AD compound target prediction method based on knowledge graph

Abstract

The invention discloses an AD (Alzheimer's disease) compound target prediction method based on a knowledge graph. The method comprises the following steps: establishing a disease-syndrome-prescription associated knowledge graph basic structure of Alzheimer's disease; performing information extraction on the text data based on the data model to obtain data features; constructing a knowledge graph of the Alzheimer's disease based on the data features; performing feature extraction on the knowledge graph based on a preset model to obtain graph features; and based on the graph features, adopting an improved restart random walk algorithm to predict in the knowledge graph to obtain a potential target. By using the method, the biological context factors of the drug-target relationship can be comprehensively considered, and the target prediction precision is improved. The method can be widely applied to the field of biological medicine.

Description

Technical Field

[0001] The present invention relates to the field of biomedicine, and in particular to a method for predicting AD compound targets based on a knowledge graph. Background Art

[0002] Alzheimer's disease (AD) is a common neurodegenerative disease that seriously affects patients' cognitive function and quality of life. Currently, there are limited treatments for AD, and drug development faces many challenges.

[0003] With the widespread application of traditional Chinese medicine around the world, especially in the treatment of complex diseases such as AD, Chinese herbal medicines have attracted much attention due to their unique multi-component, multi-target, and multi-pathway regulatory mechanisms. However, this complexity also makes the target prediction and efficacy evaluation of Chinese herbal medicines particularly difficult. Traditional experimental methods are often time-consuming and labor-intensive, and it is difficult to fully reveal the efficacy mechanism of Chinese herbal medicines.

[0004] Traditional drug target prediction methods mainly rely on a single data source, which makes it difficult to fully capture the complex interactions between drugs and targets. Many traditional methods only focus on the direct interactions between drugs and targets, while ignoring the biological context in which these interactions occur, such as cellular environment, signaling pathways, molecular functions, etc., resulting in low prediction accuracy. Summary of the invention

[0005] In view of this, in order to solve the technical problem that the existing drug target prediction methods fail to fully consider the biological context factors of the drug-target relationship, thereby resulting in low prediction accuracy, the present invention proposes a method for predicting AD compound targets based on a knowledge graph, the method comprising the following steps:

[0006] Extract information from text data based on the data model, extract target nodes, relationships, attributes and other information, and obtain data features;

[0007] Construct the basic structure of the disease-syndrome-prescription association knowledge graph of Alzheimer's disease, and construct the knowledge graph in combination with the data features;

[0008] Extract features from the knowledge graph based on the preset model to capture the complex interactions between "drug-target" and obtain graph features;

[0009] Based on the graph features, an improved restarted random walk algorithm is used to perform predictions in the knowledge graph, and potential targets are obtained according to node scores.

[0010] In some embodiments, in the step of extracting features from the knowledge graph based on a preset model to capture the complex interactions between “drugs and targets” and obtain graph features, it specifically includes:

[0011] Encode the entity relationships in the knowledge graph and embed the knowledge graph into a machine-readable vector form;

[0012] The LLM model is used to score the "gene" (target) similarity in the embedded knowledge graph, the "protein interaction association" strength score between "genes", and the "traditional Chinese medicine ingredient-gene" affinity, and the scoring results are output in the form of a matrix;

[0013] At the same time, Chinese medicine components are divided into multiple communities.

[0014] In some embodiments, based on the graph features, the improved restarted random walk algorithm is used to predict the knowledge graph, and the step of obtaining potential targets according to the node scores specifically includes:

[0015] Perform random walk simulation on the constructed knowledge graph, select the starting node of the improved restart random walk algorithm, and initialize relevant parameters;

[0016] Based on the relationship and transition probability between community nodes after knowledge graph division, a transfer matrix is ​​constructed. The transfer matrix is ​​used to reflect the accuracy of node relationships and to simulate the transfer pattern between nodes.

[0017] Using the constructed transfer matrix, random walks are performed according to the set transfer probability and restart probability to simulate the node transfer process and explore the network structure, record the number of node visits, and obtain the steady-state distribution results;

[0018] Node scores were calculated based on the steady-state distribution results, and potential targets were predicted based on the node scores.

