A rapid analysis platform for investigating the cellular locations of bacteria using two-photon fluorescence lifetime imaging microscopy

Facultative intracellular pathogens are able to live inside and outside host cells. It is highly desirable to differentiate their cellular locations for the purposes of fundamental research and clinical applications. In this work, we developed a novel analysis platform that allows users to choose two analysis models: amplitude weighted lifetime (τA) and intensity weighted lifetime (τI) for fluorescence lifetime imaging microscopy (FLIM). We applied these two models to analyse FLIM images of mouse Raw macrophage cells that were infected with bacteria Shigella Sonnei, adherent and invasive E. coli (AIEC) and Lactobacillus. The results show that the fluorescence lifetimes of bacteria depend on their cellular locations. The τA model is superior in visually differentiating bacteria that are in extra- and intra-cellular and membrane-bounded locations, whereas the τI model show excellent precision. Both models show speedy performances that analysis can be performed within 0.3 s. We also compared the proposed models with a widely used commercial software tool (τC, SPC Image, Becker & Hickl GmbH), showing similar τI and τC results. The platform also allows users to perform phasor analysis with great flexibility to pinpoint the regions of interest from lifetime images as well as phasor plots. This platform holds the disruptive potential of replacing z-stack imaging for identifying intracellular bacteria.


Introduction
Fluorescence lifetime imaging microscopy (FLIM) has been developed for detecting bacterial infections in clinical applications. Many imaging methods are dependent on fluorophore-labelled tracers that interact with bacterial surface structural components such as lipopolysaccharide or bacteria enzymes/proteins such as β-lactamase [1]. On the other hand, bacterial intrinsic fluorescent molecules have also been exploited for detection, such as porphyrins, a redfluorescent by-product of bacterial haem production, and cyan-fluorescing pyoverdines, which are fluorophores specific to Pseudomonads [2]. Most interestingly, two-photon FLIM imaging of the metabolic coenzymes reduced nicotinamide adenine dinucleotides [NAD(P)H] has been used for a separate analysis of host and pathogen metabolisms during intracellular chlamydial infections [3]. More recently, FLIM of [NAD(P)H] has been used for bacterial metabolic fingerprinting in diverse culture conditions [4]. In some cases, autofluorescence from lung tissue spectrally overlaps with signals from labelled bacteria, whereas lifetime images in general give excellent contrast [5]. By applying the phasor approach, this study has generated FLIM-phasor maps for Escherichia coli, Salmonella enterica serovar Typhimurium, Pseudomonas aeruginosa, Bacillus subtilis, and Staphylococcus epidermidis at the single cell and population levels. In contrast to the Chlamydia trachomatis, which is an obligation intracellular pathogen, facultative intracellular pathogens such as Salmonella, Shigella and pathogenic E. coli are able to survive and proliferate inside and outside the host cells. To differentiate intracellular and extracellular bacteria, z-stack imaging is usually required. This technique generates Any further distribution of this work must maintain attribution to the author(s) and the title of the work, journal citation and DOI. two-dimensional images at various depths of the cell, and it is possible to reconstruct to high-resolution 3D images. It is, however, a lengthy process, thus increasing the likelihood of cellular changes occurring.
FLIM provides contrast according to the fluorescence decay time and has been proven to be a powerful method in multi-labelled cell imaging [6][7][8]. It can be integrated with a confocal microscope or two-photon excitation microscope. In contrast to fluorescence intensity, the fluorescence decay time is independent of the local concentration of fluorophores, photobleaching, the local excitation intensity and local fluorescence detection efficiency. Moreover, the fluorescence decay times of aromatic molecules often depend on their intrinsic characteristics and local environments [9] such as Ca 2+ [10,11], pH [12], viscosity [13], temperature [14], refractive index [15], or interactions with other molecules, such as collisional, quenching or energy transfer processes [16,17]. Therefore, FLIM is not only able to distinguish spectrally overlapping fluorophores, but it can also be used to probe the immediate surroundings and dynamical processes of fluorophores. For example, previously intra-cellular imaging of gold nanorods using FLIM has shown improved contrast over fluorescence intensity imaging which results from large fluorescence lifetime differences; gold nanorods have typically short lifetimes (100 ps) compared to the fluorescence lifetimes of typical fluorophores (1.0∼4.0 ns) [18]. Furthermore, FLIM imaging can assess the energy transfer between gold nanorods to adjacent fluorophores, and FRET-FLIM has been successfully employed in resolving the cell take-up of gold nanorods and intracellular pathways [19,20].
Commercially available FLIM analysis tools usually provide initial quick analysis such as first moment analysis, and curve-fitting routines for further detailed analysis that requires end-users to choose fitting models (mono-, bi-or multi-exponential) and perform the analysis based on whether the reducedchi squared is within a specific user-selected criterion. Such exponential models, however, cannot be defined properly in complex biological systems and the fitting routine is not mathematically unique, which can lead to ambiguous interpretations. This is why more and more FLIM researchers are applying the phasor approach [11,21,22] to avoid complications in analysis and interpretations. Although some commercial tools do allow users to choose the areas of interest [23,24], they are not free. The IRF is determined by the laser, the detector and the temporal dispersion of the time-correlated single-photon counting (TCSPC) electronics used in FLIM experiments. To avoid complications, traditional software tools might use a synthesized IRF to perform the analysis if the IRF is not available or measured beforehand. In this paper, we aim to report a new analysis platform that is modelfitting free based on newly developed algorithms that unlocks the limitations of fitting routines, therefore we can directly calculate fluorescence lifetimes without resolving all parameters. We will combine the proposed analysis methods with the phasor analysis. Users are able to choose regions of interests from either lifetime images or phasor plots and perform cross comparison studies for easy and rapidly differentiating intracellular, extracellular and membrane-bounded bacteria of diverse species. The platform is envisaged to facilitate studies on bacteria-host interactions. The innovative aspects of this work include: 1. Fluorescence lifetime is used as an indicator to locate intracellular bacteria to investigate the lifetime of bacteria at different cellular locations.
2. Experimental results showed that the proposed amplitude weighted lifetime analysis method is rapid and can provide better contrast in our research for identifying cellular locations of bacteria compared to other analysis models.

