KGHC: a knowledge graph for hepatocellular carcinoma

The schema of KGHC comes from the Unified Medical Language System (UMLS) [14]. UMLS is a metathesaurus, the largest collection of biomedical dictionaries containing 2.9 million entities and 11.4 million entity names and synonyms [15]. UMLS contains complex taxonomy, including physical objects, events, or even medical equipments. Since our knowledge graph focuses on hepatocellular carcinoma, we only choose parts of UMLS related to our work. By dividing the taxonomy of UMLS, we have summarized nine concepts for our knowledge graph driven by the requirements of analyzing. The concepts of our knowledge graph include: drug, DNA, RNA, gene, protein, cell, disease, phenotypic abnormality and therapeutic-technique. The concept of drug in our knowledge graph include chemical. Then, according to the concepts, we filter the relationship from the UMLS Semantic Network. There are totally 22 relationships in KGHC.

According to the main source of biomedical information at present, we extracted the data from unstructured data and structured data. In data extraction section, we first used SemRep to extract entities, attributes and relations from unstructured information in literature and Internet. Then, we extracted the triples from structured information in SemMedDB. Finally, we applied BioIE filter out the noise in extracted data. The details are described as follows.

The entities which extracted by SemRep may contain some noise. Taking the extracted entities are incomplete as an instance, it always affects the quality of the knowledge graph. For example, B7–1 gene is recognized as gene, and quinone reductase (QR) induction is recognized as induction. To ensure the high-quality of the data in our built knowledge graph, we applied BioIE to automatically extract the multi-type entities from biomedical literature (such as disease, drug, protein, gene, DNA, RNA and cell).

We proposed the Attention-based named entity recognition model (Att-BiLSTM-CRF) [27] and used it in BioIE. Compared with the traditional BiLSTM-CRF [28] model, Att-BiLSTM-CRF can solve the problem of the inconsistent labels. Attention mechanism in the model is used to learn contextualized embedding, and it can ensure the consistency of entity label and the accuracy of entity recognition. Figure 2 shows the architecture of the Att-BiLSTM-CRF model. A document D = (X₁, …, X_t, …, X_m) containing m sentences as an input, and each sentence is expressed as (x₁, …, x_t, …, x_n), where n is the number of the words in sentence [29]. The first layer of the model is the embedding layer, which the concatenation of the character embedding, word embedding and addition features (i. e., POS information and chunking information, et al.) as input is fed into the BiLSTM layer. The BiLSTM layer is used to extract sentence features automatically. It is consisted of a forward LSTM which computes a representation \( {\overrightarrow{h}}_t \) of the sequence from left to right, and a backward LSTM which computes a representation \( {\overleftarrow{h}}_t \) of the same sequence in reverse [30]. And the concatenation of \( {h}_t=\left[{\overrightarrow{h}}_t;{\overleftarrow{h}}_t\right] \) is the output of the BiLSTM layer.

In attention layer, we apply the attention mechanism to focuses on the related tokens in the different sentences of a document to address the tagging inconsistency problem. Specifically, the attention layer is used to capture similar word attention at the document-level. The attention matrix A is used to calculate the similarity between the current target word and all words in the document. The attention matrix A, which can be described as a_{i, j}, can be computed by formula (1).

The similarity between x_i and x_j can be calculated by the following four alternatives, (i.e. manhattan distance, euclildean distance, cosine distance and perceptron), where W is a weight matrix [27].

Then formula (3) is used to calculate a document-level global vector G, where H is the output of the BiLSTM layer.

Next, to predict confidence scores for the word, a tanh layer is constructed on top of the attention layer. At last, instead of decoding each label independently, the CRF layer is added to decode the best tag path in all possible tag paths. We trained the model on the dataset of CDR and NLPBA. And to verify the effectiveness of the model, we selected a single category (chemical compound) recognition on the CHEMDNER dataset provided by BioCreative IV for comparative experiments. As shown in Table 2, our method achieves an F-score of 90.84% with no addition feature engineering, which is a state-of-the-art result.

In this work, BioIE is used to extract the entities and attributes of hepatocellular carcinoma from structured data and unstructured data. Specifically, given a sentence which has been extracted the triples from structure data and unstructured data, BioIE extracts the entities and attributes (as shown in Table 1) from this sentence. Figure 3 shows an example of BioIE. However, different data extraction method extracts different information. Take the sentence “Combined modality doxorubicin-based chemotherapy and chitosan-mediated p53 gene therapy using double-walled microspheres for treatment of human hepatocellular carcinoma.” as example, BioIE extracts the entities p53 gene and hepatocellular carcinoma and SemRep extracts the triples associated_with(chemotherapy, hepatocellular carcinoma) and associated_with(gene, hepatocellular carcinoma) from sentence. So it is important to align entities and triples between BioIE and SemRep. Therefore, we proposed a rule-based filter method.

We used SemRep and BioIE to extract the entities and attributes from the same sentence. If the text names of entities in triple are extracted by both SemRep and BioIE, and the attributes of entities (including text name, entity_start_index, entity_end_index, sentence and source) are the same, this triple is retained. Otherwise, it is removed. For example, in the above sentence, SemRep extracts the entities gene and hepatocellular carcinoma and BioIE extracts the entities p53 gene and hepatocellular carcinoma. The attributes of entity are different. So we removed the triples associated_with (gene, hepatocellular carcinoma). The entity filter method can filter the extracted data in KGHC. Corresponding, in order to ensure the accuracy of the extracted relationship, we manually filter the relations between the entities after using BioIE to filter the data.

