A simple correction for cell autofluorescence for ... - Wiley Online Library

Cytometry Part B (Clinical Cytometry) 70B:91–103 (2006)

A Simple Correction for Cell Autofluorescence for Multiparameter Cell-based Analysis of Human Solid Tumors Charles A. Smith, Agnese Pollice, David Emlet, and Stanley E. Shackney* Laboratory of Cancer Cell Biology and Genetics, Department of Human Oncology, Allegheny Singer Research Institute, Allegheny General Hospital, Pittsburgh, Pennsylvania

Background: Corrections that have been proposed to minimize the unwanted contribution of cell autofluorescence to the total fluorescence signal often require either specialized instrumentation or the sacrifice of a data channel so as to perform a measurement that can be used to correct for autofluorescence in individual cells. Here we propose a simple cell by cell correction for autofluorescence that is suitable for multiparameter laser scanning cytometry (LSC) studies in human solid tumors that relies on the ratio of mean autofluorescence to mean total cell fluorescence (mean Flauto/mean Fltotal). This approach assumes a correlation between the autofluorescence component and the total signal in individual cells. This correction does not require specialized instrumentation, and does not sacrifice a data channel in multiparameter studies. A potential disadvantage is that errors may be introduced by the assumption of a correlation between the two components of the total fluorescence signal in individual cells in samples in which no such correlation exists. Methods: Distributions of cell autofluorescence and total Her-2/neu cell fluorescence were obtained separately by LSC in three human breast cancer cell lines and in three samples of primary human lung cancer. In the breast cancer cell lines, autofluorescence measurements and Her-2/neu measurements were also obtained on the same cells. Results: We show that there is a partial correlation between autofluorescence and total Her-2/neu/FITC fluorescence in individual cells in the three breast cancer cell lines. We also show that the results of a ratio-based autofluorescence correction agree with those based on a true cell by cell correction. Computer simulation studies suggest that in samples with no correlation between the autofluorescence component and the true probe/dye fluorescence component, the ratio correction produces robust estimates of the mean true fluorescence signal, with relatively small but systematic underestimates of the coefficient of variation of such measurements under conditions commonly encountered in the measurement of human solid tumors. Conclusions: A simple cell by cell correction for autofluorescence based on the ratio of mean Flauto to mean Fltotal can be applied in cell samples in which there is a correlation between cell autofluorescence and true probe/dye fluorescence in individual cells. In cell samples that lack this correlation, or in which it is not known whether such a correlation exists, this correction can be used with the reservation that there is a systematic but relatively small underestimation of the degree of variability of the measurements. q 2006 International Society for Analytical Cytology Key terms: autofluorescence; immunofluorescence; cell-based measurements; solid tumors

The unwanted contribution of cell autofluorescence to cellular immunofluorescence measurements in the green and orange portions of the energy spectrum is a well recognized problem in quantitative cytometry. Various approaches have been developed to correct for cell autofluorescence. Mathematical models have been developed for estimating true fluorescence histograms from autofluorescence and total fluorescence histograms (1). Approaches to cell by cell corrections for autofluorescence include phase-gated

q 2006 International Society for Analytical Cytology

Grant sponsor: NIH; Grant number: CA83204; Grant sponsor: PADOH; Grant number: ME01-738. *Correspondence to: Stanley E. Shackney, Allegheny General Hospital, 320 East North Avenue, Pittsburgh, PA 15212, USA. E-mail: [email protected] Received 8 December 2004; Accepted 21 October 2005 Published online 2 February 2006 in Wiley InterScience (www. interscience.wiley.com). DOI: 10.1002/cyto.b.20090

92

SMITH ET AL.

methods that are based on the short lifetime of autofluorescence signals in comparison with those of commonly used organic fluorescent dyes (2,3), or methods that treat autofluorescence as a separate color in multiparameter analysis. The latter approach includes methods that rely on a separate autofluorescence-related measurement obtained simultaneously on each cell (4–6). Analysis of spectral imaging data can resolve the contributions to an image of individual fluorogenic or chromogenic components (7–9), and this approach can be used to identify and reduce background autofluorescence (9). Cell by cell corrections for fluorescence noise, of which autofluorescence is a major component, are highly desirable for quantitative multiparameter cellular fluorescence studies. However, the benefits of a cellbased autofluorescence correction can be offset by the requirement of specialized instrumentation (lifetimegated corrections, spectral imaging), or by a requirement to use a data channel for the collection of a cell autofluorescence measurement that might otherwise be used to collect data on other cell constituents. Agents such as trypan blue quench green autofluorescence but emit at 650 nm and above, sacrificing measurements in the red and far red portion of the spectrum (10), and therefore offer no real advantage here. Crystal violet, a quenching agent that has been applied to the study of autofluorescent alveolar macrophages (11,12), also produces high levels of fluorescence at 650 nm and above (Pollice A, unpublished observations). Thus, for older flow cytometry and laser scanning cytometry instruments, where the number of available data channels may be limited, the use of an autofluorescence correction that preserves multiparameter data channels would be highly useful. When the average cell autofluorescence signal is low in comparison with the average total fluorescence signal, and variations in the autofluorescence signal from cell to cell are small, a simple cell by cell subtraction of an average autofluorescence signal might suffice. Among the examples shown in Figure 1, the human breast cancer cell lines JC-1939, MCF-7, and SKBR-3 (panels A-C) appear to meet these criteria. In Figure 1, regression lines for stained cells (black) and autofluorescent cells (gray) were fitted to log transforms of the data, and plotted both in the linear domain and log domain for each sample. However, at least two of the samples obtained from primary human lung cancers (panels E and F) unstained cells exhibit high levels of autofluorescence that account for over 50% of the total fluorescence signal of the stained cells. Such large variations in levels of cell autofluorescence are sample-related, and are thought to arise from high cellular levels of such molecular species as tryptophan, NAD(P)H, and various flavoproteins (13–16). The spectral properties of these mixtures of cell constituents are complex, and the degree to which the autofluorescence signal can interfere with the measurement of the true signal in various regions of the spectrum can be substantial in some samples. In this article, we explore the consequences of using a simple cell-based autofluorescence correction based on the ratio of mean cell autofluorescence to total cell fluo-

