A drug-target relationship prediction method based on comparative knowledge enhancement

Abstract

The present invention belongs to the field of drug research technology, and specifically relates to a drug-target relationship prediction method based on contrast knowledge enhancement, comprising the following steps: aggregating a knowledge graph with a bipartite graph and then aggregating it with attention-aware node features to obtain a main view; destroying the relationship between nodes and entities in the original knowledge graph, and then restoring the original relationship from Gaussian noise, and reconstructing task-related subgraphs using the restored relationship to obtain a reconstructed knowledge graph; aggregating entities of the reconstructed knowledge graph into nodes to generate a knowledge-aware contrast view; performing data enhancement on nodes of the knowledge-aware contrast view using an attention-aware mechanism to obtain a BGCL-KDE model; updating the BGCL-KDE model parameters, and using the updated BGCL-KDE model to predict drug-target relationships. The method provided by the present invention can ensure that the generated knowledge graph has a positive impact on the node knowledge that is highly relevant to the drug discovery task.

Description

Technical Field

[0001] The present invention belongs to the technical field of drug research, and in particular relates to a drug-target relationship prediction method based on comparative knowledge enhancement. Background Art

[0002] At present, the research on bipartite graph learning in the field of drug discovery mainly focuses on using graph neural networks (GNNs) to encode drug molecules and targets into low-dimensional vector representations. However, GNN-based methods usually use the Laplacian matrix to average all nodes and fail to fully utilize the attention perception matrix to distinguish the importance difference between drug molecules and targets, resulting in inaccurate node embedding and affecting the performance of drug discovery models. Although the attention weight mechanism based on adaptive training can learn the interaction between drug molecules and targets, it has high computational cost and is not intuitive enough.

[0003] To address this challenge, researchers have integrated knowledge graphs (KGs) into comprehensive information networks of drug molecules and targets, enhanced the semantic information of nodes with different types of entities, and constructed knowledge-aware graph learning. DKN integrates KG embedding into nodes through deep knowledge-aware networks, and Cao et al. jointly learn recommendation and knowledge graph completion models. In addition, some path-based models extract semantically meaningful meta-paths from KGs to enhance the semantic information of nodes. KPRN generates path representations through knowledge-aware path recursive networks to enhance node semantic information. Recent studies have focused on using GNNs to capture high-order relations in KGs. KGAT learns high-order connections in KGs to enhance the information of nodes in recommendation systems. SMILES enhances node representation by leveraging multimodal knowledge graphs, while CSRec enhances node semantic information by using common sense as a supplementary KG. For efficient drug discovery, KGE_NFM incorporates KGs to learn representations of various entities in KGs and integrates multimodal information through neural factorization machines, while KGE-UNIT employs KGs and multi-task learning to enhance node information to simultaneously predict drug-target interactions. Recently, KG-diverse encodes node embeddings in KG to learn diversity-aware representations and enhance node semantic information. However, practical knowledge graphs (KGs) are often affected by noise and sparsity, resulting in connections between nodes (drug molecules and targets) and entities (non-nodes) that are irrelevant to drug discovery tasks, reducing the prediction performance of the model.

[0004] Recently, some researchers have adopted contrast learning (CL) to enhance knowledge-aware graph learning. KDAR learns bipartite graphs through an alignment contrast mechanism to align attributes in knowledge graphs, and MKGCL adopts hybrid curvature knowledge-enhanced graph contrast learning to learn bipartite graphs. Although contrastive learning (CL) can improve knowledge-aware graph learning and deal with data sparsity problems, current methods often ignore a lot of irrelevant information.

[0005] In summary, most existing methods ignore a large amount of irrelevant information in KG, which may have a harmful effect on specific downstream tasks. In recent years, diffusion models have achieved outstanding results in robust graph data generation. DiffKG generates a robust knowledge graph through a diffusion model to enhance the semantic information of nodes. However, most existing methods cannot guarantee that the generated KG can always bring benefits to the node knowledge that is strongly related to the task. Summary of the invention

[0006] In order to solve the above problems, the present invention proposes a drug-target relationship prediction method based on comparative knowledge enhancement, which can ensure that the generated knowledge graph has a positive impact on the node knowledge that is highly relevant to the drug discovery task.

