Cite this Article
Xiong Ling, Liu Yanping, Liu Ruchuan, Yuan Wei, Wang Gao, He Yi, Shuai Jianwei, Jiao Yang, Zhang Xixiang, Han Weijing, Qu Junle, Liu Liyu. Quantitative heterogeneity and subgroup classification based on motility of breast cancer cells. Chinese Physics B, 2019, 28(10): 108701  
Permissions
Quantitative heterogeneity and subgroup classification based on motility of breast cancer cells
Xiong Ling1, Liu Yanping1, 2, Liu Ruchuan1, Yuan Wei3, Wang Gao1, He Yi1, Shuai Jianwei2, Jiao Yang4, Zhang Xixiang5, Han Weijing6, Qu Junle3, ‡, Liu Liyu1, §
Chongqing Key Laboratory of Soft Condensed Matter Physics and Smart Materials, College of Physics, Chongqing University, Chongqing 401331, China
Department of Physics, Xiamen University, Xiamen 361005, China
Key Laboratory of Optoelectronic Devices and Systems of Ministry of Education and Guangdong Province, Shenzhen University, Shenzhen 518060, China
Materials Science and Engineering, Arizona State University, Tempe, Arizona 85287, USA
Division of Physical Science and Engineering, King Abdullah University of Science and Technology, Thuwal 23955-6900, Saudi Arabia
Shenzhen Shengyuan Biotechnology Co., Ltd, Shenzhen 518060, China

 

† Corresponding author. E-mail: jlqu@szu.edu.cn lyliu@cqu.edu.cn

Abstract

Cancer cell motility and its heterogeneity play an important role in metastasis, which is responsible for death of 90% of cancer patients. Here, in combination with a microfluidic technique, single-cell tracking, and systematic motility analysis, we present a rapid and quantitative approach to judge the motility heterogeneity of breast cancer cells MDA-MB-231 and MCF-7 in a well-defined three-dimensional (3D) microenvironment with controllable conditions. Following this approach, identification of highly mobile active cells in a medium with epithelial growth factor will provide a practical tool for cell invasion and metastasis investigation of multiple cancer cell types, including primary cells. Further, this approach could potentially become a speedy (∼hours) and efficient tool for basic and clinical diagnosis.

PACS: 87.19.xj;87.85.dh;87.85.gj
Keyword:cancer metastasis;motility heterogeneity;three-dimensional microenvironment;invasion potential
1. Introduction

As metastasis leads to the majority of human cancer-related deaths, tumor dissemination, including invasion and metastasis, is a great therapeutic challenge.[1, 2] During tumor cell metastasis, the sub-group cells must recognize the extracellular matrix barrier and cross the barrier to achieve distant proliferation.[3] In addition, cell motility has become a strategy for anti-tumor invasion and metastasis.[4] Therefore, cell subpopulation classification based on the motility of cells will help to identify essential cells in tumor cell metastasis. Due to the stimulation of epithelial growth factor (EGF) as a growth factor that promotes cell movement,[5–7] the cell subpopulations will change. Moreover, it has been reported that cell heterogeneity together with microenvironment heterogeneity are essential in cancer cell motility and metastasis.[8] Studies have also suggested that, besides characteristic phenotypes of tumor heterogeneity and other phenotypes, highly motile invasive cells present great heterogeneity in their motility.[9–11] In addition, the existence of heterogeneous cancer cells during the invasion process, e.g., leader and follower cells, was proven by both in vivo and in vitro experiments[12–14] and indicated that several related genes and pathways are involved.[15, 16] Therefore, an easily handled quantitative characterization/classification approach for tumor cell motility heterogeneity in 3D microenvironments[17] could provide deeper insights into the differences between cells within a tumor and promote highly mobile/invasive sub-group cell identification in cancer invasion, metastasis, and evolution. Although there are already some approaches to discriminate various cell types[18] and their separations,[19] it still remains a technical challenge to well characterize cell motility heterogeneity and identify cell subgroups due to great complexity in tumor cell heterogeneity and usually subtle differences among cells in each cell-line.

