Detection of aberrant splicing events in RNA-seq data using FRASER

Abstract

Aberrant splicing is a major cause of rare diseases. However, its prediction from genome sequence alone remains in most cases inconclusive. Recently, RNA sequencing has proven to be an effective complementary avenue to detect aberrant splicing. Here, we develop FRASER, an algorithm to detect aberrant splicing from RNA sequencing data. Unlike existing methods, FRASER captures not only alternative splicing but also intron retention events. This typically doubles the number of detected aberrant events and identified a pathogenic intron retention in MCOLN1 causing mucolipidosis. FRASER automatically controls for latent confounders, which are widespread and affect sensitivity substantially. Moreover, FRASER is based on a count distribution and multiple testing correction, thus reducing the number of calls by two orders of magnitude over commonly applied z score cutoffs, with a minor loss of sensitivity. Applying FRASER to rare disease diagnostics is demonstrated by reprioritizing a pathogenic aberrant exon truncation in TAZ from a published dataset. FRASER is easy to use and freely available.

How Machine Learning and Statistical Models Advance Molecular Diagnostics of Rare Disorders Via Analysis of RNA Sequencing Data

“…A |Δ ψ| > 0.3 is recommended as an additional filter for the identification of pathological-relevant variation, where it corresponds to the difference between the observed ψ and the expected ψ. Application of FRASER in the rare disease cohort from Kremer et al (2017) identified all three previously detected pathogenic splicing aberrations, plus an intron-retention event missed by LeafCutter, and a synonymous variant causing a splice defect missed by Kremer et al (Mertes et al, 2021). FRASER was also applied to the GEUVADIS multicenter and multi-ancestry cohort and was able to remove sample covariation for all metrics (Yépez et al, 2021b).…”

“…Bearing in mind the limitations of the approaches described, three specialized methods to systematically detect aberrant splicing were developed: FRASER, LeafCutterMD, and SPOT. FRASER (Find Rare Splicing Events in RNAseq) is an approach combining machine learning and statistical models to detect aberrant splicing from RNAseq data (Mertes et al, 2021). Using the same rationale as OUTRIDER for detecting aberrant expression, FRASER uses a denoising autoencoder automatically controlling for latent confounders.…”

“…By inserting outliers in samples from skin and brain from GTEx, FRASER obtained higher precision in detecting the simulated outliers at all recall levels than SPOT and LeafCutterMD (Mertes et al, 2019). Following the principle that rare variants in the splice regions can disrupt splicing (Rivas et al, 2015), Mertes et al (2021) also benchmarked by performing an enrichment of rare variants in the splice region among splicing outliers called by the three specialized methods. FRASER obtained a higher enrichment of rare splice-region variants and variants predicted to affect splicing [using MMsplice (Cheng et al, 2019)] than SPOT and LeafCutterMD across all tissues from the GTEx dataset (GTEx Consortium, 2017).…”

“…There are no studies about replication of outliers using multiple biopsies from the same tissue from the same patient. However, Ferraro et al (2020) and Mertes et al (2021) performed replication analyses using different tissues from the same individuals from the GTEx cohort. Ferraro et al (2020) found that a median of 5.1% expression outliers, 8.7% of splicing outliers (using SPOT), and 10.7% of mono-allelically expressed genes (using ANEVA-DOT) that were detected as aberrant in one tissue were replicated in another.…”

“…They used arbitrary thresholds (Cummings et al, 2016(Cummings et al, , 2017 and statistical methods (Kremer et al, 2016(Kremer et al, , 2017 to call transcript aberrations. Since then, the bioinformatic approaches applied to RNAseq data have been advanced with machine learning methods such as OUTRIDER or FRASER, translating into more precise outlier calling (Brechtmann et al, 2018;Mertes et al, 2021). Machine learning methods are known to provide more robust predictions than statistical models (Bzdok et al, 2018).…”