The current advances in light microscopy and engineering of fluorescent proteins (FPs) with improved properties and altered colors have provided the cellular biologist with the ability to investigate fine molecular events in living cells (1, 2). Properly combined, fluorescent molecules allow the study of protein conformational changes or inter protein interaction by means of fluorescence resonance energy transfer (FRET) (3, 4). When an excited donor fluorophore is proximate (=10 nm) and suitably oriented to an acceptor molecule, FRET occurs through a nonradiative dipole–dipole interaction. Currently, the most popular FP pairs for FRET analysis are cyan with yellow fluorescent protein (CFP/YFP) and green with red fluorescent protein (GFP/RFP) (5, 6).
Among the methodologies based on nonradiative resonance energy transfer, two approaches prevail, the fluorescence lifetime imaging microscopy (FLIM), and the fluorescence intensity-based FRET imaging. The FLIM is known to be cumbersome to require highly specialized equipment and as well to involve long acquisition time (7). To the contrary, fluorescence intensity-based FRET estimation provides higher speed of acquisition, requires lower equipment engineering, but remains challenging because of the abundant fluorescence spectral bleed-through (SBT) originating from donor and acceptor in the FRET channel. In addition to SBT, other sources of noise contaminate the FRET signals, including spectral sensitivity variations in donor and acceptor channels, autofluorescence, and detector and optical noise (8). Several methods have been developed to eliminate or decrease SBT as well as other sources of artifacts (9–11). However, despite FRET method improvements, extensive controls and complex FRET calculations are required to eliminate artifacts with a non-negligible risk to generate false positive results.
In this report, we applied two-photon spectral imaging microscopy to quantify FRET using an advanced unmixing method. Two-photon excitation sources advantageously produce no optical noise because of the usage of infrared pulse laser sources outside of the emission detection window. Spectral imaging for monitoring FRET is convenient, as simultaneously measured intensities of the donor and acceptor can be used for the ratiometric detection of FRET (12–14). Furthermore, upon linear unmixing, spectral imaging provides the opportunity to eliminate the donor SBT and nonspecific signals, such as autofluorescence from the FRET channel. However, the acceptor SBT cannot be removed from the FRET channel by linear unmixing, because both share the same spectral band.
To further simplify the FRET approach, we looked for an FP FRET pair that showed no SBT from the acceptor molecule. We found that two FRET pairs enabled selective excitation of donor but not of the acceptor by two-photon excitation. Both the CFP/YFP and YFP/mCherry have similar FRET efficiency as measured by fluorescence lifetime imaging; however, the later has better spectral separation. As a result, we show that FRET quantification using YFP/mCherry FRET pair can be achieved by means of single spectral acquisition combined with spectral unmixing and applied to intra- and intermolecular interaction quantification.
MATERIALS AND METHODS
Cell Culture and Transfection
Human embryonic kidney 293 cells also referred to as HEK-293 cells were grown in Dulbecco's modified Eagle's medium supplemented with 10% fetal calf serum and 1% penicillin–streptomycin at 37°C and 5% CO2 under humidified atmosphere. Cells were seeded 1 day before transfection in fibronectin-coated tissue culture 35-mm glass bottom plates (Mattek Corporation). Transfection of HEK-293 cells was performed at ∼80% confluency using 1 μg of plasmid mixed with Fugene 6 reagent according to the manufacturer specifications.
A 27-amino linker was used to connect the FPs CFP and YFP (15). The mCherry-YFP tandem was developed by substituting the CFP with the mCherry from the CFP-YFP construct. The human amyloid precursor protein (APP; OriGene, SC111589) and Fe65 cDNAs (OriGene, SC111588) were subcloned into mammalian expression vectors pEYFP-N1 and pmCherry-N1 (Clontech, BD Biosciences).
