Compensation: It’s not just for pretty pictures

Compensation: Compensation:

It’s not just for It’s not just for pretty picturespretty pictures

Fluorescence Spillover Compensation

• Simple in concept…Correct spillover into different parameters

• Straightforward in execution…Proper settings based on controls

• …But a lifetime to understand?Profound impact on visualizationNonintuitive aspectsSubtle interactions that can be hard to

diagnose

FITC Single Stain Control

450 500 550 600

Argon Laser FL1

FL2

FITC PE

FL2

PE

FL1

FITC

Total signal detected in

FL1

Unwanted signal

detected in FL2

= roughly 15%

True PE = Total FL2 – 15% FL1

FITC Single Stain Control

FITC Compensation Control

PE

- n

o s

tainUnwanted

signal detected in

FL2

roughly 15%

Total signal detected in

FL1

FITC CD3

FL2-15%FL1

Uncompensated Compensated

FITC Compensation Control

FITC CD3FITC CD3

PE

- n

o st

ain

PE

- n

o st

ain

Compensation in 2 colors:Mostly aesthetic

100 101 102 103 104100

101

102

103

104

100 101 102 103 104100

101

102

103

104

CD3

CD4

CD3

CompensatedUncompensated

Accurate identification and enumeration of subsets is still easy in two color experiments

Compensation:Mostly aesthetic

• Accurate discrimination of subsets is possible with uncompensated data

• However, this is true only when the expression of all antigens is uniform on each subset (e.g., CD45 / CD3 / CD4 / CD8)

• Otherwise, it may not be possible to gate on subsets (with current tools)

Compensation for more colors:It’s not just pretty pictures

• Spillover from unviewed measurement channel can alter event positions– without obvious visual evidence (no diagnostic diagonals!)

• Thus, gate positions may depend on unviewed measurement channels and be different for various tubes in a panel

• Separation of populations may require multi-dimensional surfaces (rather than lines of a polygon).

Is this data properly compensated?

100

101

102

103

100

101

102

103

Yes

No

Can’t Tell

Impact of Compensation on Visualization and Analysis of Data

• “Visualization artifacts” lead to:– Manual overcompensation– Incorrect gate settings

• Specific staining controls become essential

What causes this apparent visualization artifact (data spreading)?

Spreading due to Measurement Error

Why do these populations look funny?

101 102 103 104 105

101

102

103

104

105

PE-A: CD8

Cy7

PE

-A:

CD

20

Lymphocytes

Uncompensated Compensated

101 102 103 104 105

101

102

103

104

105

<PE-A>: CD8

<C

y7P

E-A

>:

CD

20

Lymphocytes

Multicolor Compensation

101 102 103 104 105

101

102

103

104

105

<PE-A>: CD8

<C

y7P

E-A

>:

CD

20

Lymphocytes

The Actual Spread

Imperfect Measurement Leads to Apparent Spread in Compensation

100

104

101

102

103

104

SpilloverFluorescence

100

101

102

103

Primary Fluorescence

Uncompensated

100

101

102

103

100

101

102

103

Compensated

7001100

200

(-200)

Why is there a 400-unit spread? Photon counting statistics.

Log Transformation of Data Display Leads to Manual Overcompensation


1 10 100 103 104

1

0.1

10

100

103

104

105

A B C D

Primary Fluorescence1 10 100 103 104

0

-200

-400

400

200

A B C D


Events in channel 0(out of 2446 total):

A: 30B: 475C: 933D: 1190

10.1 10 100 103 104 105

Spillover Fluorescence-100 0 100

Spillover Fluorescence

Overcompensation Cannot Correct Error-induced Spread

1 10 100 103 104

0

-200

-400

400

200

A B C D


1

0.1

10

100

103

104

105

A B C D


1 10 100 103 104

0

-400

-800

1 10 100 103 104

1

0.1

10

100

103

104

105

Primary FluorescencePrimary Fluorescence



Compensation Does NOT Introduce or Increase Error:

Compensation Only Reveals It!

