Proteins are of modular design, with structural units called domains [1]. Domains often act as the operational units responsible for many aspects of protein function, and some of them are linked to phenotypic traits and diseased states. Despite their importance in biology, domains are less studied than proteins/genes in terms of ontology annotation; something much-needed and only recently addressed by the dcGO database [2]. This database provides a systematic annotation of domains using a panel of ontologies; an ontology such as Gene Ontology (GO) [3] is controlled vocabularies but organised in a hierarchy to categorise a particular sphere of knowledge. The dcGO algorithm was initially published as an improvement to the SUPERFAMILY database [4]. The quality and utility of this resource were evaluated in the Critical Assessment of Function Annotation (CAFA) competition [5], [6]. The webserver provides several mining facilities, however, web-based facilities are limited in analytical flexibility and scalability; there is a need to have a standalone tool overcoming these limitations. Currently, there are no bioinformatics tools that are specifically designed for analysing ontologies and annotations at the domain level. Most, if not all, open-source tools (such as ‘topGO’ [7], ‘GOSemSim’ [8] and ‘ontologizer’ [9]) are gene-centric and only deal with a very limited number of ontologies, usually GO. To the best of my knowledge, these tools do not provide the support for customised analysis according to users' own ontologies and annotations. To meet these needs, I have developed ‘dcGOR’, a flexible R package that provides a basic infrastructure suitable for representing ontologies and annotations. More importantly, it supports various analytical utilities tailored to this important resource. As demonstrated below, dcGOR is capable of in-depth analyses of input domains; structural bioinformatics/genomics community is increasingly confronted with this type of analysis. With this package, users are expected to understand their domains of interest: not just in the relevance to functions, phenotypes and diseases, but also at a network level. With this package, users are also able to perform customised analysis using their own ontologies and annotations.

The dcGOR package is designed in a general way that allows for representing and analysing three bits of information: domains, ontologies and annotations. For it to be applicable in domain-centric annotations, the backend is its built-in data that is pre-compiled from the latest version of the dcGO database [2]. There are a dozen or so ontologies, such as GO, Disease Ontology (DO) [10] and Human Phenotype (HP) [11]. They are all used to annotate both SCOP domain superfamilies and families [12]. Also supported are GO annotations for domains taken from Pfam [13] and InterPro [14], and for non-coding RNAs from Rfam [15]. Table 1 lists ontologies and annotations supported in the package.

The dcGOR is exclusively implemented on the R software environment. Three S4 classes are defined: ‘InfoDataFrame’ for domains, ‘Onto’ for ontologies and ‘Anno’ for annotations. The class ‘InfoDataFrame’ is used to store domain information. Since an ontology is organised as a directed acyclic graph (DAG; a directed graph without cycles), the class ‘Onto’ represents the ontology as a directed graph in which both adjacency matrix and node/term information are defined. For annotations, the class ‘Anno’ is defined to accommodate a sparse annotation matrix and additional metadata on domains and terms. All these classes have their class-specific S4 methods. This design of data representations greatly simplifies domain ontology analyses. Table 1 outlines supported analyses: domain-based enrichment analysis, semantic similarity between pairs of annotated domains, and significance analysis for estimating a contact network.

The function dcEnrichment conducts enrichment analysis based on the hypergeometric/binomial distribution or Fisher's exact test [16]. It tests the statistical significance of the observed number of domains overlapped between an input group of domains and domains annotated by an ontology term. By default, all annotatable domains are used as the test background, but the user can specify this background. Taking as inputs a group of domains, dcEnrichment reports ontology terms that are enriched in this input domain group. To account for the ontology DAG, it also implements several algorithms that were originally applied to GO [7], [9]. The basic idea is to estimate the significance of a term after adjusting (e.g. removing) those annotations that its children terms also have. Enrichment outputs are stored as an object of S4 class ‘Eoutput’, on which methods are defined for easy view and save. Directly operating on this object, the function visEnrichment visualises the top significant terms in the context of the ontology DAG to aid intuitive interpretation.

