Automated, high-throughput derivation, characterization and differentiation of induced pluripotent stem cells

Abstract

Induced pluripotent stem cells (iPSCs) have become an essential tool for both modeling how causal genetic variants impact cellular function in disease, as well as being an emerging source of tissue for transplantation medicine. Unfortunately the preparation of somatic cells, their reprogramming and the subsequent verification of iPSC pluripotency are laborious, manual processes that limit the scale and level of reproducibility of this technology. Here we describe a modular, robotic platform for iPSC reprogramming that enables automated, high-throughput conversion of skin biopsies into iPSCs and differentiated cells with minimal manual intervention. Using this platform, we demonstrate that automated reprogramming and the pooled selection of pluripotent cells results in high quality, stable, iPSCs. These lines display less line-to-line variation than either manually produced lines or lines produced through automation followed by single colony-subcloning. The robotic platform we describe will enable the application of iPSCs to population-scale biomedical problems including the study of complex genetic diseases and the development of personalized medicines.

INTRODUCTION

The reprogramming of somatic cells into induced pluripotent stem cells (iPSCs), coupled with the development of methods for directing stem cell differentiation into relevant cell types, offers an unprecedented opportunity to study the cellular phenotypes that underlie disease^1–6. The study of these emerging “stem cell-derived disease models” has led to new mechanistic insights into a wide variety of disease conditions⁷.

Despite the opportunities now available to researchers though this technology, several limitations still exist in the application of these methods. The variation that arises between different stem cell lines during the reprogramming process can affect both functional properties of the lines and their performance in disease modeling. Most successful reports to date have relied on the evaluation of only a small number of cell lines derived from individuals harboring genetic variants of high penetrance. If stem cell technologies are to be applied to the study of common genetic variants of modest effect size, yet known to be important contributors to conditions such as schizophrenia and metabolic diseases^8,9, it will be essential to minimize both biological and technical variance. Additionally, many differentiation protocols developed for downstream applications have been optimized for only a small number of cell lines and replicating these protocols across multiple lines has proven to be an arduous and highly variable process¹⁰. In order to maximize the potential of iPSCs, new, reproducible approaches for the derivation, culture and differentiation of many cell lines in parallel will be required. Solving these issues would, undoubtedly improve experimental power for resolving the phenotypic effects contributed by a given genetic variant.

A number of factors have been reported to influence both the efficiency of reprogramming and the performance of the resulting human iPSCs including: genetic background, tissue source^11,12, reprogramming factor stoichiometry during delivery¹³, and culture related stress¹⁴. It has been difficult to quantify the level of cell line variation introduced solely by manual handling during reprogramming. Furthermore, a lack of standardization in reprogramming factors, cell culture reagents and techniques between laboratories is also likely introducing further variability^15,16. With improvements in both precision and scalability in many areas of biomedical research^17–19, particularly through the use of robotics, we reasoned that developing a fully automated, modular, platform for iPSC derivation, expansion, and differentiation would allow us to identify, and minimize, factors contributing to variability in iPSC behavior as well as provide a platform for large scale in-vitro iPSC studies.

Here we report the development of eight liquid handling platforms that automate the process of deriving, characterizing and differentiating iPSCs. Using this modular, integrated reprogramming platform we have systematically explored several factors that have been reported to be important sources of variance in the reprogramming process. Using this system we found that automated reprogramming utilizing isolation of iPSCs through cell-surface antigen expression, rather than clonal colony growth, substantially reduced the variance in performance exhibited by lines derived from a given individual.

RESULTS

System Overview

Central derivation hubs will be required to provide a seamless connection between donor samples (including both biological and clinical data) and end user scientists performing large-scale in-vitro phenotypic assays (Fig. 1A). While modular in its design, here we describe the construction of an automated derivation hub composed of eight robotic instruments the cell culture steps needed to generate and supply iPSCs from donor samples (Supplementary Fig. 1, 2A, Supplementary Movies S1-S6). Two distinct modules were used for fibroblast derivation and biohazard screening. Two further interconnected automation clusters, each composed of three individual liquid handling systems connected through two central robotic arms, were used for all subsequent steps of iPSC production. The design of these systems was such that although any one instrument may contain more specialized instruments, such as a MACS sorter, the majority of automated cell culture methods could be performed across any of the robotic units within a given cluster. Each cluster contains automated-centrifuges, microscopes, and incubators integrated via custom written software (see supplementary information for a more detailed description of each system and the methods performed) that automate particular cell culture steps. All systems were housed within HEPA-filtered, BSL II biosafety hoods. A visual guide of the entire workflow is available, highlighting each automated step through this process (Supplementary Fig. 1).