• All measurements have errors; often the error is proportional to the square-root of the measurement.

• The measurement error is present before compensation. Compensation does not increase this error, it does not change it, it does not introduce any more error.

• Compensation simply makes the error more apparent by shifting it to the low end of the log-scale.

Spread of Compensated Data

• Properly compensated data may not appear rectilinear (“rectangular”), because of measurement errors.

• This effect on compensated data is unavoidable, and it cannot be “corrected”.

• It is important to distinguish between incorrect compensation and the effects of measurement errors.

How can we identify undercompensated data?

Diagonals in the data indicate poorly compensated data: but ONLY AT 45˚! Other slopes indicate

nonlinear correlations that have nothing to do with compensation. Visual estimation is very difficult.

Proportional to intensity (spillover)

Proportional to square root of

intensity

Controls

Experimental controls fall into three categories:

Instrument setup and validation (compensation, brightness)

Staining/gating controls (Viability, FMO / ABO)

Biological

Experimental controls are necessary to properly analyze and interpret data.

Instrument Setup Controls

Typically, fluorescent beads… with a range of fluorescences from “negative” to very bright.

Use these to validate:•Laser stability & focusing•Filter performance•PMT sensitivity (voltage)•Fluidics performance•Daily variability

Consider setting target fluorescence values for beads during alignment: this allows for greatest consistency in analysis (gating) between experiments.

Compensation Controls

Single-stained samples…must be at least as bright as the reagent you are using in the experiment!

Can use any “carrier”, as long as the positive & negative populations for each control have the same fluorescence when unstained:

Cells (mix stained & unstained)Subpopulations (CD8 within total T)Beads (antibody-capture)

One compensation for every color… and one for each unique lot of a tandem (Cy5PE, Cy7PE, Cy7APC, TRPE)

Compensation Controls

Note: you can mix bead controls on some colors and cell controls on other colors.

However – the negative control for each type must match the positive control: do not use a “universal negative” unstained beads (or cells) for all controls!

Collect sufficient events to precisely estimate fluorescence medians:

Beads: At least 5-10,000Cells: At least 20-50,000

Using Beads to Compensate• Exact reagent used in experiment

• 100% positive (not ever rare)

• Small CV, very bright (precise fluorescence measurement)

• Sonicate weekly to reduce aggregates

• Some reagents may not work (Ig, non mouse, too dim, EMA/PI) – for these, use cell-based compensations

Why Bright Comp Controls?

101 102 103 104 105

FITC-A

101

102

103

104

105

Cy7PE-A

Autofluorescence

FITC spillover into Cy7PE (1%)

Unstained cells

Bright cells

Dimmer cells

Estimating a low spillover fluorescence accurately is impossible (autofluorescence).

Therefore, compensation is generally only valid for samples that are duller than the compensation control.

Insufficiently-Bright Comp Control Is …. Bad!

Note that either under- or over-compensation can result from using comp controls that are too dim!

Compensation of Multicolor FACS Data

• It is impossible to set proper compensation using visual guides (dot plots, histograms).– Use statistics (medians of gated cells)– Use automated compensation tools

• Antibody-capture beads are an excellent way to set compensation!

• Compensation controls must be matched to your experiment and at least as bright as any of your reagents.

Some Examples of Problems

• The following four examples illustrate some types of problems that can be occur related to compensation.

• In each case, compensation itself is not the problem: there is an underlying reagent, instrumentation, or analysis problem.

• However, the manifestation of this problem is an apparent incorrect compensation!

Good Instrument Alignment Is Critical!

Day 1 Day 2

Uncompensated

PE

TR-PE

While the amount of compensation did not differ, the

measurement error (correlation) decreased leading

to much better visualization of the

population!

Compensated

CFSE: A Special CaseCells stained with CFSE-DA, FITC-CD8, or unstained were collected uncompensated.

