Stochastic Cytokine Expression Induces Mixed T Helper Cell States

During early differentiation of T helper cells, stochastic cytokine expression triggers the co-expression of antagonistic transcription factors at high levels, buffered by the interplay between extracellular and intracellular signaling components.


Introduction
A multipotent progenitor cell can differentiate into a particular lineage by turning on the expression of a lineage-specific transcription factor, which coordinates the expression of a defined set of target genes. Numerous examples of such toggle-switch-like cell fate decisions have been observed in the differentiation of hematopoietic cells [1]. For example, common myeloid progenitor cells differentiate into granulocyte-monocyte progenitor versus megakaryocyte-erythrocyte progenitor cells based on expression of PU.1 versus Gata1 [2]; naive CD4 T cells differentiate into Th1 versus Th2 driven by the expression of Tbet or Gata3 [3][4][5][6]. Antagonistic transcription factors are therefore believed to be expressed exclusively in the pertinent cell types, or co-expressed at basal levels in hematopoietic progenitors prior to commitment to ''prime'' the cells for rapid deployment of transcription factors to execute a particular lineage program [7]. For instance, common myeloid progenitors can co-express low levels of PU1 and GATA1 during lineage priming [8][9][10][11][12], though their expression is mutually exclusive in the fully committed state [7].
In addition to transcription factors that reside within the cell, the signaling network governing cell differentiation often comprises extracellular components, such as cytokines that can bind to cell surface receptors leading to activation and/or repression of transcription factors. In many previous studies, where the goal has been attaining a relatively homogenous population of differentiated cells, high concentrations of cytokines were added to the culture media to bias the cellular decision process toward one particular cell fate [2,3,6,13].
In this work, we studied gene regulation during the early stage of cell differentiation to delineate the interplay between extracellular cytokines and intracellular transcription factors in single cells, using CD4 T helper cells as a model system. Contrary to previous studies where cellular fate was biased artificially [2,3,6,13], we sought to avoid this bias by exploring the spontaneous differentiation of naive CD4 T cells in the absence of exogenously added cytokines.
Tbet, encoded by Tbx21, is the master transcription factor of Th1 differentiation associated with production of the hallmark cytokine IFNc [3], whereas Gata3 is the master transcription factor of Th2 differentiation associated with IL4 production [5]. In terminally differentiated individual CD4 T cells, the expression of Tbx21 and Gata3 is mutually exclusive [14,15]. This is attributed to positive feedback loops and cross-inhibitory interactions in the regulatory network ( Figure 1A). This network consists of two types of interactions: those that depend on cytokine signaling and those that are cytokine-independent and involve only intracellular players including transcription factors. Specifically, Tbet activates Ifng [16], and extracellular IFNc can induce Tbx21 via receptor signaling [17]. Tbet also induces itself independently of signaling via cytokine receptors [18]. Similarly, Gata3 activates Il4 [19,20] and extracellular IL4 can induce Gata3 [21]. Furthermore, Gata3 can be auto-induced independently of signaling via cytokine receptors [19,22]. Finally, Tbet silences Il4 [16], Gata3 silences Ifng [23,24], and Tbet blocks the function of Gata3 through direct proteinprotein interactions [25], leading to cross-inhibitory interactions.
To quantify the number of Tbx21 and Gata3 transcripts in activated CD4 T cells, we isolated total CD4 + cells from C57BL/6 mice. CD4 cells were then activated by culturing them in wells coated with anti-CD3 and anti-CD28 antibodies, in the absence of polarizing cytokines or neutralizing antibodies against cytokines, such that CD4 T cells would choose their cell fates without being biased. We performed single-molecule fluorescent in situ hybridization (smFISH) [26] combined with immunofluorescence to quantify transcripts and protein levels in individual cells (Figures S1 and S2).

