High-Content Imaging-Based Screening of Microenvironment-Induced Changes to Stem Cells

Abstract

Effective screening methodologies for cells are challenged by the divergent and heterogeneous nature of phenotypes inherent to stem cell cultures, particularly on engineered biomaterial surfaces. In this study, we showcase a high-content, confocal imaging-based methodology to parse single-cell phenotypes by quantifying organizational signatures of specific subcellular reporter proteins and applied this profiling approach to three human stem cell types (embryonic–human embryonic stem cell [hESC], induced pluripotent–induced pluripotent stem cell [iPSC], and mesenchymal–human mesenchymal stem cell [hMSC]). We demonstrate that this method could distinguish self-renewing subpopulations of hESCs and iPSCs from heterogeneous populations. This technique can also provide insights into how incremental changes in biomaterial properties, both physiochemical and mechanical, influence stem cell fates by parsing the organization of stem cell proteins. For example, hMSCs cultured on polymeric films with varying degrees of poly(ethylene glycol) to modulate osteogenic differentiation were parsed using high-content organization of the cytoskeletal protein F-actin. In addition, hMSCs cultured on a self-assembled monolayer platform featuring compositional gradients were screened and descriptors obtained to correlate substrate variations with adipogenic lineage commitment. Taken together, high-content imaging of structurally sensitive proteins can be used as a tool to identify stem cell phenotypes at the single-cell level across a diverse range of culture conditions and microenvironments.

Introduction

Obtaining purified stem cell–derived tissues for biomedical applications presents significant challenges due to the heterogeneity inherent to stem cell cultures as well as the dynamic nature of their responses to their microenvironments. Several tools have emerged to precisely characterize how stem cells react to various stimuli (e.g., growth factors or externally applied forces) and substrate properties.^1,2 Typical screening methods focus on either the detection of lineage-specific markers or gene expression levels of whole stem cell populations,^3–6 which are difficult to scale down quantitatively to single-cell levels. In addition, due to the end-point nature of these assays, stem cells have to be cultured for several weeks before these methods can be applied to assess cellular responses.

To further accelerate the pace of stem cell studies in engineered microenvironments, high-throughput screening (HTS) approaches have been developed that allow simultaneous analysis of multiple culture conditions within a single test platform. Multiple treatment conditions can now be presented on a single plate, allowing a large array of combinatorial variations to be concurrently screened.⁷ Representative HTS platforms include high-throughput (HT) extracellular matrix microarrays,⁸ fluorescence-based polymer screening,⁹ and lab-on-chip substrates.¹⁰ Although these methodologies permit a higher number of conditions to be tested in parallel, their ability to parse cellular responses to different stimuli is still limited by the end-point nature of biological assays and low stability of imprinted arrays. Thus, a major need exists for development of profiling tools to screen early cell phenotypic responses and forecast long-term cell behaviors.

With the advent of genomics, gene microarray studies can provide differences in gene expression of thousands of transcripts across many conditions.^3,11 Although this increases the amount of information that can be obtained from stem cell cultures, the resulting data do not account for the innate heterogeneous nature of stem cell populations.¹² Gene expression of cell populations has been shown to be strikingly diverse when compared with the average gene expression of many single cells representative of a cell population.¹² Methods to distinguish individual cells from a population would be extremely advantageous for selecting cells that have potential utility in tissue engineering applications.

Our laboratory has proposed a high-content imaging-based profiling methodology that can characterize how cells respond to different microenvironments on a single-cell level.^13–16 As a first proof of concept, to characterize cell-adhesive responses to alterations in substrate topography, Saos-2 cells engineered with a green fluorescent protein (GFP) reporter for farnesylation (GFP-f) were cultured on substrates of different roughness and subsequently used to relate cell morphology to surface properties in an HT manner.¹⁶ More recently, a quantitative analysis of the early (24-h) actin morphology of human mesenchymal stem cells (hMSCs) was used to predict downstream osteogenic differentiation.¹⁴ In this study, we extended these approaches and applied our high-content image analysis to evaluate single-cell responses of various human stem cell types (adult mesenchymal, induced pluripotent, and embryonic) to extracellular stimuli and various microenvironments. Using high-resolution confocal images of biologically relevant nuclear and cytoskeletal protein markers, we identified morphometric features that are distinct to culture conditions and stem cell subtypes for a given culture group. Specifically, we examined the ability of the technique to distinguish between high-dimensional features of the nuclear organization in spontaneously differentiating human embryonic stem cells (hESCs) at the periphery versus the putatively undifferentiated hESCs at the core of proliferating colonies. Similarly, induced pluripotent stem cell (iPSC) subpopulations with varying degrees of plasticity were identified.

