Device, method, system and memory for constructing biomedical knowledge graph
By extracting and loading the characteristic attributes of diseases, drugs and genes in the biomedical knowledge graph, and using the embedded representation of the multi-head attention mechanism, the problem of insufficient characteristic attributes of the entity is solved, significantly improving the downstream task effect and accuracy of the knowledge graph.
Patent Information
- Application Number
- CN202010503855.1
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2020-06-05
- Publication Date
- 2024-11-26
- Estimated Expiration
- 2040-06-05
AI Technical Summary
Due to the lack of entity characteristic attributes, the biomedical knowledge graph limits its wide application in the computer field and the experimental effect of downstream tasks.
By extracting and loading the characteristic attributes of diseases, drugs and gene entities, using tools such as BioBERT, RDKit and Connectivity Map, semantic features, chemical properties and gene expression characteristics, conduct principal component analysis and embedded representation of multi-head attention mechanisms to improve knowledge The quality of the map and the downstream task effect.
It effectively improves the downstream task effect of the biomedical knowledge graph, especially in the inter-entity link prediction task, improves the experimental effect and accuracy of the model, and maximizes the use of data value.
Smart Images

Figure CN111782818B_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the fields of chemical informatics and bioinformatics. Specifically, the present invention relates to a device, method, system and memory for constructing a biomedical knowledge graph. Background Art
[0002] Biomedical knowledge graphs are usually directed graphs with multiple relations, where nodes represent entities and edges represent relations. This kind of graph-structured knowledge base has become a resource of great value with potential application value, contributing to precision medicine and clinical decision support, and playing an increasingly important role in healthcare practice and research. However, due to the complexity of downstream tasks and high knowledge requirements, there are still many challenges in construction and application.
[0003] The task of predicting the link between entities is currently the most widely used downstream task in biomedical knowledge graphs. It can be used for analysis and research in multiple fields. For example, the prediction of the link between drugs and diseases can be used to study the new uses of old drugs in the pharmaceutical field; the prediction of the link between drugs can be used to study the combined effects of drugs; the prediction of the link between diseases and genes can be used to study the pathogenesis of pathogenic genes, etc., which are extremely important research topics in the field of biochemistry. Although the emergence of biomedical knowledge graphs provides a new research data for these research topics, the knowledge graphs constructed from biomedical data are usually large and noisy, and the experimental results of these downstream tasks are often unsatisfactory and reach a bottleneck due to data limitations.
[0004] Therefore, in recent years, researchers have been working on how to improve the experimental results of downstream tasks of knowledge graphs and how to maximize the value of existing data to obtain the expected experimental results. With the emergence of deep learning methods, machine learning has attracted more and more attention from researchers. Data-driven analysis has become a routine procedure in many studies, including the construction and application of biomedical knowledge graphs. In many such research applications, machine learning has shown strong potential to compete with or even surpass traditional methods.
[0005] Many researchers have conducted in-depth research on how to build high-quality biomedical knowledge graphs. For example, Percha and his colleagues used the core parser of natural language processing at Stanford University to generate a dependency tree of grammatical structure for each article abstract in the Biomedical Information Retrieval System (PubMed), and built a systematic biomedical knowledge graph based on this dependency tree. Similarly, Himmelstein and his colleagues also built a heterogeneous network called HetioNet by encoding and parsing the knowledge of millions of biomedical studies. These studies make it possible to have biomedical knowledge graphs with multiple types of relationships. However, the current limited embedding techniques for nodes and relationships and the extraction of related domain knowledge limit the widespread application of these huge biomedical knowledge graphs in the computer field, thereby affecting the experimental results of downstream tasks of the knowledge graph.
[0006] At present, the application of knowledge graphs in the computer field is quite extensive, and is commonly seen in search engines, recommendation systems, and question-answering systems. The most important technology is knowledge representation learning. Through learning technologies represented by deep learning, the information of entities is represented as dense low-dimensional real-valued vectors, and then the entities, relationships, and complex associations between them are efficiently calculated in low-dimensional space. There are usually knowledge representation learning based on translation models (Trans series), knowledge representation learning based on matrix decomposition models, such as RESACL, and knowledge representation learning based on neural network models. These knowledge representation learning methods all use certain modeling methods to represent entities and vectors in low-dimensional dense vector space, and then perform calculations and reasoning. Therefore, although they have been successfully applied to knowledge graph tasks such as question answering, information extraction, and named entity ambiguity elimination outside the biomedical field, their use in biomedical knowledge graph applications is still limited by the lack of characteristic attributes of entities. Summary of the invention
[0007] To solve the above problems, the present invention is based on integrating the entity synonyms of various knowledge bases to clean up the disease and gene entities in any knowledge graph to improve the quality of the graph and improve the experimental effects of models and methods in downstream task applications.
