VRS contributors confronted numerous trade-offs in developing this specification. As these trade-offs may not be apparent to outside readers, this section highlights the most significant ones and the rationale for our design decisions, including:
Variation Rather than Variant¶
The abstract Variation class is intentionally not labeled “Variant”, despite this being the primary term used in other molecular variation exchange formats (e.g. Variant Call Format, HGVS Sequence Variant Nomenclature). This is because the term “Variant” as used in the Genetics community is intended to describe discrete changes in nucleotide / amino acid sequence. “Variation”, in contrast, captures other classes of molecular variation, including epigenetic alteration and transcript abundance. Capturing these other classes of variation is a future goal of VRS, as there are many annotations that will require these variation classes as the subject.
Allele Rather than Variant¶
The most primitive sequence assertion in VRS is the Allele entity. Colloquially, the words “allele” and “variant” have similar meanings and they are often used interchangeably. However, the VR contributors believe that it is essential to distinguish the state of the sequence from the change between states of a sequence. It is imperative that precise terms are used when modelling data. Therefore, within VRS, Allele refers to a state and “variant” refers to the change from one Allele to another.
The word “variant”, which implies change, makes it awkward to refer to the (unchanged) reference allele. Some systems will use an HGVS-like syntax (e.g., NC_000019.10:g.44906586G>G or NC_000019.10:g.44906586=) when referring to an unchanged residue. In some cases, such “variants” are even associated with allele frequencies. Similarly, a predicted consequence is better associated with an allele than with a variant.
Implementations should normalize¶
VRS STRONGLY RECOMMENDS that Alleles be normalized when generating computed identifiers. The rationale for recommending, rather than requiring, normalization is grounded in dual views of Allele objects with distinct interpretations:
- Allele as minimal representation of a change in sequence. In this view, normalization is a process that makes the representation minimal and unambiguous.
- Allele as an assertion of state. In this view, it is reasonable to want to assert state that may include (or be composed entirely of) reference bases, for which the normalization process would alter the intent.
Although this rationale applies only to Alleles, it may have have parallels with other VRS types. In addition, it is desirable for all VRS types to be treated similarly.
Furthermore, if normalization were required in order to generate Computed Identifiers, but did not apply to certain instances of VRS Variation, implementations would likely require secondary identifier mechanisms, which would undermine the intent of a global computed identifier.
The primary downside of not requiring normalization is that Variation objects might be written in non-canonical forms, thereby creating unintended degeneracy.
Therefore, normalization of all VRS Variation classes is optional in order to support the view of Allele as an assertion of state on a sequence.
Alleles are Fully Justified¶
In order to standardize the representation of sequence variation, Alleles SHOULD be fully justified from the description of the NCBI Variant Overprecision Correction Algorithm (VOCA). Furthermore, normalization rules are identical for all sequence types (DNA, RNA, and protein).
The choice of algorithm was relatively straightforward: VOCA is published, easily understood, easily implemented, and covers a wide range of cases.
The choice to fully justify is a departure from other common variation formats. The HGVS nomenclature recommendations, originally published in 1998, require that alleles be right normalized (3’ rule) on all sequence types. The Variant Call Format (VCF), released as a PDF specification in 2009, made the conflicting choice to write variants left (5’) normalized and anchored to the previous nucleotide.
Fully-justified alleles represent an alternate approach. A fully-justified representation does not make an arbitrary choice of where a variant truly occurs in a low-complexity region, but rather describes the final and unambiguous state of the resultant sequence.
Sequence ranges use an interbase coordinate system. Interbase coordinate conventions are used in this terminology because they provide conceptual consistency that is not possible with residue-based systems.
The choice of what to count–residues versus inter-residue positions–has significant semantic implications for coordinates. Because interbase coordinates and the corresponding 0-based residue-counted coordinates are numerically identical in some circumstances, uninitiated readers often conflate the choice of numerical base with the choice of residue or inter-residue counting. Whereas the choice of numerical base is inconsequential, the semantic advantages of interbase are significant.
When humans refer to a range of residues within a sequence, the most common convention is to use an interval of ordinal residue positions in the sequence. While natural for humans, this convention has several shortcomings when dealing with sequence variation.
For example, interval coordinates are interpreted as exclusive coordinates for insertions, but as inclusive coordinates for substitutions and deletions; in effect, the interpretation of coordinates depends on the variant type, which is an unfortunate coupling of distinct concepts.
The VRS collaborators investigated numerous options for modelling data, generating code, and writing the wire protocol. Required and desired selection criteria included:
- language-neutral – or at least C/C++, java, python
- high-quality tooling/libraries
- high-quality code generation
- documentation generation
- supported constructs and data types
- lists, maps, and maps of lists/maps
- nested objects
- protocol versioning (but not necessarily automatic adaptation)
Initial versions of the VRS logical model were implemented in UML, protobuf, and swagger/OpenAPI, and JSON Schema. We have implemented our schema in JSON Schema. Nonetheless, it is anticipated that some adopters of the VRS logical model may implement the specification in other protocols.
There are many packages and proposals that aspire to a canonical form for json in many languages. Despite this, there are no ratified or de facto winners. Many packages have similar names, which makes it difficult to discern whether they are related or not (often not). Although some packages look like good single-language candidates, none are ready for multi-language use. Many seem abandoned. The need for a canonical json form is evident, and there was at least one proposal for an ECMA standard.
Not using External Chromosome Declarations¶
In ChromosomeLocation, the tuple <species,chromosome name> refers an archetypal chromosome for the species. WikiData and MeSH provide such definitions (e.g., Human Chr 1 at WikiData and MeSH) and were considered, and rejected, for use in VRS. Both ontologies were anticipated to increase complexity that was not justified by the benefit to VRS. In addition, data in WikiData are crowd-sourced and therefore potentially unstable, and the species coverage in MeSH was insufficient for anticipated VRS uses.