Fig. 1.

Automated fibroblast and iPSC production. (A) The development of this robotic platform was centered around the idea that the systems described within the manuscript would act as a derivation hub that would itself interact with clinics and other outside sources to recruit samples for iPSC reprogramming. Upon the generation of repository stocks of both fibroblasts and iPSCs, these would be distributed to banking sites for repository storage and expansion. End user sites could, in turn, request custom-arrayed lies and, rather than have several robots, have one or two focused instruments that could perform assays downstream of iPSC generation. (B) Image of system for automated fibroblast production consisting of a liquid handling device, imager, centrifuge and capper/decapper contained in a biosafety cabinet, connected to an automated incubator and managed by system control software. (C) Histogram of fibroblast doubling times calculated from confluence scans of fibroblasts during expansion. (D) Comparison of fibroblast doubling times grown in low percentage serum-containing media prior to cryopreservation and after thawing in higher percentage serum-containing media (not significant, (p=0.24), paired t-test). (E) Scatterplot of doubling time vs. age of donor.

Automated Fibroblast Production

In order to generate iPSCs (from both healthy and diseased donors), a genetically diverse bank of fibroblasts was created from donated biopsies. Beyond a manual step required for the initial dissection and seeding of each biopsy/tissue sample into standard tissue culture six well plates, all further maintenance, including imaging, feeding, passaging and freezing, was performed via automation (Fig. 1B). Biopsy outgrowth was measured every five days using automated image acquisition and quantification (Supplementary Fig. 1B-C). The automated imaging system first identified wells with outgrowths before calculating a confluence value based upon the area of the culture plate occupied. Media was collected under automation and manually transferred for Mycoplasma and sterility testing on a second automated platform composed of a robotic liquid handler and plate reader. In order to prevent the spread of any potential incoming pathogens, these two initial systems operated under quarantine conditions, housed independently of downstream robotics (Supplementary Fig. 2D).

Upon completion of Mycoplasma testing, fibroblast outgrowths were enzymatically passaged using automation into new (daughter) plates for further expansion. In order to minimize population doubling and increase reprogramming potential²⁰, fibroblasts were frozen into multiple 2D barcoded cryovials after one passage and banked in order to produce low passage (p2) stocks for reprogramming, as well as larger back-up stocks one passage later. Vials were designated as either reprogramming or repository stocks through the use of an automated −80°C storage and retrieval system, with reprogramming vials remaining inside for short term storage and repository vials transferred to liquid nitrogen. The average total number of cells frozen into a single cryovial at the time of freezing was 121,437 cells which upon manual thawing and counting had an average viability of 84% ± 1.43% (mean ± SEM, n= 167).

During initial development of the automated fibroblast derivation system, a total of 640 skin tissue samples were collected with fibroblast cultures were successfully established and frozen from 89.4% of biopsies (n=572). Failures were primarily due to either bacterial or fungal contamination (4.7%, n=30) and were attributable to sample handling prior to entering the system. Failures were minimally associated with equipment or process issues (3.2%, n=21). Only 2.7% (n=17) of biopsies did not show outgrowth, which upon retrospective analysis of images from plated biopsies may have been attributable to the small size of the biopsy. Using a NanoString nCounter karyotyping assay, which can detect aneuploidy and large chromosomal gains/losses (Supplementary Fig. 3), we spot-tested 20 independent fibroblast samples. The majority of samples (19/20, 95%) showed a normal diploid karyotype (46XX or 46XY).

As the growth rate of somatic cells has been shown to be a significant determinant of reprogramming efficiency^21,22, we reasoned that parallel automated reprogramming of many fibroblast lines in a multi-well format would require controlling for variance between their growth rates. Using images collected via automated imaging during the initial fibroblast derivation under low serum conditions we calculated fibroblast growth rates and found that there was a substantial variance in growth rates prior to freezing (n = 298 cell lines) (Fig 1C–D). However, the subsequent automated thawing and expansion of fibroblasts in a medium containing a higher percentage of serum prior to reprogramming, resulted in a decline in the variance between the growth rates of each cell line without a significant change in average doubling times across the lines analyzed (p=0.24; Fig. 1D). This reduced variance greatly streamlined the process of batching cell lines with similar doubling times for subsequent reprogramming. Interestingly, we did not find an obvious correlation between donor age and fibroblast growth-rate (Fig. 1E).

Using standard tissue culture plates the first two robotic systems can successfully derive, QC, expand, and cryopreserve high quality, low passage fibroblasts, which can be banked for future use at a rate of more than 15 biopsies per week.

