How RNA-seq Works, From Blood Tube to Count Matrix
RNA sequencing measures how many copies of each RNA molecule were present in a sample at the moment it was collected. The wet lab converts RNA into a sequenceable DNA library, the sequencer reads short fragments of that library, and software assigns each fragment back to the transcript it came from. The output you work with is a matrix of counts: rows are genes or transcripts, columns are samples, and each cell is an estimate of abundance. Your genome tells you what you could express. RNA-seq tells you what was being expressed in one tissue at one time.
That last distinction is the reason RNA-seq is interesting and also the reason it is hard. A genome is static, so one good sequencing run is enough forever. A transcriptome changes with time of day, food, and sleep. It also changes with infection, exercise, and the four hours the tube spent on a bench. Every technical choice below exists to control one of those sources of variation.
What happens between the blood draw and the FASTQ file
The first step is stabilization, and it matters more than anything downstream. Whole blood RNA is dominated by globin transcripts from reticulocytes, and neutrophils begin changing their expression program within minutes of collection. Collection into a stabilizing tube that lyses cells and inactivates RNases immediately is the standard approach for whole blood. If the plan is peripheral blood mononuclear cells instead, the cells have to be separated first, which introduces hours of handling and its own expression artifacts.
Next is RNA extraction and quality assessment. The number to look at is RIN, the RNA integrity number, computed from the electrophoretic trace of the 18S and 28S ribosomal peaks on a 1 to 10 scale. We treat RIN above 8 as good, 6 to 8 as usable with caution, and below 6 as a sample whose coverage will be biased toward transcript 3’ ends. Degradation is not random, so it shows up as a systematic shift in your data rather than as noise.
Then comes selection, because about 80 to 90 percent of total RNA is ribosomal and you do not want to spend reads on it. Two options exist. Poly-A selection uses oligo-dT beads to capture transcripts with a polyadenylated tail, which gives you clean mRNA and the highest fraction of informative reads. Ribosomal depletion uses probes against rRNA and keeps everything else. That includes non-polyadenylated species such as many long non-coding RNAs, circular RNAs, and pre-mRNA. For human blood profiling we prefer poly-A selection plus globin depletion when the question is protein-coding gene expression, and rRNA depletion when the question includes regulatory RNA. Small RNAs require a separate library entirely, with size selection in the 15 to 50 nucleotide range, and dedicated pipelines exist for classes such as tRNA-derived fragments 1.
Library preparation fragments the RNA, reverse transcribes it into cDNA, and ligates sequencing adapters. One choice here has real consequences: strandedness. A stranded protocol marks which DNA strand the original RNA came from, usually by incorporating dUTP during second-strand synthesis. This lets you resolve overlapping genes on opposite strands and correctly assign antisense transcripts. Unstranded libraries lose that information permanently. Always ask for stranded, and always confirm the strandedness empirically before quantifying, because getting the flag backwards silently destroys a large fraction of your counts.
Finally, sequencing. The library is loaded onto an Illumina flow cell and read from both ends, typically 2 x 150 base pairs. Paired-end reads give you better splice junction detection and better isoform assignment than single-end reads of the same length. For standard differential expression on human samples, 25 to 30 million paired reads per sample is a reasonable target. For isoform-level or novel transcript work, 50 to 100 million is closer to right, since detecting a new isoform requires reads spanning its specific junctions 2.
The files you get and what is in them
Raw output arrives as gzipped FASTQ, two files per sample for paired-end data, named something like sample_R1.fastq.gz and sample_R2.fastq.gz. Each read occupies four lines: identifier, sequence, separator, and per-base Phred quality scores in ASCII-33 encoding. A 30 million read pair sample is roughly 5 to 8 GB compressed.
Processed output usually includes alignments in BAM format with a .bai index, plus quantification tables. If the pipeline used a lightweight quantifier you will see Salmon or kallisto output directories containing quant.sf or abundance.tsv, with per-transcript estimated counts and TPM. If it used an aligner you will see a gene-level count matrix from featureCounts or STAR --quantMode GeneCounts. Keep the FASTQs. Every quantification decision downstream depends on annotation and parameters you may want to change.
How we would analyze it
Start with quality control before anything else. Run fastqc on each FASTQ and aggregate with multiqc ., then read four things: per-base quality decay at read ends, adapter content, duplication level, and GC distribution. A bimodal GC curve usually means contamination or rRNA carryover.
For quantification we recommend Salmon with selective alignment against a decoy-aware index. Build the index from a transcriptome FASTA (GENCODE human, a recent release) concatenated with the full genome as decoy sequence, which prevents reads from intronic and intergenic DNA from being falsely assigned to transcripts.
salmon index -t gentrome.fa.gz -d decoys.txt -i salmon_idx -k 31
salmon quant -i salmon_idx -l A \
-1 sample_R1.fastq.gz -2 sample_R2.fastq.gz \
--validateMappings --gcBias --seqBias \
-p 8 -o quant/sample
The -l A flag asks Salmon to infer library type, and you should check the inferred value in lib_format_counts.json against what the protocol claimed. Expect a mapping rate of 70 to 90 percent for a good poly-A human library. Anything below 60 percent means contamination, wrong index, or degradation.
