Single-Cell Virology: On-Chip Investigation of Viral Infection Dynamics.

Microfluidic device design and use To monitor infections of single cells of sufficient statistical power to reach meaningful conclusions, we fabricated a microfluidic device capable of observing as many as 6,400 intracellular events using fluorescence detection ( Fig. S2A ). We designed the device to accommodate as many as four independent experiments by engineering four separate sample inlets, with each inlet connected to 1,600 different wells ( Figs. 2A and S2A ). Once loaded, each well was isolated from the others by using the integrated pneumatic-control lines ( Figs. 2A and S2A ). We demonstrated the effectiveness of the pneumatic seal by showing that not all cells on chip could be infected, even when a lytic infection occurred in an adjacent well ( Video S3 ). We installed the device on the stage of an inverted fluorescence microscope equipped for cell culture ( Fig. S2B ). We used a motorized stage to monitor changes in fluorescence over time automatically using time-lapse microscopy, with a snapshot taken every 30 min for 24–36 h ( Figs. 2B and S2 ). We processed images using customized MATLAB and R codes to produce time courses for GFP fluorescence ( Figs. 2B and S2 ).

Experimental design, data acquisition and data-analysis pipeline Experiment We infected HeLa S3 cells with GFP-PV at a MOI that guaranteed infection of most cells. Note that the unit of MOI used in our studies is genomes per cell instead of plaque-forming unit (pfu) per cell because of the variable nature of the genome:pfu ratio ( Korboukh et al., 2014 ). On average, 1,000 GFP-PV genomes were required to establish a single plaque ( Fig. S1B ). We performed the infection off-chip, and diluted infected cells to a density that would maximize the number of wells containing only one cell. We monitored wells for green fluorescence at 30-min intervals for a period of 24–36 h. After accounting for the empty wells and other events that precluded inclusion of a well ( Fig. S3 ), 100 to 400 single cells were available for analysis. We normalized the fluorescence intensity of each cell to the background fluorescence of its well and plotted normalized fluorescence intensity (NFI) as a function of time post-infection ( Fig. 2C ). We observed substantial variation in the kinetics of replication of GFP-PV in each cell ( Fig. 2C ). The time in which fluorescence could be detected, the rate at which fluorescence intensity rose, the maximum fluorescence intensity observed, and when the infected cell lysed, if at all, were different in each of the cells ( Figs. 2C and S3 ). We plotted the average intensity for each time point as a function of time, which yielded a curve similar to a growth curve observed by population methods ( Fig. 2D ). GFP was dispersed and diluted in the well as cells lysed, leading to a reduction in GFP fluorescence ( Fig. 2D ). The fluorescence decay reached a plateau well above background levels because not every cell producing GFP lysed over the 24-h time period monitored ( Fig. 2C and S3 ).

Data analysis

We sorted all of the time courses of successful infections into two classes: infection without lysis ( Figs. 2E and S3C ) and infection with lysis ( Fig. 2F and S3D ). We modeled the former as a sigmoidal rise and the latter as a sigmoidal rise followed by a sigmoidal decay. These models were sufficient to fit all of our experimental observations ( Fig. S3E ). We defined five phenomenological parameters to describe the sigmoidal rise that forms the foundation of our data analytics. These parameters are: (i) maximum, the maximum fluorescence intensity observed; (ii) midpoint, the time at which the intensity was half of maximum; (iii) slope, the change in intensity per hour measured at the midpoint; (iv) start point, the time at which fluorescence was first detected; and (v) infection time, the length of time required to go from the start point of fluorescence to the maximum fluorescence ( Fig. S3 ). Details of the code used for data analysis have been described ( Caglar et al., 2017 ).