High-Level Co-expression of Tbx21 and Gata3 in Individual Cells
Without exogenously imposed Th1-or Th2-biasing cues, naive CD4 T cells, essentially expressing zero copies of Tbx21 and Gata3 transcripts, turned on expression of both Tbx21 and Gata3 simultaneously in individual cells after activation ( Figure 1B-D). Simultaneous up-regulation of Tbx21 and Gata3 occurs very rapidly within 24 h, in contrast to co-expression of Tbx21 and Gata3 observed weeks after activation from naive cells followed by a reprogramming experiment [27]. Furthermore, distinct from basal co-expression in lineage priming [8][9][10][11][12], co-expression of Tbx21 and Gata3 are at high levels, such that the mean number of Gata3 transcripts per cell at 48 h is comparable to fully differentiated Th2 cells [28].
To further assess the expression levels of Tbx21 and Gata3 that we observed under non-biased condition, we compared to cells that were treated under standard polarizing conditions with supplements of IFNc and IL4 as well as neutralizing antibodies against opposing cytokines as previously described [16]. While polarized cells express Tbx21 and Gata3 in a mutually exclusive manner, we found that the expression of the up-regulated transcription factor is comparable to the high-level co-expression in cells under non-biased condition ( Figure 1E,F), indicating that that the cells under non-biased condition produce IFNc and IL4 by themselves. Furthermore, supplementing both IFNc and IL4 into the non-biased cell culture did not increase the expression of Tbx21 and Gata3 further (unpublished data), indicating that the amount of IFNc and IL4 that CD4 T cells produce has already reached saturation for signaling. In addition, high-level co-expression of Tbx21 and Gata3 in individual cells is a robust phenomenon observed over a large range of seeding cell density ( Figure S3). Interestingly, the median stoichiometry between Tbx21 and Gata3 expression was 1:1 until 24 h after activation, but Gata3 levels continued to increase after 24 h while Tbx21 levels decreased ( Figure S4). As activation time increases, the culture system presumably accumulates more Th2favoring cytokines. Since most of the significant changes in gene expression occurred within this 48 h period, we focused our analyses on this period in subsequent experiments.
To demonstrate that transcript counts serve as a good proxy for protein levels, we performed immunofluorescence against Tbet or Gata3 simultaneously with smFISH. Transcript counts and protein levels showed strong correlations in individual CD4 T cells, with a Pearson's correlation coefficient R of 0.59 (p,10 244 ) for Tbet and 0.85 (p,10 284 ) for Gata3 ( Figure 2). In addition, translational efficiency, measured by the ratio of immunofluorescence intensity over transcript count, remained constant as a function of activation time ( Figure S5).
It is interesting to note that the expression of two other transcription factors Foxp3 and RORcT, which control the other two lineages of CD4 T helper cells [induced regulatory T helper cells (iTreg) and Th17], are not excluded from cells that express Tbx21 and Gata3 ( Figure S6). We have shown that there is no strong positive or negative correlation between these four transcription factors ( Figure S6). Therefore, during early stage of CD4 T helper cell differentiation, lack of mutual exclusivity does not only apply to Tbet and Gata3, but also to Foxp3 and RORcT.

A Rare Cell Population Expresses the Cytokine Genes
We postulated that ubiquitous Tbx21 and Gata3 co-expression must be associated with both Th1 and Th2 cytokines produced by CD4 T cells upon activation [29], since no cytokines were supplied

