The Sequence Ontology (SO), a member of the Open Biomedical On-tologies (OBOs) library, was developed with the goals of standardizing the vocabulary and semantics of biological-‐sequence annotation with the goal of increased interoperability for software developers and users of genomic sequences. Here we present our recent developmen-tal approaches to address three issues of import for the SO: (1) repre-sentation of molecular sequences versus abstract sequences; (2) inte-gration with the ChEBI ontology, the Protein Ontology, the RNA Ontol-ogy, the Gene Ontology, the Chemical Information Ontology, and the Information Artifact Ontology; and (3) consistent representation of DNA, RNA, and peptide sequences and harmonizing their use toward annotation in sequence databases. We anticipate that these efforts will result in a representation of biological sequences that is more con-sistent not only internally but also with respect to its use in annota-tions in sequence databases. We further envision increased interoper-ability of the SO with other OBOs, which would benefit applications beyond sequence annotation.
The Sequence Ontology (SO), a member of the Open
Biomedical Ontologies (OBO) library
We present our recent developmental efforts seeking to
address three issues of import for the SO: (1) representation
of molecular versus abstract sequences; (2) integration with
other OBOs, particularly the Chemicals of Biological
Interest (ChEBI) ontology
The SO is a resource actively maintained by a small group
of curators responsive to the requirements and input of the
sequence-annotation community; it is currently managed via
an SVN repository, where users can download versioned
releases and revisions (http://www.sequenceontology.org/
resources/index.html). The developers of the SO have begun
efforts to harmonize the SO with other resources
There has been considerable ambiguity with regard to the
ontological nature of biological sequences, including the
categories of sequences represented in the SO. Hoehndorf
et al. have posited the existence of three types of sequence:
(1) abstract sequences are abstract entities that are
“independent of space and time: either [they] … are not located in
space and time, or they are located everywhere and at all
times”; there is, for example, only one instance of the
abstract sequence ACA; (2) syntactic sequences are sequence
representations such as those in biomedical databases and
text representations; and (3) molecular sequences are
physical chains of nucleotides or amino acids
However, there are several issues making this conceptualization as the basis for SO representation problematic. In their framing of SO concepts as generically dependent continuants, SO developers acknowledged a discordance in that sequence attributes, which are explicitly represented in the SO and used in the formal definitions of corresponding sequences, actually apply to the molecular sequences. For example, a wild_type_rescue_gene is a rescue_gene that has_quality wild_type; that is, “wild-type” describes molecular sequences, not the abstract sequences that refer to the molecular sequences. More straightforwardly, biologists fundamentally regard sequences such as genes, exons, mutations, transcripts, and peptides as molecular entities, as evidenced in, e.g., their definitions in biology textbooks, and this conceptualization is reflected in the natural-language definitions for most of the SO classes in their current official state.
We argue that since the molecular sequences are the more fundamental concepts (indeed, the generically dependent sequences depend upon them for their existence), they should be explicitly represented. That being said, there are at least a small number of SO classes whose conceptualizations as molecular sequences do not seem sensible. For example, match is defined as a “region of sequence, aligned to another sequence with some statistical significance, using an algorithm such as BLAST or SIM4”. (Annotations using this concept are typically used to provide supporting evidence to computational gene models.) For a given match, there may be a molecular sequence that directly corresponds to it (though there may not be, e.g., if gaps are permitted in this conceptualization), but since this matching occurs computationally, it seems much more sensible to represent it as a type of abstract sequence. Thus, some abstract sequences will be needed to represent the full set of concepts of the SO. Our solution to this is to represent biological sequences in two parallel ontologies, one containing the large majority of classes that can exist as molecular sequences and the other containing the corresponding abstract sequences for all of these molecular sequences and also the small number of classes that make more sense as abstract sequences.
