The Gene Ontology (GO) project started to provide semantic standards for the annotation of molecular attributes of genes and gene products [1]. The Gene Ontology is a controlled vocabulary for describing genes and gene products in terms of their associated biological processes, cellular components and molecular functions. The structural foundation of GO is formally a Directed Acyclic Graph (DAG) wherein the terms are equivalent to the nodes and the relationships to the edges of the graph [2].

GO has grown enormously. The number of organism groups participating in the GO Consortium has grown every quarter year from the initial three to roughly two dozen [3]. A lot of biological databases use GO to annotate the molecular attributes of genes and gene products [4, 5]. GO-based analysis of microarray and mass spectrometry data have been successfully realized [3]. Recently, new generation of tools based-on GO have been developed, aiming to enhance biological knowledge such as protein structure classifying [6], gene-phenotype association predicting [7] and gene network building [8]. More details are available at GO website (http://www.geneontology.org/GO.tools.shtml). Unified Medical Language System (UMLS) metathesaurus has been integrated with GO to expand UMLS into the biological domain [9].

In spite of the undoubted importance of GO, several drawbacks associated with GO and GO-based annotations have been introduced. Masseroli correctly pointed out the structural and semantic problems of GO such as metonymy, species-specific terms and multiple paths [10]. Dolan et al. evaluated the reliability of GO-based annotations [11]. Poor inter-annotator reliability of GO-based annotations for human-mouse orthologous gene pairs was reported between two gene-annotation groups, MGI and GOA. Park et al. identified syntactic errors caused by the two GO-update operations, ‘new obsoletions’ and ‘new term merges’, used in the course of GO version change [12]. They introduced GOChase to detect and correct the syntactic errors and error propagations in GO-based annotations (http://www.snubi.org/software/GOChase/).

In the present study, we further identified semantic error types in GO-based annotations; redundant, biological-domain-inconsistent and taxonomy inconsistent annotation.

The first type is “redundant annotation.” When a gene is annotated to a GO term, for instance, according to the current GO annotation paradigm, it is considered to be implicitly annotated to all parents of the term. Assigning both parent and child terms to the same gene is regarded as “redundant annotation.” In some cases, if parent and child term was annotated in specific gene product using different evidence code, these annotations hard to say completely redundant. For example, an experiment may provide enough evidence to annotate to a parent, but not to any specific child, whereas a more specific annotation may be predicted by sequence comparison or other computation. In such cases both annotation would be retained, the parent because of its experiment support and the child for specificity. So we analyze the redundant annotation to distinguish the evidence code used in parent and child term.

The second type is “biological domain-inconsistent annotation.” A GO term should avoid using species-specific definitions and rather include any term that can be applied to more than one taxonomy classes of organisms (The Gene Ontology Consortium, 2000). Some GO terms have species-specific characteristics such as nucleus (GO:0005634), specific for eukaryotes and unidirectional conjugation (GO:0009291), specific for prokaryotic specific terms. As GO-based annotation expands to various species, however, species-specific terms become increasingly problematic. For example, a gene product having UNIPROT ID O24899 from Helicobacter pylori, a kind of bacteria, is wrongly annotated to nucleus, a eukaryote-only GO term.

The third type is “taxonomy inconsistent annotation”. Recently, the GO Consortium provided terms with taxonomy restrictions, containing species-specific terms with the NCBI taxonomy group for which they are or are not appropriate (http://www.geneontology.org/GO.sensu.shtml). Forty four taxonomic groups used taxonomy restricted terms in the January 2010 GO version. Taxonomy inconsistent annotation occurs when a taxonomy restricted term is annotated to a gene that does not belong to the corresponding taxonomy group. GO consortium checks the inconsistent annotation using taxonomy restricted terms and provide reports of inconsistent annotation. But many annotations have been produced without consideration of taxonomy restricted terms. For example, we found that a eukaryote restricted GO term, Golgi apparatus (GO:0005794), was (wrongly) annotated to 27 gene products of Escherichia coli, a kind of bacteria.

In the present study, we analyzed the distributions of the semantic inconsistencies in GO-based annotations using 27 major biological databases. To understand the factors influencing such inconsistent annotations, we perform correlation analysis between the inconsistent annotations and the possible attributes for the inconsistent annotations including the usage of evidence codes (http://www.geneontology.org/GO.evidence. shtml), the number of gene products, the number of species and the number of GO terms. We developed a set of web-based utilities, GOChase-II, to correct the semantic inconsistencies in addition to the previous corrections for the syntactic errors by GOChase-I [12].

To analyze the distributions of the semantic inconsistencies in GO-based annotations we calculated the distribution of redundant annotations in the 27 biological databases (Table 1). All databases have redundant annotation. The fraction of redundant annotations in databases is distributed from 0.9% to 91% for gene products (31% in average), from 2% to 26% for GO terms (13% in average), and from 0.4% to 38% for GO annotations (12% in average). UniProtKB/Swiss-Prot shows the highest redundancy for gene product (91%) and GO annotation (38%). The database showing the highest redundancy in GO terms is Ensembl (24%). GeneDB_Pfalciparum shows the lowest numbers among the databases; 0.9% for gene products, 2.5% for GO terms and 0.4% for GO annotations. In all databases, the fractions of redundant annotation based on the same evidence code are larger than different evidence code.

We identified and corrected three types of semantic inconsistencies in GO-based annotations for gene products from 27 major biological databases. GO becomes a widely accepted ontology in biomedical field. The under-managed errors and inconsistencies may reflect its short history, its ever growing complexity, and the vast amount of the biological domain knowledge. Recently GO Consortium starts working on refining GO contents and structure [17]. The present study demonstrates that the GO community may be empowered by bioinformatics tools ensuring error-proof mechanisms concerning the GO hierarchical relationships, species-specific definitions and GO term usage guidelines.

To sum up our result in this research, there is no database free from the semantic inconsistent annotation. Among the three types of semantic inconsistent annotation, redundant annotation is most common error. About 12% of the whole annotations are redundant. Only a few biological-domain inconsistent annotations are found in the 18 biological databases because of the small number of ‘eukaryote-only’ (410) and ‘prokaryote-only’ (71) GO term.

The high correlation between IEA and inconsistent annotations (Table 3) suggests that IEA has lower reliability than others. Electronically generated associations without human judgment are labelled as IEA. GO Consortium proposes a hierarchy of reliability among evidence codes (http://www.geneontology.org/GO.evidence.shtml). In general, TAS and IDA show higher reliability. TAS and IDA have low correlation with all three types of inconsistent annotation. And most of evidence codes, which curated by human, have low correlation with all types of inconsistent annotation. This result implies that the hierarchy of reliability among evidence codes are preserved in inaccurate annotation.

The numbers of gene products and species of a database show high correlations with all types of inconsistent annotations except taxonomy-inconsistent annotation. It suggests that the complexity of database maintenance may affect the occurrence of inconsistent annotations. Therefore, it is more strongly required for such databases to implement a sound mechanism such as GOChase-II in order to avoid semantic inconsistencies caused by multiple user-groups.