How Bulk RNA-Seq Works
Bulk RNA sequencing converts the RNA in a tissue or blood sample into a table of numbers: one row per gene, one column per sample, each cell a count of sequencing reads assigned to that gene. The chemistry extracts RNA from millions of cells at once and selects the fraction you care about, usually messenger RNA. That RNA is copied into complementary DNA and fragmented. Sequencing adapters are then attached and the fragments are read on an Illumina or similar instrument. The output is a set of FASTQ files. Everything after that is software: quality trimming, alignment or pseudoalignment to a reference transcriptome, counting, normalization, and statistics. The result is a per-sample average expression profile across all the cells in the input, which is why it is called bulk.
What happens in the lab, and which choices change your data
RNA extraction comes first, and it sets a ceiling on everything downstream. The standard quality metric is RIN (RNA Integrity Number), a 1–10 score derived from a Bioanalyzer electropherogram based on the ratio and sharpness of the 28S and 18S ribosomal RNA peaks. For standard poly-A libraries you want RIN above 7. Degraded RNA produces a 3’ bias, because the poly-A tail anchors the capture and the 5’ ends of broken transcripts are lost.
Next comes RNA selection. Total RNA is roughly 80–90% ribosomal RNA, which carries almost no information about gene expression, so you have to remove it. Poly-A selection uses oligo-dT beads to pull down transcripts with a poly-adenylated tail. It is clean and efficient, and it is the right default for intact samples. Ribosomal depletion (Ribo-Zero, RiboCop) instead hybridizes probes to rRNA and removes it, keeping non-polyadenylated RNAs: most long non-coding RNAs, histone mRNAs, circular RNAs, and pre-mRNA introns. Ribo-depletion tolerates degraded or FFPE (formalin-fixed paraffin-embedded) input far better, at the cost of needing more reads for the same mRNA depth.
Then reverse transcription, fragmentation, and adapter ligation build the library. Two parameters matter for analysis. Strandedness tells you which genomic strand a read came from, usually via dUTP incorporation in the second strand. Use a stranded protocol. It resolves reads that fall in overlapping antisense genes and it is now standard. Read configuration is the other choice. Single-end reads of 50–75 bp are enough for counting genes. Paired-end reads of 100–150 bp give much better isoform assignment and are required if you care about splicing, fusions, or novel transcripts.
How many reads and how many cells
For differential expression across annotated genes in a human sample, 20–30 million reads per sample is the working standard, and 10 million will often do if you only care about moderately to highly expressed genes. Isoform-level quantification, allele-specific expression, or detection of low-abundance transcripts pushes you to 50–100 million. Adding more biological replicates almost always buys more statistical power than adding reads to the same libraries, because between-sample biological variance dominates the Poisson counting noise once you are past about 10 million reads.
Cell input is far less demanding than people expect. Standard kits ask for 100 ng to 1 µg of total RNA, which corresponds to roughly 10,000 to 100,000 mammalian cells. Low-input protocols such as SMART-Seq work from nanogram or sub-nanogram amounts, down to single cells, using template-switching to amplify full-length cDNA. The practical limit for a well-behaved bulk library is somewhere around 1,000 cells. Below that, PCR amplification cycles rise and duplicate rates go with them.
The analysis pipeline, with specific tools
Start with quality control on the raw FASTQ files. Run fastqc on each file and aggregate with multiqc . to get one HTML report. You are looking for per-base quality dropping off at the 3’ end, adapter content, and overrepresented sequences that are usually rRNA or adapter dimers. Also check the GC distribution for a second peak that suggests contamination. Trim with fastp, which handles adapter detection, quality trimming, and reporting in one pass: fastp -i R1.fq.gz -I R2.fq.gz -o R1.trim.fq.gz -O R2.trim.fq.gz --detect_adapter_for_pe -w 8.