[0008] According to one embodiment of the present invention, a method for constructing a biomedical knowledge graph is provided, comprising:
[0009] Extract all disease, drug, and gene entities from the knowledge graph;
[0010] Obtaining feature attributes for the extracted entities; and
[0011] Load the obtained feature attributes into the knowledge graph.
[0012] In one embodiment of the present invention, extracting all disease, drug and gene entities from the knowledge graph includes:
[0013] For disease entities, all disease entities that meet the standards are extracted by matching the disease names in the knowledge graph with the vocabulary, and standardized disease names and disease categories are provided for the disease entities;
[0014] For drug entities, obtain the name and SMILES sequence of the corresponding drug;
[0015] For gene entities, the genes with expression levels in the Connectivity Map and BioGPS databases were used as the vocabulary to extract all gene entities that met the standards in the map.
[0016] In one embodiment of the present invention, obtaining characteristic attributes for the extracted entity includes:
[0017] For disease entities, disease feature attributes with semantic features are extracted;
[0018] For drug entities, extracting drug characteristic attributes with drug characteristics; and
[0019] For gene entities, gene characteristic attributes that describe the expression status are extracted.
[0020] In one embodiment of the present invention, for a drug entity, the chemical properties, physical properties, chemical descriptors and molecular fingerprints of the drug are obtained through the SMILES sequence corresponding to the drug using the chemical open source toolkit RDKit as characteristic attributes of the drug entity.
[0021] In one embodiment of the present invention, loading the obtained characteristic attributes into the knowledge graph includes: performing dimensionality reduction processing on the characteristic attribute matrices of different entities respectively through principal component analysis (PCA) technology to obtain a characteristic attribute matrix with a unified dimension.
[0022] In one embodiment of the present invention, loading the obtained feature attributes into the knowledge graph includes:
[0023] The disease feature attribute matrix is X D ; The drug feature attribute matrix is X C ; The characteristic attribute matrix of gene entity is X G ; For these different entities, the characteristic attribute matrix X D , X C , X G , firstly, de-mean the values to obtain X′ D , X′ C , X′ G, and calculate the covariance matrix C D , C C , C G ;
[0024] Calculate the eigenvalue λ and corresponding eigenvector of the covariance matrix
[0025] Sort the eigenvalues in descending order, select the largest k of them, and use their corresponding k eigenvectors as column vectors to form an eigenvector matrix P, which is used as the input feature attribute matrix Y. D , Y C , Y G , project the entity feature attributes onto the selected feature vector to obtain a feature attribute matrix Y with a unified dimension of N×K D , Y C , Y G , to be loaded into the knowledge graph as input.
[0026] According to another embodiment of the present invention, a device for constructing a biomedical knowledge graph is provided, the device comprising:
[0027] processor;
[0028] a memory for storing executable instructions for the processor;
[0029] Wherein, the processor is configured to execute the steps of the above method.
[0030] According to another embodiment of the present invention, a non-transitory computer-readable storage medium is provided, on which a computer program is stored, and when the computer program is executed by a processor, the steps of the above method are implemented.
[0031] According to another embodiment of the present invention, a method for performing an inter-entity link prediction task using a biomedical knowledge graph constructed by the above method is provided, comprising:
[0032] Receive the feature attribute matrix of all entities in the knowledge graph;
[0033] The GAT formula with multi-head attention mechanism collects information from neighbor entities and learns the characteristics of the relationship between entities;
[0034] Obtain relative attention values based on the learned features of the relationship between entities;
[0035] According to the characteristics of the relationship between entities and the relative attention value, the embedded representation of the entity after the multi-head attention update is obtained, thereby obtaining the embedded representation of the updated node and relationship.