rescence in each region of the spectrum, obtained by performing measurements on separate aliquots of stained and unstained cells from the same sample. This correction implicitly assumes that the levels of the true probe/dye-generated fluorescence signal, the autofluorescence signal, and total cell fluorescence signal in each cell are correlated. Is this a reasonable assumption? We show that in the breast cancer cell line data shown in Figure 1, there is a correlation between the autofluorescence signal and the level in each cell of the total immunofluorescence signal generated by a FITC-conjugated antibody to Her-2/neu, a cell constituent that is of interest in human solid tumors. In practice, how well does the performance of our simple correction compare with that of a true cell by cell correction for cell autofluorescence? We show that the results obtained using the mean ratio correction do agree well with the results obtained using an independent method for true cell by cell correction for autofluorescence, even in samples with extensive overlap between the autofluorescence and total fluorescence measurements. Of course, one cannot assume a general correlation between cell autofluorescence and true cell fluorescence in most cell samples. We therefore used computer modeling studies to explore the consequences of applying our simple correction when, in fact, cell autofluorescence and true cell fluorescence are not correlated. Our modeling studies suggest that under most circumstances likely to be encountered in practice, the use of an average autofluorescence ratio correction still produces reasonable approximations for the true values. MATERIALS AND METHODS Preparation and Fixation of Cell Lines Breast cancer cell lines SKBR-3, and MCF-7, obtained from American Type Culture Collection (Rockville, MD) and cell line JC-1939, which was developed in our laboratory, were harvested in log phase of growth with trypsin-EDTA (GIBCO-BRL), and treated with cold (48C) 5 mM dithiothreitol (DTT) (Sigma, St. Louis, MO) for 15 min at room temperature to reduce clumping. Cell suspensions were fixed in paraformaldehyde/methanol, as previously described (17). Tumor Samples Human non-small cell lung tumor samples were obtained with informed consent under protocols approved by the Institutional Review Board of Allegheny General Hospital, Pittsburgh, PA. Freshly obtained primary lung cancer samples were mechanically disaggregated in Hanks’ Balanced Salt Solution (HBSS) (Mediatech, Herndon, VA), filtered with 200 m nylon mesh to eliminate large tissue clumps prior to fixation (Small Parts, Miami FL), washed with HBSS, treated with DTT, and fixed in paraformaldehyde/methanol. Autofluorescent Cell Samples (DNA Staining Only) Aliquots of 2 104 cells from each of the cell lines and primary tumors were filtered through 64 m nylon mesh

CELL AUTOFLUORESCENCE CORRECTION

93

FIG. 1. Dot plots showing the relationships between total Her-2/neu/FITC cell fluorescence (black dots and curves) and cell DNA content, and cell autofluorescence in the FITC data channel (gray dots and curves) and cell DNA content in, (A1) JC-1939 breast cancer cells. Data plotted in the linear domain. Mean ratio of auto fluorescence to total fluorescence is 0.29. Regression lines were fitted to log transforms of the data, and plotted in the linear domain, showing that increases in mean levels of Her-2/neu taper off at high cell DNA content levels. (A2) JC-1939 breast cancer cells as in A1, with data plotted in the log domain. Regression lines were fitted to the log transforms of the data. (B1 and B2) MCF-7 breast cancer cells. Mean ratio of auto fluorescence to total fluorescence is 0.25. Data formats as in (A). (C1 and C2) SKBR-3 breast cancer cells. Mean ratio of auto fluorescence to total fluorescence is 0.12. Data formats as in (A). Three primary lung cancer samples, (D1 and D2) BE 3449, mean ratio of auto fluorescence to total fluorescence, 0.58. Data formats as in (A). (E1 and E2). RD 3900, mean ratio of auto fluorescence to total fluorescence, 0.73. Data formats as in (A). (F1 and F2) RB 3905, mean ratio of auto fluorescence to total fluorescence, 0.80. Data formats as in (A). Units are arbitrary LSC instrument units. All data were obtained at constant photomultiplier tube gain settings and under identical protocols. [Color figure can be viewed in the online issue, which is available at www.interscience.wiley.com.]

(Small Parts, Miami, FL), centrifuged at 200 g for 2 min in phosphate buffered saline (PBS), and resuspended in 100ul of 40 ,6-diamidino-2-phenylindole (DAPI), (Sigma, St. Louis, MO) 1 ug/ml in v/v 1:1 glycerol:PBS. HybriWell chambers (22 22 0.15 mm3) (Schleicher & Schuell, Keene, NH) were affixed to pre-cleaned glass microscope slides. 100 ul of cell suspension was pipetted through one port on the HybriWell surface while allowing the air to

escape through the other port, to produce a uniformly spread cell suspension without air bubbles. MUltiparameter Immunofluorescence Staining Monoclonal antibody immunospecific for Her-2/neu, Clone CB11 (Novocastra Laboratories, New Castle upon Tyne, UK) that was directly conjugated to FITC, was

94

SMITH ET AL.

Table 1 Lasers, Emission Lines, and Filters for Four-Color LSC Emission line (nm)

Laser Diode laser (Power Technology Inc, 30 mW) Argon-ion (Cyonics Uniphase Model 2014a-20SL, 20 mW)

Fluorochrome DAPI

463/39-nm band pass

488

FITC

530/30-nm band pass 555-nm dichroic long pass 580/30-nm band pass 605-nm dichroic long pass 605-nm long pass full mirror

CY3 or PE HeNe (Cyonics Uniphase, 5mW)

