Loss of TET2 affects proliferation and drug sensitivity through altered dynamics of cell-state transitions

Summary

A persistent puzzle in cancer biology is how mutations, which neither alter growth signaling pathways nor directly interfere with drug mechanism, can still recur and persist in tumors. One example is the mutation of the DNA demethylase TET2 in acute myeloid leukemias (AMLs) that frequently persists from diagnosis through remission and relapse (Rothenberg-Thurley et al., 2018; Corces-Zimmerman et al., 2014; Nibourel et al., 2010), but whose fitness advantage in chemotherapy is unclear. Here we use isogenic human AML cell lines to show that TET2 loss-of-function alters the dynamics of transitions between differentiated and stem-like states. A conceptual mathematical model and experimental validation suggest these altered cell-state dynamics can benefit the cell population by slowing population decay during drug treatment and lowering the number of survivor cells needed to re-establish the initial population. These studies shed light on the functional and phenotypic effects of a TET2 mutation in AML and illustrate how a single gene mutation can alter a cells’ phenotypic plasticity. A record of this paper’s Transparent Peer Review process is included in the Supplemental Information.

Graphical Abstract

eTOC Blurb

Many recurrent mutations in cancer do not directly alter growth signaling or interfere with drug mechanism, yet are frequently observed in remission and relapse. Here we show that one such mutation in TET2, a commonly mutated epigenetic modifier in AML, can alter switching dynamics between stem-like and differentiated cell states, enhancing population fitness. This study illustrates how one mutation can alter phenotypic plasticity, and thereby provide an evolutionary strategy to promote cell population growth in and out of drug.

Introduction

A major challenge in cancer biology is to understand the function of recurrent mutations in the emergence of tumors or response to drug therapy. Some mutations clearly benefit cancer populations by altering regulation of growth signaling or programmed cell death (e.g. mutations in p53 or TGFβ signaling, Sanchez-Vega et al., 2018) or by directly interfering with drug effect (e.g. acquired EGFR T790M mutations in response to EGFR inhibitors, Ma, Wei and Song, 2011). However, for many observed mutations it is unclear how they affect either proliferation or drug resistance.

One example are mutations in the DNA demethylase TET2, found in ~15-20% of de novo AMLs (i.e. AMLs in patients with no clinical history of myelodysplastic syndrome (MDS) and no prior exposure to leukemogenic therapies; Cheson et al., 2003; Nibourel et al., 2010; Cancer Genome Atlas Research Network, 2013). Mutated TET2 is associated with pre-leukemic states such as clonal hematopoiesis (where a clone becomes overrepresented in the blood; Busque et al., 2012), and mutational persistence and adverse outcome in human AML (Corces-Zimmerman et al., 2014; Rothenberg-Thurley et al., 2018; Ding et al., 2012; Metzeler et al., 2011). However, due to its ability to alter DNA methylation genome-wide, the mechanisms by which TET2 mutation confers a benefit to AML cancer cell populations remain unclear. One possibility is that epigenetic variation in cancer cells increases phenotypic flexibility, enabling tumor evolution and progression (Feinberg et al., 2016; Flavahan et al., 2017). Indeed, recent studies have shown that loss of epigenetic effectors such as KDM5A in breast cancer (Hinohara et al., 2018) can affect drug sensitivity by enabling greater cell heterogeneity. Here we investigate whether TET2 loss similarly affects cell fitness in AML using an integrated approach of mathematical modeling and experimentation in paired WT and TET2-mutant isogenic human AML cell lines. We discover that TET2 mutation alters the dynamics of transitions between stem-like and differentiated cell states, enhancing population fitness in chemotherapy and lowering the number of cells needed to establish a cell population.

Results

TET2 loss-of-function mutation renders AML cell populations more stem-cell like

To investigate the consequences of TET2 mutation, we chose to compare isogenic human myeloblast cell lines expressing wildtype or mutant TET2 (Figure 1A). In human AMLs, TET2 is often found with truncating mutations or missense mutations in its catalytic domain resulting in TET2 loss-of-function (Hirsch et al., 2018; Smith et al., 2010). Therefore, isogenic cell lines were created by knocking out TET2 (TET2^KO) in two cell lines (KG1 and Thp1, see Methods). Methylation profiling confirmed that TET2^KO has the expected effect on DNA methylation (Yamazaki et al., 2015; Asmar et al., 2013; Rasmussen et al., 2015): TET2^KO cell lines display a significantly higher degree of hypermethylation compared to their WT counterparts (TET2^WT; t-test p-values KG1 <2.2e-16 and Thp1 2.9e-4; Figure 1B).

Figure 1. TET2^KO cells are more stem-like than TET2^WT isogenic counterparts.

(A) Overview of molecular profiling performed on isogenic TET2 mutant AML cell lines. (B) Representative volcano plot and histogram of CpG methylation in TET2^KO cells relative to TET2^WT (shown for KG1). (C) TET2^KO cells show decreased expression of myeloid commitment markers (ITGAM, CORO1A) and increased expression of markers associated with stemness (CD38⁻, HLA-DRA⁻, TAL1) compared to TET2^WT in both KG1 and Thp1 cells (Supplemental Figure 1B, mean ± s.e., n=6). (D) KG1 and Thp1 (Supplemental Figure 1F) TET2^KO cells produce a higher percentage of colonies associated with oligopotent progenitors (CFUGEMM) compared to TET2^WT in methylcellulose assays (mean ± s.e., n=3).

To gain an unbiased overview of the molecular changes induced by TET2^KO, we examined the transcriptomic and epigenomic profiles of the paired isogenic cell lines. RNAseq analysis identified ~300 similarly differentially expressed transcripts in both cell lines (fold-change>2, Supplemental Figure 1A, including reduced expression of myeloid differentiation markers (ITGAM, CORO1A, Wald test, BH adjusted p-values 9.99e-9 and 7.73e-2) and increased expression of markers associated with leukemic progenitors (CD38⁻, HLA-DRA⁻, TAL1, Wald test, BH adjusted p-values 3.83e-26, 1.74e-16, and 3.90e-7, Figure 1C, Supplemental Figure 1B; Chan et al., 2012; Nishioka et al., 2013; Vagapova et al., 2018). These data are consistent with analysis of TCGA LAML (phs000178.v1.p1) and OHSU (Tyner et al., 2018) datasets, which showed that TET2 expression is strongly correlated with genes expressed in the granulocytic lineage (Fisher’s exact test adjusted p-values 6.27e-33 and 9.49e-20, Supplemental Tables 1-2, see Methods).