Implementation of the pipeline

To test the ability of the single-cell experiment to identify statistically significant differences in experimental outcomes caused by an experimental variable or perturbation, we evaluated the impact of varying the multiplicity of infection. We chose multiplicity of infection because one-step-growth experiments demand infection of all cells. The single-cell experiment does not, thus providing an assessment of the dependence of infection outcome on multiplicity of infection. We infected HeLa S3 cells with 50, 500 or 5,000 genomes per cell, corresponding to multiplicities of infection of 0.05, 0.5 or 5, respectively, using pfu. We observed 93–202 individual cells that were successfully infected and analyzed the data as described above to produce distributions for each parameter at each MOI ( Figs. 2G – 2K ). We compared the distributions for the maximum parameter under the three different conditions; the curves did not completely superimpose ( Fig. 2G ). We were conservative in our approach to statistical analysis. Instead of asking if pairwise differences existed for distributions and risking over-interpretation of the data, we asked only if pairwise differences existed in the mean of the distributions ( Table S1 ). Therefore, we used t-tests to perform statistical analysis and required P-values less than 0.05 to define a statistically significant difference between two distributions. Using these criteria, the means of the maximum parameter shown in Fig. 2G were not different ( Table S2 ). We observed statistically significant differences for unique pairwise comparisons of the means for the slope and midpoint parameters ( Figs. 2H and 2I and Table S2 ). For the start-point parameter 50 and 500 were different when compared to 5,000 but not when compared to each other ( Fig. 2J and Table S2 ). For the infection-time parameter 500 and 5,000 were different when compared to 50 but not when compared to each other ( Fig. 2K and Table S2 ). We conclude that our pipeline is sufficient to take a population-level evaluation of viral growth kinetics and deconvolve this single curve into distributions of five fundamental parameters. Further, we conclude that we can identify statistically significant differences between parameters caused by an experimental variable, even a variable as insipid as multiplicity of infection. Importantly, we performed multiple experiments with different stocks of wild-type virus and HeLa cells without observing statistically significant differences for any of the parameters ( Table S3 ). To determine if the five parameters used to describe infections on the single-cell level report on the same or different biological properties or processes of the virus, host and/or interaction thereof, we performed a pairwise comparison of the parameters ( Fig. 3 ). We reasoned that parameters governed by the same biological properties or processes will be correlated. We observed a strong correlation between the midpoint and start-point parameters ( Fig. 3J ) and weaker correlation between maximum and slope parameters ( Fig. 3C ). However, all other pairwise comparisons were not correlated ( Fig. 3 ). We conclude that parameters determined using the single-cell experiment permit greater elaboration of an infection cycle than permitted by population experiments.

Single-cell analysis reveals phenotypes for an attenuated virus masked by population methods RNA-dependent RNA polymerase (RdRp) fidelity is a well-established virulence determinant ( Arias et al., 2016 ; Arnold et al., 2005 ; Korboukh et al., 2014 ; Pfeiffer and Kirkegaard, 2005 ; Van Slyke et al., 2015 ; Vignuzzi et al., 2006 ; Zeng et al., 2013 ). Studies with PV and its RdRp were among the first to make this connection by showing that changes in nucleotide-incorporation fidelity of the RdRp produce viruses that are attenuated in vivo ( Arnold et al., 2005 ; Korboukh et al., 2014 ; Vignuzzi et al., 2006 ). However, when these same viruses were compared to wild-type using conventional methods, a difference between the mutant and wild type could not be discerned ( Korboukh et al., 2014 ; Vignuzzi et al., 2006 ). This circumstance provided us with the opportunity to determine if single-cell analysis enables distinction between a PV fidelity mutant and wild type. For this experiment, we used a PV mutant expressing a mutator RdRp, H273R ( Fig. S4B ) ( Korboukh et al., 2014 ). We introduced the H273R-encoding mutation into GFP-PV cDNA. In the context of GFP-PV, the mutation caused only a small reduction in the specific infectivity of the virus ( Fig. S1B ). We compared H273R PV to wild type in infections using 50 or 5,000 genomes per cell. At low MOI, we observed only one statistically significant difference between the mutant and wild type ( Fig. 6 and Table S6 ). This difference was for the start-time parameter ( Fig. 6G and Table S6 ). At high MOI, however, we observed numerous differences: slope, midpoint, start point and infection time ( Fig. 6 and Table S6 ). H273R PV exhibited a delayed start of replication relative to WT PV ( Fig. 6H ). Consistent with biochemistry ( Korboukh et al., 2014 ), H273R PV actually replicates faster than WT PV ( Fig. 6D ), leading to faster completion of replication ( Fig. 6J ). Because of our ability to count the number of infections established, we observed directly the reduced specific infectivity of H273R PV relative to WT PV ( Fig. 6K ). This analysis provided unprecedented, detailed insight into the differences in growth properties of a mutant virus relative to wild-type virus. Some of the observed differences could actually explain the attenuated phenotype. For example, delayed start point indicates a sub-optimal interaction of the mutant virus with host cell. In the context of a heightened antiviral state, a delayed start point could produce new bottlenecks by diminishing the yield of progeny by several logs ( Figs. S4C and S4D ). The reduction in the number of infections established at low MOI indicates a frailty of the virus population to bottlenecking events. We conclude that studies on the single-cell level will provide a more comprehensive and in-depth perspective of virus biology and virus-host interactions in cell culture that may inform our understanding of pathogenesis in vivo .

