This paper presents Graph-KD, a tool to navigate through large relational knowledge sources. Graph-KD provides methods to understand relationships between concepts using open discovery, closed discovery and knowledge inference. The purpose of the tool is the support of biomedical knowledge discovery and exploration. It is primarily intended to be used by medical researchers and presents a use case involving millions of relations from UMLS. Graph-KD is able to process even large graphs efficiently and can be accessed via a webinterface (http://biomedical.dfki.de/graph-kd).
Relational knowledge bases and ontologies are rich sources of concepts and the relationships between them which often consist of large amounts of information. These resources generally include information about directly related concepts but the information about those related indirectly can also be extremely valuable. Exploring this information can provide further insights and can help to discover new knowledge.
Various tools exist to explore knowledge graphs such as UMLS. However,
existing tools either have a different focus or/and cover only parts of the functionalities of
Graph-KD. For instance, UMLS:Similarity [
Graph-KD provides functionalities to explore a large knowledge graph using open and
closed discovery as well as knowledge inference. Open and closed discovery base on the
idea of literature-based discovery (LBD) which was introduced by Swanson [
One of the major problem of LBD is the enormous number of connections which effectively rules out checking all possibilities. Therefore, it is very important to concentrate on the most relevant connections during the early stages of research. Graph-KD supports this process by providing useful information in an easily navigable format. The following functionalities are provided: Closed Discovery applies k-shortest path to find relevant connections between two concepts. Passing two concepts of interest to Graph-KB, the tool visualizes the connections between them. In case various relation paths exist, all shortest paths will be displayed. An example of the shortest paths between the class of pharmacologic substances (nsaids) and a particular disease (kidney disease) is presented in Figure 1.
Open Discovery explores concepts and relations around a target concept. An example is provided in Figure 2. The graph shows the target concept antipyretics in combination with the target relation may-be-treated-by. The open discovery searches for the relation of interest around the target concept in closest distance. Since antipyretics is not linked to any may-be-treated-by relation in our example, the tool shows nodes in close distance which are connected via this relation.
Knowledge Inference is a technique which takes existing facts into account and tries to make assumptions about unknown information. As knowledge graphs tend to be incomplete this can be a useful feature to support the open and closed discovery process. Graph-KD integrates a rule-based inference at this point.
The backend of Graph-KD is written in python and has access to Neo4j, an open source NoSQL scalable graph database management system, which stores data in form of nodes and their corresponding typed edges. Using Neo4j’s built-in functionality the open and closed discovery can be executed with a good performance.
Graph-KD: Exploring Relational Information for Knowledge Discovery For the demo use case, Graph-KD explores information from UMLS, a large biomedical knowledge base containing millions of medical terms and relations between them. UMLS defines medical concepts including their synonyms, and unifies them to a concept unique identifier (CUI). All those concepts are linked to at least one semantic type, such as Body Part, Finding or Clinical Drug. Moreover, UMLS defines relations between concepts which include for instance isa, may-treat or contraindication-of.
For our demo UMLS 2017AB was preprocessed and diminished. This included the
removal of relations containing concepts related with itself, very general relations,
inverted relations and concepts with less meaningful semantic types. Namely, semantic
types of the semantic groups GEOG, OBJC, OCCU and ORGA, according to
Bodenreider and McGray [
The rule inference relies on transitivity rules, between hyponyms in combination with other relations (e.g. if A and B are related and B is a child of C, then we find a transitive relation between A and C).
Replication of Existing Discoveries In order to show the benefit of our tool, we explore
long range dependencies by replicating existing discoveries as presented in Preiss et
al. [
Runtime Table 2 shows the runtime for open and closed discovery. For both scenarios 200,000 randomly generated requests were sent via REST to the backend. The table shows that for approximately 8% (15,499) of all randomly selected CUI pairs a shortest path can be found (max distance 4). For those connections the average (mean) runtime is 0.15 seconds. The maximum response time is 175.02 seconds for closed and 0.51 for open discovery respectively. In addition to that, 75% of the requests are processed within less than a tenth of a second.
Roller et al.
Discovery D # closed open Raynaud disease – Fish oil 3 127 discovery discovery Somatomedin C – Arginine 3 27 connections 15,499 6,793 Migraine disorders – Magnesium 4 471 mean 0.153920 0.006258 Magnesium deficiency – Neurologic disease 3 108 min 0.001387 0.001212 Alzheimer’s disease – Indomethacin 3 105 75% 0.060400 0.001599 Alzheimer’s disease – Estrogen 3 100 max 175.018601 0.512222 Schizophrenia – Calcium-I. Phospholipase A2 4 22 Table 2: REST runtime test for 200k Table 1: Exploring existing discoveries using Graph-KD random requests 4
In this work we presented Graph-KD, a tool to explore large knowledge graphs. GraphKD provides various functionalities for knowledge discovery and includes knowledge inference methods to gain further into the data. As our example in Table 1 showed, Graph-KD can be easily applied to support literature-based discovery. Moreover, other clinical use cases are possible in which physicians explore information in the knowledge graph in order to detect potential new links between medical concepts.
This project was funded by the European Union’s Horizon 2020 research and innovation program under grant agreement No 780495 (BigMedilytics) and by the German Federal Ministry of Economics and Energy through the project MACSS (01MD16011F).