FPKM vs TPM in RNA-seq: which unit to use and when
Use TPM (transcripts per million) when you want to compare the abundance of two different genes inside one sample. Use raw or estimated counts, with a size-factor method like DESeq2’s median-of-ratios, when you want to compare one gene across samples. FPKM (fragments per kilobase of transcript per million mapped fragments) is a legacy unit from the Cufflinks era that answers the same question TPM answers, but with a denominator that does not sum to a constant across samples. Because of that single property, FPKM is harder to reason about and gives you no benefit in return, so prefer TPM and keep FPKM only for reading older papers and older files.
The math, in three lines
All three units start from the same input: an estimated number of fragments assigned to a transcript or gene, plus an effective length. Effective length is the transcript length minus the mean fragment length, plus one, adjusted for the fragment length distribution. It exists because a 300 nt transcript cannot host many 200 nt fragments, so short features generate fewer fragments than their raw length suggests.
FPKM divides by length first and by depth second:
FPKM_i = count_i / (effLen_i / 1000) / (totalMappedFragments / 1e6)
TPM reverses the order. It computes a per-base rate, then normalizes those rates to sum to one million:
rate_i = count_i / effLen_i
TPM_i = rate_i / sum(rate) * 1e6
RPKM is FPKM for single-end data, where one read equals one fragment. The two differ only in bookkeeping for paired-end reads, which are two reads from one fragment and must be counted once.
The conversion between them is exact and one-directional:
TPM_i = FPKM_i / sum(FPKM) * 1e6
You can always turn FPKM into TPM. You cannot turn TPM back into FPKM without knowing the original FPKM sum, which is almost never reported.
Why TPM is the better within-sample unit
The sum of TPM over all features in a sample is exactly one million, by construction. So a TPM value is a proportion: a transcript at 250 TPM accounts for 250 out of every million transcript molecules the library sampled. That makes the number interpretable on its own and comparable to the same feature in another sample under one assumption, which is that total transcriptional output per cell did not change much.
The sum of FPKM depends on the length distribution of what was expressed in that particular sample. The sum is not fixed. Two samples with identical proportional composition can produce different FPKM sums if one expressed more long transcripts. An FPKM of 12 in sample A and 12 in sample B therefore do not mean the same fraction of the library. The original argument for TPM rests on this, and it still holds.
Where TPM also fails
TPM is a compositional measurement, and compositional data has a hard limit. If one gene goes from 5 percent of the library to 20 percent, every other gene’s TPM falls by roughly 16 percent even though nothing changed in absolute terms. Highly expressed genes dominate the denominator. Globular shifts in transcriptome size break the assumption outright, which is why absolute transcriptome scale matters for normalization and for any downstream deconvolution of cell type proportions 1.
TPM also does not transport across protocols. Comparing TPM from a poly-A selected library to TPM from a ribosomal-depletion library is a category error, because the two libraries sample different populations of RNA molecules and the denominators are not the same universe 2. The same caution applies to comparing your TPM to a public table from a different pipeline. Independent pipelines run on the same FASTQ files produce measurably different expression estimates, with the largest disagreements in low-expression and multi-mapping genes 3. Platform matters too: short-read Illumina and Oxford Nanopore RNA-seq of the same tissue recover overlapping but non-identical gene sets 4.
For differential expression, hand counts to the model and let the model do the normalization. Comparative work on tumor models found that the choice among TPM, FPKM, and normalized counts changes which genes come out as differentially expressed. The same work found that count-based methods behave better 5. Benchmarks in other downstream settings reach a similar conclusion about the sensitivity of results to the normalization choice 6.
What we would run
For a single person’s RNA-seq, we use Salmon with a decoy-aware index built from the transcriptome plus the full genome as decoy sequence. Selective alignment against decoys reduces spurious assignment of genomic and intronic reads to transcripts.
salmon index -t gentrome.fa.gz -d decoys.txt -i salmon_idx --gencode -k 31
salmon quant -i salmon_idx -l A \
-1 sample_R1.fastq.gz -2 sample_R2.fastq.gz \
--validateMappings --gcBias --seqBias --posBias \
--numBootstraps 30 -p 16 -o quant/sample
--gcBias matters more than people expect. GC content bias varies by library prep batch, and for a longitudinal profile the batches are months apart. --numBootstraps 30 gives you inferential replicates, which let you see which transcript-level estimates are unstable because of read assignment ambiguity.
The output quant.sf has five columns. The first three are Name, Length and EffectiveLength. The last two are TPM and NumReads. Keep all five. The pair you need downstream is NumReads and EffectiveLength.
To go from transcripts to genes, use tximport in R with countsFromAbundance = "lengthScaledTPM". This produces gene-level counts that carry the average transcript length correction, so a shift in isoform usage does not masquerade as a change in gene expression.
txi <- tximport(files, type = "salmon", tx2gene = tx2gene,
countsFromAbundance = "lengthScaledTPM")
dds <- DESeqDataSetFromTximport(txi, colData, ~ condition)
Keep txi$abundance as your TPM matrix for browsing and plotting. Keep txi$counts for any statistical test. Do not feed TPM into DESeq2 or edgeR.
