RPKM Demystified: A Thorough British Guide to Reads Per Kilobase of Transcript per Million Mapped Reads

RPKM is one of the oldest normalization measures used in RNA sequencing to estimate how abundantly a gene is expressed in a sample. Short for Reads Per Kilobase of transcript per Million mapped reads, this metric was designed to make expression levels comparable across genes of different lengths and across samples with varying sequencing depth. While newer approaches have emerged, RPKM remains a foundational concept in transcriptomics, and understanding its logic, strengths, and limitations is invaluable for researchers, students, and clinicians navigating RNA-seq data.

What is RPKM?

RPKM stands for Reads Per Kilobase of transcript per Million mapped reads, and it expresses the abundance of a transcript as a function of two factors: the length of the transcript and the total number of reads sequenced in the sample. The intuition is straightforward. Longer genes have more opportunities to accumulate reads simply by virtue of their size, while samples with deeper sequencing will naturally yield more reads across all genes. By normalising for both gene length and sequencing depth, RPKM enables cross-gene comparisons within a single sample and, with cautions, across samples.

In other words, RPKM attempts to level the playing field so that a read count is not biased toward particularly long genes or toward libraries with higher sequencing depth. This makes the relative expression of two genes in the same sample more interpretable and provides a common language when comparing expression landscapes across tissue types or conditions—at least in theory.

How is RPKM calculated?

The basic idea behind RPKM is captured by a simple formula. For a given gene, C denotes the number of reads mapped to that gene, N denotes the total number of reads mapped in the entire sample, and L denotes the gene’s length in bases. The RPKM value is calculated as:

RPKM = (C × 10^9) / (N × L)

Let’s break down the components in plain terms:

• Reads mapped to the gene (C): The raw count of sequencing reads that align to the transcript or gene region of interest.
• Total mapped reads (N): The total number of reads that passed quality control and were aligned to the reference genome or transcriptome within the sample.
• Gene length (L): The annotated length of the transcript or gene, measured in bases. For a transcript, this is the full transcript length; for a gene, it may be the combined length of all transcript isoforms, depending on the annotation used.

The factor of 10^9 is simply a scaling constant that converts the units into reads per kilobase per million. If you express L in kilobases (i.e., L_kb = L / 1000) and N in millions (N_m = N / 1,000,000), the formula becomes:

RPKM = C / (L_kb × N_m)

This rephrasing highlights the two normalization axes: gene length and sequencing depth. RPKM values therefore reflect how many reads would be expected for that gene if the library had a standard depth of one million reads and if the gene were exactly one kilobase long.

RPKM in practice: from sequencing reads to expression values

Turning raw sequencing data into RPKM values involves several practical steps. Each stage can influence the final expression estimates, so understanding the workflow helps in diagnosing discrepancies and in choosing appropriate downstream analyses.

Step 1: Generating sequencing reads

Biological samples are prepared, cDNA is generated, and libraries are sequenced. The depth of sequencing (i.e., the total number of reads) will directly affect N in the RPKM calculation. Researchers often aim for a depth that balances coverage with cost, though the appropriate depth depends on the organism, tissue complexity, and the research question.

Step 2: Aligning reads to a reference

Quality control is performed to filter low-quality reads. Reads are then aligned to a reference genome or transcriptome using aligners such as STAR, HISAT2, or TopHat. The alignment step determines which reads can be assigned to genes and transcripts, influencing C in the RPKM formula.

Step 3: Counting reads per gene

Once reads are aligned, counts are tallied for each gene or transcript. Tools such as HTSeq-count or featureCounts are commonly used. A critical consideration here is whether read counts are assigned to genes as a whole or to specific isoforms. The chosen strategy affects L and the resulting RPKM values, especially for genes with multiple alternative transcripts.

Step 4: Normalisation for gene length and depth

With C, N, and L defined, RPKM values can be computed for every gene. In practice, researchers often generate a genome-wide table of RPKM values to create expression atlases, heatmaps, or supplementary data for publications. It’s important to note that RPKM is most meaningful when comparing expression patterns within a single sample or, with caution, across samples processed in an equivalent manner.

