REGENERATIVE MEDICINE

Transcription

1 REGENERATIVE MEDICINE High Content Imaging of Early Morphological Signatures Predicts Long Term Mineralization Capacity of Human Mesenchymal Stem Cells upon Osteogenic Induction ROSS A. MARKLEIN, JESSICA L. LO SURDO, IAN H. BELLAYR, SANIYA A. GODIL, RAJ K. PURI, STEVEN R. BAUER Key Words. High content imaging Mesenchymal stem cell Osteogenesis Automated microscopy Morphology Cellular and Tissue Therapies Branch, Division of Cellular and Gene Therapies, Center for Biologics Evaluation and Research, US Food and Drug Administration, Silver Spring, Maryland, USA Correspondence: Steven R. Bauer, Ph.D., Division of Cellular and Gene Therapies, Center for Biologics Evaluation and Research, US Food and Drug Administration, New Hampshire Ave. Bldg 72 Rm 3208, Silver Spring, Maryland 20993, USA. Telephone: ; Fax: ; Received May 12, 2015; accepted for publication October 30, 2015; first published online in STEM CELLS EXPRESS February 11, VC AlphaMed Press /2016/$30.00/ /stem.2322 This is an open access article under the terms of the Creative Commons Attribution-NonCommercial-NoDerivs License, which permits use and distribution in any medium, provided the original work is properly cited, the use is non-commercial and no modifications or adaptations are made. ABSTRACT Human bone marrow-derived multipotent mesenchymal stromal cells, often referred to as mesenchymal stem cells (MSCs), represent an attractive cell source for many regenerative medicine applications due to their potential for multi-lineage differentiation, immunomodulation, and paracrine factor secretion. A major complication for current MSC-based therapies is the lack of well-defined characterization methods that can robustly predict how they will perform in a particular in vitro or in vivo setting. Significant advances have been made with identifying molecular markers of MSC quality and potency using multivariate genomic and proteomic approaches, and more recently with advanced techniques incorporating high content imaging to assess highdimensional single cell morphological data. We sought to expand upon current methods of high dimensional morphological analysis by investigating whether short term cell and nuclear morphological profiles of MSCs from multiple donors (at multiple passages) correlated with long term mineralization upon osteogenic induction. Using the combined power of automated high content imaging followed by automated image analysis, we demonstrated that MSC morphology after 3 days was highly correlated with 35 day mineralization and comparable to other methods of MSC osteogenesis assessment (such as alkaline phosphatase activity). We then expanded on this initial morphological characterization and identified morphological features that were highly predictive of mineralization capacities (>90% accuracy) of MSCs from additional donors and different manufacturing techniques using linear discriminant analysis. Together, this work thoroughly demonstrates the predictive power of MSC morphology for mineralization capacity and motivates further studies into MSC morphology as a predictive marker for additional in vitro and in vivo responses. STEM CELLS 2016;34: SIGNIFICANCE STATEMENT This article presents a new approach to assess the quality of mesenchymal stem cells (MSCs) for a given osteogenic assay, as well as a means to compare the effects of different culturing and isolation techniques on MSC behavior. This automated, high content imaging approach could be used to compare characteristics of MSC lots from different laboratories and potentially identify morphological signatures that effectively predict their performance in an osteogenesis bioassay. Furthermore, this article highlights the necessity for quantifying multiple osteogenic assay outcomes as simply gene expression and alkaline phosphatase activity alone were found to be highly variable and poorly correlated with more long term, mature MSC osteogenesis based on the extent of mineralization. INTRODUCTION Human multipotent mesenchymal stromal cells (MSCs, otherwise known as mesenchymal stem cells) have received considerable attention for cellular therapies based on their capacities for differentiation, immunomodulation, and paracrine factor secretion [1]. While a number of studies have demonstrated the utility of MSCs in tissue engineering applications such as bone tissue repair [2, 3], current markers of MSC identity [4 6] do not adequately predict how they will perform in both in vivo and in vitro bioassays [7 9]. Efforts to identify molecular markers predictive STEM CELLS 2016;34: VC This article is a U.S. Government work and is in the public domain in the USA. STEM CELLS published by Wiley Periodicals, Inc. on behalf of AlphaMed Press

2 936 Early MSC Shape Predicts Mineralization Capacity VC AlphaMed Press 2016 of osteogenic potential have produced promising results [10, 11]; however, it is likely that a multivariate approach will be needed to robustly predict the osteogenic potential of MSCs from different donors and expanded using different culturing techniques. High dimensional cell morphological analysis has emerged as a means to thoroughly characterize single cell morphological profiles and has recently been shown to correlate with stem cell differentiation [12 14]. Major advantages of this type of analysis are the ability to automatically acquire thousands of images, quantify multidimensional single cell morphological features, as well as the attractiveness of performing short term cell culture and potentially eliminating cost prohibitive long term in vitro and in vivo experiments [15]. Furthermore, the characterization of cell morphology using well established, high contrast fluorescent cell labeling techniques [16, 17] can be more broadly and cost effectively implemented than traditional immunohistochemical and flow cytometry-based approaches. There is also greater potential for inter- and intra-laboratory interpretation of results as quantification of morphological features is less sensitive to differences in protocols and reagents than quantification of immunohistochemistry [17, 18]. Cell and nuclear morphology are a consequence, as well as a stimulus [19 21], of cell fate commitments and could serve as potential defining attributes for MSC potency. The effect of morphology-directed stem cell lineage specification has been demonstrated in both 2D [20, 21] and 3D [22, 23] and may serve as an early indicator of osteogenic differentiation for MSCs. MSC size has also been shown to increase with passage [8] and donor age [24] so it is possible that there are inherent morphological differences in MSC populations that can explain or predict their differences in potency. Similar to cell morphology, nuclear morphology has also been highlighted as predictive of stem cell behavior [25] and effectively serves as a phenotypic