Chapter 4 Normalization methods

Different assays require different methods.

4.1 Library-size scaling

The simplest approach divides each sample by its total counts (or reads) to account for sequencing depth.

4.1.1 Counts-per-million (CPM)

There are other versions of this type of normalization (RPKM, FPKM ,TPM,…) - all of them try to normalize for the library size (the number of reads in a given experiment) but some of them incorporate other factors as well.

CPM - normalizes for just library size.
TPM - normalzie for gene length, then library size. FPKM/RPKM - similiar but one uses reads and the other fragments. normalizes for gene length and library size at the same time. TPM is preferred.

4.2 Distribution-based normalization

Methods like quantile normalization force sample distributions to match and are common in microarrays.

4.3 RNA-seq composition-aware methods

Two widely used methods:

• TMM (edgeR): computes scaling factors to adjust effective library sizes.
• RLE / size factors (DESeq2): median-of-ratios approach.

4.3.1 edgeR: TMM

TMM looks at most genes, ignores the weird ones, figures out how different two samples really are, and uses that to scale the samples so they’re comparable.

library(edgeR)

dge <- DGEList(counts = counts)
dge <- calcNormFactors(dge, method = "TMM")
logcpm <- cpm(dge, log = TRUE, prior.count = 1)

4.3.2 DESeq2: size factors

RLE finds what “most genes” are doing compared to a typical sample. It adjusts each sample so that the median gene looks the same in all samples.

library(DESeq2)

dds <- DESeqDataSetFromMatrix(countData = counts, colData = meta, design = ~ condition)
dds <- estimateSizeFactors(dds)

norm_counts <- counts(dds, normalized = TRUE)

4.4 Single-cell note

In scRNA-seq, normalization often starts with per-cell library-size scaling followed by log1p and sometimes variance stabilization. The concepts (make samples comparable; avoid composition artifacts) are the same.