Differentially expressed genes in TET2^KO cells were found to be enriched for targets of RUNX1, a hematopoietic regulator known to promote stemness and myeloid fate decisions (Fisher’s exact test adjusted p-value 7.76e-6, Supplemental Figure 1C, Supplemental Table 3; Ran et al., 2013). Consistently, the proximal promoter of RUNX1 was significantly differentially methylated in TET2^KO (Wald test, minimum adjusted p-value 1.22e-27, Supplemental Figure 1D) with a concomitant increase in the expression the RUNX1a isoform (Supplemental Figure 1E). Methylcellulose colony forming unit assays of TET2^KO cells also show increased numbers of colonies associated with oligopotent progenitor cells compared to TET2^WT controls (CFUGEMMs, Figure 1D, Supplemental Figure 1F). Overall, these data suggested that TET2^KO populations acquire more stem-like signatures, which was further validated by comparing their DNA methylation profiles to those of normal hematopoietic progenitors and leukemic stem cells (LSCs, Supplemental Figure 1G).

The TET2^KO-mediated acquisition of stem-like signatures was much more pronounced in KG1 than Thp1 in terms of expression profile and potential to form colonies of diverse myeloid lineages in methylcellulose assays (Figure 1D, Supplemental Figure 1F). This is presumably because the monocyte-like cells, Thp1, are already more differentiated than the myeloblast-like cells, KG1, regardless of TET2 mutational status. We therefore chose to focus on KG1 cells for the remainder of this work.

TET2 mutation changes dynamics of cell-state switching

To test whether this change in stem-like molecular signatures is caused by an increase in the fraction of cells with CD34^hiCD38^lo surface marker expression (a classic LSC-like profile; Costello et al, 2000; Gerber et al., 2012; Nishioka et al., 2013; Zeijlemaker et al., 2018) in TET2^KO compared to TET2^WT cells, we measured CD34 and CD38 surface marker expression in the KG1 population using flow cytometry. Quantification of CD34/CD38 expression revealed that the fraction of putatively stem-like CD34^hiCD38^lo cells was indeed increased in TET2^KO cell populations (Figure 2A, Supplemental Figure 2A).

Figure 2. TET2^KO increases the propensity of differentiated (CD34^hiCD38^hi) Cells to switch to a more stem-like (CD34^hiCD38^lo) cell state.

(A) KG1 TET2^KO cell lines show a shift in CD38 surface marker expression. The threshold for calling cells CD38^lo or CD38^hi (vertical dashed line) is defined by negative controls, such that 99.5% of unstained cells are incorrectly called CD38^lo (Supplemental Figure 2A). (B) Schematic illustration of flow cytometry experiment to quantify CD38 expression over time. Cells from either KG1 TET2^WT or TET2^KO cell lines are sorted by CD38 expression via FACS (fluorescence-activated cell sorting) into CD38^lo or CD38^hi based on the threshold in (A), and changes in CD38 expression are assessed every 72h by flow cytometry. (C) The distributions of CD38 expression measured by flow cytometry for sorted “lo” or “hi” populations over 12d in 0uM AraC, colored by the original sorted population (n=3). (D-E) Shown is the percent of CD38^hi cells in originally pure populations of CD38^lo cells (D), or the percent CD38^lo cells in originally pure populations of CD38^hi cells (E) for KG1 TET2^WT and TET2^KO cells after 0, 3, 6, 9, and 12 days of growth in drug-free media (mean ± s.e., n=2).

To confirm that KG1 CD34^hiCD38^lo cells are more stem-like than CD34^hiCD38^hi, we fractionated TET2^WT and TET2^KO cells by CD38 expression and measured the expression of genes associated with hematopoietic differentiation. Consistent with bulk RNAseq data showing higher levels of pro-differentiation marker expression in TET2^WT cells compared to TET2^KO (Fig 1C), we found that differentiation markers increased with increasing levels of CD38 expression across both cell lines (trend seen for ITGAM, CORO1A, and TAL1, but not HOXA5; Supplemental Figure 2C). Further, growth rate decreased with increasing levels of CD38 expression (Supplemental Figure 2E) across KG1 TET2^WT and TET2^KO cells. Together, these data further support that KG1 CD34^hiCD38^lo cells are more stem-like and self-renewing, where CD34^hiCD38^hi cells are more differentiated and non-dividing.

How does TET2^KO alter the ratio between stem-like and differentiated cells? One possibility is that TET2^KO reduces the rate of differentiation. To test this, subpopulations of stem-like (L, CD34^hiCD38^lo) and differentiated (H, CD34^hiCD38^hi) cells were sorted in both TET2^WT and TET2^KO populations, and the ratio of L and H cells was monitored over time to steady-state (Figure 2D-E, Supplemental Figure 3B). Indeed, sorted stem-like L populations repopulated the differentiated state more slowly in TET2^KO (Figure 2D), and sorted TET2^KO differentiated cell populations repopulated the stem-cell like state more rapidly than TET2^WT (Figure 2E). These data suggest that TET2^KO facilitates the reversible switching from differentiated to stem-like states in AML cells and provides intuition for why TET2^KO has a larger fraction of stem-like cells in the population.

Mathematical modeling illustrates consequences of altered cell-state dynamics for survival of cell population

What are the functional consequences of having such altered cell-state dynamics, and in particular a higher fraction of stem-like cells? To address this question, we developed a simple mathematical model with two cell states (Figure 3, second panel): a stem-cell-like state L, and a more differentiated state H. This model is fully characterized by a linear, homogeneous system of ordinary differential equations (ODEs) with 6 parameters with values > 0 (see Methods). Parameters are denoted by: d_L and d_H, the death rates of the L and H states; g_L and g_H, the rates at which the L and H states proliferate; and r_LH and r_HL, the transition rates from L to H and H to L, respectively.

Figure 3. Mathematical model reveals advantageous and disadvantageous parameter regimes for cell-state switching.

