← Về thư mục

Count Matrix QC - Usage Guide

Overview

Quality control of count matrices is essential before differential expression analysis to identify outliers, batch effects, and sample quality issues.

Prerequisites

if (!require('BiocManager', quietly = TRUE))
    install.packages('BiocManager')

BiocManager::install('DESeq2')
install.packages('pheatmap')

Quick Start

Tell your AI agent what you want to do: - "Check the quality of my count matrix before DE analysis" - "Identify outlier samples in my RNA-seq data" - "Make a PCA plot to see if my samples cluster by condition"

Example Prompts

Basic QC

"Run QC on my count matrix and identify any problematic samples"

"Check library sizes and gene detection rates across my samples"

Visualization

"Create a sample correlation heatmap from my count data"

"Make a PCA plot colored by condition and batch"

Outlier Detection

"Check if any samples are outliers based on PCA and correlation"

"Should I remove sample X based on the QC metrics?"

Batch Effects

"Check if my samples show batch effects in PCA"

"Add batch correction to my DESeq2 design"

What the Agent Will Do

1. Load the count matrix and sample metadata
2. Calculate library sizes and gene detection rates
3. Normalize counts for visualization (VST or rlog)
4. Generate sample correlation heatmap
5. Create PCA plot to assess sample clustering
6. Identify any outliers or batch effects
7. Recommend next steps based on QC results

Key QC Checks

1. Library sizes - Total counts per sample
2. Gene detection - Number of genes with counts
3. Sample correlation - Replicates should cluster together
4. PCA - Samples should separate by condition, not batch
5. Outlier detection - Identify problematic samples

Quick QC in R

library(DESeq2)
library(pheatmap)

# Load data
dds <- DESeqDataSetFromMatrix(countData = counts, colData = coldata,
                              design = ~ condition)

# Filter low counts
dds <- dds[rowSums(counts(dds)) >= 10, ]

# Normalize
vsd <- vst(dds, blind = TRUE)

# Sample correlation
pheatmap(cor(assay(vsd)))

# PCA
plotPCA(vsd, intgroup = 'condition')

What to Look For

Good Signs

• Replicates cluster together in PCA
• High correlation (>0.9) between replicates
• Similar library sizes across samples
• Clear separation between conditions

Warning Signs

• Sample doesn't cluster with its group
• Low correlation with replicates
• Very different library size
• PCA driven by batch, not condition

Handling Problems

Outlier Samples

# Option 1: Remove outlier
dds <- dds[, colnames(dds) != 'outlier_sample']

# Option 2: Flag for sensitivity analysis
# Run with and without outlier

Batch Effects

# Known batch: put it in the design (never pre-correct counts for DESeq2/edgeR)
design(dds) <- ~ batch + condition

• removeBatchEffect/ComBat output is for visualization only; feeding it to the count model double-corrects.
• Unknown batch: estimate surrogate variables (sva/svaseq) or factors of unwanted variation (RUVSeq) and add to the design.
• Confounding is fatal: if batch tracks condition, regressing it out removes biology too. Cross-tabulate batch against condition first; a perfect confound is unfixable.

Sample Swaps

• Check sex genes (XIST for XX; RPS4Y1/UTY/DDX3Y for XY) against recorded sex; a mismatch is a swap until proven otherwise.

Low Library Size

• Consider excluding samples with very low depth, but document the reason.
• Do not deduplicate standard RNA-seq (high duplication is expected); deduplicate only with UMIs.

Recommended Thresholds

Metric	Good	Concerning
Library size	>5M	<1M
Genes detected (human poly-A bulk)	>12,000	<8,000
Replicate correlation	>0.95	<0.85
Mapping rate (upstream quant check)	>70%	<50%

Gene-detection numbers are organism-, tissue-, and protocol-specific; the values above are typical for human poly-A bulk RNA-seq. Calibrate against the cohort rather than treating them as universal cutoffs.

Tips

• Always run QC before differential expression analysis
• Use blind=TRUE for vst/rlog when doing unsupervised QC
• Document any samples removed and justify the decision
• Run sensitivity analysis if unsure about removing a sample
• Check for batch effects even if not explicitly part of the design