PBMC characterisation: single-cell RNA-seq clustering and cell-type annotation

Single-cell RNA-seq analysis of human peripheral blood mononuclear cells (PBMCs) using scanpy. The repository holds the full downstream workflow: quality control, normalisation, feature selection, dimensionality reduction, Leiden clustering, marker-gene testing, and annotation of clusters to cell types.

PBMCs are a standard benchmark for single-cell workflows because the expected populations are known, which makes them a good system for checking that a pipeline recovers the right biology rather than just producing clusters. The workflow follows the established scanpy PBMC analysis pattern, with the QC thresholds and marker panel set out explicitly below.

What the analysis found

Cells before / after QC filtering	fill in
Genes retained after filtering	fill in
Highly variable genes selected	fill in
Leiden clusters identified	fill in
Cell types annotated	CD4+ T, CD8+ T, B, NK, monocytes, dendritic cells, platelets
Largest population	CD4+ T cells

The dominant population is CD4+ T cells, which is what a healthy PBMC sample should give: T cells typically make up the majority of mononuclear cells in peripheral blood, with monocytes, B cells and NK cells forming smaller fractions and dendritic cells and platelets rarer still. Recovering that composition is the main sanity check on the pipeline.

Method and the decisions behind it

Quality control. Cells with fewer than 200 detected genes and genes present in fewer than 3 cells are removed. Cells are then filtered on two QC metrics: fewer than 2,500 detected genes, and under 5 per cent mitochondrial counts. The upper gene threshold removes likely doublets, and the mitochondrial cutoff removes stressed or dying cells whose transcriptome is dominated by mitochondrial reads.

Normalisation. Counts are normalised per cell to 10,000 and log1p-transformed, so that clustering reflects expression profile rather than sequencing depth.

Feature selection. Highly variable genes are selected (min_mean=0.0125, max_mean=3, min_disp=0.5) and the data scaled to a maximum of 10. Restricting to variable genes focuses the structure on biological variation and reduces noise from uniformly expressed genes.

Dimensionality reduction and clustering. PCA (arpack), then a neighbourhood graph over 40 principal components with 10 neighbours, then UMAP for visualisation and Leiden for clustering.

Marker genes and annotation. Cluster markers are ranked with rank_genes_groups, then clusters are annotated using a canonical PBMC marker panel:

Cell type	Marker
T cells	CD3D
B cells	MS4A1
NK cells	GNLY
Monocytes	CD14, LYZ
Dendritic cells	FCER1A
Platelets	PPBP

Annotation is done by inspecting a dot plot of these markers across clusters and assigning each cluster to the type whose markers it expresses, rather than relying on an automated label transfer.

Repository layout

data/
  preprocessing.py          full analysis script (QC through annotation)
  pbmc_preprocessed.h5ad    input AnnData object
fig/                        output figures

Running it

conda create -n pbmc python=3.11
conda activate pbmc
pip install scanpy anndata matplotlib pandas leidenalg

python data/preprocessing.py

The input path is set at the top of preprocessing.py.

Limitations

• Cluster-to-cell-type assignment is manual and based on a small canonical marker panel, so rare or transitional populations may be merged into larger clusters.
• Marker testing uses a t-test; a rank-based test such as Wilcoxon is generally more robust for single-cell count data and would be the natural next step.
• QC thresholds are the commonly used defaults for PBMC data rather than values tuned to this dataset, and would need revisiting for a different tissue or chemistry.

Next steps

Packaging the workflow so it runs end to end from a config file, and replacing manual annotation with a reproducible, scripted assignment step.