Dynamic FRET-FLIM based screening of signal transduction pathways

Optimizing screening conditions and FLIM analysis

We first set out to determine optimal conditions to acquire time-lapse FLIM images using the Leica SP8 FALCON system. This system is designed for high-count rate Time-Correlated Single Photon Counting (TCSPC) and records mTurquoise2 lifetimes reliably at count rates in excess of 40 MHz per detector. We spread the mTurquoise2 emission over 2 HyD (hybrid detectors) by adjusting the spectrometer settings (see M&M), effectively doubling the maximum count rate. Global fitting indicated dominant lifetime components of 3.4 and 0.6 ns, indicating the superposition of two different FRET states. Saturation of the sensor with cAMP, as induced by the treatment of cells with the direct AC activator forskolin, changed the relative magnitude of the two populations but not their lifetimes. All time-lapse images were therefore fitted with a n-Exponential Reconvolution model using two fixed lifetime components of 3.4 and 0.6 ns, and the intensities resulting from these fits were exported as tiff files.

High signal to noise (S/N) ratio of lifetime measurements requires large numbers of photons to be collected per frame from each cell. However, possible photodamage, bleaching, and the necessary throughput set upper limits to the excitation power and acquisition time. To reliably resolve small differences in cAMP concentration, we aimed to achieve a lifetime repeatability (i.e., deviations of consecutive baseline readings in the integrated signal of each cell) of less than 50 ps RMS, even for dim cells. With the conditions detailed in M&M, actual observed RMS of ~ 25 ps, n = 6500 cells, was achieved for most screens. As the lifetime span of the Epac-S^H189 sensor ranges from ~ 2.0 ns in the resting state up to 3.3 ns when maximally saturated with cAMP, S/N ratio is thus better than 40:1. It can be seen in Fig. 1 that this was sufficient to clearly resolve cell-to-cell variability in response to addition of norepinephrine (NE), which activates β-adrenergic receptors in HeLa cells. This S/N ratio also suffices to resolve cell-to-cell variability in baseline lifetimes, and thus in resting cAMP concentrations. The lifetimes of FRET sensors at resting state appeared near-normally distributed (2.34 + /− 0.05 ns, mean +/− SD, n = 154), Fig. 1E. Interestingly, a small percentage of cells with slightly increased cAMP levels were found (Fig. 1, see arrows). When imaged 2 days after culturing, these cells usually grouped together, suggesting clonal differences in baseline cAMP levels in WT HeLa cells. We also noticed that in most cells cAMP levels do not return to the initial resting values after transient stimulation with NE.

The required throughput is determined by both the temporal resolution necessary to capture cAMP dynamics (time-lapse interval) and by the number of cells to be recorded from. The latter depends on several factors, including cell-to-cell variability due to stochastic differences inherent in signal transduction cascades and on incomplete penetration of the genetic perturbations carried out in the screen. Most siRNA mediated knockdown experiments display considerable variability in gene silencing resulting in incomplete or even no detectable knockdown in a percentage of cells³⁰. Our pilot studies showed that recording from a few hundred cells in a single FOV captured most of the variation in each well. To minimize the risk that factors such as ongoing aging of the medium and increasing cell confluency might bias the results, we decided to run each entire screen, i.e. 22 knockdown conditions + controls in duplicate (60 wells in total), within 6–8 h. Under these conditions, we found near-identical lifetimes in the experiments recorded at the onset and at the end of the 6-h long screen (Fig. S2). Together, these considerations led us to conduct the screens using a 20 × dry objective, recording a single field of view with ~ 200–600 cells per well, and at 2 or 5 s time-lapse rate.

Automatic extraction of kinetic parameters

To optimize automated image analysis on a cell-by-cell base, we started by comparing algorithms for reliable segmentation of individual cells. We initially adapted standard image analysis methods by generating a dedicated Image J macro tailored to our cells. In essence, cell nuclei were detected by in vivo staining with SiR-DNA, followed by Voronoi segmentation to determine cell boundaries, which was based on the time-averaged intensity of the time-lapse images. This macro²⁹ yielded good results, i.e. a ~ 95% reliable segmentation of cells was achieved as judged by eye. However, while our experiments were in progress, a general algorithm for segmentation of cells based on deep-learning algorithms was reported²⁶, the performance of which we tested against our own developments. In several independent experiments we found Cellpose²⁶ to be superior in reliability compared to more conventional image segmentation algorithms, including our own developments (Fig. S1). It must be mentioned that Cellpose is unpractically slow for near-real time analysis, but as it delivered very good segmentation without needing nuclear staining, we adopted it for all off-line segmentation of data in this study (for details, see M&M).