Author Summary
During the development of a multicellular organism, the progenitor cells, which have the potential to become any of several different cell lineages with specialized functions, commit and differentiate into one particular lineage. This differentiation of progenitors is driven by the induction of lineage-specific transcription factors, molecules that regulate gene expression. This process is often mediated by extracellular signaling molecules, including a class of molecules called cytokines that can bind to cell surface receptors, activating and/or repressing transcription factors. Here we explored the early differentiation of naive T helper (Th) cells, an important class of T lymphocytes that help effector immune cells to defend the body against various pathogens. We measured both mRNA and protein levels of cytokines and transcription factors in individual cells. In particular, mRNA levels were measured with singlemolecule resolution. Contrary to the expression of only one set of lineage-specific transcription factors, we observed ubiquitous high-level co-expression of antagonistic transcription factors in individual cells. We found that cytokines are expressed only in a small subpopulation of cells, independent from the expression of transcription factors in individual cells. When cytokine signaling is inhibited, each cell expressed only one of the antagonistic transcription factors at high levels. This reveals a weak intracellular network that is otherwise overruled by the strong signals that emanate from extracellular cytokines. These results suggest that during the early differentiation process T helper cells acquire a mixed Th1/Th2 state, instructed by extracellular cytokines. The interplay between extracellular and intracellular signaling components unveiled in Th1/Th2 differentiation may be a common strategy for mammalian cells to buffer against noisy cytokine expression.
exogenously. We thus investigated the expression of cytokines in individual CD4 T cells. Among several Th1 and Th2 cytokines, we focused our effort on Ifng and Il4. In activated Ifng 2/2 mutant cells, we observed a 95% reduction in Tbx21 expression (mean of 33 in wild-type versus 1.6 in Ifng 2/2 mutant at 24 h) and a 93% reduction in Gata3 expression in activated Il4 2/2 mutant cells (mean of 35 in wild-type versus 2.3 in Il4 2/2 mutant at 24 h) ( Figure S7) [30,31]. Our results indicate that the hallmark cytokines Ifng or Il4 alone is responsible for achieving more than 90% of the expression of Tbx21 or Gata3. Therefore, any other cytokines minimally contribute to the expression of Tbx21 or Gata3.
Current understanding of the gene regulatory network that governs Th1/Th2 differentiation would predict that Ifng or Il4 transcripts are proportional to Tbx21 or Gata3 levels in individual cells. Surprisingly, we observed that Ifng and Il4 were expressed only in a rare cell population. While the vast majority of cells were in the OFF state and contained essentially zero copies of Ifng or Il4 transcripts, the rare ON cells expressed up to more than 1,000 transcripts, resulting in a dynamic expression range of three orders of magnitude ( Figure 3A,B, Figure S8). In cells expressing more than 200 transcripts, we could not resolve individual mRNA molecules. Instead, we extrapolated the number of transcripts from the linear relationship between the total fluorescence and number of transcripts using cells with fewer transcripts ( Figure S9). Complementary to previous studies reporting that the expression of cytokines in fully differentiated T helper cells is stochastic [32,33], our work focused on the early phase of T helper cell activation, and showed that the cell-to-cell variation in the cytokine expression is significantly larger during the early phase than that in the fully differentiated state.
In addition, there was only a weak positive correlation between Tbx21 and Ifng expression (R = 0.15, p,10 26 ), or between Gata3 and Il4 expression (R = 0.32, p,10 211 ) ( Figure 3C). There was no negative correlation between Gata3 and Ifng expression (R = 0.06, p = 0.04), or between Tbx21 and Il4 expression (R = 0.26, p,10 29 ) ( Figure S10). In addition, we performed three-color smFISH experiments probing Ifng or Il4 transcripts, while simultaneously detecting Tbx21 and Gata3. Our data showed that the distributions of Tbx21 and Gata3 expression, conditioned on the number of Ifng or Il4 transcripts, are invariant, indicating that Tbx21 and Gata3 expression do not correlate with Ifng or Il4 expression ( Figure S12). Therefore, the regulation of Ifng and Il4 expression appears to be decoupled from the expression of Tbx21 or Gata3 levels in individual cells during early differentiation. It is interesting to note that Ifng and Il4 are not the only cytokines showing such expression patterns, other cytokines such as Il13 are also expressed in a rare population with large variability in the expression level amongst individual cells ( Figure S11).
We further examined the cytokine milieu surrounding the cells by quantifying the distance of each cell from its nearest cytokine producing neighbor. By plotting the cellular Tbx21 and Gata3 levels against the distance, we have found that there is no correlation between the two variables ( Figure 3A, Figure S13), indicating that diffusion of cytokine is not rate-limiting and the cytokine milieu is well-mixed.
We also examined the correlation between the transcript and protein levels of cytokines. We found that cells that expressed a high number of cytokine transcripts also contained high levels of cytokine protein as detected by immunofluorescence. Transcriptionally inactive cells did not contain measurable levels of cytokine protein ( Figure S14).
To ensure that the rare cytokine-producing cells were not nonnaive CD4 + T cells, such as Natural Killer T (NKT) cells, we analyzed the expression of Klrb1c, which encodes the NKT cell marker NK1.1, and did not observe any NK1.1-expressing cells ( Figure S15).
To ensure that the CD4 T cells we isolated did not contain effector memory cells, we compared naive CD4 T cells isolated by CD4 + positive selection with that isolated by negative selection using the mouse naive CD4 T cell isolation kit (Miltenyi). We first measured the CD44 protein levels in the isolated cells using immunofluorescence. We found that CD44 levels were low in naive T cells isolated by CD4 + positive selection, comparable to that isolated by the negative selection kit ( Figure S15). We then cultured the naive T cells isolated by the negative selection kit, and found that there was no significant positive correlation between cytokine expression and CD44 levels in activated cells ( Figure  S16). In addition, the distribution of cytokine expression of the naive T cells isolated by the negative selection kit was identical to that selected by CD4 + positive selection ( Figure S8). These results showed that rare cytokine-producing cells are not effector memory cells, but originated from bona fide naive cells.
Taken together, we conclude that a rare naive CD4 T cell population stochastically turns on Ifng or Il4 independently of Tbet or Gata3 levels. These rare cells secrete cytokines into their surroundings and instruct other cells to ubiquitously express Tbx21 and Gata3, and may thus play a pioneer role in determining the differentiation outcome of the entire cell population.
To test our hypothesis about the instructive role of cytokines on the expression of transcription factors, we analyzed their expression over the time course. While naive CD4 T cells contain essentially zero copies of cytokine transcripts, the fraction of Ifngexpressing cells increased from 0 to 16 h and decreased moderately afterwards, whereas the fraction of IL4-producing cells increased monotonously ( Figure 3D). This pattern is consistent with the trend of the mean Tbx21 and Gata3 counts per cell ( Figure 2C). In addition, it is worth noting that at the population level, the fraction of cells transcribing Ifng or Il4 still positively correlates with the mean levels of Tbx21 or Gata3 transcripts, respectively, over time (correlation coefficient R = 0.35, p = 0.044 between Tbx21 and Ifng, correlation coefficient R = 0.98, p,1.6610 244 between Gata3 and Il4). Therefore, these results indicate that cytokines are instructive on the expression of Tbx21 and Gata3.