[0019] Based on the above scheme, the present invention provides a method for predicting AD compound targets based on knowledge graphs, which integrates multi-source data and uses large models, knowledge graphs and improved random walk algorithms to achieve efficient prediction of AD compound targets. Compared with traditional methods, the present invention integrates data from a more comprehensive perspective, not only taking into account the drug-target relationship, but also its biological context factors, and can more comprehensively capture the complex interactions between drugs and targets; at the same time, it uses visualization to enhance its interpretability, which can be better applied to the biomedical field. The present invention provides a new perspective and strategy for the prediction method of AD compound targets. BRIEF DESCRIPTION OF THE DRAWINGS

[0020] Figure 1 It is an implementation flow chart of a method for predicting AD compound target based on knowledge graph of the present invention;

[0021] Figure 2 This is a flow chart of the steps of the AD compound target prediction method based on the knowledge graph of the present invention.

[0022] Figure 3 It is a schematic diagram of the process of extracting information and constructing a knowledge graph in the present invention.

[0023] Figure 4 It is a schematic diagram of the relationship between data in the knowledge graph.

[0024] Figure 5 It is a flowchart of graph feature extraction.

[0025] Figure 6 This is a schematic diagram of the improved random walk prediction process. DETAILED DESCRIPTION

[0026] The following will be combined with the drawings in the embodiments of the present application to clearly and completely describe the technical solutions in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, not all of the embodiments. Based on the embodiments in the present application, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of this application.

[0027] It should be noted that, for the convenience of description, only the parts related to the relevant invention are shown in the drawings. In the absence of conflict, the embodiments and features in the embodiments of the present application can be combined with each other.

[0028] It should be understood that the "system", "device", "unit" and / or "module" used in this application is a method for distinguishing different components, elements, parts, portions or assemblies at different levels. However, if other words can achieve the same purpose, the word can be replaced by other expressions.

[0029] As shown in this application and claims, unless the context clearly indicates an exception, the words "a", "an", "a kind" and / or "the" do not refer to the singular, but also include the plural. Generally speaking, the terms "include" and "comprise" only indicate the inclusion of clearly identified steps and elements, and these steps and elements do not constitute an exclusive list. The method or device may also include other steps or elements. The elements defined by the sentence "includes a..." do not exclude the existence of other identical elements in the process, method, commodity or device that includes the elements.

[0030] In the description of the embodiments of the present application, "plurality" means two or more than two. The following terms "first" and "second" are used for descriptive purposes only and are not to be understood as indicating or implying relative importance or implicitly indicating the number of the indicated technical features. Thus, the features defined as "first" and "second" may explicitly or implicitly include one or more of the features.

[0031] In addition, flow charts are used in the present application to illustrate the operations performed by the system according to the embodiments of the present application. It should be understood that the preceding or following operations are not necessarily performed accurately in order. On the contrary, the various steps may be processed in reverse order or simultaneously. At the same time, other operations may also be added to these processes, or a certain step or several steps of operations may be removed from these processes.

[0032] Figure 1 This is a flow chart for the implementation of the present invention, including four sections: knowledge modeling, large model text extraction, network feature extraction, and random walk prediction.

[0033] Reference Figure 2 , is a flow chart of an optional example of the AD compound target prediction method based on the knowledge graph proposed in the present invention, which can be applied to a computer device. The prediction method proposed in this embodiment may include but is not limited to the following steps:

[0034] Step S0: Establish the basic structure of the disease-syndrome-prescription association knowledge graph of Alzheimer's disease, with data sources including ancient Chinese medical books, modern medical literature, clinical data, and public gene databases.

[0035] Step S1, extracting information from text data based on a data model to obtain data features;

[0036] Step S2, constructing a knowledge graph of Alzheimer's disease based on the data features;

[0037] Step S3: extracting features from the knowledge graph based on a preset model to obtain graph features;

[0038] Step S4: Based on the graph features, an improved restarted random walk algorithm is used to make predictions in the knowledge graph to obtain potential targets.

