BRM5012 Chap.9 Clustering and Cell Annotation
Clustering and Cell Annotation
Clustering groups cells by high-dimensional expression similarity, while UMAP and t-SNE provide lower-dimensional views of those neighbourhoods. A visible group is not automatically a cell type. This chapter explains parameter sensitivity, sample and batch overlays, positive and negative marker combinations, manual and reference-based annotation, the distinction between type and state, and cautious treatment of rare clusters.
What this chapter covers
- 01
Expression neighbourhoods
- 02
Dimensionality reduction
- 03
UMAP and t-SNE
- 04
Clustering resolution
- 05
Marker combinations
- 06
Reference annotation
- 07
Cell type and state
- 08
Rare cluster checks
Worked example · free
Resolve an ambiguous synthetic cluster
- +1Use the two positive markers P and Q as support for Type One.
- +1Use the absence of R as additional support against Type Two.
- +1Assign a Type-One-like provisional label rather than an absolute identity.
- +1Check tissue-specific references and whether the cluster appears across samples.
Key terms
- Expression neighbourhood
- A local group of cells with similar high-dimensional gene-expression profiles under the chosen analysis.
- Dimensionality reduction
- A method compressing high-dimensional data into fewer coordinates for exploration and visualisation.
- Marker combination
- A pattern of present, absent or graded gene expression used to support a cell identity.
- Cell state
- A potentially transient biological program such as activation, stress, cycling or treatment response within a cell type.
Clustering and Cell Annotation FAQ
What does each dot on a UMAP represent?
In this workflow, each dot represents a single cell. Nearby dots have similar expression profiles under the chosen preprocessing and embedding, but plot distance is not physical tissue distance and the axes are not direct biological measurements.
Can I name a cluster from one marker?
A single marker can be shared across lineages or altered by cell state. Use combinations of positive and negative markers, expression level, tissue context, sample replication and relevant references. Keep uncertain labels provisional.
How should rare clusters be handled?
Check doublets, low-quality metrics, batch restriction and marker coherence before deletion or naming. Rarity alone is not a defect because single-cell methods are valuable for discovering uncommon cell types and states hidden by bulk averages.
Assessment move
Sketch a UMAP and label what it does and does not show. Repeat that nearby points have similar expression profiles, while axes and global distances are embedding artefacts rather than anatomical coordinates. For annotation practice, require at least two supporting markers, one negative marker where useful and a competing identity. Separate stable lineage identity from activation or stress state.
Overlay sample, condition, batch and QC metrics on every cluster map and ask whether a group is reproduced across biological samples. Compare a small range of clustering resolutions and record which labels remain stable. In assessment figures, explain what each dot represents, how colours were assigned and what limitation remains. Treat automatic reference labels as hypotheses that still need tissue and marker review.
Begin with an unlabeled embedding and narrate only what it directly supports: each point is a cell and nearby points have similar processed expression profiles under the chosen settings. Avoid assigning meaning to axis direction, global distance or island shape. Overlay one variable at a time—sample, condition, batch, nCount, mitochondrial proportion and cell-cycle score—and write which pattern each explanation predicts.
Compare a small resolution range and track neighbourhoods rather than counting coloured islands. A stable marker program can persist even when a display splits or merges groups. For annotation, build an evidence card for every cluster with multiple positive markers, informative negative markers, tissue context and the leading competing label. Mark whether a signal is shared across the cluster or driven by a handful of cells.
Separate lineage evidence from activation, stress and cycle programs so states are not promoted into new cell types. For a rare group, explicitly test mixed lineage markers, high counts, poor quality and batch restriction before calling a discovery. Review automatic reference labels against the reference tissue, disease state, age and technology; low confidence may indicate that the relevant type is missing from the reference.
Compare manual and reference reasoning without treating agreement as proof. Choose the broadest label that the evidence supports and retain a provisional suffix when subtype evidence conflicts. Write captions that define the dot, the colour, the high-level similarity calculation and the biological claim.
Finish each annotation with the observation that would revise it, such as doublet scoring, another marker panel, independent atlas comparison or spatial validation. The result is an auditable hypothesis rather than a name attached to a visually separated group.
Working through Clustering and Cell Annotation in BRM5012? Sia is AskSia’s AI Biology tutor — ask any BRM5012 Clustering and Cell Annotation question and get a clear, step-by-step explanation grounded in how BRM5012 is taught and assessed. Read this chapter free, then take your hardest questions to Sia.