Filters

405

633

used for direct immunostaining at a 1:10 dilution. Saturating dilutions were determined previously by titration. Following a 1 h incubation at room temperature in the dark, cell suspensions were washed 1 with PBS and resuspended in 100 ul of DAPI. The cell suspension was placed on slides with HybriWell chambers, as described for DNA staining. Cytometry Fluorescence measurements were made using a laser scanning cytometer (LSC) (CompuCyte, Cambridge, MA) with the WinCyte (version 3.6) program, equipped with an air-cooled violet diode laser emitting at a wavelength of 405 nm, an air-cooled argon laser emitting at a wavelength of 488 nm, and a HeNe laser emitting at a wavelength of 633 nm. Additional technical details on the lasers and the filter set are provided in Table 1. Cell by cell Correlations of Autofluorescence and Her-2neu Immunofluorescence Unstained JC-1939, MCF-7, and SKBR-3 cells were deposited on pre-cleaned glass microscope slides using a Cytospin centrifuge (Shandon, Pittsburgh, PA) at a concentration of 2 104 cells, stained with 200 ul of DAPI and cover-slipped. Cell immunofluorescence in the FITC (green) channel was measured by LSC (see later). The coverslip was removed by washing with PBS. The slide was then stained with FITC-conjugated anti Her-2/neu antibody CB11, and total cell fluorescence was measured in the same cells. The x,y coordinates of the cell autofluorescence measurements were matched to those of the Her-2/neu immunofluorescence using a computer program written by two of the authors (C.A.S. and S.E.S.). Duplicate slides were evaluated in each cell line. A fixed scan area of 1.6 108 mm2 centered on the HybriWell chamber was used with all samples. All measurements were performed using a 20 objective lens. DNA was used as the contouring parameter with a threshold of 700 and 30 pixels added to the nuclear contour. Data Management Cell aggregates were identified and removed from the listmode data file, using a recently described method that combines a function provided by CompuCyte with an algorithm based on concomitant cell by cell measure-

CY5

ments of cell nuclear area, nuclear perimeter, and an LSC-generated nuclear texture parameter (18). Reference cells (JC-1939) in which levels of Her-2/neu (87,000 molecules/cell) were quantitated independently by ELISA in units of molecules/cell were included as a reference standard for labeled cells. On the basis of the mean value of this absolute reference standard, the intracellular levels of Her-2/neu could be calculated for all cells in a concomitantly run tumor sample, and expressed in units of molecules per cell. Development and Application of the Simplified Cell by Cell Correction for Autofluorescence The problem of autofluorescence is greatest in the green portion of the energy spectrum, corresponding to the region of FITC fluorescence detection. The autofluorescence signal in the orange region of the spectrum (PE/ CY3) is typically one third of that found in the green, and the autofluorescence signal produced by excitation at 633 nm by the HeNe laser in the red region of the spectrum is negligible. For illustrative purposes, we use fluorescence-based measurements obtained by LSC of Her-2/neu (FITC). We have chosen data sets from six human tumor cell samples to show the spectrum of autofluorescence-related problems that are likely to be encountered in the performance of cell-based immunofluorescence measurements in clinical studies (Fig. 1). These consist of data from three breast cancer cell lines, JC-1939, MCF-7, and SKBR-3, and data from single cell suspensions obtained from three primary human lung cancers. The breast cancer cell lines all have relatively low levels of cell autofluorescence, and exhibit total cell Her-2/neu fluorescence values ranging from low levels in JC-1939 to very high levels in SKBR-3 (Fig. 1). The lung cancer cells exhibit higher levels of cell autofluorescence than the breast cancer cell lines, and exhibit varying levels of total cell fluorescence. In this article, we will focus primarily on the correction for autofluorescence in the FITC channel, where the magnitude of the autofluorescence problem is greatest. However, the same type of correction is applicable to other regions of the spectrum as well. At first glance, the substantial overlap between the total cell fluorescence values and cell autofluorescence values obtained separately on cells from the same population in the samples shown in Figure 1, panels E and F, might raise questions regarding the reliability and useful-

95


ness of quantitative cell-based information that can be extracted from such samples. However, it may be helpful to keep in mind that in cell populations in which the autofluorescence component represents a large proportion of the total fluorescence signal within most cells, the magnitude of the true fluorescence component of the signal must, in turn, be relatively small in these cells. The smaller the true signal, the less critical the variance in its measurement might be. We will examine these issues more systematically below. Assumptions Regarding the Properties of the Distributions of Cell Autofluorescence, True Probe/Dye Fluorescence, and Total Cell Autofluorescence We have observed in a wide range of samples that the variances of both cell autofluorescence and total probe/ dye fluorescence measurements of various cell constituents increase progressively with increasing cell nuclear area and cell DNA content by LSC, as is apparent by inspection of the data in Figure 1. In Table 2, the means and standard deviations are shown for Her-2/neu content and autofluorescence separately in cells with G1, S, and G2M þ higher DNA contents for SKBR-3 cells. The standard deviation (SD) increases with increasing mean level of Her-2/neu. As a consequence, the coefficient of variation (CV) undergoes little change with increasing cell Her-2/neu content. This supports the fitting of the regression of the fluorescence signal on cell DNA content to log transforms of the data in Figure 1. The overall frequency distributions for the autofluorescence and total signals in SKBR-3 cells are shown in Figure 2. Units are expressed as Her-2/neu molecules per cell, by referencing arbitrary LSC fluorescence readout units to the mean cell fluorescence values of a concomitantly run breast cancer cell line in which mean Her-2/ neu levels were determined independently by ELISA. The mean fluorescence level of the reference cell line was taken as the reference for the mean Her-2/neu level determined by ELISA. The coefficients of variation (linear) of the autofluorescence and total fluorescence signals are similar, despite a nearly 10-fold difference in mean values, as are the standard deviations of the corresponding log distributions, further supporting a direct relationship between signal variance and magnitude of the signal. It is apparent from Figure 2 that both the autofluorescence and total fluorescence signals have more symmetrical distributions in the log domain than in the linear domain, as noted by others (1). This suggests further that the use of log normal distributions would be appropriate in our modeling studies (see later). Table 3 lists the log mean values and standard deviations for total cell fluorescence and autofluorescence for the cell samples shown in Figure 1. Units in Table 3 are expressed as Her-2/neu molecules per cell. Of note, the standard deviations in Table 3 ranged between 0.11 logs and 0.34 logs among all samples, and the standard deviation of each of the autofluorescence distributions was generally similar to or less than that of

Table 2 CV of the Autofluorescecne and Her-2/neu Measurements in SKBR-3 Cells in Relation to Cell DNA Content

