Unexpected spliced/unspliced percentages

[carbycrab]

Hello,

Thank you for creating such an innovative, streamlined package! I am using scvelo in my single cell analyses and had a question regarding the scvelo results I got. When I checked the spliced/unspliced percentages scv.pl.proportions(adata), I got majority unspliced (86%) and minority spliced (14%), which seem way off from the reported 15-25% of unspliced intronic reads in La Manno et al 2018. I was wondering whether these statistics indicate something wrong with my analysis method or I should take this as factual and trust my downstream results. Here's my method: I used velocyto to generate individual loom files for all 9 of my 10x Chromium samples, merged the data, and subset to select 18555 genes x 14813 cells of interest. The downstream analyses presented in scvelo tutorial page all ran successfully.

Thanks so much for the help!

-Angela

[WeilerP]

@carbycrab, seems like something is off but without further info/access to the data I can't offer much insight. Are you checking the proportions on the raw counts? If not, this can have unexpected behavior. See e.g. #216 or #462.

[WeilerP]

@carbycrab, okay. With raw counts I mean the unprocessed, whole counts.
Two things: scv.core.get_initial_size(adata) only returns the size (there is no saving going on, you'll need set_initial_size for that) and does so for adata.X. To get the initial size of the unspliced and spliced layer, you need to call scv.core.get_initial_size(adata, layer='unspliced') (layer=spliced, respectively). Could you please double-check that the plotted proportions are the same as reported? So something along the lines of

initial_size_unspliced = scv.core.get_initial_size(adata, layer='unspliced')
initial_size_spliced = scv.core.get_initial_size(adata, layer='spliced')

total_counts = np.sum([initial_size_unspliced, initial_size_spliced], axis=0)

# Might be worth checking the distribution of this
fraction_unspliced = initial_size_unspliced / total_counts
mean_fraction_unspliced = fraction_unspliced.mean() # This is what's being plotted

fraction_spliced = initial_size_spliced / total_counts
mean_fraction_spliced = fraction_spliced.mean()

[carbycrab]

I see. Thank you so much for the code. I have initial_size_unspliced and initial_size_spliced in adata.obs now. When I ran your code, I got:

total_counts = [ 9777 14572 11387 ...  7308  7287  2530]
fraction_unspliced = [0.85752276 0.80743892 0.83929042 ... 0.86507937 0.8497324  0.85652174]
mean_fraction_unspliced = 0.8555956615794414
fraction_spliced = [0.85752276 0.80743892 0.83929042 ... 0.86507937 0.8497324  0.85652174]
mean_fraction_spliced = 0.8555956615794414

Do you know why the unspliced and spliced percentages are exactly the same? The unspliced percentage matches the pie chart, but the spliced reflects 86% instead of 14%.

[WeilerP]

@carbycrab I suppose you copied my type on the code for the fraction of spliced counts. I updated it accordingly.

[carbycrab]

Oh thanks. Sorry, I should've checked that more carefully. Yes after running it again, the fractions match the pie chart-- 86% unspliced, 14% spliced. So, any other ideas for why this could be happening? Thanks so much.

[carbycrab]

I also tried this same method with another 10x dataset I have and obtained more normal values for spliced (82%) and unspliced (18%). This current dataset with strange results is also cell hashed. Could this be contributing to the abnormal proportions?

[Shabbirm21]

Hi @carbycrab and @WeilerP, I realize this a quite old post but I was wondering ifany of you managed to figure out why you were getting this unusual behaviour? I am also getting 60% unspliced, 35% spliced and 5% ambguos. Any clue if this is a normal behavior?

Thanks,
Madiha