Reads Per Kilobase Million (RPKM)

A paired-end analog to RPKM. While RPKM normalizes by single-end reads, FPKM normalizes by fragments, counting a concordant read pair as a single unit. This prevents the double-counting of a cDNA fragment sequenced from both ends. The calculation divides the number of mapped fragments by the transcript length in kilobases and the total mapped fragments in millions. For single-end libraries, RPKM and FPKM are mathematically equivalent.

A cross-sample comparable normalization metric that addresses a critical flaw in RPKM. TPM reverses the order of operations: it first normalizes for gene length (reads per kilobase), then scales by the sum of all length-normalized values per million. This ensures the sum of all TPMs in a sample is constant (1 million), making the proportion of reads mapping to a gene directly comparable across samples. RPKM lacks this property, as its denominator varies by library size.

The simplest library-size normalization method. CPM divides raw read counts by the total mapped reads in millions, without adjusting for gene length. It is suitable for comparisons within a single sample or when gene lengths are invariant (e.g., comparing the same gene across conditions). For differential expression analysis between genes of different lengths, CPM is biased; longer genes accumulate more reads and appear artificially overexpressed.

A between-sample normalization method implemented in the edgeR package. TMM calculates scaling factors based on the weighted trimmed mean of log expression ratios between samples, after excluding highly variable and lowly expressed genes. Unlike RPKM, which normalizes within a single sample, TMM is designed to make library sizes comparable across samples by accounting for compositional biases where a few highly expressed genes distort the total count distribution.

A robust normalization method that computes a size factor for each sample by taking the median ratio of each gene's count to its geometric mean across all samples. This approach assumes most genes are not differentially expressed and is highly resistant to outliers and high-count genes. It is a prerequisite step before the negative binomial modeling in DESeq2, providing a more statistically grounded alternative to RPKM for differential expression testing.

A nuanced correction for the fact that RNA-seq read coverage is non-uniform across transcripts due to biases in fragmentation, priming, and GC content. Tools like RSEM and Salmon compute an effective length that accounts for the empirical distribution of mapped reads, rather than using the theoretical transcript length. RPKM's use of a static kilobase denominator ignores these technical biases, making effective-length-aware metrics like TPM from transcript-level quantifiers more accurate.