G1 Mean (autofluorescence) SD (autofluorescence) CV (autofluorescence) Mean Her-2/neu (stained) SD (stained) CV (stained)

Cell DNA content S

G2Mþ

749,900

1,065,900

1,404,600

186,200

371, 800

423,500

24.8%

34.9%

30.2%

3,434,000

4,968,900

6,553,300

1,569,100 45.7%

2,156,600 43.4%

3,511,600 53.6%

the corresponding total cell fluorescence distribution for the same sample. Cell by Cell Correction for Autofluorescence It is apparent from Figure 1 that cell autofluorescence increases with cell DNA content. It may be that the autofluorescence signal increases with total cell mass (see ref. 1), for which cell DNA content might be a reasonable surrogate measurement. Whatever the basis for the increase in cell autofluorescence with increasing cell DNA content, when the autofluorescence component represents most of the total signal (as in Figures 1E and 1F), the magnitude of the autofluorescence component and that of the total fluorescence signal would be expected to track together in individual cells. When the cell autofluorescence signal and its variability from cell to cell are small enough that they are likely to have little effect on the total signal, an increase in total cell fluorescence with increasing cell DNA content (e.g., Figs. 1A–1C) would imply a separate correlation between the remaining probe/dye fluorescence component of the total signal, and cell DNA content. In this setting, since both cell autofluorescence and total cell fluorescence increase with increasing cell DNA content (Figs. 1A–1C), one might infer that there is a cell by cell correlation between the cell autofluorescence component and the probe/dye fluorescence component of the total signal. Thus, in these samples, a cell by cell correction for autofluorescence that is based upon a fixed ratio between the autofluorescence signal and the total signal might be justified. The correction itself is developed as follows: For each cell, True cell fluorescence ðFltrue Þ ¼ Total cell fluorescence ðFltotal Þ cell autofluorescence ðFlauto Þ

ð1Þ

Fltotal is measured in each cell, but Flauto, the autofluorescence component in each cell is unknown. Then, Fltrue ¼ Fltotal Flauto Fltotal/Fltotal Fltrue ¼ Fltotal ð1 Flauto =Fltotal Þ

(2)

Assuming a constant Flauto/Fltotal ratio from cell to cell that is represented by mean Flauto/mean Fltotal, (using a

96

SMITH ET AL.

FIG. 2. (A1) Frequency distribution of cell autofluorescence in SKBR-3 cells, plotted in the linear domain (A2) Frequency distribution of cell autofluorescence in SKBR-3 cells, plotted in the log domain (B1) Frequency distribution of total Her-2/neu/FITC fluorescence in SKBR-3 cells, plotted in the linear domain. Note that the X-axis scale is 10 times greater than that of the linear autofluorescence distribution shown in (A1). (B2) Frequency distribution of total Her-2/neu/FITC fluorescence in SKBR-3 cells, plotted in the log domain. Units have been converted to molecules per cell, by referencing cell fluorescence levels to the mean cell fluorescence value of a concomitantly run breast cancer cell line in which mean Her-2/neu levels were determined independently by ELISA.

separate concomitantly run unstained slide to determine mean cell autofluorescence), and substituting mean Flauto/mean Fltotal for Flauto/Fltotal in Eq. (2) Fltrue ¼ Fltotal ð1 mean Flauto =mean Fltotal Þ

ð3Þ

Comparison of Results of the Mean Flauto/mean Fltotal Correction with the Results of True Cell by Cell Correction To compare the results of the mean Flauto/mean Fltotal correction with those obtained using a true cell by cell correction on the same cells, we made use of an approach published previously by Alberti et al. (6). By their method, an average ratio of green to red autofluorescence is established in unstained cells, and cell by cell measure-

ments in the red channel are then used to estimate the contribution of autofluorescence to the total signal in the green channel in each cell in the stained sample. In our studies, the weak autofluorescence signals generated in the red region of the spectrum by excitation with the argon laser 488 nm line produced a cell by cell green autofluorescence correction that was both noisy and highly leveraged (average green/red autofluorescence ratio ¼ 30). Therefore, we chose to base our cell by cell correction for autofluorescence on the green/orange autofluorescence ratio (average ratio 3.3), with a correction for autofluorescence-free FITC crosstalk into the PE/CY3 channel. While the crosstalk correction is, itself, error prone, we found that the overall results of the cell by cell correction for autofluorescence based on

97


Table 3 Log Means and Standard Deviations of Cell Autofluorescence and Total Cell Her-2/neu Fluorescence for Cell Lines and Tumor Lines Shown in Figure 1

Sample JC-1939 MCF-7 SKBR-3 BE 3449 RD 3900 RB 3905

Cell autofluorescence Log mean

S.D.

Total cell fluorescence Log mean

S.D.

Mean FLauto/ mean FLtotal

4.55 4.58 4.78 4.49 5.31 5.59

0.17 0.13 0.15 0.11 0.28 0.34

5.09 5.18 5.72 4.72 5.44 5.69

0.19 0.18 0.21 0.19 0.27 0.32

0.29 0.25 0.12 0.58 0.73 0.80

the orange channel measurements were more satisfactory than those based on the red channel measurements. Still, while the cell by cell correction method can be taken as an independent method against which the mean Flauto/mean Fltotal correction method is to be compared, this does not necessarily imply that the approach based on separate cell by cell measurements produces superior results. Statistical Analysis Differences in the means of continuous variables were evaluated by the student t test. The F test was used to evaluate differences between sample variances. RESULTS Cell by Cell Correlation Between Cell Autofluorescence and Total Cell Her-2/neu Immunofluorescence To determine directly whether there is a correlation between autofluorescence and total fluorescence in individual cells, we performed autofluorescence measurements by laser scanning cytometry on duplicate unstained cytospin slides prepared using cells from the tissue cultured cell lines, JC-1939, MCF-7, and SKBR-3. The same slides were then stained with anti-Her-2/neu antibody conjugated to FITC (see Materials and Methods) and total fluorescence was measured on each of the same cells. In these samples, the autofluorescence signals were small in relation to the total cell fluorescence signals, and the two were sufficiently well separated that it would be reasonable to suppose that the total signal might be representative of the true probe/dye fluorescence signal. The autofluorescence and total measurements were then paired cell by cell based