[0039] In some feasible embodiments, step S0 specifically includes:

[0040] Construct a knowledge model, and establish the basic structure of the disease-syndrome-prescription association knowledge graph of AD through pharmacological analysis. The data source is shown in Table 1 below; define the structure, nodes and relationships of knowledge, thereby establishing the initial framework of the process and defining the basic structure to lay the foundation for the large model to extract text. The entity structure and relationship structure of the knowledge graph are shown in Tables 2 and 3.

[0041] Table 1 Data source description

[0042]

[0043] Table 2 Knowledge graph entity structure

[0044]

[0045] Table 3 Knowledge graph relationship structure

[0046]

[0047] In some feasible embodiments, step S1 specifically includes:

[0048] The large model prompt project (Prompt) is used to extract key information such as tags, nodes and relationships from large amounts of text data, providing a data foundation for building a knowledge graph.

[0049] S1.1. Prepare basic data sources, extract data from different sources (such as databases, file systems), and aggregate data sources. The data sources contain a large number of triple relationships.

[0050] S1.2, prepare for information extraction from the big model by calling the corresponding big model API;

[0051] S1.3. Store the collected data into the selected storage system for subsequent use.

[0052] S1.4. Fine-tune the big model through corresponding design and specific optimization tips, train the big model to generate output that meets the requirements, and effectively extract and integrate Chinese medicine compound, target and other relevant information;

[0053] S1.5, Prompt conducts multiple rounds of self-optimization learning, cyclically optimizes the information extraction results, adjusts the output information, and obtains the final result.

[0054] Reference Figure 3 , a schematic diagram of the process of data extraction and knowledge graph construction.

[0055] In some feasible embodiments, step S2 specifically includes:

[0056] S2.1. The extracted effective information is restructured and restructured using the big model to lay the foundation for the construction of the knowledge graph.

[0057] Use text mining, cleaning and feature extraction techniques.

[0058] S2.2. Build a knowledge graph based on the processed data and establish entity relationships in the graph. Add attributes to entities to enhance the detailed information of the graph.

[0059] S2.3. Periodically verify and update data through large models, repeatedly optimize the map, and improve its accuracy and application efficiency. It can be executed independently or run regularly to keep the knowledge base accurate and up-to-date.

[0060] Reference diagram of the relationship between data in the knowledge graph Figure 4 .

[0061] In some feasible embodiments, step S3 specifically includes:

[0062] See also Figure 5 ,Network graph feature extraction achieved with the help of large models.

[0063] S3.1, query node and relationship information from Neo4j and embed entities;

[0064] include:

[0065] S3.1.1、Query data;

[0066] S3.1.2, node embedding;

[0067] ① Traditional Chinese medicine ingredients, construct Morgan fingerprint embedding through SMILES attributes:

[0068] z Morgan =MorganFingerprint(SMILES,radius=2,nBits=2048)

[0069] Among them, radius is a parameter used when constructing the Morgan fingerprint, which determines the range of the atomic neighborhood to be considered. A radius of 2 means that the relationship between the atom and its two "level" neighbors is considered. nBits is the number of bits in the Morgan fingerprint, which determines the length of the fingerprint vector. 2048 bits means that the generated Morgan fingerprint is a 2048-dimensional binary vector.

[0070] ②Gene, ESM-2 is a dedicated protein sequence pre-training model that can generate high-dimensional embeddings (650M model outputs 1280-dimensional embeddings), protein embedding based on gene sequences:

[0071] z Gene =ESM-2(Sequence)

[0072] Sequence is the input gene sequence, which can be any protein or gene sequence. ESM-2 converts the gene sequence into a high-dimensional vector representation (e.g., 1280 dimensions), which captures the biological function and structural information of the gene.

[0073] ③Molecular functions, biological processes, and cellular components are embedded in Gene Ontology (GO) nodes using Onto2Vec:

[0074] Z GO =Onto2Vec(GO-Annotations)

[0075] Among them, G0-Annotations is the input gene annotation information, which describes the function, process or cell location of the gene. The Onto2Vec model converts these annotations into embedding vectors to help learn the similarities between GO annotations.

