Physicians nowadays have to consider a diverse range of data sources when treating a patient. Semantic clinical data warehouses allow to easily add new data and to pro-actively help the physician making sense of the data. In this work-in-progress paper we investigate an approach of using Linked Data as the access mechanism and a graph database for storage and query processing. We describe lessons learned from a case study of discovering similar genes where we use an existing similarity metric to derive new information, the Gene Ontology as a data source, and SAP HANA as an e cient graph database.
In this work-in-progress paper we argue that overcoming such challenges is possible using Linked Data, graph databases, and semantic algorithms (Section 2): we describe a use case for discovering similar genes (Section 3) and derive lessons learned (Section 4). We mention related work (Section 5) and conclude (Section 6). 2
See Figure 1 for the architecture. Information in the semantic clinical data warehouse is presented to the user by a visualisation and analysis tool. To store, query, and visualise arbitrary information we use a graph database and the following intuitive data model (property graph): Relevant objects such as patients, interactions, and observations are represented as vertices in the graph. Such objects have properties with values of primitive datatypes such as String and Integer, e.g., the surname of a patient. Objects are related to each other via edges in the graph, e.g., a patient is diagnosed with a disease. Such relationships also can have properties, e.g., provenance information about the algorithm or human expert that has generated the relationship. The integrator and reasoner component 1) translates an RDF graph to a property graph, 2) derives implicit information useful for data integration and decision support of users, and 3) imports the graph to the graph database. The RDF graph is crawled based on the Linked Data principles.
HTTP Get User HTTP Result
Visualisation & Analysis Tool
Graph Query
Results
Graph Database Results
Semantic Clinical Data Warehouse Graph Query Bulk Load IRneteags&roanteorr RDF
Crawler
Data Sources RDF Linked (Open) Data RDF LD-Wrapper RDF LD-Wrapper
CSV XML
This architecture has the following advantages: Already, there are large
amounts of life science data { directly or using LD wrappers { published using
such widely-adopted access mechanisms and standard vocabularies [
In this section, we apply our approach to a use case for discovering similar genes
from a plant [
Relevant data sources for our prototype { HANA Linked Data AnnSim (HLA) { are descriptions of genes4, gene annotations from experts5, and the Gene Ontology (GO) with a concept hierarchy6.
Using OpenRe ne with RDF extension, we translate the former two sources to RDF and reuse links from the GO RDF representation. Crawling such data results in one RDF graph with genes, concepts, and annotations between genes and concepts.
HLA uses as a graph database HANA Graph, an extension to the HANA
in-memory database for storing and querying of property graphs [
An importer program then maps the crawled RDF graph to a property graph and bulk loads the property graph to HANA Graph. Intuitively, the importer generates for every triple two vertices for the subject and object (if not existing), and an edge for the predicate. HANA Graph then contains genes (e.g., AT5G23810) and concepts (e.g., Amino Acid Transport) as vertices, and relationships between genes and concepts as edges. For instance, there are annotation relationships between genes and concepts as well as is-a relationships between concepts. Vertices and edges can have properties, e.g., a concept has a textual description. The graph is then extended with edges between genes describing their similarity, and edges between concepts describing their distance in the is-a concept hierarchy.
Such information we compute based on an existing algorithm, AnnSim [
GOSLIM.txt 6 http://purl.obolibrary.org/obo/go.owl
As also visible in a screencast on our paper website7, the user of HLA gets an overview of genes and their similarities to other genes; can zoom into single genes to see textual descriptions of concepts; can visit concepts along the concept hierarchy (Figure 2). Also, the user can ask for a graph view showing similarities between genes based on distances between concepts (Figure 3). For the visualisation, we used a visualisation engine called Symbiosis that can be con gured with a JSON-based template language to visualise a graph. Symbiosis uses HANA Graph and GEM for querying the data.
Correct Computation of Similarities. There is a mean squared error between the results of HLA and AnnSim of 0.09. This di erence we expect is due to newer, possibly more elaborate versions of annotations and GO (version 1.2) used by HLA. We compared the results of both approaches with the gold standard, a similarity metric based on the DNA sequence of genes (SeqSim). 7 http://people.aifb.kit.edu/bka/hla/ The mean squared error between HLA and SeqSim (0.19) is lower than between AnnSim 1.0 and SeqSim (0.36), indicating that AnnSim similarities improve with newer data sources; yet, further experiments are needed to con rm this claim.
Scalable Computation of Similarities. Table 2 gives an overview of the time for the di erent steps in the execution. Loading of data is estimated with a connection of 6.7 Mbps download speed. Although HLA takes considerably longer than AnnSim 1.0, we argue that HLA's bottlenecks can be resolved and that HLA is more promising for larger datasets.
AnnSim 1.0 uses a proprietary graph data format with reduced information that is probably fast to generate (Prepare), download, and load. HLA uses a more verbose but also more expressive graph model (RDF) and has to generate (Prepare), download, transform to property graph (Map) and load 15 times more vertices, 22 times more edges and comprehensive properties such as textual descriptions. Loading graph data to HANA Graph showed fast and the preprocessing steps we believe can be optimised by parallelisation.
AnnSim 1.0 uses program logic in C/C++ over arrays to compute the 400 similarities and displays the results to users. HLA loads the relevant data to a graph database and uses program logic in Java to issue database queries to e ciently compute the similarities and to write back the results to the data warehouse. The query language GEM was useful and intuitive for graph-traversal queries. For instance, the following GEM read query is issued using a special-type function WIPE() to the SQL interface of HANA, recursively visiting one or more edges of type rdfs:subClassOf, and returns a vertex table with all ancestors of a GO concept: RESULT uri:myResult FROM f GO:0005634 g-[@core:type = 'rdfs:subClassOf']->(1,*);.
The program logic in HLA spent more than 90% of the time to compute a speci c part of AnnSim, a min-weight perfect matching (Blossom IV). We believe we can optimise the Blossom IV execution, e.g., by running part of it directly in HANA Graph via built-in and user-de ned functions. Writing back of the results to the data warehouse took a lot of time since done using single write queries instead of a bulk load. In this case, since read queries to HANA Graph showed fast, HLA should also scale with larger datasets, in contrast to AnnSim 1.0 that does not outsource bulk loading, reading, and writing to an external database. Computing additional information can be done o ine. Interactive visualisation over HANA Graph were possible using the Symbiosis engine.
Flexible Computation and Visualisation of Similarities. Whereas AnnSim 1.0 was implemented speci cally for the problem of e ciently computing similarities of objects described in a proprietary format, HLA uses Linked Data as a uni ed data model and standard access mechanism.
New data sources can be added to HLA by providing more links to crawleable
Linked Data. We believe that e orts such as by Bio2RDF [
Other objects such as patients can be compared in HLA; AnnSim only requires objects to be annotated with concepts and concepts to be described in an is-a hierarchy. Algorithms that use other relationships and derive other information can be added to HLA. The Symbiosis engine showed that { given su cient understanding of the domain experts' problem { it is easily possible (5{10h of manual work) to provide exible visualisations over a graph-based data model.
According to Haussler et al. [
In this work-in-progress paper, in a small case study of discovering similar genes we illustrated the potential of modular access mechanisms with Linked Data, queries over a schema- exible graph database, and semantic algorithms to derive new information. Continuously adding new data sources and data items, new algorithms, and new visualisations leave exciting future work.