The population (average) levels of IFNB1
transcripts in human MDDCs were measured as a function of time following NDV infection at multiplicity of infection, MOI
0.5, as shown in (see also Supplementary Figure S1
). Our previous work 
showed that at ten hours post infection both genes have reached approximately maximal expression levels. Half-maximal DDX58
expression occurred approximately 3 hours after infection, while half-maximal IFNB1
expression occurred 4–5 hours later. The time course for synthesis of the NDV L gene at 0, 3, 6, and 10 hours post infection (Supplementary Figure S1C
) showed significant NDV genome replication well before the half-maximal expression level of IFNB1
was reached. MXA
induction (Supplementary Figure S1D
) mimicked that of DDX58
(Supplementary Figure S1B
), and IFNA1
induction (Supplementary Figure S1E
) mimicked that of IFNB1
(Supplementary Figure S1A
), as expected from their common activation pathways. However, since DDX58
is an interferon-induced gene, we would expect its expression to increase only after IFNB1
Time course of IFNB1 and DDX58 induction.
Paradoxically, the reverse temporal order is observed ( and Supplementary Figure S1
). However, close inspection of the data (Supplementary Figure S2A
) shows evidence for an initial and small first phase of IFNB1
induction that begins as early as 1 hour after infection and never reaches levels above 1% of maximum until the sizeable induction observed after 6 hours. To evaluate the robustness of this low level IFNB1
induction, we performed similar experiments with MOI
0.1, 0.5, 1.0, 2.0 and 5.0 (Supplementary Figure S2
), and obtained similar results. These findings demonstrate that an early low level of induction of IFNB1
precedes - and therefore might be responsible for - the induction of DDX58
The role of this early IFNβ in the induction of DDX58
was tested by performing population and single cell assays measuring IFNB1
in the presence and absence of antibodies that attenuate extracellular interferon signaling. Cells were infected at an MOI
0.5. As shown in , a mixture of antibodies binding interferons and the interferon receptor eliminated the induction of DDX58
and dramatically reduced induction of IFNB1
, confirming the formulation that IFNβ produced from the nearly undetectable levels of early IFNB1
is responsible for the apparently paradoxical induction of DDX58
DDX58 and IFNB1 mRNA expression level.
As only half the cells were infected, in order to determine the distribution of responses observed in individual cells, we performed single cell mRNA assays (). Notably, in the absence of blocking antibodies about half the cells showed induction of IFNB1, while all cells showed induction of DDX58. In the presence of blocking antibodies, most cells showed control levels of both transcripts. Because IFNB1 is an intronless gene, the control levels of expression reflect the 4 DNA strands encoding the gene within each cell (the sense and the anti-sense DNA strands of the gene on both alleles). In the presence of blocking antibodies, a few cells were detected that showed a modest induction of INFB1 or of DDX58 greater than the levels measured in control cells.
The results shown in showed significant cell-to-cell variability in the expression level of IFNB1,
which ranged over three orders of magnitude. One factor that can differ among cells is their degree of differentiation, which is reflected in the level of the differentiation marker CD14. In order to test whether this variability contributes the large variation in gene induction with virus infection, we sorted cells into high and low CD14 subpopulations prior to infection with virus. In both groups, a similar and broad distribution of single-cell IFNB1
gene responses was observed (Supplementary Figure S3
). These results suggest that the heterogeneous levels of IFNB1
in individual cells does not result from cell heterogeneity in differentiation and reflects both the absence of expression in uninfected cells and the noise in the transcriptional induction of this gene.
What is the source of the low level of IFNB1 detected at early time points after virus infection that is necessary to initiate the positive feedback loop and generate a full antiviral response? Given the high intercellular variation in responses, the most parsimonious hypothesis is that a few infected cells are competent to induce IFNB1 before interferon activation of JAK-STAT signaling. We explored this hypothesis and its implications by developing a formal model of the system that could lead to testable predictions.
The model was agent-based (ABM) stochastically simulating intercellular IFNβ signaling and the temporal evolution of the immune response in individual cells. The constitutive RIG-I distribution across cells, and the parameters of IFNB1
induction were taken such that only the small number of cells with large amounts of RIG-I protein responded to infection. (see Materials and Methods and
for additional details). The model can be accessed at http://tsb.mssm.edu/DCresponse2viralInfABM
shows the time courses of the average induction of IFNB1 and of DDX58 obtained in the simulation, which are consonant with the experimental results shown in . A more stringent test of the model is provided by the distributions of single cell results obtained by simulation (). The model was simulated using a constant parameter set both with and without blocking antibodies and the pattern of responses in single cells was determined. The antibody blockage was implemented in the model by introducing a degradation rate for extracellular interferon β protein, which allowed it time to bind to the cell that secreted it, but made it unlikely that it would reach a neighboring cell. The effect of receptor antibody was included by reducing the receptor binding rate. Uninfected cells can be roughly characterized as those with less than ten copies of IFNB1. The RIG-I mRNA distribution in uninfected cells was considerably shifted downward when antibodies blocked the paracrine loop (), as the paracrine loop acts as an inducer of RIG-I. Furthermore, DDX58 and IFNB1 showed significant correlation within individual cells at 11 hours, as expected when each cell is activated only by autocrine signaling.