[0036] According to another embodiment of the present invention, a system for constructing a biomedical knowledge graph is provided, comprising:
[0037] An entity extraction unit, which extracts all disease, drug and gene entities from the knowledge graph;
[0038] a characteristic attribute acquisition unit, wherein the characteristic attribute acquisition unit acquires characteristic attributes for the extracted entity; and
[0039] A feature attribute loading unit, wherein the feature attribute loading unit loads the obtained feature attributes into the knowledge graph. BRIEF DESCRIPTION OF THE DRAWINGS
[0040] In order to further illustrate the above and other advantages and features of various embodiments of the present invention, a more specific description of various embodiments of the present invention will be presented with reference to the accompanying drawings. It will be understood that these drawings only depict typical embodiments of the present invention and are therefore not to be considered as limiting the scope thereof. In the accompanying drawings, for clarity, identical or corresponding parts will be represented by identical or similar reference numerals.
[0041] Figure 1 A flowchart showing a method for constructing a biomedical knowledge graph according to an embodiment of the present invention.
[0042] Figure 2 A schematic diagram showing data evolution in the construction of a biomedical knowledge graph according to an embodiment of the present invention.
[0043] Figure 3 A schematic block diagram of a device for constructing a biomedical knowledge graph according to an embodiment of the present invention is shown. DETAILED DESCRIPTION
[0044] In the following description, the present invention is described with reference to various embodiments. However, those skilled in the art will recognize that various embodiments may be implemented without one or more specific details or with other replacement and / or additional methods, materials or components. In other cases, well-known structures, materials or operations are not shown or described in detail to avoid obscuring aspects of various embodiments of the present invention. Similarly, for the purpose of explanation, specific quantities, materials and configurations are set forth to provide a comprehensive understanding of embodiments of the present invention. However, the present invention may be implemented without specific details. In addition, it should be understood that the various embodiments shown in the drawings are illustrative representations and are not necessarily drawn to scale.
[0045] In this specification, reference to "one embodiment" or "the embodiment" means that a particular feature, structure, or characteristic described in conjunction with the embodiment is included in at least one embodiment of the present invention. The phrase "in one embodiment" appearing in various places in this specification does not necessarily all refer to the same embodiment.
[0046] The present invention provides characteristic attributes with biochemical significance for all diseases, drugs and gene entities in the biomedical knowledge graph, improves the method effect of downstream tasks (such as prediction of inter-entity links) of the knowledge graph, and includes the steps of: extracting all diseases, drugs and gene entities in the knowledge graph. Extracting characteristic attributes with biochemical significance for all entities based on background knowledge or public databases. For disease entities, the deep learning model BioBERT is used to perform semantic analysis on the disease name to obtain the semantic characteristic attributes of the disease entity; for drug entities, the molecular SMILES sequence corresponding to the drug entity is used to obtain the characteristic attributes such as the chemical properties, chemical descriptors and molecular fingerprints of the drug through the open source chemical information tool RDKit; for gene entities, the public database Connectivity Map and BioGPS are used to obtain the gene expression amount of the gene entity in different tissues or samples as the characteristic attributes of the gene entity. Finally, the obtained entity characteristic attributes are loaded into the knowledge graph, and downstream tasks such as inter-entity link prediction tasks are performed. By comparing the experimental effects before and after adding the entity characteristic attributes, it is proved that the entity characteristic attributes can effectively improve the effect of downstream tasks in the knowledge graph.
[0047] The method of the present invention is suitable for extracting characteristic attributes of diseases, drugs and gene entities in any knowledge graph, and can be applied to downstream tasks to improve the experimental effects of models and methods.
[0048] Figure 1 A flowchart showing a method for constructing a biomedical knowledge graph according to an embodiment of the present invention. Figure 2 A schematic diagram of data evolution in the construction of a biomedical knowledge graph according to an embodiment of the present invention is shown below. Figure 1 and Figure 2 , a method for constructing a biomedical knowledge graph according to an embodiment of the present invention is described in detail.
[0049] First, in step 110, all disease, drug and gene entities are extracted from the knowledge graph. Figure 2 As shown, the current knowledge graph is the existing knowledge graph.
[0050] In one embodiment of the present invention, the knowledge graph currently used is a biomedical knowledge graph, including two data sets, one of which is commercial data: Pharmaceutical Knowledge Graph Embedding (PharmKG), which comes from Aladdin Healthcare Technologies Ltd.; and the biomedical knowledge graph HetioNet disclosed in the document "Systematic integration of biomedical knowledge prioritizes drugs for repurposing." Elife 6 (2017): e26726. The detailed information of the data set is as follows:
[0051] Data name Number of genetic entities Number of drug entities Number of disease entities Number of relationships PharmKG 5811 41250 4172 721493 HetioNet 18200 1450 126 213512
[0052] Table 1 Basic information of two knowledge graphs
[0053] For disease entities, the diseases in the Disease Ontology and Medical Subject Headings (MeSH) are used as the vocabulary. The disease names in the knowledge graph are matched with the vocabulary, including simulated matching and exact matching, to extract all disease entities that meet the standards, and standardized disease names and disease categories are provided for these entities.
