Nature Editor's Overview

Within a clonal population, heterogeneous gene expression patterns prime multipotent cells toward different fates. Here, each green marble represents a single cell within a “cloud” of temporary cellular states. Outliers are more inclined to commit to different lineages. Credit: Sui Huang

Even in clonal populations, individual cells vary substantially. While studying the behavior of outliers in populations of stem cells, Huang and colleagues found a surprising result: random noise seems coordinated1.

Cell-to-cell variation has generally been attributed to noise inherent in mechanisms of individual gene expression. To test this idea, Huang and colleagues analyzed the behavior of cells expressing very high levels of the stem cell marker Sca-1. They found that the outliers possessed strikingly distinct transcriptomes. Outliers could be driven to express characteristics of distinct cell fates, for example, myeloid versus erythroid. Lineage decisions could be predicted via levels of certain transcription factors, such as GATA1 or PU.1. However, if not driven into these states, gene expression within these outlier cells reverted to mean expression levels; these outlier cells were primed, not committed, and remained multipotent.

The authors concluded that clonal heterogeneity of gene expression is not due to independent noise in the expression of individual genes, but rather is a manifestation of metastable states of a slowly fluctuating transcriptome. These fluctuations may govern reversible, stochastic priming of multipotent progenitors in cell-fate decisions.

Read below to see a panel of experts' comments (in black), and responses from the authors (in italics). Table and reference numbers refer to those in the research article.

Chang, H. H. Transcriptome-wide noise controls lineage choice in mammalian progenitor cells. Nature 453, 544-547 (2008).

Reviewer 1: Expert in biological networks and circuits

The authors present a novel insight about differentiation of stem cells: cell fate is correlated with stochastic transitions between cell subpopulations, transitions that last for many days and result in marked differences in gene expression and eventual fate. The paper shows that a multipotent murine hematopoietic cell line expressing stem-cell marker Sca-1 displays an irregular distribution of expression of Sca-1 that spans over a range of 1000 fold among individual cells. This wide distribution is independent of experimental noise, cell size and cell cycle stage. Clones that arise from any position in the distribution eventually recreate the parental distribution. Causes for this phenomenon, such as, contamination, genetic mutation, imperfect partitioning at cell division or fast random fluctuations are controlled for.

The authors present a model where the population is divided into two subpopulations where each subpopulation is a distinct functional state. Slow stochastic transitions between these two states enable the regeneration of the parental distribution from any cell of the population. The authors show that clonal heterogeneity in Sca-1 is correlated with the differentiation potential of cells and that the difference between cells at the two ends of the distribution is not only in Sca-1 expression but in a whole subset of genes which includes GATA1 – encoding an erythroid-fate determining transcription factor. The authors also find a subset of genes whose expression in cells cultured in the presence of Epo (erythropoietin) is most similar to expression in cells with low levels of Sca-1. Finally, the authors show that a clone generated from cells with low Sca-1 expression, relaxes to the parental distribution in all three aspects: Sca-1 expression, expression profile of the genome and the relative differentiation rate.

The paper is novel in showing that differentiation potential is governed by stochastic transitions within the population. It carries an important message that biological function is not always determined by the ensemble average of a cell population and that outliers that departure from the average state in a slow stochastic process are the ones carrying the significant biological functionality in the priming of cell-fate commitment.

In summary, we believe that this paper reports a novel and important biological mechanism for differentiation: it gives evidence that slow stochastic variations in stem cell state last for a long time and result in different fates. It relates the recent finding of slow fluctuations in human protein levels to the biological outcome of cell fate, and finds long lasting differences in transcriptomes in different subpopulations. It is an excellent choice for Nature, provided that the comments are addressed. (Specific comments follow general author response)

Author response: Importantly, while performing the requested experiments, we discovered that transcriptome-wide noise does more than simply control the commitment of precursor cells to differentiate along the erythroid lineage: It also controls commitment to the myeloid lineage. Thus, clonal heterogeniety not only affects the propensity to differentiate, it also governs the actual choice between alternative cell fates in a multipotent progenitor cell population. These new findings not only corroborate our previous results by reproducing our observations in another system, but also extend our conclusions: They provide a new explanation for fate-indeterminacy in a multipotential progenitor cell, and support the notion of multipotency being dependent on “multilineage priming”. We have therefore added these new data to the revised manuscript, which further strengthens our story.