For quantification you have two families of tools, and the choice depends on what you want. Pseudoalignment tools such as Salmon and kallisto map reads to a transcriptome index rather than the genome, and they run in minutes on a laptop with a few gigabytes of memory. We would use Salmon in selective-alignment mode as the default: build an index with a decoy sequence set from the genome, then run salmon quant -i index -l A -1 R1.fq.gz -2 R2.fq.gz --validateMappings --gcBias -o out. The --gcBias flag corrects a real and measurable fragment-level GC bias. The -l A flag lets Salmon infer the library strandedness, and you should check the inferred value against what the protocol claims.
Use a splice-aware genome aligner instead when you need the alignments themselves. STAR is the standard here. It produces a coordinate-sorted BAM you can inspect in IGV, call variants from, or feed to fusion callers. The cost is memory: a human index needs about 32 GB of RAM. Counting from a BAM is usually done with featureCounts from the Subread package or htseq-count. Reviews of bulk RNA-seq methodology lay out these alternative routes and where each is appropriate 1.
Both routes need the same two reference files: a genome FASTA and a matched GTF annotation. Use GENCODE for human, and record the exact release number. Changing annotation versions between samples silently changes gene counts, and it is one of the most common sources of irreproducible results.
Import transcript-level Salmon output into R with tximport, which sums transcripts to genes and carries an offset correcting for average transcript length. Then model the counts. The two standard tools are DESeq2 and edgeR, both of which fit a negative binomial model to raw counts and shrink per-gene dispersion estimates across genes. Comparative work on differential expression methods finds these approaches broadly concordant on well-powered designs, with differences concentrated in low-count genes and small sample sizes 2. Do not feed TPM or FPKM into these tools. Each of them needs integer counts, because the variance model depends on the counting statistics. TPM is for eyeballing expression within a sample. It is not for testing between samples.
What bulk measures, and what it hides
A bulk profile is a cell-type-weighted average. If a tissue sample contains 60% fibroblasts and 5% T cells, a gene expressed only in T cells shows up at roughly a twentieth of its true per-cell level. Worse, a change in measured expression between two samples can come from a change in cell composition rather than any change in transcription within cells. This is the central limitation of bulk relative to single-cell RNA-seq, which profiles individual cells and reports the distribution rather than the mean 34. In heart tissue, for example, cardiomyocytes are a minority of cells but contribute a large share of the RNA. That distorts the bulk signal in a predictable direction 5.
You can partly recover composition computationally. Deconvolution methods such as CIBERSORTx, xCell, MCP-counter, and quanTIseq estimate cell-type fractions from bulk expression using reference signature matrices. Integrated tutorials show these methods applied to characterize immune content from bulk tumor RNA-seq alongside pathway and repertoire analysis 6. Treat the outputs as relative estimates across samples rather than absolute fractions. Deconvolution is also the standard way to combine single-cell references with larger bulk cohorts, since bulk sequencing remains cheaper per sample and better powered for population-scale studies 7.
Bulk still has real advantages. It is far cheaper per sample and has much lower technical noise per gene. It captures the full dynamic range of expression without dropout. It also works on frozen or archived tissue where intact single cells are unrecoverable 8. For repeated measurement of the same person over time, bulk is the practical choice, and time-series designs have their own well-developed statistical methods that model trajectories rather than pairwise contrasts 9.
Failure modes to check before you trust a result
Check the fraction of reads that map. Below 70% for a human sample suggests contamination, the wrong reference, or heavy rRNA carryover. Check duplicate rate, which is high when library complexity is low from too little input RNA. Check the 3’ bias metric from Picard’s CollectRnaSeqMetrics, which flags degradation. Check globin content in whole-blood samples, where HBB and HBA can consume 50% or more of your reads unless you use a globin depletion kit.
Then check for batch effects before anything else. Plot a PCA of variance-stabilized counts and color the points by sequencing run, extraction date, and library prep batch. If your first principal component separates by batch rather than by the biological variable you care about, fix the design or include batch as a covariate in the model. No downstream statistic saves you from a confounded batch.
Finally, an expression measurement is not a clinical result. Gene expression changes with time of day, recent food, and exercise. It also shifts with infection and acute stress. A single bulk profile describes one moment. Interpretation for any health question belongs with a clinician who can see the rest of your picture.