This method also provided insights into how biomaterial properties, both physiochemical and mechanical, influence the high-dimensional organization of intracellular stem cell proteins. For example, differences in the actin cytoskeletal organization of hMSCs cultured on films with varying degrees of poly(ethylene glycol) (PEG) were elucidated. In addition, to screen different engineered materials, hMSCs were cultured on a self-assembled monolayer (SAM) platform that features a gradient of multiple substrate conditions within a single chip. hMSCs cultured in milieu that favored adipogenic differentiation featured nuclear mitotic apparatus protein (NuMA) signatures that were markedly different from those cultured in conditions that favored pluripotency, highlighting the possibility of using high-dimensional organizational biology as a probe for phenotypic stem cell screening.

Materials and Methods

Fabrication of SAM Gradients

Glass microscope slides were rinsed with ethanol, blown dry with nitrogen, and exposed to UV radiation for 15 min to create a clean hydroxide surface layer. Slides were then rinsed with toluene and immersed in a 2.5% solution of n-octyldimethylchlorosilane in toluene. SAMs made from n-octyldimethylchlorosilane were deposited onto clean oxide surfaces, and the SAM-coated slides were placed on a motorized stage beneath the slit aperture of a UV lamp. A range of UV exposure times was obtained by decelerating the motion of the stage. The rise in UV exposure time led to increasing amounts of ozone-derived oxidation of the n-octyldimethylchlorosilane SAM, generating a gradient in surface energy across the slide.

Preparation of Polymer Films Copolymerized with PEG

Poly(desaminotyrosyl tyrosine ethyl ester carbonate)-co-x%PEG polymers were dissolved into a 1.5% (v/v) methanol in methylene chloride solvent solution, resulting in a 1% (wt/v) polymer solution. Polymer solutions were then spin-coated onto 12-mm diameter glass coverslips. Spin coating was conducted at 3000 rpm for 30 s. Five films of increasing PEG content were prepared: 0%, 2%, 3%, 4%, and 8%. Prepared films were stored in a dessicator, and prior to culturing stem cells, films were sterilized with UV light for 900 s.

Cell Culture

National Institutes of Health (NIH)–approved hESCs (H9 line) and iPSCs were commercially obtained from WiCell Research Institute (Madison, WI). Undifferentiated hESCs and iPSCs were routinely maintained under feeder-free conditions on Matrigel-coated dishes in mTeSR-1 media (Stem Cell Technologies, Vancouver, BC, Canada), as previously described.¹⁷ For immunostaining and high-content imaging studies, clusters of hESCs or iPSCs were passaged with dispase and plated in mTeSR-1 media onto LabTek (Nunc, Naperville, IL) multiwell chambers. To induce early neural differentiation of hESCs or spontaneous differentiation of iPSCs, we switched media 1 day postplating to N2SFM (Dulbecco’s modified Eagle’s medium [DMEM]/F12 with L-glutamine, 1% N2 Supplement, 1%, nonessential amino acids, 2 µg/mL heparin, 0.5% penicillin/streptomycin) or to EB20 (DMEM/F12 with L-glutamine, 20% fetal bovine serum [FBS], 0.5% penicillin/streptomycin), respectively. Both cultures were maintained for 2 to 3 days prior to fixing and staining.

hMSCs were obtained from Texas A&M University (College Station, TX). Cells were expanded in T-75 flasks in a humidity-controlled environment under 5% CO₂ and 37 °C and fed every 3 to 4 days with growth media (basal culture condition, BA) supplemented with commercial SingleQuot’s (catalog number PT-3001; Lonza, Basel, Switzerland). Cells were received at passage 1 and used for up to four passages. Cells were subcultured upon reaching 80% confluence. Osteogenic (OS) and adipogenic (AD) induction media were reconstituted as per the manufacturer (Lonza). Mixed AD/OS media were prepared by combining Lonza hMSC AD media and Lonza hMSC OS media in a 1:1 ratio. Adipogenic media in both the AD and mix conditions were cycled with a 3-day induction followed by 1-day maintenance.

