DR

Which dimensionality reduction (DR) algorithm should I use?

While the big DR debate of 2023/2024 has subsided somewhat, these Swiss marmots remain steadfastly neutral on the topic.

All DR results are highly sensitive to:

• marker selection
• data transformation and cofactors
• algorithm-specific parameters

We use DRs mostly for QC (identifying cells that escaped our gating) and to aid with cluster labelling, and we try to avoid over-interpreting the ordination of our cells.

Practical recommendations

In practice, we have found that different algorithms emphasise different aspects of the data:

t-SNE

• Strongly emphasises local structure
• Pulls nearest neighbours tightly together
• Highlights fine-grained subpopulations

👉 We prefer t-SNE when investigating subtle heterogeneity within a population
(e.g. identifying subpopulations within Tregs)

UMAP / PaCMAP

• Better preserves aspects of the global structure
• Maintains relationships between broader cell lineages
• Produces more coherent “islands” of major populations

👉 We prefer UMAP or PaCMAP for broader datasets
(e.g. CD45⁺ cells), where distinct populations such as:

• CD4 T cells
• macrophages
• B cells

should form clearly separated clusters

💡 No single DR method is “correct” — the choice depends on your biological question.
We recommend trying multiple approaches and using them as complementary views, rather than relying on a single embedding.