MSCProfiler: a single cell image processing workflow to investigate mesenchymal stem cell heterogeneity

At the forefront of cell-based therapies are mesenchymal stem cells (MSCs), with over 1000 registered clinical studies under way [1]. These are adult tissue-derived multipotent stem cells, which can be isolated from almost all postnatal organs and tissues [2]. They have a variety of pleiotropic functions such as antiapoptosis, angiogenesis, antifibrosis and chemo-attractive properties, making them a useful tool for a wide variety of therapeutic applications [1,3,4]. An important metric to qualify MSCs as beneficial therapeutic agents is the determination of their safety and risks, if any [4]. Consequently, culturing and studying these cells become important. However, one of the hurdles faced in the translation of in vitro MSC studies is that many of the properties that are studied, especially expression of surface proteins, arise because of the artificial culture environment [5]. Expression of surface markers such as CD90, CD73, CD105 and CD44, which are routinely used for MSC characterization, are limited to only being expressed in vitro and not in situ. A major challenge in the translation of MSCs to clinical application is the inherent heterogeneity of MSCs, which influences its properties of immunomodulation and regeneration [3]. The heterogeneity, arising due to isolation, culture and expansion, warrants an extensive study on various MSC subtypes and their categorization [3,6].

To understand this heterogeneity, a cell-by-cell approach as opposed to conducting studies at a population level would be more advantageous in revealing important and unique properties of various cell types [7]. In this context, single cell analysis platforms such as imaging flow cytometry (IFC) can prove to be a powerful tool. An amalgamation of high-throughput cytometry and microscopy, IFC allows extraction of traditional flow cytometric data and images of each event from fluorescent, brightfield (BF) and laser side scatter/darkfield (DF) channels [8]. The potential of this technology has mostly allowed the exploration of nuclear translocation [9], autophagy [10,11] and detection of DNA damage [12–14]. However, unraveling the immuno-heterogeneity of cells using this technology is still under way. The high-throughput nature of this technology allows the collection of enormous amounts of data from a single cell, which can be used to answer questions of cellular heterogeneity [8] and holds immense potential for data mining and creating well-trained neural networks. However, a large amount of information remains underutilized to a great extent due to the lack of data analysis tools that can extract meaningful information from the images [9].

In this study, we have used the Amnis ImageStream image cytometer which is one of the widely used IFC instruments. Data analysis on Amnis platforms can be done using their proprietary software IDEAS or with analysis tools such as FCS Express (DeNovo Software). Apart from having the advantage of performing traditional flow cytometry data analysis, the IDEAS platform also provides users with ‘wizards’, which provide assay-specific analysis templates for feature finder, enabling sequential analysis without many complications (https://cytekbio.com/pages/imagestream). In addition, specific areas or regions of the cell image can be defined using a ‘mask’ (defined set of pixels in the region of interest) that contain features such as creating nuclear masks or cell surface marker-specific masks. It is also possible to create the masks on IDEAS based on user-defined criteria. This can be done by using the mask manager, which has 13 available functions to create the new mask. However, use of this analytical software is primarily dependent on the experience of the user and therefore brings in bias. This can especially be challenging when understanding immuno-heterogeneity, which can lead to missing out on potentially important features if one does not actively look for them.

The recent trend in the incorporation of machine learning (ML) or artificial intelligence in biology has resulted in the advancement of data analysis approaches [15]. Use of virtual or label-free staining of cells [16,17], use of BF information alone to extract quantitative features of cells [16] and use of ML to analyze IFC data for diagnostics [18] have all introduced a new paradigm. However, some of these methods might require an in-depth knowledge of deep-learning techniques, neural networks or programming languages [19]. In this study, the authors used a user-friendly, freely available software, CellProfiler (available at https://cellprofiler.org/), to create a pipeline that can be used to analyze most of the commonly available image file formats [20].