Comment 1: The way that Sca-1-low cells were selected by FACS in the paper includes the tail of the bell shaped high-Sca-1 distribution (Fig2a). This leaves the possibility that this tail is responsible for the regeneration of the parental distribution. To be convincing, cells should be sorted from the extreme low Sca-1 region, and then shown to return to the original distribution.

The right-hand tail of the Sca-1_low fraction does, as the reviewer points out, participate in the “regeneration of the parental distribution” at the population-level. But the reviewer raises an interesting question: do more extreme outliers also have the ability to regenerate the parental distribution? Prompted by this possibility, we conducted an experiment where the extreme top and bottom 2% outliers in Sca-1 expression (as opposed to 15% as in Fig. 2) were isolated and allowed to relax back to the parental population. As expected, the return of these “extreme outliers” to the original distribution indeed occurred, but over an even longer time span than when less extreme outlier were sorted (>17 days compared to ∼9 days in our original results in Fig. 2). This finding is consistent with the notion that the outlier cells represent persistent states that start from a position further away from the mean on the Sca-1- axis. More importantly, it demonstrates that extreme outliers among a clonal population are still capable of regenerating the parental population. We now describe these new results in the main text and Supplementary Information Figure S7 with additional explanation in its legend.

We would like to clarify a more subtle point. We designed our original experiment (sorting of 15% outlier cells) to demonstrate the existence of long persistence of an outlier and its slow relaxation to the parental distribution. For this purpose, it is more convincing if transient persistence of outliers is detectable even among a less extreme group of outliers (i.e., close to the center of the parental population). In other words: the less extreme group of outliers we sort, the less we push towards the “desired results”. In fact, even among the 15% outlier cells, clonal individuality is long-lasting and capable of affecting cell lineage decisions, as we demonstrate.

Comment 2: Figure 4b & relevant text regarding microarray analysis - The authors show subsets of genes that are up/down regulated in Sca-1 -low in a similar pattern to Epo-7. What is the significance of this result? Of these specific genes? The following control must be done- Does randomly pooling from a group of genes will also generate a subgroup of genes with similar characteristics?

We presented the genes in Fig.4b to conform to the conventional practice of representing micro-array data with hierarchical clustering and a listing of specific genes. The goal was to strengthen our biological finding that Sca-1_low cells are naturally primed to differentiate into erythrocytes by providing a few specific gene examples. We agree that we should have qualified this statement with further analysis of the microarray data. However, “random pooling” (bootstrapping) would not be an appropriate method to use in cluster analysis, as although it can estimate the robustness of calculated aggregate quantities (see below), it would simply reduce the number of genes being clustered in this case.

To respond to this reasonable concern, we defined a subset of ∼ 200 “differentiation marker genes” based on stringent SAM-analysis (see Material and Methods in the Supplementary Information) of the unsorted, untreated control (Untreated) and the unsorted, Epo-treated sample (Epo_7d) whose GEDI maps are shown in Fig.4a. Then, we compared the sorted Sca-1_low, Sca-1_Mid, Sca-1_High fractions and the Epo_7d sample using only these “differentiation marker genes”. We found that the Sca-1_Low cells were statistically similar to the unsorted, Epo-treated cells (p < 3×10-14; pair-wise t-test), whereas the Sca-1_High (p > 0.6) and Sca-1_Mid (p > 0.8) cells were not. This new analysis confirms the original point made in Fig. 4; these results are now summarized in the main text. In addition, inspired by the Reviewer's suggestion, we conducted a stringent bootstrap analysis to estimate the error in all of our calculated dissimilarity metric. In brief, 30% of the genes were randomly selected to calculate the pair-wise dissimilarity metric, and repeated for 10,000 runs to generate the standard deviation. These results are now presented in Table S1 of the Supplementary Information.

