Highly-efficient quantitative fluorescence resonance energy transfer measurements based on deep learning

1. Introduction

Fluorescence resonance energy transfer (i.e., fluorescent resonance energy transfer or Förster resonance energy transfer, FRET) involves a nonradiative transfer of energy from an excited donor fluorophore to a nearby acceptor fluorophore.¹ It can be used to measure the molecular distance on the scale of nanometer, which is far below the resolution of optical microscopy.^1,2 Therefore, FRET is widely used in the evaluation of molecular interactions.^3,4,5,6 There are mainly two FRET methods, intensity-based and fluorescence-lifetime FRET methods.⁷ Although intensity-based FRET is more preferable if the measurements of donor–acceptor stoichiometry and fast dynamic FRET changes are needed, it requires complicated experimental process and manual image analyses to obtain precise results, which impedes FRET experiment’s efficiency promotion and restrict its application.⁷

In experimental process, crosstalk is a major obstacle to the precision of current intensity-based quantitative FRET method.^8,9 The crosstalk mainly consists of two parts: the donor-emission crosstalk and acceptor-excitation crosstalk. The donor-emission crosstalk is the emission energy of the donor in the range of the receptor emission wavelength. The acceptor-excitation crosstalk means that the receptor can be activated in the excitation wavelength of the donor. In order to reduce the crosstalk, several methods have been proposed.^{7,10,11,12,13,14} One of the most precise quantitative FRET methods is the three-cube FRET.^7,10 During three-cube FRET experiments, in addition to the experimental group which contains both donor and acceptor, samples of donor-only group and acceptor-only group need to be prepared as the reference groups. Donor-only and acceptor-only references are measured to obtain exact value of crosstalk.

In image analysis process, measurement regions should be carefully chosen for the following reasons:

First of all, FRET is widely used to evaluate binding affinity and distance between functional molecules in living cells. The FRET efficiency (EA) is determined by the concentration of free (unbound) donor molecules [D${}_{free}$], the dissociation constant (K${}_{\text{d}})$, and the maximal FRET efficiency (E${}_{A,MAX})$.^11,15 K_d and E${}_{A,MAX}$ are two important constants which reflect the binding affinity of two molecules and the distance between the bound donor and acceptor, respectively. To obtain these two constants, the relationship between [D${}_{free}$] and E${}_A$ should be fitted with a binding curve (details could be found in the Ben-Johny, Yue and Yue 2016¹⁵) Since E${}_A$ does not varies linearly with [D${}_{free}$],¹¹ if the measurement region is composed of several parts with different [D${}_{free}$], the data point ([D${}_{free}$], E${}_{\text{A}})$ will not fall on the proposed fitting curve. Therefore, one measurement region should never contain parts with different [D${}_{free}$] values. The distribution of D${}_{free}$ in measurement regions needs to be as uniform as possible. As [D${}_{free}$] is determined by both total number of donor and acceptor, the distribution of donor and acceptor in one region both needs to be uniform. A whole cell area could not be selected as one region because usually the distribution of functional molecules targeted with donor/acceptor is nonuniform in the cell.

Second, if a small area of measurement region is used to obtain one data point for E_A curve fitting, it is hard to obtain precise results because of the material exchange between the measurement region area and surrounding area during the imaging process. Therefore, relative large and constant areas should be selected as one region to obtain a single data point.

Third, to ensure the precision of the result, fluorescent “clusters” or “cavities” in cells should be avoided. The former will lead to nonspecific FRET and the latter will reduce the signal-to-noise ratio for the final results.¹⁶

Last but not the least, background should be corrected especially in cells where the fluorescent proteins are weakly expressed, so a background region should also be selected.

For the reasons above, the precision of FRET will be affected if the measurement regions are not chosen suitably. As a summary, the selected fluorescent regions should have the following properties: (1) Uniform distribution of the donor and the acceptor; (2) Relatively large and constant areas; (3) No “clusters” and “cavities”.

As region selection before fitting of the results is time-consuming, a precise and time-efficient region selection method is very important and meaningful especially when dealing with a large number of images.

To simplify the experimental process, we check many fluorescent proteins which can be used in FRET. Photoactivatable fluorescent proteins (PAFPs) which are widely used in high-resolution optical techniques.^17,18,19 can be switched on by violet light irradiation. The donor-only reference obtained process in three-cube FRET introduced above can be omitted with the PAFPs auxiliary, which is similar to photobleaching FRET (pbFRET).^7,9,20 However, with the reference obtained in advance, PAFPs auxiliary FRET can be more robust than pbFRET.

