Screening Station, a novel laboratory automation system for physiologically relevant cell-based assays

Due to their physiological relevance


Introduction
Human-induced pluripotent stem cells (iPSCs) and human primary cells are attractive tools for in vitro disease model for evaluating the efficacy of drug candidates [1][2][3][4][5][6] .However, long-term maintenance of iPSC cultures and control of the differentiation of iPSCs require craftsmanship.Several factors have been reported to affect the characteristics changes; and 3) complex experiments such as immunostaining need to be reproducibly performed to detect multiple cellular events.
Automation that minimizes human error can solve these problems and ensure that all processes are executed accurately.Several studies have reported automation strategies for cell experiment to minimize human error [ 7 , 8 ].However, available studies have only reported the automation of individual processes, such as the exchange of the culture medium of iPSCs in microplates, live-cell imaging and immunostaining [ 9 , 10 ].No single system to date can fully automate the combination of immunostaining, including the media exchange process, live-cell imaging, and transport of experimental equipment.
Cell imaging requires high-precision 3D image acquisition for downstream high-precision image analysis.Several studies have reported automatic culture systems that can analyze cell confluency through deep learning [11][12][13] .However, these systems have trouble automating the process for multiple samples.In addition, during live-cell fluorescence imaging, it is vital that the laser does not damage the cells.One technology used to acquire high-resolution 3D images without damaging cells is Nipkow spinning-disk confocal microscopy with a microlens.One such device is the CQ1 confocal quantitative image cytometer by Yokogawa Electric Co., Ltd.Since the CQ1 is also a plate reader, it can image multiple samples at high speed.In addition, the CellPathfinder image analysis software, which can be linked to the CQ1, has built-in deep learning analysis capability to enable automatic execution of a series of processes such as measurement, data transfer, and analysis.
There is currently no fully automated system that enables parallel pharmacological evaluation of multiple samples through all processes starting from cell culture to time-lapse imaging, including deep learning analysis.To develop such a system, it is important to first automate the scheduling of experiments, as has been reported previously [14] .Because classical automation systems cannot be restarted in the event of an error during execution, it is often necessary to start experiments over from the beginning, making it difficult to fully automate experimental processes.To overcome this issue, the workflow in the Green Button Go (GBG) scheduling software (Biosero Inc.San Diego, CA) separates the transport and dispensing for each labware, thereby simplifying the workflow and making error-recovery easy.GBG is thus a flexible and fast recoverable scheduling software.
To develop a fully automated system, it is also necessary to be able to operate it from a remote location.Remote experiment automation has gained increased attention and encouragement due to the COVID-19 pandemic [15] .
Furthermore, cell image analysis to quantify outcomes such as the degree of change in morphology and differentiation of iPSCs requires the setting of many parameters, which cannot be readily performed by just anyone.The CellPathfinder analysis software solves this problem by offering four deep learning functions which can be used to automatically select necessary parameters for analysis, requiring the user to only set representative (cell) images on which to perform the analysis.
Here, we developed a new laboratory automation system called the Screening Station to integrate and automate three key processes of iPSCbased assays: long-term culture, live-cell imaging and immunofluorescence staining and imaging.We optimized the system to enable it to be controlled manned, unmanned, or remotely.We describe the process and discuss the future of automated laboratory systems below.

Apparatus
All apparatus and their configurations are shown in Fig. 1 .
To induce myogenic differentiation of FSHD-MyoD N#1, the cells were dissociated with Accutase, and 4000 cells each were added to the wells of 96-well plates (cat.no.6055300, PerkinElmer, Waltham, MA) coated with Matrigel growth factor reduced, phenol red free (cat.no.356231, Corning, NY).The next day (day 1), the culture medium was exchanged with primate embryonic stem (ES) cell medium (cat.no.RCHEMD001, ReproCELL) containing 10 μM Y-27632 and 0.5% penicillin/streptomycin.On day 2, the culture medium was exchanged with primate ES cell medium containing 10 μM Y-27632, 1 μg/mL doxycycline hyclate (Dox; cat.no.D5897, LKT Laboratories, Saint Paul, MN) and 0.5% penicillin/streptomycin.On day 3, the medium was exchanged with medium containing 5% KnockOut Serum Replacement (KSR; cat.no.10828-028, Thermo Fisher Scientific) and 0.5% penicillin/streptomycin, followed by Alpha Modified Eagle Minimum Essential Medium ( -MEM) culture medium (cat.no.21444-05, Nacalai Tesque) containing 1 μg/mL Dox, and 200 μM 2mercaptoethanol (2-ME; cat.no.2143882, Nacalai Tesque) and 0.5% penicillin/streptomycin. Dox was removed on day 5 and 2-ME was removed on day 6.The cells were fixed and stained on day 5 and day 6.Exchange of the medium was performed using the Biomek i7 automatic dispenser after automatically transferring a medium reservoir from the cool incubator, culture plates from the SCALE48W automatic CO 2 incubator, and a waste reservoir and tips from the VariStock labware stocker.Cell monitoring by time-lapse live-cell imaging was automatically performed by transferring the cell plate from the SCALE48W to the CQ1 image analyzer .Immunostaining for the myosin heavy chain (MyHC) and cell nucleus was performed using Biomek i7 and the 405LS automatic washer by automatically transferring the reservoirs with blocking buffer/antibody solution from the cool incubator, cell culture plates from the SCALE48W, and formalin plates (4% paraformaldehyde phosphate buffer solution (4% PFA/PBS; cat.no.163-20145, FUJIFILM Wako Pure Chemical)) from VariStock.Scheduling of the Screening Station was performed using GBG software.Cell analysis that included deep learning analysis was performed using CellPathfinder software (Yokogawa Electric, Tokyo, Japan).

