2.1.

Instrument

We used a multifunction microplate detector (UltraEvolution FLT, Tecan Trading, Switzerland) for this study. The instrument consists of a high-throughput optical detector with two picosecond-pulsed semiconductor lasers at $440\mathrm{nm}$ and $635\mathrm{nm}$ . Light is detected with a photo-multiplier tube (PMT), connected to a time-correlated single-photon counting (TCSPC) board (Picoquant, GmBH).

Light ( $440\mathrm{nm}$ for all experiments described here) is conveyed via optical fibers to a series of $45°$ Neutral Density 2 mirrors on a translation stage, allowing for the brightness of the laser pulse to be attenuated to avoid photon pileup,⁵ before being reflected into the sample. Light emitted from the sample is passed through an interchangeable emission filter ( $540\pm 25\mathrm{nm}$ for all studies performed here) before being transmitted to the detector. The instrument initially had light collection optics for both transmission and reflection geometry. The bottom-mounted transmission optics were replaced with a beam block (Thorlabs, NJ) to reduce reflection and autofluorescence.

For all measurements, 96-well, black, untreated glass-bottom plates (Nalge NUNC International, Rochester, NY) were used, as they were found to have low autofluorescence.

2.2.

Fluorescent Dyes

Alexafluor 430 (A430) (Invitrogen, CA)¹⁸ was chosen as a suitable fluorescent dye because the peak of its excitation spectrum $(\approx 440\mathrm{nm})$ is close to the wavelength of one of the instrument lasers. It also has a large stokes shift with the peak of the emission spectrum being $\approx 540\mathrm{nm}$ and overlapping well with the excitation spectrum of Cyananine3 (cy3) dye (Invitrogen, CA), which was chosen as a suitable acceptor.

2.3.

Cell Preparations

Three cell lines were used during this investigation: Chinese hamster ovary (CHO); PS70, CHO cells stably transfected with wild-type (WT) PS1 and WT amyloid precursor protein 770 (APP); and C410Y, CHO cells stably transfected with Familial Alzheimer’s disease associated mutant PS1 and WT APP.¹⁷

Cells were grown in Opti-MEM (Gibco) supplemented with 10% feotal bovine serum, $200\mathrm{\mu }\mathrm{g}∕\mathrm{ml}$ G418, and $2.5\mathrm{\mu }\mathrm{g}∕\mathrm{ml}$ puromycin at $37°\mathrm{C}$ with 5% ${\mathrm{CO}}_2$ . For experiments, the PS70 cells were split onto black 96-well plates coated with poly L-lysine. The cells were allowed to grow to confluency $(\approx 24\mathrm{h})$ prior to immunocytochemistry. Cells were fixed in methanol ( $10\min$ , $0°\mathrm{C}$ ), permeabilized, and nonspecific binding sites blocked (5% normal donkey serum, 0.1% Triton-X 100, $45\min$ ).

To create a model cell-based FRET assay, cells were immunostained with a mouse antitubulin primary antibody (Abcam, Cambridge, MA) overnight at $4°\mathrm{C}$ , and goat antimouse secondary antibody, labeled with A430 (1:100) (Invitrogen, Carlsbad, CA), $2\mathrm{hr}$ at room temperature. Cells in some wells were additionally stained with a donkey antigoat tertiary antibody labeled with Cy3 (1:200) (Jackson Immuno Research) for $2\mathrm{hr}$ at room temperature. This protocol creates a condition whereby A430 (donor) is brought into close proximity to Cy3 (acceptor) so that FRET can occur. Controls included unstained cells as well as cells immunostained with the acceptor or donor alone.

PS70 and C410Y cells were used in a demonstration of a conformation assay.¹⁹ Termini of PS1 were immunostained with rabbit $\alpha$ -human S182 antibody against amino acids 450–467 (C-terminus) (Sigma, St Louis, MO) and goat $\alpha$ -human PS1 directed against amino acids 14–33 (N-terminus) (Sigma). Cells were subsequently incubated with goat $\alpha$ -rabbit A430 and donkey antigoat Cy3 ( $2\mathrm{hr}$ , room temperature).