[0054] For drug entities, the API program PyPubChem is used to search for the corresponding drug in the PubChem database based on the drug name. The name and SMILES sequence of the corresponding drug are obtained with the help of the information in the PubChem database to describe the characteristic attributes of the drug.
[0055] For gene entities, we first constructed a gene expression matrix consisting of the Connectivity Map and BioGPS databases, which include the expression of different genes in the genome in different tissues or cells. Using these genes with expression as the vocabulary, we can extract all gene entities that meet the standards from the knowledge graph. Each gene entity contains the gene expression levels of 1515 different tissues or cells, which can constitute the characteristic attributes of the gene entity.
[0056] In step 120, feature attributes are obtained for the extracted entities. In an embodiment of the present invention, feature attributes with biochemical significance can be extracted for all entities based on background knowledge or public databases. In other words, based on all disease, drug, and gene entities extracted from the data set, feature attributes with domain knowledge are extracted for these entities through background knowledge or public databases.
[0057] Specifically, the three types of entities extracted: disease, drug and gene, are extracted through different methods to extract disease feature attributes with semantic features, drug feature attributes with drug properties, and gene feature attributes with expression description. Based on the above content, the data processing flow will be described in detail below.
[0058] First, extract the disease characteristic attributes with semantic features. The present invention analyzes the similarities and differences of all diseases and believes that the characteristic information contained in the disease is in its disease name, that is, semantic information. Therefore, the present invention uses a natural language processing method to extract semantic features from the name of each disease through a pre-trained word vector model, and uses the dense semantic vector as a characteristic attribute to characterize the disease. In one embodiment of the present invention, the present invention can use BioBERT, a pre-trained biomedical language representation model for biomedical text mining. The model can better extract the semantic information of the disease through text mining by using a domain-specific language representation model pre-trained on a large-scale biomedical corpus. The specific method is: through the pre-trained weights of BioBERT provided on https: / / github.com / naver / biobert-pretrained, the extracted disease names of all disease entities are used as input to fine-tune BioBERT, thereby training the characteristic attributes of the disease entity with semantic features. Through the natural language processing method in deep learning and the pre-trained word vectors in the biomedical field in BioBERT, semantic features of the disease are extracted, and different semantic features are given to each entity to ensure that the characteristic attributes of similar diseases are close in Euclidean space.
[0059] Secondly, extract the characteristic attributes of the drug with drug properties. In one embodiment of the present invention, the present invention uses the API program PyPubChem of the PubChem database to search the extracted drug entities one by one, find the standard name of the corresponding drug and the molecular SMILES sequence corresponding to the drug, and then use the chemical open source toolkit RDKit to extract a series of drug properties such as chemical properties, physical properties, chemical descriptors and molecular fingerprints of the drug using the SMILES sequence as the characteristic attributes of the drug entity. The complete chemical properties extracted from RDKit are shown in the following table.
[0060]
[0061]
[0062] Table 2 Chemical properties and their corresponding detailed descriptions in the RDKit used
[0063] Finally, extract the gene characteristic attributes that describe the expression. In an embodiment of the present invention, the expression of the gene in other different tissues or cells is used as the characteristic attribute of the gene entity, that is, the universal expression characteristics of the gene are used as the characteristics of the gene, which can reflect both the specific expression of the gene and the expression difference between genes. The specific steps are: the expression of different genes in the genome in different tissues or cells is collected through the database Connectivity Map and BioGPS, with 1323 samples and 192 samples respectively; for the gene entity extracted in step 110, find the corresponding gene expression as the characteristic attribute of the gene entity.
[0064] In step 130, the obtained feature attributes are loaded into the knowledge graph.
[0065] In an embodiment of the present invention, in order to adapt to the input form of the machine learning model, the characteristic attribute matrices of different entities can be subjected to dimensionality reduction processing respectively through principal component analysis (PCA) technology.
