Microfluidic platform enables live-cell imaging of signaling and transcription combined with multiplexed secretion measurements in the same single cells

Abstract

Innate immune cells, including macrophages and dendritic cells, protect the host from pathogenic assaults in part through secretion of a program of cytokines and chemokines (C/Cs). Cell-to-cell variability in C/C secretion appears to contribute to the regulation of the immune response, but the sources of secretion variability are largely unknown. To begin to track the biological sources that control secretion variability, we developed and validated a microfluidic device to integrate live-cell imaging of fluorescent reporter proteins with a single-cell assay of protein secretion. We used this device to image NF-κB RelA nuclear translocation dynamics and Tnf transcription dynamics in macrophages in response to stimulation with the bacterial component lipopolysaccharide (LPS), followed by quantification of secretion of TNF, CCL2, CCL3, and CCL5. We found that the timing of the initial peak of RelA signaling in part determined the relative level of TNF and CCL3 secretion, but not CCL2 and CCL5 secretion. Our results support evidence that differences in timing across cell processes partly account for cell-to-cell variability in downstream responses, but that other factors introduce variability at each biological step.

Introduction

Proteins secreted from innate immune cells play key roles in the regulation of the innate and adaptive immune response to infection and injury (1). Following sensing of pathogens via toll-like receptors (TLRs) and other pattern recognition receptors, the cytokines and chemokines (C/Cs) released by innate immune cells must be precisely tuned to protect the host without causing damage to host tissue. It has been observed that innate immune cells, including macrophages and dendritic cells, respond heterogeneously to TLR stimulation, and that this cell-to-cell variability, in combination with paracrine signaling, may be a mechanism by which the population response is regulated (2–4). For this reason, understanding the sources of cell-to-cell variability in secretion and how variability is regulated following TLR stimulation is functionally important.

Following stimulation of the TLR4 receptor with the Gram-negative bacterial outer membrane component lipopolysaccharide (LPS), signaling pathways initiate the translocation of the transcription factor nuclear factor-kappa B (NF-κB) RelA:p50 heterodimer from the cytoplasm into the nucleus to induce pro-inflammatory gene expression (5). Selective induction of NF-κB-target genes ensures proper regulation of inflammation following TLR4 stimulation (6). Interestingly, there is significant cell-to-cell variability in NF-κB RelA signaling following LPS stimulation, which suggests that this could be an important source of variability in downstream responses.

Measuring signaling and downstream responses in the same single cells provides a means to quantify the extent to which variability in specific intracellular signals determines the response (7). Previous studies have measured LPS or TNF-stimulated NF-κB RelA signaling and gene expression in the same cell via imaging a fluorescent gene expression reporter in live cells (8), in situ transcript labeling in fixed cells (9,10), or single-cell RNA sequencing (11). In other cases, NF-κB RelA signaling was connected to downstream responses including viral activation (12). More recently, several methods have been developed to measure secreted proteins from single cells (13–16), and in one case imaging of NF-κB RelA signaling dynamics was connected to cytokine secretion using a microfluidic device (17). However, due to technical challenges, few studies have measured signaling, transcription, and secretion in the same cells.

Here we developed a Single-cell Secretomic microfluidic device (SiSec chip) to enable live-cell imaging and multiplexed end-point measurements of secretion from single macrophages in order to integrate measurements of signaling, transcription, and secretion in the same single cell. We demonstrate that we can track RelA translocation and Tnf transcription dynamics via fluorescent reporters following LPS stimulation, followed by secretion measurements of four C/Cs from the same cells. We find that cell-to-cell variability in the timing of RelA signaling partly accounts for variability in transcription and secretion of TNF following TLR stimulation.

Materials and methods

Design and fabrication of microfluidic master molds

The SiSec chip is a push-down valve based microfluidic device containing six flow inlets, nine valve inlets and one outlet forming 150 microwell compartments containing traps for cell capture. In one version, all wells contain one trap to focus on single-cell measurements, while in another version, 50 wells each contain one, two or three traps in order to enable experiments on cell-cell communication. The chip integrates a base fluidic layer and a top valve layer cast from two different Si master molds. These layers are aligned and bonded, and then sandwiched to a capture antibody barcoded glass coverslip to form the fully integrated SiSec chip (Fig. 1).

Fig. 1. Overview of integrated microfluidic SiSec chip for signaling and secretion measurements.

(A) Schematic of the SiSec chip. (B) Schematic of device assembly. A valve-layer and flow-layer of PDMS are bonded together. A glass slide is separately flow patterned with an antibody barcode and then bonded to the PDMS device. (C) Diagram of a single well illustrating the area for cell capture (green) and the area for the antibody barcode (blue). Note the relative positioning of single traps (top) and two and three traps (bottom). (D) Image of the antibody barcode for capture of up to five secreted proteins.

To create the fluidic (base) layer, a two-profile mold with rectangular and rounded surface profiles was fabricated using S1813 and AZ 9260 photoresists, respectively. A 4-inch silicon wafer was first coated with Hexamethyldisilazane (HMDS, Sigma Aldrich) at 3000 rpm and baked at 200°C for 2 min. Subsequently, a positive photoresist, Microposit S1813 (Shipley) was spin coated on this wafer at 900 rpm for 5 seconds and ramped up to 3000 rpm for 1 min. A short pre-exposure bake at 100°C for 3 min was performed following UV exposure at 150mJ/cm2 using EVG 20 contact/proximity mask aligner. This wafer was developed with MF-319 for 1 min following a post-exposure bake at 120°C for 3 min. The master mold was then etched using a Bosch etch process and the remaining photoresist was then removed with Acetone and IPA resulting in a 10–12 μm feature height measured using a KLA-Tencor ASIQ profiler. In order to fill gaps in the base channel and to make them rounded, the wafer was again spin coated with AZ 9260 at 3000 rpm for 1 min followed by soft bake at 100°C for 3 min. The wafer was then aligned and exposed with a second mask to fill the gaps in the etched layer resulting in an approximately 8 μm tall rectangular fills after development with 1:1 dilution of AZ 400K: DI H₂O. In order to form rounded profiles, this wafer was gradually heated to 180°C for 15 min and then allowed to cool to room temperature, thus preventing cracking of the photoresist. The final rounded feature measured 10 μm in height.