[0007] The specific technical solutions provided by the present invention are as follows:

[0008] The present invention provides a drug-target relationship prediction method based on comparative knowledge enhancement, comprising the following steps:

[0009] Aggregate different types of entities in the drug knowledge graph into nodes in the bipartite graph to generate a knowledge-aware bipartite graph; use an attention-aware mechanism to aggregate the nodes in the knowledge-aware bipartite graph with attention-aware and knowledge-aware node features to obtain a main view;

[0010] The diffusion model is used to add noise in the diffusion stage to destroy the relationship between nodes and entities in the original drug knowledge graph. In the reverse process, the original relationship is restored from pure Gaussian noise, the true relationship between nodes and entities is iteratively identified, and the restored relationship is used to reconstruct the task-related subgraph.

[0011] The structure of the original drug knowledge graph is modified by reconstructing task-related subgraphs to reconstruct a knowledge graph that is highly relevant to the drug discovery task. Different types of entities in the reconstructed knowledge graph are aggregated into the nodes of the original bipartite graph to generate a knowledge-aware comparative view. The attention-aware mechanism is used to aggregate the nodes in the knowledge-aware comparative view with attention-aware and knowledge-aware node features to obtain the BGCL-KDE model.

[0012] The feature representations of the main view and the BGCL-KDE model are compared to obtain a contrast loss function. The optimized contrast loss function is used to update the parameters of the BGCL-KDE model, and the updated BGCL-KDE model is used to predict the drug-target relationship.

[0013] As a preferred embodiment of the present invention, the aggregation of different types of entities and nodes in the bipartite graph in the drug knowledge graph and the reconstructed knowledge graph is performed using a relationship-aware knowledge embedding layer.

[0014] As a preferred embodiment of the present invention, the method of attention-aware and knowledge-aware node feature aggregation is:

[0015] The nodes in a knowledge-aware bipartite graph or a knowledge-aware contrast view are embedded into a d-dimensional vector, and a graph embedding propagation layer is constructed to remove the nonlinear activation function and the feature transformation matrix. Then, multiple graph embedding propagation layers are used to aggregate local neighbor information to achieve attention-aware and knowledge-aware node feature aggregation.

[0016] As a preferred embodiment of the present invention, different types of entities in the drug knowledge graph are drug molecules, targets and diseases.

[0017] As a preferred embodiment of the present invention, the nodes in the bipartite graph are drug molecules and targets.

[0018] As a preferred embodiment of the present invention, a method for modifying the structure of a drug knowledge graph using a reconstructed task-related subgraph is:

[0019] The highly task-related relationships between nodes and entities in the reconstructed task-related subgraph are selected to adjust the structure of the knowledge graph to maintain the information structure of the reconstructed knowledge graph.

[0020] As a preferred embodiment of the present invention, the contrast loss function is obtained by mapping the main view and the BGCL-KDE model to the feature space and comparing the feature representations of the main view and the BGCL-KDE model.

[0021] As a preferred embodiment of the present invention, optimizing the contrast loss function is performed according to the following steps: selecting positive samples and negative samples from a test set, comparing the difference between the predicted scores of the positive sample pairs and the predicted scores of the negative sample pairs, obtaining a main task loss function, and optimizing the contrast loss function using the main task loss function to obtain a total loss function; wherein the positive sample is the representation of the same node in the main view and the BGCL-KDE model, and the negative sample is the difference between the main view representation of the node and the full-graph embedding of the BGCL-KDE model;

[0022] The method for updating the parameters of the BGCL-KDE model is to use the obtained total loss function and calculate the gradient of the parameters through back propagation to update the parameters of the model.

[0023] As a preferred embodiment of the present invention, the prediction of the drug-target relationship using the updated BGCL-KDE model is carried out according to the following steps: the data of the drug and the target are input into the updated BGCL-KDE model, the drug embedding vector and the target embedding vector are obtained according to the generation order of the comparison view, and the prediction result of the action relationship between the drug and the target is obtained by multiplying the drug embedding vector by the transpose of the target embedding vector.

[0024] As a preferred embodiment of the present invention, a graph embedding layer with an attention perception mechanism is used to obtain a drug embedding vector and a target embedding vector.

[0025] The present invention combines graph neural networks, knowledge graphs, contrastive learning and diffusion models to effectively improve the performance and robustness of bipartite graph learning in the field of drug discovery. By introducing the attention perception matrix and knowledge perception contrast view, it can better distinguish the importance differences between different drug molecules or targets, generate more accurate feature representations, and improve the interpretability and predictive performance of drug discovery models.