Comment 3: The authors show that at day 14, although Sca-1 expression has converged from Sca-1-low mid and high to the parental distribution, there remains a difference regarding the relative differentiation rate between Sca-1 low and the others (mid &high) which the authors claim to be connected somehow to GATA1 expression. The authors should show how GATA1 expression changes with time in clones generated from Sca-1-low,mid and high subpopulations as they did for Sca-1.

As requested, we performed RT-PCR analysis to obtain the GATA1 mRNA levels in the Sca-1_Low, Mid, and High fractions at 5 and 14 days after isolation. We found a gradual loss in the difference in GATA1 expression among the three sorted fractions. These results recapitulate the qualitative trend in differentiation kinetics data presented in Fig. 3b, supporting the known role of GATA1 in differentiation. These data are now presented in Supplementary Fig. S8.

We want to clarify, however, that we do not mean to imply that GATA1 alone is responsible for the entire trend in differentiation kinetics shown in Fig. 3b. As supported by the transcriptome-wide micro-array data (Fig. 4a), the decrease in propensity for erythroid-differentiation among the Sca-1_Low cells is associated with changes over many genes that return to the parental gene expression pattern during the relaxation process. Our original text might have wrongly implied a direct causative relationship between GATA1 and Sca-1. Thus, in addition to repeating the GATA1 RT-PCR time course with more time points as requested (see above), we have made appropriate changes to the text to clarify this point.

Reviewer 2: Expert in developmental haematology

Chang and colleagues have identified differences in Sca-1 expression contained within clonally-derived cell populations obtained from a multipotent murine hematopoietic cell line. Specifically, they have obtained 3 populations of cells based on Sca-1 expression (Low, Mid, and High). Growth of each of these populations in cell culture inevitably leads to reacquisition of the same spectrum of Sca-1 expression in cells as they are grown in culture, recapitulating the Sca-1 profiles seen in the parental line over time. The authors suggest that the dynamics of Sca-1 expression is due to a multi-stable system. Next, the authors show that this heterogeneity of Sca-1 expression is associated with differences in the ability to differentiate when treated with Epo and each population expresses varied transcriptional profiles including GATA1 expression. In fact, the Sca-1-Low population differentiates fastest and expresses high levels of GATA1 at the RNA and protein level. This Sca-1 low population still retains cKit expression, leading the authors to suggest that in fact, these cells are still stem cells and not committed progenitors that could be further down the differentiation path. Finally, detailed array analysis is performed to show that as Low, Mid, and High Sca-1+ cells are cultured, they initially have different transcriptomes, but over time revert back to that of the parental cell line. Mechanisms describing the cell-individuality are not detailed.

Although the phenomenon described is immensely interesting and the idea of heterogeneity being retained within even clonal populations of cells is plausible, the authors merely describe this phenomenon and in some instances, do not provide conclusive data to support their interpretation. If a mechanism was determined this would definitely aid in its novelty and interest. In its current form, I believe this manuscript should not be accepted at Nature.

Comment 1: FACS analysis was not performed in the presence of PI to exclude dying cells. This is a major issue for the first portion of their paper where they argue for the existence of a multi-modal derivation of Sca-1+ cell types. In fact, if one looks at figure 2, the low Sca-1 cells “emerge” in the Mid and High cells after longer times in culture, which could merely represent that cells are initially alive, but die off in culture. By contrast, Low cells have these dead cells and thus are seen even after 9 hours of culture. This multi-modal evolution may merely reflect live vs. dead cells. No mention of cell death is discussed, however the authors make a major point that heterogeneity of Sca-1 expression cannot be due to cell size or cell cycle. PI, 7-AAD, and AnnexinV staining should accompany this analysis.