As for image analysis process, currently, the existing image analysis techniques for fluorescent image segmentation include classical techniques and deep learning (DL) techniques.²¹ Classical image segmentation techniques include the threshold-based, edge-based, region-based techniques.²² In fluorescent image segmentation researches, the mainly used techniques are the threshold-based segmentation which is optimized by automatic threshold selection²³ and the watershed segmentation^24,25 which is optimized to morphological watershed segmentation.

DL has many successful applications in biomedical fields²⁶ such as magnetic resonance imaging (MRI),²⁷ computed tomography images,²⁸ ultrasound images,²⁹ photoacoustic imaging.³⁰ In microscopic cell level researches, DL has been used for cell segmentation.³¹ and cell nuclei detection,³² but all the algorithms are used to find the edge of cells or organelle.²¹ For the four reasons mentioned above, in FRET experiments only cell segmentation by finding out the cell boundary is not enough. An algorithm is needed to detect the fluorescent regions with the three properties listed above to increase the accuracy and precision of FRET experiments. Automatic region selection could be achieved if DL is applied to FRET. However, thousands of labeled training samples are required in traditional convolutional neural network (CNN) training,³³ which are difficult to obtain in living cell experiments. U-Net can use the available labeled samples more efficiently and is particularly useful in biomedical images.³³ Moreover, the typical output of CNN from an image or a group of images is a single class label, but the output of U-Net is the class label of each pixel, which is very suitable for measurement region selection.

In this paper, a simplified and automatic FRET (saqFRET) method is presented to improve the efficiency of FRET. Our main contribution is the DL auxiliary image analysis process, which not only speed up this process but also release ambiguous and laborious manual labeling work. Comparisons with traditional segmentation methods and cell boundary segmentation based on DL are conducted to validate the necessity of the proposed fluorescent region selection method. Besides, we also propose a new pair of fluorescent proteins for quantitative FRET measurements to simplify the experimental process. With the positive control group and negative control group experiments, comparisons with traditional method are conducted to validate the precision and efficiency of the proposed method in living cells. FRET results of the same samples at different time points were plotted to further demonstrate the feasibility of the proposed method.

2. Materials and Methods

2.1. Cell culture and transfection

HEK 293T cells were plated in confocal dishes ($\mathrm{\Phi }=35$mm) at a density of $5\times 10^4$ cells/dish in Dulbecco’s modified Eagle’s medium (DMEM) supplemented with 15% fetal calf serum (FCS), 5% Penicillin-Streptomycin, and cultured with 5% CO₂ at $37^{\circ }$C in a humidified incubator. Cell proliferation took one day before transfection.

The plasmids used for transfection are PA-mCherry1 (referred to as PA-mCherry), enhanced green fluorescent protein (EGFP), and PA-mCherry-linker-EGFP. As a monomeric red photoactivatable protein, PA-mCherry has excellent photoactivation contrast over background, advanced photostability and fast photoactivation.¹⁹ Therefore, PA-mCherry was used as the acceptor in this study, while EGFP was used as the donor for its photostability. The coding sequences of PA-mCherry and EGFP were amplified from PA-mCherry1-C1 (addgene 54495) and EGFP-pBAD (addgene 54762) using polymerase chain reaction (PCR). The PCR products were purified, digested with the restriction enzymes BamHI and XbaI, and ligated in the BamHI/XbaI sites of pcDNA3.1.

For transfection, HEK 293T cells were incubated with a mixture of plasmids and Lipofectamine${}^{\text{T}\text{M}}$ 2000 (Invitrogen, Carlsbad, CA, USA) for 4–6h before being replenished with fresh medium. HEK 293T cells co-expressing EGFP and PA-mCherry were used as the negative controls or PA-mCherry + EGFP mixtures, while HEK 293T cells co-expressing PA-mCherry-EGFP fusion protein were used as the positive controls or PA-mCherry-EGFP dimers.

In the positive control group experiments, the fluorescence intensity ratio between PA-mCherry and EGFP was 0.87 ± 0.12 (mean ± standard deviation). Accordingly, the transfection ratio between PA-mCherry and EGFP in the negative control group experiments was adjusted to obtain the same fluorescence intensity ratio between PA-mCherry and EGFP. In our experiments, the transfection ratio between PA-mCherry and EGFP in the negative control group experiments was between 2 and 3.