Spectral images were obtained using a filter-free spectral imaging system combined with a custom-built two-photon microscope. The key features of the spectral detection system comprise a prism for the dispersion of fluorescence light as well as an electron-multiplying charge-coupled device as detector. Our spectral imaging system allows for 80 spectral detection channels in the range of 400–750 nm with a sampling mode of 458 × 458 pixels and an acquisition time of ∼5 s per stack. A tunable Ti:sapphire laser (MIRA 900, Coherent) was used as the two-photon excitation source. All acquisitions were performed with a Plan-Apochromat 63×/1.4 oil objective. A detailed description of our setup is available elsewhere (16). To compare the excitation profiles of the candidates for FRET pair, we measured the spectral imaging of six FPs at different excitation wavelength ranging from 820 to 940 nm. We kept the laser power constant for each excitation wavelength in order to provide a comparative analysis of the measured fluorescence emission intensity. To investigate the dependence of emission intensity on the excitation wavelength, the emission values at the emission maxima for each FP were plotted as a function of the excitation wavelength (Fig. 1A).
Fluorescence Lifetime Imaging
FLIM experiments were performed with the same custom-built two-photon microscope used for spectral imaging. The Ti-sapphire mode-locked laser was tuned to 820 and 900 nm for CFP and YFP excitation, respectively. The CFP and YFP fluorescence emissions were collected through the band pass filters 470–500 and 535–585 nm, respectively, and detected using a photon counting photomultiplier tube (H7422P-40, Hamamatsu) and a time-correlated single photon counting module (SPC-830, Becker & Hickl). Fluorescence intensity and lifetime imaging were acquired sequentially for each sample. FRET efficiency, E was calculated as E = 1 − (donor lifetime in the presence of acceptor)/(donor lifetime in the absence of acceptor) (4).
We used, Image Mining 2.0, a custom-made image processing and analysis application built with an extensible plugin infrastructure (17). The plugin dedicated for spectral unmixing analysis was developed to offload spectral image stacks into a frame buffer for further visualization and data processing.
Algorithm-Based Method to Calculate FRET Efficiency
Given the spectrum of a fluorophore as fi(λ) where i = 1, 2, …, N represents the index of a fluorophore, and its length is M, the measured spectrum of each pixel can be written as , where Ci represents the concentration of each fluorophore. We can define I(λ) as a matrix in which F is the matrix of the references spectra
As a result, the fluorophore concentration can be calculated by finding the vector C that minimizes the least square error
such that for every i, where is the left-inverse matrix of F (18). From this analysis, we know that the computation of Ci depends on the reference spectra matrix F. However, acquiring an autofluorescence free reference spectra matrix is not a trivial task. Therefore, to alleviate the artifacts as much as possible, a three-step approach was applied:1) reference from known samples were acquired and saved in the reference spectra library;2) construction of the autofluorescence model for each image during runtime;3) use the reference spectra and autofluorescence spectrum for the FRET efficiency computation. The first two steps can be classified as matrix factorization problem. Considered for an entire multispectral image, it can be formulated as
where I is a matrix, K is the number of the pixels, and C is a matrix with N = 2. It is actually the same equation as Eq. (1); however, F is unknown. This matrix factorization problem can be solved with different methods, such as principal component analysis (PCA), vector quantization, and non-negative matrix factorization (NMF) (19, 20). We applied NMF to satisfy the assumption that each pixel of the image is the sum of registered signal and the unregistered signals, that is, . NMF method is based on an iterative procedure, which captures N main features from the samples as the primary components of the PCA procedure. A detailed description of our method applied to quantify the FRET index of YFP/mCherry tagged protein(s) is given as follows:
1Lambda stacks were obtained using our two-photon spectral imaging system from cells expressing YFP or mCherry. We assumed that the acquired spectrum for each FP was a mixture of several components;1) the pure spectra and2) the autofluorescence. Accordingly, we fixed N equal to 2. Without prior information and because the algorithm needs an initial guess, we modeled the initial value [F in Eq. 3)] of each fluorophore emission spectrum by a Gaussian shape model and the autofluorescence by a uniform random noise. In addition, all the elements in matrix C were set to zeros. After several iterations, both autofluorescence and spectra signals could be extracted. The spectra signals for YFP and mCherry spectra were saved as reference spectrum (Fig. 2; Stage 2) and used in the following steps.