101 102 103 104 105

10 1

10 2

10 3

10 4

10 5

101 102 103 104 105

10 1

10 2

10 3

10 4

10 5

PE Cy5PE

FITC FITC

CFSE has a slightly redder spectrum than FITC… must use CFSE as a comp control!

Note that this exacerbates the higher “IL4+” gate required for CD8 cells.

The undercompensation would not have been detected except by looking at the APC vs. Cy7APC graphic…

Fix/Perm Changes Cy7APC Compensation Requirement

The longer Cy7APC is in fixative, the more it “falls apart”, leading to

more APC signal

Different lots of tandems can require different compensation!

TR-PE reagent 1Median = 21,100

TR-PE reagent 2Median = 8,720

PEMedian = 484

PEMedian = 698

Compensation Required(∆PE / ∆TRPE)

2.3%

8.0%

Wrong TR-PE comp control

Compensating with the wrong TRPE

Right TR-PE comp control

Staining Controls• Staining controls are necessary to identify

cells which do or do not express a given antigen.

• The threshold for positivity may depend on the amount of fluorescence in other channels!

Staining Controls• Unstained cells or complete isotype control

stains are improper controls for determining positive vs. negative expression in multi-color experiments.

• The best control is to stain cells with all reagents except the one of interest.

FMO Control“Fluorescence Minus One”

Also known as “ABO” (all-but-one)

Identifying CD4 cells with 4 colors

100 101 102 103 104100

101

102

103

104

105

100 101 102 103 104 100 101 102 103 104

Unstained Control FMO Control Fully Stained

PE

FITC

FITCPE

Cy5PECy7PE

––––

CD3–

CD8CD45RO

CD3CD4CD8

CD45RO

Isotype BoundsFMO Bounds

PBMC were stained as shown in a 4-color experiment. Compensation was properly set for all spillovers

Complex Interactions in Compensation

100 101 102 103 104100

101

102

103

104

100 101 102 103 104 100 101 102 103 104

100

101

102

103

104

10Unstained Control FMO Control Fully Stained

PE

FITC

Correct Compensation

Incorrect Compensation

The same data is shown with correct or wrong Cy5PE->Cy7PE comp setting. Note that neither of these channels is shown here!

100 101 102 103 104100

101

102

103

104

105

100 101 102 103 104 100 101 102 103 104

PE

FITC

Unstained Control FMO Control Fully StainedFITC

PECy5PECy7PE

––––

CD3–

CD8CD45RO

CD3CD4CD8

CD45RO

Isotype Bounds

FMO Bounds

FMO controls aid even when compensation is improper

Incorrect Cy5PE into Cy7PE compensation

FMO Controls

• FMO controls are a much better way to identify positive vs. negative cells

• FMO controls can also help identify problems in compensation that are not immediately visible

• FMO controls should be used whenever accurate discrimination is essential or when antigen expression is relatively low

Quad-Gates of the Future

Log Fluorescence #1

DoublePositive

SinglePositive

SinglePositive

DoublePositive

SinglePositive

SinglePositive

LogFluor.

#2

Present

Compensation & Data Visualization

These “new” distributions are much more frequently seen nowadays, with the use of red dyes (Cy7PE, Cy7APC) and with more precise instruments.

Some users have questioned the correctness of these distributions, leading some manufacturers to try to provide “corrections”.

However, this cannot be “corrected”–what is needed is education!

100 101 102 103

100

101

102

103

Spread ofpositives

Events piling up on axes

Compensation Offset???• At least two FACS data analysis software

programs offer a “feature” to turn on a “compensation offset”–to try to make data look more like what we have expected.

• The term “Compensation Offset” simply means that random noise is added to the data to hide the “spread” in compensated data.

• This “feature” reduces sensitivity!

Compensation Offset???

• Artificially increase error does make distributions rectangular again--but at a significant expense!– Low level antigen expression will be lost– In >4 color experiments, problems multiply!

