Compound target interaction prediction method and system

Abstract

The invention relates to the technical field of pharmacology and artificial intelligence, and provides a compound-target interaction prediction method and system, and the method comprises the steps: constructing a knowledge graph containing compounds, target entries and disease entities, and the correlation of the compounds, the target entries and the disease entities; inputting the knowledge graph into a multi-modal feature extraction model, and extracting topological information and internal semantic information in the knowledge graph; wherein the multi-modal feature extraction model comprises an ICD2Vec layer, a graph Transformer, a gene ontology embedding layer and an attention mechanism-based graph neural network, and the ICD2Vec layer, the graph Transformer, the gene ontology embedding layer and the attention mechanism-based graph neural network are respectively used for extracting disease semantic features, compound semantic features, protein target semantic features and topological features of a knowledge graph; performing knowledge fusion on the extracted semantic features and topological features of the knowledge graph through a feature fusion module, and obtaining a compound-target prediction probability matrix through a knowledge graph embedding method; and according to the compound-target prediction probability, performing descending sorting on the input compound-target interaction, and outputting a target recommendation list.

Description

Technical Field

[0001] The present invention relates to the field of pharmacology and artificial intelligence technology, and more specifically, to a compound-target interaction prediction method and system. Background Art

[0002] Drug molecules exert their therapeutic effects by acting as ligands to interact with target proteins and stimulate or inhibit the biological processes of the target. Target identification is a key step in the drug development process, which aims to determine the direct target of small molecules. Although traditional in vivo or in vitro experiments can be used to predict compound-target interaction (CTI), the identification methods are accurate, but time-consuming, laborious and costly, which greatly limits their widespread application in high-throughput drug screening. Therefore, efficient and accurate computational methods have received widespread attention in the prediction of drug-target interaction (DTI).

[0003] Researchers have now developed many computational methods for predicting compound-target interactions, which can be roughly divided into network-based methods and classification-based methods. Network-based methods use complex graph theory algorithms to infer new molecular interactions by representing bioactivity information as homogeneous or heterogeneous networks. Such methods do not learn the intrinsic semantic features of nodes in the network, such as compounds and targets, and rely on the integrity of the network structure, so their performance is limited when relational data is sparse or incomplete. Classification-based methods usually use deep learning models to extract features of compounds and targets, and treat target prediction as a binary classification problem. Deep learning models are usually superior to these shallow machine learning methods in prediction accuracy due to their powerful feature learning capabilities. However, the performance of classification methods depends largely on the selection of extracted features and the size of training data.

[0004] In addition, in current target prediction technologies, the data used to drive CTI prediction models are usually based on compound-target interactions, and disease entities are often used as additional auxiliary information in biomedical heterogeneous networks. Their feature representation is mainly based on the disease-disease similarity of the ontology, but there are few methods that integrate the semantic representation of diseases composed of biomedical terms, resulting in insufficient extraction of semantic information of disease entities; secondly, existing methods often only focus on the topological features or semantic features of nodes in biological interaction networks, and are unable to integrate the advantages of network-based and classification methods, thus weakening the target prediction performance; thirdly, although there are many types of compound-target interaction prediction methods, target priority tools are still relatively scarce and cannot provide effective guidance for subsequent experimental verification. Summary of the invention

[0005] In order to overcome the defects of insufficient semantic information extraction of disease entities and low target prediction performance in the above-mentioned prior art, the present invention provides a compound-target interaction prediction method and system.

[0006] In order to solve the above technical problems, the technical solution of the present invention is as follows: A method for predicting compound-target interactions comprises the following steps: Construct a knowledge graph containing compound, target and disease entities and their interrelationships; Inputting the knowledge graph into a multimodal feature extraction model to extract topological information in the global knowledge graph and intrinsic semantic information of local entities; wherein the multimodal feature extraction model includes an ICD2Vec layer, a graph transformer, a gene ontology embedding layer, and a graph neural network based on an attention mechanism, which are respectively used to extract disease semantic features, compound semantic features, protein target semantic features, and topological features of the knowledge graph; After fusing the disease semantic features, compound semantic features, protein target semantic features and topological features of the knowledge graph through a feature fusion module, a compound-target prediction probability matrix is ​​obtained through a knowledge graph embedding method; The input compound-target interactions are sorted in descending order according to the compound-target prediction probability, and a target recommendation list is output.

