Research ArticleBIOPHYSICS

Live-cell imaging and analysis reveal cell phenotypic transition dynamics inherently missing in snapshot data

See allHide authors and affiliations

Science Advances  04 Sep 2020:
Vol. 6, no. 36, eaba9319
DOI: 10.1126/sciadv.aba9319


Recent advances in single-cell techniques catalyze an emerging field of studying how cells convert from one phenotype to another, in a step-by-step process. Two grand technical challenges, however, impede further development of the field. Fixed cell–based approaches can provide snapshots of high-dimensional expression profiles but have fundamental limits on revealing temporal information, and fluorescence-based live-cell imaging approaches provide temporal information but are technically challenging for multiplex long-term imaging. We first developed a live-cell imaging platform that tracks cellular status change through combining endogenous fluorescent labeling that minimizes perturbation to cell physiology and/or live-cell imaging of high-dimensional cell morphological and texture features. With our platform and an A549 VIM-RFP epithelial-to-mesenchymal transition (EMT) reporter cell line, live-cell trajectories reveal parallel paths of EMT missing from snapshot data due to cell-cell dynamic heterogeneity. Our results emphasize the necessity of extracting dynamical information of phenotypic transitions from multiplex live-cell imaging.


Cells of a multicellular organism can assume different phenotypes that can have markedly different morphological and gene expression patterns. A fundamental question in developmental biology is how a single fertilized egg develops into different cell types in a spatialtemporally controlled manner. Cell phenotypic transition (CPT) also takes place for differentiated cells under physiological and pathological conditions. A well-studied example is the epithelial-to-mesenchymal transition (EMT), central to many fundamental biological processes, including embryonic development and tissue regeneration, wound healing, and disease-like states such as fibrosis and tumor invasiveness (1). One additional example is artificially reprogramming differentiated cells, such as fibroblasts, into induced pluripotent stem cells and other differentiated cell types such as neurons and cardiomyocytes (2). CPT is ubiquitous in biology, and a mechanistic understanding of how a CPT proceeds emerges as a focused research area with an ultimate goal of achieving effective control of the phenotype of a cell.

Recent advances in snapshot single-cell techniques, including single-cell RNA sequencing (scRNA-seq) and imaging-based techniques, pose new questions. One such question is, how does a CPT process proceed, step by step, in the continuously changed high-dimensional gene expression (e.g., transcriptome and proteome) space? These destructive methods, however, are inherently unable to reveal the temporal dynamics of how an individual cell evolves over time during a CPT.Approaches such as pseudo-time trajectory analysis (3), ergodic rate analysis (4), and RNA velocities (5) have been developed to retrieve partial dynamical information from snapshot data.

However, some fundamental limits exist in inferring dynamical information from snapshot data (6). Figure 1 illustrates that inference from snapshot data unavoidably misses key dynamical features. A bistable system is coupled to a hidden process, e.g., epigenetic modification for a cellular system, which is slow compared to the transition process being studied (Fig. 1A). Presence of the slow variable leads to observed heterogeneous dynamics (Fig. 1B, more details of the simulation are in fig. S1). Individual trajectories show characteristic stepping dynamics, but the transition positions vary among different trajectories due to the system assuming different values for the hidden variable. Consequently, when only snapshot data are available, information about the temporal correlation of individual cell trajectories is missing, and one cannot deduce the underlying two-state dynamics (Fig. 1C).

Fig. 1 Hidden variables can conceal some dynamical features from snapshot data.

(A) A double well potential with one observable coupled to a hidden variable with dynamics much slower than that of the process under study. Arrows represent simulations of two trajectories that start from well A and jump to B. The black line represents the boundary of the two wells. (B) One-dimensional (1D) potential slices along the observable coordinate corresponding to the transitions indicated in (A). (C) Superimposed trajectories from stochastic simulation (shown as background), with two typical trajectories highlighted corresponding to the transitions in (A). The hidden variables for the green trajectory and the magenta trajectory take the values −0.47 and 0.45, respectively. a.u., arbitrary units. (D) Stacked histograms of the observable at various time points [t0, t1, t2, t3, and t4 in (C)], reflecting the snapshot data. Blue bars were sampled from points that reside in well A. Red bars were sampled from points that reside in well B. Notice that the information as to which well a sampling point resides is obtained from the full 2D potential and cannot be accurately derived from the snapshot data.

The necessity of live-cell trajectories has been illustrated in a number of studies, such as the information capacity of a signal transduction network (7), incoherent feed-forward loops to detect only fold change but not the absolute change of an input signal (8), and stepwise cellular responses to drugs (9). Two recent studies conclude a linear path for EMT from analyzing single-cell RNA-seq and proteomic data (10, 11). Discrepancy, however, exists between this conclusion and theoretical predictions of parallel paths for a multistable system like EMT (12, 13), raising a question whether the theoretical models need to be modified or the inferred linear path is an artifact from snapshot data. Live-cell imaging is needed to address such a question.

Therefore, acquiring information from long-term CPT dynamics requires tracking individual cells through live-cell imaging, typically with time-lapse fluorescent imaging. However, identifying appropriate molecular species that faithfully reflect the process for labeling and generating such labeling can be tedious and time-consuming. In addition, multiplex and frequent fluorescent image acquisition over a long period of time, e.g., days, is necessary for characterizing a CPT process but is severely limited by the number of available fluorescence channels and cytotoxicity concerns.