Stem Cell Staining and Immunocytochemistry

Stem cells were first fixed with 4% paraformaldehyde (Electron Microscopy Sciences, Hatfield, PA) for 15 min. Then, a 30-min blocking and permeabilization step was performed with a 0.1% Triton X-100 (Sigma, St. Louis, MO) / 5% normal goat serum (MP Biomedicals, Solon, OH) solution in phosphate-buffered saline (PBS; Lonza). After two washes with blocking buffer (5% NGS in PBS), primary NuMA (ab36999; Abcam, Cambridge, UK) antibodies in blocking buffer at a 1:500 ratio were added overnight at 4 °C. Three 15-min washes in blocking buffer were then performed. Secondary antibodies (Alexa Fluor; Invitrogen, Carlsbad, CA) with different fluorophores and corresponding isotype controls in blocking buffer at a 1:250 ratio were added for 2 h at room temperature. Three 15-min washes in blocking buffer were then performed. To label the actin cytoskeleton, cells were fixed and stained with Alexa Fluor 488 phalloidin (Invitrogen) per the manufacturer’s instructions. All samples were counterstained with 1 µg/mL DAPI (Sigma) in PBS. Similar staining procedures were followed for Oct4 (MAB4401; Millipore, Billerica, MA), Sox2 (MAB4343; Millipore), and CD90 (15-0909-42; eBioscience, San Diego, CA).

High-Content Imaging of Stem Cells

All samples were imaged under a 63× immersion objective (NA = 1.3) with a Leica TCS SP2 system (Leica Microsystems, Inc., Wetzlar, Germany). Average projections of 15-µm-thick image sets of hESC and iPSC cultures were acquired from images at 2-µm intervals. To image hMSCs in the SAM gradients, tile scans of the cells were taken using the NuMA protein and DAPI channels. Since surface hydrophobicity varied from left to right, hMSCs attached at different focal planes on the tile. Thus, the entire slide was divided into multiple sections spaced 5 mm apart. The full spread of the gradient was approximately 35 mm in length, so the imaging was discretized over approximately seven tiles. To image hMSCs in PEG-containing films, the films were first mounted onto microscope slides with Fluorogel (Electron Microscopy Sciences) prior to imaging.

Numerical Descriptor Acquisition and Analysis from Confocal Images

Image Pro Plus Version 7.0 (Media Cybernetics, Bethesda, MD) was used for image analysis. To isolate regions of interest (ROIs) for single-cell descriptor acquisition, nuclear and cytoskeletal masks were first made using the DAPI and actin cytoskeleton channels, respectively. These masks were then superimposed to either nuclear or cytoskeletal channels to isolate signal from NuMA and actin protein channels, respectively. Forty-three numerical descriptors were then extracted for each cell’s NuMA or actin protein signal. A list of the 43 descriptors calculated along with their definitions is provided in Supplementary Figure S1. These descriptors represent quantifiable measurements of cytoskeletal and nuclear protein morphology and organization by including shape, intensity, and texture-based features. For single-cell–based functional marker expression analyses, Oct4, Sox2, and CD90 mean signal intensity values were calculated for each cell using cytoskeletal masks as ROIs.

Descriptors from two different groups identified by the user (e.g., OCT4-expressing vs non–OCT4-expressing cells in a heterogeneous population) were exported to Matlab (MathWorks, Natick, MA) for analysis (Suppl. Fig. S2). First, principal component analysis (PCA) was used to linearly reduce 43 descriptors down to three principal components. This resulted in a plot in which each point represents a stem cell in a 3D space where each axis consists of the combined features of either NuMA or actin protein in each analyzed cell. The location of each point is unique to the descriptor values for that particular cell. To assess how different the descriptor values are across two subpopulations, a support vector machine (SVM) was used to calculate sensitivity, specificity, and accuracy using 10-fold cross-validation (Suppl. Fig. S2). Unless otherwise noted, error reported on SVM classification represents the standard deviation for N = 50 pseudoexperiments (number of iterations using original data set).

Statistical Analysis

Statistical analysis was performed on morphometric parameters using SPSS Version 16.0 (SPSS, Inc., an IBM Company, Chicago, IL) and included analysis of variance (ANOVA) with Tukey’s honestly significant difference (HSD) post hoc method and other multivariate statistical tools. The differences were considered significant for p < 0.05 unless otherwise noted. Error bars indicate the standard uncertainty around the mean.

