CITE-seq scatter plot of pairwise RNA-prot correlations

[chris-rands]

Thanks for muon. Following the CITE-seq tutorial, I'd like to do simple scatter plots and regression to show the RNA-prot correlation for particular markers, something like in the Seurat muli-modal tutorial.:

I tried sc.pl.scatter(mdata, x="prot:CD3D", y="rna:CD3D") This raises: KeyError: 'There is no key prot:CD3D in MuData .obs or in .obs of any modalities.' Of course the gene/protein names are stored in .var not .obs.

Suggestions appreciated, thank you

[gtca]

Hey @chris-rands, thanks a lot for your feedback! The reason was the lack of .pl.scatter in muon.
Now mu.pl.scatter(mdata, ...) should work.

Moreover, for this there's now a more sophisticated function under the hood to parse variables so even the following things should work:

mu.pl.scatter(mdata, x="X_umap:1", y="X_umap:2", color="CD8A")
# will produce a plot similar to mu.pl.umap(mdata, color="CD8A")

mu.pl.scatter(mdata, x="CD8A", y="CD8a_TotalSeqB", color="rna:X_umap:1")
# will set the colour to the first UMAP component in the RNA modality

It also seems to me that scverse/scanpy#311 hasn't been actually fixed — is that your experience as well?

Please let me know what you think!

[chris-rands]

Hi @gtca, thank you very much, I didn't know this would require a new feature! I have tested it and it works. One comment, removing the "rna:" or "prot:" prefixes from the feature names isn't needed in my use case since I already ran mdata.var_names_make_unique() so to get this to work I actually need to double the prefixes, e.g. mu.pl.scatter(mdata, x="prot:prot:CD3D", y="rna:rna:CD3D").

I agree that the scanpy issue does not appear fixed in the sense that the color arg cannot be a list for sc.pl.scatter` it seems

Is it possible to add a regression line to the scatter plot?

[gtca]

Hey @chris-rands, glad that it works! And thanks for your feedback.
Now it should handle the case when var_names have a modality prefix correctly.
Moreover, now the same functionality should work even when some modalities have cells that are not present in others (e.g. the 'rna' modality has more cells than 'prot' modality does).

With regression, if this was implemented e.g. in sc.pl.scatter, mu.pl.scatter would just work (or we could at least build on top of that). I am not sure this is something coming in sc.pl.scatter though. One might want to overlay other interpolations, or exclude cells with missing values (e.g. with zero CD3E expression in the plot above) from the fit, and all this can be achieved e.g. with matplotlib although would require a few more lines of code:

# Calculate coefficients for the regression line
from numpy.polynomial.polynomial import polyfit
x = mdata["rna"][:,"CD3E"].X.squeeze()
y = mdata["prot"][:,"CD3_TotalSeqB"].X.squeeze()
a, b = polyfit(x, y, deg=1)

# Plot data
mu.pl.scatter(mdata, x="CD3E", y="CD3_TotalSeqB", show=False)
# Plot regression line and display the figure
plt.plot(x, a + b * x, "-")
plt.show()