1. Introduction
Biological tissue is an ensemble of cells implementing a specific common function. These are frequently cells of different types, each type running a different “subroutine” of the main functional program. An example of this arrangement is pancreatic islet of Langerhans, in which cells secreting a glucose-lowering hormone insulin (β-cells) are positioned next to their counter-parts secreting glucagon (α-cells) that, in turn, elevates blood glucose [
1]. The unwritten biological convention is that all the cells of the same type are
roughly alike. For instance, it is believed that the mutations in an energy-sensing molecule (the so-called ATP-sensitive K
+ (K
ATP) [
2,
3,
4] channel) render
roughly all pancreatic islet β-cells incapable of sensing blood glucose. At the same time, adrenaline is believed to induce secretion of glucagon from
roughly all α-cells [
5,
6], to rescue the body from hypoglycemia.
Whilst valid in a broad sense, this assumption, however, has its limits, as different cells of the same type are not absolutely identical. This phenotypical heterogeneity may stem from an exposure to different factors of microenvironment [
7], such as neighboring cells [
8], local signals [
9] or biological polarity [
10]. Secondly, biological responses have stochastic molecular nature, which makes them subject to variability. For the examples above, energy sensing by the K
ATP channel is based on its ability to bind the molecule of ATP [
11,
12], whilst the α-cell sensitivity to adrenaline stems from how densely the β-adrenergic receptors are expressed on its membrane.
The heterogeneity of cell phenotype can be probed by flow cytometry, a high-throughput technology that reports the expression of several proteins on the membrane of the same cell at once, by staining the cell suspension with monoclonal antibodies conjugated with fluorescent markers. Deep phenotypical profiling of living cells, arranged into an intact tissue, is done via high-content microscopy [
13,
14], which also characterizes specific proteins expressed on the plasma membrane. An obvious shortcoming of these otherwise powerful profiling techniques is their reliance on cell structure, which is not equivalent to the true function of the living cells. A more function-oriented approach, real-time imaging of the intracellular levels of secondary messenger compounds, such as cAMP or Ca
2+, with fluorescent sensors, can profile the subpopulations of cells via specific pharmacological stimuli [
5,
15,
16], the so-called “marker compounds” [
17]. The real-time imaging, however, cannot match the statistical power of the approaches above, due to a lower throughput and an increased experimental duration. Imaging of a response to a single specific stimulus may take tens of minutes, and the responses to stimuli added sequentially are difficult to interpret due time-dependent drift of fluorescence.
In this paper, we aimed to develop an experimental and analytical framework for deep profiling of large populations of intact living cells based on differential response to multiple pharmacological compound(s). We specifically focus on pancreatic islets of Langerhans as a model and image the cytosolic level of the secondary messenger cAMP, as this compound (i) changes on a slow (minutes) timescale, which is advantageous for imaging of large cell populations, and (ii) cannot be transmitted between two neighboring cells, unlike cytosolic Ca2+ or electrical potential of plasma membrane.
Importantly, cAMP signaling in pancreatic islet cells is a target of physiological regulation by incretins, natural highly selective insulinotropic peptide agents secreted by the gut cells. Glucagon-like peptide-1 (GLP-1), secreted by the gut L-cells; enhances the secretion of insulin and attenuates secretion of glucagon, in a glucose-dependent manner [
18]. This mode of action makes GLP-1 an excellent antidiabetic medication as per se it is not able to induce hypoglycemia, a major problem with many antidiabetic drugs. The active form of GLP-1 (known as “7–36” form) has a short lifetime in circulation as it is rapidly inactivated by a ubiquitous enzyme, dipeptidyl peptidase-4 (DPP-4), to the inactive “9–36” form. Gastric inhibitory peptide (GIP), secreted by the gut K-cells effectively works in a way similar to that of GLP-1’s, inclusive of being inactivated by DPP-4 [
19]. The two key differences between the GLP-1 and GIP action are (i) the lack of inhibition of glucagon secretion and (ii) less reliable and therefore yet unexplored therapeutic perspectives of the latter, as exogenous GIP has only a small effect in human patients [
19].
Critical steps in imaging of intracellular concentration of cAMP ([cAMP]
i) were achieved upon development of a fluorescent genetically encoded sensors able to report submicromollar changes in [cAMP] [
20] as well as the activity of proteins kinase A (PKA) [
21], a parameter that is directly dependent on [cAMP]
i. We utilized the PKA and cAMP sensors for deep multifactor (ca.10 agonists) profiling of 100 s–10,000 s of cells obtained throughout long (multihour) recording periods on an automatic imaging system.