Single-cell analysis reveals more about an antiviral agent than its efficacy Antiviral drug discovery and characterization represent a major emphasis of molecular virology. Discovery and characterization of antiviral agents are driven by the mission to minimize the IC 50 , the concentration of drug required to inhibit virus multiplication by 50%, and to maximize the CC 50 , the concentration of drug required to cause cytotoxicity in 50% of treated cells. It was apparent to us that our single-cell-analysis platform could provide information not only on the efficacy of a drug but also on the properties of the viruses most susceptible to drug action as well as the fate and properties of the viruses that survive drug action. It was also possible that we could determine if a drug targeted a cell-dependent and/or virus-dependent aspect of the life cycle. For this study, we chose to use 2′-C-methyladenosine. Inside of cells, this nucleoside is converted into the triphosphorylated form, which then serves as a substrate for the viral RdRp and leads to termination of RNA synthesis ( Carroll et al., 2003 ). This drug was the first in the class of non-obligate chain terminators that paved the way for development of sofosbuvir, the antiviral nucleoside used in combination therapy to cure hepatitis C virus infection ( Appleby et al., 2015 ; Carroll et al., 2003 ; Murakami et al., 2010 ). We grew cells in the absence or presence of 50 μM 2′-C-methyladenosine. Experiments performed at the population level identified this concentration as the IC 50 . We infected cells with WT PV at a MOI of 5,000 genomes per cell. As expected, we observed a 50% reduction in the number of infections established in the presence of 50 μM 2′-C-methyladenosine ( Fig. 7A ). Distributions for only two parameters changed in the presence of the drug; these were for maximum ( Fig. 7B ) and slope ( Fig. 7C ). The yield from replication was reduced more than the rate of replication as might be expected for an RNA-chain terminator. A pairwise comparison of the maximum and slope parameters in the absence and presence of the drug revealed a selective loss of members of the virus population exhibiting the highest values for the maximum and slope ( Fig. 7G ). This observation of survival of the least-fit members of the viral population could only be observed using a single-cell-analysis platform. We conclude that single-cell analysis will provide greater insight into the mechanisms of action, viral targets and perhaps even cellular off-targets of antiviral therapeutics.

EXPERIMENTAL PROCEDURES Device fabrication

The isolated well array microfluidic device was designed with drafting software (AutoCAD). This device consists of four sample loading inlets (or more depending on the application) and 6,000 wells, each of which has a dimension of 120 μm × 120 μm × 100 μm (length × width × height). Designs including control layer, flow layer and wells layer ( Fig. S2A ) were printed at 20,000 dots per inch (dpi) on transparent masks. The silicon molds were fabricated with standard photolithography techniques as described in Supplemental Experimental Procedures .