Results

Overview of Single-Cell High-Content Imaging and Computational Modeling

To acquire morphological information indicative of cell state, several stem cell types cultured in various conditions were imaged via confocal microscopy. Fixed cellular samples were immunolabeled with antibodies specific to cytoskeletal and nuclear proteins, as highlighted in Figure 1A, to extract numerical descriptors from reporter proteins. First, each image was split into channels corresponding to nuclear and cytoskeletal proteins of interest (Fig. 1B). Next, single-cell segmentation was accomplished by defining ROIs for the nuclear and intracellular space. To create the nuclear ROIs, DAPI-stained images were subject to a series of image-processing steps that included Gauss filtering, contrast enhancement, and fluorescence-based thresholding followed by binarization (Fig. 1C). Similarly, the actin channel was used to create intracellular ROIs. To isolate the reporter protein’s signal at single-cell level, both nuclear and intracellular masks were superimposed onto nuclear and cytoskeletal channels, respectively.

Figure 1.

Overview of single-cell imaging, feature extraction, and computational modeling. (A) Image of a stem cell labeled with actin (green) and nuclear mitotic apparatus protein (NuMA) (teal). (B) Channels of nuclear and cytoskeletal proteins of interest are separated prior to feature extraction. (C) To define the nuclear space, a mask using the DAPI channel was generated. (D) The nuclear mask was superimposed onto the NuMA protein (teal) channel to yield 43 shape, intensity, and organizational descriptors (listed in the red dotted rectangle). (E) To perform binary classification, descriptor sets from two different conditions were acquired. (F) Dimensionality reduction (principal component analysis) was applied to generate combinations of descriptors that define the subcellular state of the two conditions. Stem cell population parsing efficiency was characterized by calculating sensitivity, specificity, and accuracy using support vector machine (SVM) classification. Scale bars: A–C = 25 µm, D = 7.5 µm.

Next, 43 numerical shape, intensity, and texture-based descriptors of each isolated reporter protein signal were acquired for each cell (Fig. 1D and Suppl. Fig. S1). Numerical descriptors of cells in two distinct groups were then obtained and merged into a “feature set” (Fig. 1E). PCA was then employed to reduce the 43 descriptors from stem cells cultured in at least two different conditions down to three dimensions, which are derived from a linear combination of the 43 original descriptors (Fig. 1F). These dimensions, termed principal components (PCs), are orthogonal from one another and account for most of the variance in the binary data set. Furthermore, to evaluate the subcellular feature differences between two selected cell subpopulations, an SVM classifier was used, which used k-fold cross-validation to define test sets and training sets for each condition. SVM output sensitivity, specificity, and accuracy for each analysis. In addition, a hyperplane that best separates the two populations in the PCA domain was generated for visualization purposes. All of the analyses presented were conducted using a 10-fold k-fold cross-validation and repeated 50 times (N = 50) to acquire reported error in the form of standard deviation.

hESC Nuclear Features Are Reflective of Phenotypic States

To realize the potential profiling capabilities of our imaging-based approach, we first identified differences in NuMA protein-based nuclear features of two subpopulations of hESC cultures: pluripotent and lineage-committed hESCs, as denoted by both pluripotency marker Oct4 expression and cellular morphology. hESCs in colonies were immunolabeled by antibodies specific for NuMA and Oct4 and were counterstained with DAPI (Fig. 2A). After labeling these colonies, we noticed that Oct4 (an hESC pluripotency marker) expression was strongest within the hESC colonies (indicative of embryoid bodies), whereas Oct4 expression noticeably weakened away from the center of the hESC colonies, as evident from Figure 2A. By visual inspection, low Oct4^low cells are mainly located outside the cell clusters (bulk region of colonies), whereas Oct4^high cells reside inside the bulk regions of colonies (Fig. 2A, third panel). The observed Oct4 expression pattern, as denoted by quantitative mean fluorescence intensity (MFI) after image-based analysis, demonstrated a bimodal distribution pattern when plotted (Fig. 2B). A k-means clustering algorithm determined the boundaries of two subpopulations, resulting in an MFI threshold value of 45 (validated by manual gating performed on isotype controls, showing a baseline MFI = 42 to cover 95% of all negative controls). Tracking the locations of the cells over the lifetime of the culture, it was further confirmed that cells with high Oct4 and low Oct4 regions were located within the colony and outside the colony, respectively. This served as a functional indicator and supervisor of two distinctive subpopulations to perform nuclear descriptor analysis and classification.

Figure 2.