3.
A new user-friendly platform for FLIM analysis has been developed (figure 1). The tool is able to A) analyse FLIM images with different lifetime models, B) allow users to pinpoint a cluster of pixels or identify lifetime populations through either phasor plots or lifetime images, and C) provide detailed lifetime distribution analysis.

Cell preparation
The mouse raw macrophage cells were routinely cultured in DMEM (Dulbecco's Modified Eagle Medium) supplemented with 10% FCS (fetal calf serum) under 5% CO 2 at 37°C. Cells were seeded on to glass coverslips in 24-well plates and cultured overnight for bacterial infection. Bacteria, which were engineered to express GFP (green fluorescent protein), were harvested from an early exponential phase and added to the cells with an MOI (multiplicity of infection)=100. After 40-60 min incubation, extracellular bacteria were removed by washing 3 times with PBS (phosphate-buffered saline). Fresh DMEM supplemented with 50 μg ml −1 of gentamicin was added to the cells for further incubation. At indicated time intervals cells were washed 3 times with PBS and fixed with 3.7% paraformaldehyde for 15 min. Cells were washed 3 times with PBS and permeabilized with 0.1% triton X-100 for 5 min. Cells were washed 3 times with PBS and stained for actin with phalloidin Alexa Flour 546 (ThermoFisher). The coverslips were then mounted for microscopy with a ProLong antifade solution (ThermoFisher).

Fluorescence intensity and lifetime imaging microscopy
FLIM was performed by using a confocal microscope (LSM510, Carl Zeiss) equipped with a time-correlated single-photon counting (TCSPC) module (SPC-830, Becker & Hickl GmbH). For z-stack imaging, an Argon laser of 488 nm was used as the single-photon excitation source and fluorescence emission was collected using a 500-550 nm bandpass filter for GFP labelled bacteria and a 565-615 nm bandpass filter for Alexa Fluor 548 labelled cell actin. A femtosecond Ti: Sapphire laser (Chameleon, Coherent) at 850 nm was used as a two-photon excitation source for FLIM imaging. The laser pulse has an 80 MHz repetition rate and a duration less than 200 fs. The emitted photons were collected through a 63× water-immersion objective lens (N.A.=1.0) and a 500-550 nm bandpass filter. FLIM data were acquired through the nondescaned mode.