Design of conceptual workflow
The hardware of the automated system incorporated the equipment needed to automate all processes, namely transport of human samples/culture, medium exchange, addition of the drug discovery screened compound, live-cell monitoring, and immunostaining ( Fig. 1 A, B).The core plate handler was positioned to deliver labware to all stations.Culture plates could be automatically transported to the SCALE48W after being manually placed in the safety cabinet.After being loaded into the stocker, reagents and labware could be automatically transported to the Biomek i7 automatic dispenser or 405LS washer, and after dispensing and washing of cells, could automatically be transported to the CQ1 for imaging ( Fig. 2 A).We also installed a HEPA filter to allow experiments to be executed in a sterile environment, an exhaust system for safe use of formaldehyde for cell fixation, and a biosafety level 2 booth.
GBG was introduced to establish various workflows.These included the periodic exchange of media, live-cell imaging, immunostaining on the final day, and the testing of multiple types of 8-10 culture plates across 24 hours, 7 days a week in parallel.Imaging data were automatically stored and analyzed using a plate barcode management system after imaging, creating a network layout that allowed the data to be freely displayed anywhere ( Fig. 2 B).
To integrate the experimental automation of cell culture and image analysis, issues related to detection (reading) and experimental manipulation need to be considered.In drug discovery, we require establishment of the same cellular state in each test to detect multiple intracellular functions or drug effects and to measure cellular state (fixation and detection or staining and detection with vital stain reagents).These objectives also require real-time detection.In terms of experimental manipulation, the requirements for detection can only be achieved using a highly reproducible experimental system for multiple samples.The proposed system enables cells to be cultured in a CO 2 incubator with excellent humidity and temperature control, and to transport multiple microplates to an automatic dispenser or the CQ1 image analyzer with reciprocating motion and minimal vibration.In addition, by simultaneously adding the necessary compound(s), exchanging medium, performing immunostaining, and imaging, the system expects highly reproducible results.

Workflow for medium exchange for myogenic differentiation culture and live-cell imaging
To replace media ( Fig. 3 A), automation tips and waste reservoirs, medium reservoirs, and culture plates were transported from VariStock, the cool incubator, and the SCALE48W, respectively, to Biomek i7 by the plate handler.Medium exchange was performed by Biomek i7 after warming the medium reservoir at 45°C for 8 minutes on the Pertier ALPS (Inheco, Munich, Germany) to bring the cooled medium to room temperature.
For live-cell imaging ( Fig. 3 B), culture plates were automatically transported from the SCALE48W to the CQ1 at the required time and imaging was performed according to the appropriate protocol.To save image data generated for each plate, a folder was created from the plate barcode ID by the CQ1 through instructions from GBG. Subsequent analysis was performed automatically.
To perform the immunofluorescence imaging assay ( Fig. 3 C), automation tips, reagent reservoirs, medium reservoir, and culture plates were transported from VariStock, the cool incubator, and the SCALE48W to Biomek i7 by the plate handler.The immunofluorescence imaging assay was executed by Biomek i7 and 405LS washer, and the resulting signal was measured by the CQ1.