In short, a technical dilemma exists: Fixed cell–based techniques provide high-dimensional expression profiles of individual cells but lack true dynamical information, while fluorescent labeling–based live-cell imaging techniques typically provide dynamical information for only a small number of dynamical variables. To tackle the substantial challenges in CPT studies, here, we develop a framework for extracting cell dynamical information through quantitative analysis of live-cell trajectories in a high-dimensional cell feature space. Our method resides on the observation that a cell state can be defined by either the expression pattern or cell morphological features. The latter broadly refers to collective cellular properties such as cell body shape, organelle distribution, etc., which are convenient for live-cell imaging, with and without labeling. Hundreds of these morphological features have been routinely used in pathology and in a number of fixed- and live-cell studies for defining and studying cell phenotypes and drug responses (9). Introduction of this framework allows one to study CPTs in the context of well-established rate theories (14). Rate theories study dynamical processes of escaping from a metastable attractor or relaxing to a newly established attractor.

We applied this framework to study transforming growth factor–β (TGF-β)–induced EMT in a human A549 derivative cell line with endogenous vimentin–red fluorescent protein (VIM-RFP) labeling [American Type Culture Collection (ATCC) CCL-185EMT]. We represented the state of a cell at a given time as a point (or equivalently a vector) in a 309-dimensional composite feature space of the cell body contour shape and texture features of vimentin, an intermediate filament, and a key mesenchymal marker. While the framework is for morphological features in general, in this study, we focus on the cell body shape and use cell shape and morphology indistinctively. Through quantifying time-lapse images, aided by a deep learning–based image analysis algorithm, we were able to define the epithelial and mesenchymal regions in the state space and unravel two parallel pathways that EMT proceeds through. We provide a Python package, multiplex trajectory recording and analysis of cellular kinetics (M-TRACK), for studying CPT in the composite feature space. The framework will provide a foundation for quantitative experimental and theoretical studies of CPT dynamics.


Human A549 cells with endogenous VIM-RFP labeling were generated using the CRISPR-Cas9 system

The A549 VIM-RFP cell line (ATCC CCL-185EMT) was generated as a model system for studying EMT. It was created using CRISPR-Cas9 technology, in which the RFP sequence was inserted just before the stop codon of one allele of the endogenous vimentin gene (Fig. 2A and fig. S2A). The VIM-RFP knock-in allele was confirmed by sequencing (fig. S2B) and colocalization immunostaining of vimentin and VIM-RFP (Fig. 2B) with a Pearson correlation coefficient 0.9 and a Manders overlap coefficient 0.93.

Fig. 2 The A549 VIM-RFP cell line (ATCC CCL-185EMT) is designed for studying EMT dynamics.

(A) Schema of CRISPR-Cas9–mediated generation of VIM-RFP knock-in allele. LHA, left homology arm; RHA, right homology arm; pA, polyadenylation signal; BSR, blasticidin resistance; exon9, exon #9 of vimentin gene. (B) Colocalization immunostaining of vimentin and endogenous VIM-RFP. Scale bar, 20 μm. (C) Matrigel invasion assay of A549 VIM-RFP cells. After TGF-β induction, cells show increased invasive capacity. Scale bar, 100 μm. Data represent mean ± SD from three repeated experiments. Student’s t test; asterisk denotes P < 0.01.

EMT studies reach a consensus that there is a continuous spectrum of EMT phenotypes (1). Previous studies report that the parental A549 cells have already undergone partial EMT with detectable basal vimentin expression and can complete a full EMT upon TGF-β treatment (11). Consistently, with recombinant human TGF-β1 treatment, the A549 VIM-RFP cells showed an increased invasive capacity reflecting the functionality of mesenchymal cells (Fig. 2C) and increased Snail1 and N-cadherin expressions (fig. S2C). After 2 days of continuous treatment, most cells underwent apparent cell shape changes from round polygon shapes to elongated spear shapes, characteristic changes of EMT. These results confirmed that under TGF-β1 treatment, the A549 VIM-RFP cells completed a full EMT to convert from a partial EMT to a complete mesenchymal phenotype, and this is the process that we focused on in this study.

A complete orthonormal basis set represents cell body contour shapes

We performed time-lapse imaging on the A549 VIM-RFP cells continuously treated with TGF-β1 for 2 days, with an imaging frequency of one frame per 5 min for morphology and one frame per 10 min for vimentin (more details in Materials and Methods). To mathematically describe how a CPT (EMT here) proceeds, one needs to choose a mathematical representation of cell status at a given time. For representing the cell shape, we adopted the active shape model that has been widely used in computer-based image analyses (15) and particularly, in cell biology studies (16, 17), but here, we use it for the purpose of forming a complete orthonormal basis set (Fig. 3). That is, we first segmented the images using our modified deep convolutional neural network procedure (Fig. 3A, see Supplemental Text for details) and tracked individual cell trajectories. Each cell shape was aligned to a reference shape and was approximated by N (=150) landmark points equally spaced along the cell contour (Fig. 3B) (18). For two-dimensional (2D) images, a cell was specified by a point z = (x1, x2, …, xN; y1, y2, …, yN) in the 2N-dimensional morphology space (Fig. 3C). By performing principal components analysis (PCA) on the dataset of a collection of single-cell trajectories, one constructs a complete orthonormal basis set with the 2N − 4 eigenvectors {a}, or the principal modes, for the morphology space. Notice that alignment fixes four degrees of freedom (center and orientation). An attractive feature of the principal modes is that they have clear physical meanings. For example, the two leading principal modes of A549 VIM-RFP cells undergoing EMT reflect cell growth along the long and short axes, respectively (Fig. 3D). Then, any cell shape that is approximated by the landmark point z′(t), which is generally time dependent in live-cell imaging, can be expressed as a linear combination of these principal modes z(t)=i=12N4ci(t)ai(Fig. 3E). Therefore, c(t) = (c1(t),…, c2N−4(t)) forms a trajectory in the shape space expanded by the basis set {a}.

Fig. 3 Single-cell trajectories are quantitatively represented in the morphology space.

