Appendix A: Single-Cell Biology Refresher

Author

Single Cell RNA-seq Workshop

This appendix is the biology behind every term Lecture 00 starts using. If you’ve taken a recent molecular biology course, skim it — there’s nothing surprising. If your last cell biology course was a while ago or you came from a quantitative background, read it carefully before Lecture 00.

Tip

Goal of this appendix. When Lecture 00 says “Cell Ranger reverse-transcribes the polyA-selected mRNA inside a droplet that also contains the cell’s UMI-tagged barcoded primer”, you should know what every word means. We’ll get there in 12 short pages.

1. The cell, in one paragraph

A cell is a membrane-bound compartment that holds a copy of an organism’s genome (its DNA), plus all the machinery to read that DNA into proteins. The genome is identical across every cell in your body — what differs between cells is which genes are turned on, and the readout of “which genes are on” is the transcriptome: the set of RNA molecules currently present in that cell.

scRNA-seq measures the transcriptome one cell at a time, which is what makes it powerful (you see cell-to-cell heterogeneity) and what makes it hard (each cell yields only a tiny amount of RNA, so the data are noisy and sparse). The first single-cell transcriptome was sequenced in 2009¹; droplet chemistries later scaled the approach to thousands of cells per run^2,3.

2. From DNA to mRNA to counts

2a. The central dogma — the 30-second version

DNA   --transcription-->   mRNA   --translation-->   protein

DNA is the long-term, stable store of genetic information. Same DNA in every one of your cells.
Transcription is the cell’s machinery (the RNA polymerase) reading a stretch of DNA and producing an mRNA copy.
mRNA (messenger RNA) is the short-lived working copy. The cell makes mRNA when it wants to express the gene, then translates it into protein.
Translation is the ribosome reading the mRNA and producing the corresponding protein.

scRNA-seq measures the mRNA layer — the working copy. So when we say “this cell expresses gene X”, we mean “this cell currently contains mRNA copies of gene X”. The amount of mRNA correlates with — but is not identical to — the amount of protein.

2b. Genes have introns; mRNA does not

Genes in the DNA contain exons (coding pieces) interrupted by introns (non-coding pieces). Transcription produces a long RNA that includes both. Splicing removes the introns and joins the exons together, producing the mature mRNA.

Note

Why this matters for scRNA-seq. Cell Ranger’s “include intronic reads” mode (default in v7+) counts reads that map to introns as well as exons. This boosts gene-detection sensitivity but slightly changes the meaning of “expression”. For most analyses, you don’t need to think about this — but it’s why you sometimes see different cell counts depending on the Cell Ranger version.

2c. The polyA tail — what 10x’s chemistry actually captures

After splicing, mRNA gets a polyA tail — a stretch of ~100–200 As tacked on the 3’ end. This tail is what 10x’s reverse-transcription primers grab onto: the primer has a string of Ts complementary to the As, and that’s the first step of converting mRNA into cDNA.

This is why 10x scRNA-seq is sometimes called 3’-end sequencing — the chemistry preferentially captures the 3’ end of each mRNA, near the polyA tail. (10x also makes a 5’ assay, used for TCR/BCR sequencing.)

3. The cell membrane and dissociation

Every cell is wrapped in a plasma membrane — a lipid bilayer. To get scRNA-seq data, you need to:

Dissociate the tissue: break the cell-cell connections so you have a suspension of free, individual cells. Usually mechanical (chopping) plus enzymatic (collagenase, trypsin) treatment.
Encapsulate each cell in a droplet (the 10x part) so each cell’s RNA stays separate from every other cell’s.
Lyse the cell inside the droplet — break the membrane so the RNA spills out and can react with the chemistry.

Warning

Dissociation is the workshop’s biggest hidden source of artifacts. Some cell types (neurons, adipocytes) don’t survive standard dissociation. Some cell-type-specific genes (immediate-early genes like FOS, JUN) get induced by the stress of dissociation itself. A “stress signature” in your dataset is often dissociation, not biology. Watch for it in QC.

