A hierarchical guided fine-grained biomedical hypothesis generation method

Abstract

The present application relates to a hierarchical guided fine-grained biomedical hypothesis generation method for predicting potential relationships between biomedical entities. Current biomedical hypothesis generation methods have the problems of incomplete modeling and redundant model parameters, and the present application aims to overcome these problems. The present application proposes a new hierarchical guided fine-grained biomedical hypothesis generation model. The model constructs a knowledge association graph based on a biomedical knowledge base, extracts association information from biomedical scientific literature, and uses a graph neural network computation module, a type interval control module, and an association prediction module to embed biomedical entities. Subsequently, the model is trained on the training set, and finally can be used to calculate the probability of the existence of associations between different biomedical entities.

Description

Technical Field

[0001] This invention relates to a hierarchical structure-guided fine-grained biomedical hypothesis generation method, belonging to the field of biomedical knowledge discovery technology. Background Technology

[0002] With the publication of a large number of scientific papers, hypothesis generation has become an important research field, especially in the biomedical field. The goal of hypothesis generation is to predict unknown scientific discoveries based on existing ones. These hypotheses can greatly advance the development of the biomedical field and provide guidance for treating challenging diseases. Hypothesis generation has consistently been a popular area of ​​work in biomedicine, aiming to derive reliable biomedical hypotheses using only information from published literature.

[0003] The ABC co-occurrence model pioneered by Swanson (Swanson D R. Fish oil, Raynaud's syndrome, and undiscovered public knowledge[J]. Perspectives in biology and medicine, 1986, 30(1):7-18.) played an important role in this task. This model infers the relationship between unreported "AC" by utilizing the relationship between reported biomedical entities "AB" and reported biological entities "BC". Swanson predicted that "dietary fish oil" would be beneficial to "improvement of Raynaud's syndrome" based on the fact that "dietary fish oil" is beneficial to "reducing blood viscosity" and "reducing blood viscosity". This prediction was later confirmed by experiments (DiGiacomo RA, Kremer JM, Shah D M. Fish-oil dietary supplementation in patients with Raynaud's phenomenon: a double-blind, controlled, prospective study[J]. The American Journal of Medicine, 1989, 86(2):158-164.). However, in an increasing number of biomedical literature contexts, the ABC model has revealed some limitations. A significant limitation is that it only considers explicit relationships between entities, ignoring other possible factors and associations. Furthermore, the model always assigns a probability of 1 to all predictions, failing to distinguish the magnitude of probabilities between different predictions, leading to inconsistent prediction quality.

[0004] With the development of machine learning, some machine learning methods have been applied to biomedical hypothesis generation tasks. Spangler et al. (Spangler S, Wilkins AD, Bachman BJ, et al. Automated hypothesis generation based on mining scientific literature [C]. Proc. of the 20th ACMSIGKDD international conference on Knowledge discovery and data mining, 2014. 1877-1886.) used a semi-supervised learning method called graph diffusion to predict kinases that can phosphorylate the protein p53. They downloaded abstract texts containing kinases from MEDLINE and weighted these documents using TF-IDF (Term Frequency–Inverse Document Frequency). Based on the document corresponding to each kinase, they calculated the average value to obtain the vector representation of different kinases. Then, they constructed a graph based on the representation distance of the kinases and labeled the known p53 kinases. After graph diffusion learning, all candidate kinases were assigned a probability that they could phosphorylate p53. This work, which combines machine learning with graph structures, did indeed obtain p53 kinase, which had not been reported in the previous literature, and was confirmed by related molecular biology experiments (Spangler S, Wilkins AD, Bachman BJ, et al. Automated hypothesis generation based on mining scientific literature [C]. Proc. of the 20th ACM SIGKDD international conference on Knowledge discovery and data mining, 2014. 1877-1886.).In addition, another model called MOLIERE (Sybrandt J, Shtutman M, Safro I. Moliere: Automatic biomedical hypothesis generation system [C]. Proc. of the 23rd ACM SIGKDD international conference on knowledge discovery and data mining, 2017, 1633-1642.) constructed a network including biomedical entities and articles, and identified the shortest paths between mining targets. Extracting thematic information from article summaries along the path using the LDA model and analyzing the correlations of thematic information can help determine the relationship between two mining targets. By mining data from literature published before 2009, MOLIERE predicted that DDX3 (Dead Box RNA Helicase 3) could be used to treat cancer, a conclusion confirmed by a 2013 article (Cruciat CM, Dolde C, DeGroot RE, et al. RNA helicase DDX3 is a regulatory subunit of casein kinase1 in Wnt–β-catenin signaling[J]. Science, 2013, 339(6126):1436-1441.). However, these traditional machine learning methods cannot effectively model the intricate relationships in biomedical literature. This is because biomedical literature encompasses a wide variety of biomedical relationships, and these traditional methods cannot capture this complexity well. Furthermore, due to the lack of highly parallelized underlying implementations, these methods require significant computational time.