[0026] The present invention proposes a drug-target relationship prediction method based on comparative knowledge enhancement, which is generally applied to various network prediction tasks, including drug discovery, information retrieval and recommendation. The technical solution provided by the present invention mainly includes the following points:

[0027] (1) This paper proposes an effective bipartite graph learning contrastive learning architecture BGCL-KDE, which introduces an attention-aware mechanism to learn the importance between nodes and achieve node feature aggregation. This architecture can generate node features with attention awareness and knowledge awareness, make full use of the beneficial information in the knowledge graph, avoid harmful effects on drug discovery tasks, and effectively learn the relationship between drugs and targets.

[0028] (2) In the process of building the BGCL-KDE model, the present invention designs a guided diffusion model to generate a knowledge-aware comparative view. The model uses the node and entity embedding of the original knowledge graph to guide the generation of highly task-related nodes, enhance the semantic information of the nodes, ensure that the generated knowledge graph brings benefits to key drug or target knowledge, effectively handle the information of abnormal drugs or targets, and improve the quality of drug and target representation learning. The prediction method provided by the present invention can improve the prediction accuracy of tasks such as drug-protein chemical interactions. BRIEF DESCRIPTION OF THE DRAWINGS

[0029] Figure 1It is the overall framework of the BGCL-KDE model. DETAILED DESCRIPTION

[0030] The technical solution of the present invention is further described below in conjunction with the accompanying drawings, but the present invention is not limited to these embodiments.

[0031] There are several shortcomings of existing technologies when dealing with the field of drug discovery:

[0032] 1. Ignoring the importance of attention mechanisms: Existing methods often do not fully utilize attention mechanisms to identify the importance of different drug-target interactions in drug discovery, resulting in low accuracy in modeling the relationship between drugs and targets.

[0033] 2. Insufficient use of knowledge graphs to enhance node semantic information: Existing methods in the field of drug discovery often ignore a large amount of irrelevant information in knowledge graphs, which affects the enhancement of semantic information of drugs and targets and limits the ability to model drugs and targets.

[0034] 3. Lack of processing of abnormal drugs or targets: Existing methods often cannot effectively process abnormal drug or target information when generating knowledge graphs, and cannot guarantee that the generated knowledge graphs can always provide good help for key drug or target knowledge.

[0035] The purpose of this paper is to propose a bipartite graph contrastive learning architecture with knowledge-aware diffusion enhancement (BGCL-KDE) to address the limitations of existing drug discovery. The overall framework of the BGCL-KDE model is as follows: Figure 1 The steps are as follows:

[0036] (1) A relation-aware knowledge embedding layer is used to aggregate different types of entities (drug molecules, targets, proteins, diseases, etc.) in the KG into nodes (drug molecules and targets) in the bipartite graph, enhance the node semantic information, and generate a knowledge-aware bipartite graph.

[0037] (2) We design a graph embedding layer with an attention-aware mechanism to embed the two types of nodes in the knowledge-aware bipartite graph into a d-dimensional representation to generate attention-aware and knowledge-aware node feature aggregations to obtain the main view.

[0038] (3) Using the node and entity embeddings of the original knowledge graph as supplementary content to guide the knowledge graph that is highly relevant to the drug discovery task To enhance the semantic information of nodes. k Generate task-related subgraphs Build a comparison view The following steps are involved:

[0039] A. Add noise in the diffusion stage to destroy the original knowledge graph G k The relationship between nodes and entities.

[0040] B. In the reverse process, using the original knowledge graph G k The node and entity embeddings are used as complementary content to recover the original relations from pure Gaussian noise and iteratively identify the true relations between nodes and entities.

[0041] C. Using the restored relationship probabilities from the original knowledge graph G k Reconstruct task-related subgraphs

[0042] D. Reconstruct the knowledge graph Different types of entities are aggregated into nodes U or V of the original bipartite graph, enhancing the node semantic information and generating a knowledge-aware comparative view. Knowledge-aware contrastive views using graph embedding layers with attention-aware mechanisms The nodes in are used for data augmentation to obtain the BGCL-KDE model.

[0043] (4) Select positive and negative sample pairs from the main view and the comparison view to train the BGCL-KDE model:

[0044] Select positive and negative samples from the dataset, where positive samples represent the representation of the same node in different views, while negative samples represent the difference between the representation of the node in the main view and the full-image embedding of the comparison view; obtain the main task loss function by comparing the difference between the prediction scores of the positive sample pairs and the prediction scores of the negative sample pairs;

[0045] Map the main view and the contrast view to the feature space, compare the feature representations of the main view and the contrast view, and obtain the contrast loss function; use the contrast loss function to minimize the distance between similar sample pairs and maximize the interval between dissimilar sample pairs, so as to learn a better representation of the data representation;

[0046] Combine the contrast loss function with the main task loss function to form a total loss function that combines the two loss functions;

[0047] The total loss function is obtained and the gradient of the parameters is calculated through back propagation, and then the parameters of the model are updated to obtain the trained BGCL-KDE model.

