Bio Small Rna Seq Differential Mirna

Version Compatibility

Reference examples tested with: DESeq2 1.42+, edgeR 4.0+, apeglm 1.24+, EnhancedVolcano 1.20+, pheatmap 1.0.12+, ggplot2 3.5+

Before using code patterns, verify installed versions match. If versions differ:

• R: packageVersion('<pkg>') then ?function_name to verify parameters

If code throws ImportError, AttributeError, or TypeError, introspect the installed package and adapt the example to match the actual API rather than retrying.

Differential miRNA Expression

"Find differentially expressed miRNAs between my conditions" -> Test a raw miRNA count matrix for expression changes, accounting for the compositional fragility that makes miRNA normalization harder than mRNA.

• R: DESeq2::DESeq() or edgeR::glmQLFTest() on RAW miRNA counts

The governing principle: a few miRNAs dominate the library, so normalization is the dominant decision

A miRNA library is not a gently varying pool of thousands of features like an mRNA library. A handful of tissue-dominant miRNAs can be more than half of all reads, and the expressed repertoire is only hundreds to low-thousands of miRNAs. Two consequences follow, and they matter more than the choice of DE engine. First, global-scaling normalizers (DESeq2 median-of-ratios, edgeR TMM) assume most features are not differentially expressed and the count distribution is roughly symmetric; when one dominant miRNA shifts between conditions it absorbs the size factor and distorts every other miRNA's normalized value, manufacturing phantom changes. The normalization choice genuinely changes which miRNAs are called DE (Garmire 2012; Tam 2015) - so filter low-count noise FIRST, inspect whether a few miRNAs dominate, and report the normalizer. Second, empirical-Bayes dispersion shrinkage borrows strength across features, so with only hundreds of miRNAs the prior is estimated from a small, noisy population and is weaker than on ~20k genes; apeglm LFC shrinkage matters more for the many low-count miRNAs.

Two reframes prevent classic mistakes. RPM is for display and cross-sample viewing, never for testing - hand RAW counts to DESeq2/edgeR, which model the count distribution themselves. And a miRNA going DOWN does not necessarily mean transcriptional repression: target-directed miRNA degradation (TDMD, via ZSWIM8) lets a highly complementary target trigger decay of the miRNA itself (Han 2020; Shi 2020), so interpret a drop as a change in steady-state level, not automatically as reduced biogenesis.

A third decision is the level of testing. Mature-miRNA-level DE answers "which miRNAs changed" with good power; isomiR-level DE is sparser (more features and zeros, weaker per-feature power, heavier multiplicity), and 5' isomiRs shift the seed and can move OPPOSITE to the canonical mature form - so never silently sum 5' isomiRs into the mature count. Collapse to mature for the standard question; test at isomiR resolution only when isomiR identity is the biology.

Decision: which normalization / method

When a perturbation moves the WHOLE pool (e.g. Dicer/Drosha loss), every internal normalizer - including CLR - forces the average change to zero and is blind to it; only external spike-ins or cell-number normalization detect a global shift (Lovén 2012).

Load the count matrix

Goal: Read raw miRNA counts and build sample metadata for testing.

Approach: Load the miRge3/miRDeep2 count CSV (raw, not RPM) and define the condition factor.

library(DESeq2)

counts <- read.csv('miR.Counts.csv', row.names = 1)   # RAW counts, not RPM
coldata <- data.frame(
    condition = factor(c('control', 'control', 'treated', 'treated')),
    row.names = colnames(counts))

DESeq2 analysis

Goal: Identify miRNAs that change between conditions with small-RNA-aware filtering and shrinkage.

Approach: Build a DESeqDataSet from rounded raw counts, prefilter at a lower threshold than mRNA, run DESeq2, then shrink LFCs with apeglm for the many low-count miRNAs.

dds <- DESeqDataSetFromMatrix(
    countData = round(counts),     # DESeq2 needs integers
    colData = coldata,
    design = ~ condition)

# Lower prefilter than mRNA: miRNA libraries have fewer total counts, and most
# miRBase entries are near-zero noise. Justify the threshold; do not test everything.
keep <- rowSums(counts(dds)) >= 10
dds <- dds[keep, ]

dds <- DESeq(dds)

# Inspect for compositional risk: a size factor far from 1, or one miRNA that is a
# large fraction of reads, is a warning that median-of-ratios may be distorted.
sizeFactors(dds)

res <- results(dds, contrast = c('condition', 'treated', 'control'))
# apeglm shrinks via a named coef; for an arbitrary/multi-level contrast not expressible
# as one coef, use type = 'ashr' instead.
res_shrunk <- lfcShrink(dds, coef = 'condition_treated_vs_control', type = 'apeglm')
res_shrunk <- res_shrunk[order(res_shrunk$padj), ]