To identify sub-groups of cells with various mobility, we combined a microfluidic chip inside a 3D microenvironment, a cell-tracking technique, and an improved method to characterize velocity auto-covariance (VACV) and power spectra of tumor cells. Consequently, cell motility heterogeneity in a set of designed and easily controllable medium conditions, i.e., an EGF-supplemented medium, were analyzed. VACV and power spectra were effective factors to reveal cells with higher migration persistence, which likely included high-invasive leader cells[12, 20] in collective cancer cell invasion. As a demonstration, we applied this method to quantify and compare motility heterogeneity between high metastatic potential MDA-MB-231 and low metastatic potential MCF-7 breast cancer cells.

2. Materials and methods
2.1. Cell culture

Human breast carcinoma MDA-MB-231 (China Infrastruture of Cell Line Resource, Beijing, China) were cultured in DMEM (GIBICO, Life Tech) supplemented with 10% Fetal bovine serum (FBS). Breast cancer cell line MCF-7 (China Infrastruture of Cell Line Resource, Beijing, China) was cultured in MEM medium containing 10% FBS and 0.01-mg/mL insulin. All of the medium were containing 1% penicillin/streptomycin (Corning). These cells are cultured in a 37-°C incubator with 5% CO2.

2.2. Microchip fabrication and cell culturing in chip

Based on soft-lithographic technology, a polydimethylsiloxane (PDMS) chip in 300- depth was designed for incubation of live cells on a microscope. As shown in Fig. 1(a), the chip[8] consisted of three parallel PDMS channels on a glass substrate, with four medium reservoirs on top. Two micro-sized pillars arrays separated the side channels, where media with various components could be added for microenvironmental control and exchange of nutrition/signaling biomolecules. In experiments, the middle channel was used for cell injection in the final density of 2 mg/mL and the PH value of 7.2 Collagen I (Corning, #354236, U.S.A) adjusted by 1-mol/L NaOH (Fluka), which forms a 3D network microenvironment[21] after gelatinization for 30 min at 37 °C. Subsequently, two side channels were filled with medium or medium containing the final concentration 25-ng/ml EGF, and the entire chip was kept in an incubator with 5.0% CO2 at 37 °C during cell tracking.

Fig. 1. The microfluidic chip structure and cell classification steps. (a) Structure hierarchical diagram of the polydimethylsiloxane (PDMS) microfluidic chips. The PDMS scaffold (top layer) is bonded to a glass substrate (bottom layer), forming three channels (middle layer) separated by two arrays of micro-sized pillars (upper sides 200- long, bottom side 400- long, and high m) that center spacing 400 m. Tumor cells are injected into the middle channel (7200- long and 3000- wide) with a collagen solution, which gelatinizes after incubation for 30 min at 37 °C and 5.0% CO2. The two shoulder channels are supplied with culture medium pools in experiments. This is also illustrated in the middle-right insert. Bottom-right insert displays a representative photo of a microfluidic chip. (b) and (c) Images of cell movement in collagen (scale bar, ) and trajectory diagram, respectively. (d) and (e) Cell classification process diagram, based on VACV with different persistent times ( ) of cells and power spectra, respectively.
2.3. Cell tracking

Cells were monitored in the bright-field mode by an inverted fluorescence microscope Ti (Nikon, Tokyo, Japan) with 20×objective. Time lapse videos were captured using a charge-coupled device (CCD) camera (HAMAMATSU, MODEL C11440-22CU) head DS-Ri 1, and the interval time is 2 minutes. Data collection and imaging analysis were performed using the ImageJ (National Institutes of Health). The acquired images were then processed with ImageJ first and CellTracker[22] (Hungarian Academia of Sciences, Hungary) to acquire the trajectories of individual cells in the x–y plane. Therefore, the trajectories were the projections of the 3D motions, a simplified representation with the well characteristics of the latter. The coordinates data were obtained in the semi-automatic tracking model with two tunable parameters, the maximal cell displacement and the cell diameter. The displacement was chosen between and , and the cell diameter was chosen between and , for different tracking cell, which will have better tracking effect by changing the parameters multiple times.