Strength of Cytokine Signaling Predominates over Other Interactions
We revisited the signaling network governing Th1/Th2 choice during early CD4 T cell differentiation. Given that Ifng is transcribed in a rare pioneer cell population independently of Tbx21 and Gata3 levels, induction of Ifng by Tbet and repression by Gata3 can be neglected during the early stages of CD4 T cell differentiation, and a similar situation applies to Il4. Since Tbx21 and Gata3 are expressed simultaneously in individual cells without mutual exclusion, we postulated that the strength of receptor signaling mediated by cytokines must dominate over the intracellular network, namely self-activation and mutual inhibition of the transcription factors, which alone would lead to mutually exclusive expression of Tbx21 and Gata3.
To demonstrate that the strength of cytokine signaling is dominant, we manipulated the amount of cytokine molecules available to cells by adding neutralizing antibodies. In the presence of saturating amounts of anti-IFNc and anti-IL4, our data recapitulated the expression patterns of Tbx21 and Gata3 in Ifng 2/2 or Il4 2/2 cells, respectively ( Figure S7), strongly indicating that this depletion strategy was specific ( Figure S17). We postulated that any other cytokines minimally contribute to the expression of Tbx21 or Gata3, and indeed our result showed that adding the antibody against another Th1 cytokine IL12 has no effect on Tbx21 or Gata3 expression ( Figure S18).
Our results showed that sequestering a cytokine by the neutralizing antibody down-regulates the corresponding transcription factor but does not up-regulate the opposing transcription factor ( Figure 1C, Figure 4A, Figure S17). For instance, at 24 h, Tbx21 is down-regulated by 95% from a mean of 33 in wild-type to 1.6 in Ifng 2/2 mutant cells, but Gata3 changes from a mean of 35 to 40 ( Figure S7), which is a statistically insignificant change. Therefore, the role of extracellular cytokine signaling in specifying lineage choice is to up-regulate the corresponding transcription factor, rather than to repress that of the alternate lineage. In addition, the contribution of IFNc to Tbx21 expression is predominant, overshadowing the signaling strength of the intracellular components, namely self-activation and mutual inhibition of the transcription factors.

