CoreKB is a web-based platform enabling users to search for reliable scientific facts concerning gene expression-cancer associations over a medical Knowledge Base (KB). CoreKB provides a streamlined interface for searching either using natural language queries or by exploiting structured facets providing autocomplete facilities. It is designed to simplify information access and search of scientific facts targeting healthcare stakeholders (i.e., clinicians, physicians, and researchers). CoreKB aims at presenting the user a comprehensive overview of the scientific evidence supporting a medical fact, fully connected with ontology-based entities and well-defined literature resources. In addition, CoreKB provides the user a quantitative comparison of the possible gene-cancer associations related to a specific fact, thus enabling users to assess the degree of agreement among the evidence support.
The World Health Organization has identified cancer prevention as a critical public health
challenge of the 21st century, with an estimated 32% increase in cancer cases by 20401. To
address this challenge, cancer research has increasingly relied on microarray and next-generation
sequencing technologies, generating vast amounts of experimental data on gene
expressioncancer interactions [
In this context, eficient and reliable methods to store and organize knowledge from various
sources have become increasingly important, and KBs have emerged as essential resources for
cancer researchers [
Moreover, the massive amount of information contained in a KB can pose hindrances in
locating and extracting specific data of interest. This is compounded by the technical nature
of the queries needed to retrieve the desired information, resulting in an additional layer of
complexity, especially for non-technologically savvy users [
Cancer research also relies on specialized KBs adopting domain-specific jargon, ontologies, and terminologies resources with a large presence of complex terms, acronyms, and morphosyntactic variants, thus making query formulation and efective search even more challenging [ 13].
As a consequence, researchers may encounter dificulties when attempting to retrieve
information of interest from KBs, thus limiting efective data discovery, re-use, and new insights.
Therefore, there is an urgent need for intuitive and user-friendly search applications [
To address these issues, we propose CoreKB, an easy-to-use web-based search platform that builds upon one of the largest KBs containing fine-grained facts about gene expressioncancer associations [14]. The KB comprises over 230,000 reliable and unreliable facts extracted semi-automatically from the scientific literature by mining PubMed.
CoreKB is publicly available at http://gda.dei.unipd.it along with a demonstration video (http://gda.dei.unipd.it/static/videos/demo.mp4) presenting its features. The platform provides a streamlined interface for searching and exploring knowledge about gene expression-cancer associations. CoreKB allows users to query the system using either natural language queries or facets, enabling search for both entities and facts easily. In this regard, CoreKB provides several features (e.g., infometrics and entity cards) to support specialized users, such as medical researchers and clinicians, who are interested in searching for knowledge about gene expressioncancer associations quickly and without requiring the exact terminology or entity identifiers.
As a distinctive trait, CoreKB embraces a fact-oriented approach, which aims at providing users with a comprehensive overview of the key information concerning each fact. The overview provided consists of aggregated data including (i) the gene class distribution among the sentences extracted from the scientific literature; (ii) the number of supporting and conflicting sentences (iii) and the number of supporting publications per year, enabling users to assess the consensus supporting a fact. In summary, CoreKB ofers a powerful platform for comprehensive knowledge discovery in precision medicine.
Despite the importance of providing a fact-level overview for a scientific claim – i.e., showing aggregated information describing a scientific fact overall, by combining data coming from several scientific publications and resources – most of the state-of-the-art methods and tools only focuses on providing point-wise sentence-level information – thus not presenting the big-picture understanding. CoreKB aims to fill this gap by providing a comprehensive, literaturesupported overview for each scientific fact. In this way, CoreKB difers from the approaches described below.
DEXTER [15], a text mining method for extracting associations from scientific literature, ofers a basic search endpoint that requires users to search for gene-cancer pairs and does not support natural language queries. OncoMX [16] provides a unified and user-friendly interface for various datasets, but its literature mining capabilities are limited. OncoSearch [17] enables users to search for sentences that mention gene expression changes in cancer and ofers advanced faceted search features. However, it lacks natural language search and does not provide entity cards or infometrics to support evidence. DisGeNET [18] has a well-designed faceted navigation system, but it only returns gene-disease associations at the sentence level, making it dificult to establish reliable associations for gene-disease pairs at the fact level. Finally, BioKB [19] provides access to the semantic content of biomedical literature but does not support natural language search and focuses on exploring and visualizing the structure of the underlying KB.
To construct the KB that serves as the foundation for CoreKB, we employed a system that collects text from scientific literature and processes it to obtain sentences. These sentences then undergo Named Entity Recognition and Disambiguation (NERD) annotation, which identifies gene-cancer pairs. The NERD annotations are then subjected to bootstrapping and deployment processes. During bootstrapping, fine-grained relationships between entities are manually annotated and used to (i) train Relation Extraction (RE) methods and (ii) populate the KB. In the deployment process, the trained RE methods are used to automatically extract facts from sentences and populate the KB.