3. Data analysis
3.1. Mean square displacement (MSD)

In order to quantify the differences, the average MSD versus time trajectories is calculated for both cells as shown in Fig. 2, and it is larger for MDA-MB-231 cells than MCF-7 cells in all respective media.

Fig. 2. Motility of the two breast cancer cell lines. (a) Trajectories of MDA-MB-231 breast tumor cells show a larger range of motion after EGF addition. (b)The average MSD vs. time for MDA-MB-231 cells and MCF-7 cells in EGF- and EGF+ medium conditions. (c)The speed histograms of MDA-MB-231 and MCF-7 cells in two medium conditions.

For a rough understanding of the ability of cells migration, we computed the MSD during a time-lapse of duration t, and which is defined in Eq. (1)

where , , and are, mean square displacement at time t, the position-vector at time t, and the position-vector at time 0, respectively. Because the cells move in the x–y plane, we could only record the x–y coordinates, the above equation (1) is changed as follows:
where and , and is the time interval of each frame, and denotes averaging over time t.

3.2. Migration speed

After obtaining migratory trajectory, we could directly estimate corresponding velocity vectors by

where is the position-vector of the cell at time t and is the time interval of each frame. According to Eq. (3), we could calculate migration speed

3.3. Velocity auto-covariance function

On the basis of velocity vectors, the velocity auto-covariance function is defined as follows:

where is the time lag between any two velocity vectors calculated from one migratory trajectory. Symbol denotes averaging over time t for individual cell.

3.4. Cells classification based on VACV

After obtaining the relationship between the velocity auto-covariance function and time, we have found that there are two kind of exponential decays of auto-covariance function, one is bi-exponential decay, another mono-exponential decay for the same type of cells in same micro-environment. To determine which decay mode the velocity auto-covariance function obeys, we take a few steps as follows: first, fitting experimental velocity auto-covariance function of all cells with Eqs. (6) and (7)

and obtaining motile parameters and , respectively (the meaning of these parameters is given in table 1), second, setting thresholds of physical quantities computed from parameters derived from fits of velocity auto-covariance function, then, according to these thresholds, classifying population of cells into two types (the detail process is given in table 2). If velocity auto-covariance function of individual cell obeys bi-exponential decay, we called this cell as active cell, otherwise, we called normal cell.

Table 1.

The meaning of the obtaining motile parameters.

.
Table 2.

The thresholds of physical quantities.

.
3.5. Power spectrum

Due to the highly correlation of values of auto-covariance function in time, a least-squares fit to those data does not return reliable estimates, what’s more, fits of VACV function cannot return migration speeds of cells. However, the power spectrum can be decoupled from time and not only return persistent time P but also return the speed of cell migration, thus a fit to power spectrum can make up the defect and return reliable values.

According to Wiener–Khinchin theorem, the power spectrum of velocities of cell migration is the Fourier transformation (FT) of the velocity auto-covariance function. There is a definition of discrete FT as follows:

where , , , , is the time interval of each frame, N is the total steps in the tracking trajectory. Thus
On the basis of velocity vectors derived from cell trajectories, we could compute experimental power spectrum that fitted by maximum likelihood method (See Supplementary material for detail) using above Eq. (9).