A Polar-Coordinate Representation of the Data
Our results also show that down-regulation of the corresponding transcription factor could be modulated depending on the amount of neutralizing antibody ( Figure 4A). To facilitate a quantitative interpretation of our data, we converted the Tbx21-Gata3 scatter plot into polar coordinates of (r,h), where r is the distance from the origin, which equals ffiffiffiffiffiffiffiffiffiffiffiffiffi t 2 zg 2 p , where t represents Tbx21 and g represents Gata3 transcript count. h is the angle with the x-axis, which equals ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi ffi tan {1 g=t ð Þ p in the range 0#h#p/2 (0u#h#90u). Therefore, a small h close to 0 means that a cell is Th1-like, and a large h close to p=2 means Th2-like ( Figure 4B). We excluded cells with r,10 (shaded region) in our analysis, because h is not robust against small fluctuations in the number of transcripts in these cells. We then computed the distribution of h, which gives us a measure on the cell fate bias. For instance, a uniform distribution of h would mean that the cells can explore any intermediate state between Th1 and Th2 with no bias to a particular cell fate; whereas a ''U''-shape distribution with probability density concentrat near 0 and p/2 would mean that the cells exhibit mutual exclusion by expressing a high level of only Tbx21 or Gata3.
To illustrate this polar-coordinate representation, we converted our data at 24 h under the non-biased condition and showed that h follows an approximately uniform distribution, a hallmark of lacking mutual exclusion ( Figure 4C, Figure S19). In other words, under non-biasing condition, CD4 T cells during early differentiation occupy any intermediate cell states between Th1 and Th2 with equal probability. We also showed that as the concentration of anti-IFNc increases, the distribution of h shifts toward p/2 (more Th2-like), whereas when the concentration of anti-IL4 increases, the distribution of h shifts toward 0 (more Th1-like) ( Figure 4D, Figures S20 and S21).

A Model Governing the Early Activation of Naive CD4 T Cells
Taking the data together, we can explain the ubiquitous coexpression of Tbx21 and Gata3 under the non-biased condition: when CD4 T cells are exposed to both IFNc and IL4 secreted by the rare cytokine producing cells, they up-regulate both Tbx21 and Gata3. The key to the absence of mutually exclusive expression of Tbx21 and Gata3 is that cytokine signaling must predominate over self-activation of Tbet and Gata3 as well as mutual repression between Tbet and Gata3. This suggests that expression of Tbx21 and Gata3 is maintained at high levels by extracellular cytokine cues, with comparatively minimal effects from the intracellular signaling components ( Figure 5A). Our data show that diffusion of cytokines is not rate-limiting ( Figure S13), and we therefore propose that during early activation, CD4 T cells are bathed in a cocktail of well-mixed cytokine molecules produced by the rare pioneer cells, thus simultaneously inducing the expression of Tbx21 and Gata3 in individual CD4 T cells ubiquitously ( Figure 5B). According to our model, we hypothesized that by eliminating the extracellular IFNc and IL4, only the intracellular signaling components will remain intact and should result in mutually exclusive expression of Tbx21 and Gata3 in individual cells. To verify this, we added both anti-IFNc and anti-IL4. We tested multiple combinations of different concentrations of anti-IFNc and anti-IL4 antibodies to find an optimum where the median of h was close to p/4 (exactly in the middle of Th1 and Th2). We observed that under such conditions, Tbx21 and Gata3, in addition to being down-regulated, are expressed in a mutually exclusive manner, such that the majority of cells are situated near the Tbx21 or Gata3 axis on the scatter plot and the distribution of h falls into a ''U''-shape distribution with higher density near 0 and p/2 ( Figure 5C-D). We further explored the robustness of this phenomenon by culturing CD4 T cells under non-biased condition for 24 h first, allowing cells to first establish high-level coexpression of Tbx21 and Gata3, and then added both anti-IFNc and anti-IL4. Strikingly, the cells were still able to adopt mutually exclusive expression of Tbx21 and Gata3 ( Figure S22). We therefore conclude that under simultaneous IFNc and IL4 deprivation, only the comparatively weak intracellular signaling components that consist of the self-activation of Tbet and Gata3 as well as their mutual repression are functional, leading to mutually exclusive expression of Tbx21 and Gata3.

