Purifying stem cell‐derived red blood cells: a high‐throughput label‐free downstream processing strategy based on microfluidic spiral inertial separation and membrane filtration

Abstract Cell‐based therapeutics, such as in vitro manufactured red blood cells (mRBCs), are different to traditional biopharmaceutical products (the final product being the cells themselves as opposed to biological molecules such as proteins) and that presents a challenge of developing new robust and economically feasible manufacturing processes, especially for sample purification. Current purification technologies have limited throughput, rely on expensive fluorescent or magnetic immunolabeling with a significant (up to 70%) cell loss and quality impairment. To address this challenge, previously characterized mechanical properties of umbilical cord blood CD34+ cells undergoing in vitro erythropoiesis were used to develop an mRBC purification strategy. The approach consists of two main stages: (a) a microfluidic separation using inertial focusing for deformability‐based sorting of enucleated cells (mRBC) from nuclei and nucleated cells resulting in 70% purity and (b) membrane filtration to enhance the purity to 99%. Herein, we propose a new route for high‐throughput (processing millions of cells/min and mls of medium/min) purification process for mRBC, leading to high mRBC purity while maintaining cell integrity and no alterations in their global gene expression profile. Further adaption of this separation approach offers a potential route for processing of a wide range of cellular products.


| INTRODUCTION
Stem cell-derived red blood cells could constitute an attractive pathogen-free and sustainable alternative for donated blood for rare blood groups and patients requiring regular transfusions (Zeuner et al., 2012). In many cases, such as sickle cell anemia, myelodysplasias and leukemia, multiple blood transfusion is regarded as the only available symptomatic treatment, and that can lead to immunization against the allogeneic red blood cells and transfusion impasses (Douay & Andreu, 2007). The efficient production of manufactured red blood cells (mRBCs) is consequently an ambitious goal for blood services around the world; however, production of a single therapeutic dose (~2 × 10 12 cells) still remains a significant challenge (Cabrita et al., 2003;Peyrard et al., 2011). Clinical application of mRBC is currently hampered by a lack of technological solutions that would allow production of mRBC at a satisfactory scale and purity in compliance with good manufacture practice (GMP) regulations within Han, 2013;Migliaccio, Whitsett, Papayannopoulou, & Sadelain, 2012;Rousseau, Giarratana, & Douay, 2014;Shah, Huang, & Cheng, 2014).
Despite considerable progress in improving the expansion rate and yield of mRBC, enucleation rates remain limited (Table S1). The endproduct of existing differentiation protocols is consequently a heterogeneous mixture of enucleated mRBC, nucleated cells that remain at earlier developmental stages and free-floating nuclei expelled during the enucleation process ( Figure 1b). In 2008, Fujimi et al. (2008) reported a differentiation strategy with CB CD34+ providing an almost complete enucleation (99.4%) (Fujimi et al., 2008). However, this was achieved by coculture with macrophages, making the protocol challenging to scale-up (Goers, Freemont, & Polizzi, 2014). With lower enucleation rates, the presence of residual nucleated cells and expelled nuclei constitute a potential danger if intended for transfusion into patients (Bouhassira, 2008;Guzniczak et al., 2017). Undifferentiated nucleated cells can give rise to teratomas (benign tumors of differentiating cells) and teratocarcinomas (malignant metastatic tumors composed of highly proliferative cells; McGowan, Campbell, & Mountford, 2018), thus they have to be removed from the sample and require adequate purification approaches. In addition, presence of freefloating nuclei in large quantities may prove particularly problematic in large-scale culture systems by fouling surfaces and entangling the desired cell product in DNA  Traditionally, cell purification is performed using fluorescent or magnetic activated cell sorting (FACS or MACS) (Schriebl, Lim, Choo, Tscheliessnig, & Jungbauer, 2010). FACS and MACS both generate highly defined, purified (>95%) cell populations with a low number of unwanted cells in the final product; however, the requirement for cellspecific ligands hinders adaptation of these methods to industrial-scale processing due to the high cost of antibodies. In addition, immunolabeling is a laborious multi-step process consisting of numerous centrifugation, washing, and incubation steps often resulting in a significant (reported up to 70%) cell loss (Schriebl et al., 2010) and postisolation cell quality impairment (Lee et al., 2018). Currently, only a limited number of fluorophore-conjugated antibody reagents are suitable for clinical processing (McIntyre, Flyg, & Fong, 2010) and the adverse effects of introducing these probes into patients are unknown, but it is generally recognized that they could potentially trigger immune and toxic responses (Willoughby et al., 2016). Various alternatives to FACS and MACS for cellular therapies, such as mRBC production, have been proposed and were recently reviewed (Masri, Hoeve, Sousa, & Willoughby, 2017 F I G U R E 1 (a) Cord blood (CB) CD34+ cells undergo in vitro differentiation into manufactured red blood cells (mRBCs) over the course of 21 days. (b) As shown in the exemplary cytospin image, the end-product of the differentiation protocol is a heterogeneous population containing enucleated mRBC, partially differentiated or undifferentiated nucleated cells as well as free-floating nuclei. The scale bar corresponds to 20 µm. (c) The proposed label-free sorting strategy for the end-product consists of two steps: first, the sample is processed in a spiral microchannel with a rectangular cross-section (170 × 30 um 2 ), six loops, one inlet, and four outlets (A, B, C, and D). Inertial focusing within spiral microchannels occurs due to balance of shear gradient lift force (F L ), wall-induced lift force (F W ) as well as Dean drag (F DD ). Particles of different sizes interact with a different section of the characteristic cross-sectional velocity profile (Dean vortices). Deformable particles experience an additional deformability-induced lift force (F D ). Cells align in the spiral channel at distinct lateral equilibrium positions, that facilitates their capture in one of the four outlets. The majority of the desired enucleated cells is captured in outlet A with some contaminant nucleated cells, which are further removed by membrane filtration [Color figure can be viewed at wileyonlinelibrary.com] GUZNICZAK ET AL.
To address the challenge of mRBC purification, we propose a labelfree approach to separate cells at high throughput based on their morphological (size) and mechanical (deformability) properties. As presented in Figure 1c, the process consists of two main steps: (a) a microfluidic separation using inertial focusing in spiral microchannel and (b) membrane filtration. Due to its simplicity in operation, low manufacturing cost and proven scalability by parallelization (allowing processing millions of cells per minute) inertial focusing in spiral channels has been recognized as an attractive approach for high-throughput cell sorting (Gou, Jia, Wang, & Sun, 2018) for a wide range of applications (for a comprehensive review, see Gou et al., 2018 (Bhagat, Kuntaegowdanahalli, & Papautsky, 2008;Gou et al., 2018). As shown in Figure 1c, smaller nuclei are positioned closer to the inner wall while larger cells are observed closer to the channel centreline. As we previously reported , there is a distinct hydro-