4. Results and discussions
4.1. EGF enhances the motility difference between high metastatic MDA-MB-231 cells and low metastatic MCF-7 cells

In order to study the motility of high-metastatic MDA-MB-231 cells and low-metastatic MCF-7 cells, the individual cell trajectories and the changes of MSD over time can be used to illustrate the average motility of the two types of cells [Fig. 2(b)]. Both the broader migration range and larger MSD of MDA-MB-231 cells (See Supplementary material, Fig. S1) clearly show the higher motility of this metastatic breast cancer cell line. In addition, the motility of MDA-MB-231 cells is significantly improved by the addition of EGF [Fig. 2(c)], which EGF is a growth factor that can induce tumor cell invasion. It is noteworthy that uniform EGF environment was constructed by introducing medium of the same concentration into the channels on both sides of the chip, which effectively avoided cell movement caused by EGF gradient.[23] Obviously, the average speed of MDA-MB-231 cells is raised in the presence of EGF, while the change of MCF-7 cells is almost unobservable, in consistent with their metastatic potential.

The velocity power spectrum is the Fourier transform of the VACV function, and is uncorrelated between various frequencies,[24] so is better for the quantification of cell motility and comparison. To determine factors in cell migration factor contributing to the high mobility of MDA-MB-231 cells, VACV and power spectrum of each cell is analyzed in details. Figure 3 illustrates the average VACVs and power spectra of two cell-lines in two medium conditions. Specifically, EGF addition differentiates the average VACVs between MDA-MD-231 and MCF-7 cells. For MDA-MB-231 cells, the average VACVs remains positive until 20 min in EGF-supplemented medium [Fig. 3(b)]. For MCF-7 cells, the average VACVs is approximately zero for two medium conditions [Figs. 2(a) and Fig. 2(b)]. This reflects the larger persistence time for MDA-MD-231 cell migrations, indicating it as an appropriate parameter to highlight the diverse mobility between these two cell-lines.

Fig. 3. Velocity auto-covariance (VACVs) and power spectra of two breast cancer cell-lines. (a) and (b) Average VACVs of MDA-MB-231 (blue) and MCF-7 (red) cells in two medium conditions, as indicated. The dashed line refers to t = 20 min. (c) and (d) The corresponding power spectra of MDA-MB-231 (blue) and MCF-7 (red) cells. Higher power in the low frequency part of the spectra is found for MDA-MB-231 cells in two medium conditions, and its highest power spectrum is found in EGF+ medium. The dashed line highlights the frequency of 0.02 min−1, which is referred to as the low frequency region in the discussion.

To further confirm and quantify the persistent time difference, velocity power spectra are analyzed. As displayed in Figs. 3(c) and 3(d), the average power spectra of two cell-lines demonstrate significant differences. At the low frequency region, i.e., min−1 (corresponding to long persistent time), power is higher for MDA-MB-231 cells, and the difference is enhanced with EGF addition. In contrast, the velocity power for MCF-7 cells shows limited changes with EGF addition and stays at a low level in the whole frequency region analyzed. Therefore, both VACV and power spectra well characterize distinct persistence in migration between two cell-lines. Moreover, they also showed that MDA-MB 231 cells were more active than MCF-7 cells.

4.2. There are subpopulations in each cell line showing different migration modes

Above results clearly display differences in motility by average between two cell lines. It is necessary to analyze the VACV of single cells to look into the details in their motility variations. At the individual cell level, persistence displays great heterogeneity, which could be used as a characteristic for cell mobility heterogeneity. In contrast to clear differences in average trajectories, significant overlap in VACV and power trajectories of individual cells smears the difference between cell-lines. Even for the same cell line, these trajectories from various medium are also overlapping. Nevertheless, we believe that the broad distributions of VACV and power spectra for individual cells could become one of the aspects to characterize heterogeneous cell mobility. Analysis of individual VACV trajectories reveals two behavioral types. Some VACV trajectories could be well fitted by mono-exponential decay, and the others are better fitted by bi-exponential decay (see data analysis section for detailed fitting protocol).