3. Results
The conditions of the adenoviral infection were optimized to allow the expression of the recombinant sensor in the predominant majority of the islet cells, whilst avoiding cell death or damage (
Figure 1b). To verify the fact that we can reliably image the changes in [cAMP]
i and PKA activity, a conventional positive control was applied using the combination of two known agonists, forskolin (10 μM) and 3-isobutyl-1-methylxantine (IBMX, 100 μM). The two chemicals reversibly act on intracellular enzymes, adenylyl cyclase and phosphodiesterase, respectively, to elevate the cytosolic concentration of cAMP and hence activate PKA. In our hands, the application of forskolin and IBMX reversibly increased the fluorescence of the YFP channel and decreased the fluorescence of the CFP channel of the ACAR3 sensor (
Figure 1c,d). The onset of the increase in the YFP/CFP ratio coincided with an increase in [cAMP]
i imaged simultaneously in a different islet positioned within the same chamber, using Green Downward cADDis sensor (
Figure 1d). The [cAMP]
i signal was, however, faster to relax to its basal value than the PKA signal, upon removal of forskolin and IBMX (
Figure 1d).
The specific agonists, GLP-1 and GIP were applied at three different concentrations, 1 pM, 100 pM and 10 nM, followed by returns to the basal condition (imaging solution supplemented with 10 mM glucose). The agonists induced reversible changes in the YFP/CFP fluorescence ratio (R) (
Figure 2a). To identify α-cells within the islets, we then applied 10 μM adrenaline, which is known to elevate [cAMP]
i selectively in these cells but, in contrast, inhibit the cAMP signaling in other types of islet cells [
32]. The adrenaline response in α-cells has been proved to have a strong correlation with expression of fluorescent markers under a tissue-specific glucagon promoter [
5].
3.1. Correcting the Time-Dependent Drift of Fluorescence
The raw traces demonstrated substantial cell-to-cell variation of the YFP/CFP fluorescence ratio,
R, as well as the presence of a slow time-dependent trend in
R kinetics (
Figure 2a,b): each time the tissue was subject to the basal, agonist-free, conditions, the apparent ratio of fluorescence was decreasing with time. We accounted for the cell-to-cell variation by normalizing
R of each individual cell to its value at the beginning of the experiment (
R/R0,
Figure 2c,d). This correction also increased the signal-to-noise ratio (SNR,
Table 1).
To correct the baseline trend, we performed a subtraction of different baseline functions (
Figure S1) from the
R/R0 timecourse. The quality criteria for the baseline correction were (i) the enhancement of the SNR and (ii) the ability to resolve the effects of every compound added. Linear and exponential correction, relying on the first and the last
regions (see Methods in
Section 2), provided a substantial improvement in the SNR (
Table 1) but failed to reveal small changes in the signal, induced by adrenaline (
Figure S1a,b). The high-degree polynomial and especially spline correction, utilizing all the
regions (“
R0” in
Figure S1c,d), had a better capability for resolving the small adrenaline effect (
Figure S1c,d). The latter corrections did however introduce several artefacts into the data. The piecewise approach implementing linear (
Figure S1e) or square (
Figure S1f) fit between each pair of neighboring
regions resulted in a significant increase of the SNR and, in the case of the linear correction, excellent resolution of the small adrenaline effect (
Figure S1e). We therefore used the piecewise linear correction as a routine throughout the study (
Figure 2e,f).
3.2. Scaling Up the Unsupervised Quantification of the Effects
Quantification of the (ant)agonist effects (
Figure 2) assumes comparing the average
R/R0 values before (red in
Figure S2a) and after (blue in
Figure S2a) the application of the (ant)agonists. To limit the manual input into the image and data processing to supervision of the ROI detection and entering the timestamps vector, we sought the ways of reusing the
regions defined during the baseline subtraction step. To that end, we algorithmically expanded the
regions, so they included both the pre- and posteffect signal (
Figure 2e), and quantified the (ant)agonist effects either by fitting the signal within each region with different functions (
Figure S2b,c,e,f) or as a crude difference between the final and initial fluorescence within each region (
Figure S2d). In our hands, the sigmoid and Hill fit provided the best approximation of the (ant)agonist effects (
Table 2). Applied to the “real word” data though, the two transcendental fits did not converge in ~2–5% of cells, which prompted us to use a less precise but more stable linear fitting algorithm (
Figure S2b) for the quantification of the (ant)agonist effects.
