This paper presents an association discovery framework for proteins based on semantic annotations from biomedical literatures. An automatic ontology-based annotation method is used to create a semantic protein annotation knowledge base. A semantic reasoning service enables realisation reasoning on original annotations to infer more accurate associations. A case study on protein-disease association discovery on a real-world colorectal cancer dataset is presented.
To bridge the gap between the biomedical science and bioinformatics, many
biomedical ontologies have been created in the past few years. Ontology-based
semantic annotation for biomedical entities are of interest to both biomedical
researchers and general public. Meanwhile, the biomedical domain has a large
and fast-growing amount of literature resources, among which MedLine1 is the
primary publication repository for biomedical research. Ontology-based
biomedical text annotation has shown promising progress and several tools have been
successfully developed and evaluated in biomedical text mining problems[
This paper proposes an integrated high performance framework that leveraging protein annotations and semantic reasoning to an informative proteinbiomedical concepts association Knowledge Base(KB). Starting from a list of proteins, the system automatically retrieves a pool of MedLine citations and annotates the proteins using pre-de ned biomedical ontologies. A realisation reasoning service is applied to infer more accurate protein association information. In our preliminary study, the focus is on the discovery of potential proteindisease associations. A case study on discovering protein-disease associations for a real-world colorectal cancer tissue protein dataset is presented.
We propose a Semantic PRotein Annotation method based on MedLine (SPRAM) which produces semantically inferred protein annotations based on biomedical literatures and ontologies (Fig.1). SPRAM starts with a list of proteins of in
In biomedical ontology annotations, very often an instance is annotated with
multiple classes with subclass relationships in the ontology. To the best of our
knowledge, existing biomedical annotation tools with semantic reasoning
functionalities only do semantic expanding[
We developed a specialised realisation reasoning service for dynamically generated annotations. Di erent to the traditional Description Logic most speci c concept reasoning, our algorithm works on a dynamic set of annotations on the y instead of assertions in a static KB. Only the most speci c annotations will be stored in the KB. The algorithm takes a set of semantic protein annotations, , and an ontology, O, that is based on. A most speci c class set, 0 , is initialised to be an empty set, ;. For each class t 2 for each protein, nd all subclasses of t in O, i.e., fCig where Ci v t 2 O. Class t is added to 0 if fCig \ = ;, i.e., t is the most speci c annotation in given ontology O. The algorithm outputs the most speci c class set 0 which will be inserted into the nascent KB.
Fig.2 shows an example of the e ect of applying realisation reasoning on the disease annotations for a protein with a UniProt ID \O43175". Class \disease", \cancer" and \carcinoma" in the original annotation set are all realised to \adenocarinoma" because the last concept is subsumed by those three concepts and it is also in the original annotation set. This is important because it more accurately represents the biomedical categorical information and reduces complexity.
Jankova et al. found 45 up-regulated proteins in colorectal cancer tissues by using
experimental protein iTRAQ analysis[
To help biologists achieve these goals, we take these 45 proteins and use our SPRAM work ow to assist discovering the potential diseases associated with these proteins. The result was: 1080 MedLine citations, 354 diseases associations based on HDO, that was reduced to 241 unique associations after realisation reasoning. SPRAM returns a set of protein-disease association to the biologists. That includes also the source reference titles and URLs. The biologist can then use this result to help validate these associations easily by tracing back to the references.
Fig.3 shows the changes of the distribution of the diseases associated with these 45 proteins before and after realisation reasoning. The distribution after realisation reasoning represents more accurate and sensible information. For example, the top distributed concept, disease, was removed and the next, cancer, was greatly reduced, thereby producing a clearer set of associations. (a) DOA distribution before reasoning (b) DOA distribution after reasoning
To nd proteins reported as colorectal cancer related, the biologist issues a query using the concept, \colorectal cancer". The semantic reasoning service rewrites this query into a union of this concept and all of its subclasses. The result shows that 6 proteins (CEA, NNE, HSP 84, NPM, 3-PGDH and UEV-1) reported in the literature as being related to colorectal cancer. 5
This paper proposes an automatic protein-oriented association discovery framework based on semantic annotations from literature. A semantic reasoning service provides realisation reasoning. We demonstrate the usage of our system on protein-disease association discovery using a real-world colorectal cancer protein dataset. In upcoming work, focus will be given to a ranking model of protein associations and customisable selection of protein-PMID mappings.