Experimental setup

The experimental setup of our platform for automated monitoring of single-cell viral infections is shown in Fig. S2B . The system was built on an inverted microscope (Eclipse Ti, Nikon, Japan) equipped with a motorized stage (ProScan III, Prior Scientific, USA) and a microscope incubation system (INUBTFP-WSKM-GM2000A, Prior Scientific, USA). The microfluidic device connected to a customized pneumatic valve system was placed inside the incubation chamber. The cell/virus solutions were injected into the well array microfluidic device through different outlets. After loading, the pneumatic valves (the pneumatic control channels were filled with DI water) were closed by compressed nitrogen (30 psi) to seal single cells into each individual well and to prevent the spread of virus into the neighboring wells. The cells were isolated in individual wells and cultured in a 37 °C humid environment with 5% CO 2 level.

Data acquisition and processing

The time-lapse images (including bright-field and fluorescence) of multiple positions of the microfluidic device were automatically acquired with a 10× objective and a charge-coupled device (CCD) camera (CoolSNAP HQ2, Photometrics, USA) controlled by software (NIS-Elements AR, Nikon, Japan) every 30 minutes for 21 hours or 33 hours starting at 3 hpi. Fig. S2 shows the stitched images of wells connected to one inlet and enlarged images of 12 wells at 4 different time points. After data acquisition, the time-lapse images were processed with a customized MATLAB script to extract fluorescence intensity of single infected cells and background fluorescence intensity of the well. Additional details are described in Supplemental Experimental Procedures .

Statistical analysis and data availability

Statistical analysis was carried out in R. We used the R package sicegar ( Caglar et al., 2017 ) to classify observed time courses of normalized GFP intensity and to extract relevant parameters such as maximum intensity, midpoint, and slope. Mean parameter values among conditions were compared using the R function pairwise.t.test, which performs pairwise comparisons of all conditions while correcting for multiple testing. All raw GFP time courses, processed data, and analysis scripts are freely available at: https://github.com/clauswilke/Guo_etal_SCV .

Experimental setup

The experimental setup of our platform for automated monitoring of single-cell viral infections is shown in Fig. S2B . The system was built on an inverted microscope (Eclipse Ti, Nikon, Japan) equipped with a motorized stage (ProScan III, Prior Scientific, USA) and a microscope incubation system (INUBTFP-WSKM-GM2000A, Prior Scientific, USA). The microfluidic device connected to a customized pneumatic valve system was placed inside the incubation chamber. The cell/virus solutions were injected into the well array microfluidic device through different outlets. After loading, the pneumatic valves (the pneumatic control channels were filled with DI water) were closed by compressed nitrogen (30 psi) to seal single cells into each individual well and to prevent the spread of virus into the neighboring wells. The cells were isolated in individual wells and cultured in a 37 °C humid environment with 5% CO 2 level.

Supplementary Material 1 2 3 Video S1. Infections in cell culture plate. Related to Figure 1 This video shows the evolution of green fluorescence over a 24-h time period in cells infected with GFP-PV and maintained in a cell-culture plate as described in Fig. 1B . We observed two waves of fluorescence. The second wave could easily be construed to be secondary infections. 4 Video S2. Infections in isolated wells. Related to Figure 1 This video shows the evolution of green fluorescence over a 24-h time period for two cells infected with GFP-PV maintained in isolated wells as described in Fig. 1B . The wells shown were selected because two waves of fluorescence could be observed here as well. The second wave in this case must be a delayed, primary infection. 5 Video S3. Sealed wells. Related to Figure 1 and Figure 2 This video shows first the brightfield image of several wells, seven of which have at least one cell. We monitored evolution of green fluorescence over a 24-h time period. We observed fluorescence in two cells by 8-h post-infection with subsequent lysis. The other cells were never infected, thus demonstrating sealed wells incapable of supporting secondary infection using the microfluidics device.