Alternatively, cells were immunostained with C8 antibody raised against amino acids 676–695 of APP (20) and an antibody to the major loop between PS1 TM6 and TM7 (mouse $\alpha$ -PS1, amino acids 263–378; Chemicon, Temecula, CA). Secondary antibodies were goat $\alpha$ -mouse A430 and donkey $\alpha$ -rabbit Cy3.

In all preparations, cells were washed three times in 1X TBS between and after rounds of staining.

2.4.

Curve Fitting

During FRET,²¹ energy is nonradiatively transferred from a donor fluorophore to an acceptor molecule. Although the transfer is nonradiative, the emission spectrum of the donor must overlap with the excitation spectrum of the acceptor. It is not necessary for the acceptor to be a fluorophore, although they are generally used. The effect of the transfer is to excite (increase the brightness) of the acceptor population while quenching (decreases in both brightness and lifetime) the donor.

The stretched exponential function¹⁰ (StrEF) differs from the simple exponential (SimEF) by the inclusion of an extra characteristic constant, $\beta$ , which represents the degree of stretch of the function,

We suggest the use of it here to describe the background autofluorescence of the instrument and sample assay as it can adequately describe a broad range of continuous and discrete multi-exponential functions.⁶ From a practical perspective, the simplicity and generality of the StrEF, compared to a complex multi-exponential function, make it easier to apply. This is especially important in a high-throughput or laboratory environment where various conditions, antibodies, or cell types may be used, provided that the StrEF adequately describes the background autofluorescence and allows discrimination of FRETing and non-FRETing conditions.

A custom-written analysis named General Multi-Exponential Fitter (GeMEF) was created using Matlab (Mathworks, Novi, MI). We used a generalized multi-exponential function, which can represent a sum of StrEF or SimEF functions:

2.5.

Calculation of FRET Strength and Efficiency

The number of molecules of a given population $\left(N_i\right)$ is proportional $N_i\propto I_0{\alpha }_i$ ; the proportion of the donor molecules that are interacting (FRET strength) is given by

2.6.

Experimental Protocols

For initial characterization of the noise distribution and background autofluorescence, wells containing $100\mathrm{\mu }\mathrm{L}$ PBS (pH 7.4) were fit with parameters $C$ , $I_0$ , ${\alpha }_b$ , ${\tau }_b$ , and ${\beta }_b$ allowed to vary while the other parameters are fixed with all other values of ${\alpha }_i=0$ . We use the subscript $b$ to refer to background.

The sensitivity of the instrument to dye concentration was assessed by varying concentrations of free A430 ( ${10}^{-8}-1\mathrm{mg}∕\mathrm{ml}$ in pH 7.4 PBS) (eight wells per condition, spread over two 96-well plates). A set of eight wells per plate contained saline alone. The level of attenuation of the excitation light was tuned to maximize photon count while avoiding pileup. For low concentrations of dye, when the laser was not strongly attenuated, the background autofluorescence was measured from the wells with no fluorophore and those parameters were fixed before fitting a mono-exponential curve to the data. In the case where laser power was significantly attenuated, the background autofluorescence was too weak to accurately measure. Under those conditions, given the relative strength of the signal, the background was assumed insignificant and ignored.

A series of conditions with varying FRET strengths but the same FRET efficiency were made from solutions of goat antimouse primary antibody $(1\mathrm{\mu }\mathrm{g}∕\mathrm{ml})$ , labeled with A430,¹⁸ which were incubated with varying amounts of biotinylated (nonfluorescent) antigoat (4 – $19\mathrm{\mu }\mathrm{g}∕\mathrm{ml}$ ) secondary antibody in $850\mathrm{\mu }\mathrm{L}$ of PBS, for $1\mathrm{hr}$ at room temperature. Donkey antigoat, labeled with Cyanine 3 (Cy3) $(1\mathrm{\mu }\mathrm{g}∕\mathrm{ml})$ , was added and incubated for a further hour at room temperature. The solutions were transferred to a 96-well plate.