High-dimensional nuclear mitotic apparatus protein (NuMA) organization correlates with Oct4 expression in human embryonic stem cell (hESC) cultures. (A) hESCs were immunolabeled with antibodies specific for NuMA (green) and Oct4 (red) and counterstained with DAPI (blue). Blue and red arrows highlight the NuMA and Oct4 signal of hESCs that would be classified as high Oct4 and low Oct4 expression, respectively. (B) k-means clustering of Oct4 fluorescence was used to determine low Oct4 and high Oct4 subpopulations. (C) Principal component analysis plot of numerical NuMA descriptors of low Oct4 versus high Oct4 cells shows that cells in each population are distinct based on 3D plots of high-dimensional descriptors. Support vector machine statistics yielded sensitivity, specificity, and accuracy of 0.93 ± 0.03, 0.86 ± 0.06, and 0.91 ± 0.01, respectively. PC, principal component. Scale bars: 100 µm.

For each cell in the high Oct4 and low Oct4 subpopulations, 43 quantitative descriptors of the NuMA protein were acquired. Dimension reduction (PCA) followed by an SVM classifier was used to classify cell subpopulations that were outside the colony (low Oct4 regions) versus inside the colony (high Oct4 regions). The PCA plot showed that subpopulations expressing high Oct4 (blue circles) and low Oct4 (red circles) were in small clusters, suggesting that the NuMA morphology of hESCs in the subpopulations is highly homogeneous (Fig. 2C). SVM classification of the two populations resulted in a sensitivity of 0.93 ± 0.03, a specificity of 0.86 ± 0.06, and an overall accuracy of 0.91 ± 0.01. Our results show that PCA together with SVM classification is capable of identifying and parsing differentiating versus pluripotent hESCs in culture.

Nuclear Features of Induced Pluripotent Stem Cells Are Unique to Sox2 Expression

Next, we expanded the utility of our imaging-based profiling methodology to iPSC phenotypes. Differences in nuclear features of iPSCs, based on different degrees of pluripotency, were classified. To determine pluripotency, Sox2, a transcription factor essential for stem cell self-renewal, was used. iPSCs were immunolabeled for antibodies specific for NuMA and Sox2 and were counterstained with DAPI (Fig. 3A). Sox2 expression (represented by MFI values) of iPSCs was quantified via imaging-based analysis of the red channel, and the MFI cutoff threshold value determined from the isotype controls (baseline MFI = 42 to cover 95% of all negative controls) was used to identify Sox2^high and Sox2^low subpopulations (Fig. 3B) in heterogeneous iPSC colonies.

Figure 3.

High-dimensional organization of nuclear mitotic apparatus protein (NuMA) in induced pluripotent stem cells (iPSCs) correlates with pluripotency marker, Sox2. (A) iPSCs were immunolabeled with antibodies specific for NuMA (green) and Sox2 (red) and counterstained with DAPI (blue). Sox2 fluorescence was used to determine subpopulations of high (Sox2^high) versus low (Sox2^low) pluripotency. Red and blue arrows highlight the NuMA and Sox2 signal of iPSCs that would be classified as Sox2^low and Sox2^high, respectively. iPSCs to the left of the red dotted line in B were binned as Sox2^low and the remainder as Sox2^high. (C) Principal component analysis plot of numerical NuMA descriptors of cells expressing high Sox2 versus low Sox2 expressing iPSCs show that cells in each population are morphologically different. Support vector machine statistics yielded sensitivity, specificity, and accuracy of 0.91 ± 0.05, 0.82 ± 0.11, and 0.87 ± 0.03, respectively. PC, principal component. Scale bars: 25 µm.

For each cell in the Sox2 expression group (Sox2^high, n = 75; Sox2^low, n = 153), 43 nuclear descriptors of the NuMA protein were attained. PCA was then used to reduce the dimensionality of nuclear features for cell subpopulations that were Sox2^high versus Sox2^low. PCA plot shows that the Sox2^high (blue circles) and Sox2^low (red circles) subpopulations primarily centralized in respective single clusters, inferring that nuclear features of the iPSCs were highly homogeneous within the same Sox2 expression group (Fig. 3C). In contrast, NuMA-based nuclear features between Sox2^high and Sox2^low subpopulations were found to be distinct, as indicated by the hyperplane in Figure 3C. SVM classification of the two populations resulted in a sensitivity of 0.91 ± 0.05, a specificity of 0.82 ± 0.11, and an overall accuracy of 0.87 ± 0.03. Similar to our hESC imaging-based analysis, PCA in combination with SVM classification was effective at identifying and parsing pluripotent (Sox2^high) iPSCs from those with low Sox2 expression, indicative of a low-pluripotency phenotype.

