Machine Learning with Optical Phase Signatures for Phenotypic Profiling of Cell Lines

Abstract

Robust and reproducible profiling of cell lines is essential for phenotypic screening assays. The goals of this study were to determine robust and reproducible optical phase signatures of cell lines for classification with machine learning and to correlate optical phase parameters to motile behavior. Digital holographic microscopy (DHM) reconstructed phase maps of cells from two pairs of cancer and non-cancer cell lines. Seventeen image parameters were extracted from each cell’s phase map, used for linear support vector machine learning, and correlated to scratch wound closure and Boyden chamber chemotaxis. The classification accuracy was between 90% and 100% for the six pairwise cell line comparisons. Several phase parameters correlated with wound closure rate and chemotaxis across the four cell lines. The level of cell confluence in culture affected phase parameters in all cell lines tested. Results indicate that optical phase features of cell lines are a robust set of quantitative data of potential utility for phenotypic screening and prediction of motile behavior. © 2019 International Society for Advancement of Cytometry

Optical phase signals imaged across a living cell reveal morphology and subcellular features that relate to aspects of cell physiology. The nucleus (1), actin cytoskeleton (2), membrane-bound organelles (3), and cytoplasm (1) contribute to the cell phase profile. Events that significantly affect a cell’s morphology or subcellular phase texture are often detected longitudinally (4) or by comparing distinct cell populations (5). For example, time-lapse imaging of single cells reveals the effects on phase signals of agents inducing apoptosis (6), cytoskeletal disruption (4), or reduced proliferation (7). In clinically relevant applications, phase features of cells have been used to distinguish cancer from non-cancer cells in suspension (8), primary cancer from metastatic cancer cells (8), and in clinical biopsy specimens fixed on slides (9,10). Biomedical applications of quantitative phase imaging were recently reviewed (11).

Digital holographic microscopy (DHM), a quantitative phase imaging technique, captures the phase structure of imaged cells and reveals significant cell-to-cell heterogeneity within cell lines (12). Variation in stage of the cell cycle (13), cell mechanical properties (14), level of confluency in vitro (15), and other factors contribute to broad distributions of phase features, even between cells of the same line and type. This variation presents a challenge for label-free identification of cell type and classification of cells into functional groups based on phase features. Such variation is notably evident in cultured breast cancer cells of the MDA-MB-231 line which adopt elongated, mesenchymal and rounded, amoeboid morphologies in vitro, indicative of a compensatory invasion strategy in vivo (16). Elongated and rounded MDA-MB-231 cells were classified using a machine learning algorithm with 94% accuracy. However, the prediction of amoeboid or mesenchymal motile behavior from single-phase images was not assessed (2). Nevertheless, subtle phase morphology and texture features might further distinguish such subpopulations and predict motile phenotype.

Aspects of cell morphology and subcellular features reflect characteristics of motile behavior. For example, adherent cells tend to move sequentially through different shape-space trajectories constructed using principal component analysis of shape descriptors extracted from time-lapse image series (17). Probabilistic modeling of the shape-space identified the type of cell movement and allowed for accurate clustering of cell responses to drugs that disrupt cytoskeletal signaling. Quantitative phase imaging (QPI) is a label-free way to track and quantify cell aggregate (18), whole-cell (19,20), and subcellular movement (21). Thus, in the phase map of a single cell, there is phenotypic information relevant to motility in the cell shape, the pixel histogram, and in individual pixels’ phase values in relation to neighboring pixels. Texture parameters of the gray-level co-occurrence matrix quantify the phase relationship of neighboring pixel pairs (22). Given the involvement of cell motility in cancer metastasis, wound healing, and immune function, phenotypic profiling of cell movement behavior based upon phase features from static phase maps would contribute to the study of these processes.

Machine learning automates and optimizes classification of cells based on quantitative metrics. Such algorithms use for phenotypic profiling (23) could aid pathologists and scientists by reducing the time spent in manual image evaluation, lowering human error, and making evaluation of large image datasets feasible. A general approach using quantitative image parameters as predictors involves tabulating predictors with known classifiers in a training dataset, parameter standardization, data reduction by principal components or a similar technique, comparison of algorithm performance by cross-validation on the training dataset, and then application of the trained algorithm to a naïve dataset to determine predictive power. Several recent examples illustrate machine learning applied to quantitative phase images of cells. Screening of blood samples using six features derived from QPI as inputs to a neural network distinguished five hematologic disorders with 80–98% accuracy (24), and separately, the shape retention of red blood cells under different storage conditions (25). Similarly, three types of mouse lymphocytes were classified with accuracy of up to 90% using parameters from optical diffraction tomography and the k-nearest neighbor’s algorithm (26). Deep learning approaches using QPI parameters have also been explored (27). Macrophage activation by lipopolysaccharide (LPS) was reliably detected using cell morphologic parameters derived from DHM phase maps (28). These studies demonstrate that parameters from QPI techniques performed on multiple cell types can classify cell subpopulations into distinct pathophysiological or functional groups.