2.2. Confocal imaging

Imaging experiments were performed on a Nikon A1R (Nikon Instech Co., Ltd., Tokyo, Japan) confocal laser scanning microscope (CLSM). Samples were placed in Nikon microscopy living cell culture environment with 5% CO₂ at $37^{\circ }$C. And the cells were kept in DMEM with 15% FCS and 5% Penicillin-Streptomycin. A $100\times$ (oil) microscope objective lens was used to obtain clear fluorescent images of the cells. By changing the field of views, different groups of results can be obtained in one dish of cells. To accelerate the process of photoactivation, a region of interest (ROI) of $64\times 64$ pixels was selected in each field of view ($512\times 512$ pixels). This ROI was scanned pixel to pixel by lasers during the process of photoactivation and photographing.

Table 1 shows the imaging lasers and filters used in this study. Here four imaging channels were defined: the activation channel, the donor channel, the FRET channel and the acceptor channel. The activation channel was used to active the photoactivatable acceptor. The donor channel (CHD) was used to measure the fluorescence intensity in the condition of donor activation and donor emission. The FRET channel (CHF) was used to measure the fluorescence intensity in the condition of donor activation and acceptor emission. The acceptor channel (CHA) was used to measure the fluorescence intensity in the condition of acceptor activation and acceptor emission.

In the fluorescence imaging experiments on the CLSM, first, a field of view with a distinct living cell was chosen, and an ROI was selected which contains a large part of a cell and a small area of background. Then the images of CHD, CHF and CHA were taken successively. This process obtained the donor crosstalk in CHF and CHA. If the fluorescence in CHA was not negligible, the acceptor had already been activated, and thus the chosen field of view should be abandoned and a new field of view should be chosen. Next, over 80% acceptors were photoactivated by violet light irradiation with a total laser scanning time of at least 1 min. After the acceptors were photoactivated, the images of CHD, CHF and CHA were taken again to obtain the images with both the donor and acceptor fluorescing. The FRET values can be calculated from CHD and CHF before photoactivation and CHD, CHF and CHA after photoactivation. Six images were obtained in each group of experiments.

The fluorescence intensity obtained from each image is calculated by the average intensity of the fluorescent region minus the average intensity of the background region.

In order to maintain the consistency of the imaging conditions, all the parameters in the same group of experiments remained unchanged, including the laser intensities, pinholes and imaging speed of each channel. EGFP has high extinction coefficient and stability. Therefore, the excitation wavelength of the donor can be hypsochromically shifted appropriately (to 405nm in this study). At this donor excitation wavelength of 405nm, PA-mCherry is hardly excited in the donor channel or the FRET channel (Figs. 2(a) and 2(d)).

2.3. Convolutional neural network

U-Net is a CNN which is suitable for image-to-image problems.³³ In this study, two U-Nets were built using Keras [https://github.com/fchollet/keras] based on TensorFlow (Google, Mountain View, CA, USA) to automatically select the fluorescent regions and the background regions, respectively. The calculations were performed on central processing unit (CPU) and graphics processing unit (GPU) separately in our work. The CPU used is Intel(R) Core(TM) i5-7500 @ 3.40 GHz 3.41 GHz. The GPUs used are cloud GPUs from floydhub [https://www.floydhub.com].

2.3.1. Architecture and training set preparation for U-Net

As shown in Fig. 1 in the red box, the U-Net consists of a contracting path and a symmetric expanding path, with shortcut connections except for the main connections. The main connections are composed of convolutional layers, batch normalization layers, and max pool layers for contracting path or upsampling layers for symmetric expanding path. Shortcut connections connect the contracting path and the symmetric expanding path to perform identity mapping.

The results of each group of experiments in Sec. 2.2 constitute a sample. Therefore, each sample contains images of CHD and CHF before photoactivation and images of CHA, CHD and CHF after photoactivation. The dataset of the proposed network consists of approximately 230 independent samples. They were randomly divided into two sets: 86% as the training set, and 14% as the test set. The reported results and images in the following content were all from the test set. All images in the dataset were transformed to $64\times 64$ pixels and 8 bits.