[0005] With the introduction of deep learning methods (LeCun Y, Bengio Y, Hinton G. Deep learning[J]. Nature, 2015, 521(7553):436-444.), many deep learning models have been used to capture information in high-dimensional spatial data, which is very helpful for modeling complex biomedical entity associations (Luo Ling. Research on several key technologies of biomedical text mining[D]; Dalian University of Technology, 2019.). Graph neural network (GNN) models are a class of deep learning models that operate on graph structures. They can be used to process graph structure information and have a powerful ability to model entity relationships. Therefore, many models containing GNNs are now being applied in biomedical hypothesis generation work (Sybrandt J, Tyagin I, Shtutman M, et al. AGATHA: automatic graph mining and transformer based hypothesis generation approach[C]. Proc. of the 29th ACM International Conference on Information & Knowledge Management, 2020. 2757-2764.; Tyagin I, Kulshrestha A, Sybrandt J, et al. Accelerating COVID-19 research with graph mining and transformer-based learning[C]. Proc. of the AAAI Conference on Artificial Intelligence, 2022. 12673-12679.; Akujuobi U, Spranger M, Palaniappan SK, et al. T-pair: Temporal node-pair embedding for automatic biomedical...). Hypothesis generation[J].IEEE Transactions on Knowledge and Data Engineering,2020,34(6):2988-3001.). Since the association information between biomedical entities can be naturally modeled as a graph structure, where nodes are biomedical entities and edges are associations between entities, graph neural networks are suitable for learning high-dimensional structured information from the original biomedical entity association graph.

[0006] AGATHA (Sybrandt J, Tyagin I, Shtutman M, et al. AGATHA: automatic graphmining and transformer based hypothesis generation approach [C]. Proc. of the 29th ACM International Conference on Information & Knowledge Management, 2020, 2757-2764.) decomposes biomedical literature data from different perspectives. It performs extensive data preprocessing on biomedical literature from PubMed, including decomposing the abstract text of biomedical literature into n-tuples, entities, and technical terms, and constructing a large graph structure. Then, it uses a graph neural network to compute the embedding representations of nodes in this graph structure, and finally inputs the computed node representations into a Transformer model to predict the probability of associations between target biomedical entities. Because biomedical associations are highly complex, external knowledge is needed to assist in modeling. However, AGATHA only uses the data from the biomedical literature itself and does not incorporate external knowledge to complete this task. Therefore, a series of works have attempted to introduce external knowledge into biomedical hypothesis generation tasks. When using graph neural networks for modeling, some works have proposed that, during training, reported associations cannot be simply treated as positive samples and unreported associations as negative samples. Akujuobi et al. (Akujuobi U, Chen J, Elhoseiny M, et al. Temporal positive-unlabeled learning for biomedical hypothesis generation via risk estimation[J]. Advances in Neural Information Processing Systems, 2020, 33:4597-4609.) proposed applying PU learning to graph neural networks, where reported associations are treated as positive samples, and unreported associations are treated as unlabeled samples. By modeling the properties of the knowledge itself as external knowledge utilized in the model, this approach has been successfully applied to predicting COVID-19 related symptoms.Furthermore, the T-pair model (Akujuobi U, Spranger M, Palaniappan SK, et al. T-pair: Temporal node-pair embedding for automatic biomedical hypothesis generation[J]. IEEE Transactions on Knowledge and Data Engineering, 2020, 34(6):2988-3001.) proposes that graph neural networks should focus on modeling the temporal changes in new knowledge discovery. Related data is modeled as a dynamic graph that changes over time. By modeling the changes in related information before and after each time step, patterns of knowledge discovery in the real world can be learned (Min Bo, Wang Kun, Du Bing, et al. The Influence of Publication Time on Knowledge Discovery in Biomedical Literature[J]. Chinese Journal of Medical Information, 2017, 26(05):49-52.). Due to the unique specialization and complexity of the biomedical field, external knowledge is needed to assist in the generation of biomedical hypotheses. Although some current work based on graph neural network methods has modeled external knowledge, the modeling of the characteristics of the biomedical field is insufficient, and precise and detailed biomedical hypotheses cannot be obtained to assist in the study of biological mechanisms. Mechanism discovery in biological mechanism research includes identifying specific molecules or genes that play a role in a disease or condition (Emmerich CH, Gamboa LM, Hofmann MC, et al. Improving target assessment in biomedical research: the GOT-IT recommendations[J]. Nature reviews Drug discovery, 2021, 20(1):64-81.). To achieve this goal, external knowledge must be utilized at a more refined level. Summary of the Invention