Automated Reprogramming

The first of two downstream robotic clusters, consisting of three primary liquid handling systems together with integrated imagers, centrifuges, incubators and a magnetic cell sorter, was designed to thaw fibroblast samples, passage and seed fibroblasts at a user-selected density, deliver reprogramming factors, perform magnetic selection of reprogrammed cells, and finally image cultures to identify nascent stem cell colonies after surface marker staining ( Fig. 2A, Supplementary Fig. 1, 2A, Movie S2-S4). These processes were performed in a completely automated manner without manual plate manipulation (Fig. 2B, Supplementary Fig. 1).

Fig. 2.

Automated reprogramming. (A) Experimental scheme for automated fibroblast thawing and reprogramming. (B) Image of robotic system for automated fibroblast thawing and mRNA transfections. (C) Representative time course of mRNA transfection, with development of colonies over 22 days. (D) Representative FACS analysis of 32 biological replicates of reprogrammed cultures from automated mRNA transfection displaying a higher proportion of cells expressing the pluripotency markers TRA-1–60⁺/SSEA4⁺ 23 days post completion reprogramming by mRNA (i) and lack of the fibroblast surface marker CD13 (ii). (E-H) Effect plots of Poisson regression analysis of factors that contribute to reprogramming success: (E) Colony count vs Age; (F) Colony count vs. Recovery media post thaw; (G) Colony count vs Confluence; (H) Colony count vs Fibroblast doubling time. Gray area and red bars indicate confidence intervals.

We initially adapted a viral reprogramming method using Sendai virus²³ to the automated liquid handling system. Automated delivery of Sendai virus to 50,000 fibroblasts, at a multiplicity of infection of 4, resulted in 2 to 10 TRA-1–60⁺ colonies per well under feeder-free conditions (n=168 reprogramming attempts; Supplementary Fig. 4A). Similar efficiencies were obtained by manual reprogramming under identical conditions. During a trial phase of automated transduction, we derived several clonal lines through manual picking of TRA-1–60⁺ colonies, before continuing their culture without further subcloning. Established cultures from these colonies expressed common markers of pluripotency, including NANOG, POU5F1, SSEA4 and TRA-1–60 (Supplementary Fig. 4B).

However, and consistent with previous results²⁴, we found that even automated Sendai viral reprogramming induced significant numbers of incompletely reprogrammed TRA-1–60⁺/SSEA4+ colonies that retained surface expression of the fibroblast marker CD13 (58.2% of cells from n=12 independent lines, 18 days post infection; Supplementary Fig. 4C). Furthermore, after differentiation in embryoid body (EB) assays (see below), we found significant clonal variability in gene expression that appeared to coincide with residual Sendai transgene expression remaining in the clones between passages 5 and 10 (Supplementary Fig. 5A-B). To determine whether this variance could be reduced by eviction of the virus, lines were incubated at 38.5°C. While this initially reduced expression of viral transgenes, following an additional 2–3 passages the lines were reanalyzed and Sendai viral gene expression had again increased (Supplementary Fig. 5C). Based on these results, we concluded that while Sendai viral reprogramming was compatible with automation, the variance introduced by incomplete viral eviction, likely requiring further subcloning for complete elimination, warranted the investigation of an additional reprogramming method.

We next automated the delivery of modified mRNAs encoding reprogramming transcription factors²⁵. Following a thaw and recovery period of fibroblasts in 12 well plate format, in order for reprogramming to take place most effectively, cells must be seeded at specific densities. In order to achieve this an automated passaging method incorporating the use of an automated counting step was used to consolidate the wells of two 12 well plates into one, Geltrex coated, 24 well plate. With a typical thaw containing 48 samples, this resulted in two 24 well plates entering each reprogramming run.

As miRNAs have previously been shown to improve reprogramming efficiency²⁶, a cocktail of miRNAs was added at day 0 (24 hours after passaging) followed by 10 daily transfections of in-vitro transcribed mRNAs encoding POU5F1, KLF4, SOX2, c-MYC, LIN-28, and nuclear GFP (nGFP) (Fig. 2B). The miRNA cocktail was also included together with the fourth mRNA transfection. Human fibroblast-conditioned media (containing recombinant B18R) was replaced daily by automated media exchanges prior to each transfection. Six days after the final transfection, cultures were transitioned to a feeder-free human embryonic stem cell (hESC) culture medium. During process development this automated, integrated system generated 1008 reprogramming attempts, split across 21 experimental production runs of 48 samples per run. Runs were typically launched at a rate of 1–2 per week over a 7 month period. From this, 334 samples were eliminated due to incomplete data for the analysis described below or from fibroblasts not derived under automation. An additional 151 attempts failed due to poor growth of fibroblasts after thawing. This left 523 individual attempts that proceeded through independent reprogramming events, of which 375 were from either fibroblasts derived under automation (110 unique donors) or 148 control BJ fibroblasts (included in each run to monitor run-to-run variation).