(A) Segmentation of single cells with the deep convolution neural networks (DCNN)/watershed method. (B) Extraction of cell outline from segmented mask of individual cells, followed by resampling using the active shape model and aligning all cell outlines to a mean cell shape for consistent digitization of cell morphology features. (C) Representation of single-cell shapes as points in the 2N – 4–dimensional morphology space. Each dot represents one cell. (D) Principal modes of morphology variation. Left: first principal mode (PC1). Right: second principal mode (PC2). The ±1 represents the corresponding coordinate value on the axis of morphology PC1 or PC2. The principal modes reflect the characteristics of cell morphology variation along the PC axes. (E) A representative cell shape (left) and its reconstruction with the first seven leading principal modes. (F) A typical single-cell trajectory in the two leading morphology PC domains (left) and its corresponding contours (triangle dots marked by arrows in the left that have the same color as the contours) at various time points (right). Each dot represents an instantaneous state of the cell in the morphology space. Color bar represents time (unit in hour).

Figure 3F shows a typical trajectory projected to the first two leading principal component (PC) modes in the morphology space and their corresponding time courses of cell contour shape changes. Over time, this cell elongated along the major axis (PC1), while shortened slightly along the minor axis (PC2), resulting in a long rod shape with an enlarged cell size. Two additional trajectories in fig. S3 further reveal that single-cell trajectories are heterogeneous with switch-like or continuous transitions while sharing similar elongation of PC1 over time.

Haralick features quantify texture feature change of cytosolic distribution of vimentin during EMT

During a CPT, cell morphology changes are accompanied by global changes in gene expression profiles (19). Specifically, in A549 VIM-RFP cells that have been treated with TGF-β for 2 days, vimentin was up-regulated with texture change from being condensed in certain regions of the cell to be dispersed throughout the cytosol (Fig. 4A), consistent with previous reports (2022). These previous studies used fixed cells and lack temporal information about the vimentin dynamics. Therefore, we recorded the change in vimentin within individual cells with time-lapse imaging.

Fig. 4 Haralick features quantify vimentin texture.

(A) Flowchart of quantification. Left: typical vimentin fluorescence images of single cell before (top) and after treatment of TGF-β (4 ng/ml) for 2 days (bottom; scale bar, 50 μm). Middle: segmented single-cell image with only pixels inside the cell mask kept for Haralick feature calculations. Right: framework of calculating the single-cell Haralick features. All pixel values inside an individual segmented cell were used to calculate GLCM and Haralick features. (B) A typical single-cell trajectory on the plane of vimentin Haralick features PC1 and PC2 (left) and on the plane of PC3 and PC4 (right). Color bar represents time (units in hours). Large triangular dots labeled with numbers correspond to 0, 12, 24, 36, and 48 hours, respectively. Each dot represents an instantaneous state of the cell in the Haralick feature space. (C) Segmented single-cell images of vimentin at various time points corresponding to the trajectory in (B) (labeled with large, triangular dots).

Texture features are widely used for image profiling in drug screening, phenotype discovery, and classification (2325). We hypothesized that the texture features of vimentin can be quantified as an indicator of EMT progression. For quantification, we used Haralick features on the basis of the co-occurrence distribution of gray levels (see Materials and Methods for detailed information of Haralick features used). After segmentation, we calculated the gray-level co-occurrence matrix (GLCM) in the mask of each cell and 13 Haralick features based on the single-cell GLCM and averaged all four directions (Fig. 4A, see Materials and Method for details). Nearly every Haralick feature shows a shift in distribution after TGF-β treatment (fig. S4).

To capture the major variation in vimentin Haralick features during EMT, we performed linear dimension reduction with PCA (more details are in Materials and Methods). Figure 4 (B and C) shows a typical single-cell trajectory in the vimentin Haralick feature space and corresponding segmented single-cell images at various time points along this trajectory, respectively. The dynamics in the vimentin feature space are again heterogeneous, as indicated here, and from two additional trajectories in fig. S5.

A label-spreading function divides the composite feature space into three regions.

Overall, we described a cell state in a 309-dimensional composite feature space of cell morphology and vimentin texture features. The cells occupy distinct regions in the composite feature space at the initial (0 to 2 hours) and final stages (46 to 48 hours) of TGF-β treatment (4 ng/ml) (Fig. 5A). Physically, upon TGF-β treatment, the cell population relaxes from an initial stationary distribution in the composite feature space into a new one, and this study focuses on the dynamics of this relaxation process.

Fig. 5 Single-cell EMT states are defined in the composite feature space.

(A) Kernel density plots in the plane of morphology PC1 and vimentin Haralick feature PC1 estimated from 3567 single-cell states from 196 trajectories (represented by dots), at 0 to 2 hours (top, 1920 single-cell states) and 46 to 48 hours after addition of TGF-β (bottom, 1647 single-cell states). (B) Distributions of cells at 0 to 2 hours and 46 to 48 hours cells after adding TGF-β in various features (morphology PC1, vimentin Haralick PC1, PC3, and PC4). The diagonal axes are plots of kernel density estimation of the 1D distribution of the corresponding features. The four PC modes capture 82.7% of morphology variance, 57.0, 8.8, and 5.2% of total vimentin feature variance, respectively. The distributions along vimentin PC2 for cells before and after treatment largely overlap with each other and were not included here. A complete combination of distributions is shown in fig. S6. (C) Scatter plot of 0- to 2-hour data (left) and 46- to 48-hour data (right) on the plane of morphology PC1 and vimentin Haralick features PC1. Color represents the region a cell resides at a given time point, as predicted by the fitted label-spreading function. E, epithelial region; I, intermediate region; M, mesenchymal region. (D) A single-cell trajectory with its residing regions predicted by the label-spreading function on the domain of morphology PC1 and vimentin Haralick features PC1 and PC3 (with regions represented by different colors).

Close examination reveals major distribution shifts along four coordinates: morphology PC1 (82.7% variance of morphology), vimentin Haralick PC1 (57.0% variance), PC3 (8.8% variance), and PC4 (5.2% variance) (Fig. 5B and fig. S6). This observation permits subsequent analyses that are restricted to these collective coordinates.

