MAJIQ HET and MAJIQ PSI

[Chris Khoury]

Hi San and Biociphers Team,

Just picking up our conversation over email in thinking about how to explain a few instances of differences between LSVs calculated by MAJIQ HET and MAJIQ PSI.  Thanks for the email where you touched upon this further in response to the MAJIQ processing options that we shared with you for PSI and HET. We appreciate that the algorithms are different and keen to learn more. If you would like me to post further background from our email for other users, let me know.

As requested, I have attached the output visuals for you to interrogate. There is a notepad file with the filename of the snapshots taken from Voila view.

There are 5 examples - you will note the PSI and HET values (median) for the LSVs. I provide examples of where they are absent in one vs the other, two where they are similar (both high and low values), and finally one where PSI reports a much higher value than HET.

Is this what you were after? I do not have a splicegraph view for Voila HET.

We don’t utilize these genes in our classifier - rather it is the observation that the two pipelines produce different results - differences we will need to at least be able to explain the “why” for in our clinical pipeline. There are some samples whose values correlate perfectly from the exact same processing and comparison method.

Cheers and look forward to discussing further,
Samantha & Christopher.

[Chris Khoury]

Hi San, Hope things aren't too busy before Christmas.

Recently saw the team's MAJIQlopedia - finally - an updated splicing database. Thank you!

Was keen to see if you had any ideas about what might be occurring in the example dataset posted herein?

Regards,

Samantha/Chris

[San Jewell]

Hi Samantha/Chris,

Appreciate it, things have been a bit busy just on other efforts, haha.

I actually didn't forget about this and I've brought up the question with someone who did more detailed work in that area of the software. However, I believe they have been busy or away, so it's taking a little longer to get a definitive response. I'll post back as soon as I have more information.

Thank you and have a great holiday.

-San

[Chris Khoury]

San, thank you for following up with your colleague. 

We look forward to hearing your recommendation/thoughts on tackling the above issue.

Cheers.

[San Jewell]

Hi Chris / Samantha,

I've come back with talking over this with one of the other developers, and I'm afraid I'll still need to ask you a few more questions. What I can say, is there there are two separate questions I believe you are asking here (1) why are some LSV quantified with one approach and not quantified with another and (2) why are some individual sample PSI different between approaches.

For (1), I was able to reproduce this difference, but only with very low coverage cases, where the LSV was nearly constitutive. As the two methods use slightly different methods of filtering (HET requiring across-sample consensus and PSI requiring just a single sample), there are tiny differences in the exact criterion for acceptance. These will only become apparent with nearly constitutive LSVs where the LSVs are close to being noise, so we don't expect and major issue from leaving these small differences in place at this time

For (2), I was not able to reproduce this behavior yet using my sample data, and I'm also slightly confused about what's presented in your first post. Perhaps something was lost between when we were communicating over email and this board post? I'm attaching the full grid of cases based on your table:

PSI is on the left and HET is on the right.

In the first row, I see PSI value matches with PSI value for the blue junction (with one figure rounding), and I see the same value highlighted in the HET plot. I don't know where the HET value in the table (0.972953) is shown at all, so it looks like this is the correct result?

In the second row, I think there may have been a mistake as the PSI snapshot is actually of the same image (nearly) and so the values are the same? I'm a bit confused about this one. In the HET screenshot there is no other sample highlighted, so we cannot compare the value.

In the third row We see 0.5 shown in the PSI snapshot and 0.5 also shown in the HET snapshot, so it also looks like this is the correct result.

---------------

In the context of the original email, the visual check was done after running $ voila tsv mode on HET and PSI outputs with permissive arguments  --show-all --show-per-sample-psi --changing-between-group-dpsi 0 --non-changing-between-group-dpsi 0 as well. I also checked on some of my own sample data, PSI on one replicate, and HET with all three replicates (v3 three others) , and I double checked that the per-sample data in mine was also exactly matching between both modes.

Given this, For issue (2) I am somewhat confused where the problem is. I may need to arrange looking closer at the input data to reproduce exactly the difference in per sample PSI that is described. Could you please confirm if I am at all understanding the issue you were having?

Thank you,

-San