[0076] ④AD endophenotypes, traditional Chinese medicine, and AD prescriptions, using BioBERT to embed descriptive text information:

[0077] z Entity =BioBERT(Description)

[0078] Among them, Description is the input descriptive text, which contains descriptive information related to drugs, diseases, genes, etc. The BioBERT model converts the text into a vector that can capture professional information and relationships in the biomedical field, and the output embedding vector is the feature representation used for subsequent tasks.

[0079] S3.2. TransE embeds all relations, including embedding assumptions and optimization objectives.

[0080] S3.3. Construct multimodal input to the large model Llama.

[0081] S3.4, input embedding mapping, uniformly mapping multimodal input embeddings to LLama embedding space.

[0082] S3.5. Use the LLM model to score the preset indicators in the embedded knowledge graph to obtain a scoring result. The preset indicators include gene similarity, protein interaction strength between genes, and affinity between Chinese medicine ingredients and genes.

[0083] Gene similarity calculation:

[0084] Compute gene similarity based on embeddings of molecular functions, biological processes, and cellular components:

[0085]

[0086] in, Indicates gene g i and g j These embeddings are generated by large models (such as ESM-2, Onto2Vec) and represent information such as gene function and structure. Gene g i The embedding vector captures the biological characteristics of the gene. Gene g j , capturing the biological characteristics of the gene. )

[0087] Strength of protein interactions between genes:

[0088] w′(g i , g j )=w(g i , g j )+β·Sim(g i , g j )

[0089] Among them, w(g i , g j ) is gene g i and g j The original protein interaction strength between them. β is a regulatory factor that controls the similarity Sim(g i , g j ) increases the intensity of the interaction.

[0090] Affinity between Chinese herbal medicine ingredients and genes:

[0091] Through the affinity values ​​on the existing paths, multimodal features are constructed, and the affinities of all paths are completed through the large model:

[0092]

[0093] in, is the predicted affinity value, which indicates the strength of interaction between the TCM components and the gene (e.g., the binding affinity of the drug to the target gene). Drug is the embedding vector of the Chinese medicine ingredients. Gene is the embedding vector of the gene. Relation Embeddings representing the relationships between TCM ingredients and genes.

[0094] The fine-tuning objective is the mean square error loss:

[0095]

[0096] in, is the affinity value predicted by the model. i is the true affinity value. That is, the affinity value between the gene and the Chinese herbal medicine component in the known experimental data. N: the total number of samples, that is, the number of samples with known Chinese herbal medicine component-target affinity in the data set.

[0097] Also included: Dividing Chinese medicine ingredient communities by K-Means.

[0098] In some feasible embodiments, step S4 specifically includes:

[0099] Reference Figure 6 , which is a process diagram of the random walk simulation algorithm.

[0100] S4.1. Perform random walk simulation on the constructed AD disease-syndrome-prescription association knowledge graph, select the starting node for restarting the random walk algorithm, and initialize relevant parameters.

[0101] The disease-syndrome-prescription association knowledge graph of AD can be represented by the set G = [G1, G1, ..., G d ], where G is a multi-relation network composed of multiple relational sub-networks, G = [V, E], where V represents a set of nodes and E represents a set of different types of edges. i , G j The edge between ∈V is denoted as (G i , G j ), node G i The set of neighbor nodes is denoted as Γ(i), G i The degree is denoted as k i .

[0102] S4.2, construct the transfer matrix based on the relationship and transition probability between community nodes after knowledge graph division;

[0103] Based on MHRW (the unbiased sampling algorithm MHRW proposed by Gjoka et al. in combination with RW), an improved algorithm MHRWER (Metropolis-Hasting Random Walk with Extended Restart, MHRWER) is proposed. While considering the influence of the degree of the current node and its neighboring nodes on the transition probability, path weights are assigned to neighboring nodes based on the path access history, and the global importance of global nodes is considered. By dynamically adjusting the path weights, the ability to explore unknown areas is enhanced while reducing repeated visits to known areas.