[0007] Technical Problem: To address the shortcomings of existing biomedical hypothesis generation methods, such as incomplete modeling and redundant model parameters, the present invention aims to provide a hierarchical structure-guided fine-grained biomedical hypothesis generation method. Under the guidance of a refined biomedical hierarchical structure, this method learns the correlation information extracted from published biomedical scientific literature to predict the implicit connections between biomedical entities.

[0008] Technical solution: To achieve the above objectives, the present invention adopts the following technical solution:

[0009] Step S1: Extract the knowledge association graph from the biomedical knowledge base and organize it into a dataset, which is then divided into a training set and a validation set.

[0010] Step S2: Construct an initialization model to embed biomedical entities; the initialization model includes a graph neural network computation module, a type interval control module, and an association prediction module; the graph neural network computation module uses a graph neural network to model the biomedical association graph; the type interval control module models the biomedical hierarchical tree; the association prediction module includes an MLP, used to connect the starting node embedding representation and the target node embedding representation, and calculate the association rationality score;

[0011] The initialization model is trained using the dataset to obtain a biomedical hypothesis generation model that can predict the association of biomedical entities.

[0012] Step S3: Use the trained model to calculate the probability that there is a correlation between different biomedical entities.

[0013] Step S1 specifically includes:

[0014] Step S101: Select a suitable biomedical knowledge base, such as PubMed, PubMed Central, ClinicalTrials.gov, DrugBank, SemMedDB, etc., to meet specific research needs.

[0015] Step S102: Construct a graph of relationships between biomedical entities, where entities are represented as nodes in the graph and relationship information from the literature is represented as edges in the graph.

[0016] Step S103: Perform entity filtering on the constructed association graph to delete invalid or no longer maintained entities, ensuring high-quality data.

[0017] Step S104: Divide the association graph into two parts: data before 2015 as the training set and data after 2015 as the validation set to evaluate the model performance.

[0018] Step S2 specifically includes:

[0019] Step S201: Construct an initialization model, including a graph neural network computation module, a type interval control module, and an association prediction module.

[0020] Step S202: Train the model parameters to obtain a biomedical hypothesis generation model that can predict the association of biomedical entities.

[0021] Step S3 specifically includes:

[0022] Step S301: Input the target biomedical entity and use the model to calculate the association probability.

[0023] Preferably, the biomedical knowledge base is SemMedDB; the dataset comes from the PREDICATION table of SemMedDB.

[0024] Preferably, the graph neural network computing module uses BioBERT to initialize the embedded representation of biomedical entities.

