Help understanding patterns in EBSeq

Question

[snamjoshi87]

EBseq allows the user to generate different expression patterns in the data based on experimental conditions using the GetPatterns() function. However, I am not able to understand exactly what these patterns signify. According to the vignette, if we have 3 conditions, a pattern of 1 1 1 signifies the three conditions have the "same latent mean expression level" while 1 1 2 would indicate that conditions one and two have the same level but condition three has a different latent mean expression level.

What is "latent mean expression level"? How would I know what latent mean expression level I might expect for each condition?

I have tried to search for this term online but could not come up with much. It seems like it has something to with latent variables in HMM but I don't understand how it applies to setup for EBseq.

Answer 1

[Rik Verdonck]

It's all about testing which hypothesis best describes the data that you see. Say you have 3 conditions: male, female and hermaphrodite. Then 1 1 1 fits those genes for which there is no sex effect, 1 2 2 are those genes that have similar expression between females and hermaphrodites, but different for males, and 1 2 3 are those genes that are different between all conditions. The more conditions you have, the more "constellations" are possible (check Bell number), but not all of them may be equally meaningful.

As far as I understood, the "latent mean" just means that in the case of 1 1 1, all three conditions are from the same NB distribution with the same mean (or, the same p and r parameter). In the case of 1 1 2, the first two conditions share the same parameters (and hence have the same "latent mean"), while the last condition fits another set of NB parameters. I guess they use the phrase "latent mean expression" because it's a property of a transcript that is only indirectly observed through mapping counts. But I'm not sure either.