[0007] Furthermore, the present invention also proposes a compound-target interaction prediction system, which applies the compound-target interaction prediction method proposed in the present invention. The system comprises: The knowledge graph module is used to construct a knowledge graph containing compound, target and disease entities and their interrelationships; A multimodal feature extraction module, which is equipped with a multimodal feature extraction model for extracting topological information in the global knowledge graph and intrinsic semantic information of local entities from the input knowledge graph data; wherein the multimodal feature extraction model includes an ICD2Vec layer, a graph transformer, a gene ontology embedding layer, and a graph neural network based on an attention mechanism, which are respectively used to extract disease semantic features, compound semantic features, protein target semantic features, and topological features of the knowledge graph; A prediction module, which is used to fuse the disease semantic features, compound semantic features, protein target semantic features and topological features of the knowledge graph through a feature fusion module, and obtain a compound-target prediction probability matrix through a knowledge graph embedding method; The output module is used to sort the input compound-target interactions in descending order according to the compound-target prediction probability and output a target recommendation list.

[0008] Compared with the prior art, the technical solution of the present invention has the following beneficial effects: The present invention extracts the semantic features of compounds, protein targets and disease entities by combining the ICD2Vec layer, graph transformer and gene ontology embedding layer, fully considers the semantic information in the entities, and uses the graph neural network based on the attention mechanism to extract the topological information of entities and relationships contained in the knowledge graph, and further integrates it with the semantic features, which can effectively enhance the feature representation of nodes, improve the model's expression ability, performance and generalization ability, and thus improve the target prediction performance; The present invention further utilizes the knowledge graph embedding method to calculate the compound-target prediction probability of the feature representation that integrates semantic features and topological information of entities and relationships, completes the compound-target association prediction task, and effectively improves the compound-target association prediction accuracy. BRIEF DESCRIPTION OF THE DRAWINGS

[0009] Figure 1 The figure is a flow chart of a method for predicting compound-target interaction according to one embodiment of the present invention.

[0010] Figure 2 It is a schematic diagram of the prediction ranking of the TarIKGC model and the STarFish model on two data sets respectively according to an embodiment of the present invention.

[0011] Figure 3 The top-10 results of the TarIKGC model and the STarFish model on two data sets are shown in accordance with an embodiment of the present invention. k Hit rate diagram.

[0012] Figure 4 FIG. 1 is a schematic diagram of measuring the binding affinity of a compound to CDK2 using SPR according to one embodiment of the present invention.

[0013] Figure 5 FIG. 1 is a schematic diagram showing the molecular binding mode of two compounds in the ATP pocket of CDK2 according to one embodiment of the present invention.

[0014] Figure 6 FIG. 4 is an architecture diagram of a compound-target interaction prediction system according to an embodiment of the present invention. DETAILED DESCRIPTION

[0015] Exemplary embodiments will be described in detail herein, examples of which are shown in the accompanying drawings. When the following description refers to the drawings, the same numbers in different drawings represent the same or similar elements unless otherwise indicated. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with the present invention. Instead, they are merely examples of devices and methods consistent with some aspects of the present invention as detailed in the appended claims.

[0016] The terms used in the present invention are only for the purpose of describing specific embodiments and are not intended to limit the present invention. The singular forms "a", "the" and "the" used in the present invention and the appended claims are also intended to include plural forms unless the context clearly indicates other meanings. It should also be understood that the term "and / or" used herein refers to and includes any or all possible combinations of one or more associated listed items.

[0017] It should be understood that although the terms first, second, third, etc. may be used in the present invention to describe various information, these information should not be limited to these terms. These terms are only used to distinguish the same type of information from each other. For example, without departing from the scope of the present invention, the first information may also be referred to as the second information, and similarly, the second information may also be referred to as the first information. Depending on the context, the word "if" as used herein may be interpreted as "at the time of" or "when" or "in response to determining".

[0018] The present invention is described in detail below with reference to the accompanying drawings and specific embodiments.

[0019] Example 1 This example proposes a method for predicting compound-target interactions. Figure 1 Shown is a flow chart of the compound-target interaction prediction method of this embodiment.

[0020] The compound-target interaction prediction method proposed in this embodiment includes the following steps: S1. Construct a knowledge graph containing compound-target interactions, compound-disease associations, and target-disease associations; S2, inputting the knowledge graph into a multimodal feature extraction model to extract topological information in the global knowledge graph and intrinsic semantic information of local entities; The multimodal feature extraction model includes an ICD2Vec layer, a graph transformer, a gene ontology embedding layer, and a graph neural network based on an attention mechanism, which are used to extract disease semantic features, compound semantic features, protein target semantic features, and topological features of the knowledge graph, respectively; S3, fusing the disease semantic features, compound semantic features, protein target semantic features and topological features of the knowledge graph through a feature fusion module, and obtaining a compound-target prediction probability matrix through a knowledge graph embedding method; S4. Sort the input compound-target interactions in descending order according to the compound-target prediction probability, and output a target recommendation list.

