MCMC reproducibility about sccoda HOT 2 CLOSED

Hello Orr,

two different runs of scCODA can indeed lead to slight deviations in the credible effects due to randomness introduced by the MCMC sampling. As you pointed out correctly, the Stroma cells are likely at the border of credibility.
To obtain fully reproducible results, you can set the seed for tensorflow with tf.random.set_seed before running scCODA. You can also set numpy's seed with np.random.seed, although this should not be necessary.

In the same way as a frequentist test, the false discovery rate defines the sensitivity of your analysis. Higher FDR will lead to more credible effects, but also a higher chance of false positive results.
There is no exact guideline on how large the nominal FDR should be, but values between 0.05 and 0.2 are reasonable. The only recommendation I can give is that for small sample sizes, you might want to increase the FDR level to obtain credible results at all, but in your case (57 samples), the default of 0.05 should be fine.

I hope that this answers your questions!

from sccoda.

Hi Johannes,
Thanks so much for your quick help! All your points make sense to me, and I appreciate the advice. Best of luck with the review process for scCODA. Best, Orr

from sccoda.