We consistently observed nascent iPSC colonies between days 16 and 22 of culture (Fig. 2C), and established cultures demonstrated a pluripotent hESC-like morphology and expressed common markers of pluripotency, including NANOG, OCT4, SOX2, SSEA4 and TRA-1–81 (Supplementary Fig. 6A). Using an automated colony-counting algorithm combined with a live Tra-1–60+ antibody, an average of seven colonies per well were identified (Supplementary Fig. 6B). During process development, we observed 221 successful reprogramming events from the total of 523 attempts using automation produced donor fibroblasts and BJ fibroblasts (range 1–40 colonies). FACS analysis of samples collected pre- and post-MACS enrichment showed samples contained a high proportion of TRA-1–60⁺/SSEA4⁺ cells, with only 9.7% retaining CD13 at this stage (23–30 days post completion of reprogramming, n= 32) (Fig. 2D).

When calculated across all conditions on a per cell basis, reprogramming efficiency was between 0.001% and 0.16% per plated somatic cell, consistent with previous results obtained under feeder-free conditions²⁷, and was slightly higher for control BJ fibroblasts (0.043%) than for adult fibroblasts (0.014%) as expected. Although we were able to reprogram fibroblasts from older donors, post-hoc analysis indicated that increasing donor age negatively influenced the number of colonies produced (Fig. 2E). Most importantly, this analysis revealed a negative impact on reprogramming efficiency, when switching from a high serum- containing media, such as FM10, to a serum-free reprogramming medium used during transfections (Fig. 2F). While not affecting the absolute iPSC colony count, the use of a low serum medium, such as M106, resulted in slower growth rates and a relatively lower reprogramming efficiency, calculated on a per cell basis (Fig. 2G). In wells where fibroblasts grew to confluence during the reprogramming process, a high cell density was negatively correlated with reprogramming efficiency (Fig. 2H).

Based on the data from these process development runs, a modified reprogramming strategy was implemented, whereby a decremental serum reduction was performed during the first 5 days of reprogramming. This methodology was applied to 216 samples over 4.5 reprogramming runs (typically 48 samples per run). From the analysis of this data we found a 81.9% overall increase in average reprogramming efficiency, with overall reprogramming success at 76.9% (166/216 total samples). Of the 102 unique samples used in these runs 86% (88/102) were reprogrammed with an average of 17.7 colonies per well. Of the 14 that did not reprogram five failed on multiple occasions. Of the remaining 9, one sample was only attempted once while the other 8 were part of a run where one of the two 24 well plates was lost due to technical issues. A total of 65 unique donor lines were run in duplicate and 61 reprogrammed successfully on both occasions. The remaining samples were run in either triplicate or as single samples.

Automated iPSC purification

To better support high-throughput purification of iPSCs from incompletely reprogrammed cells, we switched from using fluorescence activated cell sorting (FACS)²⁴, to the integration of a multi well, immuno-magnetic bead separation device (MultiMACS from Miltenyi Biotec) (Fig. 3A, Supplementary Fig. 7A-B). Spike-in experiments, using a range of iPSC-to-fibroblast ratios between 1:20 and 1:100, demonstrated that an iPSC purification of approximately 26-fold could be achieved (Supplementary Fig. 7C).

Fig. 3.

Automated iPSC purification and arraying. (A) FACS analysis for TRA-1–60/SSEA4/CD13 cells pre- and post-automated MACS purification. (B) Representative images of 96 well for bulk sorted cells from first day post sorting (1dps) to 9 days post sorting (9dps) with TRA-1–60 expression pattern captured by automated imaging. (C) Clustering of sorted samples against reference hESC and fibroblast lines based upon gene expression of pluripotency and early differentiation markers. (D) Box plot of the pluripotency scores (see Extended Methods) for reference hESC lines, iPSC lines, and fibroblast cell lines. Number of unique samples shown in parentheses. (E) Example growth rates of a robotically passaged iPSC plate over 3 days culture. (F) Summary of FACS analysis of TRA-1–60⁺/SSEA4⁺ population before and after automated passage 1:3 for control hESC lines and iPSC lines derived on the system(error bars, s.d.).