[0021] In this embodiment, a biological knowledge graph including compounds, protein targets and disease entities and their mutual associations is first constructed, and then the intrinsic semantic features of the three entities in the knowledge graph are extracted through a multimodal feature extraction model, including disease semantic features, compound semantic features and protein target semantic features. At the same time, a graph neural network based on an attention mechanism is used to synchronously extract the topological information of entities and relationships contained in the knowledge graph. Finally, a knowledge graph embedding method is used to complete the compound-target association prediction task and output a target recommendation list.

[0022] In this embodiment, the semantic features of compounds, protein targets and disease entities are extracted, the semantic information in the entities is fully considered, and the graph neural network based on the attention mechanism is used to synchronously extract the topological information of entities and relationships contained in the knowledge graph and fuse them with the semantic features. This can effectively enhance the feature representation of nodes, improve the model's expression ability, performance and generalization ability, and thus improve the target prediction performance.

[0023] In an optional embodiment, the data in the knowledge graph in step S1 is stored in the form of a triple of head entity-relationship-tail entity, and the collected compound-target interaction, compound-disease association and target-disease association data are stored as a knowledge graph.

[0024] For example, compound-target interaction, compound-disease association, and target-disease association data are collected from the TTD database; drug-target interaction information is extracted from the DrugBank database; and the structure of natural products and their related target information are collected from the NPASS database. The data stimulated from each database are then stored in the form of triples to form a knowledge graph.

[0025] In an optional embodiment, the process of constructing the knowledge graph in step S1 further includes: The compound data were molecularly normalized and filtered; the target entries were annotated using UniProt IDs; and the disease entities were annotated using ICD (International Classification of Diseases).

[0026] Among them, the compound data is subjected to molecular standardization filtering to ensure the chemical rationality and consistency of the molecular structure, including partial desalting, charge neutralization and molecular validity verification.

[0027] UniProt ID is a number used to uniquely identify each protein entry in the UniProt (Universal Protein Resource) database, usually consisting of uppercase letters and numbers. This ID is stable, and even if the protein sequence is updated, the UniProt ID remains unchanged, thus ensuring consistency and traceability between scientific research literature and databases.

[0028] ICD is an international standard classification system developed and maintained by the World Health Organization for coding, classifying and describing all known human diseases, symptoms and injuries, as well as health-related problems. Using ICD to label disease entities can ensure the rationality and consistency of data labeling.

[0029] In an optional embodiment, the ICD2Vec layer includes an ICD code converter based on the ICD2Vec framework, which is used to convert the ICD code in the input data into a disease semantic representation.

[0030] Among them, ICD2Vec is a framework that converts ICD codes into mathematical vectors, which can effectively capture nonlinear relationships such as semantic relationships between diseases.

[0031] Exemplarily, in this embodiment, the ICD code of the disease entity is converted into a 1024-dimensional disease representation through the ICD2Vec layer to obtain a disease semantic representation.

[0032] In an optional embodiment, the graph transformer performs the following steps when extracting compound semantic features: Defining the query matrix , key matrix Sum Matrix ; where subscript i Represents the first i Transformer; The input knowledge graph data is extracted through the graph neural network to generate node representation; Calculate the scaled dot product attention score based on the node representation, and obtain the semantic weight matrix through the softmax function; Calculating the Laplacian matrix and the Euclidean distance matrix of the graph, and performing symmetric normalization processing on the Laplacian matrix; The Laplace matrix after symmetric normalization is , and the Euclidean distance matrix After merging with the semantic weight matrix, the structure-aware three-dimensional feature weight matrix is ​​obtained through the self-attention head ; Its expression is:

[0033] in, For the i The first n The three-dimensional feature weight matrix output by the self-attention head, Represents the key matrix ; represents the softmax function; Represents stacked GNN layers; represents the convolutional layer; Will N The three-dimensional feature weight matrix output by each self-attention head Perform connection and projection processing to obtain the attention output vector ; Its expression is:

[0034] in, is the trainable parameter matrix, is the dimension of the vector; Indicates a connection operation; The attention output vector After being processed by a feed-forward network with a skip connection residual structure, the semantic features of the compound are output.

[0035] In this embodiment, the graph transformer is selected to autonomously learn the semantic and spatial structural features of compounds to realize the extraction of compound semantic features. Among them, the distance features between atoms are introduced into the self-attention mechanism of the graph transformer, and the spatial structural features of molecular entities are better understood by introducing the Euclidean distance matrix; while the node features in the graph transformer represent the characteristics at the atomic level, and the connectivity between different nodes in the molecular graph is reflected by introducing the normalized Laplace matrix to improve the accuracy of the semantic features.

