FAQ  •  Register  •  Login

Fluidigm vs MatLab normalizer

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 :-)
<<

AdeebR

Grand master

Posts: 169

Joined: Thu Mar 13, 2014 5:58 pm

Location: NYC

Post Thu May 07, 2015 7:36 pm

Fluidigm vs MatLab normalizer

Hello Cytofers,

At a talk today Dana Pe'er stated that the Fluidigm bead normlizer algorithm introduces a lot of error and bias into the data and was pretty emphatic that it should not be used, and that Rachel Fink's Matlab algorithm should be used instead (I was not at the talk, so I heard this second hand). I had always assumed that Fluidigm's algorithm is based on Rachel Fink's and didn't realize that they were that different.

Has anyone else heard of this or done any head-to-head comparisons of the two algorithms?
Adeeb Rahman
Icahn School of Medicine at Mount Sinai, NYC
<<

mleipold

Guru

Posts: 5796

Joined: Fri Nov 01, 2013 5:30 pm

Location: Stanford HIMC, CA, USA

Post Thu May 07, 2015 8:47 pm

Re: Fluidigm vs MatLab normalizer

Hi Adeeb,

I haven't used Fluidigm's algorithm so far, just the MATLAB one.

However, there are definitely some differences between the methods. Fluidigm has a document talking about some of this:
https://www.fluidigm.com/documents (CyTOF, "Normalization of Mass Cytometry Data using EQ Four Element Beads")

From that:

"Two algorithmic processing options are available for normalization of data collected in .fcs file format. Both methods normalize for intra- and inter-file signal drift.
1. Fluidigm (DVS) method: Available on CyTOF SW version 6.0.626 and above. This method normalizes data to a global standard determined for each lot of manufactured EQ Beads, and allows normalization of data within and across experiments as well as across instruments The information for each lot of EQ Beads
is captured in the form of a “Passport”. The software is pre-loaded with passports for all the manufactured bead lots. See Appendix A for a complete description of this normalization method. Note: Eu Beads do not have Passport values assigned and data collected using these
beads must be normalized using the MATLAB Method, described below.

2. MATLAB® Method: Available through freeware offered by Stanford University. This method normalizes data using median bead intensity calculated from across the experimental data files instead of a global predetermined standard (Finck et al, Cytometry A 83:483)."


So, one major difference is that Fluidigm/PASSPORT normalizes everything to effectively an external value, the "global standard determined for each lot of manufactured EQ Beads". In contrast, the MATLAB/Finck et al method uses "median bead intensity calculated from across the experimental data files instead of a global predetermined standard".

It's my opinion that the MATLAB method should do a better job of capturing how the machine is performing on that particular day, in that particular sample. But I have not rigorously tested this, and I would love to hear other opinions/comments.


Mike
<<

AdeebR

Grand master

Posts: 169

Joined: Thu Mar 13, 2014 5:58 pm

Location: NYC

Post Thu May 07, 2015 8:59 pm

Re: Fluidigm vs MatLab normalizer

Hi Mike,

Sure, that makes sense. I had previously wondered about that aspect of the MATLAB normalizer (only normalizing within a batch of files) and considered that a limitation, which was why I have generally favored the Fluidigm version. It is also a little more convenient from a workflow standpoint since you can normalize and concatenate files in a single step.

But if there really is error and bias introduction then this seems to be a pretty serious downside.
Adeeb Rahman
Icahn School of Medicine at Mount Sinai, NYC
<<

mleipold

Guru

Posts: 5796

Joined: Fri Nov 01, 2013 5:30 pm

Location: Stanford HIMC, CA, USA

Post Thu May 07, 2015 9:46 pm

Re: Fluidigm vs MatLab normalizer

HI Adeeb,

I guess my concern about using an external set of values for normalizing is that there are variations in machine performance on a given day.

For example, in the Finck et al paper:

Figure 2E Day D has a distinctly smaller difference between signal for 159 and 169, compared to the other 3 days.

Figure 6C, Day A, smoothed even has them swapped.


So, if I'm understanding the algorithms properly, the Fluidigm algorithm would "force" the 169 to be higher for Fig 6C (or "force" the 159 to be lower). And I'm hesitant to say that's a good idea: while there could have been something funky with the machine that day, the bottom line is, it affects 169 and 159 on your same-day cell samples in the same way.

If anyone has a dataset they want to run through both algorithms and post the results, I know many of us would be greatly interested.


Mike
<<

AdeebR

Grand master

Posts: 169

Joined: Thu Mar 13, 2014 5:58 pm

Location: NYC

Post Fri May 08, 2015 2:02 am

Re: Fluidigm vs MatLab normalizer

Ok, I took a series of FCS files and ran them through both normalizers and did a quick analysis. See attached PDF.

