Membrane microdomains enriched in GPI-anchored proteins, GSL, and cholesterol have been operationally defined in terms of detergent-insoluble, low-density membrane fractions; however, these microdomains have not been detected by other techniques (for review see Harder and Simons, 1997
). In the present study, we used imaging FRET, a method that increases the resolution of immunofluorescence microscopy to the molecular scale, to probe for microdomains enriched in a GPI-anchored protein in the apical plasma membrane of MDCK cells. We expected that if most of the GPI-anchored protein 5′ NT was in membrane microdomains, then we would be able to detect FRET between 5′ NT molecules that was consistent with clustering (Table ).
Using imaging FRET, we obtained images that showed significant cell-to-cell variation in efficiencies of energy transfer between labeled 5′ NT molecules. E strongly correlated with the surface density of 5′ NT and approached zero at low surface densities (Figs. –). These observations suggest that most of 5′ NT is not in clusters at the apical membrane of MDCK cells; if it were, we would have expected to measure high E even at low surface densities (Table ). These observations were not due to the failure of the method to detect clusters since we detected clustered donors and acceptors in a simple model system, secondary antibodies bound to primary antibodies (Fig. ).
The distinct dependence of E on the surface density of 5′ NT (Figs. –) is consistent with either an entirely random distribution of the protein or a mixture of randomly distributed and clustered 5′ NT (Table ). If some 5′ NT is in clusters and some is randomly distributed, we expect that E will not necessarily go to zero in the limit of low surface densities, and E will be sensitive to D:A (Table ). If little or no 5′ NT is in clusters, then we expect E to go to zero in the limit of low surface density to be insensitive to D:A.
We found that E
approached zero in the limit of low 5′ NT surface densities (Figs. –), and the shape of the experimental curves was similar to that predicted theoretically for a random distribution (Fig. ). E
depended strongly on the surface density of acceptor, and not of the surface density of donors for relatively low concentrations of the donor fluorophore. Regardless of D:A, data from an individual experiment tended to fall on a single curve when plotted as a function of acceptor surface density in cells expressing widely varying concentrations of protein (Fig. ). However, in some experiments (Fig. ) we observed small shifts in E
for samples labeled with different D:A ratios, which hints that some clusters may be present. We thus cannot completely rule out the possibility that though most are randomly distributed, some 5′ NT molecules are clustered. We expect that under some circumstances, FRET for mixtures of randomly distributed and clustered molecules will appear similar to FRET for a purely random distribution, particularly in the limit where fclustered
is small, and frandom
is large (see Appendix
). This could explain for instance why we did not observe a larger effect on E
in cells where we induced clustering of GPI-anchored proteins by secondary antibody-induced cross-linking, or by detergent extraction of intact cells (Figs. and ). We calculated that if 10% of the molecules were clustered, data for a mixed population would appear very similar to a pure random distribution (assuming r
). However, large differences are expected when fclustered
= 50% (see Appendix
). Based on these observations, the simplest interpretation of our data is that 5′ NT is predominantly randomly distributed under the conditions of our experiments.
It is important to emphasize that our conclusion that 5′ NT is predominantly randomly distributed depends on the presence of FRET and not its absence. In a previous study from our laboratory, no FRET was detected between labeled gD1-DAF molecules under steady-state condition, implying that gD1-DAF was dispersed, i.e., randomly distributed (Hannan et al., 1993
). However, the absence of FRET does not necessarily eliminate the possibility that molecules are clustered together, since among other possibilities the distance separating them in the cluster may be larger than can be detected by FRET. In the present study we were able to detect FRET between labeled 5′ NT molecules; this indicates that, on average, the proteins are already within 10's of Å of one another. This provides further evidence that FRET would be able to detect enrichment of 5′ NT in microdomains, i.e., lipid rafts. Reexamination of the lateral organization of gD1-DAF using our current FRET method shows that energy transfer is detected between labeled gD1-DAF molecules under steady-state conditions and is correlated with protein surface density, similar to our results for 5′ NT (Nguyen, T., A. Kenworthy, and M. Edidin, unpublished observations). The difference between our past and present results may be due to lower concentrations (surface densities) of labeling antibodies in our previous experiments, and the sensitivity of the current method to cell-to-cell variations in E
. This further emphasizes the most important advantage of imaging FRET over nonimaging FRET experiments, which typically yield average E
values for a population of cells (Hannan et al., 1993
; Matko and Edidin, 1997
): imaging FRET generates images mapping energy transfer efficiencies. Although in the current study we have focused on protein homoassociations, imaging FRET is also uniquely suited for performing “imaging biochemistry” of protein– protein and protein–lipid heterointeractions in intact cells.