In rate theories, to define a reactive event, one typically divides the configuration space describing a reaction system into reactant, intermediate, and product regions (26, 27). Specifically, for the present system, we developed a computational procedure that combines the Gaussian mixture model (GMM) analysis of the cell distributions (fig. S7, A and B) followed by fitting a label-spreading function (28) using the k-nearest neighbor (KNN) method (Material and Methods). This label-spreading function takes coordinates (a multidimensional vector) of individual cells in the composite feature space as input and predicts the label of a cell, i.e., the region identity (E, I, or M discussed below) in which it resides (see Material and Methods for details). Combining the biology characteristics of EMT, this procedure divides the four-dimensional space into epithelial (E; or more precisely partial E for A549 VIM-RFP), intermediate (I), and mesenchymal (M) regions (Fig. 5C). Figure 5D shows a trajectory that starts within the E region and then progresses to the I and then to the M regions.

Single-cell EMT trajectories follow distinct paths

Consistent with a standard definition of reactive events in rate theories, we defined an ensemble of reactive trajectories as all the single-cell trajectories similar to the one in Fig. 5D (and fig. S7C) that leave the E region and end in the M region before returning to E. Overall, we recorded NT (=196) acceptable continuous trajectories (see Materials and Methods); among them, NR (=139) are reactive trajectories (movies S1 and S2).

Single-cell trajectories in the composite feature space show clear heterogeneous transition dynamics. In one representative trajectory (Fig. 6A and fig. S8A, left), the cell transits from the E to the M region following a series of transitions first along the vimentin Haralick PC1 and then the morphology PC1. In contrast, in another trajectory (Fig. 6B and fig. S8A, right), the cell proceeds with concerted morphological and vimentin Haralick feature changes.

Fig. 6 Single-cell trajectory analyses reveal parallel paths of EMT.

(A) A typical single-cell trajectory in which the major change along the vimentin Haralick PC1 precedes the major change along the morphology PC1 (class I). (B) A typical single-cell trajectory in which the morphology PC1 and vimentin Haralick PC1 show concerted variation (class II). (C) Projection of recorded 139 reactive trajectories on 2D t-SNE space using the DTW distances. Each dot is a single-cell trajectory that undergoes EMT. Color represents labels of k-means clustering on the DTW distances. (D) Mean trajectories of class I and II trajectories, respectively. They were calculated using the soft-DTW barycenter method. (E) Plausible mechanistic model. In this network, both morphology and vimentin changes are induced by TGF-β, and they both activate each other and themselves. (F) Transition paths simulated with the model in (E). The transition paths show dynamical characteristics similar to those observed experimentally.

To systematically study the two types of distinct behaviors, we used soft-dynamic time warping (DTW) (29) to calculate the distance between different trajectories, and t-distributed stochastic neighbor embedding (t-SNE) to project the trajectory distance matrix to 2D space (30). We found that these trajectories form two communities (Fig. 6C). A k-means clustering on these trajectories separated them into two groups, consistent with the two communities in the t-SNE space. The two groups of trajectories reveal different dynamical characteristics. In one group (class I), vimentin Haralick PC1 varies first, followed by marked change of morphology PC1. In the other group (class II), for most trajectories, the morphology PC1 and vimentin Haralick PC1 change concertedly, while for only a small percentage, the morphology PC1 changes earlier than vimentin Haralick PC1 (fig. S8B). Distinction between the two groups of trajectories is apparent from the scattered plot in the morphology PC1/vimentin Haralick PC1 plane (fig. S8C) and the nonoverlapping mean trajectories obtained using the soft-DTW barycenter (Fig. 6D and fig. S8D) (29, 31).

To rule out the possibility that the existence of two classes of trajectories is an artifact of DTW, we analyzed cross-correlation between morphology PC1 and vimentin Haralick PC1 of individual reactive trajectories. Cross-correlation analysis calculates the time delay at which the correlation between morphology PC1 and vimentin PC1 reaches a maximum value (32). The time delay shows a stretched distribution (fig. S8E). A large portion of trajectories have vimentin Haralick PC1 change before morphology PC1 change, while another main group of trajectories has the time delay between morphology PC1 and vimentin Haralick PC1 close to zero. After separating the trajectories into two groups based on the sign of time delay, the mean trajectories of the two groups (fig. S8E), referred to as transition paths that connect the initial and final cell states, are similar to what was obtained with k-means clustering on the DTW distance.

Therefore, a main conclusion of this study is that the live-cell platform revealed two types of paths for the TGF-β–induced EMT in A549 VIM-RFP cells. Figure 6E shows a plausible mechanistic model summarizing the existing literature (see Materials and Methods for details and references therein). TGF-β activates morphological change and vimentin to induce EMT, while morphological change and vimentin expression can induce each other and themselves. Computer simulations with the model followed by k-means clustering on DTW distance reproduce the two parallel EMT paths (Fig. 6F and fig. S9, C and D). The cross-correlation analysis also showed results similar to what was observed experimentally (fig. S9E). That is, the live-cell imaging platform presented here can provide mechanistic insight for further analyses.

We also compared our analysis on trajectories with snapshot data analysis. Similar to Fig. 1, we extracted single-cell data from the live-cell trajectories every 6 hours, treated them as snapshot data with no information of temporal correlation, and performed pseudo-time analysis with Scanpy and Wishbone (33, 34). The pseudo-time results (fig. S10, A and B) show only gradual change from epithelial to mesenchymal regions. We then enforced a two-branch analysis with Wishbone (fig. S10C), but the results are difficult for biologically meaningful interpretation. These results corroborate with Fig. 1 and demonstrate that some dynamical features revealed from live-cell imaging can be missing from snapshot data due to lack of temporal correlation information in the latter.