Discussion
Using CD4 T cells as a model of cell differentiation, we observed ubiquitous high-level co-expression of antagonistic transcription factors during the early stages of CD4 T cell differentiation under non-biased condition. CD4 T cells appear to produce a sufficient amount of IFNc and IL4 for their own activation, such that Tbx21 and Gata3 are co-expressed at high levels. Strikingly, activation and cross-inhibition of Ifng and Il4 expression appear to be decoupled from Tbx21 and Gata3 levels in individual cells ( Figure 5A). Instead, Ifng and Il4 are expressed by a rare population originated from bona fide naive cells, which do not appear to be contaminating NKT or memory CD4 T cells. We therefore postulate that these naive CD4 T cells stochastically turned on expression of Ifng or Il4 and translate protein molecules ahead of the bulk population. These cytokine-producing cells, though rare, can direct the entire cell population into assuming one particular cell fate. By manipulating the amount of cytokine available to the cells, we demonstrated that signaling strength evoked by extracellular cytokines dominates over intracellular signaling components of self-activation and mutual inhibition. Therefore, put in the perspective of the debate on instruction versus selection model of immune cell differentiation [34], our results show that the role of extracellular cytokines is to instruct cells to up-regulate transcription factors during early stage of CD4 T cell differentiation. When IFNc and IL4 are sequestered from the cells, only the intracellular signaling component is intact and hence the expression of Tbx21 and Gata3 is mutually exclusive.

