Method for constructing variant document interpretation knowledge base, interpretation method and electronic device

Abstract

A method for constructing a variant literature interpretation knowledge base based on NLP, an interpretation method, and an electronic device are disclosed. The method for constructing the variant literature interpretation knowledge base includes the following steps: acquiring disease-related literature; constructing a database of entities related to gene variants based on the disease-related literature; constructing a knowledge graph of literature evidence for variant literature interpretation; extracting evidence from the literature evidence knowledge graph to obtain evidence corresponding to the entities; and constructing the variant literature interpretation knowledge base based on the evidence and the database. This allows for more comprehensive and systematic literature evidence. Furthermore, during interpretation, inputting the entity name automatically returns results based on the evidence standards or evidence types read in the literature, achieving second-level return of literature search results and significantly improving the quality and efficiency of literature search.

Description

Technical Field

[0001] This application relates to the field of bioinformatics technology, and more specifically, to a method for constructing a knowledge base for interpreting variant literature based on NLP, an interpretation method, and an electronic device. Background Technology

[0002] The rapid development of DNA (Deoxyribonucleic acid) sequencing technology has generated an astonishing amount of genomics data, opening a new chapter in precision medicine and revolutionizing the diagnostic methods for genetic diseases. Combining gene testing results with clinical interpretation guidelines developed by various authoritative institutions, and guided by evidence-based medicine principles, interpreting clinically significant gene variants has become a consensus. For example, the American College of Medical Genetics and Genomics (ACMG) has established standards and guidelines for the classification of genetic variants; the Association for Molecular Pathology, the American Society of Clinical Oncology, and the College of American Pathologists jointly developed guidelines for interpreting somatic tumor variants. Computational and query software based on existing expert-compiled variant databases has emerged, automating or semi-automating the interpretation of some gene variants. However, the number of variants included in these databases is very limited. When clinical laboratories need to issue clinical reports on detected variants, they still primarily rely on manual interpretation by professionals using variant interpretation guidelines. During the interpretation process, interpreters use the gene and variant site to be interpreted as search keywords to obtain relevant literature. By reading and analyzing this literature, they identify literature evidence that the variant site to be interpreted meets the requirements of a specific guideline. Therefore, obtaining valuable reference information through manual literature review remains a crucial step in the interpretation process.

[0003] Obtaining variant evidence through manual literature review has several drawbacks: First, it is time-consuming and labor-intensive, with limited scope and speed, making it a critical step in interpreting disease variants. Second, it demands a high level of expertise from interpreters, requiring them to possess substantial knowledge in gene and variant interpretation. Third, the sheer volume and fragmentation of literature, coupled with the multiple names and spellings of gene variant-related entities such as gene names, variant names, drug names, disease names, and phenotype names, often necessitates manual input of common keywords, hindering the acquisition of comprehensive literature. Therefore, the difficulty in quickly and comprehensively obtaining variant evidence through manual literature review constitutes a significant bottleneck in improving interpretation efficiency. Summary of the Invention

[0004] This invention aims to at least partially address one of the technical problems in related technologies. Therefore, the first objective of this invention is to propose a method for constructing a knowledge base for interpreting variant literature based on NLP (Natural Language Processing). This method automates and intelligently acquires evidence from disease variant literature, effectively improving the speed of interpreting gene variant-related data and providing more comprehensive evidence, thereby enhancing the quality and efficiency of interpreting gene testing reports.

[0005] The second objective of this invention is to propose a method for interpreting variant literature based on NLP.

[0006] The third objective of this invention is to provide an electronic device.

[0007] To achieve the above objectives, a first aspect of the present invention proposes a method for constructing a variant literature interpretation knowledge base based on NLP, comprising the following steps: acquiring disease-related literature; constructing a database of entities related to gene variants based on the disease-related literature; constructing a literature evidence knowledge graph for variant literature interpretation; extracting evidence from the literature evidence knowledge graph to obtain evidence corresponding to the entities, and constructing a variant literature interpretation knowledge base based on the evidence and the database.

[0008] According to an embodiment of the present invention, a method for constructing a variant literature interpretation knowledge base based on NLP involves acquiring disease-related literature, constructing a database of entities related to gene variants based on this literature, constructing a knowledge graph of literature evidence for variant literature interpretation, and extracting evidence from this knowledge graph to obtain evidence corresponding to the entities. The variant literature interpretation knowledge base is then constructed based on the evidence and the database. Therefore, this application represents a method for automatically reading literature and obtaining disease literature evidence based on NLP and knowledge graph technologies. This method for constructing a variant literature interpretation knowledge base makes the literature evidence more comprehensive and systematic. During interpretation, when any entity name related to gene variants is input, the method automatically returns variant evidence results based on literature reading. This achieves automation and intelligence in obtaining disease variant literature evidence, effectively improving the interpretation speed of gene variant-related information and enabling second-level return of literature search results for entities related to gene variants. This significantly improves the efficiency of literature search and, consequently, enhances the quality and efficiency of gene testing report interpretation.

[0009] According to one embodiment of the present invention, constructing a database of entities related to gene mutations based on disease-related literature includes: constructing an entity extraction model using a portion of the disease-related literature; extracting entities from the remaining literature in the disease-related literature using the entity extraction model to obtain entity names; constructing an entity alignment model; aligning the entity names with the entity names using the entity alignment model to obtain the entity standard terms corresponding to the entity names; and constructing a database of entities related to gene mutations based on the entity names and the entity standard terms corresponding to the entity names.

[0010] According to one embodiment of the present invention, an entity extraction model is constructed using a portion of disease-related literature, including: annotating the portion of the literature with entities; adding position and entity classification labels to each word in the annotated literature to obtain an entity label sequence; constructing a pre-trained model of the entity extraction model; and adjusting the pre-trained model using the entity label sequence to obtain the entity extraction model.

[0011] According to one embodiment of the present invention, constructing a pre-trained model for an entity extraction model includes: acquiring a pre-training corpus, the pre-training corpus including relevant literature in the biomedical field; encoding each word in the pre-training corpus to obtain word embedding vectors, fragment embedding vectors, and position embedding vectors; using the sum of the word embedding vectors, fragment embedding vectors, and position embedding vectors as input, and using randomly masked word vectors as labels, pre-training a natural language processing model based on a self-attention mechanism through a backpropagation algorithm to obtain a pre-trained model.

[0012] According to one embodiment of the present invention, constructing a pre-trained model for an entity extraction model further includes: using the cross-entropy of the predicted value and the label as a loss function to train the pre-trained model until the loss value output by the loss function meets a preset condition, and the pre-trained model is trained.

[0013] According to one embodiment of the present invention, a pre-trained model is trained using an entity label sequence to obtain an entity extraction model, including: constructing a fine-tuning model of the entity extraction model based on the pre-trained model; using the model weights obtained during the training of the pre-trained model as the initial weights for the entity extraction task, and using the word embedding vectors corresponding to each word in the document after entity labeling as the input, and using the position and entity classification label corresponding to each word as the output, and training the fine-tuning model through the backpropagation algorithm to obtain the entity extraction model.

[0014] According to one embodiment of the present invention, the cross-entropy of the predicted value and the label is used as the loss function to train the fine-tuning model until the loss value output by the loss function meets the preset conditions, and the fine-tuning model training is completed.

[0015] According to one embodiment of the present invention, after extracting entity names from the remaining documents in disease-related literature using an entity extraction model, the method further includes: matching the remaining documents with a preset entity dictionary and / or a preset entity writing pattern to supplement entity names not recognized by the entity extraction model.

[0016] According to one embodiment of the present invention, constructing an entity alignment model includes: obtaining entity standard terms and other entity names corresponding to the entity standard terms, and constructing a gene alignment dictionary based on the entity standard terms and other entity names; and / or, obtaining entity standard terms and constructing a regular expression for entity alignment based on the entity standard terms.

