Sage Journals: Discover world-class research

Abstract

Due to the rapid adoption and use of human induced pluripotent stem cells (iPSCs) in recent years, there is a need for new technologies that standardize the evaluation of iPSCs to allow the objective comparison of results across different experiments and groups. In this article, we present a noninvasive, fully automated, and analytical system for morphology-based evaluation of iPSC cultures that consists of time-lapse microscopy and novel image analysis software. The presented system acquires low-light phase-contrast images of iPSC growth collected during a period of several days in culture, measures geometrical- and texture-based features of iPSC colonies throughout time, and derives a set of six biologically relevant features to automatically rank the quality of the cell culture. In a study of 94 iPSC cultures, we demonstrated the accuracy of the system by comparing the automated ranking with an independent expert evaluation based on visual review of the time-lapse movies. To our knowledge, this is the first demonstration of a fully automated and objective assessment of iPSC culture quality using noninvasive methods.

Keywords

time-lapse imaging image analysis stem cell quality control noninvasive

Introduction

There has been significant progress in using induced pluripotent stem cells (iPSCs) for modeling of human disease by harvesting somatic cells from patients, reprogramming the somatic cells into a pluripotent state, and then differentiating the pluripotent cells into a specific cell type affected by the disease. Such disease-in-dish models hold great promise for treating a wide range of human diseases, and they are currently used for studying cellular disease mechanisms, drug screening, and toxicity testing.¹ Recent studies have demonstrated that recapitulating a disease phenotype in vitro is feasible for numerous monogenic diseases, confirming the transformative potential of iPSCs toward disease modeling and therapeutic testing.^2–4

After reprogramming, iPSCs are expanded and maintained in culture through cell proliferation and passaging. Once the iPSCs have reached a time point of interest or specific passage number, they are characterized using a variety of assays, such as immunofluorescence, gene expression profiling, karyotyping, and methylation analysis.⁵ Despite the diverse set of tools available to researchers today, there are limitations in the current practice of iPSC characterization. First, in most studies, characterization of iPSCs is performed at sparse time points, typically using antibody staining and requiring the sample to be destroyed. Second, the phenotypic quality of cell cultures is mainly assessed through snapshot visual examinations, which are limited in scope and accuracy. Third, iPSC experiments are prone to a high degree of variability from factors such as cell line stability, culture conditions, reagent lot differences, and user technique. As the use of iPSC technology continues to grow and thousands of new lines are created, there is a need for technologies that can standardize the evaluation of iPSC cultures to allow the objective comparison of results among multiple experiments and different laboratories. This is particularly important because the starting state of an iPSC culture is a determining factor in the eventual fate of the differentiated cells. Efforts toward standardization have been initiated by other groups; for example, recent studies in gene expression profiling have led to the development⁶ and subsequent commercialization of real-time quantitative PCR assays that evaluate iPSC pluripotency and are designed to allow comparison of results against a global “scorecard” reference.

In this article, we present for the first time a noninvasive system for morphology-based evaluation of iPSC cultures via time-lapse microscopy and automated image analysis. We focus on the period of iPSC expansion and maintenance, when iPSCs are vulnerable to sporadic changes as well as variations in handling and culture. We believe that our system enables the fully automated and objective assessment of iPSC quality, which can be extended toward other types of cells. Unlike most other characterization techniques, our morphology-based analysis is noninvasive, thereby allowing samples to be continuously monitored without being disturbed, and it enables the direct evaluation of iPSCs in almost any application.

Materials and Methods

Overview

An overview of our system is illustrated in Figure 1 . The system includes imaging hardware, analytical software, and a cloud-computing platform to efficiently manage and process large quantities of imaging data. The system acquires low-light phase-contrast images of cells collected during a period of several days in culture and in standard dishes. The iPSC culture can be performed following standard culture protocols and in standard cell-culture plates, requiring no additional handling or variation to the common cell-culture practice. The images are automatically uploaded to the cloud (i.e., a computing server). The analytical software then processes images; measures geometric- and texture-based features of iPSC colonies throughout time; and, based on those features, computes a set of six biologically relevant features: iPSC doubling time, the degree of cell compaction, colony border spikiness, sensitivity to media change, the prevalence of dead cells, and the prevalence of spontaneously differentiated cells. The output of the system is the above feature set along with an automated ranking of iPSC cell cultures to good, fair, and poor categories based on a classification tree trained on the feature set extracted from training data (i.e., a set of images manually labeled by an expert).