Here the authors introduce MSCProfiler, a completely automated workflow, to analyze IFC data from different human MSCs such as the stem cells of human exfoliated deciduous teeth (SHEDs) and Wharton's jelly-derived MSCs (WJMSCs). The authors developed the entire workflow using the SHEDs and tested the robustness of the same using WJMSCs. SHEDs were stained with antibodies that recognize surface antigens of these cells, prescribed as minimal criteria for identification of MSCs by the International Society for Cell & Gene Therapy (ISCT) [21]. The authors included two surface markers in their studies: CD44/homing cell adhesion molecule and CD73/ecto-5′-nucleotidase. Conventionally, with routine flow cytometry-based characterization of MSCs, bivariate plots of CD44 and CD73 should represent ≥95% double-positive population (refer to Supplementary Figure 1) to suffice the criteria of MSCs as per ISCT guidelines and should express ≤2% hematopoietic marker CD45 [21]. In the authors' panel, they used the above markers and included SYTOX Green as a live/dead discriminator. Live cells exclude SYTOX Green from their membranes, whereas the nucleus of dying/dead cells whose membrane integrity has been compromised take up the dye [22]. The authors focused on the extraction of feature information from individual BF and/or fluorescent images acquired on the Amnis ImageStream Mk II instrument. This workflow can identify images of live singlets based on the exact boundary of single cells from doublets or aggregates and compute the geometrical parameters such as area, aspect ratio, eccentricity, compactness and even texture features such as, inverse difference moment and entropy. The intensity of CD44 and CD73 antigens were also extracted from the fluorescent channel images and their surface distribution pattern was observed.

Materials & methods

Cell culture

Individual cell cultures of SHEDs at passage 9 (P9) and 12 (P12) and WJMSCs at P6 were maintained for this study. Each cell type were seeded in 100 mm culture dishes at a density of 5000 per cm² and grown in KnockOut Dulbecco’s modified Eagle medium (Gibco) media supplemented with 10% fetal bovine serum (Gibco), 1× Pen-Strep (Gibco) and 1× glutamine (Gibco) and grown at 37°C (humidity conditions) until they reached 90–95% confluency with media changes every 24 or 48 h as required. Cells were then washed once with plain basal media without any supplements and trypsinization was done by adding 0.25% trypsin-EDTA (Gibco) followed by incubation at 37°C for 10 min until all the cells had detached from the culture dish. Trypsin-EDTA was neutralized with basal media and the cells were collected and centrifuged at 1200 r.p.m. for 6 min to obtain the cell pellet.

Preparation of cell suspension for immunostaining

The cell pellet was resuspended in 1 ml media, and the cell count was determined. Cells were then washed twice with staining buffer (2% fetal bovine serum in phosphate-buffered saline). Two percent fetal bovine serum helps to sustain viable cells while in suspension post-trypsinization. A cell suspension of 1 × 10⁶ cells per 50 μl suspension contributed to 100 μl of reaction volume when mixed with antibodies and staining buffer. The amount of antibody added to every tube is described in Table 1.

Antibodies & immunostaining

Live SHEDs were simultaneously stained for three surface antigens with an antibody cocktail consisting of CD44, CD73 and CD45. The list of antibody–fluorochrome conjugates are provided in Supplementary Table 1. Cells were incubated for 30 min, washed twice with staining buffer and finally resuspended in 100 μl of staining buffer. Prior to data acquisition, SYTOX Green (1×) was added to Tubes 1 and 5 (Table 1) and incubated at 37°C for 15 min. SYTOX Green is a nucleic acid stain/dye which helps discriminate between live and dead cells.

To address the biggest concern of spectral spillage in a multicolor immunophenotyping experiment, we prepared the right controls for compensation. The single color controls for every antibody used including the DNA binding dye were prepared. Using the compensation algorithm, the noBF files were generated while acquiring the single color tubes. Cells were stained for all the single color compensation controls. Since MSCs would not express CD45, cells from peripheral blood were used as single color controls. For the viability dye, SYTOX Green, compensation control was prepared by giving heat shock treatment to MSCs at 70°C for 5 min followed by incubation on ice. Table 1 summarized the reaction mix per tube, using -1 × 10⁶ cells/ml cell density.