[0048] (5) The drug and target data in the training set are input into the trained BGCL-KDE model, and the embedding vectors of the drug and target are obtained according to the generation order of the comparison view. The prediction result of the action relationship between the drug and the target is obtained by multiplying the drug embedding vector by the transpose of the target embedding vector. The drug-target predicted relationship is compared with the drug-target relationship in the test set, and the quality of the model is judged by calculating the AUROC and AUPR.

[0049] Explanation of terms involved in the present invention:

[0050] Bipartite graph learning: By constructing and learning a drug-target bipartite graph, the complex interaction patterns between drugs and targets are revealed. This method helps to explore the potential associations between drugs and targets, thereby guiding research tasks such as drug design and target prediction.

[0051] Diffusion model: Diffusion model refers to a mathematical model that describes the spread and influence of drugs or information in molecular networks. By simulating the spread of information or drugs in molecular networks, diffusion models can reveal important information such as potential therapeutic mechanisms and drug action pathways.

[0052] Contrastive learning: Contrastive learning refers to learning the feature expressions of different drugs or targets in the representation space by comparing their similarities and differences, so that similar samples are closer in the representation space and dissimilar samples are farther away in the representation space.

[0053] Knowledge-aware enhancement: Knowledge-aware enhancement refers to incorporating domain knowledge, drug interaction information, etc. into the model to enhance the model's understanding of drug characteristics and mechanisms of action.

[0054] Bipartite graph G = (U, V, Y), where U = {u 1 ,…,u i ,…,u I}(|U|=I) and V={v 1 ,…,v j ,…,v J}(|V|=J) defines two types of nodes. I and J define the number of two types of nodes. Y represents the interaction matrix between the two nodes, Y=[y i,j ] I×J ∈{0,1}, represents node u i and node v j If y ij =1, it means node u i and node v j There is an interaction between them, otherwise y ij =0.

[0055] The knowledge graph can be defined as G k =(h,r,t), the triple (h,r,t) describes the semantic correlation between the head entity h and the tail entity t in the relation type r. The entities h and t describe the nodes and their related concepts in the bipartite graph G.

[0056] D-dimensional representation usually refers to mapping data or entities into a vector space with d dimensions. In the present invention, it means that drug molecules and targets will be represented as vectors with d features.

[0057] The goal of the present invention is to predict unknown drug-target relationships, that is, to predict the interactions between drugs and targets that have not yet appeared in the knowledge graph. The drug-target relationship prediction method based on comparative knowledge enhancement provided by the present invention has the following advantages:

[0058] 1. Introducing attention mechanism to enhance drug and target representation learning: The present invention introduces an attention mechanism in the drug discovery process to effectively identify and utilize the importance differences between drugs and targets, improve the ability to model the interactions between them, and enhance the expressiveness of drug and target representations.

[0059] 2. Use knowledge graphs to enhance the semantic information of drugs and targets: The present invention integrates knowledge graphs into a comprehensive information network to enhance the semantic information of drugs and targets, fully utilize the beneficial information in the knowledge graph, and avoid harmful effects on drug discovery tasks.

[0060] 3. Handling the generation of abnormal drugs or targets: The present invention generates a robust knowledge graph through a diffusion model to ensure that the generated knowledge graph brings benefits to key drug or target knowledge, effectively handles the information of abnormal drugs or targets, and improves the quality of drug and target representation learning.

[0061] The drug-target relationship prediction method based on comparative knowledge enhancement provided by the present invention is described in detail below.