[0025] Preferably, the graph neural network employs two layers of GraphSAGE that retain the 2-norm, sampling and aggregating the neighborhoods of nodes to update the embedding representation of each node, thereby modeling the topology of the entire graph.

[0026] Preferably, the biomedical hierarchical tree is derived from the UMLS semantic network.

[0027] Preferably, the type interval control module employs the minibatch-EAF algorithm to model the hidden hierarchical information between different biomedical entity associations. Modeling the positional relationships of node pairs using the minibatch-EAF algorithm pre-classifies them into two categories: homogeneous information between nodes of the same type and heterogeneous information between nodes of different types. If the entity types on both sides of a biomedical association are the same, the type interval control module will apply the internal minibatch-EAF algorithm to calculate the ancestor embedding representations of these two entity types from the biomedical hierarchical tree. If the entity types on both sides of a biomedical association are different, the type interval control module will apply the external minibatch-EAF algorithm to calculate the ancestor embeddings from the biomedical hierarchical tree. The E value output by the internal minibatch-EAF algorithm... u(in) E v(in) E a(in) E and the output of the external minibatch-EAF algorithm u(out) E v(out) E a(out) Both are treated as sequences of length 3 and are fed into two different Transformer models for computation.

[0028] Preferably, the MLP calculates the superposition interval value c to be added to the base interval value m based on the implicit representation calculated by the Transformer model.

[0029] Beneficial Effects: This method proposes a hierarchical structure-guided graph neural network learning model for generating fine-grained biomedical hypotheses. It initializes the biomedical knowledge base using a pre-trained language model, encodes biomedical entities using a graph neural network model, and distinguishes easily confused biomedical entities by modeling a biomedical hierarchical tree. The model of this invention reduces the number of parameters, improves accuracy, and is less prone to confusion when predicting easily confused biomedical entities.

[0030] This model is the first to combine a graph neural network computation module, a type margin control module, and an association prediction module for fine-grained biomedical hypothesis generation. The graph neural network computation module innovatively preserves the L2 norm of the GraphSAGE input and scales the output to that L2 norm. The type margin control module is an independently designed module that can model the differences between different biological entity types, thereby ensuring that nodes with biomedical entity types can be distinguished from each other based on the differences in the actual biomedical field. Attached Figure Description

[0031] Figure 1 This is a schematic diagram of the workflow of the biomedical hypothesis generation system in Example 1.

[0032] Figure 2 This is a schematic diagram of the system architecture of the hypothesis generation model in Example 1.

[0033] Figure 3 This is a schematic diagram of the biomedical hierarchical tree modeled from the UMLS semantic network in Example 1. Detailed Implementation

[0034] Example 1:

[0035] See Figures 1-3 This embodiment provides a hierarchical structure-guided fine-grained biomedical hypothesis generation method, which specifically includes the following steps:

[0036] Step S1: Extract the knowledge association graph from the biomedical knowledge base and organize it into a dataset, which is then divided into a training set and a validation set.

[0037] Step S2: Combining the biomedical hierarchical tree, the embedded representation of biomedical entities is calculated through the graph neural network calculation module, type interval control module, and association prediction module, and the model is trained using the dataset.

[0038] Step S3: Use the trained model to calculate the probability that there is a correlation between different biomedical entities.

[0039] Step S1 specifically includes:

[0040] Step S101: Select a suitable biomedical knowledge base, such as PubMed, PubMed Central, ClinicalTrials.gov, DrugBank, SemMedDB, etc., to meet specific research needs.

