#38 · Transcriptomics

Models predict NMD sensitivity and calibrate RNA-seq splicing estimates

The dominant thread is supervised or model-based inference of RNA-level regulatory outcomes from transcriptomic data, here focused on nonsense-mediated decay (NMD). TrunCat, a gradient-boosting classifier, predicts NMD sensitivity of protein-truncating variants from TOPMed paired whole-genome and RNA-seq data, reaching about 78% ROC-AUC. A second preprint corrects percent-spliced-in estimates for NMD-targeted transcripts using a single-parameter linear fractional model and Gaussian mixtures separating NMD-sensitive transcripts. BRAID addresses a parallel problem in splicing quantification by calibrating RNA-seq ΔPSI estimates with split conformal intervals derived from RT-PCR residuals, achieving 0.971 RT-PCR coverage across 139 cassette exons. What is new is the linkage of paired genomic and transcriptomic data to variant-level NMD prediction, and explicit parameterized correction of NMD-induced bias in splicing metrics, with uncertainty quantified against orthogonal RT-PCR.