[0017] According to one embodiment of the present invention, the regular expression includes, but is not limited to, one or more of the following expressions: c.{arbitrary length of 1 and number of symbols ≥ 0}{arbitrary length of 1 letter}>{arbitrary length of 1 letter and number of symbols ≥ 0}; p. {Any length of letter ≥ 1, with ≥ 0 symbols}{Any length of number ≥ 1}{Any length of letter ≥ 1, with ≥ 0 symbols}; rs{Any length of number ≥ 1}; chr{Any length of letter ≥ 1}-{Any length of number ≥ 1}-{Any length of letter ≥ 1, with ≥ 0 symbols}-{Any length of letter ≥ 1, with ≥ 0 symbols}; n.{Any length of number ≥ 1, with ≥ 0 symbols}{Any length of letter ≥ 1}>{Any length of letter ≥ 1, with ≥ 0 symbols}; IVS.{Any length of number ≥ 1, with ≥ 0 symbols}{Any length of letter ≥ 1}>{Any length of letter ≥ 1, with ≥ 0 symbols}; {Any length of letter ≥ 1}{Any length of number ≥ 1}{Any length of letter ≥ 1}.

[0018] According to one embodiment of the present invention, entity names are aligned using an entity alignment model to obtain the entity standard terms corresponding to the entity names, including: performing exact matching and fuzzy matching between the entity names and the entity alignment dictionary to obtain the entity standard terms corresponding to the entity names; and performing exact matching and rule matching between the entity names and regular expressions to obtain the entity standard terms corresponding to the entity names.

[0019] According to one embodiment of the present invention, the database of entities related to gene variation includes a dictionary of {entity names: entity standard terms}, a data list of (document identification information, entity standard terms) and a data list of (document identification information, entity names).

[0020] According to one embodiment of the present invention, constructing a knowledge graph of documentary evidence for interpreting variant documents includes: obtaining the judgment logic of the evidence standard or evidence type for interpreting documents in the variant interpretation guide; representing the judgment logic in the form of triples, wherein the triple is (entity, relationship between entity and evidence standard or evidence type, evidence standard or evidence type); and constructing a knowledge graph of documentary evidence with entities and evidence standards or evidence types as nodes and relationships between entities and evidence standards or evidence types as edges.

[0021] According to one embodiment of the present invention, evidence extraction is performed on a knowledge graph of documentary evidence to obtain evidence corresponding to entities, and a knowledge base for interpreting variant literature is constructed based on the evidence and the database. This includes: extracting sentences containing nodes or node meanings, along with the preceding and following sentences, from documents corresponding to entities related to gene variants in the database, generating a set of evidence sentences corresponding to the nodes; extracting evidence words indicating relationships from the set of evidence sentences; generating entity standard terms, evidence standards or evidence types, evidence sentences, and evidence words corresponding to the documents based on the set of evidence sentences and the evidence words; and constructing a knowledge base for interpreting variant literature based on document identification information and the entity standard terms, evidence standards or evidence types, evidence sentences, and evidence words corresponding to the documents.

[0022] According to one embodiment of the present invention, an entity includes one or more of genes, variants, drugs, diseases, and phenotypes.

[0023] To achieve the above objectives, a second aspect of the present invention proposes a method for interpreting variant documents based on NLP, comprising the following steps: obtaining the entity name to be interpreted; inputting the entity name into a variant document interpretation knowledge base to obtain the evidence standard or evidence type, evidence sentence and evidence word corresponding to the entity name, wherein the variant document interpretation knowledge base is constructed according to the above-described method for constructing a variant document interpretation knowledge base based on NLP.

[0024] According to an embodiment of the present invention, the NLP-based method for interpreting variant literature obtains the entity name to be interpreted and inputs it into a variant literature interpretation knowledge base to obtain the corresponding evidence standards or evidence types, evidence sentences, and evidence words. The variant literature interpretation knowledge base is constructed according to the aforementioned NLP-based variant literature interpretation knowledge base construction method. Therefore, by inputting the entity name, the corresponding evidence standards or evidence types, evidence sentences, and evidence words can be automatically obtained, enabling automation and intelligence in obtaining disease variant literature evidence. This effectively improves the speed of gene variant interpretation and provides more comprehensive literature evidence, thereby improving the quality and efficiency of gene testing report interpretation.

[0025] To achieve the above objectives, a third aspect of the present invention provides an electronic device, including a memory, a processor, and a variant document interpretation program stored in the memory and executable on the processor. When the processor executes the variant document interpretation program, it implements the above-described NLP-based variant document interpretation method.

[0026] According to the electronic device of the present invention, when the processor executes the variant literature interpretation program, the above-mentioned NLP-based variant literature interpretation method is realized. Thus, by inputting entity names, the corresponding evidence standards or evidence types, evidence sentences and evidence words can be automatically obtained. This can realize the automation and intelligence of obtaining disease variant literature evidence, effectively improve the interpretation speed of gene variant related documents, and make the literature evidence more comprehensive, thereby improving the quality and efficiency of gene testing report interpretation.

[0027] Additional aspects and advantages of the invention will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of the invention. Attached Figure Description

[0028] Figure 1 A flowchart illustrating a method for constructing a variant literature interpretation knowledge base based on NLP according to an embodiment of the present invention;

[0029] Figure 2 This is a schematic diagram illustrating the construction of an entity extraction model according to an embodiment of the present invention;

[0030] Figure 3 This is a schematic diagram of an entity extraction model based on a self-attention mechanism according to an embodiment of the present invention;

[0031] Figure 4 This is a schematic diagram illustrating the construction of an entity alignment model according to an embodiment of the present invention;

[0032] Figure 5 This is a schematic diagram of a document evidence knowledge graph according to an embodiment of the present invention;

[0033] Figure 6 This is a schematic diagram of a knowledge base interface for interpreting variant documents according to an embodiment of the present invention;

[0034] Figure 7 A flowchart of a variation-based literature interpretation method based on NLP according to an embodiment of the present invention;

[0035] Figure 8 This is a structural block diagram of an electronic device according to an embodiment of the present invention. Detailed Implementation

[0036] Embodiments of the present invention are described in detail below, examples of which are illustrated in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and intended to explain the present invention, and should not be construed as limiting the present invention.

[0037] The following description, with reference to the accompanying drawings, describes the construction method, interpretation method, and electronic device of the NLP-based variant literature interpretation knowledge base provided by embodiments of the present invention.

[0038] Natural Language Processing (NLP) is a crucial area within computer science and artificial intelligence. Due to the ambiguity and polysemy inherent in natural language text, effectively learning vocabulary and representing and extracting information and relationships is paramount. Over the past few decades, NLP has undergone a significant evolution, from basic rules and statistics to the widespread use of deep learning techniques. In recent years, deep learning NLP, exemplified by Google's Transformer framework proposed in 2017, has demonstrated remarkable promise based on self-attention mechanisms, achieving breakthroughs in areas such as machine translation, sentiment analysis, information extraction, and automated question answering. Essentially, this technology employs self-supervised learning on massive datasets to learn pre-trained feature representations for text. Researchers then use these pre-trained feature representations as a starting point for task-specific supervised machine learning. Knowledge graphs, as a human-interpretable and machine-friendly representation of knowledge, inherently possess semantic meaning and logical rules, making them crucial for knowledge reasoning.

[0039] Therefore, this application proposes a method for constructing a knowledge base for interpreting variant literature based on natural language processing and knowledge graph technology. This method extracts entities related to gene variants from the literature and aligns these entities to standard terms, accurately and comprehensively filtering out literature related to the entity names to be interpreted. Then, based on the judgment logic of the evidence standards or evidence types used in the variant interpretation guidelines, a knowledge graph of literature evidence is constructed to obtain the evidence standards or evidence types, evidence sentences, and evidence words corresponding to the entity names to be interpreted. The judgment logic obtained by querying the knowledge graph for the entity names to be interpreted, and the corresponding evidence-related sentences extracted from the literature, can provide literature evidence for machine-automated literature reading.