readout of epigenetic and transcriptional cellular events. We therefore developed a technique for characterizing MSC cellular and nuclear morphologies using high content imaging and correlated these high dimensional morphological signatures with long term in vitro mineralization results. Due to the heterogeneity of most MSC populations, high content imaging using automated microscopy was utilized to minimize the likelihood that the observed morphological signature was representative of an MSC subpopulation and to remove bias from selecting representative images. To better characterize our high dimensional morphological dataset, we performed principal component analysis (PCA) to reveal distinct overall differences in early morphological signatures of MSC populations that may correlate with mineralization. PCA has been employed recently with MSC studies to effectively reduce the dimensionality of multivariate results (such as genomics [26]) in order to more readily visualize the greatest differences between experimental groups. After identifying morphological features that correlated with mineralization capacity for multiple MSC donors and passages, we utilized linear discriminant analysis to predict the mineralization capacity of MSC celllines from additional donors and manufacturing techniques. We discovered individual morphological features (and combinations of several features) that were highly predictive of mineralization capacity (>90% accuracy) and represent attractive quality attribute candidates for selecting MSC preparations with desired in vitro and in vivo functions. Identification of morphologically distinct subpopulations with defined functions within a heterogeneous population would allow for future investigations into molecular mechanisms governing morphological phenotypes and MSC behavior. MATERIALS AND METHODS MSC Isolation and Expansion Human bone marrow-derived stem cells were obtained from eight different donors purchased from either Lonza (Walkersville, MD, (167696, , 8F3560, ) or All Cells (Emeryville, CA, (PCBM1632, PCBM1641, PCBM1655, PCBM1662) at passage 2 (see Supporting Information Table 1 for donor specifications). MSC culture and expansion conditions were chosen based on wellestablished protocols [27]. Briefly, MSCs were expanded by plating at a density of 10,500 cells/t175 flask (60 cells/cm 2 ) using standard MSC growth medium: 500 ml a-mem, 6 ml 200 mm l- glutamine, 6 ml 10,000 U/ml penicillin-streptomycin (Life Technologies, Carlsbad, CA, and 100 ml of lot-selected fetal bovine serum (FBS) (JM Bioscience, San Diego, CA, Upon reaching 80% confluence, cells were trypsinized and re-plated at 10,500 cells/t175 flask, designated as one passage. Based on the cell confluence, each passage consisted of approximately 7-9 population doublings. Cell-lines from each donor were continuously expanded for several passages (with no freezing/thawing occurring between passages), with fractions of cells being frozen throughout the expansion process at passages 3, 5, and 7 (P3, P5, and P7) for donors , , 8F3560, PCBM1632, PCBM1641, and PCBM1662. Donor was unable to expand beyond P5 and donor PCBM1655 was unable to expand beyond P3. MSCs were cryopreserved in freezing medium consisting of 30% FBS (JM Bioscience), 5% dimethyl sulfoxide (DMSO, Sigma, St. Louis, MO, 1% l-glutamine, 1% penicillin-streptomycin in a-mem (Life Technologies). Throughout this work, the passage number refers to the number of times the MSCs were trypsinized prior to cryopreservation. Further information about the donors surface marker expression, genomic, epigenetic and proteomic profiles, as well as performance in multiple bioassays (growth, adipogenesis, immunosuppression) has been published [7, 8, 26, 28 30, 53]. MSC Seeding for Morphology, Growth, and Differentiation Analysis For morphology, growth, alkaline phosphatase (ALPL), and mineralization studies, MSCs were seeded at a density of 1,000 cells/well in 12-well plates (Corning, Corning, NY, www. fishersci.com). For gene expression studies, MSCs were seeded at a density of 10,000 cells/well in 6-well plates (Corning). All groups were initially cultured for 24 hours in growth medium and switched to the appropriate experimental medium (growth or osteogenic), which was designated day 0. Osteogenic medium utilized in these studies was purchased from R&D Systems (Minneapolis, MN, and medium was replenished every 2-3 days for growth, ALPL, gene expression, and mineralization groups. Morphological analysis studies consisted of MSCs cultured in growth and osteogenic medium (n 5 4 wells for each condition) for 3 days STEM CELLS

3 Marklein, Lo Surdo, Bellayr et al. 937 Figure 1. High content imaging methodology to predict mineralization using early morphological features. Culture expanded human mesenchymal stem cells (hmscs) from multiple donors and passages are maintained for 3 days in osteogenic induction medium ( osteo ) and non-induction MSC medium ( growth ). MSCs are labeled with FITC-maleimide (green) and Hoecsht (blue) in order to visualize cell and nuclear morphologies, respectively. Following high content imaging, automated cellular and nuclear shape analysis is performed using CellProfiler in order to obtain morphological signatures for each donor/passage. The signatures for a training set of MSC samples are correlated with long-term 35 day mineralization results obtained from the same samples and predictive models are developed using discriminant analysis. MSC samples with unknown mineralization capacities (test set) are assessed using the same high content morphological analysis and multiple predictive models are validated by comparing predicted mineralization capacities with actual observed mineralization capacities. Abbreviation: hmscs, human Mesenchymal Stem Cells. (Fig. 1) and then fixed with 4% paraformaldehyde for 15 minutes. To monitor growth and mineralization in osteogenic medium, samples were cultured in an Incucyte ZOOM (Essen Bioscience, Ann Arbor, MI, ALPL groups were cultured for 14 days in osteogenic medium and then fixed with 4% paraformaldehyde for 15 minutes. Gene expression groups were likewise cultured for 14 days in osteogenic medium followed by sample collection using RLT lysis buffer (Qiagen, Boston, MA, and finally stored at 2808C prior to performing PCR. Staining for Morphology and Differentiation Studies Cell and nuclear morphology were assessed using FITCmaleimide (Invitrogen) and Hoechst (Sigma-Aldrich), respectively. Briefly, day 3 fixed samples were incubated with 20 mm FITC-maleimide for 30 minutes, washed with PBS, incubated with 1 mg/ml Hoechst for 5 minutes, and washed with PBS prior to imaging. Day 14 ALPL samples were stained using Fast Blue RR Salt/Napthol AS-MX Phosphate (Sigma-Aldrich) for 1 hour, washed with PBS, stained with 1 mg/ml Hoecsht for 5 minutes, and washed with PBS prior to imaging. VC AlphaMed Press 2016