To create the valve (top) layer, a negative photoresist SU-8 3025 was used to make 50 μm tall rectangular channels following standard photolithography process. This height prevents collapse of the valve–channel interface upon bonding the flow and valve layers and also when the channel is not pressurized. The barcode patterning mold was fabricated using S1813 positive photoresist as discussed above through a Bosch Etch process. Etched rectangular channel heights were around 10 μm tall as measured using the contact profiler.

All master molds for the flow, valve and barcode patterning layers were silanized in a vacuum chamber with 20 μl of Trichloro(1H,1H,2H,2H-perfluorooctyl)silane (Sigma Aldrich), to prevent sticking to the PDMS. Vacuum was applied for a few minutes in a degasser, leaving the chamber sealed overnight. Pressure was released gradually from the chamber to allow vapor deposition of silane. This process was repeated following every sixth casting to maintain good release of PDMS casts.

Microfluidic device fabrication and assembly

To fabricate the flow layer, 11 g of RTV 615 base:crosslinker was mixed in a 10:0.5 ratio, transferred to a tube, centrifuged for 2 min at 2000 rpm to remove air bubbles, and spin coated over the wafer for 1 min at 2000 rpm. The wafer was then baked for 2 hours at 80°C to produce a 20 μm thick PDMS membrane that was ready to be sandwiched with the top (valve) layer. The valve layer was fabricated similar to the flow layer using a standard soft lithographic approach. 22 g of RTV 615 base:crosslinker was mixed in a 10:1 ratio, cured and cut to produce a 5 mm thick slab of PDMS with features. Holes were punched in all the inlets and dust was removed with tape.

Both the layers were exposed to O₂ plasma for 90 sec and carefully aligned inside a clean hood making sure that it remained dust free. This wafer containing the spin-coated flow layer and sandwiched valve layer was heated at 110°C for 3 hours. The bonded two-layer device was then cut out of the wafer and peeled in a unidirectional manner. Dust was removed with tape, and 6 inlet and 1 outlet holes were punched through the flow layer while avoiding re-punching holes through the valve layer. The device was re-taped only on one side (on top of the valve layer with no features on it and plasma treated for 90 sec. This was immediately covered with 1% (3-Aminopropyl)triethoxysilane (APTES, Sigma Aldrich) for 15 min, and rinsed with DI H₂O prior to bonding onto the barcoded glass slide.

To create the antibody barcode for secreted protein capture, a 5 mm thick flow patterning PDMS chip was cast from the barcode patterning mold. Dust was removed from the device with tape, and then nine inlet and outlet holes were punched. The barcode chip is not plasma treated but taped five times prior to being bound to a glass coverslip. The glass coverslip was rinsed with isopropanol, blow-dried and plasma treated for 90 sec, and then further coated with 1% (3-glycidoxypropyl)trimethoxysilane (GPTMS, Sigma Aldrich) by adding 1.5 mL of GPTMS on its surface and allowing it to incubate for 20 min. The coverslip was then rinsed three times with DI water and reversibly bonded to the barcode flow patterning chip by simply placing it on the surface of the coverslip and incubating it at 80°C for 2 hours. Flow patterning of the glass coverslip was carried out using a two-step process. First, a 3 μl bolus of the capture antibody solution was flowed through to create the capture antibody barcode and, simultaneously, a 9 μl bolus of the fibronectin solution was flowed through the cell trap area. Excess antibodies/fibronectin were washed by flowing a 15 μl PBS solution following the coating step. Flow patterning was performed for 24 hours or until the tubes run dry at 2 psi reagent injection pressure. Antibody pairs used for secreted protein capture and detection are listed in Table 1.

Table 1:

Capture and detection antibody pairs with recombinant standards

Target	Vendor	Catalog No.
CCL2/MCP-1	R&D	DY479
CCL3/MIP-1α	R&D	DY450
CCL5/RANTES	R&D	DY478
TNF-α	eBioscience	88–7324-88

As a quality control and in order to visualize that the cytokine barcodes were properly aligned within the valve area, the ends of the barcode were labelled with a standard amine functionalized CF555 dye amine (Biotium Inc.). The flow patterning chip was then peeled off and the coverslip was washed with DI H₂O. The freshly APTES coated device was immediately aligned and sandwiched to the flow patterned coverslip and incubated at 37°C for 12 hours resulting in a permanent bond. The device was then stored in a refrigerator at 4°C until further use.

To create the antibody barcode for secreted protein capture, a 5 mm thick flow patterning PDMS chip was casted from the barcode patterning mold. Dust was removed from the chip with tape, and 9 inlet and outlet holes were punched. The barcode chip was taped 5–7 times prior to being bound to an epoxysilane coated glass slide. 3 μL of the capture antibodies were injected into the microchannels separately and flowed through the microfluidic channels overnight until dry (2 channels per antibody and 1 channel of BSA555 for visualization). Afterwards, the antibody barcode slides were blocked in plain BMDM media for 2 hours, then dipped in PBS and diH2O and gently blown dry with forced N2 before being bonded to the PDMS slabs creating 4mm wide wells (5mm tall) with the barcode at on the bottom (facing upward). The bonded PDMS and barcode slide were baked for 2 hours at 80°C and then recombinant standard solutions were added to the wells (30 μL per well). Recombinant standard solutions for an 8-fold standard curve starting at 5000 pg/mL were created by diluting recombinant standards in plain media (Fig. S1A). For the cross reactivity each standard solution contained only one cytokine at a high concentration (5000 pg/mL) or a low concentration (50 pg/mL) (Fig. S1B). After adding the standard solutions, the device was incubated for 12 hours.