Acquisition on the Amnis ImageStream Mk II

Cells were acquired on the IFC platform Amnis ImageStream Mk II equipped with one charged couple device (CCD) camera (six channels/detectors) using INSPIRE acquisition software, briefly described in Figure 1A. The detection channels available for use in the instrument based on the fluorochrome conjugates were V450 (Channel 1, 435–505 nm), SYTOX Green (Channel 2, 505–560 nm), phycoerythrin (PE; Channel 3, 560–595 nm), BF images (Channel 4, 595–642 nm), Peridinin Chlorophyll Protein-Cyanine 5.5 (PerCP-Cy5.5) (Channel 5, 642–745 nm), and DF images/ side scatter (Channel 6, 745–780 nm). The instrument configuration and panel design are provided in Supplementary Table 2. Unstained cells were first run to set the baseline correction by adjusting the laser powers such that no autofluorescence could be detected. Lasers were set at 10 mW for 405 nm laser (Channel 1 excitation) and 10 mW for 488 nm laser (Channel 2, 3 and 5 excitation). Sample tubes containing all four colors were run and ≥1,00,000 events were acquired. All the single color tubes were run with the BF and DF parameters turned off and were automatically saved with ‘noBF’ mentioned in their file names. These files were later used to set up the compensation matrix postacquisition. All raw data files were saved in raw image file (.rif) format.

Setting up compensation on IDEAS

As opposed to conventional flow cytometry which relies on voltage gain of individual photomultiplier tubes to collect a signal generated from antibody fluorochrome conjugate, the detectors available on the Amnis ImageStream Mk II are charged couple devices. They have a higher quantum efficiency than photomultiplier tubes, which make them ideal detectors to collect dim fluorescence signals, a prerequisite for good imaging. On this instrument, a pixel-based compensation is performed postacquisition using IDEAS software. The ‘noBF’ files of each color acquired in .rif format were loaded onto IDEAS software to create a compensation matrix to correct the spectral spillage of fluorochromes in respective channels. Once the compensation matrix was set up, they were saved in compensation matrix (.ctm) file format.

IFC data analysis

The IFC data have two components – first, the flow cytometric data conventionally represented by dot plots and, second, the individual images of each dot on the plot. As a comparative study, the IFC data were analyzed using both IDEAS and FCS Express software. The data analysis was performed using the proprietary software IDEAS v6.2 (Amnis Corp., WA, USA). Following hierarchical gating, the population of interest was first identified, after which the image data analysis began. Every cell was demarcated with a number and was visible on the screen/image gallery. ‘noBF’ files were used to set up the compensation matrix. The .rif and .ctm files were then used to create the data analysis file (.daf) and compensated image file (.cif). The downstream analysis was performed using the .daf files. Following a hierarchical gating strategy, first the focused cells were identified by plotting a histogram of normalized frequency versus gradient root mean square value of Channel 4 (BF). This was followed by single cell identification (selected apart from the aggregates/debris) gated from the aspect ratio of Channel 4 versus area of Channel 4. The next plot was to identify live cells (negative for SYTOX) from the Channel 6 (DF) versus Channel 2 plot, followed by CD45 negative cells from the Channel 6 (DF) versus Channel 3 (CD45) plot (exclusion gating strategy), and finally populations positive for both CD44 and CD73 were identified from a bivariate plot between Channel 1 (CD44) and Channel 5 (CD73).

Analysis using FCS Express

The .daf file was read by FCS Express v6 (Research Use Only), flow cytometry data analysis software. It was ensured that both the .daf file and the .cif file for each sample tube were available in the same analysis folder. Using a gating strategy similar to that described in the preceding section, singlets were identified first, followed by live cells. The CD45 negative cells were first identified and gated onto the next plot to show the CD44 and CD73 double positive cells. This software validated the gating strategy applied to the populations before proceeding with single cell image data analysis. A flowchart of the steps involved in supervised analysis platforms are shown in Figure 1B.