4 938 Early MSC Shape Predicts Mineralization Capacity Mineralization samples were incubated with osteogenic medium containing 20 mm Xylenol Orange (Sigma-Aldrich) 24 hours before imaging (day 34) to allow for binding of the dye to calcium deposits. On day 35, the Xylenol Orange solution was washed from the samples and fresh osteogenic medium was added prior to imaging. Automated Microscopy for Morphology and Differentiation Assessment Samples for morphological analysis were imaged using an inverted Nikon Ti-S microscope with automated stage (Prior Scientific, Rockland, MA and filters (Chroma Technology, Bellows Falls, VT, compatible with FITC (cell morphology) and Hoechst (nuclear morphology). For each cell-line/passage group, 40 random 10X images were taken per well for a total of 160 images acquired for each group cultured for 3 days in growth and osteogenic medium. Automated quantification of cellular and nuclear shape features was performed using the freely available image analysis software CellProfiler [31] in order to obtain a unique morphological signature (Fig. 1) for each cell consisting of 41 cellular shape features and 41 nuclear shape features (Supporting Information Table 2). Day 3 cell density measurements (quantified as nuclei per field of view [FOV]) were also extracted from the same images used to obtain the morphological signature. The CellProfiler algorithm (termed pipeline) used to analyze cell and nuclear morphology can be viewed in Supporting Information Table 3. Day 14 ALPL samples were imaged similarly except phase images were acquired to visualize the ALPL stain with corresponding Hoechst fluorescent nuclear staining imaged for the same FOV. CellProfiler was again employed to quantify the positive ALPL staining, as well as the cell density (pipeline shown in Supporting Information Table 4). Percent confluency and mineralized area were quantified using the Incucyte s built-in automated image analysis software. Whole well images were acquired every 6 hours to obtain confluency curves during osteogenic induction. Determination of time to reach 50% confluency is outlined in Supporting Information Figure 1. On day 35, whole well images of Xylenol Orange staining were acquired using the Incucyte from the same wells analyzed for confluency. Positive mineralization was quantified using a user-defined fluorescence threshold, which was applied to each whole well image to provide a percent positive mineralized area. Parameters for both confluency and mineralization processing definitions are shown in Supporting Information Table 5. Gene Expression Quantitative real-time reverse-transcriptase PCR (qrt-pcr) was used to evaluate differences in osteogenic gene expression for each cell-line/passage following osteogenic differentiation. Total RNA was purified from lysed samples using RNEasy kits (Qiagen). cdna was synthesized using the Quantiscript Reverse Transcription Kit according to manufacturer s instructions (Qiagen) at 428C for 15 minutes, followed by inactivation at 958C for 3 minutes. qrt-pcr was performed using the Taqman Gene Expression Master Mix (Life Technologies) per the manufacturer s instructions with the following primer targets (ThermoFisher, Waltham, MA, ALPL (Hs _m1), RUNX2 (runt-related transcription factor 2, Hs _m1), SPP1 (osteopontin, Hs _m1) and IBSP VC AlphaMed Press 2016 (Bone Sialoprotein-1, Hs _m1). 25 ng of cdnaequivalents was added to each reaction. PCR was performed using an Applied Biosystems 7900: 508C for 2 minutes; 958C for 10 minutes; 40 cycles of 958C for 15 seconds and 608C for 1 minute per cycle. All conditions were repeated in triplicate with 20 ml reaction volume per well. Data was analyzed using SDS 2.3 software (ThermoFisher). All target genes were normalized to GUSB (b-glucuronidase, Hs _m1) and fold change was determined following normalization to corresponding cell-line/passage undifferentiated control. PCA, Prediction Modeling, and Statistical Analysis Morphological signatures for each cell-line/passage group were determined by taking the median value of single cell measurements (minimum of 500 cells/group) of cell/nuclear shape features in both growth and osteogenic medium for a total of 164 unique shape features (82 from growth conditions and 82 from osteogenic conditions) for each group. Furthermore, we included the differential (D) of each shape feature by subtracting the median growth value for a group from the median osteogenic value for that same group to add an additional 82 morphological features to the dataset. In total, each cell-line/passage group s unique morphological signature was comprised of 246 shape feature measurements. Determination of significantly different shape features between positive versus negative mineralization groups and growth versus osteogenic medium conditions was performed using GraphPad Prism 6 (GraphPad Software, Inc. La Jolla, CA, multiple t test functionality. PCA, discriminant analysis, and hierarchical clustering were performed using JMP10 (SAS Cary, NC, software using unsupervised and supervised methods. Receiver operating characteristic (ROC) curves were generated using GraphPad Prism 6. All statistical and linear regression analyses were performed using GraphPad Prism 6. RESULTS Cell Line and Passage-Dependent Differences in Early MSC Morphology and Proliferation We first examined cell density and cell spread area to assess differences in MSC cell-lines and the effect of passage on MSC functions as changes in these two characteristics were evident in previous publications from our group [7, 8]. MSCs cultured for 3 days had 82 single cell shape features automatically extracted from images using CellProfiler (Supporting Information Table 2, example analyzed image shown in Supporting Information Fig. 2). A conserved trend of decreased cell density with passage in both osteogenic and growth media was observed for all cell-lines that expanded beyond P3 (Fig. 2A). PCBM1655 failed to expand beyond P3 and exhibited cell density similar to P7 of other donors. Cell spread area (following osteogenic induction) increased with passage for all cell-lines except PCBM1655 (Fig. 2B). In growth conditions, cell spread area similarly increased with passage for all cell-lines except PCBM1641. At P3, PCBM1655 showed a degree of spreading in growth and osteogenic conditions similar to P7 MSCs from other cell-lines. Quantification of confluency during osteogenic induction revealed a reduced proliferative capacity with passage (p <.001) as indicated by STEM CELLS