(left) If model parameters satisfy Inequality 1, increasing r_HL and/or decreasing r_LH (black arrows) slows population decay and benefits a drug-treated cancer population. (General model) Schematic representation of mathematical model with two cell states: a stem-cell-like state L and a differentiated state H. Parameters d_L and d_H represent the death rates of the L and H states; g_L and g_H the proliferation rates; and r_LH and r_HL the transition rates from L to H and H to L. (Case 1) If cells rarely die (d_L ≈ 0, d_H ≈ 0) and the proliferation rate g_L is higher than g_H, increasing r_HL is always beneficial. (Case 2) Conversely, if one state were protected from drug effect (d_H ≈ 0), increasing r_HL is only beneficial in the unlikely scenario when net production rate of L is larger than the proliferation rate g_H.

Using this model, we asked under which circumstances the observed altered cell-state dynamics – specifically an increased switch rate towards a stem-like state – would benefit a cell population. First, we focused on the impact of such altered cell-state dynamics during drug treatment (i.e. high death rates). In this case, the time to population collapse is dominated by the largest negative eigenvalue; the less negative the eigenvalue, the slower the population collapse. By computing the eigenvalues for the ODE, it can be seen that increasing the switch rate towards L slows population decay and “benefits” a drug-treated cancer population as long as:

$$(g_L-d_L)>(g_H-d_H)$$ (Inequality 1)

This inequality simply compares net production rates (proliferation rate minus death rate) and requires it to be higher at L than H. In fact, while Inequality 1 is true, increasing r_HL or decreasing r_LH will slow population decay (Figure 3, first panel; see Methods) and thus “benefit” the cell population. This finding is in line with recent studies showing that cancer cells can reversibly transit between stem and differentiated states with different drug sensitivities (Jordan et al., 2016; Su et al., 2017; Gupta et al., 2011).

The same inequality also captures other cases in which switching to a stem-like state is either beneficial or detrimental (Figure 3, third and fourth panels). If cells rarely die, such as in a non-drug-treated condition (d_L ≈ 0, d_H ≈ 0) (Case 1), inequality 1 simplifies to

$$g_L>g_H$$

Here, switching to the L state (increasing r_HL) is beneficial for the cell population when the proliferation rate g_L of L is higher than the proliferation rate g_H of H, which is likely given the lower propensity of differentiated cells to divide. In this case, we expect cells carrying mutations that increase r_HL to take over the population. Conversely, if the differentiated cell state H were protected from the adverse effects of drug treatment (d_H ≈ 0) (Case 2), inequality 1 simplifies to

$$g_L-d_L>g_H$$

Here, increasing r_HL is only beneficial if the net production rate of L is larger than the proliferation rate g_H of H, which is unlikely for high doses of drug (d_L ≫ 0). In this case, we expect cells carrying mutations that increase r_HL to be depleted from the population over time.

Overall, this conceptual mathematical model suggests that altered cell-state dynamics – specifically an increased switch rate towards a stem-like state – can indeed benefit the cell population both in presence or absence of drug treatment, as long as it falls within the parameter regime outlined by Inequality 1.

Experimental validation of modeling predictions

Does the TET2 mutation put the AML cancer population in this predicted advantageous parameter regime? Since the proliferation, death and switching rates cannot be disentangled directly from measurements, we estimated these rates by fitting our model to time course data of sorted TET2^KO and WT populations in the presence of cytosine arabinoside (AraC, a common first-line chemotherapy drug for AML; Figure 4B, Supplemental Figure 3B, see Methods). These parameter estimates confirmed that both KG1 TET2^KO and WT populations during drug treatment are in a regime where the net production of CD34^hiCD38^lo stem-like L cells outstrips that of CD34^hiCD38^hi differentiated H cells (as required by Inequality 1, Figure 4A-B). Therefore, switching to the stem-like state is more advantageous. As the TET2^KO populations have a higher r_HL switching rate and a lower r_LH switching rate, the model predicts that the TET2^KO population, as a whole, is more drug resistant. Indeed, drug treatment experiments confirmed that TET2^KO populations are less sensitive to AraC and doxorubicin than TET2^WT populations (Figure 4C, Supplemental Figure 3G). We also used the same experimental setup to determine model parameters for different drug conditions and different CD38-defined subpopulations (Supplemental Figure 3B,E, Figure 2). Our key results – Inequality 1 holds, TET2^KO cells have a higher transit rate r_HL, and TET2^WT cells have a higher rate r_LH – held true across experiments (Supplemental Figure 3J).

Figure 4. TET2^KO cells’ altered cell-state dynamics enable longer-term drug survival in chemotherapy.

(A) Schematic illustration of model in 1uM AraC treatment. Numbers are representative estimated rates of transition, proliferation, and death (see Supplemental Figure 3H-I). (B) The percent of KG1 CD38^lo cells in originally pure populations of sorted CD38^hi cells in 1uM AraC after 0, 3, and 6 days (mean ± s.e., n=3). Dashed lines: predictions from the model. (C) Shown is fold-change cell viability relative to TET2^WT cells after 72h exposure to varying concentrations of AraC (mean ± s.e., n=5). (D) Model predictions for drug survival given increasing values of the transition parameter r_HL for a toy model that fulfills Inequality 1. Inset: survival as a function of the ratio of r_HL/r_LH. (E-F) TET2^KO CD38^hi cells were isolated by FACS and incubated with or without effector for 72h in the presence of 1uM AraC. (E) The CD38 distributions after 72h in DMSO, DS, or IFNg treatment (n=3, TET2^WT in Supplemental Figure 4A). (F) The percent of living cells in the population relative to control after 72h in DMSO, DS, or IFNg treatment (mean ± s.e., n=3). (G) Model predictions for colony outgrowth from a small population given sorted CD38^lo and CD38^hi (left) or unsorted cells (right; see Methods, mean of 10 runs). (H) Number of colonies in methylcellulose assays after seeding equal numbers of TET2^WT and TET2^KO cells (mean ± s.e., n=3). (I) Number of colonies in methylcellulose assays for cells sorted by CD38 expression (mean ± s.e., n=3).