2.3.2. Selection of fluorescent-region

For each group of images, the training data were labeled with two masks, i.e., one for images before photoactivation, and the other for images after photoactivation. These two masks corresponded to the same region with compensation for the outline and cell content movements during photoactivation. The masks indicated regions with more than 100 pixels and uniform intensity. Fluorescent “clusters” or “cavities” in cells were avoided when the masks were determined. The dataset of the images from both the CLSM and the masks was augmented by image rotation around the center. After rotation from $0^{\circ }$ to $360^{\circ }$ with an interval of $30^{\circ }$, the data was expanded by 12 times. The U-Nets learn to select regions with the same properties as the masks by updating their internal parameters during the training phases to minimize the error between the predictions and the associated masks. In the test phase, the separate set of data with masks as labels is used to evaluate how well the U-Net generalize to data not seen during the training phase. The U-Net was trained for 30 epoches. Stochastic gradient descent with a batch size of 8, Adam optimizer with a learning rate of 0.0001,³⁴ an activation function of sigmoid and a loss function of dice coefficient loss were used.

2.3.3. Selection of background-region

The method for background-region selection is similar to fluorescent-region selection, except that only one mask for each group was needed since deformations of the samples and content flowing in the cells have no influence on the background. In the image analysis, the background-region should also be carefully selected especially for low-intensity images. The blurring or light spreading around the cell will increase the background value, and therefore decrease the obtained fluorescence intensity and bring errors to the final results.

2.4. System design

The flow chart of saqFRET is shown in Fig. 1. No additional donor-only or acceptor-only reference group experiment is needed in saqFRET because of the use of the photoactivatable acceptor and the optimized excitation wavelength of the donor. The imaging experiment on the CLSM is shown in the green box in Fig. 1.

As shown in Secs. 2.2 and 2.3, each group of experiments obtained six images, i.e., the images of CHA, CHD and CHF before and after photoactivation, respectively. These images and the manual selected masks of the measurement regions comprise a CNN input sample. Two U-Nets are trained, one for the fluorescent-region selection and the other for the background-region selection. Then, with background values subtracted, the average values in the fluorescent regions are calculated as D_CHD, D_CHF, DA_CHD, DA_CHF and DA_CHA, respectively, where D stands for only the donor fluorescing and DA stands for both the donor and acceptor fluorescing. These values are used in the calculation of FRET.

2.5. FRET algorithms

There are two criteria in the evaluation of FRET. One is the energy transfer efficiency (E),² and the other is the FRET ratio (FR).^10,35,36 E and FR could be converted to each other through the extinction coefficient or the quantum yield of the donor or the acceptor according to the definition of FR. In this study, an FR calculation algorithm is proposed for saqFRET. There are many methods for the calculation of the FRET value. Because photoactivatable acceptor is used in the proposed experiment to correct the donor crosstalk, we compare the proposed algorithm with pbFRET algorithm, which is widely used in biological researches. In the traditional pbFRET algorithm, the FR is calculated as follows.^37,38,39,40 :

where DA_CHD stands for the fluorescence intensity of the donor when both the donor and acceptor are activated, while D_CHD stands for the fluorescence intensity of the donor in the presence of nonactivated or photobleached acceptor, i.e., reference. Here, the value of D_CHD is obtained from another time point, i.e., from another fluorescence status compared to DA_CHD. For example, in pbFRET, D_CHD is obtained after the acceptor has been photobleached; and in saqFRET, it is obtained before the acceptor is photoactivated. The numerator and the denominator are in different fluorescence status. If any fluorescent protein photodegradation happens, calculation of FR using Eq. (1) leads to relatively large error.

In saqFRET, FR is calculated as follows :

where DA_CHF stands for the fluorescence intensity obtained on the FRET channel when both the donor and acceptor are activated, while D_CHF stands for the fluorescence intensity obtained on the FRET channel in the presence of nonactivated or nonphotobleached acceptor. $\frac{\mathrm{\text{D}}\text{\_}\mathrm{\text{CHF}}}{\mathrm{\text{D}}\text{\_}\mathrm{\text{CHD}}}$ is the ratio of the crosstalk of the donor on the FRET channel and the fluorescence intensity of the donor on the donor channel.