5 Marklein, Lo Surdo, Bellayr et al. 939 Figure 2. Eight mesenchymal stem cell lines express passage dependent increase in cell area and decrease in proliferative capacity. (A): Average cell density per field of view (Nuclei/FOV) determined at 3 days in both growth or osteogenic medium using nuclear Hoechst stain. Initial cell density indicated by dashed line. Bar colors indicate cell passage. (B): Cell spread area quantified after 3 days in growth or osteogenic medium. (C): Percent confluency of samples after 150 hours in osteogenic induction medium assessed by whole well automated Incucyte imaging software. (D): Culture time to reach 50% confluence in osteogenic medium using whole well automated Incucyte imaging software. All values presented as mean 6 standard deviations of n 5 4 wells for each donor/passage. *, p <.0001; 1, p <.001; $, p <.01 significantly different from donor-matched p3 samples. Abreviation: FOV, field of view. percent confluency at 150 hours and the time to reach 50% confluency (Fig. 2C, 2D). Reduced proliferative capacity and increased cell size during culture in growth medium was previously shown with these donors to correlate with a decrease in MSC adipogenic potential [7]. Linear regression was performed to identify the day 3 shape features that correlated with MSC growth in osteogenic medium. Supporting Information Figure 3 demonstrates that the day 3 cell area and minor axis length (in osteogenic medium) were more informative of future MSC growth (7 days and beyond) than initial MSC cell density measurements after 3 days of osteogenic induction. Osteogenic Differentiation Gene Expression Profiles Vary Between Cell-Line Donors and Do Not Exhibit Passage Dependence Osteogenic differentiation of MSCs is associated with highly orchestrated expression of genes such as ALPL, RUNX2, SPP1, and IBSP and an increase in their expression following osteogenic induction would demonstrate osteogenic potential [32, VC AlphaMed Press 2016

6 940 Early MSC Shape Predicts Mineralization Capacity Figure 3. Day 14 osteogenic gene expression does not reveal consistent cell line or passage dependent behavior. Mesenchymal stem cells from all donors and passages were cultured for 14 days in osteogenic induction medium and assessed for expression of ALPL, RUNX2, SPP1, and IBSP. Bar colors indicate cell passage. Fold change is relative to undifferentiated day 0 noninduced controls for each cell-line/passage group (shown as a dotted line) and presented as mean 6 standard deviation of n 5 3 wells. *, p <.0001; 1, p <.01 significantly different from donor-matched p3. Abbreviations: ALPL, alkaline phosphatase; IBSP, Bone Sialoprotein-1; RUNX2, runt-related transcription factor 2; SPP1, Osteopontin. 33]. Day 14 osteogenic-induced MSCs were therefore assessed for expression of these osteogenic markers relative to noninduced day 0 expression to monitor the progression of osteogenesis (Fig. 3). ALPL expression was downregulated for almost every cell line/passage at Day 14 (with the exception of P7); however, there was no repeatable pattern of ALPL expression related to passage. For example, ALPL expression after osteogenic induction was significantly higher at P7 compared to P3 (p <.0001) for cell-lines 8F3560, , and PCBM1662, whereas the other cell-lines showed no significant differences in ALPL expression between passages. The majority of cell-lines similarly downregulated RUNX2 with only , 8F3560, , , and PCBM1662 demonstrating any upregulation. Cell-lines , , and significantly decreased in RUNX2 expression with passage (p <.0001), and PCBM1662 demonstrated an increase in RUNX2 with passage (p <.01), whereas all other cell-lines showed no significant changes with increasing passage. SPP1 expression was upregulated over fivefold in all but one group (PCBM1632 P3), and significant decreases in SPP1 expression were observed with later passages for donors 8F3560, , , and PCBM1662 (p <.01). However, in some of these instances the decrease was not monotonic: 8F3560 first increased at P5 before decreasing at P7 and PCBM1662 decreased first at P5 then increased at P7. IBSP expression VC AlphaMed Press 2016 was upregulated in all donors except PCBM1641 and PCBM1655, but the upregulation trends were not consistent as PCBM1632 and decreased with passage and 8F3560, , , and PCBM1662 increased with passage. Differences in the basal expression of osteogenic genes have been shown for cell-lines from different donors and tissue sources [33], so we also compared DC T values for all groups for each gene. Again, we did not find any passagedependent decreases in osteogenic gene expression using this quantification method (data not shown). In summary, no consistent patterns in expression level changes could be correlated with changes in passage. ALPL Activity Decreases with Passage and Correlates with MSC Proliferation ALPL expression is often used as an early indicator for osteogenic differentiation of MSCs and decreases in expression with passage are associated with a loss of osteogenic potential [34, 35]. At day 14, MSC proliferation and ALPL activity were quantified to assess the progression of osteogenic differentiation. This allowed for determination of a relationship between MSC proliferation and ALPL activity as cell proliferation is a hallmark of the earlier stages of osteogenesis [36]. Figure 4A, 4B show representative samples used to quantify cell density while representative ALPL quantitation images are STEM CELLS

7 Marklein, Lo Surdo, Bellayr et al. 941 Figure 4. Consistent decreases in alkaline phosphatase and cell proliferation with passage for all cell lines evident at day 14 using automated microscopy image analysis. Mesenchymal stem cells stimulated with osteogenic medium for 14 days were stained for nuclei (A) and alkaline phosphatase (D) with corresponding post-thresholding images (B, E) analyzed using CellProfiler. Bar colors indicate cell passage. Cell proliferation was assessed via quantification of cell density per FOV (C) and positive alkaline phosphatase expression area (F) shown as mean 6 standard deviation of n 5 4 wells. *, p <.0001 significantly different from donor-matched p3, scale bar mm. Abbreviation: AP, Alkaline Phosphatase; FOV, field of view. shown in Figure 4D, 4E. Using these automated imaging tools, we observed decreases in proliferation (Fig. 4C) and ALPL activity (Fig. 4F) with increased passage for all cell-lines. The cell-lines with the highest ALPL activity (PCBM1632, 8F3560, ) also exhibited the highest degree of proliferation, and a strong correlation (R ) was found between proliferation and ALPL activity (Supporting Information Fig. 4). However, there were some notable exceptions to this trend, such as P3, which proliferated similarly to other groups with high ALPL activity (PCBM1632, 8F3560, , PCBM1641), but demonstrated significantly less ALPL activity (p <.0001). Conversely, P3 MSCs showed high ALPL activity without the observed high proliferation of other celllines with high ALPL activity. Again, MSCs from cell-lines that did not expand to P7 (127756, PCBM1655) proliferated similarly to P7 MSCs from other cell-lines and possessed correspondingly low ALPL activity. Decrease in