Closed ScreachingFire closed 1 year ago
Correction: The first code section appears in get_model_pars()
which is called within vst()
theta_mm/theta_mle<0.001
, gets around the problem of defining an explicit cutoff for "theta_mle" to call something poisson. For some poisson like genes (often with mean < 0.01), it is possible to get an overdispersion factor (MLE-based) like 100 due to convergence issues, but they are effectively poisson. This is what the check ensures.
Hello!
After estimating the NB model parameters within
vst()
, this section of code appears afterwards:I'm just curious about the reasoning behind this section. Prior to learning the model parameters on a subset of genes, there is already a section that filters out genes that do not seem to exhibit overdispersion:
So I'm confused as to why you need to set the theta of some genes to Infinity when the predicted theta is much larger than the model-learned theta. What does this large difference between the thetas imply such that this is needed?
Any insight into this would be helpful! Thank you!