[0036] The normalized Laplacian matrix in this embodiment is defined as: ;in, is the adjacency matrix, is the degree matrix.

[0037] Optionally, the attention output vector During the feedforward network processing, multiple skip connection residual structures, dropout layers and normalization layers are used to improve the performance and stability of the model.

[0038] In an optional embodiment, the gene ontology embedding layer performs the following steps when extracting semantic features of protein targets: After the input knowledge graph data is annotated with GO functions, it is input into the residual multilayer perceptron network (ResMLP) to obtain the semantic features of protein targets; wherein the multilayer perceptron in the residual multilayer perceptron network is defined as:

[0039] in, Indicates input with GO annotation; is the bias term; is a trainable parameter matrix; Represents the ReLU function; represents batch normalization; represents a dropout layer; Represents the output of the current multi-layer perceptron; All multi-layer perceptrons are residually connected and the final output is mapped to a common entity embedding space to obtain the semantic features of protein targets. ; Its expression is: .

[0040] The gene ontology embedding layer in this embodiment uses GO functional annotations to generate protein target semantic feature representations in a biological context through a residual multi-layer perceptron network.

[0041] Among them, GO functional annotation (Gene Ontology Annotation) is a standardized classification system used to describe the functions of genes and gene products (such as proteins). Specifically, it includes the steps of assigning GO terms to gene sequences.

[0042] Furthermore, feature extraction is performed through a residual multilayer perceptron network, in which information can be transferred between residual-connected multilayer perceptrons (MLP), which can effectively alleviate the gradient vanishing problem.

[0043] In an optional embodiment, the graph neural network based on the attention mechanism performs the following steps when extracting topological features of the knowledge graph: The established knowledge graph is input into the graph neural network, and the graph neural network is used to aggregate the node features and relationship features corresponding to the central node and its neighboring nodes based on the attention mechanism to obtain the topological structure features of the central entity; The updating process of the topological structure feature of the central entity includes: Initial embedding of With its structural level representation and relationship hierarchy representation Merge and update entity embedding for the next iteration , until the iteration is completed; its expression is: ; in, Represents the central entity In the l The structural level representation in the +1 iteration is updated by iterative multi-layer aggregation from neighboring entities; its expression is:

[0044]

[0045] in, Represents the tail entity In the l Features in iterations; Represents the central entity A dataset of neighboring entities, The tail entity of the adjacent entity Relationship Entity Indicated in l The trainable weight matrix in iterations; is the activation function; Indicated in l Structure-aware attention weights in iterations; Represents the trailing entity of another adjacent entity and relationship entities ; Represents the tail entity In the l Features in iterations; Representing Entities In the l The relational hierarchy representation in the +1 iteration calculates the attention score by aggregating all the connection relations of the central entity and uses it for message passing, which is updated by the sum of weighted edge embeddings; its expression is:

[0046]

[0047] in, Representing relationship entities In the l Display features in iterations; Indicated in l The trainable weight matrix in iterations is used to linearly transform the aggregated neighbor embeddings; Indicated in l Relation-aware attention weights in iterations.

[0048] In this embodiment, the knowledge graph data is first formatted into a graph structure, where entities are nodes and relationships are edges, and an initial feature vector is assigned to each node. Then, the node features and topological structure are processed through a stacked graph neural network, and the graph neural network based on the attention mechanism is used to capture the topological information. The structure-aware attention weights are used to aggregate neighboring nodes, so that the entity embedding is updated according to the importance of neighbor entities and relationships. The relationship embedding is also dynamically updated with the trainable parameters of the model, which follows the message passing rules in the graph neural network. Finally, the entities and relationships in the knowledge graph are projected into the embedding space, and entities with similar neighborhoods will be placed in close positions.

[0049] In this embodiment, a graph neural network is used to extract the topological information of entities and relationships contained in the knowledge graph, which can effectively integrate the entity features and relationship features of the knowledge graph, aggregate and update nodes through the neighborhood information of the nodes, and use its powerful information propagation ability to learn the semantic relationships and potential information between data. It can well learn the node information, node relationship information and global structure information in knowledge reasoning, thereby making up for the missing information in the knowledge graph and improving the prediction performance of compound-target interactions.

[0050] In an optional embodiment, the method of obtaining a compound-target prediction probability matrix through a knowledge graph embedding method includes: using a DistMult semantic matching model to generate a prediction probability for each possible triple in the knowledge graph to generate a compound-target prediction probability matrix.

[0051] Among them, the DistMult semantic matching model is a knowledge graph representation learning model, and its scoring function only captures the pairwise interactions between the head entity and the tail entity components along the same dimension, namely: ,This scoring function calculates the interaction between entities and relationships through the inner product. The higher the prediction score, the stronger the rationality of the interaction between the compound and the target.

