponding process is as follows: ； in, Represented by the node Local representation of and nodes The global representation of After splicing, it is transformed by a specific type of matrix Fusion Compute Node The final feature representation of .

8. The inductive TCM target relationship discovery method based on multi-scale HGT according to claim 7, characterized in that: In step 6, a binary cross entropy loss is constructed based on the predicted value, and the relationship between the corresponding process is as follows: ； in, represents the binary cross entropy loss, represents the true label, represents the logarithmic function, represents the generated predicted interaction probability.

9. An inductive TCM target relationship discovery system based on multi-scale HGT, characterized in that: The system uses the inductive TCM target relationship discovery method based on multi-scale HGT as described in any one of claims 1 to 8 above, and the system comprises: Building blocks for: Local attention is constructed based on the topological robust local attention mechanism, global attention is constructed based on the multi-relation global cross attention mechanism, and feature fusion module is constructed based on the feature fusion mechanism. Local attention, global attention, feature fusion module and prediction module constitute the prediction model; Among them, the global attention includes a feed-forward network, and the prediction module includes a linear layer; Data processing module for: Based on the original data of traditional Chinese medicine, drug molecules, targets, and diseases, a heterogeneous information graph is constructed, and node initialization operations are performed on the heterogeneous information graph to obtain node initialization features; Feature extraction module for: Input the node initialization features into the topological robust local attention for local representation to obtain the local representation of the node; Input the node initialization features into the global attention to perform global node representation and obtain the global representation of the node; Feature processing module for: The local representation of the node and the global representation of the node are input into the feature fusion module for concatenation and feature fusion processing in sequence to obtain the final feature representation; Relationship prediction module, used to: The final feature representation is input into the linear layer for prediction to obtain the predicted value; Based on the predicted value, a binary cross entropy loss is constructed, and the prediction model is optimized using the binary cross entropy loss to obtain an optimized prediction model; The optimized prediction model is used to obtain the final prediction results.

Images