Projection to custom reference atlas

[shwong-tw]

Dear Developer,

I wonder whether ProjecTILs (v3.6.0) is expected to work well in this scenario: query (NK+T cells) project to custom reference (all cell types)?

With this scenario I got the projected result shown in figure below, where I expect the contours to sit in purple/orange regions.
It would be much appreciated if you could point me to potential issues you could spot with these limited information.
(It is difficult to provide the data and code to reproduce this figure, but if there's no immediate clues I will try to come up more concrete codes.)

Thank you very much!

[shwong-tw]

Hello developer,

I'm adding one more thing I tried today.
So I tried to subset the custom reference and keep only T / NK cells.
With make.reference, it generates UMAP shown in the right panel below, which resembles a lot the UMAP I posted yesterday.
Therefore I stepped into make.reference, use the new pca generated by prcomp_seurat, but generate UMAP using Seurat::RunUMAP(..., reduction = "pca", n.neighbors = 15, min.dist= 0.3, umap.method = "umap-learn", seed.use = 123) to mimic the behavior of run.umap.2 (which uses umap::umap). This gives the UMAP shown on the left figure below.

I haven't got time to further compare where the difference results from,
but would you see this as an expected behavior and kindly point me to which setting decided this behavior?

Thank you very much!

[mass-a]

Hello @shwong-tw
I think your scenario is a perfectly reasonable way of using ProjecTILs.
In the projection plots from your first post the projected points are mostly outside of the reference. This makes me suspect that your object may contain two different sets of UMAP coordinates; is that possible? do you have for example UMAP coordinates stored in the object metadata? (Seurat may decide to take the coordinates from there instead of the embeddings slots).