Each fact in the KB is associated with a probability distribution that reflects the likelihood of a specific gene-cancer association. These probabilities are used to perform reliability tests, which determine whether facts are reliable or unreliable based on the quantity and level of consensus of collected evidence. By adopting a probabilistic approach, the system can capture the inherent uncertainty in scientific discourse and help users understand the strength of the evidence supporting a particular gene-cancer association. The KB contains 23,879 genes and
Researchers & Clinicians
CORE System NERD + RE
Reliability Testing + KB population
Index
Request 11,530 cancer diseases, encompassing a total of 230,000 fine-grained facts. These facts are supported by 1,037,845 sentences taken from 251,038 research articles [14].
CoreKB consists of several components including (i) a React-based web interface for the frontend; (ii) a Django-based business logic for the back-end providing support for REST APIs and services; (iii) a PostgreSQL database aligned with an instance of the Virtuoso RDF triple store for storing the KB data; (iv) a Redis server broker for eficient in-memory data store and access and (v) a Python-based search component that performs NERD on entity mentions in user queries and then performs a structured search in the database to retrieve the matching facts and the related information. To this end, the search component relies on a Redis in-memory dictionary of entities returning for each entity name and synonyms the corresponding unique identifier.
When a user-provided query is entered in the system, the search component assigns a score to each entity that is maximum in case of an exact match or proportional to the number of matching terms in case of a partial match. To avoid favoring long-named entities at the disadvantage of those with short names, the score is discounted proportionally to the entity name’s length. If a single entity is recognized, all the related facts are ordered by scientific evidence support. Similarly, if multiple entities are identified, the facts regarding the most matching gene-cancer pairs are promoted in the ranking of results. It’s worth noting that CoreKB is not limited to human gene-cancer pairs, instead, it contains facts regarding any living organism. Nevertheless, human-specific genes are ranked on top.
The CoreKB platform enables users to search for factual information related to gene-cancer associations supported by scientific literature, as well as entities such as genes and cancer diseases. Users can search using free-text or structured search facilities with autocomplete features. The interface allows users to switch between the two search modes via the settings menu and provides clickable sample queries to demonstrate the system’s capabilities. The search results are presented in cards that display gene class, cancer label, statistics on supporting and conflicting sentences, key examples of supporting and conflicting sentences, and bibliometrics on the number of supporting publications per year. The reliability of the fact is indicated by a circle that is colored diferently according to the informativeness and reliability of the fact i.e., green for reliable facts, gray for facts missing proper support, and red for unreliable facts. On the right side of each fact card, specific information about the gene and cancer involved is presented, including links to the corresponding entries in the National Center for Biotechnology Information (NCBI)2 and Linked Life Data3 platforms. Users can access additional quantitative and aggregated information on dedicated landing pages, including related facts and supporting and conflicting sentences. The landing page for a particular gene displays all the available information about the gene, including its symbol, full name, type, synonyms, designations, summary, and gene class distribution for diferent cancer diseases. The sentences involving the gene are presented in a table that users can filter and sort arbitrarily. Action buttons enable users to expand or collapse each card and choose the columns to be shown in the sentence table.
Figure 3 shows the landing page for a given entity, that is, in this case the human gene ERBB2. The interface comprises two major cards (A, B). The first card (A) displays comprehensive information about the gene, including its symbol (1), full name (3), type (4), synonyms (5), designations (6), last modified date (7), summary (8), and the gene class distribution – i.e., the role of the gene in a specific gene-cancer association which can assume one of alternatives 2https://www.ncbi.nlm.nih.gov/gene/ 3http://linkedlifedata.com/
B oncogene, biomarker, tumor suppressor gene – for several cancer diseases presented both as a textual statement (9) and depicted also in a horizontal bar chart (10). It is worth noting that all the numbers reported in the textual statement are clickable and allows the users to visualize on a new tab the specific gene-cancer associations of interest (e.g., for which cancer diseases the ERBB2 gene is indicated to be an oncogene). To save space on the interface, long textual information such as the gene summary is truncated by default, but it can be expanded or collapsed by clicking on the dedicated button after the ellipsis.
The second card (B) provides a list of sentences related to the gene. The sentences are presented in a dynamic table that allows users to filter and sort them according to their preferences. For instance, in Figure 3, only the sentences where ERBB2 is reported as a biomarker are displayed because of the user-provided value biomarker for the Gene class column (13). In case of long sentences, users can either (i) resize columns; (ii) going over with the mouse on a sentence to view the full sentence text in a small tool-tip or (iii) click on the sentence to visualize it in a separate pop-up. Moreover, there are two buttons reported on the top-right corner of each card so that users can expand/collapse each card alternatively to fit the full-screen size (2, 12) and choose the columns to be shown in the sentence table (11).