[0062] The drug-target relationship prediction method based on comparative knowledge enhancement provided by the present invention is specifically carried out according to the following steps:

[0063] Step 1: Aggregate different types of entities in the drug knowledge graph into nodes in the bipartite graph to generate a knowledge-aware bipartite graph; use the attention-aware mechanism to embed the nodes in the knowledge-aware bipartite graph into a d-dimensional representation so that the attention-aware and knowledge-aware node features are aggregated to obtain the main view. Specifically, it includes:

[0064] 1. Knowledge graph aggregation

[0065] Given a bipartite graph G = (U, V, Y) and an associated knowledge graph G k=(h,r,t), using the relationship-aware knowledge embedding layer to aggregate different types of entities in the KG into nodes U or V to enhance the semantic information of the nodes and generate a knowledge-aware bipartite graph G * =(U * ,V * ,Y * ), effectively capturing various relationships in the KG connection structure. The relationship-aware knowledge embedding layer between nodes and their connected entities can be defined by formula (1):

[0066]

[0067] Among them, z i ∈R d and z e ∈R d The embeddings of node i and entity e connected to it are defined respectively. i Through different types of relationships e,i In the knowledge graph, the neighbor entities of node i are defined, and the function Norm defines the normalization, z i Defines the enhanced knowledge graph embedding of node i. Using a(e,r e,i ,i) to estimate the entity specificity and relationship specificity between entity e and node i, which can distinguish the semantics of the relationship between node i and entity e and can be defined by formula (2):

[0068]

[0069] Here, σ defines a nonlinear activation function, W∈R d×2d is a parameterized weight matrix, Define the attention vector, N i Through different types of relationships e,i The neighbor entities of node i are defined in the knowledge graph. In this way, a(e,r e,i ,i) The semantics of the relationship between node i and entity e can be distinguished during the knowledge aggregation process. Therefore, the semantic information of node i can be enhanced through the knowledge graph.

[0070] 2. Attention-aware bipartite graph learning

[0071] After obtaining the KG-enhanced bipartite graph G * =(U * ,V * ,Y * ), the present invention proposes a graph embedding layer with an attention-aware mechanism to embed the two types of nodes U and V into a d-dimensional representation for generating attention-aware and knowledge-aware node feature aggregation to obtain the main view. In particular, the graph embedding layer embeds the two types of nodes u and V into a d-dimensional representation to generate attention-aware and knowledge-aware node feature aggregation to obtain the main view.i and v j As an embedding vector and Define the embedding matrix as Z (u) ∈R I×d and Z (v) ∈R J×d , respectively representing the embedding of two types of nodes U and V. Therefore, the present invention constructs a simplified graph embedding propagation layer, which removes the nonlinear activation function and feature conversion matrix, and adopts an attention-aware mechanism to embed two types of nodes, as defined in formula (3):

[0072]

[0073] in, Respectively represent the distance from adjacent nodes to the central node u i and v j Knowledge-aware node feature aggregation, represents the i-th row of the Laplacian matrix, Denotes the jth column of the Laplacian matrix. Attention-aware normalized Laplacian matrix The definition of is shown in formula (4):

[0074]

[0075] Where A defines the adjacency matrix of the knowledge-aware bipartite graph. (u) and D (v) The diagonal matrices of two node types U and V are defined respectively. The present invention uses an attention perception matrix S to define the similarity between the two node interactions, which can be obtained by formula (5). The present invention uses a hyperparameter δ to control the contribution of the attention perception matrix S.

[0076]

[0077] In formula (5), Con p (u,v) defines the number of common p-order neighbor nodes between node u and node v generated by random walk with restart strategy, which is calculated by the length M of the sampling path and the number R of the sampling paths. |N u | and |N v| defines the number of neighbor nodes of node u / v. Therefore, S(u,v) retains the importance between node u and node v, thereby identifying the importance between the two types of nodes and clarifying the graph embedding propagation layer at the node level through the contribution of each neighbor node. Therefore, the interpretability of the graph message passing layer is enhanced. The graph message passing layer of the present invention avoids repeated parameter updates during training, which improves the efficiency of the model compared to the adaptive training attention weight mechanism because the attention perception matrix S only needs to be performed once.

[0078] The present invention utilizes multiple embedding propagation layers to aggregate local neighbor information to refine the two node embeddings defined as formula (6):

[0079]

[0080] Here, and Represents node u i and node v j In the embedding of the lth embedding propagation layer, and Represents node u i and node v j Then, the present invention uses node u i and node v j The inner product between the final embeddings of i v j The prediction score is as shown in formula (7):

[0081]

[0082] In formula (7), T represents the matrix transpose, yi,j refers to u i The positive sample node (or negative sample node) of v j The prediction score of .

[0083] Step 2: Use the diffusion model to add noise in its diffusion stage to destroy the relationship between nodes and entities in the original drug knowledge graph; restore the original relationship from pure Gaussian noise in the reverse process, iteratively identify the true relationship between nodes and entities, and use the restored relationship to reconstruct the task-related subgraph. Specifically include:

[0084] 1. Diffusion process

[0085] Using the original knowledge graph G k Node and entity embeddings of to enhance the semantic information of nodes.