Figure 1.

Overview of the presented system for induced pluripotent stem cell (iPSC) evaluation based on time-lapse microscopy and automated image analysis. (A) Images of iPSCs in 12-well plates are captured at six different x–y locations per well every 30 min for 3 to 4 days. Images were uploaded in real time to a cloud-based system for data review and high-throughput image analysis. (B) High-level description of the image analysis software: The input is a time-lapse sequence of images, and raw geometric- and texture-based features are measured from the input images by geometric graph representation (in yellow) and cell stage classification, respectively. The measurements are consolidated into a set of six biologically relevant features, resulting in the final output of our automated, noninvasive evaluation: a ranking of iPSC quality defined by good, fair, or poor.

For the purpose of development and validation of the software, we collected time-lapse image data of 94 iPSC cultures (12 separate clones among four iPSC lines) with six observations of each culture, resulting in a total of 564 time-lapse data sets. We used half of the data for training the classification tree to maximize correlation with the expert ranking of each data set, and we used the other half for validation. In the following sections of this article, we describe the details of the system, experiments, and results.

Imaging and Experiment Setup

The imaging system is constructed using 10× phase-contrast imaging objectives (Nikon), high-precision multiwell plate scanners and z-focus motors (Thorlabs), and 1.3 megapixel 2/3-in complementary metal oxide semiconductor monochrome cameras (Pixelink). The microscope is connected to a Windows-based personal computer, and image capture is controlled with custom-developed software. Images are saved in TIFF format with no additional compression. The multiwell plate scanner was configured to work with Tokai Hit stage-top incubators (model WSKM) that include precise temperature regulation via a thermocouple placed in one of the wells and temperature feedback control. Images are acquired using 625-nm low-light phase-contrast illumination.

A typical imaging experiment of iPSC culture involves imaging of cells every 30 min at multiple x–y locations per well to obtain an adequate sample of the cells, for 3–4 days or until the cells reach about 95% confluence. The images are acquired with seven images per z-stack to capture a range of focus. Media changes on the culture are performed during time intervals when the time-lapse images are not being captured.

In a single experiment that includes 12-well scanning and six observations per well, up to 100,000 images can be captured. To manage this large volume of data, directly after capture, images are automatically uploaded to a secure cloud-based system for subsequent processing and analysis. For each data set, a time-lapse movie is also generated with the optimal focal plane extracted from the z-stack.

Image Analysis Software

To analyze the iPSC image sequences, we developed robust and automated image analysis algorithms that capture cell dynamics and morphology throughout time. Different textural patterns of cells are observed in stem-cell colony formation. We have identified seven distinct cell-stage patterns, as shown in Figure 2a : single cells, medium compaction, full compaction, dead cells, differentiated cells, debris, and background. Medium-compaction cells are typically visible in small colonies via the presence of bright phase-contrast halos in between neighboring cells, whereas full-compaction cells are typically visible in medium- to large-sized colonies when cells are flat and in full contact with neighbors. Differentiated cells mostly appear as flat large cells without bright halos. Dead cells appear as bright regions with irregular texture. The prevalence and distribution of each class in the image provide valuable information about the stage, growth pattern, and quality of cells in the culture. For example, in some cases, medium-compacted cells can be a sign of stress or lack of proper culture conditions, whereas full-compacted cells can be a sign of healthy or good-quality cells. Differentiated cells are a sign of poor culture conditions and exhibit a different morphology than iPSCs. Taking advantage of machine-learning techniques in dealing with such data for which an explicit model is impossible to define, we used an extended version of our previously proposed method⁷ to compute textural and intensity-based features from the images, and we built and trained a classifier to predict the label given the computed features. As presented in Maddah and Loewke,⁷ every image is decomposed to overlapping blocks. For every block, the intensity histogram and local binary pattern⁸ are calculated and concatenated, resulting in a one-dimensional feature vector per block. In the training image set in which the images are labeled manually, each block (and therefore each feature vector) is associated with a label. We generate a training matrix from this set, scaled to range between −1 and 1, to train a linear support vector machine (SVM) classifier^9,10 with a Gaussian kernel. Parameters of the classifier are then optimized within a search range for each parameter to obtain the highest classification accuracy in predicting labels of training images using fivefold cross-validation, achieving 98% accuracy. Here, we followed the same approach but with two additional classes for background and debris, and larger block size. Once trained, the classifier is then applied to classify each foreground pixel in a given image to one of the seven cell stages. For regions that are classified as dead cells, further processing is performed to infer the type of region underneath them. For example, if a dead-cell region is surrounded by pixels classified as full compaction, that dead-cell region takes a new label that indicates dead cell on top of full-compaction region. If the dead-cell region is mostly surrounded by the background, it keeps the class label of dead cells. Figure 2b shows the result of classification for four different frames (see also Supplementary Movie S4). Cell-stage classification enables us not only to quantify the compaction pattern throughout time but also to estimate the number of cells in an automated and noninvasive way from phase-contrast images. Given the output of cell-stage classification and an average cell density per class, the total number of cells is estimated by summing over the area of each class weighted by their corresponding cell density.⁷ We then fit the exponential function, y = y₀2^(t−t0)/T, to estimate the cell-doubling time, T, from the image series,⁷ as shown in Figure 3a .