[0066] For example, the disease feature attribute matrix is X D ; The drug feature attribute matrix is X C ; The characteristic attribute matrix of gene entity is X G ; For these different entities, the characteristic attribute matrix X D , X C , X G , firstly, de-mean the values to obtain X′ D , X′ C , X′ G , and calculate the covariance matrix C D , C C , C G , the formula is as follows:
[0067]
[0068] X=(x1,x2,...,x m ) (2)
[0069]
[0070]
[0071] Where N D is the number of disease entities, N C is the number of drug entities, N G is the number of gene entities; Represents dimension N D * 128 matrix; Represents dimension NC * Matrix of 181; Represents dimension N G *1515 matrix.
[0072] Secondly, calculate the eigenvalue λ and the corresponding eigenvector of the covariance matrix
[0073]
[0074] Then, sort the eigenvalues in descending order, select the largest k of them, and then use the corresponding k (k is set to 100 by default) eigenvectors as column vectors to form the eigenvector matrix P as the input feature attribute matrix Y D , Y C , Y G . Project entity feature attributes onto the selected feature vector.
[0075] Y=PX (6)
[0076] Finally, we can get the feature attribute matrix Y with a unified dimension of N×K D , Y C , Y G , to be loaded into the knowledge graph as input to perform downstream tasks.
[0077] In an embodiment of the present invention, the downstream task may be an inter-entity link prediction task.
[0078] After the disease, drug and gene entities are reduced in dimension by PCA, the K-dimensional feature attribute matrix Y can be obtained. D , Y C , Y G , can be loaded into the knowledge graph as input to perform downstream tasks. The task and model applied by the present invention is a knowledge graph relationship prediction method based on the attention mechanism for the task of link prediction between entities. The specific algorithm flow is as follows.
[0079] Knowledge graphs are usually represented by a set of entities and a set of relationships. For example, a triple is composed of two nodes and the edge between them. Therefore, the learning goal of this model is the effective representation of entities and relationships and the scoring function. That is, when a triple t is given as input, the scoring function f(t) can give the probability that t is a true triple.
[0080] First, the model is based on the graph attention network (GAT), with the input being the feature attribute matrix of all entities in the knowledge graph, and the output being the new embedding updated by aggregating the information of neighboring entities. Here, the information of all neighboring entities of each entity in the knowledge graph is assigned different weights to collect information from neighboring entities, and different weights can be assigned according to the different degrees of importance of neighboring nodes. However, it is not enough to simply update the embedding of entities through GAT, because it ignores the features of relationships (edges), which is also the most important part of the knowledge graph. Based on this, the present invention proposes a new embedding method that integrates relationship and neighbor node features in the attention mechanism. The specific formula is derived as follows.
[0081] The feature attribute matrix Y of all entities in a given knowledge graph D , Y C , Y G In this case, the GAT formula using the multi-head attention mechanism collects information from neighbor entities and also needs to learn the characteristics of the relationship between entities:
[0082]
[0083] b ijk =LeakyReLU(W2c ijk ) (8)
[0084] where y i With y j is the characteristic attribute of entity i, j in the knowledge graph, r k is the embedded representation of the relationship between entities i and j in the knowledge graph, W1 and W2 are the parameters trained in the machine learning model, and || is the connection operator.
[0085] According to the learned relationship between entities b ijk Get the relative attention values α ijk :
[0086]
[0087] in is the neighbor entity of entity i, is the relationship between entities i and n, r, b ijk represents entities i, j and the relationship k between them; similarly b inr Represents entities i, n and the relationship r between them
[0088] We can get the embedding representation of the entity after multi-head attention update:
[0089]
[0090] Where m is the number of multi-head attention mechanisms and || is the connection operator.
[0091] Finally, the updated embedded representations of nodes and relationships can be learned, and the ConvKB model is used as a decoder to complete the decoding and prediction process.
[0092] During the training process, the loss function used is
[0093]
[0094] in γ>0 is the parameter of Hinge loss function, S is a valid triplet, and S′ is an invalid triplet.
[0095]
[0096] in W is the parameter for model training.
[0097] In the embodiment of the present invention, the model mainly focuses on the following hyperparameters, and other parameters can refer to the actual code.