Further analysis of the model suggested two additional predictions. First, the model suggested that a population with a higher fraction of differentiated cell states will show reduced population survival in drug treatment (Figure 4D). To test this conjecture, two effectors known to enrich AML cell populations for the differentiated state (without altering cell death, Supplemental Figure 4C) were used: the inflammatory stimulus interferon gamma (IFNg) and the aldehyde dehydrogenase inhibitor disulfiram (DS; Amici et al., 2018; Xu et al., 2017). We confirmed that treatment with IFNg or DS enriched both TET2^KO and WT populations for the differentiated cell state after 72 hours of exposure (Figure 4E, Supplemental Figure 4A-B). Moreover, both effectors increased the efficacy of AraC treatment in both TET2^KO and WT populations, while several other effectors not known to affect differentiation did not alter AraC sensitivity (Figure 4F, Supplemental Figure 4C, Supplemental Table 4).

Second, the model predicted that the increased ability of TET2^KO to revert to a stem-like cell state (with higher proliferation potential) will result in improved TET2^KO population outgrowth compared to TET2^WT (Figure 4G). To test this prediction, we compared the number of colonies formed by TET2^KO and TET2^WT populations in methylcellulose assays, which assesses the ability of isolated cells to reform a colony (Figure 4H). As predicted, sorted stem-like cell populations showed increased cell colony numbers in both TET2^KO and TET2^WT populations. Moreover, unsorted TET2^KO populations formed approximately 2x more colonies than their TET2^WT counterparts (Figure 4I), highlighting their increased potential for population renewal. Thus, our results suggest that TET2^KO populations not only have improved survival in the setting of therapeutic perturbations, but also have a higher likelihood to regrow an AML population from few surviving cells.

Discussion

Taken together, our results reveal that mutation of the epigenetic modifier TET2 in AML cells alters the transition dynamics between stem-like and differentiated cell states. These altered cell-state dynamics confer several benefits to the population in and out of drug, and provide a rationale for why the detection of TET2 mutants in patients in remission is strongly associated with a higher chance of relapse (Rothenberg-Thurley et al., 2018; Ding et al., 2012).

The key results from this study of a human CD34+CD38+ myeloid cell line (KG1) are consistent with past murine studies showing that TET2 mutation alters hematopoietic stem cell self-renewal and differentiation (Moran-Crusio et al., 2011). Future studies will be necessary to determine the generalizability of the cell state-switching phenomenon in other hematopoietic cell types and in vivo. Additionally, our study was limited to cell states defined by well-known surface markers. Incorporating powerful, unbiased approaches such as scRNAseq and analyses measuring RNA velocity (La Manno et al., 2018) could potentially uncover novel cell states and state switching dynamics that are relevant for drug survival.

An interesting question raised by this study is the degree to which the state switching rates are cell-intrinsic or extrinsically determined. In the process of confirming our model, we indeed found that the transition of TET2^KO cells from CD38^hi to CD38^lo could be altered by effectors like DS and IFNg (Figure 4F). To characterize the extent to which cell-state switching rates are affected by signaling between the cells themselves, we grew KG1 TET2^WT cells in media conditioned by TET2^KO cells and vice versa. The distribution of CD38 expression in TET2^WT cells was not affected by TET2^KO conditioned media, but CD38 expression clearly increased for TET2^KO cells grown in TET2^WT conditioned media with decreased switching rates from CD38^hi to CD38^lo (Supplemental Figure 4D-F). Together, these results suggest that the switching rates of TET2^KO cells can be influenced by external perturbations, while switching rates of TET2^WT cells are more stable. Future studies will be needed to identify key soluble factors that distinguish TET2^KO conditioned media from TET2^WT and, more generally, to fully test the extent to which cell state switching rates are intrinsically determined.

This work provides further evidence that mutations which modulate switching rates between more and less drug-resistant states may provide an “evolutionary shortcut” to counteract the adverse effect of drug treatment (Jordan et al., 2016; Su et al., 2017; Gupta et al., 2011). The mechanisms driving this dynamic tumor heterogeneity will require deeper study, however, future therapeutic strategies incorporating this knowledge may better counteract tumor evolution in response to treatment. For example, a natural extension of the work presented here would be to use the model and understanding of cancer cell state switching dynamics to determine ideal dosing strategies (dosage, time course, pulsed administration, and so on) to help therapeutically manage patient outcomes. A recently published review made similar suggestions for future therapeutic development in melanoma (Bai et al., 2019).

Our results have implications for AML mutations beyond TET2: the conceptual mathematical model shows that as long as stem-like and differentiated cell states differ in their net production during drug treatment (i.e., Inequality 1 is fulfilled), any other mutation increasing the transition rate towards the stem-like state will be beneficial as well. Thus, we conjecture that mutations in other epigenetic modifiers might confer a fitness advantage to drug-treated AML populations through similar mechanisms. In cancer, the target space for mutations that directly interfere with drug function is likely smaller than the target space for dysregulation of cell states; this is known to be the case in bacteria, where the number of genes conferring increased antibiotic tolerance far exceeds the number of genes conferring drug resistance (Girgis, Harris, and Tavazoie, 2012; Brauner et al., 2016). Future studies characterizing cell states and transition dynamics will provide new insight into why certain mutations are selected for, and persist in, cancer, and may lead to new approaches which increase the efficacy of current cancer therapies.

Key Changes Prompted by Reviewer Comments

The manuscript was extended with supplemental figures showing the robustness of switching parameters to CD38 expression, the biological significance of CD38-defined subpopulations, and the effect of conditioned media on cell-state switching. An experiment showing sorted subpopulations returning to steady-state is now highlighted in Figure 2C-E. Figure legends and the Methods section were expanded and clarified extensively, especially those pertaining to the selection of the CD38 threshold and the range of fit parameters. The Introduction was revised to include more literary context. The Discussion section was expanded to include further limitations of our study and potential future directions in therapeutically managing patient outcome. For context, the complete Transparent Peer Review Record is included within the Supplemental Information.

STAR Methods

Resource availability

Lead Contacts

Further information and requests for resources and reagents should be directed to and will be fulfilled by Steven J. Altschuler (steven.altschuler@ucsf.edu).

Materials Availability

This study did not generate new unique reagents.

Data and Code Availability

The RNA sequencing, DNA methylation, and flow cytometry datasets generated during this study along with original code to generate all figures are available at doi:10.17632/xmvz47rpg6.1. All original code to generate parameter fits or run the Gillespie algorithm is freely available for download at https://github.com/AltschulerWu-Lab/tet2-dynamics.

