The ever-increasing number of publicly available biomedical articles calls for automatic information extraction from digitized publications. We have implemented a pipeline which, by exploiting text mining and semantic technologies, helps researchers easily access semantic content of thousands of abstracts and full text articles from PubMed and Elsevier. The text mining component analyzes the articles content and extracts relations between a wide variety of concepts, extending the scope from proteins, chemicals and pathologies to biological processes and molecular functions. Moreover, the relations are extracted along with the context which speci es localization of the detected events, preconditions, temporal and logic order, mutual dependency and/or exclusion. Extracted knowledge is stored in a knowledge base publicly available for both, human and machine access, via web interface and SPARQL endpoint. To address the data accessibility, reusability and interoperability, all the extracted relations are standardized using unique resource identi ers (URIs) and a custom ontology based on Genia ontology.
Information extraction from biomedical literature is becoming a common
practice due to the huge amounts of available textual data, and technological
maturity which allows to gain insight into scienti c content. Text analysis evolved
from spotting relevant concepts in the text [
Knowledge discovery systems and platforms vary in scope. Many of them
are focused on speci c sub-domains. For example, EVEX [
The paper is organized as follows. Section 2 gives an overview of the BioKB platform. In Section 3 we describe the text mining component. Section 4 focuses on semantic technologies employed by BioKB. Description of the web interface follows in Section 5. Section 6 o ers a discussion, while conclusions are presented in Section 7. 2
1 Not to be confused with two other independent systems: http://www.cs. cmu.edu/~biokb/ and http://www.bioinf.mvm.ed.ac.uk/twiki/bin/view/TWiki/ BioKbPlugin.
Systems architecture is illustrated in Figure 1. Publications retrieved from
PubMed and PubMed Central are indexed by a Solr instance; each publication
is processed by the text mining component; results are converted to RDF
(NQuads) and stored in a triple store. To allow both human and machine access to
the knowledge base, SPARQL endpoint provides machine readable access while
a web application allows users to browse the content of the knowledge base. The
web application is developed in Python 3 using the Flask framework and the
SPARQLWrapper library to query the triple store. We use the vis.js [
The main steps executed by the text mining component of BioKB are:
i) named entity recognition; ii) syntactic parsing; iii) semantic interpretation
(see Figure 2). They are brie y described in the following subsections.
During the Named Entity Recognition (NER) stage, biomedical concepts are
identi ed in the text. Our choice of a NER engine was driven by two major
requirements: a) capability to identify multiple concept types (bio-entities) to
avoid using and synchronizing multiple NER tools within one pipeline; b) ability
of the engine to map entity name to its unique identi er in a dedicated database.
The latter is known as normalization process and is indispensable in order to
ensure database and semantic graph coherence. One of the systems which meets
our criteria is Re ect [
Availability of the trigger dictionary is another prerequisite for semantic
analysis. Triggers are words or expressions used to describe a biological process. For
example, mediates, increased, expression, binding and coactivating are examples
of triggers in the phrase in Figure 2. Our trigger dictionary is derived from Genia
annotated corpus [
With the entities and triggers in place, we can proceed toward syntactic
analysis. In order to maximize the probability of identi cation of
\subject-predicateobject" triples (e.g., \RFLAT-1 activates RATES "), only the sentences with
at least two entities and one trigger are processed. For syntactic analysis we use
Stanford parser [
In order to ensure transfer between syntax to semantics, we opt for the rule-based approach. It consists of assigning semantic roles to entities which are syntactic arguments of a trigger. As a result, relations are typed (mostly, the type is inherited from the type of their trigger) and, whenever applicable, directed. For example, direction of a regulatory event is from semantic subject (cause) toward semantic object (theme). On the contrary, relations of type binding and correlation are naturally not directed. We collect syntactic arguments of the triggers via the depth- rst search (DFS) of the sentence graph. The rules are applied on the ensemble of trigger and its dependencies. For example, syntactic subject of mediates, PGC-1, is the semantic subject of the regulatory event whose predicate is mediates. Sometimes nodes are merged in favor of a more straightforward semantic interpretation. Thus, increased and expression are jointly interpreted as Positive regulation, loosing their individual correspondence to Positive regulation and Gene expression relation types.