PEG Content of Polymer Films Governs Minute Changes in Cytoskeletal Organization of hMSCs

Next, hMSCs were used as a test case to adapt our high-content imaging-based method to cultures on synthetic biomaterials with systematically varied hydrophobicity and protein adsorption behaviors. To achieve differences in hydrophobicity, polyethylene glycol-co-poly(desaminotyrosyl tyrosine ethyl ester [DTE] carbonate) copolymers with varying degrees of PEG were fabricated into films. Copolymers with higher PEG content resulted in decreased hydrophobicity. Several literature accounts suggest that PEG has a vital role in various stem cell functions. For example, Briggs et al.¹⁸ showed that hMSCs cultured in poly(DTE carbonate) films with low PEG content exhibited increased osteogenic marker expression. Higher levels of PEG content have also been attributed to highly selective protein adsorption and cell motility.¹⁹ Although the influence of PEG content on cell function has been widely studied, high-content differences in the organization of cytoskeletal proteins such as F-actin have not been examined.

Cells were cultured on tyrosine-derived films with varying weight percentages of PEG (termed 2%, 3%, 4%, and 8% vs 0% controls). Representative images of the F-actin cytoskeleton for each condition show that there are differences in organization as PEG content increases (Fig. 4A), including a decrease in actin stress fibers (green) and cell size. To quantify these differences, actin mean intensity (actin signal per unit area) and area (size) were acquired and analyzed as traditional low-content features (Fig. 4B). As expected, the overall area has a decreasing trend with increasing PEG content due to increasing hydrophilicity and nonfouling polymer film properties; however, a one-way ANOVA yields no statistical significance (p > 0.05) between the various PEG-containing conditions (2% PEG and above). In addition, when comparing just the signal intensity across all conditions, no statistical significance was found across all the conditions (0% PEG and above). Therefore, using conventional low-content image features such as cell area and cell intensity, we are unable to readily distinguish across PEG-containing conditions. To identify morphological changes, higher content analyses capable of identifying more sensitive changes in morphology are necessary.

Figure 4.

High-content analysis of cytoskeletal organization of human mesenchymal stem cells (hMSCs) cultured on polymer films copolymerized with poly(ethylene glycol) (PEG). (A) Representative images of the actin cytoskeleton of hMSCs cultured on films copolymerized with different degrees of PEG. (B) Plots of area and signal intensity of the actin yield little information about morphological differences across the different PEG-containing biomaterial conditions (error bars represent standard deviation for n = 30 cells per condition). (C) Principal component analysis (PCA) plot of actin descriptors of hMSCs cultured in 2% versus 8% PEG copolymer films yields almost complete parsing between the conditions. (D) PCA plot of two conditions yielding morphologically similar cells, 4% versus 8% PEG copolymer films, yields an improved classification of cell populations based on high-content analysis. (E) Summary of PCA accuracy of condition A (x-axis) versus condition B (points in plot) is shown. Scale bar: 25 µm.

Using our imaging-based profiling approach, a PCA plot comparing copolymers with 2% versus 8% PEG resulted in almost complete separation (0.97 ± 0.01 sensitivity, 0.87 ± 0.02 specificity, and an accuracy of 0.92 ± 0.01) (Fig. 4C). Since differences in actin morphology can be easily visualized, this supports the validity of our proposed classifier. To assess the sensitivity of this classifier, PCA was employed to compare two conditions that cannot otherwise be discerned using a low-content approach (Fig. 4D). Sensitivity, specificity, and accuracy of copolymers with 4% versus 8% PEG yielded values of 0.83 ± 0.02, 0.74 ± 0.02, and 0.78 ± 0.01, respectively. Next, classification of condition A (x-axis) versus condition B (points on plot) of all substrate combinations was employed, and the resulting accuracy was plotted (see Fig. 4E). Using a 10-fold cross-validation, SVM was able to correctly classify differences between most conditions with >80% accuracy and differences between all conditions with >60% accuracy. Sensitivity, specificity, and accuracy with error reported as standard deviation of N = 50 pseudoexperiments for all substrate combinations can be found in Supplementary Figure S3.