[0086] First, the original knowledge graph G is destroyed by adding noise in the diffusion stage. k Initial state relationship x between the middle node and the entity 0 , as shown in formula (8):

[0087]

[0088] Here, t∈{1…T} defines the diffusion step, x 0 represents the initial distribution representation of the original user embedding, x t represents the final state relationship after all diffusion steps, N defines the Gaussian distribution, I is the unit matrix, and a t Represents the initial state relationship x 0 Relative to x obtained after all diffusion steps t The remaining part, t' is the value from 1 to t, β t ∈(0,1) is used to control the scale of the Gaussian noise added at each step t. The initial state x 0 =r i The original relationship between node i and its connected entity e can be expressed by a matrix, as shown in formula (9):

[0089]

[0090] where r i It is defined that node i has a relationship with entities in entity set ε. Then node i has a relationship with entity e, and vice versa. Finally, state x T It converges to the standard Gaussian distribution when T→∞.

[0091] 2. Reverse process

[0092] For the destroyed original knowledge graph G k , the present invention in the reverse process from pure Gaussian noise x T Iteratively restore the initial state x 0 In order to guide the generated KG to always be conducive to task-related node knowledge, the present invention uses the node embedding and entity embedding of the original knowledge graph as supplementary content to guide the generation of task-related node KG. The diffusion model learns from x t Restore to x t-1 To eliminate the added noise, it can be defined by formula (10):

[0093] p θ (x t-1 |x t , z i , z e )=N(x t-1 ;uθ (x t , t, z i , z e ),∑ θ (x t , t, z i , z e )) (10)

[0094] Among them, u θ (x t ,t,z i ,z e ) and ∑ θ (x t ,t,z i ,z e ) defines the mean and covariance of the Gaussian distribution generated by the neural network parameterized by θ. i and z e They represent the node and entity embeddings in the original knowledge graph, respectively, and are used to guide the generation of task-related node knowledge graphs. t-1 Yes X t The previous state, p θ (x t-1 |x t ,z i ,z e ) means learning from x through neural network t Restore to x t-1 to remove the added noise.

[0095] In order to learn the additional noise at time step t through the neural network, the present invention can mean u θ Reparameterize as shown in formula (11):

[0096]

[0097] The present invention realizes ∈ by multi-layer perceptron θ (x t ,t,z i ,z e ), where x t and step embedding t as input to predict the initial state, and the predicted initial state is marked as The present invention utilizes the original knowledge graph G k Node and entity embeddings to guide Embedding refactoring so that node and entity embeddings are adjustable Since the node and entity embeddings serve as auxiliary semantic signals, it can ensure that the generated data is always beneficial to the diffusion enhancement model. k The initial state x between the midpoint and the entity 0The maximum likelihood ELBO is used to update the model parameters.

[0098] Step 3: Use the reconstructed task-related subgraph to modify the structure of the original drug knowledge graph to reconstruct a knowledge graph that is highly relevant to the drug discovery task; aggregate different types of entities in the reconstructed knowledge graph into nodes to generate a knowledge-aware comparative view; use the attention-aware mechanism to perform data enhancement on the nodes in the knowledge-aware comparative view to obtain the BGCL-KDE model. Specifically, it includes the following steps:

[0099] In obtaining reconstruction Afterwards, the present invention adopts To modify the structure of the knowledge graph to reconstruct a knowledge graph that is highly relevant to the task Specifically, the present invention selects the original relationship (marked as ) to adjust the KG structure by taking the top q relationships between node i and entity e that are highly relevant to the task, so that the reconstructed knowledge graph can be maintained. The information structure can effectively alleviate the noise problem.

[0100] Then, the present invention uses formula (1) to reconstruct the knowledge graph Different types of entities are aggregated into nodes U or V of the original bipartite graph, enhancing the node semantic information and generating a knowledge-aware comparative view.

[0101] Finally, use formula (6) to calculate the nodes and nodes Generate Embedding and It is used for data enhancement to obtain the final contrast view, denoted as the BGCL-KDE model.

