Differential miRNA Expression - Usage Guide

Overview

Identify differentially expressed miRNAs between conditions with DESeq2 or edgeR, accounting for the fact that miRNA data is compositionally harder than mRNA. A handful of hyper-abundant miRNAs can be more than half of all reads, so global-scaling normalization is fragile and the normalization choice - not the DE engine - is the dominant decision. The workflow filters near-zero noise first, inspects whether a few miRNAs dominate, tests RAW counts (RPM is display only), and shrinks fold-changes with apeglm for the many low-count miRNAs. Biofluid data has no trustworthy endogenous normalizer and needs spike-ins plus hemolysis and batch control, and a miRNA dropping can reflect target-directed degradation (TDMD) rather than transcriptional repression.

Prerequisites

BiocManager::install(c('DESeq2', 'edgeR', 'apeglm', 'EnhancedVolcano', 'pheatmap'))
# Optional compositional sensitivity analysis: BiocManager::install('ALDEx2')

Quick Start

Tell your AI agent: - "Run DESeq2 on my raw miRge3 count matrix between treatment and control" - "Check whether a few dominant miRNAs are distorting my normalization" - "Cross-check the result with edgeR" - "My data is plasma - which normalizer should I use?" - "Shrink the fold-changes with apeglm and report base mean for each hit"

Example Prompts

Basic DE Analysis

"Find DE miRNAs between tumor and normal from my raw count matrix"

"My size factors are far from 1 and one miRNA is 60% of reads - what normalization should I use?"

"Run edgeR quasi-likelihood as a cross-check of the DESeq2 calls"

Compositional and Biofluid Caveats

"Run an ALDEx2 compositional sensitivity analysis alongside DESeq2"

"Normalize my plasma miRNA-seq with cel-miR-39 spike-ins and flag hemolysis"

"A perturbation may have changed the whole miRNA pool - how do I even detect that?"

Visualization and Export

"Make a volcano plot and a heatmap of significant miRNAs"

"Export significant miRNAs with their base mean and shrunk log2FC"

What the Agent Will Do

Load the RAW count matrix (not RPM) and build sample metadata
Prefilter at a lower threshold than mRNA and inspect for compositional dominance (size factors, top-miRNA fraction)
Run DESeq2 (or edgeR TMM) and apply apeglm LFC shrinkage for low-count miRNAs
Filter on FDR and effect size while keeping expression level visible, to avoid calling tiny miRNAs DE
Visualize and export, and flag biofluid/TDMD caveats where relevant

Tips

The normalization choice changes the result more than the DE engine does - filter first, inspect dominance, and report the normalizer
RPM is for display only; feed RAW counts to DESeq2/edgeR
Use a lower prefilter than mRNA (miRNAs have fewer total counts and most miRBase entries are noise), but justify the threshold
apeglm shrinkage matters more here because few features weaken the dispersion prior
edgeR names its FDR column FDR, not padj; pass group to filterByExpr
Biofluids have no endogenous normalizer - use cel-miR-39 spike-ins, flag hemolysis (miR-451a:miR-23a-3p), and model batch
A global pool shift (e.g. Dicer/Drosha loss) is invisible to all internal normalizers; only spike-ins or cell-number normalization detect it
A miRNA going down can be TDMD (target-driven decay), not transcriptional repression - check pri/pre-miRNA before claiming repression
Decide mature-level vs isomiR-level DE: collapse to mature for a well-powered "which miRNAs changed" question; isomiR-level is sparser and 5' isomiRs can move opposite the canonical form, so never silently sum them
apeglm shrinks a named coef; use ashr for arbitrary or multi-level contrasts

mirge3-analysis - Produces the raw count matrix
mirdeep2-analysis - Alternative quantification
target-prediction - Targets of DE miRNAs
differential-expression/deseq2-basics - General DESeq2 mechanics
differential-expression/edger-basics - General edgeR mechanics