Step 5: Quality checks and interpretation

After obtaining RPKM values, visual checks such as scatter plots of RPKM between samples, MA plots, and density plots help identify anomalies. Differences in library composition, technical bias, or annotation inconsistencies may show up as unexpected shifts in the expression landscape, prompting further investigation.

RPKM vs FPKM vs TPM: understanding the differences

Over time, RNA-seq practitioners have refined how gene expression is quantified. Three related metrics often appear in discussions: RPKM, FPKM, and TPM. While they share the same underlying idea of normalising for gene length and sequencing depth, they differ in their counting approach and intended comparisons.

RPKM vs FPKM

RPKM (reads per kilobase per million mapped reads) uses raw read counts (C) for single-end data or for regions that are treated as independent. FPKM (fragments per kilobase per million mapped fragments) is conceptually similar but is designed for paired-end sequencing, where each paired read pair represents a single fragment. In practice, FPKM replaces C with the number of fragments and remains divided by gene length and total mapped reads. For many legacy datasets generated with paired-end sequencing, FPKM values are reported, and the same caution about cross-sample comparability applies.

TPM: a popular alternative

TPM stands for transcripts per million. The key distinction is the order of operations. TPM first normalises for gene length to obtain reads per kilobase (RPK), then scales by the sum of all RPK values so that the total across all genes in a sample equals one million. This makes TPM more directly comparable across samples than RPKM, particularly for cross-sample abundance estimates within a single tissue or condition. In modern practice, TPM is often preferred when comparing expression levels between samples, while many pipelines operate on raw counts for differential expression analyses using methods such as DESeq2 or edgeR.

Limitations and caveats of RPKM

RPKM was a landmark in normalisation, but it has notable limitations that researchers should recognise when interpreting results.

• Cross-sample comparability: RPKM values are not strictly comparable across samples if library compositions differ substantially. Differences in RNA composition, sequencing biases, or highly expressed genes can distort read distributions, making direct RPKM-to-RPKM comparisons problematic.
• Gene length ambiguity: RPKM relies on accurate gene length annotations. Alternative transcript isoforms and incomplete or inconsistent annotations can bias L and therefore the final RPKM.
• Single-end vs paired-end: For paired-end data, FPKM is often used instead of RPKM to account for fragment-level counts. Mixing single-end and paired-end data without proper harmonisation can lead to misleading comparisons.
• Low counts and variability: Genes with low counts can produce unstable RPKM estimates. Filtering out low-expressed genes before downstream analysis is a common and prudent practice.
• Alternative metrics: Modern differential expression analyses typically operate on raw counts and use robust statistical models that consider dispersion and library size, offering better control of false positives than RPKM-based methods.

Appropriate contexts for RPKM

Despite its limitations, RPKM remains useful in certain scenarios. It can be a convenient summary statistic for exploratory data analysis, visualisation, or educational materials that illustrate the effect of gene length and sequencing depth on read counts. In historical datasets, RPKM was frequently used for generating gene expression atlases and for comparing expression within samples or across experiments that share identical library preparation and sequencing depth.

A modern workflow: when to use RPKM, and when to prefer alternatives

For current RNA-seq analyses, a practical approach balances the strengths of different metrics:

• Differential expression analyses: Use raw counts from feature counting with tools such as DESeq2 or edgeR, which model count data and account for library size and dispersion. This approach provides robust statistical testing for differential expression between conditions.
• Within-sample abundance estimates: If you need a sense of transcript abundance within a single sample or to compare highly similar samples with consistent library composition, TPM or properly normalised RPKM values can be informative.
• Cross-sample comparisons with caution: If you must compare across samples, ensure consistent library preparation, sequencing depth, and annotation. Consider normalising to TPM or using counts with advanced methods that mitigate compositional bias.

Practical tips for researchers using RPKM

• Stay consistent with annotation: Use the same gene models across all samples to avoid length-related artefacts. Mismatched annotations can artificially alter L and C, skewing RPKM.
• Filter low-expressed genes: Remove genes with very low counts before calculating RPKM or performing downstream analyses to improve robustness and interpretability.
• Be mindful of library composition: In samples with extreme expression of a few genes, RPKM may not reflect overall biology. Consider TPM or methods that explicitly address compositional biases.
• Document the process: Record the exact steps, including alignment software, version, counting strategy (gene- or transcript-level), and annotation, so that others can reproduce the RPKM calculations if needed.
• Use visual diagnostics: MA plots, density plots, and correlation analyses can reveal inconsistencies that might affect RPKM interpretations and highlight the need for alternative normalisation strategies.

