Single-cell analysis: normalisation, embedding and naming cell states
Single-cell RNA analysis uses normalisation, feature selection, dimensional reduction and clustering to explore candidate cell groups. Check how analysis choices affect those groups, interpret embedding limits and record the evidence for provisional cell labels.
What a learner can do afterwards
- Calculate total-count scaling under stated assumptions and choose a variable gene from supplied values
- Explain what a two-dimensional embedding preserves and what it distorts
- Assign an identity to a cluster from marker genes and state the evidence a label needs
1 · Read
In a droplet UMI assay, a barcode identifies a droplet. Quality checks assess whether its profile represents one usable cell. Counts estimate detected RNA molecules after grouping duplicate reads. A zero means no molecule of that gene was counted: RNA may be absent or may have escaped detection. A positive count can include ambient RNA or material from a doublet. Neither entry alone establishes cell identity.
Cell totals reflect capture and sampling as well as biological RNA content. Total-count scaling divides each gene count by its cell total and multiplies by a shared target. It can help when totals differ mainly for technical reasons and expression proportions are broadly comparable. It cannot recover absolute RNA content or replace checks for doublets, ambient RNA and batch effects, which are differences tied to sample processing.
Assume two cells have the same RNA composition and differ only in capture. A gene has 15 of 150 counts in one cell and 30 of 300 in the other. Scale to 1,000: 15 / 150 × 1,000 = 100 and 30 / 300 × 1,000 = 100. Both assign one tenth of their counts to this gene. Matching scaled values do not prove equal amounts of RNA in real cells.
For this gene, scaling gives 100 in each cell under the stated capture-only assumption.
For grouping, select genes with variation beyond expected technical noise, accounting for mean expression. A constant gene cannot separate cells by its values. Equal means can hide different variation, so high mean expression alone is not a selection rule. Marker testing is a later comparison between a cluster and other cells. Its eligible genes need not be restricted to the features used for grouping.
In this workflow, PCA summarises processed expression values for selected genes using linear combinations called components. Each cell receives component scores. A neighbour graph links cells with similar scores under a chosen distance and neighbour setting. Graph clustering partitions those connections. It need not use the two coordinates displayed by UMAP.
UMAP aims to retain local neighbourhood structure in a low-dimensional view, but can distort local relationships too. Its axes, island gaps and island sizes are not calibrated biological measures. A large gap alone does not establish different cell types. Check the analysis representation and marker evidence before interpreting a separation.
Higher clustering resolution often makes finer partitions. Compare their markers, sample composition and membership across reasonable settings. A split may reflect a biological state or a technical effect. A broad identity can remain supported after splitting. Continuous variation need not form discrete types. Stability is supporting evidence, not proof of biology.
Use several enriched markers and a tissue- and species-matched reference for a provisional label. If a reference links M1 and M2 to type R, their joint enrichment supports an R-like label. Repetition across samples and independent assays strengthen it; staining is one possible check. Pathway enrichment can describe marker functions using genes eligible for marker selection as background. It cannot establish identity alone.
Check what counts measure, justify scaling and feature choices, interpret groups beyond their UMAP layout, and record evidence for provisional labels.
Lesson, questions, diagrams, worked solutions and flashcards written for this dataset.
Four readings and the flashcards were written for this dataset.
More to read
Compare proportions without claiming absolute amounts
A UMI count matrix records detected RNA molecules assigned to genes and retained droplet barcodes. It does not list every molecule originally present in a cell. Capture and sequencing can miss RNA, while ambient material or a doublet can add RNA that does not belong to one intended cell.
Total-count scaling expresses each gene relative to its cell's observed total, multiplied by a shared target. Two cells can therefore have matching scaled values even when their raw totals differ. This is useful when the totals differ mainly because of capture and sampling, provided the expression proportions are broadly comparable.
The same operation can hide a biological difference in total RNA content. Scaling each cell to the same target removes information about that total from the scaled representation. It does not establish how much RNA either cell originally contained.
Keep the raw counts and record the normalisation method and its assumptions. Assess doublets and ambient RNA separately. Also check whether processing batches explain patterns that might otherwise receive biological labels.
Separate grouping features from marker evidence
A gene with a high average count need not distinguish cells. If its processed values are nearly constant, those values offer little information about differences among cells. Variable-feature selection instead considers variation in relation to mean expression and expected technical noise.
In one common workflow, PCA summarises processed expression values for selected genes. Each cell receives scores on components that combine several genes. A neighbour graph links cells with similar scores, and graph clustering partitions those connections. The choice of features, components, distance and neighbours can affect the partition.
Marker testing comes after proposing groups. It compares a cluster with stated comparison cells and may examine eligible genes beyond the original grouping features. Report that comparison and the gene set eligible for testing. A gene used to construct a partition is not independent confirmation that the partition has a particular identity.
If a marker list is used for pathway enrichment, the background should reflect genes eligible to enter that list. The earlier enrichment lesson's warnings about selection effects, overlapping annotations and multiple testing still apply. An enriched pathway can help describe a group's expression pattern without naming its cell type.
Check which conclusions survive a different partition
An analysis can partition cells at several levels of detail. Raising clustering resolution often divides a broad group into smaller groups. That change does not automatically prove new cell types or erase the evidence for the broad identity.
Ask what distinguishes the smaller groups. Compare marker patterns, their distribution across biological samples and their membership under reasonable analysis settings. A split that mainly follows sample processing needs a different explanation from one supported by repeatable biological differences. Some variation is continuous and does not require a set of discrete type names.
A UMAP view can help inspect the data, but the graph partition and its two-dimensional display are different objects. UMAP aims to retain local neighbourhood structure and can distort relationships. The size of a gap or island is not a calibrated biological measurement.
Record conclusions separately. A broad identity may be supported while finer state labels remain tentative. Several markers and a relevant reference support a provisional identity; repetition and independent assays can strengthen it. Agreement across settings helps assess a claim but does not establish biological meaning by itself.
RNA detection is not a direct transcription or protein measurement
Transcription makes RNA. For protein-coding genes, translation uses messenger RNA to make a polypeptide. An RNA sequencing measurement does not directly measure either the rate of transcription or the abundance of protein. RNA can remain after its production, and protein production and loss introduce further differences.
A detected RNA molecule also needs a defensible cell assignment. A droplet barcode identifies a compartment rather than proving that exactly one cell supplied all its material. Ambient RNA can enter that compartment, and a doublet can combine two cells' profiles.
A zero creates a different uncertainty. It states that no molecule of a gene was counted. RNA may have been absent, or it may have been present but missed. The matrix alone does not justify either assumption: that all zeros are biological absences, or that all zeros conceal expression.
Use evidence across genes and cells, alongside protocol checks. A mixed marker profile can justify investigating a doublet without diagnosing one with certainty. After quality checks, a proposed identity still needs biological support. State what was measured and which explanations remain unresolved.
2 · Watch
به انگلیسیTake it off screen
Where it sits
Learn first
Optional connections
8 questions wait behind this lesson, each with its answer explained. Every answer feeds the sky: stars light as they are learned, and dim when it is time to come back.