[0040] It should be noted that since entities related to gene variation may include one or more of entities such as genes, variations, drugs, diseases, and phenotypes, for ease of explanation, the following embodiments use gene entities and variation entities as examples to illustrate this application. It should be understood that this application is not limited thereto.

[0041] Figure 1 This is a flowchart illustrating a method for constructing a variant literature interpretation knowledge base based on NLP according to an embodiment of the present invention. (Refer to...) Figure 1 As shown, the method for constructing this NLP-based variant literature interpretation knowledge base may include the following steps:

[0042] Step S101: Obtain disease-related literature.

[0043] Disease-related literature serves as the data source for the database to be built. The more relevant literature obtained, the more comprehensive the information that can be searched. Therefore, as much disease-related literature as possible should be acquired to improve the comprehensiveness of the documentary evidence. After obtaining disease-related literature, document identification information can be assigned to the documents. The diseases can include genetic diseases, etc., without limitation.

[0044] Step S102: Construct a database of entities related to gene variations based on disease-related literature.

[0045] Specifically, this step involves constructing a database based on the disease-related literature obtained in step S101. This database is a database of entities related to gene variations, and its data format can be data lists and dictionaries. The data content can cover information such as literature identification information, gene names, variation names, standard gene terms, and standard variation terms. Specifically, NLP technology can be applied to build an end-to-end model that takes entity names as input and outputs all literature containing both and related entities. This constructs a database of entities related to gene variations, enabling the input of a corresponding entity name to return all relevant literature identification information. That is, inputting any (gene name, variation name) pair will correspond to all corresponding (standard gene terms, standard variation terms) pairs, and further to all equivalent (gene name, variation name) pairs. This allows for the filtering of all relevant literature based on entity meaning, significantly expanding the range of literature that can be searched by accurately matching a single (gene name, variation name) pair, and providing more comprehensive literature reference information for interpreting the variation site.

[0046] In one embodiment, constructing a database of entities related to gene mutations based on disease-related literature includes: constructing an entity extraction model using a portion of the disease-related literature; extracting entities from the remaining disease-related literature using the entity extraction model to obtain entity names; constructing an entity alignment model; aligning the entity names using the entity alignment model to obtain the entity standard terms corresponding to the entity names; and constructing a database of entities related to gene mutations based on the entity names and the entity standard terms corresponding to the entity names.

[0047] In other words, constructing the aforementioned end-to-end model includes building an entity extraction model and an entity alignment model. During the construction process, a small number of documents are first selected from the acquired literature to build an entity extraction model. Then, the entity extraction model is used to extract entities from the remaining documents. The extracted gene and variant names are aligned according to standard gene and variant terminology, respectively, and stored in the database. The gene and variant names to be interpreted are also aligned before literature searching. In this way, various literature descriptions (such as synonyms) that differ from the descriptions of the genes and variants to be interpreted can be used as search objects, enabling the identification of as many relevant documents as possible, thus contributing to the comprehensiveness of the literature evidence.

[0048] In one embodiment, constructing an entity extraction model using a portion of disease-related literature may include: annotating the portion of literature with entities; adding position and entity classification labels to each word in the annotated literature to obtain an entity label sequence; constructing a pre-trained model for the entity extraction model; and adjusting the pre-trained model using the entity label sequence to obtain the entity extraction model.

[0049] Specifically, the entity extraction model can be found in the appendix. Figure 2 The entity extraction model extracts entities from CNLL format text through self-attention pre-training, self-attention fine-tuning, gene dictionary lookup, and mutation pattern matching. Its core is an NLP model based on the self-attention mechanism, which uses transfer learning to achieve the process of taking document text as input and extracting entity names from the document as output.

[0050] Specifically, when constructing an entity extraction model using a subset of disease-related literature, the process begins by labeling a portion of the literature as gene and variant entities to obtain labeled corpus. Position and entity classification labels are then added to the labeled corpus to obtain entity label sequences. Position information can be represented in the form of "BIO" (Begin, Inside, Other) or "BIES" (Begin, Inside, End, Single), while entity classification labels can be gene or variant. Subsequently, a pre-trained model for the entity extraction model is constructed, and this pre-trained model is fine-tuned using the entity label sequences to obtain the final entity extraction model.

[0051] In one embodiment, constructing a pre-trained model for an entity extraction model includes: acquiring a pre-training corpus, which includes relevant literature in the biomedical field; encoding each word in the pre-training corpus to obtain word embedding vectors, fragment embedding vectors, and position embedding vectors; using the sum of the word embedding vectors, fragment embedding vectors, and position embedding vectors as input, and using randomly masked word vectors as labels, pre-training a natural language processing model based on a self-attention mechanism using a backpropagation algorithm to obtain the pre-trained model.

[0052] In one embodiment, constructing a pre-trained model for an entity extraction model further includes: training the pre-trained model using the cross-entropy of the predicted value and the label as a loss function until the loss value output by the loss function meets a preset condition, at which point the pre-trained model training is complete.

[0053] In practical implementation, a pre-trained model of biomedical literature can be constructed first. Figure 3 This is a schematic diagram of an entity extraction model based on a self-attention mechanism according to an embodiment of the present invention, as shown below. Figure 3 As shown, the pre-trained feature extractor is composed of encoders consisting of stacked transformers. Each encoder can have multiple attention heads, and different encoders and attention heads can extract semantic information at different levels from the corpus represented by features. During pre-training, the model can choose between single-sentence prediction or next-sentence prediction (NSP). A sentence in single-sentence prediction can be represented as: [CLS] sentence; a sentence pair in next-sentence prediction is represented as: [CLS] sentence 1 [SEP] sentence 2 [SEP]. In the training of this pre-trained model, the final hidden state with the input [CLS] label is used as a classifier to predict the feature information of a single sentence or whether the two sentences in a sentence pair appear sequentially in the corpus. Each word in the training data sentence is mapped to the sum of its corresponding word embedding vector, fragment embedding vector, and position embedding vector. Randomly masked sub-word vectors are used as labels, and pre-training is performed using the backpropagation algorithm. The encoding vector corresponding to each word obtained from the attention layer is the weight for calculating the relevance of that word to all words, using the formula:

[0054]

[0055] Where Q, K, and V represent Query, Key, and Value, respectively, and L represents the length of the entire input.

[0056] Then, a pre-trained model for entity extraction based on the self-attention mechanism is trained and optimized. Specifically, the cross-entropy of the predicted value and the label can be used as the loss function to train the model. In a specific example, the model can be optimized by adjusting hyperparameters such as the number of network layers, the length of the input sentence, the batch size, and the number of training rounds.

[0057] In one embodiment, training a pre-trained model using entity label sequences to obtain an entity extraction model includes: constructing a fine-tuning model of the entity extraction model based on the pre-trained model; using the model weights obtained during the training of the pre-trained model as the initial weights for the entity extraction task, and using the word embedding vectors corresponding to each word in the labeled documents as inputs, and the positions and entity classification labels corresponding to each word as outputs, and training the fine-tuning model through the backpropagation algorithm to obtain the entity extraction model.

[0058] In one embodiment, the cross-entropy of the predicted value and the label is used as the loss function to train the fine-tuning model until the loss value output by the loss function meets the preset conditions, at which point the fine-tuning model training is complete.

