Transcripts Per Million (TPM)

TPM's defining characteristic is its order of operations: it normalizes for gene length before sequencing depth. Raw counts are first divided by gene length in kilobases to produce reads per kilobase (RPK). This RPK value is then scaled by a per-sample factor so that the sum of all TPMs equals one million. This contrasts with FPKM/RPKM, which divide by length last, making TPM the preferred metric for cross-sample comparisons because the proportion of reads mapped to a gene is directly comparable.

In any given sample, the sum of all TPM values across all genes is always 1,000,000. This invariant property means TPM represents the relative molar concentration of transcripts. A gene with a TPM of 500 in sample A and 500 in sample B constitutes exactly 0.05% of the total transcript pool in both samples, regardless of differences in total sequencing depth. This makes TPM ideal for comparing transcript proportions across heterogeneous samples.

TPM values are directly interpretable as molecular fractions. If Gene X has a TPM of 100,000, it represents 10% of the total mRNA molecules in that sample. This proportional nature is critical for analyses where the compositional nature of RNA-seq data matters, such as when comparing samples with vastly different total RNA output or when a few highly expressed genes dominate the library. TPM avoids the compositional artifacts that plague raw counts and RPKM.

While RPKM and FPKM also normalize for length and depth, their inconsistent sum across samples prevents direct proportional comparison. A gene with an RPKM of 50 in two samples does not necessarily represent the same fraction of the transcriptome. TPM was explicitly designed to resolve this flaw. The conversion from RPKM to TPM is straightforward: divide each gene's RPKM by the sum of all RPKMs in the sample and multiply by 10^6. Most modern bioinformatics pipelines, including Salmon and kallisto, output TPM natively.

TPM normalization relies on the concept of effective length, not just the genomic length of a gene. Effective length accounts for the fact that RNA-seq fragments cannot be unambiguously assigned to a transcript if they are shorter than the read length. Tools like RSEM and Salmon compute an effective length by subtracting the mean fragment length from the transcript length, ensuring that the normalization factor accurately reflects the number of mappable positions. This is especially important for short transcripts and isoforms.

TPM is not a panacea. It is a within-sample normalization and does not correct for batch effects or compositional bias between samples. In experiments where a few genes are massively differentially expressed, the constant-sum constraint can induce spurious differential expression in unchanged genes. For such cases, tools like DESeq2 and edgeR use more sophisticated normalization (e.g., median-of-ratios, TMM) that are robust to compositional outliers. TPM is best used for visualization, clustering, and as input to exploratory analyses, not as the final normalization for statistical testing.

The simpler precursor to TPM that normalizes solely for sequencing depth, ignoring gene length. CPM is calculated as (raw counts / total mapped reads) × 10⁶. While useful for filtering lowly expressed genes and comparing the same gene across samples, CPM fails to correct for gene length bias, making it unsuitable for comparing expression levels between different genes within a single sample. TPM addresses this limitation by adding a gene-length normalization step first.

A legacy normalization unit that normalizes for gene length first, then sequencing depth—the reverse order of TPM. RPKM divides read counts by gene length in kilobases, then by total mapped reads in millions. The critical difference: RPKM sums are inconsistent across samples, making cross-sample comparisons of transcript proportions unreliable. TPM was explicitly designed to fix this by ensuring every sample sums to exactly one million, enabling direct proportion comparisons.

The paired-end sequencing analog of RPKM. FPKM counts fragments rather than reads, where a fragment is the physical DNA piece sequenced from both ends. For single-end data, FPKM equals RPKM. Like RPKM, FPKM suffers from non-uniform sum totals across samples, making it inferior to TPM for proportion-based analyses. Many legacy datasets use FPKM, but modern pipelines increasingly prefer TPM for its mathematical consistency.

The statistical framework that identifies genes with significant expression changes between conditions. Tools like DESeq2 and edgeR operate on raw integer counts, not TPM, because they require the discrete nature of count data to model variance correctly. However, TPM is the preferred unit for visualization, clustering, and exploratory analysis after differential expression testing. Understanding when to use counts versus TPM is critical for valid biological conclusions.

A fundamental technical artifact in RNA-seq where longer transcripts produce more sequencing reads than shorter transcripts at the same expression level. This bias inflates counts for long genes, creating false positives in comparative analyses. TPM corrects for this by dividing raw counts by gene length in kilobases as the first normalization step, ensuring that expression values reflect true transcript abundance rather than transcript length.

The computational process of removing systematic technical biases from raw read counts. Effective normalization must address three sources of variation: sequencing depth (total reads per sample), gene length (longer genes generate more reads), and library composition (highly expressed genes consume disproportionate reads). TPM addresses the first two biases explicitly, while tools like TMM and DESeq2's median-of-ratios handle composition effects for differential testing.