Reproducible optical phase measurements are critical for robust phase parameter-based classification of cell phenotypes. Many factors in culture influence cell morphology, becoming random effects if not well controlled experimentally. Culture conditions affecting cell morphology include the local two- or three-dimensional physical microenvironment (29) and proximity to other cells. Methods standardization has improved reproducibility of many biological measurements, notably in standards used for mechanical testing of tissues (30). A recent effort to establish guidelines for live cell imaging reported an improved detection of motile phenotypes by lowering background variation due to technical artifacts (31). Similar studies addressing reproducibility of QPI measurements of cells (15,32) allow for easier adoption of optical phase-based approaches for research and eventually in the clinic.

The rationale for this study was based on the high sensitivity of cell morphological parameters to cell line and functional behavior. The hypothesis was that optical phase parameters collected from DHM phase maps of cells classify cell lines of similar morphology with high accuracy and predict functional motile behavior, using machine learning. The normal and cancer human cell lines are well-characterized, although from different patients and organ sites. Although variation between cell lines should be higher compared to cells from the same patient/organ, we sought to identify differences between the well-studied MDA-MB-231 and MCF-7 breast cancer cell lines with normal cell lines of extreme mesenchymal and epithelial appearance, to aid in generalization across multiple individuals and organs. To test the first part of this hypothesis, phase signatures were determined from four cell lines, two each with mesenchymal and with rounded/epithelial morphologies, and support vector machine-learning classification was performed and evaluated. To test the second part of the hypothesis, individual phase parameters were correlated with quantitative metrics of wound closure rate from the scratch wound assay, and chemotaxis from the Boyden chamber assay, using Spearman’s rank-order correlation coefficient. Further, statistical analyses identified cells closest to the population average phase signature, individual differences in phase parameters between each cell line, and variation in phase signatures associated with cell layer confluence. Results extend the utility of QPI to assess adherent cells of similar morphology but different phenotypic status (cancer versus non-cancer), suggest standard culture conditions required for determination of phase signatures, and propose a way to identify individual cells with atypical phase features. The correlation of phase parameters to motile behavior raises the potential for identifying highly motile cells based solely on single-shot, label-free phase images.

Materials and Method

Cell Culture

Two human breast cancer cell lines MDA-MB-231 and MCF-7 which were a generous gift from Dr. Zaver Bhujwalla (John Hopkins School of Medicine, Baltimore, MD) and two immortalized human gingival cell lines fibroblast (HGF) and keratinocyte (GIE-No3B11) (Applied Biological Materials, Inc., British Columbia, Canada) were cultured in standard incubation conditions of 37°C, 5% CO₂, and 100% humidity (HERAcell 150i, Thermo Fisher Scientific, Waltham, MA) in nutrient media. The cancer cell lines were cultured on 10 cm diameter tissue-culture treated Petri dishes with Dulbecco’s modified Eagle’s medium (Lot # SLBW4140, Sigma-Aldrich, St. Louis, MO) supplemented with 10% Fetalgrow (Rocky Mountain Biologicals, Missoula, Montana) and 1% penicillin–streptomycin (Corning Inc., Corning, New York). The HGF and GIE cell lines were cultured in Prigrow 3 and Prigrow 4, respectively (Applied Biological Materials). Nutrient media for gingival cell lines were supplemented with 10% fetal bovine serum (FBS) and 1% penicillin–streptomycin. For DHM imaging and the scratch wound assay, cells were seeded on 35-mm diameter glass-bottomed Petri dishes. For confluency experiments, cells were seeded at 96, 160, and 255 cells/mm² for MDA-MB-231 and HGF, 191, 318 and 382 cells/mm² for MCF-7 and GIE for respective confluence levels of approximately 30% (low), 50% (medium), and 70% (high) after 24 h. To prepare for imaging with DHM, cells in glass-bottomed Petri dishes were removed from the incubator, fed with 200 μl of prewarmed, fresh nutrient media, and a sterile cover slip was placed over the glass-bottomed surface. Imaging was performed for 15–20 min per dish to avoid effects of the ambient environment on cultured cell morphology.

