Read Depth in RNA-Seq: What It Means and How Much You Need

Read depth in RNA-seq is the number of sequencing reads produced for a sample, usually quoted in millions of reads, and more usefully the number of reads that end up assigned to each gene or transcript. Unlike whole-genome sequencing, where depth is a single average coverage figure across a mostly uniform target, RNA-seq depth is distributed across a transcriptome whose abundances span five or six orders of magnitude. A few highly expressed genes absorb a large fraction of your library, so the same 30 million reads can give you 4,000-fold coverage of a mitochondrial transcript and zero reads for a transcription factor expressed at one copy per cell. The practical question is never “what is my depth” but “at what expression level does my depth stop supporting a conclusion”.

For a personal transcriptome from whole blood or a tissue biopsy, where the goal is to quantify most protein-coding genes rather than to discover novel isoforms, we would specify 30 to 50 million paired-end reads at 100 bp or 150 bp per sample, with ribosomal RNA depletion rather than poly(A) selection if you care about non-polyadenylated RNA. That range sits well past the point where adding reads changes which genes are detected, and it leaves headroom for the fraction of the library lost to rRNA, globin, and duplicates.

How depth is defined and calculated

The cleanest definition of depth for a sequencing experiment is the Lander-Waterman expectation: depth equals the number of reads times read length divided by the size of the target ¹. For a 3.2 Gb human genome, 30 billion bases of sequence gives 30x average coverage, meaning each base is covered by 30 reads on average. That calculation is well behaved for DNA because the target is fixed and roughly uniformly sampled.

Applying it to RNA requires a different denominator. The “target” is the set of transcripts present in the sample weighted by their abundance, which you do not know before you sequence. So RNA-seq depth is reported as total reads rather than fold coverage, and per-gene depth is reported as counts. If you want a fold-coverage number for a specific transcript, compute it directly: reads assigned to the transcript times read length divided by transcript length. A gene at 1,000 counts with 100 bp reads and a 2 kb transcript is covered about 50x.

The numbers you want from your own data are these. Library size is the total reads in the FASTQ file, which you can get with echo $(zcat sample_R1.fastq.gz | wc -l) / 4 | bc. Assigned reads is what survives alignment and counting, reported by STAR in Log.final.out as uniquely mapped reads, or by featureCounts in the .summary file as the Assigned line. Effective depth is that assigned number, and it is typically 60 to 85 percent of the raw library after you subtract multi-mappers, rRNA, unmapped reads, and reads falling outside annotated features. A library advertised as 40 million reads often delivers 28 million usable counts.

Per-gene depth comes from the count matrix. With a quantifier like salmon quant -l A -1 R1.fastq.gz -2 R2.fastq.gz --validateMappings --gcBias -o out, the quant.sf file gives you NumReads and TPM per transcript. A quick look at the count distribution tells you more than the headline number: in a typical 30 million read blood library, roughly 12,000 to 14,000 genes have 10 or more counts, another 5,000 sit between 1 and 9, and the rest are zero.

What depth buys you, and where the curve flattens

Gene detection saturates early. Somewhere between 5 and 10 million assigned reads you have already detected nearly every gene you are going to detect in a bulk sample, because additional reads pile onto transcripts that are already well covered. What continues to improve with depth is the precision of the estimate for low-abundance genes, and therefore the power to call differences between samples.

This has been measured directly. In a controlled comparison of depth and library preparation using rat liver samples, differential expression calls were largely stable down to modest depths, and library preparation protocol had a larger effect on which genes were called than depth did over a wide range ². The lesson generalizes: once you are past roughly 10 million assigned reads, protocol choices such as poly(A) versus rRNA depletion, strandedness, and input RNA quality dominate the variance more than another 10 million reads would.

Depth requirements rise sharply for three specific goals. Isoform-level quantification needs more reads than gene-level, because reads must be distributed among isoforms that share most of their exons, and the effective depth per isoform is a fraction of the gene’s total. Detecting allele-specific expression needs reads overlapping heterozygous sites, which for a typical gene means a few hundred counts before the binomial test has any power. Fusion and novel splice junction discovery depends on junction-spanning reads, a small subset of the library, and this is where short reads struggle most. Long-read platforms sequence full-length transcripts and resolve isoform structure directly at depths of a few million reads, at the cost of higher per-read error and lower throughput ³.

Coverage uniformity matters alongside raw depth. GC-rich and GC-poor transcripts are sampled unevenly by PCR amplification, and degraded RNA produces 3’ bias that inflates apparent coverage at transcript ends. These are the same biases that affect genome sequencing, where uneven coverage means a nominal average depth understates how many regions are effectively under-sampled ¹. Check this with picard CollectRnaSeqMetrics, which reports MEDIAN_5PRIME_TO_3PRIME_BIAS. Values far from 1.0 mean your depth is not where the annotation says it is.

