Repository navigation
Combining splice-effect prediction with expression-outlier detection (FRASER2 + OUTRIDER): independent evidence or double-counting? #678
Description
Title: Combining splice-effect prediction with expression-outlier detection (FRASER2 + OUTRIDER): independent evidence or double-counting?
Posting this at Julien's suggestion, so the answer is discoverable rather than sitting in
an inbox. Question was asked by email; the reply is quoted below with permission.
Question
I'm an engineer working on genomic variant interpretation. We've built a pipeline that
runs SpliceAI-style splice prediction and FRASER/OUTRIDER-style expression outlier
detection as one module inside a deterministic ACMG classification system.I've read the DROP pipeline papers closely and hit something I can't resolve from the
paper alone.Is combining splice-effect prediction with expression-outlier detection statistically
sound, or are we at risk of double-counting the same underlying signal?
Answer (Julien Gagneur)
My short answer: FRASER2 is based on a fraction of reads (split reads over all reads
related to either the donor or the acceptor), which roughly controls for expression
level. Hence one could argue this an orthogonal analysis.I would refrain to state that this is correct from a statistical point of view.
Variations in expression levels affect splicing outlier detection: if a gene is at a
lower level of expression then calling aberrant splicing is more difficult to do. This
happens in practice: when aberrant splicing triggers NMD, genes get down-regulated.
Those genes may not be called by FRASER because there is not enough reads to make the
call, yet can be called by OUTRIDER. This situation is frequent.
Activity