After incubation, the antibody barcode slide was developed by incubating with the appropriate concentration of detection antibody for 2 hours, washing excess detection antibody off with PBS+3% BSA followed by a 30 min incubation with Strep-APC. Excess Strep APC was washed off with 1x PBS and diH2O sequentially and the barcode chip dried with forced N2 gas. The barcode chip was then scanned with a four-laser microarray scanner (Molecular Devices; Genepix 4200A) for protein signal detection.

Analysis of antibody range and specificity

To generate recombinant protein standard curves, we fabricated PDMS slabs using RTV 615 base: crosslinker mix in a 10:1 ratio and baked for 2 hours at 80°C. The PDMS was cut to measure 25mm x 75mm x 5mm and 4mm sample holes were punched into the device. Dust and debris were removed with tape, and the slabs were autoclaved and exposed to O₂ plasma as described for the SiSec chip. We generated calibration curves for each antibody for using recombinant protein standards at 8 concentrations ranging from 40 to 5000 pg/ml (Fig. S1A). We found that all four antibody pairs showed a log-linear dynamic range spanning approximately two orders of magnitude with a detection limit of approximately 50 pg/ml for TNF, CCL3, and CCL2, and slightly higher for CCL5, with no significant cross reactivity (Fig. S1B).

Flow simulation

Flow profiles of the S-traps were estimated using COMSOL Multiphysics v.4.2 (COMSOL Inc., Burlington, MA) using a single-phase laminar flow module. A fine mesh was generated for the entire subdomain containing 1, 2 or 3 cell traps with an automatic triangulation method. A PARDISO solver was used to solve the stationary Navier-Stokes equation using shallow channel approximation, an initial inlet pressure of 2 psi and an outlet atmospheric pressure under no-slip boundary conditions.

Macrophage cell culture

Wild type C57BL/6J mice were purchased from Jackson Laboratories. Bone marrow derived macrophages (BMDMs) were generated as previously described (18). Briefly, bone marrow was extracted from the hind legs of the mouse with a syringe. After red blood cell lysis with ammonium-chloride-potassium lysis buffer (Lonza), cells were incubated for 4 hours at 37° C with 5% CO2 in a non-TC treated plastic petri dish with BMDM media (RPMI supplemented with 10% fetal bovine serum (FBS), 100 U/mL penicillin, 100 μg/mL streptomycin, 1% sodium pyruvate, 25 mM HEPES buffer, 2 mM L-glutamine, and 50 μM 2-mercaptoethanol). After 4 hours, the non-adherent cells were transferred to a new petri dish, and incubated with BMDM media + 20 ng/ml M-CSF (Peprotech). After 3 days, 10 ml of BMDM media + 20 ng/ml M-CSF was added to the plate. 6 days after plating, cells were harvested in PBS + 5 mM EDTA with gentle scraping. We routinely obtain >98% Cd11b+ macrophages using this protocol. The device was seeded with a cell suspension at a density of 5×10₆ cells/ml and cells were incubated for 6 hours on the microscope stage at 37°C and 5% CO₂ in order to allow the cells to spread prior to cell stimulation. All mice were housed in the Yale Animal Resources Center in specific pathogen-free conditions. All animal experiments were performed according to the approved protocols of the Yale University Institutional Animal Care and Use Committee.

RAW264.7 EGFP-RelA cells were cultured in Dulbecco’s modified Eagle’s medium (DMEM) supplemented with 25 mM HEPES buffer, 10% FBS, 100 U/mL penicillin, and 100 μg/mL streptomycin) at 37°C with 5% CO2. Cells were passaged upon reaching 80% confluency (approximately every 2 days).

Microfluidic device operation

Cells were centrifuged at 1000 rpm for 3 min and passed through a 35 μm cell strainer (BD Falcon) to remove cell clumps prior to loading in the microfluidic device. Cell suspensions were loaded onto a Tygon tube (Cole Parmer) coated with 3% BSA under sterile conditions and connected to the sample port in the microfluidic chip. The integrated SiSec chip has nine valves (Fig. S2A): the outlet valve (V_O), the side valve (V_S), and the center antibody barcode valve (V_CAb) for device priming, cell loading and well isolation; Movie S1 illustrates valve closure video for V_S and V_CAb. The other six valves focused in the inlets are the media valve (V_m), the ligand valve (V_lig), the cell valve (V_cell), the BSA valve (V_BSA), the biotin valve (V_B*) and the streptavidin valve (V_SAv), corresponding to media, ligand (LPS), cell loading, BSA, Biotinylated capture Antibodies and streptavidin-APC inlets, respectively. All other inlets are closed using the appropriate valves while flowing any specific liquid through the inlet.

Following placement of the SiSec chip on the humidified microscope stage incubator maintained at 37°C with 5% CO₂, the flow line primed with DMEM media (without phenol red) was allowed to fill the device and all inlets at 2 psi flow pressure after which all the valve lines were primed with phosphate buffer saline (PBS) at 10–12 psi. Care should be taken not to close both V_S and V_CAb at the same time as this would build up fluid pressure inside every well thus causing leaks between the flow layer and the glass coverslip.

