The Daily Dongle

biexponential cramps

Every once in a while, Version 8 of FlowJo does a funny thing to some of the parameters when biexponential scaling is enabled. It looks something like...

notice how the X parameter is squished from 0 to 10^4 above.

FlowJo has two modes for calculating biexponential scaling - global (all parameters get the same transform) or individual (each parameter gets it's own scale computed, as in the screenshot above).

The unfortunate aspect of this calculation is that once "individual" scaling is calculated, it cannot be reversed. Even the "global" scaling is a one-time-deal - once "on", it cannot be changed (except in the case form Global to Individual).

What determines whether flowjo uses Individual or Global is the method used to invoke the scaling:

- If you do it via Platforms menu, Compensate, define biexponential.. - it will always be calculated uniquely per channel.
- If you do it in your global preferences (flowjo >> preferences >> workspace tab >> define... button) then all channels are treated identically.

Let's do the global scaling exercise. Bring up the lowjo >> preferences >> workspace tab >> define... panel:

notice in the lower-right, there are three variables - these control the biexponential function.

leave the top two fields as they are, but experiment with Width Basis. With basis is the number of channels squished into linear space around 0.  I recommend trying values from -50 to -200 or so.

Here is -500 (WB too high, too squished):

Here is -100 (more like it):

-50: not bad!

-10 (default, too "low to the ground")

This titration or Width Basis value hopefully demonstrates what the variable does. Here the best looking value was -50, but slightly lower value would probably be OK too, just not too close to -10.

Population Comparison

Your Mac version has a population comparison platform, which returns differences between my sample and control using four different algorithms. How can I compare them? Which one is better?

This question demonstrates a need for better documentation of this Platform. The short answer is - there is no way to cross-compare the results of these algorithms. I will discuss them briefly below..

The four methods for sample comparison are:

SED, Overton, K-S and T(x).

SED algorithm was developed and published by Bruce Bagwell, of Verity Software House. For more details on this please contact Dr. Bagwell.

Overton's subtraction algorithm goes through each bin of two histograms and subtracts the control from the exp. The resulting differences are summed and divided by total number of events - returning % difference. This algorithm does not provide any confidence metric for the value it returns.

K-S test is a very complicated math model developed to study the confidence in two distributions being significantly different. More on that here.

Finally, the T(X) method developed by Mario Roederer provides a calculated T value as well as the maximum possible T value. T(x) is a statistic which not only provides an indication of the probability with which two distributions are different, but simultaneously provides a metric by which multiple distributions can be ranked.

Users should be cautioned that K-S and T(x) will very readily label two samples as MATHEMATICALLY different. It's up to the user to determine whether this difference is BIOLOGICALLY meaningful.

T(X) test makes this a little easier by providing a maximum possible T value. T(X) = 4 is the same significance whether the maximum value is 50 or 50,000 -- it implies that you are 4 std deviations away from expected.  The reason the max T(X) can range is that with more events (or more bins) you can achieve greater precision -- thus, greater significance.  For example, if you do a wilcoxon rank test on two groups of 3 events, the most significant p value you can achieve is about 0.03 -- when there is no overlap between the two groups.  With 4 events, the most significant p value you can achieve is about 0.004.  But a p value of 0.05 is the same significance whether the best significance you can achieve 0.03, 0.004, or 10^-44.  Similarly with T(X) -- a T(X) = 4 is the same; the max(T(X)) just tells you what the maximum significance could have been.

In sum, don't use software to find differences in your flow data. Use it to substantiate and quantitate the differences you already know about.

Powerpoint 2008 Mac - no love for PICT

FlowJo graphs/layouts are no longer ungroupable in the new version of 2008. If you find this functionality useful, stick to 2004 version.

I found that 2004 version's ungrouped slides are still ungrouped in PPT 2008.

Randomization

From tech support:

How do I process a random sampling of events form my file/gate?

There is a way to randomize, but only in version 8 of the software.  It involves exporting a certain number of events - either from the top-level node (ungated file) or one of it's subpopulations (gates).  You can select files/gates one at a time, click Workspace menu, Export. There, you'll see a number field with current event count. You can modify it, and the sampling will be "random" according to our docs, but I don't know which random algorithm is used.

New Mac Pro's and FlowJo V8

With the new crop of Macs on Apple Store, I thought I'd revisit my hardware recommendations for an analysis workstation.

Data Input:
I saw no benefits going with expensive disks - if you can get the regular sata-2 disks @ 7200RPM, that's plenty. I did not see a significant increase between that (~60MB/s thruput) and RAMdisk (~600MB/s average thruput) . The biggest slowdown in analysis is loading data via network. If you can avoid that, or pre-cache this step before analysis, it will improve your overall efficiency. Analyzing even on a gigabit network makes the analysis 2x slower vs. a standard local disk.

Network:
I have not yet met a network/server configuration that fully utilizes the capacity of a single quad-core mac. If you are serious about performance, don't analyze over a network.

CPU:
Our code has been optimized for up to 4-simultaneous processors. We saw no benefit going from 4- to 8-core workstation, thus, I would recommend getting a quad-core desktop instead of the octa-core and save ~2100USD.(almost enough to buy a second 4-core mac.).

Screens:
I would definitely recommend a big screen. Two 24 inch screens will cost less and produce more area than a single 30inch. If budget is of no concern, get two 30's :) If budget is of concern, get non-Apple displays. They work just as well.

RAM:
FlowJo doesn't use more than 2GB, being a 32 bit application. However, other programs may compete for this resource, and thus it is a good idea to have double that - about a gig for OSX and it's processes, and perhaps another gig for user apps like mail, excel, web browser, etc..  If you're on a tight budget, consider a non-Apple memory vendor like Crucial.