median_psi values from majiq v3 Modulizer

[Rong Shen]

Hi Majiq team and community,

I am new to this field and running majiq v3 (3.0.8.dev1+g7a81398df) recently to analyze my bulk RNAseq data of cell samples. I would like to ask a question about median_psi and delta median_psi values reported by deltapsi and heterogen.

I have 4 replicates for Ctrl and 4 replicates for Treatment. When I run Modulizer using deltapsi output and check the cassete.tsv file, for example one event of a gene (selected using “lsv_id”, “junction_name”), I got Treatment_median_psi (0.81140), Ctrl_median_psi (0.92520) and Ctrl-Treatment_median_psi (0.04267). You can see 0.92520 - 0.81140 != 0.04267. I checked the table from “majiq-v3 deltapsi”, the “dpsi_mean” value is 0.0427, equal to the Ctrl-Treatment_median_psi (0.04267), but it’s not derived from Crtl_raw_psi_mean (0.921) – Treatment_raw_psi_mean (0.8164), either.

I also run Modulizer using heterogen output, checking the same gene, same event, Ctrl-Treatment_het_median_psi (0.14020) is exactly equals to  Ctrl_het_median_psi (0.92780) – Treatment_het_median_psi (0.78760). But the median_psi from deltapsi and heterogen are different.

I would like to know which Ctrl-Treatment_median_psi value should I use, the one from deltapsi or heterogen? And what is the difference between them?

Many thanks,

Rong

[San Jewell]

Hi Rong,

You are correct that there is a difference in dpsi between het-mode (which uses a simple difference) and dpsi-mode (which uses a joint prior). While I could ask my labmates for more input on proper use cases, I understand that dpsi analysis is preferred for "well controlled" experiments, perhaps groups of replicates, where het mode expects higher variance (i.e. between two different tissues) If you'd like to read more about the joint prior calculation of dpsi mode, you can check under the materials and methods section here: https://elifesciences.org/articles/11752#s4 

Let me know if it makes sense,

-San

[Rong Shen]

Hi San,

Thanks very much for your response. I will read the reference paper you recommended.

As I used the same cells (for Ctrl and Treatment), would dpsi-mode makes more sense to use? If I use Modulizer with deltapsi as input, should I use Ctrl-Treatment_median_psi (0.04267) directly, or manually do the calculation using Ctrl_median_psi (0.92520) - Treatment_median_psi (0.81140)? 

Please advice if you have further comments from your labmates.

Best,

Rong 

[San Jewell]

Hi Rong,

I have asked someone in my lab and they said for your case using dpsi mode should be fine.

For using modulizer, I have noticed from some of my work that using het mode tends to capture a greater number of cases when running an analysis over highly changing events, and dpsi mode tends to capture more when looking for highly confident non-changing events, just anecdotally. I will note that if you want to leave open the option of showing the sum of what they both capture, it is valid to run modulizer with both a dpsi and het input file, and then you will get a dpsi column for each, which might allow easier filtering.

Let me know if it makes sense,

-San

[Rong Shen]

Hi San,

Thanks for sharing the insights from you and your labmates. Very helpful. I will compare the results from both. 

Could you please advice my following questions:

1. What criteria would you suggest for the selection of hits: e.g., I want to find the event with significance in cassette. 

     For Modulizer results with deltapsi input, I chose Treatment-Ctrl_median_dpsi > 0.2 and Treatment-Ctrl_probability_changing > 0.95

     For Modulizer results with het input, I chose Treatment-Ctrlhet_median_dpsi > 0.2 and Treatment-Ctrl_wo_het_ttest < 0.05

2. As I asked in the last post, for Moldulizer results with detapasi input, Treatment-Ctrl_median_dpsi != Treatment_median_dpsi - Ctrl_median_dpsi. Should I use Treatment-Ctrl_median_dpsi or calculate Treatment_median_dpsi - Ctrl_median_dpsi instead.

Many thanks,

Rong