Data are from 8 sequential 10min acquisitions of the same sample:

Set 1: concatenated without normalization using the Cytobank concat tool
Set 2: normalized with the MATLAB normalizer and then concatenated with the Cytobank tool
Set 3: normalized and concatenated simultaneously using the Fluidigm algorithm in the CyTOF software.

I did a very quick and crude analysis: gated on EQ beads and cell singlets separately. For the beads, I used the median intensity for Ce140, Eu 151/153, Ho165 and Lu175. For the cells, I used the median intensity of the positive population for every channel where I could easily gate on a clear positive (e.g., CD45, CD3, CD4, etc.)

The Fink algorithm presumably normalizes over time/across samples, but for the overall concatenated files there is minimal change in median signal intensity pre/post normalization across the mass range.

In contrast, the Fluidigm algorithm gives a big boost in signal (~2 fold relative to non-normalized), presumably because my EQ bead signal was lower than their passport target. But more concerning, the signal boost is not uniform across the mass range, with some masses getting a greater boost than others. So, if I am interpreting this correctly, it seems like with the Fluidigm algorithm the relative signal intensities of the normalized samples may no longer accurately reflect true antibody staining intensity.
Attachments
Normalizer_comparison.pdf
(117.77 KiB) Downloaded 595 times
Adeeb Rahman
Icahn School of Medicine at Mount Sinai, NYC
<<

mleipold

Guru

Posts: 5796

Joined: Fri Nov 01, 2013 5:30 pm

Location: Stanford HIMC, CA, USA

Post Fri May 08, 2015 3:32 pm

Re: Fluidigm vs MatLab normalizer

Hi Adeeb,

One point, from the "Normalizer" data sheet:

On page 7, for the MATLAB tool, it says:
"Note 2: If file concatenation is required, perform this before the normalization procedure."

That's different from how you did Set 2 of your data. I don't know how much of a change it will make, but....

Also, for the MATLAB: is there a reason you didn't just use the bead info as the normalizing reference, for the cell data?


Mike
<<

anitamkant

Master

Posts: 51

Joined: Mon Nov 18, 2013 6:30 am

Post Fri May 08, 2015 7:39 pm

Re: Fluidigm vs MatLab normalizer

Hi Adeeb,
Thanks for your post and we at Fluidigm understand your concerns.
The normalizer tool embedded in the CyTOF software is a global normalizer, while the Matlab method focuses around a particular data set. Please refer to the user guide attached.
We have received a few complaints from our customers and are in process of testing the Normalizer tool. If any bugs are found we will fix them and notify all our customers regarding future steps.
We appreciate your patience.
Attachments
ug13-02_beadnormalization_150501.pdf
(909.77 KiB) Downloaded 555 times
<<

jcvillasboas

Contributor

Posts: 41

Joined: Fri Apr 03, 2015 3:22 pm

Location: Rochester - MN

Post Thu May 21, 2015 2:31 pm

Re: Fluidigm vs MatLab normalizer

I am using a panel that contains markers tagged to 151Eu, 165Ho, 175Lu. I used EQ beads for normalization. Given that those channels are shared by some of the EQ Beads, should I be concerned about the output of my normalization procedure using the Cytobank tool? Is that a valid reason to favor normalization by MatLab for this particular sample instead? Thanks in advance. Best. JC
<<

mleipold

Guru

Posts: 5796

Joined: Fri Nov 01, 2013 5:30 pm

Location: Stanford HIMC, CA, USA

Post Fri May 22, 2015 12:13 pm

Re: Fluidigm vs MatLab normalizer

Hi,

During the acquisition process, you should be collecting Ce140 channel in addition to all your antibody channels. This is unique to the beads, and allows them to be distinguished from your cell events (If you look at Found Cells in FlowJo by plotting Ir vs Ce140, you can see the Beads as Ir-Ce+, Cells as Ir+Ce-, and Bead-Cell doublets as Ir+Ce+).

Because of this, beads and cells can be resolved, and bead-cell doublets removed. Therefore, there shouldn't be an issue with using the channels you mention for markers.


Mike
<<

jcvillasboas

Contributor

Posts: 41

Joined: Fri Apr 03, 2015 3:22 pm

Location: Rochester - MN

Post Fri May 22, 2015 5:01 pm

Re: Fluidigm vs MatLab normalizer

Mike, thank you for the reply. I am indeed collecting Ce140 and resolving the populations as you described. I was just wondering about the normalization procedure itself using a cocktail of beads that contain channels that are shared by metal-tagged antibodies. In other words, will the excess of signal from 151Eu-, 165Ho-, 175Lu-tagged cells interfere in any way with the EQ Beads normalization process using the Cytobank tool?
Next

Return to CyTOF data analysis

Who is online

Users browsing this forum: No registered users and 12 guests

cron