Preparing images for CellProfiler analysis

The individual images from Channels 1, 2, 3, 4 and 5 were selected from Export .tif images under the Tools option on IDEAS. The images come with an open microscopy environment (.ome) extension, which were converted to .tiff using Fiji/ImageJ software in an automated fashion using an in-house-built macro (Supplementary Information). These individual images were used in the image processing workflow as described in the later sections, to extract the single cellular features. The workflow was developed using CellProfiler software to analyze the single cell images of MSCs.

MSCProfiler workflow

The individual .tiff images of all the channels were then loaded into the MSCProfiler pipeline through the input modules (features that perform specific tasks). Using the NamesAndTypes module, each channel image was categorized in a way that a singlet set consisted of information on a single cell/event consisting of BF and fluorescent channel images. The automated pipeline was described by broadly classifying them into 11 steps based on the purpose they serve. Individual cell information (i.e., information from all five channels) was processed through all the modules of each component to identify and extract various features such as geometric, texture and intensity values, described in Figure 1C. To begin with, the total number of cells analyzed was 54,356 (SHEDs P9), which were selected from the ‘focused’ cell gate on the IDEAS software. Therefore, a total of 271,780 images, collected from all five channels, were run on MSCProfiler. All 271,780 images were not run through the workflow at one time. The total set was divided into 24 groups, each containing approximately 10,000 images. In order to evaluate the strength of the workflow, roughly 13,000 and 10,000 each of WJMSC (P6) and SHEDs (P12), respectively, were analyzed as well. The steps in the MSCProfiler workflow are as follows:

Results & discussion

Conventional flow cytometry-based analysis versus imaging cytometry analysis

The workflows illustrated in the Figure 1, compared supervised analytical tools with the authors' automated data analysis workflow, MSCProfiler. To understand the robustness of this workflow, they used MSCs from two sources and three different passages: WJMSCs (early passage P6) and SHEDs (later passages P9 and P12). The MSCs in all three passages showed dual expression of the cell surface markers, CD44 and CD73. However, the authors observed distinct populations losing expression of these surface markers toward the later passages (Supplementary Figure 1). The IFC results were analyzed on FCS Express software following the gating strategy as shown in Figure 2A. The next set of analysis following similar flow cytometric logic, as shown in Figure 2B, was performed using IDEAS. The FCS Express analysis revealed that out of 142,171 (the total acquired events) the number of singlets was 10,314, out of which 9733 were live cells (i.e., negative for SYTOX Green), 9714 were CD45 negative and 9404 cells were double positive for both CD44 and CD73, shown in Figure 2A. Similarly, using IDEAS software, out of 142,171 total events, 54,356 were focused cells. From the focused cells, 7900 cells were identified as singlets and 7700 were live cells. Finally, the number of cells positive for both CD44 and CD73 was found to be 7579, as seen in Figure 2B. The gate statistics and the percentages for both the software (FCS express and IDEAS) have been summarized in a tabular format in Figure 2C. The sequential output of IDEAS gating strategy is described in Figure 3A. The visual confirmation of every acquired event during the analysis stage, through the image gallery of IDEAS, served to bolster the gating strategy on the platform. For example, from the image gallery the live cells were confirmed by distinguishing from the images of the cells that had taken up SYTOX Green in Channel 2 for as shown in Figure 3A.

In Figure 4, single cell images as seen in the IDEAS image gallery are shown. In Figure 4A, the width of each cell was also calculated from the BF channel image (top right corner of each image). It displays the image gallery of different sized cells based on their BF information in Figure 4A. The authors observed that the cells could be categorized into three groups based on their cell width ranging between 16–20 μm (small), 20–26 μm (medium) and ≥27 μm (large). In Figure 4B, the CD44 and CD73 double positive expressors are shown. A lack of signal can be seen from the SYTOX and CD45 channels. In addition, to understand the dual expression pattern of the two surface markers CD44 and CD73, the corresponding channel images represented by blue and red, respectively, showed co-expression by using the masking feature of IDEAS software. Similar classification was observed in the image gallery of another passage of the SHEDs (P12) and from a different source of MSCs such as the WJMSCs (Figure 4C).

