Varcode, Isovar, and Vaxrank¶
The libraries answer different questions about the same variant. Use Varcode for predicted consequences, Isovar for RNA reconstruction and evidence, and Vaxrank for evaluating protein/peptide candidates.
| Library | Responsibility |
|---|---|
| Varcode | Generate structural/transcript hypotheses and predict their coding consequences. |
| Isovar | Reconstruct RNA-supported sequences, compare them with those hypotheses, and preserve unresolved alternatives. |
| Vaxrank | Evaluate the resulting protein/peptide candidates, retaining their evidence. |
This is the shared responsibility split, not a claim that every path is already connected. The implementation limits below matter when building a pipeline.
Varcode's part¶
Varcode models possible effects on transcripts and proteins. Its inputs can include DNA variants, germline/phase context, and externally reconstructed RNA structures. It should not independently collect BAM reads or decide which RNA assembly the patient expresses.
Use the ordinary effects() interface; no special annotator selection is needed.
Keep candidate sets when a variant has several plausible consequences. Candidate
order and effect severity are not RNA support or calibrated probabilities.
Phase evidence follows the same split. Isovar's IsovarReadPhasing reports cis
or trans only from RNA fragments that cover both variants, and Varcode treats
every other pair as unknown; see RNA phase from Isovar.
Isovar can return a structure that changes the original hypothesis, including a previously unmodeled junction. Varcode's job is to represent that structure and predict its coding consequences, not force it into the first reference isoform. RNA imports describe the existing entry points.
What must survive the handoff¶
- Nucleotide sequence and structure, coordinates, reference/annotation identity, and compatible transcript IDs—not only a protein string.
- Partial versus complete sequence, reading-frame evidence, and unresolved regions.
- Sample/library and read/fragment provenance, support, conflicts, and assumptions.
- Separate alternatives when evidence cannot distinguish them. Grouping identical proteins must preserve the contributing structures and their evidence.
RNA support for a local junction does not establish a full-length transcript or prove translation. Missing coverage is not evidence against a hypothesis. A predicted protein change and evidence that the RNA exists are separate facts.
Available today and remaining work¶
- Varcode: structural partner candidates and supplied RNA imports exist. Splice/phase/SV combinations are not exhaustively composed (#423). Change flags for imported partial proteins stay unresolved unless mapped, in-frame observed codons differ from the reference; an unobserved suffix is not a demonstrated protein deletion.
- Isovar: consult its RNA reconstruction and reconciliation guide for the inputs and evidence supported by each reconstruction path.
- Vaxrank: consult its candidate evaluation guide for supported imports, selection policies, and DNA fallback behavior.
Those downstream workflows evolve independently of Varcode. Use their own documentation for current integration coverage rather than treating Varcode's ability to represent a hypothesis as proof of an end-to-end pipeline.
A DNA-only fallback remains a prediction without RNA confirmation. It must not silently discard alternatives or relabel missing RNA as evidence of absence.
Compatible installations¶
The installed releases' dependency metadata defines the compatible package
ranges. Install or upgrade the related packages together, then run
python -m pip check in that environment. Package metadata checks dependency
constraints; tests of the workflows you use check integration behavior.
For source checkouts, consult the current Isovar requirements and Vaxrank requirements. The Varcode changelog records version-specific fixes. Historical minimum-version notes are not a compatibility guarantee for every later release.