[0041] In this embodiment, SemMedDB (Kilicoglu H, Shin D, Fiszman M, et al. SemMedDB: a PubMed-scale repository of biomedical semantic predications[J]. Bioinformatics, 2012, 28(23):3158-3160.) was selected. SemMedDB is a biomedical association database extracted from all articles indexed by MEDLINE. It uses SemRep (Rindflesch TC, Fiszman M. The interaction of domain knowledge and linguistic structure in natural language processing: interpreting hypernymic propositions in biomedical text[J]. Journal of biomedical informatics, 2003, 36(6):462-477.) as a biomedical semantic interpreter to extract the titles and abstracts of the articles. SemRep extracts triplet information (concept_a, relation, concept_b) from the corpus through pre-linguistic analysis, lexical / syntactic analysis, reference analysis, post-reference analysis, and relation analysis. Concept_a and concept_b represent conceptual information from the UMLS (Unified Medical Language System) and the relationship between these two concepts. After nearly two months of preprocessing, SemMedDB contains 57.6 million relational information extracted from 21 million articles, covering 30 types of relations and 127 biomedical entities. This dataset encompasses a wide range of biomedical information, making it highly advantageous for literature mining models that require extensive knowledge.

[0042] Step S102: Construct a graph of relationships between biomedical entities, where entities are represented as nodes in the graph and relationship information from the literature is represented as edges in the graph.

[0043] The dataset used in this example comes from the PREDICATION table in SemMedDB, which contains 115,525,915 records. By traversing all records in the PREDICATION table, a set of biomedical entities can be obtained. Based on this set, these entities are modeled as nodes in a graph structure, and the tuple (SUBJECT CUI, OBJECT CUI) is treated as an edge. Finally, a graph containing the associations of biomedical entities is constructed.

[0044] Step S103: Perform entity filtering on the constructed association graph to delete invalid or no longer maintained entities, ensuring high-quality data.

[0045] In SemMedDB, each biomedical entity in the PREDICATION table has its own corresponding ID, which can be mapped to a unique biomedical entity concept. These IDs come from two sources: UMLS and Entrez Gene. Many entity concepts in these databases are no longer retained or have been revoked, thus requiring entity filtering. For IDs from Entrez Gene, the MyGene package is used to retain all Entrez Gene entities that are still being maintained. For IDs from UMLS, the pyumls package is used to obtain information from the UMLS database; IDs not found in the database are filtered out.

[0046] Step S104: Divide the association graph into two parts: data before 2015 as the training set and data after 2015 as the validation set to evaluate the model performance.

[0047] Step S2 specifically includes:

[0048] Step S201: Construct an initialization model, including a graph neural network computation module, a type interval control module, and an association prediction module.

[0049] Graph Neural Network Computation Module:

[0050] Since each entity in the biomedical association graph requires an embedded representation to participate in hypothesis generation calculations, BioBERT (Lee J, Yoon W, Kim S, et al. BioBERT: a pre-trained biomedicallanguage representation model for biomedical text mining[J]. Bioinformatics, 2020, 36(4):1234-1240.) is used to initialize the embedded representations of these biomedical entities. The process of initializing the embedded representations of biomedical entities can be represented as follows:

[0051] E = BioBERT(V) (1)

[0052] e v =BioBERT(v) (2)

[0053] BioBERT(v) represents inputting the names of nodes v in the biomedical entity association graph into the BioBERT model for encoding, and using the encoding output by BioBERT to initialize the embedding representation of these biomedical entities. V is a vector connecting all nodes, V = [v1, v2, ..., v n E is the vector formed by concatenating the encodings of all nodes calculated by BioBERT, E = [e1, e2, ..., e2]. n ].

[0054] To enable the model to learn from published literature data, a graph neural network was used to model the biomedical association graph. For this model, a two-layer GraphSAGE (Hamilton W, Ying Z, Leskovec J. Inductive representation learning on large graphs[J]. Advances in neural information processing systems, 2017, 30.) was chosen. GraphSAGE models the topology of the entire graph by sampling and aggregating the neighborhood of each node to update its embedding representation. Simultaneously, as an inductive graph neural network, it learns an aggregator to compute the nodes in the graph structure. This setup allows the trained model to function without retraining when the number of nodes in the graph is updated.