Analysis using a novel workflow: MSCProfiler

The authors have outlined all the steps that went into building their novel pipeline: the MSCProfiler. The prime benefit of this lay in automation of the analysis workflow, starting from image quality control right up to the parameter estimation and classification of cells, described in Figure 3B. MSCProfiler can be fed with a large number of image sets depending upon the computational bandwidth available to the user. In this study, the datasets were split into batches to perform the analyses to ease out the data size. Post segmentation, the three types of features that were extracted from the imaging modalities were the geometrical and texture features and the intensity values. The ‘Focused Cells’ (54,356) annotated in this pipeline generated images of SHEDs that revealed statistically significant heterogeneous cell populations within a single passage of MSCs (SHEDs P9). Similarly, roughly 10,000 and 13,000 each of SHEDs (P12) and WJMSC (P6) cells, respectively, were also analyzed using the MSCProfiler (Supplementary Figures 2 & 3). The spread or distribution of area (based on number of pixels in a shape) in the cells (SHEDs P9), as seen in Figure 5A & B, demonstrated three different categories of populations, which were small (2200–2999 pixels), medium (3001–3999 pixels) and large (4008–4493 pixels), and most of the cells belonged to the small category. Categorization of the cell types of WJMSCs (P6) and SHEDs (P12) based on cell size are shown in Figure 5C & D. This also clearly indicates the robustness of the automated workflow using different cell types and passages, without human bias. The implication of this for the morphometric parameters was revealed further in the extraction of geometrical features. The shape quantification of the SHEDs enhanced the authors' understanding of the characterization parameters. With the goal of using the minimum necessary measurement features to characterize a single MSC adequately so that it can be unambiguously classified, the authors chose the aspect ratio, eccentricity and compactness of a single cell as discriminators. The performance of the pipeline in determining the shape measurements depended a lot on how the image objects were preprocessed. The distribution of aspect ratio (ratio of image object height vs width) did not show statistical significance among the three categories of cells (small, medium and large), as seen in Figure 6A. Analysis of the annotated pixel values using the Kruskal–Wallis multiple comparisons test demonstrated that the small cells were more circular (close to 1.0) than the medium and large cells from the same passage (p > 0.99). However, the distribution was negatively skewed in the small cells because the whisker and half-box were longer on the lower side of the median than on the upper side.

The eccentricity feature (ratio of the minor axis length to the major axis length of an image object) distributed the pixel values (between 0 and 1) and demonstrated that the small cells had lower eccentricity values than the medium and large cells in the same passage, as shown in Figure 6B. The center of distribution of the box and whisker plots was the lowest of the three distributions in the small cells (median: 0.2–0.4), while in the others the median was between 0.4 and 0.6. The distribution in all three classes was approximately symmetric, as both the half-boxes were almost the same length on both upper and lower sides. According to the statistics, small cells were more circular than the other two categories.

Cell shape measure can be best calculated from descriptors such as mean compactness (ratio of the area of an image object to the area of a circle within the same perimeter). A circle is depicted by a minimum value of 1.0. The larger the compactness, the more irregularities and complexities of the cell boundary. The small cells showed lower compactness values than the medium and large cells, as shown in Figure 6C. The center of distribution of the box and whisker plots was the lowest of the three distributions in the small cells (median: 1.0–1.1), while the medium cells had distributed values between 1.07 and 1.31 and large cells between 1.08 and 1.27. However, the distribution was positively skewed in the small cells because the whisker and half-box were longer on the upper side of the median than on the lower side.

