Intrinsic growth heterogeneity of mouse leukemia cells underlies differential susceptibility to a growth-inhibiting anticancer drug

Cancer cell populations consist of phenotypically heterogeneous cells. Growing evidence suggests that pre-existing phenotypic differences among cancer cells correlate with differential susceptibility to anticancer drugs and eventually lead to a relapse. Such phenotypic differences can arise not only externally driven by the environmental heterogeneity around individual cells but also internally by the intrinsic fluctuation of cells. However, the quantitative characteristics of intrinsic phenotypic heterogeneity emerging even under constant environments and their relevance to drug susceptibility remain elusive. Here we employed a microfluidic device, mammalian mother machine, for studying the intrinsic heterogeneity of growth dynamics of mouse lymphocytic leukemia cells (L1210) across tens of generations. The generation time of this cancer cell line had a distribution with a long tail and a heritability across generations. We determined that a minority of cell lineages exist in a slow-cycling state for multiple generations. These slow-cycling cell lineages had a higher chance of survival than the fast-cycling lineages under continuous exposure to the anticancer drug Mitomycin C. This result suggests that heritable heterogeneity in cancer cells’ growth in a population influences their susceptibility to anticancer drugs.

Phenotypic heterogeneity in a cancer-cell population is often linked to differential drug 2 susceptibility and can result in incomplete tumor remission. Although various 3 mutational causes of heterogeneous susceptibility to drugs have been documented [1][2][3], 4 growing evidence indicates that phenotypic heterogeneity arising due to non-genetic 5 causes also affects survival of individual cancer cells upon drug exposure [4][5][6][7][8][9][10][11][12]. For a drug-exposed population is correlated with the temporal over-expression of 10 resistance-marker genes [5]. Although the underlying molecular mechanisms may differ 11 depending on the drugs and cancer-cell types, phenotypic cell-to-cell variation might be 12 July 5, 2020 1/14 a general phenomenon underlying the appearance of survivors from an isogenic 13 population of cancer cells exposed to lethal stress [9]. 14 The drug tolerance of individual cell lineages could be correlated with physiological 15 states that exist prior to the drug exposure. In fact, it has been shown that cancer-cell 16 populations whose growth is suppressed, due to their entry into the stationary phase [8], 17 or due to prior exposure to growth-inhibiting chemicals [13,14], produce significantly 18 higher frequencies of survivors than do actively growing populations. It is also 19 demonstrated that a rare, long-term dormant subpopulation exists in acute 20 lymphoblastic leukemia cells and resists to anticancer treatment [15]. 21 Despite the accumulating evidence on the relationship between the heterogeneity of 22 growth phenotype and drug susceptibility, it remains unexplored what causes such 23 heterogeneity in a cancer cell population. One plausible cause is inhomogeneous 24 environments around individual cells that are omnipresent in vivo. For example, the 25 spatial organization of tumor tissue inevitably creates environmental heterogeneity 26 around individual cells [16,17]. Another but not mutually exclusive possibility is 27 intrinsic fluctuations of cellular states arising from the factors such as stochasticity in 28 gene expression and metabolism [18][19][20]. Therefore, it is important to examine the 29 extent of phenotypic heterogeneity arising intrinsically even under constant 30 environments and whether it is sufficient to alter individual cells' drug susceptibility. 31 However, quantifying the properties of intrinsic phenotypic heterogeneity in isolation 32 from externally-driven heterogeneity has been a challenge in cancer cell research. 33 In this report, we measured the fluctuations of generation time (interdivision time) 34 of individual mouse lymphocytic leukemia cells (L1210) over tens of generations in 35 constant environments using a microfluidic device. We found that a stably maintained 36 culture of L1210 cell line harbors rare, slow-cycling cell lineages. The generation time 37 was positively correlated between mother and daughter cells, which thus revealed the 38 heritability of growth phenotypes. We tested the susceptibility of the slow-cycling cell 39 lineages to an anticancer drug, Mitomycin C (MMC), and found that they tended to 40 survive longer than fast-cycling cell lineages. Therefore, our results highlight that 41 intrinsic heterogeneity in cancer cells' growth could generate a spectrum of drug 42 sensitivity in a population and may reduce the efficiency of anticancer treatment.