Then import into R with tximport, which sums transcript estimates to gene level while correcting for average transcript length. From there, the standard tool for comparing conditions is DESeq2 or edgeR, both of which model counts with a negative binomial distribution and shrink dispersion estimates across genes 3. Never feed TPM or FPKM into these models. They expect raw counts and handle normalization internally.
One correction that gets skipped: TPM is comparable within a sample, and only partly comparable across samples. If one condition has a few hugely dominant transcripts, every other gene’s TPM falls even at constant absolute abundance. Median-of-ratios normalization in DESeq2 addresses this. Raw TPM comparisons do not.
Reading a single person’s transcriptome
The hard part for an individual profile is that classical differential expression compares groups. With one person you have no group, so you need a different design. The one that works is repeated sampling of yourself under controlled conditions, which turns your own earlier samples into the comparison set. Time course designs have their own statistical methods, and the analysis choices differ meaningfully from two-condition comparisons 4.
Three sampling rules make the data interpretable. Fix the time of day, because circadian genes swing several fold. Fix the fasting state. Record anything that perturbs immune expression, since a mild viral infection will dominate a blood transcriptome for days.
RNA-seq also gains substantially from being read alongside other layers. Transcript abundance predicts protein abundance only loosely, so pairing expression with proteomics and metabolite measurements resolves questions neither layer answers alone 5. This integration is standard practice in immunology, where transcriptomic and cytometric data are routinely combined to interpret variation between individuals 6. Whether such profiles change clinical decisions is still an open research question, and interpretation of anything health-related belongs with a clinician 7.
Questions people also ask
What does RNA-seq data tell you?
It tells you the relative abundance of each transcript in the sampled tissue at the moment of collection, which genes were on or off, and which isoforms were being produced. It also reveals splicing patterns, allele-specific expression when combined with your genome, and expression of non-coding RNAs. It does not tell you protein levels or enzyme activity.
Where can I get RNA-seq data?
Public repositories are the fastest route: GEO and SRA at NCBI, ArrayExpress and ENA in Europe, plus GTEx for tissue-level human expression and TCGA for tumors. For your own data, you need a sample collected into a stabilizing tube and a lab that will return FASTQs rather than a summary report.
How do I analyze RNA-seq data?
Quality control uses FastQC and MultiQC, and quantification uses Salmon against a decoy-aware GENCODE index. Gene-level aggregation uses tximport, and modeling uses DESeq2. If you need splice junctions or novel transcripts, add a spliced aligner such as STAR and a reference-guided assembler 2.
How is RNA-seq different from a genetic test?
A genetic test reads your DNA, which is essentially fixed. RNA-seq reads the transcriptional output of that DNA, which varies by tissue and by hour, and which reflects regulatory state including epigenetic effects 8. The two answer different questions and are most useful together 9.
Footnotes
-
Jun-Hao Wang, Wen-Xin Chen, Shi-Qiang Mei, et al. tsRFun: a comprehensive platform for decoding human tsRNA expression, functions and prognostic value by high-throughput small RNA-Seq and CLIP-Seq data. Nucleic Acids Research, 2021. https://doi.org/10.1093/nar/gkab1023 ↩
-
Adam Roberts, Harold Pimentel, Cole Trapnell, et al. Identification of novel transcripts in annotated genomes using RNA-Seq. Bioinformatics, 2011. https://doi.org/10.1093/bioinformatics/btr355 ↩ ↩2
-
Sana Javed. Differential Expression Analysis of RNA-Seq Data and Co-expression Networks. Computational Biology, 2021. https://doi.org/10.1007/978-3-030-69951-2_2 ↩
-
Daniel Spies, Constance Ciaudo. Dynamics in Transcriptomics: Advancements in RNA-seq Time Course and Downstream Analysis. Computational and Structural Biotechnology Journal, 2015. https://doi.org/10.1016/j.csbj.2015.08.004 ↩
-
Georgia Charkoftaki, David C. Thompson, Jaya Prakash Golla, et al. Integrated multi-omics approach reveals a role of ALDH1A1 in lipid metabolism in human colon cancer cells. Chemico-Biological Interactions, 2019. https://doi.org/10.1016/j.cbi.2019.02.030 ↩
-
Xiaojing Chu, Bowen Zhang, Valerie A. C. M. Koeken, et al. Multi-Omics Approaches in Immunological Research. Frontiers in Immunology, 2021. https://doi.org/10.3389/fimmu.2021.668045 ↩
-
Edward Lau, Joseph C. Wu. Omics, Big Data, and Precision Medicine in Cardiovascular Sciences. Circulation Research, 2018. https://doi.org/10.1161/circresaha.118.313161 ↩
-
Jeffrey S. Buguliskis. The Epigenetic Insights of RNA-Seq. Clinical OMICs, 2016. https://doi.org/10.1089/clinomi.03.05.06 ↩
-
Audrey Minden. Introduction to Omics Technologies for Research, Diagnosis, and Drug Discovery. Biochemistry, Cell and Molecular Biology for Pharmacists and Health Professionals, 2026. https://doi.org/10.1007/978-3-032-20529-2_27 ↩
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.