[0052] This embodiment uses the DistMult semantic matching model to embed the entity and relationship representations generated by the encoder into the vector space, and uses semantic similarity to reveal potential associations and complete the original knowledge graph.

[0053] In an optional embodiment, the method further includes: The fact triples in the knowledge graph are regarded as positive sample data, all undiscovered triples are regarded as sources of negative sampling, and damaged triples are constructed as negative sample data according to different negative sample-positive sample ratios to form a training data set; The training data set is input into the multimodal feature extraction model for cross-validation evaluation training; during the cross-validation evaluation training process, for each head entity in the training triplet , retrieve all tail entities associated with it, including and , and use gradient descent to minimize the binary cross entropy loss function ; Its expression is:

[0054] in, is the scoring function used to measure the rationality of the ternary combination; are the head entity, relation entity and tail entity in the positive sample triple, represents the positive sample set; Represents the true label of the positive sample triple; is the tail entity in the negative sample triple, Represents the set of negative samples.

[0055] Further optionally, when constructing damaged triples as negative sample data according to different negative sample-positive sample ratios, it includes: According to the preset negative sample-positive sample ratio, some randomly selected positive sample triplets are The tail entity in Replace with another entity Generate negative sample triplets ; Positive sample triples And negative sample triplets Labeled as 1 and 0 respectively.

[0056] In the prediction stage, the multimodal feature extraction model trained in this embodiment is applied to extract multimodal features of the knowledge graph to obtain disease semantic features, compound semantic features, protein target semantic features and topological features of the knowledge graph, and the extracted multimodal features are fused with knowledge, and the compound-target prediction probability is calculated through the DistMult semantic matching model to form a compound-target prediction probability matrix. Finally, the input compound-target interactions are sorted in descending order based on the compound-target prediction probability, and the most likely interactions are prioritized at the top of the recommendation list to generate a target recommendation list output.

[0057] Example 2 This example applies the compound-target interaction prediction method proposed in Example 1 to evaluate the compound-target interaction prediction performance by comparing it with ten existing benchmark models.

[0058] The benchmark models used in this embodiment include seven classification-based methods: RF, KNN, STarFish, BridgeDPI, HyperPCM, BINDTI, and TransformerCPI, one network-based method: DTINet, and two integrated methods: KG-MTL and CCL-DTI.

[0059] Regarding the knowledge graph, this embodiment collects data on compound-target interactions, compound-disease associations, and target-disease associations from the TTD database, extracts drug-target interaction information from the DrugBank database, and collects the structure of natural products and their related target information from the NPASS database. The knowledge graph thus established contains 20,550 compounds, 709 diseases, and 775 targets, covering 50,169 associations of three types of relationships. In addition, 6,990 specific GO functional annotations were collected, covering molecular functions (MF), cellular components (CC), and biological processes (BP).

[0060] The association pairs in the current knowledge graph are positive samples, and negative samples are pre-sampled at three ratios to obtain the comparison results of compound-target interaction prediction performance shown in Table 1 below, which reflects the performance of the present invention and ten baseline models under different ratios of negative samples to positive samples. Among them, the prediction performance of compound-target interaction is reflected by using indicators such as accuracy (ACC), area under the receiver operating characteristic curve (AUC), area under the precision-recall curve (AUPR), recall rate (Recall), and positive-negative sample ratio (Pos: Neg).

[0061] Table 1 Comparison of compound-target interaction prediction performance

[0062] As can be seen from Table 1, the method of the present invention outperforms all baseline models in 9 of the 12 statistical indicators. Specifically, the method of the present invention performs best in all positive and negative sample ratio conditions in terms of ACC, AUC and AUPR indicators. With the increase of negative data in the training set, the method of the present invention shows an upward trend in the ACC and AUC indicators. Under the pressure of class imbalance, AUPR and recall rate decrease, which is consistent with the performance of other models. This decline is mainly due to the reduction in the number of true positives, which leads to a decrease in the confidence of the model in correctly classifying true triples.

[0063] In addition, the method of the present invention shows better recommendation performance of potential targets in terms of ranking evaluation indicators compared with STarFish, a target prediction model focused on natural products. Target prediction of natural products (NPs) is a major challenge in drug discovery.

[0064] In order to evaluate the effectiveness and generalization ability of the model, this example uses the natural product NPs-centered model STarFish as a benchmark model for target capture, and external validation uses two external datasets TeroKit and BATMAN-TCM. To ensure fairness, both models are trained on the two proposed datasets and the compound-protein interaction collection in the STarFish study. Detailed data is shown in Table 1. The goal is to determine the ability of the model to successfully identify targets associated with terpenoids and traditional Chinese medicines (TCMs) in these external datasets.