Texture features are also very important computational feature extraction descriptors. They bring about the values between shapes and individual pixel values. An important measure of variation brought about from the Haralick texture features is inverse difference moment. It is a measure of homogeneity in cells, which gets maximized when neighboring image pixels share same values (i.e., while measuring texture analysis, two pixels are considered at a single time, the reference and neighbor pixel). The spatial relationship between the reference and neighbor pixel is calculated to understand the gray level differences. There were stark differences among the three classes of cells in terms of inverse difference moment, as seen in Figure 6D (also refer to Supplementary Figure 3). The small cells were more homogeneous in terms of texture compared to the other two classes of cells- the medium and large cells. It provided high discrimination accuracy for images acquired in motion. This discriminator for local homogeneity is lower in medium and much lower in large cell types. The center of distribution of the box and whisker plots was the highest of the three distributions in the small cells (median: 0.2–0.25), while the medium cells had distributed values close to 0.2 and large cells close to 0.15.

Entropy of population analysis can reveal highly structured cellular patterns. IFC dataset distributions are being utilized for identification of malignancy in other cell types by analyzing the differences in multidimensional distributions of related entropies. The entropy and combined entropy of the small cells (9.0) were lowest compared with the medium (9.5) and large cell types (10.71). The center of distribution of the box and whisker plots in all three sets showed homogeneous entropy patterns. More entropy-based patterns could be identified from the medium and large cells, as described in Figure 6E & F. Application of these textural entropy investigations can evaluate the homogeneity and randomness of gray values within the BF cell images, essentially making it a label-free digital image analysis of single cells.

In terms of biomarker expression, CD73-PerCP Cy5.5 and CD44-V450 expression levels were compared in SHEDs (P9). CD73 intensity of expression was higher in the large cells compared with the medium and small cells, as seen from Figure 7A. The authors observed that the CD44 expression levels were higher than CD73 in all three cell types, represented in Figure 7B. To compare the distribution pattern of both the markers, the authors normalized the cell area from which CD44 and CD73 showed expression, by dividing that area by total area of the cell to obtain percentage expression. Results showed that CD44 was expressed over a larger area when compared with CD73, significantly (p < 0.0001) when analyzed using the Mann–Whitney test shown in Figure 7C. FCS Express software analyzed a population of 0.27% and 0.69% of cells which were non-expressers of CD44 and CD73 in SHEDs P9 and P12, respectively (Supplementary Figure 1). Although MSCProfiler can identify these rare cell populations, a greater number of datasets and additional criteria are required to modify the existing pipeline (part of an ongoing study).

MSC heterogeneity has been documented by many researchers. It can be derived from different sources of tissues or even arise randomly from a clonally dividing cell population. However, what is debatable is whether the appearance of such cellular heterogeneity within the MSC population follows a stochastic or a deterministic process. Identifying heterogeneity among the MSCs has always been a challenge, in which the classical approach using microscopy and standard cytometry assays gives information about their surfaceome but not an exhaustive one, as these stem cells do not express any exclusive signature surface markers. There have been attempts using deep learning and ML methods to explore functional heterogeneity and phenotypic and morphometric classification in MSCs. However, most of these studies have been conducted at the population level using microscopic imaging of MSC in vitro cultures, which restricts the sample size for analyses [23]. Few reports of in silico studies using single cell images of MSCs are under way, and all of these use comprehensive image processing algorithms with extensive knowledge of image processing computational tools [19]. MSCProfiler captures the information of MSC heterogeneity based on the texture features of the single cells from IFC and helps explain the detailed cellular features of the population. The existence of heterogeneity is proof enough to show that studies must not be limited to population level but must also look at single cell expression. This will help researchers come up with more stringent protocols for identifying MSCs best suited for clinical purposes. This brings in the rational of the present study in developing an unsupervised workflow that can assist a stem cell biologist in the appropriate identification and classification of MSCs best fit for specific translational studies.