Deep learning analysis of live-cell images of morphological changes in FSHD-MyoD N#1 cells during myogenic differentiation
Morphological changes in cells are hardly detected with bright-field images only.Accordingly, we conducted time-lapse measurement using an F-actin Labeling Probe that can be stained with live cells.A labeling probe for detection of the typical myogenic differentiation marker MyHC was not available, so instead we used F-actin as a reference gene for evaluation of FSHD differentiation.It has been confirmed that expression level is the same in undifferentiated and differentiated cells [16] .Morphological changes in FSHD-MyoD N#1 cells during myogenic differentiation were detected by time-lapse imaging.Live Cell Fluorogenic F-actin Labeling Probe (SiR-Actin Kit (CY-SC001), Cytoskeleton Inc. DENVER, CO) was added to the cells in advance, and bright-field images and fluorescent images (Ex = 652nm, Em = 674nm) were obtained by the CQ1.Use of the Live Cell Fluorogenic F-actin Labeling Probe enabled visual identification of any morphological changes ( Fig. 4 ).As shown in Fig. 4 , FSHD-MyoD N#1 muscle differentiation was most drastic within a period of 6 hours on Day 4-5.Therefore, we were able to detect rapid changes in FSHD-MyoD N#1 by taking an image every 3 h using time-lapse imaging.
Subsequently, the extent of the morphological changes in Deep Area Finder recognizes targeted objects such as cells, intracellular components and background.Meanwhile, Deep Cell Detector automatically counts cells, Deep Image Gate classifies cells into any group, and Deep Image Response quantifies the efficacy response based on the selection of negative and positive wells and dose information.In this paper, we used Deep Image Response.Deep Image Response enables recognition and digitization that is not otherwise possible by simple binarization.The Deep Image Response function learns to recognize positive and negative wells, and uses a set of generated parameters to quantify changes in cells located in the other wells based on their similarity to the positive or negative wells.There is no need to set any image recognition protocols or characteristic quantities because cell recognition is not needed.
We focused on time point 9 (48 hours), when the fluorescence value of SiR-Actin increased, and divided the analysis into two steps: confluency analysis and myogenic differentiation analysis.In confluency analysis, the deep learning algorithm was trained on images from time points 1 (0 hours) and 2 (6 hours) as negative controls and time points 9 (48 hours) and 10 (54 hours) as positive controls.In the myogenesis differentiation analysis, training was conducted using images from time points 9 and 10 as negative controls and time points 21 (90 hours) and 22 (93 hours) as positive controls.Analysis was performed using the Deep Image Response function.The positive score, which indicates how similar a well is to the positive control, increased in a time-dependent manner, allowing the degree of cell differentiation to be quantified ( Fig. 5 A,  B).Additionally, we observed intraday and day-to-day variations in the timing of morphological changes related to FSHD-MyoD N#1 myogenic differentiation ( Fig. 5 B, D).

Analysis of MyHC expression in FSHD-MyoD N#1 cells during myogenic differentiation by immunostaining
As described in our previous paper [9] , MyHC is often used as an index for myogenic differentiation and has been validated in our multipleevaluation using Hu5/KD3 cells.That paper also showed representative images of MyHC-positive myotubes using the two types of equipment, the In Cell Analyze 6000 (GE Healthcare, Marlborough, MA) and the CQ1 [ 9 ].Similarly, we also demonstrated the uniformity of myogenic differentiation conducted by the Screening Station using heatmap analyses of Hoechst-positive areas and MyHC-positive areas from the image data in each plate.To do this, we first performed formalinimmobilization.The culture plate, formalin-plate, and chip labware were automatically transported from the SCALE48W to Biomek i7.After removing the seal of the formalin-plate, 100 μL of 4% formalin was added to each well of the culture plate using Biomek i7.After 30 minutes of formalin-immobilization, the culture plate was transferred to the 405LS washer and washed three times with D-PBS (-).The culture plate was then put away in the stocker.Second, the cells were subjected to membrane permeabilization, blocking and treatment with the primary antibody mix.The culture plate was transferred to the 405LS washer and ICC-blocking buffer-2 100 μL/well was added.After incubating for 60 minutes, the culture plate was transferred to Biomek i7 and the primary antibody mix was added.After incubating for 21 hours, the culture plate was transferred to the 405LS washer and washed four times with PBST, before being transferred to Biomek i7 when the secondary antibody mix was added.After 60 minutes, the culture plate was transferred to the 405LS washer and washed three times with PBST and three times with D-PBS (-).Thereafter, the lid was discarded and the plate was sealed with a plate sealer and transported to the CQ1, where automatic imaging and analysis were performed based on the fluorescence signal.
The average MyHC-positive rate was 97.8% (coefficient of variation (CV) = 0.79%; Fig. 6 ) across the wells, indicating that it is possible to uniformly induce differentiation in stable skeletal muscles cells in a multi-well plate.We were also able to execute the experiment to an accurate schedule that was independent of the researchers' working time in the laboratory and with good results, indicating that the automation enables flexible schedules and remote experiments.The automated system also made it possible to conduct a series of experiments with a large number of plates.Further, we were able to perform experiments from home or an area of the workplace outside the laboratory.We were able to check the workflow, including the start, errors, and completion, in a timely manner through an e-mail alert sent by the GBG.We could then use the Windows Remote Desktop feature to monitor the GBG PC and perform a system recovery, or determine whether it was necessary to go into the laboratory in person.