Finally, CHO, PS70, and C410Y cells were grown in 96-well plates and immunolabelled as described above. For the model FRET assay experiment, eight wells were used for both negative controls and experimental samples. For the PS1 conformation and PS1-APP studies, 16 wells were used per condition and for negative controls. Eight wells containing unstained cells were used for background assessment.

The background assessment was conducted as described above. The lifetime of the non-FRETing donor fluorophore was then obtained by examining wells containing the donor but no acceptor. During this stage, a second exponential term was introduced and values of ${\alpha }_1,{\tau }_1$ were fit; ${\beta }_1$ was fixed to 1. Finally, the lifetime of the FRETing donor population $\left({\tau }_2\right)$ and relative amplitudes of the non-FRETing $\left({\alpha }_1\right)$ and FRETing $\left({\alpha }_2\right)$ populations were obtained from wells containing the complete assay, while ${\tau }_1$ was fixed at the previously obtained value and ${\beta }_2$ was fixed to 1; the nonacceptor wells were fit again under these conditions, as negative controls.

3.1.

Background Autofluorescence and Noise Characteristics

Wells in a 96-well plate that contain $100\mathrm{\mu }\mathrm{L}$ of saline were used to assess the signal due to instrument and plate autofluorescence and the validity of the noise model (Fig. 1 ). The goodness-of-fit measure, reduced ${\chi }^2\left(r{\chi }^2\right)$ for the SimEF, was 2.86. The time course of the residuals, divided by the expected standard deviation according to Poisson statistics $\left(\sqrt{I}\right)$ , corrects for the time dependency of the noise model. The residuals for the SimEF are not normally distributed around zero at all points during the decay. For the stretched exponential, $r{\chi }^2=1.17$ is achieved and the residuals appear to be more normal.

Fig. 1

Analysis of background fluorescence. A typical time-resolved fluorescence signal from a Nunc 96-well, glass-bottomed plate containing $100\mathrm{\mu }\mathrm{L}$ of saline is shown fitted with both simple A and stretched B exponential curves. The residuals are also shown, normalized by expected noise $\left({\sigma }_e\right)$ . Panels C and D show the normalized residual distributions fitted by a Gaussian. The width of the Gaussian $\left({\sigma }_g\right)$ and the regression value $\left(R^2\right)$ are shown. The autocorrelations (AC) (panels E, F) and the goodness of fit to $AC=0$ of all coefficients apart from the central value $\left(R^2\right)$ .

Two further tests of goodness of fit were performed to quantify the improvement to fit of using the StrEF. Fitting the distribution of the corrected residuals to a Gaussian gives an indication of how well the residuals conform to the normal distribution (Figs. 1C+D).²⁴ Ideally, the width of the fitted Gaussian $\left({\sigma }_g\right)$ and the goodness-of-fit measure $\left(R^2\right)$ should both be unity. The stretched exponential conforms better to the Gaussian ( $R^2=0.99$ compared to $R^2=0.97$ for the simple exponential). The standard deviation obtained from the width of the fitted Gaussian is close to 1 for the StrEF $({\sigma }_g=1.10)$ but larger for the simple exponential $({\sigma }_g=1.79)$ . The second test was to check for the statistical independence of various time points using the autocorrelation (AC) function²⁴ of the residuals $C\left(T\right)$ , where $T$ is the temporal shift. For the SimEF (Panel C), strong coefficients are found at shifts other than 0. The goodness of fit to $\mathrm{AC}=0$ for all nonzero lags is shown to be $R^2=0.67$ . For the StrEF, the AC function is much flatter for all nonzero shifts and $R^2=0.99$ ; this shows that the residuals are less correlated (whiter) for the StrEF. Although these data partially validate the use of the StrEF, for complete validation, the technique must be applied to a model protein–protein interaction to prove its ability to discriminate between positive and negative FRET conditions.