FLIM analysis
FLIM images were analysed by the platform using three different lifetime analysis: (1) amplitude weighted lifetime model (τ A ), (2) intensity weighted lifetime model (τ I ) and (3) mono-exponential fitting using commercial software (τ C ). The decay function was calibrated by the IRF obtained from the measurement of dried urea ((NH 2 ) 2 CO) [25]. The measured IRF or the synthesized IRF calculated from the rising edge of the fluorescence signal was used in τ C analysis [26]. For the first two methods, we use a simple model to explain how the proposed analysis models work.
Assume the true decay function, f (t), to be estimated from the measured decay y(t) and the measured IRF, IRF(t), can be expressed as where a i is the amplitude and τ i is the lifetime of the ith and p is the number of lifetime species. Traditional FLIM analysis tools usually apply curve-fitting techniques to resolve a i and where y(t) isthe measured fluorescence decayfunction. Solving this inverse problem to obtain the amplitude and lifetime components, however, is time-consuming and can be prone to errors and artefacts, especially when the photon count is low. In many applications, the analysis goals are to obtain the intensity weighted lifetime, τ I , or the amplitude-weighted lifetime, τ A , defined by [27]: to provide contrast instead of resolving all unknown parameters. Without resorting to complex iterative curve-fitting routines, there are easier ways to estimate τ I and τ A . The former has been proven [28] to be approximate to the centre-of-mass method without [29] or with the IRF considered [30], whereas the latter can be easily obtained as well with the IRF considered (the details will be reported separately). τ I and τ A are simply two different mappings, and they should be carefully used to optimise the contrast according to users' applications. In this study, we will demonstrate how they can be used to differentiate bacteria in extraand intra-cellular or membrane-bounded locations. Note that when an amplitude is dominating (a i ∼1.0), then τ I /τ A ∼1.0, meaning the measured decay for this pixel is nearly mono-exponential decay and on the phasor plot it is close to the unit circle. The measured IRF is also calibrated in the phasor analysis provided by the proposed analysis platform.

Results and discussion
3.1. Comparisons between three lifetime analysis models: τ A , τ I and commercial software (τ C ) To disclose the locations of Shi86 with macrophage cells, z-stack fluorescence imaging was performed with each layer of 1 μm thickness. Internalization of bacteria in cells has been reported before and studied using z-stack confocal microscopy [31][32][33]. Figure 2 shows three slices of the z-stack images where Shi86 were labelled with their cellular locations identified, for example in figure 2(b), (A) extracellular, (B) intracellular, and (C) membrane-bounded where the bacteria are near the cell membrane. To identify these locations, 20 z-stack images were used. The information obtained from z-stack images can be used for cross comparisons with two-photon FLIM images, such as the ones taken in the same area as figure 2(b). Figure 3 shows fluorescence lifetimes of GFP labelled Shi86 at three different cellular locations (intracellular, membrane-bounded and extracellular) and Alexa Flour 546 from macrophage cells, respectively. 106 Shi86 were analysed, and the lifetime changes of GFP were found to be related to their cellular locations. τ A analysis shows better contrast than the other models. This agrees well with the conclusions summarised in [28] that τ A analysis is suitable for investigating different species showing subtle lifetime differences or for studying samples showing a small FRET efficiency. The mean lifetime of GFP and Alexa Fluor 546 was reported to be 2.00 ns [19,34] and 2.59 ns [35]. The lifetime of Alexa Fluor 546 is found as the long lifetime component in the FLIM image with τ A =1.94±0.33 ns, τ I = 2.50±0.08 ns and τ C =2.81±0.15 ns, respectively. τ A indicates significant lifetime differences that the intracellular Shi86 has a relatively short lifetime (τ A = 1.34±0.30 ns), and the extracellular Shi86 shows a long lifetime (τ A =1.76±0.32 ns), however, the Shi86 at the membrane-bounded location (τ A =1.46±0.26 ns) is insignificantly different to intracellular Shi86. τ I has the highest precision and the same trends as τ A and τ C , however, it shows the least contrast. The τ I analysis of intracellular, membrane-bounded and extracellular Shi86 shows τ I =2.07±0.11 ns, τ I =2.14±0.11 ns and τ I =2.28±0.14 ns, whereas τ C model obtains τ C = 2.09±0.15 ns, τ C =2.12±0.11 ns and τ C =2.18± 0.11 ns. τ C analysis can also provide results calibrated with the measured IRF, giving τ C(IRF) =1.72± 0.06 ns, τ C(IRF) =1.83±0.06 ns, τ C(IRF) =1.97 ± 0.37 ns, respectively.
The tool allows users to set intensity thresholds to remove pixels with insufficient photon counts, for example, the pixels mainly collecting dimmer autofluorescence.   The threshold was set to be above 100 photons for each pixel, as our τ A and τ I analysis models require a less photon count than what multi-exponential fitting methods do [28]. Figures 5(a)-(c) shows scattering plots of the photon count versus the lifetime for both models. It clearly indicates that τ A offers better differentiation to bacterial cellular locations, compared to τ I . Figures 4(b), (c), (e) and (f) are FLIM images generated by the developed platform where the black areas mean the pixels outside the interested intensity range. Figure 4(a) is a two-photon luminescence intensity image and figures 4(b)-(d) are initial τ A , τ I , and τ C images showing Shi86 and macrophage cells obtained from the same area as figure 2(b), respectively. Figure 4(b) is  the τ A image, and it reveals multi coded colours of Shi86 depending on their cellular locations, whereas τ I and τ C provide lower contrast. Shi86 are obvious in figure 4(b) but are not clear in the intensity image ( figure 4(a)). Figures 4(e)-(g) are the same image as figures 4(b)-(d), but with different colour scales, 0.7-2.7 ns for (e) and 1.5-3.0 ns for (f) and (g), showing better contrast. It is, however, still difficult to distinguish the cellular positions of bacteria due to the subtle differences in coded lifetimes for figures 4(c), (d), (f) and (g). Moreover, the phasor plot ( figure 6) was used to distinguish populations in the lifetime image presenting clearly two clusters: Shi86 and macrophage.
In addition, the locations of bacteria can be investigated by observing the ratio of τ I /τ A . Both models give similar lifetimes of the extracellular Shi86, which makes the ratio τ I /τ A closer to 1. However, the ratio increases during phagocytosis because τ I and τ A of intracellular Shi86 are significantly different. Therefore, the ratio τ I /τ A can improve the contrast, a good indicator to reveal the locations of bacteria. Figures 7(a) and (c) show τ I /τ A images in the same area as figure 2(b) with the ratio analysis that relates to the position in the phasor plot, while figures 7(b) and (d) are their interest encircled phasor plot, respectively. In figure 7(b), the selected area is close to the semi-circle encircled by the 8-sided polygon defined by the user. This area includes some parts related to the cell membrane and extracellular Shi86 with an average ratio=1. 19. In contrast, figure 7(d) shows the area that covers inside the semi-circle with an average ratio of 1.69. This area covers mostly the intracellular Shi86.