To further verify our conclusion that most 5′ NT is randomly distributed in the cell surface would require a quantitative comparison of our data with the theoretical predictions. Variations on the analytical approach we applied here have been used to quantitatively interpret FRET data in a variety of membrane systems (e.g., Holowka and Baird, 1983a
; Dewey and Datta, 1989
; John and Jähnig, 1991
; also see Mátyus, 1992
and Clegg, 1996
for a more comprehensive list). However, our ability to extend our current analysis from a qualitative to a quantitative one is limited by several factors. First, there is some spread in the energy transfer versus acceptor surface density curves that could mask differences in the data sets. Factors contributing to this variability include inhomogeneities in the excitation across the field of view, small local variations in quenching of donor, and the inherent heterogeneity of fluorescence labeling pattern of the cell surface due to the presence of microvilli and the curvature of the apical membrane itself. Second, the lower limit of detectability of FRET with our method is conservatively 5%, although in many experiments we measured apparent E
for negative controls (labeled with donor only) of as low as 2–3% (Fig. ). This could be further improved by additional background subtraction. In the current experiments, because we were imaging confluent monolayers of cells, the only background subtraction we included was for the so-called dark current, a constant contribution due to noise from the CCD camera. With higher sensitivity we could estimate the extrapolation of E
and the acceptor surface density to zero with greater confidence. Third, there are limits to our comparisons of experimental results to theory and with the theoretical calculations themselves. Our ability to rigorously test the theoretical predictions would be improved by better estimates of acceptor surface density and r
. A physically based approximation of r
, for instance, would require more detailed structural information about 5′ NT, including the position of the epitope and the orientation of the bound antibody. To this point, we note that a number of sophisticated analyses have been developed that incorporated detailed information about the experimental system, such as the position of the donor fluorophore on the molecule of interest (e.g., Zimet et al., 1995
). This kind of information could be usefully applied in the analysis of imaging FRET data for better characterized experimental systems. The analysis would also be improved by better theoretical models for mixed populations, including specific cross-terms between the randomly distributed and clustered molecules.
It is well known that the distribution of GPI-anchored proteins in cell membranes is sensitive to fixation and labeling conditions. For example, GPI-anchored folate receptor was reported to be clustered (Rothberg et al., 1990
), but this was subsequently shown to be induced by cross-linking of secondary antibodies (Mayor et al., 1994
). Clustering of 5′ NT itself has also been shown after antibody cross-linking (Howell et al., 1987
). In our experiments, we observed very similar, dispersed labeling patterns produced by either monovalent Fab or bivalent IgG when the cells were directly fixed after labeling. In addition, the organization of the primary antibodies became more punctate (clustered) in the presence of secondary antibodies (Fig. and data not shown). Thus, our results are similar to previous reports in this regard (Rothberg et al., 1990
; Mayor et al., 1994
). However, very recent work suggests that certain fixation conditions may act to disperse preexisting clusters of the folate receptor (Wu et al., 1997
), reopening the question of how to best stabilize the native distribution of GPI-anchored proteins. This issue will require further study. Nevertheless, our current results are consistent with reports (Parton et al., 1994
; Mayor and Maxfield, 1995
; Fujimoto, 1996
; Rijnboutt et al., 1996
) of predominantly random steady-state distributions of GPI-anchored proteins, measured at the level of resolution of the electron microscope. Compared with electron microscopy, imaging FRET has the advantages of higher labeling efficiency, a larger sample size, and most importantly, increased resolution, to the molecular level.