[0059] Specifically, a task fine-tuning model based on entity extraction can be constructed first, such as... Figure 3 As shown, the model weights obtained from the pre-trained model are used as the initial weights for the entity extraction task. The input is the word embedding vector corresponding to each word in the sentence, and the output is the position and entity classification label of each word, such as "B-gene", "I-gene", "B-variant", "I-variant", "O", etc. The fine-tuned model can be learned and adjusted according to the specific task through the backpropagation algorithm to obtain the entity classification label prediction. Then, the fine-tuned model of entity extraction based on the self-attention mechanism is trained and optimized. Specifically, the cross-entropy of the predicted value and the label can be used as the loss function to train the model. In a specific example, the model can be optimized by adjusting hyperparameters such as the number of network layers, the length of the input sentence, the batch size, and the number of training epochs.

[0060] In one embodiment, after extracting entity names from the remaining documents in disease-related literature using an entity extraction model, the method further includes: matching the remaining documents with a preset entity dictionary and / or a preset entity writing pattern to supplement entity names not recognized by the entity extraction model.

[0061] Specifically, it is advisable to obtain a comprehensive dictionary of gene nomenclature and a list of variant writing patterns. The gene nomenclature dictionary can be obtained from publicly available data sources, such as gene-related databases, including but not limited to those of HGNC (HUGO Gene Nomenclature Committee) and NCBI (National Center for Biotechnology Information). After fine-tuning the model training, the test document can be predicted using an optimized entity extraction model. This involves matching and correcting the text against the obtained gene nomenclature dictionary and performing pattern matching on variant entities to supplement entities not identified by the entity extraction model.

[0062] In one embodiment, constructing an entity alignment model may include: obtaining entity standard terms and other entity names corresponding to the entity standard terms, and constructing an entity alignment dictionary based on the entity standard terms and other entity names; and / or, obtaining entity standard terms and constructing a regular expression for entity alignment based on the entity standard terms.

[0063] Specifically, on the one hand, in the process of gene alignment, to make the literature evidence more comprehensive, as many standard gene terms as possible should be obtained when constructing the entity alignment model. These standard gene terms can be selected from public databases, including but not limited to HGNC and NCBI. Correspondingly, the constructed gene alignment dictionary is associated with the obtained standard gene terms and as many aliases and pseudonyms as possible. For example, the gene whose full name is APOBEC1 complementation factor has a series of aliases (such as ACF, ASP, ACF64, ACF65, APOBEC1CF, etc.). When constructing the gene alignment dictionary, A1CF can be selected as the standard gene term, and all its aliases can be used as the names of other genes of the A1CF gene.

[0064] On the other hand, during variant alignment, standard variant terminology can be determined based on the variant rules of the HGVS (Human Genome Variation Society). It should be noted that, unlike gene nomenclature, the diversity of variant nomenclature is mainly reflected in the diversity of variant writing formats. For example, c.1427A>G, 1427A>G, 1427AG, A1427G, and c.DNA1427A>G are all equivalent to c.1427A>G. Therefore, it is preferable to construct a regular expression for variant alignment based on standard variant terminology.

[0065] The regular expression can include, but is not limited to, one or more of the following expressions: c.{Any number of length ≥ 1, and number of symbols ≥ 0}{Any letter of length ≥ 1} > {Any letter of length ≥ 1, and number of symbols ≥ 0}; p. {Any length of letter ≥ 1, with ≥ 0 symbols}{Any length of number ≥ 1}{Any length of letter ≥ 1, with ≥ 0 symbols}; rs{Any length of number ≥ 1}; chr{Any length of letter ≥ 1}-{Any length of number ≥ 1}-{Any length of letter ≥ 1, with ≥ 0 symbols}-{Any length of letter ≥ 1, with ≥ 0 symbols}; n.{Any length of number ≥ 1, with ≥ 0 symbols}{Any length of letter ≥ 1}>{Any length of letter ≥ 1, with ≥ 0 symbols}; IVS.{Any length of number ≥ 1, with ≥ 0 symbols}{Any length of letter ≥ 1}>{Any length of letter ≥ 1, with ≥ 0 symbols}; {Any length of letter ≥ 1}{Any length of number ≥ 1}{Any length of letter ≥ 1}.

[0066] In one embodiment, entity alignment is performed on entity names using an entity alignment model to obtain the entity standard terms corresponding to the entity names, including: performing exact matching and fuzzy matching between the entity names and the entity alignment dictionary to obtain the entity standard terms corresponding to the entity names; and / or, performing exact matching and rule matching between the entity names and regular expressions to obtain the entity standard terms corresponding to the entity names.

[0067] In other words, such as Figure 4 As shown, after the gene alignment dictionary is constructed, all gene names obtained through the entity extraction model are precisely matched with all gene names in the gene alignment dictionary. The gene names that successfully match (gene names, standard gene terms) are saved. Then, gene names that do not successfully match are fuzzily matched with all gene names in the gene alignment dictionary, and the gene names with the highest similarity are saved. For mutation alignment, all mutation names obtained through the entity extraction model are precisely matched with regular expression patterns, and the mutation names that successfully match (mutation names, standard mutation terms) are saved. Then, mutation names that do not successfully match are rule-matched with regular expression patterns, and the mutation names that successfully match the rule are saved.

[0068] Step S103: Construct a knowledge graph of documentary evidence for interpreting variant literature.

[0069] Specifically, since knowledge graphs, as a human-computer-friendly knowledge representation, can vividly highlight logical rules and facilitate knowledge reasoning, this application constructs a knowledge graph of documentary evidence by combining relevant literature content. Taking the construction of an ACMG documentary evidence knowledge graph as an example, the evidence standards of the concise and abstract ACMG variant interpretation guide are concretized and enriched, enabling knowledge reasoning about genes and variants, as well as automated determination of the evidence standards for the genes and variants to be investigated. It should be understood that in the embodiments of this application, the ACMG variant interpretation guide is used as a reference when constructing the knowledge graph and knowledge base, but this application is not limited to this, which is easily understood and accepted by those skilled in the art.

[0070] In one embodiment, constructing a knowledge graph of evidence for interpreting variant documents includes: obtaining the judgment logic of evidence standards or evidence types for interpreting documents in the variant interpretation guidelines; representing the judgment logic in the form of triples, wherein the triple is (entity, the relationship between the entity node and the evidence standard or evidence type, evidence standard or evidence type); and constructing a knowledge graph of document evidence with entities and evidence standards or evidence types as nodes and the relationship between entities and evidence standards or evidence types as edges.

[0071] In practical implementation, taking the ACMG variant interpretation guideline as an example, one can first explore the judgment logic of all evidence standards required for document interpretation in the ACMG variant interpretation guideline, and write it in the form of a triple of (entity, relationship between entity node and evidence standard or evidence type, evidence standard or evidence type). Figure 5 As shown, entities and evidence standards or evidence types correspond to nodes in the ACMG literature evidence knowledge graph, while the relationships between entity nodes and evidence standards or evidence types correspond to edges in the ACMG literature evidence knowledge graph (where Word represents entity standard terms in word form, Phrase represents entity standard terms in sentence form, and ACMG variant evidence criterion represents ACMG literature evidence). Then, the abstract ACMG evidence standards corresponding to each node and edge of the knowledge graph are refined and specified to expand the content of the knowledge graph. For example, the abstract description "pathogenic variant" in the ACMG variant interpretation guide actually corresponds to multiple variants, and a mapping relationship between "pathogenic variant" and multiple variants can be constructed. Finally, based on the refined knowledge graph information, a knowledge graph of all evidence standards required for reading and interpreting literature in the ACMG variant interpretation guide is optimized and constructed, serving as the ACMG literature evidence standards for querying genes and variants.

[0072] Step S104: Extract evidence from the knowledge graph of documentary evidence to obtain evidence corresponding to the entities, and construct a knowledge base for interpreting variant documents based on the evidence and the database.