Primary iPSCs enriched through MACS selection were robotically transferred into 96 well plates across a three-fold serial dilution, plated in quadruplicate (Supplementary Fig. 8A) in order to obtain an optimal cell density post-reprogramming. TRA-1–60⁺ colonies formed within 7–10 days after magnetic sorting and re-plating of single cells (Fig. 3B). With the use of a serial dilution strategy, it was possible to isolate clonal lines (Supplementary Fig. 8B) however we instead selected polyclonal wells with similar growth characteristics for downstream expansion and characterization. The mean doubling times ranged from 24–48 hours (Supplementary Fig. 8C) as calculated from replicate pools produced from a total of 142 (65 unique samples) individual reprogramming events (produced post-MACS enrichment). Of the 65 unique samples, 49 were from fibroblasts derived under automation. 320 samples had a doubling time greater than 48 hours, representing 53 unique samples and these samples, where cherry picked for consolidation (see below), were included in variance analysis.

After wells with similar confluence values were identified through automated imaging, samples were robotically consolidated into new daughter plates for expansion and quality control assays. Flow cytometry analysis showed ~80% of the cells were SSEA4+/TRA-1–60+e and expressed the pluripotency marker NANOG (Supplementary Fig. 8D-E). Sorted samples were also analyzed using a gene expression panel covering pluripotency and germ layer marker genes (Fig. 3C). Using a pluripotency analysis strategy based upon previous work²⁸, a subset of iPSCs were tested and 9 of 11 lines had scores consistent with an hESC reference panel (Fig. 3D, Supplementary Fig. 9A). Two outlying samples (10005_421 and 1005_350), while pluripotent, displayed elevated differentiation scores, retrospectively attributed to overgrowth-induced spontaneous differentiation (Supplementary Fig. 9B-C). These samples, together with sample 10005_699 (which had an elevated pluripotency score) are indicative of cultures that fail the QC criteria. Therefore, within 2 passages of reprogramming, high-purity undifferentiated iPSC lines could be established and validated by high throughput processing in 96 well plates.

Automated culture of multiple iPSC lines in parallel

The second of the two robotic clusters-similarly consisting of three primary liquid handling platforms together with an integrated imager, centrifuge and incubator- functions to expand and freeze down cells into barcoded matrix tubes; create embryoid bodies for QC analysis; and collect cell pellets for RNA and DNA isolation. With the exception of a minor manual step of sample retrieval from liquid nitrogen, all processes were completed under automation (Supplementary Fig. 1).

Although iPSCs showed a narrow range in doubling times (Supplementary Fig. 8C), to ease the parallel culture of cell lines with variable growth rates, we developed automated processes for the cryopreservation and recovery of nascent iPSCs with similar growth characteristics (Supplementary Fig. 1, Supplementary Fig. 2A (Stage 3), Supplementary Fig. 10A-B, Movie S5). Following robotic iPSC cryopreservation and thawing, iPSCs reattached, resumed proliferating and demonstrated a normal morphology (Supplementary Fig. 10C). Pre-freeze and post-thaw confluence correlation was highest one-day post thaw (Pearson’s r > 0.91) but decreased as the wells approached full confluence (r >0.71 on day 3, >0.41 on day 6). We found this decreasing correlation was largely due to overgrowth on higher confluence wells (Supplementary Fig. 10D). Expression of the cell surface markers SSEA4 and Tra-1–60 were also unaffected by freeze/thawing (Supplementary Fig. 10E).

Cells growing in 96 well plates could be successfully maintained for up to 7 days, between passages (Fig. 4F). Passage ratios ranging from 1:1 to 1:15 were successfully utilized on the system with low plate to plate variation (Supplementary Fig. 10E) and no significant impact on morphology or marker expression(Supplementary Fig. 10F-G).

Fig. 4.

Automated Embryoid body assay. (A) Image of EBs generated from iPSC ubiquitously expressing GFP using automated imaging system. (B) Representative image of the Greiner 96 well V-bottom plate with EBs is shown after passage to form EBs by automation. (C) Differentiation propensities (EC=ectoderm, ME=mesoderm, EN=endoderm) for the 10 hESC reference lines indicated by the boxes, with each colored line showing the average differentiation propensities observed over four biological iPSC lines replicates using three different methods. (D) Overall cluster analysis of gene expression analysis from EBs produced using different plate formats using the EB scorecard geneset.

Analyses of genomic DNA were performed to track line identity and to ensure that cell lines remained karyotypically normal (Supplementary Fig. 11A-B). While previous reports have stated that chromosomal abnormalities occur in approximately 20% of iPSCs²⁹, the majority of our iPSCs (89%, n=38) showed a normal diploid karyotype. Three of the abnormal lines all originated from a common fibroblast (BJ) and shared the same genomic aberration, suggesting a low-percentage heterogeneity pre-existing in the original fibroblast.