Compensation Offset???• An even more insidious problem is that overcompensation

becomes impossible to detect:

•Turn OFF this “feature” in your software10 -1 100 101 102 103

10-1

100

101

102

103

0 100 101 102 103

0

100

101

102

103

0 100 101 102 103

0

100

101

102

103

10 -1 100 101 102 10310 -1

100

101

102

103

10 -1 100 101 102 10310

-1

100

101

102

103

10 -1 100 101 102 10310 -1

100

101

102

103

Correct

5% over-compensated

Standard Transformed Offset

Is There A Solution?• The spread in compensated data is

unavoidable (basic physics)

• Can we visualize data so that the distributions are more intuitive?

• Nearly all immunophenotyping data is shown on a logarithmic scale… why?– Dynamic range of expression (4 logs)

– Often, distributions are in fact log-normal

Alternatives to a Log Scale

• Compensation reveals a linear-domain spreading in the distribution.

• This is most obvious at the low end of fluorescence, because the measurement error is small compared to bright cells.

• Can we re-scale the low end of the fluorescence scale to effect a different compression in this domain?

• What about negative values?– Remember, this is just a fluorescence from which we

subtract an estimated value with measurement error

“Bi-exponential” Scaling


1 10 100 103 104

1

0.1

10

100

103

104

105

A B C D


0

-200

-400

400

200

A B C D


S p i l l o v e r

F l u o r e s c e n c e

1 1 0 1 0 0 1 03

1 04

1 0

1 0 0

- 1 03

P r i m a r y F l u o r e s c e n c e

1 04

1 05

- 1 0

- 1 0 0

1 03 Positive Log

Negative Log

Linear

Wayne MooreDave Parks

“Bi-exponential” Transformation Makes Compensated Data More Intuitive

101 102 103 104

101

102

103

104

CD8

CD4

101 102 103 104

101

102

103

104

CCR5

CD103

0 103 104

0

103

104

CD103

CCR5

0 103 104

0

103

104

CD8

CD4

Negative Log Scale:

Events with measured

fluorescence < 0

Linear Scale: Compression of the amount of visual space

devoted to this range

Gated for CD3+ Lymphocytes;

Stained for CD3, CD4, CD8, CCR5, CD103, and 7 other

reagents

Question: What is the expression of CCR5 and

CD103 on CD4 Lymphocytes?How many events are on

the axis?

LOTS!

Transformed Distributions

Result:

No events hidden on the axes

Populations are visually

identifiable

“Bi-exponential” Transformation Makes Compensated Data More Intuitive

Only changes the visualization of data

• Does not affect gating or statistics• Cannot change the overlap (or lack thereof) of two

populations.

Supports the basic goal of graphing data: showing it in an intuitive, aesthetic manner

Note: the transformation is complex: it is different for each measurement channel and compensation matrix, and depends on the autofluorescence distribution. However, these parameters can be automatically selected by the software.

0 103 104 105

0

10 3

10 4

10 5

101 102 103 104 105

10 1

10 2

10 3

10 4

10 5

101 102 103 104 105

10 1

10 2

10 3

10 4

10 5

0 103 104 105

0

10 3

10 4

10 5

CD4

CD8

Median

Transformed

Transformation Confirms Compensation

SummaryCompensation is straightforward…. But:• Interpretation of compensated data is complicated by

“spillover spreading” arising from measurement errors• Proper compensation requires careful attention to

compensation controls: bright, matched reagents & staining conditions

• Proper analysis of compensated data requires careful attention to staining controls

• Biexponential transformation of displays is critical for analysis and interpretation of compensated data

• Bagwell & Adams: Ann NYAS 677:167 (1993)

Software compensation

• Stewart & Stewart: Cytometry 38:162 (1999)

4-color compensation, pitfalls

• Roederer: Current Protocols in Cytometry

Protocol, explanation, discussion

• Roederer: Cytometry 45:194 (2001)

Visualization artifacts, FMO controls

• http://www.drmr.com/compensation/

Compensation Resources

Compensation: It’s not just for pretty pictures

Documents

Transcript of Compensation: It’s not just for pretty pictures