>Mario, > >In principle I agree with you that the computer is more accurate in >performing compensations on the more complex experiments. I am merely >asking this so you can put everyone's (at least my) first question at ease. >Are the computers always right? Randy - If I interpret this correctly, you may be asking two different questions. 1) Is the software correctly implementing the algorithm, or are there errors in the programming? 2) Is the linear model a correct one to use? For (1), I cannot speak for the programmers, but I can say that the algorithm for generating and using the matrix for compensation is very straightforward. There are details, though, which could lead to undesirable results. For example, if the unstained cells in a particular color are piled up against the axis, using the mean as an estimate of the central tendency of the population will give a biased value. So in that case one may need to use medians. The problem with medians, though, is that they are not preserved under linear transformations (i.e. a matrix multiplication) so the result could be a little off quantitatively. I doubt, though, that it could be off enough to visually notice and be able to correct. One problem I think most people have with (2) is that, in using 2-4 colors, we have gotten away with the process of using quadrants. As one uses dyes emitting in the far red, the total amount of light (as opposed to signal to background) is rather low, so that statistical aspects of the measurements become very visible. Because of this, traditional quadrants cannot be used, and attempts to force the data into a form for their use either cannot be done or results in extremely overcompensation. Seeing this happen has convinced users, from my experience, that the software "must be wrong". It is this problem that makes large numbers of colors difficult. Dye-antibody combinations have to be selected so that low expressing antibodies don't appear on colors which statistical aspects will dominate. And we need to routinize different techniques for analysis - the background gating in FlowJo v4 is one, but approaches such as using N-1 colors to set the background on the Nth color are viable and can be utilized in any software system. Marty Marty Bigos Director, Flow Core Gladstone Institute of Virology and Immunology Building 3 SFGH Rm 509 415-695-3832 > Have you found any potential bugs in the >"system" where the computer (and I am not assuming any particular software) >is incorrect because of some unforseen influences in the input data? I know >this is vague, but while I trust in the programmers' abilities, you are >actually doing the most complex experiments with more colors and may have >found some caveats for which we must all be aware. > >Randy T. Fischer >NIH/NIAMS >Building 10, Room 6D57 >9000 Rockville Pike >Bethesda, MD 20892 >(301) 594-3537 >fischer1@mail.nih.gov > >> ---------- >> From: Mario Roederer >> Sent: Friday, June 14, 2002 5:42 PM >> To: Cytometry Mailing List >> Subject: Re: newbie help with software compensation >> >> >> (OK, who's surprised that it took me this long to weigh in?) >> >> David wrote that FCS Express allows you to change compensation values >> with sliders "to see what happens." In fact, most current interfaces >> (like doing compensation on the instrument itself) allow this. This >> "feature" is historic in nature: the fact that instruments have let >> users "manually" adjust compensation since the beginning of (FACS) >> time has caused most people to essentially demand this "feature" in >> compensation software. >> >> I think it's a very bad idea. As I proved in a paper published in >> November's Cytometry, it is IMPOSSIBLE for people to accurately >> compensate based on visual estimations (like graphical displays, dot >> plots, etc.). In other words, unless you are relying on statistics >> and ignoring the graphic, you will not properly compensate by any > > manual (slider or other) approach. >> >> Let me back off a little and reassure you that for most applications >> using FITC, PE, and perhaps a 3rd color, we can come pretty >> close--close enough that we can consider it right. However, as soon >> as you start dealing with more than 3-4 colors, and whenever you are >> dealing with the far red colors (like Cy7PE, Cy7APC, and so forth), >> it is no longer possible to properly compensate visually. It must be >> done based on statistics (e.g., median fluorescence of bright vs. dim >> populations). >> >> In fact, the best approach is to let the software calculate the >> compensation for you--no manual interface at all. Any manual >> interface (letting users adjust) will lead to incorrect compensation. >> >> Now, people will tell me that they need to "tweak" the compensation >> because it doesn't look right. The reason it doesn't look "right" is >> because we don't actually know what "right" looks like! For a better >> explanation of this, see my web pages (particularly, >> <http://www.drmr.com/compensation/>, click on "Quiz") or see the >> manuscript in November's Cytometry, or my letter to the editor in >> Clinical Cytometry (Nov or Dec). >> >> Bottom line: if you are using deep-red fluorescences, or doing more >> than 3-color compensation, then unless you use an automated approach >> to calculating compensation, then you can't get the compensation >> exactly right--this has nothing to do with ability or intelligence, >> it is an result of our inability to correctly estimate central >> tendencies of log-transformed data! (Again, all explained in my >> paper in November, but not explained in the web pages). >> >> Despite users' requests, I have strongly pushed software >> manufacturers NOT to put manual compensation interfaces in their >> programs, because I'd rather users complain than for users to have >> incorrectly-compensated data. Compensation is a very complex >> subject--surprisingly so--and it is far beyond the average user to >> understand all of the ramifications of manual compensation setting. >> I published my first paper on compensation in 1986, and I'm still >> learning & publishing on the topic. Of course, some of you may take >> that to mean that I'm pretty ... um ... thick-headed... but I'd >> rather it be taken to mean that if even an expert won't manually >> compensate his data, perhaps we should all rely on automated >> compensation algorithms and just "believe the computer." >> >> mr >> >> --
This archive was generated by hypermail 2b29 : Sun Jan 05 2003 - 19:26:13 EST