Dead cells were excluded from all analyses by gating of flow cytometric data based on Propidium Iodide (PI) staining and Forward (FSC) and Side Scatter (SSC). As shown in the Supplementary Information Fig. S10, live and dead EML cells are clearly differentiable by both PI and FSC/SSC, and thus easily removed from the analysis. In contrast, cell cycle status and cell size have been linked to continuous variation of expression levels of other proteins and thus needed to be explicitly excluded as source of the heterogeneity of Sca-1 expression levels in the text. We now include a description of the routine gating of dead cells in the Materials and Methods section of the Supplementary Information.

Comment 2: Proliferation assays performed in Figure S2 are confusing. (2a) It is hard to interpret the data because the data is not presented in the typical manner to show grow curves or kinetics. This makes it hard to appreciate some arguments (i.e. p. 4 first sentence). One would assume that by 23 hours all cells would have divided once. Also, data is not replicated with error bars. This is a major issue, because differential partitioning of ScaI could account for the drift if cells undergo even one cell division. (2b) The statement that cells have not divided by 23 hours is not supported and is not likely true because the authors state that 12 cell divisions occur in 9 days of culture (p. 3).

(2a) We chose to present the cell growth data as a rate, [fold change/day] or first derivative of cell number to demonstrate the relatively stable and constant rate of population growth in culture among the Sca-1_Low, Sca-1_Mid, and Sca-1_High fractions. The emphasis is thus on the rate in the various fractions at various times, and not on the absolute number of cells as in a typical growth curves. Additionally, to minimize variability in culture densities, all cultures were frequently split (every 2-3 days), so as to keep the average cell density within a narrow range. Thus, if presented as growth curves, they would be discontinuous and confusing.

To strengthen our conclusion that there was no statistical difference of growth rate between the sorted fractions, we now provide the mean and standard deviation of the growth rates from the 11 independent measurements in time for each sorted fraction in Figure S2b. As shown, the Sca-1_Low, Mid and High cultures all increased by approximately 3-fold in number/day, and there is no statistical difference of growth rate between these populations as seen by the overlap of the error bars.

(2b) As the Reviewer points out, with a cell division rate of 12 divisions/9 days one might expect that on average the majority of cells will have divided once within the first 23h. But due to the asynchrony of cell division cycles, the cell culture density may double in 23 hr (Figure S2) as a result of some cells dividing twice, and others not at all. Also, FACS sorting significantly slows down cell division for approximately 24 hrs in all sorted samples, as shown in Fig. S2. However, the sorted fractions recover by 48 hrs and exhibit a baseline macroscopic proliferation rate that results in a 3-fold increase in culture density per day thereafter. In retrospect, we agree that the passage “within the first 23 hrs after FACS only few cells have undergone cell division...” was a confusing way to summarize the argument that the observed heterogeneity cannot be explained entirely by uneven cell division (biased partitioning of cellular content to the two daughter cells). It has often been suggested that noise may be caused by a series of uneven partitioning of low abundance cellular molecules during cell division. However, our additional calculations and modeling show that even if all cells would divide just once within 23 hrs, the minimal departure away from perfect partitioning needed to generate the heterogeneity division would have produced aberrant multimodal distributions not consistent with the smooth unimodal distributions we observed. Since the theoretical possibility of uneven partitioning as a potential source of heterogeneity is not essential for conveying the main message of this manuscript, we have now removed the discussion and the confusing statement.

Comment 3: The authors suggest that Sca-1 low cells are in fact stem cells because they still have cKit expression compared to control cells. Although I agree that they do express cKit, it would be interesting to know if the levels are similar in Sca-1 mid and high cell types and the parental line itself. It may be that as cells differentiate they merely start to turn off both ScaI and cKit. Either result is interesting and should be added.

At the suggestion of this Reviewer, we include now in Figure S5, the c-kit expression of all three sorted fractions, the parental line, and a terminally-differentiated erythroid sample. As shown, Sca-1_Low, Sca-1_Mid, and Sca-1_High cells have nearly identical c-kit expression to each other, and comparable expression to the parental line. This indicates that within the dispersion of a clonal population, Sca-1 and c-kit are orthogonal or independent genetic dimensions. Upon differentiation, it is well-established that progenitor cells turn off both Sca-1 and c-kit, as the Reviewer mentions, and this is indeed what we observe.