[0073] In one embodiment, evidence extraction is performed on the literature evidence knowledge graph to obtain evidence corresponding to gene entities and variant entities, and a variant literature interpretation knowledge base is constructed based on the evidence and the database. This includes: extracting sentences containing nodes or node meanings and the sentences preceding and following those sentences from the literature corresponding to the database of entities related to gene variants, generating evidence sentence set information corresponding to the nodes; extracting evidence words indicating relationships from the evidence sentence set information; generating entity standard terms, evidence standards or evidence types, evidence sentences, and evidence words corresponding to the literature based on the evidence sentence set information and evidence words; and constructing a variant literature interpretation knowledge base based on the literature identification information and entity standard terms, evidence standards or evidence types, evidence sentences, and evidence words corresponding to the literature.

[0074] In practice, sentences containing the meanings of all nodes in the knowledge graph of documentary evidence, along with the three sentences preceding and following each sentence, can be extracted from the literature corresponding to entities related to gene mutations in the database. The node meanings and corresponding evidence sentence sets are then saved, and this information can be presented in the form of an evidence sentence set data table. It should be noted that the node meaning here refers to words with equivalent meanings to entity nodes in the knowledge graph of documentary evidence. From the saved node meanings and corresponding evidence sentence sets, further evidence words representing the relationships in the knowledge graph of documentary evidence are extracted. These evidence words characterize the relationship between the entity nodes connected by this edge in the knowledge graph of documentary evidence and the evidence standard or evidence type. In other words, words with equivalent meanings to words representing relationships in the knowledge graph of documentary evidence are all extraction targets. Then, based on the node meanings and corresponding evidence sentence sets, evidence words, and other information such as gene standard terms, variant standard terms, evidence standards or evidence types, and evidence words corresponding to the document, a variant document interpretation knowledge base is finally formed, containing document identification information, gene standard terms, variant standard terms, evidence standards or evidence types, evidence words, and evidence sentences. It should be noted that evidence words can include evidence vocabulary and evidence phrases. The interface of this variant document interpretation knowledge base may, but is not limited to, displaying information such as... Figure 6 As shown.

[0075] During interpretation, inputting any or a batch of (gene name, variant name) pairs will retrieve the corresponding documentary evidence standards or evidence types from the machine-automated literature reading knowledge base built on natural language processing and knowledge graph technologies. The machine-automated literature reading system, based on natural language processing and knowledge graph technologies, can update the variant literature interpretation knowledge base in real time as the literature is updated, efficiently, accurately, and comprehensively providing documentary evidence standards or evidence types related to variant sites.

[0076] The following specific embodiment further illustrates the construction method of the NLP-based variant literature interpretation knowledge base provided in this application, wherein the entities related to gene variants are gene entities and variant entities as examples.

[0077] 1. Obtain 10,000 articles related to diseases such as genetic diseases and assign them document identification information IDs.

[0078] 2. Based on the aforementioned 10,000 documents, a data list (document identification information, gene standard terminology, and variant standard terminology) is constructed. This allows for querying by inputting (gene name, variant name), which will return all relevant document identification information. The specific steps are as follows:

[0079] 2.1 Establishing an entity extraction model mainly includes two stages: building and application, as detailed below:

[0080] 2.1.1 Building an entity extraction model

[0081] (1) A small number of documents (e.g., 500) are labeled with artificial entities, of which the entities include at least two categories: genes and variants.

[0082] Here, we take two sentences from the article DOI: 10.1007 / s10048-011-0299-0 as an example (where the entity names are indicated by {}):

[0083] Sentence No. 1:

[0084] New mutations in the ATM {gene} gene and clinical data of 25 ATpatients.

[0085] Sentence No. 2:

[0086] Analysis of patient derived mRNA by cDNA sequencing confirmed the pathogenic character of c.3285-2A>G {variant}, which results in an insertion of one nucleotide and a frame shift as the consequence (p.Leu1096IlefsX26){variant}.

[0087] (2) Obtain all gene names from public data sources (such as gene-related databases HGNC, NCBI, etc.); for variants, identify all possible writing patterns. In this example, all entity names have more than 100,000 entries, and examples of variant writing patterns are as follows:

[0088] Pattern No.1: "[cgrm]\.[0-9]+[ATCGatcgu]+\>[ATCGatcgu]+$"

[0089] Pattern No.2: "IVS[ATCGatcgu \ / \>\?\(\)\[\]\;\:\*\_\-\0-9]+$"

[0090] Pattern No.3: "[p]\.[CISQMNPKDTFAGHLRWVEYX \ / \>\_\-\+0-9]+$"

[0091] Pattern No. 4: "rs[0-9]+$"

[0092] (3) Process the annotated corpus (documents) from step (1) into cnll format, that is, add a location and entity classification label to each word in the text. Taking Sentence No.1 in the above example as an example, its annotated corpus is processed into cnll format as follows:

[0093] New 0

[0094] mutation 0

[0095] in 0

[0096] the 0

[0097] ATM B-gene

[0098] gene 0

[0099] and 0

[0100] clinical 0

[0101] data 0

[0102] of 0 25 0

[0104] AT 0

[0105] patients 0

[0106] (4) Obtain summaries of a large number of biomedical literatures (such as 500,000 articles in the PubMed database) and use them as pre-training corpus for the entity extraction model. Taking Sentence No.1 in the above example as an example, encode word embedding vectors, fragment embedding vectors, and position embedding vectors for each word in the text "New", "mutation", "in"... "patients" respectively, and use the sum of these vectors as the input vector, and use the randomly masked word vectors as labels to construct a pre-trained model (see Appendix). Figure 3 After the entire text passes through an NLP model based on a self-attention mechanism, it needs to be able to predict the masked text.

[0107] (5) Training and optimizing the pre-trained model for entity extraction based on the self-attention mechanism in step (4). The model is trained using the cross-entropy of the predicted value and the label as the loss function. The model is optimized by adjusting the hyperparameters such as the number of network layers (e.g., 1 to 12 layers), the length of the input sentence (e.g., [128, 256, 512]), the batch size (e.g., [8, 16, 32, 64]), and the number of training rounds (e.g., 5 to 10).

[0108] (6) Based on the pre-trained model framework, construct a fine-tuned model. Use the model weights obtained from the pre-trained model as the initial weights for the entity extraction task. The input is the word embedding vectors corresponding to each word in the literature, and the output is the position and entity classification label corresponding to each word, such as "B-gene", "I-gene", "B-variant", "I-variant", "O", etc. The model learns through the backpropagation algorithm to obtain the label prediction.

[0109] (7) Training and optimization of the fine-tuned model for entity extraction based on the self-attention mechanism in (6). The cross-entropy of the predicted value and the label is used as the loss function to train the fine-tuned model. In this example, the number of network layers, such as 1 to 12 layers, the length of the input sentence, such as [128, 256, 512], the batch size, such as [8, 16, 32, 64], and the number of training rounds, such as 5 to 10, can be adjusted to optimize the model. The F1-score (a metric for classification problems) of the fine-tuned model for extracting genes and variant entities is above 92%.

[0110] 2.1.2 Applying Entity Extraction Model

[0111] (1) Entity extraction prediction is performed on the remaining 9500 documents using the optimized entity extraction model.

[0112] (2) Post-process the obtained prediction values, such as matching and correcting the text with the dictionary of gene names obtained during construction, performing pattern matching for variant names, and supplementing entities not recognized by the entity extraction model.

[0113] 2.2 Establish entity alignment models, mainly including gene alignment and mutation alignment (see appendix for reference). Figure 4 The specific steps are as follows:

[0114] 2.2.1 Gene Alignment

[0115] (1) Obtain gene abbreviations that are jointly recognized by HGNC and NCBI from public databases as standard gene terms.