Figure 2.

Classification of induced pluripotent stem cells (iPSCs) into seven different classes using texture information and machine learning. (A) Sample images of the seven different classes: single cell, medium compaction, full compaction, dead cells, differentiated cells, debris, and background. Medium-compaction cells are typically visible in small colonies via the presence of bright phase-contrast halos in between neighboring cells, whereas full-compaction cells are typically visible in medium- to large-sized colonies when cells are flat and in full contact with neighbors. Our classifier was trained to identify these seven classes based on their distinct textural profiles. (B) Output of automated cell-stage classification demonstrated on four images. Our classifier successfully identifies different cell stages for a wide range of scenes that can include different-sized colonies, varying amounts of dead cells and debris, and differentiated cells located outside or within colonies. For regions that are classified as dead cells, further processing is performed to infer the type of region underneath them.

Figure 3.

Automated, noninvasive measurement of induced pluripotent stem cell (iPSC) doubling time based on cell-stage classification. (A) An example of the number of cells throughout time and the corresponding exponential fit. Given the output of cell-stage classification and an average cell density per class, the total number of cells is estimated by summing over the area of each class weighted by their corresponding cell density. We estimate the doubling time from the exponential curve fitted to the cell-count data. (B) Measurements of iPSC doubling time for 12 different clones, measured among multiple experiments under normal culture conditions. By combining time-lapse imaging with state-of-the-art machine learning, we have developed a new approach for noninvasive measurement of cell-doubling time that can be easily retrained for different applications.

To robustly capture the connectivity and dynamics of colonies, we construct a geometric graph in which the nodes of the graph are defined by the uniform sampling of each cluster. The edges are formed by constructing the Delaunay triangulation between the nodes, and they are weighted by the inverse of geodesic distance between the nodes. The geodesic distance is the shortest path between the nodes constrained by the foreground regions ( Fig. 1b ). A graph-based feature that manifested itself as a robust measure of the quality of iPSCs is the percentage of weak edges in the graph, which represents the spikiness of colony borders. Poor organization of colonies (e.g., holes in the colonies) or spikiness of borders results in edges with low weight (high geodesic distance value), due to the absence of a straight line connecting the nodes that passes through the foreground pixels.

We used all the raw geometric- and texture-based features to identify metrics that correlated best with the expert ranking of training data. The resulting biologically relevant features are as follows:

Degree of cell compaction: The percentage of full-compacted cells, averaged among the frames with 70–90% confluence. This period is obtained using the total cell-cluster area and is relative to the field of view of the image.

Doubling time: As described, the doubling time is calculated from a cell-count curve estimated by cell-stage classification.

Sensitivity to media change: We observed that some iPSCs exhibit sensitivity to media change, in which the colony size retracts temporarily. To quantify this feature, we measure the magnitude of drops in the cell-cluster area. The sum of drops throughout time is used as the measure of media sensitivity.