[0104] The transition probability defined by MHRWER is:

[0105]

[0106] Among them, k i represents the degree of node i (i.e., the number of edges adjacent to node i), which is used to measure the connectivity of the node. j is a node label, indicating that in the random walk process, node j is one of the possible target nodes to which node i may transfer, and F(i) represents the neighbor set of node i. ij is the path weight, and the calculation formula is as follows:

[0107]

[0108] x ij is the path access frequency from node i to node j, π ijIt represents the affinity or protein interaction strength between Chinese medicine ingredient i and gene j (corresponding to the result calculated in the previous step according to different pathway conditions).

[0109] S4.3. Based on the transfer matrix, random walk is performed according to the transfer probability and the restart probability to simulate the node transfer process and explore the network structure, and the number of node visits is recorded.

[0110] RWR (Random Walk with Restart) is a basic random walk similarity metric. Assume that the probability of a particle randomly selecting the next neighbor node to transfer is α, and the probability of returning to the initial node is 1-α. Based on this assumption, the network transition probability matrix P can be defined, where the elements of the matrix P ij represents the probability of transferring from node i to node j. The probability vector of particles transferring to other nodes in the next step is as follows:

[0111]

[0112] in, represents the access probability vector of node i. c represents the restart probability. represents the weighted transfer matrix, which represents the probability of transferring from one node to other nodes. Represents the access probability vector of the current node. is a unit vector representing the probability distribution of the initial state.

[0113] Iteratively update w ij , dynamically adjust the node transfer probability.

[0114] S4.4. Calculate the steady-state distribution, that is, the stable probability distribution of the nodes.

[0115] Steady-state distribution calculation formula

[0116] y=(1-μ)(I-μM) -1 y0

[0117] Among them, μ is the restart parameter, which represents the probability of returning to the initial node from the current state. y0 is the initial probability distribution. The transfer matrix M is defined as follows:

[0118]

[0119] W is the weight matrix adjusted based on path weight and global importance, and its element is w ij A is the adjacency matrix, which represents the direct connection relationship between nodes. D is the diagonal matrix, which is defined as:

[0120]

[0121] This ensures that M is a normalized transition probability matrix.

[0122] During the iteration process, the transfer probability matrix M is gradually updated to determine whether the convergence condition of the algorithm or the preset maximum number of iterations is reached. If so, the final stationary distribution probability y is calculated. Otherwise, the calculation of the path weight is returned for iterative update.

[0123] Apply steady-state distribution to actual knowledge graph analysis to identify important nodes or analyze knowledge graph features.

[0124] S4.5. Based on the steady-state distribution, we deeply analyze the network characteristics, calculate the score of each node, and evaluate its possibility as a potential target.

[0125] According to the score sorting, high-potential target nodes are screened out.

[0126] In some feasible embodiments, it also includes:

[0127] With the help of visualization tools, random walk paths are displayed and potential target probabilities are visualized through heat maps, etc., in order to explain the results and enable them to play a better role in the biomedical field.

[0128] A knowledge graph-based AD compound target prediction system, comprising:

[0129] The text processing module extracts information from text data based on the data model to obtain data features;

[0130] A graph construction module, which constructs a knowledge graph of Alzheimer's disease based on the data features;

[0131] A feature extraction module extracts features from the knowledge graph based on a preset model to obtain graph features;

[0132] The random walk module uses an improved restart random walk algorithm to make predictions in the knowledge graph based on the graph features to obtain potential targets.

[0133] The contents of the above method embodiments are all applicable to the present system embodiments. The functions specifically implemented by the present system embodiments are the same as those of the above method embodiments, and the beneficial effects achieved are also the same as those achieved by the above method embodiments.

[0134] A knowledge graph-based AD compound target prediction device:

[0135] at least one processor;

[0136] at least one memory for storing at least one program;

[0137] When the at least one program is executed by the at least one processor, the at least one processor implements the AD compound target prediction method based on the knowledge graph as described above.

