Incorporate RNA-level evidence for variant effects

Background

Varcode currently predicts variant effects purely from genomic (DNA) coordinates mapped against reference gene annotations. For many structural variants, the actual protein-level consequence cannot be reliably determined from WGS alone:

• A genomic translocation may or may not produce a fusion transcript — this depends on whether the rearrangement is transcribed and how the resulting RNA is spliced
• A fusion protein is obvious from RNA-seq even when the underlying genomic breakpoint is complex or ambiguous
• RNA-level events (cryptic exons, exon truncations, intron retention, circular RNAs) have direct protein consequences that are invisible to DNA-only analysis
• UTR extensions and alternative splicing around breakpoints can only be resolved with transcript-level data

Core modeling problem: one variant, many possible isoforms

A single DNA event — especially an SV — can produce multiple distinct spliced isoforms, each with a different protein consequence. Consider a translocation whose breakpoint falls in a non-coding region: depending on which cryptic splice sites are used, the resulting transcript could include different amounts of intronic/intergenic sequence, skip different exons, or frame-shift at different points. The set of possible protein products is combinatorial.

Without RNA evidence, the best we can do is enumerate a set of plausible splice outcomes based on nearby canonical and non-canonical splice signals, scored by likelihood. This is a departure from varcode's current approach, which returns a single deterministic effect per transcript. But it's the honest representation — a SpliceDonor effect today doesn't tell you what the resulting protein looks like, just that splicing is probably disrupted.

With RNA evidence, the possibility set collapses to what was actually observed: which isoforms are expressed, at what levels, producing which proteins. RNA data constrains and resolves the ambiguity that DNA-only analysis cannot.

This principle — DNA predicts possibilities, RNA resolves them — applies retroactively to existing splice-site variant handling (see "Implications for existing code" below), not just to new SV types.

Scope

Extend varcode's effect model to incorporate RNA-derived evidence, specifically:

New RNA-informed effect types

• FusionTranscript / FusionProtein: chimeric transcript joining exons from two genes, with the resulting protein sequence
• CircularRNA: back-spliced transcript forming a circle
• CrypticExon: novel exon not present in reference annotations, often from sequence that was previously intronic or intergenic
• ExonTruncation: partial exon usage at or near a breakpoint
• IntronRetention: failure to splice out an intron, with protein consequence
• UTRExtension: extended untranslated region altering regulation

Possibility-set model for splice effects

• When predicting from DNA alone, splice-disrupting variants should produce a set of possible effects (e.g., "exon skip leading to frameshift," "cryptic donor activation leading to truncated exon," "intron retention leading to premature stop"), each with an estimated plausibility
• When RNA evidence is available, this set is narrowed to the observed isoforms, with expression-level support
• The EffectCollection for a single variant should be able to represent this uncertainty: "here are the N possible outcomes, here is which ones we observed"

Reference sequence beyond exons

• Many novel isoforms incorporate sequence from outside annotated exonic regions — intronic sequence retained by intron retention, intergenic sequence recruited by a translocation, cryptic exons from previously non-coding regions
• Effect prediction must be able to access genomic sequence around breakpoints and splice sites, not just the annotated exon sequences that current code uses

Evidence model

• Record whether an effect was predicted from DNA (with associated uncertainty) vs. observed in RNA (with transcript model IDs and read support)
• Allow both to coexist for the same variant — the DNA prediction set and the RNA observations
• Distinguish between DNA/RNA concordant variants (seen in both) and RNA-only or DNA-only events
• Link effects to supporting transcript model IDs and read evidence

Implications for existing splice-site effect handling

Today, SpliceDonor, SpliceAcceptor, and ExonicSpliceSite each return a single effect. ExonicSpliceSite has an alternate_effect field ("what happens if splicing isn't disrupted"), but this is one alternative, not a set of possibilities. We should consider evolving these to return or be expandable into a set of possible outcomes — this makes the existing model more honest and provides a natural place for RNA evidence to narrow the set.

Relevant context

• Exacto identifies all of the above RNA variant types from long-read transcriptome assemblies, providing transcript structures, RNA variant calls, and DNA/RNA integrated variant tables.
• Exacto's translate-structs command produces per-residue primary structure tables with rna_variant_call_ids and dna_variant_call_ids columns, providing the evidence linkage.
• Exacto's integration table shows one dna_variant_call_id mapping to multiple transcript_model_id entries — directly reflecting the one-DNA-event-to-many-isoforms relationship.
• Related: #252 (gene fusion effects) — this issue provides the broader RNA-evidence framework that fusion support would fit within.

Start by reading the existing SpliceDonor, SpliceAcceptor, ExonicSpliceSite, and EffectCollection implementations, then review related issue #252 and Exacto's transcript and integration-table context. Define how DNA predictions, RNA observations, transcript model IDs, read support, and multiple isoform outcomes coexist; done means the effect model supports the requested RNA-informed types and uncertainty without losing existing splice handling.