An overview of the device operation and valve configurations is presented in Fig. S2B–C. Briefly, cells were loaded into the device at 1psi flow pressure for 5 min, while having all the valves closed except for V_cell, V_O, and V_S in order that cells were only exposed to the trap area and exited through the outlet, and were not caught in the antibody barcode area. The trapped cells were allowed to rest for 6 hours prior to being stimulated with LPS. V_cell and V_O were closed during the cell resting phase to prevent additional cells from flowing into the chip and to reduce evaporation of the media through the outlet. After resting, cells were visually observed to be well spread onto the fibronectin coated trap area when compared to rounded cells that were caught initially during cell loading. After the resting phase, DMEM complete media (without phenol red) was flowed through the device at 1 psi flow in order to wash away any unattached cells. Following this, LPS 10 ng/ml (Invivogen) was flowed through the chip at 1psi for 5 min while keeping only V_lig, V_S, V_Cab, and V_O open, in order to uniformly fill all wells in the device. After stimulation, only V_CAb remained opened to isolate cells in each well, while also allowing binding of the secreted cytokines to the capture antibody barcode region. Immediately after stimulation and valve closure, the cells were imaged to observe EGFP-RelA translocation and Tnf transcriptional reporter dynamics, as described below.

Antibody barcode image acquisition and processing

Following completion of the experiment, the device was flushed with DMEM to remove any unbound cytokine. The protein barcodes were analyzed using a μELISA-like approach. All the inlet valves were closed except for the V_BSA, through which a 3% BSA blocking solution was flowed through at 1 psi for 5 min to reduce non-specific binding, followed by a mixture of biotinylated detection antibodies flowed at 1 psi for 5 min and incubated for 10 min at 37°C. This relatively short incubation time was sufficient to bind to target antigens in the channel due to rapid axial convection along the channel height as previously reported (19–21). Following incubation with the detection antibodies, the chip was again briefly blocked with 3% BSA, and then streptavidin APC was flowed onto the chip for 2 min and allowed to incubate for 15 min. Excess streptavidin APC was washed with PBS for 2 min prior to imaging. Images of the barcode region over 150 wells were acquired using a 20X 0.75 NA objective on the Nikon TiE spinning disk confocal microscope (Movie S2). Regions of interest (ROIs) capturing intensities from each barcode area were extracted from the raw images. The local neighborhood background was subtracted from each barcode intensity to obtain the final value. Wells containing zero cells were used to calculate a background threshold (BT) for each secreted protein as previously described (3). The BT was calculated as mean + 2×SD of the zero-cell barcode intensity. All cells with mean barcode intensities above the BT were counted as positive for secretion. Anti-TNF-α capture antibodies were patterned at both ends of the array in order to assess the experimental noise associated with the barcode measurements. TNF secretion intensity values were averaged for downstream analyses.

Live-cell fluorescent imaging acquisition and processing

Images were captured using a Nikon TiE spinning disk confocal microscope with an Andor iXon 888 camera. EGFP-RelA, Tnf-mCherry reporter and APC signals were acquired using the 488nm, 561nm and 647nm lasers and their corresponding filter cubes, respectively, every 6 min for up to 12 hours. The images were processed using Nikon Imaging software (NIS-AR) with custom built macro panel workflows.

To extract RelA signaling dynamics and Tnf transcription dynamics, a region of interest (ROI) was drawn around the entire cell. Tnf-mCherry reporter intensity was calculated over the whole-cell ROI. We did not include a nuclear marker in the cell due to issues with altered cell behavior and toxicity. Instead, the relative variance (RV) of EGFP-RelA intensity over the whole-cell ROI was used as a proxy for EGFP-RelA translocation, because accumulation of EGFP-RelA signal in the nucleus increases the variance of the fluorescence intensity across the cell. The normalized nuclear RelA signal was calculated from RV as follows:

$$normalizednuclearRelAsignal=\frac{RelARV–min(RelARV)}{max(RelARV)–min(RelARV)}$$

To validate that we could obtain accurate dynamics with this method, we first compared changes in EGFP-RelA RV to changes in nuclear EGFP-RelA intensity (using a Hoechst counterstain) over time in RAW cells cultured in a dish and found good correlation between the time courses (Fig. S2A; r = 0.89). We further compared changes in EGFP-RelA intensity for a manually identified nuclear ROI to changes in EGFP-RelA RV in RAW cells over time. We confirmed that RV was an effective means to identify cells with a positive EGFP-RelA translocation peak within the first 250 min. For all cells with a positive nuclear RelA signal, we confirmed that the normalized nuclear RelA signal calculated from the RV was consistent with the normalized nuclear EGFP-RelA intensity from a manually tracked ROI (Fig. S2B; Pearson r = 0.73) calculated as:

$$normalizednuclearRelAsignal=\frac{RelAROIintensity–min(RelAROIintensity)}{max(RelAROIintensity)–min(RelAROIintensity)}$$

Normalization removed the experimental variation present when integrating data acquired across multiple SiSec chips (11).

The EGFP-RelA and Tnf-mCherry signals were background corrected and normalized as described. Individual time point values were removed if they were more than three local standard deviations away from the local mean within a sliding window containing 13 time points (78 minutes). Missing data points that resulted from outlier removal (< 0.25% of all data points) were linearly interpolated from the surrounding data. EGFP-RelA and Tnf-mCherry reporter traces were smoothed using three-frame running averages to reduce the influence of high-frequency noise. Traces were analyzed for traps containing a single cell. Cells that divided during the course of the experiment were removed.

Feature extraction from EGFP-RelA and Tnf-mCherry reporter time courses

The following features were extracted from the EGFP-RelA relative variance time courses:

• Time of RelA max peak: The time at which EGFP-RelA variance is maximum. We limited this feature to values less than 170 min based on a previous study(17). If the time to max occurred more than 170 min after stimulation, the time course was analyzed separately.