4) Defining a mechanism for this phenomenon would be very interesting and should be considered for resubmission.

We agree that defining the mechanism for the clonal heterogeneity and the slow relaxation of outliers is interesting and important. There is a broad diversity of possible sources of heterogeneity, ranging from the classical stochastic fluctuations due to “gene expression noise” to some active biological process which may involve non-cell autonomous mechanisms in the control of population diversity. Since it is beyond the scope of this manuscript to provide experimental validation of any particular hypothesis at this point, we have instead focused on the rigorous quantitative and functional characterization of clonal heterogeneity. However, prompted by the Reviewer, we now provide a list of potential molecular mechanisms in the revised manuscript. Additionally, our mathematical model clearly shows that the relaxation cannot simply be explained by an Ornstein-Uhlenbeck process that would predict a first order kinetics. This supports our claim that the observations cannot be the manifestation solely of a simple stochastic process, such as gene expression noise as the single source of the stochastic heterogeneity. Thus, we caution succinctly that the mechanism underlying our observed clonal heterogeneity and slow relaxation may involve more complex mechanisms and mention the rugged “potential” landscape of complex networks as a likely cause in the manuscript.

Reviewer 3: Expert in bioengineering of transcriptional regulatory networks

In this paper, the authors report their findings on the noise properties of the stem cell marker Sca-1 in hematopoietic progenitor cells. They find that distribution of Sca-1 in these cells is consistent with a bimodal population model, and that the relaxation of sorted cells back into the stationary state confirms this. Additionally, they find that the relaxation of the high state cells back into the stationary distribution does not follow an exponential decay, but rather a sigmoidal one.

This is a very timely and interesting paper, that should be of interest to a broad range of researchers. However, I have some serious concerns over the modeling.

1) While exploring the “exponential decay” model of the time evolution of w_i, the authors report that they fit the model to experimental data. In table S3, they show that the best fit of the model gives negative values for w_1(0) for both the mid and high Sca-1 partitions. Certainly, this is unphysical. When doing the numerical fitting, they should constrain the algorithm to only search in relevant parameter space.

Our previous submission did not explain the fitting procedure clearly, hence the confusion. The negative value of the initial condition in the linear model appeared because, in an attempt to favor as much as possible the linear model as a null hypothesis, we reported the best possible fit we could obtain for the linear model. It is customary to fit the free parameter stemming from the integration constant in the ODE to the initial condition, but this resulted in a relatively poor fit. Instead, we fitted the integration constant to the asymptotic value. This provided a better fit at the cost of an unphysical negative initial value.

However, following the suggestion of this Reviewer, we now use the standard approach of fitting the initial condition. As can be seen in Table S2, all parameters now have physical values. The drawback is that, although the fit in terms of the MSE is roughly similar, the asymptotic behavior of the exponential model is significantly different from the observed values. This poor fit further highlights the inadequacy of the linear model to describe at once all aspects of the data. The new Theoretical Methods now expands our explanation of this result.

2) The authors state in the supplemental that the logistic model is a better fit. But is it really that much better of a fit that it should be deemed the correct model? Because the experimental data is so noisy, it may be hard to determine what the correct model is without some underlying biophysical theory to at least attempt to justify it. Plus, it is hard to judge the degree of the fit from the numbers given in the tables, since many of the parameters are completely non-physical. I do not know whether the error estimates are correct or not.

The Reviewer brings up an important point about model selection. The data are indeed noisy and it is difficult to determine which of the two models is better. As a way to quantify the goodness of fit, we have used the Akaike Information Criterion (AIC), which is widely employed in the literature for model selection. The AIC, which combines the MSE and a penalty term for the addition of fitting parameters, favors the logistic model over the linear model for the experimental data. Moreover, the logistic model better explains some of the qualitative features of the data, specifically the sigmoidal and the asymptotic behavior.

