Positional MEDLINE (PosMed) is a web application that quickly prioritises biomedical entities such as genes and diseases based on statistical signi cance of associations between these and a user-speci ed keyword by employing our original search engine named General and Rapid Association Study Engine (GRASE). GRASE search is modelled as an extension of SPARQL search with statistical analysis, which enables searching over semantic data including not only linked datasets but also signi cant extracted semantic links over multiple biomedical documents. PosMed was originally implemented for in silico positional cloning studies by prioritizing genes. Further applications include bioresource search with associated genetic functions or ontologies, and functional interpretation of gene variants found from exome sequencing of personal genomes. PosMed is available at http://database.riken.jp/PosMed/.
In the life sciences eld, a Semantic Web approach that employs machinereadable linked data prepared from conventional various omics datasets has been studied to understand biomedical phenomena. However, because the task of generating semantic links for our biomedical knowledge is too expensive, and such knowledge is described by a vast amount of human-readable biomedical literature, this semantic technology is still not widely adopted by biologists.
For practical use of published biomedical data on the Semantic Web, especially use of data di cult to utilise due to lack of semantic links, it is bene cial to reinforce acquisition of such data by supplying a hybrid methodology combining not only inferences over that knowledge described as linked data but also knowledge supported by statistical signi cance over a vast number of multiple raw documents.
Our implementation of this methodology is the search engine named GRASE
[
GRASE search is modelled as an extension of SPARQL search with statistical analysis, which enables searching over semantic data including not only datasets in Resource Description Framework (RDF) but also signi cant extracted semantic links over multiple biomedical documents including MEDLINE abstracts. Direct search (keyword ! entity) The GRASE search engine quickly prioritises biomedical entities such as genes, diseases, drugs and mouse strains based on statistical signi cance of associations with a user-speci ed keyword. More cona b cretely, for each entity GRASE generates a 2 2 contingency table c d consisting of the number of documents (a) where both the the keyword and the entity appear, (b) where the keyword appears but the entity does not appear, (c) where the keyword does not appear and the entity does appear, and (d) where neither the keyword nor the entity appear, then applies the Fisher's exact test to the contingency table to compute a P-value for the signi cance of the test. Inference search (entity ! entity) GRASE further infers other entities from the result of direct search by applying semantic links described by RDF triples and statistically extracted co-citation relationships over two entities e1 and e2 appearing in a common document by applying Fisher's exact test against 2 a b 2 contingency table c d consisting of the number of documents (a) where both e1 and e2 appear, (b) where e1 appears but e2 does not appear, (c) where e1 does not appear but e2 does appear, and (d) where neither e1 nor e2 appear. Therefore, an entity can be searched via a search path keyword ! entity 1 ! entity 2 and its signi cance computed as 1 (1 pd)(1 pi), where pd and pi are P-values of direct search and inference search respectively. 3
Since 2005, we have been extending datasets in PosMed to make it possible to follow an ever-changing trend of biomedical applications. Datasets currently introduced in PosMed are shown at http://database.riken.jp/PosMed/.
A typical application of PosMed is searching genes with user-speci ed
keywords and chromosomal intervals suggested by linkage analysis. So that inference
searches can be performed such as mouse gene{drug inference and mouse gene{
human gene inference, currently PosMed supports the following datasets:
{ up to 352,000 entities including not only genes in mouse, human, rat,
Arabidopsis and rice, but also drugs, metabolites, diseases and mouse strains
associated with document sets including up to 9,870,000 documents from
MEDLINE abstracts, OMIM, gene annotation, molecular interaction, Open
Biomedical Ontologies (OBO) [
In order to realise quick response, the datasets listed above are distributed over 11 computers and these work in parallel.
PosMed was used to prioritise genes in the RIKEN large-scale mouse ENU
mutagenesis project and contributed to successful identi cation of 65 responsible
genes [
One conventional problem for a mouse bioresource database is that knockout
strains are not used when the targeted gene has an unknown function and no
observed phenotype. We introduced 19,885 mouse strains registered in the
International Mouse Strain Resource (IMSR) [
PosMed successfully connects these functionally unknown genes to known genes using molecular interactions, pathway information and co-citations and as a result enables suggestion of unobserved phenotypic bioresources. PosMed not only allows users to retrieve mouse bioresources directly with the user's keywords described in bioresource annotations, but also inferentially through corresponding documents for mouse and human genes, diseases, drugs, ontologies, pathways, metabolites, molecular interactions and MEDLINE abstracts.
As an extension to other species, we newly introduced 7,207 Arabidopsis bioresources and 823 Arabidopsis phenotype observations extracted by human literature curation, so that PosMed inferentially discovers Arabidopsis bioresources as well through correspoonding documents for genes, phenotypes, ontologies, co-expressions, molecular interactions and MEDLINE abstracts. 3.3
PosMed can also be applied to functional interpretation of genetic variants detected by exome sequencing studies using a next generation sequencer. Since exome sequencing studies usually nd several hundreds or thousands of genetic variants by comparing samples and controls, prioritisation of the candidate genes using PosMed is an e ective method for further functional analysis.
Users can upload a tab-separated values le with gene IDs and their descriptions. PosMed prioritises genes listed within les by statistical relevance between the user's keywords and each gene, and displays ranked genes together with user uploaded descriptions and associated documents. 4
We proposed a Semantic Web data search methodology and tool that extends
conventional graph search like SPARQL with statistical text mining over a vast
number of documents. Not only discovering discoveries documents related to
biomedical entities when given a query as does the service GoPubMed, our
PosMed also supports biomedical entity prioritization. Among several software
tools available to prioritise positional candidate genes, PosMed was evaluated as
sepecially highly e ective in comparison with two other similar tools
GeneSniffer and SUSPECTS [