Compared to the recent advances of fixed cell–based single-cell techniques, live-cell imaging remains underdeveloped especially in studying CPTs, due to technical challenges. Specifically, the degrees of freedom specifying cell coordinates should be experimentally feasible for live-cell measurement and faithfully represent cell states. However, individual gene products typically only reflect partial dynamical information of a CPT process, and simultaneous fluorescence labeling of multiple genes is challenging. Recently, tracking cell morphological features through live-cell imaging emerges as a means of extracting temporal information about cellular processes in conjunction with expression-based cell state characterization (9, 3538). Cellular and subcellular morphologies reflect collective gene expression pattern and cell phenotype (3941). Furthermore, hundreds or more of morphology features such as cell size and shape can be conveniently extracted from bright-field images without necessity of additional fluorescence labeling. Here, we further developed a quantitative framework for recording and analyzing single-cell trajectories in a composite feature space, including cell shape and texture features and a computational package for related image analyses.

Our application to the TGF-β–induced A549 VIM-RFP EMT process demonstrates the importance of extracting dynamical information from live-cell data. A cell has a large number of molecular species that form an intricately connected network, and it interacts with a fluctuating extracellular environment, including cell-cell interactions. Consequently, even isogenetic cells show cell-cell heterogeneity, which further manifest as large trajectory-to-trajectory heterogeneity in single-cell CPT dynamics, and some dynamical features characteristic to a particular process might be unavoidably concealed from snapshot data. Our live-cell data reveals information on the two distinct types of paths of EMT with distinct vimentin dynamics. We then constructed and analyzed a mechanistic model on the basis of the previous reports that vimentin is a regulator and marker of EMT, and the model predicts these parallel paths. Further studies, however, are needed to investigate whether alternative mechanistic models can also explain the observations and examine whether these parallel paths exist in other EMT processes, in addition to TGF-β–induced EMT with the A549 VIM-RFP cells. One can also apply the framework to investigate the reverse process, mesenchymal-to-epithelial transition (MET), and examine whether MET follows different paths, as predicted from analyzing snapshot data (11).

While we present a general framework here, it has some limitations and needs further development. In this study, we restricted specification of the state of a cell by its cell shape and vimentin texture features. Application with recently developed imaging techniques, aided by machine learning–based computational algorithms, may provide additional features such as organelle texture and distributions in three dimensions from long-term label-free imaging, as demonstrated by a recent work of Sandoz et al. (42) using holotomographic microscopy to study multiorganelle dynamics. Cell cycle is a possible hidden variable contributing to the observed cell-cell heterogeneity in this study and can be tracked with proper reporters. Adding these new dimensions of information can provide finer resolution of cell state in an expanded cell state space.

Another limitation of the morphology/texture-based live-cell imaging framework is that it provides mostly indirect information on the dynamics of involved molecular species. For example, further mechanistic understanding of the observed parallel paths requires information on how the cell expression pattern changes along the paths. One can use the paths identified from live-cell imaging data, rather than currently used pseudo-trajectory approaches based on a perceived expression-similarity criterion, to time-order snapshot single-cell data, thus resolving the dilemma experienced in single-cell studies. For this purpose, one needs to establish a mapping system between the composite feature space and the expression space.

In summary, in this study, we demonstrate that live-cell imaging is necessary to reveal certain dynamical features of a CPT process concealed in snapshot data due to cell-cell heterogeneity. Meanwhile, we present a framework that facilitates recent emerging efforts of using live-cell imaging to investigate how a CPT process proceeds along continuous paths at multiplex, albeit lower-dimensional feature space, complementing fixed cell–based approaches that can provide snapshots of genome-wide expression profiles of individual cells. We expect that the framework can be generally applied since marked morphological changes typically accompany a CPT process.


Cell culture and treatment

The human non–small cell lung carcinoma lines, A549 (ATCC CCL-185) and A549 VIM-RFP (ATCC CCL-185EMT), were from ATCC. Cells were cultured in F-12K medium (Corning) with 10% fetal bovine serum (FBS) in MatTek glass bottom culture dishes (P35G-0-10-C) in a humidified atmosphere at 37°C and 5% CO2. Culture medium was changed every 3 to 5 days. During imaging, antibiotic-antimycotic (100×) (Thermo Fisher Scientific, 15240062) and 10 mM Hepes (Thermo Fisher Scientific, 15630080) were added to the culture medium.

sgRNA design and cloning into the gRNA-expression vector

The CHOPCHOP website ( was used to design high-performance single guide RNAs (sgRNAs) to target the sequence near the stop codon of the human vimentin gene. The cleavage activities of the gRNAs were validated using the T7 endonuclease 1 (T7E1) assay according to the manufacturer’s instructions (New England Biolabs, no. E3321). The sgRNA VIM-AS3 (5′-CTAAATTATCCTATATATCA-3′) was chosen in this study. To generate the sgRNA-expressing vector, VIM-AS-3 gRNA oligos were designed, phosphorylated, annealed, and cloned into the PX458 (Addgene, catalog no. 48138) vector, using Bbs I ligation. Multiple colonies were chosen for Sanger sequencing to identify the correct clones, using the primer U6 (forward): 5′-AAGTAATAATTTCTTGGGTAGTTTGCAG-3′.

Construction of VIM-RFP knock-in donor

The VIM-RFP knock-in donor was designed and constructed to contain approximately 800–base pair left and right homology arms, a Cayenne RFP gene (ATUM, no. FPB-55-609), preceded by a 22–amino acid linker, and followed by a bovine growth hormone polyadenylation signal sequence. To assist in drug-based selection of gene-edited cell clones, human elongation 1-alpha (EF1α)–blasticidin selection cassette, flanked by Loxp sites, was also cloned into the vector and positioned upstream of the right homology arm.