[0055] The specific computation process of the GraphSAGE convolutional layer is defined as follows:

[0056]

[0057]

[0058] Where AGGREGATE represents the aggregation operation, and k represents the current level. Let be the set of neighbors of node v in the graph, W be the learned weight matrix, h be the embedding representation, and σ be the activation function. The aggregation operation aims to aggregate the representations of nodes and their neighbors. If an average aggregation operation is used, this aggregation operation can be defined as:

[0059]

[0060] Where MEAN represents the average of the embedding representations of nodes, W A This is the weight matrix learned in the aggregator.

[0061] GraphSAGE models the topology of the entire graph by sampling and aggregating the neighborhoods of nodes to update the node embeddings of each node. To preserve the information from the original BioBERT encoding after multiple convolutional operations, this model records the 2-norm of the input embedding representation and scales the GraphSAGE output to the same 2-norm size. The two-layer GraphSAGE that preserves the 2-norm value is defined as follows:

[0062]

[0063]

[0064]

[0065]

[0066]

[0067]

[0068] in Let v be the embedding representation of the biomedical entity v at layer l, and σ be the sigmoid function.

[0069] Type Interval Control Module:

[0070] In the learning process of graph neural networks, the representation of an entity directly depends on the representations of its neighboring nodes. This means that if two entities share a large number of adjacent entities, their representations will become similar. Due to the characteristics of biomedical entities, diseases of the same type often have highly overlapping symptoms. This means that after learning by graph neural networks, their node representations will become similar, which can easily lead to model confusion between the two biomedical entities. In order to obtain a finely distinguishable representation, this model will model the biomedical hierarchical tree proposed by the UMLS semantic network (McCray AT. An upper-level ontology for the biomedical domain[J]. Comparative and Functional genomics,2003,4(1):80-84.), thereby using this hierarchical structure to distinguish easily confused biomedical entities and improve the quality of hypothesis generation results.

[0071] Therefore, a type spacing control module was designed in this model. This module can model the differences between different biological entity types, thereby ensuring that nodes with biomedical entity types can be distinguished from each other based on the differences in the actual biomedical field. This chapter models a biomedical hierarchical tree from the UMLS semantic network. This biomedical hierarchical structure is defined according to different biomedical entity types, and the distance in the hierarchical structure can be used to help represent the differences between different entity types in the biomedical field. Each node in the diagram represents a biomedical entity type number defined by UMLS, and each number has a corresponding biomedical entity type. For example, the number "T047" corresponds to the entity type "Disease or Syndrome" in the UMLS definition.

[0072] To leverage hierarchical information in the biomedical field, this module proposes the minibatch-EAF (minibatch Embedding Ancestor Find) algorithm to model the hidden hierarchical information between associations of different biomedical entities. For a biomedical association edge e... i,j Its two sides are biomedical entity nodes V i and V j The entity type is t x and t y Such entity type pairs can be mapped to a set of node pairs in a biomedical hierarchy tree. Since node pairs with different positional relationships in a biomedical hierarchy tree have different requirements for differentiation, the positional relationships of node pairs will be pre-classified into two categories during modeling: homogeneous information of nodes of the same type and heterogeneous information of nodes of different types.

[0073] If the entities on both sides of a biomedical association are of the same type, the type spacing control module will apply the internal minibatch-EAF algorithm to calculate the ancestor embedding representations of these two entity type nodes from the biomedical hierarchical tree. The module searches for nodes of the same type in the minibatch based on the entity type, calculates their average embedding representation, and thus obtains the ancestor embedding representation. Since in actual biomedical research, a biomedical entity may have multiple corresponding entity types. For example, the oxytocin receptor (C0069841) corresponds to the species numbers "T192" and "T116" in UMLS, representing the species "Receptor" and "Amino Acid, Peptide, or Protein," therefore, the calculation process needs to calculate all the species corresponding to the biomedical entity. Algorithm 1-1 describes the entire calculation process of the internal minibatch-EAF algorithm and how the algorithm obtains the internal ancestor embedding representation E by using a hierarchical tree to calculate homogeneous information. a(in) .