Colony-border spikiness: We use the percentage of weak edges in the graph analysis within the 70–90% confluence period as a measure of the edge spikiness of large growing colonies.

Prevalence of dead cells and debris: A high prevalence of dead cells or debris can also be a sign of low-quality iPSCs. These features are computed by the average percentage of dead cells and debris throughout time.

Prevalence of differentiated cells: Due to nonideal cell-culture conditions or poor quality of iPSC lines, the culture may contain nonpluripotent cells that have spontaneously differentiated. Detecting and quantifying the number of those cells in the culture are determinative factors in assessing the quality of cell lines and their culture. This feature is defined as the average prevalence of differentiated cells throughout time.

After creating measurements of these six features for all training data, we computed their conditional probability distributions with class labels of good, fair, and poor (see Fig. 4a ). Examples of class labels are shown in Figure 4b , which were independently assigned by an expert biologist who visually examined each time-lapse movie. Based on the probability distributions, we defined a set of ranges for each feature that would result in the highest prediction accuracy in our training data, as shown in Figure 4c . Per-position predictions (six observations in each well) were then summarized in a single prediction for each well based on the majority rule. The features with the highest predictive power are the degree of cell compaction and the doubling time, followed by the colony-border spikiness and sensitivity to media change. We found that the prevalence of dead cells and differentiated cells did not increase the accuracy of our prediction in the training data; however, these features still provide relevant information and may be applied to other data sets to identify poor-quality cultures.

Figure 4.

Three-tiered ranking system for induced pluripotent stem cell (iPSC) quality based on automated feature measurement. (A) Probability distributions for each of the six features, labeled by good, fair, and poor, for our training data. The features with highest predictive power are the degree of cell compaction and the doubling time, followed by colony-border spikiness and sensitivity to media change. (B) Example images of good, fair, and poor cultures, based on the ranking system. (C) Parameter thresholds in our ranking system. Based on the probability distributions, we defined a set of ranges for each feature that would result in the highest prediction accuracy in our training data. (D) The performance of our ranking system was 80% accuracy per position and 89% accuracy per well, evaluated on the test data. Our results successfully demonstrate one of the first applications of fully automated and objective assessment of iPSC quality.

Experiment

Cell Culture

Using the Sendai Reprogramming Kit (Life Technologies), we reprogrammed skin fibroblasts purchased from the Coriell Institute (http://ccr.coriell.org)—two healthy patients and two patients with LRRK2 mutations—into iPSCs.^11,12 We selected and tested 16 iPSC clones from the four patient fibroblast lines for consistent human pluripotent colony morphology, presence of normal karyotype, clearance of Sendai virus, and expression of pluripotent markers (Supplementary Figs. S1 and S2). All of the iPSC clones were cultured in E8 medium (Life Technologies) on human embryonic stem cell–qualified Matrigel GFR (BD Biosciences).¹³ When the cultures reached approximately 95% confluence, they were dissociated using 0.5 mM ethylenediaminetetraacetic acid–phosphate buffered saline and replated at a 1:10 split ratio with 10 μM dihydrochloride (Tocris). After the iPSC cultures were established, 12 clones were chosen and prepared for live-cell imaging experiments. The iPSC colonies were dissociated into smaller clusters of 5–20 cells and replated into 12-well plates. The multiwell plates were then loaded onto custom-built live-cell imaging systems 24 h after cell plating. Under ideal conditions, temperature and CO₂ were at 37 °C and 6.7%, respectively, whereas under stress conditions, temperature would be reduced to 32 °C, or CO₂ would be reduced to 1.0%.