Mineralization Capacity with Passage for MSCs Induced with Osteogenic Stimuli Evaluation of MSC osteogenesis using only gene expression, ALPL activity, or both can lead to misclassifying an observed response as osteogenic differentiation when the response could potentially be an artifact of the culture conditions. Therefore, we incorporated multiple osteogenic assay methods to more thoroughly demonstrate that the observed mineralization was associated with osteogenic differentiation and was not an artifact of our culture conditions [37 39]. Day 35 mineralization capacity was quantified using whole well imaging (Fig. 5A) with six replicate wells for each cell line. There was variability in the number of wells that showed mineralization in many of the replicate wells. Due to well-to-well variability for samples with any evidence of mineralization, we developed a binary classification system to define groups as positive for mineralization relative to background or negative. The distribution of all samples was used to determine a threshold staining level (Fig. 5B dashed line) that represented the 95% confidence interval of the mean for all sample wells (n total wells). Wells with values above this line were considered positive for mineralization while wells with values below the line were considered negative for mineralization. Mineralization values for all groups are shown in Figure 5C with colors corresponding to the following classification system: positive 5 1 or more positive wells (blue) and negative 5 0 positive wells (red). Determination of each cell-line s class is shown in Figure 5D. MSCs from 8F3560 P3 and P3 are shown in black as the cells had peeled from the dish prior to Day 35. Several initial cell seeding densities were attempted to address this peeling issue, however we were unable to find an optimal density that allowed for long term culture for these particular groups and therefore they are not included in further mineralization assessment. Initial cell seeding density has been shown to directly impact osteogenic differentiation of MSCs [40, 41] and choosing different seeding densities for different groups would introduce a confounding variable in our mineralization assessment. Based on our seeding and culture conditions, four cell-lines mineralized (PCBM1632, , , and 8F3560) while the other four cell-lines (PCBM1641, PCBM1655, PCBM1662, and ) were negative for mineralization at every passage. For the cell-lines that mineralized, there was an overall decrease in mineralization capacity with passage (based on number of positive wells), with differences between cell-lines evident. MSCs from donor VC AlphaMed Press 2016

8 942 Early MSC Shape Predicts Mineralization Capacity Figure 5. Cell-line and passage dependent differences in day 35 mineralization assessed using automated whole well imaging of Xylenol Orange staining. (A): Images from positive Xylenol Orange staining cultures were quantified using Incucyte automated imaging software in order to obtain a percent positive mineralized area. Contribution of well edges to mineralization signal were removed consistently for all wells (pink). (B): Scatterplot of individual whole well measurements (n 5 126) to determine mineralization cutoff value (dashed line). (C): Percent mineralized area shown for every cell-line/passage and color coded to indicate the mineralization Osteo Class shown in D (Positive Pos 5 blue, and Negative Neg 5 red). (D): Table outlining classification of each cell-line/passage group into a mineralization class based on number of positive wells (n 5 6 wells total). Cell-line/passage groups with any positive wells were classified positive, and groups with no positive wells were classified negative. Cell-line/passage groups that peeled prior to mineralization assessment are color coded black (asterisk). Scale bar mm. PCBM1632 demonstrated the highest mineralization capacity as they retained this high capacity at P5 and were the only cells capable of mineralizing at P7. Supervised Subset of Morphological Features Highly Predictive of Mineralization Capacity Using PCA We next aimed to identify early morphological features from our dataset which could serve as future predictors of MSC mineralization capacity. Figure 6 shows multiple potential predictors of mineralization capacity along with their corresponding ROC curves to highlight each feature s predictive value. Day 14 ALPL staining showed a high degree of correlation with mineralization and the area under the curve (AUC) score for the ROC curve was The difference in cell area between osteogenic and growth conditions (DArea Cell ) was slightly less predictive (AUC ), but provides an initial observation that non-mineralizing MSC groups overall possess a greater difference (D) in cell area than mineralizing MSC groups. VC AlphaMed Press 2016 The possibility of correlating multivariate morphological data was explored using PCA. Following PCA on higher dimensional morphological datasets, we observed poor predictive potential of principal component data from unsupervised (all 246 features) and supervised subsets of just the osteogenic and growth shape features (82 features each) with ROC curves following closely with the line of identity (random guess). However, when we performed PCA on a supervised set of 7 morphological features found to be significantly different between positive and negative mineralizing groups (Supporting Information Table 6), we found complete separation (AUC 1.000) based on principal component 1 (PC1), which accounts for the greatest variance in the data (60.5% ). Furthermore, two morphological features within this subset (DMinor Axis Length Cell and DZernike6_4 Nucleus ) were highly correlated with mineralization and possessed AUC scores ( and , respectively) higher than ALPL. It is also worth noting that all 7 features in the supervised PCA were differential; that is, the difference in morphology is more STEM CELLS