[0098] Hyperparameters significance Initial learning rate The initial learning rate of the model during training Embeddings dimention of the final layer The size of the last layer in the model
[0099] Table 3. Hyperparameters of the main adjustments of the model
[0100] The embodiment of the present invention is an inter-entity link prediction task, so the evaluation indicators used are mainly Hit@K and Mean Rank. This model is evaluated under filtering conditions, that is, during the test process, for a group of triples (relationships and entities), the present invention replaces the head entity or the tail entity with any other entity (a total of n-1, keeping the other entity and the relationship unchanged, and only changing one of the entities), so that (n-1) new relationship triplets are obtained, and then the entity relationship distance is calculated for these triplets, and these n-1 triplets are arranged in ascending order according to the distance, so as to calculate Hit@K and MeanRank, whether the real entity can be encountered from the first to the Kth, if encountered, it is a hit, Hit@K represents the ratio of the first K hits, and Mean Rank is calculated in the test set, and the average number of hits to hit the correct result.
[0101] The results of the evaluation are as follows.
[0102] PharmKG Hit@1 Hit@3 Hit@10 Mean Rank Attention-based-Random 0.133 0.245 0.388 0.218 Attention-based-Feature 0.152 0.280 0.424 0.244 HetioNet Hit@1 Hit@3 Hit@10 Mean Rank Attention-based-Random 0.143 0.204 0.280 0.190 Attention-based-Feature 0.165 0.241 0.324 0.210
[0103] Table 4 Model comparison experimental results
[0104] In Table 4, Attention-based-Random represents that the present invention uses randomly generated entity feature attributes in the Attention-based training process, while Attention-based-Feature represents that the present invention uses entity feature attributes with domain knowledge in the Attention-based training process.
[0105] The present invention provides a feature attribute extraction technology for all diseases, drugs and gene entities in the biomedical knowledge graph, which can effectively and extensively improve the application of machine learning models in the biomedical knowledge graph, and can effectively characterize the hidden information of all diseases, drugs and gene entities in the knowledge graph through relevant domain knowledge, and provide powerful assistance for downstream experimental verification, which can maximize the value of the data, and at the same time, the results are more accurate, which is convenient for researchers to continue the next step of experimental verification.
[0106] The present invention proposes to use natural language processing methods to extract the semantic feature attributes of disease entities in the biomedical knowledge graph through the pre-trained word vector model BioBERT to characterize the semantic information of disease entities.
[0107] The present invention proposes to use the chemical open source toolkit RDKit to extract drug characteristic attributes of drug entities in the biomedical knowledge graph, including chemical properties, chemical descriptors and molecular fingerprints, as characteristic attributes of drug entities.
[0108] The present invention proposes to use the public databases Connectivity Map and BioGPS to extract the characteristic attributes of gene entities in the biomedical knowledge graph, characterize the specific expression of gene entities, and also reflect the expression differences between genes.
[0109] The present invention can improve the experimental effect and accuracy of the model by extracting disease characteristic attributes with semantic characteristics, drug characteristic attributes with drug properties, and gene characteristic attributes describing gene expression conditions, and loading them into the machine learning model.
[0110] The present invention is more versatile and can be applied to any biomedical knowledge graph. The extracted entity feature attributes with domain knowledge can also be applied to most machine learning models, and can achieve good improvement effects, making it more convenient for users to use.
[0111] The above-described embodiments may be provided as a computer program product that may include one or more machine-readable media having machine-executable instructions stored thereon, which, when executed by one or more machines such as a computer, a computer network, or other electronic device, may cause the one or more machines to perform operations according to various embodiments of the present invention. The machine-readable medium may include, but is not limited to, a floppy disk, an optical disk, a CD-ROM (compact disk read-only memory) and a magneto-optical disk, a ROM (read-only memory), a RAM (random access memory), an EPROM (erasable programmable read-only memory), an EEPROM (electrically erasable programmable read-only memory), a magnetic or optical card, a flash memory, or other types of media / machine-readable media suitable for storing machine-executable instructions.
[0112] In addition, various embodiments may be downloaded as a computer program product, wherein the program may be transmitted from a remote computer (e.g., a server) to a requesting computer (e.g., a client) via a communication link (e.g., a modem and / or a network connection) using one or more data signals implemented and / or modulated by a carrier wave or other propagation medium. Thus, machine-readable media as used herein may include such a carrier wave, but this is not a requirement.
[0113] Figure 3 A schematic block diagram of a device for constructing a biomedical knowledge graph according to an embodiment of the present invention is shown. Figure 3 As shown, the device for constructing a biomedical knowledge graph includes an entity extraction unit 310 , a feature attribute acquisition unit 320 , and a feature attribute loading unit 330 .