Generation of A549 VIM-RFP reporter cell clones

CRISPR-Cas9 technology was used to incorporate the RFP reporter into the 3′ terminal end of the vimentin gene. Briefly, A549 cells were plated at a density of 2 × 105 cells per well in a six-well plate. After 24 hours, cells were transfected with 4.0 μg of PX458_VIM-AS3 plasmid, 4.0 μg VIM-RFP knock-in donor plasmid, and 24 μl of transfeX (ATCC ACS-4005). Blasticidin selection (10 μg/ml) was applied 24 hours after transfection. RFP-positive cells were single-cell sorted and expanded for molecular characterization.

VIM-RFP knock-in clone identification and confirmation

RFP positive A549 VIM-RFP cells were harvested, and DNA was extracted using QuickExtract (Epicentre, QE09050). Primers were designed for left homology arm and right homology arm junction polymerase chain reaction (PCR): left junction: 5′-TAGAAACTAATCTGGATTCACTCCCTCTG-3′ (forward) and 5′-ATGAAGGAGGTAGCCAGGATGTCG-3′ (reverse); right homology: 5′-ATTGCTGCCCTCTGGTTATGTGTG-3′ (forward) and 5′-ATTACACCTACAGTTAGCACCATGCG-3′ (reverse). Junction PCR was performed using the Phire Hot Start II DNA Polymerase (Thermo Fisher Scientific), and the PCR amplicons were subjected to Sanger sequencing for identification of clones that contained the expected junction sequences at both left and right homology junctions.


A549 VIM-RFP cells were washed with phosphate-buffered saline (PBS), fixed with 4% formaldehyde, and blocked with 5% normal goat serum/0.1% Triton X-100 in PBS for 30 mins. Afterward, the primary antibodies were added to the blocking buffer, and cells were incubated for 1 hour at room temperature. Cells were subsequently washed and incubated with the secondary antibodies for 1 hour and wrapped in aluminum foil. After washing, images were taken with a Nikon Ti-E microscope (Hamamatsu Flash 4.0V2). The primary antibodies were rabbit anti-vimentin (D21H3) (1:500 dilution; Cell Signaling Technologies, catalog no. 5741), mouse anti-N-cadherin (13A9) (1:100 dilution; Cell Signaling Technologies, catalog no. 14215), and mouse anti-Snail (L70G2) (1:300 dilution; Cell Signaling Technologies, catalog no. 3895). For secondary antibodies, goat anti-mouse or goat anti-rabbit Alexa Fluor 488 (Thermo Fisher Scientific, catalog no. A-21235) was used at a 1:1000 dilution.

EMT induction

A549 VIM-RFP cells were plated at a density of 1 × 104 cells/cm2 and maintained in F-12K medium (ATCC 30-2004) supplemented with 10% FBS (ATCC 30-2020). After 24 to 48 hours, culture medium was replaced with fresh medium supplemented with TGF-β (4.0 ng/ml; R&D Systems 240-B) for 1 to 3 days to induce EMT. A549 VIM-RFP cells treated with PBS were used as a control.

Matrigel invasion assay

Control and EMT-induced A549 VIM-RFP cells were seeded into inserts of Boyden chambers (BD Biosciences, San Jose, CA) that were precoated with Matrigel (1 mg/ml), at 5 × 104 cells per insert in culture medium without FBS. Inserts were then transferred to wells with culture medium containing 10% FBS as a nutritional attractor. After 24-hour incubation, invading cells on the bottom side of the insert membrane were fixed with 4% paraformaldehyde for 2 min, permeabilized with 100% methanol for 20 min, and stained with 0.05% crystal violet for 15 min at 37°C. Noninvading cells on the top side of the membrane were removed by cotton swab. Photographs were taken from five random fields per insert. Cells in the five random fields were counted.

Colocalization analysis

The colocalization analysis was performed with JACoP (plugin of ImageJ) (43, 44). The Pearson correlation coefficient and Manders overlap coefficient were calculated following the method described in (45).


Time-lapse images were taken with a Nikon Ti-E microscope (Hamamatsu Flash 4.0V2) with differential interference contrast (DIC) and tetramethyl rhodamine isothiocyanate (TRITC) channels (excitation wavelength is 555 nm and emission wavelength is 587) (20× objective, numerical aperture = 0.75). The cell culture condition was maintained with the Tokai Hit Microscope Stage Top Incubator. Cells were imaged every 5 min with the DIC channel and every 10 min with the TRITC channel. The exposure time for DIC was 100 ms and that for the TRITC channel was 30 ms. That is, each full (2 days long) single-cell trajectory contains 577 DIC images and 289 fluorescent images. While taking the images, all the imaging fields were chosen randomly.

Single-cell segmentation and tracking

We segmented single cells using a previously developed method combining deep convolution neural networks (DCNNs) and watershed (46). To quantify cell morphology, we adopted the active shape model method (15, 18). After single-cell segmentation, the cell outline was extracted and resampled into 150 points. All the single-cell outlines were aligned to a reference outline (calculated on the basis of the average of several hundred cells). The 150 points (x and y coordinates) are the 300 features of cell morphology.

For single-cell tracking, we used the TrackObjects module in CellProfiler on the segmented images using a linear assignment algorithm (47, 48). In long-term imaging, the accurate tracking of cells can be lost for several reasons, such as cells moving in or out of the field of view or inaccurate segmentation. We kept trajectories that were continuously tracked with the starting point no later than 12 hours and the end point no earlier than 30 hours after adding TGF-β. These 196 trajectories were used for subsequent PCA. Among them, 139 were identified as reactive trajectories.

