Using original labels for cluster annotation

[gboscagli]

Greetings all,

I was wondering if I could use harmonized labels from original_ann_level_[1-5] in HLCA extended to annotate my LUAD dataset (at the cluster level). I understand that, from a technical standpoint, working on labels obtained after scANVI integration and manually curated reannotation is best: but in ann_level_[1-5] I am lacking the 'Cancer' annotated cells I have in the original labels.

I am unsure between these two approaches:

• annotate clusters using original_ann_level_1 (since I just need a coarse, compartment-level annotation).
• annotate clusters using ann_level_1 and then re-annotate epithelial clusters only in healthy/tumoral.

Any feedback is highly appreciated!

Best,
Gabriele

[lisa-sikkema]

Hello Gabriele!

Thank you for your message :)
I think if I were you I would maybe do both. The original ann_level_[1-5] are not consistent across datasets and are likely of lower quality, but they are indeed more diverse in some cases.
Either way, in level 1 there is no cancer-related annotation anywhere right?

I think you could use the original annotations for discovery purposes, it might highlight cells your interested in that you could then better/further annotate yourself. I would use the final ann_level_[1-5] for actual annotations, that you might then refine with manual work.

Also good to pay attention to how many labs actually labeled cells a certain way for a given annotation you are interested in. Sometimes they just come from one lab. You can easily visually check this in cellxgene by coloring the cells by study, and then checking the composition in the annotation dropdown menu.

Let me know what you think!

[gboscagli]

Hello Lisa!

Either way, in level 1 there is no cancer-related annotation anywhere right?

In original_ann_level_1 there actually is a 'Cancer' annotation that indeed comes from just one lab (Thienpont_2018). It comprises 12k cells, 6k cells when restricted to LUAD alone. So, no dataset consistency issue, but cells are quite few with respect to the other level 1 groups (Immune/Stroma/Endothelial...), and I am not sure whether this could bias annotation procedure.

I think you could use the original annotations for discovery purposes, it might highlight cells your interested in that you could then better/further annotate yourself. I would use the final ann_level_[1-5] for actual annotations, that you might then refine with manual work.

Clear. I could try both ways and see how the labeling performs. In principle, I wanted to know if original labels were there just for 'legacy' reasons or if they could actually be used.

Thank you for your prompt response!

[lisa-sikkema]

Ah I see, you are talking about the full HLCA (core + extension).
Maybe good to know that the ann_level_[1-5] annotations for the cells in the extended HLCA (about 1.7M extra compared to the core) were themselves already annotated with label transfer. For that annotaiton I would therefore only transfer from the core I think, to keep things tractable.
For the original_ann_level_[1-5], as the name already says, these are original annotations and so not label-transferred, even in the extended HLCA, so there you could transfer from the full HLCA.

[lisa-sikkema]

Given how few cells there are, and only from a single dataset, maybe just clustering and seeing where those cells and up could be just as good.

As to the bias you mentioned that could come from low cell numbers: that should not be an issue as long as the cells cluster somewhat separately in the embedding. At least if you use the label transfer method I used (and that you can find in the tutorials), the transfer is done based on k-nearest neighbors, with k set to something like 50 if I remember correctly.

[gboscagli]

Given how few cells there are, and only from a single dataset, maybe just clustering and seeing where those cells and up could be just as good.

Do you mean running the label transfer with original_ann_level_1 and reannotating query label-transferred cancer cells as healthy/tumoral by UMAP visual inspection/marker expression?

As to the bias you mentioned that could come from low cell numbers: that should not be an issue as long as the cells cluster somewhat separately in the embedding. At least if you use the label transfer method I used (and that you can find in the tutorials), the transfer is done based on k-nearest neighbors, with k set to something like 50 if I remember correctly.

I am currently using CellTypist to transfer the labels, retraining it to the desired annotations instead of the one provided by the authors. The embedding separation thing, however, should stand invariantly of the tool, and it really makes sense.

[lisa-sikkema]

Ah sorry yes I was assuming you have a joint embedding, which is not the case if you are using CellTypist.

For my first suggestion (clustering etc.) I meant clustering the HLCA and your data jointly, which means you would need a joint annotation (which you can get with scArches, as also done in the paper and the HLCA tutorials). Same with the knn-based label transfer

[gboscagli]

I guess I'm pursuing both ways (original_ann_level from full HLCA, ann_level from core HLCA + epithelial reannotation with predefined markers). I'll keep you updated if something interesting pops up.

Thank you @lisa-sikkema!

[lisa-sikkema]

Sounds good! and yes, let me know how it goes :)