How Many Reads You Need for RNA-Seq

For bulk RNA-seq of a human sample where the goal is quantifying expression of protein-coding genes, 20 to 30 million reads that map to the transcriptome is enough, and we target 30 to 40 million paired-end reads (2×100 or 2×150 bp) per library to leave room for the fraction that gets discarded. If you want isoform-level quantification, novel transcript assembly, or allele-specific expression, plan on 60 to 100 million paired-end reads. If you are profiling whole blood without globin depletion, add 30 to 50 percent because hemoglobin transcripts will eat that share of your output. A read count on its own means very little, which is the main point of this page: the number that matters is reads assigned to features, and that number depends on library prep, RNA quality, and the tissue’s transcript diversity far more than on the number printed on the sequencing invoice.

What a read is, and what “read count” means in two different senses

A read is one sequence of bases the instrument produced from one fragment of cDNA. In paired-end sequencing, each fragment is read from both ends, producing two mates. Vendors and core facilities are inconsistent about whether “40 million reads” means 40 million fragments or 40 million single-end reads (so 20 million fragments). Ask. The difference is a factor of two in every downstream calculation you make.

“Read count” also means something different one layer up. After alignment or pseudoalignment, the count matrix holds the number of reads assigned to each gene in each sample. That per-gene count is the raw quantity differential expression tools model, typically with a negative binomial distribution that accounts for both Poisson sampling noise and biological variability between samples. The two senses are related by a hard constraint: the total reads in your library are divided among genes roughly in proportion to their abundance, so a gene at one transcript per million (TPM) gets about 30 reads in a 30-million-read library and about 3 in a 3-million-read library. Sequencing deeper makes your estimate of a lowly expressed gene’s abundance less noisy. The gene itself does not become more abundant.

Why depth requirements differ so much between questions

The published guidance ranges from 5 million to 200 million reads because the underlying question changes what counts as sufficient. Differential expression of moderately and highly expressed genes saturates early: in most human tissues, 10 million mapped reads already gives stable estimates for the top several thousand genes, and going from 10 to 30 million mainly recovers the tail. Isoform-level quantification is harder because reads must be assigned among transcripts that share most of their sequence, and the information comes from the minority of fragments spanning distinguishing junctions. Transcript discovery and assembly are hardest of all, since you need enough coverage across an entire novel transcript body to call it.

This is a structural feature of the method rather than a quirk of any one protocol. RNA-seq replaced microarrays partly because it measures over five orders of magnitude of dynamic range instead of the two or three an array manages, and partly because it detects transcripts you did not design probes for.¹ Both advantages are purchased with depth. The counting-based design also means depth requirements scale with transcriptome complexity. Non-coding RNAs are usually present at lower abundance than the coding transcripts, so an experiment that includes them needs more reads to see them at all. That is why multi-omics designs targeting long non-coding RNAs and small RNAs typically run separate libraries at their own depths rather than trying to extract everything from one.²

Consider also what fraction of your library is informative. A poly-A selected library from whole blood is dominated by HBB, HBA1, and HBA2. A ribosomal-depletion library retains intronic and non-polyadenylated RNA, spends more reads outside exons, and therefore needs more total output for the same exonic coverage. Ribosomal RNA is roughly 80 to 85 percent of total cellular RNA by mass, so depletion efficiency is the single largest determinant of how many of your reads land on something you care about.³ A 40-million-read library with 6 percent residual rRNA is a better dataset than a 60-million-read library with 35 percent.

The number we use for a personal longitudinal profile, and why

For repeated whole-blood profiling of one person over time, we run poly-A selected, stranded libraries with globin depletion. They are sequenced as 2×150 bp at 40 million read pairs. The reasoning is as follows. The scientific question is “how does this person’s expression change across time points.” The question is not “which genes differ between two groups.” Technical variability between batches is therefore the dominant error source. Sampling noise within a library does not dominate. At 40 million pairs, Poisson noise on any gene above about 1 TPM is small relative to the batch and biological variation you will see between blood draws taken two months apart. Spending 100 million reads instead buys you precision you cannot use, because the day-to-day variation in your own blood transcriptome swamps it.

Paired-end and 150 bp are less negotiable than depth. Read pairs give you fragment-level evidence for junctions and much better multi-mapping resolution, which matters for gene families and for the immunoglobulin and HLA loci that dominate immune signal in blood. Stranded libraries let you assign reads unambiguously at overlapping and antisense loci, which is worth the small loss in yield.

How to verify you got the depth you paid for

Read counts from the sequencer are the beginning of the accounting. They are not the end of it. Work through the FASTQ files yourself.

Start with raw counts. For a gzipped FASTQ, echo $(zcat sample_R1.fastq.gz | wc -l) / 4 | bc gives the number of reads in R1, which equals the number of fragments in a paired-end run. Confirm R1 and R2 match exactly.

Then quantify. We use salmon quant against a decoy-aware transcriptome index built with the GENCODE transcript FASTA plus the primary-assembly genome as decoys, run with --validateMappings --seqBias --gcBias --libType A. Salmon’s lib_format_counts.json reports inferred library type (check that it matches what you ordered) and meta_info.json reports percent_mapped. For a good poly-A human library, expect 80 to 92 percent. Below 70 percent, something is wrong: adapter contamination, rRNA carryover, genomic DNA, or a species mix-up. If you prefer alignment, STAR --quantMode GeneCounts with --outFilterMultimapNmax 20 gives you a BAM for visual inspection and a per-gene count table in one pass, at the cost of roughly 32 GB of RAM for a human index.

