Limit cell types in spillover step of xCell

[grst]

According to @dviraran, the performance of xCell can be improved
by limiting the cell types to immune cells only in the spillover step.

See also this tutorial.

As this package targets immune cells, it probably makes sense to limit the cell types per default in the
deconvolute function.

[dviraran]

Thanks Gregor.
Just to be clear - "limiting the cell types to immune cells" should be "limiting the cell types to cell types of interest". For example, If you are running it for PBMC and no macrophages are expected, its better to run xCell without them.

[grst]

ok, I see.

1. Would it already help to limit the analysis to immune cells, as a default option for the 'naive user'?

2. For the benchmark, I would consider it as 'fair' to limit xCell to the same cell types that the other methods are using. What do you think?

3. I am thinking of how to make this accessible to users. Maybe a use_cell_types argument of the deconvolute() function. Maybe you could also implement this on the xCell side as a parameter to the xCellAnalysis function that would wrap the code you show in the tutorial.

[dviraran]

1. No, because there is almost no spillover between the immune and the other cell types.

2. Yes. In my experiments, this can significantly improve results for some cell types

3. Thanks for the suggestion. I added cell.types.use argument to the xCellAnalysis function.

[grst]

Great! I'll look into this next week.

[grst]

@dviraran, I now implemented a expected_cell_types parameter for the deconvolute function in the feature/expected_cell_types branch.

At the same time I constrained xCell to the following cell types in the benchmark pipeline for the simulated and FACS data respectively:
https://github.com/grst/immune_deconvolution_benchmark/blob/836b8a631741fb63f3f547ec4a05c7f700e1a9f9/notebooks/config.R#L48

What do you think?

The results did not change a lot. NK cells improved quite a bit, though.
This is what the figures from the manuscript would look like with the modification:

[dviraran]

Thanks. Looks good to me.
The only reason it bothered me is the apparent failure of xCell in monocytes in the FACS validation (1c,d). I think that it is overcompensating with macrophages, and therefore the correlations look bad. Now its better...
Not sure what is the problem with NK cells.

Best,
Dvir

[grst]

Great, will update the preprint shortly!