Data Collection

For this study, we collected time-lapse image data of 94 iPSC cultures (12 separate clones among four iPSC lines; Supplementary Figs. S1 and S2), with six observations of each culture, resulting in a total of 564 time-lapse data sets. iPSCs were plated as small clusters of cells into 12-well plates, and they were imaged every 30 min for a period of 3–4 days until the cells reached near-confluence. The collection of data was done throughout multiple experiments, and the majority of data was collected under standard environmental conditions. For a subset of experiments (25% of the total data collected), nonideal culture conditions such as low temperature and low CO₂ levels were used to generate a broader range of iPSC quality. Media changes were performed within the capture intervals by removing the plate from the imaging system and putting it back on the system in less than 10 min. The high-precision motorized x–y stage keeps the culture plates in precisely the same location after media changes. The frequency of media change was daily for all experiments except one (6% of the total data collected), in which media changes were performed every other day to generate a stress condition. For each data set, a time-lapse movie was generated with optimal focus. Each movie of iPSC colony formation was then evaluated by an expert stem-cell biologist and given a ranking of good (Supplementary Movie S1), fair (Supplementary Movie S2), or poor (Supplementary Movie S3). Data were segregated into two buckets; half were used for developing the algorithm and training the software for automated ranking of iPSCs, and the other half were used for testing the performance of the software.

Results and Discussion

We present a system for automated iPSC evaluation that includes noninvasive time-lapse microscopy and analytical software. We collected more than 500 time-lapse movies of iPSC growth among 12 different clones. We used half of the data to develop and train a classifier of iPSC quality based on a set of six relevant biological features: iPSC doubling time, degree of cell compaction, colony-border spikiness, sensitivity to media change, the prevalence of dead cells, and the prevalence of spontaneously differentiated cells. We independently tested the software on the other half, achieving an overall classification accuracy of 80% per position (with six positions per well) and 89% accuracy per well. Of all the features we examined, the degree of cell compaction and the doubling time were shown to be the features with the highest predictive power in determining the quality of iPSC cultures. Our results show the following: (1) A higher percentage of full-compacted cells in large growing colonies is associated with high-quality iPSCs; (2) slower doubling time is associated with lower iPSC quality; (3) higher colony-border spikiness and sensitivity to media change are associated with lower-quality iPSCs; and (4) higher prevalences of dead cells and differentiated cells were found in poor-quality cultures. As shown in Figure 4d , the highest incidence of mismatch occurred on cultures that were ranked as good by manual review and fair by the algorithm, indicating that the algorithm is slightly more stringent in assigning a good rating.

The ranking showed no statistical correlation with line identity, indicating that our ranking system is primarily addressing the overall quality of the culture and environmental conditions, rather than inherent qualities associated with the individual lines that we tested. One of the features, iPSC doubling time, however, can potentially be used to identify line-specific trends. As shown in Figure 3b , by taking measurements of doubling time among multiple experiments (under normal culture conditions), we can evaluate the average doubling times as well as the variation among experiments (through the distribution of each boxplot). Based on our demonstration of measuring doubling time in an automated and noninvasive manner, we anticipate that our methods will enable future larger-scale studies on iPSC doubling time.

Our experiments with nonideal environmental conditions in the cell culture produced interesting results. We observed that low-CO₂ conditions generated changes in iPSC cell-compaction patterns as well as increased cell death, whereas low-temperature conditions slowed down iPSC growth and generated an increased amount of spontaneously differentiated cells. Performing media changes every other day resulted in an increased sensitivity to media changes, but it had no apparent effect on other features such as compaction patterns or doubling time. With time-lapse imaging and our image-analysis tools, these variations in the quality of cell cultures can be captured and quantified.

Given the successful outcome of the presented study, we plan to conduct future studies that will involve collection of data from a larger number of lines, and with experiments conducted by other labs and personnel following their own standard protocols to cover cell-line differences as well as operator variability in passaging, plating, and culture. We anticipate that our overall framework will be fully applicable among a broad range of input data, although refinement and retraining of certain algorithm modules based on a more complete data set will likely be needed. One drawback of this approach is that if a cell-culture protocol changes the morphology of cell stages significantly, then retraining of the algorithm will be needed in both cell-stage classification and cell densities per class. In addition, a limitation of our method is that our quality metrics are based on visual features, rather than a direct measurement of underlying biological processes. Future work may address this issue by performing correlative studies to examine the relationship between morphology-based features and biological assays such as gene-expression profiling or integrated cell-culture (ICC) characterization.