3.3. Exploratory Analysis of Cell Populations Based on the Response to Various Stimuli
The incretins GLP-1 and GIP, secreted by the gut L- and K-cells, respectively, are natural peptide factors that target pancreatic islet cells and induce increases in [cAMP]
i [
18]. In contrast, the catecholamine adrenaline is a body’s soluble signal that has a clear differential effect on two major islet cells subpopulations: it inhibits secretion of insulin by β-cells and induces secretion of glucagon from α-cells to rescue extreme hypoglycemia [
32]. The signals induced by incretins and adrenaline are mediated via changes in [cAMP]
i, acting by increasing or decreasing the concentration of the secondary messenger, respectively. We therefore used adrenaline as a marker compound [
17], which discriminates between β- and α-cells within the islet, thereby classifying each cell within the islet solely by its function, without any immunostaining, which would have required killing the cells and permeabilizing the cell membranes (
Figure 3). To that end, having used the PKA activity as a surrogate for [cAMP]
i (
Figure 3a), we have ranked the imaged cells by the change in the PKA activity induced by 10 μm adrenaline (
Figure 3b, arrow). Having logically sorted the heterogeneous population into α- (dashed) and β-cell populations (
Figure 3b), we observed that mouse α-cells were seemingly better responsive to GLP-1 (and GIP) than β-cells (
Figure 3b), in contrast with earlier reports on the limited impact of GLP-1 on [cAMP]
i in α-cells [
33]. Further ranking of cell responses according to the individual cell sensitivity to GIP or GLP-1 (
Figure 3c,d) suggested that the most GLP-1-sensitive cells are at the same time the most GIP-sensitive, within low, physiologically relevant concentrations of the two incretins (1 and 100 pM) (
Figure 3c).
3.4. Multiparameter Profiling of Cell Subpopulations within Islets
We have therefore computed correlations between the effects of the two incretins, GIP and GLP-1, on a per-cell basis, within α- and β-cell populations. Palpable at low and intermediate concentrations (1 pM, 100 pM), positive per-cell correlation between GIP and GLP-1 effects in the β-cell population was significantly decreased when the incretins were used at 10 nM (
Figure 4a). A similar pattern was observed for GIP and GLP-1 9–36 (
Figure 4b). Overall, despite the low potency of the inactive form of GLP-1 (9–36) (
Figure 3d), the effects of the two forms of GLP-1 appeared to associate in statistical sense, in the β-cells (
Figure 4c), inclusive of the concentration-dependent decline in correlation with GIP (cf. response to 100 pM and 10 nM among the red markers in
Figure 4a,b). For α-cells, however, the correlation between the effects of GLP-1 and GIP, at physiologically relevant concentrations of 1 and 100 pM, was negative, which, just like in the case of β-cells, was attenuated at the higher concentration of the agonists (blue in
Figure 4b).
3.5. Clusters of Islet Cells Responding to the Incretin Signals
We further probed the association of the incretin effects in two main islet cell populations, α- and β-cells, by performing the cluster analysis of the functional response to three concentrations (1 pM, 100 pM, 10 nM) of the agonists from six pancreatic islet preparations (n = 10,294 cells). The k-means cluster analysis (
Figure 5a,b) allowed mapping of five (for β-cells) and four (α-cells) functionally distinct subpopulations (
Figure 5a,b). The major contributors into the principal component 1 (PC 1) were the effects of the three concentrations of GLP-1 whereas PC 2 is mostly influenced by GLP-1 (9–36) (
Table S1,
Table S2) for both α- and β-cells.
The hierarchical clustering (
Figure 5c,d: dendrogram along the X-axis) revealed a strong association between the responses to the three concentrations of GLP-1 as well as between the responses to all concentrations GLP1 (9–36) in both β- (
Figure 5c) and α-cells (
Figure 5d). At the same time, the response to GIP displayed nonmonotonous concentration-dependent behavior (
Figure 5c,d): 1 and 100 pM clustered with the inactive form of GLP-1 (9–36) whereas the 10 nM data clustered with the effect of IBMX and forskolin, which directly stimulate adenylyl cyclase and inhibit phosphodiesterase.