High-Content Image Analysis Discerns hMSC Differentiation on Self-Assembled Monolayer Gradient

In our last case study, we applied the high-content profiling approach to a SAM gradient substrate platform, which allowed us to investigate how hMSCs respond to incremental changes of substrate properties. In this study, a COOH/OH gradient via a “click” biofunctionalization described previously,^20–22 with a varying molar ratio of COOH to OH groups on a functionalized glass slide (Fig. 5A), was used. Multipotency of hMSCs cultured on the SAM gradients for 1 week under adipogenic induction was measured via expression of CD90, a protein that decreases in expression when stem cells differentiate to osteoblasts, chondrocytes, or adipocytes.^23,24 CD90 expression across all regions of SAM gradients was low, ranging from 5% to 16% across the SAM gradient when compared with naive stem cell controls (Fig. 5B). As the gradient became more hydrophilic (more –COOH groups, less –OH groups), CD90 expression increased along the gradient, suggesting increased loss of pluripotency due to the progression of adipogenic differentiation, mediated by an increase in substrate hydrophilicity.

Figure 5.

Graded biomaterial screening: human mesenchymal stem cells (hMSCs) cultured on a self-assembled monolayer (SAM)–based hydrophobicity gradient substrate display somewhat variable levels of CD90 expression that can be better captured by analyzing differences in nuclear mitotic apparatus protein (NuMA) organization. Gradients were prepared by increasing levels of ozone-derived oxidation induced by a rise in UV exposure time. (A) This resulted in a SAM glass slide featuring a hydrophilic (COOH-rich) region that incrementally changes to a highly hydrophobic (OH-rich) end. (B) hMSCs express variable CD90 levels along the SAM gradient after a 7-day adipogenic induction. (C) Cells cultured on the hydrophilic end (expressing lowest levels of CD90) had different nuclear features from the control group, which expressed high CD90 levels and was cultured in basal growth media (with support vector machine [SVM] classification results being 0.99 ± 0.01, 0.96 ± 0.02, and 0.98 ± 0.01 for sensitivity, specificity, and accuracy, respectively). (D) SVM classification accuracy demonstrated differences in nuclear features of hMSCs cultured on the SAM gradient when compared with naive hMSC controls.

In parallel, NuMA protein descriptors were acquired for hMSCs cultured in adipogenic induction media on the SAM gradient at 72 h. It is important to note that at 72 h, traditional functional markers fail to denote the onset of lineage commitment.¹⁵ Confocal images of hMSCs immunolabeled with the NuMA protein were acquired in seven locations on the SAM gradient with 5-mm increments. PCA-based nuclear feature dimensionality reduction was then used to reduce the dimensionality of nuclear features for cell populations compared with that of naive hMSCs. Figure 5C shows a comparison made between the –COOH region 5 mm from the hydrophilic end and hMSCs cultured on the SAM chip with basal medium. Further SVM classification was performed on cells at each individual location compared with naive hMSC controls (cultured on the SAM chip in basal medium for 72 h). Classification results showed that cells cultured on the hydrophilic end (with the lowest level of CD90 expression) feature different NuMA protein morphologies versus the naive hMSC controls, with classification sensitivity, specificity, and accuracy of 0.99 ± 0.01, 0.96 ± 0.02, and 0.98 ± 0.01, respectively. Comparisons between all SAM gradient locations compared with naive hMSC controls are shown in Figure 5D, where an incremental decrease of nuclear difference (classification accuracy) was observed on gradient locations ranging from within COOH (hydrophilic)–rich regions to OH (hydrophobic)–rich regions, with the lowest difference value (65%) observed at the hydrophobic end. Sensitivity, specificity, and accuracy with error reported as standard deviation of N = 50 pseudoexperiments for all substrate combinations can be found in Supplementary Figure S4.

Discussion

Typical HTS approaches are insufficient for examining heterogeneous cell cultures as they only provide data on the cell population as a whole. This shortcoming is particularly acute for stem cells, including embryonic, induced pluripotent, and mesenchymal adult stem cells, which can commonly adopt divergent phenotypes as a function of culture time and environmental conditions. We have proposed a high-content imaging-based platform that is capable of parsing numerical morphological descriptors from images of individual stem cells cultured in varied microenvironments. By combining integrated fluorescence imaging, quantitative image analysis, and computational data mining, this approach can provide insights about the organization of various cellular proteins at a single-cell level. In this article, we demonstrate the potential of high-content imaging-based profiling to reveal subpopulations across a wide spectrum of stem cell types in different stages of development and responding to a diverse range of material configurations.

