Finding cell types with single-cell RNA sequencing

Every cell in the brain carries the same genes, but different types of cell switch on different genes and so contain different messenger RNA (mRNA) molecules. Single-cell RNA sequencing counts the mRNA molecules in thousands of individual cells, so that cell types can be recognised from the genes they express. Applied to the mouse cortex, the method has classified cells into many molecularly distinct types, each with characteristic marker genes (Zeisel et al., 2015; Tasic et al., 2016).

The laboratory part separates the tissue into single living cells, then captures each cell in a tiny droplet together with a bead that labels all of its mRNA with the same cell barcode and gives every molecule its own unique molecular identifier (UMI). After sequencing, software turns the reads into a table of molecule counts, one row per cell and one column per gene. The cells are still alive while the tissue is digested, and they can respond to the procedure: immediate early genes such as Fos and Jun, which nerve cells switch on rapidly in response to stimulation (Morgan and Curran, 1991), can then appear in the data as an artefact.

The analysis follows an established sequence: quality control to remove damaged cells, empty droplets and doublets; normalisation; selection of informative genes; principal component analysis; a neighbour graph, clustering and a UMAP picture; annotation of clusters with marker genes; and finally the comparison of conditions (Luecken and Theis, 2019). The last step holds the most common mistake in the field: treating thousands of cells as if they were thousands of independent animals.

The simulator records the true identity of every barcode, so what each decision kept, removed, merged or invented is shown. The simulation is a simplified model with 1,200 genes; in particular, the size of the stress response to dissociation, the variation in gene expression between mice and the loss of neurons with harsh dissociation are illustrative values, not measurements.

Learning objectives

• Explain what a cell barcode and a UMI are, and why droplet capture trades the number of cells against the doublet rate
• Describe how dissociation affects viability, cell-type representation and stress genes, and calculate a cell concentration and viability from a haemocytometer count
• Set quality-control thresholds on genes, UMIs and mitochondrial share from their distributions, and recognise damaged cells, empty barcodes and doublets
• Explain normalisation, highly variable genes, PCA and the elbow plot, the neighbour graph, clustering resolution and why UMAP distances are not quantitative
• Annotate clusters with marker genes such as Snap25, Slc17a7, Gad1, Aqp4, Mog and Cx3cr1
• Compare conditions with animals as replicates, and explain why cell-level tests are pseudoreplication and pseudobulk tests are not

Ethics and safety

Killing mice to collect tissue is a regulated procedure that requires approval by an Institutional Animal Ethics Committee and trained, authorised staff; in India such work is governed by the CCSEA under the Prevention of Cruelty to Animals Act 1960, with equivalent frameworks in the European Union (Directive 2010/63/EU), the United States (Animal Welfare Act and PHS Policy) and the United Kingdom (Animals (Scientific Procedures) Act 1986). The 3Rs apply: Replacement where a question can be answered without animals, Reduction by planning the number of animals from the expected variation between animals (the animal, not the cell, is the experimental unit), and Refinement of handling and euthanasia. Studies are reported following the ARRIVE guidelines, developed by the NC3Rs in the United Kingdom and endorsed by journals internationally (Percie du Sert et al., 2020), and raw reads and count matrices are deposited in a public archive with the sample metadata. Enzymes, trypan blue and fixatives are handled as the local risk assessment requires.

Key terms

• Single-cell RNA sequencing: Sequencing the mRNA of many individual cells, so that each cell gets its own gene-expression profile.
• Dissociation: Separating tissue into single cells with enzymes (proteases) and gentle pipetting (trituration).
• Immediate early genes: Genes such as Fos, Jun and Egr1 that cells switch on within minutes of stress; dissociation at 37 °C can induce them artificially.
• Viability: The percentage of cells that are alive, measured here with trypan blue, which enters only dead cells.
• Cell barcode: A DNA sequence shared by all molecules captured in one droplet, so that they can be assigned to one cell after pooling.
• UMI: Unique molecular identifier: a random DNA tag on each captured mRNA molecule, so that PCR copies are counted only once.
• Doublet: A droplet that captured two cells; its barcode looks like one cell with the genes of both.
• Counts matrix: The table of molecule (UMI) counts with one row per barcode and one column per gene.
• Mitochondrial share: The percentage of a cell's UMIs from mitochondrial genes; high values mark damaged cells.
• Normalisation: Scaling each cell's counts to the same total (here 10,000) and taking log(1 + x), so that cells sequenced to different depths can be compared.
• Highly variable genes: Genes that vary more between cells than expected for their average expression; they carry the differences between cell types.
• PCA: Principal component analysis: combines many genes into a few axes (principal components) that capture most of the variation.
• Neighbour graph and clustering: Each cell is linked to its most similar cells; a clustering algorithm (here Louvain) splits the graph into groups, more finely at higher resolution.
• UMAP: A two-dimensional picture that keeps cells that are neighbours close together; distances between islands are not quantitative.
• Pseudobulk: Adding up the counts of all cells of one type within each animal, so that animals are compared as the replicates.