RPKM in data visualisation and interpretation

When visualising RPKM values, it’s important to use appropriate scales. Log2-transformed RPKM (often with a small pseudocount to handle zeros) helps stabilise variance and makes patterns more interpretable in heatmaps, clustering, and principal component analyses. However, because RPKM can be biased by library composition, particularly when comparing across samples, visualisations should be complemented with analyses based on raw counts or TPM values where possible.

Case studies and historical context

RPKM has featured prominently in numerous early RNA-seq studies, contributing to insights into tissue-specific expression, developmental trajectories, and disease-associated transcriptional changes. In some classic demonstrations, researchers charted how gene length and sequencing depth together shape observed expression landscapes, using RPKM to illustrate the necessity of normalisation. In contemporary practice, these same studies are frequently revisited with TPM or count-based approaches to validate findings and ensure that conclusions are robust to newer normative standards.

Glossary of key terms

RPKM

Reads Per Kilobase of transcript per Million mapped reads; a normalisation method for RNA-seq expression estimates that accounts for gene length and sequencing depth.

FPKM

Fragments Per Kilobase of transcript per Million mapped fragments; the paired-end analogue of RPKM.

TPM

Transcripts Per Million; a normalised expression measure that facilitates cross-sample comparisons by standardising for library size after length normalisation.

DESeq2

A widely used software package for differential expression analysis based on count data, modelling dispersion and normalisation factors.

edgeR

Another prominent package for differential expression analysis using count data, with robust methods for small sample sizes and complex designs.

Counts

The raw number of sequencing reads mapping to a gene or transcript before normalisation.

Frequently asked questions about RPKM

Q: Can I compare RPKM values across different studies?

A: Direct cross-study comparison of RPKM is not recommended unless the studies share identical protocols, annotations, and sequencing depth. Copying RPKM values from one study to another can be misleading if library preparation or sequencing platforms differ.

Q: Should I use RPKM for differential expression?

A: For robust differential expression analysis, it is better to use raw counts with a statistical model designed for count data (such as DESeq2 or edgeR). RPKM may be useful for exploratory analyses or for presenting expression magnitudes within a single sample, but it is not ideal for identifying statistically significant changes between conditions.

Q: How do I convert RPKM to TPM?

A: A common approach is to treat RPKM values as proportional to expression levels and scale them so that the sum across all genes in a sample equals one million. A typical formula is TPM_i = RPKM_i / sum_j (RPKM_j) × 1,000,000. This provides a convenient, interpretable cross-sample scale while preserving gene-length normalisation.

Q: What about single-end versus paired-end data?

A: For single-end data, RPKM is appropriate within the same methodological framework. For paired-end data, FPKM is often used because it counts fragments (paired reads) rather than individual reads. When using mixed data types, maintain consistency and be explicit about the chosen metric in your analyses and reporting.

Final thoughts: treating RPKM with respect in modern analyses

RPKM remains a valuable historical and educational milestone in transcriptomics. It introduced the critical realisation that both gene length and sequencing depth can bias raw read counts, and it prompted the development of more sophisticated normalisation strategies. In today’s practice, researchers typically rely on raw counts for differential expression analyses and favour TPM for cross-sample abundance comparisons where appropriate. Nevertheless, understanding RPKM, its calculation, and its limitations equips researchers to interpret older data accurately, evaluate new studies with a critical eye, and communicate transparently about the methods behind expression estimates.

Concluding guidance for researchers and students

When embarking on RNA-seq data analysis, consider the following steps to use RPKM responsibly, while staying aligned with current best practices:

• Clarify the objective: within-sample abundance versus cross-sample comparisons or differential expression.
• Standardise annotations across samples to avoid length or transcript discrepancies influencing conclusions.
• Prefer raw counts for differential expression analysis using established statistical tools.
• Use TPM for cross-sample abundance comparisons when appropriate, with clear communication about the normalisation method.
• Conduct thorough quality control at each step, from read quality to alignment accuracy and counting strategies.
• Document all choices, including software versions, parameter settings, and annotation references, to enable reproducibility.