Feature extraction from images of single cells was the hallmark of this study, which led the authors to develop the MSCProfiler workflow. The aim was to identify postacquisition pattern recognition of MSCs (SHEDs) based on multiple image features of single cells such as aspect ratio, cell texture, shape, surface antigen distribution and intensities of expression of such antigens. The authors characterized the gross surfaceome of the SHEDs in this study by developing a nonsupervised image processing pipeline that can robustly segment and analyze single cell morphologies from BF standalone images. There are multiple robust, high-end screening image analysis programs available that can quantify visual cellular morphotypes by microscopy [15]. However, the critical criteria in any such image analysis tool should be lack of bias, ability to identify image-based aberrations (image blur, debris crowding, autofluorescence, saturation of pixels) and high speed of resolution of individual image objects. This pipeline used images of single cells generated from an imaging flow cytometer (Amnis ImageStream Mk II platform), which collectively gives an advantage and takes care of most of the image-based aberrations, making a single image object for acquired data files. This feature rules out the first concern of shadowing of cellular features in case of a smear or a tissue slice [24]. In this paper, the authors have described an automated protocol implemented in validated open-source software called CellProfiler [24–26] with the capacity to offer a suite of image-based measurement features which can extract quantitative information from images. In addition, the parameters extracted from this workflow can be further used to build an unbiased categorization of singe cells based on the phenotypes using ML approaches. Though such tasks can be programmed using R/Python there are more user-friendly tools like CellProfiler Analyst. In fact, with the workflow described in this paper it is possible to extract cellular parameters in file formats that are supported by CellProfiler Analyst.

Although conventional flow cytometry on its own can be a source of high-throughput screening for large datasets, IFC has been able to plug in the image data output to further enhance its throughput. The authors have strategically validated their pipeline with both conventional and image flow cytometry data and demonstrated the shortcomings in either case. Conventional flow cytometry lacks spatial information of every dot on the analysis plot, while IDEAS on the Amnis platform is proprietary and needs to be customized to meet individual needs. This has been one of the first attempts to screen for cellular heterogeneity of MSCs using morphometric features of single cells.

Conclusion

MSCProfiler is an approach to analyzing IFC data that adds valuable information from the images concurrently with conventional flow cytometric analysis. This would be the value addition to answering the respective biological questions that can be resolved using MSCs. This method is completely automated, and so there is no human bias, which was one of the crucial challenges of the other currently available techniques. It does not require extensive programming knowledge to perform such an intense analysis. Analyzing other sources of MSCs and different passages within the same type highlights the robustness of the workflow developed. However, having a high computational processing capacity to analyze thousands of image datasets should be considered as a caveat while using the MSCProfiler. As this pipeline is dependent on the computational power of the data processing system, it brings along with it the time constraint and the data storage of large file sizes (.daf file generated from the IDEAS software is ~0.1 GB). Our study used only five parameters to be analyzed by the MSCProfiler. In cases where more parameters need to be explored (6–12), other data analytical tools such as Cytominer or customized algorithms have to be employed to classify these image sets. In this study, we had to segregate the images into groups of 10,000–20,000 images for a run time of 2–4 h on the MSCProfiler.

Future perspective

This study has been one of the first attempts to screen cellular heterogeneity of MSCs derived from two different tissue sources using morphometric features of single cells. An automated workflow provided by the MSCProfiler has proved to be an unbiased approach to extracting image texture information from a huge range of single cell images that is not dependent on the instrument generating the data. Our workflow has been able to extract similar morphometric features of different types of MSCs in an unbiased manner. On the other hand, MSCProfiler can also be used to identify different cell populations with the inclusion of additional criteria such as fluorescence signals of the surface markers and correlate the extracted texture features for their identification. This is being evaluated with ongoing experiments on identification of hematopoietic stem cells or ‘blasts’ in human bone marrow samples. Most importantly, MSCProfiler does not require knowledge of any advanced computation language to apply this pipeline to the study of single cells. MSCs have a huge potential in drug discovery and cell-based therapeutic applications. While manufacturing clinical grade MSCs in an expanded scale, the availability of a quality assurance/quality control software such as the MSCProfiler can set comparable and rigorous standards for the production systems for cell culture. The stem cell heterogeneity, arising due to isolation, culture and expansion conditions, warrants a robust tool to set quality standards for these stem cells during production and make the supply more consistent and uniform in an efficient and cost-effective manner.