FitBUM distribution error about scppin HOT 5 CLOSED

That the checksum is zero means that all your p-values are in (0,1]. That your optimization fails indicates that the p-values don't follow the beta-uniform model (BUM).

Maybe inspect your p-value distribution in a histogram and check whether you see any unexpected behavior.

If you don't see anything obviously wrong you can always create node-weights from p-values in a less statistically rigorous (but still reasonable) way:

node score = log(p-value) + X
where X in (-Inf,+Inf) is an offset, which varies the size of the functional module you will get. The smaller the offset, the smaller the module you obtain.

A good starting point to play around with X is around -1. (At least in some experiments we did for our data.)

from scppin.

Thanks for getting back. I think you are right there is something wrong with the pvalue distribution. As i was using the multiple test corrected p-values, and i guess for some cell types the distribution was more heavily inclined to very low p-values than 1s and perhaps thats why the pvalues dont follow the BUM . When i reran the process for these cell types using the uncorrected pvalues it ran fine. I was wondering, in the paper and the tutorial are you using the corrected or un-corrected pvalues?
I will rerun the analysis by playing around with the X.

Thanks,
Devika

from scppin.

It is best if you use uncorrected p-values to ensure that they are all in (0,1]. That's what we did in the paper, too.

Glad to hear that it works with the uncorrected p-values.

from scppin.

Thanks Florian.
another question did you always get a high (> 0.05) p-value so that you could accept the null hypothesis of the KS test that the distributions are similar. Because i dont always get a p> 0.05 ,but get a low D statistic.

Example using un-corrected pvalues

One-sample Kolmogorov-Smirnov test

data: AT_DAT_H_pValues
D = 0.030435, p-value = 2.15e-09
alternative hypothesis: two-sided

from scppin.

In the example we discuss in the SI of our paper we do:
"The BUM is in decent agreement with the observed p-values (a one-sample Kolmogorov–Smirnov test yields a statistic of D ≈ 0.014 and a p-value of 0.086)."

We got similar results for other data we analysed.

To be honest, I wouldn't worry too much if the KS-test is not significant: you still can use the scPPIN pipeline and the scoring function. But you probably shouldn't completely trust the false discovery rates that we specify.

from scppin.