4. UMIs and barcodes — the 10x chemistry in detail

When the 10x Chromium controller makes droplets, each droplet contains:

One cell (~63% of droplets in a well-loaded run; the rest are empty or doublets)
One gel bead carrying ~750k copies of the same primer
Reverse-transcription reagents

The primer has four critical parts:

5'-[sequencer adapter]-[16-nt cell barcode]-[12-nt UMI]-[poly-T]-3'

Region	Length	What it does
Sequencer adapter	~30 nt	Lets Illumina pair the read; not biological
Cell barcode (CB)	16 nt	Identifies the droplet this molecule came from (not the cell — each cell got one CB)
UMI (Unique Molecular Identifier)	12 nt	A random tag that uniquely identifies this single mRNA molecule
Poly-T	~30 nt	Hybridizes to the mRNA’s polyA tail

When a cell’s mRNA captures onto a primer, reverse transcriptase produces a cDNA copy that inherits the cell barcode and the UMI. After PCR amplification and sequencing, every read carries:

the cell barcode → tells you which cell it came from
the UMI → tells you whether it’s a true molecule or a PCR duplicate
the cDNA sequence → tells you which gene

4a. Why UMIs

PCR amplifies some molecules more than others — by orders of magnitude. Without UMIs, the read count for a gene reflects mostly PCR luck, not how many original molecules the cell had. With UMIs, you can collapse all reads with the same (cell barcode, gene, UMI) tuple to a single count — that’s the molecule count for that gene in that cell. Doing this collapse correctly (accounting for sequencing errors in the UMI itself) is a small algorithm in its own right⁴.

Tip

This is why scRNA-seq pipelines call them UMI counts, not “read counts”. Reads are sequencing-machine output. UMIs are biological molecules.

5. The cell-by-gene matrix

Cell Ranger turns sequencing reads into a cell × gene matrix, where:

Each row is a gene (~36,000 in human, after Cell Ranger’s annotation)
Each column is a cell (a unique 10x cell barcode that passed the empty-droplet filter)
Each entry is the UMI count — how many distinct mRNA molecules of that gene Cell Ranger saw in that cell

Most entries are zero. A cell has limited mRNA; you sample some fraction of it; most genes weren’t expressed and their UMI count is 0. Typical matrices are 90–95% zeros. This is sparsity.
The matrix is stored as a sparse matrix (dgCMatrix in R, scipy.sparse.csr_matrix in Python) — only non-zero entries take memory.
The non-zero entries are integers (UMI counts), not log-fold-changes or anything else. Normalization comes later.

6. What the standard QC metrics actually measure

The three standard QC metrics — nCount_RNA, nFeature_RNA, percent.mt — each have a biological meaning:

Metric	What it counts	What it tells you
`nCount_RNA`	total UMI count per cell (sum across genes)	How “deep” this cell was sequenced. Higher = more molecules captured. Affected by both biology (transcriptionally active cells) and chemistry (longer cells stuck around).
`nFeature_RNA`	how many genes are detected (UMI ≥ 1) per cell	How “rich” this cell’s transcriptome is. Strongly correlated with nCount_RNA but reaches a saturation.
`percent.mt`	fraction of UMIs that map to mitochondrial-encoded genes	Cell health proxy. Healthy cells: ~5–15% (tissue-dependent). Stressed/dying cells leak cytoplasmic RNA but keep mitochondrial RNA → percent.mt rises.

Why mitochondrial? Your nuclear genome is in the nucleus; your mitochondria (the cell’s energy factories) have their own small genome with ~37 genes. When a cell is dying, its plasma membrane breaks down first; cytoplasmic mRNAs leak out into the suspension; but mitochondrial mRNAs stay locked inside the mitochondrion, which has its own membrane. So a cell that lost its membrane integrity has a relative excess of mitochondrial RNA → high percent.mt. This is why we filter on percent.mt < 5% (or whatever’s tissue-appropriate); systematic work on what distinguishes low-quality cells confirms that mitochondrial fraction plus library complexity carry most of the QC signal⁵.