[0116] (2) Construct an alignment dictionary corresponding to the standard gene terminology and other names of the gene. For example, the "ATM" gene in SentenceNo.1 is a standard gene term. Construct a gene alignment dictionary entry for "ATM", which includes the following entity representations:

[0117] {'ATM', 'TELO1', 'ATD', 'TEL1, telomere maintenance 1, homolog (S.cerevisiae)', 'ATDC', 'ATC', 'ATD', 'AT mutated', 'ATM serine / threoninekinase', 'TEL1', 'ATA', 'serine-protein kinase ATM', 'ataxia telangiectasiamutated (includes complementation groups A, C and D)', 'ataxia telangiectasiamutated', ' ATDC', 'TEL1, telomere maintenance 1, homolog', ' TELO1', 'ATmutated', 'ATE', ' ATC', 'AT1'}

[0118] (3) Perform exact matching between all gene names obtained through the entity extraction model and the standard gene terms in the gene alignment dictionary, and save the matched (gene name, standard gene term). Taking the gene name "ATM" extracted from Sentence No.1 as an example, "ATM" is exactly matched with the gene alignment dictionary, and the matching result is ("ATM", "ATM"), and the result is saved.

[0119] (4) If all the extracted gene names match the gene standard terms in the gene alignment dictionary through the aforementioned step (3), the gene fuzzy matching step is skipped; otherwise, fuzzy matching is required. This fuzzy matching can be string fuzzy matching, and the matching with the highest similarity (gene name, gene standard term) is saved.

[0120] 2.2.2. Mutation Alignment

[0121] (1) Determine the standard writing style of the variant entities and perform exact matching (such as regularized pattern matching) or rule matching. Taking the variant entities “c.3285-2A>G” and “p.Leu1096IlefsX26” in Sentence No.2 above as examples, determine the following two writing formats for these two variant entities respectively:

[0122] c. {Any length digit with length >= 1, may contain signs} {Any length letter with length >= 1} > {Any length letter with length >= 1, may contain signs};

[0123] p. {Letters of any length >= 1, including symbols} {Numbers of any length >= 1} {Letters of any length >= 1, including symbols};

[0124] Specifically, the mutated entity "c.3285-2A>G" begins with "c." and its alignment format conforms to the pattern c.{arbitrary length of numbers with length >= 1, including signs}{arbitrary length of letters with length >= 1}>{arbitrary length of letters with length >= 1, including signs}. Therefore, it can be aligned to "c.3285-2A>G", and the matching result ("c.3285-2A>G", "c.3285-2A>G") is saved. The mutated entity "p.Leu1096IlefsX26" begins with "p." and its alignment format conforms to p. The pattern {letters of any length >= 1, including symbols}{digits of any length >= 1}{letters of any length >= 1, including symbols} is then used to replace "X" with "*" according to the variation naming rules. Therefore, it can be aligned to "p.Leu1096Ilefs*26", and the matching result ("p.Leu1096IlefsX26", "p.Leu1096Ilefs*26") is saved.

[0125] 2.3 Construct a database of genes and variant entities, which includes, but is not limited to, the following four data storage units:

[0126] I. Dictionary of {Gene Terminology: Standard Gene Terms}

[0127] II {Variation Terminology: Standard Terms for Variation} Dictionary;

[0128] III (Document Identification Information, Gene Standard Terminology, Variation Standard Terminology) Data List;

[0129] IV (Document Identification Information, Gene Name, Variation Name) Data List.

[0130] Specifically, taking the gene designation "ATM" in Sentence No. 1 and the variant designations "c.3285-2A>G" and "p.Leu1096IlefsX26" in Sentence No. 2 as examples, the construction of Table III is explained: Assuming the document identification information of the article DOI: 10.1007 / s10048-011-0299-0 is 1, "ATM" (gene designation) corresponds to "ATM" (gene standard term), "c.3285-2A>G" (variation designation) corresponds to "c.3285-2A>G" (variation standard term), and "p.Leu1096IlefsX26" (variation designation) corresponds to "p.Leu1096Ilefs*26" (variation standard term), thus obtaining the entries in Table III (1, "ATM", ["c.3285-2A>G", "p.Leu1096Ilefs*26"]).

[0131] Therefore, by inputting any (gene name, variant name) pair, one can find all corresponding (gene standard term, variant standard term) pairs, and then all equivalent (gene name, variant name) pairs. This allows for the filtering of all relevant literature based on entity meaning, greatly expanding the range of literature that can be searched by accurately matching a single (gene name, variant name) pair. This is beneficial for providing more comprehensive literature reference information for the variant sites to be interpreted.

[0132] 3. Constructing a knowledge graph of ACMG literature evidence

[0133] 3.1 Taking the PS3 evidence standard in the ACMG variant interpretation guide as an example, we will explore its judgment logic. The PS3 evidence standard is as follows:

[0134] PS3: Well-established in vitro or in vivo functional studiessupportive of a damaging effect on the gene or gene product

[0135] For the variant to be queried, the logic for determining the PS3 evidence standard can be written as: (Well-established invitro or in vivo functional studies, supportive of a damaging effect on the gene or gene product, PS3).

[0136] 3.2 The abstract terms of PS3 in step 3.1 are concretized to refine the knowledge graph and facilitate entity linking. For example, "Well-established in vitro or in vivo functional studies" in the PS3 evidence standard can be expanded into many functional testing methods, including "cDNA sequencing" in Sentence No. 2 above, while "supportive of a damaging effect on the gene or gene product" can be expanded into various "damaging" features, such as "pathogenic" and "insertion of one nucleotide and a frame shift" in Sentence No. 2. Therefore, Sentence No. 2 provides the following triples for the ACMG literature evidence knowledge graph required to determine ("ATM", "c.3285-2A>G") as the PS3 evidence standard:

[0137] ("all established in vitro or in vivo functional studies including cDNA sequencing", "all damaging effects on the gene or gene product includingpathogenic / insertion of one nucleotide / frame shift", PS3)

[0138] 4. Extract evidence words and sentences, and then construct a knowledge base for interpreting variant literature.

[0139] 4.1 Taking the judgment ("ATM", "c.3285-2A>G") as an example from the article DOI: 10.1007 / s10048-011-0299-0:

[0140] First, by extracting the node "cDNA sequencing" that connects the ACMG literature evidence knowledge graph constructed in step 3.2 to the PS3 evidence standard, we can obtain Sentence No. 2 and its three preceding and following sentences, thereby generating evidence sentence set information. This evidence sentence set information can be represented in the form of the data table {"cDNA sequencing": evidence sentence set} as follows:

[0141] {"cDNA sequencing": ("Analysis of patient derived mRNA by cDNAsequencing confirmed the pathogenic character of c.3285-2A>G, which resultsin an insertion of one nucleotide and a frame shift as the consequence(p.Leu1096IlefsX26).", "We found a previously described nonsense mutation,c.362T>A (pLeu362X), together with a new change, c.4110-9C>G, as thepotential second disease causing mutation in a Russian patient.", "Furtheranalysis of the latter alteration revealed that it activates a cryptic splicesite resulting in an mRNA containing eight additional bases leading to aframe shift and thus confirming the transversion c.4110C>G as a pathogenicmutation.","A further new splice site mutation, c.3285-2A>G, was found inpatient 14587 in combination with a known protein truncating mutation on theother allele.", "This alteration was not detected in 294 alleles fromunaffected controls.", "We found a not previously described deletion of 5 bpin exon 57 of the ATM gene, c.5260_5264delAAGAT, in a Turkish patient which can be predicted to lead to a frame shift and premature termination of protein translation (p.Lys1754AspfsX13).", "This mutation was accompanied by a missense mutation, c.6047A>G (p.Asp2016Gly), which was previously shown to be pathogenic due to a dramatically reduced ATM protein level in a patienthomozygous for this mutation and with a protracted disease course.")}.