44
L1210 cells can grow and divide stably in the microfluidic device 45 To monitor single L1210 cells for multiple generations, we fabricated a 46 mammalian-optimized version of mother machine microfluidic 47 device [8,[21][22][23] (Fig 1A-C). The mother machine was originally developed to analyze 48 single bacterial cells [21] and then adapted for studying eukaryotic cells [8,22,23]. In 49 these devices, individual cells are placed at the closed end of the growth channels.

50
Growth and division of the cells push them towards the open end, eventually excluding 51 them from the growth channel (S1 Video). An almost identical device was used 52 previously for studying L1210 cells in a CO 2 -independent medium (L-15) [8]. In our 53 present work, we placed the device inside an on-stage CO 2 chamber and flowed 54 standard CO 2 -dependent medium (RPMI-1640) through the device. We confirmed that 55 the pH of the medium in the device was maintained at 7.5 and that it stayed robust to 56 changes in the flow rate (S1 Fig; [24]). 57 We observed hundreds of individual cell lineages simultaneously for one week by 58 time-lapse microscopy and recorded every cell division event (Fig 1D and S1 Video). To 59 check the stability of the growth of the cells in the device, we measured their division 60 rate and found it to be almost constant (9.0 ⇥ 10 2 h 1 ) throughout observation Twenty representative single-cell lineages of L1210 cells cultured in a constant environment are shown. Each horizontal line represents a single-cell lineage, and the points represent the time points at which cell division occurred. The blue cell lineages are fast-cycling cell lineages, which divided 12 times or more during the seven-day culture; the red cell lineages are slow-cycling cell lineages, which divided 11 times or less during the same period. Co-existence of the heterogeneous growth phenotype is further discussed in the main text. (E) Constant division rate. At each time point t, we estimated an instantaneous cell division probability as D(t) N (t) , where N (t) is the number of surviving cells, and D(t) is the number of cells that underwent cell division within the time-lapse interval (t, t + t]. t is 10 min throughout this study. The vertical axis is the cumulative sum of instantaneous division probability, i.e., . Blue points represent the experimental data. The black broken-line is the line of linear fit; the slope is (9.033 ± 0.003) ⇥ 10 2 h 1 . The cumulative sum of division probability increases linearly with t, which indicates that the division rate was constant throughout the observation period. (F) Constant death rate. The fraction of surviving cell lineages was plotted over time with the vertical axis on the log scale. Blue points are the experimental data. Black broken-line is the line of linear fit; the slope is (2.239 ± 0.009) ⇥ 10 3 h 1 . The surviving fraction decreased in an exponential manner, which indicated a constant death rate throughout the observation period.
( Fig 1E). One of the advantages of using the mother machine is that we can analyze observed (S1 Video), and the death rate was found to be constant (2.2 ⇥ 10 3 h 1 , 65 corresponding to roughly 2% chance of death per generation) ( Fig 1F). Cell death under 66 favorable growth conditions with low frequency has also been reported in bacteria and 67 fission yeast [21][22][23], which presumably reflects an accidental loss of cellular homeostasis 68 or stochastic triggering of signal transduction pathways leading to cell death. The   tail, which indicates that the L1210 cell population harbors cells that remain undivided 79 for extended periods (Fig 2A and B). Magenta points are the experimental data; the blue broken-line represents the curve fitted to the experimental data by a mixed shifted-exponential distribution (Eq 2); the green broken-line represents a curve of a single shifted-exponential distribution. The vertical line at ⌧ = 13.31 h indicates the threshold of generation time above which the survival function becomes smaller than 0.06. (C) Correlation between the generation times of mother and daughter cells. (D) Autocorrelation of generation time. Error bars represent standard errors. (E) Division count distribution. The number of cell divisions in each lineage that stayed alive during the seven-day culture (n = 341) was counted. The red columns represent the distribution of division counts; the blue curve represents the estimated distribution of division counts, assuming that there is no trans-generational correlation between generation times. The latter was calculated by randomly sampling generation times from the experimental data.
the observed distribution of the generation times of L1210 cells cannot be fitted to a 83 single shifted-exponential distribution due to an inflection in the decay rate at the 84 generation time of ⌧ = 14 h ( Fig 2B). Instead, we noticed that the distribution observed 85 for L1210 cells was well-explained by assuming the existence of two subpopulations 86 whose respective shifted-exponential distributions are characterized by distinct decay 87 constants 1 and 2 , i.e., the generation time distribution follows the equation: where ⌧ 0 is a fixed no-division period (= minimum cell cycle length), and a (0 < a < 1) 89 is a fraction of cells whose generation time distribution is characterized by the decay 90 constant 2 . The probability that a newborn cell remain undivided until age ⌧ is given 91 by the survival function (Fig 2): This suggests that the L1210 clonal cell population is composed of multiple (at least 93 two) subpopulations, each of which follows distinct cell division statistics. 94 We also calculated the correlation between the generation times of mother and 95 daughter cells and found a relatively strong positive correlation (Pearson's correlation Fig 2C). 97 The positive correlation between the generation times of mother and daughter cells 98 observed here is one of the strongest reported to date for eukaryotic and prokaryotic 99 cells [27][28][29][30][31][32][33]. The correlation monotonically decayed across generations, but we still 100 found a positive correlation of r = 0.20 ± 0.02 even after ten generations ( Fig 2D). This 101 result hints at the existence of long-term memory of growth states in individual cells. 102 We predicted that the existence of a slow-cycling cell subpopulation and the 103 inheritance of generation times should produce a significant heterogeneity in the 104 division frequency among cell lineages. Indeed, we found that the division counts of cell 105 lineages surviving until the end of the observation period showed a wide distribution 106 with a long left-tail ( Fig 2E). Thus, some cell lineages underwent division only rarely 107 despite the presence of favorable growth conditions. To corroborate the significance of 108 July 5, 2020 4/14 the correlation of mother-daughter division statistics within lineages, we generated 109 pseudo-lineage data by random re-sampling from the distribution shown in Fig 2A. We 110 confirmed that the left tail of the corresponding division count distribution was reduced 111 (Fig 2E). Therefore, the heavily right-skewed distribution of generation time and the 112 inheritance of the growth phenotype across generations produce the heterogeneity of 113 division frequency among cellular lineages of L1210 cells.