Experimental Model and Subject Details

KG1 (RRID: CVCL_0374; male) and Thp1/Thp1 TET2^KD (RRID: CVCL_0006; male) cells were generous gifts from the Shannon lab at UCSF and Levine lab at MSKCC, respectively. Cells were cultured in ATCC-recommended media and incubated at 37°C with 5% CO₂. All cell lines were maintained in 75ml or 250ml Suspension Culture flasks (CellTreat, Pepperell, MA) between 0.5e6 and 1e6 cells/ml.

Method Details

Generating TET2 mutants

The AML cell lines KG1 and Thp1 were selected as they express wildtype TET2 but do not express mutant FLT3, which is known to have synergistic epigenetic effects (Kunimoto et al., 2018, Shish et al., 2015). To make TET2^KO cells in the KG1 cell line, guides targeting exon 3 of TET2 (an exon with frequent indel mutations in patients with MDS; Smith et al., 2010) were designed in Benchling (5’-CACCGAGGCCAATTAAGGTGGAACC-3’). pSpCas9(BB)-2A-Puro (PX459) V2.0 was a gift from Feng Zhang (Addgene plasmid #62988). Guides were cloned into PX459 using Bbsl sites, and vectors transfected into KG1 cells with the Cell Line Nucleofector Kit R (Lonza Group, Basel, Switzerland) with protocol V-001 and according to the manufacturer’s protocol. After 48h, FITC-positive cells were isolated via FACS and expanded for 10 days. The TET2 mutation created a frameshift mutation causing truncation upstream of the conserved catalytic domain of TET2, consistent with tumors observed in patients (Weissman et al., 2012).

Drug and effector treatment

Prior to addition to cell culture media, Cytosine Beta-D-Arabinofuranoside (Sigma-Aldrich, St. Louis, MO) and 2-Deoxy-D-glucose (Sigma-Aldrich) were dissolved in H₂O, CCCP (Sigma-Aldrich), Disulfiram (Sigma-Aldrich), Etacrynic acid (Sigma-Aldrich), and Torin1 (Cell Signaling Technology, Danvers, Massachusetts) were dissolved in DMSO, and IFNg (Peprotech, Rocky Hill, NJ) was dissolved in culture media to 1000x the desired final concentration. Final DMSO (or other diluent) concentration was always 0.1%.

Cell-state transition model

We modeled a cell system with m states as a linear, homogenous system of ODEs. In matrix form this is written as

$$\frac{\mathit{d}\mathit{v}}{\mathit{d}\mathit{t}}=\mathit{Mv}$$

where v is cell-state vector of length m, and M is a m x m matrix representing cell-state transitions. We do not assume the matrix M is stochastic, which allows expansion or contraction of the total number of cells in the system—reflecting cell division or drug sensitivity—to occur. In our study, we considered two states: L (CD38^lo) or H (CD38^hi):

$$\frac{\mathit{d}}{\mathit{d}\mathit{t}}\left[\begin{array}{c}\hfill \mathit{L}\hfill \\ \hfill \mathit{H}\hfill \end{array}\right]=\left[\begin{array}{cc}\hfill ({\mathit{g}}_{\mathit{L}}-{\mathit{r}}_{\mathit{LH}}-{\mathit{d}}_{\mathit{L}})\hfill & \hfill {\mathit{r}}_{\mathit{HL}}\hfill \\ \hfill {\mathit{r}}_{\mathit{LH}}\hfill & \hfill ({\mathit{g}}_{\mathit{H}}-{\mathit{r}}_{\mathit{HL}}-{\mathit{d}}_{\mathit{H}})\hfill \end{array}\right]\left[\begin{array}{c}\hfill \mathit{L}\hfill \\ \hfill \mathit{H}\hfill \end{array}\right]$$

Where g_L and g_H are the growth rates and d_L and d_H the death rates of cell-states L and H, respectively. Here, r_LH is the transition rate from L to H, and r_HL is the transition rate from H to L. All matrix parameters are considered to be real, non-negative numbers. If we define

$$\begin{gathered} \alpha =-g_L+r_{LH}+r_{HL}-g_H+d_L+d_H \\ \beta =\sqrt{{\alpha }^2+4r_{LH}(g_H-d_H)+4r_{HL}(g_L-d_L)+4(g_H{d}_L+g_L{d}_H)-4(g_L{g}_H+d_L{d}_H)} \end{gathered}$$

then the eigenvalues of M are

$${\lambda }_1=\frac{1}{2}(-\alpha -\beta )$$

$${\lambda }_2=\frac{1}{2}(-\alpha +\beta )$$

with eigenvectors

$$\begin{gathered} \stackrel{\rightharpoonup }{{\mathit{v}}_1}=\left[-\frac{-{\mathit{g}}_{\mathit{L}}+{\mathit{r}}_{\mathit{LH}}-{\mathit{r}}_{\mathit{HL}}+{\mathit{g}}_{\mathit{H}}+{\mathit{d}}_{\mathit{L}}-{\mathit{d}}_{\mathit{H}}+\beta }{\stackrel{2{\mathit{r}}_{\mathit{HL}}}{1}}\right] \\ \stackrel{\rightharpoonup }{{\mathit{v}}_2}=\left[-\frac{-{\mathit{g}}_{\mathit{L}}+{\mathit{r}}_{\mathit{LH}}-{\mathit{r}}_{\mathit{HL}}+{\mathit{g}}_{\mathit{H}}+{\mathit{d}}_{\mathit{L}}-{\mathit{d}}_{\mathit{H}}+\beta }{\stackrel{2{\mathit{r}}_{\mathit{HL}}}{1}}\right] \end{gathered}$$

For convenience, we define the difference in production between the two states

$$\mathit{P}\equiv ({\mathit{g}}_{\mathit{L}}-{\mathit{d}}_{\mathit{L}})-({\mathit{g}}_{\mathit{H}}-{\mathit{d}}_{\mathit{H}})$$

To find the regime where changing r_HL (asympototically) increases survival rate, we focus on the larger of the other two eigenvalues, namely λ₂. Taking the derivative of λ₂ with respect to the state-switching parameters gives:

$$\begin{gathered} \frac{d{\lambda }_2}{d{r}_{HL}}=\frac{1}{2}\left(\frac{r_{LH}+r_{HL}+P}{\beta }-1\right) \\ \frac{d{\lambda }_2}{d{r}_{LH}}=\frac{1}{2}\left(\frac{r_{LH}+r_{HL}+P}{\beta }-1\right) \end{gathered}$$