DHM Imaging

A previously described off-axis, bi-telecentric DHM system was used to record digital hologram and derive reconstructed cell phase maps (33). DHM lateral resolution was 1.2 μm with 0.18 μm × 0.18 μm pixel dimensions of the lateral reconstruction. A 632 nm wavelength HeNe laser was used to generate sample and reference beams that recombined at the camera sensor plane as holograms. The holograms were captured by a 1.3 MP CMOS camera (Lumenera Corporation, Inc., Ontario, Canada) and processed into reconstructed phase maps as previously described. Principal component analysis (PCA) was initially employed in postprocessing to cancel the main hologram phase aberrations, while Zernike polynomial surface fitting was applied to cancel higher order aberrations. A detailed reconstruction method to achieve aberration-free holograms by using this DHM setup was described previously (2,33). Optical phase maps were processed in pixel units of nanometers converted to micrometers as convenient units to represent the optical pathlength through adherent cells, which were roughly 5–15 μm.

Image rocessing and Analysis

Image processing was followed by machine learning and parameter analysis (Fig. 1). A custom MATLAB code to segment cells from DHM phase maps and record cell morphology, first-order phase and second-order phase pixel texture parameters for each cell was modified from a previously described version (2) to measure additional parameters. Seventeen quantitative phase parameters were extracted (defined in the Supporting Information Table S1). Parameters included cell phase information, or three-dimensional parameters (cell phase height, phase height standard deviation, kurtosis, skew, contrast, correlation, homogeneity, and energy), and cell geometry indices, or two-dimensional parameters (cell area, perimeter, and eccentricity). In addition, each cell’s central region [observed to contain high-contrast phase features but not attributed to any single organelle, which would require additional experimental measurements and optical modeling (34)] was segmented from the DHM phase map based on selecting pixels that are greater than the average cell phase plus one standard deviation and located around the central cell area. Then, cell central region parameters were collected: the three-dimensional parameters of average phase height, maximum phase height, kurtosis and skew, and two-dimensional parameter of region area. Previous colocalization experiments of adherent MDA-MB-231 cell phase maps with fluorescently stained nuclei demonstrated highest phase values overlapping the nucleus and perinuclear region (2). We do not claim here to segment the nucleus from optical phase maps. Rather, the central region is simply defined as a region of high phase contrast with texture features of potential significance, in agreement with other studies observing complex phase texture in the central regions of cells (35,36). All cell morphology and phase parameters were referenced to saved image files of individually segmented cells to validate quantitative parameters with cells maps.

Figure 1.

Flowchart of image processing and machine-learning steps. The DHM instrument and MATLAB-based cell segmentation routine were described previously (1). The 99% tolerance interval for SSD values was calculated at 99% confidence from normalized parameter values.

Cell Phase Parameters

Phase parameters were collected from n = 2,577 cells across the four cell lines. The sum of squared deviations (SSDs) of standardized DHM parameters was calculated for each cell compared to the population average parameters from each cell line, using Eq. (1):

$$\text{SSD}=\sum _{i=1}^{17}{\left({\varphi }_i-{\overline{\varphi }}_i\right)}^2,$$ (1)

where ϕ_i is the parameter value of an individual cell normalized to the maximum value from the sample, and ${\overline{\varphi }}_i$ is the normalized parameter mean of the cell line sample. Three cells with the lowest SSD and one with the highest were identified in DHM maps and compared side-by-side. Histograms of SSD values were plotted and tested for log-normal distribution by the Shapiro–Wilk test on natural log-transformed SSD values. The number of cells were n = 229, 218, 347, and 252 for MCF-7, GIE, MDA-MD-231, and HGF cell lines, respectively. A 99% tolerance interval (TI) for SSD values was calculated at 99% confidence from log-transformed SSD values using Eq. (2):

$$99\%TI=\overline{\text{SSD}}\pm \sqrt{\nu \left(1+\frac{1}{N}\right)z_{\frac{1-p}{2}}^2/{\chi }_{1-\gamma ,\nu }^2{\sigma }_{\text{SSD}}},$$ (2)

where $\overline{\text{SSD}}$ and σ_SSD are the mean and standard deviation of the SSD sample distribution from normalized phase parameters, respectively, with ν degrees of freedom, from N cells, and with $z_{\frac{1-p}{2}}$ and ${\chi }_{1-\gamma ,\nu }^2$, as critical values from the z- and Chi-Square distribution, respectively, with P = 0.99 representing 99% tolerance, and γ = 0.01 representing 1-γ = 99% confidence. Cells with SSDs falling outside this tolerance range from experiments at different levels of cell confluency were classified as outliers. For the cell lines at low, medium, and high confluency, respectively, there were 78, 174, and 135 cells recorded for MCF-7; 114, 173, and 155 cells recorded for GIE; 107, 139, and 239 cells recorded for MDA-MB-231; and 57, 68, and 92 cells recorded for HGF.