The first experimental evidence for the existence of membrane microdomains enriched in GPI-anchored proteins was the isolation, as buoyant complexes, of detergent-insoluble membrane fractions enriched in GPI- anchored proteins, GSL, and cholesterol from MDCK cells (Brown and Rose, 1992
). The GPI-anchored protein PLAP was found to become detergent insoluble in the Golgi, a property that persisted even after the protein reached the apical membrane (Brown and Rose, 1992
). Additional components of detergent-insoluble membrane microdomains were later shown to include signal-transducing lipid-modified proteins such as nonreceptor tyrosine kinases and the caveolar marker caveolin (Stefanova et al., 1991
; Sargiacomo et al., 1993
;Arreaza et al., 1994
; Melkonian et al., 1995
). Recently, a number of studies have helped clarify the relationship between membrane microdomains enriched in GPI-anchored proteins, detergent-insoluble complexes, and caveolae, 70–100-nm invaginations of the plasma membrane that are decorated with the protein caveolin (Fra et al., 1994
; Schroeder et al., 1994
; Gorodinsky and Harris, 1995
; Mayor and Maxfield, 1995
; Schnitzer et al., 1995
; Smart et al., 1995
; Hannan and Edidin, 1996
; Ahmed et al., 1997
). These studies indicate that proteins and lipids that share the common property of detergent insolubility are not necessarily associated in intact cell membranes, but leave open the question of the size and nature of the microdomains before detergent extraction (Kurzchalia et al., 1995
; Edidin, 1997
; Harder and Simons, 1997
; Weimbs et al., 1997
). There is clearly a need for a more precise definition of what constitutes a functional membrane microdomain in intact cell membranes.
Our findings, obtained using a high-resolution imaging technique, begin to place distinct limits on the structure of microdomains enriched in GPI-anchored proteins in intact cell membranes. We report here that the GPI-anchored protein 5′ NT, which is known to associate with detergent-insoluble complexes (Mescher et al., 1981
; Schnitzer et al., 1995
; Strohmeier et al., 1997
), is also resistant to Triton X-100 extraction in MDCK cells. Yet it appears that most 5′ NT are randomly distributed and are not clustered over the <100-Å length scale of the FRET measurements. This places limits on the ways in which GPI-anchored proteins can associate with lipid rafts. For example, our data argue against a model where 5′ NT is predominantly associated with a finite number of rafts, since we would have expected to measure relatively high energy transfer because of clustering of the protein in rafts, even at low surface densities of 5′ NT expression (Fig. ). The limitations of our current measurements prevent us from ruling out the possibility that FRET between 5′ NT arises from a mixture of a large fraction of randomly distributed and a small fraction of clustered (raft-associated) molecules. However, it is interesting to consider the consequence of a purely randomly distribution of 5′ NT for the structure of lipid rafts. If this were the case, then these membrane microdomains must either be vanishingly small (in agreement with a current model [Harder and Simons, 1997
]), or alternately must comprise the entire apical membrane. Our results also have implications for the membrane organization of GPI-anchored proteins during membrane trafficking and sorting, since the biochemical properties of the membrane microdomains involved in sorting and the steady-state organization of GPI-anchored proteins are assumed to be similar. This could be further tested by directly examining the membrane organization of GPI- anchored proteins and other apically destined proteins and lipids in the Golgi. Further work will be also be required to determine whether functional associations of other lipid-modified proteins and GSL can be visualized in cell membranes, and whether these associations are mediated by lipid–lipid or protein–lipid interactions. Imaging FRET will be a powerful tool to further investigate these questions in intact cells.