[0114] The entity extraction unit 310 extracts all disease, drug and gene entities from the knowledge graph. The current knowledge graph is an existing knowledge graph.
[0115] In one embodiment of the present invention, the knowledge graph currently used is a biomedical knowledge graph, which includes two data sets, one of which is commercial data: Pharmaceutical Knowledge Graph Embedding (PharmKG), which comes from Aladdin Healthcare Technologies Ltd.; and the biomedical knowledge graph HetioNet disclosed in the document "Systematic integration of biomedical knowledge prioritizes drugs for repurposing." Elife 6 (2017): e26726.
[0116] For disease entities, the diseases in the Disease Ontology and Medical Subject Headings (MeSH) are used as the vocabulary, and the disease names in the knowledge graph are matched with the vocabulary, including simulated matching and exact matching, to extract all disease entities that meet the standards, and provide standardized disease names and disease categories for these entities. For drug entities, according to the name of the drug, the API program PyPubChem is used to search for the corresponding drug in the PubChem database, and the name and SMILES sequence of the corresponding drug are obtained with the help of the information in the PubChem database to describe the characteristic attributes of the drug. For gene entities, a gene expression matrix composed of the Connectivity Map and BioGPS databases is first constructed, which includes the expression of different genes in the genome in different tissues or cells. Using these genes with expression as the vocabulary, all gene entities that meet the standards can be extracted from the knowledge graph.
[0117] The feature attribute acquisition unit 320 acquires feature attributes for the extracted entities. In an embodiment of the present invention, feature attributes with biochemical significance can be extracted for all entities based on background knowledge or public databases. In other words, based on all disease, drug and gene entities extracted from the data set, feature attributes with domain knowledge are extracted for these entities through background knowledge or public databases.
[0118] Specifically, the three types of entities extracted: disease, drug and gene, are extracted through different methods to extract disease feature attributes with semantic features, drug feature attributes with drug properties, and gene feature attributes with expression description. Based on the above content, the data processing flow will be described in detail below.
[0119] First, extract the disease characteristic attributes with semantic features. The present invention analyzes the similarities and differences of all diseases and believes that the characteristic information contained in the disease is in its disease name, that is, semantic information. Therefore, the present invention uses a natural language processing method to extract semantic features from the name of each disease through a pre-trained word vector model, and uses the dense semantic vector as a characteristic attribute to characterize the disease. In one embodiment of the present invention, the present invention can use BioBERT, a pre-trained biomedical language representation model for biomedical text mining. The model can better extract the semantic information of the disease through text mining by using a domain-specific language representation model pre-trained on a large-scale biomedical corpus. The specific method is: through the pre-trained weights of BioBERT provided on https: / / github.com / naver / biobert-pretrained, the extracted disease names of all disease entities are used as input to fine-tune BioBERT, thereby training the characteristic attributes of the disease entity with semantic features.
[0120] Secondly, extract the characteristic attributes of the drug with drug properties. In one embodiment of the present invention, the present invention uses the API program PyPubChem of the PubChem database to search the extracted drug entities one by one, find the standard name of the corresponding drug and the molecular SMILES sequence corresponding to the drug, and then use the chemical open source toolkit RDKit to extract a series of drug properties such as chemical properties, physical properties, chemical descriptors and molecular fingerprints of the drug using the SMILES sequence as the characteristic attributes of the drug entity.
[0121] Finally, extract the gene characteristic attributes that describe the expression. In an embodiment of the present invention, the expression of the gene in other different tissues or cells is used as the characteristic attribute of the gene entity, that is, the universal expression characteristics of the gene are used as the characteristics of the gene, which can reflect both the specific expression of the gene and the expression difference between genes. The specific steps are: the expression of different genes in the genome in different tissues or cells is collected through the database Connectivity Map and BioGPS, with 1323 samples and 192 samples respectively; for the gene entity extracted in step 110, find the corresponding gene expression as the characteristic attribute of the gene entity.
[0122] The feature attribute loading unit 330 loads the obtained feature attributes into the knowledge graph.
[0123] In an embodiment of the present invention, in order to adapt to the input form of the machine learning model, the characteristic attribute matrices of different entities can be subjected to dimensionality reduction processing respectively through principal component analysis (PCA) technology.
[0124] Finally, a feature attribute matrix of uniform dimension can be obtained and loaded into the knowledge graph as input to perform downstream tasks.