Machine Learning and Principal Component Analysis

All data in this article were analyzed using the Statistics and Machine Learning Toolbox in MATLAB 2016a (The MathWorks, Inc., Natick, MA). A linear support vector algorithm was chosen based on the previous experience (2). The software allows users to define predictor and response input, set cross-validation for training, perform algorithm training, and then evaluate algorithm performance on a new test dataset. Datasets for training and fivefold cross-validation were the same as for SSD calculations, but a random subset was left out for future testing, leaving n = 219, 210, 247, and 219 cells from MCF-7, GIE, MDA-MD-231, and HGF cell lines, respectively. After the training algorithm was constructed, a naïve dataset was tested to produce a final, unbiased performance assessment. The number of cells in the naïve test sets were n = 88, 122, 100, and 91 for MCF-7, GIE, MDA-MD-231, and HGF cell lines, respectively.

There were in total six classifications by pairwise comparison across four cell lines. Accuracy, true positives, false positives, and area under curve (AUC) from receiver-operating characteristic (ROC) curves were calculated. Principal component analysis (PCA) was applied before training to evaluate the improvement in machine-learning performance with additional principal components. The optimum number of principal components was when the AUC failed to become larger with more components.

Scratch Wound Healing Assay

All four cell lines were cultured in tissue-culture treated 24-well plates with the media conditions described above. To avoid variation between well plates, a single-well plate experiment was performed, consisting of n = 6 replicate wells for each cell line, with one scratch wound per well. Wells were seeded at densities of 60,000, 70,000, 80,000, and 100,000 cells for MDA-MB-231, MCF-7, HGF, and GIE respectively, to produce confluent monolayers after 24 h. At confluency, a wound was created by scraping a sterile 200 μl plastic micropipette tip once along the bottom of each well. Media was changed to eliminate floating cells. Each well was imaged at 0, 4, 6, 8, 10, and 24 h post-wound under an inverted phase contrast microscope (CK2, Olympus, Tokyo, Japan) Wound closure rate was calculated by averaging five measurements of gap width at each timepoint and normalizing widths to wound width at 0 h. Measurements were performed using the ImageJ software (ImageJ v1.48, National Institutes of Health).

Boyden Chamber Assay

In total, 50,000 cells of each cell line were seeded in a Boyden chamber (662,638, Greiner Bio-One, Inc., NC), using six replicate chambers for each cell type, placed in a 24 well-plate. Media at 300 μl with 1% serum was added in the upper chambers for 6 h. Then, another 300 μl of serum-free media replaced the serum media while 600 μl of 10% serum-containing media was added in the lower chamber of the well plate. After 24 h all media were removed, and Boyden chambers were washed with 1X PBS, fixed with 3.7% paraformaldehyde, permeabilized with 0.1% Triton-X in PBS, and stained with a 1:1000 dilution of a 2 μg/ml 4′,6-diamidino-2-phenylindole (DAPI) (Life Technologies, Carlsbad, CA). Each chamber was imaged under a 40× objective with inverted epifluorescence microscopy (EXI-310; Accu-Scope, Commack, NY) to acquire four fields of view. The number of nuclei on the bottom surface of each Boyden chamber membrane in each of the four images were averaged and then averaged again over six replicates.

Statistical Analysis

Significance was tested by using ANOVA followed by post hoc pairwise comparisons with the Bonferroni correction. All populations were tested for normality using Levene’s test and for homoscedasticity. Also, the distribution of residuals was evaluated for nonuniform patterns, for each test. If the data did not pass these tests, the nonparametric Kruskal–Wallis test was used instead. Correlations between parameters and motility metrics from the scratch wound and Boyden chamber assays were evaluated using Spearman’s rank-order correlation coefficient. Statistical tests were performed using the Systat version 13.1 (Systat Software, Inc., Chicago, IL). The significance level was set at P < 0.05.