An additional 8 lines were subject to higher resolution SNP array analyses at both low (P8) and high passage (P20). From 2 independent fibroblast samples, three iPSC lines were derived as pooled populations with a further 5 derived as clonal lines via manual picking following automated reprogramming. Seven of the 8 lines displayed single de-novo copy number variations (CNVs) at low passage, with mosaic CNVs found in 2 of the pooled lines (Supplementary Karyotypes). Two lines (1 pick and 1 pool) displayed the development of either one or two de novo CNVs over continued culture. These numbers are in accordance with other CNV studies in iPSCs^30,31. This highlights the need for continual monitoring of any pluripotent stem cell line. Together, these results suggest that it is possible to maintain iPSCs during subsequent automated passaging without affecting pluripotency, growth characteristics or genetic stability.

Automated analysis of differentiation propensity

To quantitatively assess pluripotency and the subsequent ability of iPSC lines to differentiate, we developed automated methods for generating EBs and performing a modified version of the previously described stem cell “scorecard” assay^28,24. This method allowed us to analyze expression of gene sets after spontaneous differentiation into ectoderm (EC), mesoderm (ME), and endoderm (EN) lineages (Fig. 5A) in EBs made from both newly-derived iPSCs relative to a control panel of hESCs differentiated (Supplementary Fig. 12A) in parallel under identical conditions (Movie S6). Using an EGFP marked iPSC line, automated EB formation was optimized to seed cells from a single well of a 96 well plate into six daughter 96 well V-bottom plates (Fig. 4 A–B) which were harvested for gene expression analysis.

Fig. 5.

Reduced variation in robotically derived iPSCs. (A) Overall cluster analysis of gene expression analysis from EBs produced using different plate formats analyzed using the EB scorecard geneset. (B) Variance analysis of scorecard gene expression in EBs showing comparisons of standard deviation of gene expression values among samples derived on and off the automated system. Lines produced by the automated process showed significantly less variation in EB gene expression compared to lines produced by manual methods and later introduced onto the system (P value = 7.08E-12, Wilcoxen signed rank test). The comparison between manually derived lines and lines reprogrammed using automation followed by manual colony picking was marginally significantly different (P value = 0.023). (* = p < 0.05, *** = p < 0.001). Manual Derivation (all, n=16), Automated Derivation (all, n=21, BJ = 9, Donor = 12), Colony Picking (all, n=29, BJ = 9, Donor = 9), Automated Derivation (Donor, n=12), Automated Derivation (BJ, n=9). (C) The standard deviation in gene expression of EBs differentiated from iPSCs was independent of passage number despite being either pooled or picked iPSCs.

All reference lines tested exhibited scorecard differentiation propensities consistent with the ability to differentiate into the three germ layers (Supplementary Fig. 12B). While the scorecard values for all reference samples showed strong correlations with those previously published²⁸, a decrease in correlation was observed when EBs from our reference samples (from the identical reference lines used in that publication) were generated under different culture conditions (Supplementary Fig. 12C-E). With further analysis we found that the method used to generate EBs can introduce a large bias in differentiation potential, as lineage marker gene sets clustered differentially based upon the method used for EB formation (Fig. 4C). As such, these data highlight the need for method standardization and caution when cross-comparing EB datasets generated under varying conditions.

Reduced variation in robotically derived iPSCs

Hierarchical clustering of gene expression from the automated EB assay showed an overall consistency in the iPSCs generated by automation (Fig. 5A), with a significant reduction in variation seen when comparing entirely manually derived lines to those produced under automation ((p value = 7.08x10–12, Wilcoxen signed rank test). This was true in comparisons both within a single genotype (BJs, p values = 7.85x10–9), and between patient lines (Donor, p values = 9.28x10–11). Interestingly, iPSCs initially reprogrammed robotically before having colonies manually picked and then returned to the automation system for expansion, showed an elevated variation similar to that found in existing manually derived iPSC lines (p value = 0.023). This effect further appears to be independent of reprogramming methodology, with clonal lines derived through manual picking following robotic Sendai reprogramming, exhibiting similar variation. Thus, our findings indicate that manual clone selection is an important source of variation. Passage number, however, appears to not play a significant role in the behavior of any one cell line, whether it be a pool or pick, as the mean differentiation propensity in the scorecard assay did not significantly deviate over continued passaging (samples tested at passage 9 and 20). This suggests that lines produced by our automated process show reduced variation at early time-points after derivation compared to lines derived by current manual procedures.

Automated differentiation

To further test the differentiation capabilities of iPSC lines produced by automated methods, we generated cardiomyocytes from automation-derived iPSCs using two common protocols (Fig. 6Ai–ii, B). The lines displayed efficiencies comparable to that of published protocols and performed as well as reference lines differentiated in parallel (Fig. 6Aiii-iv, B)³².

