ML-based splice prediction for variants outside canonical splice windows

[iskandr]

Background

SpliceOutcomeSet (#262, PR #292) wraps variants that the existing classifier already flags as splice-adjacent — SpliceDonor, SpliceAcceptor, ExonicSpliceSite, IntronicSpliceSite. That covers the canonical donor/acceptor di-nucleotides, the last 3 bases of each exon, and the first 3–6 intronic bases.

But plenty of splice-altering variants sit outside that window:

• Exonic splicing enhancer (ESE) / silencer (ESS) disruption: a silent or missense variant in the middle of an exon can disrupt an ESE motif and cause the exon to be skipped in the mature transcript. varcode today emits a bare Substitution/Silent for these; SpliceOutcomeSet cannot wrap them because the classifier never flagged them.
• Deep intronic variants that create or disrupt a cryptic splice site hundreds of bases from the canonical boundary.
• Branch point disruption (~20–50 bp upstream of the acceptor).

These require ML-based prediction (SpliceAI, SpliceTransformer, MMSplice, Pangolin) or RNA evidence to detect. Without them, the possibility-set model silently under-reports splice consequences for exonic/intronic variants outside the canonical window.

Scope

• Integrate a splice-prediction scorer as an optional dependency (SpliceAI is the obvious first target).
• When a variant is outside the canonical window but the scorer flags a high-probability splice change, wrap it in a SpliceOutcomeSet (or the successor multi-outcome abstraction from the OutcomeSet generalization work).
• Keep the dependency optional; default splice_outcomes=True uses canonical-window classification only, ML-informed wrapping is a second opt-in (e.g. splice_outcomes="ml" or splice_scorer=...).

Related

• Incorporate RNA-level evidence for variant effects #259 — RNA evidence narrows the possibility set post-hoc; ML prediction widens it up front. Both contribute to the same "one DNA event → many plausible proteins" story.
• Prototype multi-effect candidates for splice-site variants #262 — tracking issue for multi-effect splice modeling.
• PR Prototype multi-effect splice outcome possibility sets (#262) #292 — initial SpliceOutcomeSet prototype; this issue is a natural follow-up.