The former will be an evolution of the Sequence
Ontology:Molecules (SOM) effort, a small ontology representing
molecules of genomic origin
The corresponding abstract sequences of these molecular sequences will remain the province of the SO; therefore, SO concepts will continue to be generically dependent continuants. This has the advantage of minimizing disruption to annotation efforts with the SO, as all current SO terms will continue to exist in the SO (whereas the concepts that are more sensible as abstract sequences will not be correspondingly represented in the MSO). In an effort toward usability, corresponding abstract and molecular sequences will be identically named but use their respective namespaces. As the current SO cross-product definitions will be transferred to their corresponding MSO concepts, SO concepts will instead be formally defined in terms of analogous MSO concepts, as will be shown in the next section. Since SO concepts will be necessarily and sufficiently defined in terms of their corresponding MSO concepts, an OWL reasoner will be able to automatically generate the hierarchy of the former from the latter, so the two parallel sequence hierarchies will not have to both be manually curated. In addition to linking to the MSO, the concepts of the SO will be connected to the CHEMINF ontology, and thus indirectly to the IAO, as described in the next section. 2.2
Many of the OBOs have been impressively developed, but lack of formal linkage among them is a serious issue, and we seek to (directly or indirectly) link both the MSO and SO to other OBOs. As for the former, among the neighboring ontologies with which we envision integration are the ChEBI ontology and the PRO, RNAO, and GO. The first of these is the primary OBO representing molecules, molecular parts, atoms, subatomic particles, and biochemical roles and applications of these entities, and all MSO concepts will be subclasses of the ChEBI class molecular entity. The current official top-level sequence term in the SO is the fuzzily named region, defined as a sequence feature with an extent greater than zero, which will be renamed to the more precise monomeric sequence, i.e., a sequence of biological monomers; this concept will be fundamentally subdivided into monomeric sequence molecule, representing sequences that are whole molecules, and monomeric subsequence, representing proper parts of monomeric sequence molecules. (We are aware that a molecule technically refers to an electrically neutral polyatomic entity and that biological sequences cannot be guaranteed to be electrically neutral (and likely are not); we are referring to a broader sense of molecules that is also reflected in the ChEBI term macromolecule, which has the concept of a molecule incorporated into its name but is also not guaranteed to be electrically neutral and is therefore not subsumed by molecule.) In biological parlance, “sequence” can refer to either a whole sequence or a proper subsequence, and this ambiguity is encapsulated in the top-level class monomeric sequence. (In fact, we can formally define this class as this union of monomeric sequence molecule and monomeric subsequence.) This fundamental subdivision of monomeric sequence allows us to more richly link the SO to ChEBI: Within the latter, sequences such as nucleic acids and peptides are represented as entire molecules but not as proper subsequences, and so this subdivision will enable us to assert the equivalency of specific existing ChEBI macromolecular classes and specific MSO subclasses of monomeric molecule. We can further link the MSO to ChEBI by defining sequence types in terms of their constituent monomers. We have created has_proper_monomeric_part as a subrelation of has_proper_part to use in such definitions, e.g.:
MSO:‘monomeric sequence’ and has_proper_monomeric_part
some CHEBI:‘amino-acid residue’ and has_proper_monomeric_part
only CHEBI:‘amino-acid residue’
The MSO will additionally be able to be linked to other
OBOs representing more specific types of biological
sequences. The PRO, an OBO which focuses on protein
classes and complexes, could link to the MSO by making its
top-level protein a subclass of MSO:peptide
sequence molecule. (As the PRO also represents
protein variants, isoforms, and modified forms, we envision
that the MSO will be further linked to the PRO relying on
our representation of sequence variation; a discussion of this
is beyond the scope of this paper, but we have done
preliminary work in a richer representation of sequence variation in
the SO
The molecular sequences represented in the MSO will also be able to be utilized by the GO: GO classes representing processes operating on sequences, particularly many subsumed by macromolecule metabolic process or regulation of macromolecule metabolic process, will be able to rely on relevant SO classes for their formal definitions. For example, the GO class RNA processing is currently informally defined as “[a]ny process involved in the conversion of one or more primary RNA transcripts into one or more mature RNA molecules”; it could be formally defined using the SO classes primary transcript and mature transcript, e.g.:
GO:‘biological_process’ and part_of
GO:‘biological_process’ and results_in_derivation_from
some MSO:‘primary transcript’ and results_in_derivation_to
some MSO:‘mature transcript’
Here, we have defined an RNA processing as a biological
process that is part of a biological process and that results in
the derivation from at least one primary transcript to at least
one mature transcript. (We have included the parthood
expression to model the involvement mentioned in the
informal definition, and we take advantage of the fact that
something is part of itself for instances of RNA processing that
are not proper parts of instances of composite RNA
processing. Also, here we have created the
occurrent-tocontinuant relations results_in_derivation_from
and results_in_derivation_to as extrapolations of
the continuant-to-continuant relation derives_from,
which is defined in the OBO Relation Ontology
While the MSO will enable integration with ChEBI, PRO, RNAO, and GO, the SO concepts will be made subclasses of the class information about a chemical entity from the CHEMINF ontology, an OBO focusing on the representation of informational chemical entities manipulated in computational algorithms and procedures, as well as the algorithms and procedures themselves; the SO will thus be indirectly connected to the IAO, as this CHEMINF class is itself a subclass of the IAO’s information content entity class, defined as “an entity that is generically dependent on some artifact and stands in relation of aboutness to some entity”. We will use denotes, a subrelation of the IAO’s fundamental is_about relation, to formally define the large majority of the concepts of the SO in terms of those of the MSO, e.g.:
Thus, an abstract transcript sequence is a chemical
information content entity that denotes a molecular transcript
sequence. Hypothetical, improbable, and even impossible
abstract sequences could be created, which may seem
problematic given that we have modeled them as information
content entities, which are defined to be “about” something.