Three QC numbers decide whether your depth is real. First, uniquely mapped percentage from the STAR log, or percent_mapped from Salmon. Second, duplication rate: high duplication at moderate depth signals low input complexity, meaning you sequenced the same few molecules repeatedly and your effective depth is much lower than your read count. Third, the fraction of counts consumed by the top 10 genes, available from picard CollectRnaSeqMetrics or computed directly from the count matrix. In whole blood without globin depletion, that fraction routinely exceeds 50 percent. Divide your nominal depth by (1 minus that fraction) to see what you effectively have for everything else.

Finally, plot a saturation curve before you decide next time’s depth. Subsample your own BAM at 10, 20, 40, 60, and 80 percent with samtools view -s, requantify, and count genes detected above 1 TPM at each level. For a typical whole-blood library the curve flattens around 20 to 25 million mapped reads for gene-level detection. That curve is specific to your tissue, your prep, and your RNA integrity, and it is the only sound basis for choosing depth.

Where RNA integrity matters more than read count

RNA degrades. A sample with an RNA integrity number (RIN) of 5 will show 3’ coverage bias, because poly-A selection captures fragments from the 3’ end of partially degraded transcripts, and no amount of extra sequencing repairs that. Check the 5’-to-3’ coverage uniformity in CollectRnaSeqMetrics (MEDIAN_5PRIME_TO_3PRIME_BIAS). If it is far from 1, your gene-level estimates are biased in a way that correlates with transcript length, and comparisons against a well-preserved sample will produce hundreds of false differences. For blood, this means collection tube and time-to-freeze dominate: PAXgene or Tempus tubes stabilize RNA at draw, and that choice affects your data more than doubling depth would.

Questions people also ask

How many million reads for RNA-seq? For human gene-level expression, 20 to 30 million mapped reads is sufficient, and ordering 30 to 40 million paired-end reads gets you there after losses. For isoform quantification or transcript discovery, 60 to 100 million pairs.

What is read count in RNA-seq? It means either the total number of sequence reads in a library, or the per-gene number of reads assigned to each gene in the count matrix used for differential expression. Context determines which.

How many cells do you need for RNA-seq? Bulk RNA-seq works from as little as 10 nanograms of total RNA with standard kits, which is roughly a few thousand cells, though 100 to 500 ng gives better library complexity. Single-cell RNA-seq is a different design: droplet platforms typically target thousands to tens of thousands of cells per run at 20,000 to 50,000 reads per cell, and population-scale single-cell studies have profiled hundreds of thousands of cells across many donors to separate individual genetic effects from cell-state effects.⁴ Rare populations need enrichment first, since sorting a few thousand cells of a specific type and sequencing them deeply detects transcriptionally distinct subsets that a whole-tissue run would miss.⁵

What is the number of reads in sequencing? It is how many separate fragment sequences the instrument produced for your sample, set by how much of a flow cell your library occupied. It is a budget, not a measure of quality, and the useful figure is how many of those reads map to features you intend to measure.

Does more depth improve differential expression results? Past roughly 20 million mapped reads, adding replicates improves statistical power considerably more than adding reads, because between-sample biological variance rather than within-sample counting noise limits detection. Standard practice for a group comparison is at least three biological replicates per condition, and more if effect sizes are small.⁶

If any expression result appears to bear on your health, take the data and the QC metrics to a clinician. RNA expression from a single blood draw is a snapshot of a highly dynamic system, and interpreting it in a clinical frame requires a physician.

Footnotes

1. Paul A McGettigan. Transcriptomics in the RNA-seq era. Current Opinion in Chemical Biology, 2013. https://doi.org/10.1016/j.cbpa.2012.12.008 ↩

2. Sang-Yoon Shin, Jin Seo Jeong, Jae Yun Lim, et al. Transcriptomic analyses of rice (Oryza sativa) genes and non-coding RNAs under nitrogen starvation using multiple omics technologies. BMC Genomics, 2018. https://doi.org/10.1186/s12864-018-4897-1 ↩

3. Varun Gupta, Jonathan R. Warner. Ribosome-omics of the human ribosome. RNA, 2014. https://doi.org/10.1261/rna.043653.113 ↩

4. Julie Jerber, Daniel D. Seaton, Anna S. E. Cuomo, et al. Population-scale single-cell RNA-seq profiling across dopaminergic neuron differentiation. Nature Genetics, 2021. https://doi.org/10.1038/s41588-021-00801-6 ↩

5. Sofia Papavasileiou, Jiezhen Mo, Daryl Boey, et al. Single‐Cell Omics Analysis of Human Basophils Reveals Two Transcriptionally Distinct Populations. Allergy, 2026. https://doi.org/10.1111/all.70209 ↩

6. Michele Araújo Pereira, Eddie Luidy Imada, Rafael Lucas Muniz Guedes. RNA‐seq: Applications and Best Practices. Applications of RNA-Seq and Omics Strategies - From Microorganisms to Human Health, 2017. https://doi.org/10.5772/intechopen.69250 ↩