Failure modes to check before you trust a number
Filter before you interpret. A gene at 0.3 TPM is within the noise floor of transcript assignment and you should not read anything into a twofold change there. A practical threshold is to keep genes with at least 10 counts in enough samples to matter, or with TPM above 0.5 to 1 across a majority of timepoints. Standard eQTL pipelines apply exactly this kind of expression and count filter before any modeling 7.
Watch the top of the distribution. Compute the fraction of TPM captured by your top 10 and top 100 genes in each sample. If hemoglobin or mitochondrial transcripts take 40 percent of the library in one draw and 12 percent in another, your TPM values are not comparable and you need to either globin-deplete or normalize with count-based size factors.
Freeze the pipeline. For a longitudinal profile the single most valuable thing you can do is run every timepoint through the same Salmon version, the same index, and the same annotation release. Switching from GENCODE v39 to v45 mid-series changes transcript definitions and produces apparent expression changes that are annotation artifacts 3.
Finally, RNA expression values are measurements. They are not clinical findings. If something in your data looks concerning, take it to a physician who can order validated diagnostic testing.
Questions people also ask
Is TPM better than FPKM? For within-sample comparison, yes, because TPM sums to a fixed one million and FPKM does not. Across samples both are compositional and both can mislead when library composition shifts 2.
Can I convert FPKM to TPM? Yes, exactly, by dividing each FPKM by the sum of all FPKM in that sample and multiplying by one million. You cannot do the reverse without the original sum.
Should I use TPM for differential expression? No. Use estimated counts with DESeq2 or edgeR, which model count noise and estimate size factors internally. The choice of measure changes which genes are called differential 5.
What about single-cell data? Length normalization is unnecessary for 3-prime tagged protocols like 10x, because each molecule yields one UMI regardless of transcript length. Single-cell pipelines use counts per cell with their own scaling 8.
Why do two pipelines give different TPM for the same gene? Aligner, annotation, and multi-mapping policy all differ. Disagreement is largest for lowly expressed genes and gene families with high sequence similarity 3.
Footnotes
-
Songjian Lu, Jiyuan Yang, Lei Yan, et al. Transcriptome size matters for single-cell RNA-seq normalization and bulk deconvolution. Nature Communications, 2025. https://doi.org/10.1038/s41467-025-56623-1 ↩
-
Shanrong Zhao, Zhan Ye, Robert Stanton. Misuse of RPKM or TPM normalization when comparing across samples and sequencing protocols. RNA, 2020. https://doi.org/10.1261/rna.074922.120 ↩ ↩2
-
Sonali Arora, Siobhan S. Pattwell, Eric C. Holland, et al. Variability in estimated gene expression among commonly used RNA-seq pipelines. Scientific Reports, 2020. https://doi.org/10.1038/s41598-020-59516-z ↩ ↩2 ↩3
-
Ekaterina V. Ilgisonis, Elena A. Ponomarenko, Svetlana N. Tarbeeva, et al. Gene-centric coverage of the human liver transcriptome: QPCR, Illumina, and Oxford Nanopore RNA-Seq. Frontiers in Molecular Biosciences, 2022. https://doi.org/10.3389/fmolb.2022.944639 ↩
-
Yingdong Zhao, Ming-Chung Li, Mariam M. Konaté, et al. TPM, FPKM, or Normalized Counts? A Comparative Study of Quantification Measures for the Analysis of RNA-seq Data from the NCI Patient-Derived Models Repository. Journal of Translational Medicine, 2021. https://doi.org/10.1186/s12967-021-02936-w ↩ ↩2
-
Hatice Büşra Lüleci, Dilara Uzuner, Müberra Fatma Cesur, et al. A benchmark of RNA-seq data normalization methods for transcriptome mapping on human genome-scale metabolic networks. npj Systems Biology and Applications, 2024. https://doi.org/10.1038/s41540-024-00448-z ↩
-
Tao Wang, Yongzhuang Liu, Junpeng Ruan, et al. A pipeline for RNA-seq based eQTL analysis with automated quality control procedures. BMC Bioinformatics, 2021. https://doi.org/10.1186/s12859-021-04307-0 ↩
-
Byungjin Hwang, Ji Hyun Lee, Duhee Bang. Single-cell RNA sequencing technologies and bioinformatics pipelines. Experimental & Molecular Medicine, 2018. https://doi.org/10.1038/s12276-018-0071-8 ↩
Pothos builds longitudinal molecular profiles of individuals: whole-genome sequencing, single-cell RNA sequencing, and plasma proteomics, delivered as a detailed report, every data file, and stored cells. Build your profile.