• Duration of RelA max peak: The amount of time the maximum peak stays above 50% of its maximum height. This descriptor was determined by fitting a Gaussian curve to the EGFP-RelA peak (all fits had R₂ > 0.5 and MSE < 0.2).

We confirmed that the value of features extracted from RV were comparable to those extracted from the absolute nuclear EGFP-RelA translocation time courses (Table S1).

The following features were extracted from the Tnf-mCherry time courses:

• Tnf peak magnitude: the maximum Tnf-mCherry reporter fluorescence intensity value after removal of outliers.

• Tnf AUC: The area under the curve of the normalized Tnf-mCherry reporter time course.

• Maximum fold-change in Tnf: the maximum Tnf-mCherry reporter fluorescence intensity divided by the initial intensity value.

• Time to max Tnf: the time at which Tnf-mCherry reporter fluorescence intensity is maximum.

Statistical analysis of data

Correlations were determined using the Pearson correlation coefficient and p-values computed by transforming the correlation to create a t statistic. Data were plotted and statistical analysis performed using MATLAB.

Results

Microfluidic device for paired measurements of dynamic signaling and secretion end points in single cells

We designed a two-layer polydimethylsiloxane (PDMS) microfluidic device for deterministic capture and culture of single cells in isolated wells (Fig. 1A). The flow layer permits precise fluid exchanges within the device while the valve-layer allows the individual chambers to be sealed in order to isolate secretion activity from cells within the chambers (Fig. 1B). Separately a glass slide was flow patterned with antibodies in order to capture secreted proteins as previously described₁₅. The final device contained 150 chambers with a 1.6 nl well volume. S-cell traps designed to deterministically capture single cells (22) were integrated into the wells in arrays of one, two, or three traps (Fig. 1C). Placing multiple traps at close proximity in the same well provided the potential for examining paracrine signaling between cells. The antibody array is physically separated from the cell traps and can be functionalized with up to 5 capture antibodies specific for secreted protein targets (Fig. 1D).

The single-cell imaging and secretion (SiSec) chip is designed to capture one, two, or three cells in each well depending on the number of traps placed within the well (Fig. 2A). Simulations of fluid flow velocity across wells showed that as the number of traps increased within the well, flow velocity across the traps was reduced due to more fluid resistance caused by the presence of the trap structures (Fig. 2B). Nevertheless, sufficient gaps between the traps allowed free cells to maneuver through the traps (both filled and empty) to reach other wells. More than half of all traps contained a single cell, while approximately 25% of the traps remained empty (Fig. 2C). This proportion was advantageous because 0-cell wells were used to calculate background intensities of the antibody barcodes (see Methods). This loading efficiency resulted in approximately 60% of all well types (1-, 2-, and 3-trap wells) having the correct number of cells loaded (Fig. 2D) with precise loading of 1- and 2-trap wells slightly greater than 60%, and 3-trap wells lower at ~40% (Fig. 2E). Overall, the integrated SiSec microfluidic device allowed for loading, capture, and isolation of single cells, as well as two- and three-cell groups.

Fig. 2. Cell capture efficiency in the integrated microfluidic device.

(A) SEM images (top) of wells with one, two, or three traps. Scale bar represents 50 μm. Fluorescent images of RAW264.7 RelA-EGF cells captured in traps (bottom). (B) Flow simulation results of surface velocity and streamline plots at an initial inlet pressure of 2 psi. illustrating flow velocity across wells. (C) Efficiency of single-cell capture per trap under indicated experimental conditions. Data presented as the mean ± standard error of the mean (s.e.m.) for four experiments. (D-E) Percent of wells containing 1-, 2-, or 3-cell traps loaded with too few or too many cells, or the correct number (i.e., ‘Just right’) (D) and a detail of correct loading by trap number (E). Data presented as mean ± standard error of the mean (s.e.m.) for two experiments.

Measurement of multiplexed secretion of cytokines and chemokines following LPS stimulation of macrophages

We functionalized the microarray with antibodies to target four chemokines and cytokines (C/Cs): C-C motif chemokine ligand 2 (CCL2), C-C motif chemokine ligand 3 (CCL3), C-C motif chemokine ligand 5 (CCL5), and tumor necrosis factor alpha (TNF-α). Anti-TNF-α capture antibodies were patterned at both ends of the array (i.e., the closest and farthest position from the cell traps) in order to assess the experimental noise associated with the barcode measurements.

Our goal was to optimize a protocol to measure signaling dynamics, transcription dynamics, and secretion output in the same single cells. Therefore, we measured secretion from a previously described RAW264.7 murine macrophage cell line expressing an endogenous Rela promoter-driven fusion of NF-κB RelA and the enhanced green fluorescent protein (EGFP-RelA) together with a Tnf promoter-driven fast-degrading mCherry fluorescent protein (Tnf-mCherry; referred to as the RAW dual reporter line)(8). RAW cells were flowed into the device and allowed to adhere for six hours, and then stimulated with 10 ng/ml LPS for 12 hours. At the end of the stimulation, LPS media was flowed out of the device and replaced with media containing a mixture of primary biotinylated detection antibodies (see Methods for details). Following detection antibody labeling, streptavidin APC was perfused into the device to label all the bound biotinylated antibodies.

LPS stimulation produced a distinct fraction of cells secreting each C/C but with highly variable amounts between cells (Fig. 3A). A threshold intensity for positive secretion was calculated from the background intensity measured in the wells with 0 cells (see Methods for details). Applying this threshold, we found that approximately 20% of RAW dual reporter cells secreted TNF-α following LPS stimulation, while between 10–25% secreted CCL2, CCL3 and CCL5 (Fig. 3A). The fraction of RAW dual reporter cells secreting each C/C following LPS stimulation was similar across experiments (Fig. 3B).