Semantic similarity is a type of comparison to assess the degree of relatedness between two entities (here domains) in meaning of their annotations [17]. Semantic similarity between domains is calculated based on their annotation by ontology terms. To do so, information content (IC) of a term is defined as the negative 10-based log-transformed frequency of domains annotated to that term. This definition considers the actual usage of a term (the frequency of annotated domains it has) to measure how specific and informative the term is. The function dcDAGdomainSim first calculates semantic similarity between terms, which is then used to derive similarity between domains. All popular IC-based semantic similarity measures [8], [17] are supported. From pairwise term similarity, dcDAGdomainSim has several methods to calculate similarity between pairs of domains, including 3 best-matching (BM) based methods: average, maximum, and complete. For a term in either domain, all these BM-based methods first calculate maximum similarity to any terms in the other domain. For more detail, the reader is referred to this review [17]. The resulting domain (semantic similarity) network is stored as an object of S4 class ‘Dnetwork’, a weighted undirected graph in which domains are nodes and their semantic similarity scores as the edge weights. Notably, the higher the semantic similarity score is, the more similar the domain pair is (the edge weight). There is no hard threshold for the semantic similarity scores, but it is advisable to focus on the edges with highest weights (e.g. the top 50% of all edges).

Given a domain network (e.g. the one resulting from dcDAGdomainSim), the function dcRWRpipeline performs random walk with restart (RWR) for estimating contact strength and significance between two input groups of domains (as seeds). It is based on the earlier work [18], but has been generalised to allow for weighting domain seeds, and done so in a single step. RWR-based contact outputs are stored as an object of S4 class ‘Coutput’, including a contact (statistical significance) network that is also a weighted undirected graph (an object of S4 class ‘Cnetwork’).

In addition to the analyses above, dcGOR also has several auxiliary functions for data load, annotation propagation, graph class conversion, and fast computation. The function dcRDataLoader is the hub for loading all kinds of package built-in data; this simplifies data use and also makes room for the future data expansion. The function dcDAGannotate is supposed to propagate annotations. According to the true-path rule, a domain annotated to a term is also annotated by all its ancestor terms (propagated to the root). This ensures that only the valid part of the ontology (in terms of domain annotations) is used properly. The function dcConverter is able to convert an object between newly defined graph classes and the one used in packages ‘igraph’ [19] and ‘dnet’ [20]. This conversion enables network visualisation. Visualisation for pairwise semantic similarity matrix is done by package ‘supraHex’ [21]. To relieve computational burden, dcGOR utilises vectorised and parallelised operations. This high-performance parallel computing is realised via executing loops in parallel, aided by two packages ‘doMC’ and ‘foreach’.

The most common use case is to analyse a list of protein domains of interest. As a proof of principle, I use two interesting lists of domains (one from SCOP, the other from Pfam) to demonstrate the functionalities supported in the dcGOR package, particularly enrichment analysis and network analysis. Also, I show how users can benefit from this package to analyse their own domains, ontologies and annotations. All these examples are reproducible following step-by-step demos on the package website, from which results can also be found.

As open-source software, the dcGOR package is freely available under the GPL-2 license (see Software S1). For ease of installation (R package dependencies), it is distributed as part of CRAN, http://cran.r-project.org/package=dcGOR. For ease of version control, it is also distributed at GitHub, https://github.com/hfang-bristol/dcGOR. The details on documentations and demos can be found at http://supfam.org/dcGOR. As missed in most R packages, online documentations and demos are user-friendly; users can see both illustrated codes and executed outputs. This will dramatically reduce the learning curve and promote the wide adoption as users can exactly reproduce what they see.

The dcGOR is a general open-source tool for ontology and annotation analysis, providing a relatively complete framework. As demonstrated, it is able to analyse three most popular domain types (SCOP, Pfam and InterPro) and Rfam RNA families as well, and to support customised analysis. For example, users can analyse domains with different definitions, such as the partner members of the InterPro consortium [14]. The package is designed to be generic to all ontologies, not merely GO (as is the case with most existing tools) but also organism-specific ontologies. The future intention is to include those in the Open Biomedical Ontologies consortium [25]. Here I only describe a handful of analyses that are routinely required for ontology analysis, but the package is scalable for further development. Other than the data expansion aforementioned, future developments will focus on developing utilities for genome function and phenotype prediction. As the standard has been set in dcGOR, it should be much easier for ontology users/developers to extend this software to meet their needs. Also, there is no reason not to apply the similar design principles for ontology analysis at the gene level.