The data which we extracted from structured and unstructured data have some noise, such as redundant, complementary, and sometimes have conflicts on some values. To ensure the accuracy of the data in the knowledge graph, we fused the data in two steps: entity mapping and entity alignment. For entity mapping, the same entity has different entity names or the same entity name represents different substances. For example, both HCC and Hepatocellular Carcinoma denote the disease hepatocellular carcinoma. In this work, we extracted the standard name and text name of the entity from SemRep and BioIE (the text name is the mention of the entity in sentence and the entity standard name is the preferred name of the entity). We use the standard name of entity as the entity name, and use the text name as an attribute of the entity to map entity. Take the sentence “Alcohol can cause HCC” as an example. HCC is the text name and Hepatocellular Carcinoma is the standard name. We use Hepatocellular Carcinoma as entity name and use HCC as an attribute of the entity.

For entity alignment, in data filter, we used BioIE to filter the entities and attributes. The entity name (standard name) and attributes (entity type) extracted by SemRep and BioIE may be different. For example, in the above sentence, the entity name of HCC is Primary carcinoma of the liver cells in SemRep and Hepatocellular Carcinoma in BioIE. We use the JaccardSimilarity [34] to calculate the similarity of entity name between BioIE and SemRep. Only when the value of JaccardSimilarity is 1, is the triple retained. Otherwise, we manually checked the consistency of entity names. We adopted a voting strategy to solve the inconsistencies (e.g. the ones of entity type), i.e., for a given entity, we tend to trust the type which has the most support.

The main storage forms of knowledge graph include Resource Description Framework (RDF) and graph database. RDF can establish links between data and query [35] through Sparql. Graph database which can store entities and relations of knowledge graph in the form of graph. Compared with the RDF, the graph database has a better readability. Neo4j is a graph database which can query and update data by using graph query language Cypher. And it provides REST structure, which can be integrated into environments based on PHP, NET, Python and JavaScript [36].

In this work, Neo4j is used to store the data. We imported the triples into Neo4j through the Neo4j-import tool. Figure 4 is a partial display of KGHC. When KGHC is opened in Neo4j, a network is displayed, in which the nodes refer to the entities and the edges refer to the relations between the entities. The biomedical researchers can use Cypher to search the entities and relations.

In data extraction section, we obtained a two-level relation knowledge graph. For example, hepatocellular carcinoma has a relationship with Hepatitis A. We extracted data from the structure data and unstructured data that is related to Hepatitis A and find that Glucagon has a relationship with Hepatitis A. So the Glucagon may be related to hepatocellular carcinoma. It may help biomedical researchers discover the substances related to hepatocellular carcinoma.

In addition, KGHC also contains a large number of attribute information. Selecting a node or edge in the network, users can see the detailed information of attributes about the triple at the bottom of the interface, as shown in Fig. 5. The detailed description of attribute is shown in Table 1. When biomedical researchers propose research hypotheses, they can obtain relevant research articles through the PMID provided by KGHC. In our opinion, KGHC can support the analysis of the hepatocellular carcinoma network and may facilitate the discovery of the molecular mechanisms behind the it.

KGHC is stored in the form of triples. It has 5028 entities and 13,296 triples. Specifically, there are 1328 drugs, 1849 proteins, 1403 diseases, 160 cells, 140 DNAs, 54 phenotypic abnormalities, 50 genes, 35 therapeutic-techniques and 9 RNAs (as shown in Fig. 6).

In addition, KGHC contains 799 triples directly related to hepatocellular carcinoma, and 12,497 triples indirectly related to it. The direct relation contains 682 entities, and the indirect relation contains 4726 entities as shown in Fig. 7. The researchers can use the direct and indirect relations to propose new hypotheses. Figure 8 shows the input source of the knowledge graph. It contains four parts: 46,172 sentences of literature, 1084 sentences of UpToDate, 5275 sentences of ClinicalTrials.​gov and 109,875 sentences of SemMedDB. Through analyzing the data, we found the following facts,

For KGHC, its data quality is of great importance. However, there is no hepatocellular carcinoma gold set currently. In data filter process, we filtered the entities and attributes by BioIE and checked the relation manually. Therefore, to assess the consistency of the entities and relations, we measured the pairwise agreement of duplicate annotations using the Jaccard score [37].

If we defined A as the set of annotations of team A, B as the set of annotations of team B, then the Jaccard agreement score could be calculated by counting the number of agreements and disagreements. If a triple relation is true, it is counted as a case of agreement. Take “Alcohol can cause hepatocellular carcinoma.” as an example. The extracted triple is cause (alcohol, hepatocellular carcinoma). If one annotator annotates this triple true, another annotates it false, then that would count as a case of disagreement. Formula (4) shows the formula of Jaccard score. We used the simple random sampling method to draw 1000 triples from our knowledge graph to calculate the consistency of the triples (i.e., entity and relation) manually. The accuracy ratio is 81.20%.