9 Marklein, Lo Surdo, Bellayr et al. 943 Figure 6. Correlation of early morphological signatures with day 35 mineralization capacities for original cell-line/passage groups. Eight mineralization predictors shown in left box with individual cell-line/passage groups color-coded based on mineralization capacity (positive 5 blue, negative 5 red). PCA was performed using unsupervised (all shape features), growth or osteogenic shape features only, and a supervised subset of 7 features (Supporting Information Table 6) found to be significantly different between positive and negative mineralization groups. Receiver operating characteristic (ROC) Curves for each mineralization predictor shown in right box with corresponding AUC values for each predictor. For PCA ROC curves, PC1 was used as it accounts for the greatest differences (variances) in morphology between all groups. Abbreviations: AUC, area under the curve; PC1, principal component 1; PCA, principal component analysis. informative of mineralization than the growth or osteogenic values of the associated features. To further emphasize the utility of differential morphological analysis, we performed hierarchical clustering and PCA on the individual growth and osteogenic morphological signatures for each group using unsupervised and supervised methods. Supporting Information Figure 5 shows that the difference in growth and osteogenic morphological signatures, as represented by the separation distance in PC space, was significantly higher in non-mineralizing groups using both unsupervised and supervised methods (p <.05). Prediction of Mineralization Capacity of MSCs from Additional Donors and Manufacturing Techniques Using Discriminant Analysis In order to confirm the predictive value of morphology, we employed discriminant analysis on a training set of MSCs and controls with known mineralization capacities to derive predictive models for a test set of MSCs with unknown mineralization capacities (Supporting Information Table 11). To strengthen our claim that morphology can predict mineralization beyond our original dataset (Supporting Information Table 1), we chose additional MSC preparations for validation that were from different donors, manufacturers, cultured in different media, and expanded using various media formulations (Supporting Information Tables 8, 9). Predictive models were built based on day 14 ALPL expression, single morphological features, as well as multiple models based on high dimensional morphological data (ranging from 3 to 246 features). The graphs in Figure 7 show the accuracy, sensitivity, and specificity for all groups in the overall test set, as well as the accuracy, sensitivity, and specificity for new donors only and expansions only within the test set. Predicted mineralization capacity for the test set groups was compared with the actual mineralization, which is shown in Supporting Information Figure 6. Although ALPL expression correlated well with mineralization in our original dataset, it was a poor predictor of mineralization in our new dataset (ALPL data shown in Supporting Information Fig. 6) with an overall accuracy of 50%, which is VC AlphaMed Press 2016

10 944 Early MSC Shape Predicts Mineralization Capacity equivalent to randomly guessing whether an MSC group would mineralize or not. The cell area differential (DArea Cell ) was more accurate overall (68%) and better at predicting mineralization of the new cells (100%) than expansions of donors from the original dataset (53%). The two morphological features that were the most significantly different between positive and negative mineralization groups in our original dataset were even better at predicting the mineralization of the training set with DMinor Axis Length Cell and DZernike6_4 Nucleus having overall accuracies of 92% and 84%, respectively. When additional morphological features found to be significantly different between positive and negative mineralization groups (Supporting Information Table 10) were included in the discriminant analysis there was high level of accuracy for models based on the top 3 and 4 features (88% and 92% accuracy); however, increasing the number of features resulted in a decrease in accuracy as models based on growth, osteogenic, and unsupervised feature sets were 76%, 68%, and 76% accurate, respectively. As was the case with the original dataset, the morphological features most significantly different between positive and negative mineralization groups for the entire dataset were differential (D). Models based on the top 3-7 features were the only models with 100% sensitivity, meaning they were able to correctly classify all groups in the training set that were positive for mineralization. The model based on DMinor Axis Length Cell was also 100% specific, indicating correct prediction of all negative mineralizing MSC preparation. Plots of individual predictors and their correlation with mineralization for all datasets (test and training) can be seen in Supporting Information Figure 7 along with their respective ROC curves. DISCUSSION Figure 7. Prediction of mineralization capacities of unknown test samples using models constructed from training set of mesenchymal stem cells with known morphological signatures and mineralization capacities. The accuracy, sensitivity, and specificity of each model s predictions are shown for the overall test set (top), new donors only (middle), and expanded cell-lines (bottom). Models based on a supervised set of morphological features are designated as DAx, where x represents the top x most significant features (based on p-values) included in the DA and are listed in Supporting Information Table 10. Training and test set cell-lines can be found in Supporting Information Table 11. Accuracy is defined as (R True Positive Mineralizers 1 R True Negative Mineralizers) / R Total Positive and Negative Mineralizers. Sensitivity is defined as R True Positive Mineralizers / R Total Positive Mineralizers. Specificity is defined as R True Negative Mineralizers / R Total Negative Mineralizers. DA, Discriminant Analysis. VC AlphaMed Press 2016 Although the International Society for Cellular Therapy first published criteria for defining MSCs nearly a decade ago, there is still a lack of universally accepted MSC characteristics that correlate with in vitro and in vivo outcomes. Current methods for assessing MSC osteogenic differentiation in vitro do not often incorporate measurement of multiple osteogenic markers such as gene expression, ALPL activity, and mineral deposition from multiple donor cell sources and passages. Identifying MSC quality attributes that correlate with in vitro osteogenic assay performance could dramatically improve our ability to identify and screen for MSC subpopulations with desired properties for bone tissue engineering. Morphological analysis of MSCs has emerged as a means to identify single cell properties that could effectively predict how cells will respond to a given stimulus or environment. Using high content imaging, we were able to thoroughly assess the morphological signature of each cell-line/passage sample by quantifying high dimensional single cell morphological data for thousands of cells. High content imaging of early MSC cell and nuclear morphologies revealed distinct morphological signatures, which correlated with their long term mineralization capacity. In this study, we have shown that the morphologies adopted by MSCs in the first 3 days of osteogenic induction can be predictive of their mineralization capacity at a later time point. Although MSC proliferation and cell size have been shown to correlate with differentiation STEM CELLS