Vimentin image analysis

Haralick features have been widely used for classifying normal and tumor cells in the lungs (49) and the subcellular features or patterns such as protein subcellular locations (50, 51). After cell segmentation, each cell was extracted, and its Haralick features were calculated using mahotas (52). Haralick features describe the texture as coarse or smooth and complexity of images, and we used the following 13 features: 1, angular second moment; 2, contrast; 3, correlation; 4, sum of squares: variance; 5, inverse difference moment; 6, sum average; 7, sum variance; 8, sum entropy; 9, entropy; 10, difference variance; 11, difference entropy; 12, information measure of correlation 1; and 13, information measure of correlation 2 (50). Haralick feature calculation was based on the GLCMs (53). The GLCM’s size was determined by the number of gray levels in the cell image. Because of cell heterogeneity, the numbers of gray levels varied in different cells. Because the GLCM has four directions (0°, 45°, 90°, and 135°), the Haralick features were averaged on all four directions to keep rotation invariance.

Scaling of single-cell trajectories

Because of cell heterogeneity, it is more informative to examine the temporal change of an individual cell relative to its initial state, such as the basal level of gene expression in signal transduction studies (8, 54, 55). For the present system, it is the initial position in the composite feature space. We used a stay point–searching algorithm (56) to find the initial stay point of each cell in the space of cell shape and vimentin Haralick features. For each trajectory, we scaled all the landmark points by the square root of the area of the initial stay point. Physically, the latter is a characteristic length of the cell, and the scaling reflects the observation that the cell size does not affect EMT (57). All the vimentin Haralick features were reset so that the values at the initial stay point assume zero. The PCs were calculated after scaling. The scaling allows one to examine the relative temporal variation of single cells.

PCA was performed on all NT trajectories, i.e., a total of Nc (49,689 cells) with 300 morphology features (Nc × 300 matrix) for linear dimensionality reduction (58). The first seven components explained more than 98% of the variance. Specifically, the first and second components explained 82.7 and 10.5% of the variance, respectively. After calculation of Haralick features for each cell, PCA was calculated on the Nc × 13 matrix for linear dimension reduction.

Procedure of defining regions in the composite feature space

We fitted the distribution on each of the four morphology/texture coordinates with a two-component (c0, c1) GMM separately (fig. S7A) (58) and used the four GMMs to define the E, I, and M states (fig. S7B).

For each single cell in the space of morphology PC1, vimentin Haralick PC1, PC3, and PC4 X(xii = 1,2,3,4), the label of each coordinate Li is defined using the GMM with the following equationsLi({xi{p(xi{ci,0)>0.5}{xi<μci,0})=0(1)Li({xi|p(xi|ci,1)>0.9}{x>μci,1})=2(2)where p(xici, …) is the posterior probability of certain component of xi, and μci,0 and μci,1 are the mean values of the two components of ith GMM. The label of complementary set is defined as 1.

With the labels defined on all four coordinates, we first defined the E state asS({L1=0}{L2=0}{L3=0}{L4=0})=E(3)the M state asS({L1=2}{L2=2}{L3=2}{L4=2})=M(4)and the I state otherwise. However, this definition suffers from one weakness in that it assigns the same weight to vimentin Haralick PC1, PC3, and PC4, although vimentin Haralick PC3 and PC4 count for less variance than the PC1 does. Because of this equal weight distribution, small fluctuations in vimentin Haralick PC3 and PC4 could lead to unstable assignment of cell states.

To solve this problem, we use the above definition as an initial estimate of cell state to fit to a label-spreading function (28, 58). When fitting the label-spreading function, we adopted the KNN method (50 neighbors) and used a high clamping factor (0.5) to assure the global and local consistency. The KNN algorithm in the label-spreading function allows one to take the different scales of vimentin Haralick PC1, PC3, and PC4, and community structure into consideration (Fig. 5C). Since PC1 is more important in defining the range of neighbors, the weight of PC1 in the definition was automatically increased. This change in definition avoids a situation in which cells belonging to a common community and close to each other in the composite feature space get assigned to different cell states.

Cross-correlation analysis

The cross-correlation was used to calculate the time delay between different signals. We observed different transition time of morphology PC1 and vimentin Haralick PC1. The cross-correlation of the two time series (of morphology PC1 and vimentin Haralick PC1) was calculated (59). The time lag between the two signals was set for when the value of cross-correlation reaches the maximum (32). We separated all the trajectories by the sign of time delay between morphology PC1 and vimentin Haralick PC1.

Stochastic simulation on double well potential (Fig. 1)

While a cellular system is far from thermodynamic equilibrium, for simplicity, we illustrated the effect of cell-cell heterogeneity due to hidden slow variables with the following model system (

The potential function isU=(o2h)4(o2h)32(o2h)2+3y4(5)where o is the observable and h is the hidden slow variable (Fig. 1A).

The simulations were performed using the following procedures:

1. Generate initial condition (o0, h0) in the left well of this potential of multiple trajectories (4711) (fig. S1) with the Metropolis-Hastings algorithm (60).

2. For each initial condition, with fixed hidden slow variable h0, propagate the observable o with Langevin simulations along the 1D potential and with a Gaussian white noise (η(t))o(t+dt)=o(t)+dt(4(o2h0)+3(o2h0)24(o2h0)3)+dtη(t)(6)where dt is 0.01.

3. Propagate each trajectory to t = 10.

Network model of EMT (Fig. 6E)

We built the network model by summarizing the existing literatures. The EMT morphology variation is mainly contributed to by generation of filament actin and E-cadherin down-regulation, which can activate the YAP/TAZ (yes-associated protein/transcriptional coactivator with PDZ-binding motif) pathway (61). The YAP/TAZ pathway can induce translocation of Smad2 (mothers against decapentaplegic homolog 2), which plays important roles in EMT (62, 63). Thus, morphology variation can activate both vimentin and itself. Vimentin can induce the EMT morphological change through regulating β1-intergrin and E-cadherin (64). Vimentin can be activated upon TGF-β induction through Slug and it also activates Slug through dephosphorylation of extracellular signal–regulated kinase, which forms a self-activation loop (65). Vimentin is required for the mediation of Slug and Axl (64, 6668), and it can induce variation of cell morphology, motility, and adhesion (68). Vimentin fibers regulate cytoskeleton architecture (64), and more vimentin fibers are assembled in A549 cells during EMT (20).