Simulations to study the effect of noise level on the fitting parameters were conducted by generating time courses $(n=100)$ of the form

Table 1

Comparison of Standard Deviations in Fitting Parameters Expected from Poisson Statistics and Measured Standard Deviations (n=96) .

3.2.

Assessment of Instrument Sensitivity to Fluorophore Concentration

Varying concentrations of free A430 were used to test the sensitivity of the instrument. At the greatest laser attenuation (ND1024), the background autofluorescence was negligible. Decay curves were fit with a mono-exponential plus an offset. The amplitude (Fig. 2A ) varies linearly with concentration. The lifetime ( $\tau$ , Fig. 2B) is stable over three orders of magnitude of concentration $({10}^{-4}–{10}^{-1})$ at $3659\pm 37\mathrm{ps}$ . At ${10}^{-5},\tau =3141\pm 167\mathrm{ps}$ . Concentrations higher than ${10}^{-1}$ could not be analyzed due to photon pile-up; at lower concentrations, the number of photons was too small to fit. At minimum laser attenuation (NDI), the background was fitted with a stretched exponential and subtracted from the signal. The intensity varied linearly with concentration (Fig. 2C). At ${10}^{-3}\mathrm{mg}∕\mathrm{ml},\tau =3659\pm 41\mathrm{ps}$ , observed at $\mathrm{ND}=1024$ . At lower concentrations, both $\tau$ and $\sigma$ increase as the signal level decreases. At the minimum measurable concentration $({10}^{-5}\mathrm{mg}∕\mathrm{ml})$ , the value of $\tau$ is 0.14% higher than the value given at high concentrations (Fig. 2D). At concentrations below ${10}^{-5}$ , the signal was too weak to fit, causing large variations in fitted values (data are not shown). These data show that amplitude varies linearly with concentrations over 4 orders of magnitude while $\tau$ remains almost completely stable. Even at low signal levels, the fitted lifetime is remarkably close to the results taken at high concentrations.

Fig. 2

Measured fluorescence signal rises monotonically with an increased fluorophore concentration. At higher laser attenuation (Panel A), the background autofluorescence was not subtracted and intensity varied linearly over 4 orders of magnitude. The lifetime (Panel B) was found to be very stable although it was found to be slightly lower with greater error at the lowest concentration. At a lower attenuation, the background autofluorescence was subtracted. The intensity was found to be linearly dependent on concentration (Panel C). The fitted lifetime did not vary significantly; the standard deviation increased at low concentrations but was small (Panel D). The amplitude of the fitted curve is shown to vary monotonically with respect to molarity. The background subtraction was performed on the data in Panels C and D. Lifetimes were found to be concentration-independent at $3747\pm 3.3\%$ standard deviation, ${\chi }^2<1.2$ for all fits.

3.3.

Simulations of the Effect of Background Autofluoresence on Measurement

To assess the impact of background autofluorescence as a confound and show that it is necessary to subtract it before fitting, a mono-exponential decay was simulated and analyzed using the standard bi-exponential model used in most FLIM experiments. The sum of a stretched exponential with parameters similar to the background function observed $(T_b=1800\mathrm{ps},{\beta }_b=0.68)$ and a mono-exponential ( ${\tau }_s=3000\mathrm{ps}$ , where the subscript $s$ means simulated) with varying relative amplitudes were fit with a bi-exponential using GeMEF. A time base similar to that generated by the instrument ( $35\mathrm{ps}\times 1330$ points) was used. The relative amplitude (Fig. 3A ) of the component with the shorter lifetime $\left({\alpha }_2\right)$ rises as the signal strength increases. The shorter lifetime $\left({\tau }_2\right)$ is closer to and lower than the simulated lifetime. It has a minimum where the signal strength is about twice the background, while ${\tau }_1$ increases monotonically with signal strength.