2We next constructed an individual autofluorescence model (Fig. 2; Stage 3) for each image. The procedure was similar to step 1, except that the autofluorescence signal was this time saved as a reference spectrum specific for each spectral data.
3For each image, C was solved with F, which is composed of YFP, mCherry, and autofluorescence spectra, according to [Eq. (2); Fig. 2; Stage 3]. Then, the FRET efficiency for each pixel was calculated using the YFP and FRET (mCherry) components in C, which represent the area under the modeled reference spectra and calibration factor of our two-photon spectral imaging system (Fig. 2; Stage 4).
The area (A) and (B) correspond to the quenched donor emission spectrum and the processed FRET spectrum, respectively, whereas QD and QA are the quantum yield of donor and acceptor and γ is the calibration factor specific to the two-photon spectral imaging system. For the FRET efficiency calculation of YFP/mCherry pair, we used the published quantum yield of 0.61 (1) for the YFP and 0.22 (21) for the mCherry. The γ calibration factor depends on the spectral sensitivity of detector for donor and acceptor. Therefore, this factor was variable according to the FRET pair. In our case, the γ value was determined using the calculated FRET efficiency of mCherry-YFP (15.9%) obtained upon lifetime measurement. The average FRET efficiency measurement was additionally performed at the single cell level upon selection of a region of interest (ROI). To exclude pixels carrying only autofluorescence, a semiautomatic ROI algorithm was developed. First, a mask indicating the region of cells was created automatically based on k-means clustering method. We performed the cell contour segmentation using the unmixed donor signal image (A) as seen from Figure 2 (Stage 3). Second, the calculation of the mean FRET efficiency values within the ROIs was performed by taking into account both the cell contour segmentation and the ROI. This method was used to calculate the FRET efficiency distribution (Fig. 2; Stage 4).
Screening for a FRET Pair Free from Acceptor Bleed-Through
To determine the FRET acceptor with the lowest SBT, we acquired the spectra of six FPs transiently expressed in HEK-293 cells using a custom-built two-photon spectral imaging microscope (16). As seen from Figure 1A, the CFP emission was detected at excitation wavelengths ranging from 820 to 920 nm, and the YFP and GFP emissions were detected upon excitation from 840 to 940 nm. The DsRed or mTomato FPs were detectable at wavelength staring from 860 nm; however, the mCherry failed to be detectable at 900 nm excitation wavelengths and was faintly detectable at 940 nm. According to our results, when using 820 and 900 nm two-photon excitations, respectively, three possible FRET pair candidates such as the CFP/YFP, YFP/mCherry, and GFP/mCherry showed low acceptor SBT. The GFP/mCherry pair was excluded as a FRET pair based on its lower energy transfer efficiency due to the lower spectral overlap between GFP emission and mCherry (22) but as well to the lower two-photon emission signal compared to YFP upon 900 nm excitation. To evaluate the FRET efficiency of the CFP/YFP and the YFP/mCherry FRET pairs, we next expressed the CFP-YFP and mCherry-YFP tandems in HEK-293 and measured FRET using two-photon FLIM. The two-photon excitation wavelength for CFP/YFP pair and YFP/mCherry pair were 820 and 900 nm, respectively. Upon covalently linking CFP to YFP and YFP to mCherry, the respective donor's lifetime were found to be significantly reduced as shown on the lifetime images and lifetime distribution plots (Figs. 1B and 1D). The average lifetime value of cells expressing the CFP or YFP alone was 3.15 ns (mean ± SEM; n = 31) and 3.20 ns (mean ± SEM; n = 32), respectively. With the CFP-YFP and mCherry-YFP tandems, the donor lifetime was reduced down to 2.65 ns (mean ± SEM; n = 41) and 2.69 ns (mean ± SEM; n = 31), respectively. The calculated energy transfer efficiency for both tandem mCherry-YFP and CFP-YFP was identical with an average value of 16%. We next acquired a full emission spectra with two-photon excitation on HEK-293 cells expressing CFP-YFP or the mCherry-YFP tandem at the wavelength of 820 or 900 nm, respectively. As seen from Figures 1C and 1E, the better spectral separation of the YFP/mCherry, when compared with the CFP/YFP FRET pair, can be illustrated by a separation of 82 and 18 nm, respectively, between the donor and acceptor emission maxima. As a result, we selected the YFP/mCherry pair for the rest of the study.
