Annotate cell types and states

Clustering groups cells with similar expression profiles. Annotation asks what those groups might be. We will start with broad PBMC labels, check several supporting and negative markers, and then refine the labels where the evidence allows it.

The tissue matters: a plausible blood-cell identity may make little sense in another sample. Use the marker patterns together with that context, and keep uncertain labels provisional. The Identify PBMC populations with scRNA-seq tutorial introduces the broad populations used here.

Open the prepared analysis¶

Open the downloaded store and its saved analysis.

This is the same prepared result as the RNA tutorial. Cluster numbers and labels below belong to this result; a new analysis may produce different clusters. The marker table compares each saved cluster with the other analyzed cells, as the run’s own run["markers"] does.

Read the marker evidence¶

Start with the default marker filters and inspect one cluster:

Each row compares expression in this cluster with the rest of the analyzed cells.

A low specificity score does not establish that a gene is absent. Read frac_exp for detection and compare it with frac_exp_rest when evaluating negative evidence.

fold_change is a ratio of the means of the values that the marker search ranks: here, library-size normalized counts. Sorted from high to low, genes that only this cluster expresses come first with inf, however faint, and genes that no cell expresses come last with NaN. With log_transform=True, or for an ADT assay normalized with CLR, the ranked values are on a log scale, so the ratio compares log-scale means and is not a fold change in expression. Compare fold changes only between results of the same assay and normalization.

Assign initial cell types¶

Plot a few familiar markers alongside the clusters. Drawing high values last makes sparse expression easier to see in these panels.

CD14 suggests a monocyte population, CD19 highlights candidate B cells, and NCAM1 (CD56) highlights an NK-like population. CD8A and CD4 help locate candidate T-cell groups, but neither is sufficient alone. In this example, CD4’s highest specificity score is in monocyte cluster 1. CD19 points to cluster 8, CD8A to cluster 7, and NCAM1 to cluster 5. These associations help locate candidate populations; they do not make any of these genes a unique lineage marker. IL3RA suggests a pDC-like population that we will check with other markers.

The default marker filters hide weak and negative evidence. For the remaining comparisons, load all marker rows once by relaxing both filters. Keep this table for the later panels.

These are starting hypotheses. Several clusters receive the same broad name:

Look for several supporting markers¶

A heatmap helps us check whether several genes support each proposed identity. Show the three leading markers per cluster, with enough width to read the cluster labels:

Read each column’s cluster number before interpreting the heatmap. The column order follows marker similarity, so it is not numerical order.

• Monocytes: cluster 1 has S100A12, VCAN, and CD14. Cluster 9 has a distinct CTSL, TCF7L2, and SMIM25 pattern. Keep its broader monocyte label while investigating its state.

• B cells: IGHD, FCER2, and TCL1A support a naive B-cell interpretation for cluster 8. IGHA1, TNFRSF13B, and IGHG1 support a memory-like B-cell population in cluster 2.

• NK and T cells: KLRF1, FGFBP2, and SPON2 support an NK interpretation for cluster 5. Cluster 6 has a GZMK, TRGC2, and KLRG1 pattern; inspect T-cell markers before interpreting its cytotoxic program. Cluster 7 has CD8B among its leading markers.

• Other T-cell groups: clusters 3 and 4 retain broad T-cell labels. Their different leading markers are a reason to inspect their states, not sufficient evidence to name a subtype.

• pDC-like cells: SERPINF1 in cluster 10 supports the initial IL3RA association, but the short heatmap panel does not settle this rare population’s identity. Keep the label provisional.

For finer T-cell labels, compare clusters 6 and 7 using a small panel:

Cluster 6 has a GZMK-associated cytotoxic T-cell pattern. Cluster 7 combines frequent CD8A and CD8B detection with CCR7, IL7R, and CD27, supporting a naive-like CD8 interpretation. GZMK is rarely detected in cluster 7. These descriptions remain hypotheses about state, not proof of a separate lineage.

Check competing identities¶

Cytotoxic genes are shared by NK cells and some T cells. Compare the candidate NK cluster 5 with T-cell clusters 6 and 7 using CD3D and CD3E, then inspect CD4, CD8A, and CD8B:

In cluster 6, CD3D and CD3E are detected in almost all cells, alongside NKG7 in most cells. This supports a T-cell interpretation despite the cytotoxic program. Calling this cluster NK from NKG7 alone would miss the T-cell evidence. For candidate NK cluster 5, check whether the T-cell markers are depleted relative to the cytotoxic markers.

In cluster 7, frequent CD8A and CD8B detection supports the CD8 T-cell label. Check CD4 and other competing lineages in the same way; low detection of one marker does not establish that a population is pure. Cluster-level averages can also hide mixtures within a group.

Save and compare the reviewed labels¶

Keep the broad labels where we have not established a finer identity. Update the groups with additional supporting evidence:

The cell table includes cells outside this run. Align the labels to the run’s selection before saving them. Rerunning this cell replaces the two annotation columns, leaving the computed clusters unchanged.

The saved naive-like CD8 T cells and memory-like B cells labels summarize the interpretations above. They are provisional teaching labels for this dataset. Keep the marker evidence and the run identity with them when recording an analysis.

What still needs further review?¶

Clusters can separate continuous states within one lineage. Compare marker evidence across partitions with Clustering and cluster evidence before treating each cluster as a distinct cell type.

Mixed lineage markers may reflect doublets, background RNA, or uncertain identity. Check per-cell evidence and Quality control across assays; a cluster-average score cannot settle the question. Rare populations deserve this review as much as large ones.

These marker statistics compare cells. Use Pseudobulk and differential expression (DE) primer for questions about differences between biological conditions. Use Mapping cells and transferring labels when transferring labels from a fixed reference.

Find marker references for your tissue¶

Use resources as evidence to compare with your measured markers, rather than a list of names to assign automatically:

• Marker collections: CellMarker, PanglaoDB, and Human Protein Atlas.

• Reference atlases and annotation tools: Azimuth, ScType, CellTypist, SingleR, and CyteType.

• Gene-list interpretation: Enrichr and ToppGene.

• Protein-marker panels: BioLegend, BD Biosciences, and OMIPs in Cytometry Part A.