The short answer to your question is yes, you can use two different tools to annotate your dataset, one dedicated to identifying and annotating cancer cells and the other for immune cells. I'm not familiar with scCancer2, but it appears to be a more general tool that does more than just cell type annotation. If you have a dataset containing cancer and immune cells similar to yours that has already been annotated, you could use that dataset with scArches to annotate your own.

In my opinion, you should first annotate the major cell types (e.g., cancer cells, lymphoid cells, myeloid cells, CAFs, etc.) by clustering the single-cell data at a lower resolution, just enough to separate the major discrete groups. If you are working with a whole-tumor sample (I mean not enriching specific cell type), this should give you a reasonable separation between tumor cells and the immune compartments. For additional confidence, you can also try copy number variation (CNV).

After identifying the major cell types, you can then dissect the cell states within each compartment. There is growing evidence that different cell types may share similar cell states, which can lead to confusion during cell-state annotation, if you attempt to resolve fine-grained states without first separating the cell types.

I am not sure what is your aim though!

Cheers!!

Nitin N.

I'm not familiar with scCancer2 and scArches, but the R package "SingleR" provides a convenient way to annotate cells using multiple references. For example, say you have a cancer cell reference dataset (cancer_ref) and an immune cell reference dataset (immune_ref). For each reference dataset, each cell in your unannotated dataset will receive an annotation if there is a reasonable match to it in the reference dataset. So, a cell may receive an annotation from cancer_ref, immune_ref, both, or neither. If both, SingleR will select the best-scoring label as the final annotation for that cell. Otherwise, the single annotation will be the final annotation. This is all described in more detail here: https://bioconductor.org/books/release/SingleRBook/using-multiple-references.html

This does not come without some caveats. If each dataset contains a discrete cell type represented in two different ways (e.g., "Example Cell" and "example_cell"), you may see some heterogeneity or messiness in your annotations. Some "Example" cells may receive the "Example Cell" label, others may receive the "example_cell" label. If you can harmonize such labels beforehand, that is ideal, but this can also be non-trivial. If there is little overlap between the reference datasets, as I imagine there might be between a cancer and immune dataset, it may not matter as much. Related to that, I'm also not sure how well this works when reference datasets have little overlap. I've only used this with mostly-overlapping reference datasets.