Fig. 3

The extent to which a stretched exponential background would lead to an apparent second exponential in the analyzed data. A stretched exponential function similar to the one observed in the instrument is added to a single exponential of varying intensities. Panels A and B show the relative amplitude of the shorter lifetime term and the returned lifetime against signal strength normalized by the amplitude of the simulated background function. Panels C and D show the case in which an offset was added, for which the fitting routine fit. The results show that the contribution of the stretched exponential background function would result in two apparent decay components and therefore false positive FRET results.

An offset might compensate for a long-lived component. A constant $\left(C\right)$ was added and the fitting routine was allowed to fit it. As the signal strength increases, the two lifetimes converge toward $\tau$ . The longer-lived component is stronger but does not dominate (Fig. 3C).

In both simulations, a significant false positive result was predicted when background autofluorescence was not accounted for, especially at the low signal strengths expected from cell-based assays.

3.4.

Measurement of FRET Strength at Constant FRET Efficiency

An important requirement of time-domain FRET data analysis is the ability to distinguish between the parameters of FRET strength and efficiency. Here we determine whether our calculations effectively distinguish these two parameters with minimal cross-talk. Donor-labeled antibody was pre-incubated with varying amounts of biotinylated secondary antibody to reduce the number of available epitopes, before being incubated with acceptor-labeled secondary. The proportion of primary antibody that is bound to secondary labeled with acceptor should be inversely proportional to the concentration of the biotinylated secondary. The inverse of the FRET strength increases linearly with no change in FRET efficiency (Fig. 4 ). These data illustrate our ability to distinguish between a change in the proportion of proteins interacting and the closeness with which they interact.

Fig. 4

Our analysis separates FRET strength and efficiency with minimal cross-talk. Goat antimouse primary antibody, pre-incubated for $1\mathrm{hr}$ with varying amounts of donkey biotin antigoat was incubated for a further hour with Cy3 labeled donkey antigoat antibody. The biotin-labeled secondary blocks the binding sites available to the labeled secondary, leading to a reduction in FRET strength (Panel A),without changing FRET efficiency (Panel B).

3.5.

Model Protein–Protein Interaction Assay

Measurements of protein interactions in cell systems using immunohistochemistry or fluorescently labeled protein expression are an important experimental use of plate readers. Here we demonstrate the use of the plate reader to detect a protein–protein interaction and show the importance of subtracting the background autofluorescence to avoid false positive results.

Sufficient signal was detected from a monolayer of adherent cells immunostained with A430, using tubulin as the antigen, in the 96-well format with the instrument. The lifetime of A430 was found to be $3794\pm 121\mathrm{ps}$ (averaged over eight wells). Controls included wells treated with either the primary or secondary antibodies alone, from which the signal was negligible.

The model protein–protein interaction described in the methods produces a circumstance in which tubulin is labeled with both A430 and Cy3, bringing the A430 and Cy3 into close proximity. The control for this experiment involved the tertiary antibody conjugated to unlabeled biotin. FRET strength $[{\alpha }_2∕({\alpha }_1+{\alpha }_2\left)\right]$ was used as a measure of the presence of a protein–protein interaction. For the negative control, FRET strength was not significantly larger than zero (Fig. 5A ) $(P=0.33)$ . For the double stained cells, FRET strength was 0.81. The lifetime of the second component was found to be $1305\pm 24\mathrm{ps}$ . If the background autofluorescence was not subtracted, a significant false positive FRET strength was observed, (Fig. 5B, $P=1.4\times {10}^{-8}$ ). The FRET strength in the model interaction dropped to 0.7414. The false positive result here illustrates the value of background correction in circumstances such as cell-based assays, in which the amount of available signal is limited compared to the background auto-fluorescence. Furthermore, the use of the StrEF is validated since it leads to the elimination of false positive results.