We agree with the Reviewer that a justification of our modeling based on an underlying theoretical understanding is desirable. Following this suggestion, in Section S4.C of the now revised Theoretical Methods, we present an expanded derivation of both the linear and the non-linear models based on a simple description of two interacting populations (corresponding to two states). In brief, if cells switch between the two states at a constant rate (independent events obeying first order kinetics), we obtain a linear model for the absolute populations. This model also translates into a linear equation for the fractions of cells in each population, which is the quantity measured by the experiments. In contrast, if we assume that the two subpopulations interact through a simple diffusive mechanism whereby signaling molecules from one cell type induce switching of the other cell type through a quadratic kinetic term, then a generalized form of the logistic model appears. In the Theoretical Methods section S4.C (b), we explain the underlying biological rationale related to the autocrine/paracrine self-promotion of differentiation that has been previously observed in cell populations. Specifically, if we assume a simple non-linear quadratic interaction between cell types, this gives rise to a (generalized) logistic equation for the temporal evolution of the fraction of cells in each subpopulation. Although this does not constitute a full-fledged biophysical model, the model makes minimal assumptions about the unknown genetic circuitry underlying the system and provides a biologically relevant framework

3) There needs to be some justification for the logistic model. Were any other models tried? What physical reason could there be for a logistic model, versus some other nonlinear model?

Please see the reply to the previous point for a discussion on derivation of the models, which we have now included in the new Theoretical Methods. We have extended Theoretical Methods section S4.C to detail the derivation of the linear and logistic models starting from a model of two interacting populations. Because of the lack of data on the specific biochemical reactions involved in the switching process, we decided against presenting additional models which could not be easily justified. Even though one could postulate other functional forms for the switching rates, any such model would be speculative since we do not have strong support for our choice of non-linear interaction. The model presented here can be considered a minimal model which assumes a biologically plausible mechanism for the non-linearity.

4) One major concern I have is the apparent correlation of the experimental data points in Fig. S11. If you look at the red data points, the “stochastic” jumps all seem to occur at the same places, especially in the center and right panels (and to a lesser, but still discernible, extent, the left). For instance, the last point in the series alway goes up, ans does the fifth from last. In fact, when I first saw the center and right panels, I thought it was the same data. Why does this occur? It might be that both sets of data were taken concurrently, so that the day to day variation of the measurement could account for it. However, this begs the question of how stable the measurements are over long periods. As the authors show in the main text, one individual trial is fairly accurate, when beads are looked at. How about when you look at beads with multiple time points over 450 hours?

The Reviewer correctly points out that in some cases there are similar day-to-day variations occurring in all samples. These variations can be either biological in nature, procedural in nature, or both. Flow cytometry analysis is very reliable for intra-session comparison, with barely detectable technical variability for aggregate quantities of a sample (one cell population). This allowed us to unambiguously show the enduring individuality of outliers and the slow relaxation. However, a comparison of absolute values between different sessions (time points) is inherently difficult to achieve due to day-to-day variability. To correct for this, we used standardized beads to remove technical (machine) noise, and the results reported were calibrated to correct for the changes of mean fluorescence among standardized beads over 450 hrs. However, despite extensive evaluation of this phenomenon, it is not possible at present to further reduce the remaining noise (day-to-day variability), possibly due to culture and reagents, since within-session replicates do not differ. Figure S18d shows a similar drift in the control population which reflects a shift in the measured fluorescence values.

Our choice of the binning algorithm to fit the data was also an attempt to overcome these daily fluctuations. Since the algorithm is probabilistic and allows for non-deterministic binning of cells into overlapping sub-populations, the algorithm is robust against shifts in the absolute fluorescence of the histograms, especially because of the large number of data points (10,000 data points per histogram). The most important feature of the process, and the focus of our modeling, is the evolution of the relative heights (weights) of the two peaks rather than their locations (means). We have used different normalization schemes and modifications of the binning algorithm and always found the theoretical fits (Fig. 2b) for the overall trend of the reconstitution of the parental population (Fig. 2a) to be robust with respect to overall shifts of the histograms. We have added Theoretical Methods section S4.C(c) to illustrate further quantitative analysis to address this issue.