[0125] Although various embodiments of the present invention are described above, it should be understood that they are presented as examples only and not as limitations. It is obvious to those skilled in the relevant art that various combinations, modifications and changes can be made thereto without departing from the spirit and scope of the present invention. Therefore, the breadth and scope of the present invention disclosed herein should not be limited by the exemplary embodiments disclosed above, but should only be defined according to the attached claims and their equivalents.
Claims
1. A method for predicting inter-entity links using biomedical knowledge graphs. include: Constructing a biomedical knowledge graph includes: extracting all disease, drug and gene entities from the knowledge graph; obtaining feature attributes for the extracted entities; and loading the obtained feature attributes into the knowledge graph, including: the disease feature attribute matrix is X D ; The drug feature attribute matrix is X C ; The characteristic attribute matrix of gene entity is X G ; For these different entities, the characteristic attribute matrix X D , X C , X G , firstly, de-mean the values to obtain X′ D , X′ C , X′ G , and calculate the covariance matrix C D , C C , C G ; Calculate the eigenvalue λ and corresponding eigenvector of the covariance matrix Sort the eigenvalues in descending order, select the largest k of them, and use their corresponding k eigenvectors as column vectors to form an eigenvector matrix P, which is used as the input feature attribute matrix Y. D , Y C , Y G , project the entity feature attributes onto the selected feature vector to obtain a feature attribute matrix Y with a unified dimension of N×K D , Y C , Y G , to be loaded into the knowledge graph as input; Receive the feature attribute matrix of all entities in the knowledge graph; The GAT formula with multi-head attention mechanism collects information from neighbor entities and learns the characteristics of the relationship between entities: b ijk =LeakyReLU(W 2 c ijk ) where y i With y f is the characteristic attribute of entity i, j in the knowledge graph, r k is the embedding representation of the relationship between entities i and j in the knowledge graph, W 1 With W 2 is the parameter trained in the machine learning model, || is the connection operator; The relative attention value α is obtained based on the learned characteristics of the relationship between entities ijk : in is the neighbor entity of entity i, is the relationship between entities i and n, r, b ijk represents entities i, j and the relationship k between them; similarly b inr It represents entities i, n and the relationship r between them; According to the characteristics of the relationship between entities and the relative attention value, the embedded representation of the entity after multi-head attention update is obtained Where m is the number of multi-head attention mechanisms and || is the connection operator, thus obtaining the embedded representation of updated nodes and relations.
2. The method for performing entity link prediction tasks using a biomedical knowledge graph as claimed in claim 1, It is characterized in that Extract all disease, drug and gene entities from the knowledge graph including: For disease entities, all disease entities that meet the standards are extracted by matching the disease names in the knowledge graph with the vocabulary, and standardized disease names and disease categories are provided for the disease entities; For drug entities, obtain the name and SMILES sequence of the corresponding drug; For gene entities, the genes with expression levels in the Connectivity Map and BioGPS databases were used as the vocabulary to extract all gene entities that met the standards in the map.
3. The method for performing inter-entity link prediction tasks using a biomedical knowledge graph as claimed in claim 1, It is characterized in that Obtaining feature attributes for the extracted entities includes: For disease entities, disease feature attributes with semantic features are extracted; For drug entities, extracting drug characteristic attributes with drug characteristics; and For gene entities, gene characteristic attributes that describe the expression status are extracted.
4. The method for performing inter-entity link prediction tasks using a biomedical knowledge graph as claimed in claim 3, It is characterized in that For drug entities, the chemical properties, physical properties, chemical descriptors and molecular fingerprints of the drug are obtained through the SMILES sequence corresponding to the drug using the chemical open source toolkit RDKit as the characteristic attributes of the drug entity.
5. The method for performing inter-entity link prediction tasks using a biomedical knowledge graph as claimed in claim 1, It is characterized in that Loading the obtained characteristic attributes into the knowledge graph includes: performing dimensionality reduction processing on the characteristic attribute matrices of different entities through principal component analysis (PCA) technology to obtain a characteristic attribute matrix with a unified dimension.
Citation Information
Patent Citations
Knowledge fusion method and device, computer device and storage medium
CN109886294A
Embedded representation processing method and device based on artificial intelligence and electronic equipment
CN110825884A
Method and device for processing process information
CN111160847A
Drug discovery methods
WO2003067504A2