Results

Optical Phase Signatures Reveal Morphology of Epithelial and Mesenchymal Cell Lines

The MCF-7 and GIE cell lines have an epithelial morphology and appear in clusters (Fig. 2A,B). Both boundary and inner MCF-7 cells appear similar, while GIE cells are more elongated at cluster borders. The HGF cell line has a mesenchymal morphology found in aligned groups (Fig. 2D). The MDA-MB-231 cell line appears as individual rounded and elongated cells (Fig. 2C). In all cultures, rounded and/or smaller cells possess higher phase profiles. In addition, cancer cell phase texture appears more punctate in the central region versus a smoother texture in non-cancer cells. Cell borders, even in clusters of interacting cells, are clearly visible in phase maps from DHM.

Figure 2.

Representative phase reconstruction maps of cells from four cell lines. (A) Human breast cancer cell MCF-7, (B) immortalized human gingival keratinocytes (GIE), (C) human breast cancer cells (MDA-MB-231), and (D) human gingival fibroblasts (HGF). Cells were imaged using DHM after 24 h of culture. Scale bar is 20 μm, grayscale bar indicates cell phase height.

Optical Phase Signatures Identify Typical and Atypical Cells

Three cells with the lowest SSD and one cell with high SSD of normalized phase parameters were identified for each cell line, and their phase maps were displayed (Fig. 3A, cells indicated by white arrows). Cells with large SSD are morphologically unlike cells with small SSD. The histograms of SSD values (Fig. 3B) are log-normal (Shapiro–Wilk tests, P > 0.05). No cells had SSD values outside a 99% tolerance interval for the log-normal distribution. The cell lines in order from least to most SSD variance are HGF, MDA-MB-231, MCF-7, and GIE.

Figure 3.

DHM phase maps for the most representative cells (A) and histograms (B) of SSD values among each population. The sum of squared deviation (SSD) shows the distance of individual cells from the population mean. Scale bar is 20 μm. Grayscale bar indicates cell phase height. Arrows indicate only the cell with SSD printed in each panel.

Optical Phase Parameters Vary Significantly across Four Cell Lines

A parallel coordinates plot provides an overview of the phase signature for each cell line and statistical comparisons (Fig. 4). Each polyline represents the median values of 17 standardized phase parameters for one cell line. Red ovals indicate groups with no significant difference by post hoc tests following ANOVA or Kruskal–Wallis tests. Marker outside red ovals are different from the other. Some phase parameters, such as sk, A, P, Cor, and sk_n, have no red ovals and are significantly different between the four cell lines. Phase mean, μ, separates only GIE from the others. The contrast texture parameter, Co, is different for MDA-MB-231 but no other cell line. Standardization brought some parameter medians that were significantly different by post hoc test together on the polyline plot, such as median A from HGF and MDA-MB-231 groups.

Figure 4.

Parallel coordinates plots of standardized mean of 17 phase signatures. Red ovals indicate no significant difference tested by post hoc comparisons with Bonferroni corrections (P < 0.0001).

Machine Learning Classifies Epithelial from Mesenchymal, Normal from Cancer Cell Lines

A support vector machine algorithm distinguishes non-cancer from cancer cells and epithelial from mesenchymal cells with 90–100% accuracy (Table 1). The GIE cell line was best distinguished from HGF cells at 100% accuracy for the test dataset. The algorithm distinguished GIE versus MDA-MB-231 and MCF-7 versus HGF with 98% accuracy. The accuracy of distinguishing MDA-MB-231 versus MCF-7 cells was slightly lower at 95% and lowest when comparing two cell lines of similar morphology: 90% between epithelial (MCF-7 vs. GIE) and mesenchymal (MDA-MB-231 vs. HGF) cell lines. Misclassification errors are also reported in Table 1.

Table 1.

Results of linear support vector machine learning for cell line classification

				MISS CLASSIFICATION
	ACCURACY (%)	TPR	auc	MDA → HGF	HGF → MDA
MDA vs. HGF	92.1	0.92	0.97	5/100	10/91
				MDA→MCF	MCF→MDA
MDA vs. MCF	90.1	0.91	0.99	5/100	12/88
				MDA→GIE	GIE→MDA
MDA vs. GIE	99.1	0.99	1.00	1/100	1/122
				HGF→MCF	MCF→HGF
HGF vs. MCF7	97.2	0.97	1.00	0/91	5/88
				HGF→GIE	GIE→HGF
HGF vs. GIE	100	1	1.00	0/91	0/122
				MCF→GIE	GIE→MCF
MCF vs. GIE	91.9	0.91	0.97	15/88	2/122

TPR, true positive rate; AUC, area under the receiver operating characteristic curve; MDA, MDA-MB-231 cell line; HGF, human gingival fibroblasts; MCF, MCF-7 cell line; GIE, GIE-No3B11 cell line. A→B indicates A misclassified as B.

