Differential Splicing - Usage Guide

Overview

Detect alternative splicing changes between conditions using rMATS-turbo (binomial LRT), leafcutter (Dirichlet-multinomial GLM), MAJIQ V3 deltapsi/HET (Bayesian posterior), SUPPA2 (empirical null), or Shiba (2025 SOTA at low coverage). Tools differ in statistical model, annotation dependence, and calibration regime - these determine which is right for a given experimental design.

Prerequisites

# Python tools (rMATS-turbo is not on PyPI; install via bioconda below)
pip install suppa pandas numpy

# R / Bioconductor
BiocManager::install(c('DESeq2'))
# leafcutter (GitHub, not Bioconductor)
# devtools::install_github('davidaknowles/leafcutter/leafcutter')

# CLI tools
conda install -c bioconda rmats regtools
# MAJIQ V3 from majiq.biociphers.org (academic license)

Quick Start

Tell your AI agent what you want to do: - "Find differential splicing between tumor and normal samples" - "Run rMATS and leafcutter in parallel and report concordant hits" - "Use MAJIQ-HET for differential splicing across a heterogeneous patient cohort" - "Compare splicing between treatment and control with low replicate count" - "Prioritize differential splicing events by combined statistical and biological significance"

Example Prompts

Standard Replicate Designs

"I have n=3 vs n=3 RNA-seq BAMs; run rMATS-turbo with FDR<0.05 and |dPSI|>0.10, then filter for >=10 junction reads per replicate."

"Use leafcutter Dirichlet-multinomial GLM on intron clusters from regtools junctions for annotation-free differential splicing."

Heterogeneous Cohorts

"I have 30 tumor and 30 normal samples from heterogeneous patients; use MAJIQ V3 HET module with posterior threshold P(|dPSI|>0.2)>0.95."

Low Replicate Count

"n=2 vs n=2 design - avoid SUPPA2; use leafcutter or Shiba with junction-imbalance correction."

Result Prioritization

"Compute combined score (-log10(FDR) * |dPSI|) and pull the top 50 events with NMD-status and protein-domain annotation."

"Cross-reference top hits with eCLIP RBP target databases to identify candidate trans-regulators."

What the Agent Will Do

Choose tools based on replicate count, cohort heterogeneity, and annotation availability
Configure stat tests (LRT, Dirichlet-multinomial GLM, Bayesian posterior, or empirical null)
Apply replicate-aware filtering (per-replicate minimum coverage)
Adjust for batch effects and confounders if covariates available
Produce ranked results with FDR, ΔPSI, and biological annotation
Recommend orthogonal validation (e.g. SpliceAI for variant-driven hits, FRASER2 for outliers)

Tips

Run two complementary tools (rMATS + leafcutter) and require concordance for high-confidence calls
For n=2 vs n=2, avoid SUPPA2 (poorly calibrated null); prefer leafcutter or Shiba
Document --b1/--b2 order - IncLevelDifference sign matches this
Check strandedness with RSeQC infer_experiment.py before quantification
Increased PSI of a poison exon decreases functional protein due to NMD - always check ORF
Report SF3B1 cancer signal as cryptic 3'ss ~10-30nt upstream of canonical
TDP-43 ALS/FTD signature: cryptic exons in UNC13A, STMN2, ATG4B
MAJIQ posterior P>0.95 ~ FDR<0.05 in many regimes but not exactly equivalent

splicing-quantification - Per-event PSI estimation
isoform-switching - DTU framework for functional consequences
sashimi-plots - Visualize significant events
outlier-splicing-detection - Single-sample-vs-cohort testing for rare disease
splice-variant-prediction - SpliceAI for variant-driven differential splicing
long-read-splicing - Differential analysis from full-isoforms
read-alignment/star-alignment - STAR 2-pass alignment required