Most of the information reported in the sentence table of Figure 3.B are clickable, including (i) the sentence; (ii) the scientific fact; (iii) the gene; (iv) the disease; (v) the PubMed link to the publication from which the sentence has been extracted and (vi) the publication journal. When the user clicks on a sentence, a pop-up appears to show all the relevant information concerning the sentence such as the sentence identifier, its textual content, the gene, the cancer disease, and their relationship. Figure 4 shows the informative pop-up for the sentence’s fact “ERBB2 is a biomarker for Stomach Carcinoma”, which is reported on top of the window card. Then the pop-up reports other information concerning the sentence, the gene, the cancer disease, and their association in separated collapsible menus. Each sentence and fact in CoreKB is univocally identified with a permanent URL enabling point-wise access to the information specific for an individual resource. Indeed, when the user clicks on the sentence identifier in Figure 4.A the corresponding landing page for the sentence is shown, as depicted in Figure 4.B. Note that in the landing page, users can access provenance information about the sentence, that is, the PubMed URL to the original publication and a link to the corresponding journal entry in the SCImago Portal4. Similarly, when users click on a fact the corresponding landing page for the scientific claim is shown providing a holistic perspective enabling to assess overall the literature consensus among the supporting/conflicting evidences.
CoreKB is a web-based search platform designed to provide users with reliable scientific facts on gene expression-cancer associations. The platform ofers natural language query support as well as structured facet search features integrated with autocomplete facilities. As a distinguishing feature, CoreKB focuses on presenting fact-oriented information rather than adopting a sentencelevel approach, where only single, independent results are presented. To this end, CoreKB combines multiple aggregated information from several literature resources either supporting or conflicting with the scientific claim of interest. This approach ofers a comprehensive overview of the scientific evidence associated with a given fact. Moreover, CoreKB provides users with a quantitative comparison of the possible gene-cancer associations, making it easier to determine if there is a consensus on a specific gene’s role. Although CoreKB focus is on gene expression-cancer associations, its model can be adapted and reused to accommodate other types of relationships with the appropriate modifications.
The platform aims to support clinicians and researchers by providing fast search, access, and consultation of reliable scientific findings along with validating literature evidence. Hence, simplifying knowledge discovery and promoting a serendipity-oriented perspective. As future work, we plan to improve CoreKB according to the requisites and feedback of clinicians, which we plan to collect by conducting a user study.
The work was supported by the ExaMode project as part of the EU H2020 program under Grant Agreement no. 825292. [13] M. Agosti, S. Marchesin, G. Silvello, Learning Unsupervised Knowledge-Enhanced Representations to Reduce the Semantic Gap in Information Retrieval, ACM Trans. Inf. Syst. 38 (2020) 38:1–38:48. URL: https://doi.org/10.1145/3417996. doi:10.1145/3417996. [14] S. Marchesin, L. Menotti, G. Silvello, O. Alonso, CORE: Gene Expression-Cancer Knowledge Base, 2023. URL: https://doi.org/10.5281/zenodo.7577127. doi:10.5281/zenodo. 7577127. [15] S. Gupta, H. Dingerdissen, K. E. Ross, Y. Hu, C. H. Wu, R. Mazumder, K. Vijay-Shanker, DEXTER: disease-expression relation extraction from text, Database J. Biol. Databases Curation 2018 (2018) bay045. [16] H. M. Dingerdissen, F. Bastian, K. Vijay-Shanker, M. Robinson-Rechavi, A. Bell, N. Gogate, S. Gupta, E. Holmes, R. Kahsay, J. Keeney, H. Kincaid, C. H. King, D. Liu, D. J. Crichton, R. Mazumder, OncoMX: A Knowledgebase for Exploring Cancer Biomarkers in the Context of Related Cancer and Healthy Data, JCO Clin. Cancer Inform. (2020) 210–220. [17] H. J. Lee, T. C. Dang, H. Lee, J. C. Park, OncoSearch: cancer gene search engine with literature evidence, Nucleic Acids Res. 42 (2014) 416–421. [18] J. P. González, J. M. Ramírez-Anguita, J. Saüch-Pitarch, F. Ronzano, E. Centeno, F. Sanz, L. I.
Furlong, The DisGeNET knowledge platform for disease genomics: 2019 update, Nucleic Acids Res. 48 (2020) D845–D855. [19] M. Biryukov, V. Grouès, V. P. Satagopam, BioKB - Text mining and semantic technologies for the biomedical content discovery, in: Proc. of the 10th International Conference on Semantic Web Applications and Tools for Health Care and Life Sciences (SWAT4LS 2017), Rome, Italy, December 4-7, 2017, volume 2042 of CEUR Workshop Proceedings, CEUR-WS.org, 2017. URL: http://ceur-ws.org/Vol-2042/paper5.pdf.