Specifying depth for a personal transcriptome

Here is what we would ask a provider for, and why. Thirty to fifty million paired-end 150 bp reads per sample gives comfortable gene-level quantification with room for isoform estimates on moderately expressed genes. Paired-end is worth the cost because mate pairs improve transcript assignment and make junction detection more reliable. Strand-specific libraries (dUTP protocol) resolve overlapping and antisense transcripts, which matters when a gene of interest sits inside another gene’s intron.

For whole blood specifically, globin mRNA can consume 50 to 70 percent of a poly(A) library, so globin depletion or rRNA depletion is not optional if you want that depth to reach the rest of the transcriptome. Ask the provider to report post-depletion assigned read counts, not raw yield.

The value of a personal transcriptome comes from reading it alongside your genome, proteins, and biomarkers rather than alone, and the depth you choose should serve that integration. Multi-omics studies consistently find that variation which is ambiguous in one layer resolves when a second layer constrains it, and study design for such integration has to account for the differing sensitivity and noise of each assay ⁴. A shallow transcriptome pairs poorly with a deep genome, because you will keep finding variants whose expression consequence you cannot measure. In cancer work, where expression evidence filters candidate variants down to those transcribed, the same logic drives depth choices ⁵.

Longitudinal sampling changes the calculus again. Two samples at 30 million reads each, taken six months apart, tell you more about your own biology than one sample at 60 million, because the within-person comparison removes most of the between-person variability that makes reference ranges wide.

Questions people also ask

What does sequencing depth mean?

Sequencing depth is how many times a given base or molecule is read. For DNA it is expressed as fold coverage of the genome. For RNA it is expressed as total reads per sample, because transcript abundance varies so widely that a single fold-coverage figure would be meaningless ¹.

How do you calculate depth of sequencing?

Multiply the number of reads by the read length and divide by the target size. For RNA-seq, use the assigned read count from your aligner or quantifier rather than the raw FASTQ count, and compute per-transcript coverage as assigned reads times read length divided by transcript length.

What does 30x sequencing depth mean?

It means each base of the target is covered by 30 reads on average, the standard specification for whole-genome sequencing where it supports reliable germline variant calling. The term does not transfer to RNA-seq. If a provider quotes “30x RNA-seq”, ask for the read count per sample instead.

How many reads do I need for RNA-seq?

Gene detection saturates near 5 to 10 million assigned reads, differential expression at gene level is stable from roughly 10 to 20 million, and isoform-level or allele-specific analysis benefits from 50 million or more ². Library preparation quality matters as much as depth across most of that range.

Should I use long reads instead?

Long-read RNA sequencing gives full-length transcript structure and resolves isoforms that short reads can only infer, which is a genuine advantage if isoform usage is your question. Throughput is lower and per-base error higher, so for accurate gene-level counts a short-read library remains the better default ³.

Anything you find in your own transcriptome that looks clinically meaningful needs a clinician to interpret in the context of your history and a validated test. RNA-seq as described here is a measurement, and it is not a diagnostic.

Oak builds longitudinal molecular profiles of individuals: whole-genome sequencing, RNA sequencing, proteomics, blood biomarkers, and continuous glucose data, integrated into one model of you. Build your profile.

Footnotes

1. David Sims, Ian Sudbery, Nicholas E. Ilott, et al. Sequencing depth and coverage: key considerations in genomic analyses. Nature Reviews Genetics, 2014. https://doi.org/10.1038/nrg3642 ↩ ↩² ↩³

2. Dongying Li, Binsheng Gong, Joshua Xu, et al. Impact of Sequencing Depth and Library Preparation on Toxicological Interpretation of RNA-Seq Data in a “Three-Sample” Scenario. Chemical Research in Toxicology, 2020. https://doi.org/10.1021/acs.chemrestox.0c00368 ↩ ↩²

3. Dóra Tombácz, Gábor Torma, Zsolt Boldogkői. Long-read transcriptomics - opportunities and challenges. npj Genomic Medicine, 2026. https://doi.org/10.1038/s41525-026-00617-5 ↩ ↩²

4. Stefan Graw, Kevin Chappell, Charity L Washam, et al. Multi-omics data integration considerations and study design for biological systems and disease. Molecular Omics, 2020. https://doi.org/10.1039/d0mo00041h ↩

5. Panagiotis Mantas, Karen A. Krogfelt. Repurposing public sarcoma multi-omics for neoantigen discovery. Cancer Immunology, Immunotherapy, 2026. https://doi.org/10.1007/s00262-026-04395-y ↩