We next examine using two-photon emission spectra, HEK-293 cells expressing YFP, mCherry, and the mCherry-YFP tandem at the excitation wavelength of 900 nm (Figs. 3A–3C). As shown in Figure 3B, we were unable to directly excite mCherry under the experimental conditions at 900 nm, but a faint emission of the mCherry was detectable at 940 nm (insert in Fig. 3B). In the case of mCherry-YFP tandem, both YFP and FRET spectral emissions could be detected (Fig. 3C). As a result, the FRET emission could be visualized as a bulge in the YFP emission spectra. As shown in Figures 3B and 3C, the emission maximum of the bulge corresponds to the mCherry emission peak matching the previously obtained spectra from mCherry overexpressing cells upon excitation at 940 nm.
Quantification of FRET by Spectral Unmixing Method
Encouraged by this result, we explored the possibility of quantifying the FRET component from the two-photon excited spectral imaging data, when using the YFP/mCherry FRET pair. We took advantage of the two-photon spectral detection that provides a full detection range of the visible spectrum free from the excitation source. This important quality facilitates accurate spectral unmixing of overlapping fluorescence emissions. We chose for this study a spectral unmixing method based on the NMF, less prone to nonspecific signal artifacts (23), which blindly separate registered signals from unregistered ones (19, 20).
As shown in Figure 2, the NMF approach was applied to distinguish the signals matching the FP YFP and mCherry spectra references from autofluorescence signals. The reference spectral signatures of the FPs were previously collected from cells overexpressing the YFP or mCherry FPs. We applied the NMF approach to build an autofluorescence spectrum model for each spectral data. The autofluorescence spectrum model was next used to resolve the linear unmixing equation together with the reference spectra (see Materials and Methods section). As a result, using the NMF-based unmixing approach, we were able to define the precise donor and FRET emission signals of the mCherry-YFP tandem construct.
Using the unmixed signals of YFP and FRET, we were able to calculate the real FRET efficiency for mCherry-YFP tandem construct. As a reference data, using the FRET efficiency of mCherry-YFP obtained from FLIM-FRET methods, we could calibrate our two-photon spectral imaging system. According to Eq. (4), calibration factor (γ) of our system for YFP/mCherry pair is ∼0.22. As seen in the final FRET efficiency image (Fig. 2; Stage 4), the mCherry-YFP tandem shows homogeneous FRET efficiency distribution within the cell as well as among the cells.
Application of FRET Spectral Imaging in Protein-Protein Interaction
To investigate the feasibility of our FRET method to quantify the interaction between independent proteins, we measured the interaction between the Fe65 protein and the Alzheimer's APP. The Fe65 protein is a member of a family of multidomain adaptor proteins that form multiprotein complexes. It has been shown that the PTB2 domain of Fe65 interacts with APP as well as with the APP intracellular fragment (AICD) released upon γ-secretase cleavage (24). Some reports show that Fe65-AICD complex translocate to the nucleus and participate in gene transcription events (25). The interaction of the APP C-terminus with the adaptor protein Fe65 mediates APP trafficking and is thought to regulate APP processing (25–27). HEK-293 cells co-expressing the APP and Fe65 proteins grafted with the YFP and the mCherry (Fig. 4A), respectively, were analyzed using the FRET efficiency to monitor the interaction between these two proteins. Although a clear emission spectra for APP-YFP could be detected at the excitation wavelength of 900 nm (Fig. 4B), no signal was measured for the Fe65-mCherry at the same wavelength (Fig. 4C), as expected from our previous results (Fig. 3B). Emission spectra from cells co-expressing APP-YFP and Fe65-mCherry showed both emission spectra originating from the YFP with a peak at 531 nm and the nonradiative energy transfer shoulder at 615 nm. Unfortunately, the cellular stoichiometry of the donor and acceptor can be variable for intermolecular FRET studies. In consequence, we monitored the FRET efficiency measured from cells co-expressing APP-YFP and Fe65-mCherry with variable ratios. As shown in Figures 4E and 4F, we found a clear correlation between the FRET efficiency value and the acceptor/donor stoichiometry. The maximum FRET efficiency plateau was found to be similar to what was obtained with the mCherry-YFP tandem, forecasting a strong protein–protein interaction between APP and Fe65.