Figure 4(a) and 4(b) present typical mono- and bi-exponential VACV trajectories, where bi-exponential trajectories display an additional component with a persistent time longer than that for mono-exponential trajectories. According to this aspect, individual cells are categorized into two subgroups: normal cells (mono-exponential VACV) and active cells (bi-exponential VACV). The general steps of cell subpopulation separation are shown in Figs. 1(b)–1(e). Normal cells show a fast decay to zero in their average VACVs, regardless of cell-line [Figs. 4(c) and 4(d)]. In contrast, average VACVs remain positive for ∼20 min and ∼10 min for active groups of MDA-MB-231 and MCF-7 cells in EGF+ medium, respectively. Furthermore, the average power spectra [Figs. 4(e) and 4(f)] for normal cells from both cell-lines remain at a low level over the whole frequency range, while those for active cells are well separated from normal cells with significant higher power in the low frequency region (i.e., min−1). Similar trends in average VACVs and power spectra for each cell group have been obtained for other medium, regardless of cell-lines (See Supplementary material, Fig. S3). This validates the above categorization of cell subgroups by persistence heterogeneity.

Fig. 4. Differentiation of active cells from normal cells by VACVs. (a) and (b) Typical VACV trajectories of individual MDA-MB-231 and MCF-7 cells: active cells (blue or red circles) and normal cells (green crosses), and the fitting curves (solid lines). (c) and (d) Comparison between active and normal cells in EGF+ medium: average VACVs of active cells (red or blue) and normal cells (green), compared to average VACV (black) of all cells, for MDA-MB-231 and MCF-7 cells, respectively. (e) and (f) Corresponding power spectra for MDA-MB-231 and MCF-7 cells. These VACVs and power spectra show similar trends for normal cells of both cell-lines.
4.3. The heterogeneity in motility positively correlates to the invasive potential of the two cancer cell lines

To further verify cell categorization and to characterize migration persistence, the power spectrum of each cell is fitted following the reported protocol[24] (See Supplementary material, Fig. S4). Figure 5(a)–5(c) and 5(d)–5(f) show histograms of persistent time from power spectra fitting for MDA-MB-231 and MCF-7 cells in EGF+ medium, respectively. Clearly, persistent times for normal cells demonstrate similar narrow-ranged distributions within the experimental time resolution (2 min), irrespective of cell types. This suggests a similar limited memory in the motion of these cells. For active cells, two distinct persistent times are identified. The other persistent time P2 (See Supplementary material, Fig. S5) is longer and indicative of “active” cells, i.e., advancing much further with a better persistence than normal cells. Consequently, the migration mode of P2 is expected to contribute significantly to the higher motility potential of invasive cancer cells. These results demonstrate that aforementioned classification of cells into two subgroups is a successful step toward clarifying their mobility heterogeneity.

Fig. 5. Persistent time of cells in EGF+ medium with and ratio of active cells. (a)–(f) Histograms of persistence times from maximum likelihood estimation fitting of power spectra for MDA-MB-231 (blue) and MCF-7 (red) cells: (a) and (d) persistent time of normal cells, (b) and (e) short persistent time component of active cells, (c) and (f) long persistent time component of active cells. (g) Ratio of active cells in four media types for MDA-MB-231 (blue) and MCF-7 (red) cells.

Next, we investigated differences in subgroup ratios between the two cell-lines. Figure 5(g) illustrates that active cells generally occupy at least 40% of the overall cells in two medium conditions. The ratio of active cells (ract) is consistently higher for high-invasive MDA-MB-231 cells in all tested medium, i.e., ract close to or above 60%. In addition, the ract ( ) for MDA-MB-231 cells in EGF-supplemented medium is higher than (∼60%) without EGF. In contrast, the ract is only around 40% for MCF-7 cells and does not rise upon EGF addition. Furthermore, in EFG+ culture medium, more cells in the MDA-MB-231 active subgroup possess a P2 longer than 10 min [Figs. 5(c) and 5(f)]. This suggests ract as a good quantitative parameter for mobility heterogeneity among each cell line. Additionally, it quantifies differences in persistence time between these two cell-lines.