[0142] 4.2 Further extract evidence terms representing head node characteristics from the data table generated in step 4.1, connecting the head node "cDNA sequencing" and the tail node PS3 in the ACMG literature evidence knowledge graph. The results include "pathogenic" and "insertion of one nucleotide frame shift", thus generating the PS3 evidence standard "c.3285-2A>G" corresponding to this article:

[0143] {"c.3285-2A>G": [("cDNA sequencing", "pathogenic", PS3), ("cDNAsequencing", "insertion of one nucleotide", PS3), ("cDNA sequencing", "frameshift", PS3)]}.

[0144] Finally, the data table (1, "ATM", "c.3285-2A>G", PS3, ["cDNA sequencing", "pathogenic", "insertion of one nucleotide", "frame shift"], ["Analysis of patient derived mRNA by cDNA sequencing confirmed the pathogenic character of c.3285-2A>G, which results in an insertion of one nucleotide and a frameshift as the consequence (p.Leu1096IlefsX26).") consisting of literature identification information, gene standard terminology, variant standard terminology, ACMG evidence criteria, evidence words, and evidence sentences) is stored in the ACMG variant literature interpretation knowledge base. Its interface is shown as follows: Figure 6 As shown.

[0145] During interpretation, inputting ("ATM", "c.3285-2A>G") will retrieve a series of results from an automated reading literature knowledge base built on natural language processing and knowledge graph technology. This includes the following knowledge: (1, "ATM", "c.3285-2A>G", ["cDNA sequencing", "pathogenic", "insertion of one nucleotide", "frameshift"], ["Analysis of patient derived mRNA by cDNA sequencing confirmed the pathogenic character of c.3285-2A>G, which results in an insertion of one nucleotide and a frame shift as the consequence (p.Leu1096IlefsX26)."]). The ACMG evidence standard is "PS3".

[0146] Therefore, by inputting any or a batch of (gene name, variant name) pairs, the corresponding ACMG evidence standard can be obtained from the machine-automated literature reading knowledge base built on natural language processing and knowledge graph technology.

[0147] According to the present invention, a method for constructing a variant literature interpretation knowledge base based on NLP involves acquiring disease-related literature, constructing a database of entities related to gene variants based on this literature, constructing an evidence knowledge graph for variant literature interpretation, and extracting evidence from the knowledge graph to obtain evidence corresponding to gene and variant entities. The variant literature interpretation knowledge base is then constructed based on this evidence and the database. Therefore, this application represents a method for automatically reading literature and obtaining disease literature evidence using natural language processing and knowledge graph technologies. This method allows for more comprehensive and systematic literature evidence. During interpretation, inputting any entity name automatically returns results based on variant evidence standards or evidence types from the literature reading, thus automating and intelligently acquiring disease variant literature evidence. This effectively improves the speed of gene variant-related interpretation, achieving second-level return of literature search results for the gene and variant in question, significantly improving the efficiency of literature search, and consequently enhancing the quality and efficiency of gene testing report interpretation.

[0148] Figure 7 This is a flowchart of a variation-based literature interpretation method according to an embodiment of the present invention. (Reference) Figure 7 As shown, this NLP-based method for interpreting variant literature may include the following steps:

[0149] Step S201: Obtain the name of the entity to be interpreted.

[0150] Step S202: Input the entity name into the variant document interpretation knowledge base to obtain the evidence standard or evidence type, evidence sentence and evidence word corresponding to the entity name. The variant document interpretation knowledge base is constructed according to the above-mentioned NLP-based variant document interpretation knowledge base construction method.

[0151] For example, a variant document interpretation knowledge base can be constructed based on the aforementioned method for building an NLP-based variant document interpretation knowledge base. This knowledge base can then be used to build a machine-automated document reading system. This system can include a human-computer interface. When a user needs to interpret an entity name, they can input the entity name to be interpreted into the machine-automated document reading system through the human-computer interface. The system then queries the knowledge base based on the entity name to be interpreted, thereby obtaining the corresponding evidence standards or evidence types, evidence sentences, and evidence words, and provides this information back to the interpreter through the human-computer interface.

[0152] According to an embodiment of the present invention, the NLP-based method for interpreting variant literature obtains the entity name to be interpreted and inputs it into a variant literature interpretation knowledge base to obtain the corresponding evidence standards or evidence types, evidence sentences, and evidence words. The variant literature interpretation knowledge base is constructed according to the aforementioned NLP-based variant literature interpretation knowledge base construction method. Therefore, by inputting the entity name, the corresponding evidence standards or evidence types, evidence sentences, and evidence words can be automatically obtained, enabling automation and intelligence in obtaining disease variant literature evidence. This effectively improves the speed of interpreting gene variant-related evidence and provides more comprehensive literature evidence, thereby improving the quality and efficiency of gene testing report interpretation.

[0153] Figure 8 This is a structural block diagram of an electronic device according to an embodiment of the present invention. (Reference) Figure 8 As shown, the electronic device includes a memory 301, a processor 302, and a variant document interpretation program stored in the memory 301 and capable of running on the processor 302. When the processor 302 executes the variant document interpretation program, it implements the above-mentioned NLP-based variant document interpretation method.

[0154] It should be noted that for the description of the electronic equipment in this application, please refer to the description of the NLP-based variant literature interpretation method in this application, which will not be repeated here.

[0155] According to the electronic device of the present invention, when the processor executes the variant literature interpretation program, the above-mentioned NLP-based variant literature interpretation method is realized. Thus, by inputting entity names, the corresponding evidence standards or evidence types, evidence sentences and evidence words can be automatically obtained. This can realize the automation and intelligence of obtaining disease variant literature evidence, effectively improve the interpretation speed of gene variant related documents, and make the literature evidence more comprehensive, thereby improving the quality and efficiency of gene testing report interpretation.

[0156] It should be noted that the logic and / or steps represented in the flowchart or otherwise described herein, for example, can be considered as a sequenced list of executable instructions for implementing logical functions, and can be specifically implemented in any computer-readable medium for use by, or in conjunction with, an instruction execution system, apparatus or device (such as a computer-based system, a processor-included system or other system that can fetch and execute instructions from, an instruction execution system, apparatus or device).

[0157] It should be understood that various parts of the present invention can be implemented in hardware, software, firmware, or a combination thereof. In the above embodiments, multiple steps or methods can be implemented in software or firmware stored in memory and executed by a suitable instruction execution system. For example, if implemented in hardware, as in another embodiment, it can be implemented using any one or a combination of the following techniques known in the art: discrete logic circuits having logic gates for implementing logical functions on data signals, application-specific integrated circuits (ASICs) having suitable combinational logic gates, programmable gate arrays (PGAs), field-programmable gate arrays (FPGAs), etc.

[0158] In the description of this specification, references to terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of the invention. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples.

[0159] Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Thus, a feature defined as "first" or "second" may explicitly or implicitly include at least one of that feature. In the description of this invention, "a plurality of" means at least two, such as two, three, etc., unless otherwise explicitly specified.

[0160] In this invention, unless otherwise explicitly specified and limited, the terms "installation," "connection," "linking," and "fixing," etc., should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral part; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; they can refer to the internal communication of two components or the interaction between two components, unless otherwise explicitly limited. Those skilled in the art can understand the specific meaning of the above terms in this invention according to the specific circumstances.

[0161] Although embodiments of the present invention have been shown and described above, it is understood that the above embodiments are exemplary and should not be construed as limiting the present invention. Those skilled in the art can make changes, modifications, substitutions and variations to the above embodiments within the scope of the present invention.