Fig. 3. Multiplexed measurements of secretion from LPS-stimulated macrophages reveal more coordination for primary BMDMs than in RAW cells.

(A) Violin plots of intensity of the secreted target measured in RAW dual reporter cell line following 12 hours of 10 ng/ml LPS stimulation. The red line indicates the background threshold calculated from 0-cell wells and the red + indicates the average intensity of the distribution. The fraction of cells secreting above background is indicated for each C/C. Data are presented from one representative experiment (n = 71). (B) Percent of cells secreting the indicated C/C above background threshold at 12 hours following 10 ng/ml LPS stimulation for 0- and 1-cell wells. Data are presented as the mean ± s.e.m. of three biological replicates. Correlation between intensities of repeated TNF antibody barcodes for all 1-cell wells from three biological replicates. R = 0.86, p < 0.0001, n = 229). (C) Heatmap of pairwise correlations of secretion intensity for all C/C pairs across single cells for RAW dual reporter cells (top) and BMDMs (bottom). Significant correlations are indicated in white text (p < 0.05).

Correlation between intensities from the two TNF-α barcodes across single RAW dual reporter cells was high (R = 0.86; Fig. 3C), although we did find that the correlation between barcodes was reduced at lower intensity values, as evidenced by the cells in which secretion was only detected on one barcode. However, there was no systematic difference between the average intensity of barcodes located close to or far from the secreting cell, suggesting that detection was not diffusion limited (Fig. S4A). We did not observe strong pairwise correlations between C/Cs secreted from the same RAW cells. The only significant correlations were observed between TNF-α and CCL3 (R = 0.31) and between TNF-α and CCL5 (R = 0.26) (Fig. 3D, top).

We also cultured and stimulated primary bone marrow derived macrophages (BMDMs) in the device. Following differentiation with M-CSF for one week, BMDMs were flowed into the device and allowed to adhere for six hours, and then stimulated with 10 ng/ml LPS for 12 hours. We also collected BMDM secretion data from the wells with 2 and 3 cells. Due to limited cell numbers, we combined 2 and 3 cells to make meaningful comparisons with the single-cell data.

Following stimulation of the BMDMs with 10 ng/ml LPS, we observed a subset of cells secreting each C/C above background, and the fraction secreting was substantially larger than the RAW dual reporter cell line (Fig. S4B). TNF was secreted by the largest fraction of cells (56%), while secretion of CCL2, CCL3, and CCL5 varied between 31–46%. In wells with 2 or 3 cells, we did not observe any increase in the fraction of cells secreting or average secretion intensity for any of the chemokines, but we did observe that multiple cells secreted slightly higher average levels of TNF-α than single cells (Fig. S4B). Correlation between intensities from the two TNF-α barcodes across single BMDM cells was high (R = 0.80; Fig. S4C). Interestingly, we observed that the pairwise correlations between C/Cs secreted from the same cells were substantially increased in BMDMs relative to RAW cells (Fig. 3D), suggesting that, despite significant cell-to-cell heterogeneity, there are sources of noise that are shared between the C/Cs that lead to more coordination in the BMDM secretion program. Overall, our results show that we can robustly measure secretion of four C/Cs in single macrophage cells in the integrated SiSec microfluidic device.

Live-cell imaging of NF-κB RelA signaling and Tnf transcription dynamics in the microfluidic device

In addition to secretion measurements, we also acquired fluorescent time-lapse images of EGFP-RelA and Tnf-mCherry reporter across all wells in the device (Fig. 4A). Following stimulation with 10 ng/ml LPS, images were taken every 6 minutes for 12 hours. We were most interested in the timing of RelA signaling, and therefore we developed a method to automate extraction of dynamic traces of RelA nuclear translocation by calculating the relative variance in EGFP intensity across the whole cell (see Methods for details). After extracting traces based on variance, we further filtered these traces to select for those with a maximum peak within 170 minutes, consistent with a previous study of LPS-stimulated RelA signaling(17), resulting in translocation traces from 26 single cells (Fig. 4B). For this cell subset, we tracked nuclear intensity over 170 minutes and confirmed that the dynamic traces obtained from relative variance matched those obtained from nuclear intensity (Fig. S3). We also extracted Tnf-mCherry intensity traces from the same cells (Fig. 4C). Overall, we observed that LPS stimulation of cells in the SiSec chip results in biologically relevant RelA signaling dynamics and increases in Tnf transcription.

Fig. 4. On-device analysis of EGFP-RelA signaling dynamics and Tnf-mCherry transcriptional reporter dynamics.

(A) Time-lapse images of EGFP-RelA and Tnf-mCherry for the same cell trapped in the microfluidic device and stimulated with 10 ng/ml LPS. (B-C) EGFP-RelA time course of nuclear translocation as measured by relative variance (B) and Tnf-mCherry reporter intensity (C) including all individual traces (black) and the average trace (red) for all cells with an EGFP-RelA peak maximum occurring < 170 minutes (top). Data are combined from two biological replicates (n = 26). (D-E) Schematic of the features extracted from EGFP-RelA traces (D) and the Tnf-mCherry traces (E). (F) Graphs of significant Pearson correlations between the time to max (left) and duration of max (right) EGFP-RelA peak with the time to max Tnf-mCherry. (G) Table of EGFP-RelA feature correlations with all Tnf-mCherry features. Significant correlations are indicated in white text (p < 0.05).