Using Wolfram∣Alpha, we find: a) $\frac{d{\lambda }_2}{d{r}_{HL}}>0$ holds if P > 0 and $r_{HL}>-(\sqrt{r_{LH}}-\sqrt{P}{)}^{{}^2}$; and b) $\frac{d{\lambda }_2}{d{r}_{LH}}\le 0$ holds if either (i) P > 0 , 0 ≤ r_LH < P , and $r_{HL}>-(\sqrt{r_{LH}}-\sqrt{P}{)}^{{}^2}$, or if (ii) P > 0 and r_LH ≥ P. In our model, we only consider non-negative, real-valued parameter values. Thus, as long as the inequality P > 0 holds, $\frac{d{\lambda }_2}{d{r}_{HL}}>0$ and $\frac{d{\lambda }_2}{d{r}_{LH}}\le 0$.

A two-state system with constant switching rates can equilibrate to a constant fraction of states (a steady state) provided the cell population does not rapidly crash. This is possible because flux into and out of each state can be balanced, and is why observations such as Figure 2D (where TET2^KO cells appear to not switch and instead retain a stem-like state) can still be an accurate reflection of cells in constant flux. However, cell proliferation, death, and switching rates cannot be disentangled directly from bulk measurements. Therefore, we estimated these rates by fitting our model to time course data of sorted TET2^KO and WT populations. To find best fit parameters for this system, cell-state compositions from the flow cytometry experiments shown in Supplemental Figure 3B & E or Figure 2D-E (using data from both CD38^hi- and CD38^lo-sorted subpopulations, described below) were used as input into the model. Prediction error (the sum of the differences between the predicted and measured number of cells in each cell-state at each timepoint) was minimized with fminsearch in MATLAB (version R2019a) with MaxIter of 5000. To eliminate “solutions” that did not converge or had poor fit, prediction error for 1000 sets of random parameter values was calculated to form a null distribution for each condition, and the 0.005-quantile for this null distribution was used as an upper limit for reasonable error values (Supplemental Figure 3H). To establish the probable range of transition parameters, best-fit parameters for 1000 random initializations (x0 in the fminsearch function) were found and plotted (Supplemental Figure 3I). Representative values for fit parameters for KG1 TET2^WT and TET2^KO cells are shown in Figure 4A. Median values for fit parameters or fit parameter ratios from converged solutions are shown in Supplemental Figure 3J and Supplemental Figure 4F.

Solving the matrix ODE gave

$$\left[\begin{array}{c}\hfill L\hfill \\ \hfill H\hfill \end{array}\right]=\left[\begin{array}{cc}\hfill \stackrel{\rightharpoonup }{v}{}_1^1\hfill & \hfill \stackrel{\rightharpoonup }{v}{}_1^2\hfill \\ \hfill \stackrel{\rightharpoonup }{v}{}_2^1\hfill & \hfill \stackrel{\rightharpoonup }{v}{}_2^2\hfill \end{array}\right]\left[\begin{array}{c}\hfill A{e}^{{\lambda }_{1}t}\hfill \\ \hfill B{e}^{{\lambda }_{2}t}\hfill \end{array}\right]$$ (12)

After solving for the constants A and B, model predictions were made using representative best fit parameters and initial conditions observed in the experiment (e.g. number of CD38^hi and CD38^lo cells in day 0 sorted cell populations; method for calling CD38^lo/hi described below).

For colony forming simulations, 100 colonies were “seeded” with 10 individual cells as colony founders with 50% in the L state and 50% in H unless otherwise noted. To simulate growth without drug treatment, rates were estimated by fitting the model to flow cytometry experimental data (described below) without AraC treatment. The Gillespie algorithm was implemented to find the number of cells in the L or H states at time t. Cells “grew” for 10 time-steps before the number of cells in each colony was counted. Colonies with more than 100 cells were considered “grown out”.

DNA methylation profiling

DNA methylation was profiled using Illumina’s Infinium MethylationEPIC BeadChip (Illumina, San Diego, CA). DNA of technical replicates for each condition was extracted using the Zymo Quick-DNA kit (Zymo Research, Irvine, CA, KG1 n=6 per condition, Thp1 n=2 per condition). Bisulfite conversion, nanodrop quantitation, array scanning, and normalization was performed by the Vincent J. Coates Genomics Sequencing Laboratory at UC Berkeley.

mRNA-seq

RNA extraction of technical replicates was performed using the Lexogen SPLIT RNA extraction kit (Lexogen, Vienna, Austria, n=6 per condition), and libraries were prepared using the QuantSeq 3’ mRNA-Seq Library Prep Kit FWD for Illumina. Samples were quantified with Invitrogen Qubit (Invitrogen, Carlsbad, CA) prior to pooling, and library size and integrity was confirmed using the Agilent Bioanalyzer with the high-sensitivity DNA kit (Agilent, Santa Clara, CA). RNA sequencing was performed using 50bp single-end sequencing on the Illumina HiSeq 4000 in the Center for Advanced Technology at UC San Francisco. A PhiX control library was used as an in-run control, spiked in at 5%.

Immunofluorescence for flow cytometry

Cells were pelleted and washed with wash buffer (HBSS + 1% BSA, filtered with a 50ml Steriflip unit (Millipore, Burlington, MA)) prior to 30m incubation in Fc Receptor Blocker (Innovex Biosciences, Richmond, CA) on ice in the dark. Cells were washed twice before resuspension in wash buffer containing conjugated antibodies for flow cytometry at their recommended concentrations (CD34-PE 555822, CD38-PE-Cy5 555461, Becton Dickinson, Franklin Lakes, NJ), and incubation for 30m on ice in the dark. Cells were washed three times and resuspended in 350uL wash buffer before measurement with flow cytometry at a flow rate of 9.0. Doublets were called based on gates drawn for FSC-A and FSC-W, and dead cells were counted based on gates drawn in FSC-A and SSC-A. All flow cytometry or FACS was performed on the Aria IIu in the Center for Advanced Technologies at UC San Francisco.