In this study, we propose an alternative FRET microscopy method to study protein–protein interaction by combining two-photon spectral imaging and using a FRET pair devoid of acceptor SBT. In classic FRET, the fluorescence emitted upon excitation of the acceptor at the donor excitation wavelength is superimposed with the FRET signature. Both multiwavelength excitation of the donor and acceptor as well as algorithm-based donor correction are required to determine the presence of FRET (9–11). When using two-photon microscopy, the difficulty to operate multiwavelength excitation imaging stems from the fact that most of two-photon microscopes are fitted with only one manually operated femtosecond laser. The use of the YFP/mCherry or CFP/YFP FRET pair avoids acceptor SBT and thus enables FRET study with a single spectral acquisition (Figs. 4G and 4H). Furthermore, this imaging approach reduces FRET calculations prone to generate artifacts.
To remove the unwanted autofluorescence signals with unknown fluorescence spectra, we applied a NMF method. Such a method was shown to be particularly well adapted to eliminate tissue autofluorescence (23). In particular, the ratio of autofluorescence and specific fluorescence intensities increased with the depth of the explored tissue. Altogether, the NMF method, combined with two-photon microscopy and the red-shifted FRET pair (YFP/mCherry), shows promising application in deep tissue FRET studies. We found that the better spectral separation seen for the YFP/mCherry pair provided advantages related to the unmixing approach for FRET quantification.
We applied the FRET efficiency to quantify donor and acceptor molecular proximity. The FRET efficiency calculation showed a robust FRET signal when using the YFP/mCherry tandem molecule. Such results enable the possibility of adapting the current available tandem molecules, built with linker sequences particular to protease or binding domains, to serve, for example, as intracellular reporters for protease activity or calcium concentration. The development of mouse transgenic models expressing a YFP/mCherry-based intracellular molecular sensor, together with two-photon spectral imaging using FRET microscopy, would improve in vivo and ex vivo deep tissue molecular studies.
To quantify donor and acceptor proximity between two independent molecules, we focused on the interaction of the APP and the adaptor protein Fe65. The carboxy-terminal PTB domains of Fe65 and APP were previously shown to associate using the yeast two hybrid systems (28). This result was as well confirmed in live cells using fluorescence intensity based FRET studies (29, 30). Our results not only confirmed those previous experiments but also quantified the extent of FRET between APP and Fe65 and measured a strong FRET dependence with the ratio of acceptor/donor molecules. Subcellular analysis of the FRET distribution points up high interaction levels of the two proteins preferentially located in the plasmalemma. It is important to note that we used γ-secretase inhibitors to prevent the production of AICD and thus its interaction with Fe65. Altogether not only our results confirmed that APP and Fe65 proteins do interact (25–27) but demonstrated the ability of the FRET microscopy method to quantify protein-protein interaction in living cells.
In summary, the FRET efficiency approach can possibly be extended to other suitable FPs or fluorescent dyes. For example, considering the two-photon spectra absorption properties of TagRFP, we believe that this protein could substitute mCherry (31). We anticipate that continued development of this method should make our FRET efficiency applicable to nonspectral microscopy upon recording fluorescence of YFP and mCherry at their emission maxima.
Grateful acknowledgment is also due to Yoonjin Ha, whose sensitive reading of the manuscript greatly improved it.