on their x, y coordinates on the slide. The correlation coefficients for cell by cell comparisons of the logs of the autofluorescence measurements versus the logs of the total fluorescence measurements are given in Table 4. All of the r values exceed 0.45, supporting the premise of proportionality between the autofluorescence measurement and the probe/dye fluorescence per cell. For each sample, the variance of the distribution of actual ratios of cell autofluorescence to total cell autofluorescence in individual cells was compared with the variance of the distribution of ratios of cell autofluorescence to total fluorescence that was obtained when the cell by cell autofluorescence measurements were paired randomly with the total cell by cell fluorescence measurements within the same data set. The results, given in Table 4, show smaller variances in the correlated cell by cell autofluorescence to total fluorescence ratios than the variances in the ratios obtained from randomly assigned pairing of the autofluorescence and total fluorescence measurements. In each case, the difference was highly significant statistically. Thus, there is a statistically significant correlation between autofluorescence and total fluorescence in individual cells in these tumor cell lines, which may reflect a correlation between the cell autofluorescence measurement and true probe/dye fluorescence component in individual cells. Results of Application of the Mean Flauto/mean Fltotal Correction Application of the mean Flauto/mean Fltotal cell by cell autofluorescence correction to the three breast cancer cell lines and the three lung cancer samples produced the results shown in Figure 3. Units are expressed in

Table 4 Correlation coefficients of total cell fluorescence and cell autofluorescence in the same cells (r), and comparison of the variance in the ratios of actual paired cell by cell measurements of cell autofluorescence and total cell fluorescence (Variance A/T, paired) with ratios of randomly paired values of cell auto fluorescence and total cell fluorescence (Variance A/T, random) Sample JC-1939 (1) JC-1939 (2) MCF-7 (1) MCF-7 (2) SKBR-3 (1) SKBR3 (2)

r 0.669 0.555 0.904 0.844 0.548

Variance A/T, paired

Variance A/T, random

P (F TEST)

1.051 1.057 1.032 1.038 1.098 1.112

1.107 1.111 1.295 1.235 1.167 1.275

1.7 1028 3.5 1018 1.45 10160 1.91 10113 3.02 108 2.33 1015

98

SMITH ET AL.

FIG. 3. Cell autofluorescence distributions (light dashed curves), total cell fluorescence distributions (heavy dashed curves), and mean Flauto/mean Fltotal ratio-corrected distributions (solid curves), plotted in the log domain, for (A) JC-1939 breast cancer cells, (B) MCF-7 breast cancer cells, (C) SKBR-3 breast cancer cells, (D) BE 3449 lung cancer cells, (E) RD 3900 lung cancer cells, and (F), RB 3905 lung cancer cells.

Her-2/neu molecules per cell equivalents. When the mean of the distribution of autofluorescence signals (light broken line) is much lower than that of the total signal (heavy broken line), the effects of the autofluorescence correction are small, as in the breast cancer cell lines (Figs. 3A–3C). In these examples, the corrected signal (solid line) changes relatively little in comparison with the total signal. When mean autofluorescence represents a large proportion of the mean total signal, as it does in the lung cancer samples shown in Figures 3E and 3F, then the means of the corrected fluorescence signals are shifted to much lower values in comparison with the total signals. In practice, the levels of the corrected signal intensity must often be interpreted in relation to some reference value that

may have biological significance. In the case of Her-2/neu, this threshold can be estimated to lie between 100,000 molecules per cell and 300,000 molecules per cell (19). In Figure 3, a reference threshold value of 200,000 molecules per cell is shown. In the examples shown in Figure 3, the corrected Her-2/neu values of most cells appear to fall below this threshold in all samples except the SKBR-3 tumor cell line. Comparison of Results of the Mean Flauto/mean Fltotal Correction with the Results of a True Cell by Cell Correction Shown in Figure 4 is a cell by cell comparison of total fluorescence per cell corrected for autofluorescence, using a method that is based on a separate autofluores-


99

FIG. 4. Cell by cell comparison of the results of the mean Flauto/mean Fltotal ratio-correction (Y axis) with those obtained with a true cell by cell correction, for (A) JC-1939 breast cancer cells, (B) MCF-7 breast cancer cells, (C) SKBR-3 breast cancer cells, (D) BE 3449 lung cancer cells, (E) RD 3900 lung cancer cells, and (F) RB 3905 lung cancer cells. [Color figure can be viewed in the online issue, which is available at www.interscience.wiley.com.]

cence measurement on each cell (see Materials and Methods), with the total fluorescence per cell corrected for autofluorescence using the mean Flauto/mean Fltotal correction in our six examples. In each sample, cells that exhibited high corrected values by one method, generally exhibited similarly high corrected values by the other, and low values by one method were correlated with low values by the other. In the lung cancer cases, both methods suggested that large proportions of cells with corrected Her2/neu values fell below the reference threshold of 200,000 cells (Table 5), The greater deviations of the slopes of the regression lines from unity in the samples shown in panels E and F may reflect limitations of either or both methods in estimating a relatively small true fluorescence signal accurately in the presence of an autofluorescence signal that is both increased and highly variable. Modeling Studies The mean Flauto/mean Fltotal correction assumes that the ratio of autofluorescence to total signal does not vary

substantially from cell to cell. Such a correlation is readily understood when the autofluorescence component of the total fluorescence signal is large. We have shown that for cell constituents such as Her-2/neu, there is still a correlation between autofluorescence and total cell fluorescence in three breast cancer cell lines even when the autofluorescence component is small (see Table 4). However, there is no reason to believe that such a correlation would hold generally for other measurements, for other samples, or under all conditions. Therefore, we have performed modeling studies to examine the consequences of applying a cell by cell correlation for autofluorescence that is based on mean Flauto/mean Fltotal, when there is no correlation whatever between autofluorescence and true probe/dye cell fluorescence in individual cells. These studies are described in Appendix 1, and the results are summarized here. As noted earlier, when the total signal is very high and the autofluorescence signal is very small, it makes little difference if the signals are correlated or not. By the same token, when the total signal is very high and the

100

SMITH ET AL.