[0074] If the entity types on both sides of the biomedical association are different, the type gap control module will apply the external minibatch-EAF algorithm to calculate the ancestor embedding E from the biomedical hierarchy tree. a(out) By finding the common ancestor of two nodes in a biomedical hierarchical tree, the association between two types of nodes in this hierarchical tree can be summarized, because the information of the most recent common ancestor in the hierarchical tree can help determine the degree of distinction between these two descendant types. Similarly, since an entity may correspond to multiple entity types, it is necessary to model all the categories corresponding to the entity during the computation process. Algorithm 1-2 describes how the external minibatch-EAF algorithm processes heterogeneous information based on the embedding average of relevant ancestor types, ultimately obtaining the external ancestor embedding representation.

[0075]

[0076]

[0077]

[0078] Then, the E output by the internal minibatch-EAF algorithm u(in) E v(in) E a(in) E and the output of the external minibatch-EAF algorithm u(out) E v(out) E a(out)Both can be considered as sequences of length 3 and fed into the Transformer model for computation. Since homogeneous and heterogeneous information are computed separately, the information calculated in these two cases will be processed by two different Transformer models. The Transformer model is chosen here primarily because the triples consisting of the source node, target node, and ancestor nodes in the biomedical hierarchy naturally possess order information; this part-to-whole order is easily modeled by the Transformer model. The Transformer model encodes the triples containing hierarchical information, and then the MLP calculates the superposition interval value c added to the base interval value m based on the implicit representation computed by the Transformer model. This process can be defined as:

[0079] c in =MLP in (Transformer in (cat in (12)

[0080] c out =MLP out (Transformer out (cat out (13)

[0081]

[0082]

[0083] Where c in c is the superposition interval value calculated for homogeneous information. out The superposition interval value is calculated for heterogeneous information. The reason for constructing this design of superimposing the basic interval is mainly to prevent the model from continuously reducing the interval value in order to converge quickly, causing the interval between the embedding representations of two nodes to become invalid. This superposition can ensure that the model can perform fine-grained dynamic modeling based on the information of the biomedical association itself while maintaining basic distinction.

[0084] Association prediction module:

[0085] Given the initial node embedding representation E u and target node embedding representation E v This model designs an association prediction module based on MLP to predict the rationality of associations. The node embeddings of two nodes in the association are concatenated and fed into the MLP to calculate the association rationality score. The association prediction module can be described as follows:

[0086]

[0087] Where ReLU is the activation function. W1 is the matrix of the first layer of the MLP, and b1 is the offset of the first layer of the MLP; W2 is the matrix of the second layer of the MLP, and b2 is the offset of the first layer of the MLP; these are initially randomly initialized and will stabilize during the learning process.

[0088] Step S202: Train the model parameters to obtain a biomedical hypothesis generation model that can predict the association of biomedical entities.

[0089] In this embodiment, the hypothesis generation model assigns a score to both positive and negative samples in the biomedical association graph, where positive samples are sampled from reported associations between nodes, and negative samples are sampled from unreported associations. Positive samples are labeled 1 and negative samples are labeled 0, and then the model predicts the plausibility score of the association.

[0090] The node representations in the diagram are jointly trained based on association prediction and type margin control. Therefore, the joint learning loss of association prediction and type margin control is considered the training objective.

[0091] To predict the associations between nodes within the graph, the cross-entropy function is used to calculate the loss L. G As shown below:

[0092]

[0093] in Then, the interval loss is used to calculate the loss for the type interval controller, as follows:

[0094] L in =max(0, m) in +c in -||E u(in) -E v(in) ||2) (18)

[0095] L out =max(0, m) out +c out -||E u(out) -E v(out) ||2) (19)

[0096] Where m in and m out It is a hyperparameter that controls the basic interval of different controllers.

[0097] Therefore, the total loss can be calculated as follows:

[0098] L = L G +λ1·Lin +λ2·L out (20)

[0099] Where λ1 and λ2 are hyperparameters representing the degree of constraint of the control type interval control module.

[0100] By training the model using the BP algorithm, a biomedical hypothesis generation model that can predict the associations of biomedical entities is finally obtained.