Stochastic Cytokine Expression in the Early Phase of T Help Cell Activation
The large variability of cytokine expression between individual cells is very striking. Previous studies on unicellular organisms such as genetically identical populations of bacteria and yeast have observed lower variations in gene expression [35][36][37][38]. We were thus intrigued by the different mechanism of gene regulation in mammalian cells, where complex chromatin modeling that is not available to unicellular organisms may play an important role. Using a two-state model of transcription, where a gene needs to transition from an inactive to an active state before transcription can occur, the steady state of transcript density can be approximated to a Gamma distribution under the limiting case of short but infrequent bursts of mRNA synthesis [39]. We therefore fitted cells expressing cytokine transcripts to Gamma distributions. From the fitting, we can deduce that the average transcriptional burst size of Ifng is 159 and that of Il4 is 176 ( Figure 6A-B). Compared to the transcription factors, which have average burst size of 18 transcripts for Tbx21 and 36 transcripts for Gata3 ( Figure 6C-D), the transcriptional bursts are much larger for cytokine genes, which are comparable to the very bursty 76-tetO construct in the work of Raj et al. [39].
As cytokine molecules are produced and secreted to the cell culture media, a uniform cytokine milieu is established because diffusion of cytokine molecules is not rate-limiting ( Figure S13), leading to up-regulation of transcription factors ubiquitously. It is interesting to note that although the production of cytokine molecules is highly heterogeneous amongst cells, the expression of transcription factors as a read-out is less variable because of averaging effect from mixing cytokine in the extracellular environment. The interplay between extracellular cytokines and intracellular transcription factors may be a common strategy for mammalian cells to buffer transcriptional noise that is otherwise intrinsic to the cells.
While cytokine expression appears to be decoupled from transcription factors in individual cells, we wondered how cytokine expression is regulated at the population level-for instance, how a cell population turns on IFNc but not IL4 when cultured under Th1-favoring conditions with supplement of antibody against IL4. We quantified the expression of Ifng in cells treated with anti-IFNc and the expression of Il4 in cells treated with anti-IL4. We found that the number of cytokine-expressing cells and hence the mean of cytokine transcripts decreased when neutralizing antibody is added to the cell culture ( Figure S23). Therefore, when cytokines are sequestered, not only the respective transcription factor gets down-regulated, but the expression of the cytokine itself is also down-regulated.
This observation suggests that although the expression of a cytokine is not positively correlated with the expression of its respective transcription factor in individual cells, the expression of cytokine in the entire cell population is still in concert with the expression level of the transcription factor. We postulate that transcription factor may be largely responsible for de-condensing the cytokine locus during early activation of CD4 T cells. While switching an inactive gene to the active state is a stochastic process in individual cells, the average of gene activation events is still deterministically controlled by the amount of transcription factors. As cell differentiation progresses, the activation of cytokine genes eventually becomes more ubiquitous and depends on the local concentration of active transcription factors, leading to higher positive correlation between a cytokine gene and the respective transcription factor in fully differentiated cells [27,33,40,41].
In the light of our work, it will be interesting to delineate the underlying molecular mechanisms governing cytokine gene expression. In addition, given sufficient technological advances, it will be interesting to perform time-lapse experiments to track stochastic cytokine expression in individual cells over a time-course to visualize how these rare cytokine-producing cells arise and evolve over time. It will also be helpful to study the single-cell transcriptome of these cells to quantify how different these cells are from other cells. Insights from such experiments will shed light on the interplay between extracellular cytokines and the intracellular transcription factor on the fate specification of single cells. We note that mixed Th1-Th2 phenotypes were also observed concurrently by two other groups, using different experimental approaches [42,43].

Strains of Mice Used
Experiments on wild-type cells were from C57BL/6 mice; experiments on Il4 2/2 cells were from B6.129P2-Il4 tm1Cgn /J mice; experiments on Ifng 2/2 cells were from B6.129S7-Ifng tm1Ts /J mice. C57BL/6, Ifng 2/2 , and Il4 2/2 mice were obtained from Jackson labs. All animals were housed at the Whitehead Institute for Biomedical Research and were maintained according to guidelines approved by the Massachusetts Institute of Technology (MIT) Committee on Animal Care.

Cell Culture
Spleens and lymph nodes of mice aged from 6 wk to 2 mo were removed, suspended in PBS supplement with 2% FCS, and gently homogenized through a nylon mesh. Red blood cells were lysed with ammonium chloride solution (StemCell Technologies). CD4 + cells were isolated by MACS purification using the CD4 microbeads (Miltenyi Biotec) in all experiments except those that explicitly mentioned negative selection. In experiments where cells were selected by depletion, MACS CD4 + T cell isolation kit II was used. The medium used throughout the cell cultures was RPMI supplemented with 10% FCS, 2 mM L-glutamine, 1% penicillin, and streptomycin.
Cells were seeded into eight-well Lab-tek 1.0 coverglass chambers that had been coated with a mixture of anti-CD3 (15 mg/ml, clone 17A2) and anti-CD28 (15 mg/ml, clone 37.51) antibodies for at least 3 h, at 1,000,000 cells per well in a total volume of 0.5 ml, except one control experiment that explicitly mentioned 250,000 cells per well. The following neutralizing antibodies were used: IFNc antibody (clone R4-6A2), IL4 antibody (clone BVD4-1D11), and IL12 antibody (clone C17.8). Cells were cultured at 37uC, 5% CO 2 . The first refresh of culture media occurred at 48 h, after which media was refreshed every 24 h. In experiments with Th1 polarization, 10 ng/ml IFNc and IL12 and 10 mg/ml anti-IL4 antibody were supplemented in the media; in experiments with Th2 polarization, 10 ng/ml IL4 and 10 mg/ml anti-IFNc antibody were supplemented in the media.