After activating the acceptor, the crosstalk of the donor on the FRET channel can be calculated as $\mathrm{\text{DA}}\text{\_}\mathrm{\text{CHD}}\times \frac{\mathrm{\text{D}}\text{\_}\mathrm{\text{CHF}}}{\mathrm{\text{D}}\text{\_}\mathrm{\text{CHD}}}$. Besides, because the acceptor does not have any crosstalks on the FRET channel in saqFRET, the actual FRET energy could be calculated as $\mathrm{\text{DA}}\text{\_}\mathrm{\text{CHF}}-\mathrm{\text{DA}}\text{\_}\mathrm{\text{CHD}}\times \frac{\mathrm{\text{D}}\text{\_}\mathrm{\text{CHF}}}{\mathrm{\text{D}}\text{\_}\mathrm{\text{CHD}}}$.

The actual FRET energy varies with the fluorescence intensity among different samples. Therefore, the FR value obtained by Eq. (2) is presented as an intensity ratio for evaluation of the FRET efficiency.

In saqFRET, if any fluorescent protein photodegradation happens, the error generated in Eq. (1) is reduced for the following reasons. First, from (D_CHF/D_CHD) the crosstalk of donor on FRET channel is obtained. This crosstalk is a constant because it is related to the spectral characteristics of fluorescent protein, which is not changed when photodegradation happens. Second, D_CHF and D_CHD are both obtained before photoactivation, so no fluorescent protein photodegradation happens. Third, with (D_CHF/D_CHD) providing the crosstalk ratio, DA_CHF and DA_CHD are both obtained after photoactivation of PA-mCherry, with protein photodegradation happening. Therefore, the proposed algorithm will reduce the fluorescent instability caused by the experimental process or the change of environment.

All results in this paper are presented as $\mathrm{\text{means}}\pm \mathrm{\text{standard}}$ deviations.

3. Results

3.1. Characterization of saqFRET pair

To simplify the FRET experimental process, EGFP and a photoactivatable fluorescent protein PA-mCherry are used as the donor and the acceptor, respectively, as shown in Fig. 2. The overlap between the emission spectrum of EGFP and the absorption spectrum of PA-mCherry after photoactivation is indicated by the blue slash line in Fig. 2(a). To reduce the crosstalk of the acceptor, the excitation wavelength of EGFP hypsochromically shifts to 405nm (purple arrow, Fig. 2(a)). In the process of photoactivation, the photoactivated proportion of PA-mCherry against the activation time can be fitted with an exponential curve.¹⁹ Around 80% PA-mCherry can be activated after violet scanning irradiation for 60s (Fig. 3(b)). It should be noted that the total scanning time on the Nikon A1R confocal laser scanning microscope during photoactivation is 60s. However, the actual activation time for every area is only several hundred milliseconds when PA-mCherry is photoactivated by 405nm laser. Thus, the photodegradation caused by 405nm laser can be ignored. It indicates that EGFP and PA-mCherry can be used as the saqFRET pair and the photoactivation time can be set to 60s in our experiments.

The crosstalk based on the donor and acceptor above was measured to evaluate the crosstalk reduction effect of this method. The measurement regions were selected by skilled experimenters. In Fig. 2(d), A_CHD, A_CHA, A_CHF stand for the fluorescent values of the images obtained from the donor channel, the acceptor channel and the FRET channel when only the acceptor is transfected, respectively. D_CHD, D_CHA, D_CHF stand for the fluorescent values of the images obtained from the donor channel, the acceptor channel and the FRET channel when only the donor is transfected respectively. As shown in Figs. 2(c) and 2(d), the acceptor crosstalk is $0.010\pm 0.006$ on the donor channel and is $0.021\pm 0.008$ on the FRET channel. The donor crosstalk on the acceptor channel is as small as $0.0001\pm 0.0016$ because the excitation laser wavelength of 561 nm is beyond the EGFP excitation spectrum. The imaging parameters of the donor-only and acceptor-only experimental groups in Fig. 2(c) are the same as those in the experiments that co-express the donor and the acceptor. Therefore, only the donor crosstalk on the FRET channel should be concerned in saqFRET (Figs. 2(c) and 2(d)), which is $0.330\pm 0.008$, while all the other crosstalks can be omitted. The crosstalk of the donor on the FRET channel can be corrected by the ratio of $\frac{\mathrm{\text{D}}\text{\_}\mathrm{\text{CHF}}}{\mathrm{\text{D}}\text{\_}\mathrm{\text{CHD}}}$ according to Eq. (2).