Questions people also ask
How do you run a bulk RNA-seq experiment?
Extract RNA and check integrity, then deplete rRNA by poly-A selection or ribo-depletion. Build a stranded paired-end library and sequence to 20–30 million reads per sample. Then trim with fastp and quantify with Salmon against a GENCODE transcriptome index. Import the results with tximport and test with DESeq2. Design with at least three biological replicates per group, and more if effect sizes are small.
How many cells do you need for bulk RNA-seq?
Standard protocols want 100 ng to 1 µg of total RNA, which is roughly 10,000 to 100,000 mammalian cells. Low-input chemistries such as SMART-Seq work down to hundreds of cells or fewer. Below about 1,000 cells you should expect higher PCR duplication and reduced library complexity.
What is the difference between bulk and single-cell RNA-seq?
Bulk gives you one averaged expression profile per sample across all cells present. Single-cell gives you a profile per cell, so you can see which cell types are present and how expression varies within them 3. Single-cell costs more per sample, has much sparser per-cell counts from dropout, and requires a viable single-cell suspension that many frozen or fibrous tissues will not give you 4.
How many reads do you need for bulk RNA-seq?
Twenty to thirty million reads per sample covers differential expression of annotated genes in human tissue. Fifty to one hundred million is appropriate for isoform-level analysis or detection of rare transcripts. Past about ten million reads, adding biological replicates gains more power than adding depth.
Footnotes
-
Amarinder Singh Thind, Isha Monga, Prasoon Kumar Thakur, et al. Demystifying emerging bulk RNA-Seq applications: the application and utility of bioinformatic methodology. Briefings in Bioinformatics, 2021. https://doi.org/10.1093/bib/bbab259 ↩
-
Maryam Kalantari-Dehaghi, Neda Ghohabi-Esfahani, Modjtaba Emadi-Baygi. From bulk RNA sequencing to spatial transcriptomics: a comparative review of differential gene expression analysis methods. Human Genomics, 2025. https://doi.org/10.1186/s40246-025-00884-w ↩
-
Jorge A. Tzec‐Interián, Daianna González‐Padilla, Elsa B. Góngora‐Castillo. Bioinformatics perspectives on transcriptomics: A comprehensive review of bulk and single‐cell RNA sequencing analyses. Quantitative Biology, 2025. https://doi.org/10.1002/qub2.78 ↩ ↩2
-
Xinmin Li, Cun-Yu Wang. From bulk, single-cell to spatial RNA sequencing. International Journal of Oral Science, 2021. https://doi.org/10.1038/s41368-021-00146-0 ↩ ↩2
-
Jana-Charlotte Hegenbarth, Giuliana Lezzoche, Leon J. De Windt, et al. Perspectives on Bulk-Tissue RNA Sequencing and Single-Cell RNA Sequencing for Cardiac Transcriptomics. Frontiers in Molecular Medicine, 2022. https://doi.org/10.3389/fmmed.2022.839338 ↩
-
Lin Yang, Jin Wang, Jennifer Altreuter, et al. Tutorial: integrative computational analysis of bulk RNA-sequencing data to characterize tumor immunity using RIMA. Nature Protocols, 2023. https://doi.org/10.1038/s41596-023-00841-8 ↩
-
Shengqiang Mao, Yilong Wang, Ningning Chao, et al. Integrated analysis of single-cell RNA-seq and bulk RNA-seq reveals immune suppression subtypes and establishes a novel signature for determining the prognosis in lung adenocarcinoma. Cellular Oncology, 2024. https://doi.org/10.1007/s13402-024-00948-4 ↩
-
Yunjin Li, Lu Ma, Duojiao Wu, et al. Advances in bulk and single-cell multi-omics approaches for systems biology and precision medicine. Briefings in Bioinformatics, 2021. https://doi.org/10.1093/bib/bbab024 ↩
-
Vera-Khlara S. Oh, Robert W. Li. Temporal Dynamic Methods for Bulk RNA-Seq Time Series Data. Genes, 2021. https://doi.org/10.3390/genes12030352 ↩
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.