[0065] like Figure 2 , 3 As shown, Figure 2 The following is a schematic diagram of the prediction rankings of the TarIKGC model and the STarFish model on two datasets. Figure 3 The top-10 performances of TarIKGC model and STarFish model on the two datasets are shown in Figure 2. k Hit rate diagram.

[0066] Depend on Figure 2 It can be seen that the quartiles of the TarIKGC model of the present invention are lower than those of STarFish in all four test sets, indicating that the TarIKGC of the present invention has a significantly higher prediction ranking for potential associations. Although the biological activity tests of the protein targets of these NPs are limited, the results show that the target reasoning ability of the TarIKGC model of the present invention is stronger.

[0067] In addition, the introduction of Top- k Evaluation indicators are used to quantitatively evaluate the prediction performance. Figure 3As shown, the TarIKGC model of the present invention is in Top- k Hit rate ( k = 5, 10, 20). For example, when trained on the STarFish dataset, TarIKGC achieved a global Top-20 hit rate of 29.8% in predicting terpenoid-target associations and a hit rate of 26.6% in predicting TCMs-target associations, which are 1.6 and 1.2 times higher than STarFish, respectively. It should be noted that the Top-20 hit rates of both models are significantly higher than those of the models trained on the proposed dataset. k The evaluation indicators all decreased. This decrease may be due to the large number of candidate targets in the STarFish dataset (1,907 targets, while the present invention has 775 targets), which increases the difficulty of model recommendation. This experiment provides a fair evaluation of model performance and demonstrates the advantages of our model over the baseline model in terms of transferability and generalization.

[0068] In order to further verify the reliability of this method, this example conducted a reverse target prediction experiment to identify potential compounds that significantly target cyclin-dependent kinase 2 (CDK2). Among them, two compounds ranked high in the recommended list and ranked third in the recommended list. It is worth noting that these two compounds have not been reported as CDK2 inhibitors before.

[0069] To verify these computational predictions, surface plasmon resonance (SPR) experiments were performed to measure the binding affinities of the two compounds to monomeric CDK2, and the following results were obtained: Figure 4 Schematic diagram of SPR measurement of the binding affinity of compounds to CDK2.

[0070] The affinities of both compounds were in the nanomolar range, as determined by steady-state approximation (SSA) affinity fitting, with equilibrium dissociation constants K D The values ​​were 0.35 μM and 4.63 μM, respectively. Molecular docking structure showed that for compound 1, its tricyclic structure was located in the hydrophobic pocket formed by amino acids such as valine and alanine, forming a close hydrophobic contact, such as Figure 5 As shown. In addition, its NH group forms a hydrogen bond with Asp86, while Lys89 donates a hydrogen bond to the oxygen atom of its ester group. The purine of compound 2 forms two key hydrogen bond interactions with Leu83 in the hinge binding site region, while the nitrogen atom of its amide group forms a hydrogen bond with Asp86. Similar hydrogen bond interactions have also been reported in previous studies on the binding mode of CDK2 inhibitors in the ATP binding site.

[0071] It can be seen that the compound-target interaction prediction method proposed in the present invention has excellent recommendation performance in discovering potential protein targets of compounds.

[0072] Example 3 This example applies the compound-target interaction prediction method proposed in Example 1 to propose a compound-target interaction prediction system. Figure 6 , which is a schematic diagram of the compound-target interaction prediction system of this embodiment.

[0073] The compound-target interaction prediction system proposed in this embodiment includes: The knowledge graph module is used to construct a knowledge graph containing compound, target and disease entities and their interrelationships; A multimodal feature extraction module, which is equipped with a multimodal feature extraction model for extracting topological information in the global knowledge graph and intrinsic semantic information of local entities from the input knowledge graph data; wherein the multimodal feature extraction model includes an ICD2Vec layer, a graph transformer, a gene ontology embedding layer, and a graph neural network based on an attention mechanism, which are respectively used to extract disease semantic features, compound semantic features, protein target semantic features, and topological features of the knowledge graph; A prediction module, which is used to fuse the disease semantic features, compound semantic features, protein target semantic features and topological features of the knowledge graph through a feature fusion module, and obtain a compound-target prediction probability matrix through a knowledge graph embedding method; The output module is used to sort the input compound-target interactions in descending order according to the compound-target prediction probability and output a target recommendation list.

[0074] It can be understood that the system of this embodiment corresponds to the method of the above-mentioned embodiment 1, and the options in the above-mentioned embodiment 1 are also applicable to this embodiment, so they will not be described repeatedly here.

