PotatoG-DKB: a potato gene-disease knowledge base mined from biological literature

Preprocessing

The data sources analyzed in this article were retrieved from PubMed (https://pubmed.ncbi.nlm.nih.gov/), a search engine maintained by the National Center for Biotechnology Information (NCBI) at the National Library of Medicine (NLM; Wei et al. (2019)). For this study, “potato and gene and disease” was used as the search string in PubMed, and 2,906 final abstracts were selected from the search results. All sources were saved in the PubMed_ID (PMID) format. PubTator Central provides automatic annotations of biological concepts from PubMed abstracts. The PubTator API (https://www.ncbi.nlm.nih.gov/research/pubtator/api.html) provides source code downloads in Perl, Python, and Java, and the annotation formats include PubTator, biocxml, and biocjson. PubTator was used to annotate the obtained sources in PubTator format, including titles, abstracts, and annotations of biological entities, saved in “.txt” format. We used rule-based syntax in sentence segmentation to avoid inaccuracies associated with punctuation-based splitting (https://docs.python.org/3/library/re.html). For example, rule-based syntax prevents the incorrect segmentation of “B. lactucae regulatory sequences in P. infestans” into “B. | lactucae regulatory sequences in P. | infestans.” A dictionary of common potato pests and diseases was then used with SpaCy’s pattern-matching functionality to expand the entity terms to enhance named entity recognition (https://spacy.io/). All sentences not containing entity information were removed, leaving a total of 21,995 sentences containing at least biological two entities.

Named entity recognition

Knowledge extraction, a crucial aspect of knowledge base construction, encompasses both named entity recognition (NER) and relationship extraction (RE). NER in the field of biology primarily targets genes, diseases, chemicals, species, and mutations. PubTator Central (PTC; Wei et al. (2019)) was used in this study for biological entity recognition (e.g., genes, diseases, chemicals, species, mutations, and cell lines). PTC, a web-based system, offers a multi-entity recognition interface—PubTator Restful API (https://www.ncbi.nlm.nih.gov/research/pubtator/api.html)—to export annotations. The named entity recognition annotations are displayed in Table 2. PMID denotes PubMed’s unique identifier, “start” and “end” indicate the boundaries of mentions in the abstract, “mention” refers to the biological entity, “category” specifies the type of biological entity, and “standard” refers to standardized identification information of the biological entities.

In PubTator, diseases are identified based on Disease Ontology (Schriml et al., 2012), which is a database of human diseases. Because of this, PubTator’s accuracy in recognizing crop-related diseases is relatively low. To address this issue, we enhanced disease and pest entity recognition using dictionary-based and pattern-matching methods, implemented through SpaCy’s entity recognition tool (https://spacy.io/). SpaCy is a free, open-source library for advanced NLP in Python.

PubTator classifies pathogen information for potatoes under “species.” To provide a more detailed representation of potato diseases and pests, we further categorized potato pathogens into fungal pathogens, bacterial pathogens, viral pathogens, and pests. Additional entities, such as abiotic stress and chromosome, were incorporated as needed to enrich the relationships within PotatoG-DKB. Ultimately, we identified data for 22 categories of entities. The detailed information on these categories is shown in Table 3.

Entities are represented in various forms in the literature, including abbreviations, word variants, and synonyms. For instance, “Phytophthora infestans” can also be referred to as “P. infestans”, “Phytophothora”, “Pi”, or “Phytophthora infestans (Mont.) de Bary”. Therefore, unique identifiers had to be established for each entity. Unique identifiers for entities such as genes, proteins, chemicals, and mutations were sourced from PubTator, while identifiers for species, fungi, bacteria, viruses, and pests were obtained from the Taxonomy database (https://www.ncbi.nlm.nih.gov/taxonomy). Entities without unique identifiers are represented as ‘--’. A potato-related entity dictionary was then constructed based on these entities and their categories. Examples from the entity dictionary are shown in Table 4, where “mention” indicates a meaningful item identified in potato-related literature, “category” refers to the type of entity, and “unique identifier” is the unique identifier of the entity in the NCBI database.

Relation extraction

Relation extraction (RE) is a subdomain of syntactic analysis aimed at revealing the semantic relationships between entities in unstructured text. In natural language processing, RE is typically treated as a supervised classification task. To achieve high accuracy in relationship extraction, a large amount of training data is required. Unfortunately, there is currently no specialized biological text corpus specific to potatoes. With the introduction of ChatGPT (OpenAI, 2022), large language models based on generative frameworks have shown promising performance across various tasks. In this study, LLM was used to extract triplet relationships. To enhance the efficiency of relationship extraction, we retained sentences containing at least two recognized entities. By constructing feature templates, we extracted the relevant relationships between entities using the Spark Language Model V3.5, a large language model released by iFlytek (https://xinghuo.xfyun.cn/sparkapi). An example of relationship extraction using this large language model is shown in Table 5. The extracted relationships, PMID, and sentences were then stored in a MySQL database.