Table 5 Comparisons of the Proportions of Cells Estimated to Have HER-2/neu Levels in Excess of 200,000 Molecules/Cell in Samples Corrected for Autofluorescence by a True Cell by Cell Correction Method, and in Samples Corrected for Autofluorescence Cell by Cell Using the Mean Flauto/mean FLtotal Correction

Sample JC-1939 MCF-7 SKBR-3 BE 3449 RD 3900 RB 3905

Cells with high Her-2/neu Cell by cell Mean ratio correction (%) correction (%) 2.7 6.8 93.3 1.2 15.2 6.2

2.3 5.0 95.1 0.9 2.6 9.2

autofluorescence signal is very high, then the true mean fluorescence signal is small, and any differences in the standard deviation (SD) due to assuming correlation when there is none are also very small. The greatest potential effect on the SD occurs when the mean autofluorescence and the mean true fluorescence signals are equal, i.e., when the signal to noise ratio is low (2:1). The modeling studies suggest that when the SD’s of each are similar, the estimate of mean corrected signal remains close to the true value, and the estimates of the SD that result from assuming a correlation when there is none, while lower than the true SD, might still be quite acceptable. DISCUSSION In this article, we have considered the consequences of applying a cell by cell autofluorescence correction based on a sample-specific ratio of mean autofluorescence to mean total fluorescence signal per cell. The advantages of using the mean Flauto/mean Fltotal correction are that it is a simple procedure that does not require special purpose instrumentation, and does not require the sacrifice of a data channel that might otherwise be available for the performance of additional potentially useful biological measurements in multiparameter studies. The correction does require that a separate unstained aliquot of cells from the same sample be run concomitantly, to obtain mean Flauto for each color channel measured, so that Eq. (3) (see Materials and Methods) can be applied cell by cell. This correction would be ideal if there were a close correlation between the autofluorescence signal and the total fluorescence signal in individual cells. We have shown that there may be a cell by cell correlation between the autofluorescence signal and the total cell Her-2/neu fluorescence signal in several human breast cancer cell lines. In these cell lines, the means of the total cell fluorescence distributions were substantially higher than those of their respective autofluorescence distributions, with little overlap between them, such that one might expect that both the means and SD’s of the total fluorescence distributions would reflect predomi-

nantly those of their respective true Her-2/neu/FTIC fluorescence distributions. Our modeling studies supported this premise. Thus, our results suggest that in these cell lines, levels of cell autofluorescence and true Her-2/neu/ FTIC fluorescence were correlated in individual cells. We compared the results of our mean cell autofluorescence/total fluorescence ratio correction with those obtained using an alternative method for cell by cell correction for autofluorescence. The reference method was based on a separate simultaneous measurement in each cell, from which the autofluorescence component of the total fluorescence signal of interest could be estimated. We found that the two methods yielded similar results. This approach would be expected to produce less satisfactory results with other measurement types or in other sample types in which there is no correlation between the autofluorescence and true probe/dye fluorescence components of the total signal in individual cells. We used a computer simulation approach to estimate the magnitude of the maximum misclassification error that might be attributable to the assumption of a correlation between autofluorescence and the total fluorescence signal in individual cells when, in fact, no such correlation exists. Our modeling studies suggest that when the means of the distributions of cell autofluorescence and true probe/dye cell fluorescence are equal (i.e., when the mean total signal is twice the mean autofluorescence signal), and when the standard deviations of these distributions are equal, the maximum expected misclassification error is in the range of 8% or less, over a wide range of SD’s. Of course, the validity of these estimates is dependent on the validity of the assumptions underlying the model, which include the use of lognormal distributions to represent the various component fluorescence distributions. Such assumptions do not take into account any fine structure that might be present in the data, such as cell cycle dependence of expression, and particularly decreases in levels of expression with cell cycle progression that might occur with some cell constituents (e.g. cyclin D1). The SD of the true probe/dye fluorescence distribution is not measured directly, and if it is substantially larger or smaller that than the SD’s encompassed by our simulations, then the error in estimating the proportion of cells that fall above or below the cutoff values used in our simulations could be larger. One obvious approach to decreasing the error in estimating the proportion of cells that belong in a given class under adverse measurement conditions would be to increase the cutoff threshold. Thus, for example, when, (a) the mean of the total fluorescence distribution is ~twice that of the autofluorescence distribution, and, (b) when there is, say, at least a twofold difference in the SD’s of the two distributions, one might choose to adopt a cutoff value that is, say, five to ten times the mean of the autofluorescence distribution. Our modeling studies suggest that with increasing disparity between the mean of the autofluorescence distribution and the mean of the true probe/dye fluorescence distribution (i.e., as the mean Flauto/mean Fltotal ratio