3.2. Precision and efficiency of automatic region selection method based on U-Net

To efficiently obtain precise average fluorescence intensity of the image from each channel, an automatic region selection method based on U-Net was proposed in this study. To evaluate the performance of this method, one positive control and one negative control were employed for comparison. The positive control is manual region selection by skilled experimenters and the negative control is random region selection. The random region selection method adopts the same background region as in the manual region selection method (Fig. 3(b)(6)), and randomly selects a fluorescent region in the cell (Fig. 3(b)(3)). Representative results with the FR distribution are shown in Fig. 3. This representative sample did not move or have flowing cell content during the experiment so a precise FR distribution could be obtained, which is not a usual case in experiments. Thus, it is selected as the sample to use the FR distribution as an assessment method. With the original cell presented in Figs. 3(a) and 3(b) shows that the regions selected by the U-Net-based method have a relatively large area (more than 100 pixels) and uniform fluorescence intensity. Two fluorescent regions are selected. They are both with large areas, with similar fluorescence intensity and are both in cytoplasm, so the FR results calculated from these two regions can be considered as one data point (Sec. 1). Besides, these regions exclude fluorescent “clusters” or “cavities” in cells, as mentioned in the Secs. 1 and 2.3 (Fig. 3(b)(2)). Meanwhile, the manual region selection method also selects the region with these properties (Fig. 3(b) (1)), but randomly selected region contains nonuniform fluorescence (Fig. 3(b)(3)). In addition to fluorescence intensity distributions, the FR distributions in cells are also nonuniform due to the existence of nuclei, organelles and physiological process (Fig. 3(c)). The FRET results calculated from the FR distribution maps are shown in Table 2. The FR distributions of the regions selected by the U-Net-based method and the manual selection method are both uniform with standard deviations of 0.059 and 0.061 respectively, while the FR standard deviation of the random region selection method is 0.144 (Fig. 3(d), Table 2). The background regions selected by the automatic region selection method are as precise as those selected by the manual region selection method (Figs. 3(b)(4) and 3(b)(5)). The FR values calculated by the U-Net-based, manual and random region selection methods are 0.486, 0.477 and 0.459, respectively (Table 2). These results show that automatic region selection based on U-Net can select regions as precisely as manual region selection by skilled experimenters while random region selection cannot.

To further validate the advantage of the proposed method over other automatic segmentation methods, Fig. 4 shows the statistical results of FR from all the test samples. The FR is calculated with regions selected by the three region-selection methods above, the existing DL-based cell segmentation method, and the traditional segmentation algorithms — threshold-based segmentation and morphological watershed segmentation. In the existing DL cell segmentation, the network is also based on U-Net, but segmentation masks of the training data are selected according to cell edges in fluorescent images. In threshold-based segmentation, the threshold is automatically obtained by Otsu’s method.²³ The morphological watershed segmentation consists of two principal stages: a coarse segmentation and a fine segmentation which extracts the precise cell boundaries. The fine segmentation stage is preceded by the gradient extraction and modification based on the combined results of the coarse segmentation.²⁵ In dimer experiments, the FR calculated with regions selected by the automatic region selection method has no significant difference from that with regions selected by the manual region selection method using Welch’s $t$-test (unpaired; two-tailed with criteria of significance: *$p<0.05$) (Fig. 4(a)). However, the FR calculated with regions selected by threshold-based segmentation, watershed segmentation, DL cell segmentation and the random region selection method all have statistically significant difference from that with regions selected by the manual region selection method at $p<0.05$ (Fig. 4(a)) using Welch’s $t$-test In mixture experiments, there is no significant difference between the FR calculated by manual region selection method and those calculated by other methods, because the FR variation in mixture experiments is very large due to different donor/acceptor ratio. Qualitative comparison among different image analysis methods on examples of dimer experiments are presented in Fig. 4(b). The results above indicate that the proposed automatic region selection method based on U-Net can select regions for precise FR calculation as the manual region selection method, while the other methods cannot.