Our high-content analyses of descriptors derived from intracellular reporter proteins support the notion that NuMA and actin are sensitive biological markers for classifying how cells interact with their environmental milieu. Stem cell signaling events are triggered by early cell attachment events, for example, when focal adhesions reinforce receptor-mediated cell adhesions to adsorbed extracellular matrix (ECM) proteins on the substrate.²⁵ Integrin-mediated mechanotransduction results in changes within the cytoskeletal organization, which in turn influences the organization of nucleoskeletal scaffolding and nuclear proteins and also triggers a cascade of signaling pathways that lead to microenvironment-driven cell behaviors, including the switch between cell differentiation and self-renewal.²⁶ Thus, cytoskeletal and nuclear protein morphologies are likely influenced by outside-in signaling pathways stemming from chemical, physical, and biological changes in the microenvironment. In addition, the organization of the NuMA protein and actin may provide information about the cell state. Specifically, for example, NuMA exhibits cell cycle–sensitive distribution as it is essential in mitotic spindle positioning and asymmetric cell division during mitosis.²⁷ Actin stress fiber thickness, density, and actin-mediated cell shape are all influenced by cell adhesion to the underlying substrate^18,19 and have also been linked with lineage commitment.²⁸

Stem cell culture systems are limited by scalable techniques to discern pluripotent hESC and iPSC subpopulations from those that are already lineage committed. Pluripotent cells, established by marker expression, exhibit highly distinguishable NuMA protein features (high sensitivity and specificity) from differentiating stem cells. This suggests that the NuMA protein organization can be used as an indicator of the hESC and iPSC phenotype, either pluripotent or lineage committed, although the latter has not been explicitly addressed in this study. Previously, we have reported that NuMA protein organizational features display distinctive patterns when hMSCs undergo osteogenic versus adipogenic differentiation.¹⁵ Follow-up studies need to be conducted to assess whether lineage commitment of iPSCs and hESCs (e.g., specific cell types from the endoderm, mesoderm, and ectoderm) can be similarly captured through profiling of the NuMA protein. The identification of early nuclear morphological signatures predictive of long-term lineage commitment in iPSCs and hESCs could significantly enhance the throughput of biomaterial screening methods since it typically takes these stem cell lines weeks or even months to differentiate into mature phenotypes.

In regenerative medicine, stem cells are induced to desirable lineages via the use of extracellular stimuli based on soluble factors and engineered substrates with controlled surface chemistry and physical properties. We sought to assess whether outside-in signaling emanating from synthetically engineered biomaterials translates to morphological differences seen in our reporter proteins. As previously shown, hMSCs cultured on tyrosine-derived biodegradable scaffolds containing varying degrees of PEG exhibited differences in actin stress fiber formation under conditions that induce changes in adhesion and differentiation.^18,19 This suggests that stem cells can function as biological probes to discern subtle variations in biomaterial properties that dictate stem cell functions.

Our high-content descriptor-based methods were able to sensitively discern differences in morphology resulting from changes in PEG molar fraction of the tyrosine-derived polycarbonate polymers. Notably, the classification efficiency was lower when comparing cell descriptors across conditions with smaller differences in PEG content. This could be due to the fact that most of the descriptors used in this study are shape based (Suppl. Fig. S5). Higher PEG content results in a decrease in cell size and actin stress fibers. In the future, texture-based descriptors resolving minute differences in the intracellular actin architecture could likely improve the degree of classification.

SAMs of alkanethiols have been modified to present well-defined and controllable surfaces presenting a wide range of chemical properties.²⁹ Specific modifications in surface chemistry alone can differentially modulate hMSC differentiation in a lineage-dependent manner.³⁰ Due to higher ECM protein adsorption at the COOH-rich region (hydrophilic end) of SAM substrates, more cells adhered when compared with the OH-rich regions (hydrophobic end), which in turn altered the rate of adipogenic differentiation on SAM gradient substrates. The fact that these variations were identified as early as 72 h via NuMA profiling highlights the promise of this platform toward early detection of differentiation-inducing materials for biomaterial discovery in regenerative medicine.

In summary, we have demonstrated that heterogeneous phenotypes of stem cell populations can be captured using high-dimensional organizational mapping of intracellular protein reporters. Although the geometries for culture systems explored here are restricted to two dimensions, the high-content imaging approach could be readily extended in the future to cell cultures in varied configurations, ranging from 3D scaffolds to hydrogels. Although the data presented here were primarily derived from cells fixed and labeled with fluorophores, in theory, this approach can be easily applied to reporter cell lines in real time in lieu of static, fixed cells. The high-content imaging methods can also be potentially coupled with newer technologies for biomolecular screening, such as emerging, single-cell gene readout assays to enable multimodal cell state analyses, as well as cell sourcing platforms for isolating cell subpopulations based on higher dimensional biological features.³¹