Single DC simulations with and without IFN-blocking antibodies.
In order to test the distributions predicted by these simulations, experiments were performed using single cell assays that simultaneously measured both IFNB1 and DDX58. The single cell data presented in did not include measurement of both transcripts within each individual cell and could not determine whether the small number of cells showing increased IFNB1 and DDX58 levels in the presence of blocking antibodies were the same subset of cells. As had been seen in the simulations, the experimentally obtained co-expression measurements in the presence of blocking antibodies showed a significant correlation between IFNB1 and DDX58 (). Moreover, in the presence of antibodies the level of RIG-I expression in cells drops overall (), as can also be seen in the simulation results (). The interpretation of these experimental results that use antibodies to block interferon signaling is that while the paracrine signaling appears to be efficiently cancelled, the autocrine loop is leaky, leading to the observed correlation and higher copy numbers of IFNB1 at 11 hours. The persistence of control levels of DDX58 and of IFNB1 in some, presumably uninfected, cells at 11 hours when antibodies were present also confirmed the effectiveness of the blockade of paracrine signaling. Overall, the experimental results are in agreement with the salient features about single cell response distribution in the absence and presence of blocking antibodies that were predicted by the model simulations.
Experimental results in individual DCs with and without blocking antibodies.
This consistency between model and experiment at later times provides some assurance in using simulations to investigate the distribution of responses at early time points. The results of such simulations () support the hypothesis that the initiation of the positive feedback loop results from a very small number of cells that are activated and release interferon at early time-points. This is made clear in , where we plot the number of bound receptors (which is connected to the rate of DDX58
induction) versus the IFNB1
copy number for specific cells at different times through the simulation. In the resulting trajectories each point represents the simulation result at time points separated by 10 minutes. Cellular responses fall into typical patterns for uninfected, infected early responder and infected late responder cells. For uninfected cells () there is no IFNB1
induction despite increasing receptor binding. For early responder cells () there is first IFNB1
induction followed by increasing numbers of bound receptors and IFNB1
through early autocrine and later paracrine signaling. For late responder cells () the behavior at early times is opposite to that of early responders with infinite slope for the trajectory instead of zero slope: at first the number of bound receptors increases without any production of IFNB
1 message, indicating paracrine signaling. Later in the simulation this leads to IFNB1
production with numbers increasing rapidly through the positive feedback loop. Notably, the ratio between early and late responders is 7
124 in the simulation, supporting the hypothesis that a small percentage of the population is responsible for activating the whole culture of cells. follows the same cell in two simulations, one without antibodies (solid line), and one with antibodies (dashed line), and highlights the effect of suppressing paracrine signaling on late responder cells. In the simulation with no antibodies the cell exhibits a clear late responder trajectory. However, when the paracrine signaling is suppressed the trajectory becomes similar to that of an early responder, where IFNB1
induction starts before any receptors are bound. This induction occurs late in the simulation and results in lower steady state values, due to partial blocking of the feedback loop.
Single cell simulations of early responder DCs.
Phase space trajectories of individual cells in simulations.
Thus far we have shown that the proposed mechanism (of a sub population of early responders efficiently activating the rest of the cells) is consistent with the experimental results. Cell-to-cell variations are essential for the process to occur so that interferon-induced genes will seem to be activated prior to significant interferon activation. To test this conjecture, we ran the simulation with decreasing levels of cell-to-cell variability, while maintaining the same average initial RIG-I concentration, and the same percentage of early responding cells To this end, the sensitivity of IFNB1
induction to the concentration of RIG-I (parameter
) was increased (see Supplementary Text S2
, Changing the Variance and Maintaining Early Responder Percentage, for details). shows the original simulation results and the results of a simulation with initial conditions with 10 times less variability in the initial DDX58
distribution. As the variance is decreased, the sensitivity of IFNB1
becomes large enough for small fluctuations in the concentration of RIG-I to significantly push a cell closer to activation. As a result, many cells are activated by their internal levels of RIG-I (which is not significantly different from the average) and so the average activation of IFNB1
occurs earlier. confirms that decreasing the variability results in similar activation times for IFNB1
, supporting our claim that variability is essential in order to induce the dynamics seen in the system. We also note that the simulation with reduced variability results in significantly higher levels of interferon, which can be harmful. As such, the early responder dynamics allows an efficient response to the viral infection while maintaining controlled levels of interferon induction.
Single DC simulation with high and low cell-to-cell variation.