11 Marklein, Lo Surdo, Bellayr et al. 945 capacity when a limited number of MSC cell-lines were analyzed [35, 42], this was not the case in our study of cell-lines from multiple donors. While subsets of our entire dataset may adhere well to these generalizable behaviors, we found that cell proliferation and cell area did not correlate well with mineralization capacity for our entire dataset (18 individual cell-lines with 8 at multiple passages). Previous methods of morphological assessment have relied exclusively on unsupervised PCA and other multivariate analyses [12 14, 25, 43]. Although inclusion of all cell morphological features in an unsupervised PCA allows for an unbiased morphological signature assessment, this includes features with high degrees of variability and with potentially poor correlation with mineralization capacity. Our results show that supervised PCA is a more useful method to correlate short term morphology with long term differentiation results. Our identification of features significantly different between mineralizing and non-mineralizing groups underscores the importance of including differential morphological features, which represent an effective normalization of osteogenic morphological data to unstimulated growth morphological data. Moreover, when considering the absolute magnitude of these differential morphological features, there was a conserved higher magnitude of the differential for groups with no mineralization capacity compared to positive mineralization groups. This indicates that while there may be a more significant difference in morphology upon osteogenic induction (compared with growth conditions), it does not necessarily imply osteogenic commitment and eventual mineralization. To our knowledge, this represents the first attempt to normalize morphological data and could serve as a useful technique for predicting other stem cell commitments (adipogenesis, chondrogenesis, neurogenesis) based on their normalized (induced vs. non-induced) morphological responses. Following the identification of morphological features that correlated with mineralization in one set of MSCs, we then demonstrated the applicability of these identified features to an even larger dataset of MSCs from different donors that were cultured and expanded using different media. Again, the most highly significant morphological features were differential, with two particular features (DMinor Axis Length Cell and DZernike6_4 Nucleus ) identified as highly significant for the entire dataset and were considerably more accurate (92% and 84% accuracy) than traditional ALPL staining (50% accurate). It was also found that inclusion of additional morphological features beyond those significantly different between mineralizing and non-mineralizing groups (Supporting Information Table 8) resulted in a decrease in model accuracy (likely due to overfitting [44]) when all osteogenic, growth, or both were used in the discriminant analysis. The majority of morphologybased predictions of MSC behavior have relied upon small numbers of MSCs donors [12 14], and our use of MSCs derived from various donors, ages, gender, and manufacturing processes further emphasizes the applicability of our particular high content imaging approach to identify and screen for MSC lots with desired mineralization capacities. The ability to accurately identify donors with mineralization potential after 3 days would allow for a significant reduction in cost and time associated with a full mineralization assay, which relies on expanding cells and differentiating them for 35 days. Non-destructive morphological assessment would allow for direct correlation of single cell morphological results, but current phase imaging techniques require longer culture times than 3 days in order to effectively predict the progenitor cell behavior [13, 14]. Analysis of the Incucyte phase images acquired for confluency quantification during osteogenic induction would allow for direct monitoring of MSC morphology prior to the onset of mineralization. It would also be possible to locally assess MSC mineralization within a well as there were noticeable foci of mineralization present in all wells that showed any evidence of mineralization (Fig. 5A as an example). Identifying unique single cell (or larger supercell [45]) morphological signatures could uncover desired subpopulations of MSCs that could be enriched and further characterized using other techniques (single cell PCR, flow cytometry, next generation sequencing). Monitoring single cell changes in morphology after osteogenic induction using time-lapse microscopy could also provide insight into identifying cells with desired mineralization potential. Another significant finding of our work was that other commonly employed techniques for assessing MSC osteogenesis (gene expression and ALPL activity) were highly variable and did not correlate with the degree of mineralization achieved by each donor/passage group. There was noticeable disparity between ALPL gene expression and activity. ALPL expression was significantly downregulated at day 14 relative to noninduced controls at day 0, while most groups showed positive staining for ALPL at day 14. Other studies involving MSC osteogenesis highlight the existence of gene expression heterogeneity in single cell [46] and clonal populations [47, 48], as well as the presence of transient osteogenic gene expression [49, 50]. While our PCR results cannot effectively discriminate between donors in terms of their mineralization capacity, they can be used to further strengthen the claim that the mineralizing groups were undergoing in vitro osteogenic differentiation and not artifactual mineralization. Specifically, all donors that exhibited any degree of mineralization at any passage (PCBM1632, , 8F3560, and ) showed upregulation in the mature osteogenic differentiation markers IBSP and SPP1. Although high levels of ALPL expression can result in phosphate (and subsequent calcium) precipitation [37, 39, 51], the fact that all donors with high levels of ALPL staining did not mineralize further indicates that our measurements were bona fide MSC-mediated mineralization. While commonly associated with osteogenesis, ALPL can also be detected in undifferentiated MSC cultures and is sometimes used as a marker for stemness in other cell types [52]. Further support that our measurements represent osteogenic activity is shown by several donor/passage groups where only a fraction of the wells mineralized (resulting in a very high coefficient of variation); however, there was consistently high ALPL expression (and low coefficient of variation). This finding emphasizes the importance of multiple bioassays for assessing MSC osteogenic capacity. CONCLUSION Understanding how distinct MSC morphological signatures relate to stem cell heterogeneity and specific biological activities could significantly improve current MSC characterization methods for clinical use in applications such as bone tissue VC AlphaMed Press 2016