| Real-time fluorescence and deformability cytometry
Cells' morphological and mechanical properties were assessed zellmechanik.com).

| Cell morphology-cytospin
To visualize cells' morphology and structure, cells were transferred onto microscope slides using a cytocentrifuge then fixed and stained using Giemsa-Wright staining (Rapid Romanowsky Stain Pack, cat. SW167/500; TCS Bioscience). Cells were harvested by centrifugation at 300g for 5 min and resuspended at 2 × 10 6 cells/ml in PBS−/− (Dulbecco's PBS buffer without calcium and magnesium; Gibco). One

| Cell processing
The current differentiation protocol involves the use of human serum as a supplement to cell culture media. It provides high concentrations of growth factors, macromolecules, carrier proteins for lipids, trace elements, attachment and spreading factors, nutrients, and hormones (Heger et al., 2018). We however found, similarly to others (Henderson et al., 2010), that microfluidic channels can clog with serum (Henderson et al., 2010) and recommend using a serum-free buffer for processing. Moreover, the presence of phenol red (pH indicator in basal medium) impairs reads from both flow cytometry and automated cell count, thus cells processed in the basal medium could not be directly sampled for the quality control tests. In this study, we Scientific) as a processing buffer. Pluronic F-68 was added to surrogate the serum protective mechanism from mechanical damages (e.g. due to shear stress generated within the spiral microchannel) (Heger et al., 2018;Tharmalingam, Ghebeh, Wuerz, & Butler, 2008;Guzniczak et al., 2018).
at circa 3-4 × 10 6 cells/ml were injected in a spiral microfluidic channel with a mid-pressure syringe pump (neMESYS 1000N; Cetoni, Germany) in 10-ml batches trough 1/16" PTFE tubing of 0.5 mm internal diameter (Thames Restek, UK). Cell concentration is a critical factor influencing focusing within the spiral microchannel. If the concentration is too high, the steric crowding effect occurs, meaning that there is physically not enough space for particles to focus in a tight single stream. To identify if the crowding effect will occur, the parameter α (number of particle diameters per channel length) can be calculated.
where W (resp. H) the width (resp. height) of the channel cross section, V F is the volume fraction of particles in the solution, a the particle diameter. For α > 1, focusing to a single stream can be challenged by steric interactions between particles (Di Carlo, 2009).
Assuming that all the particles in the input sample were of the size of the largest nucleated cells (a~5 µm), at 3-4 × 10 6 cells/ml, α varies between 0.153 and 0.255 giving an upperbound of α < 1.
As shown in Table 2, cells were examined at flow rates ranged from 200 to 1,000 µl/min (corresponding to channel Reynolds number [Re] between 33 and 168, Dean number [De] ranging between 5 and 26).
The channel Re is a dimensionless parameter, which describes the unperturbed channel flow.
Re UD Inertial forces Viscous forces , where ρ is the medium density, Uis the medium velocity, μ is the dynamic viscosity, and D h is the hydraulic diameter, defined as where H is the channel height and W is the channel width.
De is used to quantify the secondary flow within spiral microchannel, and it is defined as where R is the radius of the curvature.
The focusing behavior of cells was assessed in terms of lateral equilibrium position, measured as a distance from the cell center to the outer wall of the spiral channel in region of interest as shown in Figure S10B. Images of cells inside the spiral channel were recorded at ×10 magnification with a 4.9 mm free working distance (421251-9911-000 LD A-Plan 10× Ph1; Zeiss) using high-speed camera (CCD ProgRes ® ; Jenoptik, Germany) mounted on a microscope (Zeiss Axio Observer 3; Zeiss). Images were recorded at 130 frames per second and analyzed using a bespoke MatLab script. To enhance the purification efficiency, cells collected in outlet A of the spiral channel were passed through a 3 µm polycarbonate Isopore™ filter membrane (Merc, UK). Cell suspension was loaded into a 5 ml plastic syringe and pumped through the filter membrane at 2 ml/min using a syringe pump (neMESYS; Cetoni). Cell suspensions were passed through the filter membranes fitted onto syringe adapter and the filtrate was collected in 10-ml plastic tube. The sorting performance was assessed using the following three parameters: Cell separation efficiency was quantified by flow cytometry (BD LSR II; BD) to compare the fraction of each cell population (characterized by unique fluorescent properties) in samples collected at each outlet and after filtration. In addition, cell yield was assessed by counting the number of cells at each outlet and after filtration using MoxiZ automated cell counter (Orflo). Further data analysis was performed using GraphPad Prism 6 and FlowJo V10 CL.
In terms of actual recovery, determined by total number of cells collected versus total number of cells injected into the system, there is a slight variation (~5%) since the sample is subjected to dilution and sedimentation. Before the procedure, device is primed with running buffer without cells, the dead volume was assessed as 1.5 m, which should be discarded, since it takes around 1 min for the system to stabilize, and we collect cell suspension after this time. Since the processing time for one batch is 10 min, the first portion of collected suspension is slightly more concentrated than the very last one due to sedimentation.

| Viability-trypan blue exclusion assay
Control cells were not passed through the device but they were  Table S3 (for the convenience of the reader, also, data on donor III are included in STable 3). The study was further conducted here for cells from three different donors (see Figure 3 for exemplary data of one replica from each donor).
As presented in Figure 3 and Table S3,

| Process optimization
To identify optimal conditions to take advantage of deformability for mRBC purification, the performance of a spiral microchannel with 170 × 30 µm 2 cross-section has been tested with FACS presorted pure populations of enucleated and nucleated cells and nuclei from donor III. As mentioned previously, nucleated and enucleated cells