We next looked for correlations between the timing of EGFP-RelA and Tnf-mCherry measurements to provide information about how cell-to-cell variability in the timing of RelA signaling modulated variability in Tnf transcription across cells. To do this, we extracted features from the EGFP-RelA and Tnf-mCherry time courses, similar to previous studies (8,9,12,23). For RelA signaling, we extracted time to max EGFP-RelA peak and duration of max peak (Fig. 4D and Methods). Because we relied on relative variance to analyze EGFP-RelA nuclear translocation, we were limited to dynamic features and were not able to extract features based on changes in intensity across cells. For Tnf transcription, we extracted time to max Tnf-mCherry, Tnf-mCherry peak magnitude, area under the curve (AUC) of Tnf-mCherry intensity, and maximum fold-change of Tnf (Fig. 4E).

We found that for the subset of cells exhibiting a maximum RelA peak within 170 minutes, there was a moderate correlation between time to max Tnf-mCherry transcription intensity and both time to max EGFP-RelA peak and duration of the peak (Fig. 4F). We did not observe any significant correlations between the timing of RelA signaling and Tnf-mCherry max intensity, AUC, or max fold change (Fig. 4G), similar to a previous report(8). Our data suggest that the timing of RelA translocation moderately impacts Tnf transcription activity across cells.

Correlation between RelA signaling and secretion output across LPS-treated macrophages

Having successfully measured EGFP-RelA signaling dynamics and Tnf-mCherry transcriptional reporter dynamics, we next looked to see how the metrics extracted from these measurements correlated to secretion of four proteins across the same single cells. As previously described, we only included cells with a complete EGFP-RelA time course (up to 250 minutes) exhibiting a maximum peak within 170 minutes (Table S2). When comparing the timing of RelA signaling to secretion, we observed a significant inverse correlation between TNF secretion intensity at 12 hours and time to max EGFP-RelA peak (Fig. 5A, left; R = −0.52), but not with duration of max EGFP-RelA peak (Fig. 5D, right). We also found a significant inverse correlation between CCL3 secretion intensity at 12 hours and time to max EGFP-RelA peak (Fig. 5B, left; R = −0.59). There was no correlation between the timing of the RelA peak and CCL2 or CCL5 secretion, potentially due to the requirement of other pathways in inducing these responses (e.g., the requirement of IRF3 as a transcription factor for CCL5(24)). We also compared the mean secretion intensity for each C/C between cells exhibiting an initial EGFP-RelA translocation peak with those that did not, and we did not observe any significant difference between these groups, even for TNF and CCL3 (Fig. S5). Overall these results suggest that the timing of RelA signaling is an important determinant of the downstream response; but other factors also contribute, consistent with previous studies (17).

Fig. 5. Timing of RelA signaling is significantly correlated with TNF and CCL3 secretion.

(A-B) Pearson correlation between TNF secretion (A) and CCL3 secretion (B) and the time to max EGFP-RelA peak (left) or duration of max EGFP-RelA peak (right). Data are from one experiment for which EGFP-RelA, Tnf-mCherry, and secretion were all simultaneously recorded (n = 16). EGFP-RelA and Tnf-mCherry traces and secretion data are representative of two and three biological replicates, respectively (presented in Figs. 3 and 4).

Surprisingly, we did not find any significant correlations between features of Tnf-mCherry and average TNF secretion intensity. Taken together, our results suggest that even though variation in the timing of RelA signaling partly explains variations in Tnf transcription and TNF and CCL3 secretion, there likely exists heterogeneity in post-transcriptional processing and/or noise in the experimental measurements that leads to significant differences in these response readouts.

Discussion

It is increasingly clear that cell-to-cell heterogeneity is important for biological function, but the sources of biological heterogeneity are difficult to quantify. Here we described a novel microfluidic device to connect signaling and transcription dynamics, as measured by live-cell imaging, to end point secretion measurements in the same single cells in order to determine the fidelity of information transfer across these biological processes. We validated the robustness of secretion measurements (Fig. 3) and live-cell imaging measurements (Fig. 4) for RAW 264.7 murine macrophages stimulated with LPS.

Other nanowell and valve-based microfluidic devices have been previously reported for tracking cells longitudinally via imaging prior to measuring secretion. For example, Yamanaka et al combined imaging of cell-cell interactions, specifically between natural killer cells and their target cells, with measurements of secreted proteins in nanowells by microengraving(25). However, this device did not use valve-based control of media exchanges, and therefore did not permit quantitative tracking of signaling events that occur on relatively short time scales following stimulation. Junkin et al reported capturing 40 single cells in a high-content multicompartmental valve-based microfluidic chip to image signaling dynamics and quantify dynamic single-cell secretomic responses captured using antibody-coated beads₁₆. This device combined automated culture and stimulation with measurement of protein translocation by live-cell imaging and repeated multiplexed measurements of cytokine secretion using antibody-coated beads. However, the number of valves required to operate this elegant device reduces its scalability.

The SiSec chip is a valve-based device that is similar to the device described in Junkin et al, enabling live-cell imaging of dynamic translocation and transcription events, followed by a static endpoint secretion measurement. One advantage of the SiSec chip is that cells are deterministically captured using cell traps to improve loading efficiency. Our current cell seeding density was optimized for the single-cell traps; however, users can flow cells at a higher seeding density to capture more cells for the multi-trap chambers. Another difference is that secreted proteins are measured by antibody barcode capture in the SiSec chip. The current chip is limited to five cytokine barcodes; however, the device could be adapted to accommodate more barcodes. A final important difference is that secretion is measured at a static endpoint, rather than sampling cytokines in the media over time. Although the SiSec chip does not provide temporal information about secretion like the device by Junkin et al, the valves still allow for time-varied stimulations and the overall design and operation of the device is simpler as a result.