[0101] Step S3 specifically includes:

[0102] Step S301: Input the target biomedical entity and use the model to calculate the association probability. For example, suppose we want to generate biomedical hypotheses about heart disease. First, we build a graph of all other biomedical entities, calculate the association probability between all other biomedical entities and heart disease, remove already reported associations, and what remains are the generated biomedical hypotheses. This probability value represents the confidence in these hypotheses, i.e., how confident we are in believing that the biomedical hypothesis is true.

[0103] First, select the biomedical entity you want to calculate. Then, build a graph from this biomedical entity to all other biomedical entities. By inputting this graph into the hypothesis generation model, removing the biomedical associations that have already been reported, and sorting them from high to low according to the reasonableness of the associations, you can obtain the set of currently unreported biomedical entities related to the target biomedical entity and the corresponding association probabilities.

Claims

1. A hierarchical structure-guided fine-grained biomedical hypothesis generation method, characterized in that, Includes the following steps: Biomedical association graphs were extracted from the biomedical knowledge base and organized into a dataset; An initialization model is constructed to embed biomedical entities. This initialization model includes a graph neural network computation module, a type-space control module, and an association prediction module. The graph neural network computation module uses a graph neural network to model the biomedical association graph. The type-space control module employs a minibatch-EAF algorithm to model the hidden hierarchical information between associations of different biomedical entities. Specifically, if the entities on both sides of a biomedical association are of the same type, an internal minibatch-EAF algorithm is applied to search for nodes of the same type in the minibatch based on the entity type, calculating their average embedding representation to obtain the ancestor embedding representation. If the entity types are different, an external minibatch-EAF algorithm is applied to find the common ancestor of two nodes in the biomedical hierarchical tree, thus summarizing the association between the two types of nodes in this hierarchical tree. The output of the internal minibatch-EAF algorithm is... The external minibatch-EAF algorithm outputs ; Forming the first sequence, The first and second sequences are respectively fed into two different Transformer models for computation; the association prediction module includes an MLP, which is used to connect the starting node embedding representation and the target node embedding representation, and calculate the association rationality score. The initialization model is trained using the dataset to obtain a biomedical hypothesis generation model that can predict the association of biomedical entities; The probability of associations between different biomedical entities is calculated using a trained biomedical hypothesis generation model.

2. The hierarchical structure-guided fine-grained biomedical hypothesis generation method according to claim 1, characterized in that, The biomedical knowledge base is SemMedDB; the dataset comes from the PREDICATION table of SemMedDB.

3. The hierarchical structure-guided fine-grained biomedical hypothesis generation method according to claim 1, characterized in that, The graph neural network computing module uses BioBERT to initialize the embedded representation of biomedical entities.

4. The hierarchical structure-guided fine-grained biomedical hypothesis generation method according to claim 1, characterized in that, The graph neural network employs two layers of GraphSAGE that retain the 2-norm, sampling and aggregating the neighborhoods of nodes to update the embedding representation of each node, thereby modeling the topology of the entire graph.

5. The hierarchical structure-guided fine-grained biomedical hypothesis generation method according to claim 1, characterized in that, The biomedical hierarchical tree is derived from the UMLS semantic network.

6. The hierarchical structure-guided fine-grained biomedical hypothesis generation method according to claim 1, characterized in that, The minibatch-EAF algorithm is used to model the positional relationship of node pairs. The nodes are pre-classified into two categories: homogeneous information of nodes of the same type and heterogeneous information of nodes of different types. If the entity types on both sides of the biomedical association are the same, the type interval control module will use the internal minibatch-EAF algorithm to calculate the ancestor embedding representation of the two entity types from the biomedical hierarchical tree. If the entity types on both sides of the biomedical association are different, the type interval control module will use the external minibatch-EAF algorithm to calculate the ancestor embedding from the biomedical hierarchical tree.

7. The hierarchical structure-guided fine-grained biomedical hypothesis generation method according to claim 1, characterized in that, MLP calculates the implicit representation based on the Transformer model and adds it to the base interval value. superposition interval value .

Images