Single-Molecule Fluorescence In Situ Hybridization (smFISH)
We performed smFISH on the T cells and counted the mRNAs in individual cells as described previously [26,28]. Harvested T cells were fixed in PBS buffer with 3.7% formaldehyde for 10 min. After fixation, the cells were washed twice with PBS, permeabilized in 70% ethanol for at least 2 h, and stored at 4uC. The T cells were hybridized in the same glass chamber as cell culture. After the 70% ethanol was aspirated, the samples were washed in a solution of 25% formamide and 26SSC for 5 min. After the wash buffer was aspirated, 100 ml of hybridization solution containing labeled DNA probes in 26SSC, 1 mg/ml BSA, 10 mM VRC, 0.5 mg/ml Escherichia coli tRNA, and 0.1 g/ml dextran sulfate, with 25% formamide, were added to the sample and incubated overnight at 30uC. The next day, the samples were washed twice by adding 1 ml of wash solution consisting of 25% formamide and 26SSC. For each wash, the sample was incubated in wash solution for 30 min. Then, the sample was resuspended in 26SSC buffer. The sequences of FISH probes are available upon request.

Immunofluorescence
To simultaneously visualize mRNA and protein levels in cells, we performed immunofluorescence after FISH protocol. The cells were incubated with 26SSC, 0.2% Triton X-100, 5 mg/ml BSA, and fluorescent antibodies for 3 h at 4uC. Where a secondary antibody is required, the samples were incubated with 26SSC, 0.2% triton X-100, 5 mg/ml BSA, and the secondary antibody for 1 h at 4uC. The cells were then washed by incubating with 26SSC, 0.2% triton X-100, 5 mg/ml BSA for 1 h at 4uC. Tbet antibody is clone 4B10; Gata3 antibody is clone L50-823; IFNc antibody is polyclonal (AMC4034, Invitrogen) and a secondary goat-anti-rabbit antibody (A11034, Invitrogen) is used. We test multiple IL4 antibodies for immunofluorescence, but none of them gave satisfactory signal-to-noise ratio.

Image Acquisition
For imaging, the samples were soaked in glucose oxidase (glox) anti-fade solution, which contains 10 mM Tris (pH 7.5), 26SSC, 0.4% glucose, supplemented with glucose oxidase and catalase. A coverslip was put over the sample. All images were taken with a Nikon Ti-E inverted fluorescence microscope equipped with a 1006oil-immersion objective and a Photometrics Pixis 1024 CCD camera using MetaMorph software (Molecular Devices, Downington, PA). Stacks of images were taken automatically with 0.4 microns between the z-slices.

Image Analysis
To segment the T cells, a marker-guided watershed algorithm was used. Briefly, cell boundaries were obtained by running an edge detection algorithm on the bright-field image of the cells. To generate markers for watershed algorithm, the centroid of the region enclosed by individual cell boundaries is computed. A marker-guided watershed algorithm is then run on the distance transformation of the cell boundaries, using the markers located within the cell boundaries. The resultant cell segmentation image is then manually curated for occasional mis-segmentations.
To quantify the number of RNA molecules in each cell, a log filter is run over each optical slice of the image stack to enhance signals. A threshold is taken on the resultant image stack to pick up mRNA spots. The locations of mRNA spots are then taken to be the regional maximum pixel value of each connected region. The number of mRNA spots located within the cell boundaries of an individual cell can thus be quantified.
To quantify fluorescence signal in each cell, an optical slice corresponding to the central plane of the cells is analyzed. For each image, which covers up to 100 correctly segmented cells, the mean fluorescence per pixel of each cell is computed. The minimum of mean fluorescence is taken to be the background. Then for each cell in the image, the total fluorescence of the cell is computed as the sum of the fluorescence at each pixel subtracting the background. If this value is negative, zero is used instead. shows the probability density of cells expressing Il4 transcripts. We have shown that cultures of cells selected by negative selection also give rise to rare cells that stochastically express Ifng and Il4 at high levels. Therefore, rare cytokine-expressing cells observed in Figure 3A