Reasonably, ract could be positively correlated to the metastatic potential of cancer cells, potentially making it a useful reference. It has been reported that in collective invasion of MDA-MB-231 cells, leader and follower cells co-exist.[12, 25] The leader cells are expected to possess higher motility and better directionality in comparison to follower cells. Obviously, better persistence increases the chance of tumor cells to migrate further from the original site, and thus extraordinarily mobile leader cells are likely to evolve from the active cell subgroup. In this way, cancer cell heterogeneity in migration persistence could significantly influence invasion processes, and thus the ratio of active cells (ract) is an essential indicator. At the same time, the above approach to obtain the ratio ract has the following advantages: (i) while it evaluates mobility and heterogeneity of cells, ract focuses more on active cells with a large migration persistent time, i.e., potential leader cells in collective invasion; (ii) ract provides a relative scale from 0 to 100% that could be useful in comparison across cell-lines and types; (iii) the absolute ract is independent of other cells; (iv) acquisition of ract is fast (as short as 4 hours), in contrast to no less than 24 hours in conventional methods to evaluate cancer cell invasiveness; (v) the approach could identify active cells while keeping them alive for further investigations, including, but not limited to, the molecular mechanisms (e.g. cell contractive force, cell adhesion to collagen, secreted protease, etc.) of cancer cell invasion,[16, 26–28] impact of physiological conditions, and environmental factors (e.g. growth factor gradient, matrix fiber orientation, etc.[29, 30]

5. Conclusion

In order to quantify the motility of breast tumor cells and identify the subgroup of more mobile cells (potentially high invasive) for cancer metastasis investigation and future clinic application, we developed a rapid and unique approach in combination of the microfluidic chip, 3D tracking analysis and cell sub-group identification to distinguish the higher invasive MDA-MB-231 and the low invasive MCF-7 cells by their motility and heterogeneity. Significant heterogeneity among the motility of individual cells in each cell group is successfully quantified by the ratio ract of active cell subgroup identified with our approach. It turns out that the EGF is a good promoter to differentiate active cells from normal cells and enhances the ratio ract of high-metastatic potential cells, but not low- metastatic potential cells. The higher population of active MDA-MD-231 cells together with its much enhanced motility and persistence, are positively correlated to the high invasiveness of MDA-MD-231 cells in contrast to MCF-7 cells. Thus, the method of identify active cell population, ract, in an EGF+ media introduces a new, rapid and effective way to evaluate the invasive potential of cancer cells. In addition, the separation of active motile cells from normal cells also provides a well-defined approach for cancer invasion and metastasis investigation. We believe that the above approaches may apply to primary tumor cells in clinics and potentially provide physicians with a quantitative reference.