101


tends toward zero or unity), misclassification errors associated with the inappropriate assumption of a lack of a correlation between the autofluorescence signal and the total signal in individual cells recede in importance. As might be expected, as the mean of true probe/dye fluorescence signal comes to represent an increasing proportion of the mean total signal (i.e., the mean Flauto/mean Fltotal ratio is low), the mean and SD of the corrected distribution more closely reflect those of the true probe/ dye fluorescence distribution. However, the closer the mean Flauto/mean Fltotal ratio is to one, the more closely the SD of the corrected fluorescence distribution reflects that of the autofluorescence distribution rather than that of the true probe/dye distribution. Under these conditions, the lack of reliable information regarding the SD of the true probe/dye distribution is counterbalanced by the fact that the true mean fluorescence levels of the cell constituent of interest become vanishingly small as mean Flauto/mean Fltotal approaches 1. The cell autofluorescence levels in the lung cancer samples shown in Figure 1, panels E and F, are among the most problematic that we have encountered to date, and are not representative of most solid tumor samples. Still, our modeling studies suggest that it is possible to obtain quantitative estimates of mean levels of expression of biologically significant tumor cell constituents, and to determine population proportions of overexpressing or normally expressing/underexpressing cells within acceptable cell classification error tolerances even in such samples. Overall, the computer modeling studies suggest that the autofluorescence correction based on mean Flauto/mean Fltotal may be useful over a wide range of conditions that are likely to be encountered in multiparameter LSC analysis of human solid tumors. LITERATURE CITED 1. Corsetti JP, Sotirchos SV, Cox C, Cowles JW, Leary JF, Blumburg N. Correction of cellular autofluorescence in flow cytometry by mathematical modeling of cellular fluorescence. Cytometry 1988;9:539– 547. 2. Steinkamp JA, Lehnert NM, Keij JF, Lehnert BE. Enhanced immunofluorescence measurement resolution of surface antigens on highly autofluorescent, glutaraldehyde-fixed cells analyzed by phase-sensitive flow cytometry. Cytometry 1999;37:275–283. 3. Tadrous PJ, Siegel J, French PM, Shousha S, Lalani el-N, Stamp GW. Fluorescence lifetime imaging of unstained tissues: early results in human breast cancer. J Pathol 2003;199:309–317. 4. Steinkamp JA, Stewart CC. Dual-laser, differential fluorescence correction method for reducing cellular background autofluorescence. Cytometry 1986;7:566–574. 5. Roederer M, Murphy RF. Cell-by-cell autofluorescence correction for low signal-to-noise systems: application to epidermal growth factor endocytosis by 3T3 fibroblasts. Cytometry 1986;7:558–565. 6. Alberti S, Parks DR, Herzenberg LA. A single laser method for subtraction of cell autofluorescence in flow cytometry. Cytometry 1987;8: 114–119. 7. Tsurui H, Nishimura H, Hattori S, Hirose S, Okumura K, Shirai T. Seven-color fluorescence imaging of tissue samples based on Fourier spectroscopy and singular value decomposition. J Histochem Cytochem 2000;48:653–662. 8. Macville MV, Van der Laak JA, Speel EJ, Katzir N, Garini Y, Soenksen D, McNamara G, de Wilde PC, Hanselaar AG, Hopman AH, Ried T. Spectral imaging of multi-color chromogenic dyes in pathological specimens. Anal Cell Pathol 2001;22:133–142. 9. Ecker RC, de Martin R, Steiner GE, Schmid JA. Application of spectral imaging microscopy in cytomics and fluorescence resonance energy transfer (FRET) analysis. Cytometry A 2004;59:172–181.

10. Mosiman VL, Patterson BK, Canterero L, Goolsby CL. Reducing cellular autofluorescence in flow cytometry: an in situ method. Cytometry 1997;30:151–156. 11. Hallden G, Skold CM, Eklund A, Forslid J, Hed J. Quenching of intracellular autofluorescence in alveolar macrophages permits analysis of fluorochrome labelled surface antigens by flow cytofluorometry. J Immunol Methods 1991;142:207–214. 12. Hodge SJ, Hodge GL, Holmes M, Reynolds PN. Flow cytometric characterization of cell populations in bronchoalveolar lavage and bronchial brushings from patients with chronic obstructive pulmonary disease. Cytometry B Clin Cytom 2004;61:27–34. 13. Kunz WS. Spectral properties of fluorescent flavoproteins of isolated rat liver mitochondria. FEBS Lett 1986;195:92–96. 14. Heintzelman DL, Lotan R, Richards-Kortum RR. Characterization of the autofluorescence of polymorphonuclear leukocytes, mononuclear leukocytes and cervical epithelial cancer cells for improved spectroscopic discrimination of inflammation from dysplasia. Photochem Photobiol 2000;71:327–332. 15. Palmer GM, Keely PJ, Breslin TM, Ramanujam N. Autofluorescence spectroscopy of normal and malignant human breast cell lines. Photochem Photobiol 2003;78:462–469. 16. Kindzelskii A, Petty HR. Fluorescence spectroscopic detection of mitochondrial flavoprotein redox oscillations and transient reduction of the NADPH oxidase-associated flavoprotein in leukocytes. Eur Biophys J 2004;33:291–299. 17. Pollice A, McCoy JP, Shackney S, Smith C, Agarwal J, Burholt D, Janocko L, Hornicek F, Singh S Hartsock R. Sequential paraformaldehyde and methanol fixation for simultaneous flow cytometric analysis of DNA, cell surface proteins, and intracellular proteins. Cytometry 1992;13:432–444. 18. Shackney SE, Smith CA, Pollice AA, Brown K, Kosiban D. A suitable method for identifying cell aggregates in laser scanning cytometry listmode data for analyzing disaggregated cell suspensions obtained from human cancers. Cytometry 2004;59B:10–23. 19. Shackney SE, Shankey TV. Common patterns of genetic evolution in human solid tumors. Cytometry 1997;29:1–27.

APPENDIX 1 MODELING STUDIES For modeling purposes, we generated lognormal autofluorescence and true probe/dye fluorescence distributions of specified means and standard deviations. To convolve the two distributions, the lognormal distributions are mapped into the linear domain, and each distribution is sampled randomly using the Monte Carlo technique to obtain two independent random values for cell autofluorescence and true probe/dye fluorescence, respectively. These are added together to produce a simulated cell in which autofluorescence, true probe/dye fluorescence, and total cell fluorescence are known. The process is repeated to generate simulated cell populations of 65,000 cells, in which cell autofluorescence and total cell fluorescence are uncorrelated. The mean Flauto/mean Fltotal correction is then applied to each cell in these simulated populations, and the resulting corrected probe/dye fluorescence distribution is compared with the true probe/dye fluorescence distribution. Our starting point is two log normal distributions, where each distribution has a specified standard deviation of 0.2 log2s, (0.06 log10s), and we begin our simulations with log distributions in which the mean of the autofluorescence component and that of the true probe/dye fluorescence component are equal. The results of such a simulation study are shown in Figure A1. The mean of the convolved distribution is twice those of the autofluorescence distribution and the true probe/dye fluorescence distribution; the CV is larger in the linear domain, while its standard deviation is

102

SMITH ET AL.