12 946 Early MSC Shape Predicts Mineralization Capacity regeneration. Correlation of early morphological signatures with MSC osteogenic capacity has been investigated by many groups, and further improvement on the methods for morphological assessment and applicability across multiple donors is necessary in order to fully demonstrate its utility as a possible predictor of in vivo MSC osteogenic activity in the context of bone tissue engineering. Our development of predictive models of in vitro mineralization based on morphological data followed by successful validation with new MSC cell-lines lays the foundation for further study into identifying morphological signatures that can predict MSC potency. Our high content imaging based approach can be used to potentially screen cell-lines derived from different donors for their mineralization capacity, as well as screen biomaterials or soluble factors that may influence MSC osteogenic potential. This approach could also be utilized for predicting other stem cell fates (adipogenic, chondrogenic) and in applications involving other types of stem cells (adipose tissue-derived, umbilical cord-derived, etc.) Finally, high-dimensional morphometric analysis can be retrospectively combined with previously obtained in vitro or in vivo results in order to determine if there was a characteristic morphological signature associated with a specific outcome. ACKNOWLEDGMENTS The authors would like to acknowledge Eva Rudikoff for her assistance in culturing the MSCs and Drs. John Thomas and Cindy Osborn for reviewing the manuscript. This project was supported in part by Dr. Ross Marklein s appointment to the Research Participation Program at CBER administered by the Oak Ridge Institute for Science and Education through US Department of Education and US Food and Drug Administration. This work was also supported in part by the Food and Drug Administration Modernizing Science grant program, a BARDA grant, a grant from the Medical Countermeasures Initiative and research funds from the Division of Cell and Gene Therapies. AUTHOR CONTRIBUTIONS R.A.M.: conception and design, collection and assembly of data, data analysis and interpretation, manuscript writing, and final approval of manuscript; J.L.L.: conception and design, data analysis and interpretation, manuscript writing, and final approval of manuscript; I.H.B.: data analysis and interpretation, manuscript writing, and final approval of manuscript; S.A.G.: collection and assembly of data, data analysis and interpretation, and final approval of manuscript; R.K.P.: administrative support, provision of study material, and final approval of manuscript; S.R.B.: conception and design, financial support, administrative support, provision of study material, data analysis and interpretation, manuscript writing, and final approval of manuscript. POTENTIAL CONFLICTS OF INTEREST The authors indicate no potential conflicts of interests. REFERENCES 1 Baraniak P, McDevitt T. Stem cell paracrine actions and tissue regeneration. Regen Med 2010;5: Yeatts A, Geibel E, Fears F et al. Human mesenchymal stem cell position within scaffolds influences cell fate during dynamic culture. Biotechnol Bioeng 2012;109: Ciapetti G, Ambrosio L, Marletta G et al. Human bone marrow stromal cells: In vitro expansion and differentiation for bone engineering. Biomaterials 2006;27: Dominici M, Blanc KL, Mueller I et al. Minimal criteria for defining multipotent mesenchymal stromal cells. The International Society for Cellular Therapy position statement. Cytotherapy 2006;8: Krampera M, Galipeau J, Shi Y et al. Immunological characterization of multipotent mesenchymal stromal cells The International Society for Cellular Therapy (ISCT) working proposal. Cytotherapy 2013;15: Mendicino M, Bailey A, Wonnacott K et al. MSC-based product characterization for clinical trials: An FDA perspective. Cell Stem Cell 2014;14: Lo Surdo J, Millis B, Bauer S. Automated microscopy as a quantitative method to measure differences in adipogenic differentiation in preparations of human mesenchymal stromal cells. Cytotherapy 2013;15: Lo Surdo J, Bauer SR. Quantitative approaches to detect donor and passage differences in adipogenic potential and clonogenicity in human bone marrow-derived mesenchymal stem cells. Tissue Eng Part C Methods 2012;18: Lv F-J, Tuan R, Cheung K et al. Concise review: The surface markers and identity of human mesenchymal stem cells. STEM CELLS 2014;32: Mentink A, Hulsman M, Groen N et al. Predicting the therapeutic efficacy of MSC in bone tissue engineering using the molecular marker CADM1. Biomaterials 2013;34: Platt M, Wilder C, Wells A et al. Multipathway kinase signatures of multipotent stromal cells are predictive for osteogenic differentiation: Tissue-specific stem cells. STEM CELLS 2009;27: Treiser M, Yang E, Gordonov S et al. Cytoskeleton-based forecasting of stem cell lineage fates. Proc Natl Acad Sci USA 2010; 107: Matsuoka F, Takeuchi I, Agata H et al. Morphology-based prediction of osteogenic differentiation potential of human mesenchymal stem cells. PLoS ONE 2013;8:e Matsuoka F, Takeuchi I, Agata H et al. Characterization of time-course morphological features for efficient prediction of osteogenic potential in human mesenchymal stem cells. Biotechnol Bioeng 2014;111: Carmen J, Burger S, McCaman M et al. Developing assays to address identity, potency, purity and safety: Cell characterization in cell therapy process development. Regen Med 2012;7: Elliott J, Tona A, Plant A. Comparison of reagents for shape analysis of fixed cells by automated fluorescence microscopy. Cytometry A 2003;52: Mancia A, Elliott J, Halter M et al. Quantitative methods to characterize morphological properties of cell lines. Biotechnol Prog 2012;28: Atha D, Manne U, Grizzle W et al. Standards for immunohistochemical imaging: A protein reference device for biomarker quantitation. J Histochem Cytochem 2010;58: Gao L, McBeath R, Chen CS. Stem cell shape regulates a chondrogenic versus myogenic fate through Rac1 and N-Cadherin. STEM CELLS 2010;28: McBeath R, Pirone D, Nelson C et al. Cell shape, cytoskeletal tension, and RhoA regulate stem cell lineage commitment. Dev Cell 2004;6: Kilian KA, Bugarija B, Lahn BT et al. Geometric cues for directing the differentiation of mesenchymal stem cells. Proc Natl Acad Sci USA 2010;107: Khetan S, Guvendiren M, Legant WR et al. Degradation-mediated cellular traction directs stem cell fate in covalently crosslinked three-dimensional hydrogels. Nat Mater 2013;12: Huebsch N, Arany PR, Mao AS et al. Harnessing traction-mediated manipulation of the cell/matrix interface to control stemcell fate. Nat Mater 2010;9: Zaim M, Karaman S, Cetin G et al. Donor age and long-term culture affect differentiation and proliferation of human bone marrow mesenchymal stem cells. Ann Hematol 2012;91: VC AlphaMed Press 2016 STEM CELLS