For each individual cell (ROI), we extracted mean fluorescence intensity and donor lifetime (Fig. 1) values, along with data on ROI size and potential error conditions such as disturbances by dislodged cells and out-of-boundary conditions (detailed description in accompanying information on our Github page). These data also were used to calculate RMS noise values of intensity and lifetime signals. Moreover, after fitting the agonist induced responses of cells to a suitable model (Fig. 2 and M&M), dynamic parameters such as activation rates, peak values, decay properties and steady state value were extracted.

Next, we tested the reproducibility of our results with different batches of cells on different days. Our analysis showed excellent consistency of S/N ratio and calibration value following treatment with 25 µM forskolin. Baseline values were slightly more variable (Table S2), most likely reflecting small batch-to-batch variations in basal cAMP levels. These observations stress the importance of carrying out signaling screens within a limited time span, i.e., preferably on a single day.

Caged-cAMP assay shows importance of PDE3A in regulation of cAMP breakdown

We next set out to conduct a FLIM screen to investigate the roles of the roles of individual PDEs in breaking down cAMP. We initially studied the kinetics of cAMP changes in HeLa cells upon photorelease of caged cAMP. For that, HeLa cells stably expressing the Epac-S^H189 were seeded in 96-well plates, and using pools of 4 specific siRNAs against each isoform, individual PDEs were knocked down in duplicate wells. Cells were loaded with DMNB-caged-cAMP at 1 mM final concentration for 30 min. Uncaging with a 200 ms UV pulse caused an immediate increase in intracellular cAMP levels, and thus in donor lifetime, which subsequently returned towards its baseline level (Fig. 3). Hundreds of cells within a single FOV were imaged every 2 s for at least 140 s (or longer, if slow recovery called for that) and acquired data was stored for analysis offline.

Following segmentation, time-lapse FLIM traces for each ROI were individually fitted to a logistic (sigmoid) curve. For large numbers of cells, the data and fits were visually inspected to ensure proper fitting using a Python script (results_browser²⁹). cAMP breakdown time was then calculated from the resulting fit parameters and plotted for each PDE knockdown condition (Fig. 4). From these data, it is apparent that knockdown of PDE3A markedly affects the breakdown time in these cells (85.5 ± 2.5 s, versus 37.9 ± 0.5 s in WT cells, (mean ± SEM; p < 0.001, student t-test). Additionally, a smaller but significant effect of PDE10A knockdown on cAMP breakdown was seen (51.4 ± 0.8 s, p < 0.001).

From the data in Fig. 4, it is also apparent that the calculated decay times for individual cells show considerable variability. As the S/N ratio within a single cell is excellent (compare e.g. the variability of the initial 123 samples in the baseline in Fig. 4), the cell-to-cell variability in FLIM values likely has biological origin. Moreover, the extremely large span of the observations seen for knockdown of PDE3A and PDE10A suggest that lack of or incomplete PDE knockdown in individual cells is a further major determinant of variability in these wells. Furthermore, cell shape differences, e.g. in surface-to-volume ratios, are likely to affect cAMP clearance. This view is supported by the observation that very similar results were obtained when we repeated selected conditions, again in duplicate, a month later.

Next, we carried out a follow-up experiment to evaluate the effect of simultaneous knockdown of both PDE3A and PDE10A. Remarkably, knockdown of these two PDE genes in the same cell did not significantly slow down cAMP breakdown below the rate seen for PDE3A alone. This is perhaps unexpected because in PDE3A knockdown cells there is still a considerable rate of cAMP clearance. Therefore, next we assayed cAMP breakdown in cells pretreated with two well-characterized PDE inhibitors, the nonspecific PDE inhibitor IBMX³¹ (100 µM) and the PDE3 family specific inhibitor cilostamide³² (1 µM) administered either alone or together. Both inhibitors slowed down cAMP breakdown to rates slightly below that of combined PDE3A/PDE10A siRNA treatment, and combined they caused a further increase in cAMP clearance times (Fig. 5). It is also noteworthy that unlike PDE knockdown, inhibitor pretreatments selectively wiped out the population of cells with fastest breakdown times, consistent with the notion that high variability in breakdown speeds in the population of PDE3A and PDE10A knockdown cells reflects incomplete knockdown by siRNAs. Intriguingly, despite inhibition of all PDEs, cAMP still eventually is cleared in these HeLa cells. The mechanisms involved remain to be elucidated in further studies.

