Reproducing cell clustering

[njbrown750]

I am attempting to reproduce the clustering of T-cells and NK cells in the PanGI Atlas using R v4.4 and Seurat v5.3. I am working with the healthy TNK RDS Seurat object. I am not familiar with Python and I am having difficulty finding the correct parameters in the Jupytr notebooks to create Seurat clusters that closely match the level 3 annotation. While I can create the same number of clusters, my clusters each contain a wide mixture of the level_3_annot annotations despite trying many different parameter values. Seeing as I'm using R and a different compute cluster, how much difference in computed clusters should I expect to see? There are too many parameters for me to feasibly optimize them all. Would it be possible for me to get a list of argument values to use in Seurat to reproduce the TNK object clustering?

[amandaoliver15]

Hi! Thanks for your interest in the atlas! In short, I would expect some difference, but not to the extent of highly mixed populations in the recalculated leiden clusters.

How different is the clustering you are performing vs level 3 annotation? Could you please share 1) UMAPs showing the atlas annotations vs your Leiden clusters and 2) a barplot or crosstab comparing level 3 annotations to your clusters? This would help us understand what’s going on.

For context, many T/NK (and other) lineages were subclustered during atlas annotation. In this case, we isolated a specific cluster or groups of clusters, recalculated the UMAP using the existing coordinates, and reran Leiden clustering on this subset to resolve finer populations. For example, a single cluster in the full T/NK object might correspond to Tregs, but subclustering those cells can reveal distinct subsets, such as Helios/IKZF2⁺ or IL10⁺ Tregs.

Another factor to keep in mind is the biological complexity of T cells. Although T cell subsets are well defined by protein markers, the corresponding RNA markers often show low expression (for example, CD4 protein vs RNA). In addition, many T cell states exist along a continuum, such as the transition from naïve to memory, so depending on the clustering parameters or algorithms used, cells in intermediate states may shift between classifications. When we applied scANVI/scArches for integration and transfer learning, these transitional populations also showed higher uncertainty scores, for example in naive-memory and proliferating-mature transitions (see Supplementary Figure 8 in the paper).

[njbrown750]

Thank you for the answer. It's very helpful. Below are UMAPs and a bar plot comparing the overlap of my clusters to the existing level_3_annot metadata.

However, I've run into another tangentially related issue. I'm looking at using cells originally from the Uzzan 2022 study. The Uzzan 2022 study provided nine samples of which four were "healthy controls". In the PanGI Atlas Seurat object these four samples are labelled as "Neighbouring_inflamed" rather than "Non_pathological" in the sample_category metadata but also labelled as "control" in the donor_disease metadata. Why were these samples labelled as "Neighbouring_inflamed" instead of "Non_pathological"?