We have exploited semantic annotations of biological entities to develop a novel approach to infer new knowledge. We demonstrate this in four use cases based on the Gene Expression Ontology, an applied ontology that we developed to serve the needs of researchers involved in the analysis of genes and proteins implicated in transcriptional control of pathways/diseases. We have found that semantic annotations associated with biological entities in various commonly used data sources support the identification of related entities, thereby emulating associations that can be inferred from sequence or other structural similarities between these entities. We demonstrate how those semantic annotations can be used to make inferences about the respective biological entities.
The Gene Expression Ontology (GeXO) [
Protein features were extracted from GeXO in the form 'predicate-object.' The features form a matrix with 39680 rows corresponding to proteins and 132360 columns to features. The types of features we used are summarized in Table 1.
This feature matrix was used to compute semantic similarity among all the proteins
in the data set on the basis of the Jaccard index weighted by the information content
as described in [
For classifying false and true positives we used KEGG clusters of orthologous proteins as positive sets. The KEGG cluster annotations were removed from the data set prior to the analysis. The results in Figure 1 demonstrate the very high predictive value of semantic annotations (the results with the full data are given just as a reference). To exclude an impact of sequence information on the analysis, we removed orthology information from the data set, which affected the results only slightly. We concluded that semantic annotations are able to reveal protein similarity even in the absence of sequence information. 3
To identify recurrent patterns in the data, we used the DASS tool [
Figure 2 provides the distribution of closed sets according to the size (number of elements), or frequency (number of occurrence in the data set). The vast majority of closed sets fall within a narrow range of the size and frequency. The sets of low frequency are likely to be highly predictive due to their specificity. To have a more precise view on the predictive value of the set we focused on KEGG clusters, which define functionally distinct protein types, associated with closed sets. Figure 2 gives the distribution of closed sets according to the number of associated clusters. The highest number of sets was found to be associated with a single KEGG cluster, thus confirming the high predictive value of the closed sets. It is worth noting the very high number of closed sets without any associated KEGG cluster. In combination with the results of the ROC analysis this suggests that the closed sets could be used to classify the proteins associated to those closed sets. 4
To demonstrate the high predictive value of the closed sets, we extracted a subset containing 106 transcription factors (TFs) of 40 distinct types known or suspected to be involved in the response to the hormon gastrin.
The 106 TFs were subjected to clustering with a number of approaches on the basis of associated closed sets. The resulting clusters were used as templates for screening the total GeXO data set to identify hypothetical TFs and target genes (TGs). For an initial validation of the identified candidates, we downloaded additional information from UniProtKB (lookup for the term: gastric in http://uniprot.org) and mapped it on the identified candidates. We identified more than 1700 potential candidates including more than 460 genes with transcriptional activity (non-deep analysis, automated screening by extracting information from http://www.uniprot.org/uniprot/ with a customized Python script). Furthermore, we identified 53 known TGs and 28 genes linked to the term gastric, whereas 11 TGs and 7 gastric genes occur in more than two clustering solutions and represent (partially) supported hypotheses. Thus, these results show that we can use the closed sets concept for predicting TGs and regulators involved in response to gastrin. Additionally, we identified more than 400 novel candidates occurring in more than two clustering solutions. Evidently, not all of these candidates are directly involved in this response, but they represent a good basis for further (more detailed) analyses as well as possible wet-lab experiments.