TPM Normalization

TPM normalizes for gene length bias by first dividing raw read counts by the gene's length in kilobases, producing Reads Per Kilobase (RPK). This correction is essential because longer transcripts naturally generate more fragments during sequencing, artificially inflating their apparent expression. Without length normalization, a long, lowly expressed gene could appear more abundant than a short, highly expressed one. The RPK calculation ensures that each transcript's contribution is proportional to its molar concentration rather than its physical length.

After length correction, TPM applies a sequencing depth normalization by summing all RPK values in a sample and dividing by one million to create a scaling factor. Each gene's RPK is then divided by this factor and multiplied by 10^6. This ensures that the sum of all TPM values in a sample always equals 1,000,000, representing a complete transcriptome proportion. This property makes TPM values directly interpretable as the number of transcripts per million molecules, unlike FPKM or RPKM which do not sum to a constant.

The constant-sum property of TPM enables direct comparison across samples with different sequencing depths. Because each sample's TPM values sum to exactly one million, the proportion of transcripts allocated to each gene is independent of the total number of reads generated. This is critical for differential expression analysis where samples may have been sequenced to varying depths. A TPM of 50 for Gene A in both a 10-million-read sample and a 50-million-read sample indicates the same relative transcript abundance, assuming similar RNA composition.

While TPM and FPKM (Fragments Per Kilobase per Million) appear similar, their calculation order differs critically. FPKM normalizes for depth first, then length, resulting in values that do not sum to a constant across samples. This makes FPKM unsuitable for comparing transcript proportions between samples. TPM's order—length normalization first, then depth—produces a proportional measure. For example, a spike-in control added at equal concentration to two samples will yield identical TPM values but different FPKM values if total RNA composition differs.

Modern TPM implementations use effective length rather than total transcript length to account for the fact that RNA-seq fragments cannot map uniformly to transcript ends. Effective length is calculated as the transcript length minus the mean fragment length plus one, correcting for the reduced mappability near transcript boundaries. Tools like Salmon and Kallisto incorporate this correction during pseudoalignment, producing more accurate TPM estimates by modeling the fragment start-site distribution and sequence-specific biases that affect read coverage uniformity.

TPM values are compositional data, meaning they represent relative proportions constrained to sum to a constant. This introduces a dependency structure where a genuine increase in one highly expressed gene forces apparent decreases in all others, even if their absolute expression is unchanged. Statistical analyses must account for this using log-ratio transformations such as the centered log-ratio (CLR) or additive log-ratio (ALR) before applying methods like principal component analysis or linear regression that assume unconstrained data.