Claims

1. A method for constructing a knowledge base for interpreting variant literature based on NLP, characterized in that, Includes the following steps: Obtain disease-related literature; A database of entities related to gene mutations was constructed based on the aforementioned disease-related literature; Construct a knowledge graph of documentary evidence for interpreting variant literature; Evidence is extracted from the document evidence knowledge graph to obtain evidence corresponding to the entity, and a variant document interpretation knowledge base is constructed based on the evidence and the database. The database of entities related to gene variations constructed based on the disease-related literature includes: An entity extraction model was constructed using a portion of the disease-related literature. The entity extraction model is used to extract entities from the remaining documents in the disease-related literature to obtain entity names. Construct an entity alignment model; The entity name is aligned using the entity alignment model to obtain the entity standard terminology corresponding to the entity name. A database of entities related to gene variation is constructed based on the entity names and the corresponding entity standard terms.

2. The method for constructing a knowledge base for interpreting variant literature based on NLP according to claim 1, characterized in that, The construction of an entity extraction model using a subset of the disease-related literature includes: Entity annotation was performed on the aforementioned documents; Add position and entity classification labels to each word in the annotated document to obtain an entity label sequence; Construct a pre-trained model of the entity extraction model; The pre-trained model is adjusted using the entity label sequence to obtain the entity extraction model.

3. The method for constructing a knowledge base for interpreting variant literature based on NLP according to claim 2, characterized in that, The pre-trained model for constructing the entity extraction model includes: Obtain pre-training corpus, which includes literature related to the biomedical field; Each word in the pre-training corpus is encoded to obtain word embedding vectors, fragment embedding vectors, and position embedding vectors. The sum of the word embedding vector, the fragment embedding vector, and the position embedding vector is used as input, and the randomly masked word vectors are used as labels. A natural language processing model based on a self-attention mechanism is pre-trained using the backpropagation algorithm to obtain the pre-trained model.

4. The method for constructing a knowledge base for interpreting variant literature based on NLP according to claim 3, characterized in that, The pre-trained model for constructing the entity extraction model further includes: The pre-trained model is trained by using the cross-entropy of the predicted value and the label as a loss function until the loss value output by the loss function meets the preset conditions, at which point the pre-trained model is complete.

5. The method for constructing a knowledge base for interpreting variant literature based on NLP according to any one of claims 2-4, characterized in that, The step of training the pre-trained model using the entity label sequence to obtain the entity extraction model includes: A fine-tuning model of the entity extraction model is constructed based on the pre-trained model; The model weights obtained during the training of the pre-trained model are used as the initial weights for the entity extraction task. The word embedding vectors corresponding to each word in the entity-annotated literature are used as inputs, and the positions and entity classification labels corresponding to each word are used as outputs. The fine-tuned model is trained through the backpropagation algorithm to obtain the entity extraction model.

6. The method for constructing a knowledge base for interpreting variant literature based on NLP according to claim 5, characterized in that, The cross-entropy of the predicted value and the label is used as the loss function to train the fine-tuning model until the loss value output by the loss function meets the preset conditions, at which point the fine-tuning model training is complete.

7. The method for constructing a knowledge base for interpreting variant literature based on NLP according to claim 1, characterized in that, After extracting entity names from the remaining documents in the disease-related literature using the entity extraction model, the process further includes: The remaining documents are matched with a preset entity dictionary and / or a preset entity writing pattern to supplement entity names that are not recognized by the entity extraction model.

8. The method for constructing a knowledge base for interpreting variant literature based on NLP according to claim 1, characterized in that, The construction of the entity alignment model includes: Obtain entity standard terms and other entity names corresponding to the entity standard terms, and construct an entity alignment dictionary based on the entity standard terms and the other entity names; and / or, Obtain entity standard terms and construct a regular expression for entity alignment based on the entity standard terms.

9. The method for constructing a variant literature interpretation knowledge base based on NLP according to claim 8, characterized in that, The regular expression includes one or more of the following expressions: c. {Any length of 1 digit with ≥1 number of symbols ≥0} {Any length of 1 letter with ≥1 number of symbols ≥0} > {Any length of 1 letter with ≥1 number of symbols ≥0}; p. {Any length of letters with a length ≥ 1 and a number of symbols ≥ 0} {Any length of numbers with a length ≥ 1} {Any length of letters with a length ≥ 1 and a number of symbols ≥ 0}; rs{Any number of length ≥ 1}; chr{letters of any length ≥ 1}-{digits of any length ≥ 1}-{letters of any length ≥ 1 with ≥ 0 symbols}-{letters of any length ≥ 1 with ≥ 0 symbols}; n. {Any number of length ≥ 1, with ≥ 0 symbols} {Any letter of length ≥ 1} > {Any letter of length ≥ 1, with ≥ 0 symbols}; IVS.{Any length of 1 digits with ≥1 number of symbols ≥0}{Any length of 1 letter}>{Any length of 1 letter with ≥1 number of symbols ≥0}; {Any length of letters with a length greater than or equal to 1}{Any length of numbers with a length greater than or equal to 1}{Any length of letters with a length greater than or equal to 1}.

10. The method for constructing a variant literature interpretation knowledge base based on NLP according to claim 8, characterized in that, The step of aligning the entity names using the entity alignment model to obtain the entity standard terms corresponding to the entity names includes: Perform exact and fuzzy matching between the entity names and the entity alignment dictionary to obtain the entity standard terms corresponding to the entity names; and / or, The entity name is matched precisely and by rules with the regular expression to obtain the entity standard terminology corresponding to the entity name.

11. The method for constructing a knowledge base for interpreting variant literature based on NLP according to claim 1, characterized in that, The database of entities related to gene variation includes entity names: an entity standard terminology dictionary, document identification information: an entity standard terminology data list, and document identification information: an entity name data list.

12. The method for constructing a knowledge base for interpreting variant literature based on NLP according to claim 11, characterized in that, The constructed document evidence knowledge graph for interpreting variant documents includes: Obtain the evidence standards or evidence type determination logic used in the variation interpretation guide for document interpretation; The judgment logic is represented in the form of a triple, wherein the triple is: entity, the relationship between entity and evidence standard or evidence type, and evidence standard or evidence type; The document evidence knowledge graph is constructed using the entity and the evidence standard or evidence type as nodes and the relationship between the entity and the evidence standard or evidence type as edges.

13. The method for constructing a knowledge base for interpreting variant literature based on NLP according to claim 12, characterized in that, The step of extracting evidence from the document evidence knowledge graph to obtain evidence corresponding to the entity, and constructing a variant document interpretation knowledge base based on the evidence and the database, includes: Extract sentences containing the node or the meaning of the node, as well as the sentences before and after the node, from the literature corresponding to the database of entities related to gene variation, and generate a set of evidence sentences corresponding to the node; Extract evidence words representing the relationship from the evidence sentence set information; Generate entity standard terms, evidence standards or evidence types, evidence sentences and evidence words corresponding to the document based on the evidence sentence set information and the evidence words; The variant document interpretation knowledge base is constructed based on the document identification information and the entity standard terminology, evidence standard or evidence type, evidence sentence and evidence word corresponding to the document.

14. The method for constructing a knowledge base for interpreting variant literature based on NLP according to claim 1, characterized in that, The entities include one or more of genes, variants, drugs, diseases, and phenotypes.

15. A method for interpreting variant literature based on NLP, characterized in that, Includes the following steps: Obtain the name of the entity to be interpreted; The entity designation is input into the variant document interpretation knowledge base to obtain the evidence standard or evidence type, evidence sentence and evidence word corresponding to the entity designation, wherein the variant document interpretation knowledge base is constructed according to the construction method of the variant document interpretation knowledge base as described in any one of claims 1-14.

16. An electronic device, characterized in that, The method includes a memory, a processor, and a variant document interpretation program stored in the memory and executable on the processor. When the processor executes the variant document interpretation program, it implements the NLP-based variant document interpretation method as described in claim 15.

Images