External marker tracking by flow cytometry

For Figure 2 and Supplemental Figure 3B-C, equal numbers of CD38^hi and CD38^lo subpopulations of TET2^WT and TET2^KO cells were isolated by FACS (gate in Figure 2A, set such that <0.5% of unstained cells would be counted as CD38^hi; see Supplemental Figure 2A) and seeded separately in a round-bottom 96-well plate (CELLSTAR) at 1e5 cells/ml. For Supplemental Figure 3E-F, equal numbers of cells from the “tails” (bottom/top 5%) of the CD38 distributions of stained, unsorted TET2^WT and TET2^KO cells (see Supplemental Figure 3D) were isolated by FACS and seeded separately in a round-bottom 96-well plate (CELLSTAR) at 1e5 cells/ml. For day 0 timepoints, aliquots of freshly sorted cells were reflowed and recorded. For Supplemental Figure 3A-C, cells were treated with 0, 1, or 4uM AraC in technical triplicates. No drug was applied to the technical replicates shown in Figure 2C and Supplemental Figures 3D-F. Plates were covered with Breathe-Easy sealing membranes (Sigma-Aldrich) before incubation. For longer timepoints, media (with or without drug, as relevant) was exchanged every 3 days. At each timepoint, all cells in each replicate well were washed and stained for CD34 and CD38 (described above) and re-profiled by flow cytometry. To account for day-to-day variations in Aria IIu laser power or other settings, raw fluorescence intensity for each channel was normalized by the median intensity of unsorted, unstained controls of the appropriate cell line for each timepoint prior to analysis. For all experiments, CD38^hi and CD38^lo cell counts for analysis and model-fitting were called based on the same sorting threshold shown in Figure 2A (see Supplemental Figure 2A). The number of living cells in each cell state served as input to the model to fit parameters for the cell-state transition matrix in both cell lines for both treated and untreated conditions (described above).

Methylcellulose colony forming unit assay

Cells were counted with the TC20 automatic cell counter (Bio-Rad) with Trypan blue prior to plating in triplicate as technical replicates in MethoCult H4034 Optimum (Stemcell Technologies, Vancouver, CAN) in 35mm cell culture dishes (Eppendorf, Hamburg, DE) according to the manufacturer’s instructions. Cells were incubated for 2 weeks at 37°C with 5% CO₂. To maintain humidity, 4 MethoCult dishes were incubated amongst three lidless 35mm dishes each filled with 3ml sterile ddH₂O inside a 150mm glass petri dish with a lid (Corning, Corning, NY). Water was replenished every 3 days. Colonies were enumerated according to the manufacturer’s instructions. Results shown are representative results from two independent experiments.

Cell viability assays

For sorted population growth or effector viability assays (Supplemental Figures 2E & 4C), TET2^WT and TET2^KO cells were stained for CD34 and CD38 (described above) and segregated by CD38 expression with FACS as shown (Supplemental Figure 2D and Figure 2A, respectively). For day 0 timepoints, aliquots of freshly sorted cells were reflowed and recorded. Sorted cells were plated in round-bottom 96-well plates (CELLSTAR, Dallas, TX) at 1e5 cells/ml with replicates for each timepoint. For treated conditions, AraC and effectors were diluted and added as described above. Plates were covered with Breathe-Easy sealing membranes (Sigma-Aldrich) before incubation at 37°C with 5% CO₂. For longer timepoints, media (with or without drug, as relevant) was exchanged every 3 days. At each timepoint, all cells in each replicate well were washed, stained, and measured with the Aria IIu for the first 2000 events. The density of living cells (# of cells per ml, Supplemental Figure 2E) was calculated as the number of living cells observed per sample divided by the total length of time the sample took to reach 2000 events, and divided by the flow rate (~90 uL/minute for flow rate 9.0 on the Aria IIu).

For drug viability assays (Figure 4C, Supplemental Figure 3G), cells were plated as technical replicates at 1e5 cells/ml and treated with drug for 72h. For readout, plates were allowed to cool to room temperature before combining well-mixed cells and cell media 1:1 with room temperature CellTitre-Glo 2.0 (Promega, Madison, WI) in opaque white tissue culture plates (Corning). Reactions were allowed to proceed according to the manufacturer’s protocol, and luminescence was read out with the Biotek H4 plate reader (BioTek, Winooski, VT) in the Center for Advanced Technology at UC San Francisco. Results shown are representative results from three independent experiments.

RT-qPCR and PCR

For measuring TET2 expression, RNA was extracted from three technical replicates with the Direct-zol RNA Miniprep kit (Zymo Research) according to manufacturer’s instructions with TRI Reagent (Thermo Fisher Scientific, Waltham, MA). For measuring CD38, ITGAM, CORO1A, TAL1, and HOXA5 expression, cells were stained for CD34 and CD38 (described above) and fractionated by CD38 expression via FACS (gates shown in Supplemental Figure 2B). RNA was extracted from sorted cells (2 replicates of 300k cells) with the RNeasy Plus Mini Kit (QIAGEN, Hilden, DE) according to manufacturer’s instructions. Reverse transcription with iScript Reverse Transcription Supermix (Bio-Rad, Hercules, CA) was followed by qPCR with DyNAmo Flash SYBR Green qPCR kit (Thermo Fisher Scientific) according to the recommended protocol on the BioRad CFX Connect in the Center for Advanced Technologies at UC San Francisco. The set of genes measured was composed of genes of interest from the bulk RNAseq experiment (Supplemental Figure 1B) and relevant gene sets in the literature (GSEA: gal_leukemic_stem_cell (Gal et al., 2006), gentles_leukemic_stem_cell (Gentles et al., 2010), eppert_ce_hsc_lsc (Eppert et al., 2011)). Primers were obtained from IDT (San Jose, CA). Primer sequences for human RUNX1, TET2, HOXA5, and GAPDH were as previously described (Fujita et al., 2001; Cimmino et al., 2017; McLaughlin-Drubin et al., 2011). Primer sequences for CD38, CORO1A, ITGAM, and TAL1 were generated with the Primer Design Tool in Benchling.