[0102] Step 4: Compare the feature representation of the main view and the BGCL-KDE model to obtain the contrast loss function; use the optimized contrast loss function to update the parameters of the BGCL-KDE model, and use the updated BGCL-KDE model to predict the drug-target relationship. Specifically, it includes the following steps:

[0103] 1. Model training

[0104] After obtaining the BGCL-KDE model, select positive and negative samples from the dataset (positive samples represent the representation of the same node in different views, while negative samples represent the difference between the full-graph embedding of the node's main view representation and the comparison view), so that each sample pair contains a main view sample and a comparison view sample. Then, the main view and the comparison view are mapped to the feature space through an encoder (such as a neural network) to compare their representations. Next, by comparing the feature representations of the main view and the comparison view, the contrast loss function is used to minimize the distance between similar sample pairs while maximizing the interval between dissimilar sample pairs, thereby learning a better representation of the data representation. The specific operation process is as follows:

[0105] The present invention uses the InfoNCE loss function to compare and enhance the contrast view representation and the main view representation to update the model parameters. The present invention uses the contrast loss of node type U Defined as formula (12):

[0106]

[0107] where cosine similarity and temperature are defined as s(·) and τ, respectively. Defines the embedding of node u in the contrast view. Define the alignment between the main view and the contrast view, Define the negative pair of any two different nodes between the main view and the contrast view of a node u. The present invention defines the contrast loss of node V in the same way:

[0108] The present invention uses the main task loss to optimize the contrast loss and obtains the total loss function. The total loss function can be obtained by formula (13):

[0109]

[0110] The present invention adopts θ 1 and θ 2 Control the contribution of contrastive loss and model parameters Θ respectively. L r The main task loss function is used to measure the difference between the prediction score of the positive sample pair and the prediction score of the negative sample pair, which can be obtained by formula (14). u,i Defines the predicted score of the positive sample node v of node u, y u,j Defines the predicted score of the negative sample node v of node u.

[0111] L r =∑ (u，i，j)∈O -log(y u，i -y u，j ) (14)

[0112] The total loss function is obtained and the gradient of the parameters is calculated through back propagation, and then the parameters of the model are updated to obtain the trained BGCL-KDE model.

[0113] 2. Prediction of drug-target relationship

[0114] The drug and target data are input into the trained BGCL-KDE model to obtain the predicted drug-target relationship. The predicted results are compared with the actual drug-target relationship in the test set, and the AUROC and AUPR curves are calculated. The values ​​of AUROC and AUPR range from 0 to 1. The closer the value is to 1, the better the model performance is and the more accurate the prediction of the drug-target relationship is.

[0115] The present invention performs prediction according to the following steps and applies the above method:

[0116] 1. Construct a knowledge-aware bipartite graph: Aggregate known drug molecules and target nodes into a bipartite graph and establish relationships between nodes. This bipartite graph will contain known drug-target relationships.

[0117] 2. Embed nodes into a d-dimensional representation: Through a graph embedding layer with an attention-aware mechanism, drug molecules and target nodes are embedded into a d-dimensional representation to obtain the main view.

[0118] 3. Adding noise and reconstructing task-related subgraphs: Using the diffusion model, noise is added in the diffusion stage to destroy the known drug-target relationship, and then the original relationship is restored from the noise to reconstruct the task-related subgraphs.

[0119] 4. Generate comparative views: Use the reconstructed task-related subgraphs to modify the structure of the knowledge graph, reconstruct the task-related knowledge graph, and use the relationship-aware knowledge embedding layer to aggregate these entities into nodes to generate a knowledge-aware comparative view.

[0120] 5. Data augmentation and relationship prediction: The nodes in the contrast view are augmented through a graph embedding layer with an attention-aware mechanism to obtain the final contrast view. Ultimately, this trained contrast view can be used to predict unknown drug-target relationships.

[0121] The above is a simple example showing how to apply the given method in the drug-target relationship prediction task to improve the knowledge graph and predict potential drug-target relationships. In this way, knowledge graphs and relationship-aware embedding techniques can be used to help accelerate the drug discovery and development process.

[0122] The prediction accuracy of the bipartite graph learning contrast learning architecture BGCL-KDE model provided by the present invention was tested:

[0123] The trained model is evaluated using the test set, and the AUROC and AUPR indicators are calculated to evaluate the classification performance of the model.

[0124] The model performance evaluation results are shown in Table 1.

[0125] Table 1 Model performance evaluation results

[0126]

[0127] In the table, Metric, TransE, MuRE, KGE_NFM and KGE-UNIT are other models used for comparison.

[0128] As can be seen from Table 1, compared with other prediction models, the prediction model provided by the present invention has higher accuracy in predicting the drug-target relationship.

[0129] Those skilled in the art to which the present invention relates may make various modifications or additions to the specific embodiments described or replace them in a similar manner, but they will not deviate from the spirit of the present invention or exceed the scope defined by the attached claims.