| Process performance
Differences in hydrodynamic behavior of enucleated and nucleated cells, as well as nuclei observed at 1 ml/min flow rate, were translated and incorporated into a label-free purification process for mRBC, derived from three donors (indicated as donor I, II, and III).
The heterogeneous end-product after the differentiation protocol was injected into the spiral microchannel at 1 ml/min at a concentration of around 3 × 10 6 cells/ml. Figure 5a shows an averaged fraction of each subset in the input sample derived from donors I, II, and III. Enucleated cells constituted around 10-35% of the starting sample. As predicted, due to their deformable nature, the majority of enucleated cells were hydrodynamically directed to outlet A (closest to the outer wall).
The high-quality end-product derived from donors I and III constituted a good quality starting material for the label-free purification resulting in highest separation efficiency (>90%) and purity Populations of mRBC derived from donor I and II, at concentrations of~3 × 10 6 cells/ml were processed in the spiral microchannel with 170 × 30 µm 2 cross-section at 1 ml/min flow rate, and eluents from all outlets were collected into one vial (labeled spiral) and their quality was compared against unprocessed cells (control).
Cell integrity of the control and processed cells was investigated via trypan blue exclusion assay. Live cells are impenetrable for trypan blue, while damaged cells with impaired cell membrane integrity uptake trypan blue and they appear blue. As shown in Figure 6a, the high viability of >85% was comparable at the inlet (control) and after processing (spiral).
Cells actively respond to mechanical perturbations through the modification of gene expression (Miroshnikova, Nava, & Wickstrom, 2017). To investigate if exposing undifferentiated nucleated cells to mechanical stress engages oncogenes signaling pathways, the global gene expression patterns were investigated using poly-A selection method. Control and processed (spiral) samples were collected earlier during the differentiation process (Day 14) than samples for trypan blue assay, to ensure that nucleated cells were still transcriptionally active.
After sequencing, a standard pipeline was run that seeks to describe the variance and correlative behavior across the data before classifying those genes that are differentially expressed in each core comparison. Sample by sample correlation analysis was performed to attempt to ascertain how strongly or weakly each sample correlates across the range of gene expression values (Spearman correlation clustering, hierarchically clustered). A strong tendency for samples to cluster by sample group (donors I and II), overriding the effects of processing in the spiral microchannel (Figure 6b), was observed.
We then sought to describe the individual changes in gene expression using CuffDiff (Trapnell et al., 2013).  I II III I II III  I II III I II III   I II III I II III  I II III I II III   0  In summary, it has been confirmed that mRBC, after processing within the spiral microchannel at a sufficiently high flow rate to take advantage of the effect of F D for cells focusing, retains a high degree of viability and that there is no distinct or consistent gene expression alteration.

