Splice, germline-phase, haplotype and SV uncertainty... | openvax/varcode

Nobody has claimed this yet.

varcode models four kinds of uncertainty: splice outcome (SpliceOutcomeSet), unknown germline phase (PhaseCandidateSet), known-cis haplotypes (HaplotypeEffect) and SV alternatives (StructuralVariantEffect.candidates). Each is built by a separate code path, and when two of them apply to the same variant, one path either ignores the other or silently drops it.

All examples below are on main at 1f65898, CFTR ENST00000003084, Ensembl 81. The germline variant is a het SNV with no genotype, so phase is unknown.

Combination	What comes back	Problem
Splice donor SNV (`7:117531115 G>A`) + germline SNV in the last exon codon	`PhaseCandidateSet` containing two `SpliceOutcomeSet`s, identical in both hypotheses	The germline is never applied. `_classify_against_patient_baseline` can't apply an intronic variant to cDNA, so it falls back to reference-relative annotation (`germline.py:897-906, 919-923`), and the phase axis adds a second level of nesting with no information in it.
DEL across exons 5–6 + germline SNV in window	`PhaseCandidateSet` of `ExonLoss`, `ExonLoss`	The germline path calls the point-variant classifier for the SV, bypassing the `structural_variant` annotator. Without `germline=`, the same DEL gives `LargeDeletion`.
Splice donor SNV + cis coding SNV, with a `phase_resolver` that reports cis	Separate `Substitution` and `SpliceOutcomeSet`; no `HaplotypeEffect`	`build_haplotype_effects` skips any group that `apply_variants_to_transcript` can't build, without a warning (`phasing.py:358-360`).
DEL + cis coding SNV, with a `phase_resolver`	`LargeDeletion` and `Substitution`; no `HaplotypeEffect`	Same silent skip.
DEL starting at donor +2 and running into the intron	`Intronic`	Splice outcomes attach only to `StructuralVariantEffect` (`annotators/structural_variant.py:698`), so a deletion of donor +3..+6 has no splice candidates. A 1 bp point deletion at +3 gives a `SpliceOutcomeSet`.
DEL starting at the last base of exon 5 (removes the donor GT)	`LargeDeletion` with no splice candidates	`_breakpoint_splice_window` looks only at intronic breakpoints (`:193-213`), so a breakpoint on the exon boundary never gets splice outcomes.

RNA evidence doesn't compose either. apply_rna_evidence_to_effects only looks at the top-level effect (rna_evidence.py:192-217). For a PhaseCandidateSet wrapping SpliceOutcomeSets, observed splice isoforms are appended to the outer phase set as extras, instead of reconciling the inner splice sets.

What these have in common: each source builds its own wrapper around the output of the path before it, and the paths don't hand off to each other. A shared model would fix all of them at once. For each hypothesis (phase × splice mechanism × SV junction), build one MutantTranscript and classify it. That's also the direction proposed in #420.

Start with the interaction points named in germline.py, phasing.py, annotators/structural_variant.py, and rna_evidence.py, then read the direction in #420. Trace how each path wraps or skips effects before designing a shared hypothesis model. Done means the listed splice, phase, haplotype, SV, breakpoint, and RNA-evidence combinations compose without silently dropping information.

Written by the indexing model from the issue text.

Splice, germline-phase, haplotype and SV uncertainty don't compose

Description

Contributor guide

First steps

Research direction

Assessment