Reference
[1] Sleeman J Steeg P S 2010 Eur. J. Cancer 46 1177 https://doi.org/10.1016/j.ejca.2010.02.039
[2] Steeg P S Theodorescu D 2008 Nat. Clin. Pract. Oncol. 5 206 https://doi.org/10.1038/ncponc1066
[3] Wells A 2000 Adv. Cancer Res. 78 31 https://doi.org/10.1016/S0065-230X(08)61023-4
[4] Wells A Grahovac J Wheeler S Ma B Lauffenburger D 2013 Trends Pharmacol. Sci. 34 283 https://doi.org/10.1016/j.tips.2013.03.001
[5] Wells A Kassis J Solava J Turner T Lauffenburger D A 2002 Acta Oncol. 41 124 https://doi.org/10.1080/028418602753669481
[6] Gao H X Chen X J Du X N Guan B C Liu Y N Zhang H L 2011 Cell Calcium 50 559 https://doi.org/10.1016/j.ceca.2011.09.003
[7] Garcia R Franklin R A McCubrey J A 2006 Cell Cycle 5 2820 https://doi.org/10.4161/cc.5.23.3535
[8] Han W J Chen S H Yuan W Fan Q H Tian J X Wang X C Chen L Q Zhang X X Wei W L Liu R C Qu J L Jiao Y Austin R H Liu L Y 2016 Proc. Natl. Acad Sci. USA 113 11208 https://doi.org/10.1073/pnas.1610347113
[9] Giampieri S Manning C Hooper S Jones L Hill C S Sahai E 2009 Nat. Cell Biol. 11 1287 https://doi.org/10.1038/ncb1973
[10] Marusyk A Polyak K 2010 Biochim. Biophys. Acta-Rev. Cancer 1805 105 https://doi.org/10.1016/j.bbcan.2009.11.002
[11] Wu P H Giri A Sun S X Wirtz D 2014 Proc. Natl. Acad Sci. 111 3949 https://doi.org/10.1073/pnas.1318967111
[12] Liu L Duclos G Sun B Lee J Wu A Kam Y Sontag E D Stone H A Sturm J C Gatenby R A 2013 Proc. Natl. Acad Sci. 110 1686 https://doi.org/10.1073/pnas.1221147110
[13] Tsuji T Ibaragi S Hu G F 2009 Cancer Res. 69 7135 https://doi.org/10.1158/0008-5472.CAN-09-1618
[14] Gligorijevic B Bergman A Condeelis J 2014 PLOS Biol. 12 e1001995 https://doi.org/10.1371/journal.pbio.1001995
[15] Wang W Goswami S Sahai E Wyckoff J B Segall J E Condeelis J S 2005 Trends Cell Biol. 15 138 https://doi.org/10.1016/j.tcb.2005.01.003
[16] Friedl P Wolf K 2003 Nat. Rev. Cancer 3 362 https://doi.org/10.1038/nrc1075
[17] Han W J Yuan W Zhu J R Fan Q H Qu J Liu L Y Sci-Oncology U-C P 2016 Chin. Phy. B 25 018709 https://doi.org/10.1088/1674-1056/25/1/018709
[18] Chen P Zhang L Zhang F Liu J Bai H Tang G Lin L 2012 Biomed. Opt. Express 3 1787 https://doi.org/10.1364/BOE.3.001787
[19] Lee K H Lee K S Jung J H Chang C B Sung H J 2013 Appl. Phys. Lett. 102 141911 https://doi.org/10.1063/1.4801951
[20] Cheung K J Gabrielson E Werb Z Ewald A J 2013 Cell 155 1639 https://doi.org/10.1016/j.cell.2013.11.029
[21] Sun B 2015 Acta Phys. Sin. 64 058201 (in Chinese) https://doi.org/10.7498/aps.64.058201
[22] Piccinini F Kiss A Horvath P 2016 Bioinformatics 32 955 https://doi.org/10.1093/bioinformatics/btv686
[23] Wang S J Saadi W Lin F Nguyen C M C Jeon N L 2004 Exp. Cell Res. 300 180 https://doi.org/10.1016/j.yexcr.2004.06.030
[24] Pedersen J N Li L Gradinaru C Austin R H Cox E C Flyvbjerg H 2016 Phys. Rev. E 94 062401 https://doi.org/10.1103/PhysRevE.94.062401
[25] Chapman A del Ama L F Ferguson J Kamarashev J Wellbrock C Hurlstone A 2014 Cell Rep. 8 688 https://doi.org/10.1016/j.celrep.2014.06.045
[26] Fife C M McCarroll J A Kavallaris M 2014 Br. J. Pharmacol. 171 5507 https://doi.org/10.1111/bph.12704
[27] Levental K R Yu H Kass L Lakins J N Egeblad M Erler J T Fong S F T Csiszar K Giaccia A Weninger W Yamauchi M Gasser D L Weaver V M 2009 Cell 139 891 https://doi.org/10.1016/j.cell.2009.10.027
[28] Wolf K Wu Y I Liu Y Geiger J Tam E Overall C Stack M S Friedl P 2007 Nat. Cell Biol. 9 893 https://doi.org/10.1038/ncb1616
[29] Joyce J A Pollard J W 2009 Nat. Rev. Cancer 9 239 https://doi.org/10.1038/nrc2618
[30] Haeger A Krause M Wolf K Friedl P 2014 Bba-Gen. Subjects 1840 2386 https://doi.org/10.1016/j.bbagen.2014.03.020