Assessing prior model validity

[ndejay]

Hello,

Thanks for publishing this great method.

I am looking to generate new priors from a pilot data set consisting of a large number (>10K) of deeply sequenced cells (>100K reads/cell) from Chromium 10X technology, and had some questions about how to fine-tune model training w.r.t recapitulation of the original data.

Comparison of gamma mixed fits with original means. Is it normal for the gamma mixed fit to underestimate 0s and overestimate moderately expressed genes?

Fit a function for UMI counts dependent on read depth:

Estimation of median dispersion function for each cell type. In the example in the vignette, no relation was found between dispersion and UMI counts. In my pilot data set, I am seeing one. Is there a way to account for this?

Validation of expression probability model. The predicted number of expressed genes are off by a significant margin.

Many thanks in advance!

Nic