Between five and six principal components distilled from 17 phase parameters were needed for best machine-learning performance (Supporting Information Fig. S2). Only classification between MDA-MB-231 and HGF cell lines required six PCs, while the others performed equally well with five PCs.

Optical Phase Signatures Correlate with Scratch Wound Healing

Migration activity of four cell lines was assessed from scratch wound borders (Supporting Information Fig. S3). All cell lines began gap closure within 4 h. The HGF cell line did not uniformly close the wound gap, but individual cells migrated across the gap. Cancer cells generally migrated more resulting in faster wound closure comparing to gingival cells (Fig. 5A). Wounds were closed faster by MCF-7 than MDA-MB-231 cells in the first 10 h, yet MDA-MB-231 was the only cell lines that completed wound closure by 24 h. The gap in the GIE cell monolayer was the widest after 24 h. Phase parameters that were significantly differently among all cell lines were correlated to wound closure rate and displayed in a heat map (Fig. 5B). Green color shows a strong positive correlation of sk and eccentricity to healing extent at 10 h postscrape, while red indicates negative correlation of cell area and perimeter to healing amount, strongest at 8 h. Textural correlation and central region skew were also negatively correlated with healing amount at 24 h.

Figure 5.

Scratch wound healing assay. (A) Quantification of normalized wound closure versus culture time following the scrape from images taken at indicated time points by phase contrast microscopy provided in the Supporting Information, and (B) matrix of Spearman’s correlation coefficients of DHM phase parameters and wound closure extent at five time points postscrape.

Optical Phase Parameters Correlate with Chemotaxis

In addition to scratch wound closure, the migration activities of the four cell lines were evaluated by the Boyden chamber chemotaxis assay (Fig. 6A). The MDA-MB-231 cell line had the most cells translocated across the membrane, followed by MCF-7 and HGF cell lines. No GIE cells were found to translocate. The average number of translocated cells of these cell lines were ranked and correlated to all 17 DHM parameter ranks (Fig. 6B). Cell line ranks were flipped between the Boyden chamber assay and scratch wound closure; for example, the most chemotaxing cell line, MDA-MB-231, provided the smallest gap after 24 h of scratch wound healing, and so was ranked 1 and 4, respectively. Cell phase parameters like μ, ku, sk, Cor, Co, and Ho were highly correlated to cell response to chemotaxis.

Figure 6.

Boyden chambers assay results for four cell lines. (A) Bar graph showing number of cells migrating through membranes after 24 h and (B) Spearman’s correlation coefficients between cell translocation by chemotaxis and DHM parameters of 4 cell lines; *indicates P < 0.05 for all pairwise comparisons.

Cell Culture Density Affects Optical Phase Parameters from Cells

The reproducibility of optical phase parameters from cell lines was investigated by taking DHM measurements at three levels of confluency. Between 5 and 13 of 17 phase parameters were altered by level of confluence in culture, depending upon the cell line (Fig. 7). The skewness parameter was excluded from the bar graph of Figure 7 because it included negative values, and so was not normalized. Skewness was different between high and low levels of confluency for HGF (P< 0.05) only. More phase parameters from cell lines forming aggregates such as MCF-7 and GIE (Fig. 7B,D) were affected by level of confluence than mesenchymal cells not forming aggregates, MDA-MB-231 and HGF (Fig. 7A,C). HGF cells were least affected by the level of confluence. The number of cells with SSD values outside the 99% tolerance intervals from the distributions of Figure 3 are indicated in Figure 7. The bold highlighted table cell indicates the level of confluence of the original dataset of Figure 3, from which tolerance intervals were calculated, used to assess outliers from the new data in Figure 7.

Figure 7.

Bar graphs of normalized mean phase parameters at different levels of confluency. (A) MDA-MB-231, (B) MCF-7, (C) HGF, (D) GIE cell lines. Square brackets indicate significant differences among groups by ANOVA with P < 0.05.