Discussion
We built a system that integrates the automation of multiple processes needed to conduct cell-based assays featuring iPSCs, namely medium exchange, immunostaining, and live-cell imaging.After initial placement of the cells, reagents and labware, cell observation could be automatically performed throughout the experiment, the medium could be exchanged at the required time, and immobilization and immunostaining performed automatically and remotely.
As the system required 14 types of experimental equipment, and 11 different device drivers to integrate their usage, scheduling of the automated experiments was important.SAMI EX Scheduling Software (Beckman Coulter Inc.) and VWORKS Automation Control Software (Agilent Technologies Inc.) are well-known automated system scheduling software [ 18 , 19 ].However, we considered these software unsuitable for our complex integrated system because they did not allow for easy integration of multiple third-party drivers and performance of quick error recovery on all equipment.The over 400 device drivers of the GBG, on the other hand, made it possible to integrate numerous third-party devices.The GBG allowed us to seamlessly integrate different systems and schedule multiple devices with ease.Although executing complex workflows caused various errors, we were able to set up the GBG to display errors and send e-mail alerts to the researcher.After pausing a run to modify the method, or to fix or replace hardware, it was possible to complete the subsequent steps of the experimental process simply by pressing the retry button or continue button.The GBG also made it possible to easily and remotely recover from robot-induced errors such as Biomek i7 dispensing errors and the SCALE48W-to-the CQ1 transfer errors, as well as human errors such as misplaced labware.As a result, it was possible to fully automate the execution of medium exchange, immunostaining, and live-cell imaging within a single system.
Despite the final utility of the GBG, the system required some initial optimization.Initially, we encountered PC-related problems such as memory leaks and blue screen errors.Because our integration system had 11 drivers and dozens of workflows, the PC's memory filled up after a long period of use, causing the GBG to stop working due to a memory leak.The workflow stopped during the cell imaging step because the CQ1 was unable to communicate with the PC, which had produced a blue screen error.These issues were resolved by upgrading the PC and software.As a countermeasure to these problems, it was indispensable to set up the GBG to send error e-mails and enable recovery by remote control.We also experienced a delay in the transfer of image data via the intranet, which we resolved by using the automatic transfer function.Continued manual investigation into the causes of the errors we encountered ultimately allowed us to use the system unmanned and without errors.
Realization of error-free unmanned operation of such an experimental system enables long-term live-cell time-lapse imaging without affecting cell growth.Thus, we were able to detect timely and rapid changes in cells at all times of the day, including at night and on weekends.As data acquisition at short intervals at night is limited by the ability of the human researcher to work during these hours, unmanned operation can be an effective solution.Nevertheless, we observed intraday and day-today variations in the timing of morphological changes related to FSHD-MyoD N#1 myogenic differentiation.We predict that this variation was  Our time-lapse imaging data, deep learning analytics, and immunostaining results revealed that the cells were not initially evenly seeded across all wells.Wells that appeared to have a small number of immunostained cells were found to have fewer cells from the beginning of the time-lapse imaging procedure.Thus, rather than the problem being that the cells did not detach during the immunostaining process, we found that fewer cells were initially seeded into the wells.Identifying this fact early on can allow researchers to make prompt and important decisions, such as excluding wells whose data could distort downstream processes like deep learning analysis.Future studies should investigate ways to automate the cell seeding process to enable uniform and reproducible cell seeding.
On the other hand, imaging techniques such as cell painting [20] , which analyze a large number of cellular characteristics to classify cells, require hundreds of learning image data and complex analysis.This and other complex imaging technique should be tested in our system in the future.
The development of a laboratory system that can be operated remotely will open the door for remote joint research among scientists from around the world, and is expected to enhance the ability and speed with which researchers can validate data across studies.In addition, such a system will allow researchers with the necessary expertise and knowledge related to the experiment being conducted to adjust the parameters and settings themselves, ultimately leading to more optimal experimental outcomes.
Remote controlled laboratories have several other advantages.First, experimental procedures can be automated and errors can be fixed remotely.Second, protocols and data can easily be shared with the same type of equipment and manual experiments can be automated remotely by video sharing.Third, analysis can be performed remotely.Analysis methods that are difficult to interpret and explain can be discussed among research members and simultaneously optimized.
Despite the apparent utility of our automated Screening Station, however, we only tested the system in one series of experiments.Thus, additional studies are needed to validate its effectiveness.If the system proves as useful as we expect, continued developments including artificial intelligence (AI) drivers could lead to further automation of experiments ( Fig. 7 ) by allowing more processes to be performed remotely and efficiently.In the future, we expect that remote automation will be controlled using programming software such as KNIME and coding languages such as Python.However, as these can require time and effort to learn, software that can improve the ease with which automated laboratory systems are controlled is also needed.A recent example is Biosero Inc.'s GBG orchestrator software, which can manage entire systems from a bird's-eye view.Moreover, we expect that AI-driven systems will continue to enhance integrated automatic systems by increasing the number of operations that can be executed remotely.For example, we expect future AI-driven cell assay systems to be able to repeat a set of experiments and accumulate data, learn from experience, and use this feedback to improve the system.
Our automated system allows us to conduct a large number of experiments and obtain a large amount of data.Subsequent processing of the data, such as automatic analysis, is also important.To improve our system, in the future, we want to include a function that allows the robot system to automatically feed data back into itself.
In conclusion, we developed a flexible and remote laboratory automation system called the Screening Station that integrates automated cell experiments and cell image analysis technology.This is the world's first system to fully automate medium exchange, immunohistochemistry, and live-cell imaging processes.It provides researchers with more flexibility and enables experiments to be conducted remotely, as evidence by our ability to run experiments from home during COVID-19related restrictions.This system is expected to facilitate remote joint research among researchers around the world, as it allows all researchers to check the status of the experiments in a timely manner.It will also allow researchers to develop completely novel ideas, as they will no longer be limited by experimental time or manpower.Thus, the concept of complete remote research ( "robotic community biology ") advocated by Dr. Natsume is now becoming a reality [21] .