7. Cell types vs cell states vs lineages

Three concepts that beginning students mix up:

Concept	Definition	Example
Cell type	A stable, ontologically defined identity. Same name across labs / studies / tissues.	“B cell”, “CD14+ monocyte”, “alveolar type II pneumocyte”
Cell state	A transient condition a cell is in. Same cell type can be in many states.	“activated”, “proliferating”, “exhausted”, “dying”, “interferon-stimulated”
Cell lineage / trajectory	A path through cell-state space, often during development.	“naive → effector → memory T cell”

scRNA-seq clusters can recover cell types, cell states, and points along a lineage. Distinguishing them is the cell-type-annotation problem (Lectures 03 and 07).

Note

Why this matters. “I have 12 clusters” is a statement about the graph partitioning. “I have 12 cell types” is a biological claim that requires evidence — markers, comparison to reference atlases, biological intuition. Don’t conflate them.

8. Bulk vs single-cell RNA-seq — what changes

Bulk RNA-seq pools all the mRNA from a tissue, then sequences. The result is one expression profile per sample, averaged over potentially millions of cells of many types.

scRNA-seq sequences the mRNA from each cell separately. The result is one expression profile per cell, with cell-to-cell heterogeneity preserved.

What’s the same:

The basic measurement: count of UMIs / reads per gene
Statistical concerns: normalization, batch effects, multiple testing, replication
Many downstream questions (DE, enrichment, network analysis) — applied at a new resolution

What’s new:

Sparsity — ~90–95% zeros in droplet data. Most analyses use specialized methods designed for this
Cells are not biological replicates — cells from the same tissue from the same donor share genetics, library prep, and dissociation. Statistically, the unit is the donor, not the cell, and ignoring this produces large numbers of false discoveries⁶. (This is why pseudobulk DE in Lecture 06 is the right answer for condition contrasts)
Cell type is a latent variable — you have to recover it from the data
Integration across samples is a first-class problem, not an afterthought

9. The data lifecycle of a single-cell experiment

End to end:

Tissue   --dissociation-->   Cell suspension
    --10x Chromium-->   Droplets (cell + barcoded bead)
    --reverse transcription-->   Barcoded cDNA
    --PCR-->   Amplified library
    --Illumina sequencing-->   FASTQ files
    --Cell Ranger-->   Cell × gene UMI matrix + QC metrics
    --Seurat / Scanpy-->   Filtered, normalized, clustered, annotated object
    --downstream-->   Cell type maps, DE, integration, biological claims

Each arrow can fail. That’s why every lecture has a QC step.

10. Numbers to keep in your head

A grad student starting their first scRNA-seq study should know these orders of magnitude:

Quantity	Typical value	Why it matters
Genes in human genome (protein-coding)	~20,000	The maximum row count of any matrix
Genes annotated in Cell Ranger reference	~36,000	Includes lncRNAs, antisense, etc.
Cells per 10x channel (well)	1,000–10,000	A “10x run” usually means one channel
UMIs per cell, 10x 3’ v3	5,000–30,000	Below 1,000 = something’s wrong; above 50,000 = rare or doublet
Genes detected per cell	500–5,000	Plateaus around the cell-type’s true expressed gene set
Sparsity (% zeros)	90–95%	The reason scRNA-seq has its own toolkit
Doublets per channel, well-loaded	5–10%	Higher = the chip was over-loaded
Mitochondrial fraction in healthy cells	1–15%	Tissue-dependent; tumor cells often higher
Cost per 10x run (Chromium)	~$2–4 K	Two-condition study with 4 donors per group is ~16 channels = ~$50 K of consumables

11. Glossary of biology terms used in the workshop

These come up in every lecture; the main Glossary has the analytical terms.