Discussion

Phase parameters derived from adherent cells were able to distinguish cell lines with similar morphology and were related to functional motile behavior. Classification of cell lines of similar morphology using phase features was determined by linear support vector machine learning, requiring only five to six principal components for accuracy of 90–100%. The sum of squared deviations of normalized phase parameters identified cells of typical and atypical appearance within each cell line. Phase parameters from each cell line were compared using polylines on a parallel coordinates plot and statistical analysis to highlight differences in central tendency of each normalized parameter for each cell line. Several of the phase parameters were strongly correlated with the rate of scratch wound closure and, separately, cell chemotaxis. Finally, a subset of each cell lines’ phase parameters was influenced by the level of confluence, and SSD tolerance intervals identified cells with outlier SSD values. Taken together, these findings indicate the feasibility of developing optical phase indices of functional cell behavior useful for phenotypic profiling of cell lines.

The study design and QPI approach have intrinsic strengths and limitations. Cell phase maps from DHM are quantitative on a pixel-level, obtained without added labels, and measure living cells. Phase data from QPI are well-adapted for live-cell imaging, in which adding exogenous labels might make observations less reproducible. While DHM reduces the chance for bias from sample preparation, phase maps contain some artifacts. These artifacts can be removed in the future by image spatial filtering or using a low-coherence light source (37). Given the large number of technical approaches to achieve QPI (38), reproducible measurements across instruments have been challenging. Phase texture, for example, depends upon the lateral resolution, a function of the numerical aperture of the imaging system. The ideal phase imaging system has a lateral resolution near the diffraction limit, or else uses super-resolution approaches to improve lateral resolution (39–43).

Optical phase measurements from cells were largely reproducible across multiple datasets, with some experimental variation that was reduced by adequate sample size and controlled culture conditions. Machine learning results suggest 200 cells per group to be an adequate sample size for classification. Other similar studies have also found this number of cells aids in cell line classification (7,15,44). Confluency was constant for the machine learning dataset, and when varied, affected numerous phase parameters (Fig. 7). A dataset of cells from identical culture conditions (Fig. 7, bold boxes) had few outliers outside the 99% tolerance interval from an independent dataset (Fig. 3B), validating the approach of identifying atypical cells using optical phase signature SSD. This study suggests that five to six principal components from a 17-parameter optical phase signature serve as good predictors for cell line classification. Reduction of data dimensionality by selection of principal components is a valid way to avoid overfitting of the machine-learning algorithm while retaining the flexibility of multiparametric dataset responses to different classification problems (8). Other algorithms and statistical approaches such as neural networking and k-nearest neighbor classification (45) could be used, as well as other experimental approaches, including labeling experiments (46) and biodynamic imaging (47). Ultimately, the more challenging limitation to determining subtle alterations in cell phenotype lies in the signal measured and instrument used rather than the algorithm and statistical analysis chosen. A reference library of optical phase parameter values for cell lines in standard conditions could be built to test this assumption.

The phase parameters from this study create a phenotypic profile of the four cell lines related to their origins and characteristic appearance in two-dimensional, substrate-adherent culture. In this study, breast cancer cells (MDAMB-231 and MCF-7) and non-cancer human gingival cells (GIE, HGF) with high cell–cell affinity (MCF-7, GIE), low cell–cell affinity (MDA-MB-231), variable cell–substrate affinity (MDA-MB-231), and high cell-substrate affinity (HGF) were distinguished with high accuracy using machine learning on phase features. This result stems from the intrinsic sensitivity of optical phase morphological and sub-cellular alterations imposed through binding interactions of the cell to adjacent cells and surfaces. Indeed, the expression of VE-cadherin, which forms cell–cell connections, reverts MDA-MB-231 cells from mesenchymal to epithelial morphology and reduces migration ability, illustrating a molecular link between morphology and functional motility (48). Cell lines were distinguished using QPI and machine learning even on isolated cells in suspension (8,49), when cell–substrate attachment information is not present, indicating significant value to phase parameters relating to cell volume, dry mass, and subcellular texture. In contrast, optical phase imaging and classification of substrate-adherent cells provides additional information about cell–cell and cell–substrate interactions. Whole-slide digital pathology (50) and a high-throughput well-plate culture imaging techniques (51,52) require robust, automated methods to identify cells of clinical interest. The classification accuracy from this study of 90–100% is high, similar to other reported values using machine-learning algorithms trained on optical phase map data (8,49,53). A leave-one-out classification analysis (Supporting Information Table S2) suggested predictive power derived from most of the 17 parameters, which could be reduced to five to six principal components, in accordance with other studies (2,8,49). Taken together, these results indicate that phase parameters from the pixel histogram and gray-level co-occurrence matrix, as well as cell outline-based morphology features (8,32,53), help to classify adherent cell lines. Combining “two-dimensional” and “three-dimensional” phase parameters into signatures input to machine learning makes the algorithm more flexible, as geometric/two-dimensional parameters are likely more important for classifying cell lines of dissimilar shape, and higher-order parameters more important for classifying cell lines of similar shape.