Biomedical processes are subject to rich variety of conditions under which they could take place. We attempt to account for these by processing information conveyed by certain lexical and syntactic elements. For example, the main event in Figure 2, PCG-1 mediates positive regulation of GLUT4, is communicated along with the description of its mechanism introduced by the adverbial clause headed by trigger verb binding. By taking this bit of information into account we can logically order the events described in the sentence: (1) PCG-1 binds and coactivates MEF2C; (2) GLUT4's expression is increased (Step III of the Figure 2). 4
The choice of using semantic web technologies for this project was dictated by two main reasons. First, using an ontology to represent the hierarchy of relationships o ers di erent level of query granularity. For instance, one can ask if two entities are connected by a property regulates and be able to retrieve also results for the property increases because the two properties are linked by a sub-property relation. Additionally, the ontology and thus the hierarchy of properties can be updated without having to re-process the publications. Besides this reasoning capability, using semantic web technologies o ers full machine readable access to the complete knowledge base. Not only can the knowledge base then be used by third parties directly but it becomes possible to combine BioKB data with external sources using federated queries. 4.1
We created a simple ontology (Figures 3, 4), to represent the hierarchy of classes and properties that are used to categorize entities and relations identi ed by the text-mining component. This ontology is heavily inspired by the GENIA ontology. Our decision to allow inferences on sub-relationships resulted in the need to create a custom ontology. Indeed, in the GENIA ontology, relationships are represented by classes rather than properties. In the proposed ontology, a relationship between two bio-medical entities can be directly translated to a single triple, s p o where s and o are the entities and p is a sub-property of biokb:bioRelation, the top level property in the BioKB model. We then use the named graphs feature of Virtuoso to add metadata about this relationship. This includes information such as creation date, provenance and con dence score. 6 In the current deployment of the platform, a single instance of the open source edition of Virtuoso 7 hosts the knowledge base and provides the SPARQL endpoint. The server hosting the Virtuoso instance has the following characteristics: 128GB Ram, 8 cores, Hard Drive 500GB 10000 RPM. At the date of this publication, the size of the database is 22GB for 215 million triples. On top of the content generated by the text-mining module, the di erent ontologies mentioned in Section 4.1 have also been loaded into the triple store. The actual number of triples constituting the BioKB speci c content is about 156 million triples. Those triples are the result of the processing of more than 800 000 publications. About 10 million events were extracted from approximately 6.5 million sentences. 5
Besides the SPARQL endpoint, we created a web interface to access the BioKB content. This web application is publicly and freely available at https://biokb. lcsb.uni.lu. The home page displays a unique search eld providing autocomplete functionality for all supported bio-medicial entities. Once the user clicks on an entity, the entity page will be displayed. This page shows a textual description of the entity, the list of most common co-occurrences for this entity (as a tag cloud) and two tables with the list of relationships involving this entity as extracted by the text-mining module. Those incoming and outgoing relationships are also represented visually as an interactive graph (Figure 5). On this graph, the central node is the entity corresponding to the current page and all other nodes and edges represent the most common relationships involving this entity. For each edge, on mouse over, the label and the number of occurrences of this relationship are displayed. Each node is clickable and leads to the corresponding entity page. Each edge is also clickable and results in the display of the relationship details page (Figure 6). This page displays the list of publications where this relationship was found and the speci c sentences. On the entity page, a download button proposes an export of the result of the SPARQL DESCRIBE command in RDF/XML and in CSV.
The primary goal of our information extraction system and knowledge base is to help researchers focusing on various types of biomedical data analysis. We illustrate its functionality with two use cases related to disease network construction and enrichment.
Chronic obstructive pulmonary disease (COPD): the network veri cation challenge. Gathering disease-related factors into a large-scale network became a common practice. Such networks provide a comprehensive model which helps to elucidate mechanisms involved in pathological processes. For this network veri cation challenge, we used our system ability to provide typed, directed (if applicable) relations between various concepts. We have scanned the literature and extracted candidate relations which have been veri ed by a human expert and made part of the collaborative community curated network yielded by the 8
Challenge [
Our system is constructed with the goal of detailed knowledge extraction from textual data, its availability to human and machine. Its strength is the ability to process abstracts as well as full texts; extract semantic relations between various concept types and contextualize them in terms of location, conditions, logic and temporary order. A web interface o ers public and free access to the knowledge base while a SPARQL endpoint o ers a machine readable access.
Some aspects of the system will be further developed and there remains room for change and improvement. First of all, the benchmarking of the system accuracy needs to be performed. From the text mining perspective, it operates on the sentence level which limits its recall. Although extracted knowledge is normalized with respect to concepts and relations, various nomenclatures are used. To increase knowledge interoperability we plan to adopt Uni ed Medical Language System (UMLS) which capitalizes on straightforward communication between various systems processing biomedical and health related data. Currently triple store covers main attributes of the extracted relations, such as subjectpredicate-object while contextual aspects need to be incorporated. Future work will include enriching the scope of entity types, extending the current web application by adding, among other developments, an advanced search feature, a personalized noti cation system, a REST web service and some bibliographic management system to easily cite the publications. BioKB will also have to be continuously extended by processing more publications. 7
In this paper we described an information extraction system along with the storage database and web interface in the eld of biomedicine. The system employs text mining and semantic technologies to help discovery and accessibility of biomedical knowledge. As a proof of concept, we have shown its applicability to disease network construction and enrichment. Along with the strengths, we have pointed out the system's limitations and outlined future work directions. 8
This work was partially conducted in the scope of the eTRIKS project that received funding from the European Union and from the European Federation of Pharmaceutical Industries and Associations as an IMI JU funded project (no. 115446). The Reproducible Research Results (R3) team of the Luxembourg Centre for Systems Biomedicine is acknowledged for support of the project and for promoting reproducible research.