Term	Plain-English meaning
Transcriptome	The set of RNA molecules in a cell at one moment
mRNA	Messenger RNA — the working copy of a gene that ribosomes translate to protein
Polyadenylation	Adding a polyA tail to the 3’ end of an mRNA
Reverse transcription (RT)	Making cDNA (a complementary DNA copy) from an mRNA
PCR amplification	Making many copies of each cDNA so there’s enough to sequence
UMI (Unique Molecular Identifier)	A random tag that distinguishes original molecules from PCR duplicates
Cell barcode (CB)	A short DNA sequence shared by all reads from one droplet — identifies which “cell” each read came from
Droplet	The tiny aqueous bubble where 10x reverse-transcribes one cell’s mRNA
Gel bead	The 10x consumable that carries ~750k copies of the same barcoded RT primer
Library	The mix of barcoded, amplified cDNA molecules ready for Illumina sequencing
FASTQ	The Illumina output file format — sequence + quality scores
BCL	Illumina’s raw file format — pre-FASTQ; `cellranger mkfastq` converts BCL → FASTQ
Cell Ranger	10x’s pipeline that turns FASTQs into a cell × gene UMI matrix
Filtered matrix	Cell Ranger’s output after empty-droplet filtering — what most analyses start from
Empty droplet	A 10x droplet with no cell, just ambient RNA from the suspension
Doublet	One barcode’s worth of reads from two cells — happens when two cells get into one droplet
Ambient RNA	mRNA released by lysed/dying cells into the suspension; ends up at low level in every droplet (see SoupX)
Dissociation	Breaking tissue into a cell suspension. Mechanical + enzymatic
CMO / hashing	Chemical labels that let you pool multiple samples into one 10x channel and demultiplex bioinformatically
CITE-seq	scRNA-seq + simultaneous antibody-derived-tag (ADT) capture for surface-protein measurement
scATAC-seq	Single-cell chromatin-accessibility sequencing — measures which DNA is reachable rather than which genes are expressed
Visium	10x’s spatial transcriptomics platform — measures gene expression at ~55 µm spots while preserving tissue layout

12. Where to go for more

If this appendix moved too fast, the canonical free references are:

Molecular Biology of the Cell, 7e (Alberts et al.) — chapters 6–7 on transcription, RNA processing, and gene regulation
Single-cell best practices — the same biology in scRNA-seq context; the companion paper⁷ and the original best-practices tutorial⁸ are the field’s standard end-to-end references
10x Genomics support / Chromium — manufacturer’s documentation, very thorough on chemistry³
Scanpy tutorials, Preprocessing — Python-side walkthrough of the same concepts

Once you can read this appendix without looking anything up, you’re ready for Lecture 00.

References

Tang, F. et al. mRNA-seq whole-transcriptome analysis of a single cell. Nature Methods 6, 377–382 (2009).

Macosko, E. Z. et al. Highly parallel genome-wide expression profiling of individual cells using nanoliter droplets. Cell 161, 1202–1214 (2015).

Zheng, G. X. Y. et al. Massively parallel digital transcriptional profiling of single cells. Nature Communications 8, 14049 (2017).

Smith, T., Heger, A. & Sudbery, I. UMI-tools: Modeling sequencing errors in unique molecular identifiers to improve quantification accuracy. Genome Research 27, 491–499 (2017).

Ilicic, T. et al. Classification of low quality cells from single-cell RNA-seq data. Genome Biology 17, 29 (2016).

Squair, J. W. et al. Confronting false discoveries in single-cell differential expression. Nature Communications 12, 5692 (2021).

Heumos, L. et al. Best practices for single-cell analysis across modalities. Nature Reviews Genetics 24, 550–572 (2023).

Luecken, M. D. & Theis, F. J. Current best practices in single-cell RNA-seq analysis: A tutorial. Molecular Systems Biology 15, e8746 (2019).