Declaration of Competing Interests
The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.

Fig. 1 .
Fig. 1.Hardware configuration of the Screening Station. A. Top, side and oblique views of the hardware configuration of the Screening Station.B. List of apparatus used in the Screening Station.

Fig. 2 .
Fig. 2. Experimental equipment and computer network workflows.A. Workflow through experimental equipment of the Screening Station.B. PC and network configurations of the Screening Station.

Fig. 3 .
Fig. 3. Three workflows of the Screening Station. A. Workflow for medium exchange B. Workflow for live-cell imaging.C. Workflow for immunofluorescence imaging Images on the right show the GBG start screen user interface.

Fig. 4 .
Fig. 4. Time-lapse images of FSHD-iPSCs over 5 days.Bright-field images and fluorescent images (Ex = 652 nm, Em = 674 nm) taken of cells treated with an F-actin Labeling Probe by the CQ1.Images were captured across 22 time points.The table shows the time points and imaging times.
FSHD-MyoD N#1 cells during myogenic differentiation was quantified using the Deep Image Response function in CellPathfinder software.Cell-Pathfinder's deep learning feature comprises four functions: Deep Area Finder, Deep Cell Detector, Deep Image Gate and Deep Image Response.

Fig. 5 .
Fig. 5. Results of Deep Image Response analysis of the time-dependent degree of cell confluency and differentiation.A. Results for the time-dependent degree of cell confluency B. Result for the time-dependent degree of differentiation C. Results for the time-dependent degree of cell confluency in a different well.D. Results for the time-dependent degree of differentiation in a different well.Positive/negative training image data were used for Deep Image Response analysis.The generated parameters are used by the deep learning model to determine the similarity between each well and the positive/negative images.There is no need to establish image recognition protocols or characteristic quantities because cell recognition is not needed.

Fig. 6 .
Fig. 6.Immunostaining of FSHD-MyoD N#1 myogenic differentiation and the myosin heavy chain-positive rate in each well.Images on the left show myosin heavy chain-stained cells in green.The pie charts on the right shows the proportion of myosin heavy chain-stained cells (green) in each well.CV = coefficient of variation

Fig. 7 .
Fig. 7. Remote workflow for fullautomation of medium exchange, live-cell imaging analysis and immunostaining analysis.