114
A previous study on the growth of L1210 cells reported that the transition of 115 generation time across generations is essentially a deterministic process characterized by 116 a small set of parameters [30]. To check whether such a deterministic picture also applies 117 to our cases, we performed a correlation-dimension analysis on the generation time data. 118 However, we did not find a convergence correlates with their drug susceptibility, we applied MMC containing media in the 127 mother machine and followed the fates of individual cells. We first confirmed that MMC 128 is effective in our mother machine system; upon switching from the standard culture 129 medium to drug-containing ones, population growth rates started to decrease (Fig 3A). 130 As expected, exposure to a higher dose (200 nM) of the drug resulted in a more 131 significant decline in division probability and earlier cessation of the entire population 132 growth than exposure to a lower dose (50 nM) (Fig 3A). We also observed increases in 133 death rates in response to the drug treatment, but with a noticeable delay  The time points at which the different cell lineages died were largely heterogeneous 142 (Fig 3C and D). Since the division frequencies of the individual cell lineages were also 143 heterogeneous ( Fig 1D, 2E, 3C, and 3D), we wanted to examine whether pre-exposure 144 division frequencies correlated with susceptibility to MMC. Therefore, we categorized 145 the cell lineages into two groups: (i) those that underwent seven or more cell divisions 146 during the 96 hours preceding drug exposure and (ii) those that underwent six or fewer 147 cell divisions. The division count cut-off was determined based on the assumption that 148 cells with generation times longer than 14 h were in the slow-cycling state (see Materials 149 and Methods for details).

