FAQ  •  Register  •  Login

VISNE and Zero values in CyTOF data

Forum rules
Please be as geeky as possible. Reference, reference, reference.
Also, please note that this is a mixed bag of math-gurus and mathematically challenged, so choose your words wisely :-)
<<

nkhanbham

Master

Posts: 53

Joined: Wed Feb 25, 2015 3:03 pm

Post Fri Mar 17, 2017 7:16 pm

VISNE and Zero values in CyTOF data

I have noticed that VISNE plots generated from CyTOF data diplay unusual wispy clusters while VISNE on Flow data files gives clusters that are more rounded if that makes any sense. I have been told that the high number of zero values for all the different channels from CyTOF causes this. Has anyone else observed this with their data? How can one get around this? I have not applied the ArcSin calculation when I normalize my CyTOF files (I use the Fluidigm method) as this stretches the negative population over a couple of logs and makes biaxial plots look quite different. Would altering the zero values so they are normally distributed between -1 and 0 and then running VISNE be worth trying? I would appreciate any advice people can give.
Thanks,
Naeem
<<

mleipold

Guru

Posts: 5796

Joined: Fri Nov 01, 2013 5:30 pm

Location: Stanford HIMC, CA, USA

Post Thu Mar 23, 2017 8:59 pm

Re: VISNE and Zero values in CyTOF data

Hi Naeem,

I don't really have a good answer for your question....it's not something I've investigated closely.

You should be able to try distributing the zeroes in some fashion, between -1 and 0 as you discuss. There's also a re-processing step you can do: if you have the FCS file, you can go into FCS Processing and play with the Gaussian Negative Half-Zero Randomization.

This is normally unselected ("off"). In CyTOFv1 software, it was accessible in User Mode. In Helios, it's visible in all modes, but only clickable in Service Mode. I'm not sure about CyTOFv2 software, and I don't have one handy to check. I'm sure there are ways to do this in R or some other language too.....


In short, this allows you to take your zeroes (and only your zeroes) and randomize them around zero, according to a Gaussian distribution. You have the ability to play with the shape of the distribution by changing the Sigma and the Power Parameter. I have attached a couple slides from several years ago when I looked at this.

**Personally, I don't think this is a good idea, as it's injecting noise into your data for what I consider to be only aesthetic reasons (ie, make it look like Flow data). And, depending on the wifth of your peak-shape, you could wind up losing some resolution against Dim populations.


However, for the purposes of testing whether the pile-up on Zero is causing your viSNE "wispiness", it might be useful.


Mike
Attachments
Gaussian half-zero.pdf
(159.25 KiB) Downloaded 408 times
<<

dtelad11

Master

Posts: 129

Joined: Mon Oct 31, 2016 6:26 pm

Post Thu Mar 23, 2017 11:53 pm

Re: VISNE and Zero values in CyTOF data

I would try the hyperbolic arcsin transformation -- in our experience it leads to much cleaner analysis when it comes to automated methods such as viSNE, PhenoGraph, etc.

Alternatively, you're welcome to share one of your data files (I'm at el-addavid.amir@mssm.edu) and I'll have a look.
<<

samusik

Participant

Posts: 3

Joined: Thu Dec 08, 2016 8:56 pm

Post Fri Mar 24, 2017 12:21 am

Re: VISNE and Zero values in CyTOF data

Hi Naeem,

I would recommend trying VorteX/Force-Directed Layout, it tends to give better embeddings and smoother trajectories than viSNE

https://github.com/nolanlab/vortex/wiki/Getting-Started

Best,

NIkolay
<<

dtelad11

Master

Posts: 129

Joined: Mon Oct 31, 2016 6:26 pm

Post Mon Mar 27, 2017 2:48 pm

Re: VISNE and Zero values in CyTOF data

samusik wrote:I would recommend trying VorteX/Force-Directed Layout, it tends to give better embeddings and smoother trajectories than viSNE

A bit off topic: We are currently looking into metrics for comparing the various dimensionality reduction tools out there (t-SNE, X-shift, largeVis, etc.) in the context of CyTOF data. I haven't seen any comparisons to existing methods in Samusik et al., 2016 -- are you working on a follow-up paper, by any chance? Which metrics have you used in assessing VorteX when comparing it to viSNE?
<<

nkhanbham

Master

Posts: 53

Joined: Wed Feb 25, 2015 3:03 pm

Post Wed Mar 29, 2017 9:26 am

Re: VISNE and Zero values in CyTOF data

Thanks guys - much appreciated.
I will play around with the randomization. I have also found that increasing the number of iterations in vISNE settings (from default 1000 to 3000 or higher) improves how the populations look. So zero values may not have the impact I thought it did.
Naeem

Return to CyTOF data analysis

Who is online

Users browsing this forum: No registered users and 6 guests