We believe that our results demonstrate the first fully automated and objective assessment of iPSC cultures using noninvasive methods. We anticipate future commercial applications of our system, including characterizing the dynamic response of cells to stress and drugs, implementing quality control during the manufacture of stem-cell reagents, evaluating new protocols and testing changes to existing ones, developing new types of media and other reagents, testing reagent lots, and training new scientists and technicians on the aforementioned activities. We also anticipate that these methods can be applied to quantify the dynamics of other types of cells, including, for example, looking at growth patterns and doubling times of cancer cells.

Footnotes

Acknowledgements

We thank Azucena Lemus for assistance with cell culture and ICC characterization, and Renee Reijo Pera, Matt Burkhardt, and Mike Hasak for helpful discussions and review of the manuscript.

Declaration of Conflicting Interests

M. Maddah and K. Loewke are the shareholders of Cellogy Inc.

Funding

The work for this article was funded by Cellogy Inc.

References

Grskovic

Javaherian

Strulovici

. Induced Pluripotent Stem Cells: Opportunities for Disease Modelling and Drug Discovery. Nature Rev. Drug Discov. 2011, 10 (12), 915–929.

Dimos

Rodolfa

Niakan

. Induced Pluripotent Stem Cells Generated from Patients with ALS Can Be Differentiated into Motor Neurons. Science 2008, 321 (5893), 1218–1221.

Sun

Yazawa

Liu

. Patient-Specific Induced Pluripotent Stem Cells as a Model for Familial Dilated Cardiomyopathy. Science Trans. Med. 2012, 4 (130), 130ra47.

Nguyen

Byers

Cord

. LRRK2 Mutant iPSC-Derived DA Neurons Demonstrate Increased Susceptibility to Oxidative Stress. Cell Stem Cell 2011, 8 (3), 267–280.

Ohnuki

Takahashi

Yamanaka

Generation and Characterization of Human Induced Pluripotent Stem Cells. Curr. Prot. Stem Cell Biol. 2009, 4, 4A.2.

Bock

Kiskinis

Verstappen

. Reference Maps of Human ES and iPS Cell Variation Enable High-Throughput Characterization of Pluripotent Cell Lines. Cell 2011, 144 (3), 439–452.

Maddah

Loewke

Dynamic Morphology-Based Characterization of Stem Cells Enabled by Texture-Based Pattern Recognition from Phase-Contrast Images. Presented at the IEEE International Symposium on Biomedical Imaging, Beijing, China, April 29–May 2, 2014; Paper 54.

Wang

D. C.

Texture Classification Using Texture Spectrum. Pattern Recog. 1990, 23 (8), 905–910.

Chang

C.-C.

Lin

C.-J.

LIBSVM: A Library for Support Vector Machines. ACM Trans. Intell. Syst. Technol. 2011, 2 (1), 1–27.

10.

Mangoubi

Jeffreys

Copeland

. Non-Invasive Image Based Support Vector Machine Classification of Human Embryonic Stem Cells. In 2007 4th IEEE International Symposium on Biomedical Imaging: From Nano to Macro. Proceedings of the IEEE International Symposium on Biomedical Imaging, Washington, DC, USA, April 12–16, 2007; IEEE: Piscataway, NJ, 2007; 284–287.

11.

Ban

Nishishita

Fusaki

. Efficient Generation of Transgene-Free Human Induced Pluripotent Stem Cells (iPSCs) by Temperature-Sensitive Sendai Virus Vectors. Proc. Natl. Acad. Sci. USA 2011, 14234–14239.

12.

Macarthur

C. C.

Fontes

Ravinder

. Generation of Human-Induced Pluripotent Stem Cells by Nonintegrating RNA Sendai Virus Vector in Feeder-Free or Xeno-Free Conditions. Stem Cells Intl. 2012, 564612.

13.

Chen

Gulbranson

D. R.

Hou

. Chemically Defined Conditions for Human iPS Cell Derivation and Culture. Nature Methods 2011, 424–429.

Supplementary Material

Please find the following supplemental material available below.

For Open Access articles published under a Creative Commons License, all supplemental material carries the same license as the article it is associated with.

For non-Open Access articles published, all supplemental material carries a non-exclusive license, and permission requests for re-use of supplemental material or any part of supplemental material shall be sent directly to the copyright owner as specified in the copyright notice associated with the article.

4.96 MB

9.09 MB

8.39 MB

3.41 MB

0.00 MB

1.21 MB