However, we consider this an orthogonal issue not limited
to sequences, and there have been recent efforts to address
this issue
Sequences are annotated overwhelmingly at the genomic level. Perhaps unintuitively, these sequences are annotated at the DNA level even with the many SO concepts at the RNA and peptide levels (e.g., splice site, polypeptide domain), with the implied semantics that the RNAor peptide-level concept holds for the RNA or peptide sequence that corresponds to the DNA sequence denoted by the annotated sequence. These RNA- and peptide-level classes are informally defined as RNA and peptide sequences, respectively, as one would expect, yet they are sometimes subsumed by DNA-level concepts; for example, transcript is a subclass of gene member region. As part of our efforts toward making the SO more consistent in terms of both the ontology itself as well as its use in sequence annotations, we are addressing this conceptual tangle by consistently representing these sequence types and preparing for their use by annotators.
It is clear that the natural-language definitions of these concepts should match their formal structure, and thus, either the RNA-level definition of transcript should change, or it should not be subsumed by a DNA-level concept. We argue that these classes should be defined as they are canonically conceptualized, so transcript should be defined at the RNA level. Its classification should also reflect this, so it should be subsumed by some more generic RNA concept rather than by gene member region. Therefore, we are properly classifying these concepts.
For this classification, we have created a set of sequence classes consistently defined in terms of type of monomer. Currently, monomer type is represented by a set of polymer attributes, and sequences are attributed these qualities; for example, DNA, RNA, and peptidyl are all subclasses of polymer attribute, and, e.g., RNA chromosome is formally defined as a chromosome that has an RNA quality: ‘RNA chromosome’ subclassOf chromosome and has_quality some RNA For each type of monomer, we are creating a primary sequence class (e.g., DNA sequence, RNA sequence, peptide sequence), which was not previously explicitly represented. Rather than relying on qualities for specifying the monomer types of the sequences, we are using ChEBI classes that represent the monomers, as exemplified by the definition of peptide sequence in the previous section. A wealth of monomer types are already represented in ChEBI (including many noncanonical ones), so this strategy obviates the need for us to explicitly represent them. In addition to reducing effort on our end, it abides by the principle of orthogonality among ontologies of the OBO library. Monomeric sequences are thus subdivided along two orthogonal axes, namely, whether they are whole molecules or proper subsequences (as discussed in the previous section), and by monomer type. However, all these direct subclasses will be necessarily and sufficiently defined.
As previously stated, sequences are overwhelmingly annotated at the DNA level, but this includes the use of RNAand peptide-level classes such as splice site and polypeptide domain to mark up DNA sequences. There are several strategies we can take to address this, one of which is to explicitly represent corresponding DNA, RNA, and peptide sequences, link them accordingly, and guide annotators to proper use of these classes. We anticipate that this would be a significant change for annotators, and so as to minimize confusion, we could name these new classes as the sequences on which they are based, appended with “DNA” and “RNA”, as appropriate; for example, for polypeptide domain, we could create polypeptide domain DNA and polypeptide domain RNA, representing DNA and RNA sequences, respectively, corresponding to polypeptide domains.
There are several options as to how to link such concepts. One is to state each association as the product sequence being created from the template sequence, e.g.: ‘polypeptide domain’ subclassOf peptide sequence and created_from_template_sequence
some ‘polypeptide domain RNA’ This states that a polypeptide domain is a subclass of a peptide sequence that is created from an RNA sequence corresponding to a polypeptide domain as a template sequence, which seems odd and circular. The other option is to state each association in the reverse direction, e.g.: ‘polypeptide domain RNA’ subclassOf ‘RNA sequence’ and template_for only ‘polypeptide domain’ This formal definition seems sensible in that it is reflected in the name of the class. A disadvantage is that the template_for restriction is not existential (∃) in that not every RNA sequence corresponding to a polypeptide domain will get translated into a polypeptide domain. Rather, this would have to be made universal (∀), which we believe canonically holds. Relying on this option, we can then link the corresponding DNA and RNA sequences: ‘polypeptide domain DNA’ subclassOf ‘DNA sequence’ and template_for
only ‘polypeptide domain RNA’ Thus, polypeptide domain DNA is the class that would be used to annotate a DNA sequence that currently is annotated with polypeptide domain. However, as creation and use of explicit corresponding sequences would be a significant change to the ontology and to the annotation process, we will seek community input with regard to this.
We have presented and discussed our recent efforts in the continuing development of the SO: (1) representation of molecular versus abstract sequences; (2) integration of the SO with ChEBI, PRO, RNA GO, CHEMINF, and IAO; and (3) consistent representation and use of corresponding DNA, RNA, and peptide sequences. In addition to increasing interoperability of the SO with other OBOs, we anticipate that this work will improve the consistency of the SO both internally and and with respect to external resources; these would strengthen the SO as a tool for reasoning with regard to its use toward its primary use case of sequence annotation as well as other applications. As these discussed changes significantly alter the structure and terminology of the ontology, a measured approach must be taken to allow time to update the existing software and protocols that rely on the SO.
We gratefully acknowledge the support of this work by NIH/NHGRI R01HG004341. We also thank Colin Batchelor for his participation in helpful discussions.