Claims

1. A drug-target relationship prediction method based on comparative knowledge enhancement, characterized in that: The following steps are involved: Aggregate different types of entities in the drug knowledge graph into nodes in the bipartite graph to generate a knowledge-aware bipartite graph; use an attention-aware mechanism to aggregate the nodes in the knowledge-aware bipartite graph with attention-aware and knowledge-aware node features to obtain a main view; The diffusion model is used to add noise in its diffusion stage to destroy the relationship between nodes and entities in the original drug knowledge graph; In the reverse process, the original relations are restored from pure Gaussian noise, the real relations between nodes and entities are iteratively identified, and the restored relations are used to reconstruct the task-related subgraph; The structure of the original drug knowledge graph is modified by reconstructing task-related subgraphs to reconstruct a knowledge graph that is highly relevant to the drug discovery task. Different types of entities in the reconstructed knowledge graph are aggregated into the nodes of the original bipartite graph to generate a knowledge-aware comparative view. The attention-aware mechanism is used to aggregate the attention-aware and knowledge-aware node features of the nodes in the knowledge-aware contrast view to obtain the BGCL-KDE model. Compare the feature representations of the main view and the BGCL-KDE model to obtain a contrast loss function; use the optimized contrast loss function to update the parameters of the BGCL-KDE model, and use the updated BGCL-KDE model to predict the drug-target relationship; Optimizing the contrast loss function is performed according to the following steps: selecting positive samples and negative samples from a data set of drug-target relationships, comparing the difference between the predicted scores of the positive sample pairs and the predicted scores of the negative sample pairs, obtaining a main task loss function, and optimizing the contrast loss function using the main task loss function to obtain a total loss function; wherein the positive sample is the representation of the same node in the main view and the BGCL-KDE model, and the negative sample is the difference between the main view representation of the node and the full-image embedding of the BGCL-KDE model; The method for updating the parameters of the BGCL-KDE model is to update the parameters of the BGCL-KDE model by using the obtained total loss function and calculating the gradient of the parameters through back propagation.

2. The drug-target relationship prediction method based on comparative knowledge enhancement according to claim 1, characterized in that: The aggregation of different types of entities and nodes in the bipartite graph in the drug knowledge graph and the reconstructed knowledge graph is performed using a relationship-aware knowledge embedding layer.

3. The drug-target relationship prediction method based on comparative knowledge enhancement according to claim 1, characterized in that: The methods of attention-aware and knowledge-aware node feature aggregation are: The nodes in a knowledge-aware bipartite graph or a knowledge-aware contrast view are embedded into a d-dimensional vector, and a graph embedding propagation layer is constructed to remove the nonlinear activation function and the feature transformation matrix. Then, multiple graph embedding propagation layers are used to aggregate local neighbor information to achieve attention-aware and knowledge-aware node feature aggregation.

4. The drug-target relationship prediction method based on comparative knowledge enhancement according to claim 1, characterized in that: The method of modifying the structure of the drug knowledge graph using the reconstructed task-related subgraph is: The highly task-related relationships between nodes and entities in the reconstructed task-related subgraph are selected to adjust the structure of the knowledge graph to maintain the information structure of the reconstructed knowledge graph.

5. The drug-target relationship prediction method based on comparative knowledge enhancement according to claim 1, characterized in that: The contrast loss function is obtained by mapping the main view and the BGCL-KDE model into the feature space and comparing the feature representations of the main view and the BGCL-KDE model.

6. The drug-target relationship prediction method based on comparative knowledge enhancement according to claim 1, characterized in that: The prediction of drug-target relationship using the updated BGCL-KDE model is carried out in the following steps: the drug and target data are input into the updated BGCL-KDE model, the drug embedding vector and the target embedding vector are obtained in the order of generating the comparison view, and the prediction result of the action relationship between the drug and the target is obtained by multiplying the drug embedding vector by the transpose of the target embedding vector.

7. The drug-target relationship prediction method based on comparative knowledge enhancement according to claim 6, characterized in that: The drug embedding vector and target embedding vector are obtained using a graph embedding layer with an attention-aware mechanism.

8. The drug-target relationship prediction method based on comparative knowledge enhancement according to claim 1, characterized in that: Different types of entities in the drug knowledge graph are drug molecules, targets, and diseases.

9. The drug-target relationship prediction method based on comparative knowledge enhancement according to claim 1, characterized in that: The nodes in the bipartite graph are drug molecules and targets.

Images