Fig. 5

Comparison of apparent FRET strength for a model protein–protein interaction, in a cell-based assay, compared to a negative control. Tubulin was immunostained with A430 and either Cy3 or biotin-labeled secondary antibodies. A student’s t-test of the amplitude of the negative control shows that ${\alpha }_2∕({\alpha }_1+{\alpha }_2)=0$ is not significantly greater than zero $(P=0.33)$ (Panel A). For comparison, the results from a simple bi-exponential fit without background correction is shown (Panel B). A significant false positive result is obtained if the background is not subtracted $(P=1\times {10}^{-8})$ .

3.6.

Analysis of Protein Conformation: Alzheimer’s Disease Associated with Presenilin-1 Protein

The question of PS1 conformation has been studied using time-domain FLIM microscopy by comparing apparent lifetime changes of donor fluorophores.¹⁷ Because it is a multipass transmembrane protein, its structure has been difficult to assess; a recent fluorescence lifetime microscopy study, however, showed that antibodies directed at dual epitopes of PS1 change their relative proximity to each other in mutant, compared to WT PS1.¹⁷ We show that our current plate reader-based analysis techniques allow us to detect this subtle change.

For the negative control (Cy3 was omitted), ${\alpha }_2∕({\alpha }_1+{\alpha }_2)$ was not statistically higher than zero ( $n=16$ wells), showing that there was no confounding second lifetime due to local variations in the cellular environment. The C- and N-termini of PS1 in the mutant cell line (C410Y) showed significantly stronger FRET than the wild-type (PS70) ( $P=0.002$ , $n=16$ wells per condition) (Fig. 6A ). A comparison of FRET efficiency shows that the interterminal distance is less in the mutant C410Y variant of PS1 $(P=0.00015)$ (Fig. 6B). These data confirm previous work that the mutant variety of PS1 has a measurable conformational difference with the C- and N-termini being closer together.

Fig. 6

Results of protein conformation assay of C- and N-terminal FRETing interaction. CHO cells were stably transfected with either mutant PS1 (C410Y cell line), known to cause FAD, or WT(PS70 cell line), and human APP. The C- and N-termini of PS1 were immunolabelled with A430 and Cy3. The mutant PS1 showed statistically significantly $(P=0.002)$ smaller C- and N-terminal distance (Panel A). There is also a significantly $(P=0.0001)$ increased FRET strength (Panel B).

3.7.

Analysis of Protein–Protein Interaction: Alzheimer’s Disease Associated with Presinilin-1 and Amyloid Precursor Protein

Berezovska also addressed the question of differences in the PS1-APP interaction between WT (PS70) and mutant (C410Y) forms of PS1. Here we test our ability to distinguish whether differences in FRET are due to closeness of the two fluorophores, which would indicate a difference in the PS1-APP closeness, or differences in the number of molecules interacting, which would indicate increased binding of PS1 to APP.

We show here that using this instrument, a difference in FRET strength was detected (Fig. 7A ), which describes the proportion of donor molecules interacting with the acceptor, implying that the distance between the two proteins, when bound, is the same for both species of PS1 but that mutant PS1 has a greater propensity for binding APP. We can detect no statistically significant difference in the PS1-APP distance between the WT and mutant variants of PS1 (Fig. 7B). In addition, we notice that the FRET strength for both cell types in this experiment is particularly low $(\approx 15\%)$ . These data demonstrate that this assay can detect protein–protein interactions in intact cells.

Fig. 7

PS1-APP protein interaction assay. CHO cells stably transfixed with mutant or WT PS1 and human APP. PS1 was immunolabelled with A430 (donor) and APP was labeled with Cy3 (acceptor). There was a difference $(P=0.0005)$ in FRETing strength (Panel A), but no statistical difference $(P=0.29)$ in PS1-APP distance between the WT and mutant varieties of PS1(Panel B). This result implies that there is greater interaction between mutant PS1 and APP but that the distance between PS1 and APP is similar between the mutant and WT PS1 conditions.