Phase parameters correlate with functional motility of wound closure and chemotaxis, suggesting that morphology and subcellular phase structure relate to mechanisms of cell motility. Similarly, DHM phase values have been used to evaluate global changes to epithelial cell layer surface roughness, which correlated to altered cell motility after exposure to a drug delivery system (54). Another study recently correlated cell morphometric parameters from in-incubator time-lapse DHM with transwell migration and invasion assays (32). This study compared two melanoma cell lines, metastatic 1205Lu and non-metastatic WM793, and found a correlation between DHM motility parameters and invasive capacity of the cells. The current study analyzed four different cell lines using a single-shot DHM images. Eccentricity, texture correlation, and central region skew correlated best with wound closure rate between 10 and 24 h (Fig. 5B), whereas phase mean, kurtosis, and central region skew correlated best with transwell chemotaxis (Fig. 6), suggesting that distinct morphology and subcellular phase textures relate to these two types of motility. Cell division, single cell migration, and collective migration all aid in scratch wound closure. In contrast, translocation across the Boyden chamber membrane requires cells to move individually through pores 8 μm in diameter. Quantitative, time-lapse QPI image series acquired during the scratch wound assay could clarify if different processes were happening along the scratch wound border. Rates of cell division (55), collective migration (56), and single-cell migration (32) could be determined from phase maps of the wound border. Further investigation would also yield mechanistic insight into phase parameter sensitivity to chemotaxis by tracking multiple cells in time-series phase maps.

Statistical deviation at different levels of confluence identifies cell line optical phase sensitivity to confluence in culture. From statistical differences of phase parameters with level of confluence (Fig. 7), it can be argued that MCF-7 and GIE cells, of epithelial morphology (57), are the most sensitive to confluence in culture, and HGF, fibroblasts in origin (58), the least. A second, related approach confirmed this, by identifying 18 and 4 outliers for MCF-7 and GIE cells, respectively, using the 99% tolerance interval from the data of Figure 3, which were collected at medium confluence. Kastl et al. quantified biophysical phase-derived parameters from two pancreatic cancer cell lines with controlled changes in media osmolality and culture confluence (15). A higher refractive index but lower average dry mass and cell volume was noted with higher confluence. Correspondingly, this study noted lower adherent cell area of MDA-MDA-231 and HGF cells in conditions of ~70% confluence (Fig. 7A,C).

Since DHM is capable of tracking individual cells and cell populations longitudinally in culture, without the addition of exogenous labels, the findings of this study suggest a use of quantitative phase imaging in phenotypic screening assays. Such an approach provides more sensitive information than cell outline parameters alone. The correlation of cell phase signatures to cell motility through wound closure rate and chemotaxis assays further demonstrates the potential of QPI for phenotypic screening of drug candidates potentially altering cell migration. For example, QPI phase parameters from well-plate cultures of cell lines could be used to screen for an anti-metastasis drug that affects cancer cell chemotaxis, but not immune cell chemotaxis, or normal epithelial and fibroblast wound healing. In the future, DHM and other QPI techniques may be applied to identify drug candidates in phenotypic screens in high-throughput, live-cell assays, with key optical indices from phase signatures identifying altered states of morphology, motility, and proliferation. Further, this study points toward the possibility of developing a single algorithm to identify cell types in unknown clinical specimens across multiple patients and body, organ and tissue sites, by identifying optical phase signatures for cells of one type that are in common among these disparate sources. Functional behavior, such as motility, does not always correlate simply with morphology. The search for optical indices predictive of functional cell behavior will likely require correlated imaging and functional assays, single-cell optical and functional assessment, and validation in preclinical models and clinical biopsies.

Conclusions

In conclusion, this study demonstrates the utility of quantitative phase imaging of cell lines to manage large-scale and high-dimensional quantitative optical phase data for multicell type classification, identification of cells with atypical morphology and phase parameters, and correlation with standard motility assays. Repeated measurements at different levels of confluence determined reproducibility of optical phase signatures as well as cell line sensitivity to neighbors. A tolerance interval around the sum of squared deviations of standardized phase parameters was proposed to identify atypical cells. Results are a step toward an ideal, single classification algorithm that scores unknown cells from disparate sources along a spectrum of epithelial to mesenchymal features, using parameters from quantitative phase imaging.