[0075] Example 4 This embodiment proposes a computer device, including a memory and a processor, wherein the memory stores computer-readable instructions, and when the computer-readable instructions are executed by the processor, the processor executes all or part of the steps of the compound-target interaction prediction method proposed in Example 1.

[0076] Example 5 This embodiment proposes a storage medium having computer-readable instructions stored thereon, wherein the computer-readable instructions, when executed by a processor, implement all or part of the steps of the compound-target interaction prediction method proposed in Example 1.

[0077] Exemplarily, the storage medium includes, but is not limited to, a USB flash drive, a mobile hard disk, a read-only memory (ROM), a random access memory (RAM), a magnetic disk or an optical disk, and other media that can store program codes.

[0078] Exemplarily, the instructions, programs, code sets or instruction sets may be implemented using conventional programming languages.

[0079] Exemplarily, the processor includes but is not limited to a smart phone, a personal computer, a server, a network device, etc., and is used to execute all or part of the steps of the compound-target interaction prediction method described in Example 1.

[0080] Each embodiment of the present invention is described in a progressive manner, and the same or similar parts between the embodiments can be referred to each other, and each embodiment focuses on the differences from other embodiments. In particular, for the device embodiment, since it is basically similar to the method embodiment, the description is relatively simple, and the relevant parts can be referred to the partial description of the method embodiment. The device embodiment described above is merely exemplary, in which the modules described as separate components may or may not be physically separated, and the functions of each module can be implemented in the same or one or more software and / or hardware when implementing the scheme of the present invention. It is also possible to select some or all of the modules according to actual needs to achieve the purpose of the scheme of this embodiment.

[0081] Obviously, the above embodiments of the present invention are merely examples for clearly illustrating the present invention, and are not intended to limit the embodiments of the present invention. For those skilled in the art, other different forms of changes or modifications can be made based on the above description. It is not necessary and impossible to list all the embodiments here. Any modifications, equivalent substitutions and improvements made within the spirit and principles of the present invention should be included in the protection scope of the claims of the present invention.

Claims

1. A method for predicting compound-target interactions, characterized in that: The following steps are involved: Construct a knowledge graph containing compound, target and disease entities and their interrelationships; Inputting the knowledge graph into a multimodal feature extraction model to extract topological information in the global knowledge graph and intrinsic semantic information of local entities; wherein the multimodal feature extraction model includes an ICD2Vec layer, a graph transformer, a gene ontology embedding layer, and a graph neural network based on an attention mechanism, which are respectively used to extract disease semantic features, compound semantic features, protein target semantic features, and topological features of the knowledge graph; After fusing the disease semantic features, compound semantic features, protein target semantic features and topological features of the knowledge graph through a feature fusion module, a compound-target prediction probability matrix is ​​obtained through a knowledge graph embedding method; The input compound-target interactions are sorted in descending order according to the compound-target prediction probability, and a target recommendation list is output.

2. The compound-target interaction prediction method according to claim 1, characterized in that: In the knowledge graph, the data in the knowledge graph is stored in the form of triples; wherein the construction process of the knowledge graph also includes: The compound data were filtered by molecular standardization; The target entries were annotated using UniProt IDs; Disease entities are annotated using ICD.

3. The compound-target interaction prediction method according to claim 2, characterized in that: The ICD2Vec layer includes an ICD code converter based on the ICD2Vec framework, which is used to convert the ICD code in the input data into a disease semantic representation.

4. The compound-target interaction prediction method according to claim 1, characterized in that: The graph transformer performs the following steps when extracting compound semantic features: Defining the query matrix , key matrix Sum Matrix ; where subscript i Represents the first i Transformer; The input knowledge graph data is extracted through the graph neural network to generate node representation; Calculate the scaled dot product attention score based on the node representation, and obtain the semantic weight matrix through the softmax function; Calculating the Laplacian matrix and the Euclidean distance matrix of the graph, and performing symmetric normalization processing on the Laplacian matrix; The Laplace matrix after symmetric normalization is , and the Euclidean distance matrix After merging with the semantic weight matrix, the structure-aware three-dimensional feature weight matrix is ​​obtained through the self-attention head ; Its expression is: in, For the i The first n The three-dimensional feature weight matrix output by the self-attention head, Represents the key matrix ; represents the softmax function; Represents stacked GNN layers; represents the convolutional layer; Will N The three-dimensional feature weight matrix output by each self-attention head Perform connection and projection processing to obtain the attention output vector ; Its expression is: in, is the trainable parameter matrix, is the dimension of the vector; Indicates a connection operation; The attention output vector is processed by a feed-forward network combined with a skip connection residual structure to output compound semantic features.