The SiSec device has several additional limitations. Calibration curves suggest that the limit of detection for cytokines in the device is approximately 50–100 pg/ml (Fig. S1A). Background cross-reactivity with other proteins in the media may increase the detection threshold. For example, we observed variations in background thresholds for CCL2 and CCL3 across experiments, which might be due to constitutive secretion and binding during cell loading. However, we note that we found no significant cross reactivity between recombinant C/Cs (Fig. S1B). We are unable to rule out the possibility that sequestration of cytokines by the barcodes affects cell responses. In addition, our analysis of TNF replicate barcodes suggests that at low intensity values, variation in barcode capture affects the precision of the quantitative result (Fig. 3C). However, we were able to make observations about the sources of secretion heterogeneity between cells.

For a subset of cells with a strong initial RelA signaling peak, we found that the earlier peaks were significantly correlated with higher levels of TNF and CCL3 secretion (Fig. 5). This finding was based on a small sample size, but is consistent with existing evidence that differences in timing across cell processes account for cell-to-cell variability in downstream responses. For example, timing of processes upstream of mitochondrial outer membrane permeabilization (MOMP) predicts time to apoptosis in response to TRAIL, while the duration of MOMP itself is almost identical across cells (26). The timing of the initiation of cytokine secretion also varies considerably across a population of T cells following T cell receptor (TCR) stimulation (27), and our results suggest that this could be a result of early differences in upstream TCR signaling dynamics. Interestingly, the distribution of NF-κB translocation timing in response to TNF stimulation is regulated and a heritable property across cells (28).

RelA first peak timing also moderately correlated with the timing of maximum Tnf transcription, however it was not a robust predictor of overall Tnf transcription (Fig. 4F–G). Sung et al. developed the RAW dual reporter line that we used in our study and they similarly did not find a significant correlation between RelA peak timing and Tnf-mCherry reporter synthesis(8). In response to LPS stimulation, the timing of the first RelA peak is largely determined by the association of the macromolecular signaling complex involving MyD88(29); however both MyD88 and TRIF are important for regulating Tnf transcription (30), which might account for the poor correlation between RelA peak timing and Tnf transcription.

Our findings are complementary to conclusions from other studies examining how cell-to-cell heterogeneity in LPS-stimulated RelA signaling contributes to variability in TNF transcription and secretion. In the study by Sung et al, the strongest correlation was between the area under the curve of nuclear RelA occupancy over approximately 23 hours of tracking and total Tnf-mCherry reporter synthesis (R = 0.76). Because we did not compare absolute EGFP-RelA signaling intensity across cells, we could not test for this correlation in our data. We note that we saw less LPS-stimulated RelA translocation in our experiments relative to Sung et al., however this might be due to differences in cell culture in the microfluidic device, including the effect of cell isolation(3).

Interestingly, Junkin et al. measured RelA signaling and TNF secretion in single RAW264.7 cells and found no correlation between the amplitude of the RelA signaling peak and the TNF secretion peak (17). This observation is consistent with the requirement for other signaling pathways and transcription factors in the regulation of the LPS-mediated inflammatory response program, including IRF3 and the MAP kinases ERK and p38(24,31). Junkin et al further reported that a positive RelA signal in a single cell (i.e., RelA nuclear translocation above background control) was a good predictor of TNF positivity in the same cell. We stimulated with a 50-fold lower dose of LPS than the Junkin et al study (10 ng/ml versus 500 ng/ml), and had fewer cells with a positive RelA signal (defined as cells that exhibited a RelA signaling peak within 170 minutes). Perhaps due to these differences, when we compared cells positive and negative for RelA signaling, we found that both groups of cells had similar TNF secretion intensities (Fig. S5). We speculate that this could be due to “late” RelA signaling, combined with a MAPK response that was not measured. In any case, the high fraction of non-responders confounded our attempts to find correlations between RelA metrics and TNF outputs across all cells.

Supporting the idea that non-responding cells might confound correlations, Zhang et al demonstrated that non-responding cells confounded the information transfer of the TNF-mediated NF-kB signaling pathway within single cells (32). Consistent with this, we found that if we restricted our comparisons to those cells that exhibited RelA signaling, there was a significant correlation between the time and duration of the maximum RelA peak and TNF secretion intensity (Fig. 5C). Junkin et al did not consider the timing of RelA signaling and also did not stratify cells by the timing of their initial RelA peak and so we could not compare this result to their study.

It was somewhat surprising that we did not find any correlation between features of Tnf transcription and TNF secretion. However, another study connecting ERK signaling to downstream responses showed that cells that exhibit similar transcriptional responses can end up with large differences in protein abundances (33). These discrepancies are likely due to differences in translational and post-translational regulation, as well as experimental noise in our measurements.

Our device was designed with the potential to measure signaling and secretion for two- and three-cell groups in order to study how paracrine signaling communication between cells affects the LPS-mediated response (3,4). Although we did not have sufficient preliminary data to report in this study, we anticipate that this capability will be useful in the future.

Overall, we have developed a new microfluidic device capable of connecting live-cell readouts of signaling and transcription to secretion in the same single cells. Our proof-of-concept data indicate that these measurements will provide interesting insights into the sources of cell-to-cell variability and its regulation.

Insight, Innovation, Integration.

There is emerging evidence that cell-to-cell variability in macrophage secretion contributes to immune system function. However, the sources of variability are not well defined because it is technically challenging to perform different types of experimental measurements on the same single cell. To begin to address this, we developed a microfluidic device to image NF-κB nuclear translocation dynamics and Tnf transcription dynamics in macrophages using fluorescent live-cell reporters, followed by analysis of secretion of TNF and three other proteins. By integrating these measurements, we report evidence that the timing of NF-κB signaling in part determines secretion output.