To evaluate the efficiency of the proposed automatic region selection method in image analyses process of FRET, the time used for region selection by the manual and automatic methods is compared. With the input of images of D_CHD, D_CHF, D_CHA, DA_CHD, DA_CHF and DA_CHA, the manual region selection method needs 40s to 1min for a single sample, because all six images should be concerned. And if there are morphological changes during the photoactivation time, the selected region should be adjusted accordingly. That means that if there are 100 samples, at least 4000s are needed for the process of manual region selection. Table 3 shows the time used for the automatic region selection method based on U-Net. It indicates that the process of region selection is sped up by more than 100 times by using the proposed automatic region selection method when the number of samples is above 100. Besides, it is shown in Table 3 that GPU performs faster than CPU, and the efficiency of the automatic region selection method is more pronounced when the dataset is larger. The time consumptions of threshold-based segmentation, watershed segmentation and DL cell segmentation are close to those of our method.

3.3. Estimation of the performance of saqFRET in living cells

To test the feasibility of saqFRET in living cells, FR values of the same samples at different time points are plotted. As shown in Fig. 5, the normalized FR values of the dimer samples are recorded at 60, 90, 120 and 180s after photoactivation, respectively, to test the stability of saqFRET over time. Figure 5 indicates that the normalized FR values of the dimer samples at 60, 90, 120 and 180s have no significant difference from each other ($p>0.05$), which means that saqFRET is stable over time.

4. Discussion

In saqFRET, EGFP and PA-mCherry are used as the FRET pair to eliminate the crosstalk. Nevertheless, an improved photoactivatable fluorescent protein that can more easily be photoactivated is still needed. On the one hand, if the photoactivation time is shortened, the status of fluorescent proteins and cells will be less changed. On the other hand, the FRET experimental time will be reduced and the efficiency of saqFRET will be increased.

Except for rapid FRET quantification, there are additional benefits of the automatic region selection method based on U-Net. It can expand the selected regions to unconnected areas that also meet the requirements for fluorescent regions (Fig. 3(b)(2)). In the 64 × 64-pixel image of Fig. 3(a), the automatic region selection method selects an area of 493 pixels, while the manual region selection method selects an area of only 149 pixels (Table 2).

CNN is used to select specific regions that have the following properties: relatively large areas, uniform distribution, no fluorescent “clusters” or “cavities”. These properties are necessary to ensure the measured FRET process having uniform reaction environment and obtain precise average of fluorescence intensities in the images. In living cell fluorescence imaging, CNN has been used in cell segmentation and cell nuclei detection, but to the best of our knowledge, has not been used to select measurement regions. As shown in Fig. 4, if the whole cell area is segmented as the region to calculate FR, the result will be not as precise as the result obtained by the manual selecting regions. This region-selection application of CNN can be extended to many other biological fields to accelerate image analyses processes on the condition that enough training set can be obtained.

In this study, samples of PA-mCherry-EGFP dimer and PA-$\mathrm{\text{mCherry}}+\mathrm{\text{EGFP}}$ mixture are used to validate the feasibly of saqFRET. However, as the positive controls and negative controls of FRET, the dimer and mixture are the simplest cases because only fluorescent proteins are transfected in living cells. These fluorescent proteins tend to have uniform distribution in living cells. If physiological process is involved, region selection has a more significant effect on the precision of FR calculation, because the fluorescent distribution is more nonuniform and the functional proteins are more likely to form clusters. Clusters must be avoided in the selected regions from all the channels in case of nonspecific FRET. Therefore, our future work will be applications of saqFRET in practical biological studies.

5. Conclusions

In this paper, a highly-efficient quantitative FRET method is proposed, which simplifies and accelerates both the experimental process and the image analyses process while maintaining the precision of results. On the one hand, by employing photoactivatable acceptor to solve the crosstalk of the donor and optimization of donor excitation wavelength to solve the crosstalk of the acceptor, the experimental process to obtain precise FR is simplified. By analyzing the components of FRET process, an FR algorithm is proposed to make saqFRET more sensitive and stable. On the other hand, an automatic measurement-region selection method based on U-Net is proposed, which can accelerate the region selection process by two orders of magnitude for more than one hundred samples, while the obtained FR values have no significant difference from those using manual selection by skilled experimenters. The simplicity and automation make the proposed saqFRET method more efficient and extensible in the study of high throughput cell biology and many other FRET-used situations.

Conflicts of Interest

The authors declare no conflict of interest.

Acknowledgments

The authors thank Jinyu Wang and Huizhen Cao in the SLSTU-Nikon Biological Imaging Center, for their assistance in microscope operation. This work was supported in part by the National Natural Science Foundation of China (61871251 and 61871022) and Sichuan Science and Technology Program (2019YFSY0048).