Quantification and Statistical Analysis

Differential RNA expression analysis

mRNA reads were mapped and counted using the Integrated QuantSeq data analysis pipeline on Bluebee Platform (Bluebee, Rijswijk, Netherlands). Briefly, reads were trimmed with BBDuk, aligned to human GRCh38 with STAR, and counted with HTSeq-count. Gene filtering and differential expression analysis was performed in R (version 3.6.0; R Core Team, 2019). Genes were filtered by count such that all genes had 3 or more samples with 10 or more counts. Differential expression analysis was then performed using DESeq2 (version 1.24, Love et al., 2014) on gene counts. Genes that were found to be significantly differentially expressed (N=6 technical replicates; Wald test, BH adjusted p-value < 0.05) with an absolute Log₂fold-change in expression > 1.5X in a paired analysis of untreated cell lines were submitted to Enrichr (Kuleshov et al., 2016) for enrichment analysis. Statistical details can be found in the Results section and legends for Figure 1 and Supplemental Figure 1.

Differential DNA methylation analysis

Differential methylation analysis was performed with ChAMP (package version 2.14.0, Tian et al., 2017) in R (version 3.6.0; R Core Team, 2019). Significantly differentially methylated regions (KG1 n=6, Thp1 n=2 technical replicates per condition; Wald test, BH adjusted p-value < 0.05) were found with DMRcate (version 1.20, Peters et al., 2015), and visualization done with Gviz (version 1.28, Hahne et al., 2016). Statistical details can be found in the Results section and legend for Supplemental Figure 1.

Methylation age and stemness profiles

Quantile normalized values from the aforementioned ChAMP analysis were used as input to the DNA methylation age calculator in R as per the tutorial (horvath.genetics.ucla.edu) (Horvath, 2013). For visualization purposes, results for TET2^WT and TET2^KO cells were normalized to TET2^wt.

For the “stem-like” epigenetic signatures, we used a publicly available dataset of DNA methylation profiles from normal hematopoietic progenitors and leukemic cells sorted by CD34/CD38 expression (GEO: GSE63409; Jung et al., 2015). Linear discriminant analysis (LDA) was applied using the MASS package (version 7.3-51.4, Venables et al., 2002) in R to identify an optimal transform that increased separation of DNA methylation profiles across three key reference populations: normal HSCs, CD34⁺CD38⁻ putative LSCs, and CD34⁻ leukemia cells. The rest of the public dataset, as well as methylation data from the paired TET2^wt/TET2^ko cell lines, was projected into the lower-dimensional LDA space.

TET2 expression analysis

Expression profiles for the following datasets were queried from the Cancer Genomics Data Server (CGDS) using cdgsr (version 1.3.0, Cerami et al., 2012). Our results are, in part, based upon data generated by The Cancer Genome Atlas managed by the NCI and NHGRI. Information about TCGA can be found at http://cancergenome.nih.gov. For the datasets TCGA-LAML (dbGaP accession phs000178.v1.p1) and AML-OHSU (Tyner et al., 2018), genes with coefficient of variation < 0.1 across samples were dropped from subsequent analysis. Pearson correlation calculations were performed in R (version 3.6.0). Genes with expression that strongly correlated (cor > 0.45) or anti-correlated (cor < −0.45) with TET2 expression were submitted to Enrichr (Kuleshov et al., 2016) for enrichment analysis. Statistical details can be found in the Results section corresponding to Supplemental Tables 1-2.

KEY RESOURCES TABLE

REAGENT or RESOURCE	SOURCE	IDENTIFIER
Antibodies
PE Mouse anti-Human CD34	BD Biosciences	555822
PE-Cy5 Mouse anti-Human CD38	BD Biosciences	555461
Chemicals, Peptides, and Recombinant Proteins
Cytosine Beta-D-Arabinofuranoside	Sigma-Aldrich	C6645
Disulfiram	Sigma-Aldrich	D2950000
Human IFN-g	PeproTech	300-02
Critical Commercial Assays
CellTiter-Glo 2.0 Cell Viability Assay	Promega	G9242
QuantSeq 3’ mRNA-Seq Library Prep Kit FWD for Illumina	Lexogen	0.15.24/96
MethoCult H4034 Optimum	STEMCELL	#04034
Deposited Data
Raw and analyzed data	This paper; Mendeley data	doi:10.17632/xmvz47rpg6.1
Epigenome analysis of leukemia stem, blast and normal hematopoietic stem/progenitor cells	Jung et al., 2015; GEO	GSE63409
Experimental Models: Cell Lines
Human male: KG-1	ATCC	CCL-246; RRID: CVCL_0374
Human male: THP-1	ATCC	TIB-202; RRID: CVCL_0006
Oligonucleotides
Primer HOXA5-F: TCTCGTTGCCCTAATTCATCTTTT	McLaughlin-Drubin et al., 2011	N/A
Primer HOXA5-R: CATTCAGGACAAAGAGATGAACAGA	McLaughlin-Drubin et al., 2011	N/A
Primer CD38-F: CACCAAGCGCTTTCCCGAGACC	This paper	N/A
Primer CD38-R: GAGAGGCCCCTCCAGTGCAGAA	This paper	N/A
Primer CORO1A-F: GTGGTCCGCTCCAGCAAGTTCC	This paper	N/A
Primer CORO1A-R: CAGACCGTGGGCGCATTCTTGT	This paper	N/A
Primer ITGAM-F: CTCCGTGGACGTGGACAGCAAC	This paper	N/A
Primer ITGAM-R: AATGGCCACGTCCGTCAGCTTG	This paper	N/A
Primer TAL1-F: GGGAGCCGGATGCCTTCCCTAT	This paper	N/A
Primer TAL1-R: ACTTCATGGCCAGGCGGAGGAT	This paper	N/A
Recombinant DNA
pSpCas9(BB)-2A-Puro (PX459) V2.0	Addgene	#62988
Software and Algorithms
Code to generate all figures	This paper; Mendeley data	doi:10.17632/xmvz47rpg6.1
Code pertaining to model	This paper; GitHub	https://github.com/AltschulerWu-Lab/tet2-dynamics
DNA methylation age calculator	Horvath, 2013	https://horvath.genetics.ucla.edu/html/dnamage/
R	R Core Team, 2019	https://www.R-project.org
MATLAB	R2019a	https://www.mathworks.com

Highlights.

• TET2 knockout in human AML cell lines increases stem-like signatures

• TET2^KO cells have altered dynamics between stem-like/differentiated states

• Altered cell-state switching dynamics provides fitness advantage in and out of drug