Fig. 6.

Automated differentiation of iPSCs. (A) Directed differentiation of iPSCs into cardiomyocytes. FACS analysis shows two iPSC lines derived under automation were able to express high levels of Troponin-T following manual differentiation using a kit based assay (i-ii). Similarly, differentiation could be successfully performed using published protocols at levels comparable to reference lines (iii-iv). (B) Immunostaining of Troponin-T expression in cytospun differentiated cardiomyocytes. Scale bar = 200 μm. (C) Automated directed differentiation of Sox17 positive endodermal cells from the iPSCs and hESCs shown. The number of independent runs are indicated in parentheses. error bars are s.d. (D) The automated system can also perform media exchanges using media to direct differentiation into midbrain dopaminergic neurons. Markers for midbrain progenitors are shown for LMX1,SOX1,SOX2,NESTIN, and FOXA2. Automation was also able to maintain differentiation neurons as shown by the markers TUJ1 and TH marking dopaminergic neurons. scale bars = 100um.

Using differentiation protocols that primarily rely on defined media to guide differentiation provide an opportunity to further standardize differentiation of iPSCs. To test this, we utilized the media exchange methods developed on System 6 to perform protocol directed media exchanges on iPSCs growing in 96-well format. Using an automation compatible kit, cells expressing the endodermal marker SOX17 could be readily generated with efficiency strongly correlating to their endodermal scorecard value (Pearson correlation = 0.905) (Fig. 6C).

While the endodermal differentiation protocol lasts 5 days, we also wanted to determine if longer protocols were amenable to automated culture. To test this, we adapted a commonly used protocol for midbrain dopaminergic neurons lasting over 30 days of continuous culture³³. We found that both intermediate stage progenitors and differentiating neurons could be readily differentiated and maintained, retaining expression of markers typical for this cell type (Fig. 6D). Together these data suggest that it is possible to automate the differentiation of pluripotent stem cell lines in parallel on multiwell plates on a single module of the current system.

DISCUSSION

Here we show the establishment of fully automated processes for the generation of fibroblast banks, iPSC reprogramming, and quality control assays for assessing pluripotency and directed differentiations of iPSC lines, such as those used for phenotypic screening assays. Outside of a small number of minor manual steps such as biopsy dissection and cryovials-to-freezer transfer, we have built upon recent advances to develop a robotic system that incorporates several processing steps that improve the quality and consistency of iPSC lines over those produced by current manual methods^24,25,28. The combination of the automated delivery of modified mRNA together with the pooling and enrichment of the resulting reprogrammed cells significantly improved the consistency and quality of the resulting iPSC line as compared to traditional methods such clonal isolation of reprogrammed cells. Most importantly, we demonstrate that, without compromising quality, iPSC production can be scaled to meet the sample sizes required for in-vitro population-wide genotype-phenotype association studies.

While supporting the complete workflow for generating and characterizing iPSCs in a high-throughput manner, this platform was designed so that individual modules can be used for specific applications, such as iPSC maintenance and differentiation. To this end, we have also demonstrated that modular robotic liquid handling systems can be used to direct the differentiation of iPSC lines in multiwell plate format, enabling the development of population scale assays needed to exploit the utility of patient specific iPSCs. With this approach in mind, it is envisioned that only large core facilities would require a complete reprogramming platform with large individual labs having one system on which large numbers of iPSCs can be handled. As such, the expense to these labs would be far less and within the realm of those purchasing and maintaining equipment such as confocal microscopes.

It has allowed us to identify and control sources of functional variability between iPSCs. Importantly, by using robotics, panels of patient iPSCs can be generated with both precision and at scale, while minimizing human bias. Most notably, we found that manual isolation of newly reprogrammed iPSC colonies was in itself a substantial contributor to cell line to cell line variation. Through automation of the reprogramming process, more than a third of the variability that existed between manually selected lines was eliminated (Fig. 6B). This finding demonstrates that at very least, a substantial portion of the variation observed between manually-derived iPSC lines has purely technical origins that may obscure inherent genotypic differences.

The large scale of our experiments also allowed us to address the previously raised question of whether the origin of the somatic cells and and/or their genetic background, influenced stem cell line behavior. The production of many cell lines from both a single somatic population (BJ fibroblasts) and from multiple distinct individuals allowed us to address this issue. Our results clearly show that the level of variability between cell lines made from many donors was not different from that found with lines from a single donor. Previous studies suggest that genetic factors could be a contributing factor to functional variance between iPSCs³⁴. However, our data suggests that if these factors do contribute, they do so modestly in comparison to the technical variation that can be resolved through automation.