While analyzing these data, we noted that baseline donor lifetimes in cells pretreated with DMNB-cAMP showed considerable biological variability, ranging between 2.4 ns and 2.7 ns (Fig. S3A). In contrast, untreated cells had average lifetimes of 2.28 ns at rest and showed considerably less variability (Fig S3B). The difference increased when cells were incubated with increasing concentrations of DMNB-cAMP, indicating some leakiness (spontaneous decomposition of the caging group in the cells) of this compound. In line with this, baseline donor lifetimes in PDE3A knockdown cells and in PDE3A / PDE10A knockdown cells were significantly elevated in DMNB-cAMP treated cells (Fig. S4A), but not in untreated controls. Together, these data indicate that PDE3A also has significant activity when cAMP levels are only slightly increased in these cells. We also noted that in the vast majority of stimulated cells, cAMP levels eventually returned to their pre-stimulation value (Fig. S4C,D). A similar observation holds true for PDE knockdown cells.

We conclude that our screening platform is well suited to resolve even minor differences in cAMP clearance kinetics, and that variability between experiments carried out several weeks apart is only minor. However, pretreatment with DMNB-cAMP appears to cause a significant disturbance of baseline cAMP levels, and this effect was amplified when PDE3A was knocked down. We therefore redesigned the experimental paradigm to circumvent the confounding effect of caged cAMP.

Dominant role of PDE3A is confirmed using transient activation of GPCR signaling

Dynamic screens can also be carried out when AC is activated following stimulation of GPCRs with their cognate ligands. However, in such experiments it is much harder to dissect the contribution of PDEs in controlling the rate of cAMP clearance, because cAMP levels are also affected by the continued activity of proteins (GPCRs, G proteins and AC) upstream in the signaling cascade. Termination of Gα_s activity is believed to happen in seconds^33,34 and AC activity is strictly dependent on GTP-loaded Gα_s. However, receptor inactivation is much slower and, in most cases, not complete: a small proportion of receptors is thought to recycle to the plasma membrane where they can become reactivated by the agonist and continue to signal³⁵.

Therefore, we adopted a protocol in which cells were stimulated with a receptor agonist, followed within 10–15 s by addition of excess of a potent competitive antagonist. We chose β-adrenergic receptors as they form a well-characterized G-protein coupled receptor system^36,37, and are endogenously expressed in HeLa cells. We first stimulated HeLa cells with 40 nM of the receptor agonist isoproterenol which caused a rapid rise in cAMP levels and subsequently added the antagonist propranolol at 60 nM concentration which caused a sharp decline following the stimulation. Finally, 25 µM forskolin was added to directly stimulate AC and obtain a maximal sensor response as a control. Figure 6 shows a representative single-cell lifetime trace along with a fitted logistic curve capturing the decay kinetics.

Importantly, we noted that cAMP decay rates as determined following this experimental protocol in WT cells were approximately equal to those measured after photorelease of caged cAMP. This implicates that following addition of propranolol, all upstream steps in the signaling cascade became inactivated within seconds. It was also confirmed that when propranolol was added as first stimulus, no detectable response followed upon subsequent addition of isoproterenol. We therefore conclude that this experimental paradigm is well suited to study the role of PDEs in cAMP breakdown.

Figure 7 shows effects of individual knockdown of the same set of 22 PDEs assayed according to this protocol. Again, we find that PDE3A is the most prominent enzyme controlling cAMP breakdown in HeLa cells, followed by PDE10A. The effects of knockdown of other PDEs were not significant. Furthermore, the double knockdown of PDE isoforms 3A and 10A together is also in good agreement with the data from the first screen after photorelease of caged cAMP. Remarkably, however, the effects of PDE inhibitors IBMX and cilostamide appeared more pronounced as compared to the first screen. Eventually, in all cases cAMP levels returned towards baseline levels, indicating the activity of additional cAMP clearance mechanisms in these cells.

Finally, we tested for possible correlations between fluorescent properties of the cells, and outcome of the analysis. We found no evidence that cellular brightness affected donor baseline lifetimes for individual cells (Fig. S5). However, higher expression levels of the biosensor slightly prolonged cAMP breakdown, presumably due to the buffering effect of the Epac-sensor (Fig. S5). We conclude that variability in cellular cAMP breakdown speeds appears to dominated by true biological variability and that the agonist/antagonist stimulation paradigm is well suited to study the dynamics of cAMP turnover in genetic screens.