We constructed a potato ontology by combining entity dictionaries and relational data. It is widely acknowledged that the co-occurrence of biological entities in a sentence indicates a strong correlation. LLM can accurately extract triplet information for highly correlated entities. However, their performance is poorer for entities with weaker correlations, resulting in incomplete relationship extraction. We developed a potato ontology platform to correct erroneous results, remove irrelevant data, and manually add relationships for those unrecognized. The potato ontology curation platform is illustrated in Fig. 2. This platform resulted in a more refined potato knowledge base, which includes 5,701 relational tuples. The species, fungi, bacteria, viruses, and pests were also expanded in the potato knowledge base using taxonomy dictionaries (https://www.ncbi.nlm.nih.gov/taxonomy), resulting in the addition of 5,003 new relationships and final total of 10,704 relationship data points. The relationship data between entities is shown in Table 6.

Knowledge storage and visualization platform

Neo4j is an open-source graph database that stores structured knowledge in the form of graphs rather than traditional tables and supports semantic queries (Holzschuher & Peinl, 2013). We chose Neo4j as the graph database for PotatoG-DKB. The Potato Gene-Disease Knowledge Portal (PotatoG-DKP) was developed using Django as the front-end interface, a high-level Python web framework, with Neo4j as the graph database and a Neo4j + MySQL combination as the back-end implementation. The PotatoG-DKP supports visual query, as shown in Fig. 3. Users can input an entity of interest to query related entities and relationships. A recursive method is used to support querying extended nodes and queried data can be exported in CSV (Comma-Separated Values) and JSON (JavaScript Object Notation) formats. If users do not input any query information, the portal can display the entirety of PotatoG-DKG. To ensure data security, a full data download option is not provided, but researchers can contact us for more information if they need access to the complete data.

Named entity recognition and relationship extraction

PubTator was used for named entity recognition and the Spacy tool was used for entity identification. Using the taxonomy dictionary to expand entities, we obtained a total of 5,206 entity items, including 300 genes, 205 diseases, 300 fungi, 300 bacteria, 109 pests, 300 viruses, 300 species, 277 chemicals, and 73 proteins as well as other entities.

The Spark Large Language Model was used for relationship extraction, supplemented by manual curation. This process yielded a total of 10,704 relationship tuples. Figure 4 depicts the heatmap of typical relationships such as “gene-disease” and “fungus-disease.” Each row and column respectively represent a category of interest (e.g., genes, diseases, species), and each cell in the matrix represents the strength or frequency of the relationship between the two categories. The darker the color, the stronger the correlation between the two entities.

Potato gene-disease knowledge base

Through the potato ontology data curation platform, we obtained potato ontology data on gene-disease relationships in potato as well as other relationships between biological entities, which, together, constitute the PotatoG-DKB, as shown in Fig. 5. This knowledge base includes entities, entity_categories, entity_id, relationships, PMIDs, and sentence information describing relevant relationships. Users can browse, search, and download relevant data. Additionally, users can retrieve information based on entity items or entity categories according to their needs. The PotatoG-DKB also provides an information tracing function, allowing users to link to the original PubMed literature through PMIDs or query detailed information about entities in NCBI using entity IDs. More detailed information can be accessed and downloaded from https://www.potatogd.com.cn.

Potato gene-disease knowledge graph

The PotatoG-DKG covers 5,206 nodes and 9,443 edges. To make scientific literature directly and readily accessible to biologists and potato breeders, we developed a visualization platform, depicted in Fig. 3. The search function of the Neo4j graph database directly returns relevant relationships. For example, when searching for “A”, it may return “A->B”. In order to expand the retrieved relationships, a recursive search algorithm was adopted with a recursion depth limit of five to prevent excessive search depth and subsequent inefficiencies in querying. Figure 6 demonstrates the search results returned when using the keyword “soft rot” to search for relevant knowledge. A data download function is also provided, allowing users to export the data in both CSV and JSON formats.

Potato gene-disease knowledge portal

Relevant researchers and breeders can access and query potato gene-disease ontology data through the PotatoG-DKP. As shown in Fig. 7, the “about” section of the portal website introduces the main process of potato gene-disease knowledge base construction. The “disease and pest” section displays common potato diseases caused by fungi, bacteria, viruses, and pests (Fig. 7A). The “knowledge graph” module enables visual querying of relevant information in the ontology database. To facilitate user operations, PotatoG-DKP has a fuzzy search function, and users can also download the entity-relationship information of interest (Fig. 7B). In the “ontology” section, PotatoG-DKP provides potato gene-disease ontology data, where users can also query or download relevant data (Fig. 7C). The “text annotation” module offers text annotation, allowing users to search and read abstracts of relevant literature (Fig. 7D). If they wish to delve deeper into the specific content, they can directly access the original PubMed article through the PMID. “Links” to related databases are also provided, and users can access or download the tools used in this article through the “tools” section.