| DISCUSSION
In this study, we successfully developed a passive, high-throughput, label-free purification strategy for CB CD34+ derived red blood cells.
Using advances in the field of deformability cytometry, heterogeneous end-products of CB CD34+ in vitro erythropoiesis were characterized and label-free markers were identified for the target enucleated cells as well as contaminant nucleated cells and expelled nuclei. These label-free markers were used as a two-step purification F I G U R E 6 (a) Viability of mRBC derived from donor I and II, before (control) and after (spiral) processing in a spiral microchannel with 170 × 30 µm cross-section at 1 ml/min flow rate, measured by trypan blue exclusion assays. Bars represent mean fraction of live (plain green) or dead (checkered pattern) cells found in the samples, measured at three independent experimental occasions. (b) Heatmap of sample correlation between each pairwise combination of samples, for four replicas (number of replicas is indicated in brackets). The correlations were calculated using Spearman correlation based on all gene expression values. The level of correlation (Spearman correlation coefficient) is represented by color intensity, with strong positive correlation in red, no correlation in light pink and strong anti-correlation in blue. The plot has been hierarchically clustered on both axis using-there is no distinct gene expression alteration. (c) Bar chart showing the number of significantly (p < .05) upregulated (up) and downregulated (down) genes, between donor I (and donor II, respectively) cells processed is the spiral microchannel compared to control sample as well as donor II control and processed cells compared to control and processed cells derived from donor I. (d) Violin plots showing two most significantly (selected by the lowest p-value) upregulated and downregulated genes after processing (spiral) in comparison to control cells, for both donor I and II. Each dot is one sample, with sample groups given on the x-axis and gene expression on the y-axis. The mean and standard error for each sample group is given as a red dot and line. The spread of the samples within a sample group gives an idea of sample heterogeneity at a given gene. mRBC, manufactured red blood cell [Color figure can be viewed at wileyonlinelibrary.com] strategy: (a) using inertial focusing in a spiral microchannel where most of the target enucleated cells are covered (>90%) at relatively high purity (>70%), without compromising cell quality and (b) a membrane filtration step resulting in the removal of~99% of remaining impurities (mainly nucleated cells since >98% of nuclei were removed by the spiral microchannel). The inertial focusing strategy is based upon deformability sorting. Given the size of the overlap of the enucleated and nucleated cells, the only explanation for the shift toward the outer wall of the enucleated cells is the deformability difference; this phenomena has been previously reported and characterized, and although being a novel approach, further investigation of the underlying theoretical physics, supported by experimental data, is required . The membrane filtration step requires further optimization and development, since in this study, dead-end filtration, which is prone to membrane fouling, led to the separation efficiency of 30-50% of enucleated cells. Shah et al.
(2016) reported a positive evaluation of CB CD34+ derived mRBC as transfusion product (Shah et al., 2016), using their novel animal model to assess the potential of mRBCs to deliver oxygen to muscle tissues. To deplete undifferentiated nucleated cells before transfusion, they used a nonwoven fabric filter (Tao, Xia, Cao, & Gao, 2011).
They carried out an extensive study on the impact of filtration on the quality of mRBC and they found that cells, despite a significant cell retention on the membrane (filtration removed~75% of cells), mRBC passed through the filter remained intact and there were no difference in levels of hemoglobin expression before and after filtration.
Gene expression changes were not studied by them, nor in our work.