5. The compound-target interaction prediction method according to claim 1, characterized in that: The gene ontology embedding layer performs the following steps when extracting semantic features of protein targets: After the target entity in the input knowledge graph is annotated with GO function, it is input into the residual multi-layer perceptron network to obtain the semantic features of the protein target; wherein the multi-layer perceptron in the residual multi-layer perceptron network is defined as: in, Indicates input with GO annotation; is the bias term; is a trainable parameter matrix; Represents the ReLU function; represents batch normalization; represents a dropout layer; Represents the output of the current multi-layer perceptron; All multi-layer perceptrons are residually connected and the final output is mapped to a common entity embedding space to obtain the semantic features of protein targets. ; Its expression is: 。 6. The compound-target interaction prediction method according to claim 1, characterized in that: The graph neural network based on the attention mechanism performs the following steps when extracting topological features of the knowledge graph: The established knowledge graph is input into the graph neural network, and the graph neural network is used to aggregate the node features and relationship features corresponding to the central node and its neighboring nodes based on the attention mechanism to obtain the topological structure features of the central entity; The updating process of the topological structure feature of the central entity includes: Initial embedding of With its structural level representation and relationship hierarchy representation Merge and update entity embedding for the next iteration , until the iteration is completed; its expression is: ； in, Represents the central entity In the l The structural level representation in the +1 iteration is updated by iterative multi-layer aggregation from neighboring entities; its expression is: in, Represents the tail entity In the l Features in iterations; Represents the central entity A dataset of neighboring entities, Represents the tail entity of the adjacent entity and relationship entities ; Indicated in l The trainable weight matrix in iterations; is the activation function; Indicated in l Structure-aware attention weights in iterations; Represents the tail entity of another adjacent entity and relationship entities ; Represents the tail entity In the l Features in iterations; Representing Entities In the l The relational hierarchy representation in the +1 iteration calculates the attention score by aggregating all the connection relations of the central entity and uses it for message passing, which is updated by the sum of weighted edge embeddings; its expression is: in, Representing relationship entities In the l Display features in iterations; Indicated in l The trainable weight matrix in iterations is used to linearly transform the aggregated neighbor embeddings; Indicated in l Relation-aware attention weights in iterations.

7. The compound-target interaction prediction method according to claim 1, characterized in that: The compound-target prediction probability matrix is ​​obtained by the knowledge graph embedding method, including: The DistMult semantic matching model is used to generate a prediction probability for each possible triple in the knowledge graph and generate a compound-target prediction probability matrix.

8. The method for predicting compound-target interactions according to any one of claims 1 to 7, characterized in that: The method further comprises: The fact triples in the knowledge graph are regarded as positive sample data, all undiscovered triples are regarded as sources of negative sampling, and damaged triples are constructed as negative sample data according to different negative sample-positive sample ratios to form a training data set; The training data set is input into the multimodal feature extraction model for cross-validation evaluation training, and the binary cross entropy loss function is minimized using the gradient descent method. ; Its expression is: in, is the scoring function used to measure the rationality of the ternary combination; are the head entity, relation entity and tail entity in the positive sample triple, represents the positive sample set; Represents the true label of the positive sample triple; is the tail entity in the negative sample triple, Represents the set of negative samples.

9. The compound-target interaction prediction method according to claim 8, characterized in that: The method of constructing damaged triples as negative sample data according to different negative sample-positive sample ratios includes the following steps: According to the preset negative sample-positive sample ratio, some randomly selected positive sample triplets are The tail entity in Replace with another entity Generate negative sample triplets ; For positive sample triples And negative sample triplets Labeled as 1 and 0 respectively.

10. A compound-target interaction prediction system, using the compound-target interaction prediction method according to any one of claims 1 to 9, characterized in that: include: The knowledge graph module is used to construct a knowledge graph containing compound, target and disease entities and their interrelationships; A multimodal feature extraction module, which is equipped with a multimodal feature extraction model for extracting topological information in the global knowledge graph and intrinsic semantic information of local entities from the input knowledge graph data; wherein the multimodal feature extraction model includes an ICD2Vec layer, a graph transformer, a gene ontology embedding layer, and a graph neural network based on an attention mechanism, which are respectively used to extract disease semantic features, compound semantic features, protein target semantic features, and topological features of the knowledge graph; A prediction module, which is used to fuse the disease semantic features, compound semantic features, protein target semantic features and topological features of the knowledge graph through a feature fusion module, and obtain a compound-target prediction probability matrix through a knowledge graph embedding method; The output module is used to sort the input compound-target interactions in descending order according to the compound-target prediction probability and output a target recommendation list.

Images