Applying τ A model to other types of bacteria
Since τ A has shown to have better contrast compared with the other two models in the case of S. sonnei, we applied this model to analyse Raw cells infected by other bacteria: AIEC strains HM605 and HH427, and a Lactobacillus strain. Figure 8 shows τ A images, which were generated using the same conditions as above.
The arrows indicate the bacteria. As in the example for Shi86, the τ A model adequately differentiated intracellular, extracellular and membrane-bounded bacteria in all cases (table 1). Table 1 includes τ C analysis results using the synthesized and measured IRFs, but they do not show obvious differences. Traditional analysis tools usually use least square fitting routines to perform model-fitting analysis. For mono-exponential analysis, the fitting routine usually generates results close to τ I analysis [36]. This is in good agreement with what we obtained from table 1. τ A shows the potential to differentiate cellular locations of bacteria. Extracellular HM605, HH427 and Lactobacillus have luminescence lifetimes of 1.84±0.03 ns, 1.73±0.21 ns and 1.33±0.17 ns, respectively, whereas the intracellular bacteria have obviously shorter lifetimes of  This indicates that the lifetime of GFP labelled bacteria is shorter during internalization processes. However, τ I and τ C are not as effectively as τ A to distinguish the cellular locations of bacteria, and we will conduct more imaging experiments to investigate this further.
In summary, we have built an effective platform, which can rapidly identify cellular locations of facultative intracellular bacteria. Both models (τ A and τ I ) have speedy performances and superior clarity than intensity imaging and are theoretically faster than traditional fitting methods, as our models do not require model selections or require setting extra constraints as most traditional analysis tools do [26]. Our tool only takes 0.3 s to generate τ A images or 0.1 s (comparable to the speed of the first moment analysis of commercial software tools at 10 fps [37]) for τ I images with 2.8 GHz Intel Core i7 processor. Moreover, our direct estimation algorithms are hardware-friendly, offering even much faster analysis if they are implemented in electronics hardware [29,38]. Although commercial software tools might also provide τ A and τ I analysis functions, they usually need to perform multi-exponential fitting routines to extract all necessary parameters first and then use equation (3) to obtain τ A or τ I [23,24]. Moreover, the proposed tool offers extra analysis functions (τ A, τ I and τ I /τ A ), whereas fitting methods that have been used in most free tools only provide close-to-τ I analysis [39,40]. Different tools provide their own strategies of selecting areas of interest, but they do not offer comparable speedy analysis. Although some commercial tools are also userfriendly to allow users choosing their areas of interest [23,24], they are unfortunately not free.
The τ A model has the best contrast and this is suitable for quick imaging samples with unknown lifetimes, investigating samples with subtle lifetime differences (in this study to investigate intracellular, extracellular and membrane-bounded bacteria) or imaging samples showing a small FRET efficiency. Although, the τ I has shown excellent precision and small deviations for Shi86 and is suitable for further imaging applications that require higher precision or a higher signal-to-noise ratio. Users are able to choose the proper indicator for their applications. From the experiments and the analysis conducted, this platform holds a high potential to identify the locations of bacteria from their lifetimes without performing z-stack imaging. The imaging platform, as well as the tool developed, can be widely applied by researchers conducting FLIM measurements. Researchers interested in this tool are welcome to contact the corresponding author Dr David Li (David.Li@strath.ac. uk). The analysis tool and future updates will be available to the public through Strathclyde Pure.