While we were able to successfully reprogram many samples from subjects at advanced age, our data suggests that advanced age, as previously highlighted^35,36, is a potential inhibitor of reprogramming. However, we found that both the growth rate and confluence of cell cultures at the time of reprogramming were primary drivers of whether our automated approach succeeded in producing iPSCs in each individual case. These findings seem consistent with the observation that genetic factors that slow proliferation of cells inhibit reprogramming²².

In addition to this, we found that the reprogramming method used for producing iPSCs had a substantial effect on cell line properties. Although aspects of iPSC production using both mRNA delivery and Sendai virus infection could be automated, incomplete eviction of Sendai virus led to a substantial change in the signature of gene expression in pluripotent cells. For these reasons as well as Sendai lot-to-lot variation and availability, we switched to a modified mRNA reprogramming method as our standard protocol. However, the flexibility of the system allows for the future adoption of other reprogramming methods as these become available.

In contrast to other approaches that are focused on the generation, expansion and manipulation of only a small number of cell lines at any one time scaling, as described here, the integrated robotics system has the capacity to reprogram several hundred samples per month³⁷. The advantage to our approach of automated production over a manual process, however, is that capacity can be scaled with additional systems with only a minimal increase in personnel time. The timeline for producing seed stocks of characterized iPSC lines from fibroblasts is approximately 12–16 weeks. Although the current system may not reach cGMP standards, future systems based on this technology may allow for clinical grade production. As advances in reprogramming continue allowing easier reprogramming of alternative cell sources, adaptation of this system to input material, such as blood, is in process. Additionally, as automated single-cell isolation is possible, this high-throughput approach can also be used for the isolation, expansion and screening of targeted iPSCs after gene editing. Together this increased scale and accelerated timeline will enable many large-scale projects utilizing iPSCs to perform functional human genetics³⁸.

In the future, the increased throughput of reprogramming and reduced variability between the resulting lines will open a number of new avenues for investigation. First, this approach should allow investigation of cellular phenotypes for disease states found in conditions that are caused by diverse mutations. At the moment, most studies utilizing iPSCs for disease modeling have focused on a small number of lines originating from individuals harboring either a single or small number of highly penetrant mutations. The expanded scale and reduced variation of the automated system will lead to greatly improved statistical power in addressing the question of whether a modest effect observed in culture is a direct result of the genetic background of the subject in question.

This increased sensitivity should assist in accurately assessing the impact of common variants that influence human health. Although these variants may only contribute modestly to the overall phenotype in a given individual, their prevalence in populations highlights their importance in understanding their role in disease. The ability to accurately study such variants will represent an important next phase for in-vitro disease models. If there is one common process that eventually leads to cellular dysfunction in each particular disease, it may well be possible to devise a single strategy for inhibiting it, providing a therapeutic for all patients. In contrast, it may be that subsets of patients follow distinct paths towards disease. If this is the case, it will be necessary to identify these distinct groups of patients so that a therapeutic strategy specific to the particular path their disease follows can be devised and appropriately administered. The system described here is designed to perform standard cell culture manipulations at large scale, and this robotic technology is being adapted to generate multiple differentiated tissues of interest from panels of patients. This approach should further enable the discovery of molecular and genetic pathways that underlie traits of human development and disease by supporting the simultaneous analysis of statistically relevant numbers of patient cells, overcoming the practical limitations of current experimental approaches and reducing the inadvertent bias and variation introduced by manual methods.

Methods Summary

While a fully detailed methodology is available within the supplementary information, in brief: Written informed consent procedures, the recruitment of volunteers and biopsy collections were approved by Western Institutional Review Board. To accomplish fibroblast banking, iPSC generation, and iPSC characterization and freezing, three integrated robotic platforms were assembled from a combination of eight automated liquid handlers, five incubators, two robotic arms, four automated plate imaging systems, three cryovial decappers, three plate centrifuges, and one plate sealer. Human dermal fibroblasts were derived from donor tissue samples using the robotic system for automated culture and expansion, growth rate analysis, mycoplasma testing, and freezing. Reprogramming of fibroblasts was performed using automated transfection and feeding methods to deliver modified mRNA. Pluripotent cells were identified with live-stain TRA-1–60. iPSCs were enriched by automated depletion of non-reprogrammed fibroblasts and serially diluted three-fold in adjacent wells. Based upon growth rates and confluency level samples were consolidated before extended automated expansion, cryopreservation and analysis of gene expression and karyotyping via NanoString codesets. EBs were formed using 96 well V-bottom plates. Cell aggregates (EBs) were allowed to grow for a total of 16 days before being analyzed with custom NanoString codesets. Statistical analysis was performed using custom R and Matlab scripts and based on previously published work^28,39.