150
The cell lineages exhibiting the higher division frequency before drug exposure 151 stayed alive for the first 60 hours of exposure to 50 nM MMC, but the decay of their 152 surviving fraction accelerated thereafter (Fig 3E). On the other hand, we did not 153 observe such acceleration in the death kinetics for the slow-cycling lineages.

154
Consequently, the low and constant death rate allowed the slow-cycling cell lineages to 155 leave a larger fraction of surviving cells (ca. 40%) even after seven days of continuous 156 exposure to MMC (Fig 3E).  surviving fraction of the cell lineages exhibiting higher prior division frequencies decayed 159 more rapidly after an initial 40-hour endurance period (Fig 3F). These results suggest 160 that pre-exposure growth states affect the response to MMC treatment and subsequent 161 survival dynamics: Slow growth is detrimental to short-term survival, but advantageous 162 if the drug exposure persists.  Importantly, the cells present in the slow-cycling state died more gradually upon 181 exposure to MMC (Fig 3), which proves that the intrinsic growth heterogeneity 182 emerging in a constant environment is sufficient to produce variable susceptibility to the 183 drug. Since MMC targets DNA replication, this result might be consistent with the 184 simple view that the lack of DNA replication allows slow-cycling cells to withstand the 185 inhibitory effect by MMC. However, resistance to anticancer drugs may also result from 186 failure to undergo apoptosis or from increased efflux activity [34]. It is also plausible that, in slow-cycling cells, more intracellular resources are allocated to alternative 188 metabolic processes that alleviate the effects of the drug rather than to the processes of 189 growth and replication.

190
The trans-generational correlation of generation time of L1210 cells was previously 191 measured by Sandler et al. using a different experimental setup [30]. They reported 192 that the correlation between mother and daughter cells' generation times was nearly  The existence of a small subset of slow-cycling or non-growing quiescent cells has 205 been described in various types of cancer cell populations. The contribution of such cells 206 to anticancer drug resistance and the formation of new tumors is of great concern in 207 modern cancer biology [35][36][37][38]. Although the detection of slow/non-growing cells often 208 relies on indirect methods such as the label-retaining assay [35,37], our microfluidic 209 lineage-tracking strategy provided direct evidence of pre-existing slow-cycling L1210 210 leukemia cells that last longer upon subsequent drug exposure. It is of note that 211 generation times of the slow-cycling leukemia cells characterized in this study were less 212 than 40 h in most cases; thus, they should still be referred to as proliferative rather 213 than quiescent cells. It is an intriguing question whether the slow-cycling phenotype is a 214 prerequisite for the acquisition of stronger drug resistance, possibly mediated by drastic 215 changes in metabolic states or triggering adaptive mutations [39,40].  The soft bake was carried out at 95 C for an hour. The trench channel was created with 257 a mask aligner such that the trench and the growth channels were orthogonal to each 258 other. The post-exposure bake, development, and rinsing were carried out as described 259 above.

288
Data processing and statistical analysis 289 We tracked the images manually using ImageJ [41] to record cell division and death (1  m  n) by constructing m-dimensional vector T m (j) = (T 1 , T 2 , · · · , T m ), using the 321 first m entries of T(j). A set of points T m (j) represents an attractor of a dynamical 322 system that rules the generation of the time series, and its correlation dimension D can 323 be estimated with the correlation integral C m (r), which is defined as July 5, 2020 9/14 where N is the total number of surviving cell lineages, and I(·) is the Heaviside function 325 that takes either the value 0 or 1: If the generation time series are produced by a stochastic process, then, C m (r) ⇠ r m for 327 all m, whereas C m (r) ⇠ r D for m D if the process is deterministic. For each 328 embedded dimension m (1  m  9), the C m (r) of generation-time series data was 329 calculated for varying r, and log C m (r) was plotted against log r. For each plot, local 330 linear fitting on 300 successive data points was carried out by using ranges of five data 331 points (moving linear regression), and the maximum value of the slopes of the fitted 332 lines was reported as the estimated correlation dimension.

333
Supporting information 334 S1 Video. Growth and division of L1210 cells in the mammalian mother 335 machine microfluidic device (S1Video.mp4).