Next, we formulated a mathematical model corresponding to the networkdMdt=αm+βm+vmV2V2+Kvm2+cmM4M4+Km4δmM(7)dVdt=αv+βv+mvM2M2+Kmv2+cvV4V4+Kv4δvV(8)where αm (=0.1) and αv (=0.12) are the basal generate rates of morphology variable and vimentin, separately; βm and βv are the generate rates of morphology change and vimentin activated by TGF-β, respectively; vm (=2.0) and mv (=1.0) are the activation coefficients of vimentin and morphology to each other; cm (= 3.8) and cv (= 4.0) are the self-activation coefficients of morphology and vimentin, respectively; and Kvm (=8.0), Kmv (=8.0), Kv (=2.4), and Km (=2.4) are the half-maximal effective concentrations of the Hill function. δm (1.0) and δv (=1.0) are degradation rates of morphology and vimentin, respectively.

The Langevin simulations were performed as follows:

1. We set βm and βv as 0 for simulating the control condition (i.e., without TGF-β treatment). Initialize a trajectory with random point (M0 and V0) sampled from a uniform distribution within a range of [0,5). Run simulations with the following equationM(t+dt)=M(t)+dt(αm+βm+vmV2V2+Kvm2+cmM4M4+Km4δmM)+dtη(t)(9)dt(αv+βv+mvM2M2+Kmv2+cvV4V4+Kv4δvV)+dtη(t)(10)where dt was set to be 0.01, and η(t) was the normal Gaussian white noise. The duration of simulation was set to 100. At the end of the simulation, the cell state relaxed to the basin of the epithelial state (fig. S9A), which was then set as the initial condition under TGF-β treatment.

2. After generating multiple initial conditions from the first step, we increased the values of βm(to 0.6) and βv(to 1.0) to simulate the condition of TGF-β treatment. If the cell gets into the range that its distance to the attractor of the mesenchymal state (Fig. 6F and fig. S9B) is less than one, then this trajectory was considered to be a trajectory of EMT.

3. After getting Nsimu (=185) reactive EMT trajectories, we performed analysis with the simulated trajectories similar to what we did with the experimental trajectories.

We obtained the steady-state probability distribution Pss by solving the diffusion equation (using Matlab 2018a PDEtool)P(M,V,t)t=(P(M,V,t)F(M,V))+D22P(M,V,t)(11)where F(M,V)=(dMdt,dVdt), and D is the diffusion coefficient. With the steady-state probability distribution, we obtained the quasi-potential of EMT defined as U ∝ − ln (Pss) (69, 70). Without TGF-β, there exists a deep basin as epithelial state and a shallow basin as mesenchymal state in the quasi-potential landscape (fig. S9A). After TGF-β treatment, the landscape is changed, on which the mesenchymal basin becomes deep, and a valley is formed where the vimentin level is high (Fig. 6E and fig. S9B).

Snapshot data analysis

The snapshot data were extracted from live-cell trajectories with a time step of 6 hours. We processed the data with Scanpy and Wishbone (33, 34). The root cell in Scanpy or the start cell in Wishbone was set as the cell that is closest to the average of data at 0 hour in feature space. The number of neighbors was set to 50 in constructing a neighborhood graph in Scanpy. In addition, the number of way points in Wishbone was set to 200.

Computer package M-TRACK

The M-TRACK program is written in Python 3 and provided with a graphical user interface. It provides tools for analyses of cell morphology with the active shape model, distribution and texture features of protein or gene florescence in single cell, and single-cell trajectories in the PC domain. The input files include the original gray-level images, segmented cell mask, and a database file of tracking results from Cellprofiler. The computer package can be downloaded from GitHub ( Part of the source code is adapted from CellTool (18).

Statistical analysis

Statistical analyses were performed mainly with Python package, including SciPy and scikit-learn (58, 59). Student’s t test was used to calculate the statistical difference between different groups of samples. The samples for imaging were randomly selected to avoid bias.


Supplementary material for this article is available at

This is an open-access article distributed under the terms of the Creative Commons Attribution-NonCommercial license, which permits use, distribution, and reproduction in any medium, so long as the resultant use is not for commercial advantage and provided the original work is properly cited.


Acknowledgments: We would like to thank Y. Jiang and X. He on the helpful discussions of the active shape model. Funding: This work was partially supported by the NSF (DMS-1462049), the National Cancer Institute (R37CA232209), the National Institute of Diabetes and Digestive and Kidney Diseases (R01DK119232) to J.X., and the NIH supported microscopy resources in the Center for Biologic Imaging at University of Pittsburgh (1S10OD019973-01). Author contributions: J.X. conceived the project. D.D., S.K., M.S.E., and W.S. generated and characterized the cell line. W.W. performed experiments with support from C.T.W. and S.C.W. on imaging and from J.Z. on cell culturing. W.W. and J.X. analyzed the data with input from J.Z. Y.D. generated the GUI of the computer package. W.W. and J.X. wrote the manuscript, with input from J.Z., S.C.W., D.D., and W.S. Competing interests: D.D., S.K., M.S.E., and W.S. are employees of the ATCC, which offers the A549 VIM-RFP cell line for sale commercially. The remaining authors declare that they have no competing interests. The A549 VIM-RFP cells can be provided by ATCC pending a completed material transfer agreement. A549 VIM-RFP cells are available through ATCC website (ATCC@CCL-185EMT). Data and materials availability: The computer package, together with the image data files, can be downloaded from GitHub ( All data needed to evaluate the conclusions in the paper are present in the paper and/or the Supplementary Materials. Additional data related to this paper may be requested from the authors.
View Abstract

Stay Connected to Science Advances

Navigate This Article