FIG. A1. (A) Simulated log distributions of cell autofluorescence and true probe/dye fluorescence with equal means and SD’s (solid curve) and the convolved distribution representing the total cell fluorescence distribution (broken line). Note that the variance of the convolved distribution is less than that of the parent distributions. (B) Overlay of the mean Flauto/mean Fltotal ratio-corrected total cell fluorescence distribution over the original true probe/dye fluorescence distribution. Vertical lines represent various cutoffs for distinguishing fluorescent from non-fluorescent cells, including one at the mean of both distributions, cutoffs IP-L and IP-R, at the left and right intersection points of the two distributions, respectively, and T1 and T2, cutoffs that are far removed from the intersection points. (C) A graph of percent error in classifying the proportion of the cell population that lies above or below the cutoff due to the difference in SD’s of the two distributions, as a function cutoff position, expressed as a ratio of cutoff level to the mean. [Color figure can be viewed in the online issue, which is available at www.interscience.wiley.com.]

smaller in the log domain (Fig. A1(A)). The means of the corrected distribution and the original probe/dye fluorescence distribution are the same (Fig. A1(B)); the standard deviation of the corrected distribution is the same as that of the convolved distribution, and is less than that of the original true distribution (Fig. A1(B)). In these model populations, it is the difference in the standard deviations that is responsible for the error that is generated by the use of a fixed ratio to correct for autofluorescence in a cell population in which autofluorescence and true cell fluorescence are, in fact, not correlated. To quantitate the maximum expected error due to the discrepancy in the SDs of the two distributions, we examine the difference in the proportions of cells that lie above a given cutoff value, over a range of different cutoff values. Thus, for example, for a cutoff value that is well below the mean of the corrected (or original) distribution, almost all cells in both distributions lie above the cutoff (cutoff T1, Fig. A1(B)), and the error due to the difference in SD’s is exceedingly small (Fig. A1(C)). Conversely, for a cutoff value that is well above the mean of the corrected (or original) distribution (cutoff T2, Fig. A1(B)), almost all cells in both distributions lie below the cutoff, and the error due to the difference in SD’s is also exceedingly small. For a cutoff value that is positioned at the mean of both distributions, the proportions of cells on either side of the cutoff are equal, and the error is zero (see Fig. A1(C)). Positive and negative errors are greatest at the

left and right intersection points of the two distributions, respectively (IP-L and IP-R, Figs. A1(B) and A1(C)). At the left intersection point, the corrected distribution predicts that 8% more cells lie above the cutoff than the actual percentage calculated from the original distribution. The absolute value of the maximum expected error under a given set of simulation conditions always lies at one of the points of intersection of the two distributions, and, as long as the corrected distribution remains symmetrical, the absolute values of the maximum errors at IP-L and IP-R are approximately equal and are opposite in sign. When the ratio of the mean of the autofluorescence distribution to the mean of the true probe/dye fluorescence distribution is less than 1 (e.g., 0.1 in Figures A2(A1) and A2(A2)) or greater than 1 (e.g., 10 in Figs. A2(C1) and A2(C2)), then the maximum observed error is smaller than when the means of the autofluorescence and true probe/ dye fluorescence distributions are equal (Fig. A2(D)). The diminution in the maximum observed error reflects the greater contribution of the more prominent signal component to the mean and SD of the convolved total cell fluorescence distribution. Because the specified SD’s of the two input distributions are identical in this case, the SD of the corrected total cell fluorescence distribution are close to those of the true probe/dye fluorescence signal, whether or not the latter is the more prominent component of the cell fluorescence signal.


103

FIG. A2. (A1) Simulated log distributions of cell autofluorescence and true probe/dye fluorescence with unequal means and equal SD’s of 0.06 (solid curves), and the convolved distribution representing the total cell fluorescence distribution (broken line). The mean of the autofluorescence distribution is one tenth that of the true probe/dye fluorescence distribution. The mean and SD of convolved distribution are close to those of the true probe/dye fluorescence distribution (A2) Overlay of the mean Flauto/mean Fltotal ratio-corrected total cell fluorescence distribution over the original true probe/dye fluorescence distribution shown in A1. The mean of the corrected distribution is the same as that of the true probe/dye fluorescence distribution, and the SD’s are very similar. The error at the IP-L cutoff is 2%. (B1) Simulated log distributions of cell autofluorescence and true probe/ dye fluorescence with equal means and equal SD’s of 0.06 (solid curves), and the convolved distribution representing the total cell fluorescence distribution (broken line). The mean of convolved distribution is twice those of the autofluorescence and true probe/dye fluorescence distributions, and the difference in SD’s is greater than in panel A1. (B2) Overlay of the mean Flauto / mean Fltotal ratio-corrected total cell fluorescence distribution over the original true probe/dye fluorescence distribution. The mean of the corrected distribution is the same as that of the true probe/dye fluorescence distribution, and the difference in SD’s is greater than in panel A2. This difference, and the error at IP-L (8%) are maximal when the means of the autofluorescence and true probe/dye fluorescence distributions are equal (see panel D). (C1) Simulated log distributions of cell autofluorescence and true probe/dye fluorescence with unequal means and equal SD’s of 0.06 (solid curves). The mean of the autofluorescence distribution is ten times that of the true probe/dye fluorescence distribution. The mean and SD of convolved distribution are close to those of the autofluorescence distribution. (C2) Overlay of the mean Flauto/mean Fltotal ratio-corrected total cell fluorescence distribution over the original true probe/dye fluorescence distribution shown in C1. The mean of the corrected distribution is the same as that of the true probe/dye fluorescence distribution, and the SD’s are very similar. The error at the IP-L cutoff is 2%. (D) When the SD’s of the autofluorescence and true probe/dye fluorescence are equal, the percent error in classifying the proportion of the cell population that lies above the cutoff at IP-L or IP-R is maximal when the means of the two distributions are equal. [Color figure can be viewed in the online issue, which is available at www.interscience.wiley.com.]

We have examined the consequences of large differences in both the means and standard deviations of the autofluorescence and true probe/dye fluorescence distributions. With disparities in the SD’s of the input distributions, the behavior of the model became more com-

plex. Overall, estimates of mean true fluorescence remained very close to the input values, and the maximum observable classification error remained below 20% for pairs of disparate SD’s that ranged between 0.09 and 0.36 logs.