In the demonstrated approach, >3 × 10 6 cells/min are processed by a single device when operating at the optimal flow rate. The downstream processing method proposed in this study has the capacity for further scale-up by two means: increasing cell sample concentration and system parallelization. However, the current cell sample concentration seems reasonable for processing cells that are routinely cultured within a similar concentration range in large volumes. At present, mRBC culture is routinely carried in static culture conditions, facilitating maximal cell concentration at around 5 × 10 6 cells/ml (Rousseau et al., 2014). Volumetric throughput in the device presented here is 1 ml/min in a single layer system, which again is compatible with the state-of-the art bioreactor sizes (Rousseau et al., 2014), though larger volumes are likely to be required for commercial production. Throughput could be improved by parallelization and/or stacking, for example, like recent work by Warkiani that reached 240 ml/min or 350 L/day, with the authors reporting further parallelization was possible to triple the throughput (Warkiani, Tay, Guan, & Han, 2015). Stacking microfluidic devices (stack of 20 devices reported; Miller, Jimenez, & Bridle, 2016) is a common practice resulting in a rapid and efficient throughput improvement. Further clarification is needed on exact requirements for industrial-scale production, though, given the lack of impact on the cells of this approach, processing time is more likely to influence the economics of the process rather than cell quality. Since the device operates at elevated follow rate to reveal the differential equilibrium position determined by deformability, one of the pragmatic challenges would be to identify a suitable pumping system, withstanding high pressures (up to 30 bars) and operating in a continuous mode. Currently, cell suspensions are introduced into the device in 10 ml batches using a mid-pressure syringe pump.
Membrane filtration alone is less effective in processing the mRBC than the combined process consisting of processing in spiral microchannel followed by filtration. Particle separation by means of filtration is a widely applied technique within field of bioprocessing (Masri et al., 2017). Membrane filtration uses an average pore size where particles larger than the pore size cannot pass through. Traditional membrane filtration suffers from several drawbacks, with the main one being clogging. Clogged membrane filters degrade in performance over time and the "filter cake" may pose contamination hazards (Seo, Lean, & Kole, 2007). Membrane filtration is especially problematic for mRBC purification due to presence of large quantities of free-floating nuclei. DNA is known for being "sticky" molecule and causing fouling of surfaces .
Inertial focusing in spiral microchannels has been proposed as "membrane-free" filtration, capable of continuous and highthroughput separation based on size and deformability (Bhagat et al., 2008;Guzniczak et al., 2020). By processing the sample in spiral microchannel, a majority (>98%) of the nuclei are depleted, prolonging the life-span of the filter membrane. The dead-end membrane filtration used in this study requires further optimization and development to improve cell recovery.
All current protocols for the manufacture of RBC from stem cells face the same technological challenge of low enucleation rate. The most efficient solution is coculture with macrophages, which eliminate the expelled nuclei by the means of phagocytosis. This is an organic solution but it comes with its own technological costs, such as finding ways to retrieve the macrophages from the culture and the complexity of a coculture system with feeder layer. In the most optimistic scenario, even if the enucleation rate is improved to reach the desirable 100% and nucleated cells are not present in the endproduct of the in vitro erythropoiesis, in the absence of macrophages, the expelled nuclei will still remain within the sample. Having a robust label-free procedure for mRBC purification at high-throughput with no impact on cell quality will consequently be of significant importance for bringing mRBC a step closer to clinical use.
CB CD34+ cells are a limited and variable source of mRBC and as verified in this study, starting cell material derived from different donors give a final product characterized by different phenotypes and mechanotypes, thus implementation of universal downstream protocols is currently challenging. The field of stem cell-derived therapeutic products is maturing and with introduction of iPSC (Lapillonne et al., 2010) and immortalized erythroid cell lines (Trakarnsanga et al., 2017), it should be possible to produce large quantities of standardized mRBC and integrate technology proposed here into the formulation step of the cellular product derivation process.
To conclude, this study presents a much-needed label-free highthroughput (millions of cells/min, ml of medium/min) scalable and continuous cell sorting approach for novel stem-cell-derived 2042 | therapeutic products. In addition, the capability to sort multiple cell types simultaneously based on their size and deformability, at highthroughputs, within one system and without the compromising effect of fluorescent labels could be highly relevant for isolation of various cells of interest from heterogeneous samples.