[0138] The contents of the above method embodiments are all applicable to the present device embodiments. The functions specifically implemented by the present device embodiments are the same as those of the above method embodiments, and the beneficial effects achieved are also the same as those achieved by the above method embodiments.

[0139] A storage medium stores processor-executable instructions, which, when executed by the processor, are used to implement a knowledge graph-based AD compound target prediction method as described above.

[0140] The contents of the above method embodiments are all applicable to the present storage medium embodiments. The functions specifically implemented by the present storage medium embodiments are the same as those of the above method embodiments, and the beneficial effects achieved are also the same as those achieved by the above method embodiments.

[0141] The above is a specific description of the preferred implementation of the present invention, but the invention is not limited to the embodiments. Those skilled in the art may make various equivalent modifications or substitutions without violating the spirit of the present invention. These equivalent modifications or substitutions are all included in the scope defined by the claims of this application.

Claims

1. A method for predicting AD compound targets based on knowledge graph, characterized in that: The following steps are involved: Extract information from text data based on the data model to obtain data features; Constructing a knowledge graph of Alzheimer's disease based on the data features; Extract features from the knowledge graph based on a preset model to obtain graph features; Based on the graph features, an improved restarted random walk algorithm is used to make predictions in the knowledge graph to obtain potential targets.

2. According to claim 1, a method for predicting AD compound targets based on knowledge graph, characterized in that: Also includes: The random walk paths and the probabilities of potential targets are displayed.

3. The AD compound target prediction method based on knowledge graph according to claim 1, characterized in that: The step of extracting information from text data based on the data model to obtain data features specifically includes: Get text data from the data source; Call and set the data model API; Extracting information from the text data based on the data model to obtain data features; Storing the data features in a graph database; The data features include nodes, relationships and attributes.

4. The AD compound target prediction method based on knowledge graph according to claim 1, characterized in that: The step of extracting features from the knowledge graph based on a preset model to obtain graph features specifically includes: Encoding the entity relationships in the knowledge graph and embedding them into vector form; Use the LLM model to score the preset indicators in the embedded knowledge graph and obtain the scoring results; The preset indicators include gene similarity, protein interaction strength between genes and affinity between Chinese medicine ingredients and genes.

5. The AD compound target prediction method based on knowledge graph according to claim 4, characterized in that: Also includes: Divide communities into Chinese herbal medicine ingredients.

6. The AD compound target prediction method based on knowledge graph according to claim 1, characterized in that: The step of using an improved restarted random walk algorithm to predict the knowledge graph based on the graph features to obtain potential targets specifically includes: A random walk simulation is performed on the knowledge graph, a starting node is selected and relevant parameters are initialized. Construct a transition matrix based on the relationship and transition probability between community nodes after knowledge graph division; Based on the transfer matrix, random walk is performed according to the transfer probability and the restart probability to simulate the node transfer process and explore the network structure, record the number of node visits, and generate a steady-state distribution result; A node score is calculated based on the steady-state distribution result, and a potential target is predicted according to the node score.

7. The AD compound target prediction method based on knowledge graph according to claim 6, characterized in that: The formula of the transition probability is expressed as follows: Among them, w ij is the path weight; k i represents the degree of node i; G j is a node label, indicating that in the random walk process, node j is one of the possible target nodes when starting from node i; Γ(i) represents the neighbor set of node i.

8. A knowledge graph-based AD compound target prediction system, characterized in that: include: The text processing module extracts information from text data based on the data model to obtain data features; A graph construction module, which constructs a knowledge graph of Alzheimer's disease based on the data features; A feature extraction module extracts features from the knowledge graph based on a preset model to obtain graph features; The random walk module uses an improved restart random walk algorithm to make predictions in the knowledge graph based on the graph features to obtain potential targets.

9. A knowledge graph-based AD compound target prediction device, characterized in that: include: at least one processor; at least one